Re: Move remaining flag_unsafe_math_optimizations using simplify and match

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

On Thu, 27 Aug 2015, Andreas Schwab wrote:

> "Hurugalawadi, Naveen" <Naveen.Hurugalawadi@caviumnetworks.com> writes:
>
> > 	* fold-const.c (fold_binary_loc) : Move Optimize
> > 	root(x)*root(y)	as root(x*y) to match.pd.
> > 	Move Optimize expN(x)*expN(y) as expN(x+y) to match.pd.
> > 	Move Optimize pow(x,y)*pow(x,z) as pow(x,y+z) to match.pd.
> > 	Move Optimize a/root(b/c) into a*root(c/b) to match.pd.
> > 	Move Optimize x/expN(y) into x*expN(-y) to match.pd.
> >
> > 	* match.pd (mult (root:s @0) (root:s @1)): New simplifier.
> > 	(mult (POW:s @0 @1) (POW:s @0 @2)) : New simplifier.
> > 	(mult (exps:s @0) (exps:s @1)) : New simplifier.
> > 	(rdiv @0 (root:s (rdiv:s @1 @2))) : New simplifier.
> > 	(rdiv @0 (exps:s @1)) : New simplifier.
>
> FAIL: gcc.dg/builtins-11.c (test for excess errors)
> Excess errors:
> builtins-11.c:(.text+0x52): undefined reference to `link_error'

Indeed, generic-match ends up testing for sqrt(x)*sqrt(y) before testing 
for sqrt(x)@1*@1, so it simplifies it to sqrt(x*x)->abs(x) instead of 
plain x. Changing the genmatch machinery to better respect the order of 
patterns in match.pd might not be trivial without sacrificing performance, 
a workaround might be to add a special pattern sqrt(x)*sqrt(x), or to 
disable some patterns for GENERIC so CSE has a chance to run first.

--
Marc Glisse