[PATCH] D79164: [CostModel] getCFInstrCost

[Florian Hahn via Phabricator via llvm-commits]

fhahn added a comment.

It seems like this introduces a 9% performance regression for libquantum from SPEC2006 on AArch64 with -O3 LTO & PGO. I think AArch64 uses the implementation from `TargetTransformInfoImpl.h`, which overestimates the cost of branches on bigger cores.

In D79164#2128516 <https://reviews.llvm.org/D79164#2128516>, @maxim-kuvyrkov wrote:

> Hi Sam,
>  Linaro benchmarking CI flagged this patch as it regresses SPEC2k6's 462.libquantum by 10% at -O2, -O3 and "-O3 -flto" for aarch64-linux-gnu.  Impact on other benchmarks is below 2%, which is within noise.
>
> For -O2 profile changes are:
>  <benchmark, exe/symbol>,   %time, %size, time1, time2, size1, size2
>  462.libquantum,libquantum_base.default                       ,111,101,9859,10962,33140,33428
>  462.libquantum,[.] quantum_toffoli                           ,113,125,5616,6360,320,400
>  462.libquantum,[.] quantum_sigma_x                           ,100,100,2498,2499,224,224
>  462.libquantum,[.] quantum_cnot                              ,130,135,1150,1495,288,388
>
> After this patch quantum_toffoli is 13% slower and 25% bigger, and quantum_cnot is 30% slower and 35% bigger.
>
> Would you please investigate this?  Such big performance and code size regressions might expose a corner-case in your patch.


This should be fixed by 1ccc49924aeb <https://reviews.llvm.org/rG1ccc49924aeb7bb5fbd2eb956243d16db3a02b08>


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D79164/new/

https://reviews.llvm.org/D79164