[llvm] [LV][EVL] Emit vp.merge intrinsic to enable out-loop reduction in EVL vectorization. (PR #101641)

[Mel Chen via llvm-commits]

================
@@ -1694,6 +1694,21 @@ class VPWidenIntrinsicRecipe : public VPRecipeWithIRFlags {
         MayWriteToMemory(MayWriteToMemory),
         MayHaveSideEffects(MayHaveSideEffects) {}
 
+  VPWidenIntrinsicRecipe(Intrinsic::ID VectorIntrinsicID,
+                         std::initializer_list<VPValue *> CallArguments,
+                         Type *Ty, DebugLoc DL = {})
+      : VPRecipeWithIRFlags(VPDef::VPWidenIntrinsicSC, CallArguments),
+        VectorIntrinsicID(VectorIntrinsicID), ResultTy(Ty) {
+    LLVMContext &Ctx = Ty->getContext();
+    AttributeList Attrs = Intrinsic::getAttributes(Ctx, VectorIntrinsicID);
+    MemoryEffects ME = Attrs.getMemoryEffects();
+    MayReadFromMemory = ME.onlyWritesMemory();
+    MayWriteToMemory = ME.onlyReadsMemory();
+    MayHaveSideEffects = MayWriteToMemory ||
+                         Attrs.hasFnAttr(Attribute::NoUnwind) ||
+                         !Attrs.hasFnAttr(Attribute::WillReturn);
+  }
----------------
Mel-Chen wrote:

Yes, separated to #113890

https://github.com/llvm/llvm-project/pull/101641