1441521 - [MIPS] Optimize comparation with small constants

Status: RESOLVED FIXED

(Core :: JavaScript Engine: JIT, enhancement, P5)

Comment 1

[Dragan Mladjenovic]

Attachment: bug1441521.patch

This patch tries to embed small integers into slti/sltiu immediate instructions instead loading them into register while branching or doing cmpSet.

Comment 2

[yuyin]

hi, Mladjenovic

thanks for this patch, it looks good for me.

I just want to build it and run octane to see how the performance improvement, but 'js' program failed because of :

MacroAssembler-mips-shared-inl.h:1023
MOZ_RELEASE_ASSERT(!JitOptions.spectreIndexMasking);

and lots of jit-tests failed because of this.

Do you have the same phenomenon?

Comment 3

[yuyin]

Sorry, I mean this patch is look good to me.

But there have some other issues which cause jit-test failed before this patch.

Comment 4

[Dragan Mladjenovic]

(In reply to yuyin from comment #2)
> hi, Mladjenovic
> 
> thanks for this patch, it looks good for me.
> 
> I just want to build it and run octane to see how the performance
> improvement, but 'js' program failed because of :
> 
> MacroAssembler-mips-shared-inl.h:1023
> MOZ_RELEASE_ASSERT(!JitOptions.spectreIndexMasking);
> 
> and lots of jit-tests failed because of this.
> 
> Do you have the same phenomenon?

Sorry for the late replay.

The blame for that is all on me. You have to run js-shell with "--spectre-mitigations=off" to get past this.
I've not pushed Spectree mitigation primitives implementation for MIPS. I'm still waiting for the Spectree mitigation work to stabilize in order for me to gather any data on performance overhead. (Octane benchmark doesn't seem to be affected in my early benchmarks, the again it maybe isn't suitable to test this) I could post the patch as early as tomorrow? I would be grateful If you could then post the results of benchmarks on Loongson with and without the spectree mitigation enabled.

Comment 5

[yuyin]

Thank you.

and please fell free to ask me if you want get any results of benchmarks on Loongson. You can also send email to me directly if you like at anytime

(In reply to Dragan Mladjenovic from comment #4)

> 
> The blame for that is all on me. You have to run js-shell with
> "--spectre-mitigations=off" to get past this.
> I've not pushed Spectree mitigation primitives implementation for MIPS. I'm
> still waiting for the Spectree mitigation work to stabilize in order for me
> to gather any data on performance overhead. (Octane benchmark doesn't seem
> to be affected in my early benchmarks, the again it maybe isn't suitable to
> test this) I could post the patch as early as tomorrow? I would be grateful
> If you could then post the results of benchmarks on Loongson with and
> without the spectree mitigation enabled.

Comment 6

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/mozilla-inbound/rev/5417950f6fc2
[MIPS] Optimize comparation with small constants. r=yuyin

Comment 7

[Cristina Coroiu [:ccoroiu]]

https://hg.mozilla.org/mozilla-central/rev/5417950f6fc2