3 files changed, 36 insertions, 21 deletions

diff --git a/src/core/arm/interpreter/armemu.cpp b/src/core/arm/interpreter/armemu.cpp
index 578d71380..b2f671f94 100644
--- a/src/core/arm/interpreter/armemu.cpp
+++ b/src/core/arm/interpreter/armemu.cpp
@@ -1670,7 +1670,7 @@ mainswitch:
                             op1 *= op2;
                             //printf("SMLA_INST:BB,op1=0x%x, op2=0x%x. Rn=0x%x\n", op1, op2, Rn);
                             if (AddOverflow(op1, Rn, op1 + Rn))
-                                SETS;
+                                SETQ;
                             state->Reg[BITS (16, 19)] = op1 + Rn;
                             break;
                         }
@@ -1682,7 +1682,7 @@ mainswitch:
                             ARMword result = op1 + op2;
                             if (AddOverflow(op1, op2, result)) {
                                 result = POS (result) ? 0x80000000 : 0x7fffffff;
-                                SETS;
+                                SETQ;
                             }
                             state->Reg[BITS (12, 15)] = result;
                             break;
@@ -1795,7 +1795,7 @@ mainswitch:
                                 ARMword Rn = state->Reg[BITS(12, 15)];
 
                                 if (AddOverflow((ARMword)result, Rn, (ARMword)(result + Rn)))
-                                    SETS;
+                                    SETQ;
                                 result += Rn;
                             }
                             state->Reg[BITS (16, 19)] = (ARMword)result;
@@ -1811,7 +1811,7 @@ mainswitch:
                             if (SubOverflow
                                     (op1, op2, result)) {
                                 result = POS (result) ? 0x80000000 : 0x7fffffff;
-                                SETS;
+                                SETQ;
                             }
 
                             state->Reg[BITS (12, 15)] = result;
@@ -1934,13 +1934,13 @@ mainswitch:
 
                             if (AddOverflow
                                     (op2, op2, op2d)) {
-                                SETS;
+                                SETQ;
                                 op2d = POS (op2d) ? 0x80000000 : 0x7fffffff;
                             }
 
                             result = op1 + op2d;
                             if (AddOverflow(op1, op2d, result)) {
-                                SETS;
+                                SETQ;
                                 result = POS (result) ? 0x80000000 : 0x7fffffff;
                             }
 
@@ -2053,13 +2053,13 @@ mainswitch:
                             ARMword result;
 
                             if (AddOverflow(op2, op2, op2d)) {
-                                SETS;
+                                SETQ;
                                 op2d = POS (op2d) ? 0x80000000 : 0x7fffffff;
                             }
 
                             result = op1 - op2d;
                             if (SubOverflow(op1, op2d, result)) {
-                                SETS;
+                                SETQ;
                                 result = POS (result) ? 0x80000000 : 0x7fffffff;
                             }
 
@@ -6478,22 +6478,28 @@ L_stm_s_takeabort:
                 const s16 rn_lo = (rn_val & 0xFFFF);
                 const s16 rn_hi = ((rn_val >> 16) & 0xFFFF);
 
-                // SMUAD
-                if ((instr & 0xf0d0) == 0xf010) {
-                    state->Reg[rd_idx] = (rn_lo * rm_lo) + (rn_hi * rm_hi);
-                }
-                // SMUSD
-                else if ((instr & 0xf0d0) == 0xf050) {
-                    state->Reg[rd_idx] = (rn_lo * rm_lo) - (rn_hi * rm_hi);
-                }
-                // SMLAD
-                else if ((instr & 0xd0) == 0x10) {
-                    state->Reg[rd_idx] = (rn_lo * rm_lo) + (rn_hi * rm_hi) + (s32)state->Reg[ra_idx];
+                const u32 product1 = (rn_lo * rm_lo);
+                const u32 product2 = (rn_hi * rm_hi);
+
+                // SMUAD and SMLAD
+                if (BIT(6) == 0) {
+                    state->Reg[rd_idx] = product1 + product2;
+
+                    if (BITS(12, 15) != 15) {
+                        state->Reg[rd_idx] += state->Reg[ra_idx];
+                        ARMul_AddOverflowQ(state, product1 + product2, state->Reg[ra_idx]);
+                    }
+
+                    ARMul_AddOverflowQ(state, product1, product2);
                 }
-                // SMLSD
+                // SMUSD and SMLSD
                 else {
-                    state->Reg[rd_idx] = ((rn_lo * rm_lo) - (rn_hi * rm_hi)) + (s32)state->Reg[ra_idx];
+                    state->Reg[rd_idx] = product1 - product2;
+                    
+                    if (BITS(12, 15) != 15)
+                        state->Reg[rd_idx] += state->Reg[ra_idx];
                 }
+
                 return 1;
             }
             break;
diff --git a/src/core/arm/interpreter/armsupp.cpp b/src/core/arm/interpreter/armsupp.cpp
index b31c0ea24..6774f8a74 100644
--- a/src/core/arm/interpreter/armsupp.cpp
+++ b/src/core/arm/interpreter/armsupp.cpp
@@ -444,6 +444,14 @@ ARMul_AddOverflow (ARMul_State * state, ARMword a, ARMword b, ARMword result)
     ASSIGNV (AddOverflow (a, b, result));
 }
 
+/* Assigns the Q flag if the given result is considered an overflow from the addition of a and b  */
+void ARMul_AddOverflowQ(ARMul_State* state, ARMword a, ARMword b)
+{
+    u32 result = a + b;
+    if (((result ^ a) & (u32)0x80000000) && ((a ^ b) & (u32)0x80000000) == 0)
+        SETQ;
+}
+
 /* Assigns the C flag after an subtraction of a and b to give result.  */
 
 void
diff --git a/src/core/arm/skyeye_common/armemu.h b/src/core/arm/skyeye_common/armemu.h
index e1b286f0f..3ea14b5a3 100644
--- a/src/core/arm/skyeye_common/armemu.h
+++ b/src/core/arm/skyeye_common/armemu.h
@@ -602,6 +602,7 @@ extern ARMword ARMul_SwitchMode (ARMul_State *, ARMword, ARMword);
 extern void ARMul_MSRCpsr (ARMul_State *, ARMword, ARMword);
 extern void ARMul_SubOverflow (ARMul_State *, ARMword, ARMword, ARMword);
 extern void ARMul_AddOverflow (ARMul_State *, ARMword, ARMword, ARMword);
+extern void ARMul_AddOverflowQ(ARMul_State*, ARMword, ARMword);
 extern void ARMul_SubCarry (ARMul_State *, ARMword, ARMword, ARMword);
 extern void ARMul_AddCarry (ARMul_State *, ARMword, ARMword, ARMword);
 extern tdstate ARMul_ThumbDecode (ARMul_State *, ARMword, ARMword, ARMword *);