cmd/gc: why did json get slower with convT2I optimization?

[rsc]

CL 6337058 made various things a little faster but made the JSON benchmark 30% slower.
Why?

[nigeltao]

Comment 1:

As a data point, http://golang.org/cl/6356053 adds an extra argument to convT2I
but does not cache itab lookups. That extra argument is simply ignored. The JSONDecode
benchmark shows no dramatic change, compared to the status quo before CL 6337058
"cmd/gc: cache itab lookup in convT2I" was submitted.
benchmark                 old ns/op    new ns/op    delta
BenchmarkBinaryTree17    5902507000   5918781000   +0.28%
BenchmarkFannkuch11      3926283000   3936780000   +0.27%
BenchmarkGobDecode         22264160     22356350   +0.41%
BenchmarkGobEncode         12815560     12704780   -0.86%
BenchmarkGzip             568959600    568881800   -0.01%
BenchmarkGunzip           178905400    179608400   +0.39%
BenchmarkJSONEncode        89589000     89617050   +0.03%
BenchmarkJSONDecode       310999600    316518000   +1.77%
BenchmarkMandelbrot200      7016514      7018210   +0.02%
BenchmarkParse              7865130      7985445   +1.53%
BenchmarkRevcomp         1284725000   1254541000   -2.35%
BenchmarkTemplate         567248200    571401600   +0.73%

[DanielMorsing]

Comment 2:

It might be interesting to see if introducing the global symbol has any effect on the
benchmarks.
Dave Cheney did some benchmarks for linux/arm, which doesn't have the performance
regression. That makes me think that there is some sort of alignment or size issue.
Maybe there is value to running these benchmarks under 386, since it has the same word
size.

[nigeltao]

Comment 3:

JSONDecode program times in seconds, 5 iterations, with and sans itab cache, on my
desktop.
amd64 with/sans: 2.520 / 1.980; ratio = 1.27
386   with/sans: 3.550 / 3.770; ratio = 0.94
The "with itab cache" path avoids the modulus calculation in hashing.
http://golang.org/cl/6256062/ suggests that the soft mod is relatively
expensive on ARM.

[nigeltao]

Comment 4:

Another data point: there may be confounding factors, but with itab cache is faster than
sans itab cache when decoding a deep JSON tree. With is slower than sans when decoding a
broad JSON tree.
With:
BenchmarkJSONDecodeBroad         500       4188660 ns/op       5.74 MB/s
BenchmarkJSONDecodeDeep      500       6037996 ns/op       3.49 MB/s
Sans:
BenchmarkJSONDecodeBroad         500       3925550 ns/op       6.13 MB/s
BenchmarkJSONDecodeDeep      500       7444352 ns/op       2.83 MB/s
GOARCH=amd64. example_test.go attached.

Attachments:

1. example_test.go (1018 bytes)

[davecheney]

Comment 5:

> The "with itab cache" path avoids the modulus calculation in hashing.
http://golang.org/cl/6256062/ suggests that the soft mod is relatively
expensive on ARM.
It sure is. Simple changes like making the itab cache 1024 elements wide, thus avoiding
soft mod, give substantial speedups.

[nigeltao]

Comment 6:

Another data point (linux/amd64 desktop). Starting from revision c533f48701cb, which is
the one before "cmd/gc: cache itab lookup in convT2I"
http://code.google.com/p/go/source/detail?r=ef432c94ce9c6ed6a0c343ec0a8124237f81a522
This innocuous looking diff (which introduces no new global symbols):
----------------
$ hg identify
c533f48701cb+
$ hg diff -U 5
diff -r c533f48701cb test/bench/go1/json_test.go
--- a/test/bench/go1/json_test.go   Mon Jul 02 15:30:00 2012 -0700
+++ b/test/bench/go1/json_test.go   Mon Jul 09 14:50:03 2012 +1000
@@ -82,6 +82,9 @@
 func BenchmarkJSONDecode(b *testing.B) {
    b.SetBytes(int64(len(jsonbytes)))
    for i := 0; i < b.N; i++ {
        jsondec()
    }
+   if false {
+       print([]int(nil))
+   }
 }
----------------
has this effect on the benchmark:
----------------
benchmark              old ns/op    new ns/op    delta
BenchmarkJSONDecode    312624800    403244800  +28.99%
----------------
The generated code for func BenchmarkJSONDecode is identical (the if-false is removed by
dead code elimination) except for the stack requirement. Note the $8 versus $24 in the
"6g -S" before/after:
0230 (json_test.go:82) TEXT    BenchmarkJSONDecode+0(SB),$8-8
0230 (json_test.go:82) TEXT    BenchmarkJSONDecode+0(SB),$24-8

[davecheney]

Comment 7:

@nigeltao, very interesting diagnosis.
Ultimately this sounds like bad fortune caused the parts of the inner loop of the json
decoder to straddle a stack boundary. 
Would you suggest that there is a bug in the dead code elimination logic, in that it is
not properly adjusting the stack accounting ?

[minux]

Comment 8:

On Monday, July 9, 2012, dfc wrote:
Very probable.
The stack frame issue is possibly issue #2734.

[rsc]

Comment 9:

Labels changed: added go1.1maybe.

[rsc]

Comment 10:

Labels changed: removed go1.1maybe.

[rsc]

Comment 11:

Labels changed: added size-l.

[rsc]

Comment 12:

Labels changed: added go1.1maybe, removed go1.1.

[robpike]

Comment 13:

Labels changed: added go1.2maybe, removed go1.1maybe.

[randall77]

Comment 14:

Should get fixed with contiguous stacks, maybe for Go 1.3.  See the attached graph.

Labels changed: added go1.3, removed go1.2maybe.

Owner changed to @randall77.

Attachments:

1. both.png (6967 bytes)

[robpike]

Comment 15:

Labels changed: removed go1.3.