factor:performance tuning

.github/workflows/benchmarks.yml

@@ -75,5 +75,9 @@ jobs:
           mode: instrumentation
           run: |
             echo "Running benchmarks for ${{ matrix.benchmark-target.package }}"
-            cargo codspeed run -p ${{ matrix.benchmark-target.package }} > /dev/null
+            if [[ "${{ matrix.benchmark-target.package }}" == "uu_factor" ]]; then
+              cargo codspeed run -p ${{ matrix.benchmark-target.package }}
+            else
+              cargo codspeed run -p ${{ matrix.benchmark-target.package }} > /dev/null
+            fi
           token: ${{ secrets.CODSPEED_TOKEN }}

.vscode/cspell.dictionaries/jargon.wordlist.txt

@@ -164,6 +164,8 @@ wordlists
 xattrs
 xpass
 
+jemalloc
+jemallocator
 # * abbreviations
 AMPM
 ampm

Cargo.toml

@@ -331,6 +331,8 @@ icu_locale = "2.0.0"
 icu_provider = "2.0.0"
 indicatif = "0.18.0"
 itertools = "0.14.0"
+jemalloc-ctl = "0.5"
+jemallocator = "0.5"
 jiff = { version = "0.2.10", default-features = false, features = [
   "std",
   "alloc",
@@ -347,6 +349,7 @@ nix = { version = "0.30", default-features = false }
 nom = "8.0.0"
 notify = { version = "=8.2.0", features = ["macos_kqueue"] }
 num-bigint = "0.4.4"
+num-integer = "0.1"
 num-prime = "0.4.4"
 num-traits = "0.2.19"
 onig = { version = "~6.5.1", default-features = false }

src/uu/factor/Cargo.toml

@@ -34,6 +34,10 @@ divan = { workspace = true }
 rand = { workspace = true }
 uucore = { workspace = true, features = ["benchmark"] }
 
+[target.'cfg(any(target_os = "linux", target_os = "macos", target_os = "freebsd"))'.dev-dependencies]
+jemalloc-ctl = { workspace = true }
+jemallocator = { workspace = true }
+
 [lib]
 path = "src/factor.rs"
 

src/uu/factor/benches/factor_bench.rs

@@ -9,16 +9,42 @@ use divan::{Bencher, black_box};
 use uu_factor::uumain;
 use uucore::benchmark::run_util_function;
 
+#[cfg(any(target_os = "linux", target_os = "macos", target_os = "freebsd"))]
+use jemallocator::Jemalloc;
+
+#[cfg(any(target_os = "linux", target_os = "macos", target_os = "freebsd"))]
+#[global_allocator]
+static ALLOC: Jemalloc = Jemalloc;
+
+#[cfg(any(target_os = "linux", target_os = "macos", target_os = "freebsd"))]
+fn log_jemalloc_stats(label: &str) {
+    use jemalloc_ctl::{epoch, stats};
+
+    epoch::advance().unwrap();
+    let allocated = stats::allocated::read().unwrap();
+    let resident = stats::resident::read().unwrap();
+
+    println!(
+        "jemalloc {label}: allocated={} bytes, resident={} bytes",
+        allocated, resident
+    );
+}
+
+#[cfg(not(any(target_os = "linux", target_os = "macos", target_os = "freebsd")))]
+fn log_jemalloc_stats(_label: &str) {}
+
 /// Benchmark multiple u64 digits
 #[divan::bench(args = [(2)])]
 fn factor_multiple_u64s(bencher: Bencher, start_num: u64) {
     bencher
         // this is a range of 5000 different u128 integers
         .with_inputs(|| (start_num, start_num + 2500))
         .bench_values(|(start_u64, end_u64)| {
+            log_jemalloc_stats("before factor_multiple_u64s");
             for u64_digit in start_u64..=end_u64 {
                 black_box(run_util_function(uumain, &[&u64_digit.to_string()]));
             }
+            log_jemalloc_stats("after factor_multiple_u64s");
         });
 }