[Bug ada/59772] New: Floating point constants are not correctly encoded on AVR target

[Bug ada/59772] New: Floating point constants are not correctly encoded on AVR target

[j-etienne at users dot sourceforge.net]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59772

            Bug ID: 59772
           Summary: Floating point constants are not correctly encoded on
                    AVR target
           Product: gcc
           Version: unknown
            Status: UNCONFIRMED
          Severity: critical
          Priority: P3
         Component: ada
          Assignee: unassigned at gcc dot gnu.org
          Reporter: j-etienne at users dot sourceforge.net

In code compiled for AVR target, the floating point constants are not correct
in the generated code.

Sample code (specification first, then body) :
-------------
package Float_Test is

   function Div_Test (Value : Float) return Float;

   function Perimeter_Test (Value : Float) return Float;

end Float_Test;

package body Float_Test is

   K_Denominator : constant Float := Float (4096);
   K_Pi : constant Float := 3.1415927;

   function Div_Test (Value : Float) return Float
   is
   begin
      return (Value / K_Denominator);
   end Div_Test;

   function Perimeter_Test (Value : Float) return Float
   is
   begin
      return (Value * K_Pi);
   end Perimeter_Test;

end Float_Test;
-------------

When I commpile this sample code, with the following command :

avr-gcc -Wall -Wextra -fno-strict-aliasing -fwrapv
-fno-aggressive-loop-optimizations -mmcu=at90usb1287 -Os -S -gnatp -I
~/Programmation/Unix/gcc-4.8.2/gcc/ada float_test.adb -fdump-tree-original
-gnatD

It generates wrong values for both floating point constants :
-------------
    .file    "float_test.adb"
__SP_H__ = 0x3e
__SP_L__ = 0x3d
__SREG__ = 0x3f
__RAMPZ__ = 0x3b
__tmp_reg__ = 0
__zero_reg__ = 1
.global    __divsf3
    .text
.global    float_test__div_test
    .type    float_test__div_test, @function
float_test__div_test:
/* prologue: function */
/* frame size = 0 */
/* stack size = 0 */
.L__stack_usage = 0
    ldi r18,0
    ldi r19,0
    movw r20,r18
    call __divsf3
    ret
    .size    float_test__div_test, .-float_test__div_test
.global    __mulsf3
.global    float_test__perimeter_test
    .type    float_test__perimeter_test, @function
float_test__perimeter_test:
/* prologue: function */
/* frame size = 0 */
/* stack size = 0 */
.L__stack_usage = 0
    ldi r18,0
    ldi r19,lo8(-80)
    ldi r20,lo8(125)
    ldi r21,lo8(58)
    call __mulsf3
    ret
    .size    float_test__perimeter_test, .-float_test__perimeter_test
.global    float_test_E
    .data
    .type    float_test_E, @object
    .size    float_test_E, 2
float_test_E:
    .zero    2
    .ident    "GCC: (GNU) 4.8.2"
.global __do_copy_data
-------------

The denominator provided as argument to divsf3 is 0.0 instead of 4096, and the
constant term provided as argument to mulsf3 is 0.000967741 instead of
3.12415927

I'm using a cross GCC for AVR target hosted on MacOSX 10.9. avr-gcc -v gives :
Using built-in specs.
COLLECT_GCC=avr-gcc
COLLECT_LTO_WRAPPER=/opt/avr-ada-test/libexec/gcc/avr/4.8.2/lto-wrapper
Target: avr
Configured with: ../gcc-4.8.2/configure --prefix=/opt/avr-ada-test --target=avr
--program-prefix=avr- --disable-shared --disable-nls --disable-libssp
--with-system-zlib --disable-libada --enable-languages=ada,c --enable-cpp
--with-dwarf2 --enable-version-specific-runtime-libs --with-gmp=/opt/local
--with-mpfr=/opt/local --with-mpc=/opt/local
Thread model: single
gcc version 4.8.2 (GCC) 

Please note that I reproduced the same behaviour on a x86-linux host.
I first stumbled upon this bug when migrating my application to AVR GCC 4.7.2
using the RTS from AVR-Ada.
With GCC 4.8.2, I tried several "system.ads" locations : the one from the AVR
Ada RTS, from the native compiler, from the GCC source tree, and always got the
same result.

I noticed that in the GNAT debug file "float_test.adb.dg" , the constants are
correct :

-------------
   float_test__k_denominator : constant float := [8388608.0*2**(-11)];
   float_test__k_pi : constant float := [13176795.0*2**(-22)];

   function float_test__div_test (value : float) return float is
   begin
      return (value / float_test__k_denominator);
   end float_test__div_test;

   function float_test__perimeter_test (value : float) return float is
   begin
      return (value * float_test__k_pi);
   end float_test__perimeter_test;
-------------

While in the "original" tree dump "float_test.adb.003t.original", they are not
(and have the values that are eventually found in the machine code) :

-------------
Float_Test.Div_Test (const float value)
{
  return (float) value / 0.0;


Float_Test.Perimeter_Test (const float value)
{
  return (float) value * 9.677410125732421875e-4;
-------------

The encoding was correct with GCC 4.5

[Bug ada/59772] [4.7/4.8/4.9 regression ] floating-point constants are not correctly encoded

[ebotcazou at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59772

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Target|                            |avr-*-*
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2014-01-12
                 CC|                            |ebotcazou at gcc dot gnu.org
            Version|unknown                     |4.7.3
   Target Milestone|---                         |4.7.4
            Summary|Floating point constants    |[4.7/4.8/4.9 regression ]
                   |are not correctly encoded   |floating-point constants
                   |on AVR target               |are not correctly encoded
     Ever confirmed|0                           |1

--- Comment #1 from Eric Botcazou <ebotcazou at gcc dot gnu.org> ---
Note that there is no official GNAT port for the AVR in the FSF tree, but we
will nevertheless fix the bug.

[Bug ada/59772] [4.7/4.8/4.9 regression] floating-point constants are not correctly encoded

[ebotcazou at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59772

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |ASSIGNED
                 CC|ebotcazou at gcc dot gnu.org       |
           Assignee|unassigned at gcc dot gnu.org      |ebotcazou at gcc dot gnu.org
            Summary|[4.7/4.8/4.9 regression ]   |[4.7/4.8/4.9 regression]
                   |floating-point constants    |floating-point constants
                   |are not correctly encoded   |are not correctly encoded

--- Comment #2 from Eric Botcazou <ebotcazou at gcc dot gnu.org> ---
Fixing.

[Bug ada/59772] [4.7/4.8/4.9 regression] floating-point constants are not correctly encoded

[ebotcazou at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59772

--- Comment #3 from Eric Botcazou <ebotcazou at gcc dot gnu.org> ---
Author: ebotcazou
Date: Sun Jan 12 14:29:12 2014
New Revision: 206565

URL: http://gcc.gnu.org/viewcvs?rev=206565&root=gcc&view=rev
Log:
    PR ada/59772
    * gcc-interface/cuintp.c (build_cst_from_int): Use 32-bit integer type
    as intermediate type.
    (UI_To_gnu): Likewise.

Modified:
    trunk/gcc/ada/ChangeLog
    trunk/gcc/ada/gcc-interface/cuintp.c

[Bug ada/59772] [4.7/4.8/4.9 regression] floating-point constants are not correctly encoded

[ebotcazou at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59772

--- Comment #4 from Eric Botcazou <ebotcazou at gcc dot gnu.org> ---
Author: ebotcazou
Date: Sun Jan 12 14:29:44 2014
New Revision: 206566

URL: http://gcc.gnu.org/viewcvs?rev=206566&root=gcc&view=rev
Log:
    PR ada/59772
    * gcc-interface/cuintp.c (build_cst_from_int): Use 32-bit integer type
    as intermediate type.
    (UI_To_gnu): Likewise.

Modified:
    branches/gcc-4_8-branch/gcc/ada/ChangeLog
    branches/gcc-4_8-branch/gcc/ada/gcc-interface/cuintp.c

[Bug ada/59772] [4.7/4.8/4.9 regression] floating-point constants are not correctly encoded

[ebotcazou at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59772

--- Comment #5 from Eric Botcazou <ebotcazou at gcc dot gnu.org> ---
Author: ebotcazou
Date: Sun Jan 12 14:30:19 2014
New Revision: 206567

URL: http://gcc.gnu.org/viewcvs?rev=206567&root=gcc&view=rev
Log:
    PR ada/59772
    * gcc-interface/cuintp.c (build_cst_from_int): Use 32-bit integer type
    as intermediate type.
    (UI_To_gnu): Likewise.

Modified:
    branches/gcc-4_7-branch/gcc/ada/ChangeLog
    branches/gcc-4_7-branch/gcc/ada/gcc-interface/cuintp.c

[Bug ada/59772] [4.7/4.8/4.9 regression] floating-point constants are not correctly encoded

[ebotcazou at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59772

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |RESOLVED
         Resolution|---                         |FIXED

--- Comment #6 from Eric Botcazou <ebotcazou at gcc dot gnu.org> ---
.