Merge pull request #354 from lioncash/usaduflow

armemu: Fix underflows in USAD8/USADA8

src/core/arm/interpreter/armemu.cpp

@@ -6672,10 +6672,10 @@ L_stm_s_takeabort:
                 const u32 rm_val = state->Reg[rm_idx];
                 const u32 rn_val = state->Reg[rn_idx];
 
-                const u8 diff1 = (u8)std::labs((rn_val & 0xFF) - (rm_val & 0xFF));
-                const u8 diff2 = (u8)std::labs(((rn_val >> 8) & 0xFF) - ((rm_val >> 8) & 0xFF));
-                const u8 diff3 = (u8)std::labs(((rn_val >> 16) & 0xFF) - ((rm_val >> 16) & 0xFF));
-                const u8 diff4 = (u8)std::labs(((rn_val >> 24) & 0xFF) - ((rm_val >> 24) & 0xFF));
+                const u8 diff1 = ARMul_UnsignedAbsoluteDifference(rn_val & 0xFF, rm_val & 0xFF);
+                const u8 diff2 = ARMul_UnsignedAbsoluteDifference((rn_val >> 8) & 0xFF, (rm_val >> 8) & 0xFF);
+                const u8 diff3 = ARMul_UnsignedAbsoluteDifference((rn_val >> 16) & 0xFF, (rm_val >> 16) & 0xFF);
+                const u8 diff4 = ARMul_UnsignedAbsoluteDifference((rn_val >> 24) & 0xFF, (rm_val >> 24) & 0xFF);
 
                 u32 finalDif = (diff1 + diff2 + diff3 + diff4);
 

src/core/arm/interpreter/armsupp.cpp

@@ -392,6 +392,15 @@ ARMul_NthReg (ARMword instr, unsigned number)
     return (bit - 1);
 }
 
+/* Unsigned sum of absolute difference */
+u8 ARMul_UnsignedAbsoluteDifference(u8 left, u8 right)
+{
+	if (left > right)
+		return left - right;
+
+	return right - left;
+}
+
 /* Assigns the N and Z flags depending on the value of result.  */
 
 void

src/core/arm/skyeye_common/armemu.h

@@ -600,6 +600,7 @@ extern ARMword ARMul_SwitchMode (ARMul_State *, ARMword, ARMword);
 extern ARMword ARMul_Align (ARMul_State *, ARMword, ARMword);
 extern ARMword ARMul_SwitchMode (ARMul_State *, ARMword, ARMword);
 extern void ARMul_MSRCpsr (ARMul_State *, ARMword, ARMword);
+extern u8 ARMul_UnsignedAbsoluteDifference(u8, u8);
 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);