[Bug ada/46490] New: For four major i386 BSDs, GNAT fails FP to static integer conversion with -O2,-O3 optimization

[Bug ada/46490] New: For four major i386 BSDs, GNAT fails FP to static integer conversion with -O2,-O3 optimization

[gnugcc at marino dot st]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46490

           Summary: For four major i386 BSDs, GNAT fails FP to static
                    integer conversion with -O2,-O3 optimization
           Product: gcc
           Version: 4.6.0
               URL: http://www.dragonlace.net
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: ada
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: gnugcc@marino.st
              Host: Same as Target
            Target: FreeBSD, OpenBSD, NetBSD, DragonFlyBSD: i386
             Build: Same as Target


Created attachment 22404
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=22404
Nested_C460007 program (1 of 1)

I have successfully ported GNAT to all four major BSDs on two architectures,
i386 and x86_64 (8 ports total).
There is a failure of an ACATS test, cb460007, that fails on all four BSD
platforms, but only for the i386 arch.
All of the x86_64 ports pass the test just fine.

I have discovered the test passes on i386 if optimization level 0 or 1 is
specified.
Furthermore, if the test is re-written such that the program is contained
within a single procedure, the
test doesn't fail with any optimization levels.

I have provided two test programs.  The SlimC46007 program is cut-down version
of the ACATS test c460007, and tests case 9 and 10 only (conversion of floating
point to static integer).  The Nested_C460007 program does the same thing, but
without the extra package.


Output of running SlimC46007 to demonstrate effect of optimization levels.

> gnatmake -O2 slimc460007.adb
gnatgcc -c -O2 slimc460007.adb
gnatgcc -c -O2 c460007_1.adb
gnatbind -x slimc460007.ali
gnatlink slimc460007.ali -O2

> ./slimc460007
Test C460007: Rounding for type conversions of real operand to integer target
CASE  9 FAILED (static target subtype)
CASE 10 FAILED (static target subtype)

> rm *.o *.ali
> gnatmake slimc460007.adb
gnatgcc -c slimc460007.adb
gnatgcc -c c460007_1.adb
gnatbind -x slimc460007.ali
gnatlink slimc460007.ali

> ./slimc460007
Test C460007: Rounding for type conversions of real operand to integer target
BOTH TESTS PASSED.




Output of running Nested_C460007 where O-level has no effect
> gnatmake -O2 nested_c460007.adb
gnatgcc -c -O2 nested_c460007.adb
gnatbind -x nested_c460007.ali
gnatlink nested_c460007.ali -O2

> ./nested_c460007
BOTH TESTS PASSED.

> rm *.o *.ali
> gnatmake nested_c460007.adb
gnatgcc -c nested_c460007.adb
gnatbind -x nested_c460007.ali
gnatlink nested_c460007.ali

> ./nested_c460007
BOTH TESTS PASSED.



Note: I've been tracking this failure for months.  It's been part of 4.6.0 as
long as I can remember.
It is the only test failure for the i386 FreeBSD GNAT and the i386 DragonFly
BSD GNAT.  The i386 NetBSD GNAT is doesn't seem to support stack-checking, so
omitting those stack-check ACATS and DejaGNU test, this is the only failure on
that platform as well.


The same test was run on all 4 listed platforms with the exact same results.


i386 DragonFlyBSD 2.6  (13 Nov 2010)
====================================
> gnatgcc -v
Using built-in specs.
COLLECT_GCC=gnatgcc
COLLECT_LTO_WRAPPER=/usr/local/libexec/gcc/i386-backplane-dragonfly2.6/4.6.0/lto-wrapper
Target: i386-backplane-dragonfly2.6
Configured with: /home/marino/gnat_port/gen_x/work/gcc-4.6-20101113/configure
--enable-languages=c,ada --build=i386-backplane-dragonfly2.6
--program-prefix=gnat --prefix=/usr/local --with-system-zlib
--with-gmp=/usr/pkg --with-mpfr=/usr/pkg --with-mpc=/usr/pkg
--with-libiconv-prefix=/usr/pkg --enable-threads=single --enable-shared
--disable-bootstrap --disable-libgomp --disable-libssp --disable-nls
Thread model: single
gcc version 4.6.0 20101113 (experimental) -=> GNAT AUX [DragonFly BSD] (GCC)



i386 FreeBSD 8.0    (30 Oct 2010)
=================================
> gnatgcc -v
Using built-in specs.
COLLECT_GCC=gnatgcc
COLLECT_LTO_WRAPPER=/usr/local/libexec/gcc/i386-portbld-freebsd8.0/4.6.0/lto-wrapper
Target: i386-portbld-freebsd8.0
Configured with: /usr/home/marino/workzone/port/work/gcc-4.6-20101030/configure
--enable-languages=c,ada --build=i386-portbld-freebsd8.0 --program-prefix=gnat
--prefix=/usr/local --with-system-zlib --with-gmp=/usr/local
--with-mpfr=/usr/local --with-mpc=/usr/local --with-libiconv-prefix=/usr/local
--enable-shared --disable-bootstrap --disable-libgomp --disable-libssp
--disable-nls
Thread model: posix
gcc version 4.6.0 20101030 (experimental) -=> GNAT AUX [FreeBSD] (GCC)



i386 NetBSD 5.1RC4  (30 Oct 2010)
=================================
> gnatgcc -v
Using built-in specs.
COLLECT_GCC=gnatgcc
COLLECT_LTO_WRAPPER=/usr/local/libexec/gcc/i386-pc-netbsdelf5.1/4.6.0/lto-wrapper
Target: i386-pc-netbsdelf5.1
Configured with: /workzone/gnat_port/gen_2/work/gcc-4.6-20101030/configure
--enable-languages=c,ada --build=i386-pc-netbsdelf5.1 --program-prefix=gnat
--prefix=/usr/local --with-system-zlib --with-gmp=/usr/pkg --with-mpfr=/usr/pkg
--with-mpc=/usr/pkg --with-libiconv-prefix=/usr/pkg --enable-threads=yes
--enable-shared --disable-bootstrap --disable-libgomp --disable-libssp
--disable-nls
Thread model: posix
gcc version 4.6.0 20101030 (experimental) -=> GNAT AUX [NetBSD] (GCC)



i386 OpenBSD 4.7    (30 Oct 2010)
=================================
> gnatgcc -v
Using built-in specs.
COLLECT_GCC=gnatgcc
COLLECT_LTO_WRAPPER=/usr/local/libexec/gcc/i386-pc-openbsd4.7/4.6.0/lto-wrapper
Target: i386-pc-openbsd4.7
Configured with: /home/marino/gnat_port/gen_3/work/gcc-4.6-20101030/configure
--enable-languages=c,ada --build=i386-pc-openbsd4.7 --program-prefix=gnat
--prefix=/usr/local --with-system-zlib --with-gmp=/usr/local
--with-mpfr=/usr/local --with-mpc=/usr/local --with-libiconv-prefix=/usr/local
--enable-threads=yes --enable-shared --disable-bootstrap --disable-libgomp
--disable-libssp --disable-nls
Thread model: posix
gcc version 4.6.0 20101030 (experimental) -=> GNAT AUX [OpenBSD] (GCC)

[Bug ada/46490] For four major i386 BSDs, GNAT fails FP to static integer conversion with -O2,-O3 optimization

[gnugcc at marino dot st]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46490

--- Comment #1 from John Marino <gnugcc at marino dot st> 2010-11-15 19:53:38 UTC ---
Created attachment 22405
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=22405
SlimC46007 program (1 of 3)

[Bug ada/46490] For four major i386 BSDs, GNAT fails FP to static integer conversion with -O2,-O3 optimization

[gnugcc at marino dot st]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46490

--- Comment #2 from John Marino <gnugcc at marino dot st> 2010-11-15 19:55:34 UTC ---
Created attachment 22406
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=22406
SlimC46007 program (2 of 3)

[Bug ada/46490] For four major i386 BSDs, GNAT fails FP to static integer conversion with -O2,-O3 optimization

[gnugcc at marino dot st]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46490

--- Comment #3 from John Marino <gnugcc at marino dot st> 2010-11-15 19:56:33 UTC ---
Created attachment 22407
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=22407
SlimC46007 program (3 of 3)

[Bug ada/46490] For four major i386 BSDs, GNAT fails FP to static integer conversion with -O2,-O3 optimization

[ebotcazou at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46490

Eric Botcazou <ebotcazou at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2010.11.16 08:22:33
                 CC|                            |ebotcazou at gcc dot
                   |                            |gnu.org
     Ever Confirmed|0                           |1

--- Comment #4 from Eric Botcazou <ebotcazou at gcc dot gnu.org> 2010-11-16 08:22:33 UTC ---
There is a known issue for FreeBSD.  GNAT requires 64-bit precision on the x87
so it resets it in ada/init.c:__gnat_init_float.  This runs afoul of:

/* FreeBSD sets the rounding precision of the FPU to 53 bits.  Let the
   compiler get the contents of <float.h> and std::numeric_limits correct.  */
#undef TARGET_96_ROUND_53_LONG_DOUBLE
#define TARGET_96_ROUND_53_LONG_DOUBLE (!TARGET_64BIT)

in config/i386/freebsd.h so you'd need to force it to 0 unconditionally for
GNAT.


For the stack checking, you should try to define:

/* Static stack checking is supported by means of probes.  */
#define STACK_CHECK_STATIC_BUILTIN 1

like in config/i386/freebsd.h.

[Bug ada/46490] For four major i386 BSDs, GNAT fails FP to static integer conversion with -O2,-O3 optimization

[gnugcc at marino dot st]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46490

--- Comment #5 from John Marino <gnugcc at marino dot st> 2010-11-16 10:35:54 UTC ---
Hi Eric,
Thanks for you comment, but I don't think that is it for several reasons:

1) I am aware of both those quirks, and my codebase is patched with both of
those suggestions.

2) There's no problem with stack checking on FreeBSD i386.  There are no
failures whatsoever on either ACATS or gnat.dejagnu except for this particular
case of c460007.  NetBSD stack-checking doesn't work, but that's outside this
bug report.

3) The test passes with optimization level 0 or 1.  It fails with O2 and O3
settings.

4) There are no other rounding issues.  If TARGET_96_ROUND_53_LONG_DOUBLE is
not set correctly, multiple tests fail.

5) It's occurring identically on all four i386 platforms which are pretty
divergent.


This looks like a problem in the optimization stage.
The two test cases I provided are interesting as you can see the test will pass
if the EQUAL function is nested inside the main procedure rather than being
called as a separate package (also pointing to a problem in optimization I
would think.)

This warrants a closer look, it's not a configuration thing.
Thanks,
John

[Bug ada/46490] ACATS cb460007 fails at -O2 or above

[ebotcazou at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46490

Eric Botcazou <ebotcazou at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |wrong-code
             Target|FreeBSD, OpenBSD, NetBSD,   |i386-*-*
                   |DragonFlyBSD: i386          |
               Host|Same as Target              |
            Summary|For four major i386 BSDs,   |ACATS cb460007 fails at -O2
                   |GNAT fails FP to static     |or above
                   |integer conversion with     |
                   |-O2,-O3 optimization        |
              Build|Same as Target              |

--- Comment #6 from Eric Botcazou <ebotcazou at gcc dot gnu.org> 2010-11-16 11:16:40 UTC ---
> 5) It's occurring identically on all four i386 platforms which are pretty
> divergent.
> 
> This looks like a problem in the optimization stage.

OK, the summary is slightly misleading then, if this has nothing to do with
BSD.

Do you know if this fails with earlier versions?  Let me try on Linux.

[Bug ada/46490] ACATS cb460007 fails at -O2 or above

[gnugcc at marino dot st]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46490

--- Comment #7 from John Marino <gnugcc at marino dot st> 2010-11-16 11:25:31 UTC ---
Eric, 
Actually I believe it is limited to the BSDs, although I can't explain why.  I
also ported GNAT to x86 OpenSolaris (SXCE 130) and that one passed all ACATS.  
I also have not seen this failure on any of the Linux testsuite reports for
4.6.0.

This error may have also been present in the 4.5.0 branch, but there were
several other failures back then so I wasn't paying particular attention to it
at the time.

FYI, here are recent test reports if you find this helpful:

DragonFly32 : http://www.dragonlace.net/gnataux/dragonfly32/
FreeBSD32   : http://www.dragonlace.net/gnataux/freebsd32/
OpenBSD32   : http://www.dragonlace.net/gnataux/openbsd32/
NetBSD32    : http://www.dragonlace.net/gnataux/netbsd32/

Also, remember x86_64 works fine, there is no issue with -O2,-O3 with this test
on that architecture.

[Bug ada/46490] ACATS c460007 fails at -O2 or above

[ebotcazou at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46490

Eric Botcazou <ebotcazou at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |ASSIGNED
         AssignedTo|unassigned at gcc dot       |ebotcazou at gcc dot
                   |gnu.org                     |gnu.org

--- Comment #8 from Eric Botcazou <ebotcazou at gcc dot gnu.org> 2010-11-16 11:46:51 UTC ---
> Actually I believe it is limited to the BSDs, although I can't explain why.

You need -mtune=i386 to reproduce on Linux for example.

> FYI, here are recent test reports if you find this helpful:
> 
> DragonFly32 : http://www.dragonlace.net/gnataux/dragonfly32/
> FreeBSD32   : http://www.dragonlace.net/gnataux/freebsd32/
> OpenBSD32   : http://www.dragonlace.net/gnataux/openbsd32/
> NetBSD32    : http://www.dragonlace.net/gnataux/netbsd32/

Thanks.  If the OpenBSD port is SJLJ, you can add an XFAIL for *-*-openbsd* to
gnat.dg/test_raise_from_pure.adb.

[Bug rtl-optimization/46490] [4.6 regression] ACATS c460007 fails at -O2 or above

[ebotcazou at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46490

Eric Botcazou <ebotcazou at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
          Component|ada                         |rtl-optimization
      Known to work|                            |4.5.1
   Target Milestone|---                         |4.6.0
            Summary|ACATS c460007 fails at -O2  |[4.6 regression] ACATS
                   |or above                    |c460007 fails at -O2 or
                   |                            |above

--- Comment #9 from Eric Botcazou <ebotcazou at gcc dot gnu.org> 2010-11-16 17:07:00 UTC ---
Recategorizing.

[Bug rtl-optimization/46490] [4.6 regression] ACATS c460007 fails at -O2 or above

[ebotcazou at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46490

--- Comment #10 from Eric Botcazou <ebotcazou at gcc dot gnu.org> 2010-11-16 22:30:25 UTC ---
Author: ebotcazou
Date: Tue Nov 16 22:30:19 2010
New Revision: 166830

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=166830
Log:
    PR rtl-optimization/46490
    * combine.c (expand_compound_operation): Fix thinko.

Added:
    trunk/gcc/testsuite/gnat.dg/opt12.adb
    trunk/gcc/testsuite/gnat.dg/opt12_pkg.adb
    trunk/gcc/testsuite/gnat.dg/opt12_pkg.ads
Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/combine.c
    trunk/gcc/testsuite/ChangeLog

[Bug rtl-optimization/46490] [4.6 regression] ACATS c460007 fails at -O2 or above

[ebotcazou at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46490

Eric Botcazou <ebotcazou at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |RESOLVED
                URL|http://www.dragonlace.net   |http://gcc.gnu.org/ml/gcc-p
                   |                            |atches/2010-11/msg01736.htm
                   |                            |l
         Resolution|                            |FIXED

--- Comment #11 from Eric Botcazou <ebotcazou at gcc dot gnu.org> 2010-11-16 22:32:39 UTC ---
This should work again.

[Bug rtl-optimization/46490] [4.6 regression] ACATS c460007 fails at -O2 or above

[gnugcc at marino dot st]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46490

--- Comment #12 from John Marino <gnugcc at marino dot st> 2010-11-16 23:07:46 UTC ---
That's great news, Eric!

Many thanks, both for spending some long hours resolving this the same day it
was reported, and for the tip on the raise_from_pure test.  This results in two
GNAT ports have zero ACATS errors.

John