On Thu, 14 Feb 2013, Jakub Jelinek wrote:
> On Wed, Feb 13, 2013 at 01:04:04PM +0100, Richard Biener wrote:
> > 2013-02-13 Richard Biener <rguent...@suse.de>
> >
> > PR lto/50494
> > * varasm.c (output_constant_def_1): Get the decl representing
> > the constant as argument.
> > (output_constant_def): Wrap output_constant_def_1.
> > (make_decl_rtl): Use output_constant_def_1 with the decl
> > representing the constant.
> > (build_constant_desc): Optionally re-use a decl already
> > representing the constant.
> > (tree_output_constant_def): Adjust.
>
> Looks good to me, except formatting nit:
>
> > + /* Like output_constant_def but create a new decl representing the
> > + constant if necessary. */
> > +
> > + rtx
> > + output_constant_def (tree exp, int defer)
> > + {
> > + return output_constant_def_1 (exp, NULL_TREE, defer);
>
> Twice too many spaces.
>
> But I'd say we should also add some varasm.c function that
> increase_alignment should call for
> (TREE_CODE (decl) == VAR_DECL && DECL_IN_CONSTANT_POOL (decl))
> vars upon increasing alignment of the decl, which would do something like:
> void
> adjust_const_pool_alignment (tree decl)
> {
> struct constant_descriptor_tree *desc, key;
>
> gcc_assert (TREE_CODE (decl) == VAR_DECL && DECL_IN_CONSTANT_POOL (decl));
> key.value = DECL_INITIAL (decl);
> key.hash = const_hash_1 (DECL_INITIAL (decl));
> desc = (struct constant_descriptor_tree *)
> htab_find_with_hash (const_desc_htab, &key, key.hash);
> if (desc && MEM_P (desc) && MEM_ALIGN (desc->rtl) < DECL_ALIGN (decl))
> set_mem_align (desc->rtl, DECL_ALIGN (decl));
> }
> (completely untested). Because, if we force the constant to be aligned
> more than it would be by default, the RTL optimizers should be told about
> that too.
At the moment the vectorizer refuses to bump alignment for anything
with TREE_ASM_WRITTEN set. I suppose that should rather be checking
DECL_RTL_SET_P. Btw, isn't the constants descriptor ->rtl always
that of DECL_RTL of the decl?
But yes, we should have a proper abstraction for can_adjust_alignment
and adjust_alignment. But we should use that for all decls, not
just DECL_IN_CONSTANT_POOL.
Thanks,
Richard.
