Removed the assumption that "final" objects have a fixed number of inline slots
https://bugs.webkit.org/show_bug.cgi?id=98332
Reviewed by Filip Pizlo.
This is a step toward object size inference.
I replaced the inline storage capacity constant with a data member per
structure, set the the maximum supported value for the constant to 100,
then fixed what broke. (Note that even though this patch increases the
theoretical maximum inline capacity, it doesn't change any actual inline
capacity.)
* dfg/DFGSpeculativeJIT32_64.cpp:
(JSC::DFG::SpeculativeJIT::compile):
* dfg/DFGSpeculativeJIT64.cpp:
(JSC::DFG::SpeculativeJIT::compile):
* jit/JITPropertyAccess.cpp:
(JSC::JIT::compileGetDirectOffset): These functions just get a rename:
the constant they need is the first out of line offset along the offset
number line, which is not necessarily the same thing (and is, in this
patch, never the same thing) as the inline capacity of any given object.
(JSC::JIT::emit_op_get_by_pname):
* jit/JITPropertyAccess32_64.cpp: This function changes functionality,
since it needs to convert from the abstract offset number line to an
actual offset in memory, and it can't assume that inline and out-of-line
offsets are contiguous on the number line.
(JSC::JIT::compileGetDirectOffset): Updated for rename.
(JSC::JIT::emit_op_get_by_pname): Same as emit_op_get_by_pname above.
* llint/LowLevelInterpreter.asm: Updated to mirror changes in PropertyOffset.h,
since we duplicate values from there.
* llint/LowLevelInterpreter32_64.asm:
* llint/LowLevelInterpreter64.asm: Just like the JIT, most things are just
renames, and get_by_pname changes to do more math. I also standardized
offset calculations to use a hard-coded "-2", to match the JIT. This
isn't really better, but it makes global search and replace easier,
should we choose to refactor this code not to hard-code constants.
I also renamed loadPropertyAtVariableOffsetKnownNotFinal to
loadPropertyAtVariableOffsetKnownNotInline in order to sever the assumption
that inline capacity is tied to object type, and I changed the 64bit LLInt
to use this -- not using this previously seems to have been an oversight.
* runtime/JSObject.cpp:
(JSC::JSObject::visitChildren):
(JSC::JSFinalObject::visitChildren):
* runtime/JSObject.h:
(JSC::JSObject::offsetForLocation):
(JSNonFinalObject):
(JSC::JSFinalObject::createStructure):
(JSFinalObject):
(JSC::JSFinalObject::finishCreation): Updated for above changes.
* runtime/JSPropertyNameIterator.h:
(JSPropertyNameIterator):
(JSC::JSPropertyNameIterator::finishCreation): Store the inline capacity
of our object, since it's not a constant.
(JSC::JSPropertyNameIterator::getOffset): Removed. This function was
wrong. Luckily, it was also unused, since the C++ interpreter is gone.
* runtime/PropertyMapHashTable.h:
(PropertyTable): Use a helper function instead of hard-coding assumptions
about object types.
(JSC::PropertyTable::nextOffset):
* runtime/PropertyOffset.h:
(JSC):
(JSC::checkOffset):
(JSC::validateOffset):
(JSC::isInlineOffset):
(JSC::numberOfSlotsForLastOffset):
(JSC::propertyOffsetFor): Refactored these functions to take inline capacity
as an argument, since it's not fixed at compile time anymore.
* runtime/Structure.cpp:
(JSC::Structure::Structure):
(JSC::Structure::flattenDictionaryStructure):
(JSC::Structure::putSpecificValue):
* runtime/Structure.h:
(Structure):
(JSC::Structure::outOfLineCapacity):
(JSC::Structure::hasInlineStorage):
(JSC::Structure::inlineCapacity):
(JSC::Structure::inlineSize):
(JSC::Structure::firstValidOffset):
(JSC::Structure::lastValidOffset):
(JSC::Structure::create): Removed some hard-coded assumptions about inline
capacity and object type, and replaced with more liberal use of helper functions.
git-svn-id: http://svn.webkit.org/repository/webkit/trunk@130359 268f45cc-cd09-0410-ab3c-d52691b4dbfc
15 files changed
