https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69355
Jakub Jelinek <jakub at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |dj at gcc dot gnu.org,
| |law at gcc dot gnu.org
--- Comment #7 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
-fno-lifetime-dse -fno-devirtualize -fno-devirtualize-speculatively doesn't
help either. The number of abstraction levels is very big, so it is hard to
spot where exactly things go wrong.
But, I've tracked the problem to the tree-dfa.c change of that commit only,
reverting:
https://gcc.gnu.org/viewcvs/gcc/trunk/gcc/tree-dfa.c?limit_changes=0&r1=214748&r2=214747&pathrev=214748
fixes it. This makes a change on XprLiteral structure, which contains a single
long double member, therefore the structure has XFmode, which has precision of
80, but bitsize 128 (on x86_64, 96 on i?86).
Using the whole size of the access rather than only the precision looks like
the right thing to me though, I mean various targets don't lay the bits without
any gaps.
So, IMNSHO either that change should be reverted, or limited to partial int
modes?
