On Fri, 15 Jan 2016, Jan Hubicka wrote:
> Hi,
> this is last of my patches to improve partitionability of programs.
> This patch compacts BLOCK_VARs lists and trows away TYPE_DECLs for -g0
> and redundant TYPE_DECLs on all levels. It does make noticeable difference
> on firefox, but I managed to erase the numbers (can re-test them if requested)
>
> Bootstrapped/regtested x86_64-linux, OK?
Hmm. So I wonder why remove_unused_locals does not do this then
or if it does not because it would change code-gen based on -g/-g0
I wonder why your patch wouldn't introduce that issue with LTO.
That said - improve remove_unused_locals instead please?
Thanks,
Richard.
> * tree.c (needed_in_block_vars_p): New function.
> (free_lang_data_in_decl): Use it.
> Index: tree.c
> ===================================================================
> --- tree.c (revision 232407)
> +++ tree.c (working copy)
> @@ -5343,6 +5343,42 @@ need_assembler_name_p (tree decl)
> return true;
> }
>
> +/* Return true if DECL should stay in BLOCK_VARs list. */
> +
> +static inline bool
> +needed_in_block_vars_p (const_tree decl)
> +{
> + if (TREE_CODE (decl) == VAR_DECL && TREE_USED (decl)
> + && (!TREE_STATIC (decl) || DECL_EXTERNAL (decl)))
> + return true;
> + if (TREE_CODE (decl) == RESULT_DECL)
> + return true;
> + /* Dwarf2out output non-fortran and non-ada const decls when the
> + enum is output and it will not consider language to be ada or fortran
> + in LTO. */
> + if (TREE_CODE (decl) == CONST_DECL)
> + return false;
> + if (debug_info_level >= DINFO_LEVEL_TERSE)
> + return false;
> + /* Strip builtins from the translation-unit BLOCK. We still have targets
> + without builtin_decl_explicit support and also builtins are shared
> + nodes and thus we can't use TREE_CHAIN in multiple lists. */
> + if (TREE_CODE (decl) == FUNCTION_DECL
> + && DECL_BUILT_IN (decl))
> + return false;
> + if (DECL_IGNORED_P (decl))
> + return false;
> + if (TREE_CODE (decl) == TYPE_DECL
> + && is_redundant_typedef (decl))
> + return false;
> + gcc_checking_assert (TREE_CODE (decl) == TYPE_DECL
> + || TREE_CODE (decl) == VAR_DECL
> + || TREE_CODE (decl) == FUNCTION_DECL
> + || TREE_CODE (decl) == LABEL_DECL
> + || TREE_CODE (decl) == RESULT_DECL);
> + return true;
> +}
> +
>
> /* Reset all language specific information still present in symbol
> DECL. */
> @@ -5445,15 +5481,11 @@ free_lang_data_in_decl (tree decl)
> && DECL_INITIAL (decl)
> && TREE_CODE (DECL_INITIAL (decl)) == BLOCK)
> {
> - /* Strip builtins from the translation-unit BLOCK. We still have
> targets
> - without builtin_decl_explicit support and also builtins are shared
> - nodes and thus we can't use TREE_CHAIN in multiple lists. */
> tree *nextp = &BLOCK_VARS (DECL_INITIAL (decl));
> while (*nextp)
> {
> tree var = *nextp;
> - if (TREE_CODE (var) == FUNCTION_DECL
> - && DECL_BUILT_IN (var))
> + if (needed_in_block_vars_p (var))
> *nextp = TREE_CHAIN (var);
> else
> nextp = &TREE_CHAIN (var);
>
>
--
Richard Biener <[email protected]>
SUSE LINUX GmbH, GF: Felix Imendoerffer, Jane Smithard, Graham Norton, HRB
21284 (AG Nuernberg)
