path: root/sysdeps/x86_64/multiarch/memcmp-evex-movbe.S

/* memcmp/wmemcmp optimized with 256-bit EVEX instructions.
   Copyright (C) 2021-2024 Free Software Foundation, Inc.
   This file is part of the GNU C Library.

   The GNU C Library is free software; you can redistribute it and/or
   modify it under the terms of the GNU Lesser General Public
   License as published by the Free Software Foundation; either
   version 2.1 of the License, or (at your option) any later version.

   The GNU C Library is distributed in the hope that it will be useful,
   but WITHOUT ANY WARRANTY; without even the implied warranty of
   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
   Lesser General Public License for more details.

   You should have received a copy of the GNU Lesser General Public
   License along with the GNU C Library; if not, see
   <https://www.gnu.org/licenses/>.  */

#include <isa-level.h>

#if ISA_SHOULD_BUILD (4)


/* memcmp/wmemcmp is implemented as:
   1. Use ymm vector compares when possible. The only case where
      vector compares is not possible for when size < CHAR_PER_VEC
      and loading from either s1 or s2 would cause a page cross.
   2. For size from 2 to 7 bytes on page cross, load as big endian
      with movbe and bswap to avoid branches.
   3. Use xmm vector compare when size >= 4 bytes for memcmp or
      size >= 8 bytes for wmemcmp.
   4. Optimistically compare up to first 4 * CHAR_PER_VEC one at a
      to check for early mismatches. Only do this if its guaranteed the
      work is not wasted.
   5. If size is 8 * VEC_SIZE or less, unroll the loop.
   6. Compare 4 * VEC_SIZE at a time with the aligned first memory
      area.
   7. Use 2 vector compares when size is 2 * CHAR_PER_VEC or less.
   8. Use 4 vector compares when size is 4 * CHAR_PER_VEC or less.
   9. Use 8 vector compares when size is 8 * CHAR_PER_VEC or less.

When possible the implementation tries to optimize for frontend in the
following ways:
Throughput:
    1. All code sections that fit are able to run optimally out of the
       LSD.
    2. All code sections that fit are able to run optimally out of the
       DSB
    3. Basic blocks are contained in minimum number of fetch blocks
       necessary.

Latency:
    1. Logically connected basic blocks are put in the same
       cache-line.
    2. Logically connected basic blocks that do not fit in the same
       cache-line are put in adjacent lines. This can get beneficial
       L2 spatial prefetching and L1 next-line prefetching.  */

# include <sysdep.h>

# ifndef MEMCMP
#  define MEMCMP	__memcmp_evex_movbe
# endif

# ifndef VEC_SIZE
#  include "x86-evex256-vecs.h"
# endif

# ifdef USE_AS_WMEMCMP
#  define VMOVU_MASK	vmovdqu32
#  define CHAR_SIZE	4
#  define VPCMP	vpcmpd
#  define VPCMPEQ	vpcmpeqd
#  define VPTEST	vptestmd

#  define USE_WIDE_CHAR
# else
#  define VMOVU_MASK	vmovdqu8
#  define CHAR_SIZE	1
#  define VPCMP	vpcmpub
#  define VPCMPEQ	vpcmpeqb
#  define VPTEST	vptestmb
# endif

# include "reg-macros.h"

# define PAGE_SIZE	4096
# define CHAR_PER_VEC	(VEC_SIZE / CHAR_SIZE)


/* Warning!
           wmemcmp has to use SIGNED comparison for elements.
           memcmp has to use UNSIGNED comparison for elements.
*/

	.section SECTION(.text), "ax", @progbits
/* Cache align memcmp entry. This allows for much more thorough
   frontend optimization.  */
ENTRY_P2ALIGN (MEMCMP, 6)
# ifdef __ILP32__
	/* Clear the upper 32 bits.  */
	movl	%edx, %edx
# endif
	cmp	$CHAR_PER_VEC, %RDX_LP
	/* Fall through for [0, VEC_SIZE] as its the hottest.  */
	ja	L(more_1x_vec)

	/* Create mask of bytes that are guaranteed to be valid because
	   of length (edx). Using masked movs allows us to skip checks
	   for page crosses/zero size.  */
	mov	$-1, %VRAX
	bzhi	%VRDX, %VRAX, %VRAX
	/* NB: A `jz` might be useful here. Page-faults that are
	   invalidated by predicate execution (the evex mask) can be
	   very slow.  The expectation is this is not the norm so and
	   "most" code will not regularly call 'memcmp' with length = 0
	   and memory that is not wired up.  */
	KMOV	%VRAX, %k2



	/* Safe to load full ymm with mask.  */
	VMOVU_MASK (%rsi), %VMM(2){%k2}{z}
	/* Slightly different method for VEC_SIZE == 64 to save a bit of
	   code size. This allows us to fit L(return_vec_0) entirely in
	   the first cache line.  */
# if VEC_SIZE == 64
	VPCMPEQ	(%rdi), %VMM(2), %k1{%k2}
	KMOV	%k1, %VRCX
	sub	%VRCX, %VRAX
# else
	VPCMP	$4, (%rdi), %VMM(2), %k1{%k2}
	KMOV	%k1, %VRAX
	test	%VRAX, %VRAX
# endif
	jnz	L(return_vec_0)
	ret

	.p2align 4,, 11
L(return_vec_0):
	bsf	%VRAX, %VRAX
# ifdef USE_AS_WMEMCMP
	movl	(%rdi, %rax, CHAR_SIZE), %ecx
	xorl	%edx, %edx
	cmpl	(%rsi, %rax, CHAR_SIZE), %ecx
	/* NB: no partial register stall here because xorl zero idiom
	   above.  */
	setg	%dl
	leal	-1(%rdx, %rdx), %eax
# else
	movzbl	(%rsi, %rax), %ecx
#  if VEC_SIZE == 64
	movb	(%rdi, %rax), %al
#  else
	movzbl	(%rdi, %rax), %eax
#  endif
	subl	%ecx, %eax
# endif
	ret

	.p2align 4,, 11
L(more_1x_vec):
	/* From VEC to 2 * VEC.  No branch when size == VEC_SIZE.  */
	VMOVU	(%rsi), %VMM(1)
	/* Use compare not equals to directly check for mismatch.  */
	VPCMP	$4, (%rdi), %VMM(1), %k1
	KMOV	%k1, %VRAX
	/* NB: eax must be destination register if going to
	   L(return_vec_[0,2]). For L(return_vec_3) destination
	   register must be ecx.  */
	test	%VRAX, %VRAX
	jnz	L(return_vec_0)

	cmpq	$(CHAR_PER_VEC * 2), %rdx
	jbe	L(last_1x_vec)

	/* Check second VEC no matter what.  */
	VMOVU	VEC_SIZE(%rsi), %VMM(2)
	VPCMP	$4, VEC_SIZE(%rdi), %VMM(2), %k1
	KMOV	%k1, %VRAX
	test	%VRAX, %VRAX
	jnz	L(return_vec_1)

	/* Less than 4 * VEC.  */
	cmpq	$(CHAR_PER_VEC * 4), %rdx
	jbe	L(last_2x_vec)

	/* Check third and fourth VEC no matter what.  */
	VMOVU	(VEC_SIZE * 2)(%rsi), %VMM(3)
	VPCMP	$4, (VEC_SIZE * 2)(%rdi), %VMM(3), %k1
	KMOV	%k1, %VRAX
	test	%VRAX, %VRAX
	jnz	L(return_vec_2)

	VMOVU	(VEC_SIZE * 3)(%rsi), %VMM(4)
	VPCMP	$4, (VEC_SIZE * 3)(%rdi), %VMM(4), %k1
	KMOV	%k1, %VRCX
	test	%VRCX, %VRCX
	jnz	L(return_vec_3)

	/* Go to 4x VEC loop.  */
	cmpq	$(CHAR_PER_VEC * 8), %rdx
	ja	L(more_8x_vec)

	/* Handle remainder of size = 4 * VEC + 1 to 8 * VEC without any
	   branches.  */

	/* Load first two VEC from s2 before adjusting addresses.  */
	VMOVU	-(VEC_SIZE * 4)(%rsi, %rdx, CHAR_SIZE), %VMM(1)
	VMOVU	-(VEC_SIZE * 3)(%rsi, %rdx, CHAR_SIZE), %VMM(2)
	leaq	-(4 * VEC_SIZE)(%rdi, %rdx, CHAR_SIZE), %rdi
	leaq	-(4 * VEC_SIZE)(%rsi, %rdx, CHAR_SIZE), %rsi

	/* Wait to load from s1 until addressed adjust due to
	   unlamination of microfusion with complex address mode.  */

	/* vpxor will be all 0s if s1 and s2 are equal. Otherwise it
	   will have some 1s.  */
	vpxorq	(%rdi), %VMM(1), %VMM(1)
	vpxorq	(VEC_SIZE)(%rdi), %VMM(2), %VMM(2)

	VMOVU	(VEC_SIZE * 2)(%rsi), %VMM(3)
	vpxorq	(VEC_SIZE * 2)(%rdi), %VMM(3), %VMM(3)

	VMOVU	(VEC_SIZE * 3)(%rsi), %VMM(4)
	/* Ternary logic to xor (VEC_SIZE * 3)(%rdi) with VEC(4) while
	   oring with VEC(1). Result is stored in VEC(4).  */
	vpternlogd $0xde, (VEC_SIZE * 3)(%rdi), %VMM(1), %VMM(4)

	/* Or together VEC(2), VEC(3), and VEC(4) into VEC(4).  */
	vpternlogd $0xfe, %VMM(2), %VMM(3), %VMM(4)

	/* Test VEC(4) against itself. Store any CHAR mismatches in k1.
	 */
	VPTEST	%VMM(4), %VMM(4), %k1
	/* k1 must go to ecx for L(return_vec_0_1_2_3).  */
	KMOV	%k1, %VRCX
	test	%VRCX, %VRCX
	jnz	L(return_vec_0_1_2_3)
	/* NB: eax must be zero to reach here.  */
	ret


	.p2align 4,, 9
L(8x_end_return_vec_0_1_2_3):
	movq	%rdx, %rdi
L(8x_return_vec_0_1_2_3):
	/* L(loop_4x_vec) leaves result in `k1` for VEC_SIZE == 64.  */
# if VEC_SIZE == 64
	KMOV	%k1, %VRCX
# endif
	addq	%rdi, %rsi
L(return_vec_0_1_2_3):
	VPTEST	%VMM(1), %VMM(1), %k0
	KMOV	%k0, %VRAX
	test	%VRAX, %VRAX
	jnz	L(return_vec_0)

	VPTEST	%VMM(2), %VMM(2), %k0
	KMOV	%k0, %VRAX
	test	%VRAX, %VRAX
	jnz	L(return_vec_1)

	VPTEST	%VMM(3), %VMM(3), %k0
	KMOV	%k0, %VRAX
	test	%VRAX, %VRAX
	jnz	L(return_vec_2)
	.p2align 4,, 2
L(return_vec_3):
	/* bsf saves 1 byte from tzcnt. This keep L(return_vec_3) in one
	   fetch block and the entire L(*return_vec_0_1_2_3) in 1 cache
	   line.  */
	bsf	%VRCX, %VRCX
# ifdef USE_AS_WMEMCMP
	movl	(VEC_SIZE * 3)(%rdi, %rcx, CHAR_SIZE), %eax
	xorl	%edx, %edx
	cmpl	(VEC_SIZE * 3)(%rsi, %rcx, CHAR_SIZE), %eax
	setg	%dl
	leal	-1(%rdx, %rdx), %eax
# else
	movzbl	(VEC_SIZE * 3)(%rdi, %rcx), %eax
	movzbl	(VEC_SIZE * 3)(%rsi, %rcx), %ecx
	subl	%ecx, %eax
# endif
	ret


	.p2align 4,, 8
L(return_vec_1):
	/* bsf saves 1 byte over tzcnt and keeps L(return_vec_1) in one
	   fetch block.  */
	bsf	%VRAX, %VRAX
# ifdef USE_AS_WMEMCMP
	movl	VEC_SIZE(%rdi, %rax, CHAR_SIZE), %ecx
	xorl	%edx, %edx
	cmpl	VEC_SIZE(%rsi, %rax, CHAR_SIZE), %ecx
	setg	%dl
	leal	-1(%rdx, %rdx), %eax
# else
	movzbl	VEC_SIZE(%rsi, %rax), %ecx
	movzbl	VEC_SIZE(%rdi, %rax), %eax
	subl	%ecx, %eax
# endif
	ret

	.p2align 4,, 7
L(return_vec_2):
	/* bsf saves 1 byte over tzcnt and keeps L(return_vec_2) in one
	   fetch block.  */
	bsf	%VRAX, %VRAX
# ifdef USE_AS_WMEMCMP
	movl	(VEC_SIZE * 2)(%rdi, %rax, CHAR_SIZE), %ecx
	xorl	%edx, %edx
	cmpl	(VEC_SIZE * 2)(%rsi, %rax, CHAR_SIZE), %ecx
	setg	%dl
	leal	-1(%rdx, %rdx), %eax
# else
	movzbl	(VEC_SIZE * 2)(%rsi, %rax), %ecx
	movzbl	(VEC_SIZE * 2)(%rdi, %rax), %eax
	subl	%ecx, %eax
# endif
	ret

	.p2align 4,, 8
L(more_8x_vec):
	/* Set end of s1 in rdx.  */
	leaq	-(VEC_SIZE * 4)(%rdi, %rdx, CHAR_SIZE), %rdx
	/* rsi stores s2 - s1. This allows loop to only update one
	   pointer.  */
	subq	%rdi, %rsi
	/* Align s1 pointer.  */
	andq	$-VEC_SIZE, %rdi
	/* Adjust because first 4x vec where check already.  */
	subq	$-(VEC_SIZE * 4), %rdi

	.p2align 4
L(loop_4x_vec):
	VMOVU	(%rsi, %rdi), %VMM(1)
	vpxorq	(%rdi), %VMM(1), %VMM(1)
	VMOVU	VEC_SIZE(%rsi, %rdi), %VMM(2)
	vpxorq	VEC_SIZE(%rdi), %VMM(2), %VMM(2)
	VMOVU	(VEC_SIZE * 2)(%rsi, %rdi), %VMM(3)
	vpxorq	(VEC_SIZE * 2)(%rdi), %VMM(3), %VMM(3)
	VMOVU	(VEC_SIZE * 3)(%rsi, %rdi), %VMM(4)
	vpternlogd $0xde, (VEC_SIZE * 3)(%rdi), %VMM(1), %VMM(4)
	vpternlogd $0xfe, %VMM(2), %VMM(3), %VMM(4)
	VPTEST	%VMM(4), %VMM(4), %k1
	/* If VEC_SIZE == 64 just branch with KTEST. We have free port0
	   space and it allows the loop to fit in 2x cache lines
	   instead of 3.  */
# if VEC_SIZE == 64
	KTEST	%k1, %k1
# else
	KMOV	%k1, %VRCX
	test	%VRCX, %VRCX
# endif
	jnz	L(8x_return_vec_0_1_2_3)
	subq	$-(VEC_SIZE * 4), %rdi
	cmpq	%rdx, %rdi
	jb	L(loop_4x_vec)
	subq	%rdx, %rdi
	/* rdi has 4 * VEC_SIZE - remaining length.  */
	cmpl	$(VEC_SIZE * 3), %edi
	jge	L(8x_last_1x_vec)
	/* Load regardless of branch.  */
	VMOVU	(VEC_SIZE * 2)(%rsi, %rdx), %VMM(3)

	/* Separate logic as we can only use testb for VEC_SIZE == 64.
	 */
# if VEC_SIZE == 64
	testb	%dil, %dil
	js	L(8x_last_2x_vec)
# else
	cmpl	$(VEC_SIZE * 2), %edi
	jge	L(8x_last_2x_vec)
# endif

	vpxorq	(VEC_SIZE * 2)(%rdx), %VMM(3), %VMM(3)

	VMOVU	(%rsi, %rdx), %VMM(1)
	vpxorq	(%rdx), %VMM(1), %VMM(1)

	VMOVU	VEC_SIZE(%rsi, %rdx), %VMM(2)
	vpxorq	VEC_SIZE(%rdx), %VMM(2), %VMM(2)
	VMOVU	(VEC_SIZE * 3)(%rsi, %rdx), %VMM(4)
	vpternlogd $0xde, (VEC_SIZE * 3)(%rdx), %VMM(1), %VMM(4)
	vpternlogd $0xfe, %VMM(2), %VMM(3), %VMM(4)
	VPTEST	%VMM(4), %VMM(4), %k1
	/* L(8x_end_return_vec_0_1_2_3) expects bitmask to still be in
	   `k1`  if VEC_SIZE == 64.  */
# if VEC_SIZE == 64
	KTEST	%k1, %k1
# else
	KMOV	%k1, %VRCX
	test	%VRCX, %VRCX
# endif
	jnz	L(8x_end_return_vec_0_1_2_3)
	/* NB: eax must be zero to reach here.  */
	ret

	/* Only entry is from L(more_8x_vec).  */
	.p2align 4,, 6
L(8x_last_2x_vec):
	VPCMP	$4, (VEC_SIZE * 2)(%rdx), %VMM(3), %k1
	KMOV	%k1, %VRAX
	test	%VRAX, %VRAX
	jnz	L(8x_return_vec_2)
	.p2align 4,, 5
L(8x_last_1x_vec):
	VMOVU	(VEC_SIZE * 3)(%rsi, %rdx), %VMM(1)
	VPCMP	$4, (VEC_SIZE * 3)(%rdx), %VMM(1), %k1
	KMOV	%k1, %VRAX
	test	%VRAX, %VRAX
	jnz	L(8x_return_vec_3)
	ret

	/* Not ideally aligned (at offset +9 bytes in fetch block) but
	   not aligning keeps it in the same cache line as
	   L(8x_last_1x/2x_vec) so likely worth it. As well, saves code
	   size.  */
	.p2align 4,, 4
L(8x_return_vec_2):
	subq	$VEC_SIZE, %rdx
L(8x_return_vec_3):
	bsf	%VRAX, %VRAX
# ifdef USE_AS_WMEMCMP
	leaq	(%rdx, %rax, CHAR_SIZE), %rax
	movl	(VEC_SIZE * 3)(%rax), %ecx
	xorl	%edx, %edx
	cmpl	(VEC_SIZE * 3)(%rsi, %rax), %ecx
	setg	%dl
	leal	-1(%rdx, %rdx), %eax
# else
	addq	%rdx, %rax
	movzbl	(VEC_SIZE * 3)(%rsi, %rax), %ecx
	movzbl	(VEC_SIZE * 3)(%rax), %eax
	subl	%ecx, %eax
# endif
	ret

	.p2align 4,, 8
L(last_2x_vec):
	/* Check second to last VEC.  */
	VMOVU	-(VEC_SIZE * 2)(%rsi, %rdx, CHAR_SIZE), %VMM(1)
	VPCMP	$4, -(VEC_SIZE * 2)(%rdi, %rdx, CHAR_SIZE), %VMM(1), %k1
	KMOV	%k1, %VRAX
	test	%VRAX, %VRAX
	jnz	L(return_vec_1_end)

	/* Check last VEC.  */
	.p2align 4,, 8
L(last_1x_vec):
	VMOVU	-(VEC_SIZE * 1)(%rsi, %rdx, CHAR_SIZE), %VMM(1)
	VPCMP	$4, -(VEC_SIZE * 1)(%rdi, %rdx, CHAR_SIZE), %VMM(1), %k1
	KMOV	%k1, %VRAX
	test	%VRAX, %VRAX
	jnz	L(return_vec_0_end)
	ret


	/* Don't fully align. Takes 2-fetch blocks either way and
	   aligning will cause code to spill into another cacheline.
	 */
	.p2align 4,, 3
L(return_vec_1_end):
	/* Use bsf to save code size. This is necessary to have
	   L(one_or_less) fit in aligning bytes between.  */
	bsf	%VRAX, %VRAX
	addl	%edx, %eax
# ifdef USE_AS_WMEMCMP
	movl	-(VEC_SIZE * 2)(%rdi, %rax, CHAR_SIZE), %ecx
	xorl	%edx, %edx
	cmpl	-(VEC_SIZE * 2)(%rsi, %rax, CHAR_SIZE), %ecx
	setg	%dl
	leal	-1(%rdx, %rdx), %eax
# else
	movzbl	-(VEC_SIZE * 2)(%rsi, %rax), %ecx
	movzbl	-(VEC_SIZE * 2)(%rdi, %rax), %eax
	subl	%ecx, %eax
# endif
	ret

	.p2align 4,, 2
	/* Don't align. Takes 2-fetch blocks either way and aligning
	   will cause code to spill into another cacheline.  */
L(return_vec_0_end):
	bsf	%VRAX, %VRAX
	addl	%edx, %eax
# ifdef USE_AS_WMEMCMP
	movl	-VEC_SIZE(%rdi, %rax, CHAR_SIZE), %ecx
	xorl	%edx, %edx
	cmpl	-VEC_SIZE(%rsi, %rax, CHAR_SIZE), %ecx
	setg	%dl
	leal	-1(%rdx, %rdx), %eax
# else
	movzbl	-VEC_SIZE(%rsi, %rax), %ecx
	movzbl	-VEC_SIZE(%rdi, %rax), %eax
	subl	%ecx, %eax
# endif
	ret
	/* evex256: 2-byte until next cache line. evex512: 46-bytes
	   until next cache line.  */
END (MEMCMP)
#endif