DAG: Avoid forming shufflevector from a single extract_vector_elt

This avoids regressions in a future AMDGPU commit. Previously we
would have a build_vector (extract_vector_elt x), undef with free
access to the elements bloated into a shuffle of one element + undef,
which has much worse combine support than the extract.

Alternatively could check aggressivelyPreferBuildVectorSources, but
I'm not sure it's really different than isExtractVecEltCheap.

Author: arsenm

llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

@@ -23807,6 +23807,10 @@ SDValue DAGCombiner::reduceBuildVecToShuffle(SDNode *N) {
   SmallVector<SDValue, 8> VecIn;
   VecIn.push_back(SDValue());
 
+  // If we have a single extract_element with a constant index, track the index
+  // value.
+  unsigned OneConstExtractIndex = ~0u;
+
   for (unsigned i = 0; i != NumElems; ++i) {
     SDValue Op = N->getOperand(i);
 
@@ -23824,23 +23828,27 @@ SDValue DAGCombiner::reduceBuildVecToShuffle(SDNode *N) {
 
     // Not an undef or zero. If the input is something other than an
     // EXTRACT_VECTOR_ELT with an in-range constant index, bail out.
-    if (Op.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
-        !isa<ConstantSDNode>(Op.getOperand(1)))
+    if (Op.getOpcode() != ISD::EXTRACT_VECTOR_ELT)
       return SDValue();
-    SDValue ExtractedFromVec = Op.getOperand(0);
 
+    SDValue ExtractedFromVec = Op.getOperand(0);
     if (ExtractedFromVec.getValueType().isScalableVector())
       return SDValue();
+    auto *ExtractIdx = dyn_cast<ConstantSDNode>(Op.getOperand(1));
+    if (!ExtractIdx)
+      return SDValue();
 
-    const APInt &ExtractIdx = Op.getConstantOperandAPInt(1);
-    if (ExtractIdx.uge(ExtractedFromVec.getValueType().getVectorNumElements()))
+    if (ExtractIdx->getAsAPIntVal().uge(
+            ExtractedFromVec.getValueType().getVectorNumElements()))
       return SDValue();
 
     // All inputs must have the same element type as the output.
     if (VT.getVectorElementType() !=
         ExtractedFromVec.getValueType().getVectorElementType())
       return SDValue();
 
+    OneConstExtractIndex = ExtractIdx->getZExtValue();
+
     // Have we seen this input vector before?
     // The vectors are expected to be tiny (usually 1 or 2 elements), so using
     // a map back from SDValues to numbers isn't worth it.
@@ -23863,6 +23871,13 @@ SDValue DAGCombiner::reduceBuildVecToShuffle(SDNode *N) {
   // VecIn accordingly.
   bool DidSplitVec = false;
   if (VecIn.size() == 2) {
+    // If we only found a single constant indexed extract_vector_elt feeding the
+    // build_vector, do not produce a more complicated shuffle if the extract is
+    // cheap.
+    if (TLI.isOperationLegalOrCustom(ISD::EXTRACT_VECTOR_ELT, VT) &&
+        TLI.isExtractVecEltCheap(VT, OneConstExtractIndex))
+      return SDValue();
+
     unsigned MaxIndex = 0;
     unsigned NearestPow2 = 0;
     SDValue Vec = VecIn.back();

llvm/test/CodeGen/AMDGPU/insert_vector_dynelt.ll

@@ -452,11 +452,11 @@ define amdgpu_kernel void @byte8_inselt(ptr addrspace(1) %out, <8 x i8> %vec, i3
 ; GCN-NEXT:    s_and_b32 s6, s4, 0x1010101
 ; GCN-NEXT:    s_andn2_b64 s[2:3], s[2:3], s[4:5]
 ; GCN-NEXT:    s_or_b64 s[2:3], s[6:7], s[2:3]
-; GCN-NEXT:    v_mov_b32_e32 v3, s1
-; GCN-NEXT:    v_mov_b32_e32 v0, s2
-; GCN-NEXT:    v_mov_b32_e32 v1, s3
-; GCN-NEXT:    v_mov_b32_e32 v2, s0
-; GCN-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
+; GCN-NEXT:    v_mov_b32_e32 v0, s0
+; GCN-NEXT:    v_mov_b32_e32 v2, s2
+; GCN-NEXT:    v_mov_b32_e32 v1, s1
+; GCN-NEXT:    v_mov_b32_e32 v3, s3
+; GCN-NEXT:    flat_store_dwordx2 v[0:1], v[2:3]
 ; GCN-NEXT:    s_endpgm
 entry:
   %v = insertelement <8 x i8> %vec, i8 1, i32 %sel

llvm/test/CodeGen/X86/avx512-build-vector.ll

@@ -14,11 +14,9 @@ define <16 x i32> @test2(<16 x i32> %x) {
 define <16 x float> @test3(<4 x float> %a) {
 ; CHECK-LABEL: test3:
 ; CHECK:       ## %bb.0:
-; CHECK-NEXT:    ## kill: def $xmm0 killed $xmm0 def $zmm0
-; CHECK-NEXT:    vpmovsxbd {{.*#+}} zmm2 = [0,1,2,3,4,18,16,7,8,9,10,11,12,13,14,15]
-; CHECK-NEXT:    vxorps %xmm1, %xmm1, %xmm1
-; CHECK-NEXT:    vpermt2ps %zmm0, %zmm2, %zmm1
-; CHECK-NEXT:    vmovaps %zmm1, %zmm0
+; CHECK-NEXT:    vpshufb {{.*#+}} xmm0 = zero,zero,zero,zero,xmm0[8,9,10,11,0,1,2,3],zero,zero,zero,zero
+; CHECK-NEXT:    vpxor %xmm1, %xmm1, %xmm1
+; CHECK-NEXT:    vinserti32x4 $1, %xmm0, %zmm1, %zmm0
 ; CHECK-NEXT:    retq
   %b = extractelement <4 x float> %a, i32 2
   %c = insertelement <16 x float> <float 0.000000e+00, float 0.000000e+00, float 0.000000e+00, float 0.000000e+00, float 0.000000e+00, float undef, float 0.000000e+00, float 0.000000e+00, float 0.000000e+00, float 0.000000e+00, float 0.000000e+00, float 0.000000e+00, float 0.000000e+00, float 0.000000e+00, float 0.000000e+00, float 0.000000e+00>, float %b, i32 5

llvm/test/CodeGen/X86/insertelement-duplicates.ll

@@ -31,18 +31,16 @@ define void @PR15298(ptr nocapture %source, ptr nocapture %dest) nounwind noinli
 ; AVX-32:       # %bb.0: # %L.entry
 ; AVX-32-NEXT:    movl {{[0-9]+}}(%esp), %eax
 ; AVX-32-NEXT:    movl {{[0-9]+}}(%esp), %ecx
-; AVX-32-NEXT:    vbroadcastss 304(%ecx), %xmm0
-; AVX-32-NEXT:    vxorps %xmm1, %xmm1, %xmm1
-; AVX-32-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2],ymm1[3,4,5,6,7]
+; AVX-32-NEXT:    vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; AVX-32-NEXT:    vshufps {{.*#+}} xmm0 = xmm0[1,0,0,1]
 ; AVX-32-NEXT:    vmovups %ymm0, 608(%eax)
 ; AVX-32-NEXT:    vzeroupper
 ; AVX-32-NEXT:    retl
 ;
 ; AVX-64-LABEL: PR15298:
 ; AVX-64:       # %bb.0: # %L.entry
-; AVX-64-NEXT:    vbroadcastss 304(%rdi), %xmm0
-; AVX-64-NEXT:    vxorps %xmm1, %xmm1, %xmm1
-; AVX-64-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2],ymm1[3,4,5,6,7]
+; AVX-64-NEXT:    vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; AVX-64-NEXT:    vshufps {{.*#+}} xmm0 = xmm0[1,0,0,1]
 ; AVX-64-NEXT:    vmovups %ymm0, 608(%rsi)
 ; AVX-64-NEXT:    vzeroupper
 ; AVX-64-NEXT:    retq

llvm/test/CodeGen/X86/sse-align-12.ll

@@ -40,8 +40,8 @@ define <4 x float> @b(ptr %y, <4 x float> %z) nounwind {
 define <2 x double> @c(ptr %y) nounwind {
 ; CHECK-LABEL: c:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    movups (%rdi), %xmm0
-; CHECK-NEXT:    shufps {{.*#+}} xmm0 = xmm0[2,3,0,1]
+; CHECK-NEXT:    movsd {{.*#+}} xmm0 = mem[0],zero
+; CHECK-NEXT:    movhps {{.*#+}} xmm0 = xmm0[0,1],mem[0,1]
 ; CHECK-NEXT:    retq
   %x = load <2 x double>, ptr %y, align 8
   %a = extractelement <2 x double> %x, i32 0