Don't perform extra go env calls in go/packages.
Load only needed go env vars in golangci-lint.
Stay in sync by enabled analyzers in go vet: remove nilness and
atomicalign analyzers, add errorsas analyzer.
Don't build SSA for govet.
Standalone govet runs 25% faster than before. All runs can be 5-10% faster
than before.
Relates: #208
`adjust_format()` should use OS and not ARCH. In Windows, `$OS` will be calculated to be `"windows"` while `$ARCH` will be `"386"` or `"amd64"` (most of the time). `$ARCH` will surely never be `"windows"`.
This false-positive is not present in the upstream stand-alone 'unused'
2019.1.1 program that golangci-lint uses.
pkg/lint.ContextLoader.filterPackages() did two things:
1. It removed synthetic "testmain" packages (packages with .Name=="main"
and .PkgPath ending with ".test")
2. It removed pruned subsumed copies of packages; if a package with files
"a.go" and "a_test.go", it results in packages.Load giving us two
packages:
- ID=".../a" GoFiles=[a.go]
- ID=".../a [.../a.test]" GoFiles=[a.go a_test.go]
The first package is subsumed in the second, and leaving it around
results in duplicated work, and confuses the 'deadcode' linter.
However, the 'unused' linter relies on both the ".../a" and
".../a [.../a.test]" packages being present. Pruning them causes it to
panic in some situations, which lead to this workaround:
af6baa5dc1
While that workaround got it to not panic, it causes incorrect results.
So, split filterPackages() in to two functions: filterTestMainPackages()
and filterDuplicatePackages(). The linter.Context.Packages list only
gets filterTestMainPackages() called on it, while linter.Context.Program
and linter.Context.SSAProgram get both filters applied.
With the source of the panic fixed, roll back a few now-unnecessary
commits in go-tools.
Significantly improve CPU and memory usage when not using SSA-powered linters.
Improve readability of go/packages errors.
Improve debugging capabilities and write doc about debugging.
