compile UMULLs and some fixes

src/ARMJIT_x64/ARMJIT_ALU.cpp

@@ -301,10 +301,11 @@ void Compiler::A_Comp_MUL_MLA()
     Comp_MulOp(S, add, rd, rm, rs, rn);
 }
 
-void Compiler::A_Comp_SMULL_SMLAL()
+void Compiler::A_Comp_Mul_Long()
 {
     bool S = CurInstr.Instr & (1 << 20);
     bool add = CurInstr.Instr & (1 << 21);
+    bool sign = CurInstr.Instr & (1 << 22);
     OpArg rd = MapReg(CurInstr.A_Reg(16));
     OpArg rm = MapReg(CurInstr.A_Reg(0));
     OpArg rs = MapReg(CurInstr.A_Reg(8));
@@ -318,18 +319,34 @@ void Compiler::A_Comp_SMULL_SMLAL()
         MOV(32, R(RSCRATCH3), rs);
         TEST(32, R(RSCRATCH3), R(RSCRATCH3));
         FixupBranch zeroBSR = J_CC(CC_Z);
+        if (sign)
+        {
             BSR(32, RSCRATCH2, R(RSCRATCH3));
             NOT(32, R(RSCRATCH3));
             BSR(32, RSCRATCH, R(RSCRATCH3));
             CMP(32, R(RSCRATCH2), R(RSCRATCH));
             CMOVcc(32, RSCRATCH, R(RSCRATCH2), CC_L);
+        }
+        else
+        {
+            BSR(32, RSCRATCH, R(RSCRATCH3));
+        }
+        
         SHR(32, R(RSCRATCH), Imm8(3));
         SetJumpTarget(zeroBSR); // fortunately that's even right
         Comp_AddCycles_CI(RSCRATCH, 2);
     }
 
+    if (sign)
+    {
         MOVSX(64, 32, RSCRATCH2, rm);
         MOVSX(64, 32, RSCRATCH3, rs);
+    }
+    else
+    {
+        MOV(32, R(RSCRATCH2), rm);
+        MOV(32, R(RSCRATCH3), rs);
+    }
     if (add)
     {
         MOV(32, R(RSCRATCH), rd);

src/ARMJIT_x64/ARMJIT_Compiler.cpp

@@ -300,7 +300,7 @@ const Compiler::CompileFunc A_Comp[ARMInstrInfo::ak_Count] =
     // CMN
     F(A_Comp_CmpOp), F(A_Comp_CmpOp), F(A_Comp_CmpOp), F(A_Comp_CmpOp), F(A_Comp_CmpOp), F(A_Comp_CmpOp), F(A_Comp_CmpOp), F(A_Comp_CmpOp), F(A_Comp_CmpOp),
     // Mul
-    F(A_Comp_MUL_MLA), F(A_Comp_MUL_MLA), NULL, NULL, NULL, F(A_Comp_SMULL_SMLAL), NULL, NULL, NULL, NULL, NULL,
+    F(A_Comp_MUL_MLA), F(A_Comp_MUL_MLA), F(A_Comp_Mul_Long), F(A_Comp_Mul_Long), F(A_Comp_Mul_Long), F(A_Comp_Mul_Long), NULL, NULL, NULL, NULL, NULL,
     // ARMv5 stuff
     F(A_Comp_CLZ), NULL, NULL, NULL, NULL,
     // STR
@@ -628,7 +628,7 @@ void Compiler::Comp_AddCycles_CI(Gen::X64Reg i, int add)
     }
     else
     {
-        ConstantCycles += i + cycles;
+        ConstantCycles += cycles;
         SUB(32, MDisp(RCPU, offsetof(ARM, Cycles)), R(i));
     }
 }

src/ARMJIT_x64/ARMJIT_Compiler.h

@@ -89,7 +89,7 @@ public:
     void A_Comp_CmpOp();
 
     void A_Comp_MUL_MLA();
-    void A_Comp_SMULL_SMLAL();
+    void A_Comp_Mul_Long();
 
     void A_Comp_CLZ();
     

src/ARMJIT_x64/ARMJIT_LoadStore.cpp

@@ -423,6 +423,7 @@ void Compiler::Comp_MemAccess(int rd, int rn, const ComplexOperand& op2, int siz
                 
                 if (flags & memop_SubtractOffset)
                 {
+                    if (R(finalAddr) != rnMapped)
                         MOV(32, R(finalAddr), rnMapped);
                     if (!offset.IsZero())
                         SUB(32, R(finalAddr), offset);