https://gcc.gnu.org/bugzilla/show_bug.cgi?id=128211
--- Comment #2 from Dhruv Chawla <dhruvc at gcc dot gnu.org> --- (In reply to Drea Pinski from comment #1) > Reading this it really does seem like really really bad luck. > Pretty much, yeah. It is like https://gcc.gnu.org/bugzilla/show_bug.cgi?id=122455#c13 where store-to-load forwarding for pair-fused ops is the culprit, though it is slightly different here. > The only thing I can think of is maybe we should not save/restore using > stp/ldp after all. But doesn't the ABI require that? > > Also I am not seeing the regression on the hardware I have access to. So > this seems like something you should really report to the hw designers too. I was able to repro it on both Grace (Neoverse-v2) and DGX Spark (Cortex-X925). What hardware are you using? With the min-repro, you should see results like: Neoverse-V2 (Grace) Cortex-X925 count in x22 5.61 ns/iter 6.80 ns/iter count in x24 1.82 ns/iter 2.19 ns/iter
