glibc/sysdeps/x86_64/multiarch/memset-vec-unaligned-erms.S
Noah Goldstein e59ced2384 x86: Optimize memset-vec-unaligned-erms.S
No bug.

Optimization are

1. change control flow for L(more_2x_vec) to fall through to loop and
   jump for L(less_4x_vec) and L(less_8x_vec). This uses less code
   size and saves jumps for length > 4x VEC_SIZE.

2. For EVEX/AVX512 move L(less_vec) closer to entry.

3. Avoid complex address mode for length > 2x VEC_SIZE

4. Slightly better aligning code for the loop from the perspective of
   code size and uops.

5. Align targets so they make full use of their fetch block and if
   possible cache line.

6. Try and reduce total number of icache lines that will need to be
   pulled in for a given length.

7. Include "local" version of stosb target. For AVX2/EVEX/AVX512
   jumping to the stosb target in the sse2 code section will almost
   certainly be to a new page. The new version does increase code size
   marginally by duplicating the target but should get better iTLB
   behavior as a result.

test-memset, test-wmemset, and test-bzero are all passing.

Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com>
Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
2021-10-12 13:38:02 -05:00

420 lines
11 KiB
ArmAsm

/* memset/bzero with unaligned store and rep stosb
Copyright (C) 2016-2021 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/>. */
/* memset is implemented as:
1. Use overlapping store to avoid branch.
2. If size is less than VEC, use integer register stores.
3. If size is from VEC_SIZE to 2 * VEC_SIZE, use 2 VEC stores.
4. If size is from 2 * VEC_SIZE to 4 * VEC_SIZE, use 4 VEC stores.
5. If size is more to 4 * VEC_SIZE, align to 4 * VEC_SIZE with
4 VEC stores and store 4 * VEC at a time until done. */
#include <sysdep.h>
#ifndef MEMSET_CHK_SYMBOL
# define MEMSET_CHK_SYMBOL(p,s) MEMSET_SYMBOL(p, s)
#endif
#ifndef WMEMSET_CHK_SYMBOL
# define WMEMSET_CHK_SYMBOL(p,s) WMEMSET_SYMBOL(p, s)
#endif
#ifndef XMM0
# define XMM0 xmm0
#endif
#ifndef YMM0
# define YMM0 ymm0
#endif
#ifndef VZEROUPPER
# if VEC_SIZE > 16
# define VZEROUPPER vzeroupper
# define VZEROUPPER_SHORT_RETURN vzeroupper; ret
# else
# define VZEROUPPER
# endif
#endif
#ifndef VZEROUPPER_SHORT_RETURN
# define VZEROUPPER_SHORT_RETURN rep; ret
#endif
#ifndef MOVQ
# if VEC_SIZE > 16
# define MOVQ vmovq
# else
# define MOVQ movq
# endif
#endif
#if VEC_SIZE == 64
# define LOOP_4X_OFFSET (VEC_SIZE * 4)
#else
# define LOOP_4X_OFFSET (0)
#endif
#if defined USE_WITH_EVEX || defined USE_WITH_AVX512
# define END_REG rcx
# define LOOP_REG rdi
#else
# define END_REG rdi
# define LOOP_REG rdx
#endif
#define PAGE_SIZE 4096
/* Macro to calculate size of small memset block for aligning
purposes. */
#define SMALL_MEMSET_ALIGN(mov_sz, ret_sz) (2 * (mov_sz) + (ret_sz) + 1)
#ifndef SECTION
# error SECTION is not defined!
#endif
.section SECTION(.text),"ax",@progbits
#if VEC_SIZE == 16 && IS_IN (libc)
ENTRY (__bzero)
mov %RDI_LP, %RAX_LP /* Set return value. */
mov %RSI_LP, %RDX_LP /* Set n. */
xorl %esi, %esi
pxor %XMM0, %XMM0
jmp L(entry_from_bzero)
END (__bzero)
weak_alias (__bzero, bzero)
#endif
#if IS_IN (libc)
# if defined SHARED
ENTRY_CHK (WMEMSET_CHK_SYMBOL (__wmemset_chk, unaligned))
cmp %RDX_LP, %RCX_LP
jb HIDDEN_JUMPTARGET (__chk_fail)
END_CHK (WMEMSET_CHK_SYMBOL (__wmemset_chk, unaligned))
# endif
ENTRY (WMEMSET_SYMBOL (__wmemset, unaligned))
shl $2, %RDX_LP
WMEMSET_VDUP_TO_VEC0_AND_SET_RETURN (%esi, %rdi)
jmp L(entry_from_bzero)
END (WMEMSET_SYMBOL (__wmemset, unaligned))
#endif
#if defined SHARED && IS_IN (libc)
ENTRY_CHK (MEMSET_CHK_SYMBOL (__memset_chk, unaligned))
cmp %RDX_LP, %RCX_LP
jb HIDDEN_JUMPTARGET (__chk_fail)
END_CHK (MEMSET_CHK_SYMBOL (__memset_chk, unaligned))
#endif
ENTRY (MEMSET_SYMBOL (__memset, unaligned))
MEMSET_VDUP_TO_VEC0_AND_SET_RETURN (%esi, %rdi)
# ifdef __ILP32__
/* Clear the upper 32 bits. */
mov %edx, %edx
# endif
L(entry_from_bzero):
cmpq $VEC_SIZE, %rdx
jb L(less_vec)
cmpq $(VEC_SIZE * 2), %rdx
ja L(more_2x_vec)
/* From VEC and to 2 * VEC. No branch when size == VEC_SIZE. */
VMOVU %VEC(0), -VEC_SIZE(%rdi,%rdx)
VMOVU %VEC(0), (%rdi)
VZEROUPPER_RETURN
#if defined USE_MULTIARCH && IS_IN (libc)
END (MEMSET_SYMBOL (__memset, unaligned))
# if VEC_SIZE == 16
ENTRY (__memset_chk_erms)
cmp %RDX_LP, %RCX_LP
jb HIDDEN_JUMPTARGET (__chk_fail)
END (__memset_chk_erms)
/* Only used to measure performance of REP STOSB. */
ENTRY (__memset_erms)
/* Skip zero length. */
test %RDX_LP, %RDX_LP
jnz L(stosb)
movq %rdi, %rax
ret
# else
/* Provide a hidden symbol to debugger. */
.hidden MEMSET_SYMBOL (__memset, erms)
ENTRY (MEMSET_SYMBOL (__memset, erms))
# endif
L(stosb):
mov %RDX_LP, %RCX_LP
movzbl %sil, %eax
mov %RDI_LP, %RDX_LP
rep stosb
mov %RDX_LP, %RAX_LP
VZEROUPPER_RETURN
# if VEC_SIZE == 16
END (__memset_erms)
# else
END (MEMSET_SYMBOL (__memset, erms))
# endif
# if defined SHARED && IS_IN (libc)
ENTRY_CHK (MEMSET_CHK_SYMBOL (__memset_chk, unaligned_erms))
cmp %RDX_LP, %RCX_LP
jb HIDDEN_JUMPTARGET (__chk_fail)
END_CHK (MEMSET_CHK_SYMBOL (__memset_chk, unaligned_erms))
# endif
ENTRY_P2ALIGN (MEMSET_SYMBOL (__memset, unaligned_erms), 6)
MEMSET_VDUP_TO_VEC0_AND_SET_RETURN (%esi, %rdi)
# ifdef __ILP32__
/* Clear the upper 32 bits. */
mov %edx, %edx
# endif
cmp $VEC_SIZE, %RDX_LP
jb L(less_vec)
cmp $(VEC_SIZE * 2), %RDX_LP
ja L(stosb_more_2x_vec)
/* From VEC and to 2 * VEC. No branch when size == VEC_SIZE.
*/
VMOVU %VEC(0), (%rax)
VMOVU %VEC(0), -VEC_SIZE(%rax, %rdx)
VZEROUPPER_RETURN
#endif
.p2align 4,, 10
L(last_2x_vec):
#ifdef USE_LESS_VEC_MASK_STORE
VMOVU %VEC(0), (VEC_SIZE * 2 + LOOP_4X_OFFSET)(%rcx)
VMOVU %VEC(0), (VEC_SIZE * 3 + LOOP_4X_OFFSET)(%rcx)
#else
VMOVU %VEC(0), (VEC_SIZE * -2)(%rdi)
VMOVU %VEC(0), (VEC_SIZE * -1)(%rdi)
#endif
VZEROUPPER_RETURN
/* If have AVX512 mask instructions put L(less_vec) close to
entry as it doesn't take much space and is likely a hot target.
*/
#ifdef USE_LESS_VEC_MASK_STORE
.p2align 4,, 10
L(less_vec):
/* Less than 1 VEC. */
# if VEC_SIZE != 16 && VEC_SIZE != 32 && VEC_SIZE != 64
# error Unsupported VEC_SIZE!
# endif
/* Clear high bits from edi. Only keeping bits relevant to page
cross check. Note that we are using rax which is set in
MEMSET_VDUP_TO_VEC0_AND_SET_RETURN as ptr from here on out. */
andl $(PAGE_SIZE - 1), %edi
/* Check if VEC_SIZE store cross page. Mask stores suffer
serious performance degradation when it has to fault supress.
*/
cmpl $(PAGE_SIZE - VEC_SIZE), %edi
/* This is generally considered a cold target. */
ja L(cross_page)
# if VEC_SIZE > 32
movq $-1, %rcx
bzhiq %rdx, %rcx, %rcx
kmovq %rcx, %k1
# else
movl $-1, %ecx
bzhil %edx, %ecx, %ecx
kmovd %ecx, %k1
# endif
vmovdqu8 %VEC(0), (%rax){%k1}
VZEROUPPER_RETURN
# if defined USE_MULTIARCH && IS_IN (libc)
/* Include L(stosb_local) here if including L(less_vec) between
L(stosb_more_2x_vec) and ENTRY. This is to cache align the
L(stosb_more_2x_vec) target. */
.p2align 4,, 10
L(stosb_local):
movzbl %sil, %eax
mov %RDX_LP, %RCX_LP
mov %RDI_LP, %RDX_LP
rep stosb
mov %RDX_LP, %RAX_LP
VZEROUPPER_RETURN
# endif
#endif
#if defined USE_MULTIARCH && IS_IN (libc)
.p2align 4
L(stosb_more_2x_vec):
cmp __x86_rep_stosb_threshold(%rip), %RDX_LP
ja L(stosb_local)
#endif
/* Fallthrough goes to L(loop_4x_vec). Tests for memset (2x, 4x]
and (4x, 8x] jump to target. */
L(more_2x_vec):
/* Two different methods of setting up pointers / compare. The
two methods are based on the fact that EVEX/AVX512 mov
instructions take more bytes then AVX2/SSE2 mov instructions. As
well that EVEX/AVX512 machines also have fast LEA_BID. Both
setup and END_REG to avoid complex address mode. For EVEX/AVX512
this saves code size and keeps a few targets in one fetch block.
For AVX2/SSE2 this helps prevent AGU bottlenecks. */
#if defined USE_WITH_EVEX || defined USE_WITH_AVX512
/* If EVEX/AVX512 compute END_REG - (VEC_SIZE * 4 +
LOOP_4X_OFFSET) with LEA_BID. */
/* END_REG is rcx for EVEX/AVX512. */
leaq -(VEC_SIZE * 4 + LOOP_4X_OFFSET)(%rdi, %rdx), %END_REG
#endif
/* Stores to first 2x VEC before cmp as any path forward will
require it. */
VMOVU %VEC(0), (%rax)
VMOVU %VEC(0), VEC_SIZE(%rax)
#if !(defined USE_WITH_EVEX || defined USE_WITH_AVX512)
/* If AVX2/SSE2 compute END_REG (rdi) with ALU. */
addq %rdx, %END_REG
#endif
cmpq $(VEC_SIZE * 4), %rdx
jbe L(last_2x_vec)
/* Store next 2x vec regardless. */
VMOVU %VEC(0), (VEC_SIZE * 2)(%rax)
VMOVU %VEC(0), (VEC_SIZE * 3)(%rax)
#if defined USE_WITH_EVEX || defined USE_WITH_AVX512
/* If LOOP_4X_OFFSET don't readjust LOOP_REG (rdi), just add
extra offset to addresses in loop. Used for AVX512 to save space
as no way to get (VEC_SIZE * 4) in imm8. */
# if LOOP_4X_OFFSET == 0
subq $-(VEC_SIZE * 4), %LOOP_REG
# endif
/* Avoid imm32 compare here to save code size. */
cmpq %rdi, %rcx
#else
addq $-(VEC_SIZE * 4), %END_REG
cmpq $(VEC_SIZE * 8), %rdx
#endif
jbe L(last_4x_vec)
#if !(defined USE_WITH_EVEX || defined USE_WITH_AVX512)
/* Set LOOP_REG (rdx). */
leaq (VEC_SIZE * 4)(%rax), %LOOP_REG
#endif
/* Align dst for loop. */
andq $(VEC_SIZE * -2), %LOOP_REG
.p2align 4
L(loop):
VMOVA %VEC(0), LOOP_4X_OFFSET(%LOOP_REG)
VMOVA %VEC(0), (VEC_SIZE + LOOP_4X_OFFSET)(%LOOP_REG)
VMOVA %VEC(0), (VEC_SIZE * 2 + LOOP_4X_OFFSET)(%LOOP_REG)
VMOVA %VEC(0), (VEC_SIZE * 3 + LOOP_4X_OFFSET)(%LOOP_REG)
subq $-(VEC_SIZE * 4), %LOOP_REG
cmpq %END_REG, %LOOP_REG
jb L(loop)
.p2align 4,, MOV_SIZE
L(last_4x_vec):
VMOVU %VEC(0), LOOP_4X_OFFSET(%END_REG)
VMOVU %VEC(0), (VEC_SIZE + LOOP_4X_OFFSET)(%END_REG)
VMOVU %VEC(0), (VEC_SIZE * 2 + LOOP_4X_OFFSET)(%END_REG)
VMOVU %VEC(0), (VEC_SIZE * 3 + LOOP_4X_OFFSET)(%END_REG)
L(return):
#if VEC_SIZE > 16
ZERO_UPPER_VEC_REGISTERS_RETURN
#else
ret
#endif
.p2align 4,, 10
#ifndef USE_LESS_VEC_MASK_STORE
# if defined USE_MULTIARCH && IS_IN (libc)
/* If no USE_LESS_VEC_MASK put L(stosb_local) here. Will be in
range for 2-byte jump encoding. */
L(stosb_local):
movzbl %sil, %eax
mov %RDX_LP, %RCX_LP
mov %RDI_LP, %RDX_LP
rep stosb
mov %RDX_LP, %RAX_LP
VZEROUPPER_RETURN
# endif
/* Define L(less_vec) only if not otherwise defined. */
.p2align 4
L(less_vec):
#endif
L(cross_page):
#if VEC_SIZE > 32
cmpl $32, %edx
jae L(between_32_63)
#endif
#if VEC_SIZE > 16
cmpl $16, %edx
jae L(between_16_31)
#endif
MOVQ %XMM0, %rdi
cmpl $8, %edx
jae L(between_8_15)
cmpl $4, %edx
jae L(between_4_7)
cmpl $1, %edx
ja L(between_2_3)
jb L(return)
movb %sil, (%rax)
VZEROUPPER_RETURN
/* Align small targets only if not doing so would cross a fetch
line. */
#if VEC_SIZE > 32
.p2align 4,, SMALL_MEMSET_ALIGN(MOV_SIZE, RET_SIZE)
/* From 32 to 63. No branch when size == 32. */
L(between_32_63):
VMOVU %YMM0, (%rax)
VMOVU %YMM0, -32(%rax, %rdx)
VZEROUPPER_RETURN
#endif
#if VEC_SIZE >= 32
.p2align 4,, SMALL_MEMSET_ALIGN(MOV_SIZE, RET_SIZE)
L(between_16_31):
/* From 16 to 31. No branch when size == 16. */
VMOVU %XMM0, (%rax)
VMOVU %XMM0, -16(%rax, %rdx)
VZEROUPPER_RETURN
#endif
.p2align 4,, SMALL_MEMSET_ALIGN(3, RET_SIZE)
L(between_8_15):
/* From 8 to 15. No branch when size == 8. */
movq %rdi, (%rax)
movq %rdi, -8(%rax, %rdx)
VZEROUPPER_RETURN
.p2align 4,, SMALL_MEMSET_ALIGN(2, RET_SIZE)
L(between_4_7):
/* From 4 to 7. No branch when size == 4. */
movl %edi, (%rax)
movl %edi, -4(%rax, %rdx)
VZEROUPPER_RETURN
.p2align 4,, SMALL_MEMSET_ALIGN(3, RET_SIZE)
L(between_2_3):
/* From 2 to 3. No branch when size == 2. */
movw %di, (%rax)
movb %dil, -1(%rax, %rdx)
VZEROUPPER_RETURN
END (MEMSET_SYMBOL (__memset, unaligned_erms))