[llvm] [LV] Fix runtime-VF logic when generating RT-checks (PR #130118)

[David Sherwood via llvm-commits]

================
@@ -1924,21 +1924,17 @@ class GeneratedRTChecks {
                                  "vector.memcheck");
 
       auto DiffChecks = RtPtrChecking.getDiffChecks();
-      if (DiffChecks) {
-        Value *RuntimeVF = nullptr;
-        MemRuntimeCheckCond = addDiffRuntimeChecks(
-            MemCheckBlock->getTerminator(), *DiffChecks, MemCheckExp,
-            [VF, &RuntimeVF](IRBuilderBase &B, unsigned Bits) {
-              if (!RuntimeVF)
-                RuntimeVF = getRuntimeVF(B, B.getIntNTy(Bits), VF);
-              return RuntimeVF;
-            },
-            IC);
-      } else {
-        MemRuntimeCheckCond = addRuntimeChecks(
-            MemCheckBlock->getTerminator(), L, RtPtrChecking.getChecks(),
-            MemCheckExp, VectorizerParams::HoistRuntimeChecks);
-      }
+      MemRuntimeCheckCond =
+          DiffChecks
+              ? addDiffRuntimeChecks(
+                    MemCheckBlock->getTerminator(), *DiffChecks, MemCheckExp,
+                    [VF](IRBuilderBase &B, unsigned Bits) {
----------------
david-arm wrote:

@fhahn I thought that at first too, but when I looked into this it does look like a bug to me, but perhaps I'm missing something. The issue here is in `addDiffRuntimeChecks` this code:

```
  for (const auto &[SrcStart, SinkStart, AccessSize, NeedsFreeze] : Checks) {
    Type *Ty = SinkStart->getType();
    // Compute VF * IC * AccessSize.
    auto *VFTimesUFTimesSize =
        ChkBuilder.CreateMul(GetVF(ChkBuilder, Ty->getScalarSizeInBits()),
                             ConstantInt::get(Ty, IC * AccessSize));
```

we are potentially passing in different values for the `Bits` argument, i.e. for each value of `Bits` we should be creating a different runtime VF. See the bit in the lambda function `getRuntimeVF(B, B.getIntNTy(Bits), VF)`. So if on the first instance Bits=32, we create a 32-bit runtime VF and cache that. Then on the second iteration we might pass in Bits=64 and still return the 32-bit runtime VF. I don't even know how the `ChkBuilder.CreateMul` worked given the mismatched types.

https://github.com/llvm/llvm-project/pull/130118