[Clang][X86] Replace F16C vcvtph2ps/256 intrinsics with (convert|shuffle)vector builtins (#152911)

The following intrinsics were replaced by a combination of
`__builtin_shufflevector` and `__builtin_convertvector`:
- `__builtin_ia32_vcvtph2ps`
- `__builtin_ia32_vcvtph2ps256`

Fixes #152749

Author: moorabbit

clang/include/clang/Basic/BuiltinsX86.td

@@ -757,14 +757,6 @@ let Features = "f16c", Attributes = [NoThrow, Const, RequiredVectorWidth<256>] i
   def vcvtps2ph256 : X86Builtin<"_Vector<8, short>(_Vector<8, float>, _Constant int)">;
 }
 
-let Features = "f16c", Attributes = [NoThrow, Const, RequiredVectorWidth<128>] in {
-  def vcvtph2ps : X86Builtin<"_Vector<4, float>(_Vector<8, short>)">;
-}
-
-let Features = "f16c", Attributes = [NoThrow, Const, RequiredVectorWidth<256>] in {
-  def vcvtph2ps256 : X86Builtin<"_Vector<8, float>(_Vector<8, short>)">;
-}
-
 let Features = "rdrnd", Attributes = [NoThrow] in {
   def rdrand16_step : X86Builtin<"unsigned int(unsigned short *)">;
   def rdrand32_step : X86Builtin<"unsigned int(unsigned int *)">;

clang/lib/CodeGen/TargetBuiltins/X86.cpp

@@ -2841,8 +2841,6 @@ Value *CodeGenFunction::EmitX86BuiltinExpr(unsigned BuiltinID,
     return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 7);
 
   // f16c half2float intrinsics
-  case X86::BI__builtin_ia32_vcvtph2ps:
-  case X86::BI__builtin_ia32_vcvtph2ps256:
   case X86::BI__builtin_ia32_vcvtph2ps_mask:
   case X86::BI__builtin_ia32_vcvtph2ps256_mask:
   case X86::BI__builtin_ia32_vcvtph2ps512_mask: {

clang/lib/Headers/f16cintrin.h

@@ -38,9 +38,7 @@
 static __inline float __DEFAULT_FN_ATTRS128
 _cvtsh_ss(unsigned short __a)
 {
-  __v8hi __v = {(short)__a, 0, 0, 0, 0, 0, 0, 0};
-  __v4sf __r = __builtin_ia32_vcvtph2ps(__v);
-  return __r[0];
+  return (float)__builtin_bit_cast(__fp16, __a);
 }
 
 /// Converts a 32-bit single-precision float value to a 16-bit
@@ -109,7 +107,10 @@ _cvtsh_ss(unsigned short __a)
 static __inline __m128 __DEFAULT_FN_ATTRS128
 _mm_cvtph_ps(__m128i __a)
 {
-  return (__m128)__builtin_ia32_vcvtph2ps((__v8hi)__a);
+  typedef __fp16 __v4fp16 __attribute__((__vector_size__(8)));
+
+  __v4hi __v = __builtin_shufflevector((__v8hi)__a, (__v8hi)__a, 0, 1, 2, 3);
+  return (__m128) __builtin_convertvector((__v4fp16)__v, __v4sf);
 }
 
 /// Converts a 256-bit vector of [8 x float] into a 128-bit vector
@@ -153,7 +154,9 @@ _mm_cvtph_ps(__m128i __a)
 static __inline __m256 __DEFAULT_FN_ATTRS256
 _mm256_cvtph_ps(__m128i __a)
 {
-  return (__m256)__builtin_ia32_vcvtph2ps256((__v8hi)__a);
+  typedef __fp16 __v8fp16 __attribute__((__vector_size__(16), __aligned__(16)));
+
+  return (__m256) __builtin_convertvector((__v8fp16)__a, __v8sf);
 }
 
 #undef __DEFAULT_FN_ATTRS128

clang/test/CodeGen/X86/f16c-builtins-constrained.c

@@ -8,17 +8,8 @@
 
 float test_cvtsh_ss(unsigned short a) {
   // CHECK-LABEL: test_cvtsh_ss
-  // CHECK: insertelement <8 x i16> poison, i16 %{{.*}}, i32 0
-  // CHECK: insertelement <8 x i16> %{{.*}}, i16 0, i32 1
-  // CHECK: insertelement <8 x i16> %{{.*}}, i16 0, i32 2
-  // CHECK: insertelement <8 x i16> %{{.*}}, i16 0, i32 3
-  // CHECK: insertelement <8 x i16> %{{.*}}, i16 0, i32 4
-  // CHECK: insertelement <8 x i16> %{{.*}}, i16 0, i32 5
-  // CHECK: insertelement <8 x i16> %{{.*}}, i16 0, i32 6
-  // CHECK: insertelement <8 x i16> %{{.*}}, i16 0, i32 7
-  // CHECK: shufflevector <8 x i16> %{{.*}}, <8 x i16> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
-  // CHECK: call <4 x float> @llvm.experimental.constrained.fpext.v4f32.v4f16(<4 x half> %{{.*}}, metadata !"fpexcept.strict")
-  // CHECK: extractelement <4 x float> %{{.*}}, i32 0
+  // CHECK: [[CONV:%.*]] = call {{.*}}float @llvm.experimental.constrained.fpext.f32.f16(half %{{.*}}, metadata !"fpexcept.strict")
+  // CHECK: ret float [[CONV]]
   return _cvtsh_ss(a);
 }
 
@@ -38,7 +29,7 @@ unsigned short test_cvtss_sh(float a) {
 
 __m128 test_mm_cvtph_ps(__m128i a) {
   // CHECK-LABEL: test_mm_cvtph_ps
-  // CHECK: shufflevector <8 x i16> %{{.*}}, <8 x i16> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  // CHECK: shufflevector <8 x i16> %{{.*}}, <8 x i16> %{{.*}}, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
   // CHECK: call {{.*}}<4 x float> @llvm.experimental.constrained.fpext.v4f32.v4f16(<4 x half> %{{.*}}, metadata !"fpexcept.strict")
   return _mm_cvtph_ps(a);
 }

clang/test/CodeGen/X86/f16c-builtins.c

@@ -8,17 +8,8 @@
 
 float test_cvtsh_ss(unsigned short a) {
   // CHECK-LABEL: test_cvtsh_ss
-  // CHECK: insertelement <8 x i16> poison, i16 %{{.*}}, i32 0
-  // CHECK: insertelement <8 x i16> %{{.*}}, i16 0, i32 1
-  // CHECK: insertelement <8 x i16> %{{.*}}, i16 0, i32 2
-  // CHECK: insertelement <8 x i16> %{{.*}}, i16 0, i32 3
-  // CHECK: insertelement <8 x i16> %{{.*}}, i16 0, i32 4
-  // CHECK: insertelement <8 x i16> %{{.*}}, i16 0, i32 5
-  // CHECK: insertelement <8 x i16> %{{.*}}, i16 0, i32 6
-  // CHECK: insertelement <8 x i16> %{{.*}}, i16 0, i32 7
-  // CHECK: shufflevector <8 x i16> %{{.*}}, <8 x i16> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
-  // CHECK: fpext <4 x half> %{{.*}} to <4 x float>
-  // CHECK: extractelement <4 x float> %{{.*}}, i32 0
+  // CHECK: [[CONV:%.*]] = fpext half %{{.*}} to float
+  // CHECK: ret float [[CONV]]
   return _cvtsh_ss(a);
 }
 
@@ -35,7 +26,7 @@ unsigned short test_cvtss_sh(float a) {
 
 __m128 test_mm_cvtph_ps(__m128i a) {
   // CHECK-LABEL: test_mm_cvtph_ps
-  // CHECK: shufflevector <8 x i16> %{{.*}}, <8 x i16> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  // CHECK: shufflevector <8 x i16> %{{.*}}, <8 x i16> %{{.*}}, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
   // CHECK: fpext <4 x half> %{{.*}} to <4 x float>
   return _mm_cvtph_ps(a);
 }