Thread: Re: [TCLCORE] [Ext] Re: [=] for concise expressions (was Re: TIP 672 Implementation Complete - Read (Page 2)

SourceForge Headquarters 1320 Columbia Street Suite 310 San Diego, CA 92101 +1 (858) 422-6466

Hi Eric

I really like the tcl::unsupported::assemble approach.
It opens the way to add more commands/sugar to tcl itself without
having to hard code it in C.

Now we have to agree on the syntax. May be a wiki page with a summary of the different proposals?

Regards
rene

Von: EricT <tw...@gm...>
Gesendet: Freitag, 7. November 2025 01:03
An: Zaumseil René <RZa...@kk...>; tcl...@li...
Betreff: Re: [TCLCORE] [Ext] Re: [=] for concise expressions (was Re: TIP 672 Implementation Complete - Ready for Sponsorship)

Hi Rene,

I was experimenting with your TIP 674 prototype and realized both your let command and Colin's expression evaluator share something important: they could both be implemented in pure Tcl if tcl::unsupported::assemble were promoted to supported status with handle-based caching.

This would avoid another round of "expr wars" where lack of consensus leaves us with nothing. Instead of debating whose syntax is better, we'd each have the power to create our own Design-Specific Language. With bytecode caching, pure Tcl implementations achieve C extension performance - Colin's evaluator runs at 1.8 microseconds, nearly as fast as expr.

Why does this matter for your TIP? If let performs just as well in Tcl as in C, why require a C implementation at all? C extensions take weeks to develop, need binaries for every platform/architecture combination, and ultimately call the same bytecode infrastructure that assemble uses. A supported bytecode API would make both proposals - and many others - viable without C.

Perhaps we should advocate together for making the bytecode compilation infrastructure officially supported, rather than competing for TIP approval on individual commands?

Best,

Eric

On Wed, Nov 5, 2025 at 11:31 PM Zaumseil René via Tcl-Core <tcl...@li...<mailto:tcl...@li...>> wrote:
Hello

I also like the = approach as a new expr command without need of $ for variables.
But I'm not sure about the currently proposed syntax.
Is it "= 1 + 2" or "= 1+2" (with or without spaces)?
If it is the later then I would propose to use some of the syntax of the "let" from tip 674.

Some syntax ideas:
"= 1+2" return 3, simple case
"= a 1+2 b 2+4 c a+b" return {3 6 9} and set a,b,c
"= 1+2 = 2+4" return {3 6}
And may be some others…

Regards
rene
Von: EricT <tw...@gm...<mailto:tw...@gm...>>
Gesendet: Mittwoch, 5. November 2025 23:22
An: Colin Macleod <col...@ya...<mailto:col...@ya...>>; tcl...@li...<mailto:tcl...@li...>
Betreff: [Ext] Re: [TCLCORE] [=] for concise expressions (was Re: TIP 672 Implementation Complete - Ready for Sponsorship)

Hi Colin,

I've successfully modified your amazing code to handle arrays. In doing so, I also found 2 other issues, one is with your Boolean check, the other with your function name check, both because of [string is] issues.

- Boolean check: `$token eq "false" || $token eq "true"` (was `[string is boolean $token]` - treated 'f','n', 't', 'y', etc. variables as boolean false, no, true, yes, ...)

- Function check: `[regexp {^[[:alpha:]]} $token]` (was `[string is alpha $token]` - broke log10, atan2)

here's the code for arrays:

    # Function call or array reference?

    set nexttok [lindex $::tokens $::tokpos]

    if {$nexttok eq "(" && [regexp {^[[:alpha:]]} $token]} {

        set fun [namespace which tcl::mathfunc::$token]

        if {$fun ne {}} {

            # It's a function

            incr ::tokpos

            set opcodes "push $fun; "

            append opcodes [parseFuncArgs]

            return $opcodes

        } else {

            # Not a function, assume array reference

            incr ::tokpos

            set opcodes "push $token; "

            # Parse the index expression - leaves VALUE on stack

            append opcodes [parse 0]

            # Expect closing paren

            set closing [lindex $::tokens $::tokpos]

            if {$closing ne ")"} {

                error "Calc: expected ')' but found '$closing'"

            }

            # Stack now has: [arrayname, indexvalue]

            incr ::tokpos

            append opcodes "loadArrayStk; "

            return $opcodes

        }

    }

In addition, there has indeed been some changes in the bytecode, land and lor are no longer supported in 9.0 although they work in 8.6.

I had an AI generate some 117 test cases, which all pass on 8.6 and 111 on 9.x (the land/lor not being tested in 9.x).

Colin, with your permission, I can post the code as a new file, with all the test cases, say on a repository at github.

I think a new TIP is worth considering; one that promotes assemble to a supported form, with a compile and handle approach to avoid the time parsing the ascii byte code text. I think that this would be great for your = command, but also quite useful for others who might want to create their own little languages.

By doing it this way, it remains pure tcl, and avoids all the problems with different systems and hardware that a binary extension would create. In the end, I believe your code can achieve performance parity with expr. Not only does it remove half the [expr {...}] baggage, but all the $'s too! So much easier on these old eyes.

Regards,

Eric

On Tue, Nov 4, 2025 at 1:06 PM EricT <tw...@gm...<mailto:tw...@gm...>> wrote:
Hi Colin,

Hmmm, why can't you do bareword on $a(b) as a(b) you just need to do an uplevel to see if a is a variable, if not, it would have to be a function. True?

% tcl::unsupported::disassemble script {set a [expr {$b($c)}] }
snip
  Command 2: "expr {$b($c)}..."
    (2) push1 1 # "b"
    (4) push1 2 # "c"
    (6) loadStk
    (7) loadArrayStk
    (8) tryCvtToNumeric
    (9) storeStk
    (10) done
This doesn't look too much different from what you are producing.

I think what's really needed here is a TIP that would open up the bytecode a bit so you don't need to use an unsupported command. And then maybe even have a new command to take the string byte code you are now producing and return a handle to a cached version that was probably equivalent to the existing bytecode. Then your cache array would be

 set cache($exp) $handle

Instead of it having to parse the text, it could be as fast as bytecode. You'd likely be just as fast as expr, and safe as well, since you can't pass a string command in where the bareword is required:

 % set x {[pwd]}
[pwd]
 % = sqrt(x)
exp= |sqrt(x)| code= |push ::tcl::mathfunc::sqrt; push x; loadStk; invokeStk 2; | ifexist: 0
expected floating-point number but got "[pwd]"
I think you really have something here, perhaps this is the best answer yet to slay the expr dragon!

Regards,

Eric

On Tue, Nov 4, 2025 at 6:52 AM Colin Macleod via Tcl-Core <tcl...@li...<mailto:tcl...@li...>> wrote:

Hi Eric,

That's very neat!

Yes, a pure Tcl version could go into TclLib. I still think it may be worth trying a C implementation though.  The work-around that's needed for array references [= 2* $a(b)] would defeat the caching, so it would be good to speed up the parsing if possible.  Also I think your caching may be equivalent to doing byte-compilation, in which case it may make sense to use the framework which already exists for that.

Colin.
On 04/11/2025 01:18, EricT wrote:
that is:

if {[info exist ::cache($exp)]} {
    tailcall ::tcl::unsupported::assemble $::cache($exp)
}

(hate gmail!)

On Mon, Nov 3, 2025 at 5:17 PM EricT <tw...@gm...<mailto:tw...@gm...>> wrote:
and silly of me, it should be:
if {[info exist ::cache($exp)]} {
    tailcall ::tcl::unsupported::assemble $::cache($exp)
}

On Mon, Nov 3, 2025 at 4:50 PM EricT <tw...@gm...<mailto:tw...@gm...>> wrote:
With a debug line back in plus the tailcall:

proc = args {
    set exp [join $args]
    if { [info exist ::cache($exp)] } {
         return [tailcall  ::tcl::unsupported::assemble $::cache($exp)]
    }
    set tokens [tokenise $exp]
    deb1 "TOKENS = '$tokens'"
    set code [compile $tokens]
    deb1 "GENERATED CODE:\n$code\n"
    puts "exp= |$exp| code= |$code| ifexist: [info exist ::cache($exp)]"
    set ::cache($exp) $code
    uplevel [list ::tcl::unsupported::assemble $code]
}

% set a 5
5
 % set b 10
10
 % = a + b
exp= |a + b| code= |push a; loadStk; push b; loadStk; add; | ifexist: 0
15
% = a + b
15

% time {= a + b} 1000
1.73 microseconds per iteration
Faster still!

I thought the uplevel was needed to be able to get the local variables, seems not.

 % proc foo arg {set a 5; set b 10; set c [= a+b+arg]}
 % foo 5
exp= |a+b+arg| code= |push a; loadStk; push b; loadStk; add; push arg; loadStk; add; | ifexist: 0
20
% foo 5
20

 % proc foo arg {global xxx; set a 5; set b 10; set c [= a+b+arg+xxx]}

 % set xxx 100
100
 % foo 200
315
 % time {foo 200} 10000
2.1775 microseconds per iteration

 % parray cache
cache(a + b)       = push a; loadStk; push b; loadStk; add;
cache(a+b+arg)     = push a; loadStk; push b; loadStk; add; push arg; loadStk; add;
cache(a+b+arg+xxx) = push a; loadStk; push b; loadStk; add; push arg; loadStk; add; push xxx; loadStk; add;

Very Impressive, great job Colin! Great catch Don!

Eric

On Mon, Nov 3, 2025 at 4:22 PM Donald Porter via Tcl-Core <tcl...@li...<mailto:tcl...@li...>> wrote:
Check what effect replacing [uplevel] with [tailcall] has.

On Nov 3, 2025, at 7:13 PM, EricT <tw...@gm...<mailto:tw...@gm...>> wrote:

Subject: Your bytecode expression evaluator - impressive results with caching!

Hey Colin:

I took a look at your bytecode-based expression evaluator and was intrigued by the approach. I made a small modification to add caching and the results are really impressive. Here's what I changed:

proc = args {
    set exp [join $args]
    if {[info exist ::cache($exp)]} {
        return [uplevel [list ::tcl::unsupported::assemble $::cache($exp)]]
    }
    set tokens [tokenise $exp]
    deb1 "TOKENS = '$tokens'"
    set code [compile $tokens]
    deb1 "GENERATED CODE:\n$code\n"
    set ::cache($exp) $code
    uplevel [list ::tcl::unsupported::assemble $code]
}

The cache is just a simple array lookup - one line to store, one line to retrieve. But the performance impact is huge:

Performance Tests

Without caching
% time {= 1 + 2} 1000
24.937 microseconds per iteration

With caching
% time {= 1 + 2} 1000
1.8 microseconds per iteration

That's a 13x speedup! The tokenize and parse steps were eating about 92% of the execution time.

The Real Magic: Bare Variables + Caching

What really impressed me is how well your bare variable feature synergizes with caching:

% set a 5
5
% set b 6
6
% = a + b
11
% time {= a + b} 1000
2.079 microseconds per iteration

Now change the variable values
% set a 10
10
% = a + b
16
% time {= a + b} 1000
2.188 microseconds per iteration

The cache entry stays valid even when the variable values change! Why? Because the bytecode stores variable names, not values:

push a; loadStk; push b; loadStk; add;

The loadStk instruction does runtime lookup, so:
- Cache key is stable: "a + b"
- Works for any values of a and b
- One cache entry handles all value combinations

Compare this to if we used $-substitution:

= $a + $b    # With a=5, b=6 becomes "5 + 6"
= $a + $b    # With a=10, b=6 becomes "10 + 6" - different cache key!

Every value change would create a new cache entry or worse, a cache miss.

Comparison to Other Approaches

Tcl's expr:              about 0.40 microseconds
Direct C evaluator:      about 0.53 microseconds
Your cached approach:    about 1.80 microseconds
Your uncached approach:  about 24.9 microseconds

With caching, you're only 3-4x slower than a direct C evaluator.

My Assessment

Your design is excellent. The bare variable feature isn't just syntax sugar - it's essential for good cache performance. The synergy between:

1. Bare variables leading to stable cache keys
2. Runtime lookup keeping cache hot
3. Simple caching providing dramatic speedup

makes this really elegant.

My recommendation: Keep it in Tcl! The implementation is clean, performance is excellent (1.8 microseconds is plenty fast), and converting to C would add significant complexity for minimal gain (maybe getting to about 1.0 microseconds).

The Tcl prototype with caching is actually the right solution here. Sometimes the prototype IS the product!

Excellent work on this. The bytecode approach really shines with caching enabled.

On Sun, Nov 2, 2025 at 10:14 AM Colin Macleod via Tcl-Core <tcl...@li...<mailto:tcl...@li...>> wrote:

Hi again,

I've now made a slightly more serious prototype, see https://cmacleod.me.uk/tcl/expr_ng

This is a modified version of the prototype I wrote for tip 676. It's still in Tcl, but doesn't use `expr`.  It tokenises and parses the input, then generates TAL bytecode and uses ::tcl::unsupported::assemble to run that.  A few examples:

(bin) 100 % set a [= 3.0/4]
0.75
(bin) 101 % set b [= sin(a*10)]
0.9379999767747389
(bin) 102 % set c [= (b-a)*100]
18.79999767747389
(bin) 103 % namespace eval nn {set d [= 10**3]}
1000
(bin) 104 % set e [= a?nn::d:b]
1000
(bin) 105 % = {3 + [pwd]}
Calc: expected start of expression but found '[pwd]'
(bin) 106 % = {3 + $q}
Calc: expected start of expression but found '$q'
(bin) 107 % = sin (12)
-0.5365729180004349

(bin) 108 % array set rr {one 1 two 2 three 3}
(bin) 110 % = a * rr(two)
Calc: expected operator but found '('
(bin) 111 % = a * $rr(two)
1.5

- You can use $ to get an array value substituted before the `=` code sees the expression.

(bin) 112 % string repeat ! [= nn::d / 15]
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!

Colin.
On 02/11/2025 09:04, Donal Fellows wrote:
Doing the job properly would definitely involve changing the expression parser, with my suggested fix being to turn all bare words not otherwise recognised as constants or in positions that look like function calls (it's a parser with some lookahead) into simple variable reads (NB: C resolves such ambiguities within itself differently, but that's one of the nastiest parts of the language). We would need to retain $ support for resolving ambiguity (e.g., array reads vs function calls; you can't safely inspect the interpreter to resolve it at the time of compiling the expression due to traces and unknown handlers) as well as compatibility, but that's doable as it is a change only in cases that are currently errors.

Adding assignment is quite a bit trickier, as that needs a new major syntax class to describe the left side of the assignment. I suggest omitting that from consideration at this stage.

Donal.

-------- Original message --------
From: Colin Macleod via Tcl-Core <tcl...@li...><mailto:tcl...@li...>
Date: 02/11/2025 08:13 (GMT+00:00)
To: Pietro Cerutti <ga...@ga...><mailto:ga...@ga...>
Cc: tcl...@li...<mailto:tcl...@li...>, av...@lo...<mailto:av...@lo...>
Subject: Re: [TCLCORE] Fwd: TIP 672 Implementation Complete - Ready for Sponsorship

Indeed, this toy implementation doesn't handle that:

% = sin (12)
can't read "sin": no such variable

I'm not sure that's serious, but it could be fixed in a C implementation.
_______________________________________________
Tcl-Core mailing list
Tcl...@li...<mailto:Tcl...@li...>
https://lists.sourceforge.net/lists/listinfo/tcl-core
_______________________________________________
Tcl-Core mailing list
Tcl...@li...<mailto:Tcl...@li...>
https://lists.sourceforge.net/lists/listinfo/tcl-core

_______________________________________________
Tcl-Core mailing list
Tcl...@li...<mailto:Tcl...@li...>
https://lists.sourceforge.net/lists/listinfo/tcl-core

_______________________________________________

Tcl-Core mailing list

Tcl...@li...<mailto:Tcl...@li...>

https://lists.sourceforge.net/lists/listinfo/tcl-core
_______________________________________________
Tcl-Core mailing list
Tcl...@li...<mailto:Tcl...@li...>
https://lists.sourceforge.net/lists/listinfo/tcl-core
_______________________________________________
Tcl-Core mailing list
Tcl...@li...<mailto:Tcl...@li...>
https://lists.sourceforge.net/lists/listinfo/tcl-core

Thread: Re: [TCLCORE] [Ext] Re: [=] for concise expressions (was Re: TIP 672 Implementation Complete - Read (Page 2)

The Tool Command Language implementation

tcl-core