586886 - JM: Remove VMFrame::inlineCallCount

Status: RESOLVED FIXED

(Core :: JavaScript Engine, defect)

Description

[David Anderson [:dvander] - inactive, e-mail if emergency]

We can remove inlineCallCount by combining it with the stack quota check, simplifying the stack quota check at the same time.

Comment 1

[David Anderson [:dvander] - inactive, e-mail if emergency]

Attachment: part 1: cleanup, add entryFp

Precursor patch removes correctness dependency on iCC, introducing VMFrame::entryFp which is the stopping point for unwinding frames.

It also cleans up a bunch of confusing invariants in InvokeHelpers.

Comment 2

[David Anderson [:dvander] - inactive, e-mail if emergency]

Attachment: part 1.1: add entryFp for x86, x64

Comment 3

[David Anderson [:dvander] - inactive, e-mail if emergency]

Largest frame size on SS is 656 bytes, average is 184.
Largest frame size on v8 is 1344 bytes, average is 175.

Multiplying our current max-inline-callcount (3000) by these numbers, either as-is or generously increased, results in a huge portion of our contiguous stack.

On Windows, we still have to check the committed region of the contiguous stack. Luke points out that we can check MIN(committed, maxDepth) to take a slow path, and figure out whether we've over-recursed or just need to commit more from there. This will keep the fast single-check on Windows without reporting more memory usage.

Comment 4

[David Anderson [:dvander] - inactive, e-mail if emergency]

Attachment: part 2: remove VMFrame::inlineCallCount

Comment 5

[David Anderson [:dvander] - inactive, e-mail if emergency]

Attachment: part 3: combine commit and overrecursion checks

Comment 6

[David Anderson [:dvander] - inactive, e-mail if emergency]

So far this is a 1% SS win, 5% v8 win on my machine.

Comment 7

[David Anderson [:dvander] - inactive, e-mail if emergency]

On graphs this was a 1% SS win, 3% v8 win.

http://hg.mozilla.org/projects/jaegermonkey/rev/d7b81bd995c3
http://hg.mozilla.org/projects/jaegermonkey/rev/6a6a72c1961a

Comment 8

[Jacob Bramley [:jbramley]]

According to "hg bisect", d7b81bd995c3 broke most of the trace-tests for ARM.

Comment 9

[Jacob Bramley [:jbramley]]

ARM fixes: http://hg.mozilla.org/projects/jaegermonkey/rev/0bb64c9ca0b6

The VMFrame construction code in JaegerTrampoline hadn't been updated for ARM.