Hello,
As described in the PR, the ICE reason was the typo made when introducing
calls to add_hard_reg_set. Fixed by the first attached patch, bootstrapped
and tested on both ia64 and x86_64, committed as obvious.
The test case is very sensitive to the scheduler decisions (e.g. it didn't
fail on trunk but only on the revision reported for me), so instead of
adding the test I have put in the code two asserts checking that we can
always schedule the fence instruction as is. This hunk was tested together
with the first but committed separately.
The first patch can be safely committed to 4.8, the second can stay on
trunk only. Jakub, will it be fine with you?