458077 - Port nanojit to PowerPC

Status: VERIFIED FIXED

(Tamarin Graveyard :: Virtual Machine, defect)

Description

[Edwin Smith]

first priority is 32bit mac osx

Comment 1

[Vladimir Rainish]

Attachment: Machine dependent content has been removed from Assembler.cpp

Machine dependent content has been moved out of Assembler.cpp to Nativei386.cpp and NativeThumb.cpp to simplify future ports to Power PC and additional platform

Comment 2

[Edwin Smith]

Since there are a lot of patches being developed in parallel for nanojit, it would be better to not do any unnecessary reformatting until we reach a stable point.  Assembler.cpp in particular has lots of diff blocks that are just spacing and indentation.  lets revert those changes.

Once the major patches are all done and landed we can do a single reformat patch.

I think printActivationState() should remain cpu-independent, its not clear why ARM needs to be ifdef'd in there at all?  is it just a bug being worked around?

Lastly, and i hate to inflict this, but we should rebase this work so it's based on a recent tamarin-redux stable point, instead of tamarin-central.  (tamarin-redux contains all of what is in tamarin-central so it should just be a forward merge).  the latest tamarin-redux added a separate NativeARM.cpp to keep ARM stuff separate from Thumb.

Comment 3

[Rick Reitmaier]

Attachment: ver1

reapply vlad's changes to latest redux and minimize the diffs.

Tested and running on mac only.

Comment 4

[Edwin Smith]

Comment on attachment 342129 [details] [diff] [review]
ver1

watch out for unintended xcode project file changes, that compatibilityVersion attribute keeps coming and going...

Comment 5

[Rick Reitmaier]

compiles for winmobile and x64, but haven't run tests against it.

Comment 6

[Rick Reitmaier]

pushed...and whoops forgot to exclude the xcode project file.  And yes its annoying.

Comment 7

[Edwin Smith]

Looks like you broke the ARM build.

Comment 8

[Rick Reitmaier]

Sorry to cause pain didn't recognize that autobuilds for ARM weren't including NJ.

Comment 9

[Edwin Smith]

partly working prototype can be found in 
http://hg.mozilla.org/users/edwsmith_adobe.com/ppcjit

todo:

* ppc64
* saved fpu registers (F14-F31)
* use multi-load/store for saved regs
* patching far branches >24bit
* use float multiply-add instructions (even without intermediate rounding?)
* avoid condition register contention with cror
* save/restore CR2-4? (currently not used)
* use dedicated stack area for saved regs?
* use R2 (TOC) as scratch on mac? which platforms dedicate TOC?
* debug the test suite & esc
* LIR_x,xt,xf,loop, frag patching
* fold addressing modes
* i2d & u2d helper functions? is alignment okay with "lfd f, 12(sp)"?

Comment 10

[Edwin Smith]

just hit 0 failures in test/acceptance

performance on dual cpu 2.7Ghz G5 powermac:

avm: ../../../redux/objdir-r/shell/avmshell -Ojit     (MIR)
avm2: ../../../ppcjit/objdir-r/shell/avmshell -Ojit   (nanojit)
iterations: 1

test                                                   MIR    LIR      %sp  metric

misc/boids.as                                         7552    6578    14.8    time
misc/boidshack.as                                     1911    1765     8.3    time
misc/gameoflife.as                                   11718   10747     9.0    time
misc/primes.as                                       52341   48318     8.3    time
scimark/FFT.as                                        9122    7514    21.4    time
scimark/LU.as                                        10003    7278    37.4    time
scimark/MonteCarlo.as                                11143    6669    67.1    time
scimark/SOR.as                                        6621    6018    10.0    time
scimark/SparseCompRow.as                              4378    4396    -0.4    time

Comment 11

[Edwin Smith]

Attachment: cleanup patch to prepare for adding ppc support

This just cleans up some code to better support big endian and 64bit RegisterMask, but does not add PPC support.

The big-endian fixes might be useful for the Sparc port too.

Comment 12

[Edwin Smith]

Attachment: nanojit backend for 32bit PowerPC

needs more range limit testing, and ignore the jitcount/jitmax debug cruft.  (see above todo list).

Comment 13

[Rick Reitmaier]

Comment on attachment 350889 [details] [diff] [review]
nanojit backend for 32bit PowerPC

Looks good so far.  There are a few TODO() calls still in there which I suspect need work.  Also I'm wondering if we can use the API porting layer all the time; i.e. get rid of icache_flush

Comment 14

[Edwin Smith]

yeah, the icache_flush code needs to be cleaned up, and also the code to mark memory as executable also needs cleanup.  we can model a NJ porting layer the same way as we're doing with VMPI.  (NJPI?)

Comment 15

[Lars T Hansen]

(In reply to comment #14)
> yeah, the icache_flush code needs to be cleaned up, and also the code to mark
> memory as executable also needs cleanup.  we can model a NJ porting layer the
> same way as we're doing with VMPI.  (NJPI?)

Configurability & portability are already very low, please stick to a single platform layer :-)

Comment 16

[Edwin Smith]

Comment on attachment 350886 [details] [diff] [review]
cleanup patch to prepare for adding ppc support

pushed to tamarin-redux as 1180:988439d954d9

Comment 17

[Edwin Smith]

Attachment: hopefully final patch, only TODOs are trace specific, PEDANTIC is happy.

This should be final (sandbox build running now).

since the last patch:

* added & debugged PEDANTIC mode
* fixed remaining TODOs that apply to tamarin
* added cmpwi and cmplwi optimization
* cleaned up nPatchBranch

performance vs MIR

avm: ../../../redux/objdir-r/shell/avmshell -Ojit 
avm2: ../../../ppcjit/objdir-r/shell/avmshell -Ojit
iterations: 1

test                                                   avm    avm2     %sp  metric

misc/boids.as                                         7764    7378     5.2    misc/boidshack.as                                     2062    1884     9.4
misc/gameoflife.as                                   12594   10706    17.6
misc/primes.as                                       54065   46614    16.0
scimark/FFT.as                                        9146    8228    11.2
scimark/LU.as                                        10059    7997    25.8
scimark/MonteCarlo.as                                11192    6887    62.5
scimark/SOR.as                                        6550    6610    -0.9
scimark/SparseCompRow.as                              4447    4507    -1.3

Comment 18

[Rick Reitmaier]

Comment on attachment 351430 [details] [diff] [review]
hopefully final patch, only TODOs are trace specific, PEDANTIC is happy.

Marking + for review of all non-ppc changes, since I don't want to hold up the submit.  

Will eye-ball the ppc files later.

Comment 19

[Edwin Smith]

pushed to tamarin-redux as 1198:63bf07dcdb6d

Will leave this bug open pending final review

Comment 20

[Edwin Smith]

changes are now in tamarin-redux.  deleted working repository at http://hg.mozilla.org/users/edwsmith_adobe.com/ppcjit

Comment 21

[Cristiano "Zeus" Contin]

any chance to porting to mozilla-central?

Comment 22

[Rick Reitmaier]

> any chance to porting to mozilla-central?

If you indirectly mean SpiderMonkey then yes, 
I imagine the code will eventually land in that 
repo.  BTW, if you like ppc have a look at 
bug #468445