Tcl Source Code

This patch modifies the use of ExceptionRange by the
compiler/engine in two ways:
 - moves the computation of the ending offset of the
range to compile time (instead of runtime)
 - exploits the order in which nested exception ranges
are stored to reduce the number of runtime comparisons
- it traverses the list of exception ranges in reverse
order, thus insuring that the deepest range is found
first.

Logged In: YES 
user_id=148712

Most of these ideas are now in the HEAD

Logged In: YES 
user_id=72656

pushing this back to miguel due to the number of changes he 
has made to TEBC recently.

File Added - 18961: newRangeNPeep.patch

Logged In: YES 
user_id=148712

I keep packing stuff in this patch ... The new
newRangeNPeep.patch adds

- peep-hole optimisation in INST_FOREACH4: avoids (creating,
pushing, popping, testing, destroying) an object and saves
executing an INST_JUMP_FALSE. 

- peep-hole optimisation in several instructions: avoids
pushing an object that will be popped right away and saves
executing an INST_POP.

The first column is the patched version:

000 VERSIONS:                    1:8.4a5 2:8.4a5
001 LOOP for (to 1000)              1328    1433
002 LOOP for, iterate list          1977    1993
003 LOOP for, iterate string        2631    2618
004 LOOP foreach, iterate list       872    1149
005 LOOP foreach, iterate string    1104    1385
006 LOOP while (to 1000)            1294    1427
007 LOOP while 1 (to 1000)          1005    1088
007 BENCHMARKS                   1:8.4a5 2:8.4a5

Remark that this (and many other optimisations) would be
simpler and more efficient if we settled on not being able
to run older style bytecode, and providing a "translator" in
tbcload. BTW, are the .tbc files versioned, or does proComp
only use the 8.0 compiler avoiding the newer instructions?

File Added - 18927: newRange2.diff

Logged In: YES 
user_id=148712

Improved patch, modifying also some ExceptionRange related
tools in tclCompile.c.

As a side effect, I applied the "loop rotation optimisation"
to [foreach], eliminating one jump from the loop.

File Added - 18891: newRange.diff

Logged In: YES 
user_id=148712

New patch: this one does only modifies the runtime lookup
algorithm, but does not touch the compiler. 

Checking with Jeff if .tbc code will not be broken, i.e., if
it satisfies the assumption:
  "nested ranges are always *after* their containing ranges"

The gains are illustrated with tcllib patch #526404
(tclbench):

TCL_INTERP: 1:8.4a4 2:8.4a4
STARTED 2002-03-06 11:38:10 (runbench.tcl v1.13)
Benchmark 1:8.4a4 /CVS/tcl_SF_clean/unix/tclsh0
c 00:00:00 elapsed
126 milliseconds
Benchmark 2:8.4a4 /CVS/tcl_SF_clean/unix/tclsh-newRange
c 00:00:00 elapsed
105 milliseconds
000 VERSIONS:            1:8.4a4 2:8.4a4
001 CATCH error, complex      58      47
002 CATCH no catch used        3       3
003 CATCH return error        28      28
004 CATCH return ok            4       3
004 BENCHMARKS           1:8.4a4 2:8.4a4
FINISHED 2002-03-06 11:38:10

Logged In: YES 
user_id=148712

The speed effect is just about nil (but positive); even so,
I still like it: it looks cleaner to my eyes.

OTOH, it may break tclPro's compiled bytecodes. Better leave
that alone for now.

Logged In: YES 
user_id=148712

The speed effect of the patch seems to be negative ...

File Added - 9785: except.diff