From 3d5101ddb7a4004459ca3f894caa47cfe9208be6 Mon Sep 17 00:00:00 2001

From: "H.J. Lu" <hjl.tools@gmail.com>

Date: Fri, 5 Mar 2021 07:15:03 -0800

Subject: [PATCH] x86-64: Add memset family functions with 256-bit EVEX

Update ifunc-memset.h/ifunc-wmemset.h to select the function optimized

with 256-bit EVEX instructions using YMM16-YMM31 registers to avoid RTM

abort with usable AVX512VL and AVX512BW since VZEROUPPER isn't needed at

function exit.

(cherry picked from commit 1b968b6b9b3aac702ac2f133e0dd16cfdbb415ee)

---

 sysdeps/x86_64/multiarch/Makefile             |  1 +

 sysdeps/x86_64/multiarch/ifunc-impl-list.c    | 22 +++++++++++++++++

 sysdeps/x86_64/multiarch/ifunc-memset.h       | 24 +++++++++++++++----

 sysdeps/x86_64/multiarch/ifunc-wmemset.h      | 13 ++++++----

 .../multiarch/memset-evex-unaligned-erms.S    | 24 +++++++++++++++++++

 .../multiarch/memset-vec-unaligned-erms.S     | 20 +++++++++++-----

 6 files changed, 90 insertions(+), 14 deletions(-)

 create mode 100644 sysdeps/x86_64/multiarch/memset-evex-unaligned-erms.S

diff --git a/sysdeps/x86_64/multiarch/Makefile b/sysdeps/x86_64/multiarch/Makefile

index 4563fc56..1cc0a10e 100644

--- a/sysdeps/x86_64/multiarch/Makefile

+++ b/sysdeps/x86_64/multiarch/Makefile

@@ -43,6 +43,7 @@ sysdep_routines += strncat-c stpncpy-c strncpy-c \

 		   memchr-evex \

 		   memmove-evex-unaligned-erms \

 		   memrchr-evex \

+		   memset-evex-unaligned-erms \

 		   rawmemchr-evex \

 		   stpcpy-evex \

 		   stpncpy-evex \

diff --git a/sysdeps/x86_64/multiarch/ifunc-impl-list.c b/sysdeps/x86_64/multiarch/ifunc-impl-list.c

index 6bd3abfc..7cf83485 100644

--- a/sysdeps/x86_64/multiarch/ifunc-impl-list.c

+++ b/sysdeps/x86_64/multiarch/ifunc-impl-list.c

@@ -160,6 +160,14 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,

 	      IFUNC_IMPL_ADD (array, i, __memset_chk,

 			      CPU_FEATURE_USABLE (AVX2),

 			      __memset_chk_avx2_unaligned_erms)

+	      IFUNC_IMPL_ADD (array, i, __memset_chk,

+			      (CPU_FEATURE_USABLE (AVX512VL)

+			       && CPU_FEATURE_USABLE (AVX512BW)),

+			      __memset_chk_evex_unaligned)

+	      IFUNC_IMPL_ADD (array, i, __memset_chk,

+			      (CPU_FEATURE_USABLE (AVX512VL)

+			       && CPU_FEATURE_USABLE (AVX512BW)),

+			      __memset_chk_evex_unaligned_erms)

 	      IFUNC_IMPL_ADD (array, i, __memset_chk,

 			      CPU_FEATURE_USABLE (AVX512F),

 			      __memset_chk_avx512_unaligned_erms)

@@ -185,6 +193,14 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,

 	      IFUNC_IMPL_ADD (array, i, memset,

 			      CPU_FEATURE_USABLE (AVX2),

 			      __memset_avx2_unaligned_erms)

+	      IFUNC_IMPL_ADD (array, i, memset,

+			      (CPU_FEATURE_USABLE (AVX512VL)

+			       && CPU_FEATURE_USABLE (AVX512BW)),

+			      __memset_evex_unaligned)

+	      IFUNC_IMPL_ADD (array, i, memset,

+			      (CPU_FEATURE_USABLE (AVX512VL)

+			       && CPU_FEATURE_USABLE (AVX512BW)),

+			      __memset_evex_unaligned_erms)

 	      IFUNC_IMPL_ADD (array, i, memset,

 			      CPU_FEATURE_USABLE (AVX512F),

 			      __memset_avx512_unaligned_erms)

@@ -555,6 +571,9 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,

 	      IFUNC_IMPL_ADD (array, i, wmemset,

 			      CPU_FEATURE_USABLE (AVX2),

 			      __wmemset_avx2_unaligned)

+	      IFUNC_IMPL_ADD (array, i, wmemset,

+			      CPU_FEATURE_USABLE (AVX512VL),

+			      __wmemset_evex_unaligned)

 	      IFUNC_IMPL_ADD (array, i, wmemset,

 			      CPU_FEATURE_USABLE (AVX512F),

 			      __wmemset_avx512_unaligned))

@@ -723,6 +742,9 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,

 	      IFUNC_IMPL_ADD (array, i, __wmemset_chk,

 			      CPU_FEATURE_USABLE (AVX2),

 			      __wmemset_chk_avx2_unaligned)

+	      IFUNC_IMPL_ADD (array, i, __wmemset_chk,

+			      CPU_FEATURE_USABLE (AVX512VL),

+			      __wmemset_chk_evex_unaligned)

 	      IFUNC_IMPL_ADD (array, i, __wmemset_chk,

 			      CPU_FEATURE_USABLE (AVX512F),

 			      __wmemset_chk_avx512_unaligned))

diff --git a/sysdeps/x86_64/multiarch/ifunc-memset.h b/sysdeps/x86_64/multiarch/ifunc-memset.h

index 708bd72e..6f31f4dc 100644

--- a/sysdeps/x86_64/multiarch/ifunc-memset.h

+++ b/sysdeps/x86_64/multiarch/ifunc-memset.h

@@ -27,6 +27,10 @@ extern __typeof (REDIRECT_NAME) OPTIMIZE (sse2_unaligned_erms)

 extern __typeof (REDIRECT_NAME) OPTIMIZE (avx2_unaligned) attribute_hidden;

 extern __typeof (REDIRECT_NAME) OPTIMIZE (avx2_unaligned_erms)

   attribute_hidden;

+extern __typeof (REDIRECT_NAME) OPTIMIZE (evex_unaligned)

+  attribute_hidden;

+extern __typeof (REDIRECT_NAME) OPTIMIZE (evex_unaligned_erms)

+  attribute_hidden;

 extern __typeof (REDIRECT_NAME) OPTIMIZE (avx512_unaligned)

   attribute_hidden;

 extern __typeof (REDIRECT_NAME) OPTIMIZE (avx512_unaligned_erms)

@@ -56,10 +60,22 @@ IFUNC_SELECTOR (void)

   if (CPU_FEATURE_USABLE_P (cpu_features, AVX2))

     {

-      if (CPU_FEATURE_USABLE_P (cpu_features, ERMS))

-	return OPTIMIZE (avx2_unaligned_erms);

-      else

-	return OPTIMIZE (avx2_unaligned);

+      if (CPU_FEATURE_USABLE_P (cpu_features, AVX512VL)

+	  && CPU_FEATURE_USABLE_P (cpu_features, AVX512BW))

+	{

+	  if (CPU_FEATURE_USABLE_P (cpu_features, ERMS))

+	    return OPTIMIZE (evex_unaligned_erms);

+

+	  return OPTIMIZE (evex_unaligned);

+	}

+

+      if (!CPU_FEATURES_ARCH_P (cpu_features, Prefer_No_VZEROUPPER))

+	{

+	  if (CPU_FEATURE_USABLE_P (cpu_features, ERMS))

+	    return OPTIMIZE (avx2_unaligned_erms);

+

+	  return OPTIMIZE (avx2_unaligned);

+	}

     }

   if (CPU_FEATURE_USABLE_P (cpu_features, ERMS))

diff --git a/sysdeps/x86_64/multiarch/ifunc-wmemset.h b/sysdeps/x86_64/multiarch/ifunc-wmemset.h

index eb242210..9290c4bf 100644

--- a/sysdeps/x86_64/multiarch/ifunc-wmemset.h

+++ b/sysdeps/x86_64/multiarch/ifunc-wmemset.h

@@ -20,6 +20,7 @@

 extern __typeof (REDIRECT_NAME) OPTIMIZE (sse2_unaligned) attribute_hidden;

 extern __typeof (REDIRECT_NAME) OPTIMIZE (avx2_unaligned) attribute_hidden;

+extern __typeof (REDIRECT_NAME) OPTIMIZE (evex_unaligned) attribute_hidden;

 extern __typeof (REDIRECT_NAME) OPTIMIZE (avx512_unaligned) attribute_hidden;

 static inline void *

@@ -27,14 +28,18 @@ 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_FEATURES_ARCH_P (cpu_features, AVX_Fast_Unaligned_Load))

     {

       if (CPU_FEATURE_USABLE_P (cpu_features, AVX512F)

-	  && !CPU_FEATURES_ARCH_P (cpu_features, Prefer_No_AVX512))

+	  && !CPU_FEATURES_ARCH_P (cpu_features, Prefer_No_AVX512)

+	  && !CPU_FEATURES_ARCH_P (cpu_features, Prefer_No_VZEROUPPER))

 	return OPTIMIZE (avx512_unaligned);

-      else

+

+      if (CPU_FEATURE_USABLE_P (cpu_features, AVX512VL))

+	return OPTIMIZE (evex_unaligned);

+

+      if (!CPU_FEATURES_ARCH_P (cpu_features, Prefer_No_VZEROUPPER))

 	return OPTIMIZE (avx2_unaligned);

     }

diff --git a/sysdeps/x86_64/multiarch/memset-evex-unaligned-erms.S b/sysdeps/x86_64/multiarch/memset-evex-unaligned-erms.S

new file mode 100644

index 00000000..ae0a4d6e

--- /dev/null

+++ b/sysdeps/x86_64/multiarch/memset-evex-unaligned-erms.S

@@ -0,0 +1,24 @@

+#if IS_IN (libc)

+# define VEC_SIZE	32

+# define XMM0		xmm16

+# define YMM0		ymm16

+# define VEC0		ymm16

+# define VEC(i)		VEC##i

+# define VMOVU		vmovdqu64

+# define VMOVA		vmovdqa64

+# define VZEROUPPER

+

+# define MEMSET_VDUP_TO_VEC0_AND_SET_RETURN(d, r) \

+  movq r, %rax; \

+  vpbroadcastb d, %VEC0

+

+# define WMEMSET_VDUP_TO_VEC0_AND_SET_RETURN(d, r) \

+  movq r, %rax; \

+  vpbroadcastd d, %VEC0

+

+# define SECTION(p)		p##.evex

+# define MEMSET_SYMBOL(p,s)	p##_evex_##s

+# define WMEMSET_SYMBOL(p,s)	p##_evex_##s

+

+# include "memset-vec-unaligned-erms.S"

+#endif

diff --git a/sysdeps/x86_64/multiarch/memset-vec-unaligned-erms.S b/sysdeps/x86_64/multiarch/memset-vec-unaligned-erms.S

index 9a0fd818..71e91a8f 100644

--- a/sysdeps/x86_64/multiarch/memset-vec-unaligned-erms.S

+++ b/sysdeps/x86_64/multiarch/memset-vec-unaligned-erms.S

@@ -34,6 +34,14 @@

 # define WMEMSET_CHK_SYMBOL(p,s)	WMEMSET_SYMBOL(p, s)

 #endif

+#ifndef XMM0

+# define XMM0				xmm0

+#endif

+

+#ifndef YMM0

+# define YMM0				ymm0

+#endif

+

 #ifndef VZEROUPPER

 # if VEC_SIZE > 16

 #  define VZEROUPPER			vzeroupper

@@ -67,7 +75,7 @@

 ENTRY (__bzero)

 	mov	%RDI_LP, %RAX_LP /* Set return value.  */

 	mov	%RSI_LP, %RDX_LP /* Set n.  */

-	pxor	%xmm0, %xmm0

+	pxor	%XMM0, %XMM0

 	jmp	L(entry_from_bzero)

 END (__bzero)

 weak_alias (__bzero, bzero)

@@ -223,7 +231,7 @@ L(less_vec):

 	cmpb	$16, %dl

 	jae	L(between_16_31)

 # endif

-	MOVQ	%xmm0, %rcx

+	MOVQ	%XMM0, %rcx

 	cmpb	$8, %dl

 	jae	L(between_8_15)

 	cmpb	$4, %dl

@@ -238,16 +246,16 @@ L(less_vec):

 # if VEC_SIZE > 32

 	/* From 32 to 63.  No branch when size == 32.  */

 L(between_32_63):

-	vmovdqu	%ymm0, -32(%rdi,%rdx)

-	vmovdqu	%ymm0, (%rdi)

+	VMOVU	%YMM0, -32(%rdi,%rdx)

+	VMOVU	%YMM0, (%rdi)

 	VZEROUPPER

 	ret

 # endif

 # if VEC_SIZE > 16

 	/* From 16 to 31.  No branch when size == 16.  */

 L(between_16_31):

-	vmovdqu	%xmm0, -16(%rdi,%rdx)

-	vmovdqu	%xmm0, (%rdi)

+	VMOVU	%XMM0, -16(%rdi,%rdx)

+	VMOVU	%XMM0, (%rdi)

 	VZEROUPPER

 	ret

 # endif

-- 

GitLab