glibc/sysdeps/powerpc/powerpc64/power8/strcpy.S

414 lines
10 KiB
ArmAsm

/* Optimized strcpy/stpcpy implementation for PowerPC64/POWER8.
Copyright (C) 2015-2020 Free Software Foundation, Inc.
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
License along with the GNU C Library; if not, see
<https://www.gnu.org/licenses/>. */
#include <sysdep.h>
#ifdef USE_AS_STPCPY
# ifndef STPCPY
# define FUNC_NAME __stpcpy
# else
# define FUNC_NAME STPCPY
# endif
#else
# ifndef STRCPY
# define FUNC_NAME strcpy
# else
# define FUNC_NAME STRCPY
# endif
#endif /* !USE_AS_STPCPY */
/* Implements the function
char * [r3] strcpy (char *dest [r3], const char *src [r4])
or
char * [r3] stpcpy (char *dest [r3], const char *src [r4])
if USE_AS_STPCPY is defined.
The implementation uses unaligned doubleword access to avoid specialized
code paths depending of data alignment. Although recent powerpc64 uses
64K as default, the page cross handling assumes minimum page size of
4k. */
.machine power8
ENTRY_TOCLESS (FUNC_NAME, 4)
li r0,0 /* Doubleword with null chars to use
with cmpb. */
/* Check if the [src]+15 will cross a 4K page by checking if the bit
indicating the page size changes. Basically:
uint64_t srcin = (uint64_t)src;
uint64_t ob = srcin & 4096UL;
uint64_t nb = (srcin+15UL) & 4096UL;
if (ob ^ nb)
goto pagecross; */
addi r9,r4,15
xor r9,r9,r4
rlwinm. r9,r9,0,19,19
bne L(pagecross)
/* For short string (less than 16 bytes), just calculate its size as
strlen and issues a memcpy if null is found. */
mr r7,r4
ld r12,0(r7) /* Load doubleword from memory. */
cmpb r10,r12,r0 /* Check for null bytes in DWORD1. */
cmpdi cr7,r10,0 /* If r10 == 0, no null's have been found. */
bne cr7,L(done)
ldu r8,8(r7)
cmpb r10,r8,r0
cmpdi cr7,r10,0
bne cr7,L(done)
b L(loop_before)
.align 4
L(pagecross):
clrrdi r7,r4,3 /* Align the address to doubleword boundary. */
rlwinm r6,r4,3,26,28 /* Calculate padding. */
li r5,-1 /* MASK = 0xffffffffffffffff. */
ld r12,0(r7) /* Load doubleword from memory. */
#ifdef __LITTLE_ENDIAN__
sld r5,r5,r6
#else
srd r5,r5,r6 /* MASK = MASK >> padding. */
#endif
orc r9,r12,r5 /* Mask bits that are not part of the string. */
cmpb r10,r9,r0 /* Check for null bytes in DWORD1. */
cmpdi cr7,r10,0 /* If r10 == 0, no null's have been found. */
bne cr7,L(done)
ldu r6,8(r7)
cmpb r10,r6,r0
cmpdi cr7,r10,0
bne cr7,L(done)
ld r12,0(r7)
cmpb r10,r12,r0
cmpdi cr7,r10,0
bne cr7,L(done)
ldu r6,8(r7)
cmpb r10,r6,r0
cmpdi cr7,r10,0
bne cr7,L(done)
/* We checked for 24 - x bytes, with x being the source alignment
(0 <= x <= 16), and no zero has been found. Start the loop
copy with doubleword aligned address. */
mr r7,r4
ld r12, 0(r7)
ldu r8, 8(r7)
L(loop_before):
/* Save the two doublewords read from source and align the source
to 16 bytes for the loop. */
mr r11,r3
std r12,0(r11)
std r8,8(r11)
addi r11,r11,16
rldicl r9,r4,0,60
subf r7,r9,r7
subf r11,r9,r11
/* Source is adjusted to 16B alignment and destination r11 is
also moved based on that adjustment. Now check if r11 is
also 16B aligned to move to vectorized loop. */
andi. r6, r11, 0xF
bne L(loop_start)
/* Prepare for the loop. */
subf r4, r9, r4 /* Adjust r4 based on alignment. */
li r7, 16 /* Load required offsets. */
li r8, 32
li r9, 48
vspltisb v0, 0
addi r4, r4, 16
/* Are we 64-byte aligned? If so, jump to the vectorized loop.
Else copy 16B till r4 is 64B aligned. */
andi. r6, r4, 63
beq L(qw_loop)
lvx v6, 0, r4 /* Load 16 bytes from memory. */
vcmpequb. v5, v0, v6 /* Check for null. */
bne cr6, L(qw_done)
stvx v6, 0, r11 /* Store 16 bytes. */
addi r4, r4, 16 /* Increment the address. */
addi r11, r11, 16
andi. r6, r4, 63
beq L(qw_loop)
lvx v6, 0, r4
vcmpequb. v5, v0, v6
bne cr6, L(qw_done)
stvx v6, 0, r11
addi r4, r4, 16
addi r11, r11, 16
andi. r6, r4, 63
beq L(qw_loop)
lvx v6, 0, r4
vcmpequb. v5, v0, v6
bne cr6, L(qw_done)
stvx v6, 0, r11
addi r4, r4, 16
addi r11, r11, 16
.align 4
L(qw_loop):
lvx v1, r4, r0 /* Load 4 quadwords. */
lvx v2, r4, r7
lvx v3, r4, r8
lvx v4, r4, r9
vminub v5, v1, v2 /* Compare and merge into one VR for speed. */
vminub v8, v3, v4
vminub v7, v5, v8
vcmpequb. v7, v7, v0 /* Check for NULLs. */
bne cr6, L(qw_loop_done)
stvx v1, r11, r0 /* Store 4 quadwords. */
stvx v2, r11, r7
stvx v3, r11, r8
stvx v4, r11, r9
addi r4, r4, 64 /* Adjust address for the next iteration. */
addi r11, r11, 64 /* Adjust address for the next iteration. */
lvx v1, r4, r0 /* Load 4 quadwords. */
lvx v2, r4, r7
lvx v3, r4, r8
lvx v4, r4, r9
vminub v5, v1, v2 /* Compare and merge into one VR for speed. */
vminub v8, v3, v4
vminub v7, v5, v8
vcmpequb. v7, v7, v0 /* Check for NULLs. */
bne cr6, L(qw_loop_done)
stvx v1, r11, r0 /* Store 4 quadwords. */
stvx v2, r11, r7
stvx v3, r11, r8
stvx v4, r11, r9
addi r4, r4, 64 /* Adjust address for the next iteration. */
addi r11, r11, 64 /* Adjust address for the next iteration. */
lvx v1, r4, r0 /* Load 4 quadwords. */
lvx v2, r4, r7
lvx v3, r4, r8
lvx v4, r4, r9
vminub v5, v1, v2 /* Compare and merge into one VR for speed. */
vminub v8, v3, v4
vminub v7, v5, v8
vcmpequb. v7, v7, v0 /* Check for NULLs. */
bne cr6, L(qw_loop_done)
stvx v1, r11, r0 /* Store 4 quadwords. */
stvx v2, r11, r7
stvx v3, r11, r8
stvx v4, r11, r9
addi r4, r4, 64 /* Adjust address for the next iteration. */
addi r11, r11, 64 /* Adjust address for the next iteration. */
b L(qw_loop)
.align 4
L(qw_loop_done):
/* Null found in one of the 4 loads. */
vcmpequb. v7, v1, v0
vor v6, v1, v1
bne cr6, L(qw_done)
/* Not on the first 16B, So store it. */
stvx v1, r11, r0
addi r4, r4, 16
addi r11, r11, 16
vcmpequb. v7, v2, v0
vor v6, v2, v2
bne cr6, L(qw_done)
/* Not on the second 16B, So store it. */
stvx v2, r11, r0
addi r4, r4, 16
addi r11, r11, 16
vcmpequb. v7, v3, v0
vor v6, v3, v3
bne cr6, L(qw_done)
/* Not on the third 16B, So store it. */
stvx v6, r11, r0
addi r4, r4, 16
addi r11, r11, 16
vor v6, v4, v4
.align 4
L(qw_done):
mr r7, r4
/* Move the result to GPR. */
#ifdef __LITTLE_ENDIAN__
vsldoi v4, v6, v0, 8
mfvrd r12, v4
#else
mfvrd r12, v6
#endif
/* Check for null in the first 8 bytes. */
cmpb r10, r12, r0
cmpdi cr6, r10, 0
bne cr6, L(done2)
/* Null found in second doubleword. */
#ifdef __LITTLE_ENDIAN__
mfvrd r6, v6
#else
vsldoi v6, v6, v0, 8
mfvrd r6, v6
#endif
cmpb r10, r6, r0
addi r7, r7, 8
b L(done2)
.align 5
L(loop):
std r12, 0(r11)
std r6, 8(r11)
addi r11,r11,16
L(loop_start):
/* Load two doublewords, compare and merge in a
single register for speed. This is an attempt
to speed up the null-checking process for bigger strings. */
ld r12, 8(r7)
ldu r6, 16(r7)
cmpb r10,r12,r0
cmpb r9,r6,r0
or r8,r9,r10 /* Merge everything in one doubleword. */
cmpdi cr7,r8,0
beq cr7,L(loop)
/* OK, one (or both) of the doublewords contains a null byte. Check
the first doubleword and decrement the address in case the first
doubleword really contains a null byte. */
addi r4,r7,-8
cmpdi cr6,r10,0
addi r7,r7,-8
bne cr6,L(done2)
/* The null byte must be in the second doubleword. Adjust the address
again and move the result of cmpb to r10 so we can calculate the
length. */
mr r10,r9
addi r7,r7,8
b L(done2)
/* r10 has the output of the cmpb instruction, that is, it contains
0xff in the same position as the null byte in the original
doubleword from the string. Use that to calculate the length. */
L(done):
mr r11,r3
L(done2):
#ifdef __LITTLE_ENDIAN__
addi r9, r10, -1 /* Form a mask from trailing zeros. */
andc r9, r9, r10
popcntd r6, r9 /* Count the bits in the mask. */
#else
cntlzd r6,r10 /* Count leading zeros before the match. */
#endif
subf r5,r4,r7
srdi r6,r6,3 /* Convert leading/trailing zeros to bytes. */
add r8,r5,r6 /* Compute final length. */
#ifdef USE_AS_STPCPY
/* stpcpy returns the dest address plus the size not counting the
final '\0'. */
add r3,r11,r8
#endif
addi r8,r8,1 /* Final '/0'. */
cmpldi cr6,r8,8
mtocrf 0x01,r8
ble cr6,L(copy_LE_8)
cmpldi cr1,r8,16
blt cr1,8f
/* Handle copies of 0~31 bytes. */
.align 4
L(copy_LT_32):
/* At least 6 bytes to go. */
blt cr1,8f
/* Copy 16 bytes. */
ld r6,0(r4)
ld r8,8(r4)
addi r4,r4,16
std r6,0(r11)
std r8,8(r11)
addi r11,r11,16
8: /* Copy 8 bytes. */
bf 28,L(tail4)
ld r6,0(r4)
addi r4,r4,8
std r6,0(r11)
addi r11,r11,8
.align 4
/* Copies 4~7 bytes. */
L(tail4):
bf 29,L(tail2)
lwz r6,0(r4)
stw r6,0(r11)
bf 30,L(tail5)
lhz r7,4(r4)
sth r7,4(r11)
bflr 31
lbz r8,6(r4)
stb r8,6(r11)
blr
.align 4
/* Copies 2~3 bytes. */
L(tail2):
bf 30,1f
lhz r6,0(r4)
sth r6,0(r11)
bflr 31
lbz r7,2(r4)
stb r7,2(r11)
blr
.align 4
L(tail5):
bf 31,1f
lbz r6,4(r4)
stb r6,4(r11)
blr
.align 4
1:
bflr 31
lbz r6,0(r4)
stb r6,0(r11)
blr
/* Handles copies of 0~8 bytes. */
.align 4
L(copy_LE_8):
bne cr6,L(tail4)
ld r6,0(r4)
std r6,0(r11)
blr
END (FUNC_NAME)
#ifndef USE_AS_STPCPY
libc_hidden_builtin_def (strcpy)
#endif