suzanne.soy/racket
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Factor out the concept of interesting callers.

Browse Source
This commit is contained in:
Vincent St-Amour 2012-09-18 16:14:28 -04:00
parent c66654e9c8
commit bba3d66fb1
1 changed files with 35 additions and 22 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

57
collects/typed-racket/optimizer/tool/mzc.rkt
View File

@ -222,36 +222,49 @@
(list evt))
site))))
;; Some sites are especially interesting if we have profile data.
;; Some callers are especially interesting if we have profile data.
;; If the function under consideration takes a large portion of the
;; total time for a given call site, and is not inlined there, may
;; be worth reporting.
;; returns: `(,caller-profile-node . ,call-site-log-entries) OR #f
(define interesting-sites
;; total time of a given caller, we consider this case interesting.
;; This serves as a building block for more interesting patterns, such
;; as `key-sites' below.
;; returns: caller-profile-node OR #f
(define interesting-callers
(and profile-entry
(filter values
(for/list ([site (in-list inlining-sites)]
;; Not inlined enough at that call site.
#:when (counts-as-a-missed-opt? site))
(for/list ([edge (node-callers profile-entry)])
;; Does this edge take a "large enough" proportion of
;; the caller's total time?
(define caller-node (edge-caller edge))
(and (> (edge-caller-time edge)
(* (node-total caller-node) 0.5))
caller-node)))))
;; As above, but consed in front of the inlining info for that caller.
(define interesting-callers+sites
(and profile-entry
(filter values
(for/list ([site (in-list inlining-sites)])
(match (inlining-event-where-loc
(inliner-log-entry-inlining-event (car site)))
[`(,caller-path ,caller-line ,caller-col)
(define caller-node
(pos->node (cons caller-line caller-col)))
(define edge
(for/first ([e (node-callers profile-entry)]
#:when (eq? (edge-caller e)
caller-node))
e))
;; Does this edge take a "large enough" proportion of
;; the caller's total time?
(and edge caller-node
(> (edge-caller-time edge)
(* (node-total caller-node) 0.5))
(and (memq caller-node interesting-callers)
(cons caller-node site))]
[_ ; can't parse that, give up
#f])))))
;; If the function under consideration takes a large portion of the
;; total time for a given call site, and is not inlined there, we can
;; recommend that the user take a closer look at that specific site.
;; returns: `(,caller-profile-node . ,call-site-log-entries) OR #f
(define key-sites
(and profile-entry
(for/list ([site (in-list interesting-callers+sites)]
;; Not inlined enough at that call site.
#:when (counts-as-a-missed-opt? (cdr site)))
site)))
(define pruned-log (apply append inlining-sites))
(define recommendation
@ -289,14 +302,14 @@
(when (and profile (not inside-hot-function?))
(prune))
(cond [(and profile (not (null? interesting-sites)))
(cond [(and profile (not (null? key-sites)))
;; Inlining was not satisfactory for some call sites where we
;; accounted for a good portion of the caller's total time.
(emit-near-miss
(format "Key call site~a: ~a"
(if (> (length interesting-sites) 1) "s" "")
(if (> (length key-sites) 1) "s" "")
(string-join
(for/list ([site (in-list interesting-sites)])
(for/list ([site (in-list key-sites)])
(define node (car site))
(format "~a ~a:~a"
(node-id node)
@ -304,7 +317,7 @@
(node-col node)))
", "))
;; only compute badness for the interesting sites
(group-badness (apply append (map cdr interesting-sites))))]
(group-badness (apply append (map cdr key-sites))))]
[(counts-as-a-missed-opt? pruned-log)
;; Overall inlining ratio is not satisfactory.
(emit-near-miss #f (group-badness pruned-log))]