https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93088
Jakub Jelinek <jakub at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |jakub at gcc dot gnu.org
--- Comment #1 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
The problem is that pend is uninitialized in the testcase and we end up with:
(insn 6 70 71 14 (set (reg/v:DI 90 [ pend ])
(reg/v:DI 89 [ p ])) "pr56348.c":29:5 66 {*movdi_internal}
(nil))
and
(insn 9 88 63 12 (set (reg/v:DI 89 [ p ])
(reg/v:DI 90 [ pend ])) "pr56348.c":28:3 66 {*movdi_internal}
(nil))
as the only setters of the two pseudos (reg:DI 89 and reg:DI 90) and on that
kind of thing find_single_def_src just loops forever, alternating between the
two.
Hangs since r277923, but I think has been latent since r146700 aka
https://gcc.gnu.org/ml/gcc-patches/2009-04/msg00246.html
Guess a fix could be either to add some counter and punt after following e.g.
128 single definition pseudos with REG_P SET_SRCs, or add a hash_set or bitmap
or whatever to record pseudos we've seen already and punt when we notice a
loop, or combine both ways, start with handling a few without a bitmap/hash_set
and only after certain amount keep track on what we've seen and punt when we
see something again.
Though, even with a hash_set/bitmap, having unbounded search might result in
bad compilation complexity, if we have many loops with similar pseudos that
require extremely deep searches.
