From eff1cea8904406d8bfbb9d9611f3b9e0d07a6837 Mon Sep 17 00:00:00 2001
From: Yang Liu <liuyang22@iscas.ac.cn>
Date: Thu, 6 Apr 2023 00:28:33 +0800
Subject: [RV64_DYNAREC] Added more opcodes from SSE tests (#669)

* [RV64_DYNAREC] Added 66 0F 14 UNPCKLPD opcode

* [RV64_DYNAREC] Added 66 0F 57 XORPD opcode

* [RV64_DYNAREC] Added 66 0F 54 ANDPD opcode

* [RV64_DYNAREC] Added 66 0F 62 PUNPCKLDQ opcode

* [RV64_DYNAREC] Added 66 0F FD PADDW opcode

* [RV64_DYNAREC] Added 66 0F D4 PADDQ opcode

* [RV64_DYNAREC] Added 66 0F 55 ANDNPD opcode
---
 src/dynarec/rv64/dynarec_rv64_660f.c   | 61 ++++++++++++++++++++++++++++++++--
 src/dynarec/rv64/dynarec_rv64_helper.h | 19 ++++++++++-
 2 files changed, 77 insertions(+), 3 deletions(-)

(limited to 'src')

diff --git a/src/dynarec/rv64/dynarec_rv64_660f.c b/src/dynarec/rv64/dynarec_rv64_660f.c
index bac50efc..05b13987 100644
--- a/src/dynarec/rv64/dynarec_rv64_660f.c
+++ b/src/dynarec/rv64/dynarec_rv64_660f.c
@@ -49,7 +49,13 @@ uintptr_t dynarec64_660F(dynarec_rv64_t* dyn, uintptr_t addr, uintptr_t ip, int
     MAYUSE(j64);
     
     switch(opcode) {
-
+        case 0x14:
+            INST_NAME("UNPCKLPD Gx, Ex");
+            nextop = F8;
+            GETEXSD(d0, 0);
+            GETGX(x3);
+            FSD(d0, x3, 8);
+            break;
         case 0x1F:
             INST_NAME("NOP (multibyte)");
             nextop = F8;
@@ -121,6 +127,42 @@ uintptr_t dynarec64_660F(dynarec_rv64_t* dyn, uintptr_t addr, uintptr_t ip, int
                 ORI(xFlags, xFlags, 1<<F_ZF);
             }
             break;
+        case 0x54:
+            INST_NAME("ANDPD Gx, Ex");
+            nextop = F8;
+            GETEX(x1, 0);
+            GETGX(x2);
+            SSE_LOOP_Q(x3, x4, AND(x3, x3, x4));
+            break;
+        case 0x55:
+            INST_NAME("ANDNPD Gx, Ex");
+            nextop = F8;
+            GETEX(x1, 0);
+            GETGX(x2);
+            SSE_LOOP_Q(x3, x4, NOT(x3, x3); AND(x3, x3, x4));
+            break;
+        case 0x57:
+            INST_NAME("XORPD Gx, Ex");
+            nextop = F8;
+            GETEX(x1, 0);
+            GETGX(x2);
+            SSE_LOOP_Q(x3, x4, XOR(x3, x3, x4));
+            break;
+        case 0x62:
+            INST_NAME("PUNPCKLDQ Gx,Ex");
+            nextop = F8;
+            GETEX(x1, 0);
+            GETGX(x2);
+            // GX->ud[3] = EX->ud[1];
+            LWU(x3, x1, fixedaddress+1*4);
+            SW(x3, x2, 3*4);
+            // GX->ud[2] = GX->ud[1];
+            LWU(x3, x2, 1*4);
+            SW(x3, x2, 2*4);
+            // GX->ud[1] = EX->ud[0];
+            LWU(x3, x1, fixedaddress+0*4);
+            SW(x3, x2, 1*4);
+            break;
         case 0x6C:
             INST_NAME("PUNPCKLQDQ Gx,Ex");
             nextop = F8;
@@ -308,6 +350,13 @@ uintptr_t dynarec64_660F(dynarec_rv64_t* dyn, uintptr_t addr, uintptr_t ip, int
             AND(x1, x1, x5);
             OR(gd, gd, x1);
             break;
+        case 0xD4:
+            INST_NAME("PADDQ Gx,Ex");
+            nextop = F8;
+            GETGX(x1);
+            GETEX(x2, 0);
+            SSE_LOOP_Q(x3, x4, ADD(x3, x3, x4));
+            break;
         case 0xDB:
             INST_NAME("PAND Gx,Ex");
             nextop = F8;
@@ -343,12 +392,20 @@ uintptr_t dynarec64_660F(dynarec_rv64_t* dyn, uintptr_t addr, uintptr_t ip, int
                 SSE_LOOP_Q(x3, x4, XOR(x3, x3, x4));
             }
             break;
+        case 0xFD:
+            INST_NAME("PADDW Gx,Ex");
+            nextop = F8;
+            nextop = F8;
+            GETGX(x1);
+            GETEX(x2, 0);
+            SSE_LOOP_WQ(x3, x4, ADDW(x3, x3, x4));
+            break;
         case 0xFE:
             INST_NAME("PADDD Gx,Ex");
             nextop = F8;
             GETGX(x1);
             GETEX(x2, 0);
-            SSE_LOOP_D(x3, x4, ADDW(x3, x3, x4));
+            SSE_LOOP_DQ(x3, x4, ADDW(x3, x3, x4));
             break;
         default:
             DEFAULT;
diff --git a/src/dynarec/rv64/dynarec_rv64_helper.h b/src/dynarec/rv64/dynarec_rv64_helper.h
index 4fb77d50..565fe018 100644
--- a/src/dynarec/rv64/dynarec_rv64_helper.h
+++ b/src/dynarec/rv64/dynarec_rv64_helper.h
@@ -335,12 +335,29 @@
     F;                                  \
     SW(GX1, gback, i*4);
 
+#define SSE_LOOP_W_ITEM(GX1, EX1, F, i) \
+    LHU(GX1, gback, i*2);               \
+    LHU(EX1, wback, fixedaddress+i*2);  \
+    F;                                  \
+    SH(GX1, gback, i*2);
+
 // Loop for SSE opcode that use 32bits value and write to GX.
 #define SSE_LOOP_D(GX1, EX1, F)     \
     SSE_LOOP_D_ITEM(GX1, EX1, F, 0) \
     SSE_LOOP_D_ITEM(GX1, EX1, F, 1) \
     SSE_LOOP_D_ITEM(GX1, EX1, F, 2) \
-    SSE_LOOP_D_ITEM(GX1, EX1, F, 3) \
+    SSE_LOOP_D_ITEM(GX1, EX1, F, 3)
+
+#define SSE_LOOP_DQ(GX1, EX1, F)    \
+    SSE_LOOP_D_ITEM(GX1, EX1, F, 0) \
+    SSE_LOOP_D_ITEM(GX1, EX1, F, 1)
+
+#define SSE_LOOP_WQ(GX1, EX1, F)    \
+    SSE_LOOP_W_ITEM(GX1, EX1, F, 0) \
+    SSE_LOOP_W_ITEM(GX1, EX1, F, 1) \
+    SSE_LOOP_W_ITEM(GX1, EX1, F, 2) \
+    SSE_LOOP_W_ITEM(GX1, EX1, F, 3)
+
 
 #define SSE_LOOP_DS_ITEM(EX1, F, i)     \
     LWU(EX1, wback, fixedaddress+i*4);  \
-- 
cgit 1.4.1