[PATCH v1 5/6] x86_64: Add avx2 optimized __memcmpeq in memcmpeq-avx2.S

[Noah Goldstein <goldstein.w.n@gmail.com>]

No bug. This commit adds new optimized __memcmpeq implementation for
avx2.

The primary optimizations are:

1) skipping the logic to find the difference of the first mismatched
byte.

2) not updating src/dst addresses as the non-equals logic does not
need to be reused by different areas.
---
 sysdeps/x86_64/multiarch/ifunc-impl-list.c   |   2 -
 sysdeps/x86_64/multiarch/ifunc-memcmpeq.h    |   2 +-
 sysdeps/x86_64/multiarch/memcmpeq-avx2-rtm.S |   4 +-
 sysdeps/x86_64/multiarch/memcmpeq-avx2.S     | 309 ++++++++++++++++++-
 4 files changed, 308 insertions(+), 9 deletions(-)

diff --git a/sysdeps/x86_64/multiarch/ifunc-impl-list.c b/sysdeps/x86_64/multiarch/ifunc-impl-list.c
index f7f3806d1d..535450f52c 100644
--- a/sysdeps/x86_64/multiarch/ifunc-impl-list.c
+++ b/sysdeps/x86_64/multiarch/ifunc-impl-list.c
@@ -42,13 +42,11 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
   IFUNC_IMPL (i, name, __memcmpeq,
 	      IFUNC_IMPL_ADD (array, i, __memcmpeq,
 			      (CPU_FEATURE_USABLE (AVX2)
-                   && CPU_FEATURE_USABLE (MOVBE)
 			       && CPU_FEATURE_USABLE (BMI2)),
 			      __memcmpeq_avx2)
 	      IFUNC_IMPL_ADD (array, i, __memcmpeq,
 			      (CPU_FEATURE_USABLE (AVX2)
 			       && CPU_FEATURE_USABLE (BMI2)
-                   && CPU_FEATURE_USABLE (MOVBE)
 			       && CPU_FEATURE_USABLE (RTM)),
 			      __memcmpeq_avx2_rtm)
 	      IFUNC_IMPL_ADD (array, i, __memcmpeq,
diff --git a/sysdeps/x86_64/multiarch/ifunc-memcmpeq.h b/sysdeps/x86_64/multiarch/ifunc-memcmpeq.h
index 3319a9568a..e596c5048b 100644
--- a/sysdeps/x86_64/multiarch/ifunc-memcmpeq.h
+++ b/sysdeps/x86_64/multiarch/ifunc-memcmpeq.h
@@ -31,10 +31,10 @@ IFUNC_SELECTOR (void)
 
   if (CPU_FEATURE_USABLE_P (cpu_features, AVX2)
       && CPU_FEATURE_USABLE_P (cpu_features, BMI2)
-      && CPU_FEATURE_USABLE_P (cpu_features, MOVBE)
       && CPU_FEATURES_ARCH_P (cpu_features, AVX_Fast_Unaligned_Load))
     {
       if (CPU_FEATURE_USABLE_P (cpu_features, AVX512VL)
+	  && CPU_FEATURE_USABLE_P (cpu_features, MOVBE)
 	  && CPU_FEATURE_USABLE_P (cpu_features, AVX512BW))
 	return OPTIMIZE1 (evex);
 
diff --git a/sysdeps/x86_64/multiarch/memcmpeq-avx2-rtm.S b/sysdeps/x86_64/multiarch/memcmpeq-avx2-rtm.S
index 24b6a0c9ff..3264a4a76c 100644
--- a/sysdeps/x86_64/multiarch/memcmpeq-avx2-rtm.S
+++ b/sysdeps/x86_64/multiarch/memcmpeq-avx2-rtm.S
@@ -1,5 +1,5 @@
-#ifndef MEMCMP
-# define MEMCMP __memcmpeq_avx2_rtm
+#ifndef MEMCMPEQ
+# define MEMCMPEQ __memcmpeq_avx2_rtm
 #endif
 
 #define ZERO_UPPER_VEC_REGISTERS_RETURN \
diff --git a/sysdeps/x86_64/multiarch/memcmpeq-avx2.S b/sysdeps/x86_64/multiarch/memcmpeq-avx2.S
index 0181ea0d8d..0bf59fb8fa 100644
--- a/sysdeps/x86_64/multiarch/memcmpeq-avx2.S
+++ b/sysdeps/x86_64/multiarch/memcmpeq-avx2.S
@@ -16,8 +16,309 @@
    License along with the GNU C Library; if not, see
    <https://www.gnu.org/licenses/>.  */
 
-#ifndef MEMCMP
-# define MEMCMP	__memcmpeq_avx2
-#endif
+#if IS_IN (libc)
+
+/* __memcmpeq is implemented as:
+   1. Use ymm vector compares when possible. The only case where
+      vector compares is not possible for when size < VEC_SIZE
+      and loading from either s1 or s2 would cause a page cross.
+   2. Use xmm vector compare when size >= 8 bytes.
+   3. Optimistically compare up to first 4 * VEC_SIZE one at a
+      to check for early mismatches. Only do this if its guranteed the
+      work is not wasted.
+   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 MEMCMPEQ
+#  define MEMCMPEQ	__memcmpeq_avx2
+# endif
+
+# define VPCMPEQ	vpcmpeqb
+
+# ifndef VZEROUPPER
+#  define VZEROUPPER	vzeroupper
+# endif
+
+# ifndef SECTION
+#  define SECTION(p)	p##.avx
+# endif
+
+# define VEC_SIZE 32
+# define PAGE_SIZE	4096
+
+	.section SECTION(.text), "ax", @progbits
+ENTRY_P2ALIGN (MEMCMPEQ, 6)
+# ifdef __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.  */
+	vmovdqu	(%rsi), %ymm1
+	VPCMPEQ	(%rdi), %ymm1, %ymm1
+	vpmovmskb %ymm1, %eax
+	incl	%eax
+	jnz	L(return_neq0)
+	cmpq	$(VEC_SIZE * 2), %rdx
+	jbe	L(last_1x_vec)
+
+	/* Check second VEC no matter what.  */
+	vmovdqu	VEC_SIZE(%rsi), %ymm2
+	VPCMPEQ	VEC_SIZE(%rdi), %ymm2, %ymm2
+	vpmovmskb %ymm2, %eax
+	/* If all 4 VEC where equal eax will be all 1s so incl will overflow
+	   and set zero flag.  */
+	incl	%eax
+	jnz	L(return_neq0)
+
+	/* Less than 4 * VEC.  */
+	cmpq	$(VEC_SIZE * 4), %rdx
+	jbe	L(last_2x_vec)
+
+	/* Check third and fourth VEC no matter what.  */
+	vmovdqu	(VEC_SIZE * 2)(%rsi), %ymm3
+	VPCMPEQ	(VEC_SIZE * 2)(%rdi), %ymm3, %ymm3
+	vpmovmskb %ymm3, %eax
+	incl	%eax
+	jnz	L(return_neq0)
+
+	vmovdqu	(VEC_SIZE * 3)(%rsi), %ymm4
+	VPCMPEQ	(VEC_SIZE * 3)(%rdi), %ymm4, %ymm4
+	vpmovmskb %ymm4, %eax
+	incl	%eax
+	jnz	L(return_neq0)
+
+	/* Go to 4x VEC loop.  */
+	cmpq	$(VEC_SIZE * 8), %rdx
+	ja	L(more_8x_vec)
+
+	/* Handle remainder of size = 4 * VEC + 1 to 8 * VEC without any
+	   branches.  */
+
+	/* Adjust rsi and rdi to avoid indexed address mode. This end up
+	   saving a 16 bytes of code, prevents unlamination, and bottlenecks in
+	   the AGU.  */
+	addq	%rdx, %rsi
+	vmovdqu	-(VEC_SIZE * 4)(%rsi), %ymm1
+	vmovdqu	-(VEC_SIZE * 3)(%rsi), %ymm2
+	addq	%rdx, %rdi
+
+	VPCMPEQ	-(VEC_SIZE * 4)(%rdi), %ymm1, %ymm1
+	VPCMPEQ	-(VEC_SIZE * 3)(%rdi), %ymm2, %ymm2
+
+	vmovdqu	-(VEC_SIZE * 2)(%rsi), %ymm3
+	VPCMPEQ	-(VEC_SIZE * 2)(%rdi), %ymm3, %ymm3
+	vmovdqu	-VEC_SIZE(%rsi), %ymm4
+	VPCMPEQ	-VEC_SIZE(%rdi), %ymm4, %ymm4
+
+	/* Reduce VEC0 - VEC4.  */
+	vpand	%ymm1, %ymm2, %ymm2
+	vpand	%ymm3, %ymm4, %ymm4
+	vpand	%ymm2, %ymm4, %ymm4
+	vpmovmskb %ymm4, %eax
+	incl	%eax
+L(return_neq0):
+L(return_vzeroupper):
+	ZERO_UPPER_VEC_REGISTERS_RETURN
 
-#include "memcmp-avx2-movbe.S"
+	/* NB: p2align 5 here will ensure the L(loop_4x_vec) is also 32 byte
+	   aligned.  */
+	.p2align 5
+L(less_vec):
+	/* Check if one or less char. This is necessary for size = 0 but is
+	   also faster for size = 1.  */
+	cmpl	$1, %edx
+	jbe	L(one_or_less)
+
+	/* Check if loading one VEC from either s1 or s2 could cause a page
+	   cross. This can have false positives but is by far the fastest
+	   method.  */
+	movl	%edi, %eax
+	orl	%esi, %eax
+	andl	$(PAGE_SIZE - 1), %eax
+	cmpl	$(PAGE_SIZE - VEC_SIZE), %eax
+	jg	L(page_cross_less_vec)
+
+	/* No page cross possible.  */
+	vmovdqu	(%rsi), %ymm2
+	VPCMPEQ	(%rdi), %ymm2, %ymm2
+	vpmovmskb %ymm2, %eax
+	incl	%eax
+	/* Result will be zero if s1 and s2 match. Otherwise first set bit
+	   will be first mismatch.  */
+	bzhil	%edx, %eax, %eax
+	VZEROUPPER_RETURN
+
+	/* Relatively cold but placing close to L(less_vec) for 2 byte jump
+	   encoding.  */
+	.p2align 4
+L(one_or_less):
+	jb	L(zero)
+	movzbl	(%rsi), %ecx
+	movzbl	(%rdi), %eax
+	subl	%ecx, %eax
+	/* No ymm register was touched.  */
+	ret
+	/* Within the same 16 byte block is L(one_or_less).  */
+L(zero):
+	xorl	%eax, %eax
+	ret
+
+	.p2align 4
+L(last_1x_vec):
+	vmovdqu	-(VEC_SIZE * 1)(%rsi, %rdx), %ymm1
+	VPCMPEQ	-(VEC_SIZE * 1)(%rdi, %rdx), %ymm1, %ymm1
+	vpmovmskb %ymm1, %eax
+	incl	%eax
+	VZEROUPPER_RETURN
+
+	.p2align 4
+L(last_2x_vec):
+	vmovdqu	-(VEC_SIZE * 2)(%rsi, %rdx), %ymm1
+	VPCMPEQ	-(VEC_SIZE * 2)(%rdi, %rdx), %ymm1, %ymm1
+	vmovdqu	-(VEC_SIZE * 1)(%rsi, %rdx), %ymm2
+	VPCMPEQ	-(VEC_SIZE * 1)(%rdi, %rdx), %ymm2, %ymm2
+	vpand	%ymm1, %ymm2, %ymm2
+	vpmovmskb %ymm2, %eax
+	incl	%eax
+	VZEROUPPER_RETURN
+
+	.p2align 4
+L(more_8x_vec):
+	/* Set end of s1 in rdx.  */
+	leaq	-(VEC_SIZE * 4)(%rdi, %rdx), %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):
+	/* rsi has s2 - s1 so get correct address by adding s1 (in rdi).  */
+	vmovdqu	(%rsi, %rdi), %ymm1
+	VPCMPEQ	(%rdi), %ymm1, %ymm1
+
+	vmovdqu	VEC_SIZE(%rsi, %rdi), %ymm2
+	VPCMPEQ	VEC_SIZE(%rdi), %ymm2, %ymm2
+
+	vmovdqu	(VEC_SIZE * 2)(%rsi, %rdi), %ymm3
+	VPCMPEQ	(VEC_SIZE * 2)(%rdi), %ymm3, %ymm3
+
+	vmovdqu	(VEC_SIZE * 3)(%rsi, %rdi), %ymm4
+	VPCMPEQ	(VEC_SIZE * 3)(%rdi), %ymm4, %ymm4
+
+	vpand	%ymm1, %ymm2, %ymm2
+	vpand	%ymm3, %ymm4, %ymm4
+	vpand	%ymm2, %ymm4, %ymm4
+	vpmovmskb %ymm4, %eax
+	incl	%eax
+	jnz	L(return_neq1)
+	subq	$-(VEC_SIZE * 4), %rdi
+	/* Check if s1 pointer at end.  */
+	cmpq	%rdx, %rdi
+	jb	L(loop_4x_vec)
+
+	vmovdqu	(VEC_SIZE * 3)(%rsi, %rdx), %ymm4
+	VPCMPEQ	(VEC_SIZE * 3)(%rdx), %ymm4, %ymm4
+	subq	%rdx, %rdi
+	/* rdi has 4 * VEC_SIZE - remaining length.  */
+	cmpl	$(VEC_SIZE * 3), %edi
+	jae	L(8x_last_1x_vec)
+	/* Load regardless of branch.  */
+	vmovdqu	(VEC_SIZE * 2)(%rsi, %rdx), %ymm3
+	VPCMPEQ	(VEC_SIZE * 2)(%rdx), %ymm3, %ymm3
+	cmpl	$(VEC_SIZE * 2), %edi
+	jae	L(8x_last_2x_vec)
+	/* Check last 4 VEC.  */
+	vmovdqu	VEC_SIZE(%rsi, %rdx), %ymm1
+	VPCMPEQ	VEC_SIZE(%rdx), %ymm1, %ymm1
+
+	vmovdqu	(%rsi, %rdx), %ymm2
+	VPCMPEQ	(%rdx), %ymm2, %ymm2
+
+	vpand	%ymm3, %ymm4, %ymm4
+	vpand	%ymm1, %ymm2, %ymm3
+L(8x_last_2x_vec):
+	vpand	%ymm3, %ymm4, %ymm4
+L(8x_last_1x_vec):
+	vpmovmskb %ymm4, %eax
+	/* Restore s1 pointer to rdi.  */
+	incl	%eax
+L(return_neq1):
+	VZEROUPPER_RETURN
+
+	/* Relatively cold case as page cross are unexpected.  */
+	.p2align 4
+L(page_cross_less_vec):
+	cmpl	$16, %edx
+	jae	L(between_16_31)
+	cmpl	$8, %edx
+	ja	L(between_9_15)
+	cmpl	$4, %edx
+	jb	L(between_2_3)
+	/* From 4 to 8 bytes.  No branch when size == 4.  */
+	movl	(%rdi), %eax
+	subl	(%rsi), %eax
+	movl	-4(%rdi, %rdx), %ecx
+	movl	-4(%rsi, %rdx), %edi
+	subl	%edi, %ecx
+	orl	%ecx, %eax
+	ret
+
+	.p2align 4,, 8
+L(between_16_31):
+	/* From 16 to 31 bytes.  No branch when size == 16.  */
+
+	/* Safe to use xmm[0, 15] as no vzeroupper is needed so RTM safe.
+	 */
+	vmovdqu	(%rsi), %xmm1
+	vpcmpeqb (%rdi), %xmm1, %xmm1
+	vmovdqu	-16(%rsi, %rdx), %xmm2
+	vpcmpeqb -16(%rdi, %rdx), %xmm2, %xmm2
+	vpand	%xmm1, %xmm2, %xmm2
+	vpmovmskb %xmm2, %eax
+	notw	%ax
+	/* No ymm register was touched.  */
+	ret
+
+	.p2align 4,, 8
+L(between_9_15):
+	/* From 9 to 15 bytes.  */
+	movq	(%rdi), %rax
+	subq	(%rsi), %rax
+	movq	-8(%rdi, %rdx), %rcx
+	movq	-8(%rsi, %rdx), %rdi
+	subq	%rdi, %rcx
+	orq	%rcx, %rax
+	/* edx is guranteed to be a non-zero int.  */
+	cmovnz	%edx, %eax
+	ret
+
+	/* Don't align. This is cold and aligning here will cause code
+	   to spill into next cache line.  */
+L(between_2_3):
+	/* From 2 to 3 bytes.  No branch when size == 2.  */
+	movzwl	(%rdi), %eax
+	movzwl	(%rsi), %ecx
+	subl	%ecx, %eax
+	movzbl	-1(%rdi, %rdx), %ecx
+	/* All machines that support evex will insert a "merging uop"
+	   avoiding any serious partial register stalls.  */
+	subb	-1(%rsi, %rdx), %cl
+	orl	%ecx, %eax
+	/* No ymm register was touched.  */
+	ret
+
+    /* 2 Bytes from next cache line. */
+END (MEMCMPEQ)
+#endif
-- 
2.25.1