On Mon, Jan 23, 2012 at 6:33 AM, Andrew Pinski <andrew.pin...@caviumnetworks.com> wrote: > Hi, > The problem here is not really jump threading getting in the way of > PHI-OPT rather there is no cleanup of the IR after jump thread but > before phi-opt. > This patch adds a pass_phi_only_cprop right after the first vrp and > pass_merge_phi right before the first and last phiopts. > > Since VRP does jump threading just like DOM does, we need a > pass_phi_only_cprop right after it. And the merge PHI pass should be > done so that going into PHI-OPT we have a CFG which there is only one > PHI for a case like: > bb0: > if (a) > goto L1; > bb3: > if (b) goto L2; > L1: > x = PHI <b(3), c(0)> > L2: > x= PHI<x(L1), d(3)> > > PHI-OPT does not handle PHIs like this that well. > > OK? Bootstrapped and tested on x86_64-linux-gnu with no regressions.
The testcase already works for me on trunk, without your patch. You are adding three passes - one of it phi_only_cprop which I'd rather remove ... I suspect VRP simply exposes copy propagation opportunities, and that's expected. mergephi before phiopt makes sense, but then we should move the first mergephi instead of inserting another one. Richard. > Thanks, > Andrew Pinski > > ChangeLog: > * passes.c (init_optimization_passes): Add pass_phi_only_cprop after > the first vpr pass. > Add pass_merge_phi before the first and last phiopt passes. > > testsuite/ChangeLog: > * gcc.dg/tree-ssa/phi-opt-7.c: New testcase.
