public inbox for gcc-bugs@sourceware.org
help / color / mirror / Atom feed
From: "changyp6 at gmail dot com" <gcc-bugzilla@gcc.gnu.org>
To: gcc-bugs@gcc.gnu.org
Subject: [Bug c/102124] New: GCC 11.2.1 -ftree-loop-vectorize Causing Data To Lose Sign Bit
Date: Mon, 30 Aug 2021 08:07:16 +0000	[thread overview]
Message-ID: <bug-102124-4@http.gcc.gnu.org/bugzilla/> (raw)

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102124

            Bug ID: 102124
           Summary: GCC 11.2.1 -ftree-loop-vectorize Causing Data To Lose
                    Sign Bit
           Product: gcc
           Version: 11.2.1
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
          Assignee: unassigned at gcc dot gnu.org
          Reporter: changyp6 at gmail dot com
  Target Milestone: ---

Created attachment 51374
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=51374&action=edit
Test program for gcc 11.2.1 on AARCH64

Description of problem:
When I'm building libgcrypt 1.9.4 with GCC 11.2.1 on my AARCH64 box(armv8.2
cortex-a76), I'm using -O3 compile options. however, -O3 generated code failed
to pass "basic" test of libgcrypt, it fails on "gcry_cipher_checktag" function.
After investigation, I found that, the problem occurs in buf_eq_const()
function in file cipher/bufhelp.h of libgcrypt-1.9.4

 362 /* Constant-time compare of two buffers.  Returns 1 if buffers are equal,
 363    and 0 if buffers differ.  */
 364 static inline int
 365 buf_eq_const(const void *_a, const void *_b, size_t len)
 366 {
 367   const byte *a = _a;
 368   const byte *b = _b;
 369   int ab, ba;
 370   size_t i;
 371 
 372   /* Constant-time compare. */
 373   for (i = 0, ab = 0, ba = 0; i < len; i++)
 374     {
 375       /* If a[i] != b[i], either ab or ba will be negative. */
 376       ab |= a[i] - b[i];
 377       ba |= b[i] - a[i];
 378     }
 379 
 380   /* 'ab | ba' is negative when buffers are not equal. */
 382   return (ab | ba) >= 0;
 383 }

The calculation of 2 different array becomes >= 0 on the return value, however,
it should be negative value.

After I change -O3 to -O2, this function works again.
Then I compile libgcrypt 1.9.4 with -O2 plus additional GCC options which are
added by -O3 to locate the actual option that causing this issue, finally I
found that, if "-ftree-loop-vectorize" is used to compile this code, the
calculated result is a positive value, if removing "-ftree-loop-vectorize", the
calculated result is negative.

Then I downgraded GCC to 10.x, -ftree-loop-vectorize won't cause such issue.
So I'm sure this is a GCC bug.

It seems that "-ftree-loop-vectorize" causing "|=" operation ignore the "sign
bit" of "a[i] - b[i]" or "b[i] - a[i]".

I have summarized a test-case, which is attached


I have also compiled a cross-toolchain, using gcc git version
(98e482761b083dbc35ae59704ee1eeb0b8eeb5d1), which is also gcc 11.2.1, this git
version also has such issue.



Version-Release number of selected component (if applicable):
GCC 11.2.1

Additional Info:
GCC 11 for x86 / x86_64 doesn't have such issue.
GCC 10.x for aarch64 doesn't have such issue.

I have also submitted this bug to Fedora bugzilla
https://bugzilla.redhat.com/show_bug.cgi?id=1998964

             reply	other threads:[~2021-08-30  8:07 UTC|newest]

Thread overview: 12+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2021-08-30  8:07 changyp6 at gmail dot com [this message]
2021-08-30  8:11 ` [Bug c/102124] " changyp6 at gmail dot com
2021-08-30  8:32 ` [Bug target/102124] [11/12 Regression] -ftree-loop-vectorize Causing Data To Lose Sign Bit on AARCH64 Platform jakub at gcc dot gnu.org
2021-08-30  8:44 ` jakub at gcc dot gnu.org
2021-08-30  9:28 ` changyp6 at gmail dot com
2021-08-30 10:05 ` jakub at gcc dot gnu.org
2021-08-30 10:07 ` [Bug tree-optimization/102124] " jakub at gcc dot gnu.org
2021-08-30 10:15 ` jakub at gcc dot gnu.org
2021-09-01  4:19 ` changyp6 at gmail dot com
2021-09-01 11:40 ` cvs-commit at gcc dot gnu.org
2021-09-01 11:41 ` cvs-commit at gcc dot gnu.org
2021-09-01 11:48 ` jakub at gcc dot gnu.org

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=bug-102124-4@http.gcc.gnu.org/bugzilla/ \
    --to=gcc-bugzilla@gcc.gnu.org \
    --cc=gcc-bugs@gcc.gnu.org \
    /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).