[llvm] [LV] Compute value of escaped induction based on the computed end value. (PR #110576)

[Florian Hahn via llvm-commits]

================
@@ -2747,17 +2747,24 @@ void InnerLoopVectorizer::fixupIVUsers(PHINode *OrigPhi,
       if (isa_and_nonnull<FPMathOperator>(II.getInductionBinOp()))
         B.setFastMathFlags(II.getInductionBinOp()->getFastMathFlags());
 
-      Value *CountMinusOne = B.CreateSub(
-          VectorTripCount, ConstantInt::get(VectorTripCount->getType(), 1));
-      CountMinusOne->setName("cmo");
-
       VPValue *StepVPV = Plan.getSCEVExpansion(II.getStep());
       assert(StepVPV && "step must have been expanded during VPlan execution");
       Value *Step = StepVPV->isLiveIn() ? StepVPV->getLiveInIRValue()
                                         : State.get(StepVPV, VPLane(0));
-      Value *Escape =
-          emitTransformedIndex(B, CountMinusOne, II.getStartValue(), Step,
-                               II.getKind(), II.getInductionBinOp());
+      Value *Escape = nullptr;
+      if (EndValue->getType()->isIntegerTy())
+        Escape = B.CreateSub(EndValue, Step);
+      else if (EndValue->getType()->isPointerTy())
+        Escape = B.CreatePtrAdd(EndValue, B.CreateNeg(Step));
+      else if (EndValue->getType()->isFloatingPointTy()) {
+        Escape = B.CreateBinOp(II.getInductionBinOp()->getOpcode() ==
----------------
fhahn wrote:

Unfortunately I am not sure there's a way to do this more concisely overall. Negating the step would work for both integer and pointers, but then we still would need to select between Add/PtrAdd. And add(neg()) requires an additional instructions for integers.

For floating points, perhaps it might be possible to normalize the representation of currently `fsub step` to `fadd -step` in general, but we would still need to handle FP values separately. I left things as is for now

https://github.com/llvm/llvm-project/pull/110576