Attachment "README" to
ticket [519535ffff]
added by
msofer
2002-02-19 03:41:49.
This is a very rough/experimental patch!
I - Changes in this patch:
* Incorporation of new functions with signature
Tcl_Obj * TclCoreFun (Tcl_Interp *interp, int objc,
Tcl_Obj *CONST objv[], int argc, ClientData *argv);
These will implement the core functionality of Tcl; the standard
calls act as front ends for them.
The usage is to allow receiving pointers to C structures that are
determined at proc-compile time - for example, to variables.
* Modification of the main variable access functions in tclVar.c to
follow this structure. Factoring out TclLookupVar.
* Definition of new instructions to
- lookup variable pointers given their index (locals) or name
- push non-Tcl_Obj items on the Tcl stack
- call a TclCoreFun
* test trial: [gets] is compiled using these interfaces, with a 20%
speed gain in the benchmarks:
000 VERSIONS: patched HEAD
001 READ 595K, gets 136482 175674
006 READ 3050b, gets 742 953
011 READ bin 595K, gets 145248 183942
016 READ bin 3050b, gets 803 1007
II - Notes:
* INST_GET_SCALAR, INST_SET_SCALAR, INST_INCR_SCALAR and
INST_FOREACH were changed to use directly the core functions
instead of going through the interface.
* the (incr) functions' interaction with the new 64bit patch has to
be thought out; as is here, it will probably bomb - the patch
assumes that you can store a "long" in a "ClientData"
* all hope of portability of the compiled code is gone, as it uses
pointers. A new save/load facility for compiled code has to be
thought out; it can be implemented via stubs tables.
III - Outlook:
* The details need some working out; however, it is conceivable
that the new engine will only address variables through a
combination of (INST_LOOKUP, INST_CALL_CFUN).
* This opens the path for compiling all instructions, performing
sub-command selection at compile time, and many other opts.
