compiletest: Run tests in their own thread to avoid blocking the scheduler

The calls to waitpid are interacting badly with the message passing that goes
on between schedulers and causing us to have very little parallelism in
the test suite. I don't fully understand the sequence of events that causes
the problem here but clearly blocking on waitpid is something that a
well-behaved task should not be doing.

Unfortunately this adds quite a bit of overhead to each test: one thread, two
tasks, three stacks, so there's a tradeoff. The time to execute run-pass on
my 4-core machine goes from ~750s to ~300s.

1 files changed, 16 insertions, 2 deletions

diff --git a/src/compiletest/runtest.rs b/src/compiletest/runtest.rs
index 0fb64152d37..2346aba3bcb 100644
--- a/src/compiletest/runtest.rs
+++ b/src/compiletest/runtest.rs
@@ -20,16 +20,30 @@ use procsrv;
 use util;
 use util::logv;
 
+use std::cell::Cell;
 use std::io;
 use std::os;
 use std::str;
+use std::task::{spawn_sched, SingleThreaded};
 use std::vec;
 
 use extra::test::MetricMap;
 
 pub fn run(config: config, testfile: ~str) {
-    let mut _mm = MetricMap::new();
-    run_metrics(config, testfile, &mut _mm);
+    let config = Cell::new(config);
+    let testfile = Cell::new(testfile);
+    // FIXME #6436: Creating another thread to run the test because this
+    // is going to call waitpid. The new scheduler has some strange
+    // interaction between the blocking tasks and 'friend' schedulers
+    // that destroys parallelism if we let normal schedulers block.
+    // It should be possible to remove this spawn once std::run is
+    // rewritten to be non-blocking.
+    do spawn_sched(SingleThreaded) {
+        let config = config.take();
+        let testfile = testfile.take();
+        let mut _mm = MetricMap::new();
+        run_metrics(config, testfile, &mut _mm);
+    }
 }
 
 pub fn run_metrics(config: config, testfile: ~str, mm: &mut MetricMap) {