diff --git a/src/cpu/386.c b/src/cpu/386.c
index b5a1e61e3..c7e31de22 100644
--- a/src/cpu/386.c
+++ b/src/cpu/386.c
@@ -25,6 +25,7 @@
 #include <86box/fdd.h>
 #include <86box/fdc.h>
 #include <86box/machine.h>
+#include <86box/plat_fallthrough.h>
 #include <86box/gdbstub.h>
 #ifndef OPS_286_386
 #    define OPS_286_386
@@ -262,11 +263,10 @@ exec386_2386(int32_t cycs)
             CHECK_READ_CS(MIN(ol, 4));
             ins_fetch_fault = cpu_386_check_instruction_fault();
 
-            if (!cpu_state.abrt && ins_fetch_fault) {
-                x86gen();
+            /* Breakpoint fault has priority over other faults. */
+            if (ins_fetch_fault) {
                 ins_fetch_fault = 0;
-                /* No instructions executed at this point. */
-                goto block_ended;
+                cpu_state.abrt = 1;
             }
 
             if (!cpu_state.abrt) {
@@ -279,7 +279,8 @@ exec386_2386(int32_t cycs)
                 trap |= !!(cpu_state.flags & T_FLAG);
 
                 cpu_state.pc++;
-                x86_opcodes[(opcode | cpu_state.op32) & 0x3ff](fetchdat);
+                cpu_state.eflags &= ~(RF_FLAG);
+                x86_2386_opcodes[(opcode | cpu_state.op32) & 0x3ff](fetchdat);
                 if (x86_was_reset)
                     break;
             }
@@ -319,8 +320,7 @@ block_ended:
 #endif
                     }
                 }
-                if (!x86_was_reset && ins_fetch_fault)
-                    x86gen();   /* This is supposed to be the first one serviced by the processor according to the manual. */
+according to the manual. */
             } else if (trap) {
                 flags_rebuild();
                 if (trap & 2) dr[6] |= 0x8000;
diff --git a/src/cpu/386_common.c b/src/cpu/386_common.c
index f9ca3408d..d74b181bf 100644
--- a/src/cpu/386_common.c
+++ b/src/cpu/386_common.c
@@ -1412,7 +1412,7 @@ x86_int(int num)
     cpu_state.pc = cpu_state.oldpc;
 
     if (msw & 1)
-        is486 ? pmodeint(num, 0) : pmodeint_2386(num, 0);
+        cpu_use_exec ? pmodeint(num, 0) : pmodeint_2386(num, 0);
     else {
         addr = (num << 2) + idt.base;
 
@@ -1445,7 +1445,7 @@ x86_int(int num)
             oxpc = cpu_state.pc;
 #endif
             cpu_state.pc = readmemw(0, addr);
-            is486 ? loadcs(readmemw(0, addr + 2)) : loadcs_2386(readmemw(0, addr + 2));
+            cpu_use_exec ? loadcs(readmemw(0, addr + 2)) : loadcs_2386(readmemw(0, addr + 2));
         }
     }
 
@@ -1462,7 +1462,7 @@ x86_int_sw(int num)
     cycles -= timing_int;
 
     if (msw & 1)
-        is486 ? pmodeint(num, 1) : pmodeint_2386(num, 1);
+        cpu_use_exec ? pmodeint(num, 1) : pmodeint_2386(num, 1);
     else {
         addr = (num << 2) + idt.base;
 
@@ -1487,7 +1487,7 @@ x86_int_sw(int num)
             oxpc = cpu_state.pc;
 #endif
             cpu_state.pc = readmemw(0, addr);
-            is486 ? loadcs(readmemw(0, addr + 2)) : loadcs_2386(readmemw(0, addr + 2));
+            cpu_use_exec ? loadcs(readmemw(0, addr + 2)) : loadcs_2386(readmemw(0, addr + 2));
             cycles -= timing_int_rm;
         }
     }
@@ -1529,7 +1529,7 @@ x86_int_sw_rm(int num)
     cpu_state.eflags &= ~VIF_FLAG;
     cpu_state.flags &= ~T_FLAG;
     cpu_state.pc = new_pc;
-    is486 ? loadcs(new_cs) : loadcs_2386(new_cs);
+    cpu_use_exec ? loadcs(new_cs) : loadcs_2386(new_cs);
 #ifndef USE_NEW_DYNAREC
     oxpc = cpu_state.pc;
 #endif
diff --git a/src/cpu/386_common.h b/src/cpu/386_common.h
index 35d4f7cc8..a98a3e930 100644
--- a/src/cpu/386_common.h
+++ b/src/cpu/386_common.h
@@ -225,19 +225,37 @@ int checkio(uint32_t port, int mask);
 static __inline uint8_t
 fastreadb(uint32_t a)
 {
-    return readmembl_2386(a);
+    uint8_t ret;
+    read_type = 1;
+    ret = readmembl_2386(a);
+    read_type = 4;
+    if (cpu_state.abrt)
+        return 0;
+    return ret;
 }
 
 static __inline uint16_t
 fastreadw(uint32_t a)
 {
-    return readmemwl_2386(a);
+    uint16_t ret;
+    read_type = 1;
+    ret = readmemwl_2386(a);
+    read_type = 4;
+    if (cpu_state.abrt)
+        return 0;
+    return ret;
 }
 
 static __inline uint32_t
 fastreadl(uint32_t a)
 {
-    return readmemll_2386(a);
+    uint32_t ret;
+    read_type = 1;
+    ret = readmemll_2386(a);
+    read_type = 4;
+    if (cpu_state.abrt)
+        return 0;
+    return ret;
 }
 #else
 static __inline uint8_t
@@ -342,31 +360,41 @@ extern int opcode_length[256];
 static __inline uint16_t
 fastreadw_fetch(uint32_t a)
 {
-    uint16_t val;
+    uint16_t ret;
 
     if ((a & 0xFFF) > 0xFFE) {
-        val = fastreadb(a);
-        if (opcode_length[val & 0xff] > 1)
-            val |= ((uint16_t) fastreadb(a + 1) << 8);
-        return val;
+        ret = fastreadb(a);
+        if (!cpu_state.abrt && (opcode_length[ret & 0xff] > 1))
+            ret |= ((uint16_t) fastreadb(a + 1) << 8);
+    } else if (cpu_state.abrt)
+        ret = 0;
+    else {
+        read_type = 1;
+        ret = readmemwl_2386(a);
+        read_type = 4;
     }
 
-    return readmemwl_2386(a);
+    return ret;
 }
 
 static __inline uint32_t
 fastreadl_fetch(uint32_t a)
 {
-    uint32_t val;
+    uint32_t ret;
 
     if (cpu_16bitbus || ((a & 0xFFF) > 0xFFC)) {
-        val = fastreadw_fetch(a);
-        if (opcode_length[val & 0xff] > 2)
-            val |= ((uint32_t) fastreadw(a + 2) << 16);
-        return val;
+        ret = fastreadw_fetch(a);
+        if (!cpu_state.abrt && (opcode_length[ret & 0xff] > 2))
+            ret |= ((uint32_t) fastreadw(a + 2) << 16);
+    } else if (cpu_state.abrt)
+        ret = 0;
+    else {
+        read_type = 1;
+        ret = readmemll_2386(a);
+        read_type = 4;
     }
 
-    return readmemll_2386(a);
+    return ret;
 }
 #else
 static __inline uint16_t
diff --git a/src/cpu/386_ops.h b/src/cpu/386_ops.h
index 1bb3c167f..3e0d191f2 100644
--- a/src/cpu/386_ops.h
+++ b/src/cpu/386_ops.h
@@ -186,7 +186,11 @@ extern void x386_dynarec_log(const char *fmt, ...);
 #else
 #    include "x86_ops_flag.h"
 #endif
-#include "x86_ops_fpu.h"
+#ifdef OPS_286_386
+#    include "x86_ops_fpu_2386.h"
+#else
+#    include "x86_ops_fpu.h"
+#endif
 #include "x86_ops_inc_dec.h"
 #include "x86_ops_int.h"
 #include "x86_ops_io.h"
@@ -216,7 +220,11 @@ extern void x386_dynarec_log(const char *fmt, ...);
 #endif
 #include "x86_ops_mul.h"
 #include "x86_ops_pmode.h"
-#include "x86_ops_prefix.h"
+#ifdef OPS_286_386
+#    include "x86_ops_prefix_2386.h"
+#else
+#    include "x86_ops_prefix.h"
+#endif
 #ifdef IS_DYNAREC
 #    include "x86_ops_rep_dyn.h"
 #else
diff --git a/src/cpu/cpu.h b/src/cpu/cpu.h
index c829a0e17..db11b6135 100644
--- a/src/cpu/cpu.h
+++ b/src/cpu/cpu.h
@@ -174,6 +174,7 @@ typedef struct {
 #define VIP_FLAG   0x0010 /* in EFLAGS */
 #define VID_FLAG   0x0020 /* in EFLAGS */
 
+#define EM_FLAG    0x00004 /* in CR0 */
 #define WP_FLAG    0x10000 /* in CR0 */
 
 #define CR4_VME    (1 << 0) /* Virtual 8086 Mode Extensions */
diff --git a/src/cpu/x86.c b/src/cpu/x86.c
index 64ff6be4c..8e4a5b547 100644
--- a/src/cpu/x86.c
+++ b/src/cpu/x86.c
@@ -275,6 +275,12 @@ reset_common(int hard)
     cr4              = 0;
     cpu_state.eflags = 0;
     cgate32          = 0;
+    if (is386 && !is486) {
+        for (uint8_t i = 0; i < 4; i++)
+            dr[i] = 0x00000000;
+        dr[6] = 0xffff1ff0;
+        dr[7] = 0x00000400;
+    }
     if (is286) {
         if (is486)
             loadcs(0xF000);
diff --git a/src/cpu/x86_ops_fpu_2386.h b/src/cpu/x86_ops_fpu_2386.h
new file mode 100644
index 000000000..52c24d995
--- /dev/null
+++ b/src/cpu/x86_ops_fpu_2386.h
@@ -0,0 +1,113 @@
+/* Copyright holders: Sarah Walker
+   see COPYING for more details
+*/
+static int
+opESCAPE_d8_a16(uint32_t fetchdat)
+{
+    return x86_2386_opcodes_d8_a16[(fetchdat >> 3) & 0x1f](fetchdat);
+}
+static int
+opESCAPE_d8_a32(uint32_t fetchdat)
+{
+    return x86_2386_opcodes_d8_a32[(fetchdat >> 3) & 0x1f](fetchdat);
+}
+
+static int
+opESCAPE_d9_a16(uint32_t fetchdat)
+{
+    return x86_2386_opcodes_d9_a16[fetchdat & 0xff](fetchdat);
+}
+static int
+opESCAPE_d9_a32(uint32_t fetchdat)
+{
+    return x86_2386_opcodes_d9_a32[fetchdat & 0xff](fetchdat);
+}
+
+static int
+opESCAPE_da_a16(uint32_t fetchdat)
+{
+    return x86_2386_opcodes_da_a16[fetchdat & 0xff](fetchdat);
+}
+static int
+opESCAPE_da_a32(uint32_t fetchdat)
+{
+    return x86_2386_opcodes_da_a32[fetchdat & 0xff](fetchdat);
+}
+
+static int
+opESCAPE_db_a16(uint32_t fetchdat)
+{
+    return x86_2386_opcodes_db_a16[fetchdat & 0xff](fetchdat);
+}
+static int
+opESCAPE_db_a32(uint32_t fetchdat)
+{
+    return x86_2386_opcodes_db_a32[fetchdat & 0xff](fetchdat);
+}
+
+static int
+opESCAPE_dc_a16(uint32_t fetchdat)
+{
+    return x86_2386_opcodes_dc_a16[(fetchdat >> 3) & 0x1f](fetchdat);
+}
+static int
+opESCAPE_dc_a32(uint32_t fetchdat)
+{
+    return x86_2386_opcodes_dc_a32[(fetchdat >> 3) & 0x1f](fetchdat);
+}
+
+static int
+opESCAPE_dd_a16(uint32_t fetchdat)
+{
+    return x86_2386_opcodes_dd_a16[fetchdat & 0xff](fetchdat);
+}
+static int
+opESCAPE_dd_a32(uint32_t fetchdat)
+{
+    return x86_2386_opcodes_dd_a32[fetchdat & 0xff](fetchdat);
+}
+
+static int
+opESCAPE_de_a16(uint32_t fetchdat)
+{
+    return x86_2386_opcodes_de_a16[fetchdat & 0xff](fetchdat);
+}
+static int
+opESCAPE_de_a32(uint32_t fetchdat)
+{
+    return x86_2386_opcodes_de_a32[fetchdat & 0xff](fetchdat);
+}
+
+static int
+opESCAPE_df_a16(uint32_t fetchdat)
+{
+    return x86_2386_opcodes_df_a16[fetchdat & 0xff](fetchdat);
+}
+static int
+opESCAPE_df_a32(uint32_t fetchdat)
+{
+    return x86_2386_opcodes_df_a32[fetchdat & 0xff](fetchdat);
+}
+
+static int
+opWAIT(uint32_t fetchdat)
+{
+    if ((cr0 & 0xa) == 0xa) {
+        x86_int(7);
+        return 1;
+    }
+
+#if 0
+    if (!cpu_use_dynarec && fpu_softfloat) {
+#endif
+    if (fpu_softfloat) {
+        if (fpu_state.swd & FPU_SW_Summary) {
+            if (cr0 & 0x20) {
+                x86_int(16);
+                return 1;
+            }
+        }
+    }
+    CLOCK_CYCLES(4);
+    return 0;
+}
diff --git a/src/cpu/x86_ops_misc.h b/src/cpu/x86_ops_misc.h
index cbd2b3fbe..87dddefd8 100644
--- a/src/cpu/x86_ops_misc.h
+++ b/src/cpu/x86_ops_misc.h
@@ -726,6 +726,22 @@ opHLT(uint32_t fetchdat)
     return 0;
 }
 
+#ifdef OPS_286_386
+static int
+opLOCK(uint32_t fetchdat)
+{
+    fetchdat = fastreadl(cs + cpu_state.pc);
+    if (cpu_state.abrt)
+        return 0;
+    cpu_state.pc++;
+
+    ILLEGAL_ON((fetchdat & 0xff) == 0x90);
+
+    CLOCK_CYCLES(4);
+    PREFETCH_PREFIX();
+    return x86_2386_opcodes[(fetchdat & 0xff) | cpu_state.op32](fetchdat >> 8);
+}
+#else
 static int
 opLOCK(uint32_t fetchdat)
 {
@@ -740,6 +756,7 @@ opLOCK(uint32_t fetchdat)
     PREFETCH_PREFIX();
     return x86_opcodes[(fetchdat & 0xff) | cpu_state.op32](fetchdat >> 8);
 }
+#endif
 
 static int
 opBOUND_w_a16(uint32_t fetchdat)
diff --git a/src/cpu/x86_ops_mov_ctrl_2386.h b/src/cpu/x86_ops_mov_ctrl_2386.h
index c57dc8226..cae6c9957 100644
--- a/src/cpu/x86_ops_mov_ctrl_2386.h
+++ b/src/cpu/x86_ops_mov_ctrl_2386.h
@@ -82,18 +82,41 @@ opMOV_r_CRx_a32(uint32_t fetchdat)
 static int
 opMOV_r_DRx_a16(uint32_t fetchdat)
 {
-    if ((CPL || (cpu_state.eflags & VM_FLAG)) && (cr0 & 1)) {
+    if ((CPL > 0) && (cr0 & 1)) {
         x86gpf(NULL, 0);
         return 1;
     }
-    fetch_ea_16(fetchdat);
-    if (cpu_reg == 4 || cpu_reg == 5) {
-        if (cr4 & 0x8)
-            x86illegal();
-        else
-            cpu_reg += 2;
+    if ((dr[7] & 0x2000) && !(cpu_state.eflags & RF_FLAG)) {
+        trap |= 1;
+        return 1;
+    }
+    fetch_ea_16(fetchdat);
+    switch (cpu_reg) {
+        case 0 ... 3:
+            cpu_state.regs[cpu_rm].l = dr[cpu_reg];
+            break;
+        case 4:
+            if (cr4 & 0x8) {
+                x86illegal();
+                return 1;
+            }
+            fallthrough;
+        case 6:
+            cpu_state.regs[cpu_rm].l = dr[6];
+            break;
+        case 5:
+            if (cr4 & 0x8) {
+                x86illegal();
+                return 1;
+            }
+            fallthrough;
+        case 7:
+            cpu_state.regs[cpu_rm].l = dr[7];
+            break;
+        default:
+            x86illegal();
+            return 1;
     }
-    cpu_state.regs[cpu_rm].l = dr[cpu_reg] | (cpu_reg == 6 ? 0xffff0ff0u : 0);
     CLOCK_CYCLES(6);
     PREFETCH_RUN(6, 2, rmdat, 0, 0, 0, 0, 0);
     return 0;
@@ -101,18 +124,41 @@ opMOV_r_DRx_a16(uint32_t fetchdat)
 static int
 opMOV_r_DRx_a32(uint32_t fetchdat)
 {
-    if ((CPL || (cpu_state.eflags & VM_FLAG)) && (cr0 & 1)) {
+    if ((CPL > 0) && (cr0 & 1)) {
         x86gpf(NULL, 0);
         return 1;
     }
-    fetch_ea_32(fetchdat);
-    if (cpu_reg == 4 || cpu_reg == 5) {
-        if (cr4 & 0x8)
-            x86illegal();
-        else
-            cpu_reg += 2;
+    if ((dr[7] & 0x2000) && !(cpu_state.eflags & RF_FLAG)) {
+        trap |= 1;
+        return 1;
+    }
+    fetch_ea_32(fetchdat);
+    switch (cpu_reg) {
+        case 0 ... 3:
+            cpu_state.regs[cpu_rm].l = dr[cpu_reg];
+            break;
+        case 4:
+            if (cr4 & 0x8) {
+                x86illegal();
+                return 1;
+            }
+            fallthrough;
+        case 6:
+            cpu_state.regs[cpu_rm].l = dr[6];
+            break;
+        case 5:
+            if (cr4 & 0x8) {
+                x86illegal();
+                return 1;
+            }
+            fallthrough;
+        case 7:
+            cpu_state.regs[cpu_rm].l = dr[7];
+            break;
+        default:
+            x86illegal();
+            return 1;
     }
-    cpu_state.regs[cpu_rm].l = dr[cpu_reg] | (cpu_reg == 6 ? 0xffff0ff0u : 0);
     CLOCK_CYCLES(6);
     PREFETCH_RUN(6, 2, rmdat, 0, 0, 0, 0, 1);
     return 0;
@@ -236,24 +282,41 @@ opMOV_CRx_r_a32(uint32_t fetchdat)
 static int
 opMOV_DRx_r_a16(uint32_t fetchdat)
 {
-    if ((CPL || (cpu_state.eflags & VM_FLAG)) && (cr0 & 1)) {
+    if ((CPL > 0) && (cr0 & 1)) {
         x86gpf(NULL, 0);
         return 1;
     }
-    if ((dr[6] & 0x2000) && !(cpu_state.eflags & RF_FLAG)) {
-        dr[7] |= 0x2000;
-        dr[6] &= ~0x2000;
-        x86gen();
+    if ((dr[7] & 0x2000) && !(cpu_state.eflags & RF_FLAG)) {
+        trap |= 1;
         return 1;
     }
     fetch_ea_16(fetchdat);
-    if (cpu_reg == 4 || cpu_reg == 5) {
-        if (cr4 & 0x8)
+    switch (cpu_reg) {
+        case 0 ... 3:
+            dr[cpu_reg] = cpu_state.regs[cpu_rm].l;
+            break;
+        case 4:
+            if (cr4 & 0x8) {
+                x86illegal();
+                return 1;
+            }
+            fallthrough;
+        case 6:
+            dr[6] = (dr[6] & 0xffff0ff0) | (cpu_state.regs[cpu_rm].l & 0x0000f00f);
+            break;
+        case 5:
+            if (cr4 & 0x8) {
+                x86illegal();
+                return 1;
+            }
+            fallthrough;
+        case 7:
+            dr[7] = cpu_state.regs[cpu_rm].l | 0x00000400;
+            break;
+        default:
             x86illegal();
-        else
-            cpu_reg += 2;
+            return 1;
     }
-    dr[cpu_reg] = cpu_state.regs[cpu_rm].l;
     CLOCK_CYCLES(6);
     PREFETCH_RUN(6, 2, rmdat, 0, 0, 0, 0, 0);
     CPU_BLOCK_END();
@@ -262,18 +325,41 @@ opMOV_DRx_r_a16(uint32_t fetchdat)
 static int
 opMOV_DRx_r_a32(uint32_t fetchdat)
 {
-    if ((CPL || (cpu_state.eflags & VM_FLAG)) && (cr0 & 1)) {
+    if ((CPL > 0) && (cr0 & 1)) {
         x86gpf(NULL, 0);
         return 1;
     }
-    fetch_ea_16(fetchdat);
-    if (cpu_reg == 4 || cpu_reg == 5) {
-        if (cr4 & 0x8)
-            x86illegal();
-        else
-            cpu_reg += 2;
+    if ((dr[7] & 0x2000) && !(cpu_state.eflags & RF_FLAG)) {
+        trap |= 1;
+        return 1;
+    }
+    fetch_ea_32(fetchdat);
+    switch (cpu_reg) {
+        case 0 ... 3:
+            dr[cpu_reg] = cpu_state.regs[cpu_rm].l;
+            break;
+        case 4:
+            if (cr4 & 0x8) {
+                x86illegal();
+                return 1;
+            }
+            fallthrough;
+        case 6:
+            dr[6] = (dr[6] & 0xffff0ff0) | (cpu_state.regs[cpu_rm].l & 0x0000f00f);
+            break;
+        case 5:
+            if (cr4 & 0x8) {
+                x86illegal();
+                return 1;
+            }
+            fallthrough;
+        case 7:
+            dr[7] = cpu_state.regs[cpu_rm].l | 0x00000400;
+            break;
+        default:
+            x86illegal();
+            return 1;
     }
-    dr[cpu_reg] = cpu_state.regs[cpu_rm].l;
     CLOCK_CYCLES(6);
     PREFETCH_RUN(6, 2, rmdat, 0, 0, 0, 0, 1);
     CPU_BLOCK_END();
diff --git a/src/cpu/x86_ops_mov_seg.h b/src/cpu/x86_ops_mov_seg.h
index 2a798db5c..7fcc92312 100644
--- a/src/cpu/x86_ops_mov_seg.h
+++ b/src/cpu/x86_ops_mov_seg.h
@@ -195,7 +195,11 @@ opMOV_seg_w_a16(uint32_t fetchdat)
             cpu_state.pc++;
             if (cpu_state.abrt)
                 return 1;
+#ifdef OPS_286_386
+            x86_2386_opcodes[(fetchdat & 0xff) | cpu_state.op32](fetchdat >> 8);
+#else
             x86_opcodes[(fetchdat & 0xff) | cpu_state.op32](fetchdat >> 8);
+#endif
             return 1;
         case 0x20: /*FS*/
             op_loadseg(new_seg, &cpu_state.seg_fs);
@@ -240,7 +244,11 @@ opMOV_seg_w_a32(uint32_t fetchdat)
             cpu_state.pc++;
             if (cpu_state.abrt)
                 return 1;
+#ifdef OPS_286_386
+            x86_2386_opcodes[(fetchdat & 0xff) | cpu_state.op32](fetchdat >> 8);
+#else
             x86_opcodes[(fetchdat & 0xff) | cpu_state.op32](fetchdat >> 8);
+#endif
             return 1;
         case 0x20: /*FS*/
             op_loadseg(new_seg, &cpu_state.seg_fs);
diff --git a/src/cpu/x86_ops_prefix_2386.h b/src/cpu/x86_ops_prefix_2386.h
new file mode 100644
index 000000000..7c87f0bf3
--- /dev/null
+++ b/src/cpu/x86_ops_prefix_2386.h
@@ -0,0 +1,179 @@
+#define op_seg(name, seg, opcode_table, normal_opcode_table)                  \
+    static int op##name##_w_a16(uint32_t fetchdat)                            \
+    {                                                                         \
+        fetchdat = fastreadl(cs + cpu_state.pc);                              \
+        if (cpu_state.abrt)                                                   \
+            return 1;                                                         \
+        cpu_state.pc++;                                                       \
+                                                                              \
+        cpu_state.ea_seg = &seg;                                              \
+        cpu_state.ssegs  = 1;                                                 \
+        CLOCK_CYCLES(4);                                                      \
+        PREFETCH_PREFIX();                                                    \
+                                                                              \
+        if (opcode_table[fetchdat & 0xff])                                    \
+            return opcode_table[fetchdat & 0xff](fetchdat >> 8);              \
+        return normal_opcode_table[fetchdat & 0xff](fetchdat >> 8);           \
+    }                                                                         \
+                                                                              \
+    static int op##name##_l_a16(uint32_t fetchdat)                            \
+    {                                                                         \
+        fetchdat = fastreadl(cs + cpu_state.pc);                              \
+        if (cpu_state.abrt)                                                   \
+            return 1;                                                         \
+        cpu_state.pc++;                                                       \
+                                                                              \
+        cpu_state.ea_seg = &seg;                                              \
+        cpu_state.ssegs  = 1;                                                 \
+        CLOCK_CYCLES(4);                                                      \
+        PREFETCH_PREFIX();                                                    \
+                                                                              \
+        if (opcode_table[(fetchdat & 0xff) | 0x100])                          \
+            return opcode_table[(fetchdat & 0xff) | 0x100](fetchdat >> 8);    \
+        return normal_opcode_table[(fetchdat & 0xff) | 0x100](fetchdat >> 8); \
+    }                                                                         \
+                                                                              \
+    static int op##name##_w_a32(uint32_t fetchdat)                            \
+    {                                                                         \
+        fetchdat = fastreadl(cs + cpu_state.pc);                              \
+        if (cpu_state.abrt)                                                   \
+            return 1;                                                         \
+        cpu_state.pc++;                                                       \
+                                                                              \
+        cpu_state.ea_seg = &seg;                                              \
+        cpu_state.ssegs  = 1;                                                 \
+        CLOCK_CYCLES(4);                                                      \
+        PREFETCH_PREFIX();                                                    \
+                                                                              \
+        if (opcode_table[(fetchdat & 0xff) | 0x200])                          \
+            return opcode_table[(fetchdat & 0xff) | 0x200](fetchdat >> 8);    \
+        return normal_opcode_table[(fetchdat & 0xff) | 0x200](fetchdat >> 8); \
+    }                                                                         \
+                                                                              \
+    static int op##name##_l_a32(uint32_t fetchdat)                            \
+    {                                                                         \
+        fetchdat = fastreadl(cs + cpu_state.pc);                              \
+        if (cpu_state.abrt)                                                   \
+            return 1;                                                         \
+        cpu_state.pc++;                                                       \
+                                                                              \
+        cpu_state.ea_seg = &seg;                                              \
+        cpu_state.ssegs  = 1;                                                 \
+        CLOCK_CYCLES(4);                                                      \
+        PREFETCH_PREFIX();                                                    \
+                                                                              \
+        if (opcode_table[(fetchdat & 0xff) | 0x300])                          \
+            return opcode_table[(fetchdat & 0xff) | 0x300](fetchdat >> 8);    \
+        return normal_opcode_table[(fetchdat & 0xff) | 0x300](fetchdat >> 8); \
+    }
+
+// clang-format off
+op_seg(CS, cpu_state.seg_cs, x86_2386_opcodes, x86_2386_opcodes)
+op_seg(DS, cpu_state.seg_ds, x86_2386_opcodes, x86_2386_opcodes)
+op_seg(ES, cpu_state.seg_es, x86_2386_opcodes, x86_2386_opcodes)
+op_seg(FS, cpu_state.seg_fs, x86_2386_opcodes, x86_2386_opcodes)
+op_seg(GS, cpu_state.seg_gs, x86_2386_opcodes, x86_2386_opcodes)
+op_seg(SS, cpu_state.seg_ss, x86_2386_opcodes, x86_2386_opcodes)
+
+op_seg(CS_REPE, cpu_state.seg_cs, x86_2386_opcodes_REPE, x86_2386_opcodes)
+op_seg(DS_REPE, cpu_state.seg_ds, x86_2386_opcodes_REPE, x86_2386_opcodes)
+op_seg(ES_REPE, cpu_state.seg_es, x86_2386_opcodes_REPE, x86_2386_opcodes)
+op_seg(FS_REPE, cpu_state.seg_fs, x86_2386_opcodes_REPE, x86_2386_opcodes)
+op_seg(GS_REPE, cpu_state.seg_gs, x86_2386_opcodes_REPE, x86_2386_opcodes)
+op_seg(SS_REPE, cpu_state.seg_ss, x86_2386_opcodes_REPE, x86_2386_opcodes)
+
+op_seg(CS_REPNE, cpu_state.seg_cs, x86_2386_opcodes_REPNE, x86_2386_opcodes)
+op_seg(DS_REPNE, cpu_state.seg_ds, x86_2386_opcodes_REPNE, x86_2386_opcodes)
+op_seg(ES_REPNE, cpu_state.seg_es, x86_2386_opcodes_REPNE, x86_2386_opcodes)
+op_seg(FS_REPNE, cpu_state.seg_fs, x86_2386_opcodes_REPNE, x86_2386_opcodes)
+op_seg(GS_REPNE, cpu_state.seg_gs, x86_2386_opcodes_REPNE, x86_2386_opcodes)
+op_seg(SS_REPNE, cpu_state.seg_ss, x86_2386_opcodes_REPNE, x86_2386_opcodes)
+    // clang-format on
+
+static int
+op_66(uint32_t fetchdat) /*Data size select*/
+{
+    fetchdat = fastreadl(cs + cpu_state.pc);
+    if (cpu_state.abrt)
+        return 1;
+    cpu_state.pc++;
+
+    cpu_state.op32 = ((use32 & 0x100) ^ 0x100) | (cpu_state.op32 & 0x200);
+    CLOCK_CYCLES(2);
+    PREFETCH_PREFIX();
+    return x86_2386_opcodes[(fetchdat & 0xff) | cpu_state.op32](fetchdat >> 8);
+}
+static int
+op_67(uint32_t fetchdat) /*Address size select*/
+{
+    fetchdat = fastreadl(cs + cpu_state.pc);
+    if (cpu_state.abrt)
+        return 1;
+    cpu_state.pc++;
+
+    cpu_state.op32 = ((use32 & 0x200) ^ 0x200) | (cpu_state.op32 & 0x100);
+    CLOCK_CYCLES(2);
+    PREFETCH_PREFIX();
+    return x86_2386_opcodes[(fetchdat & 0xff) | cpu_state.op32](fetchdat >> 8);
+}
+
+static int
+op_66_REPE(uint32_t fetchdat) /*Data size select*/
+{
+    fetchdat = fastreadl(cs + cpu_state.pc);
+    if (cpu_state.abrt)
+        return 1;
+    cpu_state.pc++;
+
+    cpu_state.op32 = ((use32 & 0x100) ^ 0x100) | (cpu_state.op32 & 0x200);
+    CLOCK_CYCLES(2);
+    PREFETCH_PREFIX();
+    if (x86_2386_opcodes_REPE[(fetchdat & 0xff) | cpu_state.op32])
+        return x86_2386_opcodes_REPE[(fetchdat & 0xff) | cpu_state.op32](fetchdat >> 8);
+    return x86_2386_opcodes[(fetchdat & 0xff) | cpu_state.op32](fetchdat >> 8);
+}
+static int
+op_67_REPE(uint32_t fetchdat) /*Address size select*/
+{
+    fetchdat = fastreadl(cs + cpu_state.pc);
+    if (cpu_state.abrt)
+        return 1;
+    cpu_state.pc++;
+
+    cpu_state.op32 = ((use32 & 0x200) ^ 0x200) | (cpu_state.op32 & 0x100);
+    CLOCK_CYCLES(2);
+    PREFETCH_PREFIX();
+    if (x86_2386_opcodes_REPE[(fetchdat & 0xff) | cpu_state.op32])
+        return x86_2386_opcodes_REPE[(fetchdat & 0xff) | cpu_state.op32](fetchdat >> 8);
+    return x86_2386_opcodes[(fetchdat & 0xff) | cpu_state.op32](fetchdat >> 8);
+}
+static int
+op_66_REPNE(uint32_t fetchdat) /*Data size select*/
+{
+    fetchdat = fastreadl(cs + cpu_state.pc);
+    if (cpu_state.abrt)
+        return 1;
+    cpu_state.pc++;
+
+    cpu_state.op32 = ((use32 & 0x100) ^ 0x100) | (cpu_state.op32 & 0x200);
+    CLOCK_CYCLES(2);
+    PREFETCH_PREFIX();
+    if (x86_2386_opcodes_REPNE[(fetchdat & 0xff) | cpu_state.op32])
+        return x86_2386_opcodes_REPNE[(fetchdat & 0xff) | cpu_state.op32](fetchdat >> 8);
+    return x86_2386_opcodes[(fetchdat & 0xff) | cpu_state.op32](fetchdat >> 8);
+}
+static int
+op_67_REPNE(uint32_t fetchdat) /*Address size select*/
+{
+    fetchdat = fastreadl(cs + cpu_state.pc);
+    if (cpu_state.abrt)
+        return 1;
+    cpu_state.pc++;
+
+    cpu_state.op32 = ((use32 & 0x200) ^ 0x200) | (cpu_state.op32 & 0x100);
+    CLOCK_CYCLES(2);
+    PREFETCH_PREFIX();
+    if (x86_2386_opcodes_REPNE[(fetchdat & 0xff) | cpu_state.op32])
+        return x86_2386_opcodes_REPNE[(fetchdat & 0xff) | cpu_state.op32](fetchdat >> 8);
+    return x86_2386_opcodes[(fetchdat & 0xff) | cpu_state.op32](fetchdat >> 8);
+}
diff --git a/src/cpu/x86_ops_rep_2386.h b/src/cpu/x86_ops_rep_2386.h
index b6f64e90d..fe5048340 100644
--- a/src/cpu/x86_ops_rep_2386.h
+++ b/src/cpu/x86_ops_rep_2386.h
@@ -836,7 +836,7 @@ REP_OPS_CMPS_SCAS(a32_E, ECX, ESI, EDI, 1)
 static int
 opREPNE(uint32_t fetchdat)
 {
-    fetchdat = fastreadl(cs + cpu_state.pc);
+    fetchdat = fastreadl_fetch(cs + cpu_state.pc);
     if (cpu_state.abrt)
         return 1;
     cpu_state.pc++;
@@ -850,7 +850,7 @@ opREPNE(uint32_t fetchdat)
 static int
 opREPE(uint32_t fetchdat)
 {
-    fetchdat = fastreadl(cs + cpu_state.pc);
+    fetchdat = fastreadl_fetch(cs + cpu_state.pc);
     if (cpu_state.abrt)
         return 1;
     cpu_state.pc++;
diff --git a/src/cpu/x86_ops_stack.h b/src/cpu/x86_ops_stack.h
index 13eb883d3..fbf603ddb 100644
--- a/src/cpu/x86_ops_stack.h
+++ b/src/cpu/x86_ops_stack.h
@@ -667,7 +667,11 @@ opPOP_SS_w(uint32_t fetchdat)
     cpu_state.pc++;
     if (cpu_state.abrt)
         return 1;
+#ifdef OPS_286_386
+    x86_2386_opcodes[(fetchdat & 0xff) | cpu_state.op32](fetchdat >> 8);
+#else
     x86_opcodes[(fetchdat & 0xff) | cpu_state.op32](fetchdat >> 8);
+#endif
 
     return 1;
 }
@@ -695,7 +699,11 @@ opPOP_SS_l(uint32_t fetchdat)
     cpu_state.pc++;
     if (cpu_state.abrt)
         return 1;
+#ifdef OPS_286_386
+    x86_2386_opcodes[(fetchdat & 0xff) | cpu_state.op32](fetchdat >> 8);
+#else
     x86_opcodes[(fetchdat & 0xff) | cpu_state.op32](fetchdat >> 8);
+#endif
 
     return 1;
 }
diff --git a/src/include/86box/mem.h b/src/include/86box/mem.h
index 15afcb160..bc949834f 100644
--- a/src/include/86box/mem.h
+++ b/src/include/86box/mem.h
@@ -303,6 +303,7 @@ extern int     mmu_perm;
 extern uint8_t high_page; /* if a high (> 4 gb) page was detected */
 
 extern uint32_t pages_sz; /* #pages in table */
+extern int      read_type;
 
 extern int mem_a20_state;
 extern int mem_a20_alt;
diff --git a/src/mem/mmu_2386.c b/src/mem/mmu_2386.c
index 9ed26edfa..7c286b605 100644
--- a/src/mem/mmu_2386.c
+++ b/src/mem/mmu_2386.c
@@ -39,52 +39,43 @@
 #include <86box/rom.h>
 #include <86box/gdbstub.h>
 
-/* Set trap for data address breakpoints. */
-void
-mem_debug_check_addr(uint32_t addr, int write)
-{
-    int i = 0;
-    int set_trap = 0;
+/* As below, 1 = exec, 4 = read. */
+int    read_type = 4;
 
-    if (!(dr[7] & 0xFF))
+/* Set trap for data address breakpoints - 1 = exec, 2 = write, 4 = read. */
+void
+mem_debug_check_addr(uint32_t addr, int flags)
+{
+    uint32_t bp_addr;
+    uint32_t bp_mask;
+    uint32_t len_type_pair;
+    int bp_enabled;
+    uint8_t match_flags[4] = { 0, 2, 0, 6 };
+    char *bp_types[5] = { "N/A 0", "EXEC ", "WRITE", "N/A 3", "READ " };
+
+    if (cpu_state.abrt || ((flags == 1) && (cpu_state.eflags & RF_FLAG)))
         return;
 
-    for (i = 0; i < 4; i++) {
-        uint32_t dr_addr = dr[i];
-        int breakpoint_enabled = !!(dr[7] & (0x3 << (2 * i)));
-        int len_type_pair = ((dr[7] >> 16) & (0xF << (4 * i))) >> (4 * i);
-        if (!breakpoint_enabled)
-            continue;
-        if (!write && (len_type_pair & 3) != 3)
-            continue;
-        if ((len_type_pair & 3) != 1)
-            continue;
-        
-        switch ((len_type_pair >> 2) & 3)
-        {
-            case 0x00:
-                if (dr_addr == addr) {
-                    set_trap = 1;
-                    dr[6] |= (1 << i);
-                }
-                break;
-            case 0x01:
-                if ((dr_addr & ~1) == addr || ((dr_addr & ~1) + 1) == (addr + 1)) {
-                    set_trap = 1;
-                    dr[6] |= (1 << i);
-                }
-                break;
-            case 0x03:
-                dr_addr &= ~3;
-                if (addr >= dr_addr && addr < (dr_addr + 4)) {
-                    set_trap = 1;
-                    dr[6] |= (1 << i);
-                }
-                break;
+    if (dr[7] & 0x000000ff)  for (uint8_t i = 0; i < 4; i++) {
+        bp_addr = dr[i];
+        bp_enabled = (dr[7] >> (i << 1)) & 0x03;
+        len_type_pair = (dr[7] >> (16 + (i << 2))) & 0x0f;
+        bp_mask = ~((len_type_pair >> 2) & 0x03);
+
+        if ((flags & match_flags[len_type_pair & 0x03]) && ((bp_addr & bp_mask) == (addr & bp_mask))) {
+            /*
+               From the Intel i386 documemntation:
+
+               (Note that the processor sets Bn regardless of whether Gn or
+               Ln is set. If more than one breakpoint condition occurs at one time and if
+               the breakpoint trap occurs due to an enabled condition other than n, Bn may
+               be set, even though neither Gn nor Ln is set.)
+             */
+            dr[6] |= (1 << i);
+            if (bp_enabled)
+                trap |= (read_type == 1) ? 8 : 4;
         }
     }
-    if (set_trap)
-        trap |= 4;
 }
 
 uint8_t
@@ -291,7 +282,7 @@ readmembl_2386(uint32_t addr)
 
     GDBSTUB_MEM_ACCESS(addr, GDBSTUB_MEM_READ, 1);
 
-    mem_debug_check_addr(addr, 0);
+    mem_debug_check_addr(addr, read_type);
     addr64           = (uint64_t) addr;
     mem_logical_addr = addr;
 
@@ -319,7 +310,7 @@ writemembl_2386(uint32_t addr, uint8_t val)
     mem_mapping_t *map;
     uint64_t       a;
 
-    mem_debug_check_addr(addr, 1);
+    mem_debug_check_addr(addr, 2);
     GDBSTUB_MEM_ACCESS(addr, GDBSTUB_MEM_WRITE, 1);
 
     addr64           = (uint64_t) addr;
@@ -397,8 +388,8 @@ readmemwl_2386(uint32_t addr)
 
     addr64a[0] = addr;
     addr64a[1] = addr + 1;
-    mem_debug_check_addr(addr, 0);
-    mem_debug_check_addr(addr + 1, 0);
+    mem_debug_check_addr(addr, read_type);
+    mem_debug_check_addr(addr + 1, read_type);
     GDBSTUB_MEM_ACCESS_FAST(addr64a, GDBSTUB_MEM_READ, 2);
 
     mem_logical_addr = addr;
@@ -419,7 +410,8 @@ readmemwl_2386(uint32_t addr)
                 }
             }
 
-            return readmembl_no_mmut(addr, addr64a[0]) | (((uint16_t) readmembl_no_mmut(addr + 1, addr64a[1])) << 8);
+            return readmembl_no_mmut_2386(addr, addr64a[0]) |
+                   (((uint16_t) readmembl_no_mmut_2386(addr + 1, addr64a[1])) << 8);
         }
     }
 
@@ -454,8 +446,8 @@ writememwl_2386(uint32_t addr, uint16_t val)
 
     addr64a[0] = addr;
     addr64a[1] = addr + 1;
-    mem_debug_check_addr(addr, 1);
-    mem_debug_check_addr(addr + 1, 1);
+    mem_debug_check_addr(addr, 2);
+    mem_debug_check_addr(addr + 1, 2);
     GDBSTUB_MEM_ACCESS_FAST(addr64a, GDBSTUB_MEM_WRITE, 2);
 
     mem_logical_addr = addr;
@@ -482,8 +474,8 @@ writememwl_2386(uint32_t addr, uint16_t val)
 
             /* No need to waste precious CPU host cycles on mmutranslate's that were already done, just pass
                their result as a parameter to be used if needed. */
-            writemembl_no_mmut(addr, addr64a[0], val);
-            writemembl_no_mmut(addr + 1, addr64a[1], val >> 8);
+            writemembl_no_mmut_2386(addr, addr64a[0], val);
+            writemembl_no_mmut_2386(addr + 1, addr64a[1], val >> 8);
             return;
         }
     }
@@ -531,7 +523,8 @@ readmemwl_no_mmut_2386(uint32_t addr, uint32_t *a64)
                     return 0xffff;
             }
 
-            return readmembl_no_mmut(addr, a64[0]) | (((uint16_t) readmembl_no_mmut(addr + 1, a64[1])) << 8);
+            return readmembl_no_mmut_2386(addr, a64[0]) |
+                   (((uint16_t) readmembl_no_mmut_2386(addr + 1, a64[1])) << 8);
         }
     }
 
@@ -574,8 +567,8 @@ writememwl_no_mmut_2386(uint32_t addr, uint32_t *a64, uint16_t val)
                     return;
             }
 
-            writemembl_no_mmut(addr, a64[0], val);
-            writemembl_no_mmut(addr + 1, a64[1], val >> 8);
+            writemembl_no_mmut_2386(addr, a64[0], val);
+            writemembl_no_mmut_2386(addr + 1, a64[1], val >> 8);
             return;
         }
     }
@@ -611,7 +604,7 @@ readmemll_2386(uint32_t addr)
 
     for (i = 0; i < 4; i++) {
         addr64a[i] = (uint64_t) (addr + i);
-        mem_debug_check_addr(addr + i, 0);
+        mem_debug_check_addr(addr + i, read_type);
     }
     GDBSTUB_MEM_ACCESS_FAST(addr64a, GDBSTUB_MEM_READ, 4);
 
@@ -619,8 +612,8 @@ readmemll_2386(uint32_t addr)
 
     high_page = 0;
 
-    if (addr & 3) {
-        if (!cpu_cyrix_alignment || (addr & 7) > 4)
+    if (cpu_16bitbus || (addr & 3)) {
+        if ((addr & 3) && (!cpu_cyrix_alignment || (addr & 7) > 4))
             cycles -= timing_misaligned;
         if ((addr & 0xfff) > 0xffc) {
             if (cr0 >> 31) {
@@ -647,7 +640,8 @@ readmemll_2386(uint32_t addr)
 
             /* No need to waste precious CPU host cycles on mmutranslate's that were already done, just pass
                their result as a parameter to be used if needed. */
-            return readmemwl_no_mmut(addr, addr64a) | (((uint32_t) readmemwl_no_mmut(addr + 2, &(addr64a[2]))) << 16);
+            return readmemwl_no_mmut_2386(addr, addr64a) |
+                   (((uint32_t) readmemwl_no_mmut(addr + 2, &(addr64a[2]))) << 16);
         }
     }
 
@@ -684,7 +678,7 @@ writememll_2386(uint32_t addr, uint32_t val)
 
     for (i = 0; i < 4; i++) {
         addr64a[i] = (uint64_t) (addr + i);
-        mem_debug_check_addr(addr + i, 1);
+        mem_debug_check_addr(addr + i, 2);
     }
     GDBSTUB_MEM_ACCESS_FAST(addr64a, GDBSTUB_MEM_WRITE, 4);
 
@@ -692,8 +686,8 @@ writememll_2386(uint32_t addr, uint32_t val)
 
     high_page = 0;
 
-    if (addr & 3) {
-        if (!cpu_cyrix_alignment || (addr & 7) > 4)
+    if (cpu_16bitbus || (addr & 3)) {
+        if ((addr & 3) && (!cpu_cyrix_alignment || (addr & 7) > 4))
             cycles -= timing_misaligned;
         if ((addr & 0xfff) > 0xffc) {
             if (cr0 >> 31) {
@@ -724,8 +718,8 @@ writememll_2386(uint32_t addr, uint32_t val)
 
             /* No need to waste precious CPU host cycles on mmutranslate's that were already done, just pass
                their result as a parameter to be used if needed. */
-            writememwl_no_mmut(addr, &(addr64a[0]), val);
-            writememwl_no_mmut(addr + 2, &(addr64a[2]), val >> 16);
+            writememwl_no_mmut_2386(addr, &(addr64a[0]), val);
+            writememwl_no_mmut_2386(addr + 2, &(addr64a[2]), val >> 16);
             return;
         }
     }
@@ -770,8 +764,8 @@ readmemll_no_mmut_2386(uint32_t addr, uint32_t *a64)
 
     mem_logical_addr = addr;
 
-    if (addr & 3) {
-        if (!cpu_cyrix_alignment || (addr & 7) > 4)
+    if (cpu_16bitbus || (addr & 3)) {
+        if ((addr & 3) && (!cpu_cyrix_alignment || (addr & 7) > 4))
             cycles -= timing_misaligned;
         if ((addr & 0xfff) > 0xffc) {
             if (cr0 >> 31) {
@@ -779,7 +773,8 @@ readmemll_no_mmut_2386(uint32_t addr, uint32_t *a64)
                     return 0xffffffff;
             }
 
-            return readmemwl_no_mmut(addr, a64) | ((uint32_t) (readmemwl_no_mmut(addr + 2, &(a64[2]))) << 16);
+            return readmemwl_no_mmut_2386(addr, a64) |
+                   ((uint32_t) (readmemwl_no_mmut_2386(addr + 2, &(a64[2]))) << 16);
         }
     }
 
@@ -815,8 +810,8 @@ writememll_no_mmut_2386(uint32_t addr, uint32_t *a64, uint32_t val)
 
     mem_logical_addr = addr;
 
-    if (addr & 3) {
-        if (!cpu_cyrix_alignment || (addr & 7) > 4)
+    if (cpu_16bitbus || (addr & 3)) {
+        if ((addr & 3) && (!cpu_cyrix_alignment || (addr & 7) > 4))
             cycles -= timing_misaligned;
         if ((addr & 0xfff) > 0xffc) {
             if (cr0 >> 31) {
@@ -824,8 +819,8 @@ writememll_no_mmut_2386(uint32_t addr, uint32_t *a64, uint32_t val)
                     return;
             }
 
-            writememwl_no_mmut(addr, &(a64[0]), val);
-            writememwl_no_mmut(addr + 2, &(a64[2]), val >> 16);
+            writememwl_no_mmut_2386(addr, &(a64[0]), val);
+            writememwl_no_mmut_2386(addr + 2, &(a64[2]), val >> 16);
             return;
         }
     }
@@ -867,7 +862,7 @@ readmemql_2386(uint32_t addr)
 
     for (i = 0; i < 8; i++) {
         addr64a[i] = (uint64_t) (addr + i);
-        mem_debug_check_addr(addr + i, 0);
+        mem_debug_check_addr(addr + i, read_type);
     }
     GDBSTUB_MEM_ACCESS_FAST(addr64a, GDBSTUB_MEM_READ, 8);
 
@@ -902,7 +897,8 @@ readmemql_2386(uint32_t addr)
 
             /* No need to waste precious CPU host cycles on mmutranslate's that were already done, just pass
                their result as a parameter to be used if needed. */
-            return readmemll_no_mmut(addr, addr64a) | (((uint64_t) readmemll_no_mmut(addr + 4, &(addr64a[4]))) << 32);
+            return readmemll_no_mmut_2386(addr, addr64a) |
+                   (((uint64_t) readmemll_no_mmut_2386(addr + 4, &(addr64a[4]))) << 32);
         }
     }
 
@@ -932,7 +928,7 @@ writememql_2386(uint32_t addr, uint64_t val)
 
     for (i = 0; i < 8; i++) {
         addr64a[i] = (uint64_t) (addr + i);
-        mem_debug_check_addr(addr + i, 1);
+        mem_debug_check_addr(addr + i, 2);
     }
     GDBSTUB_MEM_ACCESS_FAST(addr64a, GDBSTUB_MEM_WRITE, 8);
 
@@ -971,8 +967,8 @@ writememql_2386(uint32_t addr, uint64_t val)
 
             /* No need to waste precious CPU host cycles on mmutranslate's that were already done, just pass
                their result as a parameter to be used if needed. */
-            writememll_no_mmut(addr, addr64a, val);
-            writememll_no_mmut(addr + 4, &(addr64a[4]), val >> 32);
+            writememll_no_mmut_2386(addr, addr64a, val);
+            writememll_no_mmut_2386(addr + 4, &(addr64a[4]), val >> 32);
             return;
         }
     }
@@ -1019,7 +1015,7 @@ do_mmutranslate_2386(uint32_t addr, uint32_t *a64, int num, int write)
     uint32_t last_addr = addr + (num - 1);
     uint64_t a = 0x0000000000000000ULL;
 
-    mem_debug_check_addr(addr, write);
+    mem_debug_check_addr(addr, write ? 2 : read_type);
 
     for (i = 0; i < num; i++)
 	a64[i] = (uint64_t) addr;
diff --git a/src/mem/rom.c b/src/mem/rom.c
index 4a20e8ebc..f9718b7ce 100644
--- a/src/mem/rom.c
+++ b/src/mem/rom.c
@@ -296,6 +296,12 @@ rom_load_linear_inverted(const char *fn, uint32_t addr, int sz, int off, uint8_t
             fatal("rom_load_linear_inverted(): Error reading the upper half of the data\n");
         if (fread(ptr + addr, sz >> 1, 1, fp) > (sz >> 1))
             fatal("rom_load_linear_inverted(): Error reading the lower half of the data\n");
+        if (sz == 0x40000) {
+            if (fread(ptr + addr + 0x30000, 1, sz >> 1, fp) > (sz >> 1))
+                fatal("rom_load_linear_inverted(): Error reading the upper half of the data\n");
+            if (fread(ptr + addr + 0x20000, sz >> 1, 1, fp) > (sz >> 1))
+                fatal("rom_load_linear_inverted(): Error reading the lower half of the data\n");
+        }
     }
 
     (void) fclose(fp);