[llvm] [VPlan] Expand VPWidenPointerInductionRecipe into separate recipes (PR #148274)

[Luke Lau via llvm-commits]

================
@@ -854,6 +856,11 @@ Value *VPInstruction::generate(VPTransformState &State) {
     Value *Addend = State.get(getOperand(1), VPLane(0));
     return Builder.CreatePtrAdd(Ptr, Addend, Name, getGEPNoWrapFlags());
   }
+  case VPInstruction::WidePtrAdd: {
+    Value *Ptr = State.get(getOperand(0), true);
+    Value *Addend = State.get(getOperand(1), vputils::onlyFirstLaneUsed(this));
----------------
lukel97 wrote:

Oddly enough, `vputils::onlyFirstLaneUsed` triggers on two test cases Transforms/LoopVectorize/AArch64/sve-widen-gep.ll and Transforms/LoopVectorize/X86/pr48340.ll.

In at least pr48340 it comes from an unrolled pointer induction where the second unrolled gep isn't used? 

```
vector.body:
  EMIT-SCALAR vp<%index> = phi [ ir<0>, ir-bb<vector.ph> ], [ vp<%index.next>, vector.body ]
  EMIT-SCALAR vp<%pointer.phi> = phi [ ir<%p>, ir-bb<vector.ph> ], [ vp<%ptr.ind>, vector.body ]
  EMIT vp<%2> = mul ir<4>, ir<3>
  EMIT vp<%3> = broadcast vp<%2>
  EMIT vp<%4> = step-vector i64
  EMIT vp<%5> = add vp<%3>, vp<%4>
  EMIT vp<%6> = mul vp<%5>, ir<1024>
  EMIT vp<%vector.gep> = wide-ptradd vp<%pointer.phi>, vp<%6>
  EMIT vp<%7> = mul ir<1024>, ir<16>
  EMIT vp<%8> = mul ir<4>, ir<0>
  EMIT vp<%9> = broadcast vp<%8>
  EMIT vp<%10> = step-vector i64
  EMIT vp<%11> = add vp<%9>, vp<%10>
  EMIT vp<%12> = mul vp<%11>, ir<1024>
  EMIT vp<%vector.gep>.1 = wide-ptradd vp<%pointer.phi>, vp<%12>
  WIDEN ir<%v> = load vp<%vector.gep>
  EMIT vp<%index.next> = add nuw vp<%index>, ir<16>
  EMIT vp<%ptr.ind> = ptradd vp<%pointer.phi>, vp<%7>
  EMIT branch-on-count vp<%index.next>, ir<%n.vec>
Successor(s): middle.block, vector.body
```

So `onlyFirstLaneUsed` returns true, and we need to continue to generate a scalar for it to avoid a (mild) regression:

```
-; CHECK-NEXT:    [[VECTOR_GEP4:%.*]] = getelementptr i8, ptr [[POINTER_PHI]], i64 0
+; CHECK-NEXT:    [[VECTOR_GEP4:%.*]] = getelementptr i8, ptr [[POINTER_PHI]], <4 x i64> <i64 0, i64 1024, i64 2048, i64 3072>
```

https://github.com/llvm/llvm-project/pull/148274