Generate the code from a macro instead of duplicating boiler
plate code. Operate on Heap::Base instead of Managed.
Change-Id: I84c5a705980899be3e5b931a093645e50d3923bf
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
We need to move the Data objects out of the Managed
objects, to avoid lots of trouble because inner classes
can't be forward declared in C++.
Instead move them all into a Heap namespace.
Change-Id: I736af60702b68a1759f4643aa16d64108693dea2
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
This decouples things a bit better and helps moving
over to directly store heapobject pointers in other
objects.
Change-Id: I798f922e018b0a3ca6f8768e4a810187f34d82f6
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
This is a step towards storing direct heap object pointers for the values
on the JS stack, to avoid the costly indirection for data access.
Change-Id: Ibb57ed6cf52a7088bbc95ee04ae3a4cb25b8c045
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
Try to free our large items for every 8 megabytes of additional large item
memory we allocate. This helps in particular on 32-bit builds and was noticable
in tst_qqmlecmascript::push_and_shift().
Change-Id: I4cc0b188e58ccaf32026e38c7aaf1cfadc83148b
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
As the example in QTBUG-42051 demonstrates, QML may produce a lot of binding
objects and each of them produce a QV4::Persistent. During the mark phase we
may run out of JS stack space. One fix (for the future) is to reduce the number
of QV4::Persistent objects, but in the meantime we can also reduce the pressure
on the stack by draining it earlier.
Task-number: QTBUG-42051
Change-Id: Iea73f8a869048ea0bf3f4a64dbd24b6fb8c68f6b
Reviewed-by: Ulf Hermann <ulf.hermann@digia.com>
Instead of allocating the data directly, centralize the object and its ::Data
allocation in one place in the memory manager. This is in preparation for
additional pointer indirection later.
Change-Id: I7880e1e7354b3258b6a8965be378cd09c9467d25
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
No need to differentiate in the name anymore, as the data
structures all inherit from each other now.
Change-Id: Ia41f50ce4e521f9626d874311ceb57e0e194888b
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
This prepares for moving over to a d pointer scheme,
where Managed subclasses don't hold any data directly. This
is required to be able to move over to a modern GC.
Change-Id: I3f59633ac07a7da461bd2d4f0f9f3a8e3b0baf02
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
Executing an allocation heavy testcase of JavaScriptCore on my
i7 Sandy Bridge notebook 6.85% is spent inside the ::engine()
call as gcc 4.8.2 of Debian didn't inline the call. Inline the
call sites by hand. I removed the protected ::engine() as it is
now unused.
$ time qmljs JavaScriptCore/tests/perf/bench-allocate-nonretained.js
before (best run of three)
real 0m2.234s
user 0m2.228s
sys 0m0.008s
after (worse run of three)
real 0m2.097s
user 0m2.088s
sys 0m0.008s
Change-Id: I20b73b3b3dac630eb1d5e7d66bcb50c839630567
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
This profiler tracks every memory allocation and deallocation, by the
MemoryManager as well as the V4 VM, and exposes them as a stream of
events to the profiler service.
Change-Id: I85297d498f0a7eb55df5d7829c4b7307de980519
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
This allows us to avoid calling a destructor on objects
that don't require one. After the memberData change this
should be most objects.
Also fix a bug where we didn't call the destroy() method on
large objects, potentially leaking memory.
Change-Id: I1708055d568d85b0a3876899d35e8c3eb92dd222
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
With this modification the CHUNKSIZE is setable by QV4_MM_MAX_CHUNK_SIZE
environment variable so the memory usage which is important for embedded
devices can be finetuned.
Change-Id: I3cd75158f2255651edd341873de035c1222e3c92
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
Small regression from commit 7ae796cb14.
Processing the deletables might result in the onDestruction emission, which in
turn may end up in GC allocation and thus GC runs. That in turn may result
in m_deletables processing, which at this point is nested then. For that to
work we need to set m_deletables back to zero _before_ beginning with the
iteration.
Fixes tst_qqmlecmascript with aggressive gc.
Change-Id: Ibb310b30cd496644557f4c1bb23318b18ee8f36c
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
Cap the max chunk size at 2MB. This value still still doesn't
affect the v8 benchmark noticably, but should avoid extreme
memory usage in some corner cases
Task-number: QTBUG-37134
Change-Id: If2050374c4a7df7ff74194d64880e2d660ea26fd
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
This option doesn't make sense anymore, as we memset the object to
0 on destruction anyway.
Change-Id: Ie40563394f9cacda1b35fde114c9a6043f53d460
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
The method is only required for the QObjectWrapper, but
there we can instead simply register the deletable in the
destroy method.
Change-Id: I944319d327859cce33ac31a174e8d56fc4babfc1
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
Use QV4_MM_STATS to get some statistics about each GC run
written to stderr.
Change-Id: Idc30c06e1c3ca8353a2f16615cb26df13a8f23ed
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
This was meant as a debugging help after we took the exact
GC into use. Turns out it doesn't help debug any issues,
aggressive garbage collection is much more helpful in
finding GC issues.
Change-Id: Idfff26d3a4a15e570615009c1e57921bca6297ca
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
A smaller maximum block size helps limit RAM usage on constrained
systems, at the cost of more frequent garbage collection.
Change-Id: Iba07d9cc628e3178dfaad35664c631844540bc9d
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
Remove SafeValue, it was used to port over to an exact GC. Since
we now have that, we can now safely merge it with QV4::Value
again. Also rename SafeString to StringValue for better naming
consistency.
Change-Id: I8553d1bec5134c53996f6b0d758738a0ec8a2e4d
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
Keep the basic methods in ManagedVTable, but have
the Object related stuff in an ObjectVTable class.
Change-Id: I9b068acf3caef813686227b8d935e7df1a7d1a6e
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
[ChangeLog][Platform Specific Changes][BSD] Fixed compile errors on
DragonFly BSD and potentially FreeBSD.
Task-number: QTBUG-35867
Change-Id: Iea90b93672c34f8a4b56e9afc4dbfb82cc993548
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
When proc is not mounted pthread_getattr_np fails, so default
to 1MB stack in getStackLimit and to exactGC in MemoryManager
Change-Id: Ic7515fd420f2d39a656808d24a3915a657722891
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
Object construction shouldn't need to zero initialize itself,
let's rather do this in the GC, where we can use fast memset's.
Change-Id: I2f9efa1729183b0d737de5a84f92af319b2c5631
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
This saves one pointer per object, and willmake other optimizations
easier in the future.
Change-Id: I1324cad31998896b5dc76af3c8a7ee9d86283bfe
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
This finally gives proper memory management for ExecutionContexts.
So far they had been garbage collected but where still allocated
using standard malloc/free(). This allows us to collect the
contexts faster and speed up context creation.
Change-Id: I02e642391d55eaa59ab3f4c2720a2ac71259caf4
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
This is required, so we can track ExecutionContexts
through the regular memory manager.
Change-Id: I1bd2e2ef275e6e9e1f364a35affbb991f4377b05
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
Disable JIT and avoid unsupported functions under WinRT.
Also add MSVC's ARM flag to the double conversion white list.
Change-Id: I22ec340a20b113fdeefb802ac61812f78a527895
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
Some test cases were badly trashing the GC, running it
way too often. Fix this by only running the GC after
we allocated more than 50% of the amount of available
items.
Change-Id: Icfe6d0730552ef2298963de706213ebc9e3e342c
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
Don't use recursive function calls anymore. Instead, push marked
objects onto the JS stack, and then pop them off when their children
are being marked.
Should reduce stack memory usage, and improves performance by ~5%.
Change-Id: I2d37d97579144fcba87ec8e9fd545dd220c01fbb
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
Keep conservative GC as a fallback for testing
Enable all tests again that were skipped due to
GC issues.
Change-Id: I8e0fa728207bdd39a96d0acf95e27841157d8402
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
This will simplify finding the remaining direct usages of
QV4::Value that need fixing.
Change-Id: I223099727436d5748027c84c53d9dfc4028e38ed
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
This was copied from qtscript's:
src/3rdparty/javascriptcore/JavaScriptCore/runtime/Collector.cpp
Change-Id: Iaa5549e6915aa4aa4bf532fbe7c9117a0858ff02
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
Lars Knoll
Simon Hausmann
Lars Knoll
Simon Hausmann
Jani Heikkinen
Holger Hans Peter Freyther
Ulf Hermann
Karim Pinter
Michael Brasser
Alex Blasche
Albert Astals Cid
Friedemann Kleint
Sérgio Martins
Andrew Knight
Kai Koehne