AuroraMiddleware/glibc
1
0
Fork 0
mirror of https://sourceware.org/git/glibc.git synced 2024-12-04 10:50:07 +00:00
Code Issues Packages Projects Releases Wiki Activity
c13e078308
glibc/sysdeps/powerpc/powerpc64/power7/memcmp.S

1060 lines
26 KiB
ArmAsm
Raw Normal View History

power7 string compare optimizations
2010-06-15 00:13:24 +00:00
/* Optimized memcmp implementation for POWER7/PowerPC64.
Update copyright dates with scripts/update-copyrights.
2015-01-02 16:28:19 +00:00
Copyright (C) 2010-2015 Free Software Foundation, Inc.
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
This file is part of the GNU C Library.
The GNU C Library is free software; you can redistribute it and/or
modify it under the terms of the GNU Lesser General Public
License as published by the Free Software Foundation; either
version 2.1 of the License, or (at your option) any later version.
The GNU C Library is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
Lesser General Public License for more details.
You should have received a copy of the GNU Lesser General Public
Replace FSF snail mail address with URLs.
2012-02-09 23:18:22 +00:00
License along with the GNU C Library; if not, see
<http://www.gnu.org/licenses/>. */
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
#include <sysdep.h>
/* int [r3] memcmp (const char *s1 [r3],
const char *s2 [r4],
size_t size [r5]) */
.machine power7
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
EALIGN (memcmp, 4, 0)
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
CALL_MCOUNT 3
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
#define rRTN r3
#define rSTR1 r3 /* first string arg */
#define rSTR2 r4 /* second string arg */
#define rN r5 /* max string length */
#define rWORD1 r6 /* current word in s1 */
#define rWORD2 r7 /* current word in s2 */
#define rWORD3 r8 /* next word in s1 */
#define rWORD4 r9 /* next word in s2 */
#define rWORD5 r10 /* next word in s1 */
#define rWORD6 r11 /* next word in s2 */
#define rOFF8 r20 /* 8 bytes offset. */
#define rOFF16 r21 /* 16 bytes offset. */
#define rOFF24 r22 /* 24 bytes offset. */
#define rOFF32 r23 /* 24 bytes offset. */
#define rWORD6_SHIFT r24 /* Left rotation temp for rWORD8. */
#define rWORD4_SHIFT r25 /* Left rotation temp for rWORD6. */
#define rWORD2_SHIFT r26 /* Left rotation temp for rWORD4. */
#define rWORD8_SHIFT r27 /* Left rotation temp for rWORD2. */
#define rSHR r28 /* Unaligned shift right count. */
#define rSHL r29 /* Unaligned shift left count. */
#define rWORD7 r30 /* next word in s1 */
#define rWORD8 r31 /* next word in s2 */
#define rWORD8SAVE (-8)
#define rWORD7SAVE (-16)
#define rOFF8SAVE (-24)
#define rOFF16SAVE (-32)
#define rOFF24SAVE (-40)
#define rOFF32SAVE (-48)
#define rSHRSAVE (-56)
#define rSHLSAVE (-64)
#define rWORD8SHIFTSAVE (-72)
#define rWORD2SHIFTSAVE (-80)
#define rWORD4SHIFTSAVE (-88)
#define rWORD6SHIFTSAVE (-96)
#ifdef __LITTLE_ENDIAN__
# define LD ldbrx
#else
# define LD ldx
#endif
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
xor r0, rSTR2, rSTR1
cmpldi cr6, rN, 0
cmpldi cr1, rN, 12
clrldi. r0, r0, 61
clrldi r12, rSTR1, 61
cmpldi cr5, r12, 0
beq- cr6, L(zeroLength)
dcbt 0, rSTR1
dcbt 0, rSTR2
Fix spelling errors in sysdeps/powerpc files.
2013-01-07 17:20:53 +00:00
/* If less than 8 bytes or not aligned, use the unaligned
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
byte loop. */
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
blt cr1, L(bytealigned)
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
std rWORD8, rWORD8SAVE(r1)
cfi_offset(rWORD8, rWORD8SAVE)
std rWORD7, rWORD7SAVE(r1)
cfi_offset(rWORD7, rWORD7SAVE)
std rOFF8, rOFF8SAVE(r1)
cfi_offset(rWORD7, rOFF8SAVE)
std rOFF16, rOFF16SAVE(r1)
cfi_offset(rWORD7, rOFF16SAVE)
std rOFF24, rOFF24SAVE(r1)
cfi_offset(rWORD7, rOFF24SAVE)
std rOFF32, rOFF32SAVE(r1)
cfi_offset(rWORD7, rOFF32SAVE)
li rOFF8,8
li rOFF16,16
li rOFF24,24
li rOFF32,32
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
bne L(unaligned)
/* At this point we know both strings have the same alignment and the
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
compare length is at least 8 bytes. r12 contains the low order
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
3 bits of rSTR1 and cr5 contains the result of the logical compare
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
of r12 to 0. If r12 == 0 then we are already double word
aligned and can perform the DW aligned loop.
Fix whitespaces.
2010-06-15 00:15:33 +00:00
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
Otherwise we know the two strings have the same alignment (but not
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
yet DW). So we force the string addresses to the next lower DW
boundary and special case this first DW using shift left to
Fix spelling errors in sysdeps/powerpc files.
2013-01-07 17:20:53 +00:00
eliminate bits preceding the first byte. Since we want to join the
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
normal (DW aligned) compare loop, starting at the second double word,
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
we need to adjust the length (rN) and special case the loop
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
versioning for the first DW. This ensures that the loop count is
correct and the first DW (shifted) is in the expected register pair. */
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
.align 4
L(samealignment):
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
clrrdi rSTR1, rSTR1, 3
clrrdi rSTR2, rSTR2, 3
beq cr5, L(DWaligned)
add rN, rN, r12
sldi rWORD6, r12, 3
srdi r0, rN, 5 /* Divide by 32 */
andi. r12, rN, 24 /* Get the DW remainder */
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD1, 0, rSTR1
LD rWORD2, 0, rSTR2
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpldi cr1, r12, 16
cmpldi cr7, rN, 32
clrldi rN, rN, 61
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
beq L(dPs4)
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
mtctr r0
bgt cr1, L(dPs3)
beq cr1, L(dPs2)
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
/* Remainder is 8 */
.align 3
L(dsP1):
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
sld rWORD5, rWORD1, rWORD6
sld rWORD6, rWORD2, rWORD6
cmpld cr5, rWORD5, rWORD6
blt cr7, L(dP1x)
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
/* Do something useful in this cycle since we have to branch anyway. */
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD1, rOFF8, rSTR1
LD rWORD2, rOFF8, rSTR2
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpld cr7, rWORD1, rWORD2
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
b L(dP1e)
/* Remainder is 16 */
.align 4
L(dPs2):
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
sld rWORD5, rWORD1, rWORD6
sld rWORD6, rWORD2, rWORD6
cmpld cr6, rWORD5, rWORD6
blt cr7, L(dP2x)
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
/* Do something useful in this cycle since we have to branch anyway. */
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD7, rOFF8, rSTR1
LD rWORD8, rOFF8, rSTR2
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpld cr5, rWORD7, rWORD8
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
b L(dP2e)
/* Remainder is 24 */
.align 4
L(dPs3):
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
sld rWORD3, rWORD1, rWORD6
sld rWORD4, rWORD2, rWORD6
cmpld cr1, rWORD3, rWORD4
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
b L(dP3e)
/* Count is a multiple of 32, remainder is 0 */
.align 4
L(dPs4):
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
mtctr r0
sld rWORD1, rWORD1, rWORD6
sld rWORD2, rWORD2, rWORD6
cmpld cr7, rWORD1, rWORD2
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
b L(dP4e)
/* At this point we know both strings are double word aligned and the
compare length is at least 8 bytes. */
.align 4
L(DWaligned):
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
andi. r12, rN, 24 /* Get the DW remainder */
srdi r0, rN, 5 /* Divide by 32 */
cmpldi cr1, r12, 16
cmpldi cr7, rN, 32
clrldi rN, rN, 61
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
beq L(dP4)
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
bgt cr1, L(dP3)
beq cr1, L(dP2)
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
/* Remainder is 8 */
.align 4
L(dP1):
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
mtctr r0
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
/* Normally we'd use rWORD7/rWORD8 here, but since we might exit early
Fix spelling errors in sysdeps/powerpc files.
2013-01-07 17:20:53 +00:00
(8-15 byte compare), we want to use only volatile registers. This
means we can avoid restoring non-volatile registers since we did not
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
change any on the early exit path. The key here is the non-early
Fix whitespaces.
2010-06-15 00:15:33 +00:00
exit path only cares about the condition code (cr5), not about which
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
register pair was used. */
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD5, 0, rSTR1
LD rWORD6, 0, rSTR2
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpld cr5, rWORD5, rWORD6
blt cr7, L(dP1x)
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD1, rOFF8, rSTR1
LD rWORD2, rOFF8, rSTR2
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpld cr7, rWORD1, rWORD2
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
L(dP1e):
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD3, rOFF16, rSTR1
LD rWORD4, rOFF16, rSTR2
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpld cr1, rWORD3, rWORD4
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD5, rOFF24, rSTR1
LD rWORD6, rOFF24, rSTR2
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpld cr6, rWORD5, rWORD6
bne cr5, L(dLcr5x)
bne cr7, L(dLcr7x)
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD7, rOFF32, rSTR1
LD rWORD8, rOFF32, rSTR2
addi rSTR1, rSTR1, 32
addi rSTR2, rSTR2, 32
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
bne cr1, L(dLcr1)
cmpld cr5, rWORD7, rWORD8
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
bdnz L(dLoop)
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
bne cr6, L(dLcr6)
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
ld rWORD8, rWORD8SAVE(r1)
ld rWORD7, rWORD7SAVE(r1)
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
.align 3
L(dP1x):
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
sldi. r12, rN, 3
bne cr5, L(dLcr5x)
subfic rN, r12, 64 /* Shift count is 64 - (rN * 8). */
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
bne L(d00)
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
ld rOFF8, rOFF8SAVE(r1)
ld rOFF16, rOFF16SAVE(r1)
ld rOFF24, rOFF24SAVE(r1)
ld rOFF32, rOFF32SAVE(r1)
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
li rRTN, 0
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
blr
/* Remainder is 16 */
.align 4
L(dP2):
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
mtctr r0
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD5, 0, rSTR1
LD rWORD6, 0, rSTR2
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpld cr6, rWORD5, rWORD6
blt cr7, L(dP2x)
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD7, rOFF8, rSTR1
LD rWORD8, rOFF8, rSTR2
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpld cr5, rWORD7, rWORD8
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
L(dP2e):
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD1, rOFF16, rSTR1
LD rWORD2, rOFF16, rSTR2
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpld cr7, rWORD1, rWORD2
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD3, rOFF24, rSTR1
LD rWORD4, rOFF24, rSTR2
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpld cr1, rWORD3, rWORD4
addi rSTR1, rSTR1, 8
addi rSTR2, rSTR2, 8
bne cr6, L(dLcr6)
bne cr5, L(dLcr5)
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
b L(dLoop2)
.align 4
L(dP2x):
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD3, rOFF8, rSTR1
LD rWORD4, rOFF8, rSTR2
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpld cr1, rWORD3, rWORD4
sldi. r12, rN, 3
bne cr6, L(dLcr6x)
addi rSTR1, rSTR1, 8
addi rSTR2, rSTR2, 8
bne cr1, L(dLcr1x)
subfic rN, r12, 64 /* Shift count is 64 - (rN * 8). */
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
bne L(d00)
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
ld rOFF8, rOFF8SAVE(r1)
ld rOFF16, rOFF16SAVE(r1)
ld rOFF24, rOFF24SAVE(r1)
ld rOFF32, rOFF32SAVE(r1)
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
li rRTN, 0
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
blr
/* Remainder is 24 */
.align 4
L(dP3):
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
mtctr r0
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD3, 0, rSTR1
LD rWORD4, 0, rSTR2
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpld cr1, rWORD3, rWORD4
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
L(dP3e):
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD5, rOFF8, rSTR1
LD rWORD6, rOFF8, rSTR2
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpld cr6, rWORD5, rWORD6
blt cr7, L(dP3x)
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD7, rOFF16, rSTR1
LD rWORD8, rOFF16, rSTR2
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpld cr5, rWORD7, rWORD8
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD1, rOFF24, rSTR1
LD rWORD2, rOFF24, rSTR2
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpld cr7, rWORD1, rWORD2
addi rSTR1, rSTR1, 16
addi rSTR2, rSTR2, 16
bne cr1, L(dLcr1)
bne cr6, L(dLcr6)
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
b L(dLoop1)
/* Again we are on a early exit path (24-31 byte compare), we want to
Fix spelling errors in sysdeps/powerpc files.
2013-01-07 17:20:53 +00:00
only use volatile registers and avoid restoring non-volatile
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
registers. */
.align 4
L(dP3x):
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD1, rOFF16, rSTR1
LD rWORD2, rOFF16, rSTR2
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpld cr7, rWORD1, rWORD2
sldi. r12, rN, 3
bne cr1, L(dLcr1x)
addi rSTR1, rSTR1, 16
addi rSTR2, rSTR2, 16
bne cr6, L(dLcr6x)
subfic rN, r12, 64 /* Shift count is 64 - (rN * 8). */
bne cr7, L(dLcr7x)
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
bne L(d00)
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
ld rOFF8, rOFF8SAVE(r1)
ld rOFF16, rOFF16SAVE(r1)
ld rOFF24, rOFF24SAVE(r1)
ld rOFF32, rOFF32SAVE(r1)
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
li rRTN, 0
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
blr
/* Count is a multiple of 32, remainder is 0 */
.align 4
L(dP4):
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
mtctr r0
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD1, 0, rSTR1
LD rWORD2, 0, rSTR2
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpld cr7, rWORD1, rWORD2
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
L(dP4e):
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD3, rOFF8, rSTR1
LD rWORD4, rOFF8, rSTR2
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpld cr1, rWORD3, rWORD4
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD5, rOFF16, rSTR1
LD rWORD6, rOFF16, rSTR2
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpld cr6, rWORD5, rWORD6
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD7, rOFF24, rSTR1
LD rWORD8, rOFF24, rSTR2
addi rSTR1, rSTR1, 24
addi rSTR2, rSTR2, 24
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpld cr5, rWORD7, rWORD8
bne cr7, L(dLcr7)
bne cr1, L(dLcr1)
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
bdz- L(d24) /* Adjust CTR as we start with +4 */
/* This is the primary loop */
.align 4
L(dLoop):
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD1, rOFF8, rSTR1
LD rWORD2, rOFF8, rSTR2
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpld cr1, rWORD3, rWORD4
bne cr6, L(dLcr6)
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
L(dLoop1):
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD3, rOFF16, rSTR1
LD rWORD4, rOFF16, rSTR2
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpld cr6, rWORD5, rWORD6
bne cr5, L(dLcr5)
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
L(dLoop2):
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD5, rOFF24, rSTR1
LD rWORD6, rOFF24, rSTR2
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpld cr5, rWORD7, rWORD8
bne cr7, L(dLcr7)
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
L(dLoop3):
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD7, rOFF32, rSTR1
LD rWORD8, rOFF32, rSTR2
addi rSTR1, rSTR1, 32
addi rSTR2, rSTR2, 32
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
bne cr1, L(dLcr1)
cmpld cr7, rWORD1, rWORD2
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
bdnz L(dLoop)
L(dL4):
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpld cr1, rWORD3, rWORD4
bne cr6, L(dLcr6)
cmpld cr6, rWORD5, rWORD6
bne cr5, L(dLcr5)
cmpld cr5, rWORD7, rWORD8
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
L(d44):
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
bne cr7, L(dLcr7)
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
L(d34):
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
bne cr1, L(dLcr1)
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
L(d24):
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
bne cr6, L(dLcr6)
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
L(d14):
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
sldi. r12, rN, 3
bne cr5, L(dLcr5)
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
L(d04):
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
ld rWORD8, rWORD8SAVE(r1)
ld rWORD7, rWORD7SAVE(r1)
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
subfic rN, r12, 64 /* Shift count is 64 - (rN * 8). */
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
beq L(duzeroLength)
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
/* At this point we have a remainder of 1 to 7 bytes to compare. Since
we are aligned it is safe to load the whole double word, and use
Fix spelling errors in sysdeps/powerpc files.
2013-01-07 17:20:53 +00:00
shift right double to eliminate bits beyond the compare length. */
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
L(d00):
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD1, rOFF8, rSTR1
LD rWORD2, rOFF8, rSTR2
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
srd rWORD1, rWORD1, rN
srd rWORD2, rWORD2, rN
cmpld cr7, rWORD1, rWORD2
bne cr7, L(dLcr7x)
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
ld rOFF8, rOFF8SAVE(r1)
ld rOFF16, rOFF16SAVE(r1)
ld rOFF24, rOFF24SAVE(r1)
ld rOFF32, rOFF32SAVE(r1)
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
li rRTN, 0
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
blr
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
.align 4
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
L(dLcr7):
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
ld rWORD8, rWORD8SAVE(r1)
ld rWORD7, rWORD7SAVE(r1)
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
L(dLcr7x):
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
ld rOFF8, rOFF8SAVE(r1)
ld rOFF16, rOFF16SAVE(r1)
ld rOFF24, rOFF24SAVE(r1)
ld rOFF32, rOFF32SAVE(r1)
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
li rRTN, 1
bgtlr cr7
li rRTN, -1
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
blr
.align 4
L(dLcr1):
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
ld rWORD8, rWORD8SAVE(r1)
ld rWORD7, rWORD7SAVE(r1)
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
L(dLcr1x):
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
ld rOFF8, rOFF8SAVE(r1)
ld rOFF16, rOFF16SAVE(r1)
ld rOFF24, rOFF24SAVE(r1)
ld rOFF32, rOFF32SAVE(r1)
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
li rRTN, 1
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
bgtlr cr1
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
li rRTN, -1
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
blr
.align 4
L(dLcr6):
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
ld rWORD8, rWORD8SAVE(r1)
ld rWORD7, rWORD7SAVE(r1)
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
L(dLcr6x):
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
ld rOFF8, rOFF8SAVE(r1)
ld rOFF16, rOFF16SAVE(r1)
ld rOFF24, rOFF24SAVE(r1)
ld rOFF32, rOFF32SAVE(r1)
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
li rRTN, 1
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
bgtlr cr6
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
li rRTN, -1
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
blr
.align 4
L(dLcr5):
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
ld rWORD8, rWORD8SAVE(r1)
ld rWORD7, rWORD7SAVE(r1)
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
L(dLcr5x):
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
ld rOFF8, rOFF8SAVE(r1)
ld rOFF16, rOFF16SAVE(r1)
ld rOFF24, rOFF24SAVE(r1)
ld rOFF32, rOFF32SAVE(r1)
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
li rRTN, 1
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
bgtlr cr5
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
li rRTN, -1
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
blr
.align 4
L(bytealigned):
mtctr rN
/* We need to prime this loop. This loop is swing modulo scheduled
Fix whitespaces.
2010-06-15 00:15:33 +00:00
to avoid pipe delays. The dependent instruction latencies (load to
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
compare to conditional branch) is 2 to 3 cycles. In this loop each
dispatch group ends in a branch and takes 1 cycle. Effectively
Fix whitespaces.
2010-06-15 00:15:33 +00:00
the first iteration of the loop only serves to load operands and
branches based on compares are delayed until the next loop.
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
So we must precondition some registers and condition codes so that
we don't exit the loop early on the first iteration. */
Fix whitespaces.
2010-06-15 00:15:33 +00:00
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
lbz rWORD1, 0(rSTR1)
lbz rWORD2, 0(rSTR2)
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
bdz L(b11)
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpld cr7, rWORD1, rWORD2
lbz rWORD3, 1(rSTR1)
lbz rWORD4, 1(rSTR2)
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
bdz L(b12)
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpld cr1, rWORD3, rWORD4
lbzu rWORD5, 2(rSTR1)
lbzu rWORD6, 2(rSTR2)
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
bdz L(b13)
.align 4
L(bLoop):
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
lbzu rWORD1, 1(rSTR1)
lbzu rWORD2, 1(rSTR2)
bne cr7, L(bLcr7)
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpld cr6, rWORD5, rWORD6
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
bdz L(b3i)
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
lbzu rWORD3, 1(rSTR1)
lbzu rWORD4, 1(rSTR2)
bne cr1, L(bLcr1)
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpld cr7, rWORD1, rWORD2
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
bdz L(b2i)
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
lbzu rWORD5, 1(rSTR1)
lbzu rWORD6, 1(rSTR2)
bne cr6, L(bLcr6)
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpld cr1, rWORD3, rWORD4
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
bdnz L(bLoop)
/* We speculatively loading bytes before we have tested the previous
bytes. But we must avoid overrunning the length (in the ctr) to
Fix whitespaces.
2010-06-15 00:15:33 +00:00
prevent these speculative loads from causing a segfault. In this
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
case the loop will exit early (before the all pending bytes are
tested. In this case we must complete the pending operations
before returning. */
L(b1i):
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
bne cr7, L(bLcr7)
bne cr1, L(bLcr1)
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
b L(bx56)
.align 4
L(b2i):
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
bne cr6, L(bLcr6)
bne cr7, L(bLcr7)
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
b L(bx34)
.align 4
L(b3i):
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
bne cr1, L(bLcr1)
bne cr6, L(bLcr6)
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
b L(bx12)
.align 4
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
L(bLcr7):
li rRTN, 1
bgtlr cr7
li rRTN, -1
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
blr
L(bLcr1):
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
li rRTN, 1
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
bgtlr cr1
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
li rRTN, -1
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
blr
L(bLcr6):
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
li rRTN, 1
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
bgtlr cr6
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
li rRTN, -1
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
blr
L(b13):
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
bne cr7, L(bx12)
bne cr1, L(bx34)
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
L(bx56):
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
sub rRTN, rWORD5, rWORD6
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
blr
nop
L(b12):
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
bne cr7, L(bx12)
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
L(bx34):
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
sub rRTN, rWORD3, rWORD4
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
blr
L(b11):
L(bx12):
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
sub rRTN, rWORD1, rWORD2
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
blr
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
Fix whitespaces.
2010-06-15 00:15:33 +00:00
.align 4
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
L(zeroLength):
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
li rRTN, 0
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
blr
.align 4
/* At this point we know the strings have different alignment and the
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
compare length is at least 8 bytes. r12 contains the low order
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
3 bits of rSTR1 and cr5 contains the result of the logical compare
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
of r12 to 0. If r12 == 0 then rStr1 is double word
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
aligned and can perform the DWunaligned loop.
Fix whitespaces.
2010-06-15 00:15:33 +00:00
Fix spelling errors in sysdeps/powerpc files.
2013-01-07 17:20:53 +00:00
Otherwise we know that rSTR1 is not already DW aligned yet.
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
So we can force the string addresses to the next lower DW
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
boundary and special case this first DW using shift left to
Fix spelling errors in sysdeps/powerpc files.
2013-01-07 17:20:53 +00:00
eliminate bits preceding the first byte. Since we want to join the
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
normal (DWaligned) compare loop, starting at the second double word,
we need to adjust the length (rN) and special case the loop
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
versioning for the first DW. This ensures that the loop count is
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
correct and the first DW (shifted) is in the expected resister pair. */
L(unaligned):
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
std rSHL, rSHLSAVE(r1)
cfi_offset(rSHL, rSHLSAVE)
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
clrldi rSHL, rSTR2, 61
beq cr6, L(duzeroLength)
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
std rSHR, rSHRSAVE(r1)
cfi_offset(rSHR, rSHRSAVE)
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
beq cr5, L(DWunaligned)
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
std rWORD8_SHIFT, rWORD8SHIFTSAVE(r1)
cfi_offset(rWORD8_SHIFT, rWORD8SHIFTSAVE)
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
/* Adjust the logical start of rSTR2 to compensate for the extra bits
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
in the 1st rSTR1 DW. */
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
sub rWORD8_SHIFT, rSTR2, r12
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
/* But do not attempt to address the DW before that DW that contains
the actual start of rSTR2. */
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
clrrdi rSTR2, rSTR2, 3
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
std rWORD2_SHIFT, rWORD2SHIFTSAVE(r1)
cfi_offset(rWORD2_SHIFT, rWORD2SHIFTSAVE)
Fix spelling errors in sysdeps/powerpc files.
2013-01-07 17:20:53 +00:00
/* Compute the left/right shift counts for the unaligned rSTR2,
Fix whitespaces.
2010-06-15 00:15:33 +00:00
compensating for the logical (DW aligned) start of rSTR1. */
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
clrldi rSHL, rWORD8_SHIFT, 61
clrrdi rSTR1, rSTR1, 3
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
std rWORD4_SHIFT, rWORD4SHIFTSAVE(r1)
cfi_offset(rWORD4_SHIFT, rWORD4SHIFTSAVE)
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
sldi rSHL, rSHL, 3
cmpld cr5, rWORD8_SHIFT, rSTR2
add rN, rN, r12
sldi rWORD6, r12, 3
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
std rWORD6_SHIFT, rWORD6SHIFTSAVE(r1)
cfi_offset(rWORD6_SHIFT, rWORD6SHIFTSAVE)
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
subfic rSHR, rSHL, 64
srdi r0, rN, 5 /* Divide by 32 */
andi. r12, rN, 24 /* Get the DW remainder */
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
/* We normally need to load 2 DWs to start the unaligned rSTR2, but in
this special case those bits may be discarded anyway. Also we
must avoid loading a DW where none of the bits are part of rSTR2 as
this may cross a page boundary and cause a page fault. */
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
li rWORD8, 0
blt cr5, L(dus0)
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD8, 0, rSTR2
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
addi rSTR2, rSTR2, 8
sld rWORD8, rWORD8, rSHL
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
L(dus0):
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD1, 0, rSTR1
LD rWORD2, 0, rSTR2
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpldi cr1, r12, 16
cmpldi cr7, rN, 32
srd r12, rWORD2, rSHR
clrldi rN, rN, 61
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
beq L(duPs4)
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
mtctr r0
or rWORD8, r12, rWORD8
bgt cr1, L(duPs3)
beq cr1, L(duPs2)
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
/* Remainder is 8 */
.align 4
L(dusP1):
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
sld rWORD8_SHIFT, rWORD2, rSHL
sld rWORD7, rWORD1, rWORD6
sld rWORD8, rWORD8, rWORD6
bge cr7, L(duP1e)
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
/* At this point we exit early with the first double word compare
complete and remainder of 0 to 7 bytes. See L(du14) for details on
how we handle the remaining bytes. */
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpld cr5, rWORD7, rWORD8
sldi. rN, rN, 3
bne cr5, L(duLcr5)
cmpld cr7, rN, rSHR
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
beq L(duZeroReturn)
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
li r0, 0
ble cr7, L(dutrim)
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD2, rOFF8, rSTR2
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
srd r0, rWORD2, rSHR
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
b L(dutrim)
/* Remainder is 16 */
.align 4
L(duPs2):
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
sld rWORD6_SHIFT, rWORD2, rSHL
sld rWORD5, rWORD1, rWORD6
sld rWORD6, rWORD8, rWORD6
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
b L(duP2e)
/* Remainder is 24 */
.align 4
L(duPs3):
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
sld rWORD4_SHIFT, rWORD2, rSHL
sld rWORD3, rWORD1, rWORD6
sld rWORD4, rWORD8, rWORD6
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
b L(duP3e)
/* Count is a multiple of 32, remainder is 0 */
.align 4
L(duPs4):
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
mtctr r0
or rWORD8, r12, rWORD8
sld rWORD2_SHIFT, rWORD2, rSHL
sld rWORD1, rWORD1, rWORD6
sld rWORD2, rWORD8, rWORD6
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
b L(duP4e)
/* At this point we know rSTR1 is double word aligned and the
compare length is at least 8 bytes. */
.align 4
L(DWunaligned):
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
std rWORD8_SHIFT, rWORD8SHIFTSAVE(r1)
cfi_offset(rWORD8_SHIFT, rWORD8SHIFTSAVE)
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
clrrdi rSTR2, rSTR2, 3
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
std rWORD2_SHIFT, rWORD2SHIFTSAVE(r1)
cfi_offset(rWORD2_SHIFT, rWORD2SHIFTSAVE)
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
srdi r0, rN, 5 /* Divide by 32 */
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
std rWORD4_SHIFT, rWORD4SHIFTSAVE(r1)
cfi_offset(rWORD4_SHIFT, rWORD4SHIFTSAVE)
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
andi. r12, rN, 24 /* Get the DW remainder */
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
std rWORD6_SHIFT, rWORD6SHIFTSAVE(r1)
cfi_offset(rWORD6_SHIFT, rWORD6SHIFTSAVE)
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
sldi rSHL, rSHL, 3
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD6, 0, rSTR2
LD rWORD8, rOFF8, rSTR2
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
addi rSTR2, rSTR2, 8
cmpldi cr1, r12, 16
cmpldi cr7, rN, 32
clrldi rN, rN, 61
subfic rSHR, rSHL, 64
sld rWORD6_SHIFT, rWORD6, rSHL
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
beq L(duP4)
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
mtctr r0
bgt cr1, L(duP3)
beq cr1, L(duP2)
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
/* Remainder is 8 */
.align 4
L(duP1):
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
srd r12, rWORD8, rSHR
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD7, 0, rSTR1
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
sld rWORD8_SHIFT, rWORD8, rSHL
or rWORD8, r12, rWORD6_SHIFT
blt cr7, L(duP1x)
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
L(duP1e):
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD1, rOFF8, rSTR1
LD rWORD2, rOFF8, rSTR2
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpld cr5, rWORD7, rWORD8
srd r0, rWORD2, rSHR
sld rWORD2_SHIFT, rWORD2, rSHL
or rWORD2, r0, rWORD8_SHIFT
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD3, rOFF16, rSTR1
LD rWORD4, rOFF16, rSTR2
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpld cr7, rWORD1, rWORD2
srd r12, rWORD4, rSHR
sld rWORD4_SHIFT, rWORD4, rSHL
bne cr5, L(duLcr5)
or rWORD4, r12, rWORD2_SHIFT
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD5, rOFF24, rSTR1
LD rWORD6, rOFF24, rSTR2
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpld cr1, rWORD3, rWORD4
srd r0, rWORD6, rSHR
sld rWORD6_SHIFT, rWORD6, rSHL
bne cr7, L(duLcr7)
or rWORD6, r0, rWORD4_SHIFT
cmpld cr6, rWORD5, rWORD6
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
b L(duLoop3)
.align 4
/* At this point we exit early with the first double word compare
complete and remainder of 0 to 7 bytes. See L(du14) for details on
how we handle the remaining bytes. */
L(duP1x):
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpld cr5, rWORD7, rWORD8
sldi. rN, rN, 3
bne cr5, L(duLcr5)
cmpld cr7, rN, rSHR
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
beq L(duZeroReturn)
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
li r0, 0
ble cr7, L(dutrim)
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD2, rOFF8, rSTR2
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
srd r0, rWORD2, rSHR
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
b L(dutrim)
/* Remainder is 16 */
.align 4
L(duP2):
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
srd r0, rWORD8, rSHR
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD5, 0, rSTR1
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
or rWORD6, r0, rWORD6_SHIFT
sld rWORD6_SHIFT, rWORD8, rSHL
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
L(duP2e):
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD7, rOFF8, rSTR1
LD rWORD8, rOFF8, rSTR2
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpld cr6, rWORD5, rWORD6
srd r12, rWORD8, rSHR
sld rWORD8_SHIFT, rWORD8, rSHL
or rWORD8, r12, rWORD6_SHIFT
blt cr7, L(duP2x)
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD1, rOFF16, rSTR1
LD rWORD2, rOFF16, rSTR2
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpld cr5, rWORD7, rWORD8
bne cr6, L(duLcr6)
srd r0, rWORD2, rSHR
sld rWORD2_SHIFT, rWORD2, rSHL
or rWORD2, r0, rWORD8_SHIFT
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD3, rOFF24, rSTR1
LD rWORD4, rOFF24, rSTR2
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpld cr7, rWORD1, rWORD2
bne cr5, L(duLcr5)
srd r12, rWORD4, rSHR
sld rWORD4_SHIFT, rWORD4, rSHL
or rWORD4, r12, rWORD2_SHIFT
addi rSTR1, rSTR1, 8
addi rSTR2, rSTR2, 8
cmpld cr1, rWORD3, rWORD4
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
b L(duLoop2)
.align 4
L(duP2x):
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpld cr5, rWORD7, rWORD8
addi rSTR1, rSTR1, 8
addi rSTR2, rSTR2, 8
bne cr6, L(duLcr6)
sldi. rN, rN, 3
bne cr5, L(duLcr5)
cmpld cr7, rN, rSHR
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
beq L(duZeroReturn)
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
li r0, 0
ble cr7, L(dutrim)
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD2, rOFF8, rSTR2
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
srd r0, rWORD2, rSHR
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
b L(dutrim)
/* Remainder is 24 */
.align 4
L(duP3):
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
srd r12, rWORD8, rSHR
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD3, 0, rSTR1
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
sld rWORD4_SHIFT, rWORD8, rSHL
or rWORD4, r12, rWORD6_SHIFT
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
L(duP3e):
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD5, rOFF8, rSTR1
LD rWORD6, rOFF8, rSTR2
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpld cr1, rWORD3, rWORD4
srd r0, rWORD6, rSHR
sld rWORD6_SHIFT, rWORD6, rSHL
or rWORD6, r0, rWORD4_SHIFT
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD7, rOFF16, rSTR1
LD rWORD8, rOFF16, rSTR2
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpld cr6, rWORD5, rWORD6
bne cr1, L(duLcr1)
srd r12, rWORD8, rSHR
sld rWORD8_SHIFT, rWORD8, rSHL
or rWORD8, r12, rWORD6_SHIFT
blt cr7, L(duP3x)
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD1, rOFF24, rSTR1
LD rWORD2, rOFF24, rSTR2
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpld cr5, rWORD7, rWORD8
bne cr6, L(duLcr6)
srd r0, rWORD2, rSHR
sld rWORD2_SHIFT, rWORD2, rSHL
or rWORD2, r0, rWORD8_SHIFT
addi rSTR1, rSTR1, 16
addi rSTR2, rSTR2, 16
cmpld cr7, rWORD1, rWORD2
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
b L(duLoop1)
.align 4
L(duP3x):
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
addi rSTR1, rSTR1, 16
addi rSTR2, rSTR2, 16
cmpld cr5, rWORD7, rWORD8
bne cr6, L(duLcr6)
sldi. rN, rN, 3
bne cr5, L(duLcr5)
cmpld cr7, rN, rSHR
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
beq L(duZeroReturn)
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
li r0, 0
ble cr7, L(dutrim)
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD2, rOFF8, rSTR2
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
srd r0, rWORD2, rSHR
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
b L(dutrim)
/* Count is a multiple of 32, remainder is 0 */
.align 4
L(duP4):
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
mtctr r0
srd r0, rWORD8, rSHR
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD1, 0, rSTR1
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
sld rWORD2_SHIFT, rWORD8, rSHL
or rWORD2, r0, rWORD6_SHIFT
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
L(duP4e):
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD3, rOFF8, rSTR1
LD rWORD4, rOFF8, rSTR2
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpld cr7, rWORD1, rWORD2
srd r12, rWORD4, rSHR
sld rWORD4_SHIFT, rWORD4, rSHL
or rWORD4, r12, rWORD2_SHIFT
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD5, rOFF16, rSTR1
LD rWORD6, rOFF16, rSTR2
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpld cr1, rWORD3, rWORD4
bne cr7, L(duLcr7)
srd r0, rWORD6, rSHR
sld rWORD6_SHIFT, rWORD6, rSHL
or rWORD6, r0, rWORD4_SHIFT
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD7, rOFF24, rSTR1
LD rWORD8, rOFF24, rSTR2
addi rSTR1, rSTR1, 24
addi rSTR2, rSTR2, 24
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpld cr6, rWORD5, rWORD6
bne cr1, L(duLcr1)
srd r12, rWORD8, rSHR
sld rWORD8_SHIFT, rWORD8, rSHL
or rWORD8, r12, rWORD6_SHIFT
cmpld cr5, rWORD7, rWORD8
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
bdz L(du24) /* Adjust CTR as we start with +4 */
/* This is the primary loop */
.align 4
L(duLoop):
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD1, rOFF8, rSTR1
LD rWORD2, rOFF8, rSTR2
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpld cr1, rWORD3, rWORD4
bne cr6, L(duLcr6)
srd r0, rWORD2, rSHR
sld rWORD2_SHIFT, rWORD2, rSHL
or rWORD2, r0, rWORD8_SHIFT
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
L(duLoop1):
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD3, rOFF16, rSTR1
LD rWORD4, rOFF16, rSTR2
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpld cr6, rWORD5, rWORD6
bne cr5, L(duLcr5)
srd r12, rWORD4, rSHR
sld rWORD4_SHIFT, rWORD4, rSHL
or rWORD4, r12, rWORD2_SHIFT
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
L(duLoop2):
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD5, rOFF24, rSTR1
LD rWORD6, rOFF24, rSTR2
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpld cr5, rWORD7, rWORD8
bne cr7, L(duLcr7)
srd r0, rWORD6, rSHR
sld rWORD6_SHIFT, rWORD6, rSHL
or rWORD6, r0, rWORD4_SHIFT
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
L(duLoop3):
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD7, rOFF32, rSTR1
LD rWORD8, rOFF32, rSTR2
addi rSTR1, rSTR1, 32
addi rSTR2, rSTR2, 32
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpld cr7, rWORD1, rWORD2
bne cr1, L(duLcr1)
srd r12, rWORD8, rSHR
sld rWORD8_SHIFT, rWORD8, rSHL
or rWORD8, r12, rWORD6_SHIFT
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
bdnz L(duLoop)
L(duL4):
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
cmpld cr1, rWORD3, rWORD4
bne cr6, L(duLcr6)
cmpld cr6, rWORD5, rWORD6
bne cr5, L(duLcr5)
cmpld cr5, rWORD7, rWORD8
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
L(du44):
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
bne cr7, L(duLcr7)
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
L(du34):
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
bne cr1, L(duLcr1)
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
L(du24):
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
bne cr6, L(duLcr6)
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
L(du14):
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
sldi. rN, rN, 3
bne cr5, L(duLcr5)
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
/* At this point we have a remainder of 1 to 7 bytes to compare. We use
Fix spelling errors in sysdeps/powerpc files.
2013-01-07 17:20:53 +00:00
shift right double to eliminate bits beyond the compare length.
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
Fix whitespaces.
2010-06-15 00:15:33 +00:00
However it may not be safe to load rWORD2 which may be beyond the
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
string length. So we compare the bit length of the remainder to
the right shift count (rSHR). If the bit count is less than or equal
we do not need to load rWORD2 (all significant bits are already in
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
rWORD8_SHIFT). */
cmpld cr7, rN, rSHR
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
beq L(duZeroReturn)
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
li r0, 0
ble cr7, L(dutrim)
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD2, rOFF8, rSTR2
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
srd r0, rWORD2, rSHR
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
.align 4
L(dutrim):
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
LD rWORD1, rOFF8, rSTR1
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
ld rWORD8, -8(r1)
subfic rN, rN, 64 /* Shift count is 64 - (rN * 8). */
or rWORD2, r0, rWORD8_SHIFT
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
ld rWORD7, rWORD7SAVE(r1)
ld rSHL, rSHLSAVE(r1)
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
srd rWORD1, rWORD1, rN
srd rWORD2, rWORD2, rN
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
ld rSHR, rSHRSAVE(r1)
ld rWORD8_SHIFT, rWORD8SHIFTSAVE(r1)
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
li rRTN, 0
cmpld cr7, rWORD1, rWORD2
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
ld rWORD2_SHIFT, rWORD2SHIFTSAVE(r1)
ld rWORD4_SHIFT, rWORD4SHIFTSAVE(r1)
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
beq cr7, L(dureturn24)
li rRTN, 1
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
ld rWORD6_SHIFT, rWORD6SHIFTSAVE(r1)
ld rOFF8, rOFF8SAVE(r1)
ld rOFF16, rOFF16SAVE(r1)
ld rOFF24, rOFF24SAVE(r1)
ld rOFF32, rOFF32SAVE(r1)
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
bgtlr cr7
li rRTN, -1
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
blr
.align 4
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
L(duLcr7):
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
ld rWORD8, rWORD8SAVE(r1)
ld rWORD7, rWORD7SAVE(r1)
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
li rRTN, 1
bgt cr7, L(dureturn29)
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
ld rSHL, rSHLSAVE(r1)
ld rSHR, rSHRSAVE(r1)
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
li rRTN, -1
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
b L(dureturn27)
.align 4
L(duLcr1):
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
ld rWORD8, rWORD8SAVE(r1)
ld rWORD7, rWORD7SAVE(r1)
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
li rRTN, 1
bgt cr1, L(dureturn29)
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
ld rSHL, rSHLSAVE(r1)
ld rSHR, rSHRSAVE(r1)
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
li rRTN, -1
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
b L(dureturn27)
.align 4
L(duLcr6):
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
ld rWORD8, rWORD8SAVE(r1)
ld rWORD7, rWORD7SAVE(r1)
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
li rRTN, 1
bgt cr6, L(dureturn29)
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
ld rSHL, rSHLSAVE(r1)
ld rSHR, rSHRSAVE(r1)
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
li rRTN, -1
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
b L(dureturn27)
.align 4
L(duLcr5):
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
ld rWORD8, rWORD8SAVE(r1)
ld rWORD7, rWORD7SAVE(r1)
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
li rRTN, 1
bgt cr5, L(dureturn29)
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
ld rSHL, rSHLSAVE(r1)
ld rSHR, rSHRSAVE(r1)
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
li rRTN, -1
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
b L(dureturn27)
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
.align 3
L(duZeroReturn):
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
li rRTN, 0
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
.align 4
L(dureturn):
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
ld rWORD8, rWORD8SAVE(r1)
ld rWORD7, rWORD7SAVE(r1)
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
L(dureturn29):
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
ld rSHL, rSHLSAVE(r1)
ld rSHR, rSHRSAVE(r1)
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
L(dureturn27):
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
ld rWORD8_SHIFT, rWORD8SHIFTSAVE(r1)
ld rWORD2_SHIFT, rWORD2SHIFTSAVE(r1)
ld rWORD4_SHIFT, rWORD4SHIFTSAVE(r1)
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
L(dureturn24):
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
ld rWORD6_SHIFT, rWORD6SHIFTSAVE(r1)
ld rOFF8, rOFF8SAVE(r1)
ld rOFF16, rOFF16SAVE(r1)
ld rOFF24, rOFF24SAVE(r1)
ld rOFF32, rOFF32SAVE(r1)
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
blr
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
L(duzeroLength):
powerpc: Fix POWER7/PPC64 performance regression on LE This patch fixes a performance regression on the POWER7/PPC64 memcmp porting for Little Endian. The LE code uses 'ldbrx' instruction to read the memory on byte reversed form, however ISA 2.06 just provide the indexed form which uses a register value as additional index, instead of a fixed value enconded in the instruction. And the port strategy for LE uses r0 index value and update the address value on each compare loop interation. For large compare size values, it adds 8 more instructions plus some more depending of trailing size. This patch fixes it by adding pre-calculate indexes to remove the address update on loops and tailing sizes. For large sizes it shows a considerable gain, with double performance pairing with BE.
2015-01-12 01:33:17 +00:00
ld rOFF8, rOFF8SAVE(r1)
ld rOFF16, rOFF16SAVE(r1)
ld rOFF24, rOFF24SAVE(r1)
ld rOFF32, rOFF32SAVE(r1)
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
li rRTN, 0
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
blr
Remove powerpc64 bounded-pointers code.
2013-03-06 00:10:21 +00:00
END (memcmp)
power7 string compare optimizations
2010-06-15 00:13:24 +00:00
libc_hidden_builtin_def (memcmp)
PowerPC LE memcmp http://sourceware.org/ml/libc-alpha/2013-08/msg00102.html This is a rather large patch due to formatting and renaming. The formatting changes were to make it possible to compare power7 and power4 versions of memcmp. Using different register defines came about while I was wrestling with the code, trying to find spare registers at one stage. I found it much simpler if we refer to a reg by the same name throughout a function, so it's better if short-term multiple use regs like rTMP are referred to using their register number. I made the cr field usage changes when attempting to reload rWORDn regs in the exit path to byte swap before comparing when little-endian. That proved a bad idea due to the pipelining involved in the main loop; Offsets to reload the regs were different first time around the loop.. Anyway, I left the cr field usage changes in place for consistency. Aside from these more-or-less cosmetic changes, I fixed a number of places where an early exit path restores regs unnecessarily, removed some dead code, and optimised one or two exits. * sysdeps/powerpc/powerpc64/power7/memcmp.S: Add little-endian support. Formatting. Consistently use rXXX register defines or rN defines. Use early exit labels that avoid restoring unused non-volatile regs. Make cr field use more consistent with rWORDn compares. Rename regs used as shift registers for unaligned loop, using rN defines for short lifetime/multiple use regs. * sysdeps/powerpc/powerpc64/power4/memcmp.S: Likewise. * sysdeps/powerpc/powerpc32/power7/memcmp.S: Likewise. Exit with addi 1,1,64 to pop stack frame. Simplify return value code. * sysdeps/powerpc/powerpc32/power4/memcmp.S: Likewise.
2013-08-17 09:16:47 +00:00
weak_alias (memcmp, bcmp)
Reference in New Issue Copy Permalink