README.md

Performance Optimization Guide for Embedded Systems

This guide documents best practices for writing high-performance code targeting ARM Cortex-M microcontrollers, with specific focus on real-time motor control applications.

Table of Contents

1. Compiler Optimization Fundamentals
2. Writing Optimization-Friendly Code
3. Debug Mode Performance
4. Analyzing Generated Code
5. Measuring Performance
6. Common Pitfalls

---

Compiler Optimization Fundamentals

Optimization Levels

Flag	Description	Use Case
-O0	No optimization	Default debug, full debuggability
-Og	Debug-friendly optimization	Recommended for debug builds
-O1	Basic optimization	Faster compile, moderate speed
-O2	Standard optimization	Good balance of speed/size
-O3	Aggressive optimization	Maximum speed, may increase size
-Os	Size optimization	Flash-constrained systems

Critical Flags for Embedded

# Recommended flags for ARM Cortex-M4F
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} \
    -mcpu=cortex-m4 \
    -mfpu=fpv4-sp-d16 \
    -mfloat-abi=hard \
    -mthumb \
    -ffunction-sections \
    -fdata-sections")

# Linker flags to remove unused code
set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} -Wl,--gc-sections")

Fast-Math Considerations

# Aggressive floating-point optimizations for whole translation units, keeping NaN/Inf checks
add_compile_options(-ffast-math -fno-finite-math-only)

Effects of -ffast-math:

• No errno from math functions, so sqrtf becomes a single vsqrt.f32
• Allows reordering of operations and reciprocal multiplication instead of division
• Enables FMA (Fused Multiply-Add) instructions
• -ffinite-math-only (part of -ffast-math) assumes no NaN or Infinity and folds std::isnan/std::isfinite
• May break IEEE 754 compliance

Use only when: You control all inputs and don't need strict IEEE behavior. Keep -fno-finite-math-only unless no code relies on NaN or Infinity checks.

Set these options for whole translation units, never per function: see Optimization Options Must Match Across Calls.

---

Writing Optimization-Friendly Code

1. Avoid Virtual Functions in Hot Paths

Bad (virtual dispatch overhead):

class ITrigonometry {
public:
    virtual float Sine(float angle) const = 0;
};

// In hot path:
float sin_val = trig->Sine(angle);  // vtable lookup + indirect call

Good (static dispatch):

struct FastTrigonometry {
    static inline float Sine(float angle) noexcept {
        // Direct call, can be inlined
        return LookupTable[index];
    }
};

// In hot path:
float sin_val = FastTrigonometry::Sine(angle);  // Inlined

2. Avoid std::optional in Performance-Critical Code

Bad (generates has_value() checks):

std::optional<float> setPoint;

void Process(float input) {
    if (!setPoint.has_value())  // Extra branch + memory access
        return;
    // ...
}

Good (simple flag):

float setPointValue = 0.0f;
bool hasSetPoint = false;

void Process(float input) {
    if (!hasSetPoint) [[unlikely]]
        return;
    // ...
}

3. Use constexpr and inline Aggressively

// Computed at compile time
inline constexpr std::array<float, 512> sineLUT = []() {
    std::array<float, 512> table{};
    for (size_t i = 0; i < 512; ++i)
        table[i] = std::sin(2.0f * M_PI * i / 512.0f);
    return table;
}();

4. Use Compiler Attributes

// Force inlining even without optimization
#define ALWAYS_INLINE __attribute__((always_inline)) inline

// Mark hot functions for better code placement
#define HOT_FUNCTION __attribute__((hot))

// Combined macro for critical functions
#define OPTIMIZE_FOR_SPEED __attribute__((always_inline, hot)) inline

Optimization Options Must Match Across Calls

GCC does not inline a function into a caller whose optimization options differ from its own. #pragma GCC optimize and __attribute__((optimize(...))) give the functions they cover options of their own, so every small helper such a function calls — element access, accessors, unit wrappers — stays an out-of-line call, and the functions themselves are no longer inlined into their callers. Choose the optimization level and floating-point flags for whole translation units instead.

5. Prefer Fixed-Size Types

// Good: Explicit sizes, portable
uint32_t counter;
int16_t current_mA;
float voltage_V;

// Avoid: Implementation-defined sizes
int counter;
short current;

6. Minimize Stack Usage

// Bad: Large stack allocation
void Calculate() {
    float buffer[1024];  // 4KB on stack!
    // ...
}

// Good: Static or class member
class Calculator {
    static float buffer[1024];  // In .bss section
    // ...
};

7. Use FMA When Possible

The compiler generates FMA (Fused Multiply-Add) instructions with -ffast-math:

// This pattern:
result = a * b + c;

// Becomes single instruction:
// vfma.f32 s0, s1, s2  (1 cycle instead of 2)

---

Debug Mode Performance

Problem

By default, Debug builds (-O0) disable all optimizations, making code 3-10x slower than Release. This is problematic for:

• Real-time control loops (FOC, PID)
• Interrupt service routines
• Communication protocols with timing requirements

Solution 1: Use -Og for Debug Builds

# In CMakeLists.txt
set(CMAKE_CXX_FLAGS_DEBUG "-Og -g" CACHE STRING "Debug flags" FORCE)

-Og provides:

• Basic inlining
• Dead code elimination
• Register allocation
• Still debuggable (variable inspection works)

Solution 2: Per-Translation-Unit Options

# Build the performance-critical sources optimized while the rest stays debuggable
set_source_files_properties(Controller.cpp PROPERTIES COMPILE_OPTIONS "$<$<CONFIG:Debug>:-O2>")

Every function in a translation unit shares its options, so inlining is unaffected. Do not use #pragma GCC optimize or __attribute__((optimize(...))) instead: see Optimization Options Must Match Across Calls.

---

Analyzing Generated Code

Disassembly with objdump

# Basic disassembly
arm-none-eabi-objdump -d firmware.elf > disassembly.txt

# With C++ demangling
arm-none-eabi-objdump -d -C firmware.elf > disassembly.txt

# Specific function (grep pattern)
arm-none-eabi-objdump -d -C firmware.elf | grep -A 100 "FunctionName"

# From static library
arm-none-eabi-objdump -d -C libfoo.a | grep -A 50 "ClassName::Method"

# Include source interleaved (requires -g)
arm-none-eabi-objdump -d -S -C firmware.elf > disassembly_with_source.txt

Size Analysis

# Section sizes
arm-none-eabi-size firmware.elf

# Detailed symbol sizes (sorted by size)
arm-none-eabi-nm --size-sort -C firmware.elf

# Top 20 largest symbols
arm-none-eabi-nm --size-sort -C firmware.elf | tail -20

Reading Assembly Output

Key ARM Cortex-M4F instructions to look for:

Instruction	Meaning	Cycles
vfma.f32	Fused multiply-add	1
vmul.f32	Multiply	1
vadd.f32	Add	1
vdiv.f32	Division	14
vsqrt.f32	Square root	14
blx r3	Indirect call (virtual)	3+
bl <addr>	Direct call	1+N
push/pop	Stack operations	1-2

Signs of Poor Optimization

; Bad: Excessive stack operations
push    {r4, r5, r6, r7, r8, r9, r10, r11, lr}
sub     sp, #104        ; Large stack frame

; Bad: Virtual dispatch
ldr     r3, [r0, #0]    ; Load vtable pointer
ldr     r3, [r3, #4]    ; Load function pointer
blx     r3              ; Indirect call

; Bad: Repeated memory loads
ldr     r3, [r7, #4]    ; Same address loaded
; ... some code ...
ldr     r3, [r7, #4]    ; Again!

Signs of Good Optimization

; Good: Minimal stack usage
push    {r4, r5, lr}
sub     sp, #16

; Good: FMA instructions
vfma.f32  s0, s1, s2

; Good: Conditional execution (no branches)
vcmpe.f32 s0, s1
it        gt
vmovgt.f32 s0, s1

; Good: Loop unrolling
vldr    s0, [r0, #0]
vldr    s1, [r0, #4]
vldr    s2, [r0, #8]
vldr    s3, [r0, #12]

---

Measuring Performance

Cycle Counter (DWT)

// Enable cycle counter (do once at startup)
CoreDebug->DEMCR |= CoreDebug_DEMCR_TRCENA_Msk;
DWT->CYCCNT = 0;
DWT->CTRL |= DWT_CTRL_CYCCNTENA_Msk;

// Measure cycles
uint32_t start = DWT->CYCCNT;
CriticalFunction();
uint32_t cycles = DWT->CYCCNT - start;

GPIO Toggle Method

// Simple but requires oscilloscope
GPIO_SetPin(DEBUG_PIN);
CriticalFunction();
GPIO_ClearPin(DEBUG_PIN);
// Measure pulse width on scope

Timer-Based Measurement

// Using SysTick or hardware timer
uint32_t start = SysTick->VAL;
CriticalFunction();
uint32_t elapsed = start - SysTick->VAL;  // SysTick counts down

Typical Cycle Budgets (120 MHz Cortex-M4)

Control Loop Rate	Available Cycles
10 kHz	12,000 cycles
20 kHz	6,000 cycles
40 kHz	3,000 cycles
100 kHz	1,200 cycles

FOC typical requirements: 200-400 cycles (optimized), 800-1500 cycles (unoptimized)

---

Common Pitfalls

1. Heap Allocation

// NEVER in embedded hot paths
auto ptr = std::make_unique<Data>();  // malloc!
std::vector<float> buffer;            // malloc!
std::string message;                  // malloc!

2. Exception Handling Overhead

// Compile with: -fno-exceptions -fno-rtti
// Avoid try/catch in embedded code

3. printf/iostream in ISR

// NEVER in interrupt handlers
void ISR_Handler() {
    printf("Debug: %f\n", value);  // ~10,000+ cycles!
}

4. Floating-Point in Integer-Only Code

// Bad: Promotes to float
int result = value * 1.5;

// Good: Integer-only
int result = value * 3 / 2;

5. Unaligned Access

// Potential unaligned access (may cause fault or slow access)
struct __attribute__((packed)) BadStruct {
    uint8_t a;
    uint32_t b;  // Unaligned!
};

// Good: Natural alignment
struct GoodStruct {
    uint32_t b;
    uint8_t a;
    uint8_t padding[3];
};

---

Quick Reference Card

Optimization Flags (whole translation units)

add_compile_options(-O3)                                 # Maximum speed
add_compile_options(-Os)                                 # Minimum size
add_compile_options(-ffast-math -fno-finite-math-only)   # Aggressive FP, NaN/Inf checks kept

Function Attributes

__attribute__((always_inline))       // Force inline
__attribute__((noinline))            // Prevent inline
__attribute__((hot))                 // Optimize for speed
__attribute__((cold))                // Optimize for size
__attribute__((pure))                // No side effects
__attribute__((const))               // Pure + no memory reads
__attribute__((flatten))             // Inline all callees

Branch Hints

if (condition) [[likely]] { }        // C++20
if (condition) [[unlikely]] { }      // C++20
if (__builtin_expect(condition, 1))  // GCC

Useful objdump Commands

# Full disassembly with source
arm-none-eabi-objdump -d -S -C file.elf

# Just .text section
arm-none-eabi-objdump -d -j .text file.elf

# Show relocations
arm-none-eabi-objdump -d -r file.o

# Section headers
arm-none-eabi-objdump -h file.elf

---

References

• ARM Cortex-M4 Technical Reference Manual
• GCC Optimization Options
• ARM Compiler armasm User Guide
• Embedded Artistry - Embedded C++ Guidelines

---