[Bug target/49211] New: MMIX: Code generation broken, when using global variables

[Bug target/49211] New: MMIX: Code generation broken, when using global variables

[nils@script-solution.de]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49211

           Summary: MMIX: Code generation broken, when using global
                    variables
           Product: gcc
           Version: 4.6.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: nils@script-solution.de
              Host: Ubuntu on x86_64
            Target: MMIX
             Build: gcc-4.6.0 with binutils-2.21


Created attachment 24386
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=24386
Contains the object-file and the executable

I hope it is no missunderstanding, but as it seems, the code generation is
wrong when using global variables (I'm wondering because it is such a basic
thing).

For example, take the following C file:
----
struct foo {
    int x,y;
};

static struct foo f;
static int x;

static void func1(int *i) {
}
static void func2(struct foo *pf) {
}

int main(void) {
    int a;
    f.x = 0;
    f.y = 1;
    x = 2;
    func1(&x);
    func1(&a);
    func2(&f);
    return 0;
}
----

I'm using the following commands to build the executable:
$ mmix-elf-escape-gcc -o test.o -c test.c
$ mmix-elf-escape-gcc -nostdlib -o test test.o

The disassembly for the main-function is:
----
0000000000000040 <main>:
  40:    27fefe10     subu $254,$254,16
  44:    affdfe08     stou $253,$254,8
  48:    23fdfe10     addu $253,$254,16
  4c:    fe000004     get $0,rJ
  50:    e3010000     setl $1,0x0
  54:    ab01fe00     sttu $1,$254,0
  58:    e3010001     setl $1,0x1
  5c:    ab01fe04     sttu $1,$254,4
  60:    e3010002     setl $1,0x2
  64:    ab01fe08     sttu $1,$254,8
  68:    2303fe08     addu $3,$254,8
  6c:    f302ffe5     pushj $2,0 <_start.>
  70:    f6040000     put rJ,$0
  74:    2701fd0c     subu $1,$253,12
  78:    c1030100     set $3,$1
  7c:    f302ffe1     pushj $2,0 <_start.>
  80:    f6040000     put rJ,$0
  84:    2303fe00     addu $3,$254,0
  88:    f302ffe6     pushj $2,20 <func2>
  8c:    f6040000     put rJ,$0
  90:    e3000000     setl $0,0x0
  94:    8dfdfe08     ldo $253,$254,8
  98:    23fefe10     addu $254,$254,16
  9c:    f8010000     pop 1,0
----

That means, the code expects, that f is located at $254, offset 0. But $254 is
the frame-pointer according to
http://www.bitrange.com/mmix/mmixfest-2001/mmixabi.html. The example shows a
few other usages, that are wrong for the same reason. Note that it is correct
for the local variable.

For the case that it helps, I've attached the object file and the executable.

I hope that the fix is easy. It would be really nice to know shortly how to
resolve that, because before that, I think one can't do anything useful with
gcc for MMIX :/

[Bug target/49211] MMIX: Code generation broken, when using global variables

[nils@script-solution.de]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49211

Nils Asmussen <nils@script-solution.de> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Severity|normal                      |blocker

[Bug target/49211] MMIX: Code generation broken, when using global variables

[hp at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49211

Hans-Peter Nilsson <hp at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |hp at gcc dot gnu.org

--- Comment #1 from Hans-Peter Nilsson <hp at gcc dot gnu.org> 2011-05-30 14:24:52 UTC ---
With -nostdlib, you're basically disabling the startup code that is supposed to
setup everything. If it still doesn't work without -nostdlib, then I'll have a
look.

[Bug target/49211] MMIX: Code generation broken, when using global variables

[hp at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49211

Hans-Peter Nilsson <hp at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |WAITING
   Last reconfirmed|                            |2011.05.30 14:25:18
     Ever Confirmed|0                           |1

[Bug target/49211] MMIX: Code generation broken, when using global variables

[hp at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49211

--- Comment #2 from Hans-Peter Nilsson <hp at gcc dot gnu.org> 2011-05-30 14:29:32 UTC ---
See gcc/config/mmix/crtn.asm, where the global registers used by the ABI are
allocated.

[Bug target/49211] MMIX: Code generation broken, when using global variables

[nils@script-solution.de]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49211

--- Comment #3 from Nils Asmussen <nils@script-solution.de> 2011-05-30 14:47:40 UTC ---
Thanks for your reply!

Perhaps I miss something, but I've no idea how that should work. Regardless of
whether $254 is initialized previously or not (in my case, it's a bootloader,
so there is no stdlib and no crt*, but I have to do that myself), using the
stack-pointer to access global variables can't work, right? Lets suppose, that
the offset in main is correct. That would mean, that x is located at $254 + 8.
But $254 is changed (of course) at the beginning and end of a function (and
nowhere else). If func2 accesses x as well, gcc generates code to access it at
$254 + 8. So, it expects it to be always at offset 8 from $254, but of course
that can't work, because $254 changes.

[Bug target/49211] MMIX: Code generation broken, when using global variables

[hp at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49211

Hans-Peter Nilsson <hp at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|WAITING                     |RESOLVED
         Resolution|                            |INVALID

--- Comment #4 from Hans-Peter Nilsson <hp at gcc dot gnu.org> 2011-05-30 14:59:06 UTC ---
(In reply to comment #3)
> Perhaps I miss something, but I've no idea how that should work.

Writing a bootloader, you really should.

> Regardless of
> whether $254 is initialized previously or not (in my case, it's a bootloader,
> so there is no stdlib and no crt*, but I have to do that myself), using the
> stack-pointer to access global variables can't work, right?

The key words are "setup" and "allocated" not "initialized".

Again, I suggest you have a look at crtn.asm, in particular these lines:

% This must be the last file on the link-line, allocating global registers
% from the top.

% Register $254 is the stack-pointer.
sp GREG

So, if you leave that out, the global-register-allocation machinery in the
linker will allocate registers starting from the topmost, which is $254 ($255
being a scratch register).

[Bug target/49211] MMIX: Code generation broken, when using global variables

[nils@script-solution.de]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49211

--- Comment #5 from Nils Asmussen <nils@script-solution.de> 2011-05-30 15:22:37 UTC ---
> Again, I suggest you have a look at crtn.asm, in particular these lines:
> 
> % This must be the last file on the link-line, allocating global registers
> % from the top.
> 
> % Register $254 is the stack-pointer.
> sp GREG
> 
> So, if you leave that out, the global-register-allocation machinery in the
> linker will allocate registers starting from the topmost, which is $254 ($255
> being a scratch register).

Oh, ok. I thought in a different direction :/
Then it's perfectly clear. Thanks!