main.cpp(357) : warning C4291: 'void *QQmlJS::VM::Managed::operator new(size_t,QQmlJS::VM::MemoryManager *)' : no matching operator delete found; memory will not be freed if initialization throws an exception
d:\dev\v4vm\src\v4\qv4managed.h(112) : see declaration of 'QQmlJS::VM::Managed::operator new'
Change-Id: Idd7f54f257ae93fdf04ecbf3f938e3b2d981bf89
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
Now the compiler cannot only optimise for this, but also stop complaining
about callers not returning any value.
Change-Id: I71d98721f70849178613096408e959d7e24dca8a
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
Due to our large amount of temps we also end up creating large stack frames and
thus add large constants to the stack pointer. That affects the encoding of the
immediates and MacroAssemblerARMv7 ASSERTs out for values that require
encoding.
This is unlikely to get fixed upstream and it's infact impossible to create a
testcase with JSC JIT due to the fact that it barely uses the stack frame.
I'd rather not patch the upstream file as it is a condition hard to find and
a patch easy to drop by accident. Instead this patch adds a simple workaround
that comes are low cost: Just load the immediate into a register and do the
addition.
Change-Id: Ia551a15d2f5f6243b295a9bfd19df778467189ec
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
Access local temps through a newly allocated LocalsRegister instead of the
regular frame pointer register. In the new exception handling we're going to
re-enter our function in the middle and want to access the same local temps,
but we can't do that through the stack frame pointer then, because that one
will _have_ to continue to point to the local stack frame in order for
unwinding to work properly.
Also the callee saved registers are now stored right below the stack
frame pointer instead of at the bottom of the stack. This way they
can be described easily in the unwind info as always relative to the
canonical frame address.
Change-Id: I53ef6291d99396577a72ceb9246f7ca3d99e5137
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
It tells us quite explicitly how it wants a default constructor
for VoidType :)
Change-Id: I854370c869f179da7f842fbf675e05678285630d
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
The test suite does things like "1 instanceof 1" and expects a
type error to be thrown. Therefore we should not assert(!"unreachable")
when instanceof is called with a numberic constant but just fall back
to the run-time implementation, which does the right thing.
Change-Id: Iced93e679d56f4491d38c50b669e12dd160c220c
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
Use branchTest32 to implement the test for the sign bit in right
shift operations.
Change-Id: I07b3ead4d32761ee3d5f529259be5b5987b7ec5a
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
Ideally these checks should get inlined in the generated
assembly.
Change-Id: I4f63f7235a7d3bbdf8413df9f7d674104ff95b07
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
-1 >> 0 should return UINT_MAX, as the result is an
unsigned int according to spec. The only way the result
of the inline shr operation can be signed is by shifting
0 bytes. But the easiest implementation is to test the
result for signed-ness and then fall back to the slow
implementation.
Change-Id: Ic4614006d06cf01376ef95b6f23ca2c7216a2812
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
Cast the void* to a quintptr, to make sure we call the right
QByteArray::number overload depending on the size of a pointer
on the archicture. Otherwise we generate 0x1324 strings that
we can't successfully replace with the function names in the
disassembler output.
Change-Id: Iddc82534487d93547b597d39286b92ffdff6da6c
Reviewed-by: Erik Verbruggen <erik.verbruggen@digia.com>
On arm the pointer to the storage of the VM::Value to return is passed in r0.
The value in that register is destroyed soon after, so later when we want to
access it in visitRet() we'll get garbage.
To solve this we behave similar to gcc now, which upon function entry saves the
values of the registers used for parameter passing onto the stack. Except that
on arm we now do this before pushing the link register, which makes the stack
frame look identical to ia32. (old ebp / return address / arg 0 / arg 1 / ...)
With that we can theoretically access the pointer to the return value storage.
In practice we also need to change meaning of the addressForArgument() helper
function to only return the address of arguments on the stack. But that makes
sense since Address() is meaningless for values passed in registers.
Also tightened the #ifdef in visitRet() for determining whether to use the
return value register or not. That wasn't strictly necessary, but makes
the condition a bit clearer.
Change-Id: I6fbef6645275ebaa75484d666b4bbfd073f945a5
Reviewed-by: Erik Verbruggen <erik.verbruggen@digia.com>
Whether we should do ret(n) or ret should depend on whether the caller
provides the pointer to the return value as hidden first parameter or not.
That's the case on ia32 but not on x86-64 or arm, where the first parameter
register is used instead. So the correct preprocessor macro to use here
is ARGUMENTS_IN_REGISTERS instead of VALUE_FITS_IN_REGISTER.
Change-Id: I3a8a8fa316896848baca37626f87ed98c096e14a
Reviewed-by: Erik Verbruggen <erik.verbruggen@digia.com>
Simplify the code for determining whether to push a function call parameter
onto the stack or into a register.
Change-Id: I3ab9230b8c0a3b2466c3000d89faf4fd79f927eb
Reviewed-by: Erik Verbruggen <erik.verbruggen@digia.com>
On ARM registers are in JSC::ARMRegisters instead of JSC::X86Registers :)
Change-Id: Ib11f0b3caa84a5015905f0a7937b4250c6f76c78
Reviewed-by: Erik Verbruggen <erik.verbruggen@digia.com>
Make sure to calculate the frame size when entering the function the
same way as when leaving it, otherwise the stack pointer adjustment
is wrong and we get nice crashes.
Change-Id: I19f953c3243cf6f1448ad95cad7587fbdca2ae6d
Reviewed-by: Erik Verbruggen <erik.verbruggen@digia.com>
This makes V4 reference all variables in scopes
that are not the global scope by index. The JIT and
the interpreter walk up the scope chain to get the
correct reference.
Variables are only resolved by name for the global
scope, if the scope contains an eval statement (as
eval can define new variables) and inside with and
catch scopes.
Change-Id: Ib9f9d1a03d50124130aefd169eeb071533ba3520
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
When doing var foo = bar = 42; then we would assign 42 to bar and bar to foo,
resulting in the wrong value for foo if bar was read-only for example.
The spec says in 11.13.1.6 that the rval is to be returned, so we just do
that via the temp we already have.
Change-Id: I44ea895abe4796af10c371baac22c2b26f37b519
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
Introduce a Reference type next to PointerToValue for which we can ensure that
it's non-null using an assert. Otherwise implemented push(PointerToValue)
to push a null pointer if the temp is null, instead of asserting.
Change-Id: I70f15e39dd80a6b2c65630060cba35f3417c0634
Reviewed-by: Lars Knoll <lars.knoll@digia.com>
Instead use Q_UNIMPLEMENTED() that just prints a warning. This
really is missing functionality that in the meanwhile shouldn't cause
crashes due to failing assertions.
Change-Id: I85314d04e35af35b95dc81e9cbdd659d13f43798
Reviewed-by: Jędrzej Nowacki <jedrzej.nowacki@digia.com>
Remove unused code, inline where we only use a method
once.
Change-Id: I1896efc3f4d309082aff2f80f944e19c1ede2f50
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
This makes all runtime structures fully non virtual.
Change-Id: I804568ca9bc33d4be0324ed542df8eab5892c0eb
Reviewed-by: Simon Hausmann <simon.hausmann@digia.com>
Erik Verbruggen
Simon Hausmann
Lars Knoll