Re: [PATCH] Fix PR 18965, a c++ and gimplifier fix

[Andrew Pinski <pinskia at physics dot uc dot edu>]

On Dec 15, 2004, at 2:41 PM, Roger Sayle wrote:

> On Wed, 15 Dec 2004, Jason Merrill wrote:
> > On Wed, 15 Dec 2004 06:06:21 -0700 (MST), Roger Sayle 
> > <roger@eyesopen.com> wrote:
> > > And the gimplifier.c change is OK, if you also fix the spelling error
> > > in the comment you move, "indicies" -> "indices".
> >
> > I don't like the gimplifier change.  How hard is it really to fully
> > support RANGE_EXPRs?
>
> I suspect that adding full support for RANGE_EXPRs is unsuitable for
> current current stage3.  Andrew's regression fix avoids the middle-end
> ICEing in the common case that can be handled without much effort, and
> in addition to C++ should help GCC's other front-ends (including gpc 
> and
> PL/1) that may not possess the same level of understanding of the 4.0's
> gimplification process.

It was not hard and is very simple as I made RANGE_EXPR make a loop
so we get run into a compile time issue and a memory problem which
we are trying to fix with RANGE_EXPR. I also hear this is needed to
fix PR 18191 also which I implemented it and not worry about not being
suitable for stage 3/regression only.


> > It's not clear to me that this special case code buys us anything.
>
> Handling the upper bound equal to the lower bound buys us exactly the
> same regression fix as the "workaround" in the C++ front-end, but has
> the added benefit of fixing this issue across front-ends.

I kept the special case so we get this "optimization" at -O0 also.


OK? Bootstrapped and tested on powerpc-darwin with no regressions.

Thanks,
Andrew Pinski

ChangeLog:
	* gimplify.c (gimplify_init_ctor_eval): Handle fully RANGE_EXPR
	in a CONSTRUCTOR by creating a loop.

Attachment: fixRANGE.diff.txt
Description: Text document