[llvm-branch-commits] [llvm] [LoopVectorize][NFC] Refactor widening decision logic (PR #140722)

[David Sherwood via llvm-branch-commits]

================
@@ -2988,30 +2990,32 @@ bool LoopVectorizationCostModel::interleavedAccessCanBeWidened(
                           : TTI.isLegalMaskedStore(Ty, Alignment, AS);
 }
 
-bool LoopVectorizationCostModel::memoryInstructionCanBeWidened(
-    Instruction *I, ElementCount VF) {
+LoopVectorizationCostModel::InstWidening
+LoopVectorizationCostModel::memoryInstructionCanBeWidened(Instruction *I,
+                                                          ElementCount VF) {
   // Get and ensure we have a valid memory instruction.
   assert((isa<LoadInst, StoreInst>(I)) && "Invalid memory instruction");
 
   auto *Ptr = getLoadStorePointerOperand(I);
   auto *ScalarTy = getLoadStoreType(I);
 
   // In order to be widened, the pointer should be consecutive, first of all.
-  if (!Legal->isConsecutivePtr(ScalarTy, Ptr))
-    return false;
+  auto Stride = Legal->isConsecutivePtr(ScalarTy, Ptr);
+  if (!Stride)
+    return CM_Unknown;
 
   // If the instruction is a store located in a predicated block, it will be
   // scalarized.
   if (isScalarWithPredication(I, VF))
-    return false;
+    return CM_Unknown;
 
   // If the instruction's allocated size doesn't equal it's type size, it
   // requires padding and will be scalarized.
   auto &DL = I->getDataLayout();
   if (hasIrregularType(ScalarTy, DL))
-    return false;
+    return CM_Unknown;
 
-  return true;
+  return Stride == 1 ? CM_Widen : CM_Widen_Reverse;
----------------
david-arm wrote:

Is it worth adding an assert that Stride is 1 or -1, since you're assuming that when the stride is not 1 it must be -1?

https://github.com/llvm/llvm-project/pull/140722