http: send Content-Length when possible

[tellnes]

This changes the behavior for http to send send a Content-Length header instead of using chunked encoding when we know the size of the body when sending the headers.

There is tree commits here. The first one is the initial implementation. The second also sends Content-Length: 0 when there is no body. The last reorganizes some code. I'll squash them before merge.

This broke a lot less tests than I feared.
Fixes #1044

[mscdex]

Do you have performance numbers on how this affects things? I would guess it would be faster since it's not having to do extra encoding....

[tellnes]

I've run the benchmark for http, but I'm not that familiar with the inner workings of those. This is run on my MBP. I do not know how sensitive those are to me surfing Github while it was running.

https://gist.github.com/tellnes/05012feaede66610cf24

It looks like something is going up while other things is going down. Maybe someone with more experience in this area can shim in?

cc @iojs/collaborators

[mscdex]

You might do something like this from the project root that contains the new changes:

./iojs benchmark/compare.js -r -g ./iojs ./iojs-prev -- http

(where ./iojs-prev is the path to a binary before the changes -- I usually use a full path including the /out/Release/iojs part to make sure a symlink doesn't mistakenly point to the new executable). This will make the output easier to compare.

[bnoordhuis]

Can you create the _contentLength field in the OutgoingMessage constructor? Adding properties post facto causes a hidden class transition.

EDIT: Sorry, scratch that. I see you're already doing that.

[bnoordhuis]

LGTM, no comments.

[rvagg]

nice work, LGTM

[tellnes]

Here is some numbers: https://gist.github.com/tellnes/44f09e07ac28acf8fc32

[mikeal]

+1

[mscdex]

LGTM