summaryrefslogtreecommitdiff
path: root/benchmark
diff options
context:
space:
mode:
authorLudovic Court`es <ludovic.courtes@laas.fr>2005-04-29 14:12:12 +0000
committerLudovic Courtès <ludo@gnu.org>2008-04-25 19:09:30 +0200
commit2d80426a3ec7de15a194d0baed0e9f4be8659b92 (patch)
tree8100ff8d82d6a1fec507e67f23987c0da58a9703 /benchmark
parent238e7a11a8ec5aa2406b31620d3e56409639d4cf (diff)
Improved the VM's efficiency. The VM is as fast as the interpreter. :-(
* benchmark/lib.scm: New file. * benchmark/measure.scm: New file. * README: Added useful pointers to various threads. * doc/guile-vm.texi: Fixed the description of `load-program' (it now expects _immediate_ integers). * src/*.[ch]: Use immediate integers whereever possible, as in the original code. For `CONS', use `scm_cell' rather than `scm_cons'. git-archimport-id: lcourtes@laas.fr--2004-libre/guile-vm--revival--0.6--patch-6
Diffstat (limited to 'benchmark')
-rw-r--r--benchmark/lib.scm119
-rwxr-xr-xbenchmark/measure.scm49
2 files changed, 168 insertions, 0 deletions
diff --git a/benchmark/lib.scm b/benchmark/lib.scm
new file mode 100644
index 000000000..f272842b3
--- /dev/null
+++ b/benchmark/lib.scm
@@ -0,0 +1,119 @@
+;; -*- Scheme -*-
+;;
+;; A library of dumb functions that may be used to benchmark Guile-VM.
+
+
+(define (fibo x)
+ (if (= 1 x)
+ 1
+ (+ x
+ (fibo (1- x)))))
+
+(define (g-c-d x y)
+ (if (= x y)
+ x
+ (if (< x y)
+ (g-c-d x (- y x))
+ (g-c-d (- x y) y))))
+
+(define (loop how-long)
+ ;; This one shows that procedure calls are no faster than within the
+ ;; interpreter: the VM yields no performance improvement.
+ (if (= 0 how-long)
+ 0
+ (loop (1- how-long))))
+
+;; Disassembly of `loop'
+;;
+; Disassembly of #<objcode 302360b0>:
+
+; nlocs = 0 nexts = 0
+
+; 0 (make-int8 64) ;; 64
+; 2 (link "=")
+; 5 (link "loop")
+; 11 (link "1-")
+; 15 (vector 3)
+; 17 (make-int8:0) ;; 0
+; 18 (load-symbol "how-long") ;; how-long
+; 28 (make-false) ;; #f
+; 29 (make-int8:0) ;; 0
+; 30 (list 3)
+; 32 (list 2)
+; 34 (list 1)
+; 36 (make-int8 8) ;; 8
+; 38 (make-int8 2) ;; 2
+; 40 (make-int8 6) ;; 6
+; 42 (cons)
+; 43 (cons)
+; 44 (make-int8 23) ;; 23
+; 46 (make-int8 4) ;; 4
+; 48 (make-int8 12) ;; 12
+; 50 (cons)
+; 51 (cons)
+; 52 (make-int8 25) ;; 25
+; 54 (make-int8 4) ;; 4
+; 56 (make-int8 6) ;; 6
+; 42 (cons)
+; 43 (cons)
+; 44 (make-int8 23) ;; 23
+; 46 (make-int8 4) ;; 4
+; 48 (make-int8 12) ;; 12
+; 50 (cons)
+; 51 (cons)
+; 52 (make-int8 25) ;; 25
+; 54 (make-int8 4) ;; 4
+; 56 (make-int8 6) ;; 6
+; 58 (cons)
+; 59 (cons)
+; 60 (list 4)
+; 62 load-program ##{201}#
+; 89 (link "loop")
+; 95 (variable-set)
+; 96 (void)
+; 97 (return)
+
+; Bytecode ##{201}#:
+
+; 0 (object-ref 0)
+; 2 (variable-ref)
+; 3 (make-int8:0) ;; 0
+; 4 (local-ref 0)
+; 6 (call 2)
+; 8 (br-if-not 0 2) ;; -> 13
+; 11 (make-int8:0) ;; 0
+; 12 (return)
+; 13 (object-ref 1)
+; 15 (variable-ref)
+; 16 (object-ref 2)
+; 18 (variable-ref)
+; 19 (local-ref 0)
+; 21 (call 1)
+; 23 (tail-call 1)
+
+
+(define (loopi how-long)
+ ;; Same as `loop'.
+ (let loopi ((how-long how-long))
+ (if (= 0 how-long)
+ 0
+ (loopi (1- how-long)))))
+
+
+(define (do-cons x)
+ ;; This one shows that the built-in `cons' instruction yields a significant
+ ;; improvement (speedup: 1.4).
+ (let loop ((x x)
+ (result '()))
+ (if (<= x 0)
+ result
+ (loop (1- x) (cons x result)))))
+
+(define (copy-list lst)
+ ;; Speedup: 1.3.
+ (let loop ((lst lst)
+ (result '()))
+ (if (null? lst)
+ result
+ (loop (cdr lst)
+ (cons (car lst) result)))))
diff --git a/benchmark/measure.scm b/benchmark/measure.scm
new file mode 100755
index 000000000..0fe4b8efa
--- /dev/null
+++ b/benchmark/measure.scm
@@ -0,0 +1,49 @@
+#!/bin/sh
+# aside from this initial boilerplate, this is actually -*- scheme -*- code
+main='(module-ref (resolve-module '\''(measure)) '\'main')'
+exec ${GUILE-guile} -l $0 -c "(apply $main (cdr (command-line)))" "$@"
+!#
+
+;; A simple interpreter vs. VM performance comparison tool
+;;
+
+(define-module (measure)
+ :export (measure)
+ :use-module (system vm core)
+ :use-module (system base compile)
+ :use-module (system base language))
+
+(define (time-for-eval sexp eval)
+ (let ((before (tms:utime (times))))
+ (eval sexp (current-module))
+ (let ((elapsed (- (tms:utime (times)) before)))
+ (format #t "elapsed time: ~a~%" elapsed)
+ elapsed)))
+
+(define *scheme* (lookup-language 'scheme))
+
+(define (measure . args)
+ (if (< (length args) 2)
+ (begin
+ (format #t "Usage: measure SEXP FILE-TO-LOAD...~%")
+ (format #t "~%")
+ (format #t "Example: measure '(loop 23424)' lib.scm~%~%")
+ (exit 1)))
+ (for-each load (cdr args))
+ (let* ((sexp (with-input-from-string (car args)
+ (lambda ()
+ (read))))
+ (time-interpreted (time-for-eval sexp eval))
+ (objcode (compile-in sexp (current-module) *scheme*))
+ (time-compiled (time-for-eval objcode
+ (let ((vm (the-vm))
+ (prog (objcode->program objcode)))
+ (lambda (o e)
+ (vm prog))))))
+ (format #t "interpreted: ~a~%" time-interpreted)
+ (format #t "compiled: ~a~%" time-compiled)
+ (format #t "speedup: ~a~%"
+ (exact->inexact (/ time-interpreted time-compiled)))
+ 0))
+
+(define main measure)