Skip to content

Commit cea9b9d

Browse files
BaiXilinmahesh-attarde
BaiXilin
authored and
mahesh-attarde
committed
[x86][AVX-VNNI] Fix VPDPBXXD Argument Type (llvm#159222)
Fixed intrinsic VPDP[SS,SU,UU]D[,S]_128/256/512's argument types to match with the ISA. Fixes part of llvm#97271.
1 parent c0ae86d commit cea9b9d

File tree

18 files changed

+1586
-807
lines changed

18 files changed

+1586
-807
lines changed

‎clang/include/clang/Basic/BuiltinsX86.td‎

Lines changed: 18 additions & 18 deletions
Original file line numberDiff line numberDiff line change
@@ -1109,51 +1109,51 @@ let Features = "avx512vnni", Attributes = [NoThrow, Const, RequiredVectorWidth<5
11091109
}
11101110

11111111
let Features = "avxvnniint8|avx10.2", Attributes = [NoThrow, Const, RequiredVectorWidth<128>] in {
1112-
def vpdpbssd128 : X86Builtin<"_Vector<4, int>(_Vector<4, int>, _Vector<4, int>, _Vector<4, int>)">;
1112+
def vpdpbssd128 : X86Builtin<"_Vector<4, int>(_Vector<4, int>, _Vector<16, char>, _Vector<16, char>)">;
11131113
}
11141114

11151115
let Features = "avxvnniint8|avx10.2", Attributes = [NoThrow, Const, RequiredVectorWidth<256>] in {
1116-
def vpdpbssd256 : X86Builtin<"_Vector<8, int>(_Vector<8, int>, _Vector<8, int>, _Vector<8, int>)">;
1116+
def vpdpbssd256 : X86Builtin<"_Vector<8, int>(_Vector<8, int>, _Vector<32, char>, _Vector<32, char>)">;
11171117
}
11181118

11191119
let Features = "avxvnniint8|avx10.2", Attributes = [NoThrow, Const, RequiredVectorWidth<128>] in {
1120-
def vpdpbssds128 : X86Builtin<"_Vector<4, int>(_Vector<4, int>, _Vector<4, int>, _Vector<4, int>)">;
1120+
def vpdpbssds128 : X86Builtin<"_Vector<4, int>(_Vector<4, int>, _Vector<16, char>, _Vector<16, char>)">;
11211121
}
11221122

11231123
let Features = "avxvnniint8|avx10.2", Attributes = [NoThrow, Const, RequiredVectorWidth<256>] in {
1124-
def vpdpbssds256 : X86Builtin<"_Vector<8, int>(_Vector<8, int>, _Vector<8, int>, _Vector<8, int>)">;
1124+
def vpdpbssds256 : X86Builtin<"_Vector<8, int>(_Vector<8, int>, _Vector<32, char>, _Vector<32, char>)">;
11251125
}
11261126

11271127
let Features = "avxvnniint8|avx10.2", Attributes = [NoThrow, Const, RequiredVectorWidth<128>] in {
1128-
def vpdpbsud128 : X86Builtin<"_Vector<4, int>(_Vector<4, int>, _Vector<4, int>, _Vector<4, int>)">;
1128+
def vpdpbsud128 : X86Builtin<"_Vector<4, int>(_Vector<4, int>, _Vector<16, char>, _Vector<16, unsigned char>)">;
11291129
}
11301130

11311131
let Features = "avxvnniint8|avx10.2", Attributes = [NoThrow, Const, RequiredVectorWidth<256>] in {
1132-
def vpdpbsud256 : X86Builtin<"_Vector<8, int>(_Vector<8, int>, _Vector<8, int>, _Vector<8, int>)">;
1132+
def vpdpbsud256 : X86Builtin<"_Vector<8, int>(_Vector<8, int>, _Vector<32, char>, _Vector<32, unsigned char>)">;
11331133
}
11341134

11351135
let Features = "avxvnniint8|avx10.2", Attributes = [NoThrow, Const, RequiredVectorWidth<128>] in {
1136-
def vpdpbsuds128 : X86Builtin<"_Vector<4, int>(_Vector<4, int>, _Vector<4, int>, _Vector<4, int>)">;
1136+
def vpdpbsuds128 : X86Builtin<"_Vector<4, int>(_Vector<4, int>, _Vector<16, char>, _Vector<16, unsigned char>)">;
11371137
}
11381138

11391139
let Features = "avxvnniint8|avx10.2", Attributes = [NoThrow, Const, RequiredVectorWidth<256>] in {
1140-
def vpdpbsuds256 : X86Builtin<"_Vector<8, int>(_Vector<8, int>, _Vector<8, int>, _Vector<8, int>)">;
1140+
def vpdpbsuds256 : X86Builtin<"_Vector<8, int>(_Vector<8, int>, _Vector<32, char>, _Vector<32, unsigned char>)">;
11411141
}
11421142

11431143
let Features = "avxvnniint8|avx10.2", Attributes = [NoThrow, Const, RequiredVectorWidth<128>] in {
1144-
def vpdpbuud128 : X86Builtin<"_Vector<4, int>(_Vector<4, int>, _Vector<4, int>, _Vector<4, int>)">;
1144+
def vpdpbuud128 : X86Builtin<"_Vector<4, int>(_Vector<4, int>, _Vector<16, unsigned char>, _Vector<16, unsigned char>)">;
11451145
}
11461146

11471147
let Features = "avxvnniint8|avx10.2", Attributes = [NoThrow, Const, RequiredVectorWidth<256>] in {
1148-
def vpdpbuud256 : X86Builtin<"_Vector<8, int>(_Vector<8, int>, _Vector<8, int>, _Vector<8, int>)">;
1148+
def vpdpbuud256 : X86Builtin<"_Vector<8, int>(_Vector<8, int>, _Vector<32, unsigned char>, _Vector<32, unsigned char>)">;
11491149
}
11501150

11511151
let Features = "avxvnniint8|avx10.2", Attributes = [NoThrow, Const, RequiredVectorWidth<128>] in {
1152-
def vpdpbuuds128 : X86Builtin<"_Vector<4, int>(_Vector<4, int>, _Vector<4, int>, _Vector<4, int>)">;
1152+
def vpdpbuuds128 : X86Builtin<"_Vector<4, int>(_Vector<4, int>, _Vector<16, unsigned char>, _Vector<16, unsigned char>)">;
11531153
}
11541154

11551155
let Features = "avxvnniint8|avx10.2", Attributes = [NoThrow, Const, RequiredVectorWidth<256>] in {
1156-
def vpdpbuuds256 : X86Builtin<"_Vector<8, int>(_Vector<8, int>, _Vector<8, int>, _Vector<8, int>)">;
1156+
def vpdpbuuds256 : X86Builtin<"_Vector<8, int>(_Vector<8, int>, _Vector<32, unsigned char>, _Vector<32, unsigned char>)">;
11571157
}
11581158

11591159
let Features = "movrs", Attributes = [NoThrow, Const] in {
@@ -4282,12 +4282,12 @@ let Features = "avx10.2", Attributes = [NoThrow, Const, RequiredVectorWidth<256>
42824282

42834283
let Features = "avx10.2", Attributes = [NoThrow, Const, RequiredVectorWidth<512>] in {
42844284
def vdpphps512 : X86Builtin<"_Vector<16, float>(_Vector<16, float>, _Vector<32, _Float16>, _Vector<32, _Float16>)">;
4285-
def vpdpbssd512 : X86Builtin<"_Vector<16, int>(_Vector<16, int>, _Vector<16, int>, _Vector<16, int>)">;
4286-
def vpdpbssds512 : X86Builtin<"_Vector<16, int>(_Vector<16, int>, _Vector<16, int>, _Vector<16, int>)">;
4287-
def vpdpbsud512 : X86Builtin<"_Vector<16, int>(_Vector<16, int>, _Vector<16, int>, _Vector<16, int>)">;
4288-
def vpdpbsuds512 : X86Builtin<"_Vector<16, int>(_Vector<16, int>, _Vector<16, int>, _Vector<16, int>)">;
4289-
def vpdpbuud512 : X86Builtin<"_Vector<16, int>(_Vector<16, int>, _Vector<16, int>, _Vector<16, int>)">;
4290-
def vpdpbuuds512 : X86Builtin<"_Vector<16, int>(_Vector<16, int>, _Vector<16, int>, _Vector<16, int>)">;
4285+
def vpdpbssd512 : X86Builtin<"_Vector<16, int>(_Vector<16, int>, _Vector<64, char>, _Vector<64, char>)">;
4286+
def vpdpbssds512 : X86Builtin<"_Vector<16, int>(_Vector<16, int>, _Vector<64, char>, _Vector<64, char>)">;
4287+
def vpdpbsud512 : X86Builtin<"_Vector<16, int>(_Vector<16, int>, _Vector<64, char>, _Vector<64, unsigned char>)">;
4288+
def vpdpbsuds512 : X86Builtin<"_Vector<16, int>(_Vector<16, int>, _Vector<64, char>, _Vector<64, unsigned char>)">;
4289+
def vpdpbuud512 : X86Builtin<"_Vector<16, int>(_Vector<16, int>, _Vector<64, unsigned char>, _Vector<64, unsigned char>)">;
4290+
def vpdpbuuds512 : X86Builtin<"_Vector<16, int>(_Vector<16, int>, _Vector<64, unsigned char>, _Vector<64, unsigned char>)">;
42914291
}
42924292

42934293
let Features = "avx10.2", Attributes = [NoThrow, RequiredVectorWidth<512>] in {

‎clang/lib/Headers/avx10_2_512niintrin.h‎

Lines changed: 12 additions & 12 deletions
Original file line numberDiff line numberDiff line change
@@ -64,8 +64,8 @@ static __inline__ __m512 __DEFAULT_FN_ATTRS _mm512_maskz_dpph_ps(__mmask16 __U,
6464
static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_dpbssd_epi32(__m512i __W,
6565
__m512i __A,
6666
__m512i __B) {
67-
return (__m512i)__builtin_ia32_vpdpbssd512((__v16si)__W, (__v16si)__A,
68-
(__v16si)__B);
67+
return (__m512i)__builtin_ia32_vpdpbssd512((__v16si)__W, (__v64qi)__A,
68+
(__v64qi)__B);
6969
}
7070

7171
static __inline__ __m512i __DEFAULT_FN_ATTRS
@@ -84,8 +84,8 @@ static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_maskz_dpbssd_epi32(
8484
static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_dpbssds_epi32(__m512i __W,
8585
__m512i __A,
8686
__m512i __B) {
87-
return (__m512i)__builtin_ia32_vpdpbssds512((__v16si)__W, (__v16si)__A,
88-
(__v16si)__B);
87+
return (__m512i)__builtin_ia32_vpdpbssds512((__v16si)__W, (__v64qi)__A,
88+
(__v64qi)__B);
8989
}
9090

9191
static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_mask_dpbssds_epi32(
@@ -104,8 +104,8 @@ static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_maskz_dpbssds_epi32(
104104
static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_dpbsud_epi32(__m512i __W,
105105
__m512i __A,
106106
__m512i __B) {
107-
return (__m512i)__builtin_ia32_vpdpbsud512((__v16si)__W, (__v16si)__A,
108-
(__v16si)__B);
107+
return (__m512i)__builtin_ia32_vpdpbsud512((__v16si)__W, (__v64qi)__A,
108+
(__v64qu)__B);
109109
}
110110

111111
static __inline__ __m512i __DEFAULT_FN_ATTRS
@@ -124,8 +124,8 @@ static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_maskz_dpbsud_epi32(
124124
static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_dpbsuds_epi32(__m512i __W,
125125
__m512i __A,
126126
__m512i __B) {
127-
return (__m512i)__builtin_ia32_vpdpbsuds512((__v16si)__W, (__v16si)__A,
128-
(__v16si)__B);
127+
return (__m512i)__builtin_ia32_vpdpbsuds512((__v16si)__W, (__v64qi)__A,
128+
(__v64qu)__B);
129129
}
130130

131131
static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_mask_dpbsuds_epi32(
@@ -144,8 +144,8 @@ static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_maskz_dpbsuds_epi32(
144144
static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_dpbuud_epi32(__m512i __W,
145145
__m512i __A,
146146
__m512i __B) {
147-
return (__m512i)__builtin_ia32_vpdpbuud512((__v16si)__W, (__v16si)__A,
148-
(__v16si)__B);
147+
return (__m512i)__builtin_ia32_vpdpbuud512((__v16si)__W, (__v64qu)__A,
148+
(__v64qu)__B);
149149
}
150150

151151
static __inline__ __m512i __DEFAULT_FN_ATTRS
@@ -164,8 +164,8 @@ static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_maskz_dpbuud_epi32(
164164
static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_dpbuuds_epi32(__m512i __W,
165165
__m512i __A,
166166
__m512i __B) {
167-
return (__m512i)__builtin_ia32_vpdpbuuds512((__v16si)__W, (__v16si)__A,
168-
(__v16si)__B);
167+
return (__m512i)__builtin_ia32_vpdpbuuds512((__v16si)__W, (__v64qu)__A,
168+
(__v64qu)__B);
169169
}
170170

171171
static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_mask_dpbuuds_epi32(

0 commit comments

Comments
 (0)