[llvm] r360596 - [X86][SSE] LowerBuildVectorv4x32 - don't insert MOVQ for undef elts

[Simon Pilgrim via llvm-commits]

Author: rksimon
Date: Mon May 13 09:10:11 2019
New Revision: 360596

URL: http://llvm.org/viewvc/llvm-project?rev=360596&view=rev
Log:
[X86][SSE] LowerBuildVectorv4x32 - don't insert MOVQ for undef elts

Fixes the regression noted in D61782 where a VZEXT_MOVL was being inserted because we weren't discriminating between 'zeroable' and 'all undef' for the upper elts.

Differential Revision: https://reviews.llvm.org/D61782

Modified:
    llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
    llvm/trunk/test/CodeGen/X86/phaddsub-undef.ll

Modified: llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/X86/X86ISelLowering.cpp?rev=360596&r1=360595&r2=360596&view=diff
==============================================================================
--- llvm/trunk/lib/Target/X86/X86ISelLowering.cpp (original)
+++ llvm/trunk/lib/Target/X86/X86ISelLowering.cpp Mon May 13 09:10:11 2019
@@ -7183,9 +7183,10 @@ static SDValue LowerBuildVectorv4x32(SDV
   }
 
   // Find all zeroable elements.
-  std::bitset<4> Zeroable;
-  for (int i=0; i < 4; ++i) {
-    SDValue Elt = Op->getOperand(i);
+  std::bitset<4> Zeroable, Undefs;
+  for (int i = 0; i < 4; ++i) {
+    SDValue Elt = Op.getOperand(i);
+    Undefs[i] = Elt.isUndef();
     Zeroable[i] = (Elt.isUndef() || X86::isZeroNode(Elt));
   }
   assert(Zeroable.size() - Zeroable.count() > 1 &&
@@ -7195,10 +7196,10 @@ static SDValue LowerBuildVectorv4x32(SDV
   // zeroable or extract_vector_elt with constant index.
   SDValue FirstNonZero;
   unsigned FirstNonZeroIdx;
-  for (unsigned i=0; i < 4; ++i) {
+  for (unsigned i = 0; i < 4; ++i) {
     if (Zeroable[i])
       continue;
-    SDValue Elt = Op->getOperand(i);
+    SDValue Elt = Op.getOperand(i);
     if (Elt.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
         !isa<ConstantSDNode>(Elt.getOperand(1)))
       return SDValue();
@@ -7237,10 +7238,12 @@ static SDValue LowerBuildVectorv4x32(SDV
 
   if (EltIdx == 4) {
     // Let the shuffle legalizer deal with blend operations.
-    SDValue VZero = getZeroVector(VT, Subtarget, DAG, SDLoc(Op));
+    SDValue VZeroOrUndef = (Zeroable == Undefs)
+                               ? DAG.getUNDEF(VT)
+                               : getZeroVector(VT, Subtarget, DAG, SDLoc(Op));
     if (V1.getSimpleValueType() != VT)
       V1 = DAG.getBitcast(VT, V1);
-    return DAG.getVectorShuffle(VT, SDLoc(V1), V1, VZero, Mask);
+    return DAG.getVectorShuffle(VT, SDLoc(V1), V1, VZeroOrUndef, Mask);
   }
 
   // See if we can lower this build_vector to a INSERTPS.

Modified: llvm/trunk/test/CodeGen/X86/phaddsub-undef.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/phaddsub-undef.ll?rev=360596&r1=360595&r2=360596&view=diff
==============================================================================
--- llvm/trunk/test/CodeGen/X86/phaddsub-undef.ll (original)
+++ llvm/trunk/test/CodeGen/X86/phaddsub-undef.ll Mon May 13 09:10:11 2019
@@ -160,26 +160,10 @@ define <16 x i32> @test16_v16i32_undef(<
 ; SSE-NEXT:    phaddd %xmm0, %xmm0
 ; SSE-NEXT:    retq
 ;
-; AVX-SLOW-LABEL: test16_v16i32_undef:
-; AVX-SLOW:       # %bb.0:
-; AVX-SLOW-NEXT:    vphaddd %xmm0, %xmm0, %xmm0
-; AVX-SLOW-NEXT:    retq
-;
-; AVX1-FAST-LABEL: test16_v16i32_undef:
-; AVX1-FAST:       # %bb.0:
-; AVX1-FAST-NEXT:    vphaddd %xmm0, %xmm0, %xmm0
-; AVX1-FAST-NEXT:    retq
-;
-; AVX2-FAST-LABEL: test16_v16i32_undef:
-; AVX2-FAST:       # %bb.0:
-; AVX2-FAST-NEXT:    vphaddd %xmm0, %xmm0, %xmm0
-; AVX2-FAST-NEXT:    retq
-;
-; AVX512-FAST-LABEL: test16_v16i32_undef:
-; AVX512-FAST:       # %bb.0:
-; AVX512-FAST-NEXT:    vphaddd %xmm0, %xmm0, %xmm0
-; AVX512-FAST-NEXT:    vmovq {{.*#+}} xmm0 = xmm0[0],zero
-; AVX512-FAST-NEXT:    retq
+; AVX-LABEL: test16_v16i32_undef:
+; AVX:       # %bb.0:
+; AVX-NEXT:    vphaddd %xmm0, %xmm0, %xmm0
+; AVX-NEXT:    retq
   %vecext = extractelement <16 x i32> %a, i32 0
   %vecext1 = extractelement <16 x i32> %a, i32 1
   %add = add i32 %vecext, %vecext1