public inbox for gcc-patches@gcc.gnu.org
 help / color / mirror / Atom feed
From: Bernhard Reutner-Fischer <rep.dot.nop@gmail.com>
To: Sandra Loosemore <sandra@codesourcery.com>
Cc: Mike Stump <mikestump@comcast.net>,
	Joseph Myers <joseph@codesourcery.com>,
		Jason Merrill <jason@redhat.com>,
	Richard Biener <richard.guenther@gmail.com>,
		gcc-patches Patches <gcc-patches@gcc.gnu.org>
Subject: Re: #pragma GCC unroll support
Date: Fri, 06 Mar 2015 12:37:00 -0000	[thread overview]
Message-ID: <CAC1BbcQ4VQZbL2bbfHd0rbpCKZOgOdKDL45Ksyd2VDo_14+1YA@mail.gmail.com> (raw)
In-Reply-To: <54F90389.8070101@codesourcery.com>

[-- Attachment #1: Type: text/plain, Size: 1555 bytes --]

On 6 March 2015 at 02:31, Sandra Loosemore <sandra@codesourcery.com> wrote:
> On 03/05/2015 04:12 PM, Mike Stump wrote:
>>
>>
>> Ping?
>>
>
> Just commenting on the documentation part:
[]

and a few coding style nits:

+++ b/gcc/c-family/c-pragma.c
@@ -1459,6 +1459,10 @@ init_pragma (void)
     cpp_register_deferred_pragma (parse_in, "GCC", "ivdep",
PRAGMA_IVDEP, false,
                                  false);

+  if (!flag_preprocess_only)
+    cpp_register_deferred_pragma (parse_in, "GCC", "unroll",
PRAGMA_UNROLL, false,
+                                 false);
+

overlong line (also for the IVDEP above)

+++ b/gcc/c/c-parser.c
+static void c_parser_while_statement (c_parser *, bool, unsigned short);
+static void c_parser_do_statement (c_parser *, bool, unsigned short);
+static void c_parser_for_statement (c_parser *, bool, unsigned short);

since we're now a C++ app I would have added a default for the
unsigned short unroll of = 0
Same for
finish_while_stmt_cond, finish_do_stmt, finish_for_cond et al.

In cp_parser_range_for() i take it you remember there is a //TODO
I am attaching an unroll-5.C which might show that this does not seem
to be implemented yet, IIUC

gcc/loop-unroll.c::decide_unrolling()
I'd put the "if (loop->unroll == 1) {continue}" earlier in the
FOR_EACH_LOOP body (we're C++ nowadays) but maybe our optimizers are
good enough to do that anyway (but i fear we're not up to that?).

I did not see c/c++ tests for both !DIR$ UNROLL and !DIR$ IVDEP, fwiw.
You seem to handle both placements proper, though.
cheers,

[-- Attachment #2: unroll-5.C --]
[-- Type: text/x-csrc, Size: 1012 bytes --]

/* { dg-do compile } */
/* { dg-options "-O2 -fdump-rtl-loop2_unroll -fdump-tree-cunrolli-details" } */

#include <string>
#include <vector>

void bar(int);
int j;

void test1() {
  unsigned long m = j;
  const std::string s("abcdefgh");
  const std::vector<int> v = {7, 6, 5, 4, 3, 2, 1, 0};
  const std::size_t v_sz = 8;

  // { dg-final { scan-tree-dump "18:\[0-9\]*: note: loop turned into non-loop" "cunrolli" } }
  #pragma GCC unroll sizeof("abcdefgh") * 4
  for (auto ch : s)
    bar(ch);

  // { dg-final { scan-tree-dump "23:\[0-9\]*: note: loop turned into non-loop" "cunrolli" } }
  #pragma GCC unroll v_sz
  for (auto i : v)
    bar(i);

  // { dg-final { scan-tree-dump "28:\[0-9\]*: note: loop turned into non-loop" "cunrolli" } }
  #pragma GCC unroll 99
  for (auto i : v)
    bar(i);
}
// { dg-final { scan-rtl-dump-not "Unable to prove that the loop iterates constant times" "loop2_unroll" } }
// { dg-final { cleanup-tree-dump "cunrolli" } }
// { dg-final { cleanup-rtl-dump "loop2_unroll" } }

  reply	other threads:[~2015-03-06 12:37 UTC|newest]

Thread overview: 25+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2014-12-22 21:51 Mike Stump
2014-12-23 20:22 ` Andi Kleen
2015-01-08 12:45 ` Richard Biener
2015-01-26 21:13   ` Mike Stump
2015-01-26 21:18     ` Mike Stump
2015-01-30  2:51       ` Joseph Myers
2015-01-30  6:12         ` Mike Stump
2015-01-30  2:54     ` Joseph Myers
2015-01-30  6:57       ` Mike Stump
2015-01-30 17:46         ` Joseph Myers
2015-01-30 17:52           ` Mike Stump
2015-01-30 18:06             ` Marek Polacek
2015-01-30 23:48               ` Bernhard Reutner-Fischer
2015-01-30 22:16                 ` Mike Stump
2015-02-02 23:22                   ` [PATCH, v0] fortran: !GCC$ unroll for DO Bernhard Reutner-Fischer
2015-02-03  8:42                     ` Tobias Burnus
2015-02-02 23:22                   ` [PATCH, RFC] fortran [was Re: #pragma GCC unroll support] Bernhard Reutner-Fischer
2015-02-03  0:08                     ` Mike Stump
2015-05-28  9:56                       ` Bernhard Reutner-Fischer
2015-05-28 12:29                         ` Mike Stump
2015-11-02 12:18                           ` Bernhard Reutner-Fischer
2015-03-05 23:13             ` #pragma GCC unroll support Mike Stump
2015-03-06  1:32               ` Sandra Loosemore
2015-03-06 12:37                 ` Bernhard Reutner-Fischer [this message]
2015-03-07  1:16               ` Joseph Myers

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=CAC1BbcQ4VQZbL2bbfHd0rbpCKZOgOdKDL45Ksyd2VDo_14+1YA@mail.gmail.com \
    --to=rep.dot.nop@gmail.com \
    --cc=gcc-patches@gcc.gnu.org \
    --cc=jason@redhat.com \
    --cc=joseph@codesourcery.com \
    --cc=mikestump@comcast.net \
    --cc=richard.guenther@gmail.com \
    --cc=sandra@codesourcery.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).