Merge pull request #350 from lioncash/qops

Implement the rest of the UQ* ops.

4 files changed, 191 insertions, 27 deletions

diff --git a/src/core/arm/dyncom/arm_dyncom_interpreter.cpp b/src/core/arm/dyncom/arm_dyncom_interpreter.cpp
index 9b355fc6e..7306794fe 100644
--- a/src/core/arm/dyncom/arm_dyncom_interpreter.cpp
+++ b/src/core/arm/dyncom/arm_dyncom_interpreter.cpp
@@ -3249,12 +3249,44 @@ ARM_INST_PTR INTERPRETER_TRANSLATE(blx_1_thumb)(unsigned int tinst, int index)
         return inst_base;
 }
 
-ARM_INST_PTR INTERPRETER_TRANSLATE(uqadd16)(unsigned int inst, int index)   { UNIMPLEMENTED_INSTRUCTION("UQADD16"); }
-ARM_INST_PTR INTERPRETER_TRANSLATE(uqadd8)(unsigned int inst, int index)    { UNIMPLEMENTED_INSTRUCTION("UQADD8"); }
-ARM_INST_PTR INTERPRETER_TRANSLATE(uqaddsubx)(unsigned int inst, int index) { UNIMPLEMENTED_INSTRUCTION("UQADDSUBX"); }
-ARM_INST_PTR INTERPRETER_TRANSLATE(uqsub16)(unsigned int inst, int index)   { UNIMPLEMENTED_INSTRUCTION("UQSUB16"); }
-ARM_INST_PTR INTERPRETER_TRANSLATE(uqsub8)(unsigned int inst, int index)    { UNIMPLEMENTED_INSTRUCTION("UQSUB8"); }
-ARM_INST_PTR INTERPRETER_TRANSLATE(uqsubaddx)(unsigned int inst, int index) { UNIMPLEMENTED_INSTRUCTION("UQSUBADDX"); }
+ARM_INST_PTR INTERPRETER_TRANSLATE(uqadd8)(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;
+
+        inst_base->cond     = BITS(inst, 28, 31);
+        inst_base->idx      = index;
+        inst_base->br       = NON_BRANCH;
+        inst_base->load_r15 = 0;
+
+        inst_cream->Rm  = BITS(inst, 0, 3);
+        inst_cream->Rn  = BITS(inst, 16, 19);
+        inst_cream->Rd  = BITS(inst, 12, 15);
+        inst_cream->op1 = BITS(inst, 20, 21);
+        inst_cream->op2 = BITS(inst, 5, 7);
+
+        return inst_base;
+}
+ARM_INST_PTR INTERPRETER_TRANSLATE(uqadd16)(unsigned int inst, int index)
+{
+        return INTERPRETER_TRANSLATE(uqadd8)(inst, index);
+}
+ARM_INST_PTR INTERPRETER_TRANSLATE(uqaddsubx)(unsigned int inst, int index)
+{
+        return INTERPRETER_TRANSLATE(uqadd8)(inst, index);
+}
+ARM_INST_PTR INTERPRETER_TRANSLATE(uqsub8)(unsigned int inst, int index)
+{
+        return INTERPRETER_TRANSLATE(uqadd8)(inst, index);
+}
+ARM_INST_PTR INTERPRETER_TRANSLATE(uqsub16)(unsigned int inst, int index)
+{
+        return INTERPRETER_TRANSLATE(uqadd8)(inst, index);
+}
+ARM_INST_PTR INTERPRETER_TRANSLATE(uqsubaddx)(unsigned int inst, int index)
+{
+        return INTERPRETER_TRANSLATE(uqadd8)(inst, index);
+}
 ARM_INST_PTR INTERPRETER_TRANSLATE(usad8)(unsigned int inst, int index)     { UNIMPLEMENTED_INSTRUCTION("USAD8"); }
 ARM_INST_PTR INTERPRETER_TRANSLATE(usada8)(unsigned int inst, int index)    { UNIMPLEMENTED_INSTRUCTION("USADA8"); }
 ARM_INST_PTR INTERPRETER_TRANSLATE(usat)(unsigned int inst, int index)      { UNIMPLEMENTED_INSTRUCTION("USAT"); }
@@ -6876,12 +6908,69 @@ unsigned InterpreterMainLoop(ARMul_State* state)
                 goto DISPATCH;
         }
 
-        UQADD16_INST:
         UQADD8_INST:
+        UQADD16_INST:
         UQADDSUBX_INST:
-        UQSUB16_INST:
         UQSUB8_INST:
+        UQSUB16_INST:
         UQSUBADDX_INST:
+        {
+                INC_ICOUNTER;
+
+                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;
+                        const u32 rm_val = RM;
+                        const u32 rn_val = RN;
+
+                        u16 lo_val = 0;
+                        u16 hi_val = 0;
+                        
+                        // UQADD16
+                        if (op2 == 0x00) {
+                                lo_val = ARMul_UnsignedSaturatedAdd16(rn_val & 0xFFFF, rm_val & 0xFFFF);
+                                hi_val = ARMul_UnsignedSaturatedAdd16((rn_val >> 16) & 0xFFFF, (rm_val >> 16) & 0xFFFF);
+                        }
+                        // UQASX
+                        else if (op2 == 0x01) {
+                                lo_val = ARMul_UnsignedSaturatedSub16(rn_val & 0xFFFF, (rm_val >> 16) & 0xFFFF);
+                                hi_val = ARMul_UnsignedSaturatedAdd16((rn_val >> 16) & 0xFFFF, rm_val & 0xFFFF);
+                        }
+                        // UQSAX
+                        else if (op2 == 0x02) {
+                                lo_val = ARMul_UnsignedSaturatedAdd16(rn_val & 0xFFFF, (rm_val >> 16) & 0xFFFF);
+                                hi_val = ARMul_UnsignedSaturatedSub16((rn_val >> 16) & 0xFFFF, rm_val & 0xFFFF);
+                        }
+                        // UQSUB16
+                        else if (op2 == 0x03) {
+                                lo_val = ARMul_UnsignedSaturatedSub16(rn_val & 0xFFFF, rm_val & 0xFFFF);
+                                hi_val = ARMul_UnsignedSaturatedSub16((rn_val >> 16) & 0xFFFF, (rm_val >> 16) & 0xFFFF);
+                        }
+                        // UQADD8
+                        else if (op2 == 0x04) {
+                                lo_val = ARMul_UnsignedSaturatedAdd8(rn_val, rm_val) |
+                                         ARMul_UnsignedSaturatedAdd8(rn_val >> 8,  rm_val >> 8) << 8;
+                                hi_val = ARMul_UnsignedSaturatedAdd8(rn_val >> 16, rm_val >> 16) |
+                                         ARMul_UnsignedSaturatedAdd8(rn_val >> 24, rm_val >> 24) << 8;
+                        }
+                        // UQSUB8
+                        else {
+                                lo_val = ARMul_UnsignedSaturatedSub8(rn_val, rm_val) |
+                                         ARMul_UnsignedSaturatedSub8(rn_val >> 8,  rm_val >> 8) << 8;
+                                hi_val = ARMul_UnsignedSaturatedSub8(rn_val >> 16, rm_val >> 16) |
+                                         ARMul_UnsignedSaturatedSub8(rn_val >> 24, rm_val >> 24) << 8;
+                        }
+                        
+                        RD = ((lo_val & 0xFFFF) | hi_val << 16);
+                }
+
+                cpu->Reg[15] += GET_INST_SIZE(cpu);
+                INC_PC(sizeof(generic_arm_inst));
+                FETCH_INST;
+                GOTO_NEXT_INST;
+        }
+
         USAD8_INST:
         USADA8_INST:
         USAT_INST:
diff --git a/src/core/arm/interpreter/armemu.cpp b/src/core/arm/interpreter/armemu.cpp
index 9b680c1e2..5d26456c7 100644
--- a/src/core/arm/interpreter/armemu.cpp
+++ b/src/core/arm/interpreter/armemu.cpp
@@ -6117,26 +6117,55 @@ L_stm_s_takeabort:
         }
             printf("Unhandled v6 insn: uasx/usax\n");
             break;
-        case 0x66:
-                        if ((instr & 0x0FF00FF0) == 0x06600FF0) { //uqsub8
-                u32 rd = (instr >> 12) & 0xF;
-                u32 rm = (instr >> 16) & 0xF;
-                u32 rn = (instr >> 0) & 0xF;
-                u32 subfrom = state->Reg[rm];
-                u32 tosub = state->Reg[rn];
-
-                u8 b1 = (u8)((u8)(subfrom)-(u8)(tosub));
-                if (b1 > (u8)(subfrom)) b1 = 0;
-                u8 b2 = (u8)((u8)(subfrom >> 8) - (u8)(tosub >> 8));
-                if (b2 > (u8)(subfrom >> 8)) b2 = 0;
-                u8 b3 = (u8)((u8)(subfrom >> 16) - (u8)(tosub >> 16));
-                if (b3 > (u8)(subfrom >> 16)) b3 = 0;
-                u8 b4 = (u8)((u8)(subfrom >> 24) - (u8)(tosub >> 24));
-                if (b4 > (u8)(subfrom >> 24)) b4 = 0;
-                state->Reg[rd] = (u32)(b1 | b2 << 8 | b3 << 16 | b4 << 24);
+        case 0x66: // UQADD16, UQASX, UQSAX, UQSUB16, UQADD8, and UQSUB8
+            {
+                const u8 rd_idx = BITS(12, 15);
+                const u8 rm_idx = BITS(0, 3);
+                const u8 rn_idx = BITS(16, 19);
+                const u8 op2    = BITS(5, 7);
+                const u32 rm_val = state->Reg[rm_idx];
+                const u32 rn_val = state->Reg[rn_idx];
+
+                u16 lo_val = 0;
+                u16 hi_val = 0;
+
+                // UQADD16
+                if (op2 == 0x00) {
+                    lo_val = ARMul_UnsignedSaturatedAdd16(rn_val & 0xFFFF, rm_val & 0xFFFF);
+                    hi_val = ARMul_UnsignedSaturatedAdd16((rn_val >> 16) & 0xFFFF, (rm_val >> 16) & 0xFFFF);
+                }
+                // UQASX
+                else if (op2 == 0x01) {
+                    lo_val = ARMul_UnsignedSaturatedSub16(rn_val & 0xFFFF, (rm_val >> 16) & 0xFFFF);
+                    hi_val = ARMul_UnsignedSaturatedAdd16((rn_val >> 16) & 0xFFFF, rm_val & 0xFFFF);
+                }
+                // UQSAX
+                else if (op2 == 0x02) {
+                    lo_val = ARMul_UnsignedSaturatedAdd16(rn_val & 0xFFFF, (rm_val >> 16) & 0xFFFF);
+                    hi_val = ARMul_UnsignedSaturatedSub16((rn_val >> 16) & 0xFFFF, rm_val & 0xFFFF);
+                }
+                // UQSUB16
+                else if (op2 == 0x03) {
+                    lo_val = ARMul_UnsignedSaturatedSub16(rn_val & 0xFFFF, rm_val & 0xFFFF);
+                    hi_val = ARMul_UnsignedSaturatedSub16((rn_val >> 16) & 0xFFFF, (rm_val >> 16) & 0xFFFF);
+                }
+                // UQADD8
+                else if (op2 == 0x04) {
+                    lo_val = ARMul_UnsignedSaturatedAdd8(rn_val, rm_val) |
+                             ARMul_UnsignedSaturatedAdd8(rn_val >> 8,  rm_val >> 8) << 8;
+                    hi_val = ARMul_UnsignedSaturatedAdd8(rn_val >> 16, rm_val >> 16) |
+                             ARMul_UnsignedSaturatedAdd8(rn_val >> 24, rm_val >> 24) << 8;
+                }
+                // UQSUB8
+                else {
+                    lo_val = ARMul_UnsignedSaturatedSub8(rn_val, rm_val) |
+                             ARMul_UnsignedSaturatedSub8(rn_val >> 8,  rm_val >> 8) << 8;
+                    hi_val = ARMul_UnsignedSaturatedSub8(rn_val >> 16, rm_val >> 16) |
+                             ARMul_UnsignedSaturatedSub8(rn_val >> 24, rm_val >> 24) << 8;
+                }
+
+                state->Reg[rd_idx] = ((lo_val & 0xFFFF) | hi_val << 16);
                 return 1;
-            } else {
-                printf ("Unhandled v6 insn: uqsub16\n");
             }
             break;
         case 0x67: // UHADD16, UHASX, UHSAX, UHSUB16, UHADD8, and UHSUB8.
diff --git a/src/core/arm/interpreter/armsupp.cpp b/src/core/arm/interpreter/armsupp.cpp
index 6774f8a74..186b1bd73 100644
--- a/src/core/arm/interpreter/armsupp.cpp
+++ b/src/core/arm/interpreter/armsupp.cpp
@@ -469,6 +469,47 @@ ARMul_SubOverflow (ARMul_State * state, ARMword a, ARMword b, ARMword result)
     ASSIGNV (SubOverflow (a, b, result));
 }
 
+/* 8-bit unsigned saturated addition */
+u8 ARMul_UnsignedSaturatedAdd8(u8 left, u8 right)
+{
+    u8 result = left + right;
+
+    if (result < left)
+        result = 0xFF;
+
+    return result;
+}
+
+/* 16-bit unsigned saturated addition */
+u16 ARMul_UnsignedSaturatedAdd16(u16 left, u16 right)
+{
+    u16 result = left + right;
+
+    if (result < left)
+        result = 0xFFFF;
+
+    return result;
+}
+
+/* 8-bit unsigned saturated subtraction */
+u8 ARMul_UnsignedSaturatedSub8(u8 left, u8 right)
+{
+    if (left <= right)
+        return 0;
+
+    return left - right;
+}
+
+/* 16-bit unsigned saturated subtraction */
+u16 ARMul_UnsignedSaturatedSub16(u16 left, u16 right)
+{
+    if (left <= right)
+        return 0;
+
+    return left - right;
+}
+
+
 /* This function does the work of generating the addresses used in an
    LDC instruction.  The code here is always post-indexed, it's up to the
    caller to get the input address correct and to handle base register
diff --git a/src/core/arm/skyeye_common/armdefs.h b/src/core/arm/skyeye_common/armdefs.h
index 34eb5aaf7..5b6e5f8cd 100644
--- a/src/core/arm/skyeye_common/armdefs.h
+++ b/src/core/arm/skyeye_common/armdefs.h
@@ -790,6 +790,11 @@ extern void ARMul_FixSPSR(ARMul_State*, ARMword, ARMword);
 extern void ARMul_ConsolePrint(ARMul_State*, const char*, ...);
 extern void ARMul_SelectProcessor(ARMul_State*, unsigned);
 
+extern u8 ARMul_UnsignedSaturatedAdd8(u8, u8);
+extern u16 ARMul_UnsignedSaturatedAdd16(u16, u16);
+extern u8 ARMul_UnsignedSaturatedSub8(u8, u8);
+extern u16 ARMul_UnsignedSaturatedSub16(u16, u16);
+
 #define DIFF_LOG 0
 #define SAVE_LOG 0