X86-64: Add _dl_runtime_resolve_avx[512]_{opt|slow} [BZ #20508]

There is transition penalty when SSE instructions are mixed with 256-bit
AVX or 512-bit AVX512 load instructions.  Since _dl_runtime_resolve_avx
and _dl_runtime_profile_avx512 save/restore 256-bit YMM/512-bit ZMM
registers, there is transition penalty when SSE instructions are used
with lazy binding on AVX and AVX512 processors.

To avoid SSE transition penalty, if only the lower 128 bits of the first
8 vector registers are non-zero, we can preserve %xmm0 - %xmm7 registers
with the zero upper bits.

For AVX and AVX512 processors which support XGETBV with ECX == 1, we can
use XGETBV with ECX == 1 to check if the upper 128 bits of YMM registers
or the upper 256 bits of ZMM registers are zero.  We can restore only the
non-zero portion of vector registers with AVX/AVX512 load instructions
which will zero-extend upper bits of vector registers.

This patch adds _dl_runtime_resolve_sse_vex which saves and restores
XMM registers with 128-bit AVX store/load instructions.  It is used to
preserve YMM/ZMM registers when only the lower 128 bits are non-zero.
_dl_runtime_resolve_avx_opt and _dl_runtime_resolve_avx512_opt are added
and used on AVX/AVX512 processors supporting XGETBV with ECX == 1 so
that we store and load only the non-zero portion of vector registers.
This avoids SSE transition penalty caused by _dl_runtime_resolve_avx and
_dl_runtime_profile_avx512 when only the lower 128 bits of vector
registers are used.

_dl_runtime_resolve_avx_slow is added and used for AVX processors which
don't support XGETBV with ECX == 1.  Since there is no SSE transition
penalty on AVX512 processors which don't support XGETBV with ECX == 1,
_dl_runtime_resolve_avx512_slow isn't provided.

	[BZ #20495]
	[BZ #20508]
	* sysdeps/x86/cpu-features.c (init_cpu_features): For Intel
	processors, set Use_dl_runtime_resolve_slow and set
	Use_dl_runtime_resolve_opt if XGETBV suports ECX == 1.
	* sysdeps/x86/cpu-features.h (bit_arch_Use_dl_runtime_resolve_opt):
	New.
	(bit_arch_Use_dl_runtime_resolve_slow): Likewise.
	(index_arch_Use_dl_runtime_resolve_opt): Likewise.
	(index_arch_Use_dl_runtime_resolve_slow): Likewise.
	* sysdeps/x86_64/dl-machine.h (elf_machine_runtime_setup): Use
	_dl_runtime_resolve_avx512_opt and _dl_runtime_resolve_avx_opt
	if Use_dl_runtime_resolve_opt is set.  Use
	_dl_runtime_resolve_slow if Use_dl_runtime_resolve_slow is set.
	* sysdeps/x86_64/dl-trampoline.S: Include <cpu-features.h>.
	(_dl_runtime_resolve_opt): New.  Defined for AVX and AVX512.
	(_dl_runtime_resolve): Add one for _dl_runtime_resolve_sse_vex.
	* sysdeps/x86_64/dl-trampoline.h (_dl_runtime_resolve_avx_slow):
	New.
	(_dl_runtime_resolve_opt): Likewise.
	(_dl_runtime_profile): Define only if _dl_runtime_profile is
	defined.

(cherry picked from commit fb0f7a6755c1bfaec38f490fbfcaa39a66ee3604)
(cherry picked from commit 4b8790c81c1a7b870a43810ec95e08a2e501123d)

5 files changed, 165 insertions, 3 deletions

diff --git a/sysdeps/x86/cpu-features.c b/sysdeps/x86/cpu-features.c
index 9ce4b495a5..11b9af2231 100644
--- a/sysdeps/x86/cpu-features.c
+++ b/sysdeps/x86/cpu-features.c
@@ -205,6 +205,20 @@ init_cpu_features (struct cpu_features *cpu_features)
       if (CPU_FEATURES_ARCH_P (cpu_features, AVX2_Usable))
 	cpu_features->feature[index_arch_AVX_Fast_Unaligned_Load]
 	  |= bit_arch_AVX_Fast_Unaligned_Load;
+
+      /* To avoid SSE transition penalty, use _dl_runtime_resolve_slow.
+         If XGETBV suports ECX == 1, use _dl_runtime_resolve_opt.  */
+      cpu_features->feature[index_arch_Use_dl_runtime_resolve_slow]
+	|= bit_arch_Use_dl_runtime_resolve_slow;
+      if (cpu_features->max_cpuid >= 0xd)
+	{
+	  unsigned int eax;
+
+	  __cpuid_count (0xd, 1, eax, ebx, ecx, edx);
+	  if ((eax & (1 << 2)) != 0)
+	    cpu_features->feature[index_arch_Use_dl_runtime_resolve_opt]
+	      |= bit_arch_Use_dl_runtime_resolve_opt;
+	}
     }
   /* This spells out "AuthenticAMD".  */
   else if (ebx == 0x68747541 && ecx == 0x444d4163 && edx == 0x69746e65)
diff --git a/sysdeps/x86/cpu-features.h b/sysdeps/x86/cpu-features.h
index 97ffe765f4..a8b5a734bd 100644
--- a/sysdeps/x86/cpu-features.h
+++ b/sysdeps/x86/cpu-features.h
@@ -37,6 +37,8 @@
 #define bit_arch_Prefer_No_VZEROUPPER		(1 << 17)
 #define bit_arch_Fast_Unaligned_Copy		(1 << 18)
 #define bit_arch_Prefer_ERMS			(1 << 19)
+#define bit_arch_Use_dl_runtime_resolve_opt	(1 << 20)
+#define bit_arch_Use_dl_runtime_resolve_slow	(1 << 21)
 
 /* CPUID Feature flags.  */
 
@@ -107,6 +109,8 @@
 # define index_arch_Prefer_No_VZEROUPPER FEATURE_INDEX_1*FEATURE_SIZE
 # define index_arch_Fast_Unaligned_Copy	FEATURE_INDEX_1*FEATURE_SIZE
 # define index_arch_Prefer_ERMS		FEATURE_INDEX_1*FEATURE_SIZE
+# define index_arch_Use_dl_runtime_resolve_opt FEATURE_INDEX_1*FEATURE_SIZE
+# define index_arch_Use_dl_runtime_resolve_slow FEATURE_INDEX_1*FEATURE_SIZE
 
 
 # if defined (_LIBC) && !IS_IN (nonlib)
@@ -277,6 +281,8 @@ extern const struct cpu_features *__get_cpu_features (void)
 # define index_arch_Prefer_No_VZEROUPPER FEATURE_INDEX_1
 # define index_arch_Fast_Unaligned_Copy	FEATURE_INDEX_1
 # define index_arch_Prefer_ERMS		FEATURE_INDEX_1
+# define index_arch_Use_dl_runtime_resolve_opt FEATURE_INDEX_1
+# define index_arch_Use_dl_runtime_resolve_slow FEATURE_INDEX_1
 
 #endif	/* !__ASSEMBLER__ */
 
diff --git a/sysdeps/x86_64/dl-machine.h b/sysdeps/x86_64/dl-machine.h
index ed0c1a8efd..c0f0fa16a2 100644
--- a/sysdeps/x86_64/dl-machine.h
+++ b/sysdeps/x86_64/dl-machine.h
@@ -68,7 +68,10 @@ elf_machine_runtime_setup (struct link_map *l, int lazy, int profile)
   Elf64_Addr *got;
   extern void _dl_runtime_resolve_sse (ElfW(Word)) attribute_hidden;
   extern void _dl_runtime_resolve_avx (ElfW(Word)) attribute_hidden;
+  extern void _dl_runtime_resolve_avx_slow (ElfW(Word)) attribute_hidden;
+  extern void _dl_runtime_resolve_avx_opt (ElfW(Word)) attribute_hidden;
   extern void _dl_runtime_resolve_avx512 (ElfW(Word)) attribute_hidden;
+  extern void _dl_runtime_resolve_avx512_opt (ElfW(Word)) attribute_hidden;
   extern void _dl_runtime_profile_sse (ElfW(Word)) attribute_hidden;
   extern void _dl_runtime_profile_avx (ElfW(Word)) attribute_hidden;
   extern void _dl_runtime_profile_avx512 (ElfW(Word)) attribute_hidden;
@@ -118,9 +121,26 @@ elf_machine_runtime_setup (struct link_map *l, int lazy, int profile)
 	     indicated by the offset on the stack, and then jump to
 	     the resolved address.  */
 	  if (HAS_ARCH_FEATURE (AVX512F_Usable))
-	    *(ElfW(Addr) *) (got + 2) = (ElfW(Addr)) &_dl_runtime_resolve_avx512;
+	    {
+	      if (HAS_ARCH_FEATURE (Use_dl_runtime_resolve_opt))
+		*(ElfW(Addr) *) (got + 2)
+		  = (ElfW(Addr)) &_dl_runtime_resolve_avx512_opt;
+	      else
+		*(ElfW(Addr) *) (got + 2)
+		  = (ElfW(Addr)) &_dl_runtime_resolve_avx512;
+	    }
 	  else if (HAS_ARCH_FEATURE (AVX_Usable))
-	    *(ElfW(Addr) *) (got + 2) = (ElfW(Addr)) &_dl_runtime_resolve_avx;
+	    {
+	      if (HAS_ARCH_FEATURE (Use_dl_runtime_resolve_opt))
+		*(ElfW(Addr) *) (got + 2)
+		  = (ElfW(Addr)) &_dl_runtime_resolve_avx_opt;
+	      else if (HAS_ARCH_FEATURE (Use_dl_runtime_resolve_slow))
+		*(ElfW(Addr) *) (got + 2)
+		  = (ElfW(Addr)) &_dl_runtime_resolve_avx_slow;
+	      else
+		*(ElfW(Addr) *) (got + 2)
+		  = (ElfW(Addr)) &_dl_runtime_resolve_avx;
+	    }
 	  else
 	    *(ElfW(Addr) *) (got + 2) = (ElfW(Addr)) &_dl_runtime_resolve_sse;
 	}
diff --git a/sysdeps/x86_64/dl-trampoline.S b/sysdeps/x86_64/dl-trampoline.S
index 12f1a5cf84..39f595e1e1 100644
--- a/sysdeps/x86_64/dl-trampoline.S
+++ b/sysdeps/x86_64/dl-trampoline.S
@@ -18,6 +18,7 @@
 
 #include <config.h>
 #include <sysdep.h>
+#include <cpu-features.h>
 #include <link-defines.h>
 
 #ifndef DL_STACK_ALIGNMENT
@@ -86,9 +87,11 @@
 #endif
 #define VEC(i)			zmm##i
 #define _dl_runtime_resolve	_dl_runtime_resolve_avx512
+#define _dl_runtime_resolve_opt	_dl_runtime_resolve_avx512_opt
 #define _dl_runtime_profile	_dl_runtime_profile_avx512
 #include "dl-trampoline.h"
 #undef _dl_runtime_resolve
+#undef _dl_runtime_resolve_opt
 #undef _dl_runtime_profile
 #undef VEC
 #undef VMOV
@@ -104,9 +107,11 @@
 #endif
 #define VEC(i)			ymm##i
 #define _dl_runtime_resolve	_dl_runtime_resolve_avx
+#define _dl_runtime_resolve_opt	_dl_runtime_resolve_avx_opt
 #define _dl_runtime_profile	_dl_runtime_profile_avx
 #include "dl-trampoline.h"
 #undef _dl_runtime_resolve
+#undef _dl_runtime_resolve_opt
 #undef _dl_runtime_profile
 #undef VEC
 #undef VMOV
@@ -126,3 +131,18 @@
 #define _dl_runtime_profile	_dl_runtime_profile_sse
 #undef RESTORE_AVX
 #include "dl-trampoline.h"
+#undef _dl_runtime_resolve
+#undef _dl_runtime_profile
+#undef VMOV
+#undef VMOVA
+
+/* Used by _dl_runtime_resolve_avx_opt/_dl_runtime_resolve_avx512_opt
+   to preserve the full vector registers with zero upper bits.  */
+#define VMOVA			vmovdqa
+#if DL_RUNTIME_RESOLVE_REALIGN_STACK || VEC_SIZE <= DL_STACK_ALIGNMENT
+# define VMOV			vmovdqa
+#else
+# define VMOV			vmovdqu
+#endif
+#define _dl_runtime_resolve	_dl_runtime_resolve_sse_vex
+#include "dl-trampoline.h"
diff --git a/sysdeps/x86_64/dl-trampoline.h b/sysdeps/x86_64/dl-trampoline.h
index b90836ab13..abe4471c1d 100644
--- a/sysdeps/x86_64/dl-trampoline.h
+++ b/sysdeps/x86_64/dl-trampoline.h
@@ -50,6 +50,105 @@
 #endif
 
 	.text
+#ifdef _dl_runtime_resolve_opt
+/* Use the smallest vector registers to preserve the full YMM/ZMM
+   registers to avoid SSE transition penalty.  */
+
+# if VEC_SIZE == 32
+/* Check if the upper 128 bits in %ymm0 - %ymm7 registers are non-zero
+   and preserve %xmm0 - %xmm7 registers with the zero upper bits.  Since
+   there is no SSE transition penalty on AVX512 processors which don't
+   support XGETBV with ECX == 1, _dl_runtime_resolve_avx512_slow isn't
+   provided.   */
+	.globl _dl_runtime_resolve_avx_slow
+	.hidden _dl_runtime_resolve_avx_slow
+	.type _dl_runtime_resolve_avx_slow, @function
+	.align 16
+_dl_runtime_resolve_avx_slow:
+	cfi_startproc
+	cfi_adjust_cfa_offset(16) # Incorporate PLT
+	vorpd %ymm0, %ymm1, %ymm8
+	vorpd %ymm2, %ymm3, %ymm9
+	vorpd %ymm4, %ymm5, %ymm10
+	vorpd %ymm6, %ymm7, %ymm11
+	vorpd %ymm8, %ymm9, %ymm9
+	vorpd %ymm10, %ymm11, %ymm10
+	vpcmpeqd %xmm8, %xmm8, %xmm8
+	vorpd %ymm9, %ymm10, %ymm10
+	vptest %ymm10, %ymm8
+	# Preserve %ymm0 - %ymm7 registers if the upper 128 bits of any
+	# %ymm0 - %ymm7 registers aren't zero.
+	PRESERVE_BND_REGS_PREFIX
+	jnc _dl_runtime_resolve_avx
+	# Use vzeroupper to avoid SSE transition penalty.
+	vzeroupper
+	# Preserve %xmm0 - %xmm7 registers with the zero upper 128 bits
+	# when the upper 128 bits of %ymm0 - %ymm7 registers are zero.
+	PRESERVE_BND_REGS_PREFIX
+	jmp _dl_runtime_resolve_sse_vex
+	cfi_adjust_cfa_offset(-16) # Restore PLT adjustment
+	cfi_endproc
+	.size _dl_runtime_resolve_avx_slow, .-_dl_runtime_resolve_avx_slow
+# endif
+
+/* Use XGETBV with ECX == 1 to check which bits in vector registers are
+   non-zero and only preserve the non-zero lower bits with zero upper
+   bits.  */
+	.globl _dl_runtime_resolve_opt
+	.hidden _dl_runtime_resolve_opt
+	.type _dl_runtime_resolve_opt, @function
+	.align 16
+_dl_runtime_resolve_opt:
+	cfi_startproc
+	cfi_adjust_cfa_offset(16) # Incorporate PLT
+	pushq %rax
+	cfi_adjust_cfa_offset(8)
+	cfi_rel_offset(%rax, 0)
+	pushq %rcx
+	cfi_adjust_cfa_offset(8)
+	cfi_rel_offset(%rcx, 0)
+	pushq %rdx
+	cfi_adjust_cfa_offset(8)
+	cfi_rel_offset(%rdx, 0)
+	movl $1, %ecx
+	xgetbv
+	movl %eax, %r11d
+	popq %rdx
+	cfi_adjust_cfa_offset(-8)
+	cfi_restore (%rdx)
+	popq %rcx
+	cfi_adjust_cfa_offset(-8)
+	cfi_restore (%rcx)
+	popq %rax
+	cfi_adjust_cfa_offset(-8)
+	cfi_restore (%rax)
+# if VEC_SIZE == 32
+	# For YMM registers, check if YMM state is in use.
+	andl $bit_YMM_state, %r11d
+	# Preserve %xmm0 - %xmm7 registers with the zero upper 128 bits if
+	# YMM state isn't in use.
+	PRESERVE_BND_REGS_PREFIX
+	jz _dl_runtime_resolve_sse_vex
+# elif VEC_SIZE == 64
+	# For ZMM registers, check if YMM state and ZMM state are in
+	# use.
+	andl $(bit_YMM_state | bit_ZMM0_15_state), %r11d
+	cmpl $bit_YMM_state, %r11d
+	# Preserve %xmm0 - %xmm7 registers with the zero upper 384 bits if
+	# neither YMM state nor ZMM state are in use.
+	PRESERVE_BND_REGS_PREFIX
+	jl _dl_runtime_resolve_sse_vex
+	# Preserve %ymm0 - %ymm7 registers with the zero upper 256 bits if
+	# ZMM state isn't in use.
+	PRESERVE_BND_REGS_PREFIX
+	je _dl_runtime_resolve_avx
+# else
+#  error Unsupported VEC_SIZE!
+# endif
+	cfi_adjust_cfa_offset(-16) # Restore PLT adjustment
+	cfi_endproc
+	.size _dl_runtime_resolve_opt, .-_dl_runtime_resolve_opt
+#endif
 	.globl _dl_runtime_resolve
 	.hidden _dl_runtime_resolve
 	.type _dl_runtime_resolve, @function
@@ -162,7 +261,10 @@ _dl_runtime_resolve:
 	.size _dl_runtime_resolve, .-_dl_runtime_resolve
 
 
-#ifndef PROF
+/* To preserve %xmm0 - %xmm7 registers, dl-trampoline.h is included
+   twice, for _dl_runtime_resolve_sse and _dl_runtime_resolve_sse_vex.
+   But we don't need another _dl_runtime_profile for XMM registers.  */
+#if !defined PROF && defined _dl_runtime_profile
 # if (LR_VECTOR_OFFSET % VEC_SIZE) != 0
 #  error LR_VECTOR_OFFSET must be multples of VEC_SIZE
 # endif