fixup! [CIR][LowerToLLVM] Fix crash in PtrStrideOp lowering

[smeenai]

After recent rebases (likely as a result of [1]), the LLVMIR stride
operand can be generated by an unrealized_conversion_cast whereas the
CIR operand is coming directly from a block argument (and thus has no
defining op), so we need to separately check the CIR operand for the
existence of a defining op.

[1] llvm/llvm-project@2d50029

[smeenai]

This might be related to llvm/llvm-project#104668, which was reverted and relanded with a fix (with some further follow-ups going in). I'll make this a draft for now and come back to this during the next rebase.

[bcardosolopes]

Ready to move out of draft (and merge)!

[smeenai]

Ready to move out of draft (and merge)!

I'll see if this is still needed after the rebase I'm working on right now and revive it if needed :)

[ChuanqiXu9]

I verified this is needed after the new trunk.

[smeenai]

I spent some time trying to understand this more. If you go back to the original introduction of this commit in 9720c61, both the LLVMIR and CIR versions of the stride are coming directly from a block argument, so indexOp check fails. After the rebase, the LLVMIR version comes from an unrealized_conversion_cast but the CIR version still comes directly from the block argument, so the indexOp check succeeds and we run into the issue this commit is fixing.

I think that's related to llvm/llvm-project#101514. I'd misunderstood the follow-ups (such as llvm/llvm-project#107109 and llvm/llvm-project#106760) as removing unrealized_conversion_cast entirely (which got me confused about the need for this change), whereas they're actually related to removing it from the output IR but consistently having it present in the intermediate stages. I updated the description and I'll land it when CI passes.