As of #18660, go get in GOPATH mode allows any Unicode letter

To be more precise, #18660 was about allowing Unicode letters within the "subdir" part of github.com import paths like "github.com/user/repo/subdir" (but not within "user" or "repo" parts).

To my knowledge, Unicode letters were allowed in vanity import paths since Go 1 or close to thereof. I tried Go 1.2.2 just now (the oldest release with binaries at https://golang.org/dl/), and it doesn't reject Unicode letters in a vanity import path:

$ ./bin/go version
go version go1.2.2 linux/amd64
$ ./bin/go get dmitri.shuralyov.com/temp/go-get-issue-unicode/испытание
$ echo $?
0

There is a relevant passage in the cmd/go/internal/module docs by @rsc that mentions that the eventual plan is to allow Unicode letters.

I assume this is now targeting Go 1.14? Should it be a release-blocker?

Change https://golang.org/cl/212198 mentions this issue: module: allow unicode Letters in CheckImportPath

Following up here as requested in @rsc's comment on golang.org/cl/212198.

I'm not sure what we did historically, but it looks like now, we do two checks when go-getting in GOPATH mode: first, we call cmd/go/internal/get.CheckImportPath, which is a copy of the definition of CheckImportPath in golang.org/x/mod/module, except that it additionally allows unicode letters as components of import paths (through an edit in its copy of pathOK).

As I understand from reading the code, the get.CheckImportPath call is the only check we do for "vanity" import paths. After that check, when we look up which VCS system is being used, for VCS-backed paths, the paths need to match against regexps is vcsPaths. Only the GitHub import path regexp allows non-ascii Unicode letters (not in the repo name, but it paths within a repo). And as I understand it we don't do any stricter checks for non-VCS-backed import paths. So in effect, non-ascii Unicode letters are only allowed in non-VCS-backed import paths and in (some) components of Github import paths.

There are a lot of issues to solve once we start allowing Unicode.
The special case for GitHub path elements after the module path is unfortunate and should probably be removed. Then we can reexamine the much larger problem as a whole. Let's not privilege GitHub users any longer.

Let's make import path elements checked with pathOK starting at the beginning of the Go 1.16 cycle. If a very important existing package comes up we can reconsider.

Change https://golang.org/cl/250942 mentions this issue: cmd/go/internal/get: make pathOK identical to that in golang.org/x/mod

Some more context: I'm looking into doing this as the Go 1.16 tree has opened, but it's not clear which path to take:

GOPATH mode lets through unicode letters in github import paths, so using pathOK to check import paths will break users in GOPATH mode using these repos.

In module mode, the proxy disables non-ASCII unicode letters, but if GOPROXY=direct, the go command will use the more lenient checking that allows non-ASCII unicode letters:

# Without proxy
$ GO111MODULE=on GOPROXY=direct go get github.com/matloob/unicode/𝕞𝕒𝕥𝕝𝕠𝕠𝕓
go: found github.com/matloob/unicode/𝕞𝕒𝕥𝕝𝕠𝕠𝕓 in github.com/matloob/unicode v0.0.0-20200828144222-b203c46e6021

# With proxy
$ GO111MODULE=on go get github.com/matloob/unicode/𝕞𝕒𝕥𝕝𝕠𝕠𝕓 
go get github.com/matloob/unicode/𝕞𝕒𝕥𝕝𝕠𝕠𝕓: malformed module path "github.com/matloob/unicode/𝕞𝕒𝕥𝕝𝕠𝕠𝕓": invalid char '𝕞'

So the question is do we just want to disable these paths in module mode, or for both module and GOPATH mode. Disabling them for module mode sounds like a reasonable idea because these modules can't be used with the proxy, but there might be GOPATH users we'd be breaking.

I'm going to try to see if I can find a GitHub index I can search through to see if any repos with go files in directories with non-ASCII unicode letters actually exist.

Change https://golang.org/cl/251878 mentions this issue: src/cmd/go/internal/get: disallow non-ASCII unicode letters from import paths

Change https://golang.org/cl/251877 mentions this issue: cmd: update vendored golang.org/x/mod

Change https://golang.org/cl/252297 mentions this issue: cmd/go: remove TestScript/get_unicode

On tip, the module mode check is now the same as the GOPATH mode check.