mirror/glibc
2
0
Fork 0
mirror of git://sourceware.org/git/glibc.git synced 2024-12-09 04:11:27 +08:00
Code Issues Packages Projects Releases Wiki Activity

x86: Optimize memrchr-evex.S

Browse Source 
The new code:
    1. prioritizes smaller user-arg lengths more.
    2. optimizes target placement more carefully
    3. reuses logic more
    4. fixes up various inefficiencies in the logic. The biggest
       case here is the `lzcnt` logic for checking returns which
       saves either a branch or multiple instructions.

The total code size saving is: 263 bytes
Geometric Mean of all benchmarks New / Old: 0.755

Regressions:
There are some regressions. Particularly where the length (user arg
length) is large but the position of the match char is near the
beginning of the string (in first VEC). This case has roughly a
20% regression.

This is because the new logic gives the hot path for immediate matches
to shorter lengths (the more common input). This case has roughly
a 35% speedup.

Full xcheck passes on x86_64.
Reviewed-by: H.J. Lu <hjl.tools@gmail.com>

(cherry picked from commit b4209615a0)
This commit is contained in:
Noah Goldstein 2022-06-06 21:11:31 -07:00 committed by Sunil K Pandey
parent 4901009dad
commit 83a986e9fb
1 changed files with 286 additions and 289 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

575
sysdeps/x86_64/multiarch/memrchr-evex.S
View File

@ -19,319 +19,316 @@
#if IS_IN (libc) #if IS_IN (libc)
# include <sysdep.h> # include <sysdep.h>
# include "evex256-vecs.h"
# if VEC_SIZE != 32
# error "VEC_SIZE != 32 unimplemented"
# endif
# define VMOVA vmovdqa64 # ifndef MEMRCHR
# define MEMRCHR __memrchr_evex
# endif
# define YMMMATCH ymm16 # define PAGE_SIZE 4096
# define VECMATCH VEC(0)
# define VEC_SIZE 32 .section SECTION(.text), "ax", @progbits
ENTRY_P2ALIGN(MEMRCHR, 6)
# ifdef __ILP32__
/* Clear upper bits. */
and %RDX_LP, %RDX_LP
# else
test %RDX_LP, %RDX_LP
# endif
jz L(zero_0)
.section .text.evex,"ax",@progbits /* Get end pointer. Minus one for two reasons. 1) It is necessary for a
ENTRY (__memrchr_evex) correct page cross check and 2) it correctly sets up end ptr to be
/* Broadcast CHAR to YMMMATCH. */ subtract by lzcnt aligned. */
vpbroadcastb %esi, %YMMMATCH leaq -1(%rdi, %rdx), %rax
vpbroadcastb %esi, %VECMATCH
sub $VEC_SIZE, %RDX_LP /* Check if we can load 1x VEC without cross a page. */
jbe L(last_vec_or_less) testl $(PAGE_SIZE - VEC_SIZE), %eax
jz L(page_cross)
add %RDX_LP, %RDI_LP /* Don't use rax for pointer here because EVEX has better encoding with
offset % VEC_SIZE == 0. */
vpcmpb $0, -(VEC_SIZE)(%rdi, %rdx), %VECMATCH, %k0
kmovd %k0, %ecx
/* Check the last VEC_SIZE bytes. */ /* Fall through for rdx (len) <= VEC_SIZE (expect small sizes). */
vpcmpb $0, (%rdi), %YMMMATCH, %k1 cmpq $VEC_SIZE, %rdx
kmovd %k1, %eax ja L(more_1x_vec)
testl %eax, %eax L(ret_vec_x0_test):
jnz L(last_vec_x0)
subq $(VEC_SIZE * 4), %rdi /* If ecx is zero (no matches) lzcnt will set it 32 (VEC_SIZE) which
movl %edi, %ecx will guarantee edx (len) is less than it. */
andl $(VEC_SIZE - 1), %ecx lzcntl %ecx, %ecx
jz L(aligned_more) cmpl %ecx, %edx
jle L(zero_0)
/* Align data for aligned loads in the loop. */ subq %rcx, %rax
addq $VEC_SIZE, %rdi
addq $VEC_SIZE, %rdx
andq $-VEC_SIZE, %rdi
subq %rcx, %rdx
.p2align 4
L(aligned_more):
subq $(VEC_SIZE * 4), %rdx
jbe L(last_4x_vec_or_less)
/* Check the last 4 * VEC_SIZE. Only one VEC_SIZE at a time
since data is only aligned to VEC_SIZE. */
vpcmpb $0, (VEC_SIZE * 3)(%rdi), %YMMMATCH, %k1
kmovd %k1, %eax
testl %eax, %eax
jnz L(last_vec_x3)
vpcmpb $0, (VEC_SIZE * 2)(%rdi), %YMMMATCH, %k2
kmovd %k2, %eax
testl %eax, %eax
jnz L(last_vec_x2)
vpcmpb $0, VEC_SIZE(%rdi), %YMMMATCH, %k3
kmovd %k3, %eax
testl %eax, %eax
jnz L(last_vec_x1)
vpcmpb $0, (%rdi), %YMMMATCH, %k4
kmovd %k4, %eax
testl %eax, %eax
jnz L(last_vec_x0)
/* Align data to 4 * VEC_SIZE for loop with fewer branches.
There are some overlaps with above if data isn't aligned
to 4 * VEC_SIZE. */
movl %edi, %ecx
andl $(VEC_SIZE * 4 - 1), %ecx
jz L(loop_4x_vec)
addq $(VEC_SIZE * 4), %rdi
addq $(VEC_SIZE * 4), %rdx
andq $-(VEC_SIZE * 4), %rdi
subq %rcx, %rdx
.p2align 4
L(loop_4x_vec):
/* Compare 4 * VEC at a time forward. */
subq $(VEC_SIZE * 4), %rdi
subq $(VEC_SIZE * 4), %rdx
jbe L(last_4x_vec_or_less)
vpcmpb $0, (%rdi), %YMMMATCH, %k1
vpcmpb $0, VEC_SIZE(%rdi), %YMMMATCH, %k2
kord %k1, %k2, %k5
vpcmpb $0, (VEC_SIZE * 2)(%rdi), %YMMMATCH, %k3
vpcmpb $0, (VEC_SIZE * 3)(%rdi), %YMMMATCH, %k4
kord %k3, %k4, %k6
kortestd %k5, %k6
jz L(loop_4x_vec)
/* There is a match. */
kmovd %k4, %eax
testl %eax, %eax
jnz L(last_vec_x3)
kmovd %k3, %eax
testl %eax, %eax
jnz L(last_vec_x2)
kmovd %k2, %eax
testl %eax, %eax
jnz L(last_vec_x1)
kmovd %k1, %eax
bsrl %eax, %eax
addq %rdi, %rax
ret ret
.p2align 4 /* Fits in aligning bytes of first cache line. */
L(last_4x_vec_or_less): L(zero_0):
addl $(VEC_SIZE * 4), %edx
cmpl $(VEC_SIZE * 2), %edx
jbe L(last_2x_vec)
vpcmpb $0, (VEC_SIZE * 3)(%rdi), %YMMMATCH, %k1
kmovd %k1, %eax
testl %eax, %eax
jnz L(last_vec_x3)
vpcmpb $0, (VEC_SIZE * 2)(%rdi), %YMMMATCH, %k2
kmovd %k2, %eax
testl %eax, %eax
jnz L(last_vec_x2)
vpcmpb $0, VEC_SIZE(%rdi), %YMMMATCH, %k3
kmovd %k3, %eax
testl %eax, %eax
jnz L(last_vec_x1_check)
cmpl $(VEC_SIZE * 3), %edx
jbe L(zero)
vpcmpb $0, (%rdi), %YMMMATCH, %k4
kmovd %k4, %eax
testl %eax, %eax
jz L(zero)
bsrl %eax, %eax
subq $(VEC_SIZE * 4), %rdx
addq %rax, %rdx
jl L(zero)
addq %rdi, %rax
ret
.p2align 4
L(last_2x_vec):
vpcmpb $0, (VEC_SIZE * 3)(%rdi), %YMMMATCH, %k1
kmovd %k1, %eax
testl %eax, %eax
jnz L(last_vec_x3_check)
cmpl $VEC_SIZE, %edx
jbe L(zero)
vpcmpb $0, (VEC_SIZE * 2)(%rdi), %YMMMATCH, %k1
kmovd %k1, %eax
testl %eax, %eax
jz L(zero)
bsrl %eax, %eax
subq $(VEC_SIZE * 2), %rdx
addq %rax, %rdx
jl L(zero)
addl $(VEC_SIZE * 2), %eax
addq %rdi, %rax
ret
.p2align 4
L(last_vec_x0):
bsrl %eax, %eax
addq %rdi, %rax
ret
.p2align 4
L(last_vec_x1):
bsrl %eax, %eax
addl $VEC_SIZE, %eax
addq %rdi, %rax
ret
.p2align 4
L(last_vec_x2):
bsrl %eax, %eax
addl $(VEC_SIZE * 2), %eax
addq %rdi, %rax
ret
.p2align 4
L(last_vec_x3):
bsrl %eax, %eax
addl $(VEC_SIZE * 3), %eax
addq %rdi, %rax
ret
.p2align 4
L(last_vec_x1_check):
bsrl %eax, %eax
subq $(VEC_SIZE * 3), %rdx
addq %rax, %rdx
jl L(zero)
addl $VEC_SIZE, %eax
addq %rdi, %rax
ret
.p2align 4
L(last_vec_x3_check):
bsrl %eax, %eax
subq $VEC_SIZE, %rdx
addq %rax, %rdx
jl L(zero)
addl $(VEC_SIZE * 3), %eax
addq %rdi, %rax
ret
.p2align 4
L(zero):
xorl %eax, %eax xorl %eax, %eax
ret ret
.p2align 4 .p2align 4,, 9
L(last_vec_or_less_aligned): L(ret_vec_x0_dec):
movl %edx, %ecx decq %rax
L(ret_vec_x0):
vpcmpb $0, (%rdi), %YMMMATCH, %k1 lzcntl %ecx, %ecx
subq %rcx, %rax
movl $1, %edx
/* Support rdx << 32. */
salq %cl, %rdx
subq $1, %rdx
kmovd %k1, %eax
/* Remove the trailing bytes. */
andl %edx, %eax
testl %eax, %eax
jz L(zero)
bsrl %eax, %eax
addq %rdi, %rax
ret ret
.p2align 4 .p2align 4,, 10
L(last_vec_or_less): L(more_1x_vec):
addl $VEC_SIZE, %edx testl %ecx, %ecx
jnz L(ret_vec_x0)
/* Check for zero length. */ /* Align rax (pointer to string). */
testl %edx, %edx andq $-VEC_SIZE, %rax
jz L(zero)
movl %edi, %ecx /* Recompute length after aligning. */
andl $(VEC_SIZE - 1), %ecx movq %rax, %rdx
jz L(last_vec_or_less_aligned)
movl %ecx, %esi /* Need no matter what. */
movl %ecx, %r8d vpcmpb $0, -(VEC_SIZE)(%rax), %VECMATCH, %k0
addl %edx, %esi kmovd %k0, %ecx
andq $-VEC_SIZE, %rdi
subl $VEC_SIZE, %esi subq %rdi, %rdx
ja L(last_vec_2x_aligned)
/* Check the last VEC. */ cmpq $(VEC_SIZE * 2), %rdx
vpcmpb $0, (%rdi), %YMMMATCH, %k1 ja L(more_2x_vec)
kmovd %k1, %eax L(last_2x_vec):
/* Remove the leading and trailing bytes. */ /* Must dec rax because L(ret_vec_x0_test) expects it. */
sarl %cl, %eax decq %rax
movl %edx, %ecx cmpl $VEC_SIZE, %edx
jbe L(ret_vec_x0_test)
movl $1, %edx testl %ecx, %ecx
sall %cl, %edx jnz L(ret_vec_x0)
subl $1, %edx
andl %edx, %eax /* Don't use rax for pointer here because EVEX has better encoding with
testl %eax, %eax offset % VEC_SIZE == 0. */
jz L(zero) vpcmpb $0, -(VEC_SIZE * 2)(%rdi, %rdx), %VECMATCH, %k0
kmovd %k0, %ecx
bsrl %eax, %eax /* NB: 64-bit lzcnt. This will naturally add 32 to position. */
addq %rdi, %rax lzcntq %rcx, %rcx
addq %r8, %rax cmpl %ecx, %edx
jle L(zero_0)
subq %rcx, %rax
ret ret
.p2align 4 /* Inexpensive place to put this regarding code size / target alignments
L(last_vec_2x_aligned): / ICache NLP. Necessary for 2-byte encoding of jump to page cross
movl %esi, %ecx case which in turn is necessary for hot path (len <= VEC_SIZE) to fit
in first cache line. */
/* Check the last VEC. */ L(page_cross):
vpcmpb $0, VEC_SIZE(%rdi), %YMMMATCH, %k1 movq %rax, %rsi
andq $-VEC_SIZE, %rsi
movl $1, %edx vpcmpb $0, (%rsi), %VECMATCH, %k0
sall %cl, %edx kmovd %k0, %r8d
subl $1, %edx /* Shift out negative alignment (because we are starting from endptr and
working backwards). */
kmovd %k1, %eax movl %eax, %ecx
/* notl because eax already has endptr - 1. (-x = ~(x - 1)). */
/* Remove the trailing bytes. */ notl %ecx
andl %edx, %eax shlxl %ecx, %r8d, %ecx
cmpq %rdi, %rsi
testl %eax, %eax ja L(more_1x_vec)
jnz L(last_vec_x1) lzcntl %ecx, %ecx
cmpl %ecx, %edx
/* Check the second last VEC. */ jle L(zero_1)
vpcmpb $0, (%rdi), %YMMMATCH, %k1 subq %rcx, %rax
movl %r8d, %ecx
kmovd %k1, %eax
/* Remove the leading bytes. Must use unsigned right shift for
bsrl below. */
shrl %cl, %eax
testl %eax, %eax
jz L(zero)
bsrl %eax, %eax
addq %rdi, %rax
addq %r8, %rax
ret ret
END (__memrchr_evex)
/* Continue creating zero labels that fit in aligning bytes and get
2-byte encoding / are in the same cache line as condition. */
L(zero_1):
xorl %eax, %eax
ret
.p2align 4,, 8
L(ret_vec_x1):
/* This will naturally add 32 to position. */
bsrl %ecx, %ecx
leaq -(VEC_SIZE * 2)(%rcx, %rax), %rax
ret
.p2align 4,, 8
L(more_2x_vec):
testl %ecx, %ecx
jnz L(ret_vec_x0_dec)
vpcmpb $0, -(VEC_SIZE * 2)(%rax), %VECMATCH, %k0
kmovd %k0, %ecx
testl %ecx, %ecx
jnz L(ret_vec_x1)
/* Need no matter what. */
vpcmpb $0, -(VEC_SIZE * 3)(%rax), %VECMATCH, %k0
kmovd %k0, %ecx
subq $(VEC_SIZE * 4), %rdx
ja L(more_4x_vec)
cmpl $(VEC_SIZE * -1), %edx
jle L(ret_vec_x2_test)
L(last_vec):
testl %ecx, %ecx
jnz L(ret_vec_x2)
/* Need no matter what. */
vpcmpb $0, -(VEC_SIZE * 4)(%rax), %VECMATCH, %k0
kmovd %k0, %ecx
lzcntl %ecx, %ecx
subq $(VEC_SIZE * 3 + 1), %rax
subq %rcx, %rax
cmpq %rax, %rdi
ja L(zero_1)
ret
.p2align 4,, 8
L(ret_vec_x2_test):
lzcntl %ecx, %ecx
subq $(VEC_SIZE * 2 + 1), %rax
subq %rcx, %rax
cmpq %rax, %rdi
ja L(zero_1)
ret
.p2align 4,, 8
L(ret_vec_x2):
bsrl %ecx, %ecx
leaq -(VEC_SIZE * 3)(%rcx, %rax), %rax
ret
.p2align 4,, 8
L(ret_vec_x3):
bsrl %ecx, %ecx
leaq -(VEC_SIZE * 4)(%rcx, %rax), %rax
ret
.p2align 4,, 8
L(more_4x_vec):
testl %ecx, %ecx
jnz L(ret_vec_x2)
vpcmpb $0, -(VEC_SIZE * 4)(%rax), %VECMATCH, %k0
kmovd %k0, %ecx
testl %ecx, %ecx
jnz L(ret_vec_x3)
/* Check if near end before re-aligning (otherwise might do an
unnecessary loop iteration). */
addq $-(VEC_SIZE * 4), %rax
cmpq $(VEC_SIZE * 4), %rdx
jbe L(last_4x_vec)
decq %rax
andq $-(VEC_SIZE * 4), %rax
movq %rdi, %rdx
/* Get endptr for loop in rdx. NB: Can't just do while rax > rdi because
lengths that overflow can be valid and break the comparison. */
andq $-(VEC_SIZE * 4), %rdx
.p2align 4
L(loop_4x_vec):
/* Store 1 were not-equals and 0 where equals in k1 (used to mask later
on). */
vpcmpb $4, (VEC_SIZE * 3)(%rax), %VECMATCH, %k1
/* VEC(2/3) will have zero-byte where we found a CHAR. */
vpxorq (VEC_SIZE * 2)(%rax), %VECMATCH, %VEC(2)
vpxorq (VEC_SIZE * 1)(%rax), %VECMATCH, %VEC(3)
vpcmpb $0, (VEC_SIZE * 0)(%rax), %VECMATCH, %k4
/* Combine VEC(2/3) with min and maskz with k1 (k1 has zero bit where
CHAR is found and VEC(2/3) have zero-byte where CHAR is found. */
vpminub %VEC(2), %VEC(3), %VEC(3){%k1}{z}
vptestnmb %VEC(3), %VEC(3), %k2
/* Any 1s and we found CHAR. */
kortestd %k2, %k4
jnz L(loop_end)
addq $-(VEC_SIZE * 4), %rax
cmpq %rdx, %rax
jne L(loop_4x_vec)
/* Need to re-adjust rdx / rax for L(last_4x_vec). */
subq $-(VEC_SIZE * 4), %rdx
movq %rdx, %rax
subl %edi, %edx
L(last_4x_vec):
/* Used no matter what. */
vpcmpb $0, (VEC_SIZE * -1)(%rax), %VECMATCH, %k0
kmovd %k0, %ecx
cmpl $(VEC_SIZE * 2), %edx
jbe L(last_2x_vec)
testl %ecx, %ecx
jnz L(ret_vec_x0_dec)
vpcmpb $0, (VEC_SIZE * -2)(%rax), %VECMATCH, %k0
kmovd %k0, %ecx
testl %ecx, %ecx
jnz L(ret_vec_x1)
/* Used no matter what. */
vpcmpb $0, (VEC_SIZE * -3)(%rax), %VECMATCH, %k0
kmovd %k0, %ecx
cmpl $(VEC_SIZE * 3), %edx
ja L(last_vec)
lzcntl %ecx, %ecx
subq $(VEC_SIZE * 2 + 1), %rax
subq %rcx, %rax
cmpq %rax, %rdi
jbe L(ret_1)
xorl %eax, %eax
L(ret_1):
ret
.p2align 4,, 6
L(loop_end):
kmovd %k1, %ecx
notl %ecx
testl %ecx, %ecx
jnz L(ret_vec_x0_end)
vptestnmb %VEC(2), %VEC(2), %k0
kmovd %k0, %ecx
testl %ecx, %ecx
jnz L(ret_vec_x1_end)
kmovd %k2, %ecx
kmovd %k4, %esi
/* Combine last 2 VEC matches. If ecx (VEC3) is zero (no CHAR in VEC3)
then it won't affect the result in esi (VEC4). If ecx is non-zero
then CHAR in VEC3 and bsrq will use that position. */
salq $32, %rcx
orq %rsi, %rcx
bsrq %rcx, %rcx
addq %rcx, %rax
ret
.p2align 4,, 4
L(ret_vec_x0_end):
addq $(VEC_SIZE), %rax
L(ret_vec_x1_end):
bsrl %ecx, %ecx
leaq (VEC_SIZE * 2)(%rax, %rcx), %rax
ret
END(MEMRCHR)
#endif #endif