[llvm] [GISel][AMDGPU] Expand ShuffleVector (PR #124527)

[Matt Arsenault via llvm-commits]

================
@@ -384,6 +384,45 @@ void CombinerHelper::applyCombineConcatVectors(
   MI.eraseFromParent();
 }
 
+bool CombinerHelper::matchCombineShuffleToBuildVector(MachineInstr &MI) const {
+  assert(MI.getOpcode() == TargetOpcode::G_SHUFFLE_VECTOR &&
+         "Invalid instruction");
+  auto &Shuffle = cast<GShuffleVector>(MI);
+
+  auto SrcVec1 = Shuffle.getSrc1Reg();
+  auto SrcVec2 = Shuffle.getSrc2Reg();
+
+  LLT SrcVec1Type = MRI.getType(SrcVec1);
+  LLT SrcVec2Type = MRI.getType(SrcVec2);
+  return SrcVec1Type.isVector() && SrcVec2Type.isVector();
+}
+
+void CombinerHelper::applyCombineShuffleToBuildVector(MachineInstr &MI) const {
+  auto &Shuffle = cast<GShuffleVector>(MI);
+
+  auto SrcVec1 = Shuffle.getSrc1Reg();
+  auto SrcVec2 = Shuffle.getSrc2Reg();
+  auto EltTy = MRI.getType(SrcVec1).getElementType();
+  auto Width = MRI.getType(SrcVec1).getNumElements();
+
+  auto Unmerge1 = Builder.buildUnmerge(EltTy, SrcVec1);
+  auto Unmerge2 = Builder.buildUnmerge(EltTy, SrcVec2);
+
+  SmallVector<Register> Extracts;
+  // Select only applicable elements from unmerged values.
+  for (auto Val : Shuffle.getMask()) {
+    if (Val == -1)
+      Extracts.push_back(Builder.buildUndef(EltTy).getReg(0));
+    else if (Val < Width)
+      Extracts.push_back(Unmerge1.getReg(Val));
+    else
+      Extracts.push_back(Unmerge2.getReg(Val - Width));
+  }
+
+  Builder.buildBuildVector(MI.getOperand(0).getReg(), Extracts);
----------------
arsenm wrote:

I thought @aemerson already did the shuffle -> build case?

Should this try to fold directly to extract_subvector like the comment suggests if it's reading consecutive elements?

https://github.com/llvm/llvm-project/pull/124527