llvm dev - Sep 2013 - [LLVMdev] Disabling special treatment of "malloc" function

Hi,

I'm using uclibc built with dragonegg-3.2 / gcc-4.6, and ran into a 
problem in which:

* uclibc's realloc malloc()'s some memory
* realloc then uses (malloc'd pointer) - some offset to find the true 
size allocated.
* MemoryDependenceAnalysis (MDA) regards load from (malloc() call + any 
offset) to be undefined, and replaces the size read with 0.
* All manner of chaos results when that value is used.

Obviously the problem is MDA assumes that load from malloc without 
intervening store is undefined, but that assumption is unreasonable when 
you're building parts of the allocator itself. It looks like LLVM makes 
a little effort to spot that situation (it won't enforce malloc's 
contract if the malloc function has a definition), but from the point of 
view of building the realloc call malloc is an external declaration just 
like for a program using libc.

For now I've added a flag to MDA such that malloc clobbers rather than 
resolving to undef; however I was wondering if there is a "proper" way
to do this with any compiler targeting LLVM?

If it's actually not solved already, I'd suggest that instead of 
assuming "derived from malloc -> undef" it could be better to leave
such
loads alone unless they are provably within the bounds of the allocation 
- so malloc(n) yields n bytes of undef, but clobbers out of bounds 
pointers derived from the same malloc.

Chris

Hi Chris,
> For now I've added a flag to MDA such that malloc clobbers rather than
> resolving to undef; however I was wondering if there is a
"proper" way to do
> this with any compiler targeting LLVM?
Clang has an option "-fno-builtin" that disables such assumptions (by
adding the attribute "nobuiltin" to the relevant callsites by the
looks of it). Is that useful to you?

Cheers.

Tim.

On 08/09/13 16:36, Tim Northover wrote:
>
>> For now I've added a flag to MDA such that malloc clobbers rather
than
>> resolving to undef; however I was wondering if there is a
"proper" way to do
>> this with any compiler targeting LLVM?
> Clang has an option "-fno-builtin" that disables such assumptions
(by
> adding the attribute "nobuiltin" to the relevant callsites by the
> looks of it). Is that useful to you?
That works; unfortunately I was using Dragonegg and would rather not 
switch at this point.

For the benefit of Googlers, it appears that Clang -fno-builtin leads to 
a call to TargetLibraryInfo::disableAllFunctions, but Dragonegg-3.2 
never calls it.

However Dragonegg-3.3 /does/ honour the GCC -fno-builtin flag (diff 
here: http://llvm.org/viewvc/llvm-project?view=revision&revision=179666).

This is a little blunt for my purposes -- I only need to disable malloc 
for this purpose, but it suffices.

(Side note: I hadn't realised a Dragonegg 3.3 release existed because 
http://dragonegg.llvm.org/ is out of date)

Chris