Mailing Lists: Apple Mailing Lists
Image of Mac OS face in stamp
 
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: libgmalloc problem

On Nov 1, 2007, at 1:48 PM, Terry Lambert wrote:

On Nov 1, 2007, at 5:16 AM, Martin Mitas wrote: 
Hi all.

Hopefully this is the right list to report this problem.

To debug our SW, we try to use libgmalloc, as described here:
http://developer.apple.com/technotes/tn2004/tn2124.html
However we encountered troubles with libgmalloc. The following
simple program is good enough to demonstrate it.


[ ... ]

$ gcc -O2 -fomit-frame-pointer problem.c
$ DYLD_INSERT_LIBRARIES=/usr/lib/libgmalloc.dylib ./a.out foo bar
Allocations will be placed on word (4 byte) boundaries.
- Small buffer overruns may not be noticed.
- Applications using AltiVec instructions may fail.
GuardMalloc-11
Segmentation fault (core dumped)

GNU Malloc's alignment boundaries are insufficiently stringent for floating pint register spills, which happen when autovectorization happens (which is enable in the compiler by default), or when explicitly using floating point.

This isn't GNU malloc; it's Guard Malloc. But if the alignment issue mentioned above is the issue here (which is quite rare with gmalloc in my experience, since data types requiring 16-byte alignment are usually allocated in multiples-of-16-bytes chunks, which results in them being properly aligned), you can work around it by setting the MALLOC_VECTOR_SIZE environment variable to 1. See the libgmalloc man page for details.

GCC on Mac OS X doesn't autovectorize by default, by the way. Terry's probably thinking of how we use the XMM registers for floating point operations by default, which means that we use the SSE instructions. But we use scalar versions of those instructions, and the compiler doesn't automatically vectorize code patterns unless you explicitly tell it to.

That said, this crash doesn't have anything to do with floating point, and it doesn't appear to have anything to do with misalignment. It doesn't crash on Leopard, but on 10.4.10, it crashes while dereferencing $eax, when $eax is "argv". Debugging it a bit and staring at the gmalloc source code for 10.4.10, I'm guessing that gcc's __builtin_frame_address() doesn't work very well when building with -fomit-frame-pointer. You can avoid this code by setting the undocumented MALLOC_NO_BACKTRACE environment variable to 1. When I do that, your code runs to completion.

Hope this helps,
Eric

_______________________________________________
Do not post admin requests to the list. They will be ignored.
Darwin-dev mailing list      (email@hidden)
Help/Unsubscribe/Update your Subscription:
http://lists.apple.com/mailman/options/darwin-dev/email@hidden

This email sent to email@hidden
References: 
 >libgmalloc problem (From: "Martin Mitas" <email@hidden>)
 >Re: libgmalloc problem (From: Terry Lambert <email@hidden>)



Visit the Apple Store online or at retail locations.
1-800-MY-APPLE

Contact Apple | Terms of Use | Privacy Policy

Copyright © 2007 Apple Inc. All rights reserved.