x86-64: Add memcmp family functions with 256-bit EVEX

Update ifunc-memcmp.h to select the function optimized with 256-bit EVEX
instructions using YMM16-YMM31 registers to avoid RTM abort with usable
AVX512VL, AVX512BW and MOVBE since VZEROUPPER isn't needed at function
exit.

5 files changed, 467 insertions, 4 deletions

diff --git a/sysdeps/x86_64/multiarch/Makefile b/sysdeps/x86_64/multiarch/Makefile
index 1cc0a10e12..9d79b138e9 100644
--- a/sysdeps/x86_64/multiarch/Makefile
+++ b/sysdeps/x86_64/multiarch/Makefile
@@ -41,6 +41,7 @@ sysdep_routines += strncat-c stpncpy-c strncpy-c \
 		   memset-avx2-unaligned-erms \
 		   memset-avx512-unaligned-erms \
 		   memchr-evex \
+		   memcmp-evex-movbe \
 		   memmove-evex-unaligned-erms \
 		   memrchr-evex \
 		   memset-evex-unaligned-erms \
@@ -81,7 +82,8 @@ sysdep_routines += wmemcmp-sse4 wmemcmp-ssse3 wmemcmp-c \
 		   wcsncmp-evex \
 		   wcsnlen-evex \
 		   wcsrchr-evex \
-		   wmemchr-evex
+		   wmemchr-evex \
+		   wmemcmp-evex-movbe
 endif
 
 ifeq ($(subdir),debug)
diff --git a/sysdeps/x86_64/multiarch/ifunc-impl-list.c b/sysdeps/x86_64/multiarch/ifunc-impl-list.c
index aac8e601df..96344a71e4 100644
--- a/sysdeps/x86_64/multiarch/ifunc-impl-list.c
+++ b/sysdeps/x86_64/multiarch/ifunc-impl-list.c
@@ -56,6 +56,11 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
 			      (CPU_FEATURE_USABLE (AVX2)
 			       && CPU_FEATURE_USABLE (MOVBE)),
 			      __memcmp_avx2_movbe)
+	      IFUNC_IMPL_ADD (array, i, memcmp,
+			      (CPU_FEATURE_USABLE (AVX512VL)
+			       && CPU_FEATURE_USABLE (AVX512BW)
+			       && CPU_FEATURE_USABLE (MOVBE)),
+			      __memcmp_evex_movbe)
 	      IFUNC_IMPL_ADD (array, i, memcmp, CPU_FEATURE_USABLE (SSE4_1),
 			      __memcmp_sse4_1)
 	      IFUNC_IMPL_ADD (array, i, memcmp, CPU_FEATURE_USABLE (SSSE3),
@@ -558,6 +563,11 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
 			      (CPU_FEATURE_USABLE (AVX2)
 			       && CPU_FEATURE_USABLE (MOVBE)),
 			      __wmemcmp_avx2_movbe)
+	      IFUNC_IMPL_ADD (array, i, wmemcmp,
+			      (CPU_FEATURE_USABLE (AVX512VL)
+			       && CPU_FEATURE_USABLE (AVX512BW)
+			       && CPU_FEATURE_USABLE (MOVBE)),
+			      __wmemcmp_evex_movbe)
 	      IFUNC_IMPL_ADD (array, i, wmemcmp, CPU_FEATURE_USABLE (SSE4_1),
 			      __wmemcmp_sse4_1)
 	      IFUNC_IMPL_ADD (array, i, wmemcmp, CPU_FEATURE_USABLE (SSSE3),
diff --git a/sysdeps/x86_64/multiarch/ifunc-memcmp.h b/sysdeps/x86_64/multiarch/ifunc-memcmp.h
index d5df541ec4..5ac41a19b8 100644
--- a/sysdeps/x86_64/multiarch/ifunc-memcmp.h
+++ b/sysdeps/x86_64/multiarch/ifunc-memcmp.h
@@ -23,17 +23,24 @@ extern __typeof (REDIRECT_NAME) OPTIMIZE (sse2) attribute_hidden;
 extern __typeof (REDIRECT_NAME) OPTIMIZE (ssse3) attribute_hidden;
 extern __typeof (REDIRECT_NAME) OPTIMIZE (sse4_1) attribute_hidden;
 extern __typeof (REDIRECT_NAME) OPTIMIZE (avx2_movbe) attribute_hidden;
+extern __typeof (REDIRECT_NAME) OPTIMIZE (evex_movbe) attribute_hidden;
 
 static inline void *
 IFUNC_SELECTOR (void)
 {
   const struct cpu_features* cpu_features = __get_cpu_features ();
 
-  if (!CPU_FEATURES_ARCH_P (cpu_features, Prefer_No_VZEROUPPER)
-      && CPU_FEATURE_USABLE_P (cpu_features, AVX2)
+  if (CPU_FEATURE_USABLE_P (cpu_features, AVX2)
       && CPU_FEATURE_USABLE_P (cpu_features, MOVBE)
       && CPU_FEATURES_ARCH_P (cpu_features, AVX_Fast_Unaligned_Load))
-    return OPTIMIZE (avx2_movbe);
+    {
+      if (CPU_FEATURE_USABLE_P (cpu_features, AVX512VL)
+	  && CPU_FEATURE_USABLE_P (cpu_features, AVX512BW))
+	return OPTIMIZE (evex_movbe);
+
+      if (!CPU_FEATURES_ARCH_P (cpu_features, Prefer_No_VZEROUPPER))
+	return OPTIMIZE (avx2_movbe);
+    }
 
   if (CPU_FEATURE_USABLE_P (cpu_features, SSE4_1))
     return OPTIMIZE (sse4_1);
diff --git a/sysdeps/x86_64/multiarch/memcmp-evex-movbe.S b/sysdeps/x86_64/multiarch/memcmp-evex-movbe.S
new file mode 100644
index 0000000000..9c093972e1
--- /dev/null
+++ b/sysdeps/x86_64/multiarch/memcmp-evex-movbe.S
@@ -0,0 +1,440 @@
+/* memcmp/wmemcmp optimized with 256-bit EVEX instructions.
+   Copyright (C) 2021 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/>.  */
+
+#if IS_IN (libc)
+
+/* memcmp/wmemcmp is implemented as:
+   1. For size from 2 to 7 bytes, load as big endian with movbe and bswap
+      to avoid branches.
+   2. Use overlapping compare to avoid branch.
+   3. Use vector compare when size >= 4 bytes for memcmp or size >= 8
+      bytes for wmemcmp.
+   4. If size is 8 * VEC_SIZE or less, unroll the loop.
+   5. Compare 4 * VEC_SIZE at a time with the aligned first memory
+      area.
+   6. Use 2 vector compares when size is 2 * VEC_SIZE or less.
+   7. Use 4 vector compares when size is 4 * VEC_SIZE or less.
+   8. Use 8 vector compares when size is 8 * VEC_SIZE or less.  */
+
+# include <sysdep.h>
+
+# ifndef MEMCMP
+#  define MEMCMP	__memcmp_evex_movbe
+# endif
+
+# define VMOVU		vmovdqu64
+
+# ifdef USE_AS_WMEMCMP
+#  define VPCMPEQ	vpcmpeqd
+# else
+#  define VPCMPEQ	vpcmpeqb
+# endif
+
+# define XMM1		xmm17
+# define XMM2		xmm18
+# define YMM1		ymm17
+# define YMM2		ymm18
+# define YMM3		ymm19
+# define YMM4		ymm20
+# define YMM5		ymm21
+# define YMM6		ymm22
+
+# define VEC_SIZE 32
+# ifdef USE_AS_WMEMCMP
+#  define VEC_MASK 0xff
+#  define XMM_MASK 0xf
+# else
+#  define VEC_MASK 0xffffffff
+#  define XMM_MASK 0xffff
+# endif
+
+/* Warning!
+           wmemcmp has to use SIGNED comparison for elements.
+           memcmp has to use UNSIGNED comparison for elemnts.
+*/
+
+	.section .text.evex,"ax",@progbits
+ENTRY (MEMCMP)
+# ifdef USE_AS_WMEMCMP
+	shl	$2, %RDX_LP
+# elif defined __ILP32__
+	/* Clear the upper 32 bits.  */
+	movl	%edx, %edx
+# endif
+	cmp	$VEC_SIZE, %RDX_LP
+	jb	L(less_vec)
+
+	/* From VEC to 2 * VEC.  No branch when size == VEC_SIZE.  */
+	VMOVU	(%rsi), %YMM2
+	VPCMPEQ (%rdi), %YMM2, %k1
+	kmovd	%k1, %eax
+	subl    $VEC_MASK, %eax
+	jnz	L(first_vec)
+
+	cmpq	$(VEC_SIZE * 2), %rdx
+	jbe	L(last_vec)
+
+	/* More than 2 * VEC.  */
+	cmpq	$(VEC_SIZE * 8), %rdx
+	ja	L(more_8x_vec)
+	cmpq	$(VEC_SIZE * 4), %rdx
+	jb	L(last_4x_vec)
+
+	/* From 4 * VEC to 8 * VEC, inclusively. */
+	VMOVU	(%rsi), %YMM1
+	VPCMPEQ (%rdi), %YMM1, %k1
+
+	VMOVU	VEC_SIZE(%rsi), %YMM2
+	VPCMPEQ VEC_SIZE(%rdi), %YMM2, %k2
+
+	VMOVU	(VEC_SIZE * 2)(%rsi), %YMM3
+	VPCMPEQ (VEC_SIZE * 2)(%rdi), %YMM3, %k3
+
+	VMOVU	(VEC_SIZE * 3)(%rsi), %YMM4
+	VPCMPEQ (VEC_SIZE * 3)(%rdi), %YMM4, %k4
+
+	kandd	%k1, %k2, %k5
+	kandd	%k3, %k4, %k6
+	kandd	%k5, %k6, %k6
+
+	kmovd	%k6, %eax
+	cmpl	$VEC_MASK, %eax
+	jne	L(4x_vec_end)
+
+	leaq	-(4 * VEC_SIZE)(%rdi, %rdx), %rdi
+	leaq	-(4 * VEC_SIZE)(%rsi, %rdx), %rsi
+	VMOVU	(%rsi), %YMM1
+	VPCMPEQ (%rdi), %YMM1, %k1
+
+	VMOVU	VEC_SIZE(%rsi), %YMM2
+	VPCMPEQ VEC_SIZE(%rdi), %YMM2, %k2
+	kandd	%k1, %k2, %k5
+
+	VMOVU	(VEC_SIZE * 2)(%rsi), %YMM3
+	VPCMPEQ (VEC_SIZE * 2)(%rdi), %YMM3, %k3
+	kandd	%k3, %k5, %k5
+
+	VMOVU	(VEC_SIZE * 3)(%rsi), %YMM4
+	VPCMPEQ (VEC_SIZE * 3)(%rdi), %YMM4, %k4
+	kandd	%k4, %k5, %k5
+
+	kmovd	%k5, %eax
+	cmpl	$VEC_MASK, %eax
+	jne	L(4x_vec_end)
+	xorl	%eax, %eax
+	ret
+
+	.p2align 4
+L(last_2x_vec):
+	/* From VEC to 2 * VEC.  No branch when size == VEC_SIZE.  */
+	VMOVU	(%rsi), %YMM2
+	VPCMPEQ (%rdi), %YMM2, %k2
+	kmovd	%k2, %eax
+	subl    $VEC_MASK, %eax
+	jnz	L(first_vec)
+
+L(last_vec):
+	/* Use overlapping loads to avoid branches.  */
+	leaq	-VEC_SIZE(%rdi, %rdx), %rdi
+	leaq	-VEC_SIZE(%rsi, %rdx), %rsi
+	VMOVU	(%rsi), %YMM2
+	VPCMPEQ (%rdi), %YMM2, %k2
+	kmovd	%k2, %eax
+	subl    $VEC_MASK, %eax
+	jnz	L(first_vec)
+	ret
+
+	.p2align 4
+L(first_vec):
+	/* A byte or int32 is different within 16 or 32 bytes.  */
+	tzcntl	%eax, %ecx
+# ifdef USE_AS_WMEMCMP
+	xorl	%eax, %eax
+	movl	(%rdi, %rcx, 4), %edx
+	cmpl	(%rsi, %rcx, 4), %edx
+L(wmemcmp_return):
+	setl	%al
+	negl	%eax
+	orl	$1, %eax
+# else
+	movzbl	(%rdi, %rcx), %eax
+	movzbl	(%rsi, %rcx), %edx
+	sub	%edx, %eax
+# endif
+	ret
+
+# ifdef USE_AS_WMEMCMP
+	.p2align 4
+L(4):
+	xorl	%eax, %eax
+	movl	(%rdi), %edx
+	cmpl	(%rsi), %edx
+	jne	L(wmemcmp_return)
+	ret
+# else
+	.p2align 4
+L(between_4_7):
+	/* Load as big endian with overlapping movbe to avoid branches.  */
+	movbe	(%rdi), %eax
+	movbe	(%rsi), %ecx
+	shlq	$32, %rax
+	shlq	$32, %rcx
+	movbe	-4(%rdi, %rdx), %edi
+	movbe	-4(%rsi, %rdx), %esi
+	orq	%rdi, %rax
+	orq	%rsi, %rcx
+	subq	%rcx, %rax
+	je	L(exit)
+	sbbl	%eax, %eax
+	orl	$1, %eax
+	ret
+
+	.p2align 4
+L(exit):
+	ret
+
+	.p2align 4
+L(between_2_3):
+	/* Load as big endian to avoid branches.  */
+	movzwl	(%rdi), %eax
+	movzwl	(%rsi), %ecx
+	shll	$8, %eax
+	shll	$8, %ecx
+	bswap	%eax
+	bswap	%ecx
+	movb	-1(%rdi, %rdx), %al
+	movb	-1(%rsi, %rdx), %cl
+	/* Subtraction is okay because the upper 8 bits are zero.  */
+	subl	%ecx, %eax
+	ret
+
+	.p2align 4
+L(1):
+	movzbl	(%rdi), %eax
+	movzbl	(%rsi), %ecx
+	subl	%ecx, %eax
+	ret
+# endif
+
+	.p2align 4
+L(zero):
+	xorl	%eax, %eax
+	ret
+
+	.p2align 4
+L(less_vec):
+# ifdef USE_AS_WMEMCMP
+	/* It can only be 0, 4, 8, 12, 16, 20, 24, 28 bytes.  */
+	cmpb	$4, %dl
+	je	L(4)
+	jb	L(zero)
+# else
+	cmpb	$1, %dl
+	je	L(1)
+	jb	L(zero)
+	cmpb	$4, %dl
+	jb	L(between_2_3)
+	cmpb	$8, %dl
+	jb	L(between_4_7)
+# endif
+	cmpb	$16, %dl
+	jae	L(between_16_31)
+	/* It is between 8 and 15 bytes.  */
+	vmovq	(%rdi), %XMM1
+	vmovq	(%rsi), %XMM2
+	VPCMPEQ %XMM1, %XMM2, %k2
+	kmovw	%k2, %eax
+	subl    $XMM_MASK, %eax
+	jnz	L(first_vec)
+	/* Use overlapping loads to avoid branches.  */
+	leaq	-8(%rdi, %rdx), %rdi
+	leaq	-8(%rsi, %rdx), %rsi
+	vmovq	(%rdi), %XMM1
+	vmovq	(%rsi), %XMM2
+	VPCMPEQ %XMM1, %XMM2, %k2
+	kmovw	%k2, %eax
+	subl    $XMM_MASK, %eax
+	jnz	L(first_vec)
+	ret
+
+	.p2align 4
+L(between_16_31):
+	/* From 16 to 31 bytes.  No branch when size == 16.  */
+	VMOVU	(%rsi), %XMM2
+	VPCMPEQ (%rdi), %XMM2, %k2
+	kmovw	%k2, %eax
+	subl    $XMM_MASK, %eax
+	jnz	L(first_vec)
+
+	/* Use overlapping loads to avoid branches.  */
+	leaq	-16(%rdi, %rdx), %rdi
+	leaq	-16(%rsi, %rdx), %rsi
+	VMOVU	(%rsi), %XMM2
+	VPCMPEQ (%rdi), %XMM2, %k2
+	kmovw	%k2, %eax
+	subl    $XMM_MASK, %eax
+	jnz	L(first_vec)
+	ret
+
+	.p2align 4
+L(more_8x_vec):
+	/* More than 8 * VEC.  Check the first VEC.  */
+	VMOVU	(%rsi), %YMM2
+	VPCMPEQ (%rdi), %YMM2, %k2
+	kmovd	%k2, %eax
+	subl    $VEC_MASK, %eax
+	jnz	L(first_vec)
+
+	/* Align the first memory area for aligned loads in the loop.
+	   Compute how much the first memory area is misaligned.  */
+	movq	%rdi, %rcx
+	andl	$(VEC_SIZE - 1), %ecx
+	/* Get the negative of offset for alignment.  */
+	subq	$VEC_SIZE, %rcx
+	/* Adjust the second memory area.  */
+	subq	%rcx, %rsi
+	/* Adjust the first memory area which should be aligned now.  */
+	subq	%rcx, %rdi
+	/* Adjust length.  */
+	addq	%rcx, %rdx
+
+L(loop_4x_vec):
+	/* Compare 4 * VEC at a time forward.  */
+	VMOVU	(%rsi), %YMM1
+	VPCMPEQ (%rdi), %YMM1, %k1
+
+	VMOVU	VEC_SIZE(%rsi), %YMM2
+	VPCMPEQ VEC_SIZE(%rdi), %YMM2, %k2
+	kandd	%k2, %k1, %k5
+
+	VMOVU	(VEC_SIZE * 2)(%rsi), %YMM3
+	VPCMPEQ (VEC_SIZE * 2)(%rdi), %YMM3, %k3
+	kandd	%k3, %k5, %k5
+
+	VMOVU	(VEC_SIZE * 3)(%rsi), %YMM4
+	VPCMPEQ (VEC_SIZE * 3)(%rdi), %YMM4, %k4
+	kandd	%k4, %k5, %k5
+
+	kmovd	%k5, %eax
+	cmpl	$VEC_MASK, %eax
+	jne	L(4x_vec_end)
+
+	addq	$(VEC_SIZE * 4), %rdi
+	addq	$(VEC_SIZE * 4), %rsi
+
+	subq	$(VEC_SIZE * 4), %rdx
+	cmpq	$(VEC_SIZE * 4), %rdx
+	jae	L(loop_4x_vec)
+
+	/* Less than 4 * VEC.  */
+	cmpq	$VEC_SIZE, %rdx
+	jbe	L(last_vec)
+	cmpq	$(VEC_SIZE * 2), %rdx
+	jbe	L(last_2x_vec)
+
+L(last_4x_vec):
+	/* From 2 * VEC to 4 * VEC. */
+	VMOVU	(%rsi), %YMM2
+	VPCMPEQ (%rdi), %YMM2, %k2
+	kmovd	%k2, %eax
+	subl    $VEC_MASK, %eax
+	jnz	L(first_vec)
+
+	addq	$VEC_SIZE, %rdi
+	addq	$VEC_SIZE, %rsi
+	VMOVU	(%rsi), %YMM2
+	VPCMPEQ (%rdi), %YMM2, %k2
+	kmovd	%k2, %eax
+	subl    $VEC_MASK, %eax
+	jnz	L(first_vec)
+
+	/* Use overlapping loads to avoid branches.  */
+	leaq	-(3 * VEC_SIZE)(%rdi, %rdx), %rdi
+	leaq	-(3 * VEC_SIZE)(%rsi, %rdx), %rsi
+	VMOVU	(%rsi), %YMM2
+	VPCMPEQ (%rdi), %YMM2, %k2
+	kmovd	%k2, %eax
+	subl    $VEC_MASK, %eax
+	jnz	L(first_vec)
+
+	addq	$VEC_SIZE, %rdi
+	addq	$VEC_SIZE, %rsi
+	VMOVU	(%rsi), %YMM2
+	VPCMPEQ (%rdi), %YMM2, %k2
+	kmovd	%k2, %eax
+	subl    $VEC_MASK, %eax
+	jnz	L(first_vec)
+	ret
+
+	.p2align 4
+L(4x_vec_end):
+	kmovd	%k1, %eax
+	subl	$VEC_MASK, %eax
+	jnz	L(first_vec)
+	kmovd	%k2, %eax
+	subl	$VEC_MASK, %eax
+	jnz	L(first_vec_x1)
+	kmovd	%k3, %eax
+	subl	$VEC_MASK, %eax
+	jnz	L(first_vec_x2)
+	kmovd	%k4, %eax
+	subl	$VEC_MASK, %eax
+	tzcntl	%eax, %ecx
+# ifdef USE_AS_WMEMCMP
+	xorl	%eax, %eax
+	movl	(VEC_SIZE * 3)(%rdi, %rcx, 4), %edx
+	cmpl	(VEC_SIZE * 3)(%rsi, %rcx, 4), %edx
+	jmp	L(wmemcmp_return)
+# else
+	movzbl	(VEC_SIZE * 3)(%rdi, %rcx), %eax
+	movzbl	(VEC_SIZE * 3)(%rsi, %rcx), %edx
+	sub	%edx, %eax
+# endif
+	ret
+
+	.p2align 4
+L(first_vec_x1):
+	tzcntl	%eax, %ecx
+# ifdef USE_AS_WMEMCMP
+	xorl	%eax, %eax
+	movl	VEC_SIZE(%rdi, %rcx, 4), %edx
+	cmpl	VEC_SIZE(%rsi, %rcx, 4), %edx
+	jmp	L(wmemcmp_return)
+# else
+	movzbl	VEC_SIZE(%rdi, %rcx), %eax
+	movzbl	VEC_SIZE(%rsi, %rcx), %edx
+	sub	%edx, %eax
+# endif
+	ret
+
+	.p2align 4
+L(first_vec_x2):
+	tzcntl	%eax, %ecx
+# ifdef USE_AS_WMEMCMP
+	xorl	%eax, %eax
+	movl	(VEC_SIZE * 2)(%rdi, %rcx, 4), %edx
+	cmpl	(VEC_SIZE * 2)(%rsi, %rcx, 4), %edx
+	jmp	L(wmemcmp_return)
+# else
+	movzbl	(VEC_SIZE * 2)(%rdi, %rcx), %eax
+	movzbl	(VEC_SIZE * 2)(%rsi, %rcx), %edx
+	sub	%edx, %eax
+# endif
+	ret
+END (MEMCMP)
+#endif
diff --git a/sysdeps/x86_64/multiarch/wmemcmp-evex-movbe.S b/sysdeps/x86_64/multiarch/wmemcmp-evex-movbe.S
new file mode 100644
index 0000000000..4726d74aa1
--- /dev/null
+++ b/sysdeps/x86_64/multiarch/wmemcmp-evex-movbe.S
@@ -0,0 +1,4 @@
+#define MEMCMP __wmemcmp_evex_movbe
+#define USE_AS_WMEMCMP 1
+
+#include "memcmp-evex-movbe.S"