Added code generation for register and immediate PUSH/POP, RET, relative CALL and JMP instructions.

src/386_dynarec.c

@@ -1191,7 +1191,7 @@ void exec386(int cycs)
 
                 cycdiff=0;
                 oldcyc=cycles;
-//                if (output && CACHE_ON()) pclog("Block %04x:%04x %04x:%08x %i\n", CS, pc, SS,ESP, codeblock_page_dirty[0x1de]);
+//                if (output && CACHE_ON()) pclog("Block %04x:%04x %04x:%08x\n", CS, pc, SS,ESP);
                 if (!CACHE_ON()) /*Interpret block*/
                 {
                         cpu_block_end = 0;
@@ -1265,7 +1265,7 @@ void exec386(int cycs)
                           and physical address. The physical address check will
                           also catch any page faults at this stage*/
                         valid_block = (block->pc == cs + pc) && (block->_cs == cs) &&
-                                      (block->use32 == use32) && (block->phys == phys_addr);
+                                      (block->use32 == use32) && (block->phys == phys_addr) && (block->stack32 == stack32);
 
                         if (valid_block && (block->page_mask & page->dirty_mask))
                         {

src/codegen.h

@@ -44,6 +44,7 @@ typedef struct codeblock_t
         uint32_t endpc;
         uint32_t phys, phys_2;
         uint32_t use32;
+        int stack32;
         int pnt;
         int ins;
         uint64_t page_mask, page_mask2;

src/codegen_ops.c

@@ -10,8 +10,10 @@
 #include "codegen_ops_x86.h"
 
 #include "codegen_ops_arith.h"
+#include "codegen_ops_jump.h"
 #include "codegen_ops_logic.h"
 #include "codegen_ops_mov.h"
+#include "codegen_ops_stack.h"
 #include "codegen_ops_xchg.h"
 
 RecompOpFn recomp_opcodes[512] =
@@ -24,8 +26,8 @@ RecompOpFn recomp_opcodes[512] =
 /*30*/  ropXOR_b_rmw,   ropXOR_w_rmw,   ropXOR_b_rm,    ropXOR_w_rm,    ropXOR_AL_imm,  ropXOR_AX_imm,  NULL,           NULL,           ropCMP_b_rmw,   ropCMP_w_rmw,   ropCMP_b_rm,    ropCMP_w_rm,    ropCMP_AL_imm,  ropCMP_AX_imm,  NULL,           NULL,
 
 /*40*/  ropINC_rw,      ropINC_rw,      ropINC_rw,      ropINC_rw,      ropINC_rw,      ropINC_rw,      ropINC_rw,      ropINC_rw,      ropDEC_rw,      ropDEC_rw,      ropDEC_rw,      ropDEC_rw,      ropDEC_rw,      ropDEC_rw,      ropDEC_rw,      ropDEC_rw,
-/*50*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*50*/  ropPUSH_16,     ropPUSH_16,     ropPUSH_16,     ropPUSH_16,     ropPUSH_16,     ropPUSH_16,     ropPUSH_16,     ropPUSH_16,     ropPOP_16,      ropPOP_16,      ropPOP_16,      ropPOP_16,      ropPOP_16,      ropPOP_16,      ropPOP_16,      ropPOP_16,
-/*60*/  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,           ropPUSH_imm_16, NULL,           ropPUSH_imm_b16,NULL,           NULL,           NULL,           NULL,           NULL,
 /*70*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
 
 /*80*/  rop80,          rop81_w,        rop80,          rop83_w,        ropTEST_b_rm,   ropTEST_w_rm,   ropXCHG_b,      ropXCHG_w,      ropMOV_b_r,     ropMOV_w_r,     ropMOV_r_b,     ropMOV_r_w,     NULL,           NULL,           NULL,           NULL,
@@ -33,9 +35,9 @@ RecompOpFn recomp_opcodes[512] =
 /*a0*/  ropMOV_AL_a,    ropMOV_AX_a,    ropMOV_a_AL,    ropMOV_a_AX,    NULL,           NULL,           NULL,           NULL,           ropTEST_AL_imm, ropTEST_AX_imm, NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
 /*b0*/  ropMOV_rb_imm,  ropMOV_rb_imm,  ropMOV_rb_imm,  ropMOV_rb_imm,  ropMOV_rb_imm,  ropMOV_rb_imm,  ropMOV_rb_imm,  ropMOV_rb_imm,  ropMOV_rw_imm,  ropMOV_rw_imm,  ropMOV_rw_imm,  ropMOV_rw_imm,  ropMOV_rw_imm,  ropMOV_rw_imm,  ropMOV_rw_imm,  ropMOV_rw_imm,
 
-/*c0*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           ropMOV_b_imm,   ropMOV_w_imm,   NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*c0*/  NULL,           NULL,           ropRET_imm_16,  ropRET_16,      NULL,           NULL,           ropMOV_b_imm,   ropMOV_w_imm,   NULL,           ropLEAVE_16,    NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
 /*d0*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
-/*e0*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*e0*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           ropCALL_r16,    ropJMP_r16,     NULL,           ropJMP_r8,      NULL,           NULL,           NULL,           NULL,
 /*f0*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
 
         /*32-bit data*/
@@ -46,8 +48,8 @@ RecompOpFn recomp_opcodes[512] =
 /*30*/  ropXOR_b_rmw,   ropXOR_l_rmw,   ropXOR_b_rm,    ropXOR_l_rm,    ropXOR_AL_imm,  ropXOR_EAX_imm, NULL,           NULL,           ropCMP_b_rmw,   ropCMP_l_rmw,   ropCMP_b_rm,    ropCMP_l_rm,    ropCMP_AL_imm,  ropCMP_EAX_imm, NULL,           NULL,
 
 /*40*/  ropINC_rl,      ropINC_rl,      ropINC_rl,      ropINC_rl,      ropINC_rl,      ropINC_rl,      ropINC_rl,      ropINC_rl,      ropDEC_rl,      ropDEC_rl,      ropDEC_rl,      ropDEC_rl,      ropDEC_rl,      ropDEC_rl,      ropDEC_rl,      ropDEC_rl,
-/*50*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*50*/  ropPUSH_32,     ropPUSH_32,     ropPUSH_32,     ropPUSH_32,     ropPUSH_32,     ropPUSH_32,     ropPUSH_32,     ropPUSH_32,     ropPOP_32,      ropPOP_32,      ropPOP_32,      ropPOP_32,      ropPOP_32,      ropPOP_32,      ropPOP_32,      ropPOP_32,
-/*60*/  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,           ropPUSH_imm_32, NULL,           ropPUSH_imm_b32,NULL,           NULL,           NULL,           NULL,           NULL,
 /*70*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
 
 /*80*/  rop80,          rop81_l,        rop80,          rop83_l,        ropTEST_b_rm,   ropTEST_l_rm,   ropXCHG_b,      ropXCHG_l,      ropMOV_b_r,     ropMOV_l_r,     ropMOV_r_b,     ropMOV_r_l,     NULL,           NULL,           NULL,           NULL,
@@ -55,9 +57,9 @@ RecompOpFn recomp_opcodes[512] =
 /*a0*/  ropMOV_AL_a,    ropMOV_EAX_a,   ropMOV_a_AL,    ropMOV_a_EAX,   NULL,           NULL,           NULL,           NULL,           ropTEST_AL_imm, ropTEST_EAX_imm,NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
 /*b0*/  ropMOV_rb_imm,  ropMOV_rb_imm,  ropMOV_rb_imm,  ropMOV_rb_imm,  ropMOV_rb_imm,  ropMOV_rb_imm,  ropMOV_rb_imm,  ropMOV_rb_imm,  ropMOV_rl_imm,  ropMOV_rl_imm,  ropMOV_rl_imm,  ropMOV_rl_imm,  ropMOV_rl_imm,  ropMOV_rl_imm,  ropMOV_rl_imm,  ropMOV_rl_imm,
 
-/*c0*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           ropMOV_b_imm,   ropMOV_l_imm,   NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*c0*/  NULL,           NULL,           ropRET_imm_32,  ropRET_32,      NULL,           NULL,           ropMOV_b_imm,   ropMOV_l_imm,   NULL,           ropLEAVE_32,    NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
 /*d0*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
-/*e0*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
+/*e0*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           ropCALL_r32,    ropJMP_r32,     NULL,           ropJMP_r8,      NULL,           NULL,           NULL,           NULL,
 /*f0*/  NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,           NULL,
 };
 

src/codegen_ops_jump.h

@@ -0,0 +1,29 @@
+static uint32_t ropJMP_r8(uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc, codeblock_t *block)
+{
+        uint32_t offset = fetchdat & 0xff;
+
+        if (offset & 0x80)
+                offset |= 0xffffff00;
+
+        STORE_IMM_ADDR_L((uintptr_t)&pc, op_pc+1+offset);
+        
+        return -1;
+}
+
+static uint32_t ropJMP_r16(uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc, codeblock_t *block)
+{
+        uint16_t offset = fetchdat & 0xffff;
+
+        STORE_IMM_ADDR_L((uintptr_t)&pc, (op_pc+2+offset) & 0xffff);
+        
+        return -1;
+}
+
+static uint32_t ropJMP_r32(uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc, codeblock_t *block)
+{
+        uint32_t offset = fastreadl(cs + op_pc);
+
+        STORE_IMM_ADDR_L((uintptr_t)&pc, op_pc+4+offset);
+        
+        return -1;
+}

src/codegen_ops_stack.h

@@ -0,0 +1,219 @@
+static uint32_t ropPUSH_16(uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc, codeblock_t *block)
+{
+        int host_reg;
+
+        STORE_IMM_ADDR_L((uintptr_t)&oldpc, op_old_pc);
+        LOAD_STACK_TO_EA(-2);
+        host_reg = LOAD_REG_W(opcode & 7);
+        MEM_STORE_ADDR_EA_W(&_ss, host_reg);
+        SP_MODIFY(-2);
+        
+        return op_pc;
+}
+static uint32_t ropPUSH_32(uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc, codeblock_t *block)
+{
+        int host_reg;
+
+        STORE_IMM_ADDR_L((uintptr_t)&oldpc, op_old_pc);
+        LOAD_STACK_TO_EA(-4);
+        host_reg = LOAD_REG_L(opcode & 7);
+        MEM_STORE_ADDR_EA_L(&_ss, host_reg);
+        SP_MODIFY(-4);
+
+        return op_pc;
+}
+
+static uint32_t ropPUSH_imm_16(uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc, codeblock_t *block)
+{
+        uint16_t imm = fetchdat & 0xffff;
+        int host_reg;
+
+        STORE_IMM_ADDR_L((uintptr_t)&oldpc, op_old_pc);
+        LOAD_STACK_TO_EA(-2);
+        host_reg = LOAD_REG_IMM(imm);
+        MEM_STORE_ADDR_EA_W(&_ss, host_reg);
+        SP_MODIFY(-2);
+        
+        return op_pc+2;
+}
+static uint32_t ropPUSH_imm_32(uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc, codeblock_t *block)
+{
+        uint32_t imm = fastreadl(cs + op_pc);
+        int host_reg;
+
+        STORE_IMM_ADDR_L((uintptr_t)&oldpc, op_old_pc);
+        LOAD_STACK_TO_EA(-4);
+        host_reg = LOAD_REG_IMM(imm);
+        MEM_STORE_ADDR_EA_L(&_ss, host_reg);
+        SP_MODIFY(-4);
+        
+        return op_pc+4;
+}
+
+static uint32_t ropPUSH_imm_b16(uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc, codeblock_t *block)
+{
+        uint16_t imm = fetchdat & 0xff;
+        int host_reg;
+
+        if (imm & 0x80)
+                imm |= 0xff00;
+
+        STORE_IMM_ADDR_L((uintptr_t)&oldpc, op_old_pc);
+        LOAD_STACK_TO_EA(-2);
+        host_reg = LOAD_REG_IMM(imm);
+        MEM_STORE_ADDR_EA_W(&_ss, host_reg);
+        SP_MODIFY(-2);
+        
+        return op_pc+1;
+}
+static uint32_t ropPUSH_imm_b32(uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc, codeblock_t *block)
+{
+        uint32_t imm = fetchdat & 0xff;
+        int host_reg;
+
+        if (imm & 0x80)
+                imm |= 0xffffff00;
+
+        STORE_IMM_ADDR_L((uintptr_t)&oldpc, op_old_pc);
+        LOAD_STACK_TO_EA(-4);
+        host_reg = LOAD_REG_IMM(imm);
+        MEM_STORE_ADDR_EA_L(&_ss, host_reg);
+        SP_MODIFY(-4);
+        
+        return op_pc+1;
+}
+
+static uint32_t ropPOP_16(uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc, codeblock_t *block)
+{
+        int host_reg;
+
+        STORE_IMM_ADDR_L((uintptr_t)&oldpc, op_old_pc);
+        LOAD_STACK_TO_EA(0);
+        MEM_LOAD_ADDR_EA_W(&_ss);
+        SP_MODIFY(2);
+        STORE_REG_TARGET_W_RELEASE(0, opcode & 7);
+        
+        return op_pc;
+}
+static uint32_t ropPOP_32(uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc, codeblock_t *block)
+{
+        int host_reg;
+
+        STORE_IMM_ADDR_L((uintptr_t)&oldpc, op_old_pc);
+        LOAD_STACK_TO_EA(0);
+        MEM_LOAD_ADDR_EA_L(&_ss);
+        SP_MODIFY(4);
+        STORE_REG_TARGET_L_RELEASE(0, opcode & 7);
+        
+        return op_pc;
+}
+
+static uint32_t ropRET_16(uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc, codeblock_t *block)
+{
+        int host_reg;
+
+        STORE_IMM_ADDR_L((uintptr_t)&oldpc, op_old_pc);
+        LOAD_STACK_TO_EA(0);
+        MEM_LOAD_ADDR_EA_W(&_ss);
+        STORE_HOST_REG_ADDR(&pc, 0);
+        SP_MODIFY(2);
+        
+        return -1;
+}
+static uint32_t ropRET_32(uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc, codeblock_t *block)
+{
+        int host_reg;
+
+        STORE_IMM_ADDR_L((uintptr_t)&oldpc, op_old_pc);
+        LOAD_STACK_TO_EA(0);
+        MEM_LOAD_ADDR_EA_L(&_ss);
+        STORE_HOST_REG_ADDR(&pc, 0);
+        SP_MODIFY(4);
+        
+        return -1;
+}
+
+static uint32_t ropRET_imm_16(uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc, codeblock_t *block)
+{
+        uint16_t offset = fetchdat & 0xffff;
+        int host_reg;
+
+        STORE_IMM_ADDR_L((uintptr_t)&oldpc, op_old_pc);
+        LOAD_STACK_TO_EA(0);
+        MEM_LOAD_ADDR_EA_W(&_ss);
+        STORE_HOST_REG_ADDR(&pc, 0);
+        SP_MODIFY(2+offset);
+        
+        return -1;
+}
+static uint32_t ropRET_imm_32(uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc, codeblock_t *block)
+{
+        uint16_t offset = fetchdat & 0xffff;
+        int host_reg;
+
+        STORE_IMM_ADDR_L((uintptr_t)&oldpc, op_old_pc);
+        LOAD_STACK_TO_EA(0);
+        MEM_LOAD_ADDR_EA_L(&_ss);
+        STORE_HOST_REG_ADDR(&pc, 0);
+        SP_MODIFY(4+offset);
+        
+        return -1;
+}
+
+static uint32_t ropCALL_r16(uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc, codeblock_t *block)
+{
+        uint16_t offset = fetchdat & 0xffff;
+        int host_reg;
+
+        STORE_IMM_ADDR_L((uintptr_t)&oldpc, op_old_pc);
+        LOAD_STACK_TO_EA(-2);
+        host_reg = LOAD_REG_IMM(op_pc+2);
+        MEM_STORE_ADDR_EA_W(&_ss, host_reg);
+        SP_MODIFY(-2);
+        STORE_IMM_ADDR_L(&pc, (op_pc+2+offset) & 0xffff);
+        
+        return -1;
+}
+static uint32_t ropCALL_r32(uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc, codeblock_t *block)
+{
+        uint32_t offset = fastreadl(cs + op_pc);
+        int host_reg;
+
+        STORE_IMM_ADDR_L((uintptr_t)&oldpc, op_old_pc);
+        LOAD_STACK_TO_EA(-4);
+        host_reg = LOAD_REG_IMM(op_pc+4);
+        MEM_STORE_ADDR_EA_L(&_ss, host_reg);
+        SP_MODIFY(-4);
+        STORE_IMM_ADDR_L(&pc, op_pc+4+offset);
+        
+        return -1;
+}
+
+static uint32_t ropLEAVE_16(uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc, codeblock_t *block)
+{
+        int host_reg;
+
+        STORE_IMM_ADDR_L((uintptr_t)&oldpc, op_old_pc);
+        LOAD_EBP_TO_EA(0);
+        MEM_LOAD_ADDR_EA_W(&_ss);
+        host_reg = LOAD_REG_W(REG_BP); /*SP = BP + 2*/
+        STORE_REG_TARGET_W_RELEASE(0, REG_BP); /*BP = POP_W()*/
+        ADD_HOST_REG_IMM_W(host_reg, 2);
+        STORE_REG_TARGET_W_RELEASE(host_reg, REG_SP);
+
+        return op_pc;
+}
+static uint32_t ropLEAVE_32(uint8_t opcode, uint32_t fetchdat, uint32_t op_32, uint32_t op_pc, codeblock_t *block)
+{
+        int host_reg;
+
+        STORE_IMM_ADDR_L((uintptr_t)&oldpc, op_old_pc);
+        LOAD_EBP_TO_EA(0);
+        MEM_LOAD_ADDR_EA_L(&_ss);
+        host_reg = LOAD_REG_L(REG_EBP); /*ESP = EBP + 4*/
+        STORE_REG_TARGET_L_RELEASE(0, REG_EBP); /*EBP = POP_L()*/
+        ADD_HOST_REG_IMM(host_reg, 4);
+        STORE_REG_TARGET_L_RELEASE(host_reg, REG_ESP);
+
+        return op_pc;
+}

src/codegen_ops_x86.h

@@ -915,3 +915,102 @@ static x86seg *FETCH_EA(x86seg *op_ea_seg, uint32_t fetchdat, int op_ssegs, uint
                 return FETCH_EA_32(op_ea_seg, fetchdat, op_ssegs, op_pc, 0);
         return FETCH_EA_16(op_ea_seg, fetchdat, op_ssegs, op_pc);
 }
+
+
+static void LOAD_STACK_TO_EA(int off)
+{
+        if (stack32)
+        {
+                addbyte(0x8b); /*MOVL EAX,[ESP]*/
+                addbyte(0x05 | (REG_EAX << 3));
+                addlong((uint32_t)&ESP);
+                if (off)
+                {
+                        addbyte(0x83); /*ADD EAX, off*/
+                        addbyte(0xc0 | (0 << 3) | REG_EAX);
+                        addbyte(off);
+                }
+        }
+        else
+        {
+                addbyte(0x0f); /*MOVZX EAX,W[ESP]*/
+                addbyte(0xb7);
+                addbyte(0x05 | (REG_EAX << 3));
+                addlong((uint32_t)&ESP);
+                if (off)
+                {
+                        addbyte(0x66); /*ADD AX, off*/
+                        addbyte(0x05);
+                        addword(off);
+                }
+        }
+}
+
+static void LOAD_EBP_TO_EA(int off)
+{
+        if (stack32)
+        {
+                addbyte(0x8b); /*MOVL EAX,[EBP]*/
+                addbyte(0x05 | (REG_EAX << 3));
+                addlong((uint32_t)&EBP);
+                if (off)
+                {
+                        addbyte(0x83); /*ADD EAX, off*/
+                        addbyte(0xc0 | (0 << 3) | REG_EAX);
+                        addbyte(off);
+                }
+        }
+        else
+        {
+                addbyte(0x0f); /*MOVZX EAX,W[EBP]*/
+                addbyte(0xb7);
+                addbyte(0x05 | (REG_EAX << 3));
+                addlong((uint32_t)&EBP);
+                if (off)
+                {
+                        addbyte(0x66); /*ADD AX, off*/
+                        addbyte(0x05);
+                        addword(off);
+                }
+        }
+}
+
+static void SP_MODIFY(int off)
+{
+        if (stack32)
+        {
+                if (off < 0x80)
+                {
+                        addbyte(0x83); /*ADD [ESP], off*/
+                        addbyte(0x05);
+                        addlong((uint32_t)&ESP);
+                        addbyte(off);
+                }
+                else
+                {
+                        addbyte(0x81); /*ADD [ESP], off*/
+                        addbyte(0x05);
+                        addlong((uint32_t)&ESP);
+                        addlong(off);
+                }
+        }
+        else
+        {
+                if (off < 0x80)
+                {
+                        addbyte(0x66); /*ADD [SP], off*/
+                        addbyte(0x83);
+                        addbyte(0x05);
+                        addlong((uint32_t)&SP);
+                        addbyte(off);
+                }
+                else
+                {
+                        addbyte(0x66); /*ADD [SP], off*/
+                        addbyte(0x81);
+                        addbyte(0x05);
+                        addlong((uint32_t)&SP);
+                        addword(off);
+                }
+        }
+}

src/codegen_x86.c

@@ -189,6 +189,7 @@ void codegen_block_init(uint32_t phys_addr)
         block->pnt = block_current;
         block->phys = phys_addr;
         block->use32 = use32;
+        block->stack32 = stack32;
         block->next = block->prev = NULL;
         
         block_pos = BLOCK_GPF_OFFSET;
@@ -803,7 +804,8 @@ generate_call:
                 uint32_t new_pc = recomp_opcodes[(opcode | op_32) & 0x1ff](opcode, fetchdat, op_32, op_pc, block);
                 if (new_pc)
                 {
-                        STORE_IMM_ADDR_L((uintptr_t)&pc, new_pc);
+                        if (new_pc != -1)
+                                STORE_IMM_ADDR_L((uintptr_t)&pc, new_pc);
 
                         codegen_block_ins++;
                         block->ins++;

src/x86_ops_mov_seg.h

@@ -154,7 +154,8 @@ static int opMOV_seg_w_a16(uint32_t fetchdat)
                 fetchdat = fastreadl(cs + pc);
                 pc++;
                 if (abrt) return 1;
-                return x86_opcodes[(fetchdat & 0xff) | op32](fetchdat >> 8);
+                x86_opcodes[(fetchdat & 0xff) | op32](fetchdat >> 8);
+                return 1;
                 case 0x20: /*FS*/
                 loadseg(new_seg, &_fs);
                 break;
@@ -192,7 +193,8 @@ static int opMOV_seg_w_a32(uint32_t fetchdat)
                 fetchdat = fastreadl(cs + pc);
                 pc++;
                 if (abrt) return 1;
-                return x86_opcodes[(fetchdat & 0xff) | op32](fetchdat >> 8);
+                x86_opcodes[(fetchdat & 0xff) | op32](fetchdat >> 8);
+                return 1;
                 case 0x20: /*FS*/
                 loadseg(new_seg, &_fs);
                 break;
@@ -277,7 +279,7 @@ static int opLSS_w_a16(uint32_t fetchdat)
         regs[reg].w = addr;
  
         CLOCK_CYCLES(7);
-        return 0;
+        return 1;
 }
 static int opLSS_w_a32(uint32_t fetchdat)
 {
@@ -291,7 +293,7 @@ static int opLSS_w_a32(uint32_t fetchdat)
         regs[reg].w = addr;
  
         CLOCK_CYCLES(7);
-        return 0;
+        return 1;
 }
 static int opLSS_l_a16(uint32_t fetchdat)
 {
@@ -306,7 +308,7 @@ static int opLSS_l_a16(uint32_t fetchdat)
         regs[reg].l = addr;
  
         CLOCK_CYCLES(7);
-        return 0;
+        return 1;
 }
 static int opLSS_l_a32(uint32_t fetchdat)
 {
@@ -321,7 +323,7 @@ static int opLSS_l_a32(uint32_t fetchdat)
         regs[reg].l = addr;
  
         CLOCK_CYCLES(7);
-        return 0;
+        return 1;
 }
 
 #define opLsel(name, sel)                                                       \