[llvm] [AArch64][LSR] Prioritize Insns in isLSRCostLess (PR #226525)

[John Brawn via llvm-commits]

john-brawn-arm wrote:

> I'm not sure if I would trust results from the llvm-test-suite a huge amount at this point. It needs very careful setup and a lot of the benchmarks are too short-running to provide reliable results, the noise levels can be too high. There are better lower-noise benchmarks out there.

Do you have any recommendations for better benchmarks to use?

> I think a few of us have looked at using instruction count but it hasn't performed well enough to pursue. A lot of the test changes in this patch are simple loops but still look worse, with more instructions even if is far from reaching regpressure limits.

Most (maybe all?) of the cases where changing to prioritizing instruction count makes things worse are due to the preferred addressing mode being none, which I'll be changing next. https://github.com/john-brawn-arm/llvm-project/commit/7bea0e1bc607310b75ec386a2f48af363d2979a1 shows the results of setting the preferred addressing mode to all, and many of the test changes are just reverting tests to what they were before this commit. Possibly it makes sense to change both these things at the same time, instead of one then the other.

https://github.com/llvm/llvm-project/pull/226525