Bug 3380 - Dynamically generate GL dispatch functions for PowerPC
Summary: Dynamically generate GL dispatch functions for PowerPC
Status: NEW
Alias: None
Product: Mesa
Classification: Unclassified
Component: GLX (show other bugs)
Version: 5.0.2
Hardware: Other Linux (All)
: high enhancement
Assignee: Default DRI bug account
QA Contact:
URL:
Whiteboard:
Keywords:
Depends on:
Blocks:
 
Reported: 2005-05-23 13:57 UTC by Ian Romanick
Modified: 2013-03-15 14:21 UTC (History)
3 users (show)

See Also:
i915 platform:
i915 features:


Attachments
early version of the generating script (8.24 KB, application/x-python)
2005-12-05 17:29 UTC, Zack Rusin
Details
script generating dispatch code for ppc32 (7.62 KB, application/x-python)
2005-12-07 15:55 UTC, Zack Rusin
Details
Mesa patch to use generated PPC assembly (2.67 KB, patch)
2005-12-08 03:34 UTC, Michel Dänzer
Details | Splinter Review

Note You need to log in before you can comment on or make changes to this bug.
Description Ian Romanick 2005-05-23 13:57:43 UTC
Dynamic generation of GL dispatch functions is currently only supported on x86
and SPARC.  PowerPC and PPC64 should also be supported.
Comment 1 Zack Rusin 2005-12-05 17:29:35 UTC
Created attachment 3995 [details]
early version of the generating script

I had a minute right now so i did the generating script. I'll try to integrate
it, add ppc64 support and in general just clean it up some time this week when
I'll have a little bit more time. I'm attaching in case someone has more spare
time than I do. Ah, and the offset computation in BRANCH_OFFSET is more
pseudo-code than assembly right now ;)
Comment 2 Zack Rusin 2005-12-07 15:55:26 UTC
Created attachment 4014 [details]
script generating dispatch code for ppc32 

k, as far as ppc32 generation, this should be it.
Comment 3 Michel Dänzer 2005-12-07 20:08:16 UTC
I managed to build libGL with this, but when running glxinfo, it dies with
SIGILL in glGetString (the first gl[^X] function called).
Comment 4 Michel Dänzer 2005-12-08 03:34:03 UTC
Created attachment 4026 [details] [review]
Mesa patch to use generated PPC assembly

Testing instructions:

* Save the current version of the generator script as
Mesa/src/mesa/glapi/gl_ppc_asm.py
* Apply the Mesa patch
* make -C Mesa/src/mesa/glapi
* Rebuild libGL (you may have to manually remove at least
Mesa/src/mesa/main/dispatch.o) and watch glxinfo die with SIGILL ;)
Comment 5 Benjamin Herrenschmidt 2006-03-01 17:05:40 UTC
Ok, I suppose I can fix that :) But first I need to understand how the whole 
thing is supposed to work... It's not very clear between python stuff 
generating asm stubs, magic macros, etc etc... (btw. I think there are bugs 
there but I'll have to dbl check when I understand what the stuff is actually 
supposed to do in the first place). 
 
Can somebody explain me how it works without the asm stuff (the pure C stuff) 
and what the "optimized" version is supposed to do ? I have enabled 
-DGLX_USE_TLS and the resulting stubs look not too bad, but I could probably 
make them a bit better with asm (like doing a proper tail call). 
 
However, I suppose that if I could acutally generate code at runtime, I could 
do even better and generating a branch island that loads the absolute address 
of the target and branch to it (basically about 4 or 5 insns in 32 bits, maybe 
a bit more in 64 bits). But first I need to get a bit more familiar with how 
this all fits together in the mesa source, so far I'm just confused :) 
 
Comment 6 Marcin Kurek 2006-10-06 05:32:00 UTC
Added myself to cc as this one looks promising.
Comment 7 Benjamin Herrenschmidt 2006-10-06 14:45:56 UTC
Sorry, I didn't go back on that yet... I read the code for about 2 hours and
just got more confused about what is it supposed to do. Some comments would have
been useful. I might get back to it at one point, I'm a bit too busy at the moment
Comment 8 chemtech 2013-03-15 14:21:56 UTC
Ian Romanick 
Do you still experience this issue with newer soft ?
Please check the status of your issue.


Use of freedesktop.org services, including Bugzilla, is subject to our Code of Conduct. How we collect and use information is described in our Privacy Policy.