#import "Basic";
#import "Compiler";
#import "Math";

// TODO Comparing implementaitons using dump

#run test_math_ext();

test_math_ext :: () { set_build_options_dc(.{do_output=false});
// main :: () {

    write_strings("=====================\n", "--- Test Math_Ext ---\n");

    // Different signals: only works if signaled variable is higher.
    /*
    #run cena();
    cena :: () {
        a: s64 = -232;
        b: u32 = 4;
        c := a+b;
        print("\n\n--- --- ---\ntttt : % : % + % = %\n--- --- ---\n\n", type_of(c), a, b, c);
    }
    */

    #import "Random";
    test_add :: (x: $Tx, y: $Ty, r: $Tr, t: Type, o: bool) -> errors_found: int {
        tr, to := add(cast(Tx)x, cast(Ty)y);
        print("add(%): % + % = % : %\n", t, x, y, r, o);
        error := 0;
        if r != tr          { error += 1; print(" > incorrect result value: got % expected %\n", tr, r); };
        if t != type_of(tr) { error += 1; print(" > incorrect result type: got % expected %\n", type_of(tr), t); };
        if o != to          { error += 1; print(" > incorrect overflow flag: got % expected %\n", to, o); };
        return error;
    }

    errors := 0;

    // test_add(cast(u8)1,       cast(u8)2,   3, u8, false);
    // test_add(cast(u8)255,       cast(u8)1,   255, u8, true);

    errors += test_add(cast( s8) S8_MAX,  cast( s8)1,  S8_MAX,     s8, true);
    errors += test_add(cast(s16)S16_MAX,  cast( u8)1, S16_MAX,    s16, true);
    errors += test_add(cast(s32)S32_MAX,  cast(s32)1, S32_MAX,    s32, true);
    errors += test_add(cast(s64)S64_MAX,  cast(u32)1, S64_MAX,    s64, true);

    errors += test_add(cast( u8) U8_MAX,  cast( u8)1,  U8_MAX,     u8, true);
    errors += test_add(cast(u16)U16_MAX,  cast(u16)1, U16_MAX,    u16, true);
    errors += test_add(cast(u32)U32_MAX,  cast(u32)1, U32_MAX,    u32, true);
    errors += test_add(cast(u64)U64_MAX,  cast(u64)1, U64_MAX,    u64, true);
        
    // errors += test_add(cast(s32)66,       cast(s64)-2,   64, s64, false);
    // errors += test_add(cast(u32)66,       cast(s64)4,    70, s64, false);
    // errors += test_add(cast(s32)S32_MAX,  cast(s64)1,    2147483648, s64, false);
    // errors += test_add(cast(s32)S32_MAX,  cast(s32)1,    S32_MAX, s32, true);
    // errors += test_add(cast(s64)S64_MAX,  cast(s64)0,    S64_MAX, s64, false);
    // errors += test_add(cast(s64)9223372036854775806,  cast(s64)1,    S64_MAX, s64, false);
    // errors += test_add(cast(s64)9223372036854775806,  cast(s64)2,    S64_MAX, s64, true);

    // errors += test_add(cast(u8)7, cast(u8)1, 8, u8, false);
    // errors += test_add(cast(u8)U8_MAX, cast(u8)1, U8_MAX, u8, true);

    // errors += test_add(cast(u16)10, cast(u8)3, 13, u16, false);
    // errors += test_add(cast(u8)1, cast(u16)U16_MAX, U16_MAX, u16, true);

    if errors > 0 print("# Found % %!\n", errors, ifx errors == 1 then "error" else "errors"); else print("  No errors found.\n");

    
/* PERFORMANCE TEST
    best_generic: float;
    best_asm: float;
    for 0..100 {
        size, time_generic, time_asm := performance_test();
        perf_generic    := cast(float)size/cast(float)to_microseconds(time_generic);
        perf_asm        := cast(float)size/cast(float)to_microseconds(time_asm);
        best_generic    = max(best_generic, perf_generic);
        best_asm        = max(best_asm, perf_asm);
    }

    print("method  : ops/usec\ngeneric : %\nasm     : %\n", best_generic, best_asm);

    performance_test :: () -> sum_size: s64, time_generic: Apollo_Time, time_asm: Apollo_Time {
    
        SUM_SIZE := 2000000;
        numbers: [..] s64;
        array_reserve(*numbers, SUM_SIZE);
        
        for 0..SUM_SIZE-1 {
            array_add(*numbers, cast(s64)random_get());
        }

        sum := 0;
        start := current_time_monotonic();
        for numbers sum = add_bad(sum, it);
        time := current_time_monotonic() - start;

        sum_asm := 0;
        start_asm := current_time_monotonic();
        for numbers sum_asm = add(sum_asm, it);
        time_asm := current_time_monotonic() - start_asm;

        assert(sum == sum_asm);

        return SUM_SIZE, time, time_asm;
    }
*/
}

add :: (x: $Tx, y: $Ty) -> result: $Tr, saturated: bool //#dump
    #modify {
        type_info_x := cast(*Type_Info)Tx;
        type_info_y := cast(*Type_Info)Ty;
        if type_info_x.type != .INTEGER || type_info_y.type != .INTEGER return false, "Non integers values passed.";
        tx := cast(*Type_Info_Integer)type_info_x;
        ty := cast(*Type_Info_Integer)type_info_y;

        largest_type :=
            ifx tx.runtime_size > ty.runtime_size then Tx else
            ifx ty.runtime_size > tx.runtime_size then Ty else
            ifx tx.signed == ty.signed then Tx else
            void;

        // Only allow to add different signedness values if largest type is the signed one (as in JAI).
        if tx.signed == ty.signed {
            Tx = largest_type;
            Ty = largest_type;
            Tr = largest_type;
        }
        else if tx.signed && Tx == largest_type {
            Ty = largest_type;
            Tr = largest_type;
        }
        else if ty.signed && Ty == largest_type {
            Tx = largest_type;
            Tr = largest_type;
        }
        else return false, "Number signedness mismatch.";

        print(">tx:ty:%:%\n", Tx, Ty);
        return true;
    }
{

    #if CPU != .X64 {
        
        #if Tr == s8 || Tr == s16 || Tr == s32 || Tr == s64 {

            #if Tr ==  s8 { MAX ::  S8_MAX; MIN ::  S8_MIN; }
            #if Tr == s16 { MAX :: S16_MAX; MIN :: S16_MIN; }
            #if Tr == s32 { MAX :: S32_MAX; MIN :: S32_MIN; }
            #if Tr == s64 { MAX :: S64_MAX; MIN :: S64_MIN; }
        
            if (y > 0 && x > MAX - y) then return MAX, true;
            if (y < 0 && x < MIN - y) then return MIN, true;
            return x + y, false;
            
        } else {

            #if Tr ==  u8 { MAX ::  U8_MAX; }
            #if Tr == u16 { MAX :: U16_MAX; }
            #if Tr == u32 { MAX :: U32_MAX; }
            #if Tr == u64 { MAX :: U64_MAX; }
        
            if (x > MAX - y) then return MAX, true;
            return x + y, false;
            
        }
        
    } else {

        #import "String";
        result: Tr = ---;
        saturated: bool = ---;


        S_ADD_ASM :: #string DONE
        #asm {
            // Calculate limit based on x's sign.
            mov     limit: gpr, MAX;
            mov     sign: gpr, x;
            shr     sign, BITS;
            add.SIZE     limit, sign;

            mov     result, x;
            add.SIZE   result, y;
            seto    saturated;
            cmovo   result, limit;
        }
        DONE

        #if Tr == s8
            #insert #run replace(replace(replace(S_ADD_ASM, ".SIZE", ".b"), "MAX", "127"), "BITS", "7");
        #if Tr == s16
            #insert #run replace(replace(replace(S_ADD_ASM, ".SIZE", ".w"), "MAX", "32767"), "BITS", "15");
        #if Tr == s32
            #insert #run replace(replace(replace(S_ADD_ASM, ".SIZE", ".d"), "MAX", "2147483647"), "BITS", "31");
        #if Tr == s64
            #insert #run replace(replace(replace(S_ADD_ASM, ".SIZE", ".q"), "MAX", "9223372036854775807"), "BITS", "63");

        
        U_ADD_ASM :: #string DONE
        #asm { // s8
            mov     limit: gpr, MAX;
            mov     result, x;
            add.SIZE    result, y;
            setc    saturated;
            cmovc   result, limit;
        }
        DONE

        #if Tr == u8
            #insert #run replace(replace(U_ADD_ASM, ".SIZE", ".b"), "MAX", "255");
        #if Tr == u16
            #insert #run replace(replace(U_ADD_ASM, ".SIZE", ".w"), "MAX", "65535");
        #if Tr == u32
            #insert #run replace(replace(U_ADD_ASM, ".SIZE", ".d"), "MAX", "4294967295");
        #if Tr == u64
            #insert #run replace(replace(U_ADD_ASM, ".SIZE", ".q"), "MAX", "18446744073709551615");


        return result, saturated;
        
    }
}

sub :: (x: $Tx, y: $Ty) -> result: $Tr, overflow: bool //#dump
    #modify {
        return true;
    }
{
    #if CPU != .X64 {
        if (y < 0 && x > S64_MAX + y) then return S64_MAX, true;
        if (y > 0 && x < S64_MIN + y) then return S64_MIN, true;
        return x - y, false;
    } else {
        return x + y, false; // TODO Implement me please.
    }
}