[PATCH][middle-end/88784] Middle end is missing some optimizations about unsigned

[Richard Biener]

On Fri, 28 Jun 2019, Andrew Pinski wrote:

> On Thu, Jun 27, 2019 at 9:55 PM Li Jia He <helijia@linux.ibm.com> wrote:
> >
> >
> >
> > On 2019/6/27 11:48 PM, Jeff Law wrote:
> > > On 6/27/19 12:11 AM, Li Jia He wrote:
> > >> Hi,
> > >>
> > >> According to the optimizable case described by Qi Feng on
> > >> issue 88784, we can combine the cases into the following:
> > >>
> > >> 1. x >  y  &&  x != XXX_MIN  -->   x > y
> > >> 2. x >  y  &&  x == XXX_MIN  -->   false
> > >> 3. x <= y  &&  x == XXX_MIN  -->   x == XXX_MIN
> > >>
> > >> 4. x <  y  &&  x != XXX_MAX  -->   x < y
> > >> 5. x <  y  &&  x == XXX_MAX  -->   false
> > >> 6. x >= y  &&  x == XXX_MAX  -->   x == XXX_MAX
> > >>
> > >> 7. x >  y  ||  x != XXX_MIN  -->   x != XXX_MIN
> > >> 8. x <= y  ||  x != XXX_MIN  -->   true
> > >> 9. x <= y  ||  x == XXX_MIN  -->   x <= y
> > >>
> > >> 10. x <  y  ||  x != XXX_MAX  -->   x != UXXX_MAX
> > >> 11. x >= y  ||  x != XXX_MAX  -->   true
> > >> 12. x >= y  ||  x == XXX_MAX  -->   x >= y
> > >>
> > >> Note: XXX_MIN represents the minimum value of type x.
> > >>        XXX_MAX represents the maximum value of type x.
> > >>
> > >> Here we don't need to care about whether the operation is
> > >> signed or unsigned.  For example, in the below equation:
> > >>
> > >> 'x >  y  &&  x != XXX_MIN  -->   x > y'
> > >>
> > >> If the x type is signed int and XXX_MIN is INT_MIN, we can
> > >> optimize it to 'x > y'.  However, if the type of x is unsigned
> > >> int and XXX_MIN is 0, we can still optimize it to 'x > y'.
> > >>
> > >> The regression testing for the patch was done on GCC mainline on
> > >>
> > >>      powerpc64le-unknown-linux-gnu (Power 9 LE)
> > >>
> > >> with no regressions.  Is it OK for trunk ?
> > >>
> > >> Thanks,
> > >> Lijia He
> > >>
> > >> gcc/ChangeLog
> > >>
> > >> 2019-06-27  Li Jia He  <helijia@linux.ibm.com>
> > >>          Qi Feng  <ffengqi@linux.ibm.com>
> > >>
> > >>      PR middle-end/88784
> > >>      * gimple-fold.c (and_comparisons_contain_equal_operands): New function.
> > >>      (and_comparisons_1): Use and_comparisons_contain_equal_operands.
> > >>      (or_comparisons_contain_equal_operands): New function.
> > >>      (or_comparisons_1): Use or_comparisons_contain_equal_operands.
> > > Would this be better done via match.pd?  ISTM this transformation would
> > > be well suited for that framework.
> >
> > Hi, Jeff
> >
> > I did this because of the following test case:
> > `
> > _Bool comp(unsigned x, unsigned y)
> > {
> >    return x > y && x != 0;
> > }
> > `
> > The gimple file dumped on the power platform is:
> > `
> > comp (unsigned int x, unsigned int y)
> > {
> >    _Bool D.2837;
> >    int iftmp.0;
> >
> >    if (x > y) goto <D.2841>; else goto <D.2839>;
> >    <D.2841>:
> >    if (x != 0) goto <D.2842>; else goto <D.2839>;
> >    <D.2842>:
> >    iftmp.0 = 1;
> >    goto <D.2840>;
> >    <D.2839>:
> >    iftmp.0 = 0;
> >    <D.2840>:
> >    D.2837 = (_Bool) iftmp.0;
> >    return D.2837;
> > }
> > `
> > However, the gimple file dumped on x86 is
> > `
> > comp (unsigned int x, unsigned int y)
> > {
> >    _Bool D.2837;
> >
> >    _1 = x > y;
> >    _2 = x != 0;
> >    _3 = _1 & _2;
> >    _4 = (int) _3;
> >    D.2837 = (_Bool) _4;
> >    return D.2837;
> > }
> > `
> >
> > The reason for the inconsistency between these two behaviors is param
> > logical-op-non-short-circuit.  If we add the pattern to the match.pd
> > file, we can only optimize the situation in which the statement is in
> > the same basic block (logical-op-non-short-circuit=1, x86).  But for
> > a cross-basic block (logical-op-non-short-circuit=0, power), match.pd
> > can't handle this situation.
> >
> > Another reason is that I found out maybe_fold_and_comparisons and
> > maybe_fold_or_comparisons are not only called by ifcombine pass but
> > also by reassoc pass. Using this method can basically unify param
> > logical-op-non-short-circuit=0 or 1.
> 
> 
> As mentioned before ifcombine pass should be using gimple-match
> instead of fold_build.  Try converting ifcombine over to gimple-match
> infrastructure and add these to match.pd.

Yes, I mentioned that in the PR.  The issue is that at the moment
to combine x > y with x <= y you'd have to build GENERIC trees
for both or temporary GIMPLE assign with a SSA def (and then feed
that into the GENERIC or GIMPLE match.pd path).

maybe_fold_and/or_comparisons handle two exploded binary expressions
while the current match.pd entries handle at most one exploded one
(the outermost then, either AND or OR).  But it would be definitely
doable to auto-generate maybe_fold_and/or_comparisons from match.pd
patterns which is what I'd ultimatively suggest to do (in some more
generalized form maybe).  Either with a separate genmatch invocation
or as part of the --gimple processing (not sure what is more feasible
here).

I told Li Jia He that I don't expect him to do this work.

Note I didn't review the actual patch yet.

Thanks,
Richard.