cmd/go: guarantee all test output is on stdout

In past releases, whether test output appears on stdout or stderr
has varied depending on exactly how go test was invoked and
also (indefensibly) on the number of CPUs available.
Standardize on standard output for all test output.
This is easy to explain and makes go test | go tool test2json work nicely.

Change-Id: I605641213fbc6c7ff49e1fd38a0f732045a8383d
Reviewed-on: https://go-review.googlesource.com/76871
Run-TryBot: Russ Cox <[email protected]>
TryBot-Result: Gobot Gobot <[email protected]>
Reviewed-by: Brad Fitzpatrick <[email protected]>

Author: rsc

src/cmd/go/internal/test/test.go

@@ -82,20 +82,21 @@ and even 'go test .').
 
 In local directory mode, go test compiles and tests the package sources
 found in the current directory and then runs the resulting test binary.
-In this mode, the test binary runs with standard output and standard error
-connected directly to the go command's own standard output and standard
-error, and test result caching (discussed below) is disabled.
-After the package test finishes, go test prints to standard output a
-summary line showing the test status ('ok' or 'FAIL'), package name,
-and elapsed time.
+In this mode, caching (discussed below) is disabled. After the package test
+finishes, go test prints a summary line showing the test status ('ok' or 'FAIL'),
+package name, and elapsed time.
 
 In package list mode, go test compiles and tests each of the packages
 listed on the command line. If a package test passes, go test prints only
 the final 'ok' summary line. If a package test fails, go test prints the
 full test output. If invoked with the -bench or -v flag, go test prints
 the full output even for passing package tests, in order to display the
-requested benchmark results or verbose logging. In package list mode,
-go test prints all test output and summary lines to standard output.
+requested benchmark results or verbose logging.
+
+All test output and summary lines are printed to the go command's standard
+output, even if the test printed them to its own standard error.
+(The go command's standard error is reserved for printing errors building
+the tests.)
 
 In package list mode, go test also caches successful package test results.
 If go test has cached a previous test run using the same test binary and
@@ -1208,36 +1209,17 @@ func (c *runCache) builderRunTest(b *work.Builder, a *work.Action) error {
 	if len(pkgArgs) == 0 || testBench {
 		// Stream test output (no buffering) when no package has
 		// been given on the command line (implicit current directory)
-		// or when benchmarking. Allowing stderr to pass through to
-		// stderr here is a bit of an historical mistake, but now a
-		// documented one. Except in this case, all output is merged
-		// to one stream written to stdout.
+		// or when benchmarking.
 		cmd.Stdout = os.Stdout
-		cmd.Stderr = os.Stderr
 	} else {
-		// If we're only running a single package under test
-		// or if parallelism is set to 1, and if we're displaying
-		// all output (testShowPass), we can hurry the output along,
-		// echoing it as soon as it comes in. We still have to copy
-		// to &buf for caching the result. This special case was
-		// introduced in Go 1.5 and is intentionally undocumented:
-		// the rationale is that the exact details of output buffering
-		// are up to the go command and subject to change.
-		// NOTE(rsc): Originally this special case also had the effect of
-		// allowing stderr to pass through to os.Stderr, unlike
-		// the normal buffering that merges stdout and stderr into stdout.
-		// This had the truly mysterious result that on a multiprocessor,
-		// "go test -v math" could print to stderr,
-		// "go test -v math strings" could not, and
-		// "go test -p=1 -v math strings" could once again.
-		// While I'm willing to let the buffer flush timing
-		// fluctuate based on minor details like this,
-		// allowing the file descriptor to which output is sent
-		// to change as well seems like a serious mistake.
-		// Go 1.10 changed this code to allow the less aggressive
-		// buffering but still merge all output to standard output.
-		// I'd really like to remove this special case entirely,
-		// but it is surely very helpful to see progress being made
+		// If we're only running a single package under test or if parallelism is
+		// set to 1, and if we're displaying all output (testShowPass), we can
+		// hurry the output along, echoing it as soon as it comes in.
+		// We still have to copy to &buf for caching the result. This special
+		// case was introduced in Go 1.5 and is intentionally undocumented:
+		// the exact details of output buffering are up to the go command and
+		// subject to change. It would be nice to remove this special case
+		// entirely, but it is surely very helpful to see progress being made
 		// when tests are run on slow single-CPU ARM systems.
 		if testShowPass && (len(pkgs) == 1 || cfg.BuildP == 1) {
 			// Write both to stdout and buf, for possible saving
@@ -1246,8 +1228,8 @@ func (c *runCache) builderRunTest(b *work.Builder, a *work.Action) error {
 		} else {
 			cmd.Stdout = &buf
 		}
-		cmd.Stderr = cmd.Stdout
 	}
+	cmd.Stderr = cmd.Stdout
 
 	// If there are any local SWIG dependencies, we want to load
 	// the shared library from the build directory.