- squirrelfish-dev - lists.macosforge.org

3 instructions
by Fern Moon 12 Jun '10

12 Jun '10

I am testing the interpreter in the webkit and the JIT is disabled. I always find that the following 3 instructions get compiled, but never get executed in the interpreter. [ 0] enter [ 1] sub r0, r-10, r-9 [ 6] ret r0 What are those instructions for ? Will them get executed someplace else? Thanks Fern

2 1

Optimizing property lookup across scope objects
by Kent Hansen 07 May '10

07 May '10

Hi, Currently SquirrelFish doesn't optimize/cache access to properties across scope objects if the scope object is "dynamic", i.e. can gain properties at any time. This makes sense since a new property could be introduced in the scope object that shadows a previously cached one. For bindings to native objects, such as QObjects, we have the case of a JS object that's not fully dynamic (in the sense that properties can't be added to it ad hoc), but the only way to achieve the scope optimization seems to be to pre-populate a static scope object's symbol table with the object's available properties (in order for the bytecompiler to determine that the optimization is safe). Would it make sense to introduce a notion of a "semi-static" scope object, in which case a special bytecode is produced that 1) does the fully scoped lookup the first time; 2) patches itself to cache / look up the property directly subsequently (knowing that it's safe to skip the "semi-static" objects where the lookup failed the first time)? An alternative approach: Always cache the object where the property was resolved, and have hooks at property creation and deletion time that take care of invalidating the cache when properties are added to / removed from an object that's in some scope chain. This is more generic (works for all types of objects), but is probably more expensive (works under the assumption that property creation+deletion is a lot rarer than assigning to an existing property). Example use case: with (qtObject) { for (i = 0; i < 10000; ++i) Math.sin(i); } In this case, qtObject could have a property called Math that would shadow the Global Object's Math. But after the first lookup, we know whether this is the case. If it's not, it's safe to cache the Global Object's Math. On the other hand, if qtObject _does_ have a Math property, it can be cached / looked up directly in qtObject henceforth, since it's never going away (native Qt properties can't be deleted). Here's a benchmark I've been using: a = 123; o = {}; with (o) { (function() { for (var i = 0; i < 100000; ++i) { a; a; a; a; a; a; } })(); } Commenting out the "with" makes it 5x faster, because JSC is then permitted to do the caching. I'd like to achieve that if "o" were a QObject wrapper with no property "a", the benchmark would be about as fast with "with" as without. Kent

2 4

Does Squiirel fish JIT supports armv7 ?
by s k 01 May '10

01 May '10

Hi All, I am planning to port squirrelfish for ARMv7. Does jit and assembler os squirrelfish support arm v7 architecture? If not so, how can i implement? Any Ideas regarding this.....? I too want to contribute some code for the arm v7 support. After going through webkit buglist. JIT supports only armv4 version? Thanks & Regards, Suresh Kumar

3 3

SquirrelFis Internals
by s k 23 Apr '10

23 Apr '10

Hi Squirrelfish Developers, Hope you are all fine. This is skra, new entrant in the field of JavaScript engine. While googling about JavaScript engines, i heard of the name "SquirrelFish Extreme- an ultimate fastest JavaScript engine". Thanks to all for contributing their ideas. Just iam having some technical queries about squirrelfish. How can we benchmark squirrelfish as fastest. How can we compare with spider monkey or V8. How can i know the internals of squirrelfish. After perusing the source code, I understand the hierarchy of working procedure(as for my understanding) JavaScript Src - > parser -> yarr -> WREC > icu > compiler(using generator) produces byte code. Generated byte code is converted to machine code using JIT and assembler. If my understanding is wrong please correct me. What is the use of wtf and pcre module. How runtime module comes into picture. Thanks & Regards, skar......[?]

2 1

Improving timeout responsiveness
by Kent Hansen 23 Apr '10

23 Apr '10

Hi, Currently JSC doesn't take into account that a native function call can take an arbitrary time to complete. This means that TimeoutChecker::didTimeout() can be called less frequently than desired (see https://bugs.webkit.org/show_bug.cgi?id=38041 for how this also affects the Qt API). Does anyone have ideas other than brute-forcing the check interval to 1? Has using a high-resolution timer in the inner JSC loop been considered? I guess the two-level timeout check (cheap (loop) counter, followed by elapsed time check) was done for performance reasons. The alternatives I see are 1) Replace the two-level check by a single, reasonably precise check based on elapsed time (if this is possible without killing performance); or 2) Provide a way for native function calls to affect the tick counter. 2) could be done by extending the NativeFunction signature to take an unsigned& ticks as argument. Long-running functions would do the timing explicitly, and use a helper function to convert the elapsed time to ticks. Regards, Kent

1 0

Constant propagation
by Kent Hansen 20 Apr '10

20 Apr '10

Hi, I don't see constant propagation mentioned at http://trac.webkit.org/wiki/SquirrelFishPerfIdeas. Has it been considered? (On a related note, is the above page up-to-date?) With JSC as of r57470: var a = 3; a + 7; produces the bytecode ... [ 7] mov r-15, 3(@k1) [ 10] add r0, r-15, 7(@k2) ... Worth pursuing? Reasonably feasible to implement in the compiler? Regards, Kent

4 5

Reviewer for Tiger vs Leopard VMTags issue?
by Kent Hansen 07 Apr '10

07 Apr '10

Hi, I'm looking for a reviewer for https://bugs.webkit.org/show_bug.cgi?id=34888 Rather than spamming more individuals, I thought I'd spam this list first to see if there are any takers. Regards, Kent

1 0

Re: [squirrelfish] C API for reporting extra memory cost
by Gavin Barraclough 04 Mar '10

04 Mar '10

> Any opinions on exposing this functionality in the C API? Is it too > low-level / implementation-dependent? Hi Kent, My feeling would be that this is a little implementation dependent. These functions really only exist because our GC only supports fixed sized objects – but we are interested in looking at some point at changing our GC implementation to be able to able to allocate arbitrary sized objects on the GC heap. This would implicitly allow us to better track the additional cost. We could add API to report extra emory cost anyway, and if the heap implementation changed so that we no longer supported this then it would become a no-op, and be deprecated over time. > If exposing reportExtraMemoryCost() as-is is acceptable, I'll go > ahead and create the patch. Otherwise, suggestions for alternative > approaches are welcome. :) Depending on the lifecycle of the additional data, one solution might be to add a version of "JSObjectMake", which rather than taking a void* argument could take a size parameter, allocate the storage, and have an output argument to return the additional storage allocated. We could implicitly report the extra cost of this storage, and in the future if we do have the ability in the future to store arbitrary sized objects in the GC heap, then this method could allocate the extra data and the JSObject in a single allocation out of the GC heap, saving the extra malloc. What are your thoughts Geoff? cheers, G.

4 5

C API for reporting extra memory cost
by Kent Hansen 02 Mar '10

02 Mar '10

Hi, With the C API there doesn't seem to be a way to indicate that an object has allocated memory outside the JS heap. Internally, e.g. JSArray calls a function reportExtraMemoryCost() to give the GC a hint that it might want to run sooner rather than later. It would be nice if this were possible when implementing a custom class through the C API as well. Any opinions on exposing this functionality in the C API? Is it too low-level / implementation-dependent? The stop-gap solution I'm employing is to have the application keep track of extra allocations itself and call JSGarbageCollect() once it reaches a threshold, essentially duplicating what reportExtraMemoryCost() does but with less information for making the decision on when to trigger the collection. On the other hand, one could argue that this API is outside the scope of JSC and instead one should rely on e.g. OS notifications that the system is running low on memory, and only call JSGarbageCollect() at that point. That depends on platform support though, I'm not sure for which platforms it's possible. If exposing reportExtraMemoryCost() as-is is acceptable, I'll go ahead and create the patch. Otherwise, suggestions for alternative approaches are welcome. :) Regards, Kent

1 0

Question about building SquirrelFish Extreme
by Huahui Wu 19 Jan '10

19 Jan '10

Hello, My name is Huahui and I am a Software Engineer in Google/Android. I am investigating the possibility to use SquirrelFish Exteme in Android's browser since it seems to have the best performance as in http://webkit.org/blog/214/introducing-squirrelfish-extreme/. However, I built and tested the latest Webkit vs. V8 with SunSpider and didn't find much performance difference. I was using a 2.4GHz Intel Core 2 Duo MacBook Pro, the JSC with SquirrelFish (Extreme?) spent 448ms, and V8 spent 446ms. I am wondering if that's because I was running the old SquirrelFish. My guess is, checking out the nightly of webkit and building it by default will generate the old version SquirrelFish instead of SquirrelFish Extreme. Am I correct? If so, could anyone help me with how to enable the extreme version of SquirrelFish? Thank you! - Huahui

3 3