[X86][F16C] Remove cvtph2ps intrinsics and use generic half2float conversion (PR37554)

[RKSimon]

Summary

This removes everything but int_x86_avx512_mask_vcvtph2ps_512 which provides the SAE variant, but even this can use the fpext generic if the rounding control is the default.

I have a new f16c-intrinsics-upgrade.ll file as well but for some reason git diff is going weird when I try to include it, basically all the ph2ps tests are cut+paste from f16c-intrinsics.ll

[craig.topper]

Should we check that this generates constrained.fpext in strict mode?

[RKSimon]

In D75162#1896365, @craig.topper wrote:

Should we check that this generates constrained.fpext in strict mode?

Sure - where is the best place(s) to test this?

[craig.topper]

In D75162#1896469, @RKSimon wrote:

In D75162#1896365, @craig.topper wrote:

Should we check that this generates constrained.fpext in strict mode?

Sure - where is the best place(s) to test this?

For the intrinsics we've tested this way so far, they're in sse-builtins-constrained.c and similar. I think we have separate files for compares that I should merge into the others. They got commited separately as I was working on compares when someone else was working on the others so I made a separate file to avoid the merge conflict.

[craig.topper]

I've posted https://reviews.llvm.org/D75304 to make the backend recognize the strict version of this pattern.

[RKSimon]

Add strict fp codegen tests - this means the patch is now dependent on D75304

[craig.topper]

LGTM

[hoyFB]

llvm/include/llvm/IR/IntrinsicsX86.td

2551

Hello, we have some tools generating these intrinsics being removed here. Can you suggest what we should generate on LLVM IR with this change?

[craig.topper]

llvm/include/llvm/IR/IntrinsicsX86.td

2551

I think the diffs to test/CodeGen/f16c-builtins.c should show the difference. You can also follow the code in the changes to AutoUpgrade.cpp

clang/lib/CodeGen/CGBuiltin.cpp

 }
 
 Value *CodeGenFunction::EmitX86CpuIs(const CallExpr *E) {
   const Expr *CPUExpr = E->getArg(0)->IgnoreParenCasts();
   StringRef CPUStr = cast<clang::StringLiteral>(CPUExpr)->getString();
   return EmitX86CpuIs(CPUStr);
 }
 
+// Convert F16 halfs to floats.
+static Value *EmitX86CvtF16ToFloatExpr(CodeGenFunction &CGF,
+                                       ArrayRef<Value *> Ops,
+                                       llvm::Type *DstTy) {
+  assert((Ops.size() == 1 || Ops.size() == 3 || Ops.size() == 4) &&
+         "Unknown cvtph2ps intrinsic");
+
+  // If the SAE intrinsic doesn't use default rounding then we can't upgrade.
+  if (Ops.size() == 4 && cast<llvm::ConstantInt>(Ops[3])->getZExtValue() != 4) {
+    Intrinsic::ID IID = Intrinsic::x86_avx512_mask_vcvtph2ps_512;
+    Function *F =
+        CGF.CGM.getIntrinsic(IID, {DstTy, Ops[0]->getType(), Ops[1]->getType(),
+                                   Ops[2]->getType(), Ops[3]->getType()});
+    return CGF.Builder.CreateCall(F, {Ops[0], Ops[1], Ops[2], Ops[3]});
+  }
+
+  unsigned NumDstElts = DstTy->getVectorNumElements();
+  Value *Src = Ops[0];
+
+  // Extract the subvector.
+  if (NumDstElts != Src->getType()->getVectorNumElements()) {
+    assert(NumDstElts == 4 && "Unexpected vector size");
+    uint32_t ShuffleMask[4] = {0, 1, 2, 3};
+    Src = CGF.Builder.CreateShuffleVector(Src, UndefValue::get(Src->getType()),
+                                          ShuffleMask);
+  }
+
+  // Bitcast from vXi16 to vXf16.
+  llvm::Type *HalfTy = llvm::VectorType::get(
+      llvm::Type::getHalfTy(CGF.getLLVMContext()), NumDstElts);
+  Src = CGF.Builder.CreateBitCast(Src, HalfTy);
+
+  // Perform the fp-extension.
+  Value *Res = CGF.Builder.CreateFPExt(Src, DstTy, "cvtph2ps");
+
+  if (Ops.size() >= 3)
+    Res = EmitX86Select(CGF, Ops[2], Res, Ops[1]);
+  return Res;
+}
+
 // Convert a BF16 to a float.
 static Value *EmitX86CvtBF16ToFloatExpr(CodeGenFunction &CGF,
                                         const CallExpr *E,
                                         ArrayRef<Value *> Ops) {
   llvm::Type *Int32Ty = CGF.Builder.getInt32Ty();
   Value *ZeroExt = CGF.Builder.CreateZExt(Ops[0], Int32Ty);
   Value *Shl = CGF.Builder.CreateShl(ZeroExt, 16);
   llvm::Type *ResultType = CGF.ConvertType(E->getType());
case X86::BI__builtin_ia32_cmpneqsd:
     return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 4);
   case X86::BI__builtin_ia32_cmpnltsd:
     return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 5);
   case X86::BI__builtin_ia32_cmpnlesd:
     return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 6);
   case X86::BI__builtin_ia32_cmpordsd:
     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:
+    return EmitX86CvtF16ToFloatExpr(*this, Ops, ConvertType(E->getType()));
+
 // AVX512 bf16 intrinsics
   case X86::BI__builtin_ia32_cvtneps2bf16_128_mask: {
     Ops[2] = getMaskVecValue(*this, Ops[2],
                              Ops[0]->getType()->getVectorNumElements());
     Intrinsic::ID IID = Intrinsic::x86_avx512bf16_mask_cvtneps2bf16_128;
     return Builder.CreateCall(CGM.getIntrinsic(IID), Ops);
   }
   case X86::BI__builtin_ia32_cvtsbf162ss_32:

clang/test/CodeGen/avx512f-builtins.c

 
 __m256 test_mm512_maskz_cvtpd_ps (__mmask8 __U, __m512d __A)
 {
   // CHECK-LABEL: @test_mm512_maskz_cvtpd_ps 
   // CHECK: @llvm.x86.avx512.mask.cvtpd2ps.512
   return _mm512_maskz_cvtpd_ps (__U,__A);
 }
 
+__m512 test_mm512_cvtph_ps (__m256i __A)
+{
+  // CHECK-LABEL: @test_mm512_cvtph_ps 
+  // CHECK: bitcast <4 x i64> %{{.*}} to <16 x i16>
+  // CHECK: bitcast <16 x i16> %{{.*}} to <16 x half>
+  // CHECK: fpext <16 x half> %{{.*}} to <16 x float>
+  return _mm512_cvtph_ps (__A);
+}
+
 __m512 test_mm512_mask_cvtph_ps (__m512 __W, __mmask16 __U, __m256i __A)
 {
   // CHECK-LABEL: @test_mm512_mask_cvtph_ps 
-  // CHECK: @llvm.x86.avx512.mask.vcvtph2ps.512
+  // CHECK: bitcast <4 x i64> %{{.*}} to <16 x i16>
+  // CHECK: bitcast <16 x i16> %{{.*}} to <16 x half>
+  // CHECK: fpext <16 x half> %{{.*}} to <16 x float>
+  // CHECK: select <16 x i1> %{{.*}}, <16 x float> %{{.*}}, <16 x float> %{{.*}}
   return _mm512_mask_cvtph_ps (__W,__U,__A);
 }
 
 __m512 test_mm512_maskz_cvtph_ps (__mmask16 __U, __m256i __A)
 {
   // CHECK-LABEL: @test_mm512_maskz_cvtph_ps 
-  // CHECK: @llvm.x86.avx512.mask.vcvtph2ps.512
+  // CHECK: bitcast <4 x i64> %{{.*}} to <16 x i16>
+  // CHECK: bitcast <16 x i16> %{{.*}} to <16 x half>
+  // CHECK: fpext <16 x half> %{{.*}} to <16 x float>
+  // CHECK: select <16 x i1> %{{.*}}, <16 x float> %{{.*}}, <16 x float> %{{.*}}
   return _mm512_maskz_cvtph_ps (__U,__A);
 }
 
 __m256i test_mm512_mask_cvttpd_epi32 (__m256i __W, __mmask8 __U, __m512d __A)
 {
   // CHECK-LABEL: @test_mm512_mask_cvttpd_epi32 
   // CHECK: @llvm.x86.avx512.mask.cvttpd2dq.512
   return _mm512_mask_cvttpd_epi32 (__W,__U,__A);

clang/test/CodeGen/avx512vl-builtins.c

 __m256 test_mm256_maskz_mov_ps(__mmask8 __U, __m256 __A) {
   // CHECK-LABEL: @test_mm256_maskz_mov_ps
   // CHECK: select <8 x i1> %{{.*}}, <8 x float> %{{.*}}, <8 x float> %{{.*}}
   return _mm256_maskz_mov_ps(__U, __A);
 }
 
 __m128 test_mm_mask_cvtph_ps(__m128 __W, __mmask8 __U, __m128i __A) {
   // CHECK-LABEL: @test_mm_mask_cvtph_ps
-  // CHECK: @llvm.x86.avx512.mask.vcvtph2ps.128
+  // CHECK: bitcast <2 x i64> %{{.*}} to <8 x i16>
+  // CHECK: shufflevector <8 x i16> %{{.*}}, <8 x i16> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  // CHECK: bitcast <4 x i16> %{{.*}} to <4 x half>
+  // CHECK: fpext <4 x half> %{{.*}} to <4 x float>
+  // CHECK: select <4 x i1> %{{.*}}, <4 x float> %{{.*}}, <4 x float> %{{.*}}
   return _mm_mask_cvtph_ps(__W, __U, __A);
 }
 
 __m128 test_mm_maskz_cvtph_ps(__mmask8 __U, __m128i __A) {
   // CHECK-LABEL: @test_mm_maskz_cvtph_ps
-  // CHECK: @llvm.x86.avx512.mask.vcvtph2ps.128
+  // CHECK: bitcast <2 x i64> %{{.*}} to <8 x i16>
+  // CHECK: shufflevector <8 x i16> %{{.*}}, <8 x i16> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  // CHECK: bitcast <4 x i16> %{{.*}} to <4 x half>
+  // CHECK: fpext <4 x half> %{{.*}} to <4 x float>
+  // CHECK: select <4 x i1> %{{.*}}, <4 x float> %{{.*}}, <4 x float> %{{.*}}
   return _mm_maskz_cvtph_ps(__U, __A);
 }
 
 __m256 test_mm256_mask_cvtph_ps(__m256 __W, __mmask8 __U, __m128i __A) {
   // CHECK-LABEL: @test_mm256_mask_cvtph_ps
-  // CHECK: @llvm.x86.avx512.mask.vcvtph2ps.256
+  // CHECK: bitcast <2 x i64> %{{.*}} to <8 x i16>
+  // CHECK: bitcast <8 x i16> %{{.*}} to <8 x half>
+  // CHECK: fpext <8 x half> %{{.*}} to <8 x float>
+  // CHECK: select <8 x i1> %{{.*}}, <8 x float> %{{.*}}, <8 x float> %{{.*}}
   return _mm256_mask_cvtph_ps(__W, __U, __A);
 }
 
 __m256 test_mm256_maskz_cvtph_ps(__mmask8 __U, __m128i __A) {
   // CHECK-LABEL: @test_mm256_maskz_cvtph_ps
-  // CHECK: @llvm.x86.avx512.mask.vcvtph2ps.256
+  // CHECK: bitcast <2 x i64> %{{.*}} to <8 x i16>
+  // CHECK: bitcast <8 x i16> %{{.*}} to <8 x half>
+  // CHECK: fpext <8 x half> %{{.*}} to <8 x float>
+  // CHECK: select <8 x i1> %{{.*}}, <8 x float> %{{.*}}, <8 x float> %{{.*}}
   return _mm256_maskz_cvtph_ps(__U, __A);
 }
 
 __m128i test_mm_mask_cvtps_ph(__m128i __W, __mmask8 __U, __m128 __A) {
   // CHECK-LABEL: @test_mm_mask_cvtps_ph
   // CHECK: @llvm.x86.avx512.mask.vcvtps2ph.128
   return _mm_mask_cvtps_ph(__W, __U, __A, _MM_FROUND_TO_ZERO | _MM_FROUND_NO_EXC);
 }

clang/test/CodeGen/f16c-builtins.c

 // RUN: %clang_cc1 -ffreestanding %s -triple=x86_64-apple-darwin -target-feature +f16c -emit-llvm -o - -Wall -Werror | FileCheck %s
 
 
 #include <immintrin.h>
 
 float test_cvtsh_ss(unsigned short a) {
   // CHECK-LABEL: test_cvtsh_ss
   // CHECK: insertelement <8 x i16> undef, 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: call <4 x float> @llvm.x86.vcvtph2ps.128(<8 x i16> %{{.*}})
+  // CHECK: shufflevector <8 x i16> %{{.*}}, <8 x i16> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  // CHECK: bitcast <4 x i16> %{{.*}} to <4 x half>
+  // CHECK: fpext <4 x half> %{{.*}} to <4 x float>
   // CHECK: extractelement <4 x float> %{{.*}}, i32 0
   return _cvtsh_ss(a);
 }
 
 unsigned short test_cvtss_sh(float a) {
   // CHECK-LABEL: test_cvtss_sh
   // CHECK: insertelement <4 x float> undef, float %{{.*}}, i32 0
   // CHECK: insertelement <4 x float> %{{.*}}, float 0.000000e+00, i32 1
   // CHECK: insertelement <4 x float> %{{.*}}, float 0.000000e+00, i32 2
   // CHECK: insertelement <4 x float> %{{.*}}, float 0.000000e+00, i32 3
   // CHECK: call <8 x i16> @llvm.x86.vcvtps2ph.128(<4 x float> %{{.*}}, i32 0)
   // CHECK: extractelement <8 x i16> %{{.*}}, i32 0
   return _cvtss_sh(a, 0);
 }
 
 __m128 test_mm_cvtph_ps(__m128i a) {
   // CHECK-LABEL: test_mm_cvtph_ps
-  // CHECK: call <4 x float> @llvm.x86.vcvtph2ps.128(<8 x i16> %{{.*}})
+  // CHECK: bitcast <2 x i64> %{{.*}} to <8 x i16>
+  // CHECK: shufflevector <8 x i16> %{{.*}}, <8 x i16> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  // CHECK: bitcast <4 x i16> %{{.*}} to <4 x half>
+  // CHECK: fpext <4 x half> %{{.*}} to <4 x float>
   return _mm_cvtph_ps(a);
 }
 
 __m256 test_mm256_cvtph_ps(__m128i a) {
   // CHECK-LABEL: test_mm256_cvtph_ps
-  // CHECK: call <8 x float> @llvm.x86.vcvtph2ps.256(<8 x i16> %{{.*}})
+  // CHECK: bitcast <2 x i64> %{{.*}} to <8 x i16>
+  // CHECK: bitcast <8 x i16> %{{.*}} to <8 x half>
+  // CHECK: fpext <8 x half> %{{.*}} to <8 x float>
   return _mm256_cvtph_ps(a);
 }
 
 __m128i test_mm_cvtps_ph(__m128 a) {
   // CHECK-LABEL: test_mm_cvtps_ph
   // CHECK: call <8 x i16> @llvm.x86.vcvtps2ph.128(<4 x float> %{{.*}}, i32 0)
   return _mm_cvtps_ph(a, 0);
 }
 
 __m128i test_mm256_cvtps_ph(__m256 a) {
   // CHECK-LABEL: test_mm256_cvtps_ph
   // CHECK: call <8 x i16> @llvm.x86.vcvtps2ph.256(<8 x float> %{{.*}}, i32 0)
   return _mm256_cvtps_ph(a, 0);
 }

llvm/include/llvm/IR/IntrinsicsX86.td

def int_x86_rdpkru : GCCBuiltin <"__builtin_ia32_rdpkru">,
               Intrinsic<[llvm_i32_ty], [], []>;
   def int_x86_wrpkru : GCCBuiltin<"__builtin_ia32_wrpkru">,
               Intrinsic<[], [llvm_i32_ty], []>;
 }
 //===----------------------------------------------------------------------===//
 // Half float conversion
 
 let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86.".
-  def int_x86_vcvtph2ps_128 : GCCBuiltin<"__builtin_ia32_vcvtph2ps">,
-              Intrinsic<[llvm_v4f32_ty], [llvm_v8i16_ty], [IntrNoMem]>;
-  def int_x86_vcvtph2ps_256 : GCCBuiltin<"__builtin_ia32_vcvtph2ps256">,
-              Intrinsic<[llvm_v8f32_ty], [llvm_v8i16_ty], [IntrNoMem]>;
   def int_x86_vcvtps2ph_128 : GCCBuiltin<"__builtin_ia32_vcvtps2ph">,
               Intrinsic<[llvm_v8i16_ty], [llvm_v4f32_ty, llvm_i32_ty],
                         [IntrNoMem, ImmArg<1>]>;
   def int_x86_vcvtps2ph_256 : GCCBuiltin<"__builtin_ia32_vcvtps2ph256">,
               Intrinsic<[llvm_v8i16_ty], [llvm_v8f32_ty, llvm_i32_ty],
                         [IntrNoMem, ImmArg<1>]>;
-  def int_x86_avx512_mask_vcvtph2ps_512 : GCCBuiltin<"__builtin_ia32_vcvtph2ps512_mask">,
+  def int_x86_avx512_mask_vcvtph2ps_512 :
               Intrinsic<[llvm_v16f32_ty], [llvm_v16i16_ty, llvm_v16f32_ty,
                                            llvm_i16_ty, llvm_i32_ty],
                         [IntrNoMem, ImmArg<3>]>;
-  def int_x86_avx512_mask_vcvtph2ps_256 : GCCBuiltin<"__builtin_ia32_vcvtph2ps256_mask">,
-              Intrinsic<[llvm_v8f32_ty], [llvm_v8i16_ty, llvm_v8f32_ty,
-                                           llvm_i8_ty], [IntrNoMem]>;
-  def int_x86_avx512_mask_vcvtph2ps_128 : GCCBuiltin<"__builtin_ia32_vcvtph2ps_mask">,
-              Intrinsic<[llvm_v4f32_ty], [llvm_v8i16_ty, llvm_v4f32_ty,
-                                           llvm_i8_ty], [IntrNoMem]>;
   def int_x86_avx512_mask_vcvtps2ph_512 : GCCBuiltin<"__builtin_ia32_vcvtps2ph512_mask">,
               Intrinsic<[llvm_v16i16_ty], [llvm_v16f32_ty, llvm_i32_ty,
                                            llvm_v16i16_ty, llvm_i16_ty],
                         [IntrNoMem, ImmArg<1>]>;
   def int_x86_avx512_mask_vcvtps2ph_256 : GCCBuiltin<"__builtin_ia32_vcvtps2ph256_mask">,
               Intrinsic<[llvm_v8i16_ty], [llvm_v8f32_ty, llvm_i32_ty,
                                            llvm_v8i16_ty, llvm_i8_ty],
                         [IntrNoMem, ImmArg<1>]>;

llvm/lib/IR/AutoUpgrade.cpp

if (Name == "addcarryx.u32" || // Added in 8.0
       Name.startswith("avx512.mask.psub.") || // Added in 4.0
       Name.startswith("avx512.mask.pmull.") || // Added in 4.0
       Name.startswith("avx512.mask.cvtdq2pd.") || // Added in 4.0
       Name.startswith("avx512.mask.cvtudq2pd.") || // Added in 4.0
       Name.startswith("avx512.mask.cvtudq2ps.") || // Added in 7.0 updated 9.0
       Name.startswith("avx512.mask.cvtqq2pd.") || // Added in 7.0 updated 9.0
       Name.startswith("avx512.mask.cvtuqq2pd.") || // Added in 7.0 updated 9.0
       Name.startswith("avx512.mask.cvtdq2ps.") || // Added in 7.0 updated 9.0
+      Name == "avx512.mask.vcvtph2ps.128" || // Added in 11.0
+      Name == "avx512.mask.vcvtph2ps.256" || // Added in 11.0
       Name == "avx512.mask.cvtqq2ps.256" || // Added in 9.0
       Name == "avx512.mask.cvtqq2ps.512" || // Added in 9.0
       Name == "avx512.mask.cvtuqq2ps.256" || // Added in 9.0
       Name == "avx512.mask.cvtuqq2ps.512" || // Added in 9.0
       Name == "avx512.mask.cvtpd2dq.256" || // Added in 7.0
       Name == "avx512.mask.cvtpd2ps.256" || // Added in 7.0
       Name == "avx512.mask.cvttpd2dq.256" || // Added in 7.0
       Name == "avx512.mask.cvttps2dq.128" || // Added in 7.0
if (Name == "addcarryx.u32" || // Added in 8.0
       Name == "sse2.cvtsi642sd" || // Added in 7.0
       Name == "sse2.cvtss2sd" || // Added in 7.0
       Name == "sse2.cvtdq2pd" || // Added in 3.9
       Name == "sse2.cvtdq2ps" || // Added in 7.0
       Name == "sse2.cvtps2pd" || // Added in 3.9
       Name == "avx.cvtdq2.pd.256" || // Added in 3.9
       Name == "avx.cvtdq2.ps.256" || // Added in 7.0
       Name == "avx.cvt.ps2.pd.256" || // Added in 3.9
+      Name.startswith("vcvtph2ps.") || // Added in 11.0
       Name.startswith("avx.vinsertf128.") || // Added in 3.7
       Name == "avx2.vinserti128" || // Added in 3.7
       Name.startswith("avx512.mask.insert") || // Added in 4.0
       Name.startswith("avx.vextractf128.") || // Added in 3.7
       Name == "avx2.vextracti128" || // Added in 3.7
       Name.startswith("avx512.mask.vextract") || // Added in 4.0
       Name.startswith("sse4a.movnt.") || // Added in 3.9
       Name.startswith("avx.movnt.") || // Added in 3.2
if (IsX86 && (Name.startswith("sse2.pcmp") ||
       } else {
         Rep = IsUnsigned ? Builder.CreateUIToFP(Rep, DstTy, "cvt")
                          : Builder.CreateSIToFP(Rep, DstTy, "cvt");
       }
 
       if (CI->getNumArgOperands() >= 3)
         Rep = EmitX86Select(Builder, CI->getArgOperand(2), Rep,
                             CI->getArgOperand(1));
+    } else if (IsX86 && (Name.startswith("avx512.mask.vcvtph2ps.") ||
+                         Name.startswith("vcvtph2ps."))) {
+      Type *DstTy = CI->getType();
+      Rep = CI->getArgOperand(0);
+      Type *SrcTy = Rep->getType();
+      unsigned NumDstElts = DstTy->getVectorNumElements();
+      if (NumDstElts != SrcTy->getVectorNumElements()) {
+        assert(NumDstElts == 4 && "Unexpected vector size");
+        uint32_t ShuffleMask[4] = {0, 1, 2, 3};
+        Rep = Builder.CreateShuffleVector(Rep, Rep, ShuffleMask);
+      }
+      Rep = Builder.CreateBitCast(
+          Rep, VectorType::get(Type::getHalfTy(C), NumDstElts));
+      Rep = Builder.CreateFPExt(Rep, DstTy, "cvtph2ps");
+      if (CI->getNumArgOperands() >= 3)
+        Rep = EmitX86Select(Builder, CI->getArgOperand(2), Rep,
+                            CI->getArgOperand(1));
     } else if (IsX86 && (Name.startswith("avx512.mask.loadu."))) {
       Rep = UpgradeMaskedLoad(Builder, CI->getArgOperand(0),
                               CI->getArgOperand(1), CI->getArgOperand(2),
                               /*Aligned*/false);
     } else if (IsX86 && (Name.startswith("avx512.mask.load."))) {
       Rep = UpgradeMaskedLoad(Builder, CI->getArgOperand(0),
                               CI->getArgOperand(1),CI->getArgOperand(2),
                               /*Aligned*/true);

llvm/lib/Target/X86/X86IntrinsicsInfo.h

static const IntrinsicData  IntrinsicsWithoutChain[] = {
   X86_INTRINSIC_DATA(avx512_mask_sqrt_sd, INTR_TYPE_SCALAR_MASK,
                      X86ISD::FSQRTS, X86ISD::FSQRTS_RND),
   X86_INTRINSIC_DATA(avx512_mask_sqrt_ss, INTR_TYPE_SCALAR_MASK,
                      X86ISD::FSQRTS, X86ISD::FSQRTS_RND),
   X86_INTRINSIC_DATA(avx512_mask_sub_sd_round, INTR_TYPE_SCALAR_MASK,
                      X86ISD::FSUBS, X86ISD::FSUBS_RND),
   X86_INTRINSIC_DATA(avx512_mask_sub_ss_round, INTR_TYPE_SCALAR_MASK,
                      X86ISD::FSUBS, X86ISD::FSUBS_RND),
-  X86_INTRINSIC_DATA(avx512_mask_vcvtph2ps_128, INTR_TYPE_1OP_MASK,
-                     X86ISD::CVTPH2PS, 0),
-  X86_INTRINSIC_DATA(avx512_mask_vcvtph2ps_256, INTR_TYPE_1OP_MASK,
-                     X86ISD::CVTPH2PS, 0),
   X86_INTRINSIC_DATA(avx512_mask_vcvtph2ps_512, INTR_TYPE_1OP_MASK_SAE,
                      X86ISD::CVTPH2PS, X86ISD::CVTPH2PS_SAE),
   X86_INTRINSIC_DATA(avx512_mask_vcvtps2ph_128, CVTPS2PH_MASK,
                      X86ISD::CVTPS2PH, X86ISD::MCVTPS2PH),
   X86_INTRINSIC_DATA(avx512_mask_vcvtps2ph_256, CVTPS2PH_MASK,
                      X86ISD::CVTPS2PH, X86ISD::MCVTPS2PH),
   X86_INTRINSIC_DATA(avx512_mask_vcvtps2ph_512, CVTPS2PH_MASK,
                      X86ISD::CVTPS2PH, X86ISD::MCVTPS2PH),
static const IntrinsicData  IntrinsicsWithoutChain[] = {
   X86_INTRINSIC_DATA(ssse3_phsub_w_128, INTR_TYPE_2OP, X86ISD::HSUB, 0),
   X86_INTRINSIC_DATA(ssse3_pmadd_ub_sw_128, INTR_TYPE_2OP, X86ISD::VPMADDUBSW, 0),
   X86_INTRINSIC_DATA(ssse3_pmul_hr_sw_128, INTR_TYPE_2OP, X86ISD::MULHRS, 0),
   X86_INTRINSIC_DATA(ssse3_pshuf_b_128, INTR_TYPE_2OP, X86ISD::PSHUFB, 0),
   X86_INTRINSIC_DATA(subborrow_32,      ADX, X86ISD::SBB, X86ISD::SUB),
   X86_INTRINSIC_DATA(subborrow_64,      ADX, X86ISD::SBB, X86ISD::SUB),
   X86_INTRINSIC_DATA(tbm_bextri_u32,    BEXTRI, X86ISD::BEXTR, 0),
   X86_INTRINSIC_DATA(tbm_bextri_u64,    BEXTRI, X86ISD::BEXTR, 0),
-  X86_INTRINSIC_DATA(vcvtph2ps_128,     INTR_TYPE_1OP, X86ISD::CVTPH2PS, 0),
-  X86_INTRINSIC_DATA(vcvtph2ps_256,     INTR_TYPE_1OP, X86ISD::CVTPH2PS, 0),
   X86_INTRINSIC_DATA(vcvtps2ph_128,     INTR_TYPE_2OP, X86ISD::CVTPS2PH, 0),
   X86_INTRINSIC_DATA(vcvtps2ph_256,     INTR_TYPE_2OP, X86ISD::CVTPS2PH, 0),
 
   X86_INTRINSIC_DATA(vgf2p8affineinvqb_128, INTR_TYPE_3OP,
                      X86ISD::GF2P8AFFINEINVQB, 0),
   X86_INTRINSIC_DATA(vgf2p8affineinvqb_256, INTR_TYPE_3OP,
                      X86ISD::GF2P8AFFINEINVQB, 0),
   X86_INTRINSIC_DATA(vgf2p8affineinvqb_512, INTR_TYPE_3OP,

llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp

if (auto *MaskC = dyn_cast<ConstantInt>(II->getArgOperand(1))) {
           Mask &= Mask - 1;
         }
 
         return replaceInstUsesWith(CI, ConstantInt::get(II->getType(), Result));
       }
     }
     break;
 
-  case Intrinsic::x86_vcvtph2ps_128:
-  case Intrinsic::x86_vcvtph2ps_256: {
-    auto Arg = II->getArgOperand(0);
-    auto ArgType = cast<VectorType>(Arg->getType());
-    auto RetType = cast<VectorType>(II->getType());
-    unsigned ArgWidth = ArgType->getNumElements();
-    unsigned RetWidth = RetType->getNumElements();
-    assert(RetWidth <= ArgWidth && "Unexpected input/return vector widths");
-    assert(ArgType->isIntOrIntVectorTy() &&
-           ArgType->getScalarSizeInBits() == 16 &&
-           "CVTPH2PS input type should be 16-bit integer vector");
-    assert(RetType->getScalarType()->isFloatTy() &&
-           "CVTPH2PS output type should be 32-bit float vector");
-
-    // Constant folding: Convert to generic half to single conversion.
-    if (isa<ConstantAggregateZero>(Arg))
-      return replaceInstUsesWith(*II, ConstantAggregateZero::get(RetType));
-
-    if (isa<ConstantDataVector>(Arg)) {
-      auto VectorHalfAsShorts = Arg;
-      if (RetWidth < ArgWidth) {
-        SmallVector<uint32_t, 8> SubVecMask;
-        for (unsigned i = 0; i != RetWidth; ++i)
-          SubVecMask.push_back((int)i);
-        VectorHalfAsShorts = Builder.CreateShuffleVector(
-            Arg, UndefValue::get(ArgType), SubVecMask);
-      }
-
-      auto VectorHalfType =
-          VectorType::get(Type::getHalfTy(II->getContext()), RetWidth);
-      auto VectorHalfs =
-          Builder.CreateBitCast(VectorHalfAsShorts, VectorHalfType);
-      auto VectorFloats = Builder.CreateFPExt(VectorHalfs, RetType);
-      return replaceInstUsesWith(*II, VectorFloats);
-    }
-
-    // We only use the lowest lanes of the argument.
-    if (Value *V = SimplifyDemandedVectorEltsLow(Arg, ArgWidth, RetWidth)) {
-      II->setArgOperand(0, V);
-      return II;
-    }
-    break;
-  }
-
   case Intrinsic::x86_sse_cvtss2si:
   case Intrinsic::x86_sse_cvtss2si64:
   case Intrinsic::x86_sse_cvttss2si:
   case Intrinsic::x86_sse_cvttss2si64:
   case Intrinsic::x86_sse2_cvtsd2si:
   case Intrinsic::x86_sse2_cvtsd2si64:
   case Intrinsic::x86_sse2_cvttsd2si:
   case Intrinsic::x86_sse2_cvttsd2si64:

llvm/test/CodeGen/X86/avx512-intrinsics-fast-isel.ll

 ; X64-NEXT:    vcvtusi2ss %rdi, %xmm0, %xmm0
 ; X64-NEXT:    retq
 entry:
   %conv.i = uitofp i64 %__B to float
   %vecins.i = insertelement <4 x float> %__A, float %conv.i, i32 0
   ret <4 x float> %vecins.i
 }
 
+define <16 x float> @test_mm512_cvtph_ps(<4 x i64> %__A) {
+; CHECK-LABEL: test_mm512_cvtph_ps:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    vcvtph2ps %ymm0, %zmm0
+; CHECK-NEXT:    ret{{[l|q]}}
+entry:
+  %0 = bitcast <4 x i64> %__A to <16 x i16>
+  %1 = bitcast <16 x i16> %0 to <16 x half>
+  %2 = fpext <16 x half> %1 to <16 x float>
+  ret <16 x float> %2
+}
+
+define <16 x float> @test_mm512_mask_cvtph_ps(<16 x float> %__W, i16 zeroext %__U, <4 x i64> %__A) {
+; X86-LABEL: test_mm512_mask_cvtph_ps:
+; X86:       # %bb.0: # %entry
+; X86-NEXT:    movzwl {{[0-9]+}}(%esp), %eax
+; X86-NEXT:    kmovw %eax, %k1
+; X86-NEXT:    vcvtph2ps %ymm1, %zmm0 {%k1}
+; X86-NEXT:    retl
+;
+; X64-LABEL: test_mm512_mask_cvtph_ps:
+; X64:       # %bb.0: # %entry
+; X64-NEXT:    kmovw %edi, %k1
+; X64-NEXT:    vcvtph2ps %ymm1, %zmm0 {%k1}
+; X64-NEXT:    retq
+entry:
+  %0 = bitcast <4 x i64> %__A to <16 x i16>
+  %1 = bitcast <16 x i16> %0 to <16 x half>
+  %2 = bitcast i16 %__U to <16 x i1>
+  %3 = fpext <16 x half> %1 to <16 x float>
+  %4 = select <16 x i1> %2, <16 x float> %3, <16 x float> %__W
+  ret <16 x float> %4
+}
+
+define <16 x float> @test_mm512_maskz_cvtph_ps(i16 zeroext %__U, <4 x i64> %__A) {
+; X86-LABEL: test_mm512_maskz_cvtph_ps:
+; X86:       # %bb.0: # %entry
+; X86-NEXT:    movzwl {{[0-9]+}}(%esp), %eax
+; X86-NEXT:    kmovw %eax, %k1
+; X86-NEXT:    vcvtph2ps %ymm0, %zmm0 {%k1} {z}
+; X86-NEXT:    retl
+;
+; X64-LABEL: test_mm512_maskz_cvtph_ps:
+; X64:       # %bb.0: # %entry
+; X64-NEXT:    kmovw %edi, %k1
+; X64-NEXT:    vcvtph2ps %ymm0, %zmm0 {%k1} {z}
+; X64-NEXT:    retq
+entry:
+  %0 = bitcast <4 x i64> %__A to <16 x i16>
+  %1 = bitcast <16 x i16> %0 to <16 x half>
+  %2 = bitcast i16 %__U to <16 x i1>
+  %3 = fpext <16 x half> %1 to <16 x float>
+  %4 = select <16 x i1> %2, <16 x float> %3, <16 x float> zeroinitializer
+  ret <16 x float> %4
+}
+
 define <8 x double> @test_mm512_cvtps_pd(<8 x float> %__A) {
 ; CHECK-LABEL: test_mm512_cvtps_pd:
 ; CHECK:       # %bb.0: # %entry
 ; CHECK-NEXT:    vcvtps2pd %ymm0, %zmm0
 ; CHECK-NEXT:    ret{{[l|q]}}
 entry:
   %conv.i = fpext <8 x float> %__A to <8 x double>
   ret <8 x double> %conv.i

llvm/test/CodeGen/X86/avx512-intrinsics-upgrade.ll

 ; X64-NEXT:    vaddpd %zmm2, %zmm1, %zmm0 ## encoding: [0x62,0xf1,0xf5,0x48,0x58,0xc2]
 ; X64-NEXT:    retq ## encoding: [0xc3]
   %res = call <8 x double> @llvm.x86.avx512.mask.cvtudq2pd.512(<8 x i32> %x0, <8 x double> %x1, i8 %x2)
   %res1 = call <8 x double> @llvm.x86.avx512.mask.cvtudq2pd.512(<8 x i32> %x0, <8 x double> %x1, i8 -1)
   %res2 = fadd <8 x double> %res, %res1
   ret <8 x double> %res2
 }
 
+define <16 x float> @test_x86_vcvtph2ps_512(<16 x i16> %a0) {
+; CHECK-LABEL: test_x86_vcvtph2ps_512:
+; CHECK:       ## %bb.0:
+; CHECK-NEXT:    vcvtph2ps %ymm0, %zmm0 ## encoding: [0x62,0xf2,0x7d,0x48,0x13,0xc0]
+; CHECK-NEXT:    ret{{[l|q]}} ## encoding: [0xc3]
+  %res = call <16 x float> @llvm.x86.avx512.mask.vcvtph2ps.512(<16 x i16> %a0, <16 x float> zeroinitializer, i16 -1, i32 4)
+  ret <16 x float> %res
+}
+
+define <16 x float> @test_x86_vcvtph2ps_512_sae(<16 x i16> %a0) {
+; CHECK-LABEL: test_x86_vcvtph2ps_512_sae:
+; CHECK:       ## %bb.0:
+; CHECK-NEXT:    vcvtph2ps {sae}, %ymm0, %zmm0 ## encoding: [0x62,0xf2,0x7d,0x18,0x13,0xc0]
+; CHECK-NEXT:    ret{{[l|q]}} ## encoding: [0xc3]
+  %res = call <16 x float> @llvm.x86.avx512.mask.vcvtph2ps.512(<16 x i16> %a0, <16 x float> zeroinitializer, i16 -1, i32 8)
+  ret <16 x float> %res
+}
+
+define <16 x float> @test_x86_vcvtph2ps_512_rrk(<16 x i16> %a0,<16 x float> %a1, i16 %mask) {
+; X86-LABEL: test_x86_vcvtph2ps_512_rrk:
+; X86:       ## %bb.0:
+; X86-NEXT:    kmovw {{[0-9]+}}(%esp), %k1 ## encoding: [0xc5,0xf8,0x90,0x4c,0x24,0x04]
+; X86-NEXT:    vcvtph2ps %ymm0, %zmm1 {%k1} ## encoding: [0x62,0xf2,0x7d,0x49,0x13,0xc8]
+; X86-NEXT:    vmovaps %zmm1, %zmm0 ## encoding: [0x62,0xf1,0x7c,0x48,0x28,0xc1]
+; X86-NEXT:    retl ## encoding: [0xc3]
+;
+; X64-LABEL: test_x86_vcvtph2ps_512_rrk:
+; X64:       ## %bb.0:
+; X64-NEXT:    kmovw %edi, %k1 ## encoding: [0xc5,0xf8,0x92,0xcf]
+; X64-NEXT:    vcvtph2ps %ymm0, %zmm1 {%k1} ## encoding: [0x62,0xf2,0x7d,0x49,0x13,0xc8]
+; X64-NEXT:    vmovaps %zmm1, %zmm0 ## encoding: [0x62,0xf1,0x7c,0x48,0x28,0xc1]
+; X64-NEXT:    retq ## encoding: [0xc3]
+  %res = call <16 x float> @llvm.x86.avx512.mask.vcvtph2ps.512(<16 x i16> %a0, <16 x float> %a1, i16 %mask, i32 4)
+  ret <16 x float> %res
+}
+
+define <16 x float> @test_x86_vcvtph2ps_512_sae_rrkz(<16 x i16> %a0, i16 %mask) {
+; X86-LABEL: test_x86_vcvtph2ps_512_sae_rrkz:
+; X86:       ## %bb.0:
+; X86-NEXT:    kmovw {{[0-9]+}}(%esp), %k1 ## encoding: [0xc5,0xf8,0x90,0x4c,0x24,0x04]
+; X86-NEXT:    vcvtph2ps {sae}, %ymm0, %zmm0 {%k1} {z} ## encoding: [0x62,0xf2,0x7d,0x99,0x13,0xc0]
+; X86-NEXT:    retl ## encoding: [0xc3]
+;
+; X64-LABEL: test_x86_vcvtph2ps_512_sae_rrkz:
+; X64:       ## %bb.0:
+; X64-NEXT:    kmovw %edi, %k1 ## encoding: [0xc5,0xf8,0x92,0xcf]
+; X64-NEXT:    vcvtph2ps {sae}, %ymm0, %zmm0 {%k1} {z} ## encoding: [0x62,0xf2,0x7d,0x99,0x13,0xc0]
+; X64-NEXT:    retq ## encoding: [0xc3]
+  %res = call <16 x float> @llvm.x86.avx512.mask.vcvtph2ps.512(<16 x i16> %a0, <16 x float> zeroinitializer, i16 %mask, i32 8)
+  ret <16 x float> %res
+}
+
+define <16 x float> @test_x86_vcvtph2ps_512_rrkz(<16 x i16> %a0, i16 %mask) {
+; X86-LABEL: test_x86_vcvtph2ps_512_rrkz:
+; X86:       ## %bb.0:
+; X86-NEXT:    kmovw {{[0-9]+}}(%esp), %k1 ## encoding: [0xc5,0xf8,0x90,0x4c,0x24,0x04]
+; X86-NEXT:    vcvtph2ps %ymm0, %zmm0 {%k1} {z} ## encoding: [0x62,0xf2,0x7d,0xc9,0x13,0xc0]
+; X86-NEXT:    retl ## encoding: [0xc3]
+;
+; X64-LABEL: test_x86_vcvtph2ps_512_rrkz:
+; X64:       ## %bb.0:
+; X64-NEXT:    kmovw %edi, %k1 ## encoding: [0xc5,0xf8,0x92,0xcf]
+; X64-NEXT:    vcvtph2ps %ymm0, %zmm0 {%k1} {z} ## encoding: [0x62,0xf2,0x7d,0xc9,0x13,0xc0]
+; X64-NEXT:    retq ## encoding: [0xc3]
+  %res = call <16 x float> @llvm.x86.avx512.mask.vcvtph2ps.512(<16 x i16> %a0, <16 x float> zeroinitializer, i16 %mask, i32 4)
+  ret <16 x float> %res
+}
+
+declare <16 x float> @llvm.x86.avx512.mask.vcvtph2ps.512(<16 x i16>, <16 x float>, i16, i32) nounwind readonly
+
 define <8 x i64> @test_valign_q(<8 x i64> %a, <8 x i64> %b) {
 ; CHECK-LABEL: test_valign_q:
 ; CHECK:       ## %bb.0:
 ; CHECK-NEXT:    valignq $2, %zmm1, %zmm0, %zmm0 ## encoding: [0x62,0xf3,0xfd,0x48,0x03,0xc1,0x02]
 ; CHECK-NEXT:    ## zmm0 = zmm1[2,3,4,5,6,7],zmm0[0,1]
 ; CHECK-NEXT:    ret{{[l|q]}} ## encoding: [0xc3]
   %res = call <8 x i64> @llvm.x86.avx512.mask.valign.q.512(<8 x i64> %a, <8 x i64> %b, i32 2, <8 x i64> zeroinitializer, i8 -1)
   ret <8 x i64> %res

 
 ; Test case to make sure we can print shuffle decode comments for constant pool loads.
 define <16 x float>@test_int_x86_avx512_mask_vpermilvar_ps_512_constant_pool(<16 x float> %x0, <16 x i32> %x1, <16 x float> %x2, i16 %x3) {
 ; X86-LABEL: test_int_x86_avx512_mask_vpermilvar_ps_512_constant_pool:
 ; X86:       ## %bb.0:
 ; X86-NEXT:    kmovw {{[0-9]+}}(%esp), %k1 ## encoding: [0xc5,0xf8,0x90,0x4c,0x24,0x04]
 ; X86-NEXT:    vpermilps {{.*#+}} zmm2 {%k1} = zmm0[2,3,0,1,7,6,5,4,9,8,11,10,12,13,14,15]
 ; X86-NEXT:    ## encoding: [0x62,0xf2,0x7d,0x49,0x0c,0x15,A,A,A,A]
-; X86-NEXT:    ## fixup A - offset: 6, value: LCPI211_0, kind: FK_Data_4
+; X86-NEXT:    ## fixup A - offset: 6, value: LCPI216_0, kind: FK_Data_4
 ; X86-NEXT:    vpermilps {{.*#+}} zmm1 {%k1} {z} = zmm0[1,0,3,2,4,5,6,7,9,8,11,10,12,13,14,15]
 ; X86-NEXT:    ## encoding: [0x62,0xf2,0x7d,0xc9,0x0c,0x0d,A,A,A,A]
-; X86-NEXT:    ## fixup A - offset: 6, value: LCPI211_1, kind: FK_Data_4
+; X86-NEXT:    ## fixup A - offset: 6, value: LCPI216_1, kind: FK_Data_4
 ; X86-NEXT:    vaddps %zmm1, %zmm2, %zmm1 ## encoding: [0x62,0xf1,0x6c,0x48,0x58,0xc9]
 ; X86-NEXT:    vpermilps {{.*#+}} zmm0 = zmm0[1,0,3,2,4,5,6,7,10,11,8,9,14,15,13,12]
 ; X86-NEXT:    ## encoding: [0x62,0xf2,0x7d,0x48,0x0c,0x05,A,A,A,A]
-; X86-NEXT:    ## fixup A - offset: 6, value: LCPI211_2, kind: FK_Data_4
+; X86-NEXT:    ## fixup A - offset: 6, value: LCPI216_2, kind: FK_Data_4
 ; X86-NEXT:    vaddps %zmm1, %zmm0, %zmm0 ## encoding: [0x62,0xf1,0x7c,0x48,0x58,0xc1]
 ; X86-NEXT:    retl ## encoding: [0xc3]
 ;
 ; X64-LABEL: test_int_x86_avx512_mask_vpermilvar_ps_512_constant_pool:
 ; X64:       ## %bb.0:
 ; X64-NEXT:    kmovw %edi, %k1 ## encoding: [0xc5,0xf8,0x92,0xcf]
 ; X64-NEXT:    vpermilps {{.*#+}} zmm2 {%k1} = zmm0[2,3,0,1,7,6,5,4,9,8,11,10,12,13,14,15]
 ; X64-NEXT:    ## encoding: [0x62,0xf2,0x7d,0x49,0x0c,0x15,A,A,A,A]
-; X64-NEXT:    ## fixup A - offset: 6, value: LCPI211_0-4, kind: reloc_riprel_4byte
+; X64-NEXT:    ## fixup A - offset: 6, value: LCPI216_0-4, kind: reloc_riprel_4byte
 ; X64-NEXT:    vpermilps {{.*#+}} zmm1 {%k1} {z} = zmm0[1,0,3,2,4,5,6,7,9,8,11,10,12,13,14,15]
 ; X64-NEXT:    ## encoding: [0x62,0xf2,0x7d,0xc9,0x0c,0x0d,A,A,A,A]
-; X64-NEXT:    ## fixup A - offset: 6, value: LCPI211_1-4, kind: reloc_riprel_4byte
+; X64-NEXT:    ## fixup A - offset: 6, value: LCPI216_1-4, kind: reloc_riprel_4byte
 ; X64-NEXT:    vaddps %zmm1, %zmm2, %zmm1 ## encoding: [0x62,0xf1,0x6c,0x48,0x58,0xc9]
 ; X64-NEXT:    vpermilps {{.*#+}} zmm0 = zmm0[1,0,3,2,4,5,6,7,10,11,8,9,14,15,13,12]
 ; X64-NEXT:    ## encoding: [0x62,0xf2,0x7d,0x48,0x0c,0x05,A,A,A,A]
-; X64-NEXT:    ## fixup A - offset: 6, value: LCPI211_2-4, kind: reloc_riprel_4byte
+; X64-NEXT:    ## fixup A - offset: 6, value: LCPI216_2-4, kind: reloc_riprel_4byte
 ; X64-NEXT:    vaddps %zmm1, %zmm0, %zmm0 ## encoding: [0x62,0xf1,0x7c,0x48,0x58,0xc1]
 ; X64-NEXT:    retq ## encoding: [0xc3]
   %res = call <16 x float> @llvm.x86.avx512.mask.vpermilvar.ps.512(<16 x float> %x0, <16 x i32> <i32 2, i32 3, i32 0, i32 1, i32 3, i32 2, i32 1, i32 0, i32 1, i32 0, i32 3, i32 2, i32 0, i32 1, i32 2, i32 3>, <16 x float> %x2, i16 %x3)
   %res1 = call <16 x float> @llvm.x86.avx512.mask.vpermilvar.ps.512(<16 x float> %x0, <16 x i32> <i32 1, i32 0, i32 3, i32 2, i32 0, i32 1, i32 2, i32 3, i32 1, i32 0, i32 3, i32 2, i32 0, i32 1, i32 2, i32 3>, <16 x float> zeroinitializer, i16 %x3)
   %res2 = call <16 x float> @llvm.x86.avx512.mask.vpermilvar.ps.512(<16 x float> %x0, <16 x i32> <i32 1, i32 0, i32 3, i32 2, i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 1, i32 0>, <16 x float> %x2, i16 -1)
   %res3 = fadd <16 x float> %res, %res1
   %res4 = fadd <16 x float> %res2, %res3
   ret <16 x float> %res4

llvm/test/CodeGen/X86/avx512-intrinsics.ll

; CHECK-NEXT:    ret{{[l|q]}}
   %res1 = call i32 @llvm.x86.avx512.vcvtss2si32(<4 x float> %a0, i32 11)
   %res2 = call i32 @llvm.x86.avx512.vcvtss2si32(<4 x float> %a0, i32 9)
   %res3 = add i32 %res, %res1
   %res4 = add i32 %res3, %res2
   ret i32 %res4
 }
 declare i32 @llvm.x86.avx512.vcvtss2si32(<4 x float>, i32) nounwind readnone
 
-define <16 x float> @test_x86_vcvtph2ps_512(<16 x i16> %a0) {
-; CHECK-LABEL: test_x86_vcvtph2ps_512:
-; CHECK:       # %bb.0:
-; CHECK-NEXT:    vcvtph2ps %ymm0, %zmm0
-; CHECK-NEXT:    ret{{[l|q]}}
-  %res = call <16 x float> @llvm.x86.avx512.mask.vcvtph2ps.512(<16 x i16> %a0, <16 x float> zeroinitializer, i16 -1, i32 4)
-  ret <16 x float> %res
-}
-
-define <16 x float> @test_x86_vcvtph2ps_512_sae(<16 x i16> %a0) {
-; CHECK-LABEL: test_x86_vcvtph2ps_512_sae:
-; CHECK:       # %bb.0:
-; CHECK-NEXT:    vcvtph2ps {sae}, %ymm0, %zmm0
-; CHECK-NEXT:    ret{{[l|q]}}
-  %res = call <16 x float> @llvm.x86.avx512.mask.vcvtph2ps.512(<16 x i16> %a0, <16 x float> zeroinitializer, i16 -1, i32 8)
-  ret <16 x float> %res
-}
-
-define <16 x float> @test_x86_vcvtph2ps_512_rrk(<16 x i16> %a0,<16 x float> %a1, i16 %mask) {
-; X64-LABEL: test_x86_vcvtph2ps_512_rrk:
-; X64:       # %bb.0:
-; X64-NEXT:    kmovw %edi, %k1
-; X64-NEXT:    vcvtph2ps %ymm0, %zmm1 {%k1}
-; X64-NEXT:    vmovaps %zmm1, %zmm0
-; X64-NEXT:    retq
-;
-; X86-LABEL: test_x86_vcvtph2ps_512_rrk:
-; X86:       # %bb.0:
-; X86-NEXT:    kmovw {{[0-9]+}}(%esp), %k1
-; X86-NEXT:    vcvtph2ps %ymm0, %zmm1 {%k1}
-; X86-NEXT:    vmovaps %zmm1, %zmm0
-; X86-NEXT:    retl
-  %res = call <16 x float> @llvm.x86.avx512.mask.vcvtph2ps.512(<16 x i16> %a0, <16 x float> %a1, i16 %mask, i32 4)
-  ret <16 x float> %res
-}
-
-define <16 x float> @test_x86_vcvtph2ps_512_sae_rrkz(<16 x i16> %a0, i16 %mask) {
-; X64-LABEL: test_x86_vcvtph2ps_512_sae_rrkz:
-; X64:       # %bb.0:
-; X64-NEXT:    kmovw %edi, %k1
-; X64-NEXT:    vcvtph2ps {sae}, %ymm0, %zmm0 {%k1} {z}
-; X64-NEXT:    retq
-;
-; X86-LABEL: test_x86_vcvtph2ps_512_sae_rrkz:
-; X86:       # %bb.0:
-; X86-NEXT:    kmovw {{[0-9]+}}(%esp), %k1
-; X86-NEXT:    vcvtph2ps {sae}, %ymm0, %zmm0 {%k1} {z}
-; X86-NEXT:    retl
-  %res = call <16 x float> @llvm.x86.avx512.mask.vcvtph2ps.512(<16 x i16> %a0, <16 x float> zeroinitializer, i16 %mask, i32 8)
-  ret <16 x float> %res
-}
-
-define <16 x float> @test_x86_vcvtph2ps_512_rrkz(<16 x i16> %a0, i16 %mask) {
-; X64-LABEL: test_x86_vcvtph2ps_512_rrkz:
-; X64:       # %bb.0:
-; X64-NEXT:    kmovw %edi, %k1
-; X64-NEXT:    vcvtph2ps %ymm0, %zmm0 {%k1} {z}
-; X64-NEXT:    retq
-;
-; X86-LABEL: test_x86_vcvtph2ps_512_rrkz:
-; X86:       # %bb.0:
-; X86-NEXT:    kmovw {{[0-9]+}}(%esp), %k1
-; X86-NEXT:    vcvtph2ps %ymm0, %zmm0 {%k1} {z}
-; X86-NEXT:    retl
-  %res = call <16 x float> @llvm.x86.avx512.mask.vcvtph2ps.512(<16 x i16> %a0, <16 x float> zeroinitializer, i16 %mask, i32 4)
-  ret <16 x float> %res
-}
-
-declare <16 x float> @llvm.x86.avx512.mask.vcvtph2ps.512(<16 x i16>, <16 x float>, i16, i32) nounwind readonly
-
 define <16 x i16> @test_x86_vcvtps2ph_256(<16 x float> %a0, <16 x i16> %src, i16 %mask, <16 x i16> * %dst) {
 ; X64-LABEL: test_x86_vcvtps2ph_256:
 ; X64:       # %bb.0:
 ; X64-NEXT:    kmovw %edi, %k1
 ; X64-NEXT:    vcvtps2ph $2, %zmm0, %ymm2 {%k1} {z}
 ; X64-NEXT:    vcvtps2ph $2, %zmm0, %ymm1 {%k1}
 ; X64-NEXT:    vpaddw %ymm1, %ymm2, %ymm1
 ; X64-NEXT:    vcvtps2ph $2, %zmm0, (%rsi)

llvm/test/CodeGen/X86/avx512vl-intrinsics-fast-isel.ll

 ; X64-NEXT:    vzeroupper
 ; X64-NEXT:    retq
 entry:
   %0 = tail call <4 x i32> @llvm.x86.avx512.mask.cvtpd2udq.256(<4 x double> %__A, <4 x i32> zeroinitializer, i8 %__U) #8
   %1 = bitcast <4 x i32> %0 to <2 x i64>
   ret <2 x i64> %1
 }
 
+define <4 x float> @test_mm_mask_cvtph_ps(<4 x float> %__W, i8 zeroext %__U, <2 x i64> %__A) {
+; X86-LABEL: test_mm_mask_cvtph_ps:
+; X86:       # %bb.0: # %entry
+; X86-NEXT:    movb {{[0-9]+}}(%esp), %al
+; X86-NEXT:    kmovw %eax, %k1
+; X86-NEXT:    vcvtph2ps %xmm1, %xmm0 {%k1}
+; X86-NEXT:    retl
+;
+; X64-LABEL: test_mm_mask_cvtph_ps:
+; X64:       # %bb.0: # %entry
+; X64-NEXT:    kmovw %edi, %k1
+; X64-NEXT:    vcvtph2ps %xmm1, %xmm0 {%k1}
+; X64-NEXT:    retq
+entry:
+  %0 = bitcast <2 x i64> %__A to <8 x i16>
+  %1 = shufflevector <8 x i16> %0, <8 x i16> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  %2 = bitcast <4 x i16> %1 to <4 x half>
+  %3 = bitcast i8 %__U to <8 x i1>
+  %4 = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  %5 = fpext <4 x half> %2 to <4 x float>
+  %6 = select <4 x i1> %4, <4 x float> %5, <4 x float> %__W
+  ret <4 x float> %6
+}
+
+define <4 x float> @test_mm_maskz_cvtph_ps(i8 zeroext %__U, <2 x i64> %__A) {
+; X86-LABEL: test_mm_maskz_cvtph_ps:
+; X86:       # %bb.0: # %entry
+; X86-NEXT:    movb {{[0-9]+}}(%esp), %al
+; X86-NEXT:    kmovw %eax, %k1
+; X86-NEXT:    vcvtph2ps %xmm0, %xmm0 {%k1} {z}
+; X86-NEXT:    retl
+;
+; X64-LABEL: test_mm_maskz_cvtph_ps:
+; X64:       # %bb.0: # %entry
+; X64-NEXT:    kmovw %edi, %k1
+; X64-NEXT:    vcvtph2ps %xmm0, %xmm0 {%k1} {z}
+; X64-NEXT:    retq
+entry:
+  %0 = bitcast <2 x i64> %__A to <8 x i16>
+  %1 = shufflevector <8 x i16> %0, <8 x i16> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  %2 = bitcast <4 x i16> %1 to <4 x half>
+  %3 = bitcast i8 %__U to <8 x i1>
+  %4 = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  %5 = fpext <4 x half> %2 to <4 x float>
+  %6 = select <4 x i1> %4, <4 x float> %5, <4 x float> zeroinitializer
+  ret <4 x float> %6
+}
+
+define <8 x float> @test_mm256_mask_cvtph_ps(<8 x float> %__W, i8 zeroext %__U, <2 x i64> %__A) {
+; X86-LABEL: test_mm256_mask_cvtph_ps:
+; X86:       # %bb.0: # %entry
+; X86-NEXT:    movb {{[0-9]+}}(%esp), %al
+; X86-NEXT:    kmovw %eax, %k1
+; X86-NEXT:    vcvtph2ps %xmm1, %ymm0 {%k1}
+; X86-NEXT:    retl
+;
+; X64-LABEL: test_mm256_mask_cvtph_ps:
+; X64:       # %bb.0: # %entry
+; X64-NEXT:    kmovw %edi, %k1
+; X64-NEXT:    vcvtph2ps %xmm1, %ymm0 {%k1}
+; X64-NEXT:    retq
+entry:
+  %0 = bitcast <2 x i64> %__A to <8 x i16>
+  %1 = bitcast <8 x i16> %0 to <8 x half>
+  %2 = bitcast i8 %__U to <8 x i1>
+  %3 = fpext <8 x half> %1 to <8 x float>
+  %4 = select <8 x i1> %2, <8 x float> %3, <8 x float> %__W
+  ret <8 x float> %4
+}
+
+define <8 x float> @test_mm256_maskz_cvtph_ps(i8 zeroext %__U, <2 x i64> %__A) {
+; X86-LABEL: test_mm256_maskz_cvtph_ps:
+; X86:       # %bb.0: # %entry
+; X86-NEXT:    movb {{[0-9]+}}(%esp), %al
+; X86-NEXT:    kmovw %eax, %k1
+; X86-NEXT:    vcvtph2ps %xmm0, %ymm0 {%k1} {z}
+; X86-NEXT:    retl
+;
+; X64-LABEL: test_mm256_maskz_cvtph_ps:
+; X64:       # %bb.0: # %entry
+; X64-NEXT:    kmovw %edi, %k1
+; X64-NEXT:    vcvtph2ps %xmm0, %ymm0 {%k1} {z}
+; X64-NEXT:    retq
+entry:
+  %0 = bitcast <2 x i64> %__A to <8 x i16>
+  %1 = bitcast <8 x i16> %0 to <8 x half>
+  %2 = bitcast i8 %__U to <8 x i1>
+  %3 = fpext <8 x half> %1 to <8 x float>
+  %4 = select <8 x i1> %2, <8 x float> %3, <8 x float> zeroinitializer
+  ret <8 x float> %4
+}
+
 define <2 x i64> @test_mm_mask_cvtps_epi32(<2 x i64> %__W, i8 zeroext %__U, <4 x float> %__A) {
 ; X86-LABEL: test_mm_mask_cvtps_epi32:
 ; X86:       # %bb.0: # %entry
 ; X86-NEXT:    movb {{[0-9]+}}(%esp), %al
 ; X86-NEXT:    kmovw %eax, %k1
 ; X86-NEXT:    vcvtps2dq %xmm1, %xmm0 {%k1}
 ; X86-NEXT:    retl
 ;

llvm/test/CodeGen/X86/avx512vl-intrinsics-upgrade.ll

 ; X64-NEXT:    vaddps %ymm2, %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xf4,0x58,0xc2]
 ; X64-NEXT:    retq # encoding: [0xc3]
   %res = call <8 x float> @llvm.x86.avx512.mask.cvtdq2ps.256(<8 x i32> %x0, <8 x float> %x1, i8 %x2)
   %res1 = call <8 x float> @llvm.x86.avx512.mask.cvtdq2ps.256(<8 x i32> %x0, <8 x float> %x1, i8 -1)
   %res2 = fadd <8 x float> %res, %res1
   ret <8 x float> %res2
 }
 
+define <4 x float> @test_x86_vcvtph2ps_128(<8 x i16> %a0) {
+; CHECK-LABEL: test_x86_vcvtph2ps_128:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vcvtph2ps %xmm0, %xmm0 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x79,0x13,0xc0]
+; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
+  %res = call <4 x float> @llvm.x86.avx512.mask.vcvtph2ps.128(<8 x i16> %a0, <4 x float> zeroinitializer, i8 -1)
+  ret <4 x float> %res
+}
+
+define <4 x float> @test_x86_vcvtph2ps_128_rrk(<8 x i16> %a0,<4 x float> %a1, i8 %mask) {
+; X86-LABEL: test_x86_vcvtph2ps_128_rrk:
+; X86:       # %bb.0:
+; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
+; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
+; X86-NEXT:    vcvtph2ps %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x09,0x13,0xc8]
+; X86-NEXT:    vmovaps %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc1]
+; X86-NEXT:    retl # encoding: [0xc3]
+;
+; X64-LABEL: test_x86_vcvtph2ps_128_rrk:
+; X64:       # %bb.0:
+; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
+; X64-NEXT:    vcvtph2ps %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x09,0x13,0xc8]
+; X64-NEXT:    vmovaps %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc1]
+; X64-NEXT:    retq # encoding: [0xc3]
+  %res = call <4 x float> @llvm.x86.avx512.mask.vcvtph2ps.128(<8 x i16> %a0, <4 x float> %a1, i8 %mask)
+  ret <4 x float> %res
+}
+
+define <4 x float> @test_x86_vcvtph2ps_128_rrkz(<8 x i16> %a0, i8 %mask) {
+; X86-LABEL: test_x86_vcvtph2ps_128_rrkz:
+; X86:       # %bb.0:
+; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
+; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
+; X86-NEXT:    vcvtph2ps %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7d,0x89,0x13,0xc0]
+; X86-NEXT:    retl # encoding: [0xc3]
+;
+; X64-LABEL: test_x86_vcvtph2ps_128_rrkz:
+; X64:       # %bb.0:
+; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
+; X64-NEXT:    vcvtph2ps %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7d,0x89,0x13,0xc0]
+; X64-NEXT:    retq # encoding: [0xc3]
+  %res = call <4 x float> @llvm.x86.avx512.mask.vcvtph2ps.128(<8 x i16> %a0, <4 x float> zeroinitializer, i8 %mask)
+  ret <4 x float> %res
+}
+
+declare <4 x float> @llvm.x86.avx512.mask.vcvtph2ps.128(<8 x i16>, <4 x float>, i8) nounwind readonly
+
+define <8 x float> @test_x86_vcvtph2ps_256(<8 x i16> %a0) {
+; CHECK-LABEL: test_x86_vcvtph2ps_256:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vcvtph2ps %xmm0, %ymm0 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x7d,0x13,0xc0]
+; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
+  %res = call <8 x float> @llvm.x86.avx512.mask.vcvtph2ps.256(<8 x i16> %a0, <8 x float> zeroinitializer, i8 -1)
+  ret <8 x float> %res
+}
+
+define <8 x float> @test_x86_vcvtph2ps_256_rrk(<8 x i16> %a0,<8 x float> %a1, i8 %mask) {
+; X86-LABEL: test_x86_vcvtph2ps_256_rrk:
+; X86:       # %bb.0:
+; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
+; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
+; X86-NEXT:    vcvtph2ps %xmm0, %ymm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x29,0x13,0xc8]
+; X86-NEXT:    vmovaps %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfc,0x28,0xc1]
+; X86-NEXT:    retl # encoding: [0xc3]
+;
+; X64-LABEL: test_x86_vcvtph2ps_256_rrk:
+; X64:       # %bb.0:
+; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
+; X64-NEXT:    vcvtph2ps %xmm0, %ymm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x29,0x13,0xc8]
+; X64-NEXT:    vmovaps %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfc,0x28,0xc1]
+; X64-NEXT:    retq # encoding: [0xc3]
+  %res = call <8 x float> @llvm.x86.avx512.mask.vcvtph2ps.256(<8 x i16> %a0, <8 x float> %a1, i8 %mask)
+  ret <8 x float> %res
+}
+
+define <8 x float> @test_x86_vcvtph2ps_256_rrkz(<8 x i16> %a0, i8 %mask) {
+; X86-LABEL: test_x86_vcvtph2ps_256_rrkz:
+; X86:       # %bb.0:
+; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
+; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
+; X86-NEXT:    vcvtph2ps %xmm0, %ymm0 {%k1} {z} # encoding: [0x62,0xf2,0x7d,0xa9,0x13,0xc0]
+; X86-NEXT:    retl # encoding: [0xc3]
+;
+; X64-LABEL: test_x86_vcvtph2ps_256_rrkz:
+; X64:       # %bb.0:
+; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
+; X64-NEXT:    vcvtph2ps %xmm0, %ymm0 {%k1} {z} # encoding: [0x62,0xf2,0x7d,0xa9,0x13,0xc0]
+; X64-NEXT:    retq # encoding: [0xc3]
+  %res = call <8 x float> @llvm.x86.avx512.mask.vcvtph2ps.256(<8 x i16> %a0, <8 x float> zeroinitializer, i8 %mask)
+  ret <8 x float> %res
+}
+
+declare <8 x float> @llvm.x86.avx512.mask.vcvtph2ps.256(<8 x i16>, <8 x float>, i8) nounwind readonly
+
 declare <4 x i32> @llvm.x86.avx512.mask.cvtpd2dq.256(<4 x double>, <4 x i32>, i8)
 
 define <4 x i32>@test_int_x86_avx512_mask_cvt_pd2dq_256(<4 x double> %x0, <4 x i32> %x1, i8 %x2) {
 ; X86-LABEL: test_int_x86_avx512_mask_cvt_pd2dq_256:
 ; X86:       # %bb.0:
 ; X86-NEXT:    vcvtpd2dq %ymm0, %xmm2 # EVEX TO VEX Compression encoding: [0xc5,0xff,0xe6,0xd0]
 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]

llvm/test/CodeGen/X86/avx512vl-intrinsics.ll

; X64-NEXT:    retq # encoding: [0xc3]
   %2 = bitcast i8 %x4 to <8 x i1>
   %extract = shufflevector <8 x i1> %2, <8 x i1> %2, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
   %3 = select <4 x i1> %extract, <4 x i64> %1, <4 x i64> zeroinitializer
   %4 = call <4 x i64> @llvm.x86.avx512.pternlog.q.256(<4 x i64> %x0, <4 x i64> %x1, <4 x i64> %x2, i32 33)
   %res2 = add <4 x i64> %3, %4
   ret <4 x i64> %res2
 }
 
-define <4 x float> @test_x86_vcvtph2ps_128(<8 x i16> %a0) {
-; CHECK-LABEL: test_x86_vcvtph2ps_128:
-; CHECK:       # %bb.0:
-; CHECK-NEXT:    vcvtph2ps %xmm0, %xmm0 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x79,0x13,0xc0]
-; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
-  %res = call <4 x float> @llvm.x86.avx512.mask.vcvtph2ps.128(<8 x i16> %a0, <4 x float> zeroinitializer, i8 -1)
-  ret <4 x float> %res
-}
-
-define <4 x float> @test_x86_vcvtph2ps_128_rrk(<8 x i16> %a0,<4 x float> %a1, i8 %mask) {
-; X86-LABEL: test_x86_vcvtph2ps_128_rrk:
-; X86:       # %bb.0:
-; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
-; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
-; X86-NEXT:    vcvtph2ps %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x09,0x13,0xc8]
-; X86-NEXT:    vmovaps %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc1]
-; X86-NEXT:    retl # encoding: [0xc3]
-;
-; X64-LABEL: test_x86_vcvtph2ps_128_rrk:
-; X64:       # %bb.0:
-; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
-; X64-NEXT:    vcvtph2ps %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x09,0x13,0xc8]
-; X64-NEXT:    vmovaps %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xf8,0x28,0xc1]
-; X64-NEXT:    retq # encoding: [0xc3]
-  %res = call <4 x float> @llvm.x86.avx512.mask.vcvtph2ps.128(<8 x i16> %a0, <4 x float> %a1, i8 %mask)
-  ret <4 x float> %res
-}
-
-
-define <4 x float> @test_x86_vcvtph2ps_128_rrkz(<8 x i16> %a0, i8 %mask) {
-; X86-LABEL: test_x86_vcvtph2ps_128_rrkz:
-; X86:       # %bb.0:
-; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
-; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
-; X86-NEXT:    vcvtph2ps %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7d,0x89,0x13,0xc0]
-; X86-NEXT:    retl # encoding: [0xc3]
-;
-; X64-LABEL: test_x86_vcvtph2ps_128_rrkz:
-; X64:       # %bb.0:
-; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
-; X64-NEXT:    vcvtph2ps %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x7d,0x89,0x13,0xc0]
-; X64-NEXT:    retq # encoding: [0xc3]
-  %res = call <4 x float> @llvm.x86.avx512.mask.vcvtph2ps.128(<8 x i16> %a0, <4 x float> zeroinitializer, i8 %mask)
-  ret <4 x float> %res
-}
-
-declare <4 x float> @llvm.x86.avx512.mask.vcvtph2ps.128(<8 x i16>, <4 x float>, i8) nounwind readonly
-
-define <8 x float> @test_x86_vcvtph2ps_256(<8 x i16> %a0) {
-; CHECK-LABEL: test_x86_vcvtph2ps_256:
-; CHECK:       # %bb.0:
-; CHECK-NEXT:    vcvtph2ps %xmm0, %ymm0 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x7d,0x13,0xc0]
-; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
-  %res = call <8 x float> @llvm.x86.avx512.mask.vcvtph2ps.256(<8 x i16> %a0, <8 x float> zeroinitializer, i8 -1)
-  ret <8 x float> %res
-}
-
-define <8 x float> @test_x86_vcvtph2ps_256_rrk(<8 x i16> %a0,<8 x float> %a1, i8 %mask) {
-; X86-LABEL: test_x86_vcvtph2ps_256_rrk:
-; X86:       # %bb.0:
-; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
-; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
-; X86-NEXT:    vcvtph2ps %xmm0, %ymm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x29,0x13,0xc8]
-; X86-NEXT:    vmovaps %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfc,0x28,0xc1]
-; X86-NEXT:    retl # encoding: [0xc3]
-;
-; X64-LABEL: test_x86_vcvtph2ps_256_rrk:
-; X64:       # %bb.0:
-; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
-; X64-NEXT:    vcvtph2ps %xmm0, %ymm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x29,0x13,0xc8]
-; X64-NEXT:    vmovaps %ymm1, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfc,0x28,0xc1]
-; X64-NEXT:    retq # encoding: [0xc3]
-  %res = call <8 x float> @llvm.x86.avx512.mask.vcvtph2ps.256(<8 x i16> %a0, <8 x float> %a1, i8 %mask)
-  ret <8 x float> %res
-}
-
-define <8 x float> @test_x86_vcvtph2ps_256_rrkz(<8 x i16> %a0, i8 %mask) {
-; X86-LABEL: test_x86_vcvtph2ps_256_rrkz:
-; X86:       # %bb.0:
-; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
-; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
-; X86-NEXT:    vcvtph2ps %xmm0, %ymm0 {%k1} {z} # encoding: [0x62,0xf2,0x7d,0xa9,0x13,0xc0]
-; X86-NEXT:    retl # encoding: [0xc3]
-;
-; X64-LABEL: test_x86_vcvtph2ps_256_rrkz:
-; X64:       # %bb.0:
-; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
-; X64-NEXT:    vcvtph2ps %xmm0, %ymm0 {%k1} {z} # encoding: [0x62,0xf2,0x7d,0xa9,0x13,0xc0]
-; X64-NEXT:    retq # encoding: [0xc3]
-  %res = call <8 x float> @llvm.x86.avx512.mask.vcvtph2ps.256(<8 x i16> %a0, <8 x float> zeroinitializer, i8 %mask)
-  ret <8 x float> %res
-}
-
-declare <8 x float> @llvm.x86.avx512.mask.vcvtph2ps.256(<8 x i16>, <8 x float>, i8) nounwind readonly
-
 define <8 x i16> @test_x86_vcvtps2ph_128(<4 x float> %a0, i8 %mask, <8 x i16> %src) {
 ; X86-LABEL: test_x86_vcvtps2ph_128:
 ; X86:       # %bb.0:
 ; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax # encoding: [0x0f,0xb6,0x44,0x24,0x04]
 ; X86-NEXT:    kmovw %eax, %k1 # encoding: [0xc5,0xf8,0x92,0xc8]
 ; X86-NEXT:    vcvtps2ph $2, %xmm0, %xmm2 # EVEX TO VEX Compression encoding: [0xc4,0xe3,0x79,0x1d,0xc2,0x02]
 ; X86-NEXT:    vcvtps2ph $2, %xmm0, %xmm3 {%k1} {z} # encoding: [0x62,0xf3,0x7d,0x89,0x1d,0xc3,0x02]
 ; X86-NEXT:    vcvtps2ph $2, %xmm0, %xmm1 {%k1} # encoding: [0x62,0xf3,0x7d,0x09,0x1d,0xc1,0x02]

llvm/test/CodeGen/X86/f16c-intrinsics-fast-isel.ll

; X64-NEXT:    retq
   %ins0 = insertelement <8 x i16> undef, i16 %a0, i32 0
   %ins1 = insertelement <8 x i16> %ins0, i16 0, i32 1
   %ins2 = insertelement <8 x i16> %ins1, i16 0, i32 2
   %ins3 = insertelement <8 x i16> %ins2, i16 0, i32 3
   %ins4 = insertelement <8 x i16> %ins3, i16 0, i32 4
   %ins5 = insertelement <8 x i16> %ins4, i16 0, i32 5
   %ins6 = insertelement <8 x i16> %ins5, i16 0, i32 6
   %ins7 = insertelement <8 x i16> %ins6, i16 0, i32 7
-  %cvt = call <4 x float> @llvm.x86.vcvtph2ps.128(<8 x i16> %ins7)
+  %shuffle = shufflevector <8 x i16> %ins7, <8 x i16> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  %bc = bitcast <4 x i16> %shuffle to <4 x half>
+  %cvt = fpext <4 x half> %bc to <4 x float>
   %res = extractelement <4 x float> %cvt, i32 0
   ret float %res
 }
 
 define i16 @test_cvtss_sh(float %a0) nounwind {
 ; X32-LABEL: test_cvtss_sh:
 ; X32:       # %bb.0:
 ; X32-NEXT:    vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero

 ; X32-NEXT:    vcvtph2ps %xmm0, %xmm0
 ; X32-NEXT:    retl
 ;
 ; X64-LABEL: test_mm_cvtph_ps:
 ; X64:       # %bb.0:
 ; X64-NEXT:    vcvtph2ps %xmm0, %xmm0
 ; X64-NEXT:    retq
   %arg0 = bitcast <2 x i64> %a0 to <8 x i16>
-  %res = call <4 x float> @llvm.x86.vcvtph2ps.128(<8 x i16> %arg0)
+  %shuffle = shufflevector <8 x i16> %arg0, <8 x i16> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  %bc = bitcast <4 x i16> %shuffle to <4 x half>
+  %res = fpext <4 x half> %bc to <4 x float>
   ret <4 x float> %res
 }
 
 define <8 x float> @test_mm256_cvtph_ps(<2 x i64> %a0) nounwind {
 ; X32-LABEL: test_mm256_cvtph_ps:
 ; X32:       # %bb.0:
 ; X32-NEXT:    vcvtph2ps %xmm0, %ymm0
 ; X32-NEXT:    retl
 ;
 ; X64-LABEL: test_mm256_cvtph_ps:
 ; X64:       # %bb.0:
 ; X64-NEXT:    vcvtph2ps %xmm0, %ymm0
 ; X64-NEXT:    retq
   %arg0 = bitcast <2 x i64> %a0 to <8 x i16>
-  %res = call <8 x float> @llvm.x86.vcvtph2ps.256(<8 x i16> %arg0)
+  %bc = bitcast <8 x i16> %arg0 to <8 x half>
+  %res = fpext <8 x half> %bc to <8 x float>
   ret <8 x float> %res
 }
 
 define <2 x i64> @test_mm_cvtps_ph(<4 x float> %a0) nounwind {
 ; X32-LABEL: test_mm_cvtps_ph:
 ; X32:       # %bb.0:
 ; X32-NEXT:    vcvtps2ph $0, %xmm0, %xmm0
 ; X32-NEXT:    retl

llvm/test/Transforms/InstCombine/X86/x86-f16c.ll

 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
 ; RUN: opt < %s -instcombine -S | FileCheck %s
 
 declare <4 x float> @llvm.x86.vcvtph2ps.128(<8 x i16>)
 declare <8 x float> @llvm.x86.vcvtph2ps.256(<8 x i16>)
 
 ;
-; Vector Demanded Bits
+; Vector Demanded Elts
 ;
 
 ; Only bottom 4 elements required.
 define <4 x float> @demand_vcvtph2ps_128(<8 x i16> %A) {
 ; CHECK-LABEL: @demand_vcvtph2ps_128(
-; CHECK-NEXT:    [[TMP1:%.*]] = tail call <4 x float> @llvm.x86.vcvtph2ps.128(<8 x i16> [[A:%.*]])
-; CHECK-NEXT:    ret <4 x float> [[TMP1]]
+; CHECK-NEXT:    [[TMP1:%.*]] = shufflevector <8 x i16> [[A:%.*]], <8 x i16> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+; CHECK-NEXT:    [[TMP2:%.*]] = bitcast <4 x i16> [[TMP1]] to <4 x half>
+; CHECK-NEXT:    [[CVTPH2PS:%.*]] = fpext <4 x half> [[TMP2]] to <4 x float>
+; CHECK-NEXT:    ret <4 x float> [[CVTPH2PS]]
 ;
   %1 = shufflevector <8 x i16> %A, <8 x i16> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3>
   %2 = tail call <4 x float> @llvm.x86.vcvtph2ps.128(<8 x i16> %1)
   ret <4 x float> %2
 }
 
 ; All 8 elements required.
 define <8 x float> @demand_vcvtph2ps_256(<8 x i16> %A) {
 ; CHECK-LABEL: @demand_vcvtph2ps_256(
 ; CHECK-NEXT:    [[TMP1:%.*]] = shufflevector <8 x i16> [[A:%.*]], <8 x i16> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3>
-; CHECK-NEXT:    [[TMP2:%.*]] = tail call <8 x float> @llvm.x86.vcvtph2ps.256(<8 x i16> [[TMP1]])
-; CHECK-NEXT:    ret <8 x float> [[TMP2]]
+; CHECK-NEXT:    [[TMP2:%.*]] = bitcast <8 x i16> [[TMP1]] to <8 x half>
+; CHECK-NEXT:    [[CVTPH2PS:%.*]] = fpext <8 x half> [[TMP2]] to <8 x float>
+; CHECK-NEXT:    ret <8 x float> [[CVTPH2PS]]
 ;
   %1 = shufflevector <8 x i16> %A, <8 x i16> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3>
   %2 = tail call <8 x float> @llvm.x86.vcvtph2ps.256(<8 x i16> %1)
   ret <8 x float> %2
 }
 
 ;
 ; Constant Folding