libatomic: Improve 16-byte atomics on Intel AVX [PR104688]

As mentioned in the PR, the latest Intel SDM has added:
"Processors that enumerate support for Intel® AVX (by setting the feature flag CPUID.01H:ECX.AVX[bit 28])
guarantee that the 16-byte memory operations performed by the following instructions will always be
carried out atomically:
• MOVAPD, MOVAPS, and MOVDQA.
• VMOVAPD, VMOVAPS, and VMOVDQA when encoded with VEX.128.
• VMOVAPD, VMOVAPS, VMOVDQA32, and VMOVDQA64 when encoded with EVEX.128 and k0 (masking disabled).
(Note that these instructions require the linear addresses of their memory operands to be 16-byte
aligned.)"

The following patch deals with it just on the libatomic library side so far,
currently (since ~ 2017) we emit all the __atomic_* 16-byte builtins as
library calls since and this is something that we can hopefully backport.

The patch simply introduces yet another ifunc variant that takes priority
over the pure CMPXCHG16B one, one that checks AVX and CMPXCHG16B bits and
on non-Intel clears the AVX bit during detection for now (if AMD comes
with the same guarantee, we could revert the config/x86/init.c hunk),
which implements 16-byte atomic load as vmovdqa and 16-byte atomic store
as vmovdqa followed by mfence.

2022-03-17  Jakub Jelinek  <jakub@redhat.com>

	PR target/104688
	* Makefile.am (IFUNC_OPTIONS): Change on x86_64 to -mcx16 -mcx16.
	(libatomic_la_LIBADD): Add $(addsuffix _16_2_.lo,$(SIZEOBJS)) for
	x86_64.
	* Makefile.in: Regenerated.
	* config/x86/host-config.h (IFUNC_COND_1): For x86_64 define to
	both AVX and CMPXCHG16B bits.
	(IFUNC_COND_2): Define.
	(IFUNC_NCOND): For x86_64 define to 2 * (N == 16).
	(MAYBE_HAVE_ATOMIC_CAS_16, MAYBE_HAVE_ATOMIC_EXCHANGE_16,
	MAYBE_HAVE_ATOMIC_LDST_16): Define to IFUNC_COND_2 rather than
	IFUNC_COND_1.
	(HAVE_ATOMIC_CAS_16): Redefine to 1 whenever IFUNC_ALT != 0.
	(HAVE_ATOMIC_LDST_16): Redefine to 1 whenever IFUNC_ALT == 1.
	(atomic_compare_exchange_n): Define whenever IFUNC_ALT != 0
	on x86_64 for N == 16.
	(__atomic_load_n, __atomic_store_n): Redefine whenever IFUNC_ALT == 1
	on x86_64 for N == 16.
	(atomic_load_n, atomic_store_n): New functions.
	* config/x86/init.c (__libat_feat1_init): On x86_64 clear bit_AVX
	if CPU vendor is not Intel.

libatomic/Makefile.am

@@ -138,8 +138,9 @@ IFUNC_OPTIONS	     = -march=i586
 libatomic_la_LIBADD += $(addsuffix _8_1_.lo,$(SIZEOBJS))
 endif
 if ARCH_X86_64
-IFUNC_OPTIONS	     = -mcx16
+IFUNC_OPTIONS	     = -mcx16 -mcx16
-libatomic_la_LIBADD += $(addsuffix _16_1_.lo,$(SIZEOBJS))
+libatomic_la_LIBADD += $(addsuffix _16_1_.lo,$(SIZEOBJS)) \
+		       $(addsuffix _16_2_.lo,$(SIZEOBJS))
 endif
 endif
 

libatomic/Makefile.in

@@ -96,7 +96,9 @@ target_triplet = @target@
 @ARCH_ARM_LINUX_TRUE@@HAVE_IFUNC_TRUE@	$(addsuffix \
 @ARCH_ARM_LINUX_TRUE@@HAVE_IFUNC_TRUE@	_8_2_.lo,$(SIZEOBJS))
 @ARCH_I386_TRUE@@HAVE_IFUNC_TRUE@am__append_3 = $(addsuffix _8_1_.lo,$(SIZEOBJS))
-@ARCH_X86_64_TRUE@@HAVE_IFUNC_TRUE@am__append_4 = $(addsuffix _16_1_.lo,$(SIZEOBJS))
+@ARCH_X86_64_TRUE@@HAVE_IFUNC_TRUE@am__append_4 = $(addsuffix _16_1_.lo,$(SIZEOBJS)) \
+@ARCH_X86_64_TRUE@@HAVE_IFUNC_TRUE@		       $(addsuffix _16_2_.lo,$(SIZEOBJS))
+
 subdir = .
 ACLOCAL_M4 = $(top_srcdir)/aclocal.m4
 am__aclocal_m4_deps = $(top_srcdir)/../config/acx.m4 \
@@ -435,7 +437,7 @@ libatomic_la_LIBADD = $(foreach s,$(SIZES),$(addsuffix \
 @ARCH_AARCH64_LINUX_TRUE@@HAVE_IFUNC_TRUE@IFUNC_OPTIONS = -march=armv8-a+lse
 @ARCH_ARM_LINUX_TRUE@@HAVE_IFUNC_TRUE@IFUNC_OPTIONS = -march=armv7-a+fp -DHAVE_KERNEL64
 @ARCH_I386_TRUE@@HAVE_IFUNC_TRUE@IFUNC_OPTIONS = -march=i586
-@ARCH_X86_64_TRUE@@HAVE_IFUNC_TRUE@IFUNC_OPTIONS = -mcx16
+@ARCH_X86_64_TRUE@@HAVE_IFUNC_TRUE@IFUNC_OPTIONS = -mcx16 -mcx16
 libatomic_convenience_la_SOURCES = $(libatomic_la_SOURCES)
 libatomic_convenience_la_LIBADD = $(libatomic_la_LIBADD)
 MULTISRCTOP = 

libatomic/config/x86/host-config.h

@@ -55,31 +55,37 @@ load_feat1 (void)
 }
 
 #ifdef __x86_64__
-# define IFUNC_COND_1	(load_feat1 () & bit_CMPXCHG16B)
+# define IFUNC_COND_1	((load_feat1 () & (bit_AVX | bit_CMPXCHG16B)) \
+			 == (bit_AVX | bit_CMPXCHG16B))
+# define IFUNC_COND_2	(load_feat1 () & bit_CMPXCHG16B)
 #else
 # define IFUNC_COND_1	(load_feat1 () & bit_CMPXCHG8B)
 #endif
 
 #ifdef __x86_64__
-# define IFUNC_NCOND(N) (N == 16)
+# define IFUNC_NCOND(N) (2 * (N == 16))
 #else
 # define IFUNC_NCOND(N) (N == 8)
 #endif
 
 #ifdef __x86_64__
 # undef MAYBE_HAVE_ATOMIC_CAS_16
-# define MAYBE_HAVE_ATOMIC_CAS_16	IFUNC_COND_1
+# define MAYBE_HAVE_ATOMIC_CAS_16	IFUNC_COND_2
 # undef MAYBE_HAVE_ATOMIC_EXCHANGE_16
-# define MAYBE_HAVE_ATOMIC_EXCHANGE_16	IFUNC_COND_1
+# define MAYBE_HAVE_ATOMIC_EXCHANGE_16	IFUNC_COND_2
 # undef MAYBE_HAVE_ATOMIC_LDST_16
-# define MAYBE_HAVE_ATOMIC_LDST_16	IFUNC_COND_1
+# define MAYBE_HAVE_ATOMIC_LDST_16	IFUNC_COND_2
 /* Since load and store are implemented with CAS, they are not fast.  */
 # undef FAST_ATOMIC_LDST_16
 # define FAST_ATOMIC_LDST_16		0
-# if IFUNC_ALT == 1
+# if IFUNC_ALT != 0
 #  undef HAVE_ATOMIC_CAS_16
 #  define HAVE_ATOMIC_CAS_16 1
 # endif
+# if IFUNC_ALT == 1
+#  undef HAVE_ATOMIC_LDST_16
+#  define HAVE_ATOMIC_LDST_16 1
+# endif
 #else
 # undef MAYBE_HAVE_ATOMIC_CAS_8
 # define MAYBE_HAVE_ATOMIC_CAS_8	IFUNC_COND_1
@@ -93,7 +99,7 @@ load_feat1 (void)
 # endif
 #endif
 
-#if defined(__x86_64__) && N == 16 && IFUNC_ALT == 1
+#if defined(__x86_64__) && N == 16 && IFUNC_ALT != 0
 static inline bool
 atomic_compare_exchange_n (UTYPE *mptr, UTYPE *eptr, UTYPE newval,
                            bool weak_p UNUSED, int sm UNUSED, int fm UNUSED)
@@ -108,6 +114,29 @@ atomic_compare_exchange_n (UTYPE *mptr, UTYPE *eptr, UTYPE newval,
 # define atomic_compare_exchange_n atomic_compare_exchange_n
 #endif /* Have CAS 16 */
 
+#if defined(__x86_64__) && N == 16 && IFUNC_ALT == 1
+#define __atomic_load_n(ptr, model) \
+  (sizeof (*ptr) == 16 ? atomic_load_n (ptr, model) \
+		       : (__atomic_load_n) (ptr, model))
+#define __atomic_store_n(ptr, val, model) \
+  (sizeof (*ptr) == 16 ? atomic_store_n (ptr, val, model) \
+		       : (__atomic_store_n) (ptr, val, model))
+
+static inline UTYPE
+atomic_load_n (UTYPE *ptr, int model UNUSED)
+{
+  UTYPE ret;
+  __asm__ ("vmovdqa\t{%1, %0|%0, %1}" : "=x" (ret) : "m" (*ptr));
+  return ret;
+}
+
+static inline void
+atomic_store_n (UTYPE *ptr, UTYPE val, int model UNUSED)
+{
+  __asm__ ("vmovdqa\t{%1, %0|%0, %1}\n\tmfence" : "=m" (*ptr) : "x" (val));
+}
+#endif
+
 #endif /* HAVE_IFUNC */
 
 #include_next <host-config.h>

libatomic/config/x86/init.c

@@ -34,6 +34,18 @@ __libat_feat1_init (void)
   unsigned int eax, ebx, ecx, edx;
   FEAT1_REGISTER = 0;
   __get_cpuid (1, &eax, &ebx, &ecx, &edx);
+#ifdef __x86_64__
+  if ((FEAT1_REGISTER & (bit_AVX | bit_CMPXCHG16B))
+      == (bit_AVX | bit_CMPXCHG16B))
+    {
+      /* Intel SDM guarantees that 16-byte VMOVDQA on 16-byte aligned address
+	 is atomic, but so far we don't have this guarantee from AMD.  */
+      unsigned int ecx2 = 0;
+      __get_cpuid (0, &eax, &ebx, &ecx2, &edx);
+      if (ecx2 != signature_INTEL_ecx)
+	FEAT1_REGISTER &= ~bit_AVX;
+    }
+#endif
   /* See the load in load_feat1.  */
   __atomic_store_n (&__libat_feat1, FEAT1_REGISTER, __ATOMIC_RELAXED);
   return FEAT1_REGISTER;