cpuid cache query 2

#include <stdio.h>
#include <stdint.h>

typedef uint8_t u8;
typedef uint32_t u32;
typedef uint64_t u64;
typedef u32 b32;

typedef float r32;

struct cpuid_data {
    u32 eax;
    u32 ebx;
    u32 ecx;
    u32 edx;
};

char *cache_types[] = {
    "",
    "Data",
    "Instr",
    "Shared",
};


#if 1
#include <intrin.h>

struct cpuid_data get_cpuid(u32 leaf) {
    struct cpuid_data result = {0};

    __cpuid((u32*)&result, leaf);

    return result;
}

struct cpuid_data get_cpuid_count(u32 leaf, u32 count) {
    struct cpuid_data result = {0};

    __cpuidex((u32*)&result, leaf, count);

    return result;
}

#else

#include <cpuid.h>
struct cpuid_data get_cpuid(u32 leaf) {
    struct cpuid_data result = {0};

    __cpuid(leaf, result.eax, result.ebx, result.ecx, result.edx);

    return result;
}

struct cpuid_data get_cpuid_count(u32 leaf, u32 count) {
    struct cpuid_data result = {0};

    __cpuid_count(leaf, count, result.eax, result.ebx, result.ecx, result.edx);

    return result;
}

#endif




int main(int argc, char **argv) {
    (void)argc;
    (void)argv;

    struct cpuid_data maximum_supported = get_cpuid(0x0);

    if(maximum_supported.eax > 0x4) {
        u32 index = 0;
        while(1) {
            struct cpuid_data cpuid_cache_configuration_04 = get_cpuid_count(0x4, index);

            // NOTE(peter): When cache type is zero we exit.
            if((cpuid_cache_configuration_04.eax & 0x1f) == 0) {
                break;
            }

            u32 max_cores_per_package = (cpuid_cache_configuration_04.eax >> 26) + 1;
            u32 max_threads_per_cache = (cpuid_cache_configuration_04.eax >> 14 & 0xfff) + 1;
            // note(peter): bits 13...10 reserved
            u32 fully_associative     = cpuid_cache_configuration_04.eax >> 9 & 0x1;
            u32 self_initializing     = cpuid_cache_configuration_04.eax >> 8 & 0x1;
            u32 cache_level           = cpuid_cache_configuration_04.eax >> 5 & 0x7;
            u32 cache_type            = cpuid_cache_configuration_04.eax & 0x1f;

            u32 ways_of_associativity      = (cpuid_cache_configuration_04.ebx >> 22) + 1;
            u32 physical_line_partitions   = (cpuid_cache_configuration_04.ebx >> 12 & 0x3ff) + 1;
            u32 system_coherency_line_size = (cpuid_cache_configuration_04.ebx  & 0xfff) + 1;

            u32 sets = (cpuid_cache_configuration_04.ecx) + 1;

            u32 complex_cache_indexing = (cpuid_cache_configuration_04.edx >> 2) & 0x1;
            u32 cache_inclusiveness    = (cpuid_cache_configuration_04.edx >> 1) & 0x1;
            u32 write_back_invalidate  = (cpuid_cache_configuration_04.edx) & 0x1;

            u64 cache_size_ = ways_of_associativity * physical_line_partitions * system_coherency_line_size * sets;
            b32 cache_size_kb = (cache_size_ < (1024*1024));
            r32 cache_size = (cache_size_kb) ? cache_size_ / 1024.f : (cache_size_ / (1024.f*1024.f));

            printf("Cache index: %d\n", index);
            printf("  Package max cores: %d\n", max_cores_per_package);
            printf("  Cache max threads: %d\n", max_threads_per_cache);
            printf("  Self initializing: %d\n", self_initializing);
            printf("  %s\n", fully_associative ? "Fully associative" : "Not fully associative");
            printf("  Cache: L%d %s (%.2f %s) %d-way set associative\n", cache_level, cache_types[cache_type], cache_size, cache_size_kb ? "KB" : "MB", ways_of_associativity);
            printf("  Physical line partitions: %d\n", physical_line_partitions);
            printf("  System coherency line size: %d\n", system_coherency_line_size);
            printf("  Sets: %d\n", sets);
            printf("  Cache indexing: %s\n", complex_cache_indexing ? "A complex function is used to index the cache, potentially using all address bits." : "Direct mapped");
            printf("  Inclusive to lower caches: %s\n", cache_inclusiveness ? "True" : "False");
            printf("  WBINVD/INVD on lower caches: %s\n", write_back_invalidate ? "True" : "False");
            printf("\n");
            index++;
        }
    }

    return 0;
}