[llvm] [AArch64][GlobalISel] Improve codegen for G_VECREDUCE_{SMIN,SMAX,UMIN,UMAX} for odd-sized vectors (PR #82740)

[Dhruv Chawla via llvm-commits]

================
@@ -5420,6 +5456,35 @@ LegalizerHelper::moreElementsVector(MachineInstr &MI, unsigned TypeIdx,
     Observer.changedInstr(MI);
     return Legalized;
   }
+  case TargetOpcode::G_VECREDUCE_FADD:
+  case TargetOpcode::G_VECREDUCE_FMUL:
+  case TargetOpcode::G_VECREDUCE_ADD:
+  case TargetOpcode::G_VECREDUCE_MUL:
+  case TargetOpcode::G_VECREDUCE_AND:
+  case TargetOpcode::G_VECREDUCE_OR:
+  case TargetOpcode::G_VECREDUCE_XOR:
+  case TargetOpcode::G_VECREDUCE_SMAX:
+  case TargetOpcode::G_VECREDUCE_SMIN:
+  case TargetOpcode::G_VECREDUCE_UMAX:
+  case TargetOpcode::G_VECREDUCE_UMIN: {
+    LLT OrigTy = MRI.getType(MI.getOperand(1).getReg());
+    MachineOperand &MO = MI.getOperand(1);
+    auto NewVec = MIRBuilder.buildPadVectorWithUndefElements(MoreTy, MO);
----------------
dc03-work wrote:

Using G_BUILD_VECTOR seems to cause quite a few regressions: https://gist.github.com/dc03-work/66c179f4d380cfac3ed7b36525a583ef.

https://github.com/llvm/llvm-project/pull/82740