LoopFullUnrollPass failed due to constant inference in opaque pointer mode.

[dianqk]

I tried to disable the opaque pointer in rust-lang/rust#115339 to get the SLPVectorizer optimization.
See https://godbolt.org/z/GqdPGarsn.

According to my investigation, the success of optimization in non-opaque pointer mode relates to SROA.cpp#L1831-L1845 or InstructionCombining.cpp#L2220-L2274. Unfortunately, these two transformations have been removed during opaque pointer migration.

My plan is to restore the behavior of InstructionCombining.cpp#L2220-L2274 using alloc instead of bitcast.
Looking at the code comments, this seems to be an important optimization. So it might be worth backporting to LLVM 17?

[fhahn]

Could you share the LLVM IR before the SLP vectorizer?

[dianqk]

Could you share the LLVM IR before the SLP vectorizer?

There are two smaller IRs:

• https://llvm.godbolt.org/z/6dGGzo45o
• https://llvm.godbolt.org/z/f9774x53s

But I dont's see the value. See #65764 (comment).

[dianqk]

Could you share the LLVM IR before the SLP vectorizer?

I added an IR that looks pretty good in #65875.