[llvm] a84790e - [X86] SimplifyDemandedVectorEltsForTargetNode - reduce width of X86 conversions nodes when upper elements are not demanded. (#102882)

[via llvm-commits]

Author: Simon Pilgrim
Date: 2024-08-16T12:23:39+01:00
New Revision: a84790e5fca0429683b24f7bb52d2c4d947dc011

URL: https://github.com/llvm/llvm-project/commit/a84790e5fca0429683b24f7bb52d2c4d947dc011
DIFF: https://github.com/llvm/llvm-project/commit/a84790e5fca0429683b24f7bb52d2c4d947dc011.diff

LOG: [X86] SimplifyDemandedVectorEltsForTargetNode - reduce width of X86 conversions nodes when upper elements are not demanded. (#102882)

Fixes #83402

Added: 
    

Modified: 
    llvm/lib/Target/X86/X86ISelLowering.cpp
    llvm/test/CodeGen/X86/vector-half-conversions.ll

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index db50f132b1349..c6e1764edffc3 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -42524,6 +42524,26 @@ bool X86TargetLowering::SimplifyDemandedVectorEltsForTargetNode(
       SDValue Insert =
           insertSubVector(UndefVec, ExtOp, 0, TLO.DAG, DL, ExtSizeInBits);
       return TLO.CombineTo(Op, Insert);
+    }
+      // Conversions.
+      // TODO: Add more CVT opcodes when we have test coverage.
+    case X86ISD::CVTTP2SI:
+    case X86ISD::CVTTP2UI:
+    case X86ISD::CVTPH2PS: {
+      SDLoc DL(Op);
+      unsigned Scale = SizeInBits / ExtSizeInBits;
+      SDValue SrcOp = Op.getOperand(0);
+      MVT SrcVT = SrcOp.getSimpleValueType();
+      unsigned SrcExtSize =
+          std::max<unsigned>(SrcVT.getSizeInBits() / Scale, 128);
+      MVT ExtVT = MVT::getVectorVT(VT.getSimpleVT().getScalarType(),
+                                   ExtSizeInBits / VT.getScalarSizeInBits());
+      SDValue ExtOp = TLO.DAG.getNode(
+          Opc, DL, ExtVT, extractSubVector(SrcOp, 0, TLO.DAG, DL, SrcExtSize));
+      SDValue UndefVec = TLO.DAG.getUNDEF(VT);
+      SDValue Insert =
+          insertSubVector(UndefVec, ExtOp, 0, TLO.DAG, DL, ExtSizeInBits);
+      return TLO.CombineTo(Op, Insert);
     }
       // Zero upper elements.
     case X86ISD::VZEXT_MOVL:

diff  --git a/llvm/test/CodeGen/X86/vector-half-conversions.ll b/llvm/test/CodeGen/X86/vector-half-conversions.ll
index e87814ebb1dbe..ef0f3f3e816df 100644
--- a/llvm/test/CodeGen/X86/vector-half-conversions.ll
+++ b/llvm/test/CodeGen/X86/vector-half-conversions.ll
@@ -4990,6 +4990,7 @@ define <4 x i32> @fptosi_2f16_to_4i32(<2 x half> %a) nounwind {
   ret <4 x i32> %ext
 }
 
+; PR83402
 define <4 x i32> @fptosi_4f16_to_4i32(<4 x half> %a) nounwind {
 ; AVX-LABEL: fptosi_4f16_to_4i32:
 ; AVX:       # %bb.0:
@@ -5024,16 +5025,14 @@ define <4 x i32> @fptosi_4f16_to_4i32(<4 x half> %a) nounwind {
 ;
 ; F16C-LABEL: fptosi_4f16_to_4i32:
 ; F16C:       # %bb.0:
-; F16C-NEXT:    vcvtph2ps %xmm0, %ymm0
+; F16C-NEXT:    vcvtph2ps %xmm0, %xmm0
 ; F16C-NEXT:    vcvttps2dq %xmm0, %xmm0
-; F16C-NEXT:    vzeroupper
 ; F16C-NEXT:    retq
 ;
 ; AVX512-LABEL: fptosi_4f16_to_4i32:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    vcvtph2ps %xmm0, %ymm0
+; AVX512-NEXT:    vcvtph2ps %xmm0, %xmm0
 ; AVX512-NEXT:    vcvttps2dq %xmm0, %xmm0
-; AVX512-NEXT:    vzeroupper
 ; AVX512-NEXT:    retq
   %cvt = fptosi <4 x half> %a to <4 x i32>
   ret <4 x i32> %cvt
@@ -5213,13 +5212,12 @@ define <4 x i32> @fptoui_4f16_to_4i32(<4 x half> %a) nounwind {
 ;
 ; F16C-LABEL: fptoui_4f16_to_4i32:
 ; F16C:       # %bb.0:
-; F16C-NEXT:    vcvtph2ps %xmm0, %ymm0
-; F16C-NEXT:    vcvttps2dq %ymm0, %ymm1
-; F16C-NEXT:    vsubps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0
-; F16C-NEXT:    vcvttps2dq %ymm0, %ymm0
+; F16C-NEXT:    vcvtph2ps %xmm0, %xmm0
+; F16C-NEXT:    vcvttps2dq %xmm0, %xmm1
+; F16C-NEXT:    vsubps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
+; F16C-NEXT:    vcvttps2dq %xmm0, %xmm0
 ; F16C-NEXT:    vorps %xmm0, %xmm1, %xmm0
 ; F16C-NEXT:    vblendvps %xmm1, %xmm0, %xmm1, %xmm0
-; F16C-NEXT:    vzeroupper
 ; F16C-NEXT:    retq
 ;
 ; AVX512F-LABEL: fptoui_4f16_to_4i32:
@@ -5232,9 +5230,8 @@ define <4 x i32> @fptoui_4f16_to_4i32(<4 x half> %a) nounwind {
 ;
 ; AVX512-FASTLANE-LABEL: fptoui_4f16_to_4i32:
 ; AVX512-FASTLANE:       # %bb.0:
-; AVX512-FASTLANE-NEXT:    vcvtph2ps %xmm0, %ymm0
+; AVX512-FASTLANE-NEXT:    vcvtph2ps %xmm0, %xmm0
 ; AVX512-FASTLANE-NEXT:    vcvttps2udq %xmm0, %xmm0
-; AVX512-FASTLANE-NEXT:    vzeroupper
 ; AVX512-FASTLANE-NEXT:    retq
   %cvt = fptoui <4 x half> %a to <4 x i32>
   ret <4 x i32> %cvt