Re: [PR51752] publication safety violations in loop invariant motion pass

[Aldy Hernandez <aldyh at redhat dot com>]

On 02/28/12 11:05, Richard Henderson wrote:
> On 02/27/12 08:22, Aldy Hernandez wrote:
> > > transform by making transaction load/store stmts behave the same as
> > > potentially trapping stmts (thus, only optimize if the memory is accessed
> > > unconditional somewhere else).  That would work for PRE as well.
> > > [easiest would be to make *_could_trap_p return true for memory ops
> > > inside a transaction]
> >
> > Provided the gimple bit works, this seems reasonable, though quite a big hammer.  But given that we are nearing a release, I would be in favor of it.
> >
> > Richard Henderson, what do you think?
>
> Well, hooking could_trap_p sounds like an easy solution.
>
> Gimple bits, on the other hand, are not.  Keeping those up-to-date is
> always a real pain.  We have had several gimple bits in the history of
> the TM code, and we've gotten rid of them all because they were too
> invasive to maintain.
>
> OTOH, I have no better suggestion...

Well, my solution wasn't pretty either.

However, Mr. Guenther's solution involves recalculating the bits when 
needed, so at least they won't be out of date.

Richards, what do y'all think of this approach?  It is a big hammer as 
discussed, but keeps most of us happy for 4.7.

Tested on x86-64 Linux.

Aldy

Attachment: curr
Description: Text document