// Tests for integer saturating arighmetic (with branch-free procedures on x64).
// Expects signed values in two's complement.

#import "Basic";
#import "Compiler";
#import "Math";
#load "Math_Ext.jai";

main :: () {

    write_strings(
        "#========================#\n",
        "# Unit tests             #\n"
        );

    test_op :: (operation: string, x: $Tx, y: $Ty, result: $Tr, type: Type, saturated: bool, remainder: Tr = 0) -> errors_found: int #expand {

        print_test_call :: (operation: string) -> string {
            #import "String";
            str: string = ---;
            if operation != "div" {
                TEST_CALL :: #string DONE
                t_result, t_saturated := OP(cast(Tx)x, cast(Ty)y);
                if result != t_result print("%_%(%, %) = %0%0\n", operation, type, x, y, result, ifx saturated then " : saturated");
                DONE
                str = replace(TEST_CALL, "OP", operation);
            } else {
                TEST_CALL :: #string DONE
                t_result, t_remainder, t_saturated := OP(cast(Tx)x, cast(Ty)y);
                if result != t_result print("%_%(%, %) = % + %0%0\n", operation, type, x, y, result, remainder, ifx saturated then " : saturated");
                DONE
                str = replace(TEST_CALL, "OP", operation);
            }
            return str;
        }
        
        #insert #run print_test_call(operation);
        
        errors := 0;
        if result != t_result           { errors += 1; print(" > incorrect result value: got % expected %\n", t_result, result); };
        if type != type_of(t_result)    { errors += 1; print(" > incorrect result type: got % expected %\n", type_of(t_result), type); };
        if saturated != t_saturated     { errors += 1; print(" > incorrect saturated flag: got % expected %\n", t_saturated, saturated); };
        #if operation == "div" {
            if remainder != t_remainder { errors += 1; print(" > incorrect remainder value: got % expected %\n", t_remainder, remainder); };
        }
        return errors;
    }

    errors := 0;

    // Test signed add.
    errors += test_op("add", cast( s8) S8_MAX,  cast( s8)1,          S8_MAX,     s8, true);
    errors += test_op("add", cast(s16)S16_MAX,  cast( u8)1,         S16_MAX,    s16, true);
    errors += test_op("add", cast(s32)S32_MAX,  cast(s32)1,         S32_MAX,    s32, true);
    errors += test_op("add", cast(s64)S64_MAX,  cast(u32)1,         S64_MAX,    s64, true);

    errors += test_op("add", cast( s8) S8_MAX,  cast( s8) S8_MIN,   -1,          s8, false);
    errors += test_op("add", cast(s16)S16_MAX,  cast(s16)S16_MIN,   -1,         s16, false);
    errors += test_op("add", cast(s32)S32_MAX,  cast(s32)S32_MIN,   -1,         s32, false);
    errors += test_op("add", cast(s64)S64_MAX,  cast(s64)S64_MIN,   -1,         s64, false);

    // Test unsigned add.
    errors += test_op("add", cast( u8) U8_MAX,  cast( u8)1,          U8_MAX,     u8, true);
    errors += test_op("add", cast(u16)U16_MAX,  cast(u16)1,         U16_MAX,    u16, true);
    errors += test_op("add", cast(u32)U32_MAX,  cast(u32)1,         U32_MAX,    u32, true);
    errors += test_op("add", cast(u64)U64_MAX,  cast(u64)1,         U64_MAX,    u64, true);

    errors += test_op("add", cast( u8) U8_MAX,  cast( u8)0,          U8_MAX,     u8, false);
    errors += test_op("add", cast(u16)U16_MAX,  cast(u16)0,         U16_MAX,    u16, false);
    errors += test_op("add", cast(u32)U32_MAX,  cast(u32)0,         U32_MAX,    u32, false);
    errors += test_op("add", cast(u64)U64_MAX,  cast(u64)0,         U64_MAX,    u64, false);

    // Test signed sub.
    errors += test_op("sub", cast( s8) S8_MIN,  cast( s8)1,          S8_MIN,     s8, true);
    errors += test_op("sub", cast(s16)S16_MIN,  cast( u8)1,         S16_MIN,    s16, true);
    errors += test_op("sub", cast(s32)S32_MIN,  cast(s32)1,         S32_MIN,    s32, true);
    errors += test_op("sub", cast(s64)S64_MIN,  cast(u32)1,         S64_MIN,    s64, true);

    errors += test_op("sub", cast( s8)-1,       cast( s8) S8_MAX,    S8_MIN,     s8, false);
    errors += test_op("sub", cast(s16)-1,       cast(s16)S16_MAX,   S16_MIN,    s16, false);
    errors += test_op("sub", cast(s32)-1,       cast(s32)S32_MAX,   S32_MIN,    s32, false);
    errors += test_op("sub", cast(s64)-1,       cast(s64)S64_MAX,   S64_MIN,    s64, false);

    // Test unsigned sub.
    errors += test_op("sub", cast( u8)1,        cast( u8) U8_MAX,   0,           u8, true);
    errors += test_op("sub", cast( u8)1,        cast(u16)U16_MAX,   0,          u16, true);
    errors += test_op("sub", cast(u32)1,        cast(u32)U32_MAX,   0,          u32, true);
    errors += test_op("sub", cast(u32)1,        cast(u64)U64_MAX,   0,          u64, true);

    errors += test_op("sub", cast( u8) U8_MAX,  cast( u8)0,          U8_MAX,     u8, false);
    errors += test_op("sub", cast(u16)U16_MAX,  cast( u8)0,         U16_MAX,    u16, false);
    errors += test_op("sub", cast(u32)U32_MAX,  cast(u32)0,         U32_MAX,    u32, false);
    errors += test_op("sub", cast(u64)U64_MAX,  cast(u32)0,         U64_MAX,    u64, false);

    // Test signed mul.
    errors += test_op("mul", cast( s8) S8_MIN,  cast( s8)-1,         S8_MAX,     s8, true);
    errors += test_op("mul", cast(s16)S16_MIN,  cast( s8)-1,        S16_MAX,    s16, true);
    errors += test_op("mul", cast(s32)S32_MIN,  cast(s32)-1,        S32_MAX,    s32, true);
    errors += test_op("mul", cast(s64)S64_MIN,  cast(s32)-1,        S64_MAX,    s64, true);

    errors += test_op("mul", cast( s8) S8_MAX,  cast( s8)-2,         S8_MIN,     s8, true);
    errors += test_op("mul", cast(s16)S16_MAX,  cast( s8)-2,        S16_MIN,    s16, true);
    errors += test_op("mul", cast(s32)S32_MAX,  cast(s32)-2,        S32_MIN,    s32, true);
    errors += test_op("mul", cast(s64)S64_MAX,  cast(s32)-2,        S64_MIN,    s64, true);

    errors += test_op("mul", cast( s8)-2,       cast( s8) S8_MAX,   S8_MIN,     s8, true);
    errors += test_op("mul", cast( s8)-2,       cast(s16)S16_MAX,   S16_MIN,    s16, true);
    errors += test_op("mul", cast(s32)-2,       cast(s32)S32_MAX,   S32_MIN,    s32, true);
    errors += test_op("mul", cast(s32)-2,       cast(s64)S64_MAX,   S64_MIN,    s64, true);

    errors += test_op("mul", cast( s8) S8_MAX,  cast( s8)2,          S8_MAX,     s8, true);
    errors += test_op("mul", cast(s16)S16_MAX,  cast( s8)2,         S16_MAX,    s16, true);
    errors += test_op("mul", cast(s32)S32_MAX,  cast(s32)2,         S32_MAX,    s32, true);
    errors += test_op("mul", cast(s64)S64_MAX,  cast(s32)2,         S64_MAX,    s64, true);

    errors += test_op("mul", cast( s8) S8_MAX,  cast( s8)-1,         -S8_MAX,    s8, false);
    errors += test_op("mul", cast(s16)S16_MAX,  cast( s8)-1,        -S16_MAX,   s16, false);
    errors += test_op("mul", cast(s32)S32_MAX,  cast(s32)-1,        -S32_MAX,   s32, false);
    errors += test_op("mul", cast(s64)S64_MAX,  cast(s32)-1,        -S64_MAX,   s64, false);

    errors += test_op("mul", cast( s8) S8_MAX,  cast( s8)0,         0,           s8, false);
    errors += test_op("mul", cast(s16)S16_MAX,  cast( u8)0,         0,          s16, false);
    errors += test_op("mul", cast(s32)S32_MAX,  cast(s32)0,         0,          s32, false);
    errors += test_op("mul", cast(s64)S64_MAX,  cast(u32)0,         0,          s64, false);

    // Test unsigned mul.
    errors += test_op("mul", cast( u8) U8_MAX,  cast( u8)1,          U8_MAX,     u8, false);
    errors += test_op("mul", cast(u16)U16_MAX,  cast( u8)1,         U16_MAX,    u16, false);
    errors += test_op("mul", cast(u32)U32_MAX,  cast(u32)1,         U32_MAX,    u32, false);
    errors += test_op("mul", cast(u64)U64_MAX,  cast(u32)1,         U64_MAX,    u64, false);

    errors += test_op("mul", cast( u8) U8_MAX,  cast( u8)2,          U8_MAX,     u8, true);
    errors += test_op("mul", cast(u16)U16_MAX,  cast( u8)2,         U16_MAX,    u16, true);
    errors += test_op("mul", cast(u32)U32_MAX,  cast(u32)2,         U32_MAX,    u32, true);
    errors += test_op("mul", cast(u64)U64_MAX,  cast(u32)2,         U64_MAX,    u64, true);

    // Test signed div.
    errors += test_op("div", cast( s8) S8_MIN,  cast( s8)-1,         S8_MAX,     s8, true,  -1);
    errors += test_op("div", cast(s16)S16_MIN,  cast( s8)-1,        S16_MAX,    s16, true,  -1);
    errors += test_op("div", cast(s32)S32_MIN,  cast(s32)-1,        S32_MAX,    s32, true,  -1);
    errors += test_op("div", cast(s64)S64_MIN,  cast(s32)-1,        S64_MAX,    s64, true,  -1);

    errors += test_op("div", cast( s8) S8_MAX,  cast( s8)-2,        - S8_MAX/2,  s8, false, 1);
    errors += test_op("div", cast(s16)S16_MAX,  cast( s8)-2,        -S16_MAX/2, s16, false, 1);
    errors += test_op("div", cast(s32)S32_MAX,  cast(s32)-2,        -S32_MAX/2, s32, false, 1);
    errors += test_op("div", cast(s64)S64_MAX,  cast(s32)-2,        -S64_MAX/2, s64, false, 1);

    errors += test_op("div", cast( s8)15,       cast( s8)5,         3,          s8, false,  0);
    errors += test_op("div", cast( u8)15,       cast(s16)7,         2,          s16, false, 1);
    errors += test_op("div", cast(s16)15,       cast(s32)13,        1,          s32, false, 2);
    errors += test_op("div", cast(u16)100,      cast(s64)3,         33,         s64, false, 1);

    // Test unsigned div.
    errors += test_op("div", cast( u8) U8_MAX,  cast( u8)2,          U8_MAX/2,   u8, false, 1);
    errors += test_op("div", cast(u16)U16_MAX,  cast( u8)2,         U16_MAX/2,  u16, false, 1);
    errors += test_op("div", cast(u32)U32_MAX,  cast(u32)2,         U32_MAX/2,  u32, false, 1);
    errors += test_op("div", cast(u64)U64_MAX,  cast(u32)2,         U64_MAX/2,  u64, false, 1);

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

    write_strings(
        "#========================#\n",
        "# Benchmarks             #\n"
        );
    
    #import "Random";
    performance_test :: ($operation: string, $type: Type, print_result: bool = true) -> ops_per_us_gen: float, ops_per_us_asm: float {
        
        NUM_TESTS :: 500;
        SIZE_DATA :: 12000; // Keep it below cache size.
        best_gen := 0.0;
        best_asm := 0.0;

        numbers_x: [..] type;
        numbers_y: [..] type;
        numbers_zgen: [SIZE_DATA] type;
        numbers_zasm: [SIZE_DATA] type;
        array_reserve(*numbers_x, SIZE_DATA);
        array_reserve(*numbers_y, SIZE_DATA);

        #if type == u8  { MIN :: 0; MAX :: U8_MAX; }
        #if type == u16 { MIN :: 0; MAX :: U16_MAX; }
        #if type == u32 { MIN :: 0; MAX :: U32_MAX; }
        #if type == u64 { MIN :: 0; MAX :: U64_MAX; }
        #if type == s8  { MIN :: S8_MIN; MAX :: S8_MAX; }
        #if type == s16 { MIN :: S16_MIN; MAX :: S16_MAX; }
        #if type == s32 { MIN :: S32_MIN; MAX :: S32_MAX; }
        #if type == s64 { MIN :: S64_MIN; MAX :: S64_MAX; }

        for 0..SIZE_DATA-1 {
            x := cast(type) random_get_within_range(xx MIN, xx MAX);
            y := cast(type) random_get_within_range(xx MIN, xx MAX);
            if y == 0 && operation == "div" {
                y = 1;
            }
            array_add(*numbers_x, x);
            array_add(*numbers_y, y);
        }
        
        for 0..NUM_TESTS-1 {
        
            time_gen := current_time_monotonic();
            for 0..SIZE_DATA-1 #insert #run replace("numbers_zgen[it] = OP(numbers_x[it], numbers_y[it], true);", "OP", operation);
            time_gen = current_time_monotonic() - time_gen;

            time_asm := current_time_monotonic();
            for 0..SIZE_DATA-1 #insert #run replace("numbers_zasm[it] = OP(numbers_x[it], numbers_y[it]);", "OP", operation);
            time_asm = current_time_monotonic() - time_asm;

            for 0..SIZE_DATA-1 assert(numbers_zgen[it] == numbers_zasm[it]);

            perf_gen    := cast(float)SIZE_DATA/cast(float)to_microseconds(time_gen);
            perf_asm    := cast(float)SIZE_DATA/cast(float)to_microseconds(time_asm);
            best_gen    = max(best_gen, perf_gen);
            best_asm    = max(best_asm, perf_asm);
        }

        if print_result {
            if type == s8 || type == u8 write_string(" ");
            print("% | % | % |\n", type, best_gen, best_asm);
        }
        return best_gen, best_asm;
    }

    ff := context.print_style.default_format_float;
    ff.zero_removal = .NO;
    ff.width = 7;
    ff.trailing_width = 1;
    context.print_style.default_format_float = ff;

    write_strings(
        "    #    (ops / usec)   #\n",
        "    | generic | x64 asm |\n"
    );

    write_strings(
        "--- | ----------------- |\n",
        "    |        add        |\n"
        );
    performance_test("add", u8);
    performance_test("add", u16);
    performance_test("add", u32);
    performance_test("add", u64);
    performance_test("add", s8);
    performance_test("add", s16);
    performance_test("add", s32);
    performance_test("add", s64);

    write_strings(
        "--- | ----------------- |\n",
        "    |        sub        |\n"
        );
    performance_test("sub", u8);
    performance_test("sub", u16);
    performance_test("sub", u32);
    performance_test("sub", u64);
    performance_test("sub", s8);
    performance_test("sub", s16);
    performance_test("sub", s32);
    performance_test("sub", s64);

    write_strings(
        "--- | ----------------- |\n",
        "    |        mul        |\n"
        );
    performance_test("mul", u8);
    performance_test("mul", u16);
    performance_test("mul", u32);
    performance_test("mul", u64);
    performance_test("mul", s8);
    performance_test("mul", s16);
    performance_test("mul", s32);
    performance_test("mul", s64);

    write_strings(
        "--- | ----------------- |\n",
        "    |        div        |\n"
        );
    performance_test("div", u8);
    performance_test("div", u16);
    performance_test("div", u32);
    performance_test("div", u64);
    performance_test("div", s8);
    performance_test("div", s16);
    performance_test("div", s32);
    performance_test("div", s64);
}