Merge pull request #378 from lioncash/s8

dyncom: Implement SADD8/SSUB8

src/core/arm/dyncom/arm_dyncom_interpreter.cpp

@@ -2171,8 +2171,7 @@ ARM_INST_PTR INTERPRETER_TRANSLATE(rsc)(unsigned int inst, int index)
     }
     return inst_base;
 }
-ARM_INST_PTR INTERPRETER_TRANSLATE(sadd8)(unsigned int inst, int index)    { UNIMPLEMENTED_INSTRUCTION("SADD8"); }
+ARM_INST_PTR INTERPRETER_TRANSLATE(sadd8)(unsigned int inst, int index)
-ARM_INST_PTR INTERPRETER_TRANSLATE(sadd16)(unsigned int inst, int index)
 {
     arm_inst* const inst_base = (arm_inst*)AllocBuffer(sizeof(arm_inst) + sizeof(generic_arm_inst));
     generic_arm_inst* const inst_cream = (generic_arm_inst*)inst_base->component;
@@ -2190,10 +2189,27 @@ ARM_INST_PTR INTERPRETER_TRANSLATE(sadd16)(unsigned int inst, int index)
 
     return inst_base;
 }
+ARM_INST_PTR INTERPRETER_TRANSLATE(sadd16)(unsigned int inst, int index)
+{
+    return INTERPRETER_TRANSLATE(sadd8)(inst, index);
+}
 ARM_INST_PTR INTERPRETER_TRANSLATE(saddsubx)(unsigned int inst, int index)
 {
-    return INTERPRETER_TRANSLATE(sadd16)(inst, index);
+    return INTERPRETER_TRANSLATE(sadd8)(inst, index);
 }
+ARM_INST_PTR INTERPRETER_TRANSLATE(ssub8)(unsigned int inst, int index)
+{
+    return INTERPRETER_TRANSLATE(sadd8)(inst, index);
+}
+ARM_INST_PTR INTERPRETER_TRANSLATE(ssub16)(unsigned int inst, int index)
+{
+    return INTERPRETER_TRANSLATE(sadd8)(inst, index);
+}
+ARM_INST_PTR INTERPRETER_TRANSLATE(ssubaddx)(unsigned int inst, int index)
+{
+    return INTERPRETER_TRANSLATE(sadd8)(inst, index);
+}
+
 ARM_INST_PTR INTERPRETER_TRANSLATE(sbc)(unsigned int inst, int index)
 {
     arm_inst *inst_base = (arm_inst *)AllocBuffer(sizeof(arm_inst) + sizeof(sbc_inst));
@@ -2408,15 +2424,7 @@ ARM_INST_PTR INTERPRETER_TRANSLATE(ssat16)(unsigned int inst, int index)
 
     return inst_base;
 }
-ARM_INST_PTR INTERPRETER_TRANSLATE(ssub8)(unsigned int inst, int index)    { UNIMPLEMENTED_INSTRUCTION("SSUB8"); }
+
-ARM_INST_PTR INTERPRETER_TRANSLATE(ssub16)(unsigned int inst, int index)
-{
-    return INTERPRETER_TRANSLATE(sadd16)(inst, index);
-}
-ARM_INST_PTR INTERPRETER_TRANSLATE(ssubaddx)(unsigned int inst, int index)
-{
-    return INTERPRETER_TRANSLATE(sadd16)(inst, index);
-}
 ARM_INST_PTR INTERPRETER_TRANSLATE(stc)(unsigned int inst, int index)
 {
     arm_inst *inst_base = (arm_inst *)AllocBuffer(sizeof(arm_inst) + sizeof(stc_inst));
@@ -5039,6 +5047,7 @@ unsigned InterpreterMainLoop(ARMul_State* state) {
     }
 
     SADD8_INST:
+    SSUB8_INST:
     SADD16_INST:
     SADDSUBX_INST:
     SSUBADDX_INST:
@@ -5046,7 +5055,9 @@ unsigned InterpreterMainLoop(ARMul_State* state) {
     {
         if (inst_base->cond == 0xE || CondPassed(cpu, inst_base->cond)) {
             generic_arm_inst* const inst_cream = (generic_arm_inst*)inst_base->component;
+            const u8 op2 = inst_cream->op2;
 
+            if (op2 == 0x00 || op2 == 0x01 || op2 == 0x02 || op2 == 0x03) {
                 const s16 rn_lo = (RN & 0xFFFF);
                 const s16 rn_hi = ((RN >> 16) & 0xFFFF);
                 const s16 rm_lo = (RM & 0xFFFF);
@@ -5061,17 +5072,17 @@ unsigned InterpreterMainLoop(ARMul_State* state) {
                     hi_result = (rn_hi + rm_hi);
                 }
                 // SASX
-            else if (inst_cream->op2 == 0x01) {
+                else if (op2 == 0x01) {
                     lo_result = (rn_lo - rm_hi);
                     hi_result = (rn_hi + rm_lo);
                 }
                 // SSAX
-            else if (inst_cream->op2 == 0x02) {
+                else if (op2 == 0x02) {
                     lo_result = (rn_lo + rm_hi);
                     hi_result = (rn_hi - rm_lo);
                 }
                 // SSUB16
-            else if (inst_cream->op2 == 0x03) {
+                else if (op2 == 0x03) {
                     lo_result = (rn_lo - rm_lo);
                     hi_result = (rn_hi - rm_hi);
                 }
@@ -5094,6 +5105,48 @@ unsigned InterpreterMainLoop(ARMul_State* state) {
                     cpu->Cpsr &= ~(1 << 19);
                 }
             }
+            else if (op2 == 0x04 || op2 == 0x07) {
+                s32 lo_val1, lo_val2;
+                s32 hi_val1, hi_val2;
+
+                // SADD8
+                if (op2 == 0x04) {
+                    lo_val1 = (s32)(s8)(RN & 0xFF) + (s32)(s8)(RM & 0xFF);
+                    lo_val2 = (s32)(s8)((RN >> 8) & 0xFF)  + (s32)(s8)((RM >> 8) & 0xFF);
+                    hi_val1 = (s32)(s8)((RN >> 16) & 0xFF) + (s32)(s8)((RM >> 16) & 0xFF);
+                    hi_val2 = (s32)(s8)((RN >> 24) & 0xFF) + (s32)(s8)((RM >> 24) & 0xFF);
+                }
+                // SSUB8
+                else {
+                    lo_val1 = (s32)(s8)(RN & 0xFF) - (s32)(s8)(RM & 0xFF);
+                    lo_val2 = (s32)(s8)((RN >> 8) & 0xFF) - (s32)(s8)((RM >> 8) & 0xFF);
+                    hi_val1 = (s32)(s8)((RN >> 16) & 0xFF) - (s32)(s8)((RM >> 16) & 0xFF);
+                    hi_val2 = (s32)(s8)((RN >> 24) & 0xFF) - (s32)(s8)((RM >> 24) & 0xFF);
+                }
+
+                RD =  ((lo_val1 & 0xFF) | ((lo_val2 & 0xFF) << 8) | ((hi_val1 & 0xFF) << 16) | ((hi_val2 & 0xFF) << 24));
+
+                if (lo_val1 >= 0)
+                    cpu->Cpsr |= (1 << 16);
+                else
+                    cpu->Cpsr &= ~(1 << 16);
+
+                if (lo_val2 >= 0)
+                    cpu->Cpsr |= (1 << 17);
+                else
+                    cpu->Cpsr &= ~(1 << 17);
+
+                if (hi_val1 >= 0)
+                    cpu->Cpsr |= (1 << 18);
+                else
+                    cpu->Cpsr &= ~(1 << 18);
+
+                if (hi_val2 >= 0)
+                    cpu->Cpsr |= (1 << 19);
+                else
+                    cpu->Cpsr &= ~(1 << 19);
+            }
+        }
 
         cpu->Reg[15] += GET_INST_SIZE(cpu);
         INC_PC(sizeof(generic_arm_inst));
@@ -5407,7 +5460,7 @@ unsigned InterpreterMainLoop(ARMul_State* state) {
         FETCH_INST;
         GOTO_NEXT_INST;
     }
-    SSUB8_INST:
+
     STC_INST:
     {
         // Instruction not implemented

src/core/arm/interpreter/armemu.cpp

@@ -5881,67 +5881,45 @@ L_stm_s_takeabort:
                 const u32 rm_val = state->Reg[rm_idx];
                 const u32 rn_val = state->Reg[rn_idx];
 
-                u8 lo_val1;
+                s32 lo_val1, lo_val2;
-                u8 lo_val2;
+                s32 hi_val1, hi_val2;
-                u8 hi_val1;
-                u8 hi_val2;
 
                 // SADD8
                 if ((instr & 0xFF0) == 0xf90) {
-                    lo_val1 = (u8)((rn_val & 0xFF) + (rm_val & 0xFF));
+                    lo_val1 = (s32)(s8)(rn_val & 0xFF) + (s32)(s8)(rm_val & 0xFF);
-                    lo_val2 = (u8)(((rn_val >> 8) & 0xFF) + ((rm_val >> 8) & 0xFF));
+                    lo_val2 = (s32)(s8)((rn_val >> 8) & 0xFF)  + (s32)(s8)((rm_val >> 8) & 0xFF);
-                    hi_val1 = (u8)(((rn_val >> 16) & 0xFF) + ((rm_val >> 16) & 0xFF));
+                    hi_val1 = (s32)(s8)((rn_val >> 16) & 0xFF) + (s32)(s8)((rm_val >> 16) & 0xFF);
-                    hi_val2 = (u8)(((rn_val >> 24) & 0xFF) + ((rm_val >> 24) & 0xFF));
+                    hi_val2 = (s32)(s8)((rn_val >> 24) & 0xFF) + (s32)(s8)((rm_val >> 24) & 0xFF);
-
-                    if (lo_val1 & 0x80)
-                        state->GEFlag |= (1 << 16);
-                    else
-                        state->GEFlag &= ~(1 << 16);
-
-                    if (lo_val2 & 0x80)
-                        state->GEFlag |= (1 << 17);
-                    else
-                        state->GEFlag &= ~(1 << 17);
-
-                    if (hi_val1 & 0x80)
-                        state->GEFlag |= (1 << 18);
-                    else
-                        state->GEFlag &= ~(1 << 18);
-
-                    if (hi_val2 & 0x80)
-                        state->GEFlag |= (1 << 19);
-                    else
-                        state->GEFlag &= ~(1 << 19);
                 }
                 // SSUB8
                 else {
-                    lo_val1 = (u8)((rn_val & 0xFF) - (rm_val & 0xFF));
+                    lo_val1 = (s32)(s8)(rn_val & 0xFF) - (s32)(s8)(rm_val & 0xFF);
-                    lo_val2 = (u8)(((rn_val >> 8) & 0xFF) - ((rm_val >> 8) & 0xFF));
+                    lo_val2 = (s32)(s8)((rn_val >> 8) & 0xFF)  - (s32)(s8)((rm_val >> 8) & 0xFF);
-                    hi_val1 = (u8)(((rn_val >> 16) & 0xFF) - ((rm_val >> 16) & 0xFF));
+                    hi_val1 = (s32)(s8)((rn_val >> 16) & 0xFF) - (s32)(s8)((rm_val >> 16) & 0xFF);
-                    hi_val2 = (u8)(((rn_val >> 24) & 0xFF) - ((rm_val >> 24) & 0xFF));
+                    hi_val2 = (s32)(s8)((rn_val >> 24) & 0xFF) - (s32)(s8)((rm_val >> 24) & 0xFF);
+                }
 
-                    if (!(lo_val1 & 0x80))
+                if (lo_val1 >= 0)
                     state->GEFlag |= (1 << 16);
                 else
                     state->GEFlag &= ~(1 << 16);
 
-                    if (!(lo_val2 & 0x80))
+                if (lo_val2 >= 0)
                     state->GEFlag |= (1 << 17);
                 else
                     state->GEFlag &= ~(1 << 17);
 
-                    if (!(hi_val1 & 0x80))
+                if (hi_val1 >= 0)
                     state->GEFlag |= (1 << 18);
                 else
                     state->GEFlag &= ~(1 << 18);
 
-                    if (!(hi_val2 & 0x80))
+                if (hi_val2 >= 0)
                     state->GEFlag |= (1 << 19);
                 else
                     state->GEFlag &= ~(1 << 19);
-                }
 
-                state->Reg[rd_idx] = (lo_val1 | lo_val2 << 8 | hi_val1 << 16 | hi_val2 << 24);
+                state->Reg[rd_idx] = ((lo_val1 & 0xFF) | ((lo_val2 & 0xFF) << 8) | ((hi_val1 & 0xFF) << 16) | ((hi_val2 & 0xFF) << 24));
                 return 1;
             }
             else {