[llvm] [LV] Increase vectorize-memory-check-threshold to 256 (PR #151712)

[Igor Kirillov via llvm-commits]

igogo-x86 wrote:

@fhahn If you mean raising the threshold from 128 to 256, that didn’t affect the performance of the SPEC benchmarks or the compile time. The improvements I showed above - I haven’t prototyped them yet, so I can’t say.

But overall, this change plus the reduction in the number of runtime checks helps the cactus benchmark. This benchmark’s hottest area is within three huge functions (one of them is extremely large), and it runs twice as fast after vectorising them. My estimates show that after reducing the number of runtime checks and raising the threshold, we can vectorise two out of three functions, and they are responsible for 50% of the runtime. To vectorise the third function, we may have to raise the threshold even higher, but we can decide that later.

https://github.com/llvm/llvm-project/pull/151712