This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [PATCH] Remove special streaming of builtins


2016-07-26 22:52 GMT+03:00 Richard Biener <rguenther@suse.de>:
> On July 26, 2016 7:26:46 PM GMT+02:00, "H.J. Lu" <hjl.tools@gmail.com> wrote:
>>On Mon, Jul 25, 2016 at 4:35 AM, Richard Biener <rguenther@suse.de>
>>wrote:
>>>
>>> So I needed to fix that builtins appearing in BLOCK_VARs and the
>>solution
>>> I came up with accidentially disabled streaming via the special path.
>>> Thus the following patch removes the special-casing completely and
>>makes
>>> the BLOCK_VARs handling work the same way as for regular externs (by
>>> streaming a local copy).  We stream each builtin decl once and then
>>> refer to it via the decl index (which is cheaper than the special
>>> casing).
>>>
>>> I'm not 100% this solves for example the -fno-math-errno inlining
>>> across TUs (it certainly doesn't if you use attribute optimize with
>>> -fno-math-errno), but it eventually should by means of having two
>>> different BUILT_IN_XXX if they have different behavior.  At least
>>> if all relevant bits are set on the function _type_ rather than
>>> the decl which I think we still lto-symtab replace with one
>>> entity during WPA(?)
>>>
>>> Well.
>>>
>>> LTO bootstrapped and tested on x86_64-unknown-linux-gnu
>>(c,c++,fortran),
>>> bootstrapped on x86_64-unknown-linux-gnu (all), testing in progress.
>>>
>>> I might have not catched all fndecl compares.
>>>
>>> Will apply to trunk if testing completes.  As said, maybe followup
>>> cleanups possible, at least to lto-opts.c / lto-wrapper.
>>>
>>> Richard.
>>>
>>> 2016-07-25  Richard Biener  <rguenther@suse.de>
>>>
>>>         * cgraph.c (cgraph_node::verify_node): Compare against
>>builtin
>>>         by using DECL_BUILT_IN_CLASS and DECL_FUNCTION_CODE.
>>>         * tree-chkp.c (chkp_gimple_call_builtin_p): Likewise.
>>>         * tree-streamer.h (streamer_handle_as_builtin_p): Remove.
>>>         (streamer_get_builtin_tree): Likewise.
>>>         (streamer_write_builtin): Likewise.
>>>         * lto-streamer.h (LTO_builtin_decl): Remove.
>>>         * lto-streamer-in.c (lto_read_tree_1): Remove assert.
>>>         (lto_input_scc): Remove LTO_builtin_decl handling.
>>>         (lto_input_tree_1): Liekwise.
>>>         * lto-streamer-out.c (lto_output_tree_1): Remove special
>>>         handling of builtins.
>>>         (DFS::DFS): Likewise.
>>>         * tree-streamer-in.c (streamer_get_builtin_tree): Remove.
>>>         * tree-streamer-out.c (pack_ts_function_decl_value_fields):
>>Remove
>>>         assert.
>>>         (streamer_write_builtin): Remove.
>>>
>>>         lto/
>>>         * lto.c (compare_tree_sccs_1): Remove
>>streamer_handle_as_builtin_p uses.
>>>         (unify_scc): Likewise.
>>>         (lto_read_decls): Likewise.
>>>
>>
>>This caused:
>>
>>https://gcc.gnu.org/bugzilla/show_bug.cgi?id=72683
>
> Probably another by-decl built-in function compare in the mpx support.  I have fixed the one that triggered on a not mpx capable machine.  Ilya possibly knows where the other one(s) are lurking off his head?

I found a couple more of such comparisons.  Their replacement fixes
PR72683 and PR72657.

Thanks,
Ilya

>
> Richard.
>
>


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]