[DAGCombiner] SCALAR_TO_VECTOR(EXTRACT_VECTOR_ELT(V,C)) -> VECTOR_SHUFFLE

[RKSimon]

Summary

This patch attempts to convert a SCALAR_TO_VECTOR(EXTRACT_VECTOR_ELT) pair into a VECTOR_SHUFFLE.

This prevents many cases of (slow) spilling scalar data between the gpr + simd registers.

At present the optimization only accepts cases where there is no TRUNC of the scalar type (i.e. all types must match).

[RKSimon]

Thanks Chandler

llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

private:
     SDValue visitLOAD(SDNode *N);
     SDValue visitSTORE(SDNode *N);
     SDValue visitINSERT_VECTOR_ELT(SDNode *N);
     SDValue visitEXTRACT_VECTOR_ELT(SDNode *N);
     SDValue visitBUILD_VECTOR(SDNode *N);
     SDValue visitCONCAT_VECTORS(SDNode *N);
     SDValue visitEXTRACT_SUBVECTOR(SDNode *N);
     SDValue visitVECTOR_SHUFFLE(SDNode *N);
+    SDValue visitSCALAR_TO_VECTOR(SDNode *N);
     SDValue visitINSERT_SUBVECTOR(SDNode *N);
     SDValue visitMLOAD(SDNode *N);
     SDValue visitMSTORE(SDNode *N);
 
     SDValue XformToShuffleWithZero(SDNode *N);
     SDValue ReassociateOps(unsigned Opc, SDLoc DL, SDValue LHS, SDValue RHS);
 
     SDValue visitShiftByConstant(SDNode *N, ConstantSDNode *Amt);
SDValue DAGCombiner::visit(SDNode *N) {
   case ISD::LOAD:               return visitLOAD(N);
   case ISD::STORE:              return visitSTORE(N);
   case ISD::INSERT_VECTOR_ELT:  return visitINSERT_VECTOR_ELT(N);
   case ISD::EXTRACT_VECTOR_ELT: return visitEXTRACT_VECTOR_ELT(N);
   case ISD::BUILD_VECTOR:       return visitBUILD_VECTOR(N);
   case ISD::CONCAT_VECTORS:     return visitCONCAT_VECTORS(N);
   case ISD::EXTRACT_SUBVECTOR:  return visitEXTRACT_SUBVECTOR(N);
   case ISD::VECTOR_SHUFFLE:     return visitVECTOR_SHUFFLE(N);
+  case ISD::SCALAR_TO_VECTOR:   return visitSCALAR_TO_VECTOR(N);
   case ISD::INSERT_SUBVECTOR:   return visitINSERT_SUBVECTOR(N);
   case ISD::MLOAD:              return visitMLOAD(N);
   case ISD::MSTORE:             return visitMSTORE(N);
   }
   return SDValue();
 }
 
 SDValue DAGCombiner::combine(SDNode *N) {
if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && N->isOnlyUserOf(N0.getNode()) &&
     //   shuffle(shuffle(A, B, M0), C, M1) -> shuffle(A, C, M2)
     //   shuffle(shuffle(A, B, M0), C, M1) -> shuffle(B, C, M2)
     return DAG.getVectorShuffle(VT, SDLoc(N), SV0, SV1, &Mask[0]);
   }
 
   return SDValue();
 }
 
+SDValue DAGCombiner::visitSCALAR_TO_VECTOR(SDNode *N) {
+  SDValue InVal = N->getOperand(0);
+  EVT VT = N->getValueType(0);
+
+  // Replace a SCALAR_TO_VECTOR(EXTRACT_VECTOR_ELT(V,C0)) pattern
+  // with a VECTOR_SHUFFLE.
+  if (InVal.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
+    SDValue InVec = InVal->getOperand(0);
+    SDValue EltNo = InVal->getOperand(1);
+
+    // FIXME: We could support implicit truncation if the shuffle can be
+    // scaled to a smaller vector scalar type.
+    ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(EltNo);
+    if (C0 && VT == InVec.getValueType() &&
+        VT.getScalarType() == InVal.getValueType()) {
+      SmallVector<int, 8> NewMask(VT.getVectorNumElements(), -1);
+      int Elt = C0->getZExtValue();
+      NewMask[0] = Elt;
+
+      if (TLI.isShuffleMaskLegal(NewMask, VT))
+        return DAG.getVectorShuffle(VT, SDLoc(N), InVec, DAG.getUNDEF(VT),
+                                    NewMask);
+    }
+  }
+
+  return SDValue();
+}
+
 SDValue DAGCombiner::visitINSERT_SUBVECTOR(SDNode *N) {
   SDValue N0 = N->getOperand(0);
   SDValue N2 = N->getOperand(2);
 
   // If the input vector is a concatenation, and the insert replaces
   // one of the halves, we can optimize into a single concat_vectors.
   if (N0.getOpcode() == ISD::CONCAT_VECTORS &&
       N0->getNumOperands() == 2 && N2.getOpcode() == ISD::Constant) {

llvm/trunk/test/CodeGen/X86/mmx-arg-passing-x86-64.ll

 
 define void @t4(x86_mmx %v1, x86_mmx %v2) nounwind  {
 ; X86-64-LABEL: t4:
 ; X86-64:       ## BB#0:
 ; X86-64-NEXT:    movdq2q %xmm1, %mm0
 ; X86-64-NEXT:    movq %mm0, -{{[0-9]+}}(%rsp)
 ; X86-64-NEXT:    movdq2q %xmm0, %mm0
 ; X86-64-NEXT:    movq %mm0, -{{[0-9]+}}(%rsp)
-; X86-64-NEXT:    movq {{.*#+}} xmm0 = mem[0],zero
 ; X86-64-NEXT:    movq {{.*#+}} xmm1 = mem[0],zero
-; X86-64-NEXT:    paddb %xmm0, %xmm1
-; X86-64-NEXT:    movd %xmm1, %rax
-; X86-64-NEXT:    movd %rax, %xmm0
+; X86-64-NEXT:    movq {{.*#+}} xmm0 = mem[0],zero
+; X86-64-NEXT:    paddb %xmm1, %xmm0
 ; X86-64-NEXT:    movb $1, %al
 ; X86-64-NEXT:    jmp _pass_v8qi ## TAILCALL
   %v1a = bitcast x86_mmx %v1 to <8 x i8>
   %v2b = bitcast x86_mmx %v2 to <8 x i8>
   %tmp3 = add <8 x i8> %v1a, %v2b
   %tmp3a = bitcast <8 x i8> %tmp3 to x86_mmx
   %tmp4 = tail call i32 (...)* @pass_v8qi( x86_mmx %tmp3a ) nounwind
   ret void

llvm/trunk/test/CodeGen/X86/pr14161.ll

entry:
   ret <2 x i16> %8
 }
 
 define <2 x i16> @bad(<4 x i32>*, <4 x i8>*) {
 ; CHECK-LABEL: bad:
 ; CHECK:       # BB#0: # %entry
 ; CHECK-NEXT:    movdqa (%rdi), %xmm0
 ; CHECK-NEXT:    pminud {{.*}}(%rip), %xmm0
-; CHECK-NEXT:    pextrd $1, %xmm0, %eax
-; CHECK-NEXT:    movd %eax, %xmm0
-; CHECK-NEXT:    pmovzxwq %xmm0, %xmm0
+; CHECK-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,1,2,3]
+; CHECK-NEXT:    pmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
 ; CHECK-NEXT:    retq
 entry:
   %2 = load <4 x i32>, <4 x i32>* %0, align 16
   %3 = call <4 x i32> @llvm.x86.sse41.pminud(<4 x i32> %2, <4 x i32> <i32 127, i32 127, i32 127, i32 127>)
   %4 = extractelement <4 x i32> %3, i32 0
   %5 = extractelement <4 x i32> %3, i32 1
   %6 = extractelement <4 x i32> %3, i32 2
   %7 = extractelement <4 x i32> %3, i32 3
   %8 = bitcast i32 %4 to <2 x i16>
   %9 = bitcast i32 %5 to <2 x i16>
   ret <2 x i16> %9
 }