From: Simon Pilgrim <llvm-dev@redking.me.uk>
Date: Fri, 3 Feb 2017 17:59:58 +0000 (+0000)
Subject: [X86][SSE] Add support for combining scalar_to_vector(extract_vector_elt) into a... 
X-Git-Url: https://granicus.if.org/sourcecode?a=commitdiff_plain;h=da9f5183608bc474e9365517604fa71643c3f7dd;p=llvm

[X86][SSE] Add support for combining scalar_to_vector(extract_vector_elt) into a target shuffle.

Correctly flagging upper elements as undef.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@294020 91177308-0d34-0410-b5e6-96231b3b80d8
---

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index a7a141050d9..f4b81063bcc 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -5729,6 +5729,20 @@ static bool getFauxShuffleMask(SDValue N, SmallVectorImpl<int> &Mask,
     Ops.push_back(IsAndN ? N1 : N0);
     return true;
   }
+  case ISD::SCALAR_TO_VECTOR: {
+    // Match against a scalar_to_vector of an extract from a similar vector.
+    SDValue N0 = N.getOperand(0);
+    if (N0.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
+        N0.getOperand(0).getValueType() != VT ||
+        !isa<ConstantSDNode>(N0.getOperand(1)) ||
+        NumElts <= N0.getConstantOperandVal(1) ||
+        !N->isOnlyUserOf(N0.getNode()))
+      return false;
+    Ops.push_back(N0.getOperand(0));
+    Mask.push_back(N0.getConstantOperandVal(1));
+    Mask.append(NumElts - 1, SM_SentinelUndef);
+    return true;
+  }
   case X86ISD::PINSRB:
   case X86ISD::PINSRW: {
     SDValue InVec = N.getOperand(0);
diff --git a/test/CodeGen/X86/combine-fcopysign.ll b/test/CodeGen/X86/combine-fcopysign.ll
index 807ac4e3fc6..43e09bfe5fe 100644
--- a/test/CodeGen/X86/combine-fcopysign.ll
+++ b/test/CodeGen/X86/combine-fcopysign.ll
@@ -292,7 +292,7 @@ define <4 x float> @combine_vec_fcopysign_fptrunc_sgn(<4 x float> %x, <4 x doubl
 ; SSE-NEXT:    cvtsd2ss %xmm1, %xmm1
 ; SSE-NEXT:    andps %xmm4, %xmm1
 ; SSE-NEXT:    orps %xmm6, %xmm1
-; SSE-NEXT:    insertps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[2,3]
+; SSE-NEXT:    unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
 ; SSE-NEXT:    movaps %xmm3, %xmm1
 ; SSE-NEXT:    movhlps {{.*#+}} xmm1 = xmm1[1,1]
 ; SSE-NEXT:    andps %xmm5, %xmm1
diff --git a/test/CodeGen/X86/pr29112.ll b/test/CodeGen/X86/pr29112.ll
index 8bf704835ae..de2d5d79102 100644
--- a/test/CodeGen/X86/pr29112.ll
+++ b/test/CodeGen/X86/pr29112.ll
@@ -24,11 +24,11 @@ define <4 x float> @bar(<4 x float>* %a1p, <4 x float>* %a2p, <4 x float> %a3, <
 ; CHECK-NEXT:    vextractf32x4 $2, %zmm3, %xmm4
 ; CHECK-NEXT:    vblendps {{.*#+}} xmm4 = xmm0[0,1,2],xmm4[3]
 ; CHECK-NEXT:    vpermilps {{.*#+}} xmm5 = xmm2[3,1,2,3]
-; CHECK-NEXT:    vinsertps {{.*#+}} xmm5 = xmm5[0],xmm1[0],xmm5[2,3]
+; CHECK-NEXT:    vunpcklps {{.*#+}} xmm5 = xmm5[0],xmm1[0],xmm5[1],xmm1[1]
 ; CHECK-NEXT:    vinsertps {{.*#+}} xmm5 = xmm5[0,1],xmm2[1],xmm5[3]
 ; CHECK-NEXT:    vinsertps {{.*#+}} xmm5 = xmm5[0,1,2],xmm3[1]
 ; CHECK-NEXT:    vmovshdup {{.*#+}} xmm7 = xmm8[1,1,3,3]
-; CHECK-NEXT:    vinsertps {{.*#+}} xmm7 = xmm7[0],xmm1[0],xmm7[2,3]
+; CHECK-NEXT:    vunpcklps {{.*#+}} xmm7 = xmm7[0],xmm1[0],xmm7[1],xmm1[1]
 ; CHECK-NEXT:    vinsertps {{.*#+}} xmm10 = xmm7[0,1],xmm2[1],xmm7[3]
 ; CHECK-NEXT:    vblendps {{.*#+}} xmm7 = xmm10[0,1,2],xmm3[3]
 ; CHECK-NEXT:    vblendps {{.*#+}} xmm11 = xmm0[0,1,2],xmm3[3]
diff --git a/test/CodeGen/X86/vector-shuffle-combining-xop.ll b/test/CodeGen/X86/vector-shuffle-combining-xop.ll
index eef7d5a1366..241c63c6acd 100644
--- a/test/CodeGen/X86/vector-shuffle-combining-xop.ll
+++ b/test/CodeGen/X86/vector-shuffle-combining-xop.ll
@@ -342,23 +342,18 @@ define void @buildvector_v4f32_0404(float %a, float %b, <4 x float>* %ptr) {
   ret void
 }
 
-; FIXME: Failed to fold to vpermil2ps
 define void @buildvector_v4f32_07z6(float %a, <4 x float> %b, <4 x float>* %ptr) {
 ; X32-LABEL: buildvector_v4f32_07z6:
 ; X32:       # BB#0:
 ; X32-NEXT:    movl {{[0-9]+}}(%esp), %eax
-; X32-NEXT:    vpermilps {{.*#+}} xmm1 = xmm0[3,1,2,3]
-; X32-NEXT:    vmovss {{.*#+}} xmm2 = mem[0],zero,zero,zero
-; X32-NEXT:    vunpcklps {{.*#+}} xmm1 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]
-; X32-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],zero,xmm0[2]
+; X32-NEXT:    vmovss {{.*#+}} xmm1 = mem[0],zero,zero,zero
+; X32-NEXT:    vpermil2ps {{.*#+}} xmm0 = xmm1[0],xmm0[3],zero,xmm0[2]
 ; X32-NEXT:    vmovaps %xmm0, (%eax)
 ; X32-NEXT:    retl
 ;
 ; X64-LABEL: buildvector_v4f32_07z6:
 ; X64:       # BB#0:
-; X64-NEXT:    vpermilps {{.*#+}} xmm2 = xmm1[3,1,2,3]
-; X64-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
-; X64-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0,1],zero,xmm1[2]
+; X64-NEXT:    vpermil2ps {{.*#+}} xmm0 = xmm0[0],xmm1[3],zero,xmm1[2]
 ; X64-NEXT:    vmovaps %xmm0, (%rdi)
 ; X64-NEXT:    retq
   %b2 = extractelement <4 x float> %b, i32 2