Map decode optimisations

[bboreham]

This is branched after #2351, so look at that one first.

Here I apply the the map-reading code taken from LatestMap to other uses of ps.Map; this removes the indirection via a Go map and consequent copying.

Also call CodecDecodeSelf() instead of Decode() where possible: this avoids a runtime type lookup, so goes a little faster. I had to get a little funky with providing an implementation for the parser to run before the real codec code is generated; that is the purpose of dummySelfer.

Bonus: having less recursion through Decode() makes it easier to interpret profiles.

[2opremio]

What's the performance improvement?

[bboreham]

Same test as in #2351:

With this PR, commit 46d1705:

$ go test -run TestDecode
ok  	github.com/weaveworks/scope/report	1.197s
ok  	github.com/weaveworks/scope/report	1.363s
ok  	github.com/weaveworks/scope/report	1.200s
ok  	github.com/weaveworks/scope/report	1.308s

Compare against 1.652:
That's about a 27% speed-up, fastest to fastest.

[bboreham]

Note CI failed because you run the unit-tests before running the code-generation step.

[2opremio]

Uhm, could the new code generation dependency have messed up coveralls?

[bboreham]

The problem is that you weren't testing the right code before.
The coverage of report.codecgen.co is at 37%, and it wasn't included before, so that drags the average down.
Arguably this is meaningless, since most of the code paths in that file are for situations we don't expect. But I don't know what you can do about that.

[2opremio]

OK, fair enough

[bboreham]

Rebased so we don't see the two commits already merged. Sorry, should have done that earlier.

[bboreham]

Tests failed sometimes with a "panic: Not run on same source!" message from cover:

• https://circleci.com/gh/weaveworks/scope/6857
• https://circleci.com/gh/weaveworks/scope/6858
  and succeeded sometimes:
• https://circleci.com/gh/weaveworks/scope/6866
• https://circleci.com/gh/weaveworks/scope/6848

I suspect this is caused by codecgen embedding a random number into the code, and this can be different between the two CircleCI lanes.
More specifically, the cover tool cares about the length of the number, which can be either 3 digits or 4 digits if left to chance.
So I pushed another commit fixing this number for CircleCI builds.

Also rebased just now.