On Tue, Nov 26, 2013 at 09:25:40PM +0100, Jakub Jelinek wrote: > Hi! > > The problem on this testcase is that we have > (debug_insn 30 29 31 7 (var_location:HI D#1 (subreg:HI (reg/v:SI 93 [ p ]) > 0)) pr59166.c:20 -1 > (nil)) > and split_live_ranges_for_shrink_wrap decides to replace > SImode pseudo 93 with some other SImode pseudo. But it > uses DF_REF_LOC, which is address of the HImode subreg around the pseudo, > validate_change succeeds on it, because there is no validation inside of > debug_insns and then we crash during var-tracking because of a mode > mismatch.
thanks a lot for fixing it, I had no idea that DF_REF_REAL_LOC even existed. This change also apparently makes it possible to throw away the SUBREG check in the same function. I'll enqueue a patch removing it for the next stage1. > > The following patch uses DF_REF_REAL_LOC instead, so that it looks through > the subreg and adjusts what it should. Code inspection showed the same > issue in find_moveable_pseudos, don't have a testcase for that though. That is where I copied the substitution from. Martin > > Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk? > > 2013-11-26 Jakub Jelinek <ja...@redhat.com> > > PR rtl-optimization/59166 > * ira.c (find_moveable_pseudos): Use DF_REF_REAL_LOC instead of > DF_REF_LOC in validate_change call. > (split_live_ranges_for_shrink_wrap): Likewise. > > * gcc.dg/torture/pr59166.c: New test.
