On Fri, Jun 09, 2017 at 11:29:48AM +0200, Richard Biener wrote:
> > force_gimple_operand_gsi called by update_range_test can using match.pd
> > simplifications sometimes return INTEGER_CST (especially when cunroll
> > unrolled code isn't really optimized by forwprop/ccp and similar passes
> > before reassoc2), but that is something not acceptable to the rest of
> > the optimize_range* code, because it needs to know not just the value,
> > but also some gimple_stmt_iterator to insert related code etc.
> >
> > This patch makes sure we have a SSA_NAME even in that case.
> > Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
>
> Ok.
>
> In the end this results in a missed optimization downstream but I
> guess we can't resolve this w/o major refactoring?
The constants should be propagated into uses soon afterwards, there
is both forwprop and ccp shortly after the first reassoc pass and
e.g. dom/vrp2 after second reassoc pass that should fix it up.
Jakub
