Re: PR79286, ira combine_and_move_insns in loops

[Alan Modra <amodra@gmail.com>]

On Thu, Feb 02, 2017 at 12:18:31AM +1030, Alan Modra wrote:
> This patch cures PR79286 by restoring the REG_DEAD note test used
> prior to r235660, but modified to only exclude insns that may trap.
> I'd like to allow combine/move without a REG_DEAD note in loops
> because insns in loops often lack such notes, and I recall seeing
> quite a few cases at the time I wrote r235660 where loops benefited
> from allowing the combine/move to happen.

Ugh, the new testcase fails for x86 -m32 -Os, but not due to ira this
time but rather reload.  I haven't looked into what is going wrong in
reload yet, but the net result is the same:  The faulting mem read is
moved before the printf call.

There were no other testsuite regressions, apart from the random set
of fails I have been getting for a long time on x86_64 for
c-c++-common/ubsan/float-cast-overflow-10.c,
c-c++-common/ubsan/float-cast-overflow-2.c,
c-c++-common/ubsan/float-cast-overflow-8.c, and
c-c++-common/ubsan/overflow-mul-4.c.

What is the correct thing to do for a new testcase that fails like
this?  Add a dg-fail-if?  Assuming I or someone else can't fix the
reload fail.

The new testcase -Os failure occurs on gcc-4.x, gcc-5 and gcc-6, but
gcc-3.4 passes.

-- 
Alan Modra
Australia Development Lab, IBM