Re: [PATCH v4 1/2] x86: Optimize strlen-evex.S

["H.J. Lu" <hjl.tools@gmail.com>]

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
>