commit b05bd59823bcedee281d3fd5bd4928698ea9d69d
Author: Noah Goldstein <goldstein.w.n@gmail.com>
Date: Mon Jun 6 21:11:32 2022 -0700
x86: Optimize memrchr-avx2.S
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: 306 bytes
Geometric Mean of all benchmarks New / Old: 0.760
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
10-20% 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 15-45% speedup.
Full xcheck passes on x86_64.
Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
(cherry picked from commit af5306a735eb0966fdc2f8ccdafa8888e2df0c87)
diff --git a/sysdeps/x86_64/multiarch/memrchr-avx2-rtm.S b/sysdeps/x86_64/multiarch/memrchr-avx2-rtm.S
index cea2d2a72db7406a..5e9beeeef2677c9f 100644
--- a/sysdeps/x86_64/multiarch/memrchr-avx2-rtm.S
+++ b/sysdeps/x86_64/multiarch/memrchr-avx2-rtm.S
@@ -2,6 +2,7 @@
# define MEMRCHR __memrchr_avx2_rtm
#endif
+#define COND_VZEROUPPER COND_VZEROUPPER_XTEST
#define ZERO_UPPER_VEC_REGISTERS_RETURN \
ZERO_UPPER_VEC_REGISTERS_RETURN_XTEST
diff --git a/sysdeps/x86_64/multiarch/memrchr-avx2.S b/sysdeps/x86_64/multiarch/memrchr-avx2.S
index ac7370cb06e9a0fd..5f8e0be18cfe4fad 100644
--- a/sysdeps/x86_64/multiarch/memrchr-avx2.S
+++ b/sysdeps/x86_64/multiarch/memrchr-avx2.S
@@ -21,340 +21,318 @@
# include <sysdep.h>
# ifndef MEMRCHR
-# define MEMRCHR __memrchr_avx2
+# define MEMRCHR __memrchr_avx2
# endif
# ifndef VZEROUPPER
-# define VZEROUPPER vzeroupper
+# define VZEROUPPER vzeroupper
# endif
# ifndef SECTION
# define SECTION(p) p##.avx
# endif
-# define VEC_SIZE 32
+# define VEC_SIZE 32
+# define PAGE_SIZE 4096
+ .section SECTION(.text), "ax", @progbits
+ENTRY(MEMRCHR)
+# ifdef __ILP32__
+ /* Clear upper bits. */
+ and %RDX_LP, %RDX_LP
+# else
+ test %RDX_LP, %RDX_LP
+# endif
+ jz L(zero_0)
- .section SECTION(.text),"ax",@progbits
-ENTRY (MEMRCHR)
- /* Broadcast CHAR to YMM0. */
vmovd %esi, %xmm0
- vpbroadcastb %xmm0, %ymm0
-
- sub $VEC_SIZE, %RDX_LP
- jbe L(last_vec_or_less)
-
- add %RDX_LP, %RDI_LP
-
- /* Check the last VEC_SIZE bytes. */
- vpcmpeqb (%rdi), %ymm0, %ymm1
- vpmovmskb %ymm1, %eax
- testl %eax, %eax
- jnz L(last_vec_x0)
+ /* Get end pointer. Minus one for two reasons. 1) It is necessary for a
+ correct page cross check and 2) it correctly sets up end ptr to be
+ subtract by lzcnt aligned. */
+ leaq -1(%rdx, %rdi), %rax
- subq $(VEC_SIZE * 4), %rdi
- movl %edi, %ecx
- andl $(VEC_SIZE - 1), %ecx
- jz L(aligned_more)
+ vpbroadcastb %xmm0, %ymm0
- /* Align data for aligned loads in the loop. */
- addq $VEC_SIZE, %rdi
- addq $VEC_SIZE, %rdx
- andq $-VEC_SIZE, %rdi
- subq %rcx, %rdx
+ /* Check if we can load 1x VEC without cross a page. */
+ testl $(PAGE_SIZE - VEC_SIZE), %eax
+ jz L(page_cross)
+
+ vpcmpeqb -(VEC_SIZE - 1)(%rax), %ymm0, %ymm1
+ vpmovmskb %ymm1, %ecx
+ cmpq $VEC_SIZE, %rdx
+ ja L(more_1x_vec)
+
+L(ret_vec_x0_test):
+ /* If ecx is zero (no matches) lzcnt will set it 32 (VEC_SIZE) which
+ will gurantee edx (len) is less than it. */
+ lzcntl %ecx, %ecx
+
+ /* Hoist vzeroupper (not great for RTM) to save code size. This allows
+ all logic for edx (len) <= VEC_SIZE to fit in first cache line. */
+ COND_VZEROUPPER
+ cmpl %ecx, %edx
+ jle L(zero_0)
+ subq %rcx, %rax
+ ret
- .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. */
- vpcmpeqb (VEC_SIZE * 3)(%rdi), %ymm0, %ymm1
- vpmovmskb %ymm1, %eax
- testl %eax, %eax
- jnz L(last_vec_x3)
-
- vpcmpeqb (VEC_SIZE * 2)(%rdi), %ymm0, %ymm2
- vpmovmskb %ymm2, %eax
- testl %eax, %eax
- jnz L(last_vec_x2)
-
- vpcmpeqb VEC_SIZE(%rdi), %ymm0, %ymm3
- vpmovmskb %ymm3, %eax
- testl %eax, %eax
- jnz L(last_vec_x1)
-
- vpcmpeqb (%rdi), %ymm0, %ymm4
- vpmovmskb %ymm4, %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
+ /* Fits in aligning bytes of first cache line. */
+L(zero_0):
+ xorl %eax, %eax
+ ret
- .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)
-
- vmovdqa (%rdi), %ymm1
- vmovdqa VEC_SIZE(%rdi), %ymm2
- vmovdqa (VEC_SIZE * 2)(%rdi), %ymm3
- vmovdqa (VEC_SIZE * 3)(%rdi), %ymm4
-
- vpcmpeqb %ymm1, %ymm0, %ymm1
- vpcmpeqb %ymm2, %ymm0, %ymm2
- vpcmpeqb %ymm3, %ymm0, %ymm3
- vpcmpeqb %ymm4, %ymm0, %ymm4
-
- vpor %ymm1, %ymm2, %ymm5
- vpor %ymm3, %ymm4, %ymm6
- vpor %ymm5, %ymm6, %ymm5
-
- vpmovmskb %ymm5, %eax
- testl %eax, %eax
- jz L(loop_4x_vec)
-
- /* There is a match. */
- vpmovmskb %ymm4, %eax
- testl %eax, %eax
- jnz L(last_vec_x3)
-
- vpmovmskb %ymm3, %eax
- testl %eax, %eax
- jnz L(last_vec_x2)
-
- vpmovmskb %ymm2, %eax
- testl %eax, %eax
- jnz L(last_vec_x1)
-
- vpmovmskb %ymm1, %eax
- bsrl %eax, %eax
- addq %rdi, %rax
+ .p2align 4,, 9
+L(ret_vec_x0):
+ lzcntl %ecx, %ecx
+ subq %rcx, %rax
L(return_vzeroupper):
ZERO_UPPER_VEC_REGISTERS_RETURN
- .p2align 4
-L(last_4x_vec_or_less):
- addl $(VEC_SIZE * 4), %edx
- cmpl $(VEC_SIZE * 2), %edx
- jbe L(last_2x_vec)
-
- vpcmpeqb (VEC_SIZE * 3)(%rdi), %ymm0, %ymm1
- vpmovmskb %ymm1, %eax
- testl %eax, %eax
- jnz L(last_vec_x3)
-
- vpcmpeqb (VEC_SIZE * 2)(%rdi), %ymm0, %ymm2
- vpmovmskb %ymm2, %eax
- testl %eax, %eax
- jnz L(last_vec_x2)
-
- vpcmpeqb VEC_SIZE(%rdi), %ymm0, %ymm3
- vpmovmskb %ymm3, %eax
- testl %eax, %eax
- jnz L(last_vec_x1_check)
- cmpl $(VEC_SIZE * 3), %edx
- jbe L(zero)
-
- vpcmpeqb (%rdi), %ymm0, %ymm4
- vpmovmskb %ymm4, %eax
- testl %eax, %eax
- jz L(zero)
- bsrl %eax, %eax
- subq $(VEC_SIZE * 4), %rdx
- addq %rax, %rdx
- jl L(zero)
- addq %rdi, %rax
- VZEROUPPER_RETURN
-
- .p2align 4
+ .p2align 4,, 10
+L(more_1x_vec):
+ testl %ecx, %ecx
+ jnz L(ret_vec_x0)
+
+ /* Align rax (string pointer). */
+ andq $-VEC_SIZE, %rax
+
+ /* Recompute remaining length after aligning. */
+ movq %rax, %rdx
+ /* Need this comparison next no matter what. */
+ vpcmpeqb -(VEC_SIZE)(%rax), %ymm0, %ymm1
+ subq %rdi, %rdx
+ decq %rax
+ vpmovmskb %ymm1, %ecx
+ /* Fall through for short (hotter than length). */
+ cmpq $(VEC_SIZE * 2), %rdx
+ ja L(more_2x_vec)
L(last_2x_vec):
- vpcmpeqb (VEC_SIZE * 3)(%rdi), %ymm0, %ymm1
- vpmovmskb %ymm1, %eax
- testl %eax, %eax
- jnz L(last_vec_x3_check)
cmpl $VEC_SIZE, %edx
- jbe L(zero)
-
- vpcmpeqb (VEC_SIZE * 2)(%rdi), %ymm0, %ymm1
- vpmovmskb %ymm1, %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
- VZEROUPPER_RETURN
-
- .p2align 4
-L(last_vec_x0):
- bsrl %eax, %eax
- addq %rdi, %rax
- VZEROUPPER_RETURN
+ jbe L(ret_vec_x0_test)
+
+ testl %ecx, %ecx
+ jnz L(ret_vec_x0)
+
+ vpcmpeqb -(VEC_SIZE * 2 - 1)(%rax), %ymm0, %ymm1
+ vpmovmskb %ymm1, %ecx
+ /* 64-bit lzcnt. This will naturally add 32 to position. */
+ lzcntq %rcx, %rcx
+ COND_VZEROUPPER
+ cmpl %ecx, %edx
+ jle L(zero_0)
+ subq %rcx, %rax
+ ret
- .p2align 4
-L(last_vec_x1):
- bsrl %eax, %eax
- addl $VEC_SIZE, %eax
- addq %rdi, %rax
- VZEROUPPER_RETURN
- .p2align 4
-L(last_vec_x2):
- bsrl %eax, %eax
- addl $(VEC_SIZE * 2), %eax
- addq %rdi, %rax
+ /* Inexpensive place to put this regarding code size / target alignments
+ / ICache NLP. Necessary for 2-byte encoding of jump to page cross
+ case which in turn is necessary for hot path (len <= VEC_SIZE) to fit
+ in first cache line. */
+L(page_cross):
+ movq %rax, %rsi
+ andq $-VEC_SIZE, %rsi
+ vpcmpeqb (%rsi), %ymm0, %ymm1
+ vpmovmskb %ymm1, %ecx
+ /* Shift out negative alignment (because we are starting from endptr and
+ working backwards). */
+ movl %eax, %r8d
+ /* notl because eax already has endptr - 1. (-x = ~(x - 1)). */
+ notl %r8d
+ shlxl %r8d, %ecx, %ecx
+ cmpq %rdi, %rsi
+ ja L(more_1x_vec)
+ lzcntl %ecx, %ecx
+ COND_VZEROUPPER
+ cmpl %ecx, %edx
+ jle L(zero_0)
+ subq %rcx, %rax
+ ret
+ .p2align 4,, 11
+L(ret_vec_x1):
+ /* This will naturally add 32 to position. */
+ lzcntq %rcx, %rcx
+ subq %rcx, %rax
VZEROUPPER_RETURN
+ .p2align 4,, 10
+L(more_2x_vec):
+ testl %ecx, %ecx
+ jnz L(ret_vec_x0)
- .p2align 4
-L(last_vec_x3):
- bsrl %eax, %eax
- addl $(VEC_SIZE * 3), %eax
- addq %rdi, %rax
- ret
+ vpcmpeqb -(VEC_SIZE * 2 - 1)(%rax), %ymm0, %ymm1
+ vpmovmskb %ymm1, %ecx
+ testl %ecx, %ecx
+ jnz L(ret_vec_x1)
- .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
- VZEROUPPER_RETURN
- .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
- VZEROUPPER_RETURN
+ /* Needed no matter what. */
+ vpcmpeqb -(VEC_SIZE * 3 - 1)(%rax), %ymm0, %ymm1
+ vpmovmskb %ymm1, %ecx
- .p2align 4
-L(zero):
- xorl %eax, %eax
- VZEROUPPER_RETURN
+ 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)
+
+ /* Needed no matter what. */
+ vpcmpeqb -(VEC_SIZE * 4 - 1)(%rax), %ymm0, %ymm1
+ vpmovmskb %ymm1, %ecx
+ lzcntl %ecx, %ecx
+ subq $(VEC_SIZE * 3), %rax
+ COND_VZEROUPPER
+ subq %rcx, %rax
+ cmpq %rax, %rdi
+ ja L(zero_2)
+ ret
- .p2align 4
-L(null):
+ /* First in aligning bytes. */
+L(zero_2):
xorl %eax, %eax
ret
- .p2align 4
-L(last_vec_or_less_aligned):
- movl %edx, %ecx
+ .p2align 4,, 4
+L(ret_vec_x2_test):
+ lzcntl %ecx, %ecx
+ subq $(VEC_SIZE * 2), %rax
+ COND_VZEROUPPER
+ subq %rcx, %rax
+ cmpq %rax, %rdi
+ ja L(zero_2)
+ ret
- vpcmpeqb (%rdi), %ymm0, %ymm1
- movl $1, %edx
- /* Support rdx << 32. */
- salq %cl, %rdx
- subq $1, %rdx
+ .p2align 4,, 11
+L(ret_vec_x2):
+ /* ecx must be non-zero. */
+ bsrl %ecx, %ecx
+ leaq (VEC_SIZE * -3 + 1)(%rcx, %rax), %rax
+ VZEROUPPER_RETURN
- vpmovmskb %ymm1, %eax
+ .p2align 4,, 14
+L(ret_vec_x3):
+ /* ecx must be non-zero. */
+ bsrl %ecx, %ecx
+ leaq (VEC_SIZE * -4 + 1)(%rcx, %rax), %rax
+ VZEROUPPER_RETURN
- /* Remove the trailing bytes. */
- andl %edx, %eax
- testl %eax, %eax
- jz L(zero)
- bsrl %eax, %eax
- addq %rdi, %rax
- VZEROUPPER_RETURN
.p2align 4
-L(last_vec_or_less):
- addl $VEC_SIZE, %edx
+L(more_4x_vec):
+ testl %ecx, %ecx
+ jnz L(ret_vec_x2)
- /* Check for zero length. */
- testl %edx, %edx
- jz L(null)
+ vpcmpeqb -(VEC_SIZE * 4 - 1)(%rax), %ymm0, %ymm1
+ vpmovmskb %ymm1, %ecx
- movl %edi, %ecx
- andl $(VEC_SIZE - 1), %ecx
- jz L(last_vec_or_less_aligned)
+ testl %ecx, %ecx
+ jnz L(ret_vec_x3)
- movl %ecx, %esi
- movl %ecx, %r8d
- addl %edx, %esi
- andq $-VEC_SIZE, %rdi
+ /* Check if near end before re-aligning (otherwise might do an
+ unnecissary loop iteration). */
+ addq $-(VEC_SIZE * 4), %rax
+ cmpq $(VEC_SIZE * 4), %rdx
+ jbe L(last_4x_vec)
- subl $VEC_SIZE, %esi
- ja L(last_vec_2x_aligned)
+ /* Align rax to (VEC_SIZE - 1). */
+ orq $(VEC_SIZE * 4 - 1), %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. */
+ orq $(VEC_SIZE * 4 - 1), %rdx
- /* Check the last VEC. */
- vpcmpeqb (%rdi), %ymm0, %ymm1
- vpmovmskb %ymm1, %eax
-
- /* Remove the leading and trailing bytes. */
- sarl %cl, %eax
- movl %edx, %ecx
+ .p2align 4
+L(loop_4x_vec):
+ /* Need this comparison next no matter what. */
+ vpcmpeqb -(VEC_SIZE * 1 - 1)(%rax), %ymm0, %ymm1
+ vpcmpeqb -(VEC_SIZE * 2 - 1)(%rax), %ymm0, %ymm2
+ vpcmpeqb -(VEC_SIZE * 3 - 1)(%rax), %ymm0, %ymm3
+ vpcmpeqb -(VEC_SIZE * 4 - 1)(%rax), %ymm0, %ymm4
- movl $1, %edx
- sall %cl, %edx
- subl $1, %edx
+ vpor %ymm1, %ymm2, %ymm2
+ vpor %ymm3, %ymm4, %ymm4
+ vpor %ymm2, %ymm4, %ymm4
+ vpmovmskb %ymm4, %esi
- andl %edx, %eax
- testl %eax, %eax
- jz L(zero)
+ testl %esi, %esi
+ jnz L(loop_end)
- bsrl %eax, %eax
- addq %rdi, %rax
- addq %r8, %rax
- VZEROUPPER_RETURN
+ addq $(VEC_SIZE * -4), %rax
+ cmpq %rdx, %rax
+ jne L(loop_4x_vec)
- .p2align 4
-L(last_vec_2x_aligned):
- movl %esi, %ecx
+ subl %edi, %edx
+ incl %edx
- /* Check the last VEC. */
- vpcmpeqb VEC_SIZE(%rdi), %ymm0, %ymm1
+L(last_4x_vec):
+ /* Used no matter what. */
+ vpcmpeqb -(VEC_SIZE * 1 - 1)(%rax), %ymm0, %ymm1
+ vpmovmskb %ymm1, %ecx
- movl $1, %edx
- sall %cl, %edx
- subl $1, %edx
+ cmpl $(VEC_SIZE * 2), %edx
+ jbe L(last_2x_vec)
- vpmovmskb %ymm1, %eax
+ testl %ecx, %ecx
+ jnz L(ret_vec_x0_end)
- /* Remove the trailing bytes. */
- andl %edx, %eax
+ vpcmpeqb -(VEC_SIZE * 2 - 1)(%rax), %ymm0, %ymm1
+ vpmovmskb %ymm1, %ecx
+ testl %ecx, %ecx
+ jnz L(ret_vec_x1_end)
- testl %eax, %eax
- jnz L(last_vec_x1)
+ /* Used no matter what. */
+ vpcmpeqb -(VEC_SIZE * 3 - 1)(%rax), %ymm0, %ymm1
+ vpmovmskb %ymm1, %ecx
- /* Check the second last VEC. */
- vpcmpeqb (%rdi), %ymm0, %ymm1
+ cmpl $(VEC_SIZE * 3), %edx
+ ja L(last_vec)
+
+ lzcntl %ecx, %ecx
+ subq $(VEC_SIZE * 2), %rax
+ COND_VZEROUPPER
+ subq %rcx, %rax
+ cmpq %rax, %rdi
+ jbe L(ret0)
+ xorl %eax, %eax
+L(ret0):
+ ret
- movl %r8d, %ecx
- vpmovmskb %ymm1, %eax
+ .p2align 4
+L(loop_end):
+ vpmovmskb %ymm1, %ecx
+ testl %ecx, %ecx
+ jnz L(ret_vec_x0_end)
+
+ vpmovmskb %ymm2, %ecx
+ testl %ecx, %ecx
+ jnz L(ret_vec_x1_end)
+
+ vpmovmskb %ymm3, %ecx
+ /* 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
+ leaq (VEC_SIZE * -4 + 1)(%rcx, %rax), %rax
+ VZEROUPPER_RETURN
- /* Remove the leading bytes. Must use unsigned right shift for
- bsrl below. */
- shrl %cl, %eax
- testl %eax, %eax
- jz L(zero)
+ .p2align 4,, 4
+L(ret_vec_x1_end):
+ /* 64-bit version will automatically add 32 (VEC_SIZE). */
+ lzcntq %rcx, %rcx
+ subq %rcx, %rax
+ VZEROUPPER_RETURN
- bsrl %eax, %eax
- addq %rdi, %rax
- addq %r8, %rax
+ .p2align 4,, 4
+L(ret_vec_x0_end):
+ lzcntl %ecx, %ecx
+ subq %rcx, %rax
VZEROUPPER_RETURN
-END (MEMRCHR)
+
+ /* 2 bytes until next cache line. */
+END(MEMRCHR)
#endif