CogVM binaries as per VMMaker.oscog-eem.1834/r3684
General:
Correct undue sign extension wile promoting 32 to 64 bits int in
fetchLong64:ofObject: This was causing trouble in non spur object memory (V3).
Fix slowdown of update on 64-bit x64 by using int variables for the tides in
sqExternalSemaphores.c and inline assembler for the sqCompareAndSwap in
sqAtomicOps.h.
Fix slips in offset time primitives. They need to pop argumentCount + 1, not
argumentCount.
Define an optional primitive as the fast primitive fail code so that unselected
optional primitives are in the prmitiveTable as fast primitive fails.
Spur:
Have WideString>>at: fail to answer an out-of-range character in the
interpreter primitive (as well as in the JIT).
Fix bug in following state on primitive failure. The old code would
always follow to depth 1, even if the accessor depth was 0.
Hard-code primitiveSize's depth to 0 (accessing length in at:[put:]
and size causes the stack depth computation to answer 1 instead of
0 for these primitives.
Fix assert and dequeueMourner for case where mournQueue is nil.
- fixed a bug in receiver accessing in immutability primitive for
mirror primitive.
- change primitiveSetOrHasIdentityHash to patch the class table if
the new hash is set to a behavior (the primitive knows it's a
behavior if the second (optional) argument is true)
For example:
FullBlockClosure tryPrimitive: 161 withArgs: {38.true}
No longer follow the method and context fields in a closure in
activateNewClosureMethod:numArgs:mayContextSwitch:; the caller will
have failed if these are forwarders, so no need to check again.
Cogit:
Add a primitive that answers pc map data for methods which can be
used to better decorate methods in the VM Profiler. Refactor the pc
map enumeration facilities so that the Sista pic data primitive can
share the same enumerator. Do this by collapsing the
isBackwardBranch and annotation parameters into a single parameter.
Follow selectors in the openPICList post Spur become.
Fix the ARM's caller-saved register mask now that we can name all the actual
registers.
Reworked machine code generation of immutability so for common stores it
uses a single trampoline for both store checks and immutability checks.
- improved support for register allocation: branch merge
successfully compiled with register moved instead of spilling.
Sista Cogit:
Don't bother to add counters to conditional jumps implementing and:
and or:. Added the remote inst var access bytecode in sista V1
bytecode set without interfering with existing code.
Plugins:
Upgrade LargeIntegersPlugin to v2.0
LargeInteger primitives now deal with 32-bits digits. No change to image code.
Memory is 8 bytes aligned on Spur. When storing 32/64 bits large
integers values, allways fill the eight bytes whatever the
effectivily used size, rather than bother with dissertion of size.
Generate integer type checking as C macros rather than direct/indirect
interpreterProxy function call in plugins. This, and 32-bit accessing mean
singificantly faster large integer arithmetic.
Fix primAlienReplace to use positiveMachineIntegerValueOf: instead of
positive32BitValueOf:.
BitBltPlugin operates on 32bit word units, therefore it's better to declare its
operands as 'unsigned int' rather than sqInt. On 32bit VM, this doesn't change
anything, but on 64bits spur, it makes this snippet work:
| wideString source pos blt expectedWideString |
source := #[1 64 255 14 1 64 48 251].
expectedWideString := WideString fromByteArray: source.
wideString := WideString new: source size // 4.
pos := 0.
blt := (BitBlt
toForm: (Form new hackBits: wideString))
sourceForm: (Form new hackBits: source).
blt
combinationRule: Form over;
sourceX: 0;
sourceY: pos // 4;
height: wideString byteSize // 4;
width: 4;
destX: 0;
destY: 0;
copyBits.
wideString restoreEndianness.
self assert: wideString = expectedWideString
Hence it fixes loading/diffing MCZ with wide character.
Mac OS Builds:
Don't link with -fvisibility=hidden; it breaks external plugins. Use
-fvisibility=default instead. This fixes e.g. UnixOSProcessPlugin on Mac OS X.
Windows Builds:
The Windows VMs are no longer "dpiAware". If you want one-to-one pixel mapping check the README for ins ructions. It's a simple edit.
Locked
results.csv (880 bytes)