diff --git a/benches/buffer.rs b/benches/buffer.rs
new file mode 100644
index 00000000..b32be433
--- /dev/null
+++ b/benches/buffer.rs
@@ -0,0 +1,71 @@
+#![feature(test, maybe_uninit_uninit_array_transpose)]
+extern crate test;
+
+use std::mem::MaybeUninit;
+
+// Call getrandom on a zero-initialized stack buffer
+#[inline(always)]
+fn bench_getrandom<const N: usize>() {
+    let mut buf = [0u8; N];
+    getrandom::getrandom(&mut buf).unwrap();
+    test::black_box(&buf as &[u8]);
+}
+
+// Call getrandom_uninit on an uninitialized stack buffer
+#[inline(always)]
+fn bench_getrandom_uninit<const N: usize>() {
+    let mut uninit = [MaybeUninit::uninit(); N];
+    let buf: &[u8] = getrandom::getrandom_uninit(&mut uninit).unwrap();
+    test::black_box(buf);
+}
+
+// We benchmark using #[inline(never)] "inner" functions for two reasons:
+//  - Avoiding inlining reduces a source of variance when running benchmarks.
+//  - It is _much_ easier to get the assembly or IR for the inner loop.
+//
+// For example, using cargo-show-asm (https://github.com/pacak/cargo-show-asm),
+// we can get the assembly for a particular benchmark's inner loop by running:
+//   cargo asm --bench buffer --release buffer::p384::bench_getrandom::inner
+macro_rules! bench {
+    ( $name:ident, $size:expr ) => {
+        pub mod $name {
+            #[bench]
+            pub fn bench_getrandom(b: &mut test::Bencher) {
+                #[inline(never)]
+                fn inner() {
+                    super::bench_getrandom::<{ $size }>()
+                }
+
+                b.bytes = $size as u64;
+                b.iter(inner);
+            }
+            #[bench]
+            pub fn bench_getrandom_uninit(b: &mut test::Bencher) {
+                #[inline(never)]
+                fn inner() {
+                    super::bench_getrandom_uninit::<{ $size }>()
+                }
+
+                b.bytes = $size as u64;
+                b.iter(inner);
+            }
+        }
+    };
+}
+
+// 16 bytes (128 bits) is the size of an 128-bit AES key/nonce.
+bench!(aes128, 128 / 8);
+
+// 32 bytes (256 bits) is the seed sized used for rand::thread_rng
+// and the `random` value in a ClientHello/ServerHello for TLS.
+// This is also the size of a 256-bit AES/HMAC/P-256/Curve25519 key
+// and/or nonce.
+bench!(p256, 256 / 8);
+
+// A P-384/HMAC-384 key and/or nonce.
+bench!(p384, 384 / 8);
+
+// Initializing larger buffers is not the primary use case of this library, as
+// this should normally be done by a userspace CSPRNG. However, we have a test
+// here to see the effects of a lower (amortized) syscall overhead.
+bench!(page, 4096);
diff --git a/benches/mod.rs b/benches/mod.rs
deleted file mode 100644
index 23b8c74c..00000000
--- a/benches/mod.rs
+++ /dev/null
@@ -1,64 +0,0 @@
-#![feature(test)]
-#![feature(maybe_uninit_as_bytes)]
-
-extern crate test;
-
-use std::mem::MaybeUninit;
-
-// Used to benchmark the throughput of getrandom in an optimal scenario.
-// The buffer is hot, and does not require initialization.
-#[inline(always)]
-fn bench_getrandom<const N: usize>(b: &mut test::Bencher) {
-    b.bytes = N as u64;
-    b.iter(|| {
-        let mut buf = [0u8; N];
-        getrandom::getrandom(&mut buf[..]).unwrap();
-        test::black_box(buf);
-    });
-}
-
-// Used to benchmark the throughput of getrandom is a slightly less optimal
-// scenario. The buffer is still hot, but requires initialization.
-#[inline(always)]
-fn bench_getrandom_uninit<const N: usize>(b: &mut test::Bencher) {
-    b.bytes = N as u64;
-    b.iter(|| {
-        let mut buf: MaybeUninit<[u8; N]> = MaybeUninit::uninit();
-        let _ = getrandom::getrandom_uninit(buf.as_bytes_mut()).unwrap();
-        let buf: [u8; N] = unsafe { buf.assume_init() };
-        test::black_box(buf)
-    });
-}
-
-macro_rules! bench {
-    ( $name:ident, $size:expr ) => {
-        pub mod $name {
-            #[bench]
-            pub fn bench_getrandom(b: &mut test::Bencher) {
-                super::bench_getrandom::<{ $size }>(b);
-            }
-
-            #[bench]
-            pub fn bench_getrandom_uninit(b: &mut test::Bencher) {
-                super::bench_getrandom_uninit::<{ $size }>(b);
-            }
-        }
-    };
-}
-
-// 16 bytes (128 bits) is the size of an 128-bit AES key/nonce.
-bench!(aes128, 128 / 8);
-
-// 32 bytes (256 bits) is the seed sized used for rand::thread_rng
-// and the `random` value in a ClientHello/ServerHello for TLS.
-// This is also the size of a 256-bit AES/HMAC/P-256/Curve25519 key
-// and/or nonce.
-bench!(p256, 256 / 8);
-
-// A P-384/HMAC-384 key and/or nonce.
-bench!(p384, 384 / 8);
-
-// Initializing larger buffers is not the primary use case of this library, as
-// this should normally be done by a userspace CSPRNG. However, we have a test
-// here to see the effects of a lower (amortized) syscall overhead.
-bench!(page, 4096);