ESP32 Indirect function call performance test

esp32_perf_test.c

#include <stdint.h>
#include <stdio.h>
#include <stdbool.h>
#include <unistd.h>
#include <limits.h>
#include "esp_timer.h"
#include "esp_task_wdt.h"

#define TEST_COUNT 10000000

int fn_1(int a) {
    return a * 8;
}

int fn_4(int a) {
    return a * 2;
}

int fn_8(int a) {
    return a;
}

int fn_12(int a) {
    return (a >> 1) + (a >> 3) + (a >> 5) + (a >> 7) + (a >> 9) + (a >> 11) + (a >> 13) + (a >> 15) + (a >> 17);
}

int fn_16(int a) {
    return a >> 1;
}

int direct_if(int fn) {
    int i, b = 0;
    if (fn == 1) {
        for (i = 0; i < TEST_COUNT; ++i)
            b = fn_1(b + i);
    } else if (fn == 4) {
        for (i = 0; i < TEST_COUNT; ++i)
            b = fn_4(b + i);
    } else if (fn == 8) {
        for (i = 0; i < TEST_COUNT; ++i)
            b = fn_8(b + i);
    } else if (fn == 12) {
        for (i = 0; i < TEST_COUNT; ++i)
            b = fn_12(b + i);
    } else if (fn == 16) {
        for (i = 0; i < TEST_COUNT; ++i)
            b = fn_16(b + i);
    }
    return b;
}

int direct_switch(int fn) {
    int i, b = 0;
    switch (fn) {
        case 1:
            for (i = 0; i < TEST_COUNT; ++i)
                b = fn_1(b + i);
            break;
        case 4:
            for (i = 0; i < TEST_COUNT; ++i)
                b = fn_4(b + i);
            break;
        case 8:
            for (i = 0; i < TEST_COUNT; ++i)
                b = fn_8(b + i);
            break;
        case 12:
            for (i = 0; i < TEST_COUNT; ++i)
                b = fn_12(b + i);
            break;
        case 16:
            for (i = 0; i < TEST_COUNT; ++i)
                b = fn_16(b + i);
    }
    return b;
}

typedef int (*fn_t)(int);

enum {
    TestedFBits = 5
};
fn_t FPointers[TestedFBits] = { fn_1, fn_4, fn_8, fn_12, fn_16 };
int FIndexes[TestedFBits] = { 1, 4, 8, 12, 16 };

int indirect_version(fn_t fn) {
    int i, b = 0;
    for (i = 0; i < TEST_COUNT; ++i) {
        b = fn(b + i);
    }
    return b;
}

int GlobalVar;

void app_main(void) {
    int64_t start_clock, end_clock;
    start_clock = esp_timer_get_time();
    for (int8_t i = 0; i < TestedFBits; i++) {
        int bits = FIndexes[i];
        GlobalVar += direct_if(bits);
    }
    end_clock = esp_timer_get_time();
    printf("direct_if took: %lld micros\n", end_clock - start_clock);
    start_clock = esp_timer_get_time();
    for (int8_t i = 0; i < TestedFBits; i++) {
        int bits = FIndexes[i];
        GlobalVar += direct_switch(bits);
    }
    end_clock = esp_timer_get_time();
    printf("direct_switch took: %lld micros\n", end_clock - start_clock);
    start_clock = esp_timer_get_time();
    for (int8_t i = 0; i < TestedFBits; i++) {
        GlobalVar += indirect_version(FPointers[i]);
    }
    end_clock = esp_timer_get_time();
    printf("indirect_version took: %lld micros\n", end_clock - start_clock);
    while (1) {
        GlobalVar++;
    }
}

__DATA__
gcc 11.2
-O2:
direct_if took: 1254457 micros
direct_switch took: 1672600 micros
indirect_version took: 4390699 micros

-Os:
direct_if took: 1965483 micros
direct_switch took: 2132584 micros
indirect_version took: 4599714 micros