Tweak the perf-record invocation.

Specifically: triple the sampling freqency, and collect both cycles and
instructions.

Also mention Hotspot for viewing perf profiles.

Author: nnethercote

collector/README.md

@@ -229,7 +229,8 @@ except that `$PROFILER` is one of the following.
     where execution time is spent and finding hot functions.
   - **Slowdown**. Negligible.
   - **Output**. Binary output is written to files with a `perf` prefix. Those
-    files can be read with `perf-report` and other similar `perf` commands.
+    files can be read with `perf-report` and other similar `perf` commands, or
+    with the excellent [Hotspot](https://github.com/KDAB/hotspot) viewer.
 - `oprofile`: Profile with [OProfile](http://oprofile.sourceforge.net/), a
   sampling profiler.
   - **Purpose**. OProfile is a general-purpose profiler, good for seeing

collector/src/bin/rustc-fake.rs

@@ -104,7 +104,8 @@ fn main() {
                 cmd.arg("record")
                     .arg("--call-graph=dwarf")
                     .arg("--output=perf")
-                    .arg("--freq=99")
+                    .arg("--freq=299")
+                    .arg("--event=cycles:u,instructions:u")
                     .arg(&rustc)
                     .args(&args);