std: reduce the generic code instantiated by fail!().

This splits the vast majority of the code path taken by
`fail!()` (`begin_unwind`) into a separate non-generic inline(never)
function, so that uses of `fail!()` only monomorphise a small amount of
code, reducing code bloat and making very small crates compile faster.

Author: huonw

src/libstd/rt/unwind.rs

@@ -383,16 +383,31 @@ pub fn begin_unwind_raw(msg: *u8, file: *u8, line: uint) -> ! {
 }
 
 /// This is the entry point of unwinding for fail!() and assert!().
-#[inline(never)] #[cold] // this is the slow path, please never inline this
+#[inline(never)] #[cold] // avoid code bloat at the call sites as much as possible
 pub fn begin_unwind<M: Any + Send>(msg: M, file: &'static str, line: uint) -> ! {
-    // Note that this should be the only allocation performed in this block.
+    // Note that this should be the only allocation performed in this code path.
     // Currently this means that fail!() on OOM will invoke this code path,
     // but then again we're not really ready for failing on OOM anyway. If
     // we do start doing this, then we should propagate this allocation to
     // be performed in the parent of this task instead of the task that's
     // failing.
-    let msg = ~msg as ~Any;
 
+    // see below for why we do the `Any` coercion here.
+    begin_unwind_inner(~msg, file, line)
+}
+
+
+/// The core of the unwinding.
+///
+/// This is non-generic to avoid instantiation bloat in other crates
+/// (which makes compilation of small crates noticably slower). (Note:
+/// we need the `Any` object anyway, we're not just creating it to
+/// avoid being generic.)
+///
+/// Do this split took the LLVM IR line counts of `fn main() { fail!()
+/// }` from ~1900/3700 (-O/no opts) to 180/590.
+#[inline(never)] #[cold] // this is the slow path, please never inline this
+fn begin_unwind_inner(msg: ~Any, file: &'static str, line: uint) -> ! {
     let mut task;
     {
         let msg_s = match msg.as_ref::<&'static str>() {