[llvm] [VPlan] Introduce CSE pass (PR #151872)

Florian Hahn via llvm-commits llvm-commits at lists.llvm.org
Wed Aug 20 02:59:40 PDT 2025


================
@@ -1765,6 +1765,113 @@ void VPlanTransforms::clearReductionWrapFlags(VPlan &Plan) {
   }
 }
 
+namespace {
+struct VPCSEDenseMapInfo : public DenseMapInfo<VPSingleDefRecipe *> {
+  static bool isSentinel(const VPSingleDefRecipe *Def) {
+    return Def == getEmptyKey() || Def == getTombstoneKey();
+  }
+
+  /// Get any instruction opcode or intrinsic ID data embedded in recipe \p R.
+  /// Returns an optional pair, where the first element indicates whether it is
+  /// an intrinsic ID.
+  static std::optional<std::pair<bool, unsigned>>
+  getOpcodeOrIntrinsicID(const VPSingleDefRecipe *R) {
+    return TypeSwitch<const VPSingleDefRecipe *,
+                      std::optional<std::pair<bool, unsigned>>>(R)
+        .Case<VPInstruction, VPWidenRecipe, VPWidenCastRecipe,
+              VPWidenSelectRecipe, VPHistogramRecipe, VPPartialReductionRecipe,
+              VPReplicateRecipe>(
+            [](auto *I) { return std::make_pair(false, I->getOpcode()); })
+        .Case<VPWidenIntrinsicRecipe>([](auto *I) {
+          return std::make_pair(true, I->getVectorIntrinsicID());
+        })
+        .Default([](auto *) { return std::nullopt; });
+  }
+
+  /// During CSE, we can only handle certain recipes that don't read from
+  /// memory: if they read from memory, there could be an intervening write to
+  /// memory before the next instance is CSE'd, leading to an incorrect result.
+  /// We can extend the list of handled recipes in the future, provided we
+  /// account for the data embedded in them while checking for equality or
+  /// hashing.
+  static bool canHandle(const VPSingleDefRecipe *Def) {
+    return isa<VPInstruction, VPWidenRecipe, VPWidenCastRecipe,
+               VPWidenSelectRecipe, VPHistogramRecipe, VPReplicateRecipe,
+               VPWidenIntrinsicRecipe>(Def) &&
+           !Def->mayReadFromMemory();
+  }
+
+  /// Hash the underlying data of \p Def.
+  static unsigned getHashValue(const VPSingleDefRecipe *Def) {
+    const VPlan *Plan = Def->getParent()->getPlan();
+    VPTypeAnalysis TypeInfo(*Plan);
+    hash_code Result = hash_combine(
+        Def->getVPDefID(), getOpcodeOrIntrinsicID(Def),
+        TypeInfo.inferScalarType(Def), vputils::isSingleScalar(Def),
+        hash_combine_range(Def->operands()));
+    if (isa<VPReplicateRecipe>(Def))
+      return hash_combine(Result, Def->getUnderlyingInstr());
----------------
fhahn wrote:

I don't think that's needed.

The issue with `extractvalue`/`insertvalue` is that the index the extract/insert is not a proper operand in LLVM IR, hence also not in VPlan.

Better to just not handle `extractvalue`/`insertvalue` opcodes

https://github.com/llvm/llvm-project/pull/151872


More information about the llvm-commits mailing list