549509 - JM: jsval representation profiling

Status: RESOLVED FIXED

(Core :: JavaScript Engine, defect)

Description

[David Mandelin [:dmandelin]]

Motivation: Consider JSOP_MUL. We know we should have some fast paths, but which inputs do we need to cover? int*int? double*double? Others? We need to know the frequency with which each tuple of input jsval tags occurs. 

Here's what we need:

For each opcode, record the tuple of input tags that occur at runtime [ i_1, i_2, ..., i_n] for an opcode that uses n inputs. Print the results as a histogram of the input tag tuple for each opcode. This feature should be in the tree so that we can rerun it at any time.

I'm calling this "representation profiling" rather than "type profiling" because this is about the representations used by the interpreter, not semantic type properties of the values. It's a subtle distinction but it may be clearer if this profiling is compared to the type profiling bug.

It seems that the output tags might also be interesting, but I can't think of an immediate use for them.

Comment 1

[David Mandelin [:dmandelin]]

Attachment: SunSpider results

Each line is:

  op, opname, input_1, ..., input_n, count

The file is sorted by op. So this block:

 31,mod,double,int,3180
 31,mod,int,int,223331

means that for |mod|, the inputs are (double, int) 3180 times and (int, int) 223331 times.

This will change when we get the new jsval formats, because we'll have 32-bit ints instead of just 31 bits.

Comment 2

[David Mandelin [:dmandelin]]

Attachment: Patch

This is the instrumentation patch.

Comment 3

[Luke Wagner [:luke]]

Comment on attachment 430497 [details] [diff] [review]
Patch

Nicely done

>diff --git a/js/src/jsapi.h b/js/src/jsapi.h
>+#define JS_REPRMETER

Just a reminder to remove this little guy hiding at the top of jsapi.h.

>+    enum repr {

Perhaps, if you don't feel strongly otherwise, the type name could be capitalized?

>+        OpInputHistogram::AddPtr p = opinputs.lookupForAdd(opinput);
>+        if (!p)
>+            opinputs.add(p, opinput, 1);
>+        else
>+            opinputs.put(opinput, p->value + 1);
>+    }

Not that efficiency matters, but 'put' does a second lookup.  I think what you want is:

OpInputHistogram::AddPtr p = opinputs.lookupForAd(opinput);
if (p)
  p->value++;
else
  opinputs.add(p, opinput, 1);

Comment 4

[David Mandelin [:dmandelin]]

http://hg.mozilla.org/tracemonkey/rev/a79ea749498a

Comment 5

[Ed Lee :Mardak]

Entries with count over 1mil:

86,getlocal,20810089
84,getarg,6108183
20,lt,int,int,5482894
55,getelem,array:dense,int,4995047
219,int8,4875481
87,setlocal,int,4685771
63,one,3674171
27,add,int,int,3560338
17,bitand,int,int,3437635
59,name,2435188
87,setlocal,double,2377583
25,rsh,int,int,1979685
56,setelem,array:dense,int,int,1831824
24,lsh,int,int,1699570
28,sub,int,int,1629786
88,uint16,1610496
29,mul,double,double,1483490
154,getgvar,1389074
27,add,double,double,1138757
62,zero,1136940
29,mul,int,int,1058658
228,trace,1027598

Total count ~105mil, so these entries represent ~75%.

Comment 6

[Ed Lee :Mardak]

Attachment: SunSpider results sorted within opcode

Comment 7

[Ed Lee :Mardak]

(In reply to comment #6)
> Created an attachment (id=431979) [details]
> SunSpider results sorted within opcode

data.split("\n").reduce(function(a, l) (l=l.split(","), l[0] = l[0] - 0, a[l[0]] ? a[l[0]].push(l) : a[l[0]] = [l], a), []).map(function(o) o.sort(function(a, b) a[a.length - 1] - b[b.length - 1]).join("\n")).join("\n\n").replace(/\n{2,}/g, "\n\n")

Comment 8

[Brendan Eich [:brendan]]

Comment on attachment 430497 [details] [diff] [review]
Patch

>+#ifdef JS_REPRMETER
>+# define METER_REPR(fp)         (reprmeter::MeterRepr(fp))
>+#else
>+# define METER_REPR(fp)

Avoid empty statement warnings (forget which compilers/options, but we do this elsewhere based on being nagged over nothing) by making the no-op macro's body ((void) 0).

/be

Comment 9

[David Mandelin [:dmandelin]]

(In reply to comment #8)
> (From update of attachment 430497 [details] [diff] [review])
> >+#ifdef JS_REPRMETER
> >+# define METER_REPR(fp)         (reprmeter::MeterRepr(fp))
> >+#else
> >+# define METER_REPR(fp)
> 
> Avoid empty statement warnings (forget which compilers/options, but we do this
> elsewhere based on being nagged over nothing) by making the no-op macro's body
> ((void) 0).

Done.

Comment 10

[Robert Sayre]

http://hg.mozilla.org/mozilla-central/rev/a79ea749498a

Comment 11

[David Mandelin [:dmandelin]]

Attachment: v8-v4 results sorted within opcode

Comment 12

[David Anderson [:dvander] - inactive, e-mail if emergency]

Attachment: top 20 stub calls for ss & v8

Added stub call profiling: http://hg.mozilla.org/users/danderson_mozilla.com/moo/rev/b79cd88ddfea

Attached are the top 20 stub calls for SS and v8.