Version 1 of Tcl invoke performance

See also Tcl IO performance, where we reached the

Interim conclusion The time spent in the original benchmark seems to be decomposable in

• 8% in startup and loop overhead (0.19/2.5)
• 20% in command call and setting a variable in the called command (0.50/2.50) Further experimentation with a C-coded command that just returns TCL_OK shows that the cost of setting the variable is almost negligible
• 72% doing the file access, reading and conversion within gets itself
• the cost of setting the variable in the called command seems to be below the measurable threshold for the larger dataset (read1a actually faster than read2 - the difference must be noise)

Further experiments

• on a core instrumented with code developed together with GPS
• running a tiny do-almost-nothing script (see below)
• using a command empty defined in C to just return Tcl_Ok;

provide the following timings for the most-exercised opcodes (measured in cpu ticks at 1.6MHz):

8.5a6 time: 1822167696, count: 20004698
----------------------------------------------------
op         %T       avgT         %ops         Nops
6       24.53     446.87         5.00      1000323      INST_INVOKE_STK1
80      22.68     413.32         5.00      1000019      INST_LIST_INDEX
29      10.04      91.47        10.00      2000048      INST_INCR_SCALAR1_IMM
105      9.77      44.52        20.00      4000452      INST_START_CMD
103      7.58     138.16         5.00      1000004      INST_LIST_INDEX_IMM
10       7.20      26.22        25.00      5000750      INST_LOAD_SCALAR1
17       5.78      52.67        10.00      2000236      INST_STORE_SCALAR1
47       5.23      95.32         5.00      1000001      INST_LT
1        4.19      38.11        10.00      2001039      INST_PUSH1
3        2.95      53.83         5.00      1000302      INST_POP

Striking observations (which require an explanation)

Taking the fastest opcode as comparison basis:

   1 INST_LOAD_SCALAR_STK1 is the fastest opcode (faster that INST_PUSH1 and INST_POP!), INST_STORE_STACK1 is pretty fast too
   2 [lindex] is amazingly slow - the non-immediate version is as slow as a command invocation
   3 command invocations are expensive (remark that only '''empty''' is invoked in the loop)
   4 comparisons (INST_LT) and basic arithmetic (INST_INCR_SCALAR_IMM1) are amazingly slow when compared to basic variable access
   5 the "pure loss" INST_START_CMD is amazingly slow

The script being run is

lappend auto_path /home/CVS/emptyFunc/
package require empty

exec /usr/bin/taskset -p 0x00000001 [pid]

proc main N {
    set y 0
    set a [list foo boo moo]
    for {set i 0} {$i < $N} {incr i} {
	empty 1
	incr y
	set z [lindex $a 1]
	set z 1
	lindex $a $z
    }
}

if {[llength $argv]} {
    main [lindex $argv 0]
} else {
    main 1000000
}