On 30/05/16 11:46, Richard Biener wrote:
This patch fixes the assert conservatively by aborting graphite code
>generation when encountering a phi with more than two arguments in
>copy_bb_and_scalar_dependences.
>
>Bootstrapped and reg-tested on x86_64.
>
>OK for trunk, 6 branch?
Did you check if simply returning false from bb_contains_loop_phi_nodes
instead of asserting works? The caller has a 'else' that is supposed
to handle condition PHIs. After all it already handles one predecessor
specially ... Thus
if (EDGE_COUNT (bb->preds) != 2)
return false;
should work here.
Unfortunately, that doesn't work. We run into another assert in
copy_cond_phi_nodes:
...
/* Cond phi nodes should have exactly two arguments. */
gcc_assert (2 == EDGE_COUNT (bb->preds));
...
Or replace this function with bb_loop_header_p (bb)
(works w/o loop info, the function seems to depend on loop info and
thus simply checking bb->loop_father->header == bb should work as well).
I think that will run into the same assert.
Thanks,
- Tom
