Re: [RFA/commit] (SPARC/LEON) fix incorrect array return value printed by "finish"

[vladimir dot mezentsev at oracle dot com]

On 04/23/2018 10:45 AM, Joel Brobecker wrote:
> Consider the code in the gdb.ada/array_return.exp testcase, which
> defines a function returning an array of 2 integers:
>
>    type Data_Small is array (1 .. 2) of Integer;
>    function Create_Small return Data_Small;
>
> When doing a "finish" from inside function Create_Small, we expect
> GDB to tell us that the return value was "(1, 1)". However, it currently
> prints the wrong value:
>
>     (gdb) finish
>     Run till exit from #0  pck.create_small () at /[...]/pck.adb:5
>     p () at /[...]/p.adb:10
>     10         Large := Create_Large;
>     Value returned is $1 = (0, 0)

 I never used  ada and it looks like a bug in ada compiler not in gdb.
Probably ada generates incorrect code for function which returns a small
array.

The similar c  test works:
% cat r.c
#include <stdio.h>

typedef int __attribute__ ((vector_size (2 * sizeof(int)))) I2;

I2 func() {
    I2 a;
    a[0] = 11;
    a[1] = 12;
    return a;
}

int main() {
    I2 x = func();
    printf("%d %d\n", x[0], x[1]);
    return 0;
}

% gcc -g r.c
% ./a.out
11 12

% /export/home/vmezents/git/obj_binutils-gdb/gdb/gdb ./a.out
GNU gdb (GDB) 8.0.50.20170512-git
Copyright (C) 2017 Free Software Foundation, Inc.
License GPLv3+: GNU GPL version 3 or later
<http://gnu.org/licenses/gpl.html>
This is free software: you are free to change and redistribute it.
There is NO WARRANTY, to the extent permitted by law.  Type "show copying"
and "show warranty" for details.
This GDB was configured as "sparc64-unknown-linux-gnu".
Type "show configuration" for configuration details.
For bug reporting instructions, please see:
<http://www.gnu.org/software/gdb/bugs/>.
Find the GDB manual and other documentation resources online at:
<http://www.gnu.org/software/gdb/documentation/>.
For help, type "help".
Type "apropos word" to search for commands related to "word"...
Resuming the execution of threads of all processes is on.
Reading symbols from ./a.out...done.
(gdb) b main
Breakpoint 1 at 0x1006a8: file r.c, line 13.
(gdb) run
Starting program: /export/home/vmezents/BUGs/25502809/a.out

Breakpoint 1, main () at r.c:13
13        I2 x = func();
(gdb) print func()
$1 = {11, 12}


See also  my comment below.

> This is a regression which I traced back to the following commit...
>
>     | commit 1933fd8ee01ad2e74a9c6341bc40f54962a8f889
>     | Date:   Fri May 19 03:06:19 2017 -0700
>     | Subject: gdb: fix TYPE_CODE_ARRAY handling in sparc targets
>
> ... which, despite what the subject says, is not really about
> TYPE_CODE_ARRAY handling, which is a bit of an implementation detail,
> but about the GNU vectors extension.
>
> The author of the patch equated TYPE_CODE_ARRAY with vectors, which
> is not correct. Vectors are TYPE_CODE_ARRAY types with the TYPE_VECTOR
> flag set. So at the very minimum, the patch should have been checking
> for both TYPE_CODE_ARRAY and TYPE_VECTOR.
>
> But, that's not the only thing that did not seem right to me. When
> looking at the ABI, and at the summary of the implementation in GCC
> of the calling conventions for that architecture:
>
>                                 size      argument     return value
>
>       small integer              <4       int. reg.      int. reg.
>       word                        4       int. reg.      int. reg.
>       double word                 8       int. reg.      int. reg.
>
>       _Complex small integer     <8       int. reg.      int. reg.
>       _Complex word               8       int. reg.      int. reg.
>       _Complex double word       16        memory        int. reg.
>
>       vector integer            <=8       int. reg.       FP reg.
>       vector integer             >8        memory         memory
>
>       float                       4       int. reg.       FP reg.
>       double                      8       int. reg.       FP reg.
>       long double                16        memory         memory
>
>       _Complex float              8        memory         FP reg.
>       _Complex double            16        memory         FP reg.
>       _Complex long double       32        memory         FP reg.
>
>       vector float              any        memory         memory
>
>       aggregate                 any        memory         memory

 I think the 4 from last 5 are incorrect (See for example
https://www.gnu.org/software/gcc/gcc-3.4/sparc-abi.html)
It should be:

      _Complex float              8        reg         FP reg.
      _Complex double            16        reg         FP reg.
      vector float              <=16        reg         reg
      aggregate                 <=16        reg         reg



 It will be good to add comments with the correct table into 
gdb/sparc-tdep.c.


-Vladimir

>
> The nice thing about the patch above is that it nicely factorized
> the code that determines how arguments are passed/returns. The bad
> news is that the implementation, particularly for the handling of
> arrays and vectors, doesn't seem to match the summary above. Hence,
> the regression we observed.
>
> So what I did was review and re-implement some of the predicate functions
> according to the summary above. Because dejagnu crashes all our Solaris
> machines real bad, I can't run the dejagnu testsuite there. So what I did
> was test the patch with AdaCore's testsuite against leon3-elf, no
> regression. I verified that this fixes the regression above while
> at the same time still passing gdb.base/gnu_vector.exp (I transposed
> that testcase to our testsuite), which is the testcase that was cited
> in the commit above as seeing some FAIL->PASS improvements.
>
> This patch also removes one assertion...
>
>       gdb_assert (sparc_integral_or_pointer_p (type)
>                   || (TYPE_CODE (type) == TYPE_CODE_ARRAY && len <= 8));
>
> ... because that assertion is really the "negative" of the other conditions
> written in the same "if, else if, else [assert]" block in this function.
> To me, this assertion forces us to maintain two versions of the same code,
> and is an unnecessary burden. In particular, the above is not the
> correct condition, and the ABI summary table above shows that we need
> a more complex condition to describe the situations where we expect
> arguments to be passed by register.
>
> gdb/ChangeLog:
>
>         * sparc-tdep.c (sparc_structure_return_p): Re-implement to
>         match the ABI as summarized in GCC's gcc/config/sparc/sparc.c.
>         (sparc_arg_by_memory_p): Renamed from sparc_arg_on_registers_p.
>         Re-implement to match the ABI as summarized in GCC's
>         gcc/config/sparc/sparc.c.  All callers updated.
>         (sparc32_store_arguments): Remove assertion.
>
> OK to commit in master?
>
> Thanks!