perf: reduce release executable size by 65.8%

.github/workflows/release.yml

@@ -23,13 +23,14 @@ jobs:
           submodules: true
       - uses: dtolnay/rust-toolchain@nightly
         with:
+          components: rust-src
           targets: ${{ matrix.target }}
       - name: Check the tag matches the crate version
         shell: bash
         run: |
           version=$(cargo metadata --no-deps --format-version 1 | jq -r '.packages[] | select(.name == "enable-loopback") | .version')
           test "v$version" = "$GITHUB_REF_NAME" || { echo "tag $GITHUB_REF_NAME does not match crate version $version"; exit 1; }
-      - run: cargo build --release --locked --target ${{ matrix.target }}
+      - run: cargo build --release --locked --target ${{ matrix.target }} -Z build-std=std,panic_abort -Z build-std-features=optimize_for_size
       - name: Name the asset and its checksum
         shell: bash
         run: |

Cargo.toml

@@ -10,3 +10,11 @@ repository = "https://github.com/libnyanpasu/enable-loopback.git"
 
 [profile.dev]
 debug = "line-tables-only"
+
+[profile.release]
+opt-level = "z"
+lto = "fat"
+codegen-units = 1
+# Native GUI callbacks cannot unwind across the Windows ABI boundary.
+panic = "abort"
+strip = "symbols"

README.md

@@ -26,6 +26,23 @@ cargo run -- --lang zh-cn
 If a global Cargo configuration uses a compiler wrapper, set `RUSTC_WRAPPER=`
 to bypass it for these commands. Do not change the global configuration.
 
+For a size-optimized release, use the same command as the release workflow:
+
+```sh
+cargo build --release --locked --target x86_64-pc-windows-msvc \
+  -Z build-std=std,panic_abort -Z build-std-features=optimize_for_size
+```
+
+Use `aarch64-pc-windows-msvc` for arm64. This requires nightly and `rust-src`
+(included in `rust-toolchain.toml`). The executable is in
+`target/<target>/release/enableLoopback.exe`. A plain `cargo build --release`
+also uses the size profile, but keeps the precompiled standard library and is
+slightly larger. Release builds use full LTO, one codegen unit, size optimization,
+stripped symbols, and abort on panic; development builds retain unwinding and
+line tables. Recoverable application errors still use the existing error dialogs.
+
+See [release size measurements](docs/release-size.md) for the measured tradeoffs.
+
 `enableLoopback.exe [--lang CODE]` opens the GUI. Supported languages:
 `en`, `zh-cn`, `zh-tw`, `ja`, `ko`, `ru`; otherwise the system language is used,
 with English as the fallback. `manual` skips automatic refresh and `-registry`

docs/release-size.md

@@ -0,0 +1,56 @@
+# Release executable size
+
+Measured on 2026-10-21 with the locked dependencies, windows-rs submodule
+`75c2364d1ba98d4b1fb9a9f28e2a9b0fc067e6e6`, and
+`rustc 1.100.0-nightly (a36d05efa 2026-09-09)`, targeting Windows x64 MSVC.
+Sizes are the actual uncompressed EXE, including its icon and manifest.
+
+| Configuration | Bytes | MiB | Reduction from baseline |
+| --- | ---: | ---: | ---: |
+| Original default release profile | 3,802,624 | 3.63 | — |
+| Full LTO, one codegen unit, `s`, abort, strip | 1,773,056 | 1.69 | 53.4% |
+| Full LTO, one codegen unit, `z`, abort, strip | 1,442,304 | 1.38 | 62.1% |
+| Same `z` profile, size-optimized standard library | 1,302,016 | 1.24 | 65.8% |
+
+`z` saves 330,752 bytes compared with `s` in this application. Rebuilding the
+standard library saves another 140,288 bytes. These measurements select the
+release profile and the release workflow's `build-std` options; they do not
+establish an absolute lower bound. Compiler updates and architectures can change
+the result.
+
+The original PE contained 2,680,320 bytes of code, 894,976 bytes of read-only data,
+and only 55,808 bytes of icon/manifest resources. Optimizing generated code
+therefore provided much more benefit than changing the artwork. The implementation
+keeps all six languages, the GUI, CLI, and existing runtime deployment model.
+No executable packer is used.
+
+Full LTO trades build time for size. Size optimization may trade throughput for
+size; interactive behavior was checked, but this is not a performance benchmark.
+`panic = "abort"` terminates on programming panics instead of unwinding; it does
+not change ordinary `Result` error handling. Native GUI callbacks already cannot
+unwind across the Windows ABI. Tests retain the test harness's unwind behavior.
+Symbol stripping reduces postmortem debugging information. The additional
+standard-library optimization requires nightly's unstable `build-std` support
+and `rust-src`; the normal release command remains available without it.
+
+Validation:
+
+- All 52 application tests passed in both development and release profiles.
+- `cargo clippy --all-targets --locked -- -D warnings` and application formatting
+  checks passed.
+- ARM64 compilation reached the final link, but this machine has only x64/x86
+  MSVC build tools and no ARM64 linker. Its EXE size and runtime behavior remain
+  unverified locally; the release workflow retains both architecture targets.
+- The smallest x64 EXE loaded all 136 real AppContainers. Native GUI checks
+  exercised checkbox edits and counter updates, refresh cancellation and refresh,
+  dirty-close confirmation and cancellation, and clean close. Firewall exemptions
+  were compared before/after and were unchanged; Save was never invoked.
+- The repository's existing GUI smoke script has stale sort-header expectations
+  and requests a SelectionItemPattern unavailable on this runtime's rows; the
+  original baseline fails those checks too. A temporary adapter under
+  `target/size-study` used the current header names, omitted native selection/sort
+  assertions, and waited for dialog dismissal animations. Selection and sorting
+  are covered by the application tests, but were not verified through native UIA.
+
+Reproduce the smallest build with the command in the README. Intermediate x64
+EXEs from this investigation are retained locally in `target/size-study/`.

rust-toolchain.toml

@@ -1,3 +1,3 @@
 [toolchain]
 channel = "nightly"
-components = ["rustfmt", "clippy"]
+components = ["rustfmt", "clippy", "rust-src"]