Closed Bug 676700 Opened 11 years ago Closed 11 years ago

Slightly optimize js::IndexToId


(Core :: JavaScript Engine, defect)

Not set





(Reporter: Waldo, Assigned: Waldo)


(Whiteboard: [inbound])


(1 file, 1 obsolete file)

Attached patch Patch (obsolete) — Splinter Review
Sparse/dense patchwork needed js::IndexToId, which I then implemented, before discovering it had already been done.  But the current implementation isn't as efficient as this one (it creates an intermediate string rather than atomize directly from a jschar buffer), so let's improve it slightly.
Attachment #550851 - Flags: review?(luke)
IIRC, js_AtomizeString(js_IntToString()) is only a few branches (and obviously a function call) worse than the proposed IndexToId.  Can you measure the perf improvement for any use of IndexToId?

Btw, there is a flagrantly redundant branch in js_AtomizeString perhaps you could sweep up...
If nothing else, it seems like we should have an inline reusable function to factor out the core backwards-filling index-to-chars algorithm.
Attached patch v2Splinter Review
It's not the js_AtomizeString call that contributes extra time.  It's that js_NumberToString must allocate a new JSString, and then that JSString immediately becomes garbage after the atom is created.  So a little time, a little more GC pressure, neither particularly great (if not too horrible generally -- although, one of the JSON parser optimizations I did was to eliminate an extra string creation like this, and it showed up in perf numbers).

I was thinking that loop was a bit duplicated, but it looked a little annoying to abstract it out.  This patch does that.
Attachment #550851 - Attachment is obsolete: true
Attachment #550851 - Flags: review?(luke)
Attachment #550912 - Flags: review?(luke)
Ohhh, I knew Atomize converted non-atom strings to JSAtoms, but I forgot that it doesn't do this to string argument passed to js_AtomizeString.  It seems like it should.  A quick scan of js_AtomizeString shows at least a couple of places that exhibit the immediately-garbage string pattern you just mentioned.
However, if the input is already atomized, then the new string is still garbage, so your approach is better.
Comment on attachment 550912 [details] [diff] [review]

Review of attachment 550912 [details] [diff] [review]:

Wow, I didn't realize it was x3!  r+ with nits.

::: js/src/jsatominlines.h
@@ +142,5 @@
> + * than end.
> + */
> +template <typename T>
> +inline mozilla::RangedPtr<T>
> +WriteIndex(uint32 index, mozilla::RangedPtr<T> end)

As much as I like short names, this seems too short for how weird it is.  How about BackfillIndexInCharBuffer?

Second idea: how about a symbolic UINT32_CHAR_BUFFER_LENGTH constant that is statically asserted correct here and can be reused/asserted in the 3 other places?
Attachment #550912 - Flags: review?(luke) → review+
Whiteboard: [inbound]
Target Milestone: --- → mozilla8
Closed: 11 years ago
Resolution: --- → FIXED
You need to log in before you can comment on or make changes to this bug.