aarch64: skip the extend of a scalar comparison's result

[fitzgen]

icmp and fcmp on scalars both lower through lower_cond_result_bool, whose every arm already clears upper bits, so extending it is a no-op.

This removes ~18,000 uxtb instructions emitted from a PCA-based subset of Sightglass (this is ~87% of uxtb instructions emitted and 0.220% of all instructions emitted). It also results in an average 0.42% faster execution (significant) in terms of cycles.

[alexcrichton]

We've generally tried to avoid lowering rules like this historically where correctness relies on other lowering rules in the system, although we also have some already for x64 so it's not the most principled stance per se. I'd be surprised though if this uextend+icmp showed up too too often in terms of materializing the result of a comparison vs feeding the uextend+icmp into a branch/select/trap/etc. Is this perhaps something where we could get the lion's share of the benefit by shifting around these rules to where conditions are lowered or similar?

[fitzgen]

We've generally tried to avoid lowering rules like this historically where correctness relies on other lowering rules in the system

I pushed another commit that folds this into the matching rule, but unfortunately, this has the effect of re-lowering the icmp/fcmp at each uextend site, rather than reusing it across many. Not sure what the best way forward here is...

[alexcrichton]

With optimizations enabled I think that would be resolved with GVN though, right?

[fitzgen]

With optimizations enabled I think that would be resolved with GVN though, right?

Yeah, and in fact I read the diff backwards (d'oh) so the filetests are asserting that the duplication doesn't happen right now, so we should be good to merge this (assuming you think it looks good)