| Issue |
143219
|
| Summary |
[flang][perf] Adding noalias for function arguments affects function specialization
|
| Labels |
flang
|
| Assignees |
vzakhari
|
| Reporter |
vzakhari
|
CPU2017/548 slowed down after https://github.com/llvm/llvm-project/commit/a0d699a8e686cba99690cf28463d14526c5bfbc8, so this change was partially reverted by https://github.com/llvm/llvm-project/commit/d16ecad968244ca42ce09cb3199dea1ec994b211
I could reproduce it on neoverse-v1 with `-Ofast -flto -mcpu=neoverse-v1 -fuse-ld=lld -mmlir -force-no-alias=false/true`.
It looks like the `digits_2` function does not pass the function specialization threshold on the Latency bonus.
Without noalias:
```
FnSpecialization: Specialization bonus {Latency = 2424 (71%)}
```
With noalias:
```
FnSpecialization: Specialization bonus {Latency = 1098 (34%)}
FnSpecialization: No possible specializations found for _QMbrute_forcePdigits_2
```
The current threshold is 40%.
It looks like the big portion of the latency bounds comes from the multiple `mod(row, 3)` expressions, which are not CSEd without noalias - so function specialization computes a lot of bonus for them when `row` is a constant. With noalias, there is a single `srem` instruction that computes `mod(row, 3)` at the beginning of the function.
I will investigate more to see if the function specialization misses to account for some bonuses.
![]()
_______________________________________________
llvm-bugs mailing list
llvm-bugs@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
