This is the mail archive of the
mailing list for the GCC project.
Re: [PATCH][RFC] Fix PR56113 more
On Fri, 1 Feb 2013, Jakub Jelinek wrote:
> On Fri, Feb 01, 2013 at 10:00:00AM +0100, Richard Biener wrote:
> > This reduces compile-time of the testcase in PR56113 (with n = 40000)
> > from 575s to 353s. It does so by reducing the quadratic algorithm
> > to impose an order on visiting dominator sons during a domwalk.
> > Steven raises the issue that there exist domwalk users that modify
> > the CFG during the walk and thus the new scheme does not work
> > (at least optimally, as the current quadratic scheme does). As
> > we are using a fixed-size sbitmap to track visited blocks existing
> > domwalks cannot add new blocks to the CFG so the worst thing that
> > can happen is that the order of dominator sons is no longer
> > optimal (I suppose with the "right" CFG manipulations even the
> > domwalk itself does not work - so I'd be hesitant to try to support
> > such domwalk users) - back to the state before any ordering
> > was imposed on the dom children visits (see rev 159100).
> I think it would be desirable to first analyze the failures Steven saw, if
> any. As you said, asan doesn't use domwalk, so it is a mystery to me.
Yeah. Now, fortunately domwalk.h is only directly included and thus
the set of optimizers using it are
tree-ssa-math-opts.c: If we did this using domwalk.c, an efficient
implementation would have
tree-ssa-pre.c:/* Local state for the eliminate domwalk. */
tree-ssa-pre.c: eliminate domwalk. */
tree-ssa-pre.c:/* At the current point of the eliminate domwalk make OP
tree-ssa-pre.c:/* Perform elimination for the basic-block B during the
I don't see any target specific ones that do not have coverage
with x86_64 multilib testing (maybe compare-elim.c? though that
doesn't really require a domwalk as it is only using the
before_dom_children hook). That said, arbitrary CFG manipulations
during domwalk certainly will not preserve "domwalk" properties
of a domwalk.
Steven - can you reproduce your failures (and on which target?)