how long do the different python-vm bytecodes take?

[Terry Reedy]

"GHUM" <haraldarminmassa at gmail.com> wrote in message 
news:0555ed9f-c5f1-4d0b-afbd-439c863e9439 at s13g2000prd.googlegroups.com...
|I looked at the virtual machine bytecode of python programs:
|
| def f(whatever):
|    text="Hallo"+whatever
|    return text
|
| import dis
| dis.dis(f)
|  2           0 LOAD_CONST               1 ('Hallo')
|              3 LOAD_FAST                0 (whatever)
|              6 BINARY_ADD
|              7 STORE_FAST               1 (text)
|
|  3          10 LOAD_FAST                1 (text)
|             13 RETURN_VALUE
|
|
| and now I am curious: how long does one LOAD_FAST take? I am thinking
| of something like back in the assembler-world, where there existed
| tables like:
|
| LDA  -> 2 cycles
| BNE  -> 2 cycles, 3 on branch
|
| of course, the "real time" is dependand on many factors, esp. the
| speed of the processor.
| But ... is there a relative scale somewhere?
|
| Somehting like
|
| LOAD_CONST      1 arbitraryUnit
| LOAD_FAST         1.9 arbitraryUnits
| RETURN_VALUE   8.0 arbitraryUnits
|
| ???
|
| my Google queries did not reveal anything usefull so far. Any hints?

There is no such table that I know of and there hardly could be, as the 
relative times will depend on hardware, OS, C compiler (including 
optimization settings), as well as the objects on the stack to be operated 
on.  Consider BINARY_ADD.  Pretty fast for 2 ints, arbitrarily slow for 
arbitrary user-class objects.  There are a few things knowns, such as _FAST 
versions being faster that other versions, and that eliminating unneeded 
codes is faster than not.  Developers usually test proposed 'speedup' 
changes on several platforms.

tjr