NetBSD Problem Report #46677

From www@NetBSD.org  Mon Jul  9 15:41:47 2012
Return-Path: <www@NetBSD.org>
Received: from mail.netbsd.org (mail.netbsd.org [149.20.53.66])
	by www.NetBSD.org (Postfix) with ESMTP id BA55163B8E6
	for <gnats-bugs@gnats.NetBSD.org>; Mon,  9 Jul 2012 15:41:46 +0000 (UTC)
Message-Id: <20120709154146.220EC63B882@www.NetBSD.org>
Date: Mon,  9 Jul 2012 15:41:46 +0000 (UTC)
From: ragge@ludd.ltu.se
Reply-To: ragge@ludd.ltu.se
To: gnats-bugs@NetBSD.org
Subject: gcc bug with builtin ffs on vax.
X-Send-Pr-Version: www-1.0

>Number:         46677
>Category:       port-vax
>Synopsis:       gcc bug with builtin ffs on vax.
>Confidential:   no
>Severity:       serious
>Priority:       medium
>Responsible:    port-vax-maintainer
>State:          closed
>Class:          sw-bug
>Submitter-Id:   net
>Arrival-Date:   Mon Jul 09 15:45:00 +0000 2012
>Closed-Date:    Wed Jul 11 13:41:36 +0000 2012
>Last-Modified:  Thu Jul 12 18:30:09 +0000 2012
>Originator:     Anders Magnusson
>Release:        6.0 beta2
>Organization:
>Environment:
NetBSD vs4090.tethuvudet.se 6.0_BETA2 NetBSD 6.0_BETA2 (GENERIC) #10: Mon Jul  9 17:02:03 CEST 2012  ragge@bakfull:/local/home/ragge/vaxsrc/60beta/sys/arch/vax/compile/obj/GENERIC vax

>Description:
The builtin ffs function fails in some situations for gcc both 4.1 or 4.5, and both native and cross-compiled.


>How-To-Repeat:
The code example below shows that the second ffs gets optimized away.

vs4090:/home/ragge >cat x.c

struct foo { long __bits[4]; };

int
firstsig(struct foo *ss)
{
        int sig;

        sig = ffs(ss->__bits[0]);
        if (sig != 0)
                return (sig);

        sig = ffs(ss->__bits[1]);
        if (sig != 0)
                return (sig + 32);

        return (0);
}

vs4090:/home/ragge >gcc -S -O2 x.c
vs4090:/home/ragge >cat x.s
#NO_APP
        .file   "x.c"
        .text
        .align 1
.globl firstsig
        .type   firstsig, @function
firstsig:
        .word 0x0
        subl2 $4,%sp
        ffs $0,$32,*4(%ap),%r0
        jneq .L2
        mnegl $1,%r0
.L2:
        incl %r0
        ret
        .size   firstsig, .-firstsig
        .ident  "GCC: (GNU) 4.1.3 20080704 prerelease (NetBSD nb3 20111107)"

>Fix:

>Release-Note:

>Audit-Trail:
From: <Paul_Koning@Dell.com>
To: <gnats-bugs@NetBSD.org>
Cc: <port-vax-maintainer@netbsd.org>, <gnats-admin@netbsd.org>,
	<netbsd-bugs@netbsd.org>
Subject: Re: port-vax/46677: gcc bug with builtin ffs on vax.
Date: Tue, 10 Jul 2012 21:30:05 +0000

 This seems to help.

 Index: gnu/dist/gcc4/gcc/config/vax/builtins.md
 =3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=
 =3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=
 =3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D
 --- gnu/dist/gcc4/gcc/config/vax/builtins.md	(revision 254596)
 +++ gnu/dist/gcc4/gcc/config/vax/builtins.md	(working copy)
 @@ -25,7 +25,7 @@
    "
  {
    rtx label =3D gen_label_rtx ();
 -  emit_insn (gen_ffssi2_internal (operands[0], operands[1], operands[0]));
 +  emit_insn (gen_ffssi2_internal (operands[0], operands[1], operands[1]));
    emit_jump_insn (gen_bne (label));
    emit_insn (gen_negsi2 (operands[0], const1_rtx));
    emit_label (label);
 @@ -36,6 +36,6 @@
  (define_insn "ffssi2_internal"
    [(set (match_operand:SI 0 "nonimmediate_operand" "=3Dg")
          (ffs:SI (match_operand:SI 1 "general_operand" "nrQ")))
 -   (set (cc0) (match_operand:SI 2 "nonimmediate_operand" "0"))]
 +   (set (cc0) (ffs:SI (match_operand:SI 2 "general_operand" "1")))]
    ""
    "ffs $0,$32,%1,%0")


 The reason is that, according to the GCC internals documentation, "parallel=
 " means the operations are done in parallel, i.e., first all the inputs are=
  computed and then all the outputs are set.  So the original code (which se=
 ts cc0 based on the output operand of the FFS) would set cc0 based on the p=
 revious value, NOT based on the result of the ffs.  The changed code says t=
 hat the condition code is set based on the ffs answer.

 It looks like it works correctly, judging by the generated assembly code.

 	paul

From: "Anders Magnusson" <ragge@netbsd.org>
To: gnats-bugs@gnats.NetBSD.org
Cc: 
Subject: PR/46677 CVS commit: src/gnu/dist/gcc4/gcc/config/vax
Date: Wed, 11 Jul 2012 13:34:45 +0000

 Module Name:	src
 Committed By:	ragge
 Date:		Wed Jul 11 13:34:45 UTC 2012

 Modified Files:
 	src/gnu/dist/gcc4/gcc/config/vax: builtins.md

 Log Message:
 Bugfix builtin ffs, fixes PR port-vax/46677, fix from Paul Koning.


 To generate a diff of this commit:
 cvs rdiff -u -r1.2 -r1.3 src/gnu/dist/gcc4/gcc/config/vax/builtins.md

 Please note that diffs are not public domain; they are subject to the
 copyright notices on the relevant files.

State-Changed-From-To: open->closed
State-Changed-By: ragge@NetBSD.org
State-Changed-When: Wed, 11 Jul 2012 13:41:36 +0000
State-Changed-Why:
Recompiled system, tested and the bug had gone. Great, thanks!


From: "Jeff Rizzo" <riz@netbsd.org>
To: gnats-bugs@gnats.NetBSD.org
Cc: 
Subject: PR/46677 CVS commit: [netbsd-6] src/gnu/dist/gcc4/gcc/config/vax
Date: Thu, 12 Jul 2012 18:29:57 +0000

 Module Name:	src
 Committed By:	riz
 Date:		Thu Jul 12 18:29:57 UTC 2012

 Modified Files:
 	src/gnu/dist/gcc4/gcc/config/vax [netbsd-6]: builtins.md

 Log Message:
 Pull up following revision(s) (requested by ragge in ticket #411):
 	gnu/dist/gcc4/gcc/config/vax/builtins.md: revision 1.3
 Bugfix builtin ffs, fixes PR port-vax/46677, fix from Paul Koning.


 To generate a diff of this commit:
 cvs rdiff -u -r1.2 -r1.2.40.1 src/gnu/dist/gcc4/gcc/config/vax/builtins.md

 Please note that diffs are not public domain; they are subject to the
 copyright notices on the relevant files.

>Unformatted:

Home	PR Database Search