diff --git a/src/codegen/codegen_ops_x86-64.h b/src/codegen/codegen_ops_x86-64.h
index a68dee7ae..906b07dc6 100644
--- a/src/codegen/codegen_ops_x86-64.h
+++ b/src/codegen/codegen_ops_x86-64.h
@@ -4363,7 +4363,7 @@ static inline int FP_LOAD_REG_INT_W(int reg)
         addbyte(0xc5);
         addbyte((uint8_t)cpu_state_offset(ST));
 
-        CALL_FUNC((uintptr_t)x87_fround);
+        CALL_FUNC((uintptr_t)x87_fround16_64);
         
         addbyte(0x93); /*XCHG EBX, EAX*/
         
@@ -4393,7 +4393,7 @@ static inline int FP_LOAD_REG_INT(int reg)
         addbyte(0xc5);
         addbyte((uint8_t)cpu_state_offset(ST));
 
-        CALL_FUNC((uintptr_t)x87_fround);
+        CALL_FUNC((uintptr_t)x87_fround32_64);
         
         addbyte(0x93); /*XCHG EBX, EAX*/
         
diff --git a/src/cpu/x87_ops.h b/src/cpu/x87_ops.h
index 17861fa57..9e06a343e 100644
--- a/src/cpu/x87_ops.h
+++ b/src/cpu/x87_ops.h
@@ -152,10 +152,72 @@ static __inline double x87_pop()
         return t;
 }
 
+static __inline int16_t x87_fround16(double b)
+{
+        int16_t a, c;
+
+        switch ((cpu_state.npxc >> 10) & 3)
+        {
+                case 0: /*Nearest*/
+                a = (int16_t)floor(b);
+                c = (int16_t)floor(b + 1.0);
+                if ((b - a) < (c - b))
+                        return a;
+                else if ((b - a) > (c - b))
+                        return c;
+                else
+                        return (a & 1) ? c : a;
+                case 1: /*Down*/
+                return (int16_t)floor(b);
+                case 2: /*Up*/
+                return (int16_t)ceil(b);
+                case 3: /*Chop*/
+                return (int16_t)b;
+        }
+
+        return 0;
+}
+
+static __inline int64_t x87_fround16_64(double b)
+{
+    return (int64_t) x87_fround16(b);
+}
+
+static __inline int32_t x87_fround32(double b)
+{
+        int32_t a, c;
+
+        switch ((cpu_state.npxc >> 10) & 3)
+        {
+                case 0: /*Nearest*/
+                a = (int32_t)floor(b);
+                c = (int32_t)floor(b + 1.0);
+                if ((b - a) < (c - b))
+                        return a;
+                else if ((b - a) > (c - b))
+                        return c;
+                else
+                        return (a & 1) ? c : a;
+                case 1: /*Down*/
+                return (int32_t)floor(b);
+                case 2: /*Up*/
+                return (int32_t)ceil(b);
+                case 3: /*Chop*/
+                return (int32_t)b;
+        }
+
+        return 0;
+}
+
+static __inline int64_t x87_fround32_64(double b)
+{
+    return (int64_t) x87_fround32(b);
+}
+
 static __inline int64_t x87_fround(double b)
 {
         int64_t a, c;
-        
+
         switch ((cpu_state.npxc >> 10) & 3)
         {
                 case 0: /*Nearest*/
diff --git a/src/cpu/x87_ops_loadstore.h b/src/cpu/x87_ops_loadstore.h
index f742196ea..5e754fc90 100644
--- a/src/cpu/x87_ops_loadstore.h
+++ b/src/cpu/x87_ops_loadstore.h
@@ -42,24 +42,20 @@ static int opFILDiw_a32(uint32_t fetchdat)
 
 static int opFISTiw_a16(uint32_t fetchdat)
 {
-        int64_t temp64;
         FP_ENTER();
         fetch_ea_16(fetchdat);
 	SEG_CHECK_WRITE(cpu_state.ea_seg);
-        temp64 = x87_fround(ST(0));
-        seteaw((int16_t)temp64);
+        seteaw(x87_fround16(ST(0)));
         CLOCK_CYCLES((fpu_type >= FPU_487SX) ? (x87_timings.fist_16) : (x87_timings.fist_16 * cpu_multi));
         return cpu_state.abrt;
 }
 #ifndef FPU_8087
 static int opFISTiw_a32(uint32_t fetchdat)
 {
-        int64_t temp64;
         FP_ENTER();
         fetch_ea_32(fetchdat);
 	SEG_CHECK_WRITE(cpu_state.ea_seg);
-        temp64 = x87_fround(ST(0));
-        seteaw((int16_t)temp64);
+        seteaw(x87_fround16(ST(0)));
         CLOCK_CYCLES((fpu_type >= FPU_487SX) ? (x87_timings.fist_16) : (x87_timings.fist_16 * cpu_multi));
         return cpu_state.abrt;
 }
@@ -67,12 +63,10 @@ static int opFISTiw_a32(uint32_t fetchdat)
 
 static int opFISTPiw_a16(uint32_t fetchdat)
 {
-        int64_t temp64;
         FP_ENTER();
         fetch_ea_16(fetchdat);
 	SEG_CHECK_WRITE(cpu_state.ea_seg);
-        temp64 = x87_fround(ST(0));
-        seteaw((int16_t)temp64); if (cpu_state.abrt) return 1;
+        seteaw(x87_fround16(ST(0))); if (cpu_state.abrt) return 1;
         x87_pop();
         CLOCK_CYCLES((fpu_type >= FPU_487SX) ? (x87_timings.fist_16) : (x87_timings.fist_16 * cpu_multi));
         return 0;
@@ -80,12 +74,10 @@ static int opFISTPiw_a16(uint32_t fetchdat)
 #ifndef FPU_8087
 static int opFISTPiw_a32(uint32_t fetchdat)
 {
-        int64_t temp64;
         FP_ENTER();
         fetch_ea_32(fetchdat);
 	SEG_CHECK_WRITE(cpu_state.ea_seg);
-        temp64 = x87_fround(ST(0));
-        seteaw((int16_t)temp64); if (cpu_state.abrt) return 1;
+        seteaw(x87_fround16(ST(0))); if (cpu_state.abrt) return 1;
         x87_pop();
         CLOCK_CYCLES((fpu_type >= FPU_487SX) ? (x87_timings.fist_16) : (x87_timings.fist_16 * cpu_multi));
         return 0;
@@ -240,24 +232,20 @@ static int opFILDil_a32(uint32_t fetchdat)
 
 static int opFISTil_a16(uint32_t fetchdat)
 {
-        int64_t temp64;
         FP_ENTER();
         fetch_ea_16(fetchdat);
 	SEG_CHECK_WRITE(cpu_state.ea_seg);
-        temp64 = x87_fround(ST(0));
-        seteal((int32_t)temp64);
+        seteal(x87_fround32(ST(0)));
         CLOCK_CYCLES((fpu_type >= FPU_487SX) ? (x87_timings.fist_32) : (x87_timings.fist_32 * cpu_multi));
         return cpu_state.abrt;
 }
 #ifndef FPU_8087
 static int opFISTil_a32(uint32_t fetchdat)
 {
-        int64_t temp64;
         FP_ENTER();
         fetch_ea_32(fetchdat);
 	SEG_CHECK_WRITE(cpu_state.ea_seg);
-        temp64 = x87_fround(ST(0));
-        seteal((int32_t)temp64);
+        seteal(x87_fround32(ST(0)));
         CLOCK_CYCLES((fpu_type >= FPU_487SX) ? (x87_timings.fist_32) : (x87_timings.fist_32 * cpu_multi));
         return cpu_state.abrt;
 }
@@ -265,12 +253,10 @@ static int opFISTil_a32(uint32_t fetchdat)
 
 static int opFISTPil_a16(uint32_t fetchdat)
 {
-        int64_t temp64;
         FP_ENTER();
         fetch_ea_16(fetchdat);
 	SEG_CHECK_WRITE(cpu_state.ea_seg);
-        temp64 = x87_fround(ST(0));
-        seteal((int32_t)temp64); if (cpu_state.abrt) return 1;
+        seteal(x87_fround32(ST(0))); if (cpu_state.abrt) return 1;
         x87_pop();
         CLOCK_CYCLES((fpu_type >= FPU_487SX) ? (x87_timings.fist_32) : (x87_timings.fist_32 * cpu_multi));
         return 0;
@@ -278,12 +264,10 @@ static int opFISTPil_a16(uint32_t fetchdat)
 #ifndef FPU_8087
 static int opFISTPil_a32(uint32_t fetchdat)
 {
-        int64_t temp64;
         FP_ENTER();
         fetch_ea_32(fetchdat);
 	SEG_CHECK_WRITE(cpu_state.ea_seg);
-        temp64 = x87_fround(ST(0));
-        seteal((int32_t)temp64); if (cpu_state.abrt) return 1;
+        seteal(x87_fround32(ST(0))); if (cpu_state.abrt) return 1;
         x87_pop();
         CLOCK_CYCLES((fpu_type >= FPU_487SX) ? (x87_timings.fist_32) : (x87_timings.fist_32 * cpu_multi));
         return 0;