I'm finding that memory profiling is inaccurate in recent versions of tip. Allocations
are underreported, and frees are underreported even further (making it look like vast
amounts of memory are still in use).

I've written a test that demonstrates this behavior - it's at
http://play.golang.org/p/QUMOyGNTIg . Ten worker goroutines allocate a total of 1
million 8kB slices. The alloc/free counts from runtime.ReadMemStats are correct, but the
figures from runtime.MemProfile (used by e.g. net/http/pprof) are inaccurate.

The test passes with go1.1.2, it passes with GOMAXPROCS=1, and it passes if there's only
one goroutine allocating memory.

It fails on recent versions of tip with GOMAXPROCS=2 with several goroutines allocating
memory.



$ ./go1.1.2/bin/go version
go version go1.1.2 linux/amd64
$ ./go1.1.2/bin/go test -cpu=1 ~/profile_test.go 
ok      command-line-arguments  3.180s
$ ./go1.1.2/bin/go test -cpu=2 ~/profile_test.go 
ok      command-line-arguments  4.916s

$ ./go1.2.094fb360be8d/bin/go version
go version devel +094fb360be8d Tue Oct 08 16:53:56 2013 +1100 linux/amd64
$ ./go1.2.094fb360be8d/bin/go test -cpu=1 ~/profile_test.go 
ok      command-line-arguments  4.233s
$ ./go1.2.094fb360be8d/bin/go test -cpu=2 ~/profile_test.go 
--- FAIL: TestProfileLeak-2 (3.92 seconds)
    profile_test.go:38: found bucket: size 8192, count 1000000
    profile_test.go:84: n1 707328, s1 5794430976
    profile_test.go:93: n2 37248, s2 305135616
    profile_test.go:97: poor accuracy when measuring accuracy: 70.7% of allocations reported
    profile_test.go:101: profiler claims garbage: 37248 slices totalling 305135616 bytes
FAIL
FAIL    command-line-arguments  3.927s

$ ./go1.2.5f680cab75f6/bin/go version
go version devel +5f680cab75f6 Fri Oct 11 11:30:03 2013 +1100 linux/amd64
$ ./go1.2.5f680cab75f6/bin/go test -cpu=1 ~/profile_test.go
ok      command-line-arguments  4.203s
$ ./go1.2.5f680cab75f6/bin/go test -cpu=2 ~/profile_test.go
--- FAIL: TestProfileLeak-2 (5.22 seconds)
    profile_test.go:38: found bucket: size 8192, count 1000000
    profile_test.go:84: n1 674240, s1 5523374080
    profile_test.go:93: n2 30848, s2 252706816
    profile_test.go:97: poor accuracy when measuring accuracy: 67.4% of allocations reported
    profile_test.go:101: profiler claims garbage: 30848 slices totalling 252706816 bytes
FAIL
FAIL    command-line-arguments  5.219s