free() hangs when passed a valid pointer created by posix_memalign() - gcc and C++11

Question

I need to allocate 16-byte-aligned memory, and from what I can tell the accepted method to do so is posix_memalign(), using the man page as reference (other StackOverflow questions indicated this to be so). The code below is simplified to exclude unrelated parts (ie other platforms), but I have kept some context (sse is just a namespace):

#include <malloc.h>
#include <stdlib.h>
float* sse::alloc(unsigned int count)
{
    void* p;
    int r = posix_memalign(&p,16,sizeof(float)*count);
    if ( r == 0 )
        return (float*)p;
    /* else output error and exit(1) - has never failed */
    else exit(1);
}

void sse::free(float* p)
{
    free(p);
}

The code that uses it is pretty self-explanatory:

int main(int argc, char* argv[])
{
    const unsigned int total = 16000;
    float *array = sse::alloc(total), *arr2 = sse::alloc(total);
    /* null ptr checks */
    // ...
    sse::free(array); sse::free(arr2);
    return 0;
}

I have commented out all non-essential code to test this, and have confirmed that it does indeed 'hang' on free(p); - when I Ctrl-C in gdb it reports the line. The behaviour is no different in valgrind, and the SSE code (using Intel intrinsics) (in place of ...) runs successfully. I have been compiling with fairly standard options: -g -O3 -std=c++11, and have tried no/less optimisation, no debug, and some unnecessary casts. Some information about my system (please ask if you would like more):

uname -a: Linux (name) 3.12.3-1-ARCH #1 SMP PREEMPT (date) x86_64 GNU/Linux
g++ --version: 4.8.2
gdb --version: 7.6.1
valgrind --version: 3.9.0

Since the man pages clearly state that free() is the correct function, I am incredibly stumped, and I would prefer to avoid writing a mechanism to use new/delete and padding by 15 bytes (for obvious reasons). If there is an alternative that I am unaware of, I am happy to try that. Also, information about potential causes of such a hang could prove useful, as it is particularly difficult to search for some of these terms (still easier than searching for 'stack overflow').

Can you post the minimal code? Your example code is missing the return from `alloc`, among other things. — David Schwartz, Jan 13 '14 at 23:26
@Neofish: But there's no `main`, so it does nothing. Post your [testcase](http://sscce.org). If you don't have one yet, you need to do more debugging on your own before resorting to asking here. — Lightness Races in Orbit, Jan 13 '14 at 23:40
(Though it would appear that you have stumbled upon a genuine mystery, so I'd be interested in the answer, when we get there.) — Lightness Races in Orbit, Jan 13 '14 at 23:44
Use _mm_malloc http://stackoverflow.com/questions/16376942/best-cross-platform-method-to-get-aligned-memory — Z boson, Jan 14 '14 at 14:35

score 13 · Accepted Answer · answered Jan 13 '14 at 23:45

13

you need to call the global free - change the free routine to:

void sse::free(float* p)
{
    ::free(p);
}

namespace rules cause it to call sse::free unless otherwise told

answered Jan 13 '14 at 23:45

Anya Shenanigans

91,618
3
107
122

Ouch, that has to sting. – Casey Jan 13 '14 at 23:47
Omg I can't believe I missed that.. I'm so ashamed. Thanks. – Neofish Jan 13 '14 at 23:47
7

@Neofish If you'd compiled *without* optimizations, the stack overflow would have been more evident; however, the compiler decided to helpfully "optimize" the infinite recursion into an infinite loop instead. – Adam Rosenfield Jan 13 '14 at 23:51
@Adam Useful to know, although it seems to make that optimisation anyway (might be my CFLAGS/etc). Can readily make it overflow by changing to a pointer-reference, or better, a C-style cast to void* raises a compile-time type error, which would make it a bit more clear. – Neofish Jan 14 '14 at 00:01

free() hangs when passed a valid pointer created by posix_memalign() - gcc and C++11

1 Answers1