Version 8 of copy-on-write

Updated 2006-09-13 19:04:56

[explain importance of this notion of immutability to proper understanding of Tcl's semantics]

Those who have a little knowledge of C (or other similar languages) have learned to pass arguments to a function by value or by reference.

Tcl scripts have such functionalities : when you need to pass the variable by value, you put a $ before the variable name. And this does not modify the original variable.

But when you need to access uplevel variables, you can use upvar in the proc you call, passing just the name of the variable as argument. Then you can access the content of a variable defined in the upper stack level. This feature works especially well with arrays (arrays cannot be passed by value).

But when you come to Tcl C API (the internals that are required to write C extensions), you must know how Tcl passes variables by value, which makes use of the copy-on-write semantics.


Basically the copy-on-write semantics are:

  • when a proc receives a Tcl_Obj, it is in fact a pointer to the original variable, which is stored in a Tcl_Obj structure.
  • when all you need is read the variable, you do not require copying the data (1). The variable content may be shared between the caller and the callee.
  • when you need to work with a modified version of this variable, you need (2) to check whether the variable is shared or not
  • if the variable is shared, then you need to duplicate its content. (3)

A surprise

Here are 2 procs, the first gets a list passed by value, the second by reference via upvar.

 # Proc 1
 proc lmul-byval {list mult} {
     foreach i $list {
         lappend out [expr {$i*$mult}]
     }
     return $out
 }

 # Proc 2
 proc lmul-byref {list mult} {
     upvar 1 $list l
     for {set i 0} {$i<[llength $l]} {incr i} {
         lset l $i [expr {[lindex $l $i] * $mult}]
     }
     return
 }

 # Timings
 proc timeit {lsize val mult} {
     for {set i 0} {$i<$lsize} {incr i} {lappend list1 $val;lappend list2 $val}
     puts "By value : [time {set list1 [lmul-byval $list1 $mult]} 10]"
     puts "By reference : [time {lmul-byref list2 $mult} 10]"
 }
 foreach size {100 1000 10000} {
     puts "TIME : $size elements"
     timeit $size 2 2
 }

Results:

 TIME : 100 elements
 By value : 41.1 microseconds per iteration
 By reference : 84.0 microseconds per iteration
 TIME : 1000 elements
 By value : 386.9 microseconds per iteration
 By reference : 764.8 microseconds per iteration
 TIME : 10000 elements
 By value : 4300.3 microseconds per iteration
 By reference : 7893.7 microseconds per iteration

As you can see, upvar is costly. This is not like in C, in which large structures are faster passed by reference via pointers than by value. In C++, objects need to be passed by reference (&obj vs. obj) because the copy constructor can be waste CPU time when you work on large vectors of objects.

Another example showing this difference:

 proc checka {l} {
     if {[llength $l] == 1} {
         if {![string is integer $l]} {error "non-numerical value (should be integer)"}
         return [list bignum $l]
     }
     return $l
 }

 proc checkb {v} {
     upvar 1 $v l
     if {[llength $l] == 1} {
         if {![string is integer $l]} {error "non-numerical value (should be integer)"}
         set l [list bignum $l]
     }
 }

In the above example, we check a list represents a bignum, and if it is a simple integer, makes a bignum of it.(assuming bignums are lists of integers with a "bignum" heading tag)

 foreach i {1 {bignum 2} {bignum 3 3 3 3 3 3 3 3 3}} {
 set a $i
 puts "By value: [time {set a [checka $a]} 100]"
 set b $i
 puts "By reference: [time {checkb b} 100]" 
 }

 By value: 2.61 microseconds per iteration
 By reference: 3.42 microseconds per iteration
 By value: 2.11 microseconds per iteration
 By reference: 2.62 microseconds per iteration
 By value: 2.1 microseconds per iteration
 By reference: 2.51 microseconds per iteration

This demonstrates that upvar is more time-wasting than passing arguments by value and returning it, for this perticular kind of tasks.


Example in C (taken from foriter - a loop command)

1 (just read the variable)

        result = Tcl_GetIntFromObj(interp, objv[2], &start);
        if (result != TCL_OK) {
            return result;
        }
        ...

2 (check if the variable is shared)

        if ( Tcl_IsShared( obj_counter ) ) {

3 (duplicate shared data)

            obj_counter = Tcl_DuplicateObj( obj_counter );

Increment reference counter (see Tcl_Obj refCount HOWTO)

            Tcl_IncrRefCount( obj_counter );

Remember that the Tcl_Obj is shared

            isSharedObjCounter=1;
        }

        ...

When the Tcl_Obj becomes useless, we can decrement refcount to allow freeing memory. Without that, the whole interpreter may crash without alerting you :-( ...

        /* ending of the loop */

        if ( isSharedObjCounter ) {
            Tcl_DecrRefCount( obj_counter );
        }

[and consequences for extension design]

   ???