Ok. I've recently converted much of partcl's test suite over to "pure
tcl". Instead of using a perl based test script, the tests are now
written in tcl, using a very small test::more like script. Now we can
more easily run the tests from tcl's own test suite, and easily use
tclsh to run our own tests.
This also turned out to be a speed win for the smaller test scripts:
only one parrot had to be invoked.
However, for larger test files (like t/cmd_expr.t), it's a definite
slowdown.
If we run:
../../parrot tcl.pbc --pir t/cmd_expr.t
we get a 9K line PIR sub. Running this through parrot dies after
about 2.5m on an runtime error on one of the tests.
Here's a snippet from near the end of the code:
start_1559:
dynamic_1559:
.local pmc command
$P1566 = new .String
$P1566 = 'foreach'
push_eh invalid_1559
command = get_hll_global '&foreach'
clear_eh
if_null command, invalid_1559
$P1565 = command($P1560, $P1561, $P1563)
goto end_1559
invalid_1559:
.local pmc interactive
interactive = get_root_global ['tcl'], '$tcl_interactive'
unless interactive goto err_command1559
.local pmc unk
unk=find_global '&unknown'
unk($P1566, $P1560, $P1561, $P1563)
goto end_1559
err_command1559:
$S0 = $P1566
$S0 = concat "invalid command name \"", $S0
$S0 .= "\""
.throw($S0)
end_1559:
.return ($P1565)
note the .local declarations there - command, interactive, unk -
they're duplicated many times throughout the code. In an effort to
make this better (both more correct and faster), I made a local
modification to the compiler to avoid re-using named pmcs. So, with
this change, we get:
start_1559:
dynamic_1559:
.local pmc command_1559
$P1566 = new .String
$P1566 = 'foreach'
push_eh invalid_1559
command_1559 = get_hll_global '&foreach'
clear_eh
if_null command_1559, invalid_1559
$P1565 = command_1559($P1560, $P1561, $P1563)
goto end_1559
invalid_1559:
.local pmc interactive_1559
interactive_1559 = get_hll_global '$tcl_interactive'
unless interactive_1559 goto err_command1559
.local pmc unk_1559
unk_1559=get_hll_global '&unknown'
unk_1559($P1566, $P1560, $P1561, $P1563)
goto end_1559
err_command1559:
$S0 = $P1566
$S0 = concat "invalid command name \"", $S0
$S0 .= "\""
.throw($S0)
end_1559:
So, this adds (rough guess) about 4600 unique variables to the sub.
This makes the run time go from about 2.5m to nearly 7m. (until the
same eventual runtime failure which isn't an issue for the purposes
of this post. =-).
So... not sure if this is more a question for the partcl end (how can
we generate PIR more suitable to parrot), or for the parrot end (how
can we more efficiently handle this kind of PIR).
I know the usual caveats about speed (get it working first), but this
is one of those times where the current speed makes it somewhat
difficult.
One potential change to parrot was suggested by Matt Diephouse on
IRC: have a way to declare "this variable is no longer used, I
promise" to limit the scope of its consideration when doing register
allocation.
Thoughts?
--
Will "Coke" Coleda
[EMAIL PROTECTED]
