[llvm] [InferAS] Infer the address space of inttoptr (PR #173244)

[Artem Belevich via llvm-commits]

================
@@ -369,15 +384,56 @@ getPointerOperands(const Value &V, const DataLayout &DL,
     return {II.getArgOperand(0)};
   }
   case Instruction::IntToPtr: {
-    assert(isNoopPtrIntCastPair(&Op, DL, TTI));
-    auto *P2I = cast<Operator>(Op.getOperand(0));
-    return {P2I->getOperand(0)};
+    if (isNoopPtrIntCastPair(&Op, DL, TTI)) {
+      auto *P2I = cast<Operator>(Op.getOperand(0));
+      return {P2I->getOperand(0)};
+    }
+    assert(isSafeToCastPtrIntPair(&Op, DL));
+    return {PtrIntCastPairs[&Op]};
   }
   default:
     llvm_unreachable("Unexpected instruction type.");
   }
 }
 
+bool InferAddressSpacesImpl::isSafeToCastPtrIntPair(
+    const Operator *I2P, const DataLayout &DL) const {
+  assert(I2P->getOpcode() == Instruction::IntToPtr);
+  if (PtrIntCastPairs.count(I2P))
+    return true;
+
+  if (I2P->getType()->isVectorTy())
+    return false;
+
+  auto *Xor = dyn_cast<Operator>(I2P->getOperand(0));
+  if (!Xor || Xor->getOpcode() != Instruction::Xor)
+    return false;
----------------
Artem-B wrote:

We do not have to copy/paste the functionality. We just need to figure out how to get instcombine to do the job it already does.

E.g. we could replace the argument of the `inttoptr(ptrvalue)` which may be operating on a pointer in originally known AS with something like this:

```
  as_inference_helper_intrinsic(
    and(bitmask_for bits_that_must_remain_unchanged, // constant for the AS we operate on
            xor(initial_ptrvalue, ptrvalue)
    ), // instcombine collapses `and` to constant zero if the upper bits are unchanged
    ptrvalue,
)
```

Later on if AS inference pass sees `as_inference_helper_intrinsic` with a constant 0 first argument, it can proceed inferring the AS. Otherwise, it just replaces the intrinsic with the `ptrvalue` (and the and/xor if they are still there) and keeps the pointer in the generic AS.

https://github.com/llvm/llvm-project/pull/173244