[llvm] [LV] Enable auto-vectorisation of loops with uncountable exits (PR #133099)

[David Sherwood via llvm-commits]

david-arm wrote:

> I tried this out on RISC-V and we get 107% more loops vectorized on one of the SPEC CPU 2017 benchmarks, with an overall geomean 5% improvement across all the benchmarks.
> 
> There's also an LNT run [here](https://lnt.lukelau.me/db_default/v4/nts/369?show_delta=yes&show_previous=yes&show_stddev=yes&show_mad=yes&show_all=yes&show_all_samples=yes&show_sample_counts=yes&show_small_diff=yes&num_comparison_runs=0&test_filter=&test_min_value_filter=&aggregation_fn=min&MW_confidence_lv=0.05&compare_to=370&submit=Update) and there's no regressions. So no issues on the RISC-V side!

OK great! I'm glad that you see no regressions for RISC-V, although it's worth saying this patch only starts vectorising early exit loops if we know something about the object we're loading from. For example, if you can prove that all loads are within the bounds of an `alloca` then it's safe to vectorise.

https://github.com/llvm/llvm-project/pull/133099