Enable precise Cachegrind collection mode

Author: Kobzol

Cargo.lock

@@ -43,3 +43,9 @@ benchlib = { path = "benchlib" }
 [target.'cfg(windows)'.dependencies]
 miow = "0.3"
 windows-sys = { version = "0.36.1", features = ["Win32_Foundation"] }
+
+[features]
+# Enable more precise Cachegrind profiles for runtime benchmarks.
+# Requires a recent Valgrind to be installed.
+# Pass DEP_VALGRIND=<path-to-valgrind>/include environment variable when building.
+precise-cachegrind = []

collector/Cargo.toml

@@ -488,7 +488,15 @@ It is also possible to profile runtime benchmarks using the following command:
 ```
 
 Currently, a `<PROFILER>` can be `cachegrind`, which will run the runtime benchmark under
-`Cachegrind`.
+`Cachegrind`. If you pass `--features precise-cachegrind`, you can get more precise profiling results.
+In this mode, Cachegrind will only record the instructions of the actual benchmark, and ignore any
+other code (e.g. benchmark initialization). To use this mode, you need to provide a path to a Valgrind
+build directory (at least Valgrind 3.22 is required), like this:
+
+```
+DEP_VALGRIND=<path-to-valgrind-install>/include cargo run --release --bin collector \
+  --features precise-cachegrind profile_runtime cachegrind <RUSTC> <BENCHMARK_NAME> 
+```
 
 ## Codegen diff
 You can use the `codegen_diff` command to display the assembly, LLVM IR or MIR difference between two

collector/README.md

@@ -15,9 +15,11 @@ env_logger = "0.10.0"
 clap = { version = "4.1", features = ["derive", "string"] }
 libc = "0.2"
 flate2 = { version = "1", optional = true }
+crabgrind = { version = "0.1.10", optional = true }
 
 [target.'cfg(target_os = "linux")'.dependencies]
 perf-event = "0.4.7"
 
 [features]
 compression = ["dep:flate2"]
+precise-cachegrind = ["dep:crabgrind"]

collector/benchlib/Cargo.toml

@@ -1,4 +1,17 @@
 pub fn profile_function<F: Fn() -> Bench, R, Bench: FnOnce() -> R>(benchmark_constructor: &F) {
     let func = benchmark_constructor();
+
+    // With the `precise-cachegrind` feature, we want to enable cachegrind recording
+    // only for the actual execution of the profiled function.
+    #[cfg(feature = "precise-cachegrind")]
+    {
+        crabgrind::cachegrind::start_instrumentation();
+    }
+
     func();
+
+    #[cfg(feature = "precise-cachegrind")]
+    {
+        crabgrind::cachegrind::stop_instrumentation();
+    }
 }

collector/benchlib/src/profile.rs

@@ -358,6 +358,10 @@ fn start_cargo_build(
         command.arg(target_dir);
     }
 
+    // Enable the precise-cachegrind feature for the benchlib dependency of the runtime group.
+    #[cfg(feature = "precise-cachegrind")]
+    command.arg("--features").arg("benchlib/precise-cachegrind");
+
     let child = command
         .spawn()
         .map_err(|error| anyhow::anyhow!("Failed to start cargo: {:?}", error))?;

collector/src/runtime/benchmark.rs

@@ -39,6 +39,13 @@ pub fn profile_runtime(
                 .arg("--branch-sim=no")
                 .arg("--cache-sim=no")
                 .arg(format!("--cachegrind-out-file={}", cgout_tmp.display()));
+
+            // Disable cachegrind profile collection at start.
+            // It will be enabled only for the profiled function using
+            // Valgrind client requests (see `benchlib/src/profile.rs`).
+            #[cfg(feature = "precise-cachegrind")]
+            cmd.arg("--instr-at-start=no");
+
             cmd.stdin(Stdio::null());
             cmd.arg(&group.binary).arg("profile").arg(benchmark);
             command_output(&mut cmd).context("Cannot run profiler")?;