问题
To quote the SICStus Prolog manual:
The usual mathematical theory behind Logic Programming forbids the creation of cyclic terms, dictating that an occurs-check should be done each time a variable is unified with a term. Unfortunately, an occurs-check would be so expensive as to render Prolog impractical as a programming language.
However, I ran these benchmarks (The Prolog ones) and saw fairly minor differences (less than 20%) in SWI Prolog between the occurs check (OC) being on and off:
OC is off: :- set_prolog_flag(occurs_check, false). in .swiplrc (restarted)
?- run_interleaved(10).
% 768,486,984 inferences, 91.483 CPU in 91.483 seconds (100% CPU, 8400298 Lips)
true.
?- run(1).
'Program' Time GC
================================
boyer 0.453 0.059
browse 0.395 0.000
chat_parser 0.693 0.000
crypt 0.481 0.000
fast_mu 0.628 0.000
flatten 0.584 0.000
meta_qsort 0.457 0.000
mu 0.523 0.000
nreverse 0.406 0.000
poly_10 0.512 0.000
prover 0.625 0.000
qsort 0.574 0.000
queens_8 0.473 0.000
query 0.494 0.000
reducer 0.595 0.000
sendmore 0.619 0.000
simple_analyzer 0.620 0.000
tak 0.486 0.000
zebra 0.529 0.000
average 0.534 0.003
true.
OC is on: :- set_prolog_flag(occurs_check, true). in .swiplrc (restarted)
?- run_interleaved(10).
% 853,189,814 inferences, 105.545 CPU in 105.580 seconds (100% CPU, 8083669 Lips)
true.
?- run(1).
'Program' Time GC
================================
boyer 0.572 0.060
browse 0.618 0.000
chat_parser 0.753 0.000
crypt 0.480 0.000
fast_mu 0.684 0.000
flatten 0.767 0.000
meta_qsort 0.659 0.000
mu 0.607 0.000
nreverse 0.547 0.000
poly_10 0.541 0.000
prover 0.705 0.000
qsort 0.660 0.000
queens_8 0.491 0.000
query 0.492 0.000
reducer 0.867 0.000
sendmore 0.629 0.000
simple_analyzer 0.757 0.000
tak 0.550 0.000
zebra 0.663 0.000
average 0.634 0.003
true.
Are these benchmarks not representative of real-world usage? (I remember reading somewhere that they were specifically chosen to be "fairly representative") Is SWI Prolog implementing some optimization tricks, that SICStus people aren't aware of, that make the cost of OC minor? If so, are they published? (I found a bunch of papers about this from the '90s, but I don't know if they are state-of-the-art)
回答1:
Cannot verify the result of less than 20%. I get an overhead of 40%:
Windows Laptop:
399 / 285 = 140%
Macintosh Laptop:
388 / 274 = 142%
Was testing newest version 8.3.17, maybe this is the reason?
来源：https://stackoverflow.com/questions/65600226/what-occurs-check-optimizations-is-swi-prolog-using