DKF: The aim of this page is to collect together stuff on how to make the pips squeak in Tcl. Anything for making the code faster (such as better idioms for compilation, discussions on reference counting, etc.) is fine, though long stuff should probably be shunted off somewhere else and only a link to it left here.
their favorite OO extension and update the page.
[time] can help answer questions like "Is this code faster than that code?"
In versions of Tcl prior to 8.4, using return to return a value was slower than simply arranging for the value to be the last thing in the procedure body. 8.4 corrected this. See Tcl 8.0 performance advice for the discussion, which was deleted from this page because it has been inaccurate for several years.
PSE: Wait a minute! What EXACTLY do you mean here?? Lets see what I get:
#! /bin/env tclsh
proc foo1 {} {return foo}
proc foo2 {} {uplevel set bar foo}
proc foo3 {} {set ::bar foo}
puts [time {foo1} 1000]
set bar {}
puts [time {foo2} 1000]
puts [time {foo3} 1000]31 microseconds per iteration 72 microseconds per iteration 15 microseconds per iteration
Hmmm... Tcl7.5 could do a return in 5 us. What happened?
DKF: The machinery to implement [return] is not all that simple, and has a fair old cost associated with it. However, in your code above you should be aware that the first example has completely different semantics to the other two (returns a value, as opposed to setting a variable in the surrounding/global context.) Compare your first example with:
proc foo {} {set foo foo}However, my workstation is sufficiently fast that measuring the difference between these is impossible (the margin of error is just too high!)
PSE: Ack. Yes. Same here. I see today that it is impossible to do meaningful benchmarks unless they are run as a script that runs the compared code in as rapid succession as possible, too much network related interference to make any sense of the results otherwise. Guess I need to do all my benchmarking on my Linux box at home and just figure that the sun boxes will give similar relative performance
I have my doubts about this optimisation. What I get (NT4.0, TclPro1.2) is:
% proc foo1 {} {return foo}
% proc foo2 {} {set foo foo}
% time {foo1} 1000000
4 microseconds per iteration
% time {foo2} 1000000
3 microseconds per iterationjp: Maybe this optimisation is obsolete, or its actual effect is too light to be under consideration?
DKF: Note that UNIX has a much smaller time granularity than Windows (certainly 95/98 and quite probably NT too.) The best way to benchmark a tiny snippet of Tcl (especially where it is platform-independent code you're testing) is to use a slow UNIX box. The slowness of its operation allows for sensitivity and the fine granularity of its timer lets you pick up what differences there are. And then you have to pray that the relative speeds scale as you move to different platforms/architectures and/or faster machines. Which is usually a good assumption if you're not shimmering between numbers and strings (where performance seems to vary a lot according to the vendor's implementation of the C library.)
Use braces around all conditionals to get the best efficiency out of the compiler. In Tcl7, one might have
#warning bad code ahead! if [...]
this should now be
if {[...]}[expr] expressions should also be braced. There will be a few cases where this won't work (building up exprs in a var).
Try to only ever evaluate constant strings. Where this is not possible, try to make sure not to change any particular string very often. Cache it if possible. Tk binding scripts are a prime example of how not to get maximum speed out of Tcl8.
TFW 2001-01-31: Here's a surprising timing reported by Tom Wilkason, <[email protected]>, on news:comp.lang.tcl on an Win 2K 375Mhz box (updated)
# Q is a local drive mounted as a share
time {cd q:/utilities;cd [file dirname [pwd]]} 1000
2163 microseconds per iteration
time {cd q:/utilities;cd q:/} 1000
2083 microseconds per iteration
# This is the same dir as above but from the raw mount
time {cd d:/source/utilities;cd [file dirname [pwd]]} 1000
601 microseconds per iteration
time {cd d:/source/utilities;cd d:/source} 1000
551 microseconds per iterationDKF 2000-02-14:
Curiouser and Curioser , since when I perform the timings with Tcl8.0.4 on my ultrasparc/Solaris box, I get the following timings...
% time {cd /home/fellowsd/arch; cd /home/fellowsd} 100000
48 microseconds per iteration
% time {cd /home/fellowsd/arch; cd [file dirname [pwd]]} 100000
129 microseconds per iteration
% time {cd /home/fellowsd/arch; cd [file join [pwd] ..]} 100000
131 microseconds per iterationSo things are not as straight-forward as they might seem. (Side note: is NT really so slow at changing directories, or is it just that the particular NT machine that the above test was done on is grossly underpowered anyway?)
DKF: You can find out just what the byte-code compiler is doing by activating internal tracing. You can set the Tcl variable tcl_traceExec to one of the following values, and internal trace functions will generate messages on stdout for every byte-code executed.
0: no execution tracing 1: trace invocations of Tcl procs only 2: trace invocations of all (not compiled away) commands 3: display each instruction executed
You can set the variable $tcl_traceCompile to one of the following values to get information during byte code compilation of a procedure or toplevel command.
0: no compile tracing 1: one line summary 2: detailed listing of byte codes
This is documented in the tclvars(n) reference , right at the obttom of the 8.0 version. Note that the output is always generated on stdout at a very low level. Mac people cannot do profiling (unless Apple fixes this in MacOSX) and Windows people will need to run from a console (which might be a problem with wish, IIRC.) Still, something that is only ever useful in some situations is almost always better than nothing at all.
DKF: Also, in Tcl 8.5 you can use [::tcl::unsupported::disassemble] to pull apart some code and show the bytecodes. No guarantees about formatting, etc, but it is very convenient and returns the result as a string so you can do further processing, show in a text widget, and so on.
This uses K to make [lreplace] work in-place.
proc K {x y} {set x}
set theList [lreplace [K $theList [set theList {}]] 7 42]See [L1 ] and Can some explain what is happening here? ,comp.lang.tcl ,2010-03-31
DKF: It seems from digging around that it depended on a discovery of mine from no later than 1999, i.e., that several Tclers were writing posts to c.l.t at the time describing it as a well-known technique. I've not yet identified what it started, but it may have started from Reversing List, comp.lang.tcl, 1998-11-21 . It's quite difficult to search back in the 1998–99 period; you keep finding irrelevant stuff from later in the way.
KBK: Donal, this deserves to be made more widely known; it just came up again as I was benchmarking Bob Techentin's code to Shuffle a list. It's incredible what a difference it made in the performance.
DKF: In 8.5 you can also do this:
set theList [lreplace $theList[set theList {}] 7 42]PYK 2012-10: MS explained on the IRC channel that $theList[set theList {}] is more effective than $theList[unset theList] because (1) you do not have to update the varname hashtable (assuming the variable is actually in a hastable); (2) you don't have to worry about the variable being at either end of a link, and (3) because [set] is bytecompiled but [unset] is not (too complicated). In short, setting to {} is faster because it does a lot less.
That depends on the fact that the low-level (bytecode) string concatenation operator knows not to do anything to an object when joining an empty string onto the end of it.
RWT: Just because I want to be able to find this again later... I've copied some notes from Jeff Hobbs. (To be updated when he rewrites it. :-)
Compress multiple Tcl commands as often as possible.
This is based on the idea that getting into the C side of Tcl as often as possible is the best way to go. Or, avoid running through Tcl_Eval as often as possible. This is often best used by compounding string calls into single regexps or magic use of regsubs. A classic example is the 9 line HTML parser from Stephen Uhler.
Use [string compare ...] when you want string comparisons in conditionals.
if {![string compare $a $b]} ...is faster than
if {$a == $b}but don't do this for numbers, because
0x3==3 is true
but
[string compare 0x3 3]==0
is not true (they aren't string equal)
Interestingly, in tcl8.2
if { [ string equal $a $b ] }is always nearly twice as slow as the two above. But in tcl8.3
if { [ string equal $a $b ] }
if { $a == $b }
if { ! [ string compare $a $b ] }are all within 4usec of eac other on solaris2.7, and 1usec on Windows NT. On solaris, the first two are always quicker than the last, and on Windows NT, the middle is slower than the other two. YMMV.
Warning! Before Tcl 8.4.x, [string compare] has did not always return a correct value! In fixing that problem, string compare is now as much as 30% slower than before. If you are wanting to compare two strings accurately, [string equal] is the best choice.
A JH example where [string compare] failed, which comes from the ActiveState ActiveTcl mailing list:
woset [~] 106 > echo 'puts [string compare \x00 \x01]' | tclsh8.3 1 woset [~] 107 > echo 'puts [string compare \x00 \x01]' | tclsh8.4 -1
don't change object representation without reason.
This starts with Tcl8.0 and the fact that lists and strings no longer have equal internal representations. Switching between these comes at a cost, and it is often easy to avoid when you become "one with the Tcl". For example, many program Tcl procs with an 'args'. This will come is as a list to your proc. Don't do
[string compare {} $args]to see if it has something in it, instead do
[llength $args].
Doing the string compare would convert the $args list to a string, then back to list when it was actually used. Keep this in mind for numbers as well, with the caveats of point 2 (incr is faster if you keep treating the var as a number).
[lindex] is now O(1) on lists that already have an internal representation. This also assists in things like [lrange] and [lreplace], and even [lappend]. If your indices are numeric, the list is significantly faster than arrays.
Note: I believe it was Paul Duffin that has presented us with a full Tcl_Obj'ified version of arrays, which will also speed up the array implementation in Tcl8.
KBK: But be aware that you may inadvertently copy a Tcl_Obj if you're not careful, and destroy the performance advantage. See Shuffle a list for an example.
It's faster than [for] and [lindex]. This effect is particularly pronounced before 8.0 (since you only had to parse a string into a list once) but it is still noticeable after.
set cities [list Rochester Minneapolis St.Paul Mankato Bemidji]
set ncities llength $cities for {set i 0} {$i<$ncities} {incr i} {
set city [lindex $cities $i]
# This is slower}
foreach city $cities {
# This is faster, and more elegant
}
Even if you need to know the item index, the following foreach idiom is faster than the for form:
set i -1; foreach city $cities {
incr i
# Normal body comes here
}The [incr] has to come first in the body if the body might do a [continue`], as the index would otherwise not always be updated to match.
[foreach] can operate on several lists in parallel:
set temperatures [list -10 -12 -14 -18 -32]
foreach city $cities temp $temperatures {
# Fast, elegant, and even cool
}
# with those temperatures, its downright chilly - glennjOnce a proc is first compiled (upon first use), any amount of whitespace and commenting in it is removed in the compiled representation. The load time for heavily commented procs will be slightly longer (Tcl has to read all that in before it knows what to avoid), but after its first call, Tcl will byte-compile it, and use that rep. Of course, if you keep changing the procs around, you're going to have problems. I believe there are also points like redefining certain Tcl core commands that invalidate all byte-compiled proc representations.
Combining points "multiple commands" and "don't change objects", we can reach the conclusion that
foreach {a b c d e} [cmdReturningList] {break}is faster than
set list [cmdReturningList] set a [lindex $list 0] set b [lindex $list 1] ....
The break is helpful to ensure that we don't actually do any loops, in case [cmdReturningList] is changed to return more than 5 args. Note that when we want the result from two similar calls, we get even better advantages out of:
foreach {a b c} [cmdRetList] {a' b' c'} [cmdRetList'] {break}Note: 8.5 onwards, you can use the more readable [lassign].
AMG: When 8.5 isn't available and I use [foreach] in this way, I usually skip the [break] when the list is guaranteed to iterate only once.
Yes, accessing stuff in the hash table does take a little longer, so if you have a number that you want to loop 1000 times on, put it into a simple variable first.
Inline Tcl code doesn't get all the optimizations that a procedure can get. For example, this command outside a proc
for {set i 0} {$i<10000} {incr i} {lappend a $i}takes three times longer (on my machine, running Tcl 8.0.4) than
proc init_me {} {
global b
for {set i 0} {$i<10000} {incr i} {lappend b $i}
}
init_meThe differences can be even more dramatic in Tcl 8.1, because the compiler was split into a parser front-end and a compiler back-end. Global code is just parsed into tokens, similarly to the way that Tcl operated before 8.0 as this offers a smaller performance hit when running scripts that only execute once (notably Tk bindings, but many other kinds of callback come into this category too.) Only [proc] and [apply] bodies are fully compiled.
Regular expressions are expensive to compile, and non-constant regexps must be compiled every time they are scanned for.
Suppose you want to check for the case where there is "a b c" with arbitrary (including none) amounts of whitespace between the letters. In 8.0, the following is the most obvious way to do it, but it is hard to write:
regexp "a\[ \t\n\r\]*b\[ \t\n\r\]*c\[ \t\n\r\]" $string
For clarity's sake it is far better to do:
set ws "\[ \t\n\r\]"
regexp "a${ws}*b${ws}*c" $stringHowever, this is not very efficient since it forces recompilation each time (unlike the first version where the string is constant, and therefore can support Tcl_Obj-based saving of compiled values.) The following does not suffer from this problem, and is probably clearer as well in practical programs:
# Only execute these lines once, at startup time
set ws "\[ \t\n\r\]"
set abcRE "a${ws}*b${ws}*c"
# Now refer to the (probably global) variable each time you want to match
regexp $abcRE $stringIf you are using 8.1 or later, you have access to the new regular expression engine, which allows several things to make this bite less (special escapes for common sequences, regular expression cache (in later versions), and implemented cacheing of REs in Tcl_Objs...)
The best solution for our example problem is probably
set pat {a\s*b\s*c} ;# Globally
regexp $pat $string ;# Each time neededSee comment below for additional speedup.
In general, do whatever you can to avoid losing the object that contains the compiled expression.
Solution: We need to change the 8.1 implementation to introduce a second level cache similar to the one in 8.0. This will allow us to avoid recompilation in cases where the same computed string value is passed in, even when it comes from a different object. We can probably use a thread-local cache, instead of a per-interp cache to improve sharing and make the regexp code more modular.
-- JC
Actually, for performance sake, the best solution is to inline the pattern:
regexp {a\s*b\s*c} $stringThis will be even faster (yes, I've verified this with timings). For simple regexps, this is more readable anyway since it keeps the regexp definition where it is being used.
Another significant way to speed up regular expressions is to avoid making the regexp retry matches (i.e., backtrack) if possible. As a simple case, consider this:
regexp {a.*b.*c} "abbbbbbbbbbbbbbbbbbbb"In this statement, the regexp engine is going to match the "a", then the "b" and then fail to match the "c". But it won't stop there, it will match every single b and using those matches will try (and fail to match the c in every position (where there are b's)). This is extremely inefficient. It is much more efficient to break this into to two searches so that you can prevent the regexp engine from all the pointless work backtracking.
This is just the tip of the iceberg, alas.
-- DL
JH: I think the above backtrack issue went away with the updated 8.1+ RE, but the problem does still exist in 8.4.16/8.5b2 for the equivalent glob case:
string match "*a*b*c*" "abbbbbbbbbbbbbbbbb"
That you end up with pointless backtracking.
In Tcl 8.1 and above, this has been fied, and [read] performance on large files is satisfactory. See the READ tests on the Tcl Benchmarks page.
One common Tcl-ism is to read an entire data file into memory in one fell swoop, instead of reading and processing each line in turn. This works well if your file is small compared to system memory, and you can afford to do all the processing in memory. You can also use the split command to break the file down into a list of individual lines, like this:
set fp [open myFile r] set data [split [read $fp] "\n"] close $fp
But for larger files (say in the megabyte range) Tcl 8.0's read can be very slow because it is using a buffer size of 4096 bytes. You can improve performance by increasing the channel buffer size with [fconfigure], but the most efficient way is to tell [read] exactly how much data it will be reading.
set fp [open "myFile" r] set size [file size "myFile"] set data [split [read $fp $size] "\n"] close $fp
Patterns for Scripted Applications , by Nat Pryce, particularly Bootstrap Script.
From <news:[email protected] > by Scott Stanton comes this Tcl 8.1 performance tip:
"... you should not use the global "env" array as a general data storage area. There is a lot of mechanism behind "env" that is needed to keep it in sync with multiple interps and the C environ array."
set someLetters {[abcxyz]+}
# The bad way to do it
foreach string $someList {set m($string) [regexp "${someLetters}0${someLetters}" $string]}
# The good way to do it
set RE "${someLetters}0${someLetters"
foreach string $someList {set m($string) [regexp $RE $string]}DKF
DKF: OK, the string performance bug has now been fixed properly in 8.1.2, and virtually all operations are now as fast in 8.1.2 as in 8.0.* Much kudos to Ajuba Solutions for fixing it so quickly and in such a non-blecherous fashion. [L3 ]
Comparing various methods of summing a list of numbers:
#! /bin/env tclsh
for {set i 15} {$i<100} {incr i} {
lappend L $i
}
set iterations 10000
set trial 0
proc sumup list {
set sum 0; foreach n $list {incr sum $n}; set sum
}
set time [time {set sum [sumup $L]} $iterations]
puts "sumup : $time, $sum"
set time [time {set sum [expr [join [concat 0 $L] +]]} $iterations]
puts "join/concat: $time, $sum"
proc reduce {f i l} {
foreach e $l {
set i [$f $i $e]
}
return $i
}
proc sum {x y} {expr {$x+$y}}
set time [time {set sum [reduce sum 0 $L]} $iterations]
puts "reduce : $time, $sum"
Results with Tcl-8.6, Intel Core 2 Duo, 3.06 GHz, OS X 10.6.8:
mathop : 7.2078723 microseconds per iteration, 4845 sumup : 12.1550874 microseconds per iteration, 4845 join/concat: 43.089464500000005 microseconds per iteration, 4845 reduce : 43.4680041 microseconds per iteration, 4845
DKF: In 8.2 (and presumably 8.1 as well) it seems that the bytecode compiler is only used some of the time (the most notable time is when you have a procedure body, of course) so any benchmark - it is usually benchmarks that hit this - that does not use the compiler will take a noticeable hit, especially when they use a loop of some kind. Avoid this by the use of procedures wherever possible.
Thus, this is slow:
for {set a 0} {$a<100000} {incr a} {expr $a + 77}But this is much faster:
proc main {} {
for {set a 0} {$a<100000} {incr a} {
expr $a + 77
}
}
mainAnd this is faster still (due to expression optimisation):
proc main {} {
for {set a 0} {$a<100000} {incr a} {
expr {$a + 77}
}
}
mainThe tip for walking down lists applies also to strings (amazingly?)
#! /bin/env tclsh
set str {
The tyrannous and bloody deed is done.
The most arch of piteous massacre
That ever yet this land was guilty of.
Dighton and Forrest, whom I did suborn
To do this ruthless piece of butchery,
Although they were flesh'd villains, bloody dogs,
Melting with tenderness and kind compassion
Wept like two children in their deaths' sad stories.
'Lo, thus' quoth Dighton, 'lay those tender babes:'
'Thus, thus,' quoth Forrest, 'girdling one another
Within their innocent alabaster arms:
Their lips were four red roses on a stalk,
Which in their summer beauty kiss'd each other.
A book of prayers on their pillow lay;
Which once,' quoth Forrest, 'almost changed my mind;
But O! the devil'--there the villain stopp'd
Whilst Dighton thus told on: 'We smothered
The most replenished sweet work of nature,
That from the prime creation e'er she framed.'
Thus both are gone with conscience and remorse;
They could not speak; and so I left them both,
To bring this tidings to the bloody king.
And here he comes.
}
set iterations 10000
set time [time {
foreach {a} [split $str {}] {
set b $a
}
} $iterations]
puts "split: $time"
set time [time {
for {set a 0} {$a < [string length $str]} {incr a} {
set b [string index $str $a]
}
} $iterations]
puts "index: $time"Results with Tcl-8.6, Intel Core 2 Duo, 3.06 GHz, OS X 10.6.8:
split: 597.7267553 microseconds per iteration index: 967.8046357000001 microseconds per iteration
DKF: OK, today's goodies are a little discussion of one of the speed optimisations in Tcl8.3.
It is part of the definition of canonically-formatted lists that when you evaluate them, the first word of the list will be the name of the command, the second word of the list will be the first argument, etc. But in 8.0, 8.1 and 8.2 this case was always handled by converting to a string (putting in all the appropriate quoting) and then evaluating (which just strips all the quoting back out again before handing off to the command handler in question.) Which is stupidly slow, especially when you have many arguments that would not otherwise have string forms.
So, for any xxx, yyy and zzz (including variable and command substitutions) the following two lines are precisely equivalent:
xxx yyy zzz eval [list xxx yyy zzz]
So, to fix this, 8.3 short-cuts this and hands off the contents of the list directly to the command handler when it knows that doing so will not change the semantics of the language. Which turns out to be exactly when the command to evaluate is a pure list (i.e. a list whose object form does not have a string representation) as that is the only time you can prove that there are no hidden gotchas lurking in the string form because anything unpleasant. This gives us a nice implementation of [lconcat] that is nearly as efficient as a C version (whose only advantage would be that they could know to pre-allocate the correct length of result list before copying.)
proc lconcat args {
set result [list]
foreach list $args {
eval [linsert $list 0 lappend result]
}
set result; # FASTER THAN RETURN
}To make this sort of thing easier to write, there is an additional optimisation for [concat] (and other places that use concat internally, like eval) so that where all its arguments are pure lists, the result is a pure list formed by sticking the contents of the arguments together in the obvious way. The only reason I didn't use that above (which would have been noticeably clearer, admittedly) is that this prevents trouble with non-pure list arguments.
DGP: Note that this performance optimization is available for [uplevel] as well as [eval], but to take advantage of it you must explicitly include the optional level argument of [uplevel]:
proc test args {
uplevel 1 $args ;# Can use pure list optimization
uplevel $args ;# Attempt to interpret $args as a level
;# generates a string rep, which prevents
;# use of the pure list optimization
}DKF:
Remember, the actual numerical performance of computers varies widely with operating system and processor speed, and comparing actual figures is very tricky. If you're going to put up benchmarking info, at least do everyone the favour of Finding Out Your Processor and Operating System Configuration...
See Shuffle a list for an example.
File I/O performance is a considerable subject in its own right.
The possible impact on performance of different ways to use indirection to call a proc is (crudely) measured in Procedure calling timed and Speed of different method dispatching methods.
See Can you run this benchmark 10 times faster for an example of how to use self-modifying scripts to improve performance.
The performance of the -command option of lsort is awful. Avoid it wherever possible; the lsort page shows how.
Has anyone taken a good look at http://shootout.alioth.debian.org/ and the reported performance of Tcl to see if either
DKF: It is often interesting (both from a performance and a debugging point-of-view) to see the bytecode for a piece of code. Here's how with 8.4:
proc traced {script {compileLevel 2} {execLevel 3}} {
proc traced_body {} "set ::tcl_traceExec 0\n\
set ::tcl_traceCompile 0\n$script"
set ::tcl_traceExec $execLevel
set ::tcl_traceCompile $compileLevel
uplevel 1 traced_body
}traced {foreach {a b c} [list 1 2 3 4 5 6] {puts [expr {$a+$b/$c}]}}What could be easier?
Dossy 2003-08-31: In dealing with arrays, I find that I often have non-trivial references to arrays. Sometimes, the array name and the element are both stored in seperate vars, so when I need to access the array value, I have two idioms:
set a(key) value set b a set c key set value [set [set b]($c)]
OR
set value [set $b\($c)]
Obviously, I'm curious which one people prefer and consider "idiomatic Tcl" and which one performs better. The first question I'm hoping to answer by posting this question to the wiki and gathering some opinions, but the latter I could do by benchmarking:
proc foo {n} {
set a(j) abc
set b a
set k j
for {set i 0} {$i < $n} {incr i} {
set $b\($k)
}
}
proc bar {n} {
set a(j) abc
set b a
set k j
for {set i 0} {$i < $n} {incr i} {
set [set b]($k)
}
}
% info patchlevel
8.4.3
% time "foo 1000000"
2504742 microseconds per iteration
% time "bar 1000000"
2523562 microseconds per iterationIt's not significant, but it seems the "$arrayName\($arrayKey)" is minimally faster.
So, if performance isn't a concern, which style do people prefer? Me, being a purist, I prefer the "set [set arrayName]($arrayKey)" way.
JMN: It hadn't occurred to me to use:
set value [set $b\($c)]
I tend use to use
set value [set ${b}($c)]for a once-off as it's a more general way to specify the extent of a variable and I like to keep it consistent with other substitution situations such as:
set x ca
puts ${x}bwhere puts $x\b of course doesn't give the desired effect.
If accessing an indirected array variable in a loop where performance matters, I'd use upvar to bring it back to 'normal' syntax. It seems significantly faster.
proc baz {n} {
set a(j) abc
set b a
set k j
upvar 0 $b arr
for {set i 0} {$i < $n} {incr i} {
set arr($k)
}
}Dossy: Julian, excellent points. And, yes, upvar does seem the way to go:
% time "baz 1000000" 471348 microseconds per iteration
That's a huge difference. Neat.
JMN: If startup time of a Tcl/Tk application is a concern, consider that 'package require' of a package that doesn't exist on your system can take a relatively long time if you have a large number of packages installed on your auto_path. Some packages try to load other packages using something like:
if {[catch {package require tcllibc}]} {On one of my systems such calls can take around 900msec
If you're using Tcl >= 8.5, consider the use of tcl modules to reduce the time taken by the system in loading packages.
1S ... Or, you could check Gathering packages' ifneeded scripts as well.
[To write up...]