stepharo wrote:
Hi eliot
Thanks!
I'm really eager to read about the new become implementation.
There is this...
http://www.mirandabanda.org/cogblog/2013/09/13/lazy-become-and-a-partial-read-barrier/
cheers -ben
I could not fully understand the following but this is exciting to
know that you removed complexity.
Stef
"The read barriers on method lookup (of the methodClass association in
super
sends, of the superclass link, of method dictionaries, method
dictionary arrays,
selectors and methods is cheap. So replace scanning classes and method
dictionaries in the class table post become with read marriewrs on
methodClass,
superclass and method dictionary etc on lookup.
The read barrier on an object from which we are going to fetch state
(such as a
class or method dictionary) is essentially free on modern machines
because the
class index and the state very likely share a cache line, and the
register code
for testing is so cheap compared to memory access. Further the read
barrier on
selectors is cheap because the method lookup cache is effective in
reducing the
number of message lookups and because nil entries need no check.
So nuke all the followNecessaryForwardingInMethod: machinery including
the
cmUsesMethodClass hack. Nuke scanClassPostBecome:effects:."
On 15/7/14 21:58, Eliot Miranda wrote:
...at http://www.mirandabanda.org/files/Cog/VM/VM.r3041/.
The principal changes are fixing a performance regression in the 3040
VMs, fixing some Spur bugs in primitiveClone/shallowCopy, and
radically simplifying the Spur become implementation,
trading the elimination of lots of post-become class table scanning
for adding read barriers to class hierarchy method lookup.
CogVM source as per VMMaker.oscog-eem.820/r3041
Put the handling of the cloning of cogged methods in the clone:
implementations,
removing it from the primitive. Add it to the pinning clone too.
Specialize the store check trampoline generation.
Move it down to the relevant object representations.
Move setting of isRemembered flag to true into SpurGenerationScavenger>>
remember:. Inline possibleRootStoreInto: (given that remember: is /not/
inlined. Call remember directly from the ceStoreCheckTrampoline, and
hence have remember: answer its argument.
Fix localization bug. Variables in initialize methods were not
considered
references (cuz there are excluded). This caused
VMMaker.oscog-eem.816's
extraction of zero/false vars to StackInterpreter>>#initialize to cause
nextPollUsecs to be localized to checkForEventsMayContextSwitch:.
Spur:
Fix bug in scanClassPostBecome:effects: with new lazy
selector following policy by... throwing it all away.
The read barriers on method lookup (of the methodClass association in
super
sends, of the superclass link, of method dictionaries, method
dictionary arrays,
selectors and methods is cheap. So replace scanning classes and method
dictionaries in the class table post become with read marriewrs on
methodClass,
superclass and method dictionary etc on lookup.
The read barrier on an object from which we are going to fetch state
(such as a
class or method dictionary) is essentially free on modern machines
because the
class index and the state very likely share a cache line, and the
register code
for testing is so cheap compared to memory access. Further the read
barrier on
selectors is cheap because the method lookup cache is effective in
reducing the
number of message lookups and because nil entries need no check.
So nuke all the followNecessaryForwardingInMethod: machinery
including the
cmUsesMethodClass hack. Nuke scanClassPostBecome:effects:.
Rip out the forwardingCount: measurement code. It causes
bad performance regressions (due to failing inlines?)
Clean up, e.g. replace followNonImmediateField:ofObject:
uses with followObjField:ofObject:.
canPinObjects can be inlined.
--
best,
Eliot
