On Thu, 17 Nov 2022 at 15:30, Jeff Law <jeffreya...@gmail.com> wrote: > > > On 11/10/22 14:36, Philipp Tomsich wrote: > > Add a split for cases where we can use two bclri (or one bclri and an > > andi) to clear two bits. > > > > gcc/ChangeLog: > > > > * config/riscv/bitmanip.md (*bclri<mode>_nottwobits): New pattern. > > (*bclridisi_nottwobits): New pattern, handling the sign-bit. > > * config/riscv/predicates.md (const_nottwobits_operand): > > New predicate. > > > > gcc/testsuite/ChangeLog: > > > > * gcc.target/riscv/zbs-bclri.c: New test. > > Don't we only have to worry about (subreg:DI (reg:SI )) to preserve the > extension constraints? Not that I think there's any value in allowing
That is the reason for the !paradoxical_subreg_p(...) check in "bclri<mode>_nottwobits"... so at pattern will always be safe. Do you see a risk on the "*bclridisi_nottwobits"? > HI/QI, mostly wanting to double check my understanding of the extension > constraints. QI will be fully covered by simm12 (e.g., using the andi instruction). For HI, we might in fact want to have a special case (similar to "*bclridisi_nottwobits") — but the reach of the simm12 seems to cover enough of 16-bit space that this hasn't shown up (yet) as enough of a problem to warrant a special case. Philipp.
