https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111601
--- Comment #24 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
As the pass seems to work on one basic block at the time (everything from
analysis to actual changes), I wonder if e.g. get_uses shouldn't punt if it
sees (non-DEBUG_INSN!) uses in other basic block but the current one (this
would still not fix this wrong-code) or if the definition's luid is higher than
use's luid (i.e. the use is in the same bb, but still from a different
iteration of the loop).
See the
loop:
*last_viable = c;
*cand = c->next;
c->next = (z_candidate *) 0;
last_viable = &c->next;
if (!*cand)
break;
c = *cand;
goto loop;
statements done in the loop body, the *last_viable = c; store is the %r9 store
to %r10
(which is initially address of some automatic variable, but &c->next aka c + 96
for second and following iteration), the c->next = 0 store is the %r5 store to
%r10 (which has been incremented by 96 first.
