Ada and NTPL

Ada and NTPL

[Laurent Guerby]

I did succeed at get tasking working on Red Hat 9 x86 and NTPL
using the ACT tree (HEAD + gcc-head as of 23Apr2003 around 20:00 GMT)
and GCC mainline. The resulting compiler gets much better results than
3.2 and results in line with the ACT 5.00 release
except for a miscompilation of a-calend.adb that gives 17 failures.

For the record here is a reduced test case:

with Ada.Calendar;
procedure P1 is
   Some_Time : Ada.Calendar.Time;
begin
   Some_Time := Ada.Calendar.Time_Of (Month => 9, Day => 16, Year => 1993);
end P1;
$ gnatmake p1
$ ./p1

raised CONSTRAINT_ERROR : a-calend.adb:418 explicit raise
$

The code failing in Ada.Calendar is:

      if        not Year   'Valid
        or else not Month  'Valid
        or else not Day    'Valid
        or else not Seconds'Valid
      then
         raise Constraint_Error;
      end if;

I did check the assembly.

The build process is as follows:

1/ replace gcc/ada in GCC by the sources from the ACT tree
2/ copy -f all files from the gcc-head ACT source tree into gcc/ada
3/ apply the following patch

Index: Make-lang.in
===================================================================
RCS file: /anoncvs/gnat/Make-lang.in,v
retrieving revision 1.131.2.10
diff -u -r1.131.2.10 Make-lang.in
--- Make-lang.in	21 Apr 2003 12:24:45 -0000	1.131.2.10
+++ Make-lang.in	23 Apr 2003 18:29:21 -0000
@@ -328,12 +328,12 @@
 	   ADA_INCLUDES="-I- -Iada/ -I$(srcdir)/ada"\
 	   CC="stage2/xgcc -Bstage2" STAGE_PREFIX=stage2/  gnattools1
 	$(MAKE) -C ada $(ADA_TOOLS_FLAGS_TO_PASS) gnattools2
-	$(MAKE) -C ada $(ADA_TOOLS_FLAGS_TO_PASS) gnattools3
+#	$(MAKE) -C ada $(ADA_TOOLS_FLAGS_TO_PASS) gnattools3
 
 regnattools: 
 	$(MAKE) -C ada $(ADA_TOOLS_FLAGS_TO_PASS) gnattools1-re
 	$(MAKE) -C ada $(ADA_TOOLS_FLAGS_TO_PASS) gnattools2
-	$(MAKE) -C ada $(ADA_TOOLS_FLAGS_TO_PASS) gnattools3
+#	$(MAKE) -C ada $(ADA_TOOLS_FLAGS_TO_PASS) gnattools3
 
 # use host-gcc host-gnatmake host-gnatbind host-gnatlink
 # put the host RTS dir first in the PATH to hide the default runtime
Index: Makefile.in
===================================================================
RCS file: /anoncvs/gnat/Makefile.in,v
retrieving revision 1.1679.2.8
diff -u -r1.1679.2.8 Makefile.in
--- Makefile.in	28 Mar 2003 10:13:17 -0000	1.1679.2.8
+++ Makefile.in	23 Apr 2003 18:29:25 -0000
@@ -1283,7 +1283,7 @@
 
 include $(fsrcdir)/Makefile.rtl
 
-GNATRTL_OBJS = $(GNATRTL_NONTASKING_OBJS) $(GNATRTL_TASKING_OBJS) g-trasym.o
+GNATRTL_OBJS = $(GNATRTL_NONTASKING_OBJS) $(GNATRTL_TASKING_OBJS) g-trasym.o memtrack.o
 
 # Files which are suitable in no run time/hi integrity mode
 
Index: 5iosinte.ads
===================================================================
RCS file: /anoncvs/gnat/5iosinte.ads,v
retrieving revision 1.31
diff -u -r1.31 5iosinte.ads
--- 5iosinte.ads	31 Oct 2002 15:44:58 -0000	1.31
+++ 5iosinte.ads	24 Apr 2003 06:58:55 -0000
@@ -507,9 +507,13 @@
    end record;
    pragma Convention (C, pthread_mutex_t);
 
+   type pthread_cond_padding_t is array (0 .. 35) of unsigned_char;
+   pragma Convention (C, pthread_cond_padding_t);
+
    type pthread_cond_t is record
       c_lock     : struct_pthread_fast_lock;
       c_waiting  : System.Address;
+      c_padding  : pthread_cond_padding_t;
    end record;
    pragma Convention (C, pthread_cond_t);
 
Then it should bootstrap without problem and run quite well on
http://perso.wanadoo.fr/guerby/ftp/acats4gnat-0.7.tgz
using "gcc version 3.2.2 20030222 (Red Hat Linux 3.2.2-5)"
as base compiler for C & Ada.

-- 
Laurent Guerby <guerby@acm.org>

Re: Ada and NTPL

[Chris Proctor]

On Thu, Apr 24, 2003 at 09:12:18AM +0200, Laurent Guerby wrote:
> I did succeed at get tasking working on Red Hat 9 x86 and NTPL
> using the ACT tree (HEAD + gcc-head as of 23Apr2003 around 20:00 GMT)
> and GCC mainline. The resulting compiler gets much better results than
> 3.2 and results in line with the ACT 5.00 release
> except for a miscompilation of a-calend.adb that gives 17 failures.
> 
> For the record here is a reduced test case:
> 
> with Ada.Calendar;
> procedure P1 is
>    Some_Time : Ada.Calendar.Time;
> begin
>    Some_Time := Ada.Calendar.Time_Of (Month => 9, Day => 16, Year => 1993);
> end P1;
> $ gnatmake p1
> $ ./p1
> 
> raised CONSTRAINT_ERROR : a-calend.adb:418 explicit raise
> $
> 
> The code failing in Ada.Calendar is:
> 
>       if        not Year   'Valid
>         or else not Month  'Valid
>         or else not Day    'Valid
>         or else not Seconds'Valid
>       then
>          raise Constraint_Error;
>       end if;
> 
> I did check the assembly.
> 

The Seconds'Valid check is what failing.
This only miscompiles with optimisation and appears to be i386 only failure
(sparc-solaris works correctly).

This is failure is documented in ada/9536 and ada/9129 and
in message thread starting at
<http://gcc.gnu.org/ml/gcc/2002-12/msg00167.html>

The 'Valid check is optimized into False no matter what the actual
value of the parameter is.

Below is a more accurate test case:

--
--  Testcase for ada/9536, ada/9129
--

with Ada.Text_IO;

procedure Ada9536 is

  type Day_Duration is new Duration range 0.0 .. 86400.0;

  function Valid_Test (D : in Day_Duration) return Boolean is
  begin
    return D'Valid;
  end Valid_Test;

begin

  if Valid_Test (0.0) then
    Ada.Text_IO.Put_Line ("PASSED");
  else
    Ada.Text_IO.Put_Line ("FAILED");
  end if;

end Ada9536;

Re: Ada and NTPL

[Geert Bosch]

On Thursday, Apr 24, 2003, at 03:12 America/New_York, Laurent Guerby 
wrote:
> I did succeed at get tasking working on Red Hat 9 x86 and NTPL
> using the ACT tree (HEAD + gcc-head as of 23Apr2003 around 20:00 GMT)
> and GCC mainline. The resulting compiler gets much better results than
> 3.2 and results in line with the ACT 5.00 release
> except for a miscompilation of a-calend.adb that gives 17 failures.
>
> For the record here is a reduced test case:
>
> with Ada.Calendar;
> procedure P1 is
>    Some_Time : Ada.Calendar.Time;
> begin
>    Some_Time := Ada.Calendar.Time_Of (Month => 9, Day => 16, Year => 
> 1993);
> end P1;
> $ gnatmake p1
> $ ./p1
>
> raised CONSTRAINT_ERROR : a-calend.adb:418 explicit raise
> $
>
> The code failing in Ada.Calendar is:
>
>       if        not Year   'Valid
>         or else not Month  'Valid
>         or else not Day    'Valid
>         or else not Seconds'Valid
>       then
>          raise Constraint_Error;
>       end if;
>

There is already a GNATS bug filed for this. I looked at it and
this is a bug triggered by a change in the GCC back end and is not
related to front end changes. This is a regression compared to 3.2.

   -Geert

Re: Ada and NTPL

[Robert Dewar]

> Note that there are about a half dozen patches from GNAT's gcc-32.dif that
> are still to be applied to the GCC mainline.  Hopefully one of them will
> solve this, but it'll probably be about a week before they all make it in
> since I'm out of town for a few days.


Note also that the file gcc-32.dif is accessible at the libre site

Re: Ada and NTPL

[Richard Kenner]

Note that there are about a half dozen patches from GNAT's gcc-32.dif that
are still to be applied to the GCC mainline.  Hopefully one of them will
solve this, but it'll probably be about a week before they all make it in
since I'm out of town for a few days.