On 10/06/2016 04:11 AM, Richard Biener wrote:
The remove_forwarder_block hunk is redundant (I've fixed a similar bug there
by checking in tree_forwarder_block_p). This function doesn't operate from
a worklist and thus shouldn't have the issue.
Agreed. I hadn't really looked closely at remove_forwarder_block, but
it has the same check for newly exposed infinite loops so my patch added
the check for new loop headers in both places. But I agree it isn't
necessary here when I look at how remove_forwarder_block is used.
The remove_forwarder_block_with_phi hunk is ok.
I'll extract/install just the remove_forwarder_block_with_phi hunk & the
testcase.
Thanks,
jeff
