perf(codegen): make -Os the default native optimization level

changelog.d/8439-size-opt-default-on.md

@@ -0,0 +1 @@
+Make `-Os` the default native optimization level. Measured on the quiet bench mini under the sweep's bench lock, `-Os` costs no runtime speed on the benchmark corpus — every delta across `churn`, `cycles`, `fib40`, `interp`, `iso_miss` and `tree_wide` was within 0.8% and none disjoint from the `-O3` arm — while #8418 measured a further 346.7 MiB off a 4,743-module generated bundle. `PERRY_LL_SIZE_OPT=0` (or `off`/`false`/`no`) restores `-O3`.

crates/perry-codegen/src/linker.rs

@@ -307,11 +307,24 @@ fn cpu_tuning_arg_for(
     }
 }
 
+/// Size optimization is **on by default**. Measured on the quiet bench mini
+/// under the sweep's bench lock, `-Os` costs no measurable runtime speed: across
+/// `churn`, `cycles`, `fib40`, `interp`, `iso_miss` and `tree_wide` every delta
+/// was within 0.8% and none was disjoint from the `-O3` arm's samples. What it
+/// buys on a dense generated bundle is large — #8418 measured a further
+/// 346.7 MiB off a 4,743-module executable beyond the structural wins.
+///
+/// So the trade is compile time (about +21% on that bundle) against shipped
+/// binary size, with runtime speed unaffected. Smaller artifacts for every user
+/// are worth a developer-side build cost, so unset means enabled.
+///
+/// `PERRY_LL_SIZE_OPT=0` (or `off`/`false`/`no`) restores `-O3` for bisection or
+/// for a build that would rather have the compile time back.
 fn size_optimization_requested(value: Option<&str>) -> bool {
-    value
-        .map(str::trim)
-        .map(str::to_ascii_lowercase)
-        .is_some_and(|value| matches!(value.as_str(), "1" | "true" | "on" | "yes"))
+    match value.map(str::trim).map(str::to_ascii_lowercase) {
+        None => true,
+        Some(value) => !matches!(value.as_str(), "0" | "false" | "off" | "no"),
+    }
 }
 
 fn build_clang_compile_plan(
@@ -330,9 +343,10 @@ fn build_clang_compile_plan(
         cpu_tuning_arg_for(requested_cpu.as_deref(), target_triple, &effective_target);
     let stderr_remarks_path = PathBuf::from(format!("{}.clang-stderr", obj_path.display()));
 
-    // Perry defaults to speed-optimized native output. Generated-bundle users
-    // can explicitly trade runtime speed for artifact size with
-    // PERRY_LL_SIZE_OPT; there is no module-size-driven policy change.
+    // Perry defaults to SIZE-optimized native output: `-Os` measured no runtime
+    // cost on the benchmark corpus (see `size_optimization_requested`), and it
+    // materially shrinks dense generated bundles. `PERRY_LL_SIZE_OPT=0` restores
+    // `-O3`. There is no module-size-driven policy change.
     let size_opt = env::var("PERRY_LL_SIZE_OPT").ok();
     let opt_flag = if size_optimization_requested(size_opt.as_deref()) {
         "-Os"

crates/perry-codegen/src/linker_tests.rs

@@ -167,8 +167,9 @@ fn compile_plan_records_effective_target_and_native_tuning() {
         false,
     );
     assert!(plan.clang_args.contains(&"-fno-math-errno".to_string()));
-    // Native compilation defaults to speed-optimized -O3.
-    assert!(plan.clang_args.contains(&"-O3".to_string()));
+    // Native compilation defaults to size-optimized -Os (see
+    // `size_optimization_requested`); `PERRY_LL_SIZE_OPT=0` restores -O3.
+    assert!(plan.clang_args.contains(&"-Os".to_string()));
     assert!(plan.clang_args.contains(&"-target".to_string()));
     assert!(plan.analysis_clang_args.contains(&"-target".to_string()));
     // Apple aarch64 pins `apple-m1` rather than `native`: the decision to emit
@@ -185,11 +186,15 @@ fn compile_plan_records_effective_target_and_native_tuning() {
 }
 
 #[test]
-fn compile_plan_defaults_to_o3() {
+fn compile_plan_defaults_to_os() {
     // Module size is deliberately absent from the compile plan: Perry's
     // runtime optimization contract does not change for large generated IR.
     // Scalability is handled by codegen-unit partitioning and structured
     // outlining before LLVM sees the function bodies.
+    //
+    // The default optimization level is `-Os`: measured on the quiet bench mini,
+    // `-Os` costs no runtime speed on the benchmark corpus while materially
+    // shrinking dense generated bundles. `PERRY_LL_SIZE_OPT=0` restores `-O3`.
     let plan = build_clang_compile_plan(
         PathBuf::from("clang"),
         PathBuf::from("/tmp/input.ll"),
@@ -198,20 +203,24 @@ fn compile_plan_defaults_to_o3() {
         false,
         false,
     );
-    assert!(plan.clang_args.contains(&"-O3".to_string()));
-    assert!(!plan.clang_args.contains(&"-Os".to_string()));
+    assert!(plan.clang_args.contains(&"-Os".to_string()));
+    assert!(!plan.clang_args.contains(&"-O3".to_string()));
     assert!(!plan.clang_args.contains(&"-O0".to_string()));
 }
 
 #[test]
-fn size_optimization_flag_is_explicit_and_truthy() {
-    for enabled in ["1", "true", "TRUE", " on ", "yes"] {
+fn size_optimization_is_on_unless_explicitly_disabled() {
+    // Unset means enabled — the default flipped once `-Os` was measured to cost
+    // no runtime speed on the benchmark corpus.
+    assert!(size_optimization_requested(None));
+    for enabled in ["1", "true", "TRUE", " on ", "yes", "", "anything-else"] {
         assert!(size_optimization_requested(Some(enabled)), "{enabled}");
     }
-    for disabled in ["", "0", "false", "off", "no", "anything-else"] {
+    // Only an explicit negative restores `-O3`, for bisection or to buy back
+    // compile time.
+    for disabled in ["0", "false", "off", "no", "OFF", " 0 "] {
         assert!(!size_optimization_requested(Some(disabled)), "{disabled}");
     }
-    assert!(!size_optimization_requested(None));
 }
 
 #[test]