| Ticket UUID: | 3302370 | ||
| Title: | Performance gap between [global] and [::] | ||
| Type: | Bug | Created on: | 2011-05-15 10:00:16 |
| Submitter: | ferrieux | Assigned to: | msofer |
| Subsystem: | 07. Variables | Severity: | |
| Priority: | 5 Medium | Last modified: | 2011-09-24 05:38:49 |
| Status: | Open | Closed by: | |
| Resolution: | None | Closed on: | |
| Version: | None | ||
| Description: | ||||
Using [global] instead of [::] in a proc gives a 7x speedup.
% proc f {} {global zzz;for {set i 0} {$i<10000000} {incr i} {incr zzz}}
% proc g {} {for {set i 0} {$i<10000000} {incr i} {incr ::zzz}}
% time f
871320 microseconds per iteration
% time g
This can be traced to the different bytecode ops used: in the tight loop, the [global] case uses a compiled local (upvarred to the global), while the FQN uses loadScalarStk repeatedly.
I would have assumed that the intrep cacheing of the resolved varname in that literal value "::zzz" would save most of the work, but something else must be increasing the cost...
6123508 microseconds per iteration
| ||||
| User Comments: | ||||
ferrieux added on 2011-05-15 17:05:00:
Hmm, sorry for the bad layout. Copypasta glitch. Summary of timings: global -> 871320 microseconds :: -> 6123508 microseconds dkf added on 2011-05-25 05:28:58:
I vaguely remember there being a whole bunch of experimentation with variable reference caching in... was it 8.4? I think after 8.4.2 things started to be turned off again because of odd bugs in corner cases, but that's quite a long time ago now so memory might be off. What follows is "thinking with my mouth open" (except its typed ;-)) It should be possible to cache a Var* in a fully-qualified name, but there's a possibility that it could cause problems with [unset] semantics. Plus the name needs to be interpreter-bound anyway. Ick! msofer added on 2011-05-25 05:36:27:
This CAN be solved easily enough: when the compiler detects a FQN, it can create a same-named (with qualifiers!) variable in the LVT that is auto-upvared to the global on proc entry. The body then always refers to the one in the LVT. The same thing could actually be done for [eval] and [uplevel] bodies, even for non qualified names. But in that case it probably requires other changes elsewhere: the whole core assumes that only proc (and lambda, and oo) bodies have an LVT, this assumption could be hidden somewhere unexpected. Another solution would be to code a fast-path for FQN variables in TEBC, reading the Tcl_Obj intrep. Currently only LVT variables have a fast inlined path, the others call the functions in tclVar.c ferrieux added on 2011-05-25 14:09:41:
The first solution, even if not generalized to uplevel+eval bodies for now, sounds like an optimum to me... FWIW (Wow ! an FQN-named compiled local -- brilliant) ferrieux added on 2011-09-24 05:38:49:
Miguel, is this doable ? | ||||
