x86: Optimize memchr-evex.S

No bug. This commit optimizes memchr-evex.S. The optimizations include
replacing some branches with cmovcc, avoiding some branches entirely
in the less_4x_vec case, making the page cross logic less strict,
saving some ALU in the alignment process, and most importantly
increasing ILP in the 4x loop. test-memchr, test-rawmemchr, and
test-wmemchr are all passing.

Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com>
Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
(cherry picked from commit 2a76821c30)
This commit is contained in:
Noah Goldstein 2021-05-03 03:03:19 -04:00 committed by H.J. Lu
parent 86250b81da
commit 68e438825d

View File

@ -26,14 +26,28 @@
# ifdef USE_AS_WMEMCHR # ifdef USE_AS_WMEMCHR
# define VPBROADCAST vpbroadcastd # define VPBROADCAST vpbroadcastd
# define VPCMP vpcmpd # define VPMINU vpminud
# define SHIFT_REG r8d # define VPCMP vpcmpd
# define VPCMPEQ vpcmpeqd
# define CHAR_SIZE 4
# else # else
# define VPBROADCAST vpbroadcastb # define VPBROADCAST vpbroadcastb
# define VPCMP vpcmpb # define VPMINU vpminub
# define SHIFT_REG ecx # define VPCMP vpcmpb
# define VPCMPEQ vpcmpeqb
# define CHAR_SIZE 1
# endif # endif
# ifdef USE_AS_RAWMEMCHR
# define RAW_PTR_REG rcx
# define ALGN_PTR_REG rdi
# else
# define RAW_PTR_REG rdi
# define ALGN_PTR_REG rcx
# endif
# define XMMZERO xmm23
# define YMMZERO ymm23
# define XMMMATCH xmm16 # define XMMMATCH xmm16
# define YMMMATCH ymm16 # define YMMMATCH ymm16
# define YMM1 ymm17 # define YMM1 ymm17
@ -44,6 +58,8 @@
# define YMM6 ymm22 # define YMM6 ymm22
# define VEC_SIZE 32 # define VEC_SIZE 32
# define CHAR_PER_VEC (VEC_SIZE / CHAR_SIZE)
# define PAGE_SIZE 4096
.section .text.evex,"ax",@progbits .section .text.evex,"ax",@progbits
ENTRY (MEMCHR) ENTRY (MEMCHR)
@ -51,11 +67,7 @@ ENTRY (MEMCHR)
/* Check for zero length. */ /* Check for zero length. */
test %RDX_LP, %RDX_LP test %RDX_LP, %RDX_LP
jz L(zero) jz L(zero)
# endif
movl %edi, %ecx
# ifdef USE_AS_WMEMCHR
shl $2, %RDX_LP
# else
# ifdef __ILP32__ # ifdef __ILP32__
/* Clear the upper 32 bits. */ /* Clear the upper 32 bits. */
movl %edx, %edx movl %edx, %edx
@ -64,318 +76,403 @@ ENTRY (MEMCHR)
/* Broadcast CHAR to YMMMATCH. */ /* Broadcast CHAR to YMMMATCH. */
VPBROADCAST %esi, %YMMMATCH VPBROADCAST %esi, %YMMMATCH
/* Check if we may cross page boundary with one vector load. */ /* Check if we may cross page boundary with one vector load. */
andl $(2 * VEC_SIZE - 1), %ecx movl %edi, %eax
cmpl $VEC_SIZE, %ecx andl $(PAGE_SIZE - 1), %eax
ja L(cros_page_boundary) cmpl $(PAGE_SIZE - VEC_SIZE), %eax
ja L(cross_page_boundary)
/* Check the first VEC_SIZE bytes. */ /* Check the first VEC_SIZE bytes. */
VPCMP $0, (%rdi), %YMMMATCH, %k1 VPCMP $0, (%rdi), %YMMMATCH, %k0
kmovd %k1, %eax kmovd %k0, %eax
testl %eax, %eax
# ifndef USE_AS_RAWMEMCHR # ifndef USE_AS_RAWMEMCHR
jnz L(first_vec_x0_check) /* If length < CHAR_PER_VEC handle special. */
/* Adjust length and check the end of data. */ cmpq $CHAR_PER_VEC, %rdx
subq $VEC_SIZE, %rdx jbe L(first_vec_x0)
jbe L(zero)
# else
jnz L(first_vec_x0)
# endif # endif
/* Align data for aligned loads in the loop. */
addq $VEC_SIZE, %rdi
andl $(VEC_SIZE - 1), %ecx
andq $-VEC_SIZE, %rdi
# ifndef USE_AS_RAWMEMCHR
/* Adjust length. */
addq %rcx, %rdx
subq $(VEC_SIZE * 4), %rdx
jbe L(last_4x_vec_or_less)
# endif
jmp L(more_4x_vec)
.p2align 4
L(cros_page_boundary):
andl $(VEC_SIZE - 1), %ecx
# ifdef USE_AS_WMEMCHR
/* NB: Divide shift count by 4 since each bit in K1 represent 4
bytes. */
movl %ecx, %SHIFT_REG
sarl $2, %SHIFT_REG
# endif
andq $-VEC_SIZE, %rdi
VPCMP $0, (%rdi), %YMMMATCH, %k1
kmovd %k1, %eax
/* Remove the leading bytes. */
sarxl %SHIFT_REG, %eax, %eax
testl %eax, %eax testl %eax, %eax
jz L(aligned_more) jz L(aligned_more)
tzcntl %eax, %eax tzcntl %eax, %eax
# ifdef USE_AS_WMEMCHR # ifdef USE_AS_WMEMCHR
/* NB: Multiply wchar_t count by 4 to get the number of bytes. */ /* NB: Multiply bytes by CHAR_SIZE to get the wchar_t count. */
sall $2, %eax leaq (%rdi, %rax, CHAR_SIZE), %rax
# endif
# ifndef USE_AS_RAWMEMCHR
/* Check the end of data. */
cmpq %rax, %rdx
jbe L(zero)
# endif
addq %rdi, %rax
addq %rcx, %rax
ret
.p2align 4
L(aligned_more):
# ifndef USE_AS_RAWMEMCHR
/* 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, %rdx
jbe L(zero)
# endif
addq $VEC_SIZE, %rdi
# ifndef USE_AS_RAWMEMCHR
subq $(VEC_SIZE * 4), %rdx
jbe L(last_4x_vec_or_less)
# endif
L(more_4x_vec):
/* Check the first 4 * VEC_SIZE. Only one VEC_SIZE at a time
since data is only aligned to VEC_SIZE. */
VPCMP $0, (%rdi), %YMMMATCH, %k1
kmovd %k1, %eax
testl %eax, %eax
jnz L(first_vec_x0)
VPCMP $0, VEC_SIZE(%rdi), %YMMMATCH, %k1
kmovd %k1, %eax
testl %eax, %eax
jnz L(first_vec_x1)
VPCMP $0, (VEC_SIZE * 2)(%rdi), %YMMMATCH, %k1
kmovd %k1, %eax
testl %eax, %eax
jnz L(first_vec_x2)
VPCMP $0, (VEC_SIZE * 3)(%rdi), %YMMMATCH, %k1
kmovd %k1, %eax
testl %eax, %eax
jnz L(first_vec_x3)
addq $(VEC_SIZE * 4), %rdi
# ifndef USE_AS_RAWMEMCHR
subq $(VEC_SIZE * 4), %rdx
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
# ifndef USE_AS_RAWMEMCHR
/* Adjust length. */
addq %rcx, %rdx
# endif
.p2align 4
L(loop_4x_vec):
/* Compare 4 * VEC at a time forward. */
VPCMP $0, (%rdi), %YMMMATCH, %k1
VPCMP $0, VEC_SIZE(%rdi), %YMMMATCH, %k2
kord %k1, %k2, %k5
VPCMP $0, (VEC_SIZE * 2)(%rdi), %YMMMATCH, %k3
VPCMP $0, (VEC_SIZE * 3)(%rdi), %YMMMATCH, %k4
kord %k3, %k4, %k6
kortestd %k5, %k6
jnz L(4x_vec_end)
addq $(VEC_SIZE * 4), %rdi
# ifdef USE_AS_RAWMEMCHR
jmp L(loop_4x_vec)
# else # else
subq $(VEC_SIZE * 4), %rdx
ja L(loop_4x_vec)
L(last_4x_vec_or_less):
/* Less than 4 * VEC and aligned to VEC_SIZE. */
addl $(VEC_SIZE * 2), %edx
jle L(last_2x_vec)
VPCMP $0, (%rdi), %YMMMATCH, %k1
kmovd %k1, %eax
testl %eax, %eax
jnz L(first_vec_x0)
VPCMP $0, VEC_SIZE(%rdi), %YMMMATCH, %k1
kmovd %k1, %eax
testl %eax, %eax
jnz L(first_vec_x1)
VPCMP $0, (VEC_SIZE * 2)(%rdi), %YMMMATCH, %k1
kmovd %k1, %eax
testl %eax, %eax
jnz L(first_vec_x2_check)
subl $VEC_SIZE, %edx
jle L(zero)
VPCMP $0, (VEC_SIZE * 3)(%rdi), %YMMMATCH, %k1
kmovd %k1, %eax
testl %eax, %eax
jnz L(first_vec_x3_check)
xorl %eax, %eax
ret
.p2align 4
L(last_2x_vec):
addl $(VEC_SIZE * 2), %edx
VPCMP $0, (%rdi), %YMMMATCH, %k1
kmovd %k1, %eax
testl %eax, %eax
jnz L(first_vec_x0_check)
subl $VEC_SIZE, %edx
jle L(zero)
VPCMP $0, VEC_SIZE(%rdi), %YMMMATCH, %k1
kmovd %k1, %eax
testl %eax, %eax
jnz L(first_vec_x1_check)
xorl %eax, %eax
ret
.p2align 4
L(first_vec_x0_check):
tzcntl %eax, %eax
# ifdef USE_AS_WMEMCHR
/* NB: Multiply wchar_t count by 4 to get the number of bytes. */
sall $2, %eax
# endif
/* Check the end of data. */
cmpq %rax, %rdx
jbe L(zero)
addq %rdi, %rax addq %rdi, %rax
ret
.p2align 4
L(first_vec_x1_check):
tzcntl %eax, %eax
# ifdef USE_AS_WMEMCHR
/* NB: Multiply wchar_t count by 4 to get the number of bytes. */
sall $2, %eax
# endif # endif
/* Check the end of data. */
cmpq %rax, %rdx
jbe L(zero)
addq $VEC_SIZE, %rax
addq %rdi, %rax
ret ret
.p2align 4 # ifndef USE_AS_RAWMEMCHR
L(first_vec_x2_check):
tzcntl %eax, %eax
# ifdef USE_AS_WMEMCHR
/* NB: Multiply wchar_t count by 4 to get the number of bytes. */
sall $2, %eax
# endif
/* Check the end of data. */
cmpq %rax, %rdx
jbe L(zero)
addq $(VEC_SIZE * 2), %rax
addq %rdi, %rax
ret
.p2align 4
L(first_vec_x3_check):
tzcntl %eax, %eax
# ifdef USE_AS_WMEMCHR
/* NB: Multiply wchar_t count by 4 to get the number of bytes. */
sall $2, %eax
# endif
/* Check the end of data. */
cmpq %rax, %rdx
jbe L(zero)
addq $(VEC_SIZE * 3), %rax
addq %rdi, %rax
ret
.p2align 4
L(zero): L(zero):
xorl %eax, %eax xorl %eax, %eax
ret ret
# endif
.p2align 4 .p2align 5
L(first_vec_x0): L(first_vec_x0):
/* Check if first match was before length. */
tzcntl %eax, %eax
xorl %ecx, %ecx
cmpl %eax, %edx
leaq (%rdi, %rax, CHAR_SIZE), %rax
cmovle %rcx, %rax
ret
# else
/* NB: first_vec_x0 is 17 bytes which will leave
cross_page_boundary (which is relatively cold) close enough
to ideal alignment. So only realign L(cross_page_boundary) if
rawmemchr. */
.p2align 4
# endif
L(cross_page_boundary):
/* Save pointer before aligning as its original value is
necessary for computer return address if byte is found or
adjusting length if it is not and this is memchr. */
movq %rdi, %rcx
/* Align data to VEC_SIZE. ALGN_PTR_REG is rcx for memchr and rdi
for rawmemchr. */
andq $-VEC_SIZE, %ALGN_PTR_REG
VPCMP $0, (%ALGN_PTR_REG), %YMMMATCH, %k0
kmovd %k0, %r8d
# ifdef USE_AS_WMEMCHR
/* NB: Divide shift count by 4 since each bit in K0 represent 4
bytes. */
sarl $2, %eax
# endif
# ifndef USE_AS_RAWMEMCHR
movl $(PAGE_SIZE / CHAR_SIZE), %esi
subl %eax, %esi
# endif
# ifdef USE_AS_WMEMCHR
andl $(CHAR_PER_VEC - 1), %eax
# endif
/* Remove the leading bytes. */
sarxl %eax, %r8d, %eax
# ifndef USE_AS_RAWMEMCHR
/* Check the end of data. */
cmpq %rsi, %rdx
jbe L(first_vec_x0)
# endif
testl %eax, %eax
jz L(cross_page_continue)
tzcntl %eax, %eax tzcntl %eax, %eax
# ifdef USE_AS_WMEMCHR # ifdef USE_AS_WMEMCHR
/* NB: Multiply wchar_t count by 4 to get the number of bytes. */ /* NB: Multiply bytes by CHAR_SIZE to get the wchar_t count. */
leaq (%rdi, %rax, 4), %rax leaq (%RAW_PTR_REG, %rax, CHAR_SIZE), %rax
# else # else
addq %rdi, %rax addq %RAW_PTR_REG, %rax
# endif # endif
ret ret
.p2align 4 .p2align 4
L(first_vec_x1): L(first_vec_x1):
tzcntl %eax, %eax tzcntl %eax, %eax
# ifdef USE_AS_WMEMCHR leaq VEC_SIZE(%rdi, %rax, CHAR_SIZE), %rax
/* NB: Multiply wchar_t count by 4 to get the number of bytes. */
leaq VEC_SIZE(%rdi, %rax, 4), %rax
# else
addq $VEC_SIZE, %rax
addq %rdi, %rax
# endif
ret ret
.p2align 4 .p2align 4
L(first_vec_x2): L(first_vec_x2):
tzcntl %eax, %eax tzcntl %eax, %eax
# ifdef USE_AS_WMEMCHR leaq (VEC_SIZE * 2)(%rdi, %rax, CHAR_SIZE), %rax
/* NB: Multiply wchar_t count by 4 to get the number of bytes. */ ret
leaq (VEC_SIZE * 2)(%rdi, %rax, 4), %rax
.p2align 4
L(first_vec_x3):
tzcntl %eax, %eax
leaq (VEC_SIZE * 3)(%rdi, %rax, CHAR_SIZE), %rax
ret
.p2align 4
L(first_vec_x4):
tzcntl %eax, %eax
leaq (VEC_SIZE * 4)(%rdi, %rax, CHAR_SIZE), %rax
ret
.p2align 5
L(aligned_more):
/* Check the first 4 * VEC_SIZE. Only one VEC_SIZE at a time
since data is only aligned to VEC_SIZE. */
# ifndef USE_AS_RAWMEMCHR
/* Align data to VEC_SIZE. */
L(cross_page_continue):
xorl %ecx, %ecx
subl %edi, %ecx
andq $-VEC_SIZE, %rdi
/* esi is for adjusting length to see if near the end. */
leal (VEC_SIZE * 5)(%rdi, %rcx), %esi
# ifdef USE_AS_WMEMCHR
/* NB: Divide bytes by 4 to get the wchar_t count. */
sarl $2, %esi
# endif
# else # else
addq $(VEC_SIZE * 2), %rax andq $-VEC_SIZE, %rdi
addq %rdi, %rax L(cross_page_continue):
# endif
/* Load first VEC regardless. */
VPCMP $0, (VEC_SIZE)(%rdi), %YMMMATCH, %k0
kmovd %k0, %eax
# ifndef USE_AS_RAWMEMCHR
/* Adjust length. If near end handle specially. */
subq %rsi, %rdx
jbe L(last_4x_vec_or_less)
# endif
testl %eax, %eax
jnz L(first_vec_x1)
VPCMP $0, (VEC_SIZE * 2)(%rdi), %YMMMATCH, %k0
kmovd %k0, %eax
testl %eax, %eax
jnz L(first_vec_x2)
VPCMP $0, (VEC_SIZE * 3)(%rdi), %YMMMATCH, %k0
kmovd %k0, %eax
testl %eax, %eax
jnz L(first_vec_x3)
VPCMP $0, (VEC_SIZE * 4)(%rdi), %YMMMATCH, %k0
kmovd %k0, %eax
testl %eax, %eax
jnz L(first_vec_x4)
# ifndef USE_AS_RAWMEMCHR
/* Check if at last CHAR_PER_VEC * 4 length. */
subq $(CHAR_PER_VEC * 4), %rdx
jbe L(last_4x_vec_or_less_cmpeq)
addq $VEC_SIZE, %rdi
/* Align data to VEC_SIZE * 4 for the loop and readjust length.
*/
# ifdef USE_AS_WMEMCHR
movl %edi, %ecx
andq $-(4 * VEC_SIZE), %rdi
andl $(VEC_SIZE * 4 - 1), %ecx
/* NB: Divide bytes by 4 to get the wchar_t count. */
sarl $2, %ecx
addq %rcx, %rdx
# else
addq %rdi, %rdx
andq $-(4 * VEC_SIZE), %rdi
subq %rdi, %rdx
# endif
# else
addq $VEC_SIZE, %rdi
andq $-(4 * VEC_SIZE), %rdi
# endif
vpxorq %XMMZERO, %XMMZERO, %XMMZERO
/* Compare 4 * VEC at a time forward. */
.p2align 4
L(loop_4x_vec):
/* It would be possible to save some instructions using 4x VPCMP
but bottleneck on port 5 makes it not woth it. */
VPCMP $4, (VEC_SIZE * 4)(%rdi), %YMMMATCH, %k1
/* xor will set bytes match esi to zero. */
vpxorq (VEC_SIZE * 5)(%rdi), %YMMMATCH, %YMM2
vpxorq (VEC_SIZE * 6)(%rdi), %YMMMATCH, %YMM3
VPCMP $0, (VEC_SIZE * 7)(%rdi), %YMMMATCH, %k3
/* Reduce VEC2 / VEC3 with min and VEC1 with zero mask. */
VPMINU %YMM2, %YMM3, %YMM3 {%k1} {z}
VPCMP $0, %YMM3, %YMMZERO, %k2
# ifdef USE_AS_RAWMEMCHR
subq $-(VEC_SIZE * 4), %rdi
kortestd %k2, %k3
jz L(loop_4x_vec)
# else
kortestd %k2, %k3
jnz L(loop_4x_vec_end)
subq $-(VEC_SIZE * 4), %rdi
subq $(CHAR_PER_VEC * 4), %rdx
ja L(loop_4x_vec)
/* Fall through into less than 4 remaining vectors of length case.
*/
VPCMP $0, (VEC_SIZE * 4)(%rdi), %YMMMATCH, %k0
kmovd %k0, %eax
addq $(VEC_SIZE * 3), %rdi
.p2align 4
L(last_4x_vec_or_less):
/* Check if first VEC contained match. */
testl %eax, %eax
jnz L(first_vec_x1_check)
/* If remaining length > CHAR_PER_VEC * 2. */
addl $(CHAR_PER_VEC * 2), %edx
jg L(last_4x_vec)
L(last_2x_vec):
/* If remaining length < CHAR_PER_VEC. */
addl $CHAR_PER_VEC, %edx
jle L(zero_end)
/* Check VEC2 and compare any match with remaining length. */
VPCMP $0, (VEC_SIZE * 2)(%rdi), %YMMMATCH, %k0
kmovd %k0, %eax
tzcntl %eax, %eax
cmpl %eax, %edx
jbe L(set_zero_end)
leaq (VEC_SIZE * 2)(%rdi, %rax, CHAR_SIZE), %rax
L(zero_end):
ret
.p2align 4
L(first_vec_x1_check):
tzcntl %eax, %eax
/* Adjust length. */
subl $-(CHAR_PER_VEC * 4), %edx
/* Check if match within remaining length. */
cmpl %eax, %edx
jbe L(set_zero_end)
/* NB: Multiply bytes by CHAR_SIZE to get the wchar_t count. */
leaq VEC_SIZE(%rdi, %rax, CHAR_SIZE), %rax
ret
L(set_zero_end):
xorl %eax, %eax
ret
.p2align 4
L(loop_4x_vec_end):
# endif
/* rawmemchr will fall through into this if match was found in
loop. */
/* k1 has not of matches with VEC1. */
kmovd %k1, %eax
# ifdef USE_AS_WMEMCHR
subl $((1 << CHAR_PER_VEC) - 1), %eax
# else
incl %eax
# endif
jnz L(last_vec_x1_return)
VPCMP $0, %YMM2, %YMMZERO, %k0
kmovd %k0, %eax
testl %eax, %eax
jnz L(last_vec_x2_return)
kmovd %k2, %eax
testl %eax, %eax
jnz L(last_vec_x3_return)
kmovd %k3, %eax
tzcntl %eax, %eax
# ifdef USE_AS_RAWMEMCHR
leaq (VEC_SIZE * 3)(%rdi, %rax, CHAR_SIZE), %rax
# else
leaq (VEC_SIZE * 7)(%rdi, %rax, CHAR_SIZE), %rax
# endif # endif
ret ret
.p2align 4 .p2align 4
L(4x_vec_end): L(last_vec_x1_return):
kmovd %k1, %eax
testl %eax, %eax
jnz L(first_vec_x0)
kmovd %k2, %eax
testl %eax, %eax
jnz L(first_vec_x1)
kmovd %k3, %eax
testl %eax, %eax
jnz L(first_vec_x2)
kmovd %k4, %eax
testl %eax, %eax
L(first_vec_x3):
tzcntl %eax, %eax tzcntl %eax, %eax
# ifdef USE_AS_WMEMCHR # ifdef USE_AS_RAWMEMCHR
/* NB: Multiply wchar_t count by 4 to get the number of bytes. */ # ifdef USE_AS_WMEMCHR
leaq (VEC_SIZE * 3)(%rdi, %rax, 4), %rax /* NB: Multiply bytes by CHAR_SIZE to get the wchar_t count. */
# else leaq (%rdi, %rax, CHAR_SIZE), %rax
addq $(VEC_SIZE * 3), %rax # else
addq %rdi, %rax addq %rdi, %rax
# endif
# else
/* NB: Multiply bytes by CHAR_SIZE to get the wchar_t count. */
leaq (VEC_SIZE * 4)(%rdi, %rax, CHAR_SIZE), %rax
# endif # endif
ret ret
.p2align 4
L(last_vec_x2_return):
tzcntl %eax, %eax
# ifdef USE_AS_RAWMEMCHR
/* NB: Multiply bytes by CHAR_SIZE to get the wchar_t count. */
leaq VEC_SIZE(%rdi, %rax, CHAR_SIZE), %rax
# else
/* NB: Multiply bytes by CHAR_SIZE to get the wchar_t count. */
leaq (VEC_SIZE * 5)(%rdi, %rax, CHAR_SIZE), %rax
# endif
ret
.p2align 4
L(last_vec_x3_return):
tzcntl %eax, %eax
# ifdef USE_AS_RAWMEMCHR
/* NB: Multiply bytes by CHAR_SIZE to get the wchar_t count. */
leaq (VEC_SIZE * 2)(%rdi, %rax, CHAR_SIZE), %rax
# else
/* NB: Multiply bytes by CHAR_SIZE to get the wchar_t count. */
leaq (VEC_SIZE * 6)(%rdi, %rax, CHAR_SIZE), %rax
# endif
ret
# ifndef USE_AS_RAWMEMCHR
L(last_4x_vec_or_less_cmpeq):
VPCMP $0, (VEC_SIZE * 5)(%rdi), %YMMMATCH, %k0
kmovd %k0, %eax
subq $-(VEC_SIZE * 4), %rdi
/* Check first VEC regardless. */
testl %eax, %eax
jnz L(first_vec_x1_check)
/* If remaining length <= CHAR_PER_VEC * 2. */
addl $(CHAR_PER_VEC * 2), %edx
jle L(last_2x_vec)
.p2align 4
L(last_4x_vec):
VPCMP $0, (VEC_SIZE * 2)(%rdi), %YMMMATCH, %k0
kmovd %k0, %eax
testl %eax, %eax
jnz L(last_vec_x2)
VPCMP $0, (VEC_SIZE * 3)(%rdi), %YMMMATCH, %k0
kmovd %k0, %eax
/* Create mask for possible matches within remaining length. */
# ifdef USE_AS_WMEMCHR
movl $((1 << (CHAR_PER_VEC * 2)) - 1), %ecx
bzhil %edx, %ecx, %ecx
# else
movq $-1, %rcx
bzhiq %rdx, %rcx, %rcx
# endif
/* Test matches in data against length match. */
andl %ecx, %eax
jnz L(last_vec_x3)
/* if remaining length <= CHAR_PER_VEC * 3 (Note this is after
remaining length was found to be > CHAR_PER_VEC * 2. */
subl $CHAR_PER_VEC, %edx
jbe L(zero_end2)
VPCMP $0, (VEC_SIZE * 4)(%rdi), %YMMMATCH, %k0
kmovd %k0, %eax
/* Shift remaining length mask for last VEC. */
# ifdef USE_AS_WMEMCHR
shrl $CHAR_PER_VEC, %ecx
# else
shrq $CHAR_PER_VEC, %rcx
# endif
andl %ecx, %eax
jz L(zero_end2)
tzcntl %eax, %eax
leaq (VEC_SIZE * 4)(%rdi, %rax, CHAR_SIZE), %rax
L(zero_end2):
ret
L(last_vec_x2):
tzcntl %eax, %eax
leaq (VEC_SIZE * 2)(%rdi, %rax, CHAR_SIZE), %rax
ret
.p2align 4
L(last_vec_x3):
tzcntl %eax, %eax
leaq (VEC_SIZE * 3)(%rdi, %rax, CHAR_SIZE), %rax
ret
# endif
END (MEMCHR) END (MEMCHR)
#endif #endif