Created attachment 1885 [details]
New PowerPC JIT compiler

Created attachment 1886 [details]
Makefile patch

patch for Makefile on SVN trunk

Created attachment 1887 [details]
New PowerPC JIT compiler, v2

use mmap + mprotect: works on linux @ G5 (32bit userspace)

Not to be a pain, but I don't think we should be mingling GPL3 stuff in here, unless we're going to move the whole codebase to GPL3.

It's just a mess to have to say "ioq3 is GPL2, unless you use THIS ONE PLATFORM, and just in case, you might want to delete that file anyhow so there's no doubt."

This is the diff of that GPL3 file between now and when they moved from GPL2...

http://sourceware.org/cgi-bin/cvsweb.cgi/src/opcodes/ppc-opc.c.diff?r1=1.96&r2=1.110&cvsroot=src

I don't have the ability to compile this right now to test...how dependent are we on the newer changes in that file?

--ryan.

After finishing vm_powerpc.c (some more optimizations, check OSX) I was planning to strip instruction list to minimum (leaving only the used instructions). I can do it based on the older ppc-opc.c without any problem.

Cool...I'll try this on a 64-bit G5 when I get a chance.   :)

--ryan.

Created attachment 1888 [details]
New PowerPC JIT compiler, v3

added some optimizations, most important one: OP_CONST+OP_CALL fusion

(In reply to comment #6)
> Cool...I'll try this on a 64-bit G5 when I get a chance.   :)

Trying won't help much, ppc64 has completely different ELF ABI and parts of compiler will have to be rewritten. Some changes are simple to implement (like link register placement), but other things I don't understand yet (special function preamble). I want the code ABI conforming so it won't break each time something in GCC changes, like with the original vm_ppc.c happened.

If anyone could give me (temporal) access to ppc64 machine, with 64-bit userspace, it would be great.

Created attachment 1890 [details]
New PowerPC JIT compiler, v4

More speed optimizations: uses registers instead of opstack. vm_powerpc.c cleaned up. vm_powerpc_asm.c unchanged. There will be no more speed improvements, only fixes and, hopefully, ppc64 support.

Created attachment 1892 [details]
New PowerPC JIT compiler, v5

vm_powerpc.c: Includes ppc64 support ! (only dedicated server tested on linux @ XServe G5). Changed stack behavior on OSX, as it is not System V conforming.
vm_powerpc_asm_c: Now it is based on revision 1.96 of ppc-opc.c - it is GPL v2. Stripped opcode list to minimum.

Created attachment 1900 [details]
New PowerPC JIT compiler, v6

vm_powerpc.c: Fixed for OS X, works, and works damn well (tested on Panther @ G4). OP_BLOCK_COPY now uses double words on ppc64.

Created attachment 1902 [details]
Makefile patch, v2

includes ppc64 changes

Created attachment 1913 [details]
New PowerPC JIT compiler, v7

vm_powerpc_asm.c: stripped almost all the unused code (there are some #defines left, but this does not increase binary size)
vm_powerpc.c: added/fixed some comments, lots of cosmetics, few speedups, but no functional changes

PING !

any ACKs/NACKs on the patch?

There's a thread on the lists, but only one person was kind enough to test it:

http://lists.ioquake.org/pipermail/ioquake3-ioquake.org/2008-October/003169.html

It seems like the patch won't be properly tested until it gets to the trunk. You may still leave vm_ppc.c temporarily so if the new one isn't working for someone, she should be able to switch to the old compiler without any problem.

I tried to ask few developers on IRC, but all of them have different reasons not to test it (-ENOTIME, borked HD, something else).

Looks good to me. Analize is spelled incorrectly.

Created attachment 1914 [details]
New PowerPC JIT compiler, v8

Fixed "Analyze" typo. Added missing "static". 

I'm sure there are more typos, I just can't see them right now.

PING ! Just wanted to remind about the patch.
Maybe someone will finally find the courage to commit it :-P

Current status of the patch:
- includes all the optimizations found in vm_ppc.c and some more
- prepared for both 32 and 64 bit PowerPCs
- supports ELF (Linux) and OSX ABIs, other ABIs (AIX ?) may be added easily if needed
- while playing, game spends between 13% and 22% of time inside generated code, with interpreter it is over 60%
- fast compilation: under 0.4 seconds for qagame on 1.3GHz G4

Tested and working on:
- OS X 10.3 (Panther) @ iBook G4, gcc 3.3 [ioq3, tremulous]
- Linux PLD @ iBook G4, gcc 4.3.2 [ioq3 + various mods, tremulous]
- Linux debian 32bit @ XServe G5, gcc 4.2 [ioq3ded, tremded]
- Linux debian 64bit @ XServe G5, gcc 4.2 [ioq3ded]

Created attachment 1918 [details]
New PowerPC JIT compiler, v9

fix: realloc vm->instructionPointers on ppc64
minor: rename VM_Malloc, VM_Free to PPC_Malloc, PPC_Free; safer constant optimization handling

applied finally :)

cool, thanks !