[Mlir-commits] [clang-tools-extra] [libcxx] [mlir] [clang] [llvm] [flang] [LV] Improve AnyOf reduction codegen. (PR #78304)

Florian Hahn llvmlistbot at llvm.org
Mon Feb 5 06:29:38 PST 2024


================
@@ -1079,16 +1070,13 @@ Value *llvm::createAnyOfTargetReduction(IRBuilderBase &Builder, Value *Src,
     NewVal = SI->getTrueValue();
   }
 
-  // Create a splat vector with the new value and compare this to the vector
-  // we want to reduce.
-  ElementCount EC = cast<VectorType>(Src->getType())->getElementCount();
-  Value *Right = Builder.CreateVectorSplat(EC, InitVal);
-  Value *Cmp =
-      Builder.CreateCmp(CmpInst::ICMP_NE, Src, Right, "rdx.select.cmp");
-
   // If any predicate is true it means that we want to select the new value.
-  Cmp = Builder.CreateOrReduce(Cmp);
-  return Builder.CreateSelect(Cmp, NewVal, InitVal, "rdx.select");
+  Value *AnyOf =
+      Src->getType()->isVectorTy() ? Builder.CreateOrReduce(Src) : Src;
+  // The compares in the loop may yield poison, which propagates through the
+  // bitwise ORs. Freeze it here before the condition is used.
+  AnyOf = Builder.CreateFreeze(AnyOf);
----------------
fhahn wrote:

> Similar to a plain "result |= value[i]" OR reduction,

Freeze won't be needed in that case; if there's already a binary OR in the input, poison from the compare gets already propagated. It is only needed when converting from the `select` form (which doesn't propagate poison from the condition to its result)

Yes, but this will need a bit of additional refactoring, in particular how `createAndCollectMergePhiForReduction` looks up the reduction result value and when `ComputeReductionResult` VPInstructions are created.

https://github.com/llvm/llvm-project/pull/78304


More information about the Mlir-commits mailing list