cs: reduce PLT_LINKLET_TIMES
overhead
This commit is contained in:
parent
41dc6f1335
commit
b1632232ca
|
@ -1,10 +1,9 @@
|
|||
(define region-times (make-eq-hashtable))
|
||||
(define region-gc-times (make-eq-hashtable))
|
||||
(define region-counts (make-eq-hashtable))
|
||||
(define region-memories (make-eq-hashtable))
|
||||
(define perf-regions (make-eq-hashtable))
|
||||
(define-record perf-region (time gc-time count memory))
|
||||
|
||||
(define current-start-time '())
|
||||
(define current-gc-start-time '())
|
||||
(define current-perf-frame #f)
|
||||
|
||||
(define-record perf-frame (start gc-start nested-delta nested-gc-delta next))
|
||||
|
||||
(define performance-thread-id (get-thread-id))
|
||||
|
||||
|
@ -37,41 +36,62 @@
|
|||
(syntax-rules ()
|
||||
[(_ label e ...) (measure-performance-region label (lambda () e ...))]))
|
||||
|
||||
(define (label->perf-region label)
|
||||
(or (hashtable-ref perf-regions label #f)
|
||||
(let ([r (make-perf-region 0 0 0 0)])
|
||||
(hashtable-set! perf-regions label r)
|
||||
r)))
|
||||
|
||||
(define (measure-performance-region label thunk)
|
||||
(cond
|
||||
[(and measure-performance?
|
||||
(eqv? (get-thread-id) performance-thread-id))
|
||||
(with-interrupts-disabled
|
||||
(set! current-start-time (cons (current-process-milliseconds) current-start-time))
|
||||
(set! current-gc-start-time (cons (current-gc-milliseconds) current-gc-start-time)))
|
||||
(set! current-perf-frame (make-perf-frame (current-process-milliseconds)
|
||||
(current-gc-milliseconds)
|
||||
0
|
||||
0
|
||||
current-perf-frame))
|
||||
(begin0
|
||||
(thunk)
|
||||
(with-interrupts-disabled
|
||||
(let ([delta (- (current-process-milliseconds) (car current-start-time))]
|
||||
[gc-delta (- (current-gc-milliseconds) (car current-gc-start-time))])
|
||||
(hashtable-update! region-times label (lambda (v) (+ v delta)) 0)
|
||||
(hashtable-update! region-gc-times label (lambda (v) (+ v gc-delta)) 0)
|
||||
(hashtable-update! region-counts label add1 0)
|
||||
(set! current-start-time (cdr current-start-time))
|
||||
(set! current-gc-start-time (cdr current-gc-start-time))
|
||||
(let loop ([l current-start-time] [gc-l current-gc-start-time])
|
||||
(unless (null? l)
|
||||
(set-car! l (+ (car l) delta))
|
||||
(set-car! gc-l (+ (car gc-l) gc-delta))
|
||||
(loop (cdr l) (cdr gc-l)))))))]
|
||||
(let ([f current-perf-frame])
|
||||
(when f ; avoid crash if thread swaps mangle stack update
|
||||
(let ([delta (- (current-process-milliseconds) (perf-frame-start f) (perf-frame-nested-delta f))]
|
||||
[gc-delta (- (current-gc-milliseconds) (perf-frame-gc-start f) (perf-frame-nested-gc-delta f))]
|
||||
[r (label->perf-region label)])
|
||||
(set-perf-region-time! r (+ (perf-region-time r) delta))
|
||||
(set-perf-region-gc-time! r (+ (perf-region-gc-time r) gc-delta))
|
||||
(set-perf-region-count! r (+ (perf-region-count r) 1))
|
||||
(let ([next (perf-frame-next f)])
|
||||
(set! current-perf-frame next)
|
||||
(when next
|
||||
(set-perf-frame-nested-delta! next (+ delta (perf-frame-nested-delta f) (perf-frame-nested-delta next)))
|
||||
(set-perf-frame-nested-gc-delta! next (+ gc-delta (perf-frame-nested-gc-delta f) (perf-frame-nested-gc-delta next)))))))))]
|
||||
[else (thunk)]))
|
||||
|
||||
(define (add-performance-memory! label delta)
|
||||
(when measure-performance?
|
||||
(with-interrupts-disabled
|
||||
(hashtable-update! region-memories label (lambda (v) (+ v delta)) 0))))
|
||||
(let ([r (label->perf-region label)])
|
||||
(set-perf-region-memory! r (+ (perf-region-memory r) delta))))))
|
||||
|
||||
(define (linklet-performance-init!)
|
||||
(hashtable-set! region-times 'boot
|
||||
(time->ms (sstats-cpu (statistics)))))
|
||||
(let ([r (label->perf-region 'boot)])
|
||||
(set-perf-region-time! r (time->ms (sstats-cpu (statistics))))))
|
||||
|
||||
(define (linklet-performance-report!)
|
||||
(when measure-performance?
|
||||
(let ([region-times (make-eq-hashtable)]
|
||||
[region-gc-times (make-eq-hashtable)]
|
||||
[region-counts (make-eq-hashtable)]
|
||||
[region-memories (make-eq-hashtable)])
|
||||
(hash-table-for-each perf-regions
|
||||
(lambda (k r)
|
||||
(hashtable-set! region-times k (perf-region-time r))
|
||||
(hashtable-set! region-gc-times k (perf-region-gc-time r))
|
||||
(hashtable-set! region-counts k (perf-region-count r))
|
||||
(let ([m (perf-region-memory r)])
|
||||
(unless (zero? m)
|
||||
(hashtable-set! region-memories k m)))))
|
||||
(for-each (lambda (s)
|
||||
(let ([label (if (#%memq (car s) register-allocation-passes)
|
||||
'regalloc
|
||||
|
@ -155,4 +175,4 @@
|
|||
(report 0 'total total (#%format " [~a]" gc-total) 'ms "")
|
||||
(lprintf ";;")
|
||||
(for-each (lambda (p) (report 0 (car p) (/ (cdr p) 1024 1024) "" 'MB ""))
|
||||
(ht->sorted-list region-memories)))))
|
||||
(ht->sorted-list region-memories))))))
|
||||
|
|
Loading…
Reference in New Issue
Block a user