[Bug c/54954] New: malloc optimizations not disabled by -fno-builtin

[Bug c/54954] New: malloc optimizations not disabled by -fno-builtin

[swalter at lexmark dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54954

             Bug #: 54954
           Summary: malloc optimizations not disabled by -fno-builtin
    Classification: Unclassified
           Product: gcc
           Version: 4.7.2
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: swalter@lexmark.com


Consider the code:

static int foo;

void bar ()
{
    foo = 0;
    malloc();
    assert(foo);
}

foo is not reloaded after the call to malloc, even when -fno-builtin is thrown.
 This is incorrect behavior because malloc() could call back into the current
module, where foo is modified.  A full example is attached.

(Assume that malloc calls increment_count() if I can't get a second file to
attach...)

[Bug c/54954] malloc optimizations not disabled by -fno-builtin

[swalter at lexmark dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54954

--- Comment #1 from swalter at lexmark dot com 2012-10-17 16:10:29 UTC ---
Created attachment 28462
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=28462
Bug testcase

[Bug c/54954] malloc optimizations not disabled by -fno-builtin

[swalter at lexmark dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54954

--- Comment #2 from swalter at lexmark dot com 2012-10-17 16:14:33 UTC ---
Created attachment 28463
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=28463
malloc implementation for test.c

[Bug c/54954] malloc optimizations not disabled by -fno-builtin

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54954

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |WAITING
   Last reconfirmed|                            |2012-10-18
     Ever Confirmed|0                           |1

--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> 2012-10-18 10:12:40 UTC ---
It works just fine for me.  Please make sure to really use -fno-builtin ;)

[Bug c/54954] malloc optimizations not disabled by -fno-builtin

[bren at ragh dot us]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54954

Brendan Chandler <bren at ragh dot us> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |bren at ragh dot us

--- Comment #4 from Brendan Chandler <bren at ragh dot us> 2012-10-18 11:53:01 UTC ---
Try compiling with -02 as well.

[Bug c/54954] malloc optimizations not disabled by -fno-builtin

[swalter at lexmark dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54954

--- Comment #5 from swalter at lexmark dot com 2012-10-18 13:41:13 UTC ---
Thanks for looking into this, Richard.  I should have mentioned that you'll
need to build with optimization turned on.

Expected behavior:
test.c built with -O -fno-builtin runs to successful completion

Actual behavior:
test.c built with -O -fno-builtin aborts (failed assertion)
test.c built with -O0 does not abort

[Bug c/54954] malloc optimizations not disabled by -fno-builtin

[pinskia at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54954

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

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

--- Comment #6 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
The header file malloc.h  (which is non-standard by the way) has the attribute
malloc on the malloc function call.

So this is invalid.

[Bug c/54954] malloc optimizations not disabled by -fno-builtin

[swalter at lexmark dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54954

--- Comment #7 from swalter at lexmark dot com ---
(In reply to Andrew Pinski from comment #6)
> The header file malloc.h  (which is non-standard by the way) has the
> attribute malloc on the malloc function call.
> 
> So this is invalid.

Sorry, can you spell that out for me?  You're saying that the behavior I'm
seeing is because I'm using malloc.h for the definition of malloc(), rather
than stdlib.h?  malloc.h's definition of malloc has __attribute__((malloc)), so
that allows gcc to perform the optimization even when -fno-builtin in thrown?

[Bug c/54954] malloc optimizations not disabled by -fno-builtin

[pinskia at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54954

--- Comment #8 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
(In reply to swalter from comment #7)
> (In reply to Andrew Pinski from comment #6)
> > The header file malloc.h  (which is non-standard by the way) has the
> > attribute malloc on the malloc function call.
> > 
> > So this is invalid.
> 
> Sorry, can you spell that out for me?  You're saying that the behavior I'm
> seeing is because I'm using malloc.h for the definition of malloc(), rather
> than stdlib.h?  malloc.h's definition of malloc has __attribute__((malloc)),
> so that allows gcc to perform the optimization even when -fno-builtin in
> thrown?

No I am saying the definition inside both malloc.h and stdlib.h has the
attribute on it which allows gcc to perform the optimization even when
-fno-builtin is used.  My comment about malloc.h being non-standard was an
aside comment and just pointing out the use of a non-standard header.

[Bug c/54954] malloc optimizations not disabled by -fno-builtin

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54954

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
         Resolution|INVALID                     |WORKSFORME

--- Comment #9 from Richard Biener <rguenth at gcc dot gnu.org> ---
I said it works for me because it works with -fno-builtin-malloc.

Andrew, you are wrong - the malloc attribute itself does not tell us
that the malloc call may not clobber global memory.

I've massaged the testcase to build w/o warnings:

m.c:
#include <string.h>
#include <stdio.h>

void *__libc_malloc(size_t size);

void *malloc(size_t size)
{
    increment_count();
    return (void*)__libc_malloc(size);
}

t.c:
#include <assert.h>
#include <stdlib.h>
#include <stdio.h>

static int count = 0;

void increment_count(void)
{
        count++;
}

int main (void)
{
        char *ptr;
        count = 0;
        ptr = malloc(1);
        assert(ptr);
        assert(count);
        printf("%p\n", ptr);
        free(ptr);
        return 0;
}

rguenther@murzim:/tmp> /space/rguenther/install/gcc-4.7.0/bin/gcc -o t t.c m.c
-O2 -fno-builtin-malloc
rguenther@murzim:/tmp> ./t
0x1127010
rguenther@murzim:/tmp> /space/rguenther/install/gcc-4.7.0/bin/gcc -o t t.c m.c
-O2  
rguenther@murzim:/tmp> ./t
t: t.c:18: main: Assertion `count' failed.
Aborted


Works with all GCC versions I tried.  I tried on x86_64-linux (just in case
RTL optimizers break this in which case the target architecture may be
important).