New Cog VMs available

Oops, for http://www.mirandabanda.org/files/Cog/VM/VM.r3163/ read http://www.mirandabanda.org/files/Cog/VM/VM.r3164

On Sat, Dec 6, 2014 at 1:53 PM, Eliot Miranda <[hidden email]> wrote:

... at http://www.mirandabanda.org/files/Cog/VM/VM.r3163/.

CogVM binaries as per VMMaker.oscog-eem.972/r3163

Fix a devilishly tricky bug with 32-bit Spur and PICs in the Cogit.  In 32-bit

Spur only zero is a valid cache tag, that for Characters, due to SmallIntegers

having both 2r11 and 2r01 as valid tags.  The cache check prolog collapses

these down onto 1 by anding with 1 and hence collapses Character's 2r10 to 0.

The cache tag is also tested in the PIC abort routine to distinguish between

an MNU abort and an interpret abort.  A PIC MNU entry jumps to code that first

zeros the cache tag (held in ClassReg) before calling the abort routine, which

tests ClassReg and selects the relevant abort processing.

Hence in 32-bit Spur if a send site is linked for Character and later extended

to a PIC with an interpret case, that interpret case will instead invoke MNU

processing.  This is rare; first there needs to be a send to a CHaracter, then

a send to something else which binds to something that must be interpreted.

Ouch.

The fix is simple; if a send site that was linked with a zero cache tag (which

can only happen in 32-bit Spur) link to an open PIC instead of to a closed PIC,

or in the case of an MNU do not link.

Add a debug lookup routine that only looks up (and follows forwarding pointers).

Add a -breakmnu vm argument that calls warning on an mnu of a specific selector.

Fix SmartSyntaxPlugin code generation for Spur of

- asOop: old code assumed 4 byte oops & headers

- Unsigned old code assumed 32-bit oops

Affects AsyncFilePlugin, BochsIA32Plugin & FT2Plgin.

Avoid duplication of effort in special selector send forwarding fault handling.

Improve stack trace printing for Newspeak and Glue by providing something that

assumes the penultimate literal is the selector or an AdditionalMethodState-like

thing containing it.

Reimplement primitiveCopyObject to work for both pointer and bits objects.

This to support a good Object>>clone for Spur.

Fix Spur's isWordsOrBytes[NonImm]:.  Old code answered true for CompiledMethods.

--

best,

Eliot

--

best,

Eliot

ARMv6 binaries updated accordingly.

On Dec 6, 2014, at 19:17 , Eliot Miranda wrote:

Oops, for http://www.mirandabanda.org/files/Cog/VM/VM.r3163/ read http://www.mirandabanda.org/files/Cog/VM/VM.r3164

On Sat, Dec 6, 2014 at 1:53 PM, Eliot Miranda <[hidden email]> wrote:

... at http://www.mirandabanda.org/files/Cog/VM/VM.r3163/.

CogVM binaries as per VMMaker.oscog-eem.972/r3163

Fix a devilishly tricky bug with 32-bit Spur and PICs in the Cogit.  In 32-bit

Spur only zero is a valid cache tag, that for Characters, due to SmallIntegers

having both 2r11 and 2r01 as valid tags.  The cache check prolog collapses

these down onto 1 by anding with 1 and hence collapses Character's 2r10 to 0.

The cache tag is also tested in the PIC abort routine to distinguish between

an MNU abort and an interpret abort.  A PIC MNU entry jumps to code that first

zeros the cache tag (held in ClassReg) before calling the abort routine, which

tests ClassReg and selects the relevant abort processing.

Hence in 32-bit Spur if a send site is linked for Character and later extended

to a PIC with an interpret case, that interpret case will instead invoke MNU

processing.  This is rare; first there needs to be a send to a CHaracter, then

a send to something else which binds to something that must be interpreted.

Ouch.

The fix is simple; if a send site that was linked with a zero cache tag (which

can only happen in 32-bit Spur) link to an open PIC instead of to a closed PIC,

or in the case of an MNU do not link.

Add a debug lookup routine that only looks up (and follows forwarding pointers).

Add a -breakmnu vm argument that calls warning on an mnu of a specific selector.

Fix SmartSyntaxPlugin code generation for Spur of

- asOop: old code assumed 4 byte oops & headers

- Unsigned old code assumed 32-bit oops

Affects AsyncFilePlugin, BochsIA32Plugin & FT2Plgin.

Avoid duplication of effort in special selector send forwarding fault handling.

Improve stack trace printing for Newspeak and Glue by providing something that

assumes the penultimate literal is the selector or an AdditionalMethodState-like

thing containing it.

Reimplement primitiveCopyObject to work for both pointer and bits objects.

This to support a good Object>>clone for Spur.

Fix Spur's isWordsOrBytes[NonImm]:.  Old code answered true for CompiledMethods.

--

best,

Eliot

--

best,

Eliot