Re: Android performance drop moving from LuaJIT-2.0.1 -> LuaJIT-2.0.2 ?
- From: "Kriss@xxxxxxxx" <Kriss@xxxxxxxx>
- To: luajit@xxxxxxxxxxxxx
- Date: Wed, 4 Sep 2013 21:15:39 +0100
Thank you for the fast response.
By a process of just reverting files between the two versions I have
pinpointed exactly what is going wrong. Not really sure why it is bad for
me and not anyone else?
The change is in lj_mcode.c,
/* Target an address in the static assembler code (64K aligned).
** Try addresses within a distance of
target-range/2+1MB..target+range/2-1MB.
** Use half the jump range so every address in the range can reach any
other.
*/
#if LJ_TARGET_MIPS
/* Use the middle of the 256MB-aligned region. */
uintptr_t target = ((uintptr_t)(void *)lj_vm_exit_handler & 0xf0000000u) +
0x08000000u;
#else
uintptr_t target = (uintptr_t)(void *)lj_vm_exit_handler &
~(uintptr_t)0xffff;
#endif
const uintptr_t range = (1u << (LJ_TARGET_JUMPRANGE-1)) - (1u << 21);
On Wed, Sep 4, 2013 at 6:47 PM, Mike Pall <mike-1309@xxxxxxxxxx> wrote:
> Kriss@xxxxxxxx wrote:
> > I seem to be experiencing a huge lua speed drop when using LuaJIT-2.0.2
> on
> > android, code that was previously clocking in at around 5ms now take
> 100ms
> >
> > Dropping back to LuaJIT-2.0.1 fixes the problem.
> >
> > Running LuaJIT-2.0.2 with "jit.off()" fixes the problem.
>
> That's rather strange ...
>
> > So it does seem to be something to do with a change in the JIT code.
>
> Here's the only possibly relevant change between 2.0.1 and 2.0.2:
>
> commit 3e286c1e72651f91a2f4ca5b1fa2f366ac4cd0fa
> Date: Thu Feb 28 17:51:16 2013 +0100
>
> ARM: Fix cache flush/sync for exit stubs of JIT-compiled code.
>
> diff --git a/src/lj_asm_arm.h b/src/lj_asm_arm.h
> index afb1054..196f797 100644
> --- a/src/lj_asm_arm.h
> +++ b/src/lj_asm_arm.h
> @@ -91,6 +91,7 @@ static MCode *asm_exitstub_gen(ASMState *as, ExitNo
> group)
> *mxp++ = group*EXITSTUBS_PER_GROUP;
> for (i = 0; i < EXITSTUBS_PER_GROUP; i++)
> *mxp++ = ARMI_B|((-6-i)&0x00ffffffu);
> + lj_mcode_sync(as->mcbot, mxp);
> lj_mcode_commitbot(as->J, mxp);
> as->mcbot = mxp;
> as->mclim = as->mcbot + MCLIM_REDZONE;
>
> Try to remove the lj_mcode_sync line from 2.0.2, rebuild and test
> again. If it goes away, we'll have to investigate further. Because
> the sync is really needed as it was added due to a (rare) crash.
>
> OTOH, if this is really the problem, then it would indicate your
> code is generating traces at a very high rate. Which means you
> should look at the output of -jv or the equivalent:
>
> require("jit.v").start("/sdcard/wherever/my_dump.txt")
>
> For this to work you'll need to install or bundle src/jit/v.lua
> and src/jit/vmdef.lua.
>
> > This also seems to be something to do specifically with Android rather
> than
> > ARM, I'm also running luajit on the Raspberry Pi (although built with
> > different flags) and not experiencing the same problem.
> >
> > Currently testing primarily on a Nexus7.
>
> The RPi is an ARMv6 with a tiny cache, where too frequent cache
> flushes would have little consequences. The Nexus 7 (2012) is a
> quad-core Cortex-A9 which might be rather sensitive to that.
>
> > This is part of a reasonably large project and it is of course possibly I
> > have some other strangeness somewhere that is triggering the problem.
> Still
> > the moving from 2.0.1 to 2.0.2 is my biggest clue on where the problem is
> > so I'm hoping to try and work out exactly what that change was.
>
> If the above change isn't the culprit, you could do a git bisect
> until you find the commit that triggers it. A bit tricky to use
> that for a large project, though.
>
> --Mike
>
>
--
Kriss
http://www.WetGenes.com/
Other related posts:
