[llvm] [VPlan] Extract reverse operation for reverse accesses (PR #146525)

[Mel Chen via llvm-commits]

================
@@ -2281,21 +2295,37 @@ InstructionCost VPWidenCastRecipe::computeCost(ElementCount VF,
     return TTI::CastContextHint::Normal;
   };
 
+  using namespace llvm::VPlanPatternMatch;
   VPValue *Operand = getOperand(0);
   TTI::CastContextHint CCH = TTI::CastContextHint::None;
   // For Trunc/FPTrunc, get the context from the only user.
-  if ((Opcode == Instruction::Trunc || Opcode == Instruction::FPTrunc) &&
-      !hasMoreThanOneUniqueUser() && getNumUsers() > 0) {
-    if (auto *StoreRecipe = dyn_cast<VPRecipeBase>(*user_begin()))
-      CCH = ComputeCCH(StoreRecipe);
+  if (Opcode == Instruction::Trunc || Opcode == Instruction::FPTrunc) {
+    auto GetOnlyUser = [](const VPSingleDefRecipe *R) -> VPRecipeBase * {
+      if (R->getNumUsers() == 0 || R->hasMoreThanOneUniqueUser())
+        return nullptr;
+      return dyn_cast<VPRecipeBase>(*R->user_begin());
+    };
+
+    if (VPRecipeBase *Recipe = GetOnlyUser(this)) {
+      if (match(Recipe, m_VPInstruction<VPInstruction::Reverse>(m_VPValue())))
+        Recipe = GetOnlyUser(cast<VPInstruction>(Recipe));
+      if (Recipe)
+        CCH = ComputeCCH(Recipe);
----------------
Mel-Chen wrote:

RISC-V currently doesn’t have memory access instructions like this; AArch64 seems to have some. But I’m not sure if AArch64 supports combining load + shuffle + cast into a single instruction (I haven’t found one so far). Also, the current AArch64 TTI doesn’t use CastContextHint::Reversed for its decisions. I’m thinking that if load + reverse + cast can’t be combined into a load, maybe we could just remove CastContextHint::Reversed and return CastContextHint::None instead. What do you think?
cc. @david-arm 

https://github.com/llvm/llvm-project/pull/146525