cmd/go: module graph pruning and lazy module loading

[bcmills]

I propose to change cmd/go to avoid loading transitive module dependencies
that have no observable effect on the packages to be built.

The key insights that lead to this approach are:

1. If no package in a given dependency module is ever (even transitively)
   imported by any package loaded by an invocation of the go command, then an
   incompatibility between any package in that dependency and any other package
   has no observable effect in the resulting program(s). Therefore, we can
   safely ignore the (transitive) requirements of any module that does not
   contribute any package to the build.

2. We can use the explicit requirements of the main module as a coarse filter
   on the set of modules relevant to the main module and to previous
   invocations of the go command.

Based on those insights, I propose to change the go command to retain more
transitive dependencies in go.mod files and to avoid loading go.mod files
for “irrelevant” modules, while still maintaining high reproducibility for build
and test operations.

Design doc: http://golang.org/design/36460-lazy-module-loading.

[gopherbot]

Change https://golang.org/cl/217557 mentions this issue: cmd/go: add a regression test for package import cycles guarded by build tags

[gopherbot]

Change https://golang.org/cl/220080 mentions this issue: design/36460: add design for lazy module loading

[bcmills]

A detailed design is in CL 220080.

(You can view the rendered Markdown by opening the .md file at the latest patch set and clicking the gitiles link.)

[ChrisHines]

For modules that are tidy:

The module versions recorded in the go.mod file would be exactly those listed in vendor/modules.txt, if present.

👍 I think that is a nice property to have. I have some projects for which go mod vendor is sane, but go mod tidy updates the go.mod file in undesirable ways due to test-of-test dependencies. If I understand this proposal correctly that problem would go away.

[bcmills]

I added a few examples (in our usual cmd/go script test format) to the design doc.

@ChrisHines, the third example explicitly covers the behavior for test-of-test dependencies.

[bcmills]

CC @julieqiu since this may affect how we display dependencies on pkg.go.dev.

CC @heschik, @dmitshur because this may affect the package search algorithms used by goimports and godoc.

CC @bep because I seem to recall that Hugo is doing something unusual with module dependencies independent of Go package imports. (But note that we are unlikely to actively support the use of Go modules for purposes other than compiling Go source code.)