ci: integrate callgrind benchmarks using iai-callgrind

.github/workflows/checks.yml

@@ -17,6 +17,23 @@ jobs:
         run: cargo bench --no-run
       - name: all features
         run: cargo bench --all-features --no-run
+  callgrind:
+    name: Callgrind
+    runs-on: ubuntu-latest
+    steps:
+      - name: checkout
+        uses: actions/checkout@v7
+      - name: install
+        uses: dtolnay/rust-toolchain@nightly
+      - name: default features
+        run: |
+          sudo apt-get update -y
+          sudo apt-get install -y valgrind
+          version=$(cargo metadata --format-version=1 | jq -r '.packages[] | select(.name == "iai-callgrind").version')
+          cargo install iai-callgrind-runner --version "$version"
+          cargo bench --bench callgrind --profile callgrind
+      - name: all features
+        run: cargo bench --bench callgrind --profile callgrind --all-features
   practices:
     name: Practices
     runs-on: ubuntu-latest

Cargo.toml

@@ -18,6 +18,16 @@ opt-level = 3
 debug = false
 strip = true
 
+# Dedicated profile for the iai-callgrind benchmarks. Callgrind needs the
+# function symbols in the binary to attribute event counts, so we disable
+# stripping (and keep debug info for line-level detail). Without symbols the
+# runner collects zero events for every benchmark. LTO is left on (inherited)
+# since it does not affect collection and keeps the run representative.
+[profile.callgrind]
+inherits = "bench"
+debug = 1
+strip = false
+
 [features]
 allocator-api = []
 allocator-api2 = ["dep:allocator-api2", "allocator-api"]
@@ -41,6 +51,7 @@ malloc_size_of = { version = "0.1", optional = true, default-features = false }
 [dev-dependencies]
 serde_test = "1.0"
 criterion = "0.8"
+iai-callgrind = "0.14"
 
 [[test]]
 name = "arbitrary"
@@ -75,6 +86,11 @@ name = "bench"
 path = "benches/bench.rs"
 harness = false
 
+[[bench]]
+name = "callgrind"
+path = "benches/callgrind.rs"
+harness = false
+
 [package.metadata.docs.rs]
 all-features = true
 rustdoc-args = ["--cfg", "docsrs"]