[Headers][X86] amxintrin.h - fix attributes according to Intel SDM

[e-kud]

tileloadd, tileloaddt1 and tilestored are part of amx-tile feature.

The problem is observed if __tile_loadd intrinsic is invoked, _tile_loadd_internal requiring amx-int8 is inlined into __tile_loadd that has only amx-tile.

[e-kud]

I'm not sure if tests are needed. Let me know.

[llvmbot]

@llvm/pr-subscribers-backend-x86

Author: Evgenii Kudriashov (e-kud)

Changes

tileloadd, tileloaddt1 and tilestored are part of amx-tile feature.

The problem is observed if __tile_loadd intrinsic is invoked, _tile_loadd_internal requiring amx-int8 is inlined into __tile_loadd that has only amx-tile.

---

Full diff: https://github.com/llvm/llvm-project/pull/122204.diff

1 Files Affected:

• (modified) clang/lib/Headers/amxintrin.h (+3-3)

diff --git a/clang/lib/Headers/amxintrin.h b/clang/lib/Headers/amxintrin.h
index b0140615677f27..a7da10d9951e7e 100644
--- a/clang/lib/Headers/amxintrin.h
+++ b/clang/lib/Headers/amxintrin.h
@@ -234,7 +234,7 @@ typedef int _tile1024i_1024a
     __attribute__((__vector_size__(1024), __aligned__(1024)));
 
 /// This is internal intrinsic. C/C++ user should avoid calling it directly.
-static __inline__ _tile1024i __DEFAULT_FN_ATTRS_INT8
+static __inline__ _tile1024i __DEFAULT_FN_ATTRS_TILE
 _tile_loadd_internal(unsigned short m, unsigned short n, const void *base,
                      __SIZE_TYPE__ stride) {
   return __builtin_ia32_tileloadd64_internal(m, n, base,
@@ -242,7 +242,7 @@ _tile_loadd_internal(unsigned short m, unsigned short n, const void *base,
 }
 
 /// This is internal intrinsic. C/C++ user should avoid calling it directly.
-static __inline__ _tile1024i __DEFAULT_FN_ATTRS_INT8
+static __inline__ _tile1024i __DEFAULT_FN_ATTRS_TILE
 _tile_loaddt1_internal(unsigned short m, unsigned short n, const void *base,
                        __SIZE_TYPE__ stride) {
   return __builtin_ia32_tileloaddt164_internal(m, n, base,
@@ -278,7 +278,7 @@ _tile_dpbuud_internal(unsigned short m, unsigned short n, unsigned short k,
 }
 
 /// This is internal intrinsic. C/C++ user should avoid calling it directly.
-static __inline__ void __DEFAULT_FN_ATTRS_INT8
+static __inline__ void __DEFAULT_FN_ATTRS_TILE
 _tile_stored_internal(unsigned short m, unsigned short n, void *base,
                       __SIZE_TYPE__ stride, _tile1024i tile) {
   return __builtin_ia32_tilestored64_internal(m, n, base,

[llvmbot]

@llvm/pr-subscribers-clang

Author: Evgenii Kudriashov (e-kud)

Changes

tileloadd, tileloaddt1 and tilestored are part of amx-tile feature.

The problem is observed if __tile_loadd intrinsic is invoked, _tile_loadd_internal requiring amx-int8 is inlined into __tile_loadd that has only amx-tile.

---

Full diff: https://github.com/llvm/llvm-project/pull/122204.diff

1 Files Affected:

• (modified) clang/lib/Headers/amxintrin.h (+3-3)

diff --git a/clang/lib/Headers/amxintrin.h b/clang/lib/Headers/amxintrin.h
index b0140615677f27..a7da10d9951e7e 100644
--- a/clang/lib/Headers/amxintrin.h
+++ b/clang/lib/Headers/amxintrin.h
@@ -234,7 +234,7 @@ typedef int _tile1024i_1024a
     __attribute__((__vector_size__(1024), __aligned__(1024)));
 
 /// This is internal intrinsic. C/C++ user should avoid calling it directly.
-static __inline__ _tile1024i __DEFAULT_FN_ATTRS_INT8
+static __inline__ _tile1024i __DEFAULT_FN_ATTRS_TILE
 _tile_loadd_internal(unsigned short m, unsigned short n, const void *base,
                      __SIZE_TYPE__ stride) {
   return __builtin_ia32_tileloadd64_internal(m, n, base,
@@ -242,7 +242,7 @@ _tile_loadd_internal(unsigned short m, unsigned short n, const void *base,
 }
 
 /// This is internal intrinsic. C/C++ user should avoid calling it directly.
-static __inline__ _tile1024i __DEFAULT_FN_ATTRS_INT8
+static __inline__ _tile1024i __DEFAULT_FN_ATTRS_TILE
 _tile_loaddt1_internal(unsigned short m, unsigned short n, const void *base,
                        __SIZE_TYPE__ stride) {
   return __builtin_ia32_tileloaddt164_internal(m, n, base,
@@ -278,7 +278,7 @@ _tile_dpbuud_internal(unsigned short m, unsigned short n, unsigned short k,
 }
 
 /// This is internal intrinsic. C/C++ user should avoid calling it directly.
-static __inline__ void __DEFAULT_FN_ATTRS_INT8
+static __inline__ void __DEFAULT_FN_ATTRS_TILE
 _tile_stored_internal(unsigned short m, unsigned short n, void *base,
                       __SIZE_TYPE__ stride, _tile1024i tile) {
   return __builtin_ia32_tilestored64_internal(m, n, base,

[phoebewang]

Maybe split out amx-tile only tests from clang/test/CodeGen/X86/amx_api.c as regression test?