The current parameters (initial slack = 6, kGenerousAllocationsCount = 16)
caused 15-20% performance regression on 3d-raytrace on x64
(visible both in standalone shell and in the web harness). This is very
likely some kind of an alignment problem (it goes away completely with --prof).
A similar problem can be observed on ia32 with initial slack=2.
Increased initial slack to 8 and decreased constructions count to 8.
This made the regression go away (and did not harm any of the other tests).
Review URL: http://codereview.chromium.org/3595002
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@5571 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
during snapshot size optimization.
Sorry, now I figured out that the diff implementation itself was also
incorrect. Reachable nodes must be filtered from the beginning,
otherwise, an object that is already disconnected, but not discarded
yet, will not appear as a deleted (thankfully, this bug for some
reason had appeared on the x64 port.)
BUG=868
TEST=HeapSnapshotRootPreservedAfterSorting
Review URL: http://codereview.chromium.org/3531005
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@5570 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
The lazy parsing functions took a host of arguments that can all be
derived from the SharedFunctionInfo, and the SharedFunctionInfo is
always available when parsing lazily. Change the interface to take a
single CompilationInfo or SharedFunctionInfo argument.
Also remove a flag in the parser that was always false when it was read.
Review URL: http://codereview.chromium.org/3538005
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@5561 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
It was a wart that we had three handle fields, exactly one of which
was non-null; and that we had three overloaded constructors. Instead,
introduce subclasses and virtual methods.
Remove some unused fields from class CompilationInfo.
Review URL: http://codereview.chromium.org/3566003
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@5560 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
When inspecting an AST node in the debugger, it can be difficult to
find its concrete type. Add a "node_type" virtual function to all the
concrete AST nodes.
Use a macro to generate the definition of node_type as well as Accept
and the AsXXX type testing/casting functions.
Review URL: http://codereview.chromium.org/3431026
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@5549 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
Reduces size of preparser data significantly when there are nested functions.
Also allows us to drop the "skip" fields of function entries,
that tells us how much preparse-data to skip when skipping the function source.
Review URL: http://codereview.chromium.org/3412034
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@5541 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
The object's space in Page starts after Page header and is aligned to kMapAlignment which is 32 bytes on 32-bit and 8 bytes on 64-bit.
In case of 64-bit target, the current page header size is exactly 32 bytes so we get the code magically aligned at 32 bytes but it is better to have a separate CODE_POINTER_ALIGN macro to make sure the object space in Page is aligned properly for both maps and code.
There could be a small waste of bytes sometimes (since both Page header and Code header sizes are aligned separately) but it seems the optimal one would involve cross-dependencies between .h files and not clear if it's worth it.
This is a back-port from Isolates branch.
Review URL: http://codereview.chromium.org/3461021
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@5526 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
The cause for missing functions is that some of them are created
from compiled code (see FastNewClosureStub), and thus not get
registered in profiler's code map.
My solution is to hook on GC visitor to provide JS functions
addresses to profiler, only if it is enabled.
BUG=858
TEST=
Review URL: http://codereview.chromium.org/3417019
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@5523 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
Finally sovles the problem that r5342 attempted to solve.
When adding a stub to a map's code cache we need to make
sure that this map is not used by object that do not need
this stub.
Existing solution had 2 flaws:
1. It checked that the map is cached by asking the current context.
If the object escaped into another context then NormalizedMapCache::Contains
returns false negative.
2. If a map gets evicted from the cache we should not try to modify it
even though Contains returns false.
This patch implements much less fragile solution of the same problem:
A map now has a flag (is_shared) that is set once the map is added
to a cache, stays set even after the cache eviction, and is cleared
if the object goes back to fast mode.
Added a regression test.
Review URL: http://codereview.chromium.org/3472006
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@5518 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
For some reason, the scope's arguments and arguments shadow were
variable proxies, which resulted in all references to the arguments
shadow being shared in the AST. This makes it hard to put per-node
state on the AST nodes.
I took the opportunity to remove Variable::AsVariable which has
confused people in the past, and to rename Variable::slot to the more
accurate Variable::AsSlot.
Review URL: http://codereview.chromium.org/3432022
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@5517 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
The number of inobject properties used to be derived from the number
of this property assignments in the constructor (and increased by 2 to
allow for properties added later). This very often leads to wasted inobject
slots.
This patch reclaims some of the unused inobject space by the following method:
- for each constructor function the first several objects are allocated using the initial
("generous) instance size estimation (this is called 'tracking phase').
- during the tracking phase map transitions are tracked and actual property counts are collected.
- at the end of the tracking phase instance sizes in the maps are decreased if necessary
(starting with the function's initial map and traversing the transition tree).
- all further allocation use more realistic instance size estimation.
Shrinking generously allocated objects without costly heap traversal is made possible
by initializing their inobject properties with one_pointer_filler_map (instead of undefined).
The initial slack for the generous allocation is increased from 2 to 6 which really helps some tests.
Review URL: http://codereview.chromium.org/3329019
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@5510 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
