[llvm] [VPlan] Materialize VF and VFxUF using VPInstructions. (PR #152879)

[Florian Hahn via llvm-commits]

================
@@ -3335,6 +3335,33 @@ void VPlanTransforms::materializeVectorTripCount(VPlan &Plan,
   VectorTC.replaceAllUsesWith(Res);
 }
 
+void VPlanTransforms::materializeVFAndVFxUF(VPlan &Plan, VPBasicBlock *VectorPH,
+                                            ElementCount VFEC) {
+  VPBuilder Builder(VectorPH, VectorPH->begin());
+  Type *TCTy = VPTypeAnalysis(Plan).inferScalarType(Plan.getTripCount());
+  VPValue &VF = Plan.getVF();
+  VPValue &VFxUF = Plan.getVFxUF();
+  if (VF.getNumUsers()) {
+    VPValue *RuntimeVF = Builder.createElementCount(TCTy, VFEC);
+    if (any_of(VF.users(), [&VF](VPUser *U) { return !U->usesScalars(&VF); })) {
+      auto *BC = Builder.createNaryOp(VPInstruction::Broadcast, RuntimeVF);
+      VF.replaceUsesWithIf(
+          BC, [&VF](VPUser &U, unsigned) { return !U.usesScalars(&VF); });
+    }
+    VF.replaceAllUsesWith(RuntimeVF);
+
+    VPValue *UF = Plan.getOrAddLiveIn(ConstantInt::get(TCTy, Plan.getUF()));
+    auto *MulByUF = Plan.getUF() == 1 ? RuntimeVF
----------------
fhahn wrote:

Hm, but they generate operations in slightly different order, the first does (VF * Vscale) * UF, whereas otherwise we do ((VF * UF) * vscale), the former will serve users of the runtime-vf.

Both compute the same value, but in different order. The patch keeps the sequence the same as for IR.


https://github.com/llvm/llvm-project/pull/152879