1391611 - Devirtualize MNode::kind()

Status: RESOLVED FIXED

(Core :: JavaScript Engine: JIT, enhancement)

Description

[Jan de Mooij [:jandem]]

Attachment: Patch

MNode::kind is a pretty hot virtual function to determine isResumePoint() or isDefinition(). We can devirtualize it by making MNode::block_ a tagged pointer - the low bit can store the kind.

I also |= delete|'d these functions on the MDefinition and MResumePoint derived classes, this caught a place where we called isResumePoint() on an MDefinition (this is always false).

Comment 1

[Nicolas B. Pierron [:nbp]]

Comment on attachment 8898759 [details] [diff] [review]
Patch

Review of attachment 8898759 [details] [diff] [review]:
-----------------------------------------------------------------

::: js/src/jit/MIR.h
@@ +343,3 @@
>      }
>      MBasicBlock* block() const {
> +        return reinterpret_cast<MBasicBlock*>(blockAndKind_ & ~KindMask);

Is this really worth it?

I would expect resume point to access the block()->info() pointer frequently.

@@ +527,5 @@
>      }
>  
> +    // Calling isDefinition or isResumePoint on MDefinition is unnecessary.
> +    bool isDefinition() const = delete;
> +    bool isResumePoint() const = delete;

Nice!

Comment 2

[Jan de Mooij [:jandem]]

(In reply to Nicolas B. Pierron [:nbp] from comment #1)
> Is this really worth it?
> 
> I would expect resume point to access the block()->info() pointer frequently.

I just added some logging and ran Octane. MNode::kind() was called about 10.1 million times, MNode::block() 11.5 million times. So block() is called a bit more, but getting rid of the virtual call for kind() outweighs the unmasking there IMO.

Comment 3

[Nicolas B. Pierron [:nbp]]

One more thing that I forgot. Can you specialize MDefinition::block() to avoid the un-masking as and assert that the low bots are already 0.

Comment 4

[Jan de Mooij [:jandem]]

Attachment: Followup

What do you think of this one? It optimizes block() on both MDefinition and MResumePoint to be a bit more efficient than MNode::block().

Comment 5

[Jan de Mooij [:jandem]]

Attachment: Followup

The right one.

Comment 6

[Nicolas B. Pierron [:nbp]]

Comment on attachment 8898797 [details] [diff] [review]
Followup

Review of attachment 8898797 [details] [diff] [review]:
-----------------------------------------------------------------

::: js/src/jit/MIR.h
@@ +338,5 @@
> +        MOZ_ASSERT(isResumePoint());
> +        static_assert(unsigned(Kind::ResumePoint) == 1, "Code below relies on low bit being 1");
> +        // Use a subtraction: if the caller does block()->foo, the compiler
> +        // will be able to fold it with the load.
> +        return reinterpret_cast<MBasicBlock*>(blockAndKind_ - 1);

Whoa, this is indeed nicer!

Comment 7

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/mozilla-inbound/rev/0cc6be22d7e0
Devirtualize MNode::kind(). r=nbp

Comment 8

[Phil Ringnalda (:philor)]

https://hg.mozilla.org/mozilla-central/rev/0cc6be22d7e0