vcomp: Implement _vcomp_reduction_{u,i}8 and add tests.

dlls/vcomp/main.c

@@ -736,6 +736,36 @@ void CDECL _vcomp_atomic_xor_i8(LONG64 *dest, LONG64 val)
     do old = *dest; while (interlocked_cmpxchg64(dest, old ^ val, old) != old);
 }
 
+static void CDECL _vcomp_atomic_bool_and_i8(LONG64 *dest, LONG64 val)
+{
+    LONG64 old;
+    do old = *dest; while (interlocked_cmpxchg64(dest, old && val, old) != old);
+}
+
+static void CDECL _vcomp_atomic_bool_or_i8(LONG64 *dest, LONG64 val)
+{
+    LONG64 old;
+    do old = *dest; while (interlocked_cmpxchg64(dest, old ? old : (val != 0), old) != old);
+}
+
+void CDECL _vcomp_reduction_i8(unsigned int flags, LONG64 *dest, LONG64 val)
+{
+    static void (CDECL * const funcs[])(LONG64 *, LONG64) =
+    {
+        _vcomp_atomic_add_i8,
+        _vcomp_atomic_add_i8,
+        _vcomp_atomic_mul_i8,
+        _vcomp_atomic_and_i8,
+        _vcomp_atomic_or_i8,
+        _vcomp_atomic_xor_i8,
+        _vcomp_atomic_bool_and_i8,
+        _vcomp_atomic_bool_or_i8,
+    };
+    unsigned int op = (flags >> 8) & 0xf;
+    op = min(op, sizeof(funcs)/sizeof(funcs[0]) - 1);
+    funcs[op](dest, val);
+}
+
 void CDECL _vcomp_atomic_add_r4(float *dest, float val)
 {
     int old, new;

dlls/vcomp/tests/vcomp.c

@@ -106,9 +106,11 @@ static void  (CDECL   *p_vcomp_master_end)(void);
 static void  (CDECL   *p_vcomp_reduction_i1)(unsigned int flags, char *dest, char val);
 static void  (CDECL   *p_vcomp_reduction_i2)(unsigned int flags, short *dest, short val);
 static void  (CDECL   *p_vcomp_reduction_i4)(unsigned int flags, int *dest, int val);
+static void  (CDECL   *p_vcomp_reduction_i8)(unsigned int flags, LONG64 *dest, LONG64 val);
 static void  (CDECL   *p_vcomp_reduction_u1)(unsigned int flags, unsigned char *dest, unsigned char val);
 static void  (CDECL   *p_vcomp_reduction_u2)(unsigned int flags, unsigned short *dest, unsigned short val);
 static void  (CDECL   *p_vcomp_reduction_u4)(unsigned int flags, unsigned int *dest, unsigned int val);
+static void  (CDECL   *p_vcomp_reduction_u8)(unsigned int flags, ULONG64 *dest, ULONG64 val);
 static void  (CDECL   *p_vcomp_sections_init)(int n);
 static int   (CDECL   *p_vcomp_sections_next)(void);
 static void  (CDECL   *p_vcomp_set_num_threads)(int num_threads);
@@ -353,9 +355,11 @@ static BOOL init_vcomp(void)
     VCOMP_GET_PROC(_vcomp_reduction_i1);
     VCOMP_GET_PROC(_vcomp_reduction_i2);
     VCOMP_GET_PROC(_vcomp_reduction_i4);
+    VCOMP_GET_PROC(_vcomp_reduction_i8);
     VCOMP_GET_PROC(_vcomp_reduction_u1);
     VCOMP_GET_PROC(_vcomp_reduction_u2);
     VCOMP_GET_PROC(_vcomp_reduction_u4);
+    VCOMP_GET_PROC(_vcomp_reduction_u8);
     VCOMP_GET_PROC(_vcomp_sections_init);
     VCOMP_GET_PROC(_vcomp_sections_next);
     VCOMP_GET_PROC(_vcomp_set_num_threads);
@@ -1893,7 +1897,7 @@ static void test_atomic_double(void)
 
 static void test_reduction_integer8(void)
 {
-    struct
+    static const struct
     {
         unsigned int flags;
         char v1, v2, expected;
@@ -1937,7 +1941,7 @@ static void test_reduction_integer8(void)
 
 static void test_reduction_integer16(void)
 {
-    struct
+    static const struct
     {
         unsigned int flags;
         short v1, v2, expected;
@@ -1987,7 +1991,7 @@ static void CDECL reduction_cb(int *a, int *b)
 
 static void test_reduction_integer32(void)
 {
-    struct
+    static const struct
     {
         unsigned int flags;
         int v1, v2, expected;
@@ -2061,6 +2065,49 @@ static void test_reduction_integer32(void)
     }
 }
 
+static void test_reduction_integer64(void)
+{
+    static const struct
+    {
+        unsigned int flags;
+        LONG64 v1, v2, expected;
+    }
+    tests[] =
+    {
+        { 0x000,                            0x1122334455667788,  0x7766554433221100, -0x7777777777777778 },
+        { VCOMP_REDUCTION_FLAGS_ADD,        0x1122334455667788,  0x7766554433221100, -0x7777777777777778 },
+        { VCOMP_REDUCTION_FLAGS_MUL,        0x1122334455667788,  0x7766554433221100,  0x3e963337c6000800 },
+        { VCOMP_REDUCTION_FLAGS_MUL,        0x1122334455667788, -0x7766554433221100,  0xc169ccc839fff800 },
+        { VCOMP_REDUCTION_FLAGS_AND,        0x1122334455667788,  0x7766554433221100,  0x1122114411221100 },
+        { VCOMP_REDUCTION_FLAGS_OR,         0x1122334455667788,  0x7766554433221100,  0x7766774477667788 },
+        { VCOMP_REDUCTION_FLAGS_XOR,        0x1122334455667788,  0x7766554433221100,  0x6644660066446688 },
+        { VCOMP_REDUCTION_FLAGS_BOOL_AND,                    1,                   2,                   1 },
+        { VCOMP_REDUCTION_FLAGS_BOOL_OR,                     0,                   2,                   1 },
+        { 0x800,                                             0,                   2,                   1 },
+        { 0x900,                                             0,                   2,                   1 },
+        { 0xa00,                                             0,                   2,                   1 },
+        { 0xb00,                                             0,                   2,                   1 },
+        { 0xc00,                                             0,                   2,                   1 },
+        { 0xd00,                                             0,                   2,                   1 },
+        { 0xe00,                                             0,                   2,                   1 },
+        { 0xf00,                                             0,                   2,                   1 },
+    };
+    int i;
+
+    for (i = 0; i < sizeof(tests)/sizeof(tests[0]); i++)
+    {
+        LONG64 val = tests[i].v1;
+        p_vcomp_reduction_i8(tests[i].flags, &val, tests[i].v2);
+        ok(val == tests[i].expected, "test %d: unexpectedly got %s\n", i, debugstr_longlong(val));
+    }
+    for (i = 0; i < sizeof(tests)/sizeof(tests[0]); i++)
+    {
+        ULONG64 val = tests[i].v1;
+        p_vcomp_reduction_u8(tests[i].flags, &val, tests[i].v2);
+        ok(val == tests[i].expected, "test %d: unexpectedly got %s\n", i, debugstr_longlong(val));
+    }
+}
+
 START_TEST(vcomp)
 {
     if (!init_vcomp())
@@ -2088,6 +2135,7 @@ START_TEST(vcomp)
     test_reduction_integer8();
     test_reduction_integer16();
     test_reduction_integer32();
+    test_reduction_integer64();
 
     release_vcomp();
 }

dlls/vcomp/vcomp.spec

@@ -76,13 +76,13 @@
 @ cdecl _vcomp_reduction_i1(long ptr long)
 @ cdecl _vcomp_reduction_i2(long ptr long)
 @ cdecl _vcomp_reduction_i4(long ptr long)
-@ stub _vcomp_reduction_i8
+@ cdecl _vcomp_reduction_i8(long ptr int64)
 @ stub _vcomp_reduction_r4
 @ stub _vcomp_reduction_r8
 @ cdecl _vcomp_reduction_u1(long ptr long) _vcomp_reduction_i1
 @ cdecl _vcomp_reduction_u2(long ptr long) _vcomp_reduction_i2
 @ cdecl _vcomp_reduction_u4(long ptr long) _vcomp_reduction_i4
-@ stub _vcomp_reduction_u8
+@ cdecl _vcomp_reduction_u8(long ptr int64) _vcomp_reduction_i8
 @ cdecl _vcomp_sections_init(long)
 @ cdecl _vcomp_sections_next()
 @ cdecl _vcomp_set_num_threads(long)

dlls/vcomp100/vcomp100.spec

@@ -76,13 +76,13 @@
 @ cdecl _vcomp_reduction_i1(long ptr long) vcomp._vcomp_reduction_i1
 @ cdecl _vcomp_reduction_i2(long ptr long) vcomp._vcomp_reduction_i2
 @ cdecl _vcomp_reduction_i4(long ptr long) vcomp._vcomp_reduction_i4
-@ stub _vcomp_reduction_i8
+@ cdecl _vcomp_reduction_i8(long ptr int64) vcomp._vcomp_reduction_i8
 @ stub _vcomp_reduction_r4
 @ stub _vcomp_reduction_r8
 @ cdecl _vcomp_reduction_u1(long ptr long) vcomp._vcomp_reduction_u1
 @ cdecl _vcomp_reduction_u2(long ptr long) vcomp._vcomp_reduction_u2
 @ cdecl _vcomp_reduction_u4(long ptr long) vcomp._vcomp_reduction_u4
-@ stub _vcomp_reduction_u8
+@ cdecl _vcomp_reduction_u8(long ptr int64) vcomp._vcomp_reduction_u8
 @ cdecl _vcomp_sections_init(long) vcomp._vcomp_sections_init
 @ cdecl _vcomp_sections_next() vcomp._vcomp_sections_next
 @ cdecl _vcomp_set_num_threads(long) vcomp._vcomp_set_num_threads

dlls/vcomp110/vcomp110.spec

@@ -77,13 +77,13 @@
 @ cdecl _vcomp_reduction_i1(long ptr long) vcomp._vcomp_reduction_i1
 @ cdecl _vcomp_reduction_i2(long ptr long) vcomp._vcomp_reduction_i2
 @ cdecl _vcomp_reduction_i4(long ptr long) vcomp._vcomp_reduction_i4
-@ stub _vcomp_reduction_i8
+@ cdecl _vcomp_reduction_i8(long ptr int64) vcomp._vcomp_reduction_i8
 @ stub _vcomp_reduction_r4
 @ stub _vcomp_reduction_r8
 @ cdecl _vcomp_reduction_u1(long ptr long) vcomp._vcomp_reduction_u1
 @ cdecl _vcomp_reduction_u2(long ptr long) vcomp._vcomp_reduction_u2
 @ cdecl _vcomp_reduction_u4(long ptr long) vcomp._vcomp_reduction_u4
-@ stub _vcomp_reduction_u8
+@ cdecl _vcomp_reduction_u8(long ptr int64) vcomp._vcomp_reduction_u8
 @ cdecl _vcomp_sections_init(long) vcomp._vcomp_sections_init
 @ cdecl _vcomp_sections_next() vcomp._vcomp_sections_next
 @ cdecl _vcomp_set_num_threads(long) vcomp._vcomp_set_num_threads

dlls/vcomp120/vcomp120.spec

@@ -77,13 +77,13 @@
 @ cdecl _vcomp_reduction_i1(long ptr long) vcomp._vcomp_reduction_i1
 @ cdecl _vcomp_reduction_i2(long ptr long) vcomp._vcomp_reduction_i2
 @ cdecl _vcomp_reduction_i4(long ptr long) vcomp._vcomp_reduction_i4
-@ stub _vcomp_reduction_i8
+@ cdecl _vcomp_reduction_i8(long ptr int64) vcomp._vcomp_reduction_i8
 @ stub _vcomp_reduction_r4
 @ stub _vcomp_reduction_r8
 @ cdecl _vcomp_reduction_u1(long ptr long) vcomp._vcomp_reduction_u1
 @ cdecl _vcomp_reduction_u2(long ptr long) vcomp._vcomp_reduction_u2
 @ cdecl _vcomp_reduction_u4(long ptr long) vcomp._vcomp_reduction_u4
-@ stub _vcomp_reduction_u8
+@ cdecl _vcomp_reduction_u8(long ptr int64) vcomp._vcomp_reduction_u8
 @ cdecl _vcomp_sections_init(long) vcomp._vcomp_sections_init
 @ cdecl _vcomp_sections_next() vcomp._vcomp_sections_next
 @ cdecl _vcomp_set_num_threads(long) vcomp._vcomp_set_num_threads

dlls/vcomp140/vcomp140.spec

@@ -77,13 +77,13 @@
 @ cdecl _vcomp_reduction_i1(long ptr long) vcomp._vcomp_reduction_i1
 @ cdecl _vcomp_reduction_i2(long ptr long) vcomp._vcomp_reduction_i2
 @ cdecl _vcomp_reduction_i4(long ptr long) vcomp._vcomp_reduction_i4
-@ stub _vcomp_reduction_i8
+@ cdecl _vcomp_reduction_i8(long ptr int64) vcomp._vcomp_reduction_i8
 @ stub _vcomp_reduction_r4
 @ stub _vcomp_reduction_r8
 @ cdecl _vcomp_reduction_u1(long ptr long) vcomp._vcomp_reduction_u1
 @ cdecl _vcomp_reduction_u2(long ptr long) vcomp._vcomp_reduction_u2
 @ cdecl _vcomp_reduction_u4(long ptr long) vcomp._vcomp_reduction_u4
-@ stub _vcomp_reduction_u8
+@ cdecl _vcomp_reduction_u8(long ptr int64) vcomp._vcomp_reduction_u8
 @ cdecl _vcomp_sections_init(long) vcomp._vcomp_sections_init
 @ cdecl _vcomp_sections_next() vcomp._vcomp_sections_next
 @ cdecl _vcomp_set_num_threads(long) vcomp._vcomp_set_num_threads

dlls/vcomp90/vcomp90.spec

@@ -76,13 +76,13 @@
 @ cdecl _vcomp_reduction_i1(long ptr long) vcomp._vcomp_reduction_i1
 @ cdecl _vcomp_reduction_i2(long ptr long) vcomp._vcomp_reduction_i2
 @ cdecl _vcomp_reduction_i4(long ptr long) vcomp._vcomp_reduction_i4
-@ stub _vcomp_reduction_i8
+@ cdecl _vcomp_reduction_i8(long ptr int64) vcomp._vcomp_reduction_i8
 @ stub _vcomp_reduction_r4
 @ stub _vcomp_reduction_r8
 @ cdecl _vcomp_reduction_u1(long ptr long) vcomp._vcomp_reduction_u1
 @ cdecl _vcomp_reduction_u2(long ptr long) vcomp._vcomp_reduction_u2
 @ cdecl _vcomp_reduction_u4(long ptr long) vcomp._vcomp_reduction_u4
-@ stub _vcomp_reduction_u8
+@ cdecl _vcomp_reduction_u8(long ptr int64) vcomp._vcomp_reduction_u8
 @ cdecl _vcomp_sections_init(long) vcomp._vcomp_sections_init
 @ cdecl _vcomp_sections_next() vcomp._vcomp_sections_next
 @ cdecl _vcomp_set_num_threads(long) vcomp._vcomp_set_num_threads