Re: [patch 1/2] PR27313 Discover non-trapping memory accesses

[Richard Henderson <rth at redhat dot com>]

On Mon, Apr 16, 2007 at 06:51:25PM +0200, Michael Matz wrote:
> The plan in any case was, to only note stores as shadowing other 
> accesses.

Ah.  That would indeed simplify the problem.

> I'm not sure if we ever are allowed to reorder stores,
> generally we are not.


One transformation to think of there is store promotion.

	*p = x
	y = *p
	*p = z	[notrap]
=>
	t = x;
	y = t;
	*p = z;

If we're not careful, we might leave the notrap flag set 
in the store of z.

> I also have a hard time to imagine a usefull transformation where a load 
> is moved in front of a dominating load, but that dominating load itself is 
> not moved.

One thing we'd really want to do with this transformation is
to promote *any* reference though a pointer as non-trapping.  So

	x = p->x	[trap]
	y = p->y	[notrap]

This isn't what your current code does, so your current code doesn't
have to consider it, but it would be extremely useful for Java.  With
such a transformation, it's easy to imagine Y being moved around the
dominator tree independently of X.


r~