https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80706
--- Comment #4 from Uroš Bizjak <ubizjak at gmail dot com> --- (In reply to Jakub Jelinek from comment #3) > So peephole2 does here: > - fstpl 24(%esp) > - movq 24(%esp), %xmm0 > - movq %xmm0, 32(%esp) > + fstpl 32(%esp) > movl $0, 32(%esp) > That is a nice simplification, but has one extra requirement not checked > (and hard to check) in the peephole2 patterns - that the memory slot stored > in the first store is a scratch memory not used afterwards (or overwritten > first, i.e. dead). While we have peep2_reg_dead_p predicates, we don't have > peep2_mem_dead_p and implementing that would be hard, only DSE has > infrastructure to do that, but dse2 is run before peephole2 pass. > All we could do is simplify the mem[sp+24]=st; xmm0=[sp+24]; [sp+32]=xmm0; > into mem[sp+24]=st; mem[sp+32]=st; and let the regstack pass figure out > something with it - fstl 24(%esp); fstpl 32(%esp) ?). DSE isn't run > afterwards, so it would be nice to do that earlier though. Let's keep the dangling store to a temporary here. We already loaded the value from the memory, so one extra store won't hurt that much...
