llvm dev - Jun 2013 - [LLVMdev] Telling clang/LLVM a data object is "near"

Hi,

I'm using clang for OS X development.

The data model used for x86_64 makes a distinction between near and far data...

static int x;

is assumed to be reachable within 2GB rip-relative addressing and generates code
accordingly.

extern int x;

is not assumed to be so reachable.  It generates slightly longer code that may
need to use the GOT.  The linker (either static or dynamic) has the option of
optimizing and replacing a mov instruction from GOT with a lea instruction if it
finds that the data element is accessible with a 2GB rip-relative address.
However, even with a mov->lea replacement, a longer instruction sequence is
usually needed (such as 2 instructions where one would suffice if the data is
known to be near.) + the compiler is forced to allocate a register for holding
the object's address - possibly impacting other optimizations.

The programmer may know that although the data is not static - it is still
reachable with a near address.  For instance, if it's in another source file
that is statically linked to a resulting object file with size < 2GB.


Is there a way to tell clang to assume that a piece of data is near, such as

extern __attribute((assume_near))__ int x;

and generate code accordingly (i.e. not via GOT access).
Thanks in Advance.

On Sun, Jun 23, 2013 at 10:21:12AM -0700, Ofer Porat
wrote:
> Is there a way to tell clang to assume that a piece of data is near, such
as
> 
> extern __attribute((assume_near))__ int x;
> 
> and generate code accordingly (i.e. not via GOT access).
Hidden visibility should do exactly that.

Joerg