[llvm] [AArch64][GlobalISel] Combine vecreduce(ext) to {U/S}ADDLV (PR #75832)

[David Green via llvm-commits]

================
@@ -418,6 +411,157 @@ void applyExtAddvToUdotAddv(MachineInstr &MI, MachineRegisterInfo &MRI,
   MI.eraseFromParent();
 }
 
+// Matches {U/S}ADDV(ext(x)) => {U/S}ADDLV(x)
+// Ensure that the type coming from the extend instruction is the right size
+bool matchExtUaddvToUaddlv(MachineInstr &MI, MachineRegisterInfo &MRI,
+                           std::pair<Register, bool> &MatchInfo) {
+  assert(MI.getOpcode() == TargetOpcode::G_VECREDUCE_ADD &&
+         "Expected G_VECREDUCE_ADD Opcode");
+
+  // Check if the last instruction is an extend
+  MachineInstr *ExtMI = getDefIgnoringCopies(MI.getOperand(1).getReg(), MRI);
+  auto ExtOpc = ExtMI->getOpcode();
+
+  if (ExtOpc == TargetOpcode::G_ZEXT)
+    std::get<1>(MatchInfo) = 0;
+  else if (ExtOpc == TargetOpcode::G_SEXT)
+    std::get<1>(MatchInfo) = 1;
+  else
+    return false;
+
+  // Check if the source register is a valid type
+  Register ExtSrcReg = ExtMI->getOperand(1).getReg();
+  LLT ExtSrcTy = MRI.getType(ExtSrcReg);
+  LLT DstTy = MRI.getType(MI.getOperand(0).getReg());
+  if ((DstTy.getScalarSizeInBits() == 16 &&
+       ExtSrcTy.getNumElements() % 8 == 0 && ExtSrcTy.getNumElements() < 256) ||
+      (DstTy.getScalarSizeInBits() == 32 &&
+       ExtSrcTy.getNumElements() % 4 == 0 &&
+       ExtSrcTy.getNumElements() < 65536) ||
----------------
davemgreen wrote:

It looks like getNumElements() returns a uint16_t, so these conditions will always be true. The 256 limit for i8 is probably enough.

https://github.com/llvm/llvm-project/pull/75832