Hi,
On Mon, Dec 19, 2016 at 11:09:52AM +0100, Martin Liska wrote:
> Hello.
>
> Building mariadb with -flto exposes a bug which I also used to see
> in Firefox. It's caused by IPA CP starting from r236418, where the
> pass started to propagate const VAR_DECLs. Problem is that the pass
> does not update call graph by adding IPA_REF_READ of the propagated
> variable.
>
> Patch can bootstrap on ppc64le-redhat-linux and survives regression tests.
>
> Ready to be installed?
Honza needs to have a look at this but since I have suggested this
approach, I am of course fine with it, except that...
> Martin
> From 477e81fde08d0520ce552ec8baa0349590dc683c Mon Sep 17 00:00:00 2001
> From: marxin <mli...@suse.cz>
> Date: Mon, 19 Dec 2016 11:03:34 +0100
> Subject: [PATCH] Fix IPA CP where it forgot to add a reference in cgraph
>
> gcc/ChangeLog:
>
> 2016-12-19 Martin Liska <mli...@suse.cz>
>
> * cgraphclones.c (cgraph_node::create_virtual_clone):
> Create either IPA_REF_LOAD of IPA_REF_READ depending on
> whether new_tree is a VAR_DECL or an ADDR_EXPR.
> * ipa-cp.c (create_specialized_node): Add reference just for
> ADDR_EXPRs.
> * symtab.c (symtab_node::maybe_create_reference): Remove guard
> as it's guarded in callers.
> ---
> gcc/cgraphclones.c | 6 +++++-
> gcc/ipa-cp.c | 3 ++-
> gcc/symtab.c | 2 --
> 3 files changed, 7 insertions(+), 4 deletions(-)
>
> diff --git a/gcc/cgraphclones.c b/gcc/cgraphclones.c
> index 349892dab67..93c86e6a1cc 100644
> --- a/gcc/cgraphclones.c
> +++ b/gcc/cgraphclones.c
> @@ -624,7 +624,11 @@ cgraph_node::create_virtual_clone (vec<cgraph_edge *>
> redirect_callers,
> || in_lto_p)
> new_node->unique_name = true;
> FOR_EACH_VEC_SAFE_ELT (tree_map, i, map)
> - new_node->maybe_create_reference (map->new_tree, IPA_REF_ADDR, NULL);
> + {
> + ipa_ref_use use_type
> + = TREE_CODE (map->new_tree) == VAR_DECL ? IPA_REF_ADDR : IPA_REF_LOAD;
...this test should be for ADDR_EXPR here. Or you could switch the
IPA_REF_* constants the other way round which I bet is going to have
the same effect in practice, but personally, I'd test for ADDR_EXPR.
Thanks,
Martin
