diff --git a/dlls/vcomp/main.c b/dlls/vcomp/main.c
index d4c936612ab..34e327eedb7 100644
--- a/dlls/vcomp/main.c
+++ b/dlls/vcomp/main.c
@@ -894,6 +894,46 @@ void CDECL _vcomp_atomic_sub_r8(double *dest, double val)
     while (interlocked_cmpxchg64((LONG64 *)dest, new, old) != old);
 }
 
+static void CDECL _vcomp_atomic_bool_and_r8(double *dest, double val)
+{
+    LONG64 old, new;
+    do
+    {
+        old = *(LONG64 *)dest;
+        *(double *)&new = (*(double *)&old != 0.0) ? (val != 0.0) : 0.0;
+    }
+    while (interlocked_cmpxchg64((LONG64 *)dest, new, old) != old);
+}
+
+static void CDECL _vcomp_atomic_bool_or_r8(double *dest, double val)
+{
+    LONG64 old, new;
+    do
+    {
+        old = *(LONG64 *)dest;
+        *(double *)&new = (*(double *)&old != 0.0) ? *(double *)&old : (val != 0.0);
+    }
+    while (interlocked_cmpxchg64((LONG64 *)dest, new, old) != old);
+}
+
+void CDECL _vcomp_reduction_r8(unsigned int flags, double *dest, double val)
+{
+    static void (CDECL * const funcs[])(double *, double) =
+    {
+        _vcomp_atomic_add_r8,
+        _vcomp_atomic_add_r8,
+        _vcomp_atomic_mul_r8,
+        _vcomp_atomic_bool_or_r8,
+        _vcomp_atomic_bool_or_r8,
+        _vcomp_atomic_bool_or_r8,
+        _vcomp_atomic_bool_and_r8,
+        _vcomp_atomic_bool_or_r8,
+    };
+    unsigned int op = (flags >> 8) & 0xf;
+    op = min(op, sizeof(funcs)/sizeof(funcs[0]) - 1);
+    funcs[op](dest, val);
+}
+
 int CDECL omp_get_dynamic(void)
 {
     TRACE("stub\n");
diff --git a/dlls/vcomp/tests/vcomp.c b/dlls/vcomp/tests/vcomp.c
index 15fc473a704..96a3d9c20a8 100644
--- a/dlls/vcomp/tests/vcomp.c
+++ b/dlls/vcomp/tests/vcomp.c
@@ -108,6 +108,7 @@ static void  (CDECL   *p_vcomp_reduction_i2)(unsigned int flags, short *dest, sh
 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_r4)(unsigned int flags, float *dest, float val);
+static void  (CDECL   *p_vcomp_reduction_r8)(unsigned int flags, double *dest, double 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);
@@ -358,6 +359,7 @@ static BOOL init_vcomp(void)
     VCOMP_GET_PROC(_vcomp_reduction_i4);
     VCOMP_GET_PROC(_vcomp_reduction_i8);
     VCOMP_GET_PROC(_vcomp_reduction_r4);
+    VCOMP_GET_PROC(_vcomp_reduction_r8);
     VCOMP_GET_PROC(_vcomp_reduction_u1);
     VCOMP_GET_PROC(_vcomp_reduction_u2);
     VCOMP_GET_PROC(_vcomp_reduction_u4);
@@ -2110,7 +2112,7 @@ static void test_reduction_integer64(void)
     }
 }
 
-static void test_reduction_float(void)
+static void test_reduction_float_double(void)
 {
     static const struct
     {
@@ -2161,6 +2163,13 @@ static void test_reduction_float(void)
         ok(tests[i].expected - 0.001 < val && val < tests[i].expected + 0.001,
            "test %d: expected val == %f, got %f\n", i, tests[i].expected, val);
     }
+    for (i = 0; i < sizeof(tests)/sizeof(tests[0]); i++)
+    {
+        double val = tests[i].v1;
+        p_vcomp_reduction_r8(tests[i].flags, &val, tests[i].v2);
+        ok(tests[i].expected - 0.001 < val && val < tests[i].expected + 0.001,
+           "test %d: expected val == %f, got %f\n", i, tests[i].expected, val);
+    }
 }
 
 START_TEST(vcomp)
@@ -2191,7 +2200,7 @@ START_TEST(vcomp)
     test_reduction_integer16();
     test_reduction_integer32();
     test_reduction_integer64();
-    test_reduction_float();
+    test_reduction_float_double();
 
     release_vcomp();
 }
diff --git a/dlls/vcomp/vcomp.spec b/dlls/vcomp/vcomp.spec
index 22f06cb50a1..f2231c0de24 100644
--- a/dlls/vcomp/vcomp.spec
+++ b/dlls/vcomp/vcomp.spec
@@ -78,7 +78,7 @@
 @ cdecl _vcomp_reduction_i4(long ptr long)
 @ cdecl _vcomp_reduction_i8(long ptr int64)
 @ cdecl _vcomp_reduction_r4(long ptr float)
-@ stub _vcomp_reduction_r8
+@ cdecl _vcomp_reduction_r8(long ptr double)
 @ 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
diff --git a/dlls/vcomp100/vcomp100.spec b/dlls/vcomp100/vcomp100.spec
index b87bd8b29fd..afea2f2af65 100644
--- a/dlls/vcomp100/vcomp100.spec
+++ b/dlls/vcomp100/vcomp100.spec
@@ -78,7 +78,7 @@
 @ cdecl _vcomp_reduction_i4(long ptr long) vcomp._vcomp_reduction_i4
 @ cdecl _vcomp_reduction_i8(long ptr int64) vcomp._vcomp_reduction_i8
 @ cdecl _vcomp_reduction_r4(long ptr float) vcomp._vcomp_reduction_r4
-@ stub _vcomp_reduction_r8
+@ cdecl _vcomp_reduction_r8(long ptr double) vcomp._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
diff --git a/dlls/vcomp110/vcomp110.spec b/dlls/vcomp110/vcomp110.spec
index e86d50b8ef4..cec1c99a7f2 100644
--- a/dlls/vcomp110/vcomp110.spec
+++ b/dlls/vcomp110/vcomp110.spec
@@ -79,7 +79,7 @@
 @ cdecl _vcomp_reduction_i4(long ptr long) vcomp._vcomp_reduction_i4
 @ cdecl _vcomp_reduction_i8(long ptr int64) vcomp._vcomp_reduction_i8
 @ cdecl _vcomp_reduction_r4(long ptr float) vcomp._vcomp_reduction_r4
-@ stub _vcomp_reduction_r8
+@ cdecl _vcomp_reduction_r8(long ptr double) vcomp._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
diff --git a/dlls/vcomp120/vcomp120.spec b/dlls/vcomp120/vcomp120.spec
index e86d50b8ef4..cec1c99a7f2 100644
--- a/dlls/vcomp120/vcomp120.spec
+++ b/dlls/vcomp120/vcomp120.spec
@@ -79,7 +79,7 @@
 @ cdecl _vcomp_reduction_i4(long ptr long) vcomp._vcomp_reduction_i4
 @ cdecl _vcomp_reduction_i8(long ptr int64) vcomp._vcomp_reduction_i8
 @ cdecl _vcomp_reduction_r4(long ptr float) vcomp._vcomp_reduction_r4
-@ stub _vcomp_reduction_r8
+@ cdecl _vcomp_reduction_r8(long ptr double) vcomp._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
diff --git a/dlls/vcomp140/vcomp140.spec b/dlls/vcomp140/vcomp140.spec
index e86d50b8ef4..cec1c99a7f2 100644
--- a/dlls/vcomp140/vcomp140.spec
+++ b/dlls/vcomp140/vcomp140.spec
@@ -79,7 +79,7 @@
 @ cdecl _vcomp_reduction_i4(long ptr long) vcomp._vcomp_reduction_i4
 @ cdecl _vcomp_reduction_i8(long ptr int64) vcomp._vcomp_reduction_i8
 @ cdecl _vcomp_reduction_r4(long ptr float) vcomp._vcomp_reduction_r4
-@ stub _vcomp_reduction_r8
+@ cdecl _vcomp_reduction_r8(long ptr double) vcomp._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
diff --git a/dlls/vcomp90/vcomp90.spec b/dlls/vcomp90/vcomp90.spec
index b87bd8b29fd..afea2f2af65 100644
--- a/dlls/vcomp90/vcomp90.spec
+++ b/dlls/vcomp90/vcomp90.spec
@@ -78,7 +78,7 @@
 @ cdecl _vcomp_reduction_i4(long ptr long) vcomp._vcomp_reduction_i4
 @ cdecl _vcomp_reduction_i8(long ptr int64) vcomp._vcomp_reduction_i8
 @ cdecl _vcomp_reduction_r4(long ptr float) vcomp._vcomp_reduction_r4
-@ stub _vcomp_reduction_r8
+@ cdecl _vcomp_reduction_r8(long ptr double) vcomp._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