[X86] Combine uitofp <v x i32> to <v x half>

llvm/lib/CodeGen/SelectionDAG/LegalizeVectorOps.cpp

@@ -1777,6 +1777,31 @@ void VectorLegalizer::ExpandUINT_TO_FLOAT(SDNode *Node,
   assert((BW == 64 || BW == 32) &&
          "Elements in vector-UINT_TO_FP must be 32 or 64 bits wide");
 
+  // If STRICT_/FMUL is not supported by the target (in case of f16) replace the
+  // UINT_TO_FP with a larger float and round to the smaller type
+  if ((!IsStrict && !TLI.isOperationLegalOrCustom(ISD::FMUL, DstVT)) ||
+      (IsStrict && !TLI.isOperationLegalOrCustom(ISD::STRICT_FMUL, DstVT))) {
+    EVT FPVT = BW == 32 ? MVT::f32 : MVT::f64;
+    SDValue UIToFP;
+    SDValue Result;
+    SDValue TargetZero = DAG.getIntPtrConstant(0, DL, /*isTarget=*/true);
+    EVT FloatVecVT = SrcVT.changeVectorElementType(FPVT);
+    if (IsStrict) {
+      UIToFP = DAG.getNode(ISD::STRICT_UINT_TO_FP, DL, {FloatVecVT, MVT::Other},
+                           {Node->getOperand(0), Src});
+      Result = DAG.getNode(ISD::STRICT_FP_ROUND, DL, {DstVT, MVT::Other},
+                           {Node->getOperand(0), UIToFP, TargetZero});
+      Results.push_back(Result);
+      Results.push_back(Result.getValue(1));
+    } else {
+      UIToFP = DAG.getNode(ISD::UINT_TO_FP, DL, FloatVecVT, Src);
+      Result = DAG.getNode(ISD::FP_ROUND, DL, DstVT, UIToFP, TargetZero);
+      Results.push_back(Result);
+    }
+
+    return;
+  }
+
   SDValue HalfWord = DAG.getConstant(BW / 2, DL, SrcVT);
 
   // Constants to clear the upper part of the word.

llvm/test/CodeGen/X86/uint_to_half.ll

@@ -0,0 +1,198 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx,+f16c | FileCheck %s -check-prefixes=AVX1
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx2,+f16c | FileCheck %s -check-prefixes=AVX2
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512vl | FileCheck %s -check-prefixes=AVX512
+
+define <8 x half> @test_uitofp_v8i32_v8f16(<8 x i32> %a) {
+; AVX1-LABEL: test_uitofp_v8i32_v8f16:
+; AVX1:       # %bb.0:
+; AVX1-NEXT:    vpsrld $16, %xmm0, %xmm1
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm2
+; AVX1-NEXT:    vpsrld $16, %xmm2, %xmm2
+; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm1, %ymm1
+; AVX1-NEXT:    vcvtdq2ps %ymm1, %ymm1
+; AVX1-NEXT:    vmulps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm1, %ymm1
+; AVX1-NEXT:    vandps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0
+; AVX1-NEXT:    vcvtdq2ps %ymm0, %ymm0
+; AVX1-NEXT:    vaddps %ymm0, %ymm1, %ymm0
+; AVX1-NEXT:    vcvtps2ph $4, %ymm0, %xmm0
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: test_uitofp_v8i32_v8f16:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [1258291200,1258291200,1258291200,1258291200,1258291200,1258291200,1258291200,1258291200]
+; AVX2-NEXT:    vpblendw {{.*#+}} ymm1 = ymm0[0],ymm1[1],ymm0[2],ymm1[3],ymm0[4],ymm1[5],ymm0[6],ymm1[7],ymm0[8],ymm1[9],ymm0[10],ymm1[11],ymm0[12],ymm1[13],ymm0[14],ymm1[15]
+; AVX2-NEXT:    vpsrld $16, %ymm0, %ymm0
+; AVX2-NEXT:    vpbroadcastd {{.*#+}} ymm2 = [1392508928,1392508928,1392508928,1392508928,1392508928,1392508928,1392508928,1392508928]
+; AVX2-NEXT:    vpblendw {{.*#+}} ymm0 = ymm0[0],ymm2[1],ymm0[2],ymm2[3],ymm0[4],ymm2[5],ymm0[6],ymm2[7],ymm0[8],ymm2[9],ymm0[10],ymm2[11],ymm0[12],ymm2[13],ymm0[14],ymm2[15]
+; AVX2-NEXT:    vbroadcastss {{.*#+}} ymm2 = [5.49764202E+11,5.49764202E+11,5.49764202E+11,5.49764202E+11,5.49764202E+11,5.49764202E+11,5.49764202E+11,5.49764202E+11]
+; AVX2-NEXT:    vsubps %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vaddps %ymm0, %ymm1, %ymm0
+; AVX2-NEXT:    vcvtps2ph $4, %ymm0, %xmm0
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    retq
+;
+; AVX512-LABEL: test_uitofp_v8i32_v8f16:
+; AVX512:       # %bb.0:
+; AVX512-NEXT:    vcvtudq2ps %ymm0, %ymm0
+; AVX512-NEXT:    vcvtps2ph $4, %ymm0, %xmm0
+; AVX512-NEXT:    vzeroupper
+; AVX512-NEXT:    retq
+  %vec = uitofp <8 x i32> %a to <8 x half>
+  ret <8 x half> %vec
+}
+
+define <8 x half> @test_strict_uitofp_v8i32_v8f16(<8 x i32> %a) {
+; AVX1-LABEL: test_strict_uitofp_v8i32_v8f16:
+; AVX1:       # %bb.0:
+; AVX1-NEXT:    vpsrld $16, %xmm0, %xmm1
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm2
+; AVX1-NEXT:    vpsrld $16, %xmm2, %xmm2
+; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm1, %ymm1
+; AVX1-NEXT:    vcvtdq2ps %ymm1, %ymm1
+; AVX1-NEXT:    vmulps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm1, %ymm1
+; AVX1-NEXT:    vandps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0
+; AVX1-NEXT:    vcvtdq2ps %ymm0, %ymm0
+; AVX1-NEXT:    vaddps %ymm0, %ymm1, %ymm0
+; AVX1-NEXT:    vcvtps2ph $4, %ymm0, %xmm0
+; AVX1-NEXT:    vzeroupper
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: test_strict_uitofp_v8i32_v8f16:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [1258291200,1258291200,1258291200,1258291200,1258291200,1258291200,1258291200,1258291200]
+; AVX2-NEXT:    vpblendw {{.*#+}} ymm1 = ymm0[0],ymm1[1],ymm0[2],ymm1[3],ymm0[4],ymm1[5],ymm0[6],ymm1[7],ymm0[8],ymm1[9],ymm0[10],ymm1[11],ymm0[12],ymm1[13],ymm0[14],ymm1[15]
+; AVX2-NEXT:    vpsrld $16, %ymm0, %ymm0
+; AVX2-NEXT:    vpbroadcastd {{.*#+}} ymm2 = [1392508928,1392508928,1392508928,1392508928,1392508928,1392508928,1392508928,1392508928]
+; AVX2-NEXT:    vpblendw {{.*#+}} ymm0 = ymm0[0],ymm2[1],ymm0[2],ymm2[3],ymm0[4],ymm2[5],ymm0[6],ymm2[7],ymm0[8],ymm2[9],ymm0[10],ymm2[11],ymm0[12],ymm2[13],ymm0[14],ymm2[15]
+; AVX2-NEXT:    vbroadcastss {{.*#+}} ymm2 = [5.49764202E+11,5.49764202E+11,5.49764202E+11,5.49764202E+11,5.49764202E+11,5.49764202E+11,5.49764202E+11,5.49764202E+11]
+; AVX2-NEXT:    vsubps %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vaddps %ymm0, %ymm1, %ymm0
+; AVX2-NEXT:    vcvtps2ph $4, %ymm0, %xmm0
+; AVX2-NEXT:    vzeroupper
+; AVX2-NEXT:    retq
+;
+; AVX512-LABEL: test_strict_uitofp_v8i32_v8f16:
+; AVX512:       # %bb.0:
+; AVX512-NEXT:    vcvtudq2ps %ymm0, %ymm0
+; AVX512-NEXT:    vcvtps2ph $4, %ymm0, %xmm0
+; AVX512-NEXT:    vzeroupper
+; AVX512-NEXT:    retq
+  %vec = tail call <8 x half> @llvm.experimental.constrained.uitofp.f16.i32(<8 x i32> %a, metadata !"round.dynamic", metadata !"fpexcept.strict")
+  ret <8 x half> %vec
+}
+
+define <16 x half> @test_uitofp_v16i32_v16f16(<16 x i32> %a) {
+; AVX1-LABEL: test_uitofp_v16i32_v16f16:
+; AVX1:       # %bb.0:
+; AVX1-NEXT:    vpsrld $16, %xmm0, %xmm2
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm3
+; AVX1-NEXT:    vpsrld $16, %xmm3, %xmm3
+; AVX1-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm2
+; AVX1-NEXT:    vcvtdq2ps %ymm2, %ymm2
+; AVX1-NEXT:    vbroadcastss {{.*#+}} ymm3 = [6.5536E+4,6.5536E+4,6.5536E+4,6.5536E+4,6.5536E+4,6.5536E+4,6.5536E+4,6.5536E+4]
+; AVX1-NEXT:    vmulps %ymm3, %ymm2, %ymm2
+; AVX1-NEXT:    vbroadcastss {{.*#+}} ymm4 = [65535,65535,65535,65535,65535,65535,65535,65535]
+; AVX1-NEXT:    vandps %ymm4, %ymm0, %ymm0
+; AVX1-NEXT:    vcvtdq2ps %ymm0, %ymm0
+; AVX1-NEXT:    vaddps %ymm0, %ymm2, %ymm0
+; AVX1-NEXT:    vcvtps2ph $4, %ymm0, %xmm0
+; AVX1-NEXT:    vpsrld $16, %xmm1, %xmm2
+; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm5
+; AVX1-NEXT:    vpsrld $16, %xmm5, %xmm5
+; AVX1-NEXT:    vinsertf128 $1, %xmm5, %ymm2, %ymm2
+; AVX1-NEXT:    vcvtdq2ps %ymm2, %ymm2
+; AVX1-NEXT:    vmulps %ymm3, %ymm2, %ymm2
+; AVX1-NEXT:    vandps %ymm4, %ymm1, %ymm1
+; AVX1-NEXT:    vcvtdq2ps %ymm1, %ymm1
+; AVX1-NEXT:    vaddps %ymm1, %ymm2, %ymm1
+; AVX1-NEXT:    vcvtps2ph $4, %ymm1, %xmm1
+; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: test_uitofp_v16i32_v16f16:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    vpbroadcastd {{.*#+}} ymm2 = [1258291200,1258291200,1258291200,1258291200,1258291200,1258291200,1258291200,1258291200]
+; AVX2-NEXT:    vpblendw {{.*#+}} ymm3 = ymm0[0],ymm2[1],ymm0[2],ymm2[3],ymm0[4],ymm2[5],ymm0[6],ymm2[7],ymm0[8],ymm2[9],ymm0[10],ymm2[11],ymm0[12],ymm2[13],ymm0[14],ymm2[15]
+; AVX2-NEXT:    vpsrld $16, %ymm0, %ymm0
+; AVX2-NEXT:    vpbroadcastd {{.*#+}} ymm4 = [1392508928,1392508928,1392508928,1392508928,1392508928,1392508928,1392508928,1392508928]
+; AVX2-NEXT:    vpblendw {{.*#+}} ymm0 = ymm0[0],ymm4[1],ymm0[2],ymm4[3],ymm0[4],ymm4[5],ymm0[6],ymm4[7],ymm0[8],ymm4[9],ymm0[10],ymm4[11],ymm0[12],ymm4[13],ymm0[14],ymm4[15]
+; AVX2-NEXT:    vbroadcastss {{.*#+}} ymm5 = [5.49764202E+11,5.49764202E+11,5.49764202E+11,5.49764202E+11,5.49764202E+11,5.49764202E+11,5.49764202E+11,5.49764202E+11]
+; AVX2-NEXT:    vsubps %ymm5, %ymm0, %ymm0
+; AVX2-NEXT:    vaddps %ymm0, %ymm3, %ymm0
+; AVX2-NEXT:    vcvtps2ph $4, %ymm0, %xmm0
+; AVX2-NEXT:    vpblendw {{.*#+}} ymm2 = ymm1[0],ymm2[1],ymm1[2],ymm2[3],ymm1[4],ymm2[5],ymm1[6],ymm2[7],ymm1[8],ymm2[9],ymm1[10],ymm2[11],ymm1[12],ymm2[13],ymm1[14],ymm2[15]
+; AVX2-NEXT:    vpsrld $16, %ymm1, %ymm1
+; AVX2-NEXT:    vpblendw {{.*#+}} ymm1 = ymm1[0],ymm4[1],ymm1[2],ymm4[3],ymm1[4],ymm4[5],ymm1[6],ymm4[7],ymm1[8],ymm4[9],ymm1[10],ymm4[11],ymm1[12],ymm4[13],ymm1[14],ymm4[15]
+; AVX2-NEXT:    vsubps %ymm5, %ymm1, %ymm1
+; AVX2-NEXT:    vaddps %ymm1, %ymm2, %ymm1
+; AVX2-NEXT:    vcvtps2ph $4, %ymm1, %xmm1
+; AVX2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX2-NEXT:    retq
+;
+; AVX512-LABEL: test_uitofp_v16i32_v16f16:
+; AVX512:       # %bb.0:
+; AVX512-NEXT:    vcvtudq2ps %zmm0, %zmm0
+; AVX512-NEXT:    vcvtps2ph $4, %zmm0, %ymm0
+; AVX512-NEXT:    retq
+  %vec = uitofp <16 x i32> %a to <16 x half>
+  ret <16 x half> %vec
+}
+
+define <16 x half> @test_strict_uitofp_v16i32_v16f16(<16 x i32> %a) {
+; AVX1-LABEL: test_strict_uitofp_v16i32_v16f16:
+; AVX1:       # %bb.0:
+; AVX1-NEXT:    vpsrld $16, %xmm0, %xmm2
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm3
+; AVX1-NEXT:    vpsrld $16, %xmm3, %xmm3
+; AVX1-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm2
+; AVX1-NEXT:    vcvtdq2ps %ymm2, %ymm2
+; AVX1-NEXT:    vbroadcastss {{.*#+}} ymm3 = [6.5536E+4,6.5536E+4,6.5536E+4,6.5536E+4,6.5536E+4,6.5536E+4,6.5536E+4,6.5536E+4]
+; AVX1-NEXT:    vmulps %ymm3, %ymm2, %ymm2
+; AVX1-NEXT:    vbroadcastss {{.*#+}} ymm4 = [65535,65535,65535,65535,65535,65535,65535,65535]
+; AVX1-NEXT:    vandps %ymm4, %ymm0, %ymm0
+; AVX1-NEXT:    vcvtdq2ps %ymm0, %ymm0
+; AVX1-NEXT:    vaddps %ymm0, %ymm2, %ymm0
+; AVX1-NEXT:    vcvtps2ph $4, %ymm0, %xmm0
+; AVX1-NEXT:    vpsrld $16, %xmm1, %xmm2
+; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm5
+; AVX1-NEXT:    vpsrld $16, %xmm5, %xmm5
+; AVX1-NEXT:    vinsertf128 $1, %xmm5, %ymm2, %ymm2
+; AVX1-NEXT:    vcvtdq2ps %ymm2, %ymm2
+; AVX1-NEXT:    vmulps %ymm3, %ymm2, %ymm2
+; AVX1-NEXT:    vandps %ymm4, %ymm1, %ymm1
+; AVX1-NEXT:    vcvtdq2ps %ymm1, %ymm1
+; AVX1-NEXT:    vaddps %ymm1, %ymm2, %ymm1
+; AVX1-NEXT:    vcvtps2ph $4, %ymm1, %xmm1
+; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: test_strict_uitofp_v16i32_v16f16:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    vpbroadcastd {{.*#+}} ymm2 = [1258291200,1258291200,1258291200,1258291200,1258291200,1258291200,1258291200,1258291200]
+; AVX2-NEXT:    vpblendw {{.*#+}} ymm3 = ymm0[0],ymm2[1],ymm0[2],ymm2[3],ymm0[4],ymm2[5],ymm0[6],ymm2[7],ymm0[8],ymm2[9],ymm0[10],ymm2[11],ymm0[12],ymm2[13],ymm0[14],ymm2[15]
+; AVX2-NEXT:    vpsrld $16, %ymm0, %ymm0
+; AVX2-NEXT:    vpbroadcastd {{.*#+}} ymm4 = [1392508928,1392508928,1392508928,1392508928,1392508928,1392508928,1392508928,1392508928]
+; AVX2-NEXT:    vpblendw {{.*#+}} ymm0 = ymm0[0],ymm4[1],ymm0[2],ymm4[3],ymm0[4],ymm4[5],ymm0[6],ymm4[7],ymm0[8],ymm4[9],ymm0[10],ymm4[11],ymm0[12],ymm4[13],ymm0[14],ymm4[15]
+; AVX2-NEXT:    vbroadcastss {{.*#+}} ymm5 = [5.49764202E+11,5.49764202E+11,5.49764202E+11,5.49764202E+11,5.49764202E+11,5.49764202E+11,5.49764202E+11,5.49764202E+11]
+; AVX2-NEXT:    vsubps %ymm5, %ymm0, %ymm0
+; AVX2-NEXT:    vaddps %ymm0, %ymm3, %ymm0
+; AVX2-NEXT:    vcvtps2ph $4, %ymm0, %xmm0
+; AVX2-NEXT:    vpblendw {{.*#+}} ymm2 = ymm1[0],ymm2[1],ymm1[2],ymm2[3],ymm1[4],ymm2[5],ymm1[6],ymm2[7],ymm1[8],ymm2[9],ymm1[10],ymm2[11],ymm1[12],ymm2[13],ymm1[14],ymm2[15]
+; AVX2-NEXT:    vpsrld $16, %ymm1, %ymm1
+; AVX2-NEXT:    vpblendw {{.*#+}} ymm1 = ymm1[0],ymm4[1],ymm1[2],ymm4[3],ymm1[4],ymm4[5],ymm1[6],ymm4[7],ymm1[8],ymm4[9],ymm1[10],ymm4[11],ymm1[12],ymm4[13],ymm1[14],ymm4[15]
+; AVX2-NEXT:    vsubps %ymm5, %ymm1, %ymm1
+; AVX2-NEXT:    vaddps %ymm1, %ymm2, %ymm1
+; AVX2-NEXT:    vcvtps2ph $4, %ymm1, %xmm1
+; AVX2-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; AVX2-NEXT:    retq
+;
+; AVX512-LABEL: test_strict_uitofp_v16i32_v16f16:
+; AVX512:       # %bb.0:
+; AVX512-NEXT:    vcvtudq2ps %zmm0, %zmm0
+; AVX512-NEXT:    vcvtps2ph $4, %zmm0, %ymm0
+; AVX512-NEXT:    retq
+  %vec = tail call <16 x half> @llvm.experimental.constrained.uitofp.f16.i32(<16 x i32> %a, metadata !"round.dynamic", metadata !"fpexcept.strict")
+  ret <16 x half> %vec
+}