* [PATCH v4 1/2] x86: Optimize strlen-evex.S
@ 2021-04-19 18:43 Noah Goldstein
2021-04-19 18:43 ` [PATCH v4 2/2] x86: Optimize strlen-avx2.S Noah Goldstein
2021-04-19 22:16 ` [PATCH v4 1/2] x86: Optimize strlen-evex.S H.J. Lu
0 siblings, 2 replies; 6+ messages in thread
From: Noah Goldstein @ 2021-04-19 18:43 UTC (permalink / raw)
To: libc-alpha
No bug. This commit optimizes strlen-evex.S. The
optimizations are mostly small things but they add up to roughly
10-30% performance improvement for strlen. The results for strnlen are
bit more ambiguous. test-strlen, test-strnlen, test-wcslen, and
test-wcsnlen are all passing.
Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com>
---
sysdeps/x86_64/multiarch/strlen-evex.S | 584 ++++++++++++++-----------
1 file changed, 320 insertions(+), 264 deletions(-)
diff --git a/sysdeps/x86_64/multiarch/strlen-evex.S b/sysdeps/x86_64/multiarch/strlen-evex.S
index 0583819078..36b28198c8 100644
--- a/sysdeps/x86_64/multiarch/strlen-evex.S
+++ b/sysdeps/x86_64/multiarch/strlen-evex.S
@@ -29,11 +29,13 @@
# ifdef USE_AS_WCSLEN
# define VPCMP vpcmpd
# define VPMINU vpminud
-# define SHIFT_REG r9d
+# define SHIFT_REG ecx
+# define CHAR_SIZE 4
# else
# define VPCMP vpcmpb
# define VPMINU vpminub
-# define SHIFT_REG ecx
+# define SHIFT_REG edx
+# define CHAR_SIZE 1
# endif
# define XMMZERO xmm16
@@ -46,132 +48,168 @@
# define YMM6 ymm22
# define VEC_SIZE 32
+# define PAGE_SIZE 4096
+# define LOG_PAGE_SIZE 12
+# define CHAR_PER_VEC (VEC_SIZE / CHAR_SIZE)
.section .text.evex,"ax",@progbits
ENTRY (STRLEN)
# ifdef USE_AS_STRNLEN
- /* Check for zero length. */
+ /* Check zero length. */
test %RSI_LP, %RSI_LP
jz L(zero)
-# ifdef USE_AS_WCSLEN
- shl $2, %RSI_LP
-# elif defined __ILP32__
+# ifdef __ILP32__
/* Clear the upper 32 bits. */
movl %esi, %esi
# endif
mov %RSI_LP, %R8_LP
# endif
- movl %edi, %ecx
- movq %rdi, %rdx
+ movl %edi, %eax
vpxorq %XMMZERO, %XMMZERO, %XMMZERO
-
+ /* Shift left eax to clear all bits not relevant to page cross
+ check. This saves 2 bytes of code as opposed to using andl with
+ PAGE_SIZE - 1. Then compare with PAGE_SIZE - VEC_SIZE shifted
+ left by the same amount (an imm32 either way). */
+ sall $(32 - LOG_PAGE_SIZE), %eax
/* Check if we may cross page boundary with one vector load. */
- andl $(2 * VEC_SIZE - 1), %ecx
- cmpl $VEC_SIZE, %ecx
- ja L(cros_page_boundary)
+ cmpl $((PAGE_SIZE - VEC_SIZE) << (32 - LOG_PAGE_SIZE)), %eax
+ ja L(cross_page_boundary)
/* Check the first VEC_SIZE bytes. Each bit in K0 represents a
null byte. */
VPCMP $0, (%rdi), %YMMZERO, %k0
kmovd %k0, %eax
- testl %eax, %eax
-
# ifdef USE_AS_STRNLEN
- jnz L(first_vec_x0_check)
- /* Adjust length and check the end of data. */
- subq $VEC_SIZE, %rsi
- jbe L(max)
-# else
- jnz L(first_vec_x0)
+ /* If length < CHAR_PER_VEC handle special. */
+ cmpq $CHAR_PER_VEC, %rsi
+ jbe L(first_vec_x0)
# endif
-
- /* Align data for aligned loads in the loop. */
- addq $VEC_SIZE, %rdi
- andl $(VEC_SIZE - 1), %ecx
- andq $-VEC_SIZE, %rdi
-
+ testl %eax, %eax
+ jz L(aligned_more)
+ tzcntl %eax, %eax
+ ret
# ifdef USE_AS_STRNLEN
- /* Adjust length. */
- addq %rcx, %rsi
+L(zero):
+ xorl %eax, %eax
+ ret
- subq $(VEC_SIZE * 4), %rsi
- jbe L(last_4x_vec_or_less)
+ .p2align 4
+L(first_vec_x0):
+ /* Set bit for max len so that tzcnt will return min of max len
+ and position of first match. */
+ btsq %rsi, %rax
+ tzcntl %eax, %eax
+ ret
# endif
- jmp L(more_4x_vec)
.p2align 4
-L(cros_page_boundary):
- andl $(VEC_SIZE - 1), %ecx
- andq $-VEC_SIZE, %rdi
-
-# ifdef USE_AS_WCSLEN
- /* NB: Divide shift count by 4 since each bit in K0 represent 4
- bytes. */
- movl %ecx, %SHIFT_REG
- sarl $2, %SHIFT_REG
+L(first_vec_x1):
+ tzcntl %eax, %eax
+ /* Safe to use 32 bit instructions as these are only called for
+ size = [1, 159]. */
+# ifdef USE_AS_STRNLEN
+ /* Use ecx which was computed earlier to compute correct value.
+ */
+ leal -(CHAR_PER_VEC * 4 + 1)(%rcx, %rax), %eax
+# else
+ subl %edx, %edi
+# ifdef USE_AS_WCSLEN
+ /* NB: Divide bytes by 4 to get the wchar_t count. */
+ sarl $2, %edi
+# endif
+ leal CHAR_PER_VEC(%rdi, %rax), %eax
# endif
- VPCMP $0, (%rdi), %YMMZERO, %k0
- kmovd %k0, %eax
+ ret
- /* Remove the leading bytes. */
- sarxl %SHIFT_REG, %eax, %eax
- testl %eax, %eax
- jz L(aligned_more)
+ .p2align 4
+L(first_vec_x2):
tzcntl %eax, %eax
-# ifdef USE_AS_WCSLEN
- /* NB: Multiply wchar_t count by 4 to get the number of bytes. */
- sall $2, %eax
-# endif
+ /* Safe to use 32 bit instructions as these are only called for
+ size = [1, 159]. */
# ifdef USE_AS_STRNLEN
- /* Check the end of data. */
- cmpq %rax, %rsi
- jbe L(max)
-# endif
- addq %rdi, %rax
- addq %rcx, %rax
- subq %rdx, %rax
-# ifdef USE_AS_WCSLEN
- shrq $2, %rax
+ /* Use ecx which was computed earlier to compute correct value.
+ */
+ leal -(CHAR_PER_VEC * 3 + 1)(%rcx, %rax), %eax
+# else
+ subl %edx, %edi
+# ifdef USE_AS_WCSLEN
+ /* NB: Divide bytes by 4 to get the wchar_t count. */
+ sarl $2, %edi
+# endif
+ leal (CHAR_PER_VEC * 2)(%rdi, %rax), %eax
# endif
ret
.p2align 4
-L(aligned_more):
+L(first_vec_x3):
+ tzcntl %eax, %eax
+ /* Safe to use 32 bit instructions as these are only called for
+ size = [1, 159]. */
# ifdef USE_AS_STRNLEN
- /* "rcx" is less than VEC_SIZE. Calculate "rdx + rcx - VEC_SIZE"
- with "rdx - (VEC_SIZE - rcx)" instead of "(rdx + rcx) - VEC_SIZE"
- to void possible addition overflow. */
- negq %rcx
- addq $VEC_SIZE, %rcx
-
- /* Check the end of data. */
- subq %rcx, %rsi
- jbe L(max)
+ /* Use ecx which was computed earlier to compute correct value.
+ */
+ leal -(CHAR_PER_VEC * 2 + 1)(%rcx, %rax), %eax
+# else
+ subl %edx, %edi
+# ifdef USE_AS_WCSLEN
+ /* NB: Divide bytes by 4 to get the wchar_t count. */
+ sarl $2, %edi
+# endif
+ leal (CHAR_PER_VEC * 3)(%rdi, %rax), %eax
# endif
+ ret
- addq $VEC_SIZE, %rdi
-
+ .p2align 4
+L(first_vec_x4):
+ tzcntl %eax, %eax
+ /* Safe to use 32 bit instructions as these are only called for
+ size = [1, 159]. */
# ifdef USE_AS_STRNLEN
- subq $(VEC_SIZE * 4), %rsi
- jbe L(last_4x_vec_or_less)
+ /* Use ecx which was computed earlier to compute correct value.
+ */
+ leal -(CHAR_PER_VEC + 1)(%rcx, %rax), %eax
+# else
+ subl %edx, %edi
+# ifdef USE_AS_WCSLEN
+ /* NB: Divide bytes by 4 to get the wchar_t count. */
+ sarl $2, %edi
+# endif
+ leal (CHAR_PER_VEC * 4)(%rdi, %rax), %eax
# endif
+ ret
-L(more_4x_vec):
+ .p2align 5
+L(aligned_more):
+ movq %rdi, %rdx
+ /* Align data to VEC_SIZE. */
+ andq $-(VEC_SIZE), %rdi
+L(cross_page_continue):
/* Check the first 4 * VEC_SIZE. Only one VEC_SIZE at a time
since data is only aligned to VEC_SIZE. */
- VPCMP $0, (%rdi), %YMMZERO, %k0
- kmovd %k0, %eax
- testl %eax, %eax
- jnz L(first_vec_x0)
-
+# ifdef USE_AS_STRNLEN
+ /* + CHAR_SIZE because it simplies the logic in
+ last_4x_vec_or_less. */
+ leaq (VEC_SIZE * 5 + CHAR_SIZE)(%rdi), %rcx
+ subq %rdx, %rcx
+# ifdef USE_AS_WCSLEN
+ /* NB: Divide bytes by 4 to get the wchar_t count. */
+ sarl $2, %ecx
+# endif
+# endif
+ /* Load first VEC regardless. */
VPCMP $0, VEC_SIZE(%rdi), %YMMZERO, %k0
+# ifdef USE_AS_STRNLEN
+ /* Adjust length. If near end handle specially. */
+ subq %rcx, %rsi
+ jb L(last_4x_vec_or_less)
+# endif
kmovd %k0, %eax
testl %eax, %eax
jnz L(first_vec_x1)
VPCMP $0, (VEC_SIZE * 2)(%rdi), %YMMZERO, %k0
kmovd %k0, %eax
- testl %eax, %eax
+ test %eax, %eax
jnz L(first_vec_x2)
VPCMP $0, (VEC_SIZE * 3)(%rdi), %YMMZERO, %k0
@@ -179,258 +217,276 @@ L(more_4x_vec):
testl %eax, %eax
jnz L(first_vec_x3)
- addq $(VEC_SIZE * 4), %rdi
-
-# ifdef USE_AS_STRNLEN
- subq $(VEC_SIZE * 4), %rsi
- jbe L(last_4x_vec_or_less)
-# endif
-
- /* Align data to 4 * VEC_SIZE. */
- movq %rdi, %rcx
- andl $(4 * VEC_SIZE - 1), %ecx
- andq $-(4 * VEC_SIZE), %rdi
+ VPCMP $0, (VEC_SIZE * 4)(%rdi), %YMMZERO, %k0
+ kmovd %k0, %eax
+ testl %eax, %eax
+ jnz L(first_vec_x4)
+ addq $VEC_SIZE, %rdi
# ifdef USE_AS_STRNLEN
- /* Adjust length. */
+ /* Check if at last VEC_SIZE * 4 length. */
+ cmpq $(CHAR_PER_VEC * 4 - 1), %rsi
+ jbe L(last_4x_vec_or_less_load)
+ movl %edi, %ecx
+ andl $(VEC_SIZE * 4 - 1), %ecx
+# ifdef USE_AS_WCSLEN
+ /* NB: Divide bytes by 4 to get the wchar_t count. */
+ sarl $2, %ecx
+# endif
+ /* Readjust length. */
addq %rcx, %rsi
# endif
+ /* Align data to VEC_SIZE * 4. */
+ andq $-(VEC_SIZE * 4), %rdi
+ /* Compare 4 * VEC at a time forward. */
.p2align 4
L(loop_4x_vec):
- /* Compare 4 * VEC at a time forward. */
- VMOVA (%rdi), %YMM1
- VMOVA VEC_SIZE(%rdi), %YMM2
- VMOVA (VEC_SIZE * 2)(%rdi), %YMM3
- VMOVA (VEC_SIZE * 3)(%rdi), %YMM4
-
- VPMINU %YMM1, %YMM2, %YMM5
- VPMINU %YMM3, %YMM4, %YMM6
+ /* Load first VEC regardless. */
+ VMOVA (VEC_SIZE * 4)(%rdi), %YMM1
+# ifdef USE_AS_STRNLEN
+ /* Break if at end of length. */
+ subq $(CHAR_PER_VEC * 4), %rsi
+ jb L(last_4x_vec_or_less_cmpeq)
+# endif
+ /* Save some code size by microfusing VPMINU with the load. Since
+ the matches in ymm2/ymm4 can only be returned if there where no
+ matches in ymm1/ymm3 respectively there is no issue with overlap.
+ */
+ VPMINU (VEC_SIZE * 5)(%rdi), %YMM1, %YMM2
+ VMOVA (VEC_SIZE * 6)(%rdi), %YMM3
+ VPMINU (VEC_SIZE * 7)(%rdi), %YMM3, %YMM4
+
+ VPCMP $0, %YMM2, %YMMZERO, %k0
+ VPCMP $0, %YMM4, %YMMZERO, %k1
+ subq $-(VEC_SIZE * 4), %rdi
+ kortestd %k0, %k1
+ jz L(loop_4x_vec)
+
+ /* Check if end was in first half. */
+ kmovd %k0, %eax
+ subq %rdx, %rdi
+# ifdef USE_AS_WCSLEN
+ shrq $2, %rdi
+# endif
+ testl %eax, %eax
+ jz L(second_vec_return)
- VPMINU %YMM5, %YMM6, %YMM5
- VPCMP $0, %YMM5, %YMMZERO, %k0
- ktestd %k0, %k0
- jnz L(4x_vec_end)
+ VPCMP $0, %YMM1, %YMMZERO, %k2
+ kmovd %k2, %edx
+ /* Combine VEC1 matches (edx) with VEC2 matches (eax). */
+# ifdef USE_AS_WCSLEN
+ sall $CHAR_PER_VEC, %eax
+ orl %edx, %eax
+ tzcntl %eax, %eax
+# else
+ salq $CHAR_PER_VEC, %rax
+ orq %rdx, %rax
+ tzcntq %rax, %rax
+# endif
+ addq %rdi, %rax
+ ret
- addq $(VEC_SIZE * 4), %rdi
-# ifndef USE_AS_STRNLEN
- jmp L(loop_4x_vec)
-# else
- subq $(VEC_SIZE * 4), %rsi
- ja L(loop_4x_vec)
+# ifdef USE_AS_STRNLEN
+L(last_4x_vec_or_less_load):
+ /* Depending on entry adjust rdi / prepare first VEC in YMM1. */
+ VMOVA (VEC_SIZE * 4)(%rdi), %YMM1
+L(last_4x_vec_or_less_cmpeq):
+ VPCMP $0, %YMM1, %YMMZERO, %k0
+ addq $(VEC_SIZE * 3), %rdi
L(last_4x_vec_or_less):
- /* Less than 4 * VEC and aligned to VEC_SIZE. */
- addl $(VEC_SIZE * 2), %esi
- jle L(last_2x_vec)
-
- VPCMP $0, (%rdi), %YMMZERO, %k0
kmovd %k0, %eax
+ /* If remaining length > VEC_SIZE * 2. This works if esi is off by
+ VEC_SIZE * 4. */
+ testl $(CHAR_PER_VEC * 2), %esi
+ jnz L(last_4x_vec)
+
+ /* length may have been negative or positive by an offset of
+ CHAR_PER_VEC * 4 depending on where this was called from. This
+ fixes that. */
+ andl $(CHAR_PER_VEC * 4 - 1), %esi
testl %eax, %eax
- jnz L(first_vec_x0)
+ jnz L(last_vec_x1_check)
- VPCMP $0, VEC_SIZE(%rdi), %YMMZERO, %k0
- kmovd %k0, %eax
- testl %eax, %eax
- jnz L(first_vec_x1)
+ /* Check the end of data. */
+ subl $CHAR_PER_VEC, %esi
+ jb L(max)
VPCMP $0, (VEC_SIZE * 2)(%rdi), %YMMZERO, %k0
kmovd %k0, %eax
- testl %eax, %eax
- jnz L(first_vec_x2_check)
- subl $VEC_SIZE, %esi
- jle L(max)
+ tzcntl %eax, %eax
+ /* Check the end of data. */
+ cmpl %eax, %esi
+ jb L(max)
- VPCMP $0, (VEC_SIZE * 3)(%rdi), %YMMZERO, %k0
- kmovd %k0, %eax
- testl %eax, %eax
- jnz L(first_vec_x3_check)
+ subq %rdx, %rdi
+# ifdef USE_AS_WCSLEN
+ /* NB: Divide bytes by 4 to get the wchar_t count. */
+ sarq $2, %rdi
+# endif
+ leaq (CHAR_PER_VEC * 2)(%rdi, %rax), %rax
+ ret
+L(max):
movq %r8, %rax
+ ret
+# endif
+
+ /* Placed here in strnlen so that the jcc L(last_4x_vec_or_less)
+ in the 4x VEC loop can use 2 byte encoding. */
+ .p2align 4
+L(second_vec_return):
+ VPCMP $0, %YMM3, %YMMZERO, %k0
+ /* Combine YMM3 matches (k0) with YMM4 matches (k1). */
+# ifdef USE_AS_WCSLEN
+ kunpckbw %k0, %k1, %k0
+ kmovd %k0, %eax
+ tzcntl %eax, %eax
+# else
+ kunpckdq %k0, %k1, %k0
+ kmovq %k0, %rax
+ tzcntq %rax, %rax
+# endif
+ leaq (CHAR_PER_VEC * 2)(%rdi, %rax), %rax
+ ret
+
+
+# ifdef USE_AS_STRNLEN
+L(last_vec_x1_check):
+ tzcntl %eax, %eax
+ /* Check the end of data. */
+ cmpl %eax, %esi
+ jb L(max)
+ subq %rdx, %rdi
# ifdef USE_AS_WCSLEN
- shrq $2, %rax
+ /* NB: Divide bytes by 4 to get the wchar_t count. */
+ sarq $2, %rdi
# endif
+ leaq (CHAR_PER_VEC)(%rdi, %rax), %rax
ret
.p2align 4
-L(last_2x_vec):
- addl $(VEC_SIZE * 2), %esi
+L(last_4x_vec):
+ /* Test first 2x VEC normally. */
+ testl %eax, %eax
+ jnz L(last_vec_x1)
- VPCMP $0, (%rdi), %YMMZERO, %k0
+ VPCMP $0, (VEC_SIZE * 2)(%rdi), %YMMZERO, %k0
kmovd %k0, %eax
testl %eax, %eax
- jnz L(first_vec_x0_check)
- subl $VEC_SIZE, %esi
- jle L(max)
+ jnz L(last_vec_x2)
- VPCMP $0, VEC_SIZE(%rdi), %YMMZERO, %k0
+ /* Normalize length. */
+ andl $(CHAR_PER_VEC * 4 - 1), %esi
+ VPCMP $0, (VEC_SIZE * 3)(%rdi), %YMMZERO, %k0
kmovd %k0, %eax
testl %eax, %eax
- jnz L(first_vec_x1_check)
- movq %r8, %rax
-# ifdef USE_AS_WCSLEN
- shrq $2, %rax
-# endif
- ret
+ jnz L(last_vec_x3)
- .p2align 4
-L(first_vec_x0_check):
+ /* Check the end of data. */
+ subl $(CHAR_PER_VEC * 3), %esi
+ jb L(max)
+
+ VPCMP $0, (VEC_SIZE * 4)(%rdi), %YMMZERO, %k0
+ kmovd %k0, %eax
tzcntl %eax, %eax
-# ifdef USE_AS_WCSLEN
- /* NB: Multiply wchar_t count by 4 to get the number of bytes. */
- sall $2, %eax
-# endif
/* Check the end of data. */
- cmpq %rax, %rsi
- jbe L(max)
- addq %rdi, %rax
- subq %rdx, %rax
+ cmpl %eax, %esi
+ jb L(max_end)
+
+ subq %rdx, %rdi
# ifdef USE_AS_WCSLEN
- shrq $2, %rax
+ /* NB: Divide bytes by 4 to get the wchar_t count. */
+ sarq $2, %rdi
# endif
+ leaq (CHAR_PER_VEC * 4)(%rdi, %rax), %rax
ret
.p2align 4
-L(first_vec_x1_check):
+L(last_vec_x1):
tzcntl %eax, %eax
+ subq %rdx, %rdi
# ifdef USE_AS_WCSLEN
- /* NB: Multiply wchar_t count by 4 to get the number of bytes. */
- sall $2, %eax
-# endif
- /* Check the end of data. */
- cmpq %rax, %rsi
- jbe L(max)
- addq $VEC_SIZE, %rax
- addq %rdi, %rax
- subq %rdx, %rax
-# ifdef USE_AS_WCSLEN
- shrq $2, %rax
+ /* NB: Divide bytes by 4 to get the wchar_t count. */
+ sarq $2, %rdi
# endif
+ leaq (CHAR_PER_VEC)(%rdi, %rax), %rax
ret
.p2align 4
-L(first_vec_x2_check):
+L(last_vec_x2):
tzcntl %eax, %eax
+ subq %rdx, %rdi
# ifdef USE_AS_WCSLEN
- /* NB: Multiply wchar_t count by 4 to get the number of bytes. */
- sall $2, %eax
-# endif
- /* Check the end of data. */
- cmpq %rax, %rsi
- jbe L(max)
- addq $(VEC_SIZE * 2), %rax
- addq %rdi, %rax
- subq %rdx, %rax
-# ifdef USE_AS_WCSLEN
- shrq $2, %rax
+ /* NB: Divide bytes by 4 to get the wchar_t count. */
+ sarq $2, %rdi
# endif
+ leaq (CHAR_PER_VEC * 2)(%rdi, %rax), %rax
ret
.p2align 4
-L(first_vec_x3_check):
+L(last_vec_x3):
tzcntl %eax, %eax
-# ifdef USE_AS_WCSLEN
- /* NB: Multiply wchar_t count by 4 to get the number of bytes. */
- sall $2, %eax
-# endif
+ subl $(CHAR_PER_VEC * 2), %esi
/* Check the end of data. */
- cmpq %rax, %rsi
- jbe L(max)
- addq $(VEC_SIZE * 3), %rax
- addq %rdi, %rax
- subq %rdx, %rax
+ cmpl %eax, %esi
+ jb L(max_end)
+ subq %rdx, %rdi
# ifdef USE_AS_WCSLEN
- shrq $2, %rax
+ /* NB: Divide bytes by 4 to get the wchar_t count. */
+ sarq $2, %rdi
# endif
+ leaq (CHAR_PER_VEC * 3)(%rdi, %rax), %rax
ret
-
- .p2align 4
-L(max):
+L(max_end):
movq %r8, %rax
-# ifdef USE_AS_WCSLEN
- shrq $2, %rax
-# endif
- ret
-
- .p2align 4
-L(zero):
- xorl %eax, %eax
ret
# endif
+ /* Cold case for crossing page with first load. */
.p2align 4
-L(first_vec_x0):
- tzcntl %eax, %eax
-# ifdef USE_AS_WCSLEN
- /* NB: Multiply wchar_t count by 4 to get the number of bytes. */
- sall $2, %eax
-# endif
- addq %rdi, %rax
- subq %rdx, %rax
+L(cross_page_boundary):
+ movq %rdi, %rdx
+ /* Align data to VEC_SIZE. */
+ andq $-VEC_SIZE, %rdi
+ VPCMP $0, (%rdi), %YMMZERO, %k0
+ kmovd %k0, %eax
+ /* Remove the leading bytes. */
# ifdef USE_AS_WCSLEN
- shrq $2, %rax
+ /* NB: Divide shift count by 4 since each bit in K0 represent 4
+ bytes. */
+ movl %edx, %ecx
+ shrl $2, %ecx
+ andl $(CHAR_PER_VEC - 1), %ecx
# endif
- ret
-
- .p2align 4
-L(first_vec_x1):
+ /* SHIFT_REG is ecx for USE_AS_WCSLEN and edx otherwise. */
+ sarxl %SHIFT_REG, %eax, %eax
+ testl %eax, %eax
+# ifndef USE_AS_STRNLEN
+ jz L(cross_page_continue)
tzcntl %eax, %eax
-# ifdef USE_AS_WCSLEN
- /* NB: Multiply wchar_t count by 4 to get the number of bytes. */
- sall $2, %eax
-# endif
- addq $VEC_SIZE, %rax
- addq %rdi, %rax
- subq %rdx, %rax
-# ifdef USE_AS_WCSLEN
- shrq $2, %rax
-# endif
ret
-
- .p2align 4
-L(first_vec_x2):
- tzcntl %eax, %eax
-# ifdef USE_AS_WCSLEN
- /* NB: Multiply wchar_t count by 4 to get the number of bytes. */
- sall $2, %eax
-# endif
- addq $(VEC_SIZE * 2), %rax
- addq %rdi, %rax
- subq %rdx, %rax
-# ifdef USE_AS_WCSLEN
- shrq $2, %rax
-# endif
+# else
+ jnz L(cross_page_less_vec)
+# ifndef USE_AS_WCSLEN
+ movl %edx, %ecx
+ andl $(CHAR_PER_VEC - 1), %ecx
+# endif
+ movl $CHAR_PER_VEC, %eax
+ subl %ecx, %eax
+ /* Check the end of data. */
+ cmpq %rax, %rsi
+ ja L(cross_page_continue)
+ movl %esi, %eax
ret
-
- .p2align 4
-L(4x_vec_end):
- VPCMP $0, %YMM1, %YMMZERO, %k0
- kmovd %k0, %eax
- testl %eax, %eax
- jnz L(first_vec_x0)
- VPCMP $0, %YMM2, %YMMZERO, %k1
- kmovd %k1, %eax
- testl %eax, %eax
- jnz L(first_vec_x1)
- VPCMP $0, %YMM3, %YMMZERO, %k2
- kmovd %k2, %eax
- testl %eax, %eax
- jnz L(first_vec_x2)
- VPCMP $0, %YMM4, %YMMZERO, %k3
- kmovd %k3, %eax
-L(first_vec_x3):
+L(cross_page_less_vec):
tzcntl %eax, %eax
-# ifdef USE_AS_WCSLEN
- /* NB: Multiply wchar_t count by 4 to get the number of bytes. */
- sall $2, %eax
-# endif
- addq $(VEC_SIZE * 3), %rax
- addq %rdi, %rax
- subq %rdx, %rax
-# ifdef USE_AS_WCSLEN
- shrq $2, %rax
-# endif
+ /* Select min of length and position of first null. */
+ cmpq %rax, %rsi
+ cmovb %esi, %eax
ret
+# endif
END (STRLEN)
#endif
--
2.29.2
^ permalink raw reply [flat|nested] 6+ messages in thread
* [PATCH v4 2/2] x86: Optimize strlen-avx2.S
2021-04-19 18:43 [PATCH v4 1/2] x86: Optimize strlen-evex.S Noah Goldstein
@ 2021-04-19 18:43 ` Noah Goldstein
2021-04-19 22:16 ` H.J. Lu
2021-04-19 22:16 ` [PATCH v4 1/2] x86: Optimize strlen-evex.S H.J. Lu
1 sibling, 1 reply; 6+ messages in thread
From: Noah Goldstein @ 2021-04-19 18:43 UTC (permalink / raw)
To: libc-alpha
No bug. This commit optimizes strlen-avx2.S. The optimizations are
mostly small things but they add up to roughly 10-30% performance
improvement for strlen. The results for strnlen are bit more
ambiguous. test-strlen, test-strnlen, test-wcslen, and test-wcsnlen
are all passing.
Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com>
---
sysdeps/x86_64/multiarch/ifunc-impl-list.c | 16 +-
sysdeps/x86_64/multiarch/strlen-avx2.S | 535 +++++++++++++--------
2 files changed, 337 insertions(+), 214 deletions(-)
diff --git a/sysdeps/x86_64/multiarch/ifunc-impl-list.c b/sysdeps/x86_64/multiarch/ifunc-impl-list.c
index 0b0927b124..4613bead7c 100644
--- a/sysdeps/x86_64/multiarch/ifunc-impl-list.c
+++ b/sysdeps/x86_64/multiarch/ifunc-impl-list.c
@@ -285,10 +285,12 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
/* Support sysdeps/x86_64/multiarch/strlen.c. */
IFUNC_IMPL (i, name, strlen,
IFUNC_IMPL_ADD (array, i, strlen,
- CPU_FEATURE_USABLE (AVX2),
+ (CPU_FEATURE_USABLE (AVX2)
+ && CPU_FEATURE_USABLE (BMI2)),
__strlen_avx2)
IFUNC_IMPL_ADD (array, i, strlen,
(CPU_FEATURE_USABLE (AVX2)
+ && CPU_FEATURE_USABLE (BMI2)
&& CPU_FEATURE_USABLE (RTM)),
__strlen_avx2_rtm)
IFUNC_IMPL_ADD (array, i, strlen,
@@ -301,10 +303,12 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
/* Support sysdeps/x86_64/multiarch/strnlen.c. */
IFUNC_IMPL (i, name, strnlen,
IFUNC_IMPL_ADD (array, i, strnlen,
- CPU_FEATURE_USABLE (AVX2),
+ (CPU_FEATURE_USABLE (AVX2)
+ && CPU_FEATURE_USABLE (BMI2)),
__strnlen_avx2)
IFUNC_IMPL_ADD (array, i, strnlen,
(CPU_FEATURE_USABLE (AVX2)
+ && CPU_FEATURE_USABLE (BMI2)
&& CPU_FEATURE_USABLE (RTM)),
__strnlen_avx2_rtm)
IFUNC_IMPL_ADD (array, i, strnlen,
@@ -646,10 +650,12 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
/* Support sysdeps/x86_64/multiarch/wcslen.c. */
IFUNC_IMPL (i, name, wcslen,
IFUNC_IMPL_ADD (array, i, wcslen,
- CPU_FEATURE_USABLE (AVX2),
+ (CPU_FEATURE_USABLE (AVX2)
+ && CPU_FEATURE_USABLE (BMI2)),
__wcslen_avx2)
IFUNC_IMPL_ADD (array, i, wcslen,
(CPU_FEATURE_USABLE (AVX2)
+ && CPU_FEATURE_USABLE (BMI2)
&& CPU_FEATURE_USABLE (RTM)),
__wcslen_avx2_rtm)
IFUNC_IMPL_ADD (array, i, wcslen,
@@ -662,10 +668,12 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
/* Support sysdeps/x86_64/multiarch/wcsnlen.c. */
IFUNC_IMPL (i, name, wcsnlen,
IFUNC_IMPL_ADD (array, i, wcsnlen,
- CPU_FEATURE_USABLE (AVX2),
+ (CPU_FEATURE_USABLE (AVX2)
+ && CPU_FEATURE_USABLE (BMI2)),
__wcsnlen_avx2)
IFUNC_IMPL_ADD (array, i, wcsnlen,
(CPU_FEATURE_USABLE (AVX2)
+ && CPU_FEATURE_USABLE (BMI2)
&& CPU_FEATURE_USABLE (RTM)),
__wcsnlen_avx2_rtm)
IFUNC_IMPL_ADD (array, i, wcsnlen,
diff --git a/sysdeps/x86_64/multiarch/strlen-avx2.S b/sysdeps/x86_64/multiarch/strlen-avx2.S
index 1caae9e6bc..eb92f587b4 100644
--- a/sysdeps/x86_64/multiarch/strlen-avx2.S
+++ b/sysdeps/x86_64/multiarch/strlen-avx2.S
@@ -27,9 +27,11 @@
# ifdef USE_AS_WCSLEN
# define VPCMPEQ vpcmpeqd
# define VPMINU vpminud
+# define CHAR_SIZE 4
# else
# define VPCMPEQ vpcmpeqb
# define VPMINU vpminub
+# define CHAR_SIZE 1
# endif
# ifndef VZEROUPPER
@@ -41,349 +43,462 @@
# endif
# define VEC_SIZE 32
+# define PAGE_SIZE 4096
+# define LOG_PAGE_SIZE 12
.section SECTION(.text),"ax",@progbits
ENTRY (STRLEN)
# ifdef USE_AS_STRNLEN
- /* Check for zero length. */
+ /* Check zero length. */
test %RSI_LP, %RSI_LP
jz L(zero)
+ /* Store max len in R8_LP before adjusting if using WCSLEN. */
+ mov %RSI_LP, %R8_LP
# ifdef USE_AS_WCSLEN
shl $2, %RSI_LP
# elif defined __ILP32__
/* Clear the upper 32 bits. */
movl %esi, %esi
# endif
- mov %RSI_LP, %R8_LP
# endif
- movl %edi, %ecx
+ movl %edi, %eax
movq %rdi, %rdx
vpxor %xmm0, %xmm0, %xmm0
-
+ /* Shift left eax to clear all bits not relevant to page cross
+ check. This saves 2 bytes of code as opposed to using andl with
+ PAGE_SIZE - 1. Then compare with PAGE_SIZE - VEC_SIZE shifted
+ left by the same amount (an imm32 either way). */
+ sall $(32 - LOG_PAGE_SIZE), %eax
/* Check if we may cross page boundary with one vector load. */
- andl $(2 * VEC_SIZE - 1), %ecx
- cmpl $VEC_SIZE, %ecx
- ja L(cros_page_boundary)
+ cmpl $((PAGE_SIZE - VEC_SIZE) << (32 - LOG_PAGE_SIZE)), %eax
+ ja L(cross_page_boundary)
/* Check the first VEC_SIZE bytes. */
- VPCMPEQ (%rdi), %ymm0, %ymm1
- vpmovmskb %ymm1, %eax
- testl %eax, %eax
-
+ VPCMPEQ (%rdi), %ymm0, %ymm1
+ vpmovmskb %ymm1, %eax
# ifdef USE_AS_STRNLEN
- jnz L(first_vec_x0_check)
- /* Adjust length and check the end of data. */
- subq $VEC_SIZE, %rsi
- jbe L(max)
-# else
- jnz L(first_vec_x0)
+ /* If length < VEC_SIZE handle special. */
+ cmpq $VEC_SIZE, %rsi
+ jbe L(first_vec_x0)
# endif
-
- /* Align data for aligned loads in the loop. */
- addq $VEC_SIZE, %rdi
- andl $(VEC_SIZE - 1), %ecx
- andq $-VEC_SIZE, %rdi
+ /* If empty continue to aligned_more. Otherwise return bit
+ position of first match. */
+ testl %eax, %eax
+ jz L(aligned_more)
+ tzcntl %eax, %eax
+# ifdef USE_AS_WCSLEN
+ shrl $2, %eax
+# endif
+ VZEROUPPER_RETURN
# ifdef USE_AS_STRNLEN
- /* Adjust length. */
- addq %rcx, %rsi
+L(zero):
+ xorl %eax, %eax
+ ret
- subq $(VEC_SIZE * 4), %rsi
- jbe L(last_4x_vec_or_less)
+ .p2align 4
+L(first_vec_x0):
+ /* Set bit for max len so that tzcnt will return min of max len
+ and position of first match. */
+ btsq %rsi, %rax
+ tzcntl %eax, %eax
+# ifdef USE_AS_WCSLEN
+ shrl $2, %eax
+# endif
+ VZEROUPPER_RETURN
# endif
- jmp L(more_4x_vec)
.p2align 4
-L(cros_page_boundary):
- andl $(VEC_SIZE - 1), %ecx
- andq $-VEC_SIZE, %rdi
- VPCMPEQ (%rdi), %ymm0, %ymm1
- vpmovmskb %ymm1, %eax
- /* Remove the leading bytes. */
- sarl %cl, %eax
- testl %eax, %eax
- jz L(aligned_more)
+L(first_vec_x1):
tzcntl %eax, %eax
+ /* Safe to use 32 bit instructions as these are only called for
+ size = [1, 159]. */
# ifdef USE_AS_STRNLEN
- /* Check the end of data. */
- cmpq %rax, %rsi
- jbe L(max)
+ /* Use ecx which was computed earlier to compute correct value.
+ */
+ subl $(VEC_SIZE * 4 + 1), %ecx
+ addl %ecx, %eax
+# else
+ subl %edx, %edi
+ incl %edi
+ addl %edi, %eax
# endif
- addq %rdi, %rax
- addq %rcx, %rax
- subq %rdx, %rax
# ifdef USE_AS_WCSLEN
- shrq $2, %rax
+ shrl $2, %eax
# endif
-L(return_vzeroupper):
- ZERO_UPPER_VEC_REGISTERS_RETURN
+ VZEROUPPER_RETURN
.p2align 4
-L(aligned_more):
+L(first_vec_x2):
+ tzcntl %eax, %eax
+ /* Safe to use 32 bit instructions as these are only called for
+ size = [1, 159]. */
# ifdef USE_AS_STRNLEN
- /* "rcx" is less than VEC_SIZE. Calculate "rdx + rcx - VEC_SIZE"
- with "rdx - (VEC_SIZE - rcx)" instead of "(rdx + rcx) - VEC_SIZE"
- to void possible addition overflow. */
- negq %rcx
- addq $VEC_SIZE, %rcx
-
- /* Check the end of data. */
- subq %rcx, %rsi
- jbe L(max)
+ /* Use ecx which was computed earlier to compute correct value.
+ */
+ subl $(VEC_SIZE * 3 + 1), %ecx
+ addl %ecx, %eax
+# else
+ subl %edx, %edi
+ addl $(VEC_SIZE + 1), %edi
+ addl %edi, %eax
# endif
+# ifdef USE_AS_WCSLEN
+ shrl $2, %eax
+# endif
+ VZEROUPPER_RETURN
- addq $VEC_SIZE, %rdi
+ .p2align 4
+L(first_vec_x3):
+ tzcntl %eax, %eax
+ /* Safe to use 32 bit instructions as these are only called for
+ size = [1, 159]. */
+# ifdef USE_AS_STRNLEN
+ /* Use ecx which was computed earlier to compute correct value.
+ */
+ subl $(VEC_SIZE * 2 + 1), %ecx
+ addl %ecx, %eax
+# else
+ subl %edx, %edi
+ addl $(VEC_SIZE * 2 + 1), %edi
+ addl %edi, %eax
+# endif
+# ifdef USE_AS_WCSLEN
+ shrl $2, %eax
+# endif
+ VZEROUPPER_RETURN
+ .p2align 4
+L(first_vec_x4):
+ tzcntl %eax, %eax
+ /* Safe to use 32 bit instructions as these are only called for
+ size = [1, 159]. */
# ifdef USE_AS_STRNLEN
- subq $(VEC_SIZE * 4), %rsi
- jbe L(last_4x_vec_or_less)
+ /* Use ecx which was computed earlier to compute correct value.
+ */
+ subl $(VEC_SIZE + 1), %ecx
+ addl %ecx, %eax
+# else
+ subl %edx, %edi
+ addl $(VEC_SIZE * 3 + 1), %edi
+ addl %edi, %eax
# endif
+# ifdef USE_AS_WCSLEN
+ shrl $2, %eax
+# endif
+ VZEROUPPER_RETURN
-L(more_4x_vec):
+ .p2align 5
+L(aligned_more):
+ /* Align data to VEC_SIZE - 1. This is the same number of
+ instructions as using andq with -VEC_SIZE but saves 4 bytes of
+ code on the x4 check. */
+ orq $(VEC_SIZE - 1), %rdi
+L(cross_page_continue):
/* Check the first 4 * VEC_SIZE. Only one VEC_SIZE at a time
since data is only aligned to VEC_SIZE. */
- VPCMPEQ (%rdi), %ymm0, %ymm1
- vpmovmskb %ymm1, %eax
- testl %eax, %eax
- jnz L(first_vec_x0)
-
- VPCMPEQ VEC_SIZE(%rdi), %ymm0, %ymm1
- vpmovmskb %ymm1, %eax
+# ifdef USE_AS_STRNLEN
+ /* + 1 because rdi is aligned to VEC_SIZE - 1. + CHAR_SIZE because
+ it simplies the logic in last_4x_vec_or_less. */
+ leaq (VEC_SIZE * 4 + CHAR_SIZE + 1)(%rdi), %rcx
+ subq %rdx, %rcx
+# endif
+ /* Load first VEC regardless. */
+ VPCMPEQ 1(%rdi), %ymm0, %ymm1
+# ifdef USE_AS_STRNLEN
+ /* Adjust length. If near end handle specially. */
+ subq %rcx, %rsi
+ jb L(last_4x_vec_or_less)
+# endif
+ vpmovmskb %ymm1, %eax
testl %eax, %eax
jnz L(first_vec_x1)
- VPCMPEQ (VEC_SIZE * 2)(%rdi), %ymm0, %ymm1
- vpmovmskb %ymm1, %eax
+ VPCMPEQ (VEC_SIZE + 1)(%rdi), %ymm0, %ymm1
+ vpmovmskb %ymm1, %eax
testl %eax, %eax
jnz L(first_vec_x2)
- VPCMPEQ (VEC_SIZE * 3)(%rdi), %ymm0, %ymm1
- vpmovmskb %ymm1, %eax
+ VPCMPEQ (VEC_SIZE * 2 + 1)(%rdi), %ymm0, %ymm1
+ vpmovmskb %ymm1, %eax
testl %eax, %eax
jnz L(first_vec_x3)
- addq $(VEC_SIZE * 4), %rdi
-
-# ifdef USE_AS_STRNLEN
- subq $(VEC_SIZE * 4), %rsi
- jbe L(last_4x_vec_or_less)
-# endif
-
- /* Align data to 4 * VEC_SIZE. */
- movq %rdi, %rcx
- andl $(4 * VEC_SIZE - 1), %ecx
- andq $-(4 * VEC_SIZE), %rdi
+ VPCMPEQ (VEC_SIZE * 3 + 1)(%rdi), %ymm0, %ymm1
+ vpmovmskb %ymm1, %eax
+ testl %eax, %eax
+ jnz L(first_vec_x4)
+ /* Align data to VEC_SIZE * 4 - 1. */
# ifdef USE_AS_STRNLEN
- /* Adjust length. */
+ /* Before adjusting length check if at last VEC_SIZE * 4. */
+ cmpq $(VEC_SIZE * 4 - 1), %rsi
+ jbe L(last_4x_vec_or_less_load)
+ incq %rdi
+ movl %edi, %ecx
+ orq $(VEC_SIZE * 4 - 1), %rdi
+ andl $(VEC_SIZE * 4 - 1), %ecx
+ /* Readjust length. */
addq %rcx, %rsi
+# else
+ incq %rdi
+ orq $(VEC_SIZE * 4 - 1), %rdi
# endif
-
+ /* Compare 4 * VEC at a time forward. */
.p2align 4
L(loop_4x_vec):
- /* Compare 4 * VEC at a time forward. */
- vmovdqa (%rdi), %ymm1
- vmovdqa VEC_SIZE(%rdi), %ymm2
- vmovdqa (VEC_SIZE * 2)(%rdi), %ymm3
- vmovdqa (VEC_SIZE * 3)(%rdi), %ymm4
- VPMINU %ymm1, %ymm2, %ymm5
- VPMINU %ymm3, %ymm4, %ymm6
- VPMINU %ymm5, %ymm6, %ymm5
-
- VPCMPEQ %ymm5, %ymm0, %ymm5
- vpmovmskb %ymm5, %eax
- testl %eax, %eax
- jnz L(4x_vec_end)
-
- addq $(VEC_SIZE * 4), %rdi
-
-# ifndef USE_AS_STRNLEN
- jmp L(loop_4x_vec)
-# else
+# ifdef USE_AS_STRNLEN
+ /* Break if at end of length. */
subq $(VEC_SIZE * 4), %rsi
- ja L(loop_4x_vec)
-
-L(last_4x_vec_or_less):
- /* Less than 4 * VEC and aligned to VEC_SIZE. */
- addl $(VEC_SIZE * 2), %esi
- jle L(last_2x_vec)
+ jb L(last_4x_vec_or_less_cmpeq)
+# endif
+ /* Save some code size by microfusing VPMINU with the load. Since
+ the matches in ymm2/ymm4 can only be returned if there where no
+ matches in ymm1/ymm3 respectively there is no issue with overlap.
+ */
+ vmovdqa 1(%rdi), %ymm1
+ VPMINU (VEC_SIZE + 1)(%rdi), %ymm1, %ymm2
+ vmovdqa (VEC_SIZE * 2 + 1)(%rdi), %ymm3
+ VPMINU (VEC_SIZE * 3 + 1)(%rdi), %ymm3, %ymm4
+
+ VPMINU %ymm2, %ymm4, %ymm5
+ VPCMPEQ %ymm5, %ymm0, %ymm5
+ vpmovmskb %ymm5, %ecx
- VPCMPEQ (%rdi), %ymm0, %ymm1
- vpmovmskb %ymm1, %eax
- testl %eax, %eax
- jnz L(first_vec_x0)
+ subq $-(VEC_SIZE * 4), %rdi
+ testl %ecx, %ecx
+ jz L(loop_4x_vec)
- VPCMPEQ VEC_SIZE(%rdi), %ymm0, %ymm1
- vpmovmskb %ymm1, %eax
- testl %eax, %eax
- jnz L(first_vec_x1)
- VPCMPEQ (VEC_SIZE * 2)(%rdi), %ymm0, %ymm1
- vpmovmskb %ymm1, %eax
+ VPCMPEQ %ymm1, %ymm0, %ymm1
+ vpmovmskb %ymm1, %eax
+ subq %rdx, %rdi
testl %eax, %eax
+ jnz L(last_vec_return_x0)
- jnz L(first_vec_x2_check)
- subl $VEC_SIZE, %esi
- jle L(max)
-
- VPCMPEQ (VEC_SIZE * 3)(%rdi), %ymm0, %ymm1
- vpmovmskb %ymm1, %eax
+ VPCMPEQ %ymm2, %ymm0, %ymm2
+ vpmovmskb %ymm2, %eax
testl %eax, %eax
-
- jnz L(first_vec_x3_check)
- movq %r8, %rax
-# ifdef USE_AS_WCSLEN
+ jnz L(last_vec_return_x1)
+
+ /* Combine last 2 VEC. */
+ VPCMPEQ %ymm3, %ymm0, %ymm3
+ vpmovmskb %ymm3, %eax
+ /* rcx has combined result from all 4 VEC. It will only be used if
+ the first 3 other VEC all did not contain a match. */
+ salq $32, %rcx
+ orq %rcx, %rax
+ tzcntq %rax, %rax
+ subq $(VEC_SIZE * 2 - 1), %rdi
+ addq %rdi, %rax
+# ifdef USE_AS_WCSLEN
shrq $2, %rax
-# endif
+# endif
VZEROUPPER_RETURN
+
+# ifdef USE_AS_STRNLEN
.p2align 4
-L(last_2x_vec):
- addl $(VEC_SIZE * 2), %esi
- VPCMPEQ (%rdi), %ymm0, %ymm1
- vpmovmskb %ymm1, %eax
- testl %eax, %eax
+L(last_4x_vec_or_less_load):
+ /* Depending on entry adjust rdi / prepare first VEC in ymm1. */
+ subq $-(VEC_SIZE * 4), %rdi
+L(last_4x_vec_or_less_cmpeq):
+ VPCMPEQ 1(%rdi), %ymm0, %ymm1
+L(last_4x_vec_or_less):
- jnz L(first_vec_x0_check)
- subl $VEC_SIZE, %esi
- jle L(max)
+ vpmovmskb %ymm1, %eax
+ /* If remaining length > VEC_SIZE * 2. This works if esi is off by
+ VEC_SIZE * 4. */
+ testl $(VEC_SIZE * 2), %esi
+ jnz L(last_4x_vec)
- VPCMPEQ VEC_SIZE(%rdi), %ymm0, %ymm1
- vpmovmskb %ymm1, %eax
+ /* length may have been negative or positive by an offset of
+ VEC_SIZE * 4 depending on where this was called from. This fixes
+ that. */
+ andl $(VEC_SIZE * 4 - 1), %esi
testl %eax, %eax
- jnz L(first_vec_x1_check)
- movq %r8, %rax
-# ifdef USE_AS_WCSLEN
- shrq $2, %rax
-# endif
- VZEROUPPER_RETURN
+ jnz L(last_vec_x1_check)
- .p2align 4
-L(first_vec_x0_check):
+ subl $VEC_SIZE, %esi
+ jb L(max)
+
+ VPCMPEQ (VEC_SIZE + 1)(%rdi), %ymm0, %ymm1
+ vpmovmskb %ymm1, %eax
tzcntl %eax, %eax
/* Check the end of data. */
- cmpq %rax, %rsi
- jbe L(max)
+ cmpl %eax, %esi
+ jb L(max)
+ subq %rdx, %rdi
+ addl $(VEC_SIZE + 1), %eax
addq %rdi, %rax
- subq %rdx, %rax
# ifdef USE_AS_WCSLEN
shrq $2, %rax
# endif
VZEROUPPER_RETURN
+# endif
.p2align 4
-L(first_vec_x1_check):
+L(last_vec_return_x0):
tzcntl %eax, %eax
- /* Check the end of data. */
- cmpq %rax, %rsi
- jbe L(max)
- addq $VEC_SIZE, %rax
+ subq $(VEC_SIZE * 4 - 1), %rdi
addq %rdi, %rax
- subq %rdx, %rax
-# ifdef USE_AS_WCSLEN
+# ifdef USE_AS_WCSLEN
shrq $2, %rax
-# endif
+# endif
VZEROUPPER_RETURN
.p2align 4
-L(first_vec_x2_check):
+L(last_vec_return_x1):
tzcntl %eax, %eax
- /* Check the end of data. */
- cmpq %rax, %rsi
- jbe L(max)
- addq $(VEC_SIZE * 2), %rax
+ subq $(VEC_SIZE * 3 - 1), %rdi
addq %rdi, %rax
- subq %rdx, %rax
-# ifdef USE_AS_WCSLEN
+# ifdef USE_AS_WCSLEN
shrq $2, %rax
-# endif
+# endif
VZEROUPPER_RETURN
+# ifdef USE_AS_STRNLEN
.p2align 4
-L(first_vec_x3_check):
+L(last_vec_x1_check):
+
tzcntl %eax, %eax
/* Check the end of data. */
- cmpq %rax, %rsi
- jbe L(max)
- addq $(VEC_SIZE * 3), %rax
+ cmpl %eax, %esi
+ jb L(max)
+ subq %rdx, %rdi
+ incl %eax
addq %rdi, %rax
- subq %rdx, %rax
# ifdef USE_AS_WCSLEN
shrq $2, %rax
# endif
VZEROUPPER_RETURN
- .p2align 4
L(max):
movq %r8, %rax
+ VZEROUPPER_RETURN
+
+ .p2align 4
+L(last_4x_vec):
+ /* Test first 2x VEC normally. */
+ testl %eax, %eax
+ jnz L(last_vec_x1)
+
+ VPCMPEQ (VEC_SIZE + 1)(%rdi), %ymm0, %ymm1
+ vpmovmskb %ymm1, %eax
+ testl %eax, %eax
+ jnz L(last_vec_x2)
+
+ /* Normalize length. */
+ andl $(VEC_SIZE * 4 - 1), %esi
+ VPCMPEQ (VEC_SIZE * 2 + 1)(%rdi), %ymm0, %ymm1
+ vpmovmskb %ymm1, %eax
+ testl %eax, %eax
+ jnz L(last_vec_x3)
+
+ subl $(VEC_SIZE * 3), %esi
+ jb L(max)
+
+ VPCMPEQ (VEC_SIZE * 3 + 1)(%rdi), %ymm0, %ymm1
+ vpmovmskb %ymm1, %eax
+ tzcntl %eax, %eax
+ /* Check the end of data. */
+ cmpl %eax, %esi
+ jb L(max)
+ subq %rdx, %rdi
+ addl $(VEC_SIZE * 3 + 1), %eax
+ addq %rdi, %rax
# ifdef USE_AS_WCSLEN
shrq $2, %rax
# endif
VZEROUPPER_RETURN
- .p2align 4
-L(zero):
- xorl %eax, %eax
- ret
-# endif
.p2align 4
-L(first_vec_x0):
+L(last_vec_x1):
+ /* essentially duplicates of first_vec_x1 but use 64 bit
+ instructions. */
tzcntl %eax, %eax
+ subq %rdx, %rdi
+ incl %eax
addq %rdi, %rax
- subq %rdx, %rax
-# ifdef USE_AS_WCSLEN
+# ifdef USE_AS_WCSLEN
shrq $2, %rax
-# endif
+# endif
VZEROUPPER_RETURN
.p2align 4
-L(first_vec_x1):
+L(last_vec_x2):
+ /* essentially duplicates of first_vec_x1 but use 64 bit
+ instructions. */
tzcntl %eax, %eax
- addq $VEC_SIZE, %rax
+ subq %rdx, %rdi
+ addl $(VEC_SIZE + 1), %eax
addq %rdi, %rax
- subq %rdx, %rax
-# ifdef USE_AS_WCSLEN
+# ifdef USE_AS_WCSLEN
shrq $2, %rax
-# endif
+# endif
VZEROUPPER_RETURN
.p2align 4
-L(first_vec_x2):
+L(last_vec_x3):
tzcntl %eax, %eax
- addq $(VEC_SIZE * 2), %rax
+ subl $(VEC_SIZE * 2), %esi
+ /* Check the end of data. */
+ cmpl %eax, %esi
+ jb L(max_end)
+ subq %rdx, %rdi
+ addl $(VEC_SIZE * 2 + 1), %eax
addq %rdi, %rax
- subq %rdx, %rax
-# ifdef USE_AS_WCSLEN
+# ifdef USE_AS_WCSLEN
shrq $2, %rax
-# endif
+# endif
+ VZEROUPPER_RETURN
+L(max_end):
+ movq %r8, %rax
VZEROUPPER_RETURN
+# endif
+ /* Cold case for crossing page with first load. */
.p2align 4
-L(4x_vec_end):
- VPCMPEQ %ymm1, %ymm0, %ymm1
- vpmovmskb %ymm1, %eax
- testl %eax, %eax
- jnz L(first_vec_x0)
- VPCMPEQ %ymm2, %ymm0, %ymm2
- vpmovmskb %ymm2, %eax
+L(cross_page_boundary):
+ /* Align data to VEC_SIZE - 1. */
+ orq $(VEC_SIZE - 1), %rdi
+ VPCMPEQ -(VEC_SIZE - 1)(%rdi), %ymm0, %ymm1
+ vpmovmskb %ymm1, %eax
+ /* Remove the leading bytes. sarxl only uses bits [5:0] of COUNT
+ so no need to manually mod rdx. */
+ sarxl %edx, %eax, %eax
+# ifdef USE_AS_STRNLEN
testl %eax, %eax
- jnz L(first_vec_x1)
- VPCMPEQ %ymm3, %ymm0, %ymm3
- vpmovmskb %ymm3, %eax
+ jnz L(cross_page_less_vec)
+ leaq 1(%rdi), %rcx
+ subq %rdx, %rcx
+ /* Check length. */
+ cmpq %rsi, %rcx
+ jb L(cross_page_continue)
+ movq %r8, %rax
+# else
testl %eax, %eax
- jnz L(first_vec_x2)
- VPCMPEQ %ymm4, %ymm0, %ymm4
- vpmovmskb %ymm4, %eax
-L(first_vec_x3):
+ jz L(cross_page_continue)
tzcntl %eax, %eax
- addq $(VEC_SIZE * 3), %rax
- addq %rdi, %rax
- subq %rdx, %rax
-# ifdef USE_AS_WCSLEN
- shrq $2, %rax
+# ifdef USE_AS_WCSLEN
+ shrl $2, %eax
+# endif
# endif
+L(return_vzeroupper):
+ ZERO_UPPER_VEC_REGISTERS_RETURN
+
+# ifdef USE_AS_STRNLEN
+ .p2align 4
+L(cross_page_less_vec):
+ tzcntl %eax, %eax
+ cmpq %rax, %rsi
+ cmovb %esi, %eax
+# ifdef USE_AS_WCSLEN
+ shrl $2, %eax
+# endif
VZEROUPPER_RETURN
+# endif
END (STRLEN)
#endif
--
2.29.2
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: [PATCH v4 2/2] x86: Optimize strlen-avx2.S
2021-04-19 18:43 ` [PATCH v4 2/2] x86: Optimize strlen-avx2.S Noah Goldstein
@ 2021-04-19 22:16 ` H.J. Lu
2021-04-19 23:39 ` Noah Goldstein
0 siblings, 1 reply; 6+ messages in thread
From: H.J. Lu @ 2021-04-19 22:16 UTC (permalink / raw)
To: Noah Goldstein; +Cc: libc-alpha, carlos
On Mon, Apr 19, 2021 at 02:43:45PM -0400, Noah Goldstein wrote:
> No bug. This commit optimizes strlen-avx2.S. The optimizations are
> mostly small things but they add up to roughly 10-30% performance
> improvement for strlen. The results for strnlen are bit more
> ambiguous. test-strlen, test-strnlen, test-wcslen, and test-wcsnlen
> are all passing.
>
> Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com>
> ---
> sysdeps/x86_64/multiarch/ifunc-impl-list.c | 16 +-
> sysdeps/x86_64/multiarch/strlen-avx2.S | 535 +++++++++++++--------
> 2 files changed, 337 insertions(+), 214 deletions(-)
>
> diff --git a/sysdeps/x86_64/multiarch/ifunc-impl-list.c b/sysdeps/x86_64/multiarch/ifunc-impl-list.c
> index 0b0927b124..4613bead7c 100644
> --- a/sysdeps/x86_64/multiarch/ifunc-impl-list.c
> +++ b/sysdeps/x86_64/multiarch/ifunc-impl-list.c
> @@ -285,10 +285,12 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
> /* Support sysdeps/x86_64/multiarch/strlen.c. */
> IFUNC_IMPL (i, name, strlen,
> IFUNC_IMPL_ADD (array, i, strlen,
> - CPU_FEATURE_USABLE (AVX2),
> + (CPU_FEATURE_USABLE (AVX2)
> + && CPU_FEATURE_USABLE (BMI2)),
> __strlen_avx2)
> IFUNC_IMPL_ADD (array, i, strlen,
> (CPU_FEATURE_USABLE (AVX2)
> + && CPU_FEATURE_USABLE (BMI2)
> && CPU_FEATURE_USABLE (RTM)),
> __strlen_avx2_rtm)
> IFUNC_IMPL_ADD (array, i, strlen,
> @@ -301,10 +303,12 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
> /* Support sysdeps/x86_64/multiarch/strnlen.c. */
> IFUNC_IMPL (i, name, strnlen,
> IFUNC_IMPL_ADD (array, i, strnlen,
> - CPU_FEATURE_USABLE (AVX2),
> + (CPU_FEATURE_USABLE (AVX2)
> + && CPU_FEATURE_USABLE (BMI2)),
> __strnlen_avx2)
> IFUNC_IMPL_ADD (array, i, strnlen,
> (CPU_FEATURE_USABLE (AVX2)
> + && CPU_FEATURE_USABLE (BMI2)
> && CPU_FEATURE_USABLE (RTM)),
> __strnlen_avx2_rtm)
> IFUNC_IMPL_ADD (array, i, strnlen,
> @@ -646,10 +650,12 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
> /* Support sysdeps/x86_64/multiarch/wcslen.c. */
> IFUNC_IMPL (i, name, wcslen,
> IFUNC_IMPL_ADD (array, i, wcslen,
> - CPU_FEATURE_USABLE (AVX2),
> + (CPU_FEATURE_USABLE (AVX2)
> + && CPU_FEATURE_USABLE (BMI2)),
> __wcslen_avx2)
> IFUNC_IMPL_ADD (array, i, wcslen,
> (CPU_FEATURE_USABLE (AVX2)
> + && CPU_FEATURE_USABLE (BMI2)
> && CPU_FEATURE_USABLE (RTM)),
> __wcslen_avx2_rtm)
> IFUNC_IMPL_ADD (array, i, wcslen,
> @@ -662,10 +668,12 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
> /* Support sysdeps/x86_64/multiarch/wcsnlen.c. */
> IFUNC_IMPL (i, name, wcsnlen,
> IFUNC_IMPL_ADD (array, i, wcsnlen,
> - CPU_FEATURE_USABLE (AVX2),
> + (CPU_FEATURE_USABLE (AVX2)
> + && CPU_FEATURE_USABLE (BMI2)),
> __wcsnlen_avx2)
> IFUNC_IMPL_ADD (array, i, wcsnlen,
> (CPU_FEATURE_USABLE (AVX2)
> + && CPU_FEATURE_USABLE (BMI2)
> && CPU_FEATURE_USABLE (RTM)),
> __wcsnlen_avx2_rtm)
> IFUNC_IMPL_ADD (array, i, wcsnlen,
> diff --git a/sysdeps/x86_64/multiarch/strlen-avx2.S b/sysdeps/x86_64/multiarch/strlen-avx2.S
> index 1caae9e6bc..eb92f587b4 100644
> --- a/sysdeps/x86_64/multiarch/strlen-avx2.S
> +++ b/sysdeps/x86_64/multiarch/strlen-avx2.S
> @@ -27,9 +27,11 @@
> # ifdef USE_AS_WCSLEN
> # define VPCMPEQ vpcmpeqd
> # define VPMINU vpminud
> +# define CHAR_SIZE 4
> # else
> # define VPCMPEQ vpcmpeqb
> # define VPMINU vpminub
> +# define CHAR_SIZE 1
> # endif
>
> # ifndef VZEROUPPER
> @@ -41,349 +43,462 @@
> # endif
>
> # define VEC_SIZE 32
> +# define PAGE_SIZE 4096
> +# define LOG_PAGE_SIZE 12
>
> .section SECTION(.text),"ax",@progbits
> ENTRY (STRLEN)
> # ifdef USE_AS_STRNLEN
> - /* Check for zero length. */
> + /* Check zero length. */
> test %RSI_LP, %RSI_LP
> jz L(zero)
> + /* Store max len in R8_LP before adjusting if using WCSLEN. */
> + mov %RSI_LP, %R8_LP
> # ifdef USE_AS_WCSLEN
> shl $2, %RSI_LP
> # elif defined __ILP32__
> /* Clear the upper 32 bits. */
> movl %esi, %esi
> # endif
> - mov %RSI_LP, %R8_LP
> # endif
> - movl %edi, %ecx
> + movl %edi, %eax
> movq %rdi, %rdx
> vpxor %xmm0, %xmm0, %xmm0
> -
> + /* Shift left eax to clear all bits not relevant to page cross
> + check. This saves 2 bytes of code as opposed to using andl with
> + PAGE_SIZE - 1. Then compare with PAGE_SIZE - VEC_SIZE shifted
> + left by the same amount (an imm32 either way). */
> + sall $(32 - LOG_PAGE_SIZE), %eax
I prefer AND over SAL for better throughput.
H.J.
> /* Check if we may cross page boundary with one vector load. */
> - andl $(2 * VEC_SIZE - 1), %ecx
> - cmpl $VEC_SIZE, %ecx
> - ja L(cros_page_boundary)
> + cmpl $((PAGE_SIZE - VEC_SIZE) << (32 - LOG_PAGE_SIZE)), %eax
> + ja L(cross_page_boundary)
>
> /* Check the first VEC_SIZE bytes. */
> - VPCMPEQ (%rdi), %ymm0, %ymm1
> - vpmovmskb %ymm1, %eax
> - testl %eax, %eax
> -
> + VPCMPEQ (%rdi), %ymm0, %ymm1
> + vpmovmskb %ymm1, %eax
> # ifdef USE_AS_STRNLEN
> - jnz L(first_vec_x0_check)
> - /* Adjust length and check the end of data. */
> - subq $VEC_SIZE, %rsi
> - jbe L(max)
> -# else
> - jnz L(first_vec_x0)
> + /* If length < VEC_SIZE handle special. */
> + cmpq $VEC_SIZE, %rsi
> + jbe L(first_vec_x0)
> # endif
> -
> - /* Align data for aligned loads in the loop. */
> - addq $VEC_SIZE, %rdi
> - andl $(VEC_SIZE - 1), %ecx
> - andq $-VEC_SIZE, %rdi
> + /* If empty continue to aligned_more. Otherwise return bit
> + position of first match. */
> + testl %eax, %eax
> + jz L(aligned_more)
> + tzcntl %eax, %eax
> +# ifdef USE_AS_WCSLEN
> + shrl $2, %eax
> +# endif
> + VZEROUPPER_RETURN
>
> # ifdef USE_AS_STRNLEN
> - /* Adjust length. */
> - addq %rcx, %rsi
> +L(zero):
> + xorl %eax, %eax
> + ret
>
> - subq $(VEC_SIZE * 4), %rsi
> - jbe L(last_4x_vec_or_less)
> + .p2align 4
> +L(first_vec_x0):
> + /* Set bit for max len so that tzcnt will return min of max len
> + and position of first match. */
> + btsq %rsi, %rax
> + tzcntl %eax, %eax
> +# ifdef USE_AS_WCSLEN
> + shrl $2, %eax
> +# endif
> + VZEROUPPER_RETURN
> # endif
> - jmp L(more_4x_vec)
>
> .p2align 4
> -L(cros_page_boundary):
> - andl $(VEC_SIZE - 1), %ecx
> - andq $-VEC_SIZE, %rdi
> - VPCMPEQ (%rdi), %ymm0, %ymm1
> - vpmovmskb %ymm1, %eax
> - /* Remove the leading bytes. */
> - sarl %cl, %eax
> - testl %eax, %eax
> - jz L(aligned_more)
> +L(first_vec_x1):
> tzcntl %eax, %eax
> + /* Safe to use 32 bit instructions as these are only called for
> + size = [1, 159]. */
> # ifdef USE_AS_STRNLEN
> - /* Check the end of data. */
> - cmpq %rax, %rsi
> - jbe L(max)
> + /* Use ecx which was computed earlier to compute correct value.
> + */
> + subl $(VEC_SIZE * 4 + 1), %ecx
> + addl %ecx, %eax
> +# else
> + subl %edx, %edi
> + incl %edi
> + addl %edi, %eax
> # endif
> - addq %rdi, %rax
> - addq %rcx, %rax
> - subq %rdx, %rax
> # ifdef USE_AS_WCSLEN
> - shrq $2, %rax
> + shrl $2, %eax
> # endif
> -L(return_vzeroupper):
> - ZERO_UPPER_VEC_REGISTERS_RETURN
> + VZEROUPPER_RETURN
>
> .p2align 4
> -L(aligned_more):
> +L(first_vec_x2):
> + tzcntl %eax, %eax
> + /* Safe to use 32 bit instructions as these are only called for
> + size = [1, 159]. */
> # ifdef USE_AS_STRNLEN
> - /* "rcx" is less than VEC_SIZE. Calculate "rdx + rcx - VEC_SIZE"
> - with "rdx - (VEC_SIZE - rcx)" instead of "(rdx + rcx) - VEC_SIZE"
> - to void possible addition overflow. */
> - negq %rcx
> - addq $VEC_SIZE, %rcx
> -
> - /* Check the end of data. */
> - subq %rcx, %rsi
> - jbe L(max)
> + /* Use ecx which was computed earlier to compute correct value.
> + */
> + subl $(VEC_SIZE * 3 + 1), %ecx
> + addl %ecx, %eax
> +# else
> + subl %edx, %edi
> + addl $(VEC_SIZE + 1), %edi
> + addl %edi, %eax
> # endif
> +# ifdef USE_AS_WCSLEN
> + shrl $2, %eax
> +# endif
> + VZEROUPPER_RETURN
>
> - addq $VEC_SIZE, %rdi
> + .p2align 4
> +L(first_vec_x3):
> + tzcntl %eax, %eax
> + /* Safe to use 32 bit instructions as these are only called for
> + size = [1, 159]. */
> +# ifdef USE_AS_STRNLEN
> + /* Use ecx which was computed earlier to compute correct value.
> + */
> + subl $(VEC_SIZE * 2 + 1), %ecx
> + addl %ecx, %eax
> +# else
> + subl %edx, %edi
> + addl $(VEC_SIZE * 2 + 1), %edi
> + addl %edi, %eax
> +# endif
> +# ifdef USE_AS_WCSLEN
> + shrl $2, %eax
> +# endif
> + VZEROUPPER_RETURN
>
> + .p2align 4
> +L(first_vec_x4):
> + tzcntl %eax, %eax
> + /* Safe to use 32 bit instructions as these are only called for
> + size = [1, 159]. */
> # ifdef USE_AS_STRNLEN
> - subq $(VEC_SIZE * 4), %rsi
> - jbe L(last_4x_vec_or_less)
> + /* Use ecx which was computed earlier to compute correct value.
> + */
> + subl $(VEC_SIZE + 1), %ecx
> + addl %ecx, %eax
> +# else
> + subl %edx, %edi
> + addl $(VEC_SIZE * 3 + 1), %edi
> + addl %edi, %eax
> # endif
> +# ifdef USE_AS_WCSLEN
> + shrl $2, %eax
> +# endif
> + VZEROUPPER_RETURN
>
> -L(more_4x_vec):
> + .p2align 5
> +L(aligned_more):
> + /* Align data to VEC_SIZE - 1. This is the same number of
> + instructions as using andq with -VEC_SIZE but saves 4 bytes of
> + code on the x4 check. */
> + orq $(VEC_SIZE - 1), %rdi
> +L(cross_page_continue):
> /* Check the first 4 * VEC_SIZE. Only one VEC_SIZE at a time
> since data is only aligned to VEC_SIZE. */
> - VPCMPEQ (%rdi), %ymm0, %ymm1
> - vpmovmskb %ymm1, %eax
> - testl %eax, %eax
> - jnz L(first_vec_x0)
> -
> - VPCMPEQ VEC_SIZE(%rdi), %ymm0, %ymm1
> - vpmovmskb %ymm1, %eax
> +# ifdef USE_AS_STRNLEN
> + /* + 1 because rdi is aligned to VEC_SIZE - 1. + CHAR_SIZE because
> + it simplies the logic in last_4x_vec_or_less. */
> + leaq (VEC_SIZE * 4 + CHAR_SIZE + 1)(%rdi), %rcx
> + subq %rdx, %rcx
> +# endif
> + /* Load first VEC regardless. */
> + VPCMPEQ 1(%rdi), %ymm0, %ymm1
> +# ifdef USE_AS_STRNLEN
> + /* Adjust length. If near end handle specially. */
> + subq %rcx, %rsi
> + jb L(last_4x_vec_or_less)
> +# endif
> + vpmovmskb %ymm1, %eax
> testl %eax, %eax
> jnz L(first_vec_x1)
>
> - VPCMPEQ (VEC_SIZE * 2)(%rdi), %ymm0, %ymm1
> - vpmovmskb %ymm1, %eax
> + VPCMPEQ (VEC_SIZE + 1)(%rdi), %ymm0, %ymm1
> + vpmovmskb %ymm1, %eax
> testl %eax, %eax
> jnz L(first_vec_x2)
>
> - VPCMPEQ (VEC_SIZE * 3)(%rdi), %ymm0, %ymm1
> - vpmovmskb %ymm1, %eax
> + VPCMPEQ (VEC_SIZE * 2 + 1)(%rdi), %ymm0, %ymm1
> + vpmovmskb %ymm1, %eax
> testl %eax, %eax
> jnz L(first_vec_x3)
>
> - addq $(VEC_SIZE * 4), %rdi
> -
> -# ifdef USE_AS_STRNLEN
> - subq $(VEC_SIZE * 4), %rsi
> - jbe L(last_4x_vec_or_less)
> -# endif
> -
> - /* Align data to 4 * VEC_SIZE. */
> - movq %rdi, %rcx
> - andl $(4 * VEC_SIZE - 1), %ecx
> - andq $-(4 * VEC_SIZE), %rdi
> + VPCMPEQ (VEC_SIZE * 3 + 1)(%rdi), %ymm0, %ymm1
> + vpmovmskb %ymm1, %eax
> + testl %eax, %eax
> + jnz L(first_vec_x4)
>
> + /* Align data to VEC_SIZE * 4 - 1. */
> # ifdef USE_AS_STRNLEN
> - /* Adjust length. */
> + /* Before adjusting length check if at last VEC_SIZE * 4. */
> + cmpq $(VEC_SIZE * 4 - 1), %rsi
> + jbe L(last_4x_vec_or_less_load)
> + incq %rdi
> + movl %edi, %ecx
> + orq $(VEC_SIZE * 4 - 1), %rdi
> + andl $(VEC_SIZE * 4 - 1), %ecx
> + /* Readjust length. */
> addq %rcx, %rsi
> +# else
> + incq %rdi
> + orq $(VEC_SIZE * 4 - 1), %rdi
> # endif
> -
> + /* Compare 4 * VEC at a time forward. */
> .p2align 4
> L(loop_4x_vec):
> - /* Compare 4 * VEC at a time forward. */
> - vmovdqa (%rdi), %ymm1
> - vmovdqa VEC_SIZE(%rdi), %ymm2
> - vmovdqa (VEC_SIZE * 2)(%rdi), %ymm3
> - vmovdqa (VEC_SIZE * 3)(%rdi), %ymm4
> - VPMINU %ymm1, %ymm2, %ymm5
> - VPMINU %ymm3, %ymm4, %ymm6
> - VPMINU %ymm5, %ymm6, %ymm5
> -
> - VPCMPEQ %ymm5, %ymm0, %ymm5
> - vpmovmskb %ymm5, %eax
> - testl %eax, %eax
> - jnz L(4x_vec_end)
> -
> - addq $(VEC_SIZE * 4), %rdi
> -
> -# ifndef USE_AS_STRNLEN
> - jmp L(loop_4x_vec)
> -# else
> +# ifdef USE_AS_STRNLEN
> + /* Break if at end of length. */
> subq $(VEC_SIZE * 4), %rsi
> - ja L(loop_4x_vec)
> -
> -L(last_4x_vec_or_less):
> - /* Less than 4 * VEC and aligned to VEC_SIZE. */
> - addl $(VEC_SIZE * 2), %esi
> - jle L(last_2x_vec)
> + jb L(last_4x_vec_or_less_cmpeq)
> +# endif
> + /* Save some code size by microfusing VPMINU with the load. Since
> + the matches in ymm2/ymm4 can only be returned if there where no
> + matches in ymm1/ymm3 respectively there is no issue with overlap.
> + */
> + vmovdqa 1(%rdi), %ymm1
> + VPMINU (VEC_SIZE + 1)(%rdi), %ymm1, %ymm2
> + vmovdqa (VEC_SIZE * 2 + 1)(%rdi), %ymm3
> + VPMINU (VEC_SIZE * 3 + 1)(%rdi), %ymm3, %ymm4
> +
> + VPMINU %ymm2, %ymm4, %ymm5
> + VPCMPEQ %ymm5, %ymm0, %ymm5
> + vpmovmskb %ymm5, %ecx
>
> - VPCMPEQ (%rdi), %ymm0, %ymm1
> - vpmovmskb %ymm1, %eax
> - testl %eax, %eax
> - jnz L(first_vec_x0)
> + subq $-(VEC_SIZE * 4), %rdi
> + testl %ecx, %ecx
> + jz L(loop_4x_vec)
>
> - VPCMPEQ VEC_SIZE(%rdi), %ymm0, %ymm1
> - vpmovmskb %ymm1, %eax
> - testl %eax, %eax
> - jnz L(first_vec_x1)
>
> - VPCMPEQ (VEC_SIZE * 2)(%rdi), %ymm0, %ymm1
> - vpmovmskb %ymm1, %eax
> + VPCMPEQ %ymm1, %ymm0, %ymm1
> + vpmovmskb %ymm1, %eax
> + subq %rdx, %rdi
> testl %eax, %eax
> + jnz L(last_vec_return_x0)
>
> - jnz L(first_vec_x2_check)
> - subl $VEC_SIZE, %esi
> - jle L(max)
> -
> - VPCMPEQ (VEC_SIZE * 3)(%rdi), %ymm0, %ymm1
> - vpmovmskb %ymm1, %eax
> + VPCMPEQ %ymm2, %ymm0, %ymm2
> + vpmovmskb %ymm2, %eax
> testl %eax, %eax
> -
> - jnz L(first_vec_x3_check)
> - movq %r8, %rax
> -# ifdef USE_AS_WCSLEN
> + jnz L(last_vec_return_x1)
> +
> + /* Combine last 2 VEC. */
> + VPCMPEQ %ymm3, %ymm0, %ymm3
> + vpmovmskb %ymm3, %eax
> + /* rcx has combined result from all 4 VEC. It will only be used if
> + the first 3 other VEC all did not contain a match. */
> + salq $32, %rcx
> + orq %rcx, %rax
> + tzcntq %rax, %rax
> + subq $(VEC_SIZE * 2 - 1), %rdi
> + addq %rdi, %rax
> +# ifdef USE_AS_WCSLEN
> shrq $2, %rax
> -# endif
> +# endif
> VZEROUPPER_RETURN
>
> +
> +# ifdef USE_AS_STRNLEN
> .p2align 4
> -L(last_2x_vec):
> - addl $(VEC_SIZE * 2), %esi
> - VPCMPEQ (%rdi), %ymm0, %ymm1
> - vpmovmskb %ymm1, %eax
> - testl %eax, %eax
> +L(last_4x_vec_or_less_load):
> + /* Depending on entry adjust rdi / prepare first VEC in ymm1. */
> + subq $-(VEC_SIZE * 4), %rdi
> +L(last_4x_vec_or_less_cmpeq):
> + VPCMPEQ 1(%rdi), %ymm0, %ymm1
> +L(last_4x_vec_or_less):
>
> - jnz L(first_vec_x0_check)
> - subl $VEC_SIZE, %esi
> - jle L(max)
> + vpmovmskb %ymm1, %eax
> + /* If remaining length > VEC_SIZE * 2. This works if esi is off by
> + VEC_SIZE * 4. */
> + testl $(VEC_SIZE * 2), %esi
> + jnz L(last_4x_vec)
>
> - VPCMPEQ VEC_SIZE(%rdi), %ymm0, %ymm1
> - vpmovmskb %ymm1, %eax
> + /* length may have been negative or positive by an offset of
> + VEC_SIZE * 4 depending on where this was called from. This fixes
> + that. */
> + andl $(VEC_SIZE * 4 - 1), %esi
> testl %eax, %eax
> - jnz L(first_vec_x1_check)
> - movq %r8, %rax
> -# ifdef USE_AS_WCSLEN
> - shrq $2, %rax
> -# endif
> - VZEROUPPER_RETURN
> + jnz L(last_vec_x1_check)
>
> - .p2align 4
> -L(first_vec_x0_check):
> + subl $VEC_SIZE, %esi
> + jb L(max)
> +
> + VPCMPEQ (VEC_SIZE + 1)(%rdi), %ymm0, %ymm1
> + vpmovmskb %ymm1, %eax
> tzcntl %eax, %eax
> /* Check the end of data. */
> - cmpq %rax, %rsi
> - jbe L(max)
> + cmpl %eax, %esi
> + jb L(max)
> + subq %rdx, %rdi
> + addl $(VEC_SIZE + 1), %eax
> addq %rdi, %rax
> - subq %rdx, %rax
> # ifdef USE_AS_WCSLEN
> shrq $2, %rax
> # endif
> VZEROUPPER_RETURN
> +# endif
>
> .p2align 4
> -L(first_vec_x1_check):
> +L(last_vec_return_x0):
> tzcntl %eax, %eax
> - /* Check the end of data. */
> - cmpq %rax, %rsi
> - jbe L(max)
> - addq $VEC_SIZE, %rax
> + subq $(VEC_SIZE * 4 - 1), %rdi
> addq %rdi, %rax
> - subq %rdx, %rax
> -# ifdef USE_AS_WCSLEN
> +# ifdef USE_AS_WCSLEN
> shrq $2, %rax
> -# endif
> +# endif
> VZEROUPPER_RETURN
>
> .p2align 4
> -L(first_vec_x2_check):
> +L(last_vec_return_x1):
> tzcntl %eax, %eax
> - /* Check the end of data. */
> - cmpq %rax, %rsi
> - jbe L(max)
> - addq $(VEC_SIZE * 2), %rax
> + subq $(VEC_SIZE * 3 - 1), %rdi
> addq %rdi, %rax
> - subq %rdx, %rax
> -# ifdef USE_AS_WCSLEN
> +# ifdef USE_AS_WCSLEN
> shrq $2, %rax
> -# endif
> +# endif
> VZEROUPPER_RETURN
>
> +# ifdef USE_AS_STRNLEN
> .p2align 4
> -L(first_vec_x3_check):
> +L(last_vec_x1_check):
> +
> tzcntl %eax, %eax
> /* Check the end of data. */
> - cmpq %rax, %rsi
> - jbe L(max)
> - addq $(VEC_SIZE * 3), %rax
> + cmpl %eax, %esi
> + jb L(max)
> + subq %rdx, %rdi
> + incl %eax
> addq %rdi, %rax
> - subq %rdx, %rax
> # ifdef USE_AS_WCSLEN
> shrq $2, %rax
> # endif
> VZEROUPPER_RETURN
>
> - .p2align 4
> L(max):
> movq %r8, %rax
> + VZEROUPPER_RETURN
> +
> + .p2align 4
> +L(last_4x_vec):
> + /* Test first 2x VEC normally. */
> + testl %eax, %eax
> + jnz L(last_vec_x1)
> +
> + VPCMPEQ (VEC_SIZE + 1)(%rdi), %ymm0, %ymm1
> + vpmovmskb %ymm1, %eax
> + testl %eax, %eax
> + jnz L(last_vec_x2)
> +
> + /* Normalize length. */
> + andl $(VEC_SIZE * 4 - 1), %esi
> + VPCMPEQ (VEC_SIZE * 2 + 1)(%rdi), %ymm0, %ymm1
> + vpmovmskb %ymm1, %eax
> + testl %eax, %eax
> + jnz L(last_vec_x3)
> +
> + subl $(VEC_SIZE * 3), %esi
> + jb L(max)
> +
> + VPCMPEQ (VEC_SIZE * 3 + 1)(%rdi), %ymm0, %ymm1
> + vpmovmskb %ymm1, %eax
> + tzcntl %eax, %eax
> + /* Check the end of data. */
> + cmpl %eax, %esi
> + jb L(max)
> + subq %rdx, %rdi
> + addl $(VEC_SIZE * 3 + 1), %eax
> + addq %rdi, %rax
> # ifdef USE_AS_WCSLEN
> shrq $2, %rax
> # endif
> VZEROUPPER_RETURN
>
> - .p2align 4
> -L(zero):
> - xorl %eax, %eax
> - ret
> -# endif
>
> .p2align 4
> -L(first_vec_x0):
> +L(last_vec_x1):
> + /* essentially duplicates of first_vec_x1 but use 64 bit
> + instructions. */
> tzcntl %eax, %eax
> + subq %rdx, %rdi
> + incl %eax
> addq %rdi, %rax
> - subq %rdx, %rax
> -# ifdef USE_AS_WCSLEN
> +# ifdef USE_AS_WCSLEN
> shrq $2, %rax
> -# endif
> +# endif
> VZEROUPPER_RETURN
>
> .p2align 4
> -L(first_vec_x1):
> +L(last_vec_x2):
> + /* essentially duplicates of first_vec_x1 but use 64 bit
> + instructions. */
> tzcntl %eax, %eax
> - addq $VEC_SIZE, %rax
> + subq %rdx, %rdi
> + addl $(VEC_SIZE + 1), %eax
> addq %rdi, %rax
> - subq %rdx, %rax
> -# ifdef USE_AS_WCSLEN
> +# ifdef USE_AS_WCSLEN
> shrq $2, %rax
> -# endif
> +# endif
> VZEROUPPER_RETURN
>
> .p2align 4
> -L(first_vec_x2):
> +L(last_vec_x3):
> tzcntl %eax, %eax
> - addq $(VEC_SIZE * 2), %rax
> + subl $(VEC_SIZE * 2), %esi
> + /* Check the end of data. */
> + cmpl %eax, %esi
> + jb L(max_end)
> + subq %rdx, %rdi
> + addl $(VEC_SIZE * 2 + 1), %eax
> addq %rdi, %rax
> - subq %rdx, %rax
> -# ifdef USE_AS_WCSLEN
> +# ifdef USE_AS_WCSLEN
> shrq $2, %rax
> -# endif
> +# endif
> + VZEROUPPER_RETURN
> +L(max_end):
> + movq %r8, %rax
> VZEROUPPER_RETURN
> +# endif
>
> + /* Cold case for crossing page with first load. */
> .p2align 4
> -L(4x_vec_end):
> - VPCMPEQ %ymm1, %ymm0, %ymm1
> - vpmovmskb %ymm1, %eax
> - testl %eax, %eax
> - jnz L(first_vec_x0)
> - VPCMPEQ %ymm2, %ymm0, %ymm2
> - vpmovmskb %ymm2, %eax
> +L(cross_page_boundary):
> + /* Align data to VEC_SIZE - 1. */
> + orq $(VEC_SIZE - 1), %rdi
> + VPCMPEQ -(VEC_SIZE - 1)(%rdi), %ymm0, %ymm1
> + vpmovmskb %ymm1, %eax
> + /* Remove the leading bytes. sarxl only uses bits [5:0] of COUNT
> + so no need to manually mod rdx. */
> + sarxl %edx, %eax, %eax
> +# ifdef USE_AS_STRNLEN
> testl %eax, %eax
> - jnz L(first_vec_x1)
> - VPCMPEQ %ymm3, %ymm0, %ymm3
> - vpmovmskb %ymm3, %eax
> + jnz L(cross_page_less_vec)
> + leaq 1(%rdi), %rcx
> + subq %rdx, %rcx
> + /* Check length. */
> + cmpq %rsi, %rcx
> + jb L(cross_page_continue)
> + movq %r8, %rax
> +# else
> testl %eax, %eax
> - jnz L(first_vec_x2)
> - VPCMPEQ %ymm4, %ymm0, %ymm4
> - vpmovmskb %ymm4, %eax
> -L(first_vec_x3):
> + jz L(cross_page_continue)
> tzcntl %eax, %eax
> - addq $(VEC_SIZE * 3), %rax
> - addq %rdi, %rax
> - subq %rdx, %rax
> -# ifdef USE_AS_WCSLEN
> - shrq $2, %rax
> +# ifdef USE_AS_WCSLEN
> + shrl $2, %eax
> +# endif
> # endif
> +L(return_vzeroupper):
> + ZERO_UPPER_VEC_REGISTERS_RETURN
> +
> +# ifdef USE_AS_STRNLEN
> + .p2align 4
> +L(cross_page_less_vec):
> + tzcntl %eax, %eax
> + cmpq %rax, %rsi
> + cmovb %esi, %eax
> +# ifdef USE_AS_WCSLEN
> + shrl $2, %eax
> +# endif
> VZEROUPPER_RETURN
> +# endif
>
> END (STRLEN)
> #endif
> --
> 2.29.2
>
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: [PATCH v4 1/2] x86: Optimize strlen-evex.S
2021-04-19 18:43 [PATCH v4 1/2] x86: Optimize strlen-evex.S Noah Goldstein
2021-04-19 18:43 ` [PATCH v4 2/2] x86: Optimize strlen-avx2.S Noah Goldstein
@ 2021-04-19 22:16 ` H.J. Lu
2021-04-19 23:38 ` Noah Goldstein
1 sibling, 1 reply; 6+ messages in thread
From: H.J. Lu @ 2021-04-19 22:16 UTC (permalink / raw)
To: Noah Goldstein; +Cc: libc-alpha, carlos, hjl.tools
On Mon, Apr 19, 2021 at 02:43:44PM -0400, Noah Goldstein wrote:
> No bug. This commit optimizes strlen-evex.S. The
> optimizations are mostly small things but they add up to roughly
> 10-30% performance improvement for strlen. The results for strnlen are
> bit more ambiguous. test-strlen, test-strnlen, test-wcslen, and
> test-wcsnlen are all passing.
>
> Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com>
> ---
> sysdeps/x86_64/multiarch/strlen-evex.S | 584 ++++++++++++++-----------
> 1 file changed, 320 insertions(+), 264 deletions(-)
>
> diff --git a/sysdeps/x86_64/multiarch/strlen-evex.S b/sysdeps/x86_64/multiarch/strlen-evex.S
> index 0583819078..36b28198c8 100644
> --- a/sysdeps/x86_64/multiarch/strlen-evex.S
> +++ b/sysdeps/x86_64/multiarch/strlen-evex.S
> @@ -29,11 +29,13 @@
> # ifdef USE_AS_WCSLEN
> # define VPCMP vpcmpd
> # define VPMINU vpminud
> -# define SHIFT_REG r9d
> +# define SHIFT_REG ecx
> +# define CHAR_SIZE 4
> # else
> # define VPCMP vpcmpb
> # define VPMINU vpminub
> -# define SHIFT_REG ecx
> +# define SHIFT_REG edx
> +# define CHAR_SIZE 1
> # endif
>
> # define XMMZERO xmm16
> @@ -46,132 +48,168 @@
> # define YMM6 ymm22
>
> # define VEC_SIZE 32
> +# define PAGE_SIZE 4096
> +# define LOG_PAGE_SIZE 12
> +# define CHAR_PER_VEC (VEC_SIZE / CHAR_SIZE)
>
> .section .text.evex,"ax",@progbits
> ENTRY (STRLEN)
> # ifdef USE_AS_STRNLEN
> - /* Check for zero length. */
> + /* Check zero length. */
> test %RSI_LP, %RSI_LP
> jz L(zero)
> -# ifdef USE_AS_WCSLEN
> - shl $2, %RSI_LP
> -# elif defined __ILP32__
> +# ifdef __ILP32__
> /* Clear the upper 32 bits. */
> movl %esi, %esi
> # endif
> mov %RSI_LP, %R8_LP
> # endif
> - movl %edi, %ecx
> - movq %rdi, %rdx
> + movl %edi, %eax
> vpxorq %XMMZERO, %XMMZERO, %XMMZERO
> -
> + /* Shift left eax to clear all bits not relevant to page cross
> + check. This saves 2 bytes of code as opposed to using andl with
> + PAGE_SIZE - 1. Then compare with PAGE_SIZE - VEC_SIZE shifted
> + left by the same amount (an imm32 either way). */
> + sall $(32 - LOG_PAGE_SIZE), %eax
I prefer AND over SAL for better throughput.
H.J.
---
> /* Check if we may cross page boundary with one vector load. */
> - andl $(2 * VEC_SIZE - 1), %ecx
> - cmpl $VEC_SIZE, %ecx
> - ja L(cros_page_boundary)
> + cmpl $((PAGE_SIZE - VEC_SIZE) << (32 - LOG_PAGE_SIZE)), %eax
> + ja L(cross_page_boundary)
>
> /* Check the first VEC_SIZE bytes. Each bit in K0 represents a
> null byte. */
> VPCMP $0, (%rdi), %YMMZERO, %k0
> kmovd %k0, %eax
> - testl %eax, %eax
> -
> # ifdef USE_AS_STRNLEN
> - jnz L(first_vec_x0_check)
> - /* Adjust length and check the end of data. */
> - subq $VEC_SIZE, %rsi
> - jbe L(max)
> -# else
> - jnz L(first_vec_x0)
> + /* If length < CHAR_PER_VEC handle special. */
> + cmpq $CHAR_PER_VEC, %rsi
> + jbe L(first_vec_x0)
> # endif
> -
> - /* Align data for aligned loads in the loop. */
> - addq $VEC_SIZE, %rdi
> - andl $(VEC_SIZE - 1), %ecx
> - andq $-VEC_SIZE, %rdi
> -
> + testl %eax, %eax
> + jz L(aligned_more)
> + tzcntl %eax, %eax
> + ret
> # ifdef USE_AS_STRNLEN
> - /* Adjust length. */
> - addq %rcx, %rsi
> +L(zero):
> + xorl %eax, %eax
> + ret
>
> - subq $(VEC_SIZE * 4), %rsi
> - jbe L(last_4x_vec_or_less)
> + .p2align 4
> +L(first_vec_x0):
> + /* Set bit for max len so that tzcnt will return min of max len
> + and position of first match. */
> + btsq %rsi, %rax
> + tzcntl %eax, %eax
> + ret
> # endif
> - jmp L(more_4x_vec)
>
> .p2align 4
> -L(cros_page_boundary):
> - andl $(VEC_SIZE - 1), %ecx
> - andq $-VEC_SIZE, %rdi
> -
> -# ifdef USE_AS_WCSLEN
> - /* NB: Divide shift count by 4 since each bit in K0 represent 4
> - bytes. */
> - movl %ecx, %SHIFT_REG
> - sarl $2, %SHIFT_REG
> +L(first_vec_x1):
> + tzcntl %eax, %eax
> + /* Safe to use 32 bit instructions as these are only called for
> + size = [1, 159]. */
> +# ifdef USE_AS_STRNLEN
> + /* Use ecx which was computed earlier to compute correct value.
> + */
> + leal -(CHAR_PER_VEC * 4 + 1)(%rcx, %rax), %eax
> +# else
> + subl %edx, %edi
> +# ifdef USE_AS_WCSLEN
> + /* NB: Divide bytes by 4 to get the wchar_t count. */
> + sarl $2, %edi
> +# endif
> + leal CHAR_PER_VEC(%rdi, %rax), %eax
> # endif
> - VPCMP $0, (%rdi), %YMMZERO, %k0
> - kmovd %k0, %eax
> + ret
>
> - /* Remove the leading bytes. */
> - sarxl %SHIFT_REG, %eax, %eax
> - testl %eax, %eax
> - jz L(aligned_more)
> + .p2align 4
> +L(first_vec_x2):
> tzcntl %eax, %eax
> -# ifdef USE_AS_WCSLEN
> - /* NB: Multiply wchar_t count by 4 to get the number of bytes. */
> - sall $2, %eax
> -# endif
> + /* Safe to use 32 bit instructions as these are only called for
> + size = [1, 159]. */
> # ifdef USE_AS_STRNLEN
> - /* Check the end of data. */
> - cmpq %rax, %rsi
> - jbe L(max)
> -# endif
> - addq %rdi, %rax
> - addq %rcx, %rax
> - subq %rdx, %rax
> -# ifdef USE_AS_WCSLEN
> - shrq $2, %rax
> + /* Use ecx which was computed earlier to compute correct value.
> + */
> + leal -(CHAR_PER_VEC * 3 + 1)(%rcx, %rax), %eax
> +# else
> + subl %edx, %edi
> +# ifdef USE_AS_WCSLEN
> + /* NB: Divide bytes by 4 to get the wchar_t count. */
> + sarl $2, %edi
> +# endif
> + leal (CHAR_PER_VEC * 2)(%rdi, %rax), %eax
> # endif
> ret
>
> .p2align 4
> -L(aligned_more):
> +L(first_vec_x3):
> + tzcntl %eax, %eax
> + /* Safe to use 32 bit instructions as these are only called for
> + size = [1, 159]. */
> # ifdef USE_AS_STRNLEN
> - /* "rcx" is less than VEC_SIZE. Calculate "rdx + rcx - VEC_SIZE"
> - with "rdx - (VEC_SIZE - rcx)" instead of "(rdx + rcx) - VEC_SIZE"
> - to void possible addition overflow. */
> - negq %rcx
> - addq $VEC_SIZE, %rcx
> -
> - /* Check the end of data. */
> - subq %rcx, %rsi
> - jbe L(max)
> + /* Use ecx which was computed earlier to compute correct value.
> + */
> + leal -(CHAR_PER_VEC * 2 + 1)(%rcx, %rax), %eax
> +# else
> + subl %edx, %edi
> +# ifdef USE_AS_WCSLEN
> + /* NB: Divide bytes by 4 to get the wchar_t count. */
> + sarl $2, %edi
> +# endif
> + leal (CHAR_PER_VEC * 3)(%rdi, %rax), %eax
> # endif
> + ret
>
> - addq $VEC_SIZE, %rdi
> -
> + .p2align 4
> +L(first_vec_x4):
> + tzcntl %eax, %eax
> + /* Safe to use 32 bit instructions as these are only called for
> + size = [1, 159]. */
> # ifdef USE_AS_STRNLEN
> - subq $(VEC_SIZE * 4), %rsi
> - jbe L(last_4x_vec_or_less)
> + /* Use ecx which was computed earlier to compute correct value.
> + */
> + leal -(CHAR_PER_VEC + 1)(%rcx, %rax), %eax
> +# else
> + subl %edx, %edi
> +# ifdef USE_AS_WCSLEN
> + /* NB: Divide bytes by 4 to get the wchar_t count. */
> + sarl $2, %edi
> +# endif
> + leal (CHAR_PER_VEC * 4)(%rdi, %rax), %eax
> # endif
> + ret
>
> -L(more_4x_vec):
> + .p2align 5
> +L(aligned_more):
> + movq %rdi, %rdx
> + /* Align data to VEC_SIZE. */
> + andq $-(VEC_SIZE), %rdi
> +L(cross_page_continue):
> /* Check the first 4 * VEC_SIZE. Only one VEC_SIZE at a time
> since data is only aligned to VEC_SIZE. */
> - VPCMP $0, (%rdi), %YMMZERO, %k0
> - kmovd %k0, %eax
> - testl %eax, %eax
> - jnz L(first_vec_x0)
> -
> +# ifdef USE_AS_STRNLEN
> + /* + CHAR_SIZE because it simplies the logic in
> + last_4x_vec_or_less. */
> + leaq (VEC_SIZE * 5 + CHAR_SIZE)(%rdi), %rcx
> + subq %rdx, %rcx
> +# ifdef USE_AS_WCSLEN
> + /* NB: Divide bytes by 4 to get the wchar_t count. */
> + sarl $2, %ecx
> +# endif
> +# endif
> + /* Load first VEC regardless. */
> VPCMP $0, VEC_SIZE(%rdi), %YMMZERO, %k0
> +# ifdef USE_AS_STRNLEN
> + /* Adjust length. If near end handle specially. */
> + subq %rcx, %rsi
> + jb L(last_4x_vec_or_less)
> +# endif
> kmovd %k0, %eax
> testl %eax, %eax
> jnz L(first_vec_x1)
>
> VPCMP $0, (VEC_SIZE * 2)(%rdi), %YMMZERO, %k0
> kmovd %k0, %eax
> - testl %eax, %eax
> + test %eax, %eax
> jnz L(first_vec_x2)
>
> VPCMP $0, (VEC_SIZE * 3)(%rdi), %YMMZERO, %k0
> @@ -179,258 +217,276 @@ L(more_4x_vec):
> testl %eax, %eax
> jnz L(first_vec_x3)
>
> - addq $(VEC_SIZE * 4), %rdi
> -
> -# ifdef USE_AS_STRNLEN
> - subq $(VEC_SIZE * 4), %rsi
> - jbe L(last_4x_vec_or_less)
> -# endif
> -
> - /* Align data to 4 * VEC_SIZE. */
> - movq %rdi, %rcx
> - andl $(4 * VEC_SIZE - 1), %ecx
> - andq $-(4 * VEC_SIZE), %rdi
> + VPCMP $0, (VEC_SIZE * 4)(%rdi), %YMMZERO, %k0
> + kmovd %k0, %eax
> + testl %eax, %eax
> + jnz L(first_vec_x4)
>
> + addq $VEC_SIZE, %rdi
> # ifdef USE_AS_STRNLEN
> - /* Adjust length. */
> + /* Check if at last VEC_SIZE * 4 length. */
> + cmpq $(CHAR_PER_VEC * 4 - 1), %rsi
> + jbe L(last_4x_vec_or_less_load)
> + movl %edi, %ecx
> + andl $(VEC_SIZE * 4 - 1), %ecx
> +# ifdef USE_AS_WCSLEN
> + /* NB: Divide bytes by 4 to get the wchar_t count. */
> + sarl $2, %ecx
> +# endif
> + /* Readjust length. */
> addq %rcx, %rsi
> # endif
> + /* Align data to VEC_SIZE * 4. */
> + andq $-(VEC_SIZE * 4), %rdi
>
> + /* Compare 4 * VEC at a time forward. */
> .p2align 4
> L(loop_4x_vec):
> - /* Compare 4 * VEC at a time forward. */
> - VMOVA (%rdi), %YMM1
> - VMOVA VEC_SIZE(%rdi), %YMM2
> - VMOVA (VEC_SIZE * 2)(%rdi), %YMM3
> - VMOVA (VEC_SIZE * 3)(%rdi), %YMM4
> -
> - VPMINU %YMM1, %YMM2, %YMM5
> - VPMINU %YMM3, %YMM4, %YMM6
> + /* Load first VEC regardless. */
> + VMOVA (VEC_SIZE * 4)(%rdi), %YMM1
> +# ifdef USE_AS_STRNLEN
> + /* Break if at end of length. */
> + subq $(CHAR_PER_VEC * 4), %rsi
> + jb L(last_4x_vec_or_less_cmpeq)
> +# endif
> + /* Save some code size by microfusing VPMINU with the load. Since
> + the matches in ymm2/ymm4 can only be returned if there where no
> + matches in ymm1/ymm3 respectively there is no issue with overlap.
> + */
> + VPMINU (VEC_SIZE * 5)(%rdi), %YMM1, %YMM2
> + VMOVA (VEC_SIZE * 6)(%rdi), %YMM3
> + VPMINU (VEC_SIZE * 7)(%rdi), %YMM3, %YMM4
> +
> + VPCMP $0, %YMM2, %YMMZERO, %k0
> + VPCMP $0, %YMM4, %YMMZERO, %k1
> + subq $-(VEC_SIZE * 4), %rdi
> + kortestd %k0, %k1
> + jz L(loop_4x_vec)
> +
> + /* Check if end was in first half. */
> + kmovd %k0, %eax
> + subq %rdx, %rdi
> +# ifdef USE_AS_WCSLEN
> + shrq $2, %rdi
> +# endif
> + testl %eax, %eax
> + jz L(second_vec_return)
>
> - VPMINU %YMM5, %YMM6, %YMM5
> - VPCMP $0, %YMM5, %YMMZERO, %k0
> - ktestd %k0, %k0
> - jnz L(4x_vec_end)
> + VPCMP $0, %YMM1, %YMMZERO, %k2
> + kmovd %k2, %edx
> + /* Combine VEC1 matches (edx) with VEC2 matches (eax). */
> +# ifdef USE_AS_WCSLEN
> + sall $CHAR_PER_VEC, %eax
> + orl %edx, %eax
> + tzcntl %eax, %eax
> +# else
> + salq $CHAR_PER_VEC, %rax
> + orq %rdx, %rax
> + tzcntq %rax, %rax
> +# endif
> + addq %rdi, %rax
> + ret
>
> - addq $(VEC_SIZE * 4), %rdi
>
> -# ifndef USE_AS_STRNLEN
> - jmp L(loop_4x_vec)
> -# else
> - subq $(VEC_SIZE * 4), %rsi
> - ja L(loop_4x_vec)
> +# ifdef USE_AS_STRNLEN
>
> +L(last_4x_vec_or_less_load):
> + /* Depending on entry adjust rdi / prepare first VEC in YMM1. */
> + VMOVA (VEC_SIZE * 4)(%rdi), %YMM1
> +L(last_4x_vec_or_less_cmpeq):
> + VPCMP $0, %YMM1, %YMMZERO, %k0
> + addq $(VEC_SIZE * 3), %rdi
> L(last_4x_vec_or_less):
> - /* Less than 4 * VEC and aligned to VEC_SIZE. */
> - addl $(VEC_SIZE * 2), %esi
> - jle L(last_2x_vec)
> -
> - VPCMP $0, (%rdi), %YMMZERO, %k0
> kmovd %k0, %eax
> + /* If remaining length > VEC_SIZE * 2. This works if esi is off by
> + VEC_SIZE * 4. */
> + testl $(CHAR_PER_VEC * 2), %esi
> + jnz L(last_4x_vec)
> +
> + /* length may have been negative or positive by an offset of
> + CHAR_PER_VEC * 4 depending on where this was called from. This
> + fixes that. */
> + andl $(CHAR_PER_VEC * 4 - 1), %esi
> testl %eax, %eax
> - jnz L(first_vec_x0)
> + jnz L(last_vec_x1_check)
>
> - VPCMP $0, VEC_SIZE(%rdi), %YMMZERO, %k0
> - kmovd %k0, %eax
> - testl %eax, %eax
> - jnz L(first_vec_x1)
> + /* Check the end of data. */
> + subl $CHAR_PER_VEC, %esi
> + jb L(max)
>
> VPCMP $0, (VEC_SIZE * 2)(%rdi), %YMMZERO, %k0
> kmovd %k0, %eax
> - testl %eax, %eax
> - jnz L(first_vec_x2_check)
> - subl $VEC_SIZE, %esi
> - jle L(max)
> + tzcntl %eax, %eax
> + /* Check the end of data. */
> + cmpl %eax, %esi
> + jb L(max)
>
> - VPCMP $0, (VEC_SIZE * 3)(%rdi), %YMMZERO, %k0
> - kmovd %k0, %eax
> - testl %eax, %eax
> - jnz L(first_vec_x3_check)
> + subq %rdx, %rdi
> +# ifdef USE_AS_WCSLEN
> + /* NB: Divide bytes by 4 to get the wchar_t count. */
> + sarq $2, %rdi
> +# endif
> + leaq (CHAR_PER_VEC * 2)(%rdi, %rax), %rax
> + ret
> +L(max):
> movq %r8, %rax
> + ret
> +# endif
> +
> + /* Placed here in strnlen so that the jcc L(last_4x_vec_or_less)
> + in the 4x VEC loop can use 2 byte encoding. */
> + .p2align 4
> +L(second_vec_return):
> + VPCMP $0, %YMM3, %YMMZERO, %k0
> + /* Combine YMM3 matches (k0) with YMM4 matches (k1). */
> +# ifdef USE_AS_WCSLEN
> + kunpckbw %k0, %k1, %k0
> + kmovd %k0, %eax
> + tzcntl %eax, %eax
> +# else
> + kunpckdq %k0, %k1, %k0
> + kmovq %k0, %rax
> + tzcntq %rax, %rax
> +# endif
> + leaq (CHAR_PER_VEC * 2)(%rdi, %rax), %rax
> + ret
> +
> +
> +# ifdef USE_AS_STRNLEN
> +L(last_vec_x1_check):
> + tzcntl %eax, %eax
> + /* Check the end of data. */
> + cmpl %eax, %esi
> + jb L(max)
> + subq %rdx, %rdi
> # ifdef USE_AS_WCSLEN
> - shrq $2, %rax
> + /* NB: Divide bytes by 4 to get the wchar_t count. */
> + sarq $2, %rdi
> # endif
> + leaq (CHAR_PER_VEC)(%rdi, %rax), %rax
> ret
>
> .p2align 4
> -L(last_2x_vec):
> - addl $(VEC_SIZE * 2), %esi
> +L(last_4x_vec):
> + /* Test first 2x VEC normally. */
> + testl %eax, %eax
> + jnz L(last_vec_x1)
>
> - VPCMP $0, (%rdi), %YMMZERO, %k0
> + VPCMP $0, (VEC_SIZE * 2)(%rdi), %YMMZERO, %k0
> kmovd %k0, %eax
> testl %eax, %eax
> - jnz L(first_vec_x0_check)
> - subl $VEC_SIZE, %esi
> - jle L(max)
> + jnz L(last_vec_x2)
>
> - VPCMP $0, VEC_SIZE(%rdi), %YMMZERO, %k0
> + /* Normalize length. */
> + andl $(CHAR_PER_VEC * 4 - 1), %esi
> + VPCMP $0, (VEC_SIZE * 3)(%rdi), %YMMZERO, %k0
> kmovd %k0, %eax
> testl %eax, %eax
> - jnz L(first_vec_x1_check)
> - movq %r8, %rax
> -# ifdef USE_AS_WCSLEN
> - shrq $2, %rax
> -# endif
> - ret
> + jnz L(last_vec_x3)
>
> - .p2align 4
> -L(first_vec_x0_check):
> + /* Check the end of data. */
> + subl $(CHAR_PER_VEC * 3), %esi
> + jb L(max)
> +
> + VPCMP $0, (VEC_SIZE * 4)(%rdi), %YMMZERO, %k0
> + kmovd %k0, %eax
> tzcntl %eax, %eax
> -# ifdef USE_AS_WCSLEN
> - /* NB: Multiply wchar_t count by 4 to get the number of bytes. */
> - sall $2, %eax
> -# endif
> /* Check the end of data. */
> - cmpq %rax, %rsi
> - jbe L(max)
> - addq %rdi, %rax
> - subq %rdx, %rax
> + cmpl %eax, %esi
> + jb L(max_end)
> +
> + subq %rdx, %rdi
> # ifdef USE_AS_WCSLEN
> - shrq $2, %rax
> + /* NB: Divide bytes by 4 to get the wchar_t count. */
> + sarq $2, %rdi
> # endif
> + leaq (CHAR_PER_VEC * 4)(%rdi, %rax), %rax
> ret
>
> .p2align 4
> -L(first_vec_x1_check):
> +L(last_vec_x1):
> tzcntl %eax, %eax
> + subq %rdx, %rdi
> # ifdef USE_AS_WCSLEN
> - /* NB: Multiply wchar_t count by 4 to get the number of bytes. */
> - sall $2, %eax
> -# endif
> - /* Check the end of data. */
> - cmpq %rax, %rsi
> - jbe L(max)
> - addq $VEC_SIZE, %rax
> - addq %rdi, %rax
> - subq %rdx, %rax
> -# ifdef USE_AS_WCSLEN
> - shrq $2, %rax
> + /* NB: Divide bytes by 4 to get the wchar_t count. */
> + sarq $2, %rdi
> # endif
> + leaq (CHAR_PER_VEC)(%rdi, %rax), %rax
> ret
>
> .p2align 4
> -L(first_vec_x2_check):
> +L(last_vec_x2):
> tzcntl %eax, %eax
> + subq %rdx, %rdi
> # ifdef USE_AS_WCSLEN
> - /* NB: Multiply wchar_t count by 4 to get the number of bytes. */
> - sall $2, %eax
> -# endif
> - /* Check the end of data. */
> - cmpq %rax, %rsi
> - jbe L(max)
> - addq $(VEC_SIZE * 2), %rax
> - addq %rdi, %rax
> - subq %rdx, %rax
> -# ifdef USE_AS_WCSLEN
> - shrq $2, %rax
> + /* NB: Divide bytes by 4 to get the wchar_t count. */
> + sarq $2, %rdi
> # endif
> + leaq (CHAR_PER_VEC * 2)(%rdi, %rax), %rax
> ret
>
> .p2align 4
> -L(first_vec_x3_check):
> +L(last_vec_x3):
> tzcntl %eax, %eax
> -# ifdef USE_AS_WCSLEN
> - /* NB: Multiply wchar_t count by 4 to get the number of bytes. */
> - sall $2, %eax
> -# endif
> + subl $(CHAR_PER_VEC * 2), %esi
> /* Check the end of data. */
> - cmpq %rax, %rsi
> - jbe L(max)
> - addq $(VEC_SIZE * 3), %rax
> - addq %rdi, %rax
> - subq %rdx, %rax
> + cmpl %eax, %esi
> + jb L(max_end)
> + subq %rdx, %rdi
> # ifdef USE_AS_WCSLEN
> - shrq $2, %rax
> + /* NB: Divide bytes by 4 to get the wchar_t count. */
> + sarq $2, %rdi
> # endif
> + leaq (CHAR_PER_VEC * 3)(%rdi, %rax), %rax
> ret
> -
> - .p2align 4
> -L(max):
> +L(max_end):
> movq %r8, %rax
> -# ifdef USE_AS_WCSLEN
> - shrq $2, %rax
> -# endif
> - ret
> -
> - .p2align 4
> -L(zero):
> - xorl %eax, %eax
> ret
> # endif
>
> + /* Cold case for crossing page with first load. */
> .p2align 4
> -L(first_vec_x0):
> - tzcntl %eax, %eax
> -# ifdef USE_AS_WCSLEN
> - /* NB: Multiply wchar_t count by 4 to get the number of bytes. */
> - sall $2, %eax
> -# endif
> - addq %rdi, %rax
> - subq %rdx, %rax
> +L(cross_page_boundary):
> + movq %rdi, %rdx
> + /* Align data to VEC_SIZE. */
> + andq $-VEC_SIZE, %rdi
> + VPCMP $0, (%rdi), %YMMZERO, %k0
> + kmovd %k0, %eax
> + /* Remove the leading bytes. */
> # ifdef USE_AS_WCSLEN
> - shrq $2, %rax
> + /* NB: Divide shift count by 4 since each bit in K0 represent 4
> + bytes. */
> + movl %edx, %ecx
> + shrl $2, %ecx
> + andl $(CHAR_PER_VEC - 1), %ecx
> # endif
> - ret
> -
> - .p2align 4
> -L(first_vec_x1):
> + /* SHIFT_REG is ecx for USE_AS_WCSLEN and edx otherwise. */
> + sarxl %SHIFT_REG, %eax, %eax
> + testl %eax, %eax
> +# ifndef USE_AS_STRNLEN
> + jz L(cross_page_continue)
> tzcntl %eax, %eax
> -# ifdef USE_AS_WCSLEN
> - /* NB: Multiply wchar_t count by 4 to get the number of bytes. */
> - sall $2, %eax
> -# endif
> - addq $VEC_SIZE, %rax
> - addq %rdi, %rax
> - subq %rdx, %rax
> -# ifdef USE_AS_WCSLEN
> - shrq $2, %rax
> -# endif
> ret
> -
> - .p2align 4
> -L(first_vec_x2):
> - tzcntl %eax, %eax
> -# ifdef USE_AS_WCSLEN
> - /* NB: Multiply wchar_t count by 4 to get the number of bytes. */
> - sall $2, %eax
> -# endif
> - addq $(VEC_SIZE * 2), %rax
> - addq %rdi, %rax
> - subq %rdx, %rax
> -# ifdef USE_AS_WCSLEN
> - shrq $2, %rax
> -# endif
> +# else
> + jnz L(cross_page_less_vec)
> +# ifndef USE_AS_WCSLEN
> + movl %edx, %ecx
> + andl $(CHAR_PER_VEC - 1), %ecx
> +# endif
> + movl $CHAR_PER_VEC, %eax
> + subl %ecx, %eax
> + /* Check the end of data. */
> + cmpq %rax, %rsi
> + ja L(cross_page_continue)
> + movl %esi, %eax
> ret
> -
> - .p2align 4
> -L(4x_vec_end):
> - VPCMP $0, %YMM1, %YMMZERO, %k0
> - kmovd %k0, %eax
> - testl %eax, %eax
> - jnz L(first_vec_x0)
> - VPCMP $0, %YMM2, %YMMZERO, %k1
> - kmovd %k1, %eax
> - testl %eax, %eax
> - jnz L(first_vec_x1)
> - VPCMP $0, %YMM3, %YMMZERO, %k2
> - kmovd %k2, %eax
> - testl %eax, %eax
> - jnz L(first_vec_x2)
> - VPCMP $0, %YMM4, %YMMZERO, %k3
> - kmovd %k3, %eax
> -L(first_vec_x3):
> +L(cross_page_less_vec):
> tzcntl %eax, %eax
> -# ifdef USE_AS_WCSLEN
> - /* NB: Multiply wchar_t count by 4 to get the number of bytes. */
> - sall $2, %eax
> -# endif
> - addq $(VEC_SIZE * 3), %rax
> - addq %rdi, %rax
> - subq %rdx, %rax
> -# ifdef USE_AS_WCSLEN
> - shrq $2, %rax
> -# endif
> + /* Select min of length and position of first null. */
> + cmpq %rax, %rsi
> + cmovb %esi, %eax
> ret
> +# endif
>
> END (STRLEN)
> #endif
> --
> 2.29.2
>
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: [PATCH v4 1/2] x86: Optimize strlen-evex.S
2021-04-19 22:16 ` [PATCH v4 1/2] x86: Optimize strlen-evex.S H.J. Lu
@ 2021-04-19 23:38 ` Noah Goldstein
0 siblings, 0 replies; 6+ messages in thread
From: Noah Goldstein @ 2021-04-19 23:38 UTC (permalink / raw)
To: H.J. Lu; +Cc: GNU C Library, Carlos O'Donell
On Mon, Apr 19, 2021 at 6:16 PM H.J. Lu <hjl.tools@gmail.com> wrote:
>
> On Mon, Apr 19, 2021 at 02:43:44PM -0400, Noah Goldstein wrote:
> > No bug. This commit optimizes strlen-evex.S. The
> > optimizations are mostly small things but they add up to roughly
> > 10-30% performance improvement for strlen. The results for strnlen are
> > bit more ambiguous. test-strlen, test-strnlen, test-wcslen, and
> > test-wcsnlen are all passing.
> >
> > Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com>
> > ---
> > sysdeps/x86_64/multiarch/strlen-evex.S | 584 ++++++++++++++-----------
> > 1 file changed, 320 insertions(+), 264 deletions(-)
> >
> > diff --git a/sysdeps/x86_64/multiarch/strlen-evex.S b/sysdeps/x86_64/multiarch/strlen-evex.S
> > index 0583819078..36b28198c8 100644
> > --- a/sysdeps/x86_64/multiarch/strlen-evex.S
> > +++ b/sysdeps/x86_64/multiarch/strlen-evex.S
> > @@ -29,11 +29,13 @@
> > # ifdef USE_AS_WCSLEN
> > # define VPCMP vpcmpd
> > # define VPMINU vpminud
> > -# define SHIFT_REG r9d
> > +# define SHIFT_REG ecx
> > +# define CHAR_SIZE 4
> > # else
> > # define VPCMP vpcmpb
> > # define VPMINU vpminub
> > -# define SHIFT_REG ecx
> > +# define SHIFT_REG edx
> > +# define CHAR_SIZE 1
> > # endif
> >
> > # define XMMZERO xmm16
> > @@ -46,132 +48,168 @@
> > # define YMM6 ymm22
> >
> > # define VEC_SIZE 32
> > +# define PAGE_SIZE 4096
> > +# define LOG_PAGE_SIZE 12
> > +# define CHAR_PER_VEC (VEC_SIZE / CHAR_SIZE)
> >
> > .section .text.evex,"ax",@progbits
> > ENTRY (STRLEN)
> > # ifdef USE_AS_STRNLEN
> > - /* Check for zero length. */
> > + /* Check zero length. */
> > test %RSI_LP, %RSI_LP
> > jz L(zero)
> > -# ifdef USE_AS_WCSLEN
> > - shl $2, %RSI_LP
> > -# elif defined __ILP32__
> > +# ifdef __ILP32__
> > /* Clear the upper 32 bits. */
> > movl %esi, %esi
> > # endif
> > mov %RSI_LP, %R8_LP
> > # endif
> > - movl %edi, %ecx
> > - movq %rdi, %rdx
> > + movl %edi, %eax
> > vpxorq %XMMZERO, %XMMZERO, %XMMZERO
> > -
> > + /* Shift left eax to clear all bits not relevant to page cross
> > + check. This saves 2 bytes of code as opposed to using andl with
> > + PAGE_SIZE - 1. Then compare with PAGE_SIZE - VEC_SIZE shifted
> > + left by the same amount (an imm32 either way). */
> > + sall $(32 - LOG_PAGE_SIZE), %eax
>
> I prefer AND over SAL for better throughput.
Done.
>
>
> H.J.
> ---
> > /* Check if we may cross page boundary with one vector load. */
> > - andl $(2 * VEC_SIZE - 1), %ecx
> > - cmpl $VEC_SIZE, %ecx
> > - ja L(cros_page_boundary)
> > + cmpl $((PAGE_SIZE - VEC_SIZE) << (32 - LOG_PAGE_SIZE)), %eax
> > + ja L(cross_page_boundary)
> >
> > /* Check the first VEC_SIZE bytes. Each bit in K0 represents a
> > null byte. */
> > VPCMP $0, (%rdi), %YMMZERO, %k0
> > kmovd %k0, %eax
> > - testl %eax, %eax
> > -
> > # ifdef USE_AS_STRNLEN
> > - jnz L(first_vec_x0_check)
> > - /* Adjust length and check the end of data. */
> > - subq $VEC_SIZE, %rsi
> > - jbe L(max)
> > -# else
> > - jnz L(first_vec_x0)
> > + /* If length < CHAR_PER_VEC handle special. */
> > + cmpq $CHAR_PER_VEC, %rsi
> > + jbe L(first_vec_x0)
> > # endif
> > -
> > - /* Align data for aligned loads in the loop. */
> > - addq $VEC_SIZE, %rdi
> > - andl $(VEC_SIZE - 1), %ecx
> > - andq $-VEC_SIZE, %rdi
> > -
> > + testl %eax, %eax
> > + jz L(aligned_more)
> > + tzcntl %eax, %eax
> > + ret
> > # ifdef USE_AS_STRNLEN
> > - /* Adjust length. */
> > - addq %rcx, %rsi
> > +L(zero):
> > + xorl %eax, %eax
> > + ret
> >
> > - subq $(VEC_SIZE * 4), %rsi
> > - jbe L(last_4x_vec_or_less)
> > + .p2align 4
> > +L(first_vec_x0):
> > + /* Set bit for max len so that tzcnt will return min of max len
> > + and position of first match. */
> > + btsq %rsi, %rax
> > + tzcntl %eax, %eax
> > + ret
> > # endif
> > - jmp L(more_4x_vec)
> >
> > .p2align 4
> > -L(cros_page_boundary):
> > - andl $(VEC_SIZE - 1), %ecx
> > - andq $-VEC_SIZE, %rdi
> > -
> > -# ifdef USE_AS_WCSLEN
> > - /* NB: Divide shift count by 4 since each bit in K0 represent 4
> > - bytes. */
> > - movl %ecx, %SHIFT_REG
> > - sarl $2, %SHIFT_REG
> > +L(first_vec_x1):
> > + tzcntl %eax, %eax
> > + /* Safe to use 32 bit instructions as these are only called for
> > + size = [1, 159]. */
> > +# ifdef USE_AS_STRNLEN
> > + /* Use ecx which was computed earlier to compute correct value.
> > + */
> > + leal -(CHAR_PER_VEC * 4 + 1)(%rcx, %rax), %eax
> > +# else
> > + subl %edx, %edi
> > +# ifdef USE_AS_WCSLEN
> > + /* NB: Divide bytes by 4 to get the wchar_t count. */
> > + sarl $2, %edi
> > +# endif
> > + leal CHAR_PER_VEC(%rdi, %rax), %eax
> > # endif
> > - VPCMP $0, (%rdi), %YMMZERO, %k0
> > - kmovd %k0, %eax
> > + ret
> >
> > - /* Remove the leading bytes. */
> > - sarxl %SHIFT_REG, %eax, %eax
> > - testl %eax, %eax
> > - jz L(aligned_more)
> > + .p2align 4
> > +L(first_vec_x2):
> > tzcntl %eax, %eax
> > -# ifdef USE_AS_WCSLEN
> > - /* NB: Multiply wchar_t count by 4 to get the number of bytes. */
> > - sall $2, %eax
> > -# endif
> > + /* Safe to use 32 bit instructions as these are only called for
> > + size = [1, 159]. */
> > # ifdef USE_AS_STRNLEN
> > - /* Check the end of data. */
> > - cmpq %rax, %rsi
> > - jbe L(max)
> > -# endif
> > - addq %rdi, %rax
> > - addq %rcx, %rax
> > - subq %rdx, %rax
> > -# ifdef USE_AS_WCSLEN
> > - shrq $2, %rax
> > + /* Use ecx which was computed earlier to compute correct value.
> > + */
> > + leal -(CHAR_PER_VEC * 3 + 1)(%rcx, %rax), %eax
> > +# else
> > + subl %edx, %edi
> > +# ifdef USE_AS_WCSLEN
> > + /* NB: Divide bytes by 4 to get the wchar_t count. */
> > + sarl $2, %edi
> > +# endif
> > + leal (CHAR_PER_VEC * 2)(%rdi, %rax), %eax
> > # endif
> > ret
> >
> > .p2align 4
> > -L(aligned_more):
> > +L(first_vec_x3):
> > + tzcntl %eax, %eax
> > + /* Safe to use 32 bit instructions as these are only called for
> > + size = [1, 159]. */
> > # ifdef USE_AS_STRNLEN
> > - /* "rcx" is less than VEC_SIZE. Calculate "rdx + rcx - VEC_SIZE"
> > - with "rdx - (VEC_SIZE - rcx)" instead of "(rdx + rcx) - VEC_SIZE"
> > - to void possible addition overflow. */
> > - negq %rcx
> > - addq $VEC_SIZE, %rcx
> > -
> > - /* Check the end of data. */
> > - subq %rcx, %rsi
> > - jbe L(max)
> > + /* Use ecx which was computed earlier to compute correct value.
> > + */
> > + leal -(CHAR_PER_VEC * 2 + 1)(%rcx, %rax), %eax
> > +# else
> > + subl %edx, %edi
> > +# ifdef USE_AS_WCSLEN
> > + /* NB: Divide bytes by 4 to get the wchar_t count. */
> > + sarl $2, %edi
> > +# endif
> > + leal (CHAR_PER_VEC * 3)(%rdi, %rax), %eax
> > # endif
> > + ret
> >
> > - addq $VEC_SIZE, %rdi
> > -
> > + .p2align 4
> > +L(first_vec_x4):
> > + tzcntl %eax, %eax
> > + /* Safe to use 32 bit instructions as these are only called for
> > + size = [1, 159]. */
> > # ifdef USE_AS_STRNLEN
> > - subq $(VEC_SIZE * 4), %rsi
> > - jbe L(last_4x_vec_or_less)
> > + /* Use ecx which was computed earlier to compute correct value.
> > + */
> > + leal -(CHAR_PER_VEC + 1)(%rcx, %rax), %eax
> > +# else
> > + subl %edx, %edi
> > +# ifdef USE_AS_WCSLEN
> > + /* NB: Divide bytes by 4 to get the wchar_t count. */
> > + sarl $2, %edi
> > +# endif
> > + leal (CHAR_PER_VEC * 4)(%rdi, %rax), %eax
> > # endif
> > + ret
> >
> > -L(more_4x_vec):
> > + .p2align 5
> > +L(aligned_more):
> > + movq %rdi, %rdx
> > + /* Align data to VEC_SIZE. */
> > + andq $-(VEC_SIZE), %rdi
> > +L(cross_page_continue):
> > /* Check the first 4 * VEC_SIZE. Only one VEC_SIZE at a time
> > since data is only aligned to VEC_SIZE. */
> > - VPCMP $0, (%rdi), %YMMZERO, %k0
> > - kmovd %k0, %eax
> > - testl %eax, %eax
> > - jnz L(first_vec_x0)
> > -
> > +# ifdef USE_AS_STRNLEN
> > + /* + CHAR_SIZE because it simplies the logic in
> > + last_4x_vec_or_less. */
> > + leaq (VEC_SIZE * 5 + CHAR_SIZE)(%rdi), %rcx
> > + subq %rdx, %rcx
> > +# ifdef USE_AS_WCSLEN
> > + /* NB: Divide bytes by 4 to get the wchar_t count. */
> > + sarl $2, %ecx
> > +# endif
> > +# endif
> > + /* Load first VEC regardless. */
> > VPCMP $0, VEC_SIZE(%rdi), %YMMZERO, %k0
> > +# ifdef USE_AS_STRNLEN
> > + /* Adjust length. If near end handle specially. */
> > + subq %rcx, %rsi
> > + jb L(last_4x_vec_or_less)
> > +# endif
> > kmovd %k0, %eax
> > testl %eax, %eax
> > jnz L(first_vec_x1)
> >
> > VPCMP $0, (VEC_SIZE * 2)(%rdi), %YMMZERO, %k0
> > kmovd %k0, %eax
> > - testl %eax, %eax
> > + test %eax, %eax
> > jnz L(first_vec_x2)
> >
> > VPCMP $0, (VEC_SIZE * 3)(%rdi), %YMMZERO, %k0
> > @@ -179,258 +217,276 @@ L(more_4x_vec):
> > testl %eax, %eax
> > jnz L(first_vec_x3)
> >
> > - addq $(VEC_SIZE * 4), %rdi
> > -
> > -# ifdef USE_AS_STRNLEN
> > - subq $(VEC_SIZE * 4), %rsi
> > - jbe L(last_4x_vec_or_less)
> > -# endif
> > -
> > - /* Align data to 4 * VEC_SIZE. */
> > - movq %rdi, %rcx
> > - andl $(4 * VEC_SIZE - 1), %ecx
> > - andq $-(4 * VEC_SIZE), %rdi
> > + VPCMP $0, (VEC_SIZE * 4)(%rdi), %YMMZERO, %k0
> > + kmovd %k0, %eax
> > + testl %eax, %eax
> > + jnz L(first_vec_x4)
> >
> > + addq $VEC_SIZE, %rdi
> > # ifdef USE_AS_STRNLEN
> > - /* Adjust length. */
> > + /* Check if at last VEC_SIZE * 4 length. */
> > + cmpq $(CHAR_PER_VEC * 4 - 1), %rsi
> > + jbe L(last_4x_vec_or_less_load)
> > + movl %edi, %ecx
> > + andl $(VEC_SIZE * 4 - 1), %ecx
> > +# ifdef USE_AS_WCSLEN
> > + /* NB: Divide bytes by 4 to get the wchar_t count. */
> > + sarl $2, %ecx
> > +# endif
> > + /* Readjust length. */
> > addq %rcx, %rsi
> > # endif
> > + /* Align data to VEC_SIZE * 4. */
> > + andq $-(VEC_SIZE * 4), %rdi
> >
> > + /* Compare 4 * VEC at a time forward. */
> > .p2align 4
> > L(loop_4x_vec):
> > - /* Compare 4 * VEC at a time forward. */
> > - VMOVA (%rdi), %YMM1
> > - VMOVA VEC_SIZE(%rdi), %YMM2
> > - VMOVA (VEC_SIZE * 2)(%rdi), %YMM3
> > - VMOVA (VEC_SIZE * 3)(%rdi), %YMM4
> > -
> > - VPMINU %YMM1, %YMM2, %YMM5
> > - VPMINU %YMM3, %YMM4, %YMM6
> > + /* Load first VEC regardless. */
> > + VMOVA (VEC_SIZE * 4)(%rdi), %YMM1
> > +# ifdef USE_AS_STRNLEN
> > + /* Break if at end of length. */
> > + subq $(CHAR_PER_VEC * 4), %rsi
> > + jb L(last_4x_vec_or_less_cmpeq)
> > +# endif
> > + /* Save some code size by microfusing VPMINU with the load. Since
> > + the matches in ymm2/ymm4 can only be returned if there where no
> > + matches in ymm1/ymm3 respectively there is no issue with overlap.
> > + */
> > + VPMINU (VEC_SIZE * 5)(%rdi), %YMM1, %YMM2
> > + VMOVA (VEC_SIZE * 6)(%rdi), %YMM3
> > + VPMINU (VEC_SIZE * 7)(%rdi), %YMM3, %YMM4
> > +
> > + VPCMP $0, %YMM2, %YMMZERO, %k0
> > + VPCMP $0, %YMM4, %YMMZERO, %k1
> > + subq $-(VEC_SIZE * 4), %rdi
> > + kortestd %k0, %k1
> > + jz L(loop_4x_vec)
> > +
> > + /* Check if end was in first half. */
> > + kmovd %k0, %eax
> > + subq %rdx, %rdi
> > +# ifdef USE_AS_WCSLEN
> > + shrq $2, %rdi
> > +# endif
> > + testl %eax, %eax
> > + jz L(second_vec_return)
> >
> > - VPMINU %YMM5, %YMM6, %YMM5
> > - VPCMP $0, %YMM5, %YMMZERO, %k0
> > - ktestd %k0, %k0
> > - jnz L(4x_vec_end)
> > + VPCMP $0, %YMM1, %YMMZERO, %k2
> > + kmovd %k2, %edx
> > + /* Combine VEC1 matches (edx) with VEC2 matches (eax). */
> > +# ifdef USE_AS_WCSLEN
> > + sall $CHAR_PER_VEC, %eax
> > + orl %edx, %eax
> > + tzcntl %eax, %eax
> > +# else
> > + salq $CHAR_PER_VEC, %rax
> > + orq %rdx, %rax
> > + tzcntq %rax, %rax
> > +# endif
> > + addq %rdi, %rax
> > + ret
> >
> > - addq $(VEC_SIZE * 4), %rdi
> >
> > -# ifndef USE_AS_STRNLEN
> > - jmp L(loop_4x_vec)
> > -# else
> > - subq $(VEC_SIZE * 4), %rsi
> > - ja L(loop_4x_vec)
> > +# ifdef USE_AS_STRNLEN
> >
> > +L(last_4x_vec_or_less_load):
> > + /* Depending on entry adjust rdi / prepare first VEC in YMM1. */
> > + VMOVA (VEC_SIZE * 4)(%rdi), %YMM1
> > +L(last_4x_vec_or_less_cmpeq):
> > + VPCMP $0, %YMM1, %YMMZERO, %k0
> > + addq $(VEC_SIZE * 3), %rdi
> > L(last_4x_vec_or_less):
> > - /* Less than 4 * VEC and aligned to VEC_SIZE. */
> > - addl $(VEC_SIZE * 2), %esi
> > - jle L(last_2x_vec)
> > -
> > - VPCMP $0, (%rdi), %YMMZERO, %k0
> > kmovd %k0, %eax
> > + /* If remaining length > VEC_SIZE * 2. This works if esi is off by
> > + VEC_SIZE * 4. */
> > + testl $(CHAR_PER_VEC * 2), %esi
> > + jnz L(last_4x_vec)
> > +
> > + /* length may have been negative or positive by an offset of
> > + CHAR_PER_VEC * 4 depending on where this was called from. This
> > + fixes that. */
> > + andl $(CHAR_PER_VEC * 4 - 1), %esi
> > testl %eax, %eax
> > - jnz L(first_vec_x0)
> > + jnz L(last_vec_x1_check)
> >
> > - VPCMP $0, VEC_SIZE(%rdi), %YMMZERO, %k0
> > - kmovd %k0, %eax
> > - testl %eax, %eax
> > - jnz L(first_vec_x1)
> > + /* Check the end of data. */
> > + subl $CHAR_PER_VEC, %esi
> > + jb L(max)
> >
> > VPCMP $0, (VEC_SIZE * 2)(%rdi), %YMMZERO, %k0
> > kmovd %k0, %eax
> > - testl %eax, %eax
> > - jnz L(first_vec_x2_check)
> > - subl $VEC_SIZE, %esi
> > - jle L(max)
> > + tzcntl %eax, %eax
> > + /* Check the end of data. */
> > + cmpl %eax, %esi
> > + jb L(max)
> >
> > - VPCMP $0, (VEC_SIZE * 3)(%rdi), %YMMZERO, %k0
> > - kmovd %k0, %eax
> > - testl %eax, %eax
> > - jnz L(first_vec_x3_check)
> > + subq %rdx, %rdi
> > +# ifdef USE_AS_WCSLEN
> > + /* NB: Divide bytes by 4 to get the wchar_t count. */
> > + sarq $2, %rdi
> > +# endif
> > + leaq (CHAR_PER_VEC * 2)(%rdi, %rax), %rax
> > + ret
> > +L(max):
> > movq %r8, %rax
> > + ret
> > +# endif
> > +
> > + /* Placed here in strnlen so that the jcc L(last_4x_vec_or_less)
> > + in the 4x VEC loop can use 2 byte encoding. */
> > + .p2align 4
> > +L(second_vec_return):
> > + VPCMP $0, %YMM3, %YMMZERO, %k0
> > + /* Combine YMM3 matches (k0) with YMM4 matches (k1). */
> > +# ifdef USE_AS_WCSLEN
> > + kunpckbw %k0, %k1, %k0
> > + kmovd %k0, %eax
> > + tzcntl %eax, %eax
> > +# else
> > + kunpckdq %k0, %k1, %k0
> > + kmovq %k0, %rax
> > + tzcntq %rax, %rax
> > +# endif
> > + leaq (CHAR_PER_VEC * 2)(%rdi, %rax), %rax
> > + ret
> > +
> > +
> > +# ifdef USE_AS_STRNLEN
> > +L(last_vec_x1_check):
> > + tzcntl %eax, %eax
> > + /* Check the end of data. */
> > + cmpl %eax, %esi
> > + jb L(max)
> > + subq %rdx, %rdi
> > # ifdef USE_AS_WCSLEN
> > - shrq $2, %rax
> > + /* NB: Divide bytes by 4 to get the wchar_t count. */
> > + sarq $2, %rdi
> > # endif
> > + leaq (CHAR_PER_VEC)(%rdi, %rax), %rax
> > ret
> >
> > .p2align 4
> > -L(last_2x_vec):
> > - addl $(VEC_SIZE * 2), %esi
> > +L(last_4x_vec):
> > + /* Test first 2x VEC normally. */
> > + testl %eax, %eax
> > + jnz L(last_vec_x1)
> >
> > - VPCMP $0, (%rdi), %YMMZERO, %k0
> > + VPCMP $0, (VEC_SIZE * 2)(%rdi), %YMMZERO, %k0
> > kmovd %k0, %eax
> > testl %eax, %eax
> > - jnz L(first_vec_x0_check)
> > - subl $VEC_SIZE, %esi
> > - jle L(max)
> > + jnz L(last_vec_x2)
> >
> > - VPCMP $0, VEC_SIZE(%rdi), %YMMZERO, %k0
> > + /* Normalize length. */
> > + andl $(CHAR_PER_VEC * 4 - 1), %esi
> > + VPCMP $0, (VEC_SIZE * 3)(%rdi), %YMMZERO, %k0
> > kmovd %k0, %eax
> > testl %eax, %eax
> > - jnz L(first_vec_x1_check)
> > - movq %r8, %rax
> > -# ifdef USE_AS_WCSLEN
> > - shrq $2, %rax
> > -# endif
> > - ret
> > + jnz L(last_vec_x3)
> >
> > - .p2align 4
> > -L(first_vec_x0_check):
> > + /* Check the end of data. */
> > + subl $(CHAR_PER_VEC * 3), %esi
> > + jb L(max)
> > +
> > + VPCMP $0, (VEC_SIZE * 4)(%rdi), %YMMZERO, %k0
> > + kmovd %k0, %eax
> > tzcntl %eax, %eax
> > -# ifdef USE_AS_WCSLEN
> > - /* NB: Multiply wchar_t count by 4 to get the number of bytes. */
> > - sall $2, %eax
> > -# endif
> > /* Check the end of data. */
> > - cmpq %rax, %rsi
> > - jbe L(max)
> > - addq %rdi, %rax
> > - subq %rdx, %rax
> > + cmpl %eax, %esi
> > + jb L(max_end)
> > +
> > + subq %rdx, %rdi
> > # ifdef USE_AS_WCSLEN
> > - shrq $2, %rax
> > + /* NB: Divide bytes by 4 to get the wchar_t count. */
> > + sarq $2, %rdi
> > # endif
> > + leaq (CHAR_PER_VEC * 4)(%rdi, %rax), %rax
> > ret
> >
> > .p2align 4
> > -L(first_vec_x1_check):
> > +L(last_vec_x1):
> > tzcntl %eax, %eax
> > + subq %rdx, %rdi
> > # ifdef USE_AS_WCSLEN
> > - /* NB: Multiply wchar_t count by 4 to get the number of bytes. */
> > - sall $2, %eax
> > -# endif
> > - /* Check the end of data. */
> > - cmpq %rax, %rsi
> > - jbe L(max)
> > - addq $VEC_SIZE, %rax
> > - addq %rdi, %rax
> > - subq %rdx, %rax
> > -# ifdef USE_AS_WCSLEN
> > - shrq $2, %rax
> > + /* NB: Divide bytes by 4 to get the wchar_t count. */
> > + sarq $2, %rdi
> > # endif
> > + leaq (CHAR_PER_VEC)(%rdi, %rax), %rax
> > ret
> >
> > .p2align 4
> > -L(first_vec_x2_check):
> > +L(last_vec_x2):
> > tzcntl %eax, %eax
> > + subq %rdx, %rdi
> > # ifdef USE_AS_WCSLEN
> > - /* NB: Multiply wchar_t count by 4 to get the number of bytes. */
> > - sall $2, %eax
> > -# endif
> > - /* Check the end of data. */
> > - cmpq %rax, %rsi
> > - jbe L(max)
> > - addq $(VEC_SIZE * 2), %rax
> > - addq %rdi, %rax
> > - subq %rdx, %rax
> > -# ifdef USE_AS_WCSLEN
> > - shrq $2, %rax
> > + /* NB: Divide bytes by 4 to get the wchar_t count. */
> > + sarq $2, %rdi
> > # endif
> > + leaq (CHAR_PER_VEC * 2)(%rdi, %rax), %rax
> > ret
> >
> > .p2align 4
> > -L(first_vec_x3_check):
> > +L(last_vec_x3):
> > tzcntl %eax, %eax
> > -# ifdef USE_AS_WCSLEN
> > - /* NB: Multiply wchar_t count by 4 to get the number of bytes. */
> > - sall $2, %eax
> > -# endif
> > + subl $(CHAR_PER_VEC * 2), %esi
> > /* Check the end of data. */
> > - cmpq %rax, %rsi
> > - jbe L(max)
> > - addq $(VEC_SIZE * 3), %rax
> > - addq %rdi, %rax
> > - subq %rdx, %rax
> > + cmpl %eax, %esi
> > + jb L(max_end)
> > + subq %rdx, %rdi
> > # ifdef USE_AS_WCSLEN
> > - shrq $2, %rax
> > + /* NB: Divide bytes by 4 to get the wchar_t count. */
> > + sarq $2, %rdi
> > # endif
> > + leaq (CHAR_PER_VEC * 3)(%rdi, %rax), %rax
> > ret
> > -
> > - .p2align 4
> > -L(max):
> > +L(max_end):
> > movq %r8, %rax
> > -# ifdef USE_AS_WCSLEN
> > - shrq $2, %rax
> > -# endif
> > - ret
> > -
> > - .p2align 4
> > -L(zero):
> > - xorl %eax, %eax
> > ret
> > # endif
> >
> > + /* Cold case for crossing page with first load. */
> > .p2align 4
> > -L(first_vec_x0):
> > - tzcntl %eax, %eax
> > -# ifdef USE_AS_WCSLEN
> > - /* NB: Multiply wchar_t count by 4 to get the number of bytes. */
> > - sall $2, %eax
> > -# endif
> > - addq %rdi, %rax
> > - subq %rdx, %rax
> > +L(cross_page_boundary):
> > + movq %rdi, %rdx
> > + /* Align data to VEC_SIZE. */
> > + andq $-VEC_SIZE, %rdi
> > + VPCMP $0, (%rdi), %YMMZERO, %k0
> > + kmovd %k0, %eax
> > + /* Remove the leading bytes. */
> > # ifdef USE_AS_WCSLEN
> > - shrq $2, %rax
> > + /* NB: Divide shift count by 4 since each bit in K0 represent 4
> > + bytes. */
> > + movl %edx, %ecx
> > + shrl $2, %ecx
> > + andl $(CHAR_PER_VEC - 1), %ecx
> > # endif
> > - ret
> > -
> > - .p2align 4
> > -L(first_vec_x1):
> > + /* SHIFT_REG is ecx for USE_AS_WCSLEN and edx otherwise. */
> > + sarxl %SHIFT_REG, %eax, %eax
> > + testl %eax, %eax
> > +# ifndef USE_AS_STRNLEN
> > + jz L(cross_page_continue)
> > tzcntl %eax, %eax
> > -# ifdef USE_AS_WCSLEN
> > - /* NB: Multiply wchar_t count by 4 to get the number of bytes. */
> > - sall $2, %eax
> > -# endif
> > - addq $VEC_SIZE, %rax
> > - addq %rdi, %rax
> > - subq %rdx, %rax
> > -# ifdef USE_AS_WCSLEN
> > - shrq $2, %rax
> > -# endif
> > ret
> > -
> > - .p2align 4
> > -L(first_vec_x2):
> > - tzcntl %eax, %eax
> > -# ifdef USE_AS_WCSLEN
> > - /* NB: Multiply wchar_t count by 4 to get the number of bytes. */
> > - sall $2, %eax
> > -# endif
> > - addq $(VEC_SIZE * 2), %rax
> > - addq %rdi, %rax
> > - subq %rdx, %rax
> > -# ifdef USE_AS_WCSLEN
> > - shrq $2, %rax
> > -# endif
> > +# else
> > + jnz L(cross_page_less_vec)
> > +# ifndef USE_AS_WCSLEN
> > + movl %edx, %ecx
> > + andl $(CHAR_PER_VEC - 1), %ecx
> > +# endif
> > + movl $CHAR_PER_VEC, %eax
> > + subl %ecx, %eax
> > + /* Check the end of data. */
> > + cmpq %rax, %rsi
> > + ja L(cross_page_continue)
> > + movl %esi, %eax
> > ret
> > -
> > - .p2align 4
> > -L(4x_vec_end):
> > - VPCMP $0, %YMM1, %YMMZERO, %k0
> > - kmovd %k0, %eax
> > - testl %eax, %eax
> > - jnz L(first_vec_x0)
> > - VPCMP $0, %YMM2, %YMMZERO, %k1
> > - kmovd %k1, %eax
> > - testl %eax, %eax
> > - jnz L(first_vec_x1)
> > - VPCMP $0, %YMM3, %YMMZERO, %k2
> > - kmovd %k2, %eax
> > - testl %eax, %eax
> > - jnz L(first_vec_x2)
> > - VPCMP $0, %YMM4, %YMMZERO, %k3
> > - kmovd %k3, %eax
> > -L(first_vec_x3):
> > +L(cross_page_less_vec):
> > tzcntl %eax, %eax
> > -# ifdef USE_AS_WCSLEN
> > - /* NB: Multiply wchar_t count by 4 to get the number of bytes. */
> > - sall $2, %eax
> > -# endif
> > - addq $(VEC_SIZE * 3), %rax
> > - addq %rdi, %rax
> > - subq %rdx, %rax
> > -# ifdef USE_AS_WCSLEN
> > - shrq $2, %rax
> > -# endif
> > + /* Select min of length and position of first null. */
> > + cmpq %rax, %rsi
> > + cmovb %esi, %eax
> > ret
> > +# endif
> >
> > END (STRLEN)
> > #endif
> > --
> > 2.29.2
> >
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: [PATCH v4 2/2] x86: Optimize strlen-avx2.S
2021-04-19 22:16 ` H.J. Lu
@ 2021-04-19 23:39 ` Noah Goldstein
0 siblings, 0 replies; 6+ messages in thread
From: Noah Goldstein @ 2021-04-19 23:39 UTC (permalink / raw)
To: H.J. Lu; +Cc: GNU C Library, Carlos O'Donell
On Mon, Apr 19, 2021 at 6:16 PM H.J. Lu <hjl.tools@gmail.com> wrote:
>
> On Mon, Apr 19, 2021 at 02:43:45PM -0400, Noah Goldstein wrote:
> > No bug. This commit optimizes strlen-avx2.S. The optimizations are
> > mostly small things but they add up to roughly 10-30% performance
> > improvement for strlen. The results for strnlen are bit more
> > ambiguous. test-strlen, test-strnlen, test-wcslen, and test-wcsnlen
> > are all passing.
> >
> > Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com>
> > ---
> > sysdeps/x86_64/multiarch/ifunc-impl-list.c | 16 +-
> > sysdeps/x86_64/multiarch/strlen-avx2.S | 535 +++++++++++++--------
> > 2 files changed, 337 insertions(+), 214 deletions(-)
> >
> > diff --git a/sysdeps/x86_64/multiarch/ifunc-impl-list.c b/sysdeps/x86_64/multiarch/ifunc-impl-list.c
> > index 0b0927b124..4613bead7c 100644
> > --- a/sysdeps/x86_64/multiarch/ifunc-impl-list.c
> > +++ b/sysdeps/x86_64/multiarch/ifunc-impl-list.c
> > @@ -285,10 +285,12 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
> > /* Support sysdeps/x86_64/multiarch/strlen.c. */
> > IFUNC_IMPL (i, name, strlen,
> > IFUNC_IMPL_ADD (array, i, strlen,
> > - CPU_FEATURE_USABLE (AVX2),
> > + (CPU_FEATURE_USABLE (AVX2)
> > + && CPU_FEATURE_USABLE (BMI2)),
> > __strlen_avx2)
> > IFUNC_IMPL_ADD (array, i, strlen,
> > (CPU_FEATURE_USABLE (AVX2)
> > + && CPU_FEATURE_USABLE (BMI2)
> > && CPU_FEATURE_USABLE (RTM)),
> > __strlen_avx2_rtm)
> > IFUNC_IMPL_ADD (array, i, strlen,
> > @@ -301,10 +303,12 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
> > /* Support sysdeps/x86_64/multiarch/strnlen.c. */
> > IFUNC_IMPL (i, name, strnlen,
> > IFUNC_IMPL_ADD (array, i, strnlen,
> > - CPU_FEATURE_USABLE (AVX2),
> > + (CPU_FEATURE_USABLE (AVX2)
> > + && CPU_FEATURE_USABLE (BMI2)),
> > __strnlen_avx2)
> > IFUNC_IMPL_ADD (array, i, strnlen,
> > (CPU_FEATURE_USABLE (AVX2)
> > + && CPU_FEATURE_USABLE (BMI2)
> > && CPU_FEATURE_USABLE (RTM)),
> > __strnlen_avx2_rtm)
> > IFUNC_IMPL_ADD (array, i, strnlen,
> > @@ -646,10 +650,12 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
> > /* Support sysdeps/x86_64/multiarch/wcslen.c. */
> > IFUNC_IMPL (i, name, wcslen,
> > IFUNC_IMPL_ADD (array, i, wcslen,
> > - CPU_FEATURE_USABLE (AVX2),
> > + (CPU_FEATURE_USABLE (AVX2)
> > + && CPU_FEATURE_USABLE (BMI2)),
> > __wcslen_avx2)
> > IFUNC_IMPL_ADD (array, i, wcslen,
> > (CPU_FEATURE_USABLE (AVX2)
> > + && CPU_FEATURE_USABLE (BMI2)
> > && CPU_FEATURE_USABLE (RTM)),
> > __wcslen_avx2_rtm)
> > IFUNC_IMPL_ADD (array, i, wcslen,
> > @@ -662,10 +668,12 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
> > /* Support sysdeps/x86_64/multiarch/wcsnlen.c. */
> > IFUNC_IMPL (i, name, wcsnlen,
> > IFUNC_IMPL_ADD (array, i, wcsnlen,
> > - CPU_FEATURE_USABLE (AVX2),
> > + (CPU_FEATURE_USABLE (AVX2)
> > + && CPU_FEATURE_USABLE (BMI2)),
> > __wcsnlen_avx2)
> > IFUNC_IMPL_ADD (array, i, wcsnlen,
> > (CPU_FEATURE_USABLE (AVX2)
> > + && CPU_FEATURE_USABLE (BMI2)
> > && CPU_FEATURE_USABLE (RTM)),
> > __wcsnlen_avx2_rtm)
> > IFUNC_IMPL_ADD (array, i, wcsnlen,
> > diff --git a/sysdeps/x86_64/multiarch/strlen-avx2.S b/sysdeps/x86_64/multiarch/strlen-avx2.S
> > index 1caae9e6bc..eb92f587b4 100644
> > --- a/sysdeps/x86_64/multiarch/strlen-avx2.S
> > +++ b/sysdeps/x86_64/multiarch/strlen-avx2.S
> > @@ -27,9 +27,11 @@
> > # ifdef USE_AS_WCSLEN
> > # define VPCMPEQ vpcmpeqd
> > # define VPMINU vpminud
> > +# define CHAR_SIZE 4
> > # else
> > # define VPCMPEQ vpcmpeqb
> > # define VPMINU vpminub
> > +# define CHAR_SIZE 1
> > # endif
> >
> > # ifndef VZEROUPPER
> > @@ -41,349 +43,462 @@
> > # endif
> >
> > # define VEC_SIZE 32
> > +# define PAGE_SIZE 4096
> > +# define LOG_PAGE_SIZE 12
> >
> > .section SECTION(.text),"ax",@progbits
> > ENTRY (STRLEN)
> > # ifdef USE_AS_STRNLEN
> > - /* Check for zero length. */
> > + /* Check zero length. */
> > test %RSI_LP, %RSI_LP
> > jz L(zero)
> > + /* Store max len in R8_LP before adjusting if using WCSLEN. */
> > + mov %RSI_LP, %R8_LP
> > # ifdef USE_AS_WCSLEN
> > shl $2, %RSI_LP
> > # elif defined __ILP32__
> > /* Clear the upper 32 bits. */
> > movl %esi, %esi
> > # endif
> > - mov %RSI_LP, %R8_LP
> > # endif
> > - movl %edi, %ecx
> > + movl %edi, %eax
> > movq %rdi, %rdx
> > vpxor %xmm0, %xmm0, %xmm0
> > -
> > + /* Shift left eax to clear all bits not relevant to page cross
> > + check. This saves 2 bytes of code as opposed to using andl with
> > + PAGE_SIZE - 1. Then compare with PAGE_SIZE - VEC_SIZE shifted
> > + left by the same amount (an imm32 either way). */
> > + sall $(32 - LOG_PAGE_SIZE), %eax
>
> I prefer AND over SAL for better throughput.
Done.
>
> H.J.
> > /* Check if we may cross page boundary with one vector load. */
> > - andl $(2 * VEC_SIZE - 1), %ecx
> > - cmpl $VEC_SIZE, %ecx
> > - ja L(cros_page_boundary)
> > + cmpl $((PAGE_SIZE - VEC_SIZE) << (32 - LOG_PAGE_SIZE)), %eax
> > + ja L(cross_page_boundary)
> >
> > /* Check the first VEC_SIZE bytes. */
> > - VPCMPEQ (%rdi), %ymm0, %ymm1
> > - vpmovmskb %ymm1, %eax
> > - testl %eax, %eax
> > -
> > + VPCMPEQ (%rdi), %ymm0, %ymm1
> > + vpmovmskb %ymm1, %eax
> > # ifdef USE_AS_STRNLEN
> > - jnz L(first_vec_x0_check)
> > - /* Adjust length and check the end of data. */
> > - subq $VEC_SIZE, %rsi
> > - jbe L(max)
> > -# else
> > - jnz L(first_vec_x0)
> > + /* If length < VEC_SIZE handle special. */
> > + cmpq $VEC_SIZE, %rsi
> > + jbe L(first_vec_x0)
> > # endif
> > -
> > - /* Align data for aligned loads in the loop. */
> > - addq $VEC_SIZE, %rdi
> > - andl $(VEC_SIZE - 1), %ecx
> > - andq $-VEC_SIZE, %rdi
> > + /* If empty continue to aligned_more. Otherwise return bit
> > + position of first match. */
> > + testl %eax, %eax
> > + jz L(aligned_more)
> > + tzcntl %eax, %eax
> > +# ifdef USE_AS_WCSLEN
> > + shrl $2, %eax
> > +# endif
> > + VZEROUPPER_RETURN
> >
> > # ifdef USE_AS_STRNLEN
> > - /* Adjust length. */
> > - addq %rcx, %rsi
> > +L(zero):
> > + xorl %eax, %eax
> > + ret
> >
> > - subq $(VEC_SIZE * 4), %rsi
> > - jbe L(last_4x_vec_or_less)
> > + .p2align 4
> > +L(first_vec_x0):
> > + /* Set bit for max len so that tzcnt will return min of max len
> > + and position of first match. */
> > + btsq %rsi, %rax
> > + tzcntl %eax, %eax
> > +# ifdef USE_AS_WCSLEN
> > + shrl $2, %eax
> > +# endif
> > + VZEROUPPER_RETURN
> > # endif
> > - jmp L(more_4x_vec)
> >
> > .p2align 4
> > -L(cros_page_boundary):
> > - andl $(VEC_SIZE - 1), %ecx
> > - andq $-VEC_SIZE, %rdi
> > - VPCMPEQ (%rdi), %ymm0, %ymm1
> > - vpmovmskb %ymm1, %eax
> > - /* Remove the leading bytes. */
> > - sarl %cl, %eax
> > - testl %eax, %eax
> > - jz L(aligned_more)
> > +L(first_vec_x1):
> > tzcntl %eax, %eax
> > + /* Safe to use 32 bit instructions as these are only called for
> > + size = [1, 159]. */
> > # ifdef USE_AS_STRNLEN
> > - /* Check the end of data. */
> > - cmpq %rax, %rsi
> > - jbe L(max)
> > + /* Use ecx which was computed earlier to compute correct value.
> > + */
> > + subl $(VEC_SIZE * 4 + 1), %ecx
> > + addl %ecx, %eax
> > +# else
> > + subl %edx, %edi
> > + incl %edi
> > + addl %edi, %eax
> > # endif
> > - addq %rdi, %rax
> > - addq %rcx, %rax
> > - subq %rdx, %rax
> > # ifdef USE_AS_WCSLEN
> > - shrq $2, %rax
> > + shrl $2, %eax
> > # endif
> > -L(return_vzeroupper):
> > - ZERO_UPPER_VEC_REGISTERS_RETURN
> > + VZEROUPPER_RETURN
> >
> > .p2align 4
> > -L(aligned_more):
> > +L(first_vec_x2):
> > + tzcntl %eax, %eax
> > + /* Safe to use 32 bit instructions as these are only called for
> > + size = [1, 159]. */
> > # ifdef USE_AS_STRNLEN
> > - /* "rcx" is less than VEC_SIZE. Calculate "rdx + rcx - VEC_SIZE"
> > - with "rdx - (VEC_SIZE - rcx)" instead of "(rdx + rcx) - VEC_SIZE"
> > - to void possible addition overflow. */
> > - negq %rcx
> > - addq $VEC_SIZE, %rcx
> > -
> > - /* Check the end of data. */
> > - subq %rcx, %rsi
> > - jbe L(max)
> > + /* Use ecx which was computed earlier to compute correct value.
> > + */
> > + subl $(VEC_SIZE * 3 + 1), %ecx
> > + addl %ecx, %eax
> > +# else
> > + subl %edx, %edi
> > + addl $(VEC_SIZE + 1), %edi
> > + addl %edi, %eax
> > # endif
> > +# ifdef USE_AS_WCSLEN
> > + shrl $2, %eax
> > +# endif
> > + VZEROUPPER_RETURN
> >
> > - addq $VEC_SIZE, %rdi
> > + .p2align 4
> > +L(first_vec_x3):
> > + tzcntl %eax, %eax
> > + /* Safe to use 32 bit instructions as these are only called for
> > + size = [1, 159]. */
> > +# ifdef USE_AS_STRNLEN
> > + /* Use ecx which was computed earlier to compute correct value.
> > + */
> > + subl $(VEC_SIZE * 2 + 1), %ecx
> > + addl %ecx, %eax
> > +# else
> > + subl %edx, %edi
> > + addl $(VEC_SIZE * 2 + 1), %edi
> > + addl %edi, %eax
> > +# endif
> > +# ifdef USE_AS_WCSLEN
> > + shrl $2, %eax
> > +# endif
> > + VZEROUPPER_RETURN
> >
> > + .p2align 4
> > +L(first_vec_x4):
> > + tzcntl %eax, %eax
> > + /* Safe to use 32 bit instructions as these are only called for
> > + size = [1, 159]. */
> > # ifdef USE_AS_STRNLEN
> > - subq $(VEC_SIZE * 4), %rsi
> > - jbe L(last_4x_vec_or_less)
> > + /* Use ecx which was computed earlier to compute correct value.
> > + */
> > + subl $(VEC_SIZE + 1), %ecx
> > + addl %ecx, %eax
> > +# else
> > + subl %edx, %edi
> > + addl $(VEC_SIZE * 3 + 1), %edi
> > + addl %edi, %eax
> > # endif
> > +# ifdef USE_AS_WCSLEN
> > + shrl $2, %eax
> > +# endif
> > + VZEROUPPER_RETURN
> >
> > -L(more_4x_vec):
> > + .p2align 5
> > +L(aligned_more):
> > + /* Align data to VEC_SIZE - 1. This is the same number of
> > + instructions as using andq with -VEC_SIZE but saves 4 bytes of
> > + code on the x4 check. */
> > + orq $(VEC_SIZE - 1), %rdi
> > +L(cross_page_continue):
> > /* Check the first 4 * VEC_SIZE. Only one VEC_SIZE at a time
> > since data is only aligned to VEC_SIZE. */
> > - VPCMPEQ (%rdi), %ymm0, %ymm1
> > - vpmovmskb %ymm1, %eax
> > - testl %eax, %eax
> > - jnz L(first_vec_x0)
> > -
> > - VPCMPEQ VEC_SIZE(%rdi), %ymm0, %ymm1
> > - vpmovmskb %ymm1, %eax
> > +# ifdef USE_AS_STRNLEN
> > + /* + 1 because rdi is aligned to VEC_SIZE - 1. + CHAR_SIZE because
> > + it simplies the logic in last_4x_vec_or_less. */
> > + leaq (VEC_SIZE * 4 + CHAR_SIZE + 1)(%rdi), %rcx
> > + subq %rdx, %rcx
> > +# endif
> > + /* Load first VEC regardless. */
> > + VPCMPEQ 1(%rdi), %ymm0, %ymm1
> > +# ifdef USE_AS_STRNLEN
> > + /* Adjust length. If near end handle specially. */
> > + subq %rcx, %rsi
> > + jb L(last_4x_vec_or_less)
> > +# endif
> > + vpmovmskb %ymm1, %eax
> > testl %eax, %eax
> > jnz L(first_vec_x1)
> >
> > - VPCMPEQ (VEC_SIZE * 2)(%rdi), %ymm0, %ymm1
> > - vpmovmskb %ymm1, %eax
> > + VPCMPEQ (VEC_SIZE + 1)(%rdi), %ymm0, %ymm1
> > + vpmovmskb %ymm1, %eax
> > testl %eax, %eax
> > jnz L(first_vec_x2)
> >
> > - VPCMPEQ (VEC_SIZE * 3)(%rdi), %ymm0, %ymm1
> > - vpmovmskb %ymm1, %eax
> > + VPCMPEQ (VEC_SIZE * 2 + 1)(%rdi), %ymm0, %ymm1
> > + vpmovmskb %ymm1, %eax
> > testl %eax, %eax
> > jnz L(first_vec_x3)
> >
> > - addq $(VEC_SIZE * 4), %rdi
> > -
> > -# ifdef USE_AS_STRNLEN
> > - subq $(VEC_SIZE * 4), %rsi
> > - jbe L(last_4x_vec_or_less)
> > -# endif
> > -
> > - /* Align data to 4 * VEC_SIZE. */
> > - movq %rdi, %rcx
> > - andl $(4 * VEC_SIZE - 1), %ecx
> > - andq $-(4 * VEC_SIZE), %rdi
> > + VPCMPEQ (VEC_SIZE * 3 + 1)(%rdi), %ymm0, %ymm1
> > + vpmovmskb %ymm1, %eax
> > + testl %eax, %eax
> > + jnz L(first_vec_x4)
> >
> > + /* Align data to VEC_SIZE * 4 - 1. */
> > # ifdef USE_AS_STRNLEN
> > - /* Adjust length. */
> > + /* Before adjusting length check if at last VEC_SIZE * 4. */
> > + cmpq $(VEC_SIZE * 4 - 1), %rsi
> > + jbe L(last_4x_vec_or_less_load)
> > + incq %rdi
> > + movl %edi, %ecx
> > + orq $(VEC_SIZE * 4 - 1), %rdi
> > + andl $(VEC_SIZE * 4 - 1), %ecx
> > + /* Readjust length. */
> > addq %rcx, %rsi
> > +# else
> > + incq %rdi
> > + orq $(VEC_SIZE * 4 - 1), %rdi
> > # endif
> > -
> > + /* Compare 4 * VEC at a time forward. */
> > .p2align 4
> > L(loop_4x_vec):
> > - /* Compare 4 * VEC at a time forward. */
> > - vmovdqa (%rdi), %ymm1
> > - vmovdqa VEC_SIZE(%rdi), %ymm2
> > - vmovdqa (VEC_SIZE * 2)(%rdi), %ymm3
> > - vmovdqa (VEC_SIZE * 3)(%rdi), %ymm4
> > - VPMINU %ymm1, %ymm2, %ymm5
> > - VPMINU %ymm3, %ymm4, %ymm6
> > - VPMINU %ymm5, %ymm6, %ymm5
> > -
> > - VPCMPEQ %ymm5, %ymm0, %ymm5
> > - vpmovmskb %ymm5, %eax
> > - testl %eax, %eax
> > - jnz L(4x_vec_end)
> > -
> > - addq $(VEC_SIZE * 4), %rdi
> > -
> > -# ifndef USE_AS_STRNLEN
> > - jmp L(loop_4x_vec)
> > -# else
> > +# ifdef USE_AS_STRNLEN
> > + /* Break if at end of length. */
> > subq $(VEC_SIZE * 4), %rsi
> > - ja L(loop_4x_vec)
> > -
> > -L(last_4x_vec_or_less):
> > - /* Less than 4 * VEC and aligned to VEC_SIZE. */
> > - addl $(VEC_SIZE * 2), %esi
> > - jle L(last_2x_vec)
> > + jb L(last_4x_vec_or_less_cmpeq)
> > +# endif
> > + /* Save some code size by microfusing VPMINU with the load. Since
> > + the matches in ymm2/ymm4 can only be returned if there where no
> > + matches in ymm1/ymm3 respectively there is no issue with overlap.
> > + */
> > + vmovdqa 1(%rdi), %ymm1
> > + VPMINU (VEC_SIZE + 1)(%rdi), %ymm1, %ymm2
> > + vmovdqa (VEC_SIZE * 2 + 1)(%rdi), %ymm3
> > + VPMINU (VEC_SIZE * 3 + 1)(%rdi), %ymm3, %ymm4
> > +
> > + VPMINU %ymm2, %ymm4, %ymm5
> > + VPCMPEQ %ymm5, %ymm0, %ymm5
> > + vpmovmskb %ymm5, %ecx
> >
> > - VPCMPEQ (%rdi), %ymm0, %ymm1
> > - vpmovmskb %ymm1, %eax
> > - testl %eax, %eax
> > - jnz L(first_vec_x0)
> > + subq $-(VEC_SIZE * 4), %rdi
> > + testl %ecx, %ecx
> > + jz L(loop_4x_vec)
> >
> > - VPCMPEQ VEC_SIZE(%rdi), %ymm0, %ymm1
> > - vpmovmskb %ymm1, %eax
> > - testl %eax, %eax
> > - jnz L(first_vec_x1)
> >
> > - VPCMPEQ (VEC_SIZE * 2)(%rdi), %ymm0, %ymm1
> > - vpmovmskb %ymm1, %eax
> > + VPCMPEQ %ymm1, %ymm0, %ymm1
> > + vpmovmskb %ymm1, %eax
> > + subq %rdx, %rdi
> > testl %eax, %eax
> > + jnz L(last_vec_return_x0)
> >
> > - jnz L(first_vec_x2_check)
> > - subl $VEC_SIZE, %esi
> > - jle L(max)
> > -
> > - VPCMPEQ (VEC_SIZE * 3)(%rdi), %ymm0, %ymm1
> > - vpmovmskb %ymm1, %eax
> > + VPCMPEQ %ymm2, %ymm0, %ymm2
> > + vpmovmskb %ymm2, %eax
> > testl %eax, %eax
> > -
> > - jnz L(first_vec_x3_check)
> > - movq %r8, %rax
> > -# ifdef USE_AS_WCSLEN
> > + jnz L(last_vec_return_x1)
> > +
> > + /* Combine last 2 VEC. */
> > + VPCMPEQ %ymm3, %ymm0, %ymm3
> > + vpmovmskb %ymm3, %eax
> > + /* rcx has combined result from all 4 VEC. It will only be used if
> > + the first 3 other VEC all did not contain a match. */
> > + salq $32, %rcx
> > + orq %rcx, %rax
> > + tzcntq %rax, %rax
> > + subq $(VEC_SIZE * 2 - 1), %rdi
> > + addq %rdi, %rax
> > +# ifdef USE_AS_WCSLEN
> > shrq $2, %rax
> > -# endif
> > +# endif
> > VZEROUPPER_RETURN
> >
> > +
> > +# ifdef USE_AS_STRNLEN
> > .p2align 4
> > -L(last_2x_vec):
> > - addl $(VEC_SIZE * 2), %esi
> > - VPCMPEQ (%rdi), %ymm0, %ymm1
> > - vpmovmskb %ymm1, %eax
> > - testl %eax, %eax
> > +L(last_4x_vec_or_less_load):
> > + /* Depending on entry adjust rdi / prepare first VEC in ymm1. */
> > + subq $-(VEC_SIZE * 4), %rdi
> > +L(last_4x_vec_or_less_cmpeq):
> > + VPCMPEQ 1(%rdi), %ymm0, %ymm1
> > +L(last_4x_vec_or_less):
> >
> > - jnz L(first_vec_x0_check)
> > - subl $VEC_SIZE, %esi
> > - jle L(max)
> > + vpmovmskb %ymm1, %eax
> > + /* If remaining length > VEC_SIZE * 2. This works if esi is off by
> > + VEC_SIZE * 4. */
> > + testl $(VEC_SIZE * 2), %esi
> > + jnz L(last_4x_vec)
> >
> > - VPCMPEQ VEC_SIZE(%rdi), %ymm0, %ymm1
> > - vpmovmskb %ymm1, %eax
> > + /* length may have been negative or positive by an offset of
> > + VEC_SIZE * 4 depending on where this was called from. This fixes
> > + that. */
> > + andl $(VEC_SIZE * 4 - 1), %esi
> > testl %eax, %eax
> > - jnz L(first_vec_x1_check)
> > - movq %r8, %rax
> > -# ifdef USE_AS_WCSLEN
> > - shrq $2, %rax
> > -# endif
> > - VZEROUPPER_RETURN
> > + jnz L(last_vec_x1_check)
> >
> > - .p2align 4
> > -L(first_vec_x0_check):
> > + subl $VEC_SIZE, %esi
> > + jb L(max)
> > +
> > + VPCMPEQ (VEC_SIZE + 1)(%rdi), %ymm0, %ymm1
> > + vpmovmskb %ymm1, %eax
> > tzcntl %eax, %eax
> > /* Check the end of data. */
> > - cmpq %rax, %rsi
> > - jbe L(max)
> > + cmpl %eax, %esi
> > + jb L(max)
> > + subq %rdx, %rdi
> > + addl $(VEC_SIZE + 1), %eax
> > addq %rdi, %rax
> > - subq %rdx, %rax
> > # ifdef USE_AS_WCSLEN
> > shrq $2, %rax
> > # endif
> > VZEROUPPER_RETURN
> > +# endif
> >
> > .p2align 4
> > -L(first_vec_x1_check):
> > +L(last_vec_return_x0):
> > tzcntl %eax, %eax
> > - /* Check the end of data. */
> > - cmpq %rax, %rsi
> > - jbe L(max)
> > - addq $VEC_SIZE, %rax
> > + subq $(VEC_SIZE * 4 - 1), %rdi
> > addq %rdi, %rax
> > - subq %rdx, %rax
> > -# ifdef USE_AS_WCSLEN
> > +# ifdef USE_AS_WCSLEN
> > shrq $2, %rax
> > -# endif
> > +# endif
> > VZEROUPPER_RETURN
> >
> > .p2align 4
> > -L(first_vec_x2_check):
> > +L(last_vec_return_x1):
> > tzcntl %eax, %eax
> > - /* Check the end of data. */
> > - cmpq %rax, %rsi
> > - jbe L(max)
> > - addq $(VEC_SIZE * 2), %rax
> > + subq $(VEC_SIZE * 3 - 1), %rdi
> > addq %rdi, %rax
> > - subq %rdx, %rax
> > -# ifdef USE_AS_WCSLEN
> > +# ifdef USE_AS_WCSLEN
> > shrq $2, %rax
> > -# endif
> > +# endif
> > VZEROUPPER_RETURN
> >
> > +# ifdef USE_AS_STRNLEN
> > .p2align 4
> > -L(first_vec_x3_check):
> > +L(last_vec_x1_check):
> > +
> > tzcntl %eax, %eax
> > /* Check the end of data. */
> > - cmpq %rax, %rsi
> > - jbe L(max)
> > - addq $(VEC_SIZE * 3), %rax
> > + cmpl %eax, %esi
> > + jb L(max)
> > + subq %rdx, %rdi
> > + incl %eax
> > addq %rdi, %rax
> > - subq %rdx, %rax
> > # ifdef USE_AS_WCSLEN
> > shrq $2, %rax
> > # endif
> > VZEROUPPER_RETURN
> >
> > - .p2align 4
> > L(max):
> > movq %r8, %rax
> > + VZEROUPPER_RETURN
> > +
> > + .p2align 4
> > +L(last_4x_vec):
> > + /* Test first 2x VEC normally. */
> > + testl %eax, %eax
> > + jnz L(last_vec_x1)
> > +
> > + VPCMPEQ (VEC_SIZE + 1)(%rdi), %ymm0, %ymm1
> > + vpmovmskb %ymm1, %eax
> > + testl %eax, %eax
> > + jnz L(last_vec_x2)
> > +
> > + /* Normalize length. */
> > + andl $(VEC_SIZE * 4 - 1), %esi
> > + VPCMPEQ (VEC_SIZE * 2 + 1)(%rdi), %ymm0, %ymm1
> > + vpmovmskb %ymm1, %eax
> > + testl %eax, %eax
> > + jnz L(last_vec_x3)
> > +
> > + subl $(VEC_SIZE * 3), %esi
> > + jb L(max)
> > +
> > + VPCMPEQ (VEC_SIZE * 3 + 1)(%rdi), %ymm0, %ymm1
> > + vpmovmskb %ymm1, %eax
> > + tzcntl %eax, %eax
> > + /* Check the end of data. */
> > + cmpl %eax, %esi
> > + jb L(max)
> > + subq %rdx, %rdi
> > + addl $(VEC_SIZE * 3 + 1), %eax
> > + addq %rdi, %rax
> > # ifdef USE_AS_WCSLEN
> > shrq $2, %rax
> > # endif
> > VZEROUPPER_RETURN
> >
> > - .p2align 4
> > -L(zero):
> > - xorl %eax, %eax
> > - ret
> > -# endif
> >
> > .p2align 4
> > -L(first_vec_x0):
> > +L(last_vec_x1):
> > + /* essentially duplicates of first_vec_x1 but use 64 bit
> > + instructions. */
> > tzcntl %eax, %eax
> > + subq %rdx, %rdi
> > + incl %eax
> > addq %rdi, %rax
> > - subq %rdx, %rax
> > -# ifdef USE_AS_WCSLEN
> > +# ifdef USE_AS_WCSLEN
> > shrq $2, %rax
> > -# endif
> > +# endif
> > VZEROUPPER_RETURN
> >
> > .p2align 4
> > -L(first_vec_x1):
> > +L(last_vec_x2):
> > + /* essentially duplicates of first_vec_x1 but use 64 bit
> > + instructions. */
> > tzcntl %eax, %eax
> > - addq $VEC_SIZE, %rax
> > + subq %rdx, %rdi
> > + addl $(VEC_SIZE + 1), %eax
> > addq %rdi, %rax
> > - subq %rdx, %rax
> > -# ifdef USE_AS_WCSLEN
> > +# ifdef USE_AS_WCSLEN
> > shrq $2, %rax
> > -# endif
> > +# endif
> > VZEROUPPER_RETURN
> >
> > .p2align 4
> > -L(first_vec_x2):
> > +L(last_vec_x3):
> > tzcntl %eax, %eax
> > - addq $(VEC_SIZE * 2), %rax
> > + subl $(VEC_SIZE * 2), %esi
> > + /* Check the end of data. */
> > + cmpl %eax, %esi
> > + jb L(max_end)
> > + subq %rdx, %rdi
> > + addl $(VEC_SIZE * 2 + 1), %eax
> > addq %rdi, %rax
> > - subq %rdx, %rax
> > -# ifdef USE_AS_WCSLEN
> > +# ifdef USE_AS_WCSLEN
> > shrq $2, %rax
> > -# endif
> > +# endif
> > + VZEROUPPER_RETURN
> > +L(max_end):
> > + movq %r8, %rax
> > VZEROUPPER_RETURN
> > +# endif
> >
> > + /* Cold case for crossing page with first load. */
> > .p2align 4
> > -L(4x_vec_end):
> > - VPCMPEQ %ymm1, %ymm0, %ymm1
> > - vpmovmskb %ymm1, %eax
> > - testl %eax, %eax
> > - jnz L(first_vec_x0)
> > - VPCMPEQ %ymm2, %ymm0, %ymm2
> > - vpmovmskb %ymm2, %eax
> > +L(cross_page_boundary):
> > + /* Align data to VEC_SIZE - 1. */
> > + orq $(VEC_SIZE - 1), %rdi
> > + VPCMPEQ -(VEC_SIZE - 1)(%rdi), %ymm0, %ymm1
> > + vpmovmskb %ymm1, %eax
> > + /* Remove the leading bytes. sarxl only uses bits [5:0] of COUNT
> > + so no need to manually mod rdx. */
> > + sarxl %edx, %eax, %eax
> > +# ifdef USE_AS_STRNLEN
> > testl %eax, %eax
> > - jnz L(first_vec_x1)
> > - VPCMPEQ %ymm3, %ymm0, %ymm3
> > - vpmovmskb %ymm3, %eax
> > + jnz L(cross_page_less_vec)
> > + leaq 1(%rdi), %rcx
> > + subq %rdx, %rcx
> > + /* Check length. */
> > + cmpq %rsi, %rcx
> > + jb L(cross_page_continue)
> > + movq %r8, %rax
> > +# else
> > testl %eax, %eax
> > - jnz L(first_vec_x2)
> > - VPCMPEQ %ymm4, %ymm0, %ymm4
> > - vpmovmskb %ymm4, %eax
> > -L(first_vec_x3):
> > + jz L(cross_page_continue)
> > tzcntl %eax, %eax
> > - addq $(VEC_SIZE * 3), %rax
> > - addq %rdi, %rax
> > - subq %rdx, %rax
> > -# ifdef USE_AS_WCSLEN
> > - shrq $2, %rax
> > +# ifdef USE_AS_WCSLEN
> > + shrl $2, %eax
> > +# endif
> > # endif
> > +L(return_vzeroupper):
> > + ZERO_UPPER_VEC_REGISTERS_RETURN
> > +
> > +# ifdef USE_AS_STRNLEN
> > + .p2align 4
> > +L(cross_page_less_vec):
> > + tzcntl %eax, %eax
> > + cmpq %rax, %rsi
> > + cmovb %esi, %eax
> > +# ifdef USE_AS_WCSLEN
> > + shrl $2, %eax
> > +# endif
> > VZEROUPPER_RETURN
> > +# endif
> >
> > END (STRLEN)
> > #endif
> > --
> > 2.29.2
> >
^ permalink raw reply [flat|nested] 6+ messages in thread
end of thread, other threads:[~2021-04-19 23:39 UTC | newest]
Thread overview: 6+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2021-04-19 18:43 [PATCH v4 1/2] x86: Optimize strlen-evex.S Noah Goldstein
2021-04-19 18:43 ` [PATCH v4 2/2] x86: Optimize strlen-avx2.S Noah Goldstein
2021-04-19 22:16 ` H.J. Lu
2021-04-19 23:39 ` Noah Goldstein
2021-04-19 22:16 ` [PATCH v4 1/2] x86: Optimize strlen-evex.S H.J. Lu
2021-04-19 23:38 ` Noah Goldstein
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).