Use unsigned short vector types for pmulhuw intrinsics

Author: RKSimon

clang/include/clang/Basic/BuiltinsX86.td

@@ -105,9 +105,9 @@ let Attributes = [Const, NoThrow, RequiredVectorWidth<128>] in {
     def vec_set_v8hi : X86Builtin<"_Vector<8, short>(_Vector<8, short>, short, _Constant int)">;
   }
 
-  let Features = "sse2", Attributes = [NoThrow, Const, Constexpr] in {
+  let Features = "sse2", Attributes = [NoThrow, Const, Constexpr, RequiredVectorWidth<128>] in {
     def pmulhw128 : X86Builtin<"_Vector<8, short>(_Vector<8, short>, _Vector<8, short>)">;
-    def pmulhuw128 : X86Builtin<"_Vector<8, short>(_Vector<8, short>, _Vector<8, short>)">;
+    def pmulhuw128 : X86Builtin<"_Vector<8, unsigned short>(_Vector<8, unsigned short>, _Vector<8, unsigned short>)">;
   }
 
   let Features = "sse3" in {
@@ -621,7 +621,7 @@ let Features = "avx2", Attributes = [NoThrow, Const, RequiredVectorWidth<256>] i
 }
 
 let Features = "avx2", Attributes = [NoThrow, Const, Constexpr, RequiredVectorWidth<256>] in {
-  def pmulhuw256 : X86Builtin<"_Vector<16, short>(_Vector<16, short>, _Vector<16, short>)">;
+  def pmulhuw256 : X86Builtin<"_Vector<16, unsigned short>(_Vector<16, unsigned short>, _Vector<16, unsigned short>)">;
   def pmulhw256 : X86Builtin<"_Vector<16, short>(_Vector<16, short>, _Vector<16, short>)">;
 }
 
@@ -1438,7 +1438,7 @@ let Features = "avx512bw,evex512", Attributes = [NoThrow, Const, RequiredVectorW
 }
 
 let Features = "avx512bw,evex512", Attributes = [NoThrow, Const, Constexpr, RequiredVectorWidth<512>] in {
-  def pmulhuw512 : X86Builtin<"_Vector<32, short>(_Vector<32, short>, _Vector<32, short>)">;
+  def pmulhuw512 : X86Builtin<"_Vector<32, unsigned short>(_Vector<32, unsigned short>, _Vector<32, unsigned short>)">;
   def pmulhw512 : X86Builtin<"_Vector<32, short>(_Vector<32, short>, _Vector<32, short>)">;
 }
 

clang/lib/Headers/avx2intrin.h

@@ -1732,7 +1732,7 @@ _mm256_mulhrs_epi16(__m256i __a, __m256i __b)
 static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
 _mm256_mulhi_epu16(__m256i __a, __m256i __b)
 {
-  return (__m256i)__builtin_ia32_pmulhuw256((__v16hi)__a, (__v16hi)__b);
+  return (__m256i)__builtin_ia32_pmulhuw256((__v16hu)__a, (__v16hu)__b);
 }
 
 /// Multiplies signed 16-bit integer elements of two 256-bit vectors of

clang/lib/Headers/avx512bwintrin.h

@@ -1116,7 +1116,7 @@ _mm512_maskz_mulhi_epi16(__mmask32 __U, __m512i __A, __m512i __B)
 static __inline__ __m512i __DEFAULT_FN_ATTRS512_CONSTEXPR
 _mm512_mulhi_epu16(__m512i __A, __m512i __B)
 {
-  return (__m512i)__builtin_ia32_pmulhuw512((__v32hi) __A, (__v32hi) __B);
+  return (__m512i)__builtin_ia32_pmulhuw512((__v32hu) __A, (__v32hu) __B);
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS512

clang/lib/Headers/emmintrin.h

@@ -2415,7 +2415,7 @@ _mm_mulhi_epi16(__m128i __a, __m128i __b) {
 ///    of each of the eight 32-bit products.
 static __inline__ __m128i __DEFAULT_FN_ATTRS_CONSTEXPR
 _mm_mulhi_epu16(__m128i __a, __m128i __b) {
-  return (__m128i)__builtin_ia32_pmulhuw128((__v8hi)__a, (__v8hi)__b);
+  return (__m128i)__builtin_ia32_pmulhuw128((__v8hu)__a, (__v8hu)__b);
 }
 
 /// Multiplies the corresponding elements of two signed [8 x i16]

clang/lib/Headers/xmmintrin.h

@@ -24,6 +24,7 @@ typedef float __m128_u __attribute__((__vector_size__(16), __aligned__(1)));
 
 /* Unsigned types */
 typedef unsigned int __v4su __attribute__((__vector_size__(16)));
+typedef unsigned short __v8hu __attribute__((__vector_size__(16)));
 
 /* This header should only be included in a hosted environment as it depends on
  * a standard library to provide allocation routines. */
@@ -2450,8 +2451,8 @@ _mm_movemask_pi8(__m64 __a)
 static __inline__ __m64 __DEFAULT_FN_ATTRS_SSE2_CONSTEXPR
 _mm_mulhi_pu16(__m64 __a, __m64 __b)
 {
-  return __trunc64(__builtin_ia32_pmulhuw128((__v8hi)__zext128(__a),
-                                             (__v8hi)__zext128(__b)));
+  return __trunc64(__builtin_ia32_pmulhuw128((__v8hu)__zext128(__a),
+                                             (__v8hu)__zext128(__b)));
 }
 
 /// Shuffles the 4 16-bit integers from a 64-bit integer vector to the