[clang] [llvm] [InstCombine] Preserve multi-dimensional array structure in GEP optimization (PR #176414)

[Nikita Popov via cfe-commits]

nikic wrote:

I've checked the test case, and I think the root cause here is actually additional unrolling. I haven't tried figuring out what exactly changes in the cost model. I see one invariant GEP index being hoisted out now, possibly that's enough to push it over the edge?

https://github.com/llvm/llvm-project/pull/176414