glibc/sysdeps/x86_64/multiarch/memset-vec-unaligned-erms.S
Noah Goldstein 46b5e98ef6 x86: Add seperate non-temporal tunable for memset
The tuning for non-temporal stores for memset vs memcpy is not always
the same. This includes both the exact value and whether non-temporal
stores are profitable at all for a given arch.

This patch add `x86_memset_non_temporal_threshold`. Currently we
disable non-temporal stores for non Intel vendors as the only
benchmarks showing its benefit have been on Intel hardware.
Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
2024-05-30 12:36:09 -05:00

460 lines
12 KiB
ArmAsm

/* memset with unaligned store and rep stosb
Copyright (C) 2016-2024 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 1 * VEC_SIZE with
4 VEC stores and store 4 * VEC at a time until done.
6. On machines ERMS feature, if size is range
[__x86_rep_stosb_threshold, __x86_memset_non_temporal_threshold)
then REP STOSB will be used.
7. If size >= __x86_memset_non_temporal_threshold, use a
non-temporal stores. */
#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 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
# define MOVD vmovd
# else
# define MOVQ movq
# define MOVD movd
# 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
# define LESS_VEC_REG rax
#else
# define END_REG rdi
# define LOOP_REG rdx
# define LESS_VEC_REG rdi
#endif
#ifdef USE_XMM_LESS_VEC
# define XMM_SMALL 1
#else
# define XMM_SMALL 0
#endif
#ifdef USE_LESS_VEC_MASK_STORE
# define SET_REG64 rcx
# define SET_REG32 ecx
# define SET_REG16 cx
# define SET_REG8 cl
#else
# define SET_REG64 rsi
# define SET_REG32 esi
# define SET_REG16 si
# define SET_REG8 sil
#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 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_SET_VEC0_AND_SET_RETURN (%esi, %rdi)
WMEMSET_VDUP_TO_VEC0_LOW()
cmpq $VEC_SIZE, %rdx
jb L(less_vec_from_wmemset)
WMEMSET_VDUP_TO_VEC0_HIGH()
jmp L(entry_from_wmemset)
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_SET_VEC0_AND_SET_RETURN (%esi, %rdi)
# ifdef __ILP32__
/* Clear the upper 32 bits. */
mov %edx, %edx
# endif
cmpq $VEC_SIZE, %rdx
jb L(less_vec)
MEMSET_VDUP_TO_VEC0_HIGH()
L(entry_from_wmemset):
cmpq $(VEC_SIZE * 2), %rdx
ja L(more_2x_vec)
/* From VEC and to 2 * VEC. No branch when size == VEC_SIZE. */
VMOVU %VMM(0), -VEC_SIZE(%rdi,%rdx)
VMOVU %VMM(0), (%rdi)
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
/* Align to ensure the L(less_vec) logic all fits in 1x cache lines. */
.p2align 6,, 47
.p2align 4
L(less_vec):
L(less_vec_from_wmemset):
/* 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 suppress. */
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 %VMM(0), (%rax){%k1}
VZEROUPPER_RETURN
#endif
#if defined USE_MULTIARCH && IS_IN (libc)
END (MEMSET_SYMBOL (__memset, unaligned))
# 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_SET_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)
MEMSET_VDUP_TO_VEC0_HIGH ()
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 %VMM(0), (%rdi)
VMOVU %VMM(0), (VEC_SIZE * -1)(%rdi, %rdx)
VZEROUPPER_RETURN
#endif
.p2align 4,, 4
L(last_2x_vec):
#ifdef USE_LESS_VEC_MASK_STORE
VMOVU %VMM(0), (VEC_SIZE * -2)(%rdi, %rdx)
VMOVU %VMM(0), (VEC_SIZE * -1)(%rdi, %rdx)
#else
VMOVU %VMM(0), (VEC_SIZE * -2)(%rdi)
VMOVU %VMM(0), (VEC_SIZE * -1)(%rdi)
#endif
VZEROUPPER_RETURN
#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):
/* Store next 2x vec regardless. */
VMOVU %VMM(0), (%rdi)
VMOVU %VMM(0), (VEC_SIZE * 1)(%rdi)
/* 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 AVX2/SSE2 compute END_REG (rdi) with ALU. */
addq %rdx, %END_REG
#endif
cmpq $(VEC_SIZE * 4), %rdx
jbe L(last_2x_vec)
#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
/* Store next 2x vec regardless. */
VMOVU %VMM(0), (VEC_SIZE * 2)(%rax)
VMOVU %VMM(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 * -1), %LOOP_REG
.p2align 4
L(loop):
VMOVA %VMM(0), LOOP_4X_OFFSET(%LOOP_REG)
VMOVA %VMM(0), (VEC_SIZE + LOOP_4X_OFFSET)(%LOOP_REG)
VMOVA %VMM(0), (VEC_SIZE * 2 + LOOP_4X_OFFSET)(%LOOP_REG)
VMOVA %VMM(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 %VMM(0), LOOP_4X_OFFSET(%END_REG)
VMOVU %VMM(0), (VEC_SIZE + LOOP_4X_OFFSET)(%END_REG)
VMOVU %VMM(0), (VEC_SIZE * 2 + LOOP_4X_OFFSET)(%END_REG)
VMOVU %VMM(0), (VEC_SIZE * 3 + LOOP_4X_OFFSET)(%END_REG)
L(return_vzeroupper):
#if VEC_SIZE > 16
ZERO_UPPER_VEC_REGISTERS_RETURN
#else
ret
#endif
#ifdef USE_WITH_AVX2
.p2align 4
#else
.p2align 4,, 4
#endif
#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):
cmp __x86_memset_non_temporal_threshold(%rip), %RDX_LP
jae L(nt_memset)
movzbl %sil, %eax
mov %RDX_LP, %RCX_LP
mov %RDI_LP, %RDX_LP
rep stosb
# if (defined USE_WITH_SSE2) || (defined USE_WITH_AVX512)
/* Use xchg to save 1-byte (this helps align targets below). */
xchg %RDX_LP, %RAX_LP
# else
mov %RDX_LP, %RAX_LP
# endif
VZEROUPPER_RETURN
#endif
#ifndef USE_LESS_VEC_MASK_STORE
/* Define L(less_vec) only if not otherwise defined. */
.p2align 4,, 12
L(less_vec):
/* Broadcast esi to partial register (i.e VEC_SIZE == 32 broadcast to
xmm). This is only does anything for AVX2. */
MEMSET_VDUP_TO_VEC0_LOW ()
L(less_vec_from_wmemset):
#endif
L(cross_page):
#if VEC_SIZE > 32
cmpl $32, %edx
jge L(between_32_63)
#endif
#if VEC_SIZE > 16
cmpl $16, %edx
jge L(between_16_31)
#endif
#ifndef USE_XMM_LESS_VEC
MOVQ %VMM_128(0), %SET_REG64
#endif
cmpl $8, %edx
jge L(between_8_15)
cmpl $4, %edx
jge L(between_4_7)
cmpl $1, %edx
jg L(between_2_3)
jl L(between_0_0)
movb %SET_REG8, (%LESS_VEC_REG)
L(between_0_0):
ret
/* 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 %VMM_256(0), (%LESS_VEC_REG)
VMOVU %VMM_256(0), -32(%LESS_VEC_REG, %rdx)
VZEROUPPER_RETURN
#endif
#if VEC_SIZE >= 32
.p2align 4,, SMALL_MEMSET_ALIGN(MOV_SIZE, 1)
L(between_16_31):
/* From 16 to 31. No branch when size == 16. */
VMOVU %VMM_128(0), (%LESS_VEC_REG)
VMOVU %VMM_128(0), -16(%LESS_VEC_REG, %rdx)
ret
#endif
/* Move size is 3 for SSE2, EVEX, and AVX512. Move size is 4 for AVX2.
*/
.p2align 4,, SMALL_MEMSET_ALIGN(3 + XMM_SMALL, 1)
L(between_8_15):
/* From 8 to 15. No branch when size == 8. */
#ifdef USE_XMM_LESS_VEC
MOVQ %VMM_128(0), (%rdi)
MOVQ %VMM_128(0), -8(%rdi, %rdx)
#else
movq %SET_REG64, (%LESS_VEC_REG)
movq %SET_REG64, -8(%LESS_VEC_REG, %rdx)
#endif
ret
/* Move size is 2 for SSE2, EVEX, and AVX512. Move size is 4 for AVX2.
*/
.p2align 4,, SMALL_MEMSET_ALIGN(2 << XMM_SMALL, 1)
L(between_4_7):
/* From 4 to 7. No branch when size == 4. */
#ifdef USE_XMM_LESS_VEC
MOVD %VMM_128(0), (%rdi)
MOVD %VMM_128(0), -4(%rdi, %rdx)
#else
movl %SET_REG32, (%LESS_VEC_REG)
movl %SET_REG32, -4(%LESS_VEC_REG, %rdx)
#endif
ret
/* 4 * XMM_SMALL for the third mov for AVX2. */
.p2align 4,, 4 * XMM_SMALL + SMALL_MEMSET_ALIGN(3, 1)
L(between_2_3):
/* From 2 to 3. No branch when size == 2. */
#ifdef USE_XMM_LESS_VEC
movb %SET_REG8, (%rdi)
movb %SET_REG8, 1(%rdi)
movb %SET_REG8, -1(%rdi, %rdx)
#else
movw %SET_REG16, (%LESS_VEC_REG)
movb %SET_REG8, -1(%LESS_VEC_REG, %rdx)
#endif
ret
#if defined USE_MULTIARCH && IS_IN (libc)
# ifdef USE_WITH_AVX512
/* Force align so the loop doesn't cross a cache-line. */
.p2align 4
# endif
.p2align 4,, 7
/* Memset using non-temporal stores. */
L(nt_memset):
VMOVU %VMM(0), (VEC_SIZE * 0)(%rdi)
leaq (VEC_SIZE * -4)(%rdi, %rdx), %rdx
/* Align DST. */
orq $(VEC_SIZE * 1 - 1), %rdi
incq %rdi
.p2align 4,, 7
L(nt_loop):
VMOVNT %VMM(0), (VEC_SIZE * 0)(%rdi)
VMOVNT %VMM(0), (VEC_SIZE * 1)(%rdi)
VMOVNT %VMM(0), (VEC_SIZE * 2)(%rdi)
VMOVNT %VMM(0), (VEC_SIZE * 3)(%rdi)
subq $(VEC_SIZE * -4), %rdi
cmpq %rdx, %rdi
jb L(nt_loop)
sfence
VMOVU %VMM(0), (VEC_SIZE * 0)(%rdx)
VMOVU %VMM(0), (VEC_SIZE * 1)(%rdx)
VMOVU %VMM(0), (VEC_SIZE * 2)(%rdx)
VMOVU %VMM(0), (VEC_SIZE * 3)(%rdx)
VZEROUPPER_RETURN
#endif
END(MEMSET_SYMBOL(__memset, unaligned_erms))