https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79149
--- Comment #14 from wilco at gcc dot gnu.org --- (In reply to Arnd Bergmann from comment #13) > (In reply to wilco from comment #12) > > Does wp512 use 64-bit types? If so, this is likely PR77308. > > Yes, as seen in the attachment it uses lots of 64-bit operations. However, > it sounds like PR77308 is ARM specific, but I see the same behavior > on most other architectures, including 64-bit ones. Quoting the > kernel patch I linked to, with stack frame sizes for the function > depending on architecture, optimization flags and compiler version > (only 4.9 and 7.0 here, there is little difference anyway) The 64-bit expansion issues in PR77308 are ARM specific indeed, but it also shows scheduling causes unnecessary high register pressure and extra spilling. Your results indicate that -fsched-pressure should really be the default. And given that it still results in more spilling compared to not scheduling, it probably needs to be made less aggressive or compute pressure more accurately.
