https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110924
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |hubicka at gcc dot gnu.org
--- Comment #9 from Richard Biener <rguenth at gcc dot gnu.org> ---
Hm, OK. So I actually thought of this issue. It's basically a hole in the
requirement of all function "exits" having a virtual use and thus virtual
PHIs being created along the paths to it. I was aware of GIMPLE_RESX and
thought about const noreturn (but that attribute combination is rejected).
IPA pure-const detects func_15 as looping 'const' and noreturn:
Function is locally const.
Function found to be noreturn: func_15
Function found to be looping const: func_15/2
Declaration updated to be looping const: func_15/2
so there's two ways to "fix" this, one - drop the assumption we have
PHIs on the paths to function exit (including in not returning regions).
That makes
virtual_operand_live::get_live_in (basic_block bb)
{
...
/* Since we don't have a virtual PHI we can now pick any of the
incoming edges liveout value. All returns from the function have
a virtual use forcing generation of virtual PHIs. */
edge_iterator ei;
edge e;
FOR_EACH_EDGE (e, ei, bb->preds)
if (liveout[e->src->index])
{
if (EDGE_PRED (bb, 0) != e)
liveout[EDGE_PRED (bb, 0)->src->index] = liveout[e->src->index];
return liveout[e->src->index];
}
instead required to check each edge and we can't simply take the value
from the immediate dominator as fallback. When we discover divergence
we need to return NULL (unknown - nobody created the actually live VOP).
The other alternative is to make sure we _do_ have the virtual operand.
Either by making sure the "invalid" combination of const + noreturn
isn't detected or by creating a virtual use in that case anyway
(and fixup code because of that inconsistency).
