[gcc/unix] Export WTF::fastMalloc, WTF::fastFree, WTF::fastCalloc, WTF::fastRealloc

[Holger Freyther]

All non darwin gcc targets build JSC with -fvisibility=hidden hiding the FastMalloc functions. For malloc analysis it is handy to be able to preload an instrumented version of the above functions. 

The proposed change is to use __attribute__((visibility("default"))) on GCC systems.

[Holger Freyther]

Created attachment 41859 [details]
First attempt to export these symbols

Use __attribute__((visibility("default"))) to export the allocation functions again. On Mac these symbols are exported via the .exp file and available to the WebCore/WebKit library.

[Holger Freyther]

Created attachment 41865 [details]
Export the symbols in libwebkit.so as well

Export the FastMalloc functions in the .so file as well. This is the last step to be able to instrument a TCmalloc enabled libwebkit with memprof.

[Darin Adler]

Comment on attachment 41859 [details]
First attempt to export these symbols

> +#if COMPILER(GCC) && !PLATFORM(DARWIN)
> +#define WTF_FAST_MALLOC_EXPORT __attribute__((visibility("default")))
> +#else
> +#define WTF_FAST_MALLOC_EXPORT
> +#endif

Why ~PLATFORM(DARWIN)"?

If you are using __attribute__ then you should also use __visibility__.

Is it really helpful to do this just for these functions? Doesn't this need to be done consistently across the rest of JavaScriptCore?

[Darin Adler]

Oh, I see now. You're saying this is good "for malloc analysis".

Still, why the !PLATFORM(DARWIN) part? Would this change cause a problem on that platform?

[Holger Freyther]

(In reply to comment #4)
> Oh, I see now. You're saying this is good "for malloc analysis".

Ironically I'm working/using a tool that went from Corel -> Eazel -> GNOME into my hands. I'm using LD_PRELOAD to load my version of malloc and this fails with a release build of WebKit. One option would have been to instrument the mmap/sbrk calls but then I just see calls to TCSystem_Alloc.

> Still, why the !PLATFORM(DARWIN) part? Would this change cause a problem on
> that platform?

My assumption is that it is going to work as visibility should be okay with any gcc4.x release. On a mac build these symbols should have "default" visibility and saying they are default should be harmles. I just tried to minimize the risk of breaking a build on a file that is included by almost everything.

[Holger Freyther]

I'm going to build on Leopard tomorrow and will send an updated patch without the !PLATFORM(DARWIN)

[Holger Freyther]

Created attachment 41945 [details]
Export fastMalloc, fastFree, fastRealloc, fastCalloc with normal linkage

The leopard build was fine, removing the !PLATFORM(DARWIN).

[Darin Adler]

Comment on attachment 41865 [details]
Export the symbols in libwebkit.so as well

rs=me

[Holger Freyther]

Landed in r50204 and r50205.