Fast thread-local storage for OpenGL drivers

[Gareth Hughes]

Roland McGrath wrote:
> 
> These people clearly haven't read all of the TLS paper, or looked at the
> GCC implementation of __thread long enough to notice -ftls-model and
> __attribute__ ((tls_model)).

This is what I was talking about.  I've read the entire document several
times, and still can't see a way that a dynamically loadable shared library
can be guaranteed to use the single-instruction Local Exec access model.  If
I'm wrong, please explain why.

> I think the TLS document intends to explain what the models mean in
> practical terms on each architecture, but I can believe it's not all
> that clear.  The GCC manual doesn't explain the access models and code
> sequences, just tells you how to tell the compiler what you want in the
> terms that the TLS document defines.  
> 
> If you want maximal flexibility, i.e. to always work with dlopen, then
> indeed you must use the "dynamic" TLS access models (GD or LD).  You can
> use the Initial Exec model if you want faster accesses at the cost of some
> flexibility.

libGL.so simply has to work with dlopen -- if for no other reason than
essentially all major 3D games (Quake3, Doom3, UT2003 etc) dlopen libGL.so
rather than linking with it.  This is not going to change.

> When compiling PIC, IE-model accesses have one additional indirection,
> i.e. loading the offset from the GOT just as the address of a global
> variable is loaded in PIC.  See the instruction sequences in the TLS spec.

I'm pretty sure all implementations of OpenGL are not compiled as PIC at
this point in time.  That's a whole other discussion, however.

> If you use static linking, these instruction sequences reduce to constants
> at link time (i.e. direct "%gs:NNN" accesses on x86).

Can you describe how I could use static linking here?  As I said, libGL.so
must be a dynamically loadable shared library.  What we want is the
single-instruction Local Exec access model.  At this point in time, my
understanding of the situation is that these are mutually exclusive
requirements.

> If you link a shared object containing IE-model access relocs, the object
> will have the DF_STATIC_TLS flag set.  By the spec, this means that dlopen
> might refuse to load it.

As I said, not being able to dlopen libGL.so is unacceptable.

> In glibc, we actually allocate some excess space in the thread-local
> storage area layout determined at startup time.  This lets a dynamically
> loaded module use static TLS if its PT_TLS segment fits in the available
> surplus.  (In sysdeps/generic/dl-tls.c, see TLS_STATIC_SURPLUS.)  If there
> is insufficient space preallocated, then loading the module will fail.  In
> fact, we put this feature there with GL in mind and can adjust the
> preallocated surplus for what is most useful in practice.

I think the set of performance critical thread-local variables is something
like two or three (depending on the implementation).  The libGL.so API
dispatcher needs fast access to one or two of these (dispatch table
pointers), while the driver backend needs fast access to all of them
(context pointer and dispatch table pointers).  The other thread-local
variables are generally not accessed in performance-critical situations.

Another issue I forgot to mention, or forgot to make clear, is that we need
to be able to access these thread-local variables in runtime generated code.
A driver's top-level API functions are often generated at runtime, and need
to be able to do things like switch dispatch tables (obviously, they'd have
direct access to the context they were associated with, and so wouldn't need
to go through the pointer in TLS).  Are we guaranteed that the __thread
variables aren't going to move around?  How would we work out what code to
generate to access a given __thread variable?

(I've included both phil-list and wine-devel, if you'd like this discussion
kept to one or other of these lists, please say so).

--
Gareth Hughes (gareth at nvidia.com)
OpenGL Developer, NVIDIA Corporation