Proposal: cmd/compile: add a go:inline directive

[OneOfOne]

The compiler (gc) has come a long way, however the inliner is sometimes lacking, for example any function with a for-loop no matter how simple it is, won't get inlined so we either have to use a goto loop instead (which is used few times in the runtime) or just have a one big ugly function.

My proposal is to add a go:inline directive that will force inlining, even if it's just for non-exported functions.

This will help make a lot of code much cleaner and allows separation of big functions that can't be currently inlined.

If the proposal is accepted, I'd like to work on it.

Another semi-common example:

func (xx *XXHash64) WriteString(s string) (int, error) {
	if len(s) == 0 {
		return 0, nil
	}
	return xx.Write([]byte(s))
}
// cannot inline (*XXHash64).WriteString: non-leaf method

Related: #17566 #14768

@dr2chase @josharian @randall77 @mvdan

[josharian]

This proposal has basically no chance of being accepted.

Go eschews knobs.

Compiler directives are widely disliked.

I don't think it can be implemented reasonably, if at all. Functions are not eligible for inlining for many reasons. These include conceptual implementation difficulty: I have no idea how to inline a function containing defer. And practical implementation difficulty: It took a dedicated person a summer to get non-leaf inlining working correctly, even with part of the implementation already written. And it's still not released because of the binary size and compilation time aspects. Supporting such a directive would mean solving a host of such problems.

I personally have long wanted a much smaller directive that causes compilation to fail if a function is deemed ineligible for inlining. This would allow me to refactor without fear that future inlining changes will cause my no-op inlining to become a silent performance regression. Even this much more limited proposal has been rejected forcefully.

All that said, if you'd like to work on improving inlining, that'd be fabulous. There are many opportunities to improve inlining decisions, to improve how we do inlining (e.g. partial inlining), and to expand the set of functions that can be unlined.

[neelance]

@josharian The flag -gcflags "-m" makes the go compiler print decisions on inlining. Maybe you can build a tool that parses this output and verifies it against some "must inline" list. Most people won't need it, but if you really do, then you could add that tool to your build pipeline.

[josharian]

@neelance yes, or alternatively parse the output of go tool nm and make sure the inlined function has been dead-code-eliminated. (Note that I want it most of all for package runtime, so "my build pipeline" is probably go test runtime.)

[neelance]

@josharian I already suspected that you want it for the stdlib. May I ask why you didn't build it yet? I get the point of not adding such a feature to the Go compiler itself, but a standalone tool doesn't do any harm if only used by individuals voluntarily. I just cringe when I hear "fear of refactoring". ;-)

[josharian]

@neelance just haven't gotten to it yet; I have little contribution time at the moment and lots I want to do. I'd be delighted if someone else wanted to jump in and do this. The recently introduce runtime function tophash would be an excellent and relevant test case. :)

[gopherbot]

Change https://golang.org/cl/57410 mentions this issue: runtime: add TestIntendedInlining

[rsc]

For the reasons Josh outlined above, we're not going to do this. Note also that even in C/C++, the "inline" directive does not really mean "this must be inlined". It is more a guideline than a rule.

[mandarjog]

Any reason why inlining is more difficult in go than other languages like java or c++.

I was trying to find the rationale behind no branches / 40 expressions or less restrictions for inlined functions.
Any pointers are appreciated.

While trying to write performant code (dynamic proto3 marshaling), hand inlining a function brought down marshal time from 5ms to 4ms. That can be significant depending on your performance goals.

[ianlancetaylor]

@mandarjog A closed issue is not the place for this discussion. General discussions about inlining in Go should use a forum; see https://golang.org/wiki/Questions . Specific discussions about the gc toolchain should take place on the golang-dev mailing list.