On Wed, Jul 6, 2011 at 6:26 AM, Richard Guenther <rguent...@suse.de> wrote: > > This fixes PR49645 - with MEM_REF the value-numbering machinery > to look through aggregate copies wasn't working reliably as > we have two representations for X, X and MEM[&X]. The following > patch fixes that by internally always using the more complicated > representation. > > The patch needs consistent DECL_HARD_REGISTER settings to avoid > generating MEM_REFs for them though and the C frontend fails > to set that flag for global variables - hence the c-decl.c part > (otherwise compile.exp 20041119-1.c ICEs). > > Bootstrapped and tested on x86_64-unknown-linux-gnu, are the > C frontend parts ok for trunk? > > Thanks, > Richard. > > 2011-07-06 Richard Guenther <rguent...@suse.de> > > PR tree-optimization/49645 > * c-decl.c (finish_decl): Also set DECL_HARD_REGISTER for global > register variables. > * tree-ssa-sccvn.c (vn_reference_op_eq): Disregard differences > in type qualification here ... > (copy_reference_ops_from_ref): ... not here. > (vn_reference_lookup_3): ... or here. > (copy_reference_ops_from_ref): Record decl bases as MEM[&decl]. > (vn_reference_lookup): Do the lookup with a valueized ao-ref. > > * g++.dg/tree-ssa/pr8781.C: Disable SRA. >
This caused: http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49662 -- H.J.
