[LV] Use VPWidenIntrinsicRecipe to vp_icmp/vp_fcmp

Author: LiqinWeng

llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp

@@ -1000,6 +1000,13 @@ void VPWidenIntrinsicRecipe::execute(VPTransformState &State) {
     // Remove EVL from Args
     Args.pop_back();
 
+    if (VPCmpIntrinsic::isVPCmp(VectorIntrinsicID)) {
+      auto &Ctx = State.Builder.getContext();
+      Value *Pred = MetadataAsValue::get(
+          Ctx, MDString::get(Ctx, CmpInst::getPredicateName(getPredicate())));
+      Args.push_back(Pred);
+    }
+
     Value *VPInst = VBuilder.createSimpleIntrinsic(
         VectorIntrinsicID, TysForDecl[0], Args, "vp.call");
 
@@ -1061,6 +1068,18 @@ InstructionCost VPWidenIntrinsicRecipe::computeCost(ElementCount VF,
     ParamTys.push_back(
         ToVectorTy(Ctx.Types.inferScalarType(getOperand(I)), VF));
 
+  if (std::optional<unsigned> FOp =
+          VPIntrinsic::getFunctionalOpcodeForVP(VectorIntrinsicID)) {
+    if (VPCmpIntrinsic::isVPCmp(VectorIntrinsicID)) {
+      Instruction *CtxI = dyn_cast_or_null<Instruction>(getUnderlyingValue());
+      Type *VectorTy = ToVectorTy(Ctx.Types.inferScalarType(getOperand(0)), VF);
+      return Ctx.TTI.getCmpSelInstrCost(FOp.value(), VectorTy, nullptr,
+                                        getPredicate(), CostKind,
+                                        {TTI::OK_AnyValue, TTI::OP_None},
+                                        {TTI::OK_AnyValue, TTI::OP_None}, CtxI);
+    }
+  }
+
   // TODO: Rework TTI interface to avoid reliance on underlying IntrinsicInst.
   FastMathFlags FMF = hasFastMathFlags() ? getFastMathFlags() : FastMathFlags();
   IntrinsicCostAttributes CostAttrs(

llvm/test/Transforms/LoopVectorize/RISCV/vplan-vp-cmp-intrinsics.ll

@@ -27,7 +27,7 @@ define void @vp_icmp(ptr noalias %a, ptr noalias %b, ptr noalias %c, i64 %N) {
 ; IF-EVL-NEXT:     CLONE ir<[[GEP2:%.+]]> = getelementptr inbounds ir<%c>, vp<[[ST]]>
 ; IF-EVL-NEXT:     vp<[[PTR2:%[0-9]+]]> = vector-pointer ir<[[GEP2]]>
 ; IF-EVL-NEXT:     WIDEN ir<[[LD2:%.+]]> = vp.load vp<[[PTR2]]>, vp<[[EVL]]>
-; IF-EVL-NEXT:     WIDEN ir<[[ICMP:%.+]]> = vp.icmp sgt ir<[[LD1]]>, ir<[[LD2]]>, vp<[[EVL]]>
+; IF-EVL-NEXT:     WIDEN-INTRINSIC ir<[[ICMP:%.+]]> = call sgt llvm.vp.icmp(ir<[[LD1]]>, ir<[[LD2]]>, vp<[[EVL]]>)
 ; IF-EVL-NEXT:     WIDEN-CAST ir<[[ZEXT:%.+]]> = zext ir<[[ICMP]]> to i32
 ; IF-EVL-NEXT:     CLONE ir<[[GEP3:%.+]]> = getelementptr inbounds ir<%a>, vp<[[ST]]>
 ; IF-EVL-NEXT:     vp<[[PTR3:%[0-9]+]]> = vector-pointer ir<[[GEP3]]>
@@ -82,7 +82,7 @@ define void @vp_fcmp(ptr noalias %a, ptr noalias %b, ptr noalias %c, i64 %N) {
 ; IF-EVL-NEXT:     CLONE ir<[[GEP2:%.+]]> = getelementptr inbounds ir<%c>, vp<[[ST]]>
 ; IF-EVL-NEXT:     vp<[[PTR2:%[0-9]+]]> = vector-pointer ir<[[GEP2]]>
 ; IF-EVL-NEXT:     WIDEN ir<[[LD2:%.+]]> = vp.load vp<[[PTR2]]>, vp<[[EVL]]>
-; IF-EVL-NEXT:     WIDEN ir<[[FCMP:%.+]]> = vp.fcmp ogt ir<[[LD1]]>, ir<[[LD2]]>, vp<[[EVL]]>
+; IF-EVL-NEXT:     WIDEN-INTRINSIC ir<[[FCMP:%.+]]> = call ogt llvm.vp.fcmp(ir<[[LD1]]>, ir<[[LD2]]>, vp<[[EVL]]>)
 ; IF-EVL-NEXT:     WIDEN-CAST ir<[[UITOFP:%.+]]> = uitofp  ir<[[FCMP]]> to float
 ; IF-EVL-NEXT:     CLONE ir<[[GEP3:%.+]]> = getelementptr inbounds ir<%a>, vp<[[ST]]>
 ; IF-EVL-NEXT:     vp<[[PTR3:%[0-9]+]]> = vector-pointer ir<[[GEP3]]>

llvm/test/Transforms/LoopVectorize/RISCV/vplan-vp-select-intrinsics.ll

@@ -27,7 +27,7 @@
  ; IF-EVL-NEXT:     CLONE ir<[[GEP2:%.+]]> = getelementptr inbounds ir<%c>, vp<[[ST]]>
  ; IF-EVL-NEXT:     vp<[[PTR2:%[0-9]+]]> = vector-pointer ir<[[GEP2]]>
  ; IF-EVL-NEXT:     WIDEN ir<[[LD2:%.+]]> = vp.load vp<[[PTR2]]>, vp<[[EVL]]>
- ; IF-EVL-NEXT:     WIDEN ir<[[CMP:%.+]]> = icmp sgt ir<[[LD1]]>, ir<[[LD2]]>
+ ; IF-EVL-NEXT:     WIDEN-INTRINSIC ir<[[CMP:%.+]]> = call sgt llvm.vp.icmp(ir<[[LD1]]>, ir<[[LD2]]>, vp<[[EVL]]>)
  ; IF-EVL-NEXT:     WIDEN-INTRINSIC ir<[[SUB:%.+]]> = call llvm.vp.sub(ir<0>, ir<[[LD2]]>, vp<[[EVL]]>)
  ; IF-EVL-NEXT:     WIDEN-INTRINSIC vp<[[SELECT:%.+]]> = call llvm.vp.select(ir<[[CMP]]>, ir<%1>, ir<%2>, vp<[[EVL]]>)
  ; IF-EVL-NEXT:     WIDEN-INTRINSIC ir<[[ADD:%.+]]> = call llvm.vp.add(vp<[[SELECT]]>, ir<[[LD1]]>, vp<[[EVL]]>)