Add visitGEPOfAlloc to restore the visitGEPOfBitcast part of the behavior of non-opaque pointers.

[dianqk]

Closes #65763.

[nikic]

We definitely shouldn't do this. From the issue, it's not clear where the problem actually is (given lack of reproducer), but I expect that this needs to be solved by teaching some transform to use offset-based reasoning for GEPs (or alternatively, by canonicalizing to i8 GEPs).

[dianqk]

We definitely shouldn't do this. From the issue, it's not clear where the problem actually is (given lack of reproducer), but I expect that this needs to be solved by teaching some transform to use offset-based reasoning for GEPs (or alternatively, by canonicalizing to i8 GEPs).

Thanks for the reply. I feel like I understand the issue a little better.
I found out earlier that SROA was not canonicalizing all the GEPs to i8 GEPs due to the escaped scene.
Since non-opaque pointers are OK, I incorrectly guessed it was a regression issue. So I didn't look at that part in depth.

I tried a lot of reproducers using llvm-reduce. An LLVM any-version reproducer may not exist.
I believe it was the various transformations of GEP that led to the constant inference that was lucky. That should be the essence of the problem.
Here just lucky in the case of non-opaque pointers.
I can execute inline before any pass. This is a reproducer of LLVM 16 not being able to optimize successfully in either mode.

Do you think bringing ahead GVN to LoopFullUnroll is the right choice? This should provide more optimization opportunities for LoopFullUnroll.

[dianqk]

Preparing an early-cse PR.