~kris/dots

srice

ref: e9b48d06a8541f3eda5c4db90382ab3c77183afb srice/doc/build-systems.md -rw-r--r-- 25.4 KiB
e9b48d06 — Kris Yotam xprofile: systemd-aware pipewire start + blueman-applet; sb-internet: tolerate missing /proc/net/wireless 2 months ago

#Build Systems Reference

Makefile patterns, compiler flags, and build acceleration tools for C/systems developers.

#1. Auto-Dependency Generation

Compiler generates .d files listing what each .o depends on. On the next make run, those files are included and stale objects are rebuilt automatically.

DEPDIR := .deps
DEPFLAGS = -MMD -MP -MF $(DEPDIR)/$*.d

SRCS := $(wildcard src/*.c)
OBJS := $(patsubst src/%.c, build/%.o, $(SRCS))

build/%.o: src/%.c | $(DEPDIR)
	$(CC) $(DEPFLAGS) $(CFLAGS) -c $< -o $@

$(DEPDIR):
	mkdir -p $@

-include $(patsubst src/%.c, $(DEPDIR)/%.d, $(SRCS))

Flag breakdown:

Flag Effect
-MMD Write a .d file with user-header deps (omits system headers)
-MP Add phony targets for each dep so deleted headers don't error
-MF <file> Write the dependency file to this path instead of stdout
-MD Like -MMD but includes system headers

The leading - on -include suppresses errors when .d files don't exist yet (first build).

#2. GNU Make Automatic Variables

Variable Expands to
$@ The target filename
$< The first prerequisite
$^ All prerequisites, deduplicated
$+ All prerequisites, with duplicates preserved
$? Prerequisites newer than the target
$* The stem matched by % in a pattern rule
$(@D) Directory part of $@
$(@F) File part of $@
$(<D) Directory part of $<
$(<F) File part of $<
$(^D) Directory parts of $^
$(^F) File parts of $^

Usage example:

build/%.o: src/%.c
	mkdir -p $(@D)
	$(CC) -c $< -o $@
	# $*  = the stem, e.g. "main" when building build/main.o

#3. Make Functions Reference

#Text manipulation

# subst: literal string replacement
$(subst from,to,text)
$(subst .c,.o,foo.c bar.c)          # foo.o bar.o

# patsubst: pattern substitution (% wildcard)
$(patsubst %.c,%.o,foo.c bar.c)     # foo.o bar.o
# Shorthand for patsubst on a variable:
OBJS := $(SRCS:.c=.o)

# filter: keep words matching pattern
$(filter %.c, foo.c foo.h bar.c)    # foo.c bar.c

# filter-out: remove words matching pattern
$(filter-out %.h, foo.c foo.h)      # foo.c

# sort: sort and deduplicate
$(sort bar foo foo baz)             # bar baz foo

# strip: remove extra whitespace
$(strip  foo  bar )                 # foo bar

# words / word / wordlist
$(words foo bar baz)                # 3
$(word 2, foo bar baz)              # bar
$(wordlist 2,3, foo bar baz)        # bar baz

# dir / notdir / basename / suffix / addprefix / addsuffix
$(dir src/foo.c)                    # src/
$(notdir src/foo.c)                 # foo.c
$(basename src/foo.c)               # src/foo
$(suffix src/foo.c)                 # .c
$(addprefix build/, foo.o bar.o)    # build/foo.o build/bar.o
$(addsuffix .o, foo bar)            # foo.o bar.o

#File discovery

# wildcard: glob expansion (unlike bare *, works inside functions)
SRCS := $(wildcard src/*.c src/**/*.c)

#Iteration and metaprogramming

# foreach: iterate over a list
DIRS := src lib test
$(foreach d, $(DIRS), $(wildcard $(d)/*.c))

# call: invoke a named function (user-defined macro)
reverse = $(2) $(1)
$(call reverse, foo, bar)           # bar foo

# Multi-arg call with define
define compile_rule
$(1)/%.o: $(2)/%.c
	$$(CC) $$(CFLAGS) -c $$< -o $$@
endef

$(eval $(call compile_rule, build, src))

#$$ escaping in eval

Inside define/eval blocks the text is expanded twice. Use $$ to produce a literal $ in the final rule:

define module_rule
$(1).o: $(1).c
	$$(CC) -c $$< -o $$@   # $$ becomes $ after first expansion
endef

$(eval $(call module_rule, main))

#4. Special Targets

Target Effect
.PHONY Declares targets that are not real files; always runs the recipe
.DELETE_ON_ERROR Deletes the target if its recipe exits with a nonzero status; prevents corrupt outputs being left in place
.SECONDARY Prevents intermediate files from being deleted after use; can list specific files or leave stem empty for all
.PRECIOUS Like .SECONDARY but also suppresses deletion on interrupt
.ONESHELL Runs all lines in a recipe in a single shell invocation; lets you use multi-line scripts without backslash continuations
.NOTPARALLEL Disables parallel execution for the current Makefile (use sparingly; prefer per-target ordering instead)
.SECONDEXPANSION Enables a second expansion pass on prerequisites; allows $$(VAR) tricks for dynamic dependency lists
.DEFAULT Recipe to run when no rule is found for a target
.SUFFIXES Controls the list of known suffixes for old-style suffix rules; set to empty to disable them 
.DELETE_ON_ERROR:
.PHONY: all clean test

# SECONDEXPANSION example
.SECONDEXPANSION:
build/%.o: $$(patsubst build/%.o,src/%.c,$@)
	$(CC) -c $< -o $@

# ONESHELL example
.ONESHELL:
gen-version:
	VER=$$(git describe --tags)
	echo "#define VERSION \"$$VER\"" > version.h

#5. Parallel Builds

make -j$(nproc)                 # use all logical CPUs
make -j$(nproc) -Otarget        # serialize output per target (readable logs)
make -j8                        # fixed job count

Output sync options for -O:

Value Behavior
none No synchronization (default)
line Output one line at a time
target Buffer all output per target, print when done
recurse Like target but for recursive make

Limiting parallelism for specific targets:

# These two targets must not run simultaneously
link: .NOTPARALLEL
link: $(OBJS)
	$(CC) $^ -o $@

Or use ordering prerequisites:

b: a        # b waits for a, even with -j
a:
	sleep 1 && echo a
b: a
	echo b

#6. Non-Recursive Make with Includes

Recursive make (calling $(MAKE) in subdirectories) loses dependency information across directories and is slow. The alternative is a single top-level Makefile that includes per-module module.mk files.

Top-level Makefile:

# Reset before each include
SRCS :=
CFLAGS_EXTRA :=

include src/lib/module.mk
include src/app/module.mk
include src/test/module.mk

OBJS := $(patsubst %.c,build/%.o,$(SRCS))

src/lib/module.mk:

# Paths relative to project root
SRCS += src/lib/alloc.c \
        src/lib/buf.c   \
        src/lib/hash.c

# Module-specific flags appended to global compile rule
CFLAGS_EXTRA += -Isrc/lib/include

Pattern rule handles all modules:

build/%.o: %.c
	@mkdir -p $(@D)
	$(CC) $(CFLAGS) $(CFLAGS_EXTRA) $(DEPFLAGS) -c $< -o $@

Benefits: single pass, full cross-directory dependency tracking, faster than recursive invocations.

#7. ccache

ccache intercepts compiler calls, hashes inputs, and returns cached object files on hits. Typical speedup on a warm cache: 5-100x.

#Install

# Arch
sudo pacman -S ccache

# Debian/Ubuntu
sudo apt install ccache

# macOS
brew install ccache

#Use in Makefile

CC  := ccache gcc
CXX := ccache g++

Or prefix via environment without changing the Makefile:

export CC="ccache gcc"
export CXX="ccache g++"
make -j$(nproc)

Or symlink method (intercepts without changing any config):

export PATH="/usr/lib/ccache:$PATH"
# /usr/lib/ccache contains gcc, g++, cc, c++ symlinks pointing to ccache

#Key config commands

ccache --show-stats          # hit rate, cache size, cache dir
ccache --show-config         # effective config
ccache --zero-stats          # reset counters
ccache --clear               # purge all cached objects

# Config (persistent)
ccache --set-config max_size=10G
ccache --set-config cache_dir=~/.cache/ccache
ccache --set-config compression=true
ccache --set-config compression_level=6

# Useful env overrides
export CCACHE_DIR=~/.cache/ccache
export CCACHE_MAXSIZE=10G
export CCACHE_COMPRESS=1
export CCACHE_SLOPPINESS=pch_defines,time_macros   # needed for PCH caching

#Cache miss causes to eliminate

  • __DATE__ / __TIME__ macros (use -DVERSION instead)
  • Absolute paths baked in by -I or __FILE__ (use -ffile-prefix-map=old=new)
  • Changing umask between builds

#8. distcc

distcc distributes compilation across a cluster. Each remote machine compiles preprocessed source sent by the local machine. Works transparently as a compiler wrapper.

#Install

sudo pacman -S distcc        # Arch
sudo apt install distcc      # Debian

#Daemon setup (on each remote build host)

# Edit /etc/conf.d/distccd or /etc/default/distcc
DISTCC_ARGS="--allow 192.168.1.0/24 --jobs 8 --log-stderr"

sudo systemctl enable --now distccd

#Local configuration

export DISTCC_HOSTS="localhost/4 192.168.1.10/8 192.168.1.11/8"
# Format: host/max_jobs

#Makefile integration

CC  := distcc gcc
CXX := distcc g++

#Pump mode (faster preprocessing)

Pump mode offloads the preprocessor step to remotes too, reducing data sent over the wire:

pump make -j32 CC="distcc gcc"

#Combine distcc + ccache

CC  := ccache distcc gcc
CXX := ccache distcc g++

ccache checks locally first; on a miss, distcc handles compilation remotely.

#Monitor

distccmon-text 1      # refresh every 1 second
distccmon-gnome       # GUI monitor if installed

#9. bear for compile_commands.json

compile_commands.json is required by clangd, clang-tidy, include-what-you-use, and most editors. bear intercepts the compiler via LD_PRELOAD and records every compilation command.

sudo pacman -S bear      # Arch
sudo apt install bear    # Debian

#Generate

bear -- make -j$(nproc)
# Writes compile_commands.json in the current directory

#Incremental update (bear 3.x)

bear --append -- make changed_module.o

#Alternatives

# cmake generates it natively
cmake -DCMAKE_EXPORT_COMPILE_COMMANDS=ON ..

# intercept-build (part of scan-build/clang-tools)
intercept-build make

# compiledb (Python, wraps make)
pip install compiledb
compiledb make

#10. mold Linker

mold is a modern linker written in C++. On large projects it links 10-20x faster than GNU ld and 2-5x faster than lld.

#Install

sudo pacman -S mold       # Arch
sudo apt install mold     # Debian 12+

# Or from source
git clone https://github.com/rui314/mold
cd mold && cmake -DCMAKE_BUILD_TYPE=Release -B build && cmake --build build -j$(nproc)
sudo cmake --install build

#Use

# Pass to compiler driver
gcc -fuse-ld=mold -o prog $(OBJS)
g++ -fuse-ld=mold -o prog $(OBJS)

# Or set in Makefile
LDFLAGS += -fuse-ld=mold

#Makefile pattern

LD      := $(CC)
LDFLAGS := -fuse-ld=mold -Wl,--as-needed

$(TARGET): $(OBJS)
	$(LD) $(LDFLAGS) $^ -o $@

#Notes

  • -fuse-ld=mold requires GCC 12+ or Clang 12+; on older compilers use -B /usr/libexec/mold
  • Not for static libraries (those use ar, not a linker)
  • ThinLTO is supported: add -flto=thin to both CFLAGS and LDFLAGS

#11. Precompiled Headers

Precompiling a large or stable header (like a project-wide common.h or a bundled stb.h) eliminates repeated parse time for every translation unit that includes it.

#Create a PCH

PCH_SRC := src/precompiled.h
PCH_OUT := build/precompiled.h.gch

$(PCH_OUT): $(PCH_SRC) | build
	$(CC) $(CFLAGS) -x c-header $< -o $@

#Use the PCH

GCC looks for a .gch file alongside the original header automatically:

build/%.o: src/%.c $(PCH_OUT)
	$(CC) $(CFLAGS) -include src/precompiled.h -c $< -o $@

The -include flag forces inclusion before anything in the source file. If build/precompiled.h.gch exists and was compiled with the same flags, GCC uses it automatically.

#Clang

Clang requires explicit -include-pch:

$(PCH_OUT): $(PCH_SRC)
	$(CC) $(CFLAGS) -x c-header $< -o $@

build/%.o: src/%.c $(PCH_OUT)
	$(CC) $(CFLAGS) -include-pch $(PCH_OUT) -c $< -o $@

#ccache + PCH

export CCACHE_SLOPPINESS=pch_defines,time_macros

Without this, ccache misses on PCH users because it is conservative about flag differences.

#12. Unity Builds

A unity build concatenates all .c files into one and compiles the single translation unit. The linker sees one object. Benefits: no repeated header parsing, better inlining across TUs, smaller binaries. Drawback: all-or-nothing rebuilds.

#Manual unity file

/* unity.c */
#include "src/alloc.c"
#include "src/buf.c"
#include "src/hash.c"
#include "src/io.c"

unity: build/unity.o
	$(CC) build/unity.o -o $@

build/unity.o: unity.c
	$(CC) $(CFLAGS) -c $< -o $@

#Generated unity file

SRCS := $(wildcard src/*.c)

build/unity.c: $(SRCS)
	@mkdir -p $(@D)
	$(file > $@,/* generated */)
	$(foreach s,$^,$(file >> $@,#include "../$(s)"))

build/unity.o: build/unity.c
	$(CC) $(CFLAGS) -c $< -o $@

#Hybrid: per-module unity + parallel compilation

# Each module gets its own unity file; modules compile in parallel
MODULES := net io fs crypto

define unity_rule
build/$(1)-unity.c: $$(wildcard src/$(1)/*.c)
	$$(file > $$@,/* $(1) */)
	$$(foreach s,$$^,$$(file >> $$@,#include "../../$$s"))

build/$(1)-unity.o: build/$(1)-unity.c
	$$(CC) $$(CFLAGS) -c $$< -o $$@
endef

$(foreach m,$(MODULES),$(eval $(call unity_rule,$(m))))

#13. Make Debugging

#Dry run and tracing

make -n                          # print commands without running them
make -n --always-make            # -n but pretend everything is out of date
make --trace                     # print why each target is rebuilt
make --debug=a                   # verbose: all decisions, including up-to-date
make --debug=v                   # verbose: variable values
make --debug=i                   # verbose: implicit rule search
make --warn-undefined-variables  # warn whenever $(UNDEFINED) expands to empty

make -p                          # dump all rules, variables, and implicit rules
make -p -f /dev/null             # dump builtins without running any Makefile

#In-Makefile diagnostics

# $(info ...) prints at parse time, no error
$(info SRCS = $(SRCS))
$(info CC   = $(CC))

# $(warning ...) prints with filename:line prefix
$(warning Building with flags: $(CFLAGS))

# $(error ...) prints and immediately aborts
ifeq ($(CC),)
$(error CC is not set)
endif

# origin tells you where a variable was defined
$(info origin of CC: $(origin CC))   # default, environment, file, command line, override, automatic

#Inspecting a specific target

make --trace target
make -d target 2>&1 | less       # very verbose, full decision log

#14. Compiler Flags Reference

#Debug build

CFLAGS_DEBUG := \
    -O0                 \   # no optimization; fast compilation
    -g3                 \   # max debug info including macros
    -fno-omit-frame-pointer \
    -fsanitize=address,undefined \
    -fstack-protector-strong \
    -DDEBUG

#Release build

CFLAGS_RELEASE := \
    -O2                 \   # safe optimizations
    -DNDEBUG            \
    -ffunction-sections \   # enables --gc-sections to strip dead code
    -fdata-sections     \
    -flto               \   # link-time optimization
    -march=native           # tune for the build machine (not for distributed binaries)

LDFLAGS_RELEASE := \
    -Wl,--gc-sections   \
    -Wl,-O1             \
    -flto

#Profiling build

CFLAGS_PROF := \
    -O2                 \
    -g                  \   # keep symbols for perf/gprof
    -pg                 \   # gprof instrumentation
    -fno-omit-frame-pointer \
    -DNDEBUG

For perf you do not need -pg, just -g -fno-omit-frame-pointer.

#Full warnings

WARN_FLAGS := \
    -Wall               \   # common warnings
    -Wextra             \   # more warnings
    -Wpedantic          \   # strict ISO C
    -Wshadow            \
    -Wconversion        \
    -Wsign-conversion   \
    -Wdouble-promotion  \
    -Wformat=2          \
    -Wundef             \
    -fno-common         \
    -Wunused            \
    -Wmissing-prototypes \   # C only
    -Wstrict-prototypes     # C only

Promote warnings to errors in CI:

CFLAGS += -Werror

#Hardening flags (production binaries)

CFLAGS_HARDEN := \
    -D_FORTIFY_SOURCE=2     \
    -fstack-protector-strong \
    -fPIE

LDFLAGS_HARDEN := \
    -pie                    \
    -Wl,-z,relro            \
    -Wl,-z,now

#15. mk (Plan 9 Make)

mk is the Plan 9 equivalent of make, designed to be simpler and cleaner.

#Key differences from GNU make

Feature GNU make mk
Shell Uses /bin/sh per recipe line Uses rc shell (Plan 9) or shell of choice
Parallel flag -j N -n N
Automatic variables $@, $<, $^ $target, $prereq, $newprereq
Pattern rules %.o: %.c %.$O: %.c
All prerequisites $^ $prereq
Variable export Manual Automatic via MKFLAGS
Dependency files Manual -include Native `< mkdepend` integration

#mk syntax

CC = gcc
CFLAGS = -O2 -Wall

%.o: %.c
    $CC $CFLAGS -c $prereq -o $target

prog: main.o util.o
    $CC $prereq -o $target

#Install on Linux

# 9base (Plan 9 userland port)
sudo pacman -S 9base     # Arch AUR
# or build from suckless.org/tools/9base

#Advantages

  • Simpler rule syntax
  • No recursive variable expansion surprises
  • mkfile can be used with the 9front toolchain directly
  • Recipe is a single string passed to a shell, so .ONESHELL is the default behavior

#16. Ninja Build System

Ninja is a low-level build system focused purely on speed. It does no rule inference; higher-level tools (CMake, Meson, GN) generate .ninja files.

#Install

sudo pacman -S ninja      # Arch
sudo apt install ninja-build   # Debian

#Key speed advantages over make

  • Parses its input files in parallel
  • Uses a dependency log (.ninja_deps, .ninja_log) instead of re-stat-ing every file
  • Avoids shell overhead: single process, no fork per recipe line
  • Implicit dep scanning built in (no need for -MMD tricks)

#Basic ninja file syntax

cc = gcc
cflags = -O2 -Wall

rule cc
  command = $cc $cflags -MMD -MF $out.d -c $in -o $out
  depfile = $out.d
  deps = gcc

rule link
  command = $cc $in -o $out

build build/main.o: cc src/main.c
build build/util.o: cc src/util.c
build prog: link build/main.o build/util.o

default prog

#Run

ninja                     # build default target
ninja -j$(nproc)          # explicit parallelism (default is all CPUs)
ninja -v                  # verbose (show full commands)
ninja -n                  # dry run
ninja -d explain          # why is each target rebuilt

#Generating ninja files

# CMake
cmake -GNinja -DCMAKE_BUILD_TYPE=Release ..
ninja

# Meson
meson setup build
cd build && ninja

# GN (Chromium ecosystem)
gn gen out/Release
ninja -C out/Release

#17. Complete Production Makefile Template

# ------------------------------------------------------------
# Project config
# ------------------------------------------------------------
TARGET  := prog
SRCDIR  := src
BLDDIR  := build
DEPDIR  := $(BLDDIR)/.deps

CC      := ccache gcc
CXX     := ccache g++
AR      := ar
LD      := $(CC)

# ------------------------------------------------------------
# Sources and objects
# ------------------------------------------------------------
SRCS    := $(shell find $(SRCDIR) -name '*.c')
OBJS    := $(patsubst $(SRCDIR)/%.c, $(BLDDIR)/%.o, $(SRCS))

# ------------------------------------------------------------
# Flags
# ------------------------------------------------------------
CFLAGS  := -std=c11 -Wall -Wextra -Wpedantic -Wshadow \
           -ffunction-sections -fdata-sections
LDFLAGS := -fuse-ld=mold -Wl,--gc-sections -Wl,--as-needed
LIBS    :=

DEPFLAGS = -MMD -MP -MF $(DEPDIR)/$*.d

ifeq ($(BUILD),debug)
    CFLAGS  += -O0 -g3 -fsanitize=address,undefined -fno-omit-frame-pointer -DDEBUG
    LDFLAGS += -fsanitize=address,undefined
else ifeq ($(BUILD),release)
    CFLAGS  += -O2 -DNDEBUG -flto -march=native
    LDFLAGS += -flto
else
    CFLAGS  += -Og -g -fno-omit-frame-pointer
endif

# ------------------------------------------------------------
# Special targets
# ------------------------------------------------------------
.DELETE_ON_ERROR:
.PHONY: all clean distclean test compile_commands

# ------------------------------------------------------------
# Default target
# ------------------------------------------------------------
all: $(BLDDIR)/$(TARGET)

# ------------------------------------------------------------
# Link
# ------------------------------------------------------------
$(BLDDIR)/$(TARGET): $(OBJS)
	$(LD) $(LDFLAGS) $^ $(LIBS) -o $@

# ------------------------------------------------------------
# Compile
# ------------------------------------------------------------
$(BLDDIR)/%.o: $(SRCDIR)/%.c | $(DEPDIR)
	@mkdir -p $(@D) $(dir $(DEPDIR)/$*.d)
	$(CC) $(CFLAGS) $(DEPFLAGS) -c $< -o $@

# ------------------------------------------------------------
# Directories
# ------------------------------------------------------------
$(BLDDIR) $(DEPDIR):
	mkdir -p $@

# ------------------------------------------------------------
# Dependencies
# ------------------------------------------------------------
-include $(wildcard $(DEPDIR)/**/*.d $(DEPDIR)/*.d)

# ------------------------------------------------------------
# compile_commands.json
# ------------------------------------------------------------
compile_commands:
	bear -- $(MAKE) clean all

# ------------------------------------------------------------
# Clean
# ------------------------------------------------------------
clean:
	rm -rf $(BLDDIR)

distclean: clean
	rm -f compile_commands.json

# ------------------------------------------------------------
# Test
# ------------------------------------------------------------
test: $(BLDDIR)/$(TARGET)
	./tests/run.sh $<

# ------------------------------------------------------------
# Debug helpers
# ------------------------------------------------------------
print-%:
	@echo '$* = $($*)'
# Usage: make print-CFLAGS

#Usage

make                       # debug-ish build (Og)
make BUILD=release -j$(nproc)
make BUILD=debug -j$(nproc)
make compile_commands
make print-CFLAGS

#18. Incremental Linking with -r and module.ro

Partial linking (-r) combines multiple object files into a single relocatable object without resolving symbols. The result can be linked again later. Useful for large projects with stable subsystems.

#Create a partial object

build/net.ro: $(NET_OBJS)
	$(LD) -r $^ -o $@

build/io.ro: $(IO_OBJS)
	$(LD) -r $^ -o $@

build/crypto.ro: $(CRYPTO_OBJS)
	$(LD) -r $^ -o $@

$(TARGET): build/net.ro build/io.ro build/crypto.ro build/main.o
	$(LD) $(LDFLAGS) $^ $(LIBS) -o $@

#Advantages

  • Subsystem objects can be compiled independently and cached
  • Only the changed subsystem's .ro is rebuilt; others are reused
  • The final link step only sees a handful of .ro files instead of hundreds of .o files, making it faster
  • Can hide internal symbols with --localize-hidden on partial objects
build/net.ro: $(NET_OBJS)
	$(LD) -r $^ -o $@.tmp
	objcopy --localize-hidden $@.tmp $@
	rm $@.tmp

This enforces subsystem boundaries: symbols marked __attribute__((visibility("hidden"))) inside net/ become local after the partial link and cannot be referenced by other modules.

#Highest Impact Optimizations Ranked by Time Saved

Ranked by observed wall-clock build time reduction on a typical C project (1-500k SLOC). Your numbers will vary based on hardware and project structure.

Rank Optimization Typical Speedup Notes
1 -j$(nproc) parallel make 4-16x Free; limited by the critical path and I/O
2 ccache (warm cache) 5-100x Near-instant for unchanged files across branches
3 mold linker 10-20x (link step) Link is often the serializing bottleneck
4 Precompiled headers 2-5x High value when headers are large and stable
5 Unity builds 2-10x Eliminates repeated header parsing; best for CI
6 distcc (cluster) Linear with machines Needs fast network and homogeneous compilers
7 Non-recursive make 1.2-2x Eliminates repeated make invocations and shell forks
8 Auto-dependency generation Correct > fast Avoids unnecessary full rebuilds on header changes
9 -O0 in debug builds 1.5-3x compile Slower binary, but much faster compilation
10 Ninja over make 1.1-1.5x Parse overhead matters most on projects with thousands of targets
11 Incremental linking (-r) 1.2-2x (link) Effective when only one subsystem changes at a time
12 -flto=thin Smaller binary, slower build Worth it for release; use lld or mold to parallelize

#Combining for maximum effect

# Development inner loop (full acceleration)
CC="ccache gcc" make -j$(nproc) -Otarget BUILD=debug

# CI clean build (unity + parallel + mold)
make unity -j$(nproc) LDFLAGS="-fuse-ld=mold"

# Release (LTO + mold + ccache)
CC="ccache gcc" make -j$(nproc) BUILD=release LDFLAGS="-fuse-ld=mold -flto"

The single highest-ROI change for most projects starting from a naive serial make: add -j$(nproc) and set CC := ccache gcc. Takes 30 seconds and cuts iteration time dramatically.