Merge pull request #1522 from nnethercote/add-samply

Add support for profiling the compiler with Samply.

Author: nnethercote

ci/check-profiling.sh

@@ -39,6 +39,9 @@ cargo build -p collector --bin rustc-fake
 # oprofile: untested... it's not used much, and might have the same problems
 # that `perf` has due to virtualized hardware.
 
+# samply: untested because it has the same problems that `perf` has due to
+# virtualized hardware.
+
 # Cachegrind.
 RUST_BACKTRACE=1 RUST_LOG=raw_cargo_messages=trace,collector=debug,rust_sysroot=debug \
     cargo run -p collector --bin collector -- \

collector/README.md

@@ -289,6 +289,13 @@ The mandatory `<PROFILER>` argument must be one of the following.
   - **Notes**. OProfile fails moderately often with this message: "operf-record
     process killed by signal 13". The failures seem to be random; re-running
     often results in success.
+- `samply`: Profile with [Samply](https://github.com/mstange/samply/), a
+  sampling profiler.
+  - **Purpose**. Samply is a general-purpose profiler, good for seeing
+    where execution time is spent and finding hot functions.
+  - **Slowdown**. Negligible.
+  - **Output**. Binary output is written to a file with a `samply` prefix.
+    That file can be loaded with `samply load`.
 - `cachegrind`: Profile with
   [Cachegrind](http://valgrind.org/docs/manual/cg-manual.html), a tracing
   profiler. Requires Valgrind 3.15 or later.

collector/src/bin/rustc-fake.rs

@@ -230,6 +230,19 @@ fn main() {
                 run_with_determinism_env(cmd);
             }
 
+            "Samply" => {
+                let mut cmd = Command::new("samply");
+                let has_samply = cmd.output().is_ok();
+                assert!(has_samply);
+                cmd.arg("record")
+                    .arg("--no-open")
+                    .arg("--save-only")
+                    .arg(&tool)
+                    .args(&args);
+
+                run_with_determinism_env(cmd);
+            }
+
             "Cachegrind" => {
                 let mut cmd = Command::new("valgrind");
                 let has_valgrind = cmd.output().is_ok();

collector/src/execute/mod.rs

@@ -51,6 +51,7 @@ impl PerfTool {
             | ProfileTool(SelfProfile)
             | ProfileTool(PerfRecord)
             | ProfileTool(Oprofile)
+            | ProfileTool(Samply)
             | ProfileTool(Cachegrind)
             | ProfileTool(Callgrind)
             | ProfileTool(Dhat)
@@ -86,6 +87,7 @@ impl PerfTool {
             | ProfileTool(SelfProfile)
             | ProfileTool(PerfRecord)
             | ProfileTool(Oprofile)
+            | ProfileTool(Samply)
             | ProfileTool(Cachegrind)
             | ProfileTool(Callgrind)
             | ProfileTool(Dhat)

collector/src/execute/profiler.rs

@@ -14,6 +14,7 @@ pub enum Profiler {
     SelfProfile,
     PerfRecord,
     Oprofile,
+    Samply,
     Cachegrind,
     Callgrind,
     Dhat,
@@ -206,6 +207,16 @@ impl<'a> Processor for ProfileProcessor<'a> {
                 fs::write(opann_file, &op_annotate_cmd.output()?.stdout)?;
             }
 
+            // Samply produces (via rustc-fake) a data file called
+            // `profile.json`. We copy it from the temp dir to the output dir,
+            // giving it a new name in the process.
+            Profiler::Samply => {
+                let tmp_samply_file = filepath(data.cwd.as_ref(), "profile.json");
+                let samply_file = filepath(self.output_dir, &out_file("samply"));
+
+                fs::copy(&tmp_samply_file, &samply_file)?;
+            }
+
             // Cachegrind produces (via rustc-fake) a data file called `cgout`.
             // We copy it from the temp dir to the output dir, giving it a new
             // name in the process, and then post-process it to produce another