cmd/go: fix -gcflags, -ldflags not applying to current directory
A flag setting like -gcflags=-e applies only to the packages
named on the command line, not to their dependencies.
The way we used to implement this was to remember the
command line arguments, reinterpret them as pattern matches
instead of package argument generators (globs), and apply them
during package load. The reason for this complexity was to
address a command-line like:
go build -gcflags=-e fmt runtime
The load of fmt will load dependencies, including runtime,
and the load of runtime will reuse the result of the earlier load.
Because we were computing the effective -gcflags for each
package during the load, we had to have a way to tell, when
encountering runtime during the load of fmt, that runtime had
been named on the command line, even though we hadn't
gotten that far. That would be easy if the only possible
arguments were import paths, but we also need to handle
go build -gcflags=-e fmt runt...
go build -gcflags=-e fmt $GOROOT/src/runtime
go build -gcflags=-e fmt $GOROOT/src/runt...
and so on.
The match predicates usually did their job well, but not
always. In particular, thanks to symlinks and case-insensitive
file systems and unusual ways to spell file paths, it's always
been possible in various corner cases to give an argument
that evalutes to the runtime package during loading but
failed to match it when reused to determine "was this package
named on the command line?"
CL 109235 fixed one instance of this problem by making
a directory pattern match case-insensitive on Windows, but that
is incorrect in some other cases and doesn't address the root problem,
namely that there will probably always be odd corner cases
where pattern matching and pattern globbing are not exactly aligned.
This CL eliminates the assumption that pattern matching
and pattern globbing are always completely in agreement,
by simply marking the packages named on the command line
after the package load returns them. This means delaying
the computation of tool flags until after the load too,
for a few different ways packages are loaded.
The different load entry points add some complexity,
which is why the original approach seemed more attractive,
but the original approach had complexity that we simply
didn't recognize at the time.
This CL then rolls back the CL 109235 pattern-matching change,
but it keeps the test introduced in that CL. That test still passes.
In addition to fixing ambiguity due to case-sensitive file systems,
this new approach also very likely fixes various ambiguities that
might arise from abuse of symbolic links.