Skip to content

The truly free, open-source, local-first, cross-platform CapCut replacement. Scriptable through a JSON-RPC API & MCP.

ci(release): the Intel Mac builds with thin LTO, so a release no longer waits an hour on its link

Jareere525d7cbb24d
diff --git a/.github/workflows/build-app.yml b/.github/workflows/build-app.yml
index 4e5d6a1b..75651847 100644
--- a/.github/workflows/build-app.yml
+++ b/.github/workflows/build-app.yml
@@ -302,6 +302,23 @@ jobs:
echo "ORT_DIR=$PWD/$stem"
} >> "$GITHUB_ENV"
+ # Thin LTO and sixteen codegen units for the Intel Mac, in place of the
+ # app profile's fat LTO and one unit. The Intel runner is the slowest
+ # target and the one every release waits on: in v0.2.7-rc.4 its
+ # build was 70 minutes, 57 of them after the last dependency compiled -
+ # fat LTO re-optimising the whole program on one thread, which no cache
+ # can skip. Thin LTO does that work in parallel, for a binary a few
+ # per cent slower on the platform with the fewest users. Every other
+ # target keeps the profile as Cargo.toml writes it. Set before the
+ # cache step, so these are part of its key (CARGO is in env-vars).
+ - name: Thin LTO (Intel macOS)
+ if: matrix.platform == 'macos' && matrix.arch == 'x86_64'
+ run: |
+ {
+ echo "CARGO_PROFILE_APP_LTO=thin"
+ echo "CARGO_PROFILE_APP_CODEGEN_UNITS=16"
+ } >> "$GITHUB_ENV"
+
# ── Windows ────────────────────────────────────────────────────────
# LLVM: libclang for bindgen on both architectures, the C/C++
# compiler on arm64, and llvm-objdump for the instruction-set scan.
diff --git a/src/Cargo.toml b/src/Cargo.toml
index 481d5d25..3342098b 100644
--- a/src/Cargo.toml
+++ b/src/Cargo.toml
@@ -104,6 +104,8 @@ codegen-units = 256
inherits = "release"
# Fat LTO: re-optimise the whole dependency graph - slint, skia, winit and
# the UI crate - as one unit, so calls across crate boundaries inline.
+# The Intel Mac release build overrides this with thin LTO, for time; see
+# "Thin LTO (Intel macOS)" in .github/workflows/build-app.yml.
lto = "fat"
# No unwind tables and no landing pads: every call site loses its unwind
# edge, which frees the optimiser and shrinks the binary. This is a GUI