From 68703e3e4708b255ec7bdab3a365532cc5122617 Mon Sep 17 00:00:00 2001
From: Khalil Estell <kammce0@gmail.com>
Date: Sun, 18 Oct 2026 14:02:29 -0700
Subject: [PATCH] :zap: (patch) Split Cortex-M LLVM deploys into three jobs

A libhal-util 6.0.0-alpha.1 deploy hit the 45 minute Cortex-M timeout:
each of its two LLVM jobs built 18-24 packages at about 110 seconds
each. Half of that time was the test_package, which rebuilds every
imported C++ module (mp-units, libhal, libhal-util) to link one program.

- The deploy's LLVM cores are built in three jobs instead of two, with
  each job's cores listed in the Plan job: M0-M4, M4F-M23 and M33-M85.
- baremetal_package.sh only runs the test_package for the first package
  it builds. Later packages pass --test-folder "".
---
 .github/scripts/baremetal_package.sh         |  9 +++++++--
 .github/workflows/package_and_upload_all.yml | 15 ++++++++-------
 README.md                                    |  7 ++++++-
 3 files changed, 21 insertions(+), 10 deletions(-)

diff --git a/.github/scripts/baremetal_package.sh b/.github/scripts/baremetal_package.sh
index 74f3fd5..efe67a4 100644
--- a/.github/scripts/baremetal_package.sh
+++ b/.github/scripts/baremetal_package.sh
@@ -82,11 +82,16 @@ pipx install conan>=$CONAN_VERSION
 conan config install https://github.com/libhal/conan-config2.git
 conan hal setup
 
-# Loop over architectures and build types
+# Loop over architectures and build types. Only the first package runs the
+# test_package: it rebuilds every imported C++ module just to link one small
+# program, which roughly doubles the cost of each package, and linking against
+# the other architectures adds little.
+TEST_FOLDER_ARGS=()
 for ARCH in "${ARCH_LIST[@]}"; do
   echo "Building for architecture: $ARCH"
 
   for BUILD_TYPE in "${BUILD_TYPES[@]}"; do
-    conan create $DIR -s:h build_type=$BUILD_TYPE -s:h os=baremetal -s:h arch=$ARCH --version $VERSION -pr:h $COMPILER_PROFILE --build=missing $EXTRA_CONAN_ARGS
+    conan create $DIR -s:h build_type=$BUILD_TYPE -s:h os=baremetal -s:h arch=$ARCH --version $VERSION -pr:h $COMPILER_PROFILE --build=missing "${TEST_FOLDER_ARGS[@]}" $EXTRA_CONAN_ARGS
+    TEST_FOLDER_ARGS=(--test-folder "")
   done
 done
diff --git a/.github/workflows/package_and_upload_all.yml b/.github/workflows/package_and_upload_all.yml
index 5a61fcf..8be8b5d 100644
--- a/.github/workflows/package_and_upload_all.yml
+++ b/.github/workflows/package_and_upload_all.yml
@@ -102,14 +102,15 @@ jobs:
               '[{name: "M sample", archs: $archs}]')
             cortex_m_gcc="cortex-m0plus,$sample"
           else
-            # Deploy: every build type and Cortex-M core. LLVM splits the
-            # cores across two jobs because one job for all of them is slow.
+            # Deploy: every build type and Cortex-M core. Building every core in
+            # one job is slow, so LLVM builds them in three parallel jobs.
             build_types="Debug,MinSizeRel,Release"
-            older="cortex-m3,cortex-m4,cortex-m4f,cortex-m7f,cortex-m7d"
-            newer="cortex-m23,cortex-m33,cortex-m33f,cortex-m35pf,cortex-m55,cortex-m85"
-            cortex_m_llvm=$(jq -nc --arg older "$armv6m$older" --arg newer "$newer" \
-              '[{name: "M0-M7", archs: $older}, {name: "M23-M85", archs: $newer}]')
-            cortex_m_gcc="cortex-m0,cortex-m0plus,cortex-m1,$older,$newer"
+            cortex_m_llvm=$(jq -nc --arg armv6m "$armv6m" '[
+              { name: "M0-M4",   archs: ($armv6m + "cortex-m3,cortex-m4") },
+              { name: "M4F-M23", archs: "cortex-m4f,cortex-m7f,cortex-m7d,cortex-m23" },
+              { name: "M33-M85", archs: "cortex-m33,cortex-m33f,cortex-m35pf,cortex-m55,cortex-m85" }
+            ]')
+            cortex_m_gcc="cortex-m0,cortex-m0plus,cortex-m1,cortex-m3,cortex-m4,cortex-m4f,cortex-m7f,cortex-m7d,cortex-m23,cortex-m33,cortex-m33f,cortex-m35pf,cortex-m55,cortex-m85"
           fi
 
           {
diff --git a/README.md b/README.md
index 749ea06..c32df77 100644
--- a/README.md
+++ b/README.md
@@ -409,7 +409,7 @@ It runs in one of two modes, picked by `version`:
 | --- | --- | --- |
 | Build types | `MinSizeRel` | `Debug`, `MinSizeRel`, `Release` |
 | Linux x86_64, Linux armv8, macOS armv8 (LLVM) | ✅ | ✅ |
-| Cortex-M (LLVM) | M0+, M3, M4F, M33F in one job | M0, M0+, M1, M3, M4, M4F, M7F, M7D and M23, M33, M33F, M35PF, M55, M85 in two jobs |
+| Cortex-M (LLVM) | M0+, M3, M4F, M33F in one job | M0, M0+, M1, M3, M4, M4F, M7F, M7D, M23, M33, M33F, M35PF, M55, M85, split evenly across three jobs |
 | Cortex-M (GCC, when `modules_support_needed` is false) | M0+, M3, M4F, M33F | every Cortex-M architecture |
 
 Check mode keeps pull requests and pushes to main cheap. Deploys still build
@@ -420,6 +420,11 @@ place and the build jobs read its outputs. ARMv6-M (M0, M0+, M1) is left out of
 Host jobs time out after 30 minutes and Cortex-M jobs after 45, so a hung
 runner can't run for hours.
 
+Within each Cortex-M job, only the first package runs its `test_package`. A
+test_package rebuilds every imported C++ module (mp-units, libhal, ...) just to
+link one program, so running it for every architecture and build type roughly
+doubled the job's time.
+
 **Usage:**
 
 ```yaml