Bit tests with a memory operand should be avoided on Intel processors.
On Atom bts with a memory operand has 10~11 cycles latency and on Core 2
it has 5~6 cycles latency, and also produces multiple uops.
If the bit base operand is a register the latency is 1 cycle instead.
This CL improves several cases from Kraken 1.1 benchmark on Atom. For example,
12% for imaging-desaturate, 8% for audio-fft, and 5% for audio-beat-detection.
Review URL: http://codereview.chromium.org//7048003
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@7987 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
Better support for 'polymorphic' JS and external arrays
Allow keyed store/load stubs to switch between external array and fast JS arrays without forcing a state transition to the generic stub.
There CL consists of two pieces of functionality. First, code stubs for fast element arrays don't immediately transition to the MEGAMORPHIC state when there's a map mismatch. Second, two ICs are cached per map for fast elements, the MONOMORPHIC version, and a new MEGAMORPHIC version that handles two or more different maps and dispatches to shared stubs to perform the array operation.
Review URL: http://codereview.chromium.org/7036016
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@7935 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
Allow keyed store/load stubs to switch between external array and fast JS arrays without forcing a state transition to the generic stub.
There CL consists of two pieces of functionality. First, code stubs for fast element arrays don't immediately transition to the MEGAMORPHIC state when there's a map mismatch. Second, two ICs are cached per map for fast elements, the MONOMORPHIC version, and a new MEGAMORPHIC version that handles two or more different maps. Currently, the only array types supported by the MEGAMORPHIC stub are fast elements for objects and JSArrays.
BUG=
TEST=
Review URL: http://codereview.chromium.org/6894003
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@7917 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
This is IA32 only for now.
Added a random value to each assembler instance (JIT cookie) to be used for constant splitting. Added safe versions of setting a register with an immediate value and for pushing an immediate value. Used these functions where user controlled immediate values could be emitted in the code stream. I also used it for immediates which are an argument number even though the number of formal arguments is currently limited to 16k.
I found no compares directly with user controlled constants.
I am not sure whether the test is that useful, but it might catch some changes missing constant splitting.
Review URL: http://codereview.chromium.org//7005031
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@7868 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
Ensure that there is always enough bytes between consequtive calls in
unoptimized code to write a call instruction at the return points
without overlapping.
This handles the case where two return points were only four bytes
apart (because the latter call was to a register).
BUG=v8:1234
Review URL: http://codereview.chromium.org/6624091
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@7089 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
Make esi available to the register allocator rather than dedicating it
permanently to the context.
The context is still passed in register esi to JavaScript and to the runtime
as part of the calling convention. Because some stubs might end up calling
JS or the runtime, it is also conservatively passed to stubs.
Roughly half the calls have been modified to use the context as an input
value in fixed register esi. The other half are marked as calls or deferred
code so esi is spilled and can be explicitly set.
It is no longer necessary to restore the context to esi after a call that
might change it.
Review URL: http://codereview.chromium.org/6452001
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@6713 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
There was an unnecessary load on every statically-resolved context lookup.
Remove it.
This revealed a hidden bug in const initializers inside 'with'. They claim
to be statically resolved (having slot type CONTEXT) but they occur in a
spot where the runtime context chain and the static scope chain do not
agree. This is fixed by special casing const initializers in the backend.
Review URL: http://codereview.chromium.org/6384020
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@6635 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
Change the ARM exit frame to have the same layout as the IA32 exit frame. This basically re-arranges the order of fp and sp and changes the sp location of the entry frame to hold the sp used by the gc and not the sp for popping the arguments. This removes the option of tearing down the frame and returning using one ldm instruction.
The main motivation for this is to avoid pushing an alignment word before generating the entry frame. The GC handling of optimized frames process the registers pushed as part of a safepoint and asumes that these are at the top of the frame, so if an alignment word is pushed this processing will be one off.
The alignment handling in the C entry stub have also been simplified. Now the value of lr is stored to a stack slot already reserved avoiding pushing it and keeping track of "frame skew".
This does result in more instructions in the exit frame on ARM, but we can look into improving this later.
Review URL: http://codereview.chromium.org/6247019
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@6448 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
The the call of the builtin in InstanceofStub was not correctly protected with an internal frame leading to the return address being handled as a pointer during GC.
Marked the Instanceof stub as allowing stub calls (the RecordWriteStub was removed some days ago).
This issue was not caught by the assertion designed for this when debug mode is run with --debug-code (which out tests always does) as generating code for Abort set the allow stub calls flag to true. This has been fixed by restoring the allow stub calls flag correctly.
Review URL: http://codereview.chromium.org/6097010
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@6218 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
inline allocation code used the expected number of properties to
calculate the number of inobject properties for an object instead of
getting the actual number from the initial map.
It is safer to use the inobject property count from the initial map in
any case because that is the amount the instances will get. I think
this disconnect got introduced when adding shrinking of objects.
Unfortuntely I haven't been able to create a simple reproduction for a
test case but this fixes the webpage that exhibits the crash. I'll see
if I can create a reproduction tomorrow.
Review URL: http://codereview.chromium.org/5278003
git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@5879 ce2b1a6d-e550-0410-aec6-3dcde31c8c00