Build Profiling-Guided Optimized Clang for the Linux kernel

pgo.sh

#!/bin/bash
# This expects a specific directory layout and contents:
#
#  $HOME/src/linux        - clone of Linux repo (https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git)
#  $HOME/src/llvm-project - clone of LLVM repo (https://github.com/llvm/llvm-project.git)
#  $HOME/src/linux-build  - place to keep kernel builds
#  $HOME/src/llvm-build   - place to keep LLVM builds
#
#!/bin/bash
set -e

BUILD="$HOME/src/llvm-build"
PROJECT="$HOME/src/llvm-project/llvm"

# Stage 1 (for when we don't already have Clang)
CLANG=$(which clang)
if [ -z "$CLANG" ] ; then
        echo "Stage 1"
        cd "$BUILD"
        OUT="$BUILD/stage1"
        rm -rf "$OUT"
        mkdir "$OUT" && cd "$OUT"
        cmake -G Ninja --log-level=NOTICE \
                -DLLVM_TARGETS_TO_BUILD=X86 -DCMAKE_BUILD_TYPE=Release \
                -DLLVM_ENABLE_PROJECTS='clang;lld;compiler-rt;polly'    \
                -DLLVM_ENABLE_LLD=ON                    \
                -DCMAKE_C_COMPILER=gcc \
                -DCMAKE_CXX_COMPILER=g++ \
                -DCMAKE_ASM_COMPILER=gcc \
                -DCMAKE_INSTALL_PREFIX="$OUT/install"   \
                "$PROJECT"
        time ninja install

        STAGE1="$BUILD/stage1/install/bin"
else
        echo "Using existing $CLANG -- to rebuild, remove it first."
        STAGE1="$(dirname "$CLANG")"
fi

# Stage 2 (to enable collecting profiling data)
echo "Stage 2"
cd "$BUILD"
OUT="$BUILD/stage2-prof-gen"
if [ ! -d "$OUT" ]; then
        mkdir "$OUT" && cd "$OUT"
        cmake -G Ninja --log-level=NOTICE \
                -DCMAKE_BUILD_TYPE=Release \
                -DLLVM_ENABLE_PROJECTS='clang;lld;compiler-rt;polly'    \
                -DLLVM_EXPERIMENTAL_TARGETS_TO_BUILD='CSKY' \
                -DLLVM_ENABLE_LTO=Thin                  \
                -DLLVM_USE_LINKER=lld                   \
                -DCMAKE_C_COMPILER="$STAGE1"/clang      \
                -DCMAKE_CXX_COMPILER="$STAGE1"/clang++  \
                -DCMAKE_RANLIB="$STAGE1"/llvm-ranlib    \
                -DCMAKE_AR="$STAGE1"/llvm-ar            \
                -DLLVM_BUILD_INSTRUMENTED=ON            \
                -DCMAKE_INSTALL_PREFIX="$OUT/install"   \
                "$PROJECT"
        time ninja install
else
        echo "Using existing $OUT -- to rebuild, remove it first."
fi
STAGE2="$OUT/install/bin"
PROFILES="$OUT/profiles"
rm -rf "$PROFILES"/*

# Train PGO
cd "$HOME"/src/linux
WORKTREE="$HOME"/src/linux-build/clang
rm -rf "$WORKTREE"
git worktree prune
git worktree add --detach "$WORKTREE" master
cd "$WORKTREE"

echo "Training x86"
make -s allmodconfig CC="$STAGE2"/clang LD="$STAGE2"/ld.lld
time make -s -j$(getconf _NPROCESSORS_ONLN) --keep-going \
        CC="$STAGE2"/clang LD="$STAGE2"/ld.lld || true

echo "Training arm64"
make -s allmodconfig \
        CC="$STAGE2"/clang LD="$STAGE2"/ld.lld CROSS_COMPILE=aarch64-linux-gnu- ARCH=arm64
time make -s -j$(getconf _NPROCESSORS_ONLN) --keep-going \
        CC="$STAGE2"/clang LD="$STAGE2"/ld.lld CROSS_COMPILE=aarch64-linux-gnu- ARCH=arm64 || true

echo "Training arm"
make -s allmodconfig \
        CC="$STAGE2"/clang LD="$STAGE2"/ld.lld CROSS_COMPILE=arm-linux-gnueabihf- ARCH=arm
time make -s -j$(getconf _NPROCESSORS_ONLN) --keep-going \
        CC="$STAGE2"/clang LD="$STAGE2"/ld.lld CROSS_COMPILE=arm-linux-gnueabihf- ARCH=arm || true

# Merge training
cd "$PROFILES"
"$STAGE2"/llvm-profdata merge -output=clang.profdata *

# Stage 3 (built with PGO profile data)
echo "Stage 3"
cd "$BUILD"
OUT="$BUILD/stage3"
rm -rf "$OUT"
mkdir "$OUT" && cd "$OUT"
cmake -G Ninja --log-level=NOTICE \
        -DCMAKE_BUILD_TYPE=Release \
        -DLLVM_ENABLE_PROJECTS='clang;lld;compiler-rt;polly'    \
        -DLLVM_EXPERIMENTAL_TARGETS_TO_BUILD='CSKY' \
        -DLLVM_ENABLE_LTO=Thin                  \
        -DLLVM_USE_LINKER=lld                   \
        -DCMAKE_C_COMPILER="$STAGE1"/clang      \
        -DCMAKE_CXX_COMPILER="$STAGE1"/clang++  \
        -DCMAKE_RANLIB="$STAGE1"/llvm-ranlib    \
        -DCMAKE_AR="$STAGE1"/llvm-ar            \
        -DLLVM_PROFDATA_FILE="$PROFILES"/clang.profdata \
        -DCMAKE_INSTALL_PREFIX="$OUT/install"   \
        "$PROJECT"
time ninja install
STAGE3="$OUT/install/bin"
echo "$STAGE3"                                                                109,1         Bo