Remove noce_mem_write_may_trap_or_fault_p in ifcvt

[Michael Matz]

Hi,

On Fri, 20 Nov 2015, Bernd Schmidt wrote:

> On 11/19/2015 12:49 AM, Jeff Law wrote:
> > On 11/18/2015 12:16 PM, Bernd Schmidt wrote:
> > > I don't think so, actually. One safe option would be to rip it out and
> > > just stop transforming this case, but let's start by looking at the code
> > > just a bit further down, calling noce_can_store_speculate. This was
> > > added later than the code we're discussing, and it tries to verify that
> > > the same memory location will unconditionally be written to at a point
> > > later than the one we're trying to convert
> > And if we dig into that thread, Ian suggests this isn't terribly
> > important anyway.
> > 
> > However, if we go even further upthread, we find an assertion from
> > Michael that this is critical for 456.hmmer and references BZ 27313.
> 
> Cc'd in case he has additional input.

The transformation I indicated in the mail introducing the cselim pass is 
indeed crucial for hmmer.  But that has nothing to do with 
noce_mem_write_may_trap_or_fault_p.  The latter is purely an RTL 
transformation, the transformation needs to happen on gimple level.  I 
agree with Bernd that the dominator walking code on the RTL side trying to 
find an unconditional write to the same address in order to validate the 
transformation looks wrong.


Ciao,
Michael.