[llvm] [AArch64][Machine-Combiner] Split loads into lanes of neon vectors into multiple vectors when possible (PR #142941)

[Jonathan Cohen via llvm-commits]

================
@@ -7367,11 +7372,241 @@ static bool getMiscPatterns(MachineInstr &Root,
   return false;
 }
 
+static bool getGatherPattern(MachineInstr &Root,
+                             SmallVectorImpl<unsigned> &Patterns,
+                             unsigned LoadLaneOpCode, unsigned NumLanes) {
+  const MachineRegisterInfo &MRI = Root.getMF()->getRegInfo();
+  const TargetRegisterInfo *TRI =
+      Root.getMF()->getSubtarget().getRegisterInfo();
+
+  // The root of the pattern must load into the last lane of the vector.
+  if (Root.getOperand(2).getImm() != NumLanes - 1)
+    return false;
+
+  // Check that we have load into all lanes except lane 0.
+  // For each load we also want to check that:
+  // 1. It has a single debug use (since we will be replacing the virtual
+  // register)
+  // 2. That the addressing mode only uses a single offset register.
+  auto *CurrInstr = MRI.getUniqueVRegDef(Root.getOperand(1).getReg());
+  auto Range = llvm::seq<unsigned>(1, NumLanes - 1);
+  SmallSet<unsigned, 4> RemainingLanes(Range.begin(), Range.end());
+  while (RemainingLanes.begin() != RemainingLanes.end() &&
----------------
jcohen-apple wrote:

Thanks

https://github.com/llvm/llvm-project/pull/142941