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;
diff --git a/src/floppy/fdc.c b/src/floppy/fdc.c
index 4da50795e..3c81dd02e 100644
--- a/src/floppy/fdc.c
+++ b/src/floppy/fdc.c
@@ -1892,9 +1892,10 @@ fdc_is_verify(fdc_t *fdc)
 
 
 int
-fdc_data(fdc_t *fdc, uint8_t data)
+fdc_data(fdc_t *fdc, uint8_t data, int last)
 {
-    int result = 0;
+    int i, result = 0;
+    int n;
 
     if (fdc->deleted & 2) {
 	/* We're in a VERIFY command, so return with 0. */
@@ -1924,27 +1925,39 @@ fdc_data(fdc_t *fdc, uint8_t data)
 		}
 	}
     } else {
-	result = dma_channel_write(fdc->dma_ch, data);
-
 	if (fdc->tc)
 		return -1;
 
-	if (result & DMA_OVER) {
-		fdc->data_ready = 1;
-		fdc->stat = 0xd0;
-		fdc->tc = 1;
-		return -1;
-	}
-
 	if (!fdc->fifo || (fdc->tfifo < 1)) {
+		dma_channel_write(fdc->dma_ch, data);
+
 		fdc->data_ready = 1;
 		fdc->stat = 0xd0;
+
+		if (result & DMA_OVER) {
+			fdc->tc = 1;
+			return -1;
+		}
 	} else {
-		fdc_fifo_buf_advance(fdc);
-		if (fdc->fifobufpos == 0) {
+		/* FIFO enabled */
+		fdc_fifo_buf_write(fdc, data);
+		if (last || (fdc->fifobufpos == 0)) {
 			/* We have wrapped around, means FIFO is over */
 			fdc->data_ready = 1;
 			fdc->stat = 0xd0;
+
+			n = (fdc->fifobufpos > 0) ? (fdc->fifobufpos - 1) : fdc->tfifo;
+			if (fdc->fifobufpos > 0)
+				fdc->fifobufpos = 0;
+
+			for (i = 0; i <= n; i++) {
+				result = dma_channel_write(fdc->dma_ch, fdc->fifobuf[i]);
+
+				if (result & DMA_OVER) {
+					fdc->tc = 1;
+					return -1;
+				}
+			}
 		}
 	}
     }
@@ -2057,10 +2070,10 @@ fdc_writeprotect(fdc_t *fdc)
 
 int fdc_getdata(fdc_t *fdc, int last)
 {
-    int data;
+    int i, data = 0;
 
     if ((fdc->flags & FDC_FLAG_PCJR) || !fdc->dma) {
-	if ((fdc->flags & FDC_FLAG_PCJR) || !fdc->fifo) {
+	if ((fdc->flags & FDC_FLAG_PCJR) || !fdc->fifo || (fdc->tfifo < 1)) {
 		data = fdc->dat;
 
 		if (!last)
@@ -2072,20 +2085,32 @@ int fdc_getdata(fdc_t *fdc, int last)
 			fdc->stat = 0xb0;
 	}
     } else {
-	data = dma_channel_read(fdc->dma_ch);
+	if (!fdc->fifo || (fdc->tfifo < 1)) {
+		data = dma_channel_read(fdc->dma_ch);
+
+		if (data & DMA_OVER)
+			fdc->tc = 1;
 
-	if (!fdc->fifo) {
 		if (!last)
 			fdc->stat = 0x90;
 	} else {
-		fdc_fifo_buf_advance(fdc);
+		if (fdc->fifobufpos == 0) {
+			for (i = 0; i <= fdc->tfifo; i++) {
+				data = dma_channel_read(fdc->dma_ch);
+				fdc->fifobuf[i] = data;
+
+				if (data & DMA_OVER) {
+					fdc->tc = 1;
+					break;
+				}
+			}
+		}
+
+		data = fdc_fifo_buf_read(fdc);
 
 		if (!last && (fdc->fifobufpos == 0))
 			fdc->stat = 0x90;
 	}
-
-	if (data & DMA_OVER)
-		fdc->tc = 1;
     }
 
     return data & 0xff;
diff --git a/src/floppy/fdd_86f.c b/src/floppy/fdd_86f.c
index 5ff02adb3..ebd7d6f7c 100644
--- a/src/floppy/fdd_86f.c
+++ b/src/floppy/fdd_86f.c
@@ -1577,7 +1577,7 @@ d86f_read_sector_data(int drive, int side)
 			} else {
 				if (dev->data_find.bytes_obtained < d86f_get_data_len(drive)) {
 					if (dev->state != STATE_16_VERIFY_DATA) {
-						read_status = fdc_data(d86f_fdc, data);
+						read_status = fdc_data(d86f_fdc, data, dev->data_find.bytes_obtained == ((d86f_get_data_len(drive)) - 1));
 						if (read_status == -1)
 							dev->dma_over++;
 					}
@@ -2139,7 +2139,7 @@ d86f_turbo_read(int drive, int side)
     } else {
 	if (dev->data_find.bytes_obtained < (128UL << dev->last_sector.id.n)) {
 		if (dev->state != STATE_16_VERIFY_DATA) {
-			read_status = fdc_data(d86f_fdc, dat);
+			read_status = fdc_data(d86f_fdc, dat, dev->data_find.bytes_obtained == ((128UL << dev->last_sector.id.n) - 1));
 			if (read_status == -1)
 				dev->dma_over++;
 		}
diff --git a/src/include/86box/fdc.h b/src/include/86box/fdc.h
index 5e2d50d06..87404e174 100644
--- a/src/include/86box/fdc.h
+++ b/src/include/86box/fdc.h
@@ -157,7 +157,7 @@ extern void	fdc_set_base(fdc_t *fdc, int base);
 extern void	fdc_set_irq(fdc_t *fdc, int irq);
 extern void	fdc_set_dma_ch(fdc_t *fdc, int dma_ch);
 extern int	fdc_getdata(fdc_t *fdc, int last);
-extern int	fdc_data(fdc_t *fdc, uint8_t data);
+extern int	fdc_data(fdc_t *fdc, uint8_t data, int last);
 
 extern void	fdc_sectorid(fdc_t *fdc, uint8_t track, uint8_t side,
 			     uint8_t sector, uint8_t size, uint8_t crc1,