benches: add bench_gil

Cargo.toml

@@ -32,6 +32,7 @@ serde = {version = "1.0", optional = true}
 
 [dev-dependencies]
 assert_approx_eq = "1.1.0"
+criterion = "0.3"
 trybuild = "1.0.23"
 rustversion = "1.0"
 proptest = { version = "0.10.1", default-features = false, features = ["std"] }
@@ -69,6 +70,10 @@ auto-initialize = []
 # Optimizes PyObject to Vec conversion and so on.
 nightly = []
 
+[[bench]]
+name = "bench_gil"
+harness = false
+
 [workspace]
 members = [
     "pyo3-macros",
@@ -83,4 +88,4 @@ members = [
 [package.metadata.docs.rs]
 no-default-features = true
 features = ["macros", "num-bigint", "num-complex", "hashbrown", "serde", "multiple-pymethods"]
-rustdoc-args = ["--cfg", "docsrs"]
+rustdoc-args = ["--cfg", "docsrs"]

benches/bench_gil.rs

@@ -0,0 +1,41 @@
+use criterion::{criterion_group, criterion_main, BatchSize, Bencher, Criterion};
+
+use pyo3::{prelude::*, GILPool};
+
+fn bench_clean_gilpool_new(b: &mut Bencher) {
+    Python::with_gil(|_py| {
+        b.iter(|| {
+            let _ = unsafe { GILPool::new() };
+        });
+    });
+}
+
+fn bench_clean_acquire_gil(b: &mut Bencher) {
+    // Acquiring first GIL will also create a "clean" GILPool, so this measures the Python overhead.
+    b.iter(|| {
+        let _ = Python::acquire_gil();
+    });
+}
+
+fn bench_dirty_acquire_gil(b: &mut Bencher) {
+    let obj = Python::with_gil(|py| py.None());
+    b.iter_batched(
+        || {
+            // Clone and drop an object so that the GILPool has work to do.
+            let _ = obj.clone();
+        },
+        |_| {
+            let _ = Python::acquire_gil();
+        },
+        BatchSize::NumBatches(1),
+    );
+}
+
+fn criterion_benchmark(c: &mut Criterion) {
+    c.bench_function("clean_gilpool_new", bench_clean_gilpool_new);
+    c.bench_function("clean_acquire_gil", bench_clean_acquire_gil);
+    c.bench_function("dirty_acquire_gil", bench_dirty_acquire_gil);
+}
+
+criterion_group!(benches, criterion_benchmark);
+criterion_main!(benches);