Re: Drop Z from X + Z < Y + Z

[Marc Glisse <marc dot glisse at inria dot fr>]

On Fri, 28 Apr 2017, Richard Biener wrote:

> On Fri, Apr 28, 2017 at 1:35 PM, Marc Glisse <marc.glisse@inria.fr> wrote:
> > Hello,
> >
> > surprisingly, this did not cause any Wstrict-overflow failure. Some of it
> > sounds more like reassoc's job, but it is convenient to handle simple cases
> > in match.pd. I think we could wait until there are reports of regressions
> > related to register pressure before adding single_use tests.
> >
> > For a std::vector<long> v, we now simplify v.size()==v.capacity() to a
> > single comparison (long)finish==(long)end_storage (I guess I could still try
> > to drop the casts to consider it really done). Handling
> > v.size()<v.capacity() seems much harder, because of C++'s questionable
> > choice to use unsigned types. I may still be able to remove the divisions,
> > I'll see if I can sprinkle some 'convert' in recent transformations.
> >
> > Bootstrap+regtest on powerpc64le-unknown-linux-gnu.
>
> +(for cmp (eq ne minus)
>
> Fat fingered 'minus' (in all places) or did you want to get fancy?
> (the transforms
> look valid even for cmp == minus)  Maybe adjust comments to reflect this.

I started with just comparisons and then noticed that minus worked as 
well. I'll adjust the comment and rename cmp to op as suggested by Jakub. 
I may have to separate the minus case when I add converts in a future 
patch.

> There are a few related cases in fold-const.c, namely X +- Y CMP X -> Y CMP 0,
> some of them also handling POINTER_PLUS_EXPR.  So I wonder if you can
> handle pointer_plus like plus and maybe move those fold-const.c patterns.

You already added (T)(P + A) - (T)(P + B) -> (T)A - (T)B, so I'll have to 
take care to avoid redundancy. Since it was meant for pointer_plus more 
than plus, you used convert, not convert?, and it is currently not 
redundant with my patch. If I integrate pointer_plus in the same 
transformation, I will probably have to use :C instead of :c, and we will 
get 2 dead paths (operand_equal_p on the first argument of one 
pointer_plus and the second argument of another pointer_plus), it might be 
easier as a separate pattern.

--
Marc Glisse