diff --git a/src/codegen.c b/src/codegen.c
index 409221f..2f36a42 100644
--- a/src/codegen.c
+++ b/src/codegen.c
@@ -416,7 +416,7 @@ void codegen_generate_call(uint8_t opcode, OpFn op, uint32_t fetchdat, uint32_t
                         break;
                         case 0xde:
                         op_table = (op_32 & 0x200) ? x86_dynarec_opcodes_de_a32 : x86_dynarec_opcodes_de_a16;
-                        recomp_op_table = NULL;//recomp_opcodes_de;
+                        recomp_op_table = recomp_opcodes_de;
                         opcode_mask = 0xff;
                         over = 1;
                         pc_off = -1;
diff --git a/src/codegen_ops.c b/src/codegen_ops.c
index 55916ee..6fe6799 100644
--- a/src/codegen_ops.c
+++ b/src/codegen_ops.c
@@ -200,3 +200,50 @@ RecompOpFn recomp_opcodes_dc[512] =
 /*e0*/  ropFSUBRr,      ropFSUBRr,      ropFSUBRr,      ropFSUBRr,      ropFSUBRr,      ropFSUBRr,      ropFSUBRr,      ropFSUBRr,      ropFSUBr,       ropFSUBr,       ropFSUBr,       ropFSUBr,       ropFSUBr,       ropFSUBr,       ropFSUBr,       ropFSUBr,
 /*f0*/  ropFDIVRr,      ropFDIVRr,      ropFDIVRr,      ropFDIVRr,      ropFDIVRr,      ropFDIVRr,      ropFDIVRr,      ropFDIVRr,      ropFDIVr,       ropFDIVr,       ropFDIVr,       ropFDIVr,       ropFDIVr,       ropFDIVr,       ropFDIVr,       ropFDIVr,
 };
+
+RecompOpFn recomp_opcodes_de[512] =
+{
+        /*16-bit data*/
+/*      00              01              02              03              04              05              06              07              08              09              0a              0b              0c              0d              0e              0f*/
+/*00*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*10*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*20*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*30*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+
+/*40*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*50*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*60*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*70*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+
+/*80*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*90*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*a0*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*b0*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+
+/*c0*/  ropFADDP,       ropFADDP,       ropFADDP,       ropFADDP,       ropFADDP,       ropFADDP,       ropFADDP,       ropFADDP,       ropFMULP,       ropFMULP,       ropFMULP,       ropFMULP,       ropFMULP,       ropFMULP,       ropFMULP,       ropFMULP,
+/*d0*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*e0*/  ropFSUBRP,      ropFSUBRP,      ropFSUBRP,      ropFSUBRP,      ropFSUBRP,      ropFSUBRP,      ropFSUBRP,      ropFSUBRP,      ropFSUBP,       ropFSUBP,       ropFSUBP,       ropFSUBP,       ropFSUBP,       ropFSUBP,       ropFSUBP,       ropFSUBP,
+/*f0*/  ropFDIVRP,      ropFDIVRP,      ropFDIVRP,      ropFDIVRP,      ropFDIVRP,      ropFDIVRP,      ropFDIVRP,      ropFDIVRP,      ropFDIVP,       ropFDIVP,       ropFDIVP,       ropFDIVP,       ropFDIVP,       ropFDIVP,       ropFDIVP,       ropFDIVP,
+
+        /*32-bit data*/
+/*      00              01              02              03              04              05              06              07              08              09              0a              0b              0c              0d              0e              0f*/
+/*00*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*10*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*20*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*30*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+
+/*40*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*50*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*60*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*70*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+
+/*80*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*90*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*a0*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*b0*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+
+/*c0*/  ropFADDP,       ropFADDP,       ropFADDP,       ropFADDP,       ropFADDP,       ropFADDP,       ropFADDP,       ropFADDP,       ropFMULP,       ropFMULP,       ropFMULP,       ropFMULP,       ropFMULP,       ropFMULP,       ropFMULP,       ropFMULP,
+/*d0*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*e0*/  ropFSUBRP,      ropFSUBRP,      ropFSUBRP,      ropFSUBRP,      ropFSUBRP,      ropFSUBRP,      ropFSUBRP,      ropFSUBRP,      ropFSUBP,       ropFSUBP,       ropFSUBP,       ropFSUBP,       ropFSUBP,       ropFSUBP,       ropFSUBP,       ropFSUBP,
+/*f0*/  ropFDIVRP,      ropFDIVRP,      ropFDIVRP,      ropFDIVRP,      ropFDIVRP,      ropFDIVRP,      ropFDIVRP,      ropFDIVRP,      ropFDIVP,       ropFDIVP,       ropFDIVP,       ropFDIVP,       ropFDIVP,       ropFDIVP,       ropFDIVP,       ropFDIVP,
+};
diff --git a/src/codegen_ops.h b/src/codegen_ops.h
index fbbaed4..432652d 100644
--- a/src/codegen_ops.h
+++ b/src/codegen_ops.h
@@ -14,9 +14,9 @@ extern RecompOpFn recomp_opcodes_d8[512];
 extern RecompOpFn recomp_opcodes_da[512];
 extern RecompOpFn recomp_opcodes_db[512];*/
 extern RecompOpFn recomp_opcodes_dc[512];
-/*extern RecompOpFn recomp_opcodes_dd[512];
+/*extern RecompOpFn recomp_opcodes_dd[512];*/
 extern RecompOpFn recomp_opcodes_de[512];
-extern RecompOpFn recomp_opcodes_df[512];
+/*extern RecompOpFn recomp_opcodes_df[512];
 extern RecompOpFn recomp_opcodes_REPE[512];
 extern RecompOpFn recomp_opcodes_REPNE[512];*/
 
diff --git a/src/codegen_ops_fpu_arith.c b/src/codegen_ops_fpu_arith.c
index 4d2237b..507f937 100644
--- a/src/codegen_ops_fpu_arith.c
+++ b/src/codegen_ops_fpu_arith.c
@@ -5,10 +5,16 @@
 #include "386_common.h"
 #include "x87.h"
 #include "codegen.h"
+#include "codegen_accumulate.h"
 #include "codegen_ir.h"
 #include "codegen_ops.h"
 #include "codegen_ops_fpu_arith.h"
 
+static void fpu_POP(ir_data_t *ir)
+{
+        uop_ADD_IMM(ir, IREG_FPU_TOP, IREG_FPU_TOP, 1);
+}
+
 uint32_t ropFADD(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc)
 {
         int src_reg = fetchdat & 7;
@@ -29,6 +35,17 @@ uint32_t ropFADDr(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t fe
 
         return op_pc;
 }
+uint32_t ropFADDP(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc)
+{
+        int dest_reg = fetchdat & 7;
+
+        uop_FP_ENTER(ir);
+        uop_FADD(ir, IREG_ST(dest_reg), IREG_ST(dest_reg), IREG_ST(0));
+        uop_MOV_IMM(ir, IREG_tag(dest_reg), TAG_VALID);
+        fpu_POP(ir);
+        
+        return op_pc;
+}
 
 uint32_t ropFDIV(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc)
 {
@@ -40,6 +57,16 @@ uint32_t ropFDIV(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t fet
 
         return op_pc;
 }
+uint32_t ropFDIVR(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc)
+{
+        int src_reg = fetchdat & 7;
+
+        uop_FP_ENTER(ir);
+        uop_FDIV(ir, IREG_ST(0), IREG_ST(src_reg), IREG_ST(0));
+        uop_MOV_IMM(ir, IREG_tag(0), TAG_VALID);
+
+        return op_pc;
+}
 uint32_t ropFDIVr(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc)
 {
         int dest_reg = fetchdat & 7;
@@ -50,16 +77,6 @@ uint32_t ropFDIVr(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t fe
 
         return op_pc;
 }
-uint32_t ropFDIVR(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc)
-{
-        int src_reg = fetchdat & 7;
-
-        uop_FP_ENTER(ir);
-        uop_FDIV(ir, IREG_ST(0), IREG_ST(src_reg), IREG_ST(0));
-        uop_MOV_IMM(ir, IREG_tag(0), TAG_VALID);
-
-        return op_pc;
-}
 uint32_t ropFDIVRr(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc)
 {
         int dest_reg = fetchdat & 7;
@@ -70,6 +87,28 @@ uint32_t ropFDIVRr(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t f
 
         return op_pc;
 }
+uint32_t ropFDIVP(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc)
+{
+        int dest_reg = fetchdat & 7;
+
+        uop_FP_ENTER(ir);
+        uop_FDIV(ir, IREG_ST(dest_reg), IREG_ST(dest_reg), IREG_ST(0));
+        uop_MOV_IMM(ir, IREG_tag(dest_reg), TAG_VALID);
+        fpu_POP(ir);
+        
+        return op_pc;
+}
+uint32_t ropFDIVRP(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc)
+{
+        int dest_reg = fetchdat & 7;
+
+        uop_FP_ENTER(ir);
+        uop_FDIV(ir, IREG_ST(dest_reg), IREG_ST(0), IREG_ST(dest_reg));
+        uop_MOV_IMM(ir, IREG_tag(dest_reg), TAG_VALID);
+        fpu_POP(ir);
+
+        return op_pc;
+}
 
 uint32_t ropFMUL(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc)
 {
@@ -91,6 +130,17 @@ uint32_t ropFMULr(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t fe
 
         return op_pc;
 }
+uint32_t ropFMULP(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc)
+{
+        int dest_reg = fetchdat & 7;
+
+        uop_FP_ENTER(ir);
+        uop_FMUL(ir, IREG_ST(dest_reg), IREG_ST(dest_reg), IREG_ST(0));
+        uop_MOV_IMM(ir, IREG_tag(dest_reg), TAG_VALID);
+        fpu_POP(ir);
+
+        return op_pc;
+}
 
 uint32_t ropFSUB(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc)
 {
@@ -102,6 +152,16 @@ uint32_t ropFSUB(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t fet
 
         return op_pc;
 }
+uint32_t ropFSUBR(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc)
+{
+        int src_reg = fetchdat & 7;
+
+        uop_FP_ENTER(ir);
+        uop_FSUB(ir, IREG_ST(0), IREG_ST(src_reg), IREG_ST(0));
+        uop_MOV_IMM(ir, IREG_tag(0), TAG_VALID);
+
+        return op_pc;
+}
 uint32_t ropFSUBr(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc)
 {
         int dest_reg = fetchdat & 7;
@@ -112,16 +172,6 @@ uint32_t ropFSUBr(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t fe
 
         return op_pc;
 }
-uint32_t ropFSUBR(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc)
-{
-        int src_reg = fetchdat & 7;
-
-        uop_FP_ENTER(ir);
-        uop_FSUB(ir, IREG_ST(0), IREG_ST(src_reg), IREG_ST(0));
-        uop_MOV_IMM(ir, IREG_tag(0), TAG_VALID);
-
-        return op_pc;
-}
 uint32_t ropFSUBRr(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc)
 {
         int dest_reg = fetchdat & 7;
@@ -132,3 +182,25 @@ uint32_t ropFSUBRr(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t f
 
         return op_pc;
 }
+uint32_t ropFSUBP(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc)
+{
+        int dest_reg = fetchdat & 7;
+
+        uop_FP_ENTER(ir);
+        uop_FSUB(ir, IREG_ST(dest_reg), IREG_ST(dest_reg), IREG_ST(0));
+        uop_MOV_IMM(ir, IREG_tag(dest_reg), TAG_VALID);
+        fpu_POP(ir);
+
+        return op_pc;
+}
+uint32_t ropFSUBRP(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc)
+{
+        int dest_reg = fetchdat & 7;
+
+        uop_FP_ENTER(ir);
+        uop_FSUB(ir, IREG_ST(dest_reg), IREG_ST(0), IREG_ST(dest_reg));
+        uop_MOV_IMM(ir, IREG_tag(dest_reg), TAG_VALID);
+        fpu_POP(ir);
+        
+        return op_pc;
+}
diff --git a/src/codegen_ops_fpu_arith.h b/src/codegen_ops_fpu_arith.h
index 063f614..4e60233 100644
--- a/src/codegen_ops_fpu_arith.h
+++ b/src/codegen_ops_fpu_arith.h
@@ -1,12 +1,18 @@
 uint32_t ropFADD(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc);
 uint32_t ropFADDr(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc);
+uint32_t ropFADDP(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc);
 uint32_t ropFDIV(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc);
-uint32_t ropFDIVr(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc);
 uint32_t ropFDIVR(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc);
+uint32_t ropFDIVr(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc);
 uint32_t ropFDIVRr(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc);
+uint32_t ropFDIVP(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc);
+uint32_t ropFDIVRP(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc);
 uint32_t ropFMUL(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc);
 uint32_t ropFMULr(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc);
+uint32_t ropFMULP(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc);
 uint32_t ropFSUB(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc);
-uint32_t ropFSUBr(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc);
 uint32_t ropFSUBR(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc);
+uint32_t ropFSUBr(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc);
 uint32_t ropFSUBRr(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc);
+uint32_t ropFSUBP(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc);
+uint32_t ropFSUBRP(codeblock_t *block, ir_data_t *ir, uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc);
diff --git a/src/x87.c b/src/x87.c
index 625e56e..44f22e0 100644
--- a/src/x87.c
+++ b/src/x87.c
@@ -70,7 +70,7 @@ void x87_dumpregs()
         }
         else
         {
-                pclog("ST(0)=%f\tST(1)=%f\tST(2)=%f\tST(3)=%f\t\n",cpu_state.ST[cpu_state.TOP],cpu_state.ST[(cpu_state.TOP+1)&7],cpu_state.ST[(cpu_state.TOP+2)&7],cpu_state.ST[(cpu_state.TOP+3)&7]);
+                pclog("ST(0)=%f\tST(1)=%f\tST(2)=%f\tST(3)=%f\t\n",cpu_state.ST[cpu_state.TOP&7],cpu_state.ST[(cpu_state.TOP+1)&7],cpu_state.ST[(cpu_state.TOP+2)&7],cpu_state.ST[(cpu_state.TOP+3)&7]);
                 pclog("ST(4)=%f\tST(5)=%f\tST(6)=%f\tST(7)=%f\t\n",cpu_state.ST[(cpu_state.TOP+4)&7],cpu_state.ST[(cpu_state.TOP+5)&7],cpu_state.ST[(cpu_state.TOP+6)&7],cpu_state.ST[(cpu_state.TOP+7)&7]);
         }
         pclog("Status = %04X  Control = %04X  Tag = %04X\n", cpu_state.npxs, cpu_state.npxc, x87_gettag());
diff --git a/src/x87_ops.h b/src/x87_ops.h
index 3756c19..dd46f91 100644
--- a/src/x87_ops.h
+++ b/src/x87_ops.h
@@ -38,8 +38,8 @@ static inline void x87_checkexceptions()
 
 static inline void x87_push(double i)
 {
-        cpu_state.TOP=(cpu_state.TOP-1)&7;
-        cpu_state.ST[cpu_state.TOP] = i;
+        cpu_state.TOP--;
+        cpu_state.ST[cpu_state.TOP&7] = i;
         cpu_state.tag[cpu_state.TOP&7] = TAG_VALID;
 }
 
@@ -53,16 +53,16 @@ static inline void x87_push_u64(uint64_t i)
                 
         td.ll = i;
 
-        cpu_state.TOP=(cpu_state.TOP-1)&7;
-        cpu_state.ST[cpu_state.TOP] = td.d;
+        cpu_state.TOP--;
+        cpu_state.ST[cpu_state.TOP&7] = td.d;
         cpu_state.tag[cpu_state.TOP&7] = TAG_VALID;
 }
 
 static inline double x87_pop()
 {
-        double t = cpu_state.ST[cpu_state.TOP];
+        double t = cpu_state.ST[cpu_state.TOP&7];
         cpu_state.tag[cpu_state.TOP&7] = TAG_EMPTY;
-        cpu_state.TOP=(cpu_state.TOP+1)&7;
+        cpu_state.TOP++;
         return t;
 }
 
diff --git a/src/x87_ops_arith.h b/src/x87_ops_arith.h
index 4561c82..d5810c6 100644
--- a/src/x87_ops_arith.h
+++ b/src/x87_ops_arith.h
@@ -11,7 +11,7 @@ static int opFADD ## name ## _a ## a_size(uint32_t fetchdat)    \
         ST(0) += use_var;                                       \
         if ((cpu_state.npxc >> 10) & 3)                                   \
                 fesetround(FE_TONEAREST);                       \
-        cpu_state.tag[cpu_state.TOP] = TAG_VALID;               \
+        cpu_state.tag[cpu_state.TOP&7] = TAG_VALID;             \
         CLOCK_CYCLES(8);                                        \
         return 0;                                               \
 }                                                               \
@@ -48,7 +48,7 @@ static int opFDIV ## name ## _a ## a_size(uint32_t fetchdat)    \
         SEG_CHECK_READ(cpu_state.ea_seg);                       \
         load_var = get(); if (cpu_state.abrt) return 1;                   \
         x87_div(ST(0), ST(0), use_var);                         \
-        cpu_state.tag[cpu_state.TOP] = TAG_VALID;               \
+        cpu_state.tag[cpu_state.TOP&7] = TAG_VALID;             \
         CLOCK_CYCLES(73);                                       \
         return 0;                                               \
 }                                                               \
@@ -60,7 +60,7 @@ static int opFDIVR ## name ## _a ## a_size(uint32_t fetchdat)   \
         SEG_CHECK_READ(cpu_state.ea_seg);                       \
         load_var = get(); if (cpu_state.abrt) return 1;                   \
         x87_div(ST(0), use_var, ST(0));                         \
-        cpu_state.tag[cpu_state.TOP] = TAG_VALID;               \
+        cpu_state.tag[cpu_state.TOP&7] = TAG_VALID;             \
         CLOCK_CYCLES(73);                                       \
         return 0;                                               \
 }                                                               \
@@ -72,7 +72,7 @@ static int opFMUL ## name ## _a ## a_size(uint32_t fetchdat)    \
         SEG_CHECK_READ(cpu_state.ea_seg);                       \
         load_var = get(); if (cpu_state.abrt) return 1;                   \
         ST(0) *= use_var;                                       \
-        cpu_state.tag[cpu_state.TOP] = TAG_VALID;               \
+        cpu_state.tag[cpu_state.TOP&7] = TAG_VALID;             \
         CLOCK_CYCLES(11);                                       \
         return 0;                                               \
 }                                                               \
@@ -84,7 +84,7 @@ static int opFSUB ## name ## _a ## a_size(uint32_t fetchdat)    \
         SEG_CHECK_READ(cpu_state.ea_seg);                       \
         load_var = get(); if (cpu_state.abrt) return 1;                   \
         ST(0) -= use_var;                                       \
-        cpu_state.tag[cpu_state.TOP] = TAG_VALID;               \
+        cpu_state.tag[cpu_state.TOP&7] = TAG_VALID;             \
         CLOCK_CYCLES(8);                                        \
         return 0;                                               \
 }                                                               \
@@ -96,7 +96,7 @@ static int opFSUBR ## name ## _a ## a_size(uint32_t fetchdat)   \
         SEG_CHECK_READ(cpu_state.ea_seg);                       \
         load_var = get(); if (cpu_state.abrt) return 1;                   \
         ST(0) = use_var - ST(0);                                \
-        cpu_state.tag[cpu_state.TOP] = TAG_VALID;               \
+        cpu_state.tag[cpu_state.TOP&7] = TAG_VALID;             \
         CLOCK_CYCLES(8);                                        \
         return 0;                                               \
 }
@@ -121,7 +121,7 @@ static int opFADD(uint32_t fetchdat)
         cpu_state.pc++;
         if (fplog) pclog("FADD\n");
         ST(0) = ST(0) + ST(fetchdat & 7);
-        cpu_state.tag[cpu_state.TOP] = TAG_VALID;
+        cpu_state.tag[cpu_state.TOP&7] = TAG_VALID;
         CLOCK_CYCLES(8);
         return 0;
 }
@@ -232,7 +232,7 @@ static int opFDIV(uint32_t fetchdat)
         cpu_state.pc++;
         if (fplog) pclog("FDIV\n");
         x87_div(ST(0), ST(0), ST(fetchdat & 7));
-        cpu_state.tag[cpu_state.TOP] = TAG_VALID;
+        cpu_state.tag[cpu_state.TOP&7] = TAG_VALID;
         CLOCK_CYCLES(73);
         return 0;
 }
@@ -264,7 +264,7 @@ static int opFDIVR(uint32_t fetchdat)
         cpu_state.pc++;
         if (fplog) pclog("FDIVR\n");
         x87_div(ST(0), ST(fetchdat&7), ST(0));
-        cpu_state.tag[cpu_state.TOP] = TAG_VALID;
+        cpu_state.tag[cpu_state.TOP&7] = TAG_VALID;
         CLOCK_CYCLES(73);
         return 0;
 }
@@ -296,7 +296,7 @@ static int opFMUL(uint32_t fetchdat)
         cpu_state.pc++;
         if (fplog) pclog("FMUL\n");
         ST(0) = ST(0) * ST(fetchdat & 7);
-        cpu_state.tag[cpu_state.TOP] = TAG_VALID;
+        cpu_state.tag[cpu_state.TOP&7] = TAG_VALID;
         CLOCK_CYCLES(16);
         return 0;
 }
@@ -328,7 +328,7 @@ static int opFSUB(uint32_t fetchdat)
         cpu_state.pc++;
         if (fplog) pclog("FSUB\n");
         ST(0) = ST(0) - ST(fetchdat & 7);
-        cpu_state.tag[cpu_state.TOP] = TAG_VALID;
+        cpu_state.tag[cpu_state.TOP&7] = TAG_VALID;
         CLOCK_CYCLES(8);
         return 0;
 }
@@ -360,7 +360,7 @@ static int opFSUBR(uint32_t fetchdat)
         cpu_state.pc++;
         if (fplog) pclog("FSUBR\n");
         ST(0) = ST(fetchdat & 7) - ST(0);
-        cpu_state.tag[cpu_state.TOP] = TAG_VALID;
+        cpu_state.tag[cpu_state.TOP&7] = TAG_VALID;
         CLOCK_CYCLES(8);
         return 0;
 }
diff --git a/src/x87_ops_loadstore.h b/src/x87_ops_loadstore.h
index 7a0b938..96c18e6 100644
--- a/src/x87_ops_loadstore.h
+++ b/src/x87_ops_loadstore.h
@@ -95,8 +95,8 @@ static int opFILDiq_a16(uint32_t fetchdat)
         temp64 = geteaq(); if (cpu_state.abrt) return 1;
         if (fplog) pclog("  %f  %08X %08X\n", (double)temp64, readmeml(easeg,cpu_state.eaaddr), readmeml(easeg,cpu_state.eaaddr+4));
         x87_push((double)temp64);
-        cpu_state.MM[cpu_state.TOP].q = temp64;
-        cpu_state.tag[cpu_state.TOP] = TAG_VALID | TAG_UINT64;
+        cpu_state.MM[cpu_state.TOP&7].q = temp64;
+        cpu_state.tag[cpu_state.TOP&7] = TAG_VALID | TAG_UINT64;
 
         CLOCK_CYCLES(10);
         return 0;
@@ -111,8 +111,8 @@ static int opFILDiq_a32(uint32_t fetchdat)
         temp64 = geteaq(); if (cpu_state.abrt) return 1;
         if (fplog) pclog("  %f  %08X %08X\n", (double)temp64, readmeml(easeg,cpu_state.eaaddr), readmeml(easeg,cpu_state.eaaddr+4));
         x87_push((double)temp64);
-        cpu_state.MM[cpu_state.TOP].q = temp64;
-        cpu_state.tag[cpu_state.TOP] = TAG_VALID | TAG_UINT64;
+        cpu_state.MM[cpu_state.TOP&7].q = temp64;
+        cpu_state.tag[cpu_state.TOP&7] = TAG_VALID | TAG_UINT64;
 
         CLOCK_CYCLES(10);
         return 0;
@@ -180,8 +180,8 @@ static int FISTPiq_a16(uint32_t fetchdat)
         fetch_ea_16(fetchdat);
         SEG_CHECK_WRITE(cpu_state.ea_seg);
         if (fplog) pclog("FISTPl %08X:%08X\n", easeg, cpu_state.eaaddr);
-        if (cpu_state.tag[cpu_state.TOP] & TAG_UINT64)
-                temp64 = cpu_state.MM[cpu_state.TOP].q;
+        if (cpu_state.tag[cpu_state.TOP&7] & TAG_UINT64)
+                temp64 = cpu_state.MM[cpu_state.TOP&7].q;
         else
                 temp64 = x87_fround(ST(0));
         seteaq(temp64); if (cpu_state.abrt) return 1;
@@ -196,8 +196,8 @@ static int FISTPiq_a32(uint32_t fetchdat)
         fetch_ea_32(fetchdat);
         SEG_CHECK_WRITE(cpu_state.ea_seg);
         if (fplog) pclog("FISTPl %08X:%08X\n", easeg, cpu_state.eaaddr);
-        if (cpu_state.tag[cpu_state.TOP] & TAG_UINT64)
-                temp64 = cpu_state.MM[cpu_state.TOP].q;
+        if (cpu_state.tag[cpu_state.TOP&7] & TAG_UINT64)
+                temp64 = cpu_state.MM[cpu_state.TOP&7].q;
         else
                 temp64 = x87_fround(ST(0));
         seteaq(temp64); if (cpu_state.abrt) return 1;
diff --git a/src/x87_ops_misc.h b/src/x87_ops_misc.h
index 34a28d0..5ca534c 100644
--- a/src/x87_ops_misc.h
+++ b/src/x87_ops_misc.h
@@ -335,8 +335,8 @@ static int opFLD(uint32_t fetchdat)
         old_tag = cpu_state.tag[(cpu_state.TOP + fetchdat) & 7];
         old_i64 = cpu_state.MM[(cpu_state.TOP + fetchdat) & 7].q;
         x87_push(ST(fetchdat&7));
-        cpu_state.tag[cpu_state.TOP] = old_tag;
-        cpu_state.MM[cpu_state.TOP].q = old_i64;
+        cpu_state.tag[cpu_state.TOP&7] = old_tag;
+        cpu_state.MM[cpu_state.TOP&7].q = old_i64;
         CLOCK_CYCLES(4);
         return 0;
 }
@@ -352,11 +352,11 @@ static int opFXCH(uint32_t fetchdat)
         td = ST(0);
         ST(0) = ST(fetchdat&7);
         ST(fetchdat&7) = td;
-        old_tag = cpu_state.tag[cpu_state.TOP];
-        cpu_state.tag[cpu_state.TOP] = cpu_state.tag[(cpu_state.TOP + fetchdat) & 7];
+        old_tag = cpu_state.tag[cpu_state.TOP&7];
+        cpu_state.tag[cpu_state.TOP&7] = cpu_state.tag[(cpu_state.TOP + fetchdat) & 7];
         cpu_state.tag[(cpu_state.TOP + fetchdat) & 7] = old_tag;
-        old_i64 = cpu_state.MM[cpu_state.TOP].q;
-        cpu_state.MM[cpu_state.TOP].q = cpu_state.MM[(cpu_state.TOP + fetchdat) & 7].q;
+        old_i64 = cpu_state.MM[cpu_state.TOP&7].q;
+        cpu_state.MM[cpu_state.TOP&7].q = cpu_state.MM[(cpu_state.TOP + fetchdat) & 7].q;
         cpu_state.MM[(cpu_state.TOP + fetchdat) & 7].q = old_i64;
         
         CLOCK_CYCLES(4);
@@ -369,7 +369,7 @@ static int opFCHS(uint32_t fetchdat)
         cpu_state.pc++;
         if (fplog) pclog("FCHS\n");
         ST(0) = -ST(0);
-        cpu_state.tag[cpu_state.TOP] = TAG_VALID;
+        cpu_state.tag[cpu_state.TOP&7] = TAG_VALID;
         CLOCK_CYCLES(6);
         return 0;
 }
@@ -380,7 +380,7 @@ static int opFABS(uint32_t fetchdat)
         cpu_state.pc++;
         if (fplog) pclog("FABS %f\n", ST(0));
         ST(0) = fabs(ST(0));
-        cpu_state.tag[cpu_state.TOP] = TAG_VALID;
+        cpu_state.tag[cpu_state.TOP&7] = TAG_VALID;
         CLOCK_CYCLES(3);
         return 0;
 }
@@ -488,7 +488,7 @@ static int opF2XM1(uint32_t fetchdat)
         cpu_state.pc++;
         if (fplog) pclog("F2XM1\n");
         ST(0) = pow(2.0, ST(0)) - 1.0;
-        cpu_state.tag[cpu_state.TOP] = TAG_VALID;
+        cpu_state.tag[cpu_state.TOP&7] = TAG_VALID;
         CLOCK_CYCLES(200);
         return 0;
 }
@@ -523,7 +523,7 @@ static int opFPTAN(uint32_t fetchdat)
         cpu_state.pc++;
         if (fplog) pclog("FPTAN\n");
         ST(0) = tan(ST(0));
-        cpu_state.tag[cpu_state.TOP] = TAG_VALID;
+        cpu_state.tag[cpu_state.TOP&7] = TAG_VALID;
         x87_push(1.0);
         cpu_state.npxs &= ~C2;
         CLOCK_CYCLES(235);
@@ -547,7 +547,7 @@ static int opFDECSTP(uint32_t fetchdat)
         FP_ENTER();
         cpu_state.pc++;
         if (fplog) pclog("FDECSTP\n");
-        cpu_state.TOP = (cpu_state.TOP - 1) & 7;
+        cpu_state.TOP--;
         CLOCK_CYCLES(4);
         return 0;
 }
@@ -557,7 +557,7 @@ static int opFINCSTP(uint32_t fetchdat)
         FP_ENTER();
         cpu_state.pc++;
         if (fplog) pclog("FDECSTP\n");
-        cpu_state.TOP = (cpu_state.TOP + 1) & 7;
+        cpu_state.TOP++;
         CLOCK_CYCLES(4);
         return 0;
 }
@@ -570,7 +570,7 @@ static int opFPREM(uint32_t fetchdat)
         if (fplog) pclog("FPREM %f %f  ", ST(0), ST(1));
         temp64 = (int64_t)(ST(0) / ST(1));
         ST(0) = ST(0) - (ST(1) * (double)temp64);
-        cpu_state.tag[cpu_state.TOP] = TAG_VALID;
+        cpu_state.tag[cpu_state.TOP&7] = TAG_VALID;
         if (fplog) pclog("%f\n", ST(0));
         cpu_state.npxs &= ~(C0|C1|C2|C3);
         if (temp64 & 4) cpu_state.npxs|=C0;
@@ -587,7 +587,7 @@ static int opFPREM1(uint32_t fetchdat)
         if (fplog) pclog("FPREM1 %f %f  ", ST(0), ST(1));
         temp64 = (int64_t)(ST(0) / ST(1));
         ST(0) = ST(0) - (ST(1) * (double)temp64);
-        cpu_state.tag[cpu_state.TOP] = TAG_VALID;
+        cpu_state.tag[cpu_state.TOP&7] = TAG_VALID;
         if (fplog) pclog("%f\n", ST(0));
         cpu_state.npxs &= ~(C0|C1|C2|C3);
         if (temp64 & 4) cpu_state.npxs|=C0;
@@ -603,7 +603,7 @@ static int opFSQRT(uint32_t fetchdat)
         cpu_state.pc++;
         if (fplog) pclog("FSQRT\n");
         ST(0) = sqrt(ST(0));
-        cpu_state.tag[cpu_state.TOP] = TAG_VALID;
+        cpu_state.tag[cpu_state.TOP&7] = TAG_VALID;
         CLOCK_CYCLES(83);
         return 0;
 }
@@ -616,7 +616,7 @@ static int opFSINCOS(uint32_t fetchdat)
         if (fplog) pclog("FSINCOS\n");
         td = ST(0);
         ST(0) = sin(td);
-        cpu_state.tag[cpu_state.TOP] = TAG_VALID;
+        cpu_state.tag[cpu_state.TOP&7] = TAG_VALID;
         x87_push(cos(td));
         cpu_state.npxs &= ~C2;
         CLOCK_CYCLES(330);
@@ -629,7 +629,7 @@ static int opFRNDINT(uint32_t fetchdat)
         cpu_state.pc++;
         if (fplog) pclog("FRNDINT %g ", ST(0));
         ST(0) = (double)x87_fround(ST(0));
-        cpu_state.tag[cpu_state.TOP] = TAG_VALID;
+        cpu_state.tag[cpu_state.TOP&7] = TAG_VALID;
         if (fplog) pclog("%g\n", ST(0));
         CLOCK_CYCLES(21);
         return 0;
@@ -643,7 +643,7 @@ static int opFSCALE(uint32_t fetchdat)
         if (fplog) pclog("FSCALE\n");
         temp64 = (int64_t)ST(1);
         ST(0) = ST(0) * pow(2.0, (double)temp64);
-        cpu_state.tag[cpu_state.TOP] = TAG_VALID;
+        cpu_state.tag[cpu_state.TOP&7] = TAG_VALID;
         CLOCK_CYCLES(30);
         return 0;
 }
@@ -654,7 +654,7 @@ static int opFSIN(uint32_t fetchdat)
         cpu_state.pc++;
         if (fplog) pclog("FSIN\n");
         ST(0) = sin(ST(0));
-        cpu_state.tag[cpu_state.TOP] = TAG_VALID;
+        cpu_state.tag[cpu_state.TOP&7] = TAG_VALID;
         cpu_state.npxs &= ~C2;
         CLOCK_CYCLES(300);
         return 0;
@@ -666,7 +666,7 @@ static int opFCOS(uint32_t fetchdat)
         cpu_state.pc++;
         if (fplog) pclog("FCOS\n");
         ST(0) = cos(ST(0));
-        cpu_state.tag[cpu_state.TOP] = TAG_VALID;
+        cpu_state.tag[cpu_state.TOP&7] = TAG_VALID;
         cpu_state.npxs &= ~C2;
         CLOCK_CYCLES(300);
         return 0;
@@ -836,8 +836,8 @@ static int opFSTCW_a32(uint32_t fetchdat)
                 if (fplog) pclog("FCMOV %f\n", ST(fetchdat & 7));                       \
                 if (cond_ ## condition)                                                 \
                 {                                                                       \
-                        cpu_state.tag[cpu_state.TOP] = cpu_state.tag[(cpu_state.TOP + fetchdat) & 7];                           \
-                        cpu_state.MM[cpu_state.TOP].q = cpu_state.MM[(cpu_state.TOP + fetchdat) & 7].q;                     \
+                        cpu_state.tag[cpu_state.TOP&7] = cpu_state.tag[(cpu_state.TOP + fetchdat) & 7];                           \
+                        cpu_state.MM[cpu_state.TOP&7].q = cpu_state.MM[(cpu_state.TOP + fetchdat) & 7].q;                     \
                         ST(0) = ST(fetchdat & 7);                                       \
                 }                                                                       \
                 CLOCK_CYCLES(4);                                                        \