| Message ID | ZewfI3MMlf/SF2/e@tucnak |
|---|---|
| State | New |
| Headers | show |
| Series | fwprop: Restore previous behavior for forward propagation of RTL with MEMs [PR114284] | expand |
> Am 09.03.2024 um 09:36 schrieb Jakub Jelinek <jakub@redhat.com>: > > Hi! > > Before the recent PR111267 r14-8319 fwprop changes, fwprop would never try > to propagate what was not considered PROFITABLE, where the profitable part > actually was partly about profitability, partly about very good reasons > not to actually propagate and partly for cases where propagation is > completely incorrect. > In particular, classify_result has: > /* Allow (subreg (mem)) -> (mem) simplifications with the following > exceptions: > 1) Propagating (mem)s into multiple uses is not profitable. > 2) Propagating (mem)s across EBBs may not be profitable if the source EBB > runs less frequently. > 3) Propagating (mem)s into paradoxical (subreg)s is not profitable. > 4) Creating new (mem/v)s is not correct, since DCE will not remove the old > ones. */ > if (single_use_p > && single_ebb_p > && SUBREG_P (old_rtx) > && !paradoxical_subreg_p (old_rtx) > && MEM_P (new_rtx) > && !MEM_VOLATILE_P (new_rtx)) > return PROFITABLE; > and didn't mark any other MEM_P (new_rtx) or rtxes which contain > a MEM in its subrtxes as PROFITABLE. Now, since r14-8319 profitable_p > method has been renamed to likely_profitable_p and has just a minor role. > Now, rule 4) above is something that isn't about profitability, but about > correct behavior, if you propagate mem/v, the code is miscompiled. > This particular case has been fixed elsewhere by Haochen in r14-9379. > But I think even the 1) and 2) and maybe 3) are a strong don't do it, > don't rely solely on rtx costs, increasing the number of loads of the same > memory, even when cached, is undesirable, canceling load hoisting can > be undesirable as well. > > So, the following patch restores previous behavior of src contains any MEMs, > in that case likely_profitable_p () is taken as the old profitable_p () > as a requirement rather than just a hint. For propagation of something > which doesn't load from memory this keeps the r14-8319 behavior. > > Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk? Ok Richard > 2024-03-09 Jakub Jelinek <jakub@redhat.com> > > PR target/114284 > * fwprop.cc (try_fwprop_subst_pattern): Don't propagate > src containing MEMs unless prop.likely_profitable_p (). > > --- gcc/fwprop.cc.jj 2024-03-08 09:07:29.371626376 +0100 > +++ gcc/fwprop.cc 2024-03-08 16:18:16.125853619 +0100 > @@ -451,6 +451,7 @@ try_fwprop_subst_pattern (obstack_waterm > > if (!prop.likely_profitable_p () > && (prop.changed_mem_p () > + || contains_mem_rtx_p (src) > || use_insn->is_asm () > || !single_set (use_rtl))) > { > > Jakub >
--- gcc/fwprop.cc.jj 2024-03-08 09:07:29.371626376 +0100 +++ gcc/fwprop.cc 2024-03-08 16:18:16.125853619 +0100 @@ -451,6 +451,7 @@ try_fwprop_subst_pattern (obstack_waterm if (!prop.likely_profitable_p () && (prop.changed_mem_p () + || contains_mem_rtx_p (src) || use_insn->is_asm () || !single_set (use_rtl))) {
Hi! Before the recent PR111267 r14-8319 fwprop changes, fwprop would never try to propagate what was not considered PROFITABLE, where the profitable part actually was partly about profitability, partly about very good reasons not to actually propagate and partly for cases where propagation is completely incorrect. In particular, classify_result has: /* Allow (subreg (mem)) -> (mem) simplifications with the following exceptions: 1) Propagating (mem)s into multiple uses is not profitable. 2) Propagating (mem)s across EBBs may not be profitable if the source EBB runs less frequently. 3) Propagating (mem)s into paradoxical (subreg)s is not profitable. 4) Creating new (mem/v)s is not correct, since DCE will not remove the old ones. */ if (single_use_p && single_ebb_p && SUBREG_P (old_rtx) && !paradoxical_subreg_p (old_rtx) && MEM_P (new_rtx) && !MEM_VOLATILE_P (new_rtx)) return PROFITABLE; and didn't mark any other MEM_P (new_rtx) or rtxes which contain a MEM in its subrtxes as PROFITABLE. Now, since r14-8319 profitable_p method has been renamed to likely_profitable_p and has just a minor role. Now, rule 4) above is something that isn't about profitability, but about correct behavior, if you propagate mem/v, the code is miscompiled. This particular case has been fixed elsewhere by Haochen in r14-9379. But I think even the 1) and 2) and maybe 3) are a strong don't do it, don't rely solely on rtx costs, increasing the number of loads of the same memory, even when cached, is undesirable, canceling load hoisting can be undesirable as well. So, the following patch restores previous behavior of src contains any MEMs, in that case likely_profitable_p () is taken as the old profitable_p () as a requirement rather than just a hint. For propagation of something which doesn't load from memory this keeps the r14-8319 behavior. Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk? 2024-03-09 Jakub Jelinek <jakub@redhat.com> PR target/114284 * fwprop.cc (try_fwprop_subst_pattern): Don't propagate src containing MEMs unless prop.likely_profitable_p (). Jakub