On Tue, Aug 2, 2011 at 6:14 AM, Kai Tietz <ktiet...@googlemail.com> wrote: > 2011/8/2 Richard Guenther <richard.guent...@gmail.com>: >> On Tue, Aug 2, 2011 at 12:17 PM, Kai Tietz <ktiet...@googlemail.com> wrote: >>> Hello, >>> >>> this patch removes in forward-propagation useless comparisons X != 0 >>> and X != ~0 for boolean-typed X. For one-bit precision typed X we >>> simplifiy X == 0 (and X != ~0) to ~X, and for X != 0 (and X == ~0) to >>> X. >>> For none one-bit precisione typed X, we simplify here X == 0 -> X ^ 1, >>> and for X != 0 -> X. We can do this as even for Ada - which has only >>> boolean-type with none-one-bit precision - the truth-value is one. >> >> This isn't a simplification but a canonicalization and thus should be >> done by fold_stmt instead (we are not propagating anything after all). >> In fact, fold_stmt should do parts of this already by means of its >> canonicalizations via fold. > > Well, it simplifies and canonicalizes. But to put this into > gimple-fold looks better. > >>> Additionally this patch changes for function >>> forward_propagate_comparison the meaning of true-result. As this >>> result wasn't used and it is benefitial to use this propagation also >> >> which is a bug - for a true return value we need to set cfg_changed to true. > > I addressed this in my updated patch (see below) > >>> in second loop in function ssa_forward_propagate_and_combine, it >>> returns true iff statement was altered. Additionally this function >>> handles now the boolean-typed simplifications. >> >> why call it twice? How should that be "beneficial"? I think that >> forward_propagate_into_comparison should instead fold the changed >> statement. > > Well, due missing fold_stmt call, there were still none-converted > comparisons. I've added here the call to fold_stmt_inplace, and it > solved the issue. > >>> For the hunk in gimple.c for function canonicalize_cond_expr_cond: >>> This change seems to show no real effect, but IMHO it makes sense to >>> add here the check for cast from boolean-type to be consitant. >> >> Probably yes. >> >> Thanks, >> Richard. > > > 2011-08-02 Kai Tietz <kti...@redhat.com> > > * gimple.c (canonicalize_cond_expr_cond): Handle cast from boolean-type. > (ssa_forward_propagate_and_combine): Interprete result of > forward_propagate_comparison. > * gcc/gimple-fold.c (fold_gimple_assign): Add canonicalization for > boolean-typed operands for comparisons. > > 2011-08-02 Kai Tietz <kti...@redhat.com> > > * gcc.dg/tree-ssa/forwprop-15.c: New testcase. > > Regression tested and bootstrapped for all languages (including Ada > and Obj-C++). Ok for apply? >
It caused: http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49947 -- H.J.
