4 files changed, 54 insertions, 1 deletions
diff --git a/src/emu/x64run66.c b/src/emu/x64run66.c
index cfc10831..0eefff20 100644
--- a/src/emu/x64run66.c
+++ b/src/emu/x64run66.c
@@ -326,6 +326,13 @@ int Run66(x64emu_t *emu, rex_t rex, int rep)
         }
         break;
     
+    case 0xA9:                             /* TEST AX,Iw */
+        if(rex.w)
+            test64(emu, R_RAX, F32S64);
+        else
+            test16(emu, R_AX, F16);
+        break;
+
     case 0xAB:                      /* (REP) STOSW */
         if(rex.w)
             tmp8s = ACCESS_FLAG(F_DF)?-8:+8;
diff --git a/src/emu/x64run660f.c b/src/emu/x64run660f.c
index 132fc7d7..b2e21c32 100644
--- a/src/emu/x64run660f.c
+++ b/src/emu/x64run660f.c
@@ -33,6 +33,7 @@ int Run660F(x64emu_t *emu, rex_t rex)
     uint8_t tmp8u;
     int8_t tmp8s;
     uint16_t tmp16u;
+    int32_t tmp32s;
     uint64_t tmp64u;
     reg64_t *oped, *opgd;
     sse_regs_t *opex, *opgx, eax1;
@@ -861,7 +862,22 @@ int Run660F(x64emu_t *emu, rex_t rex)
         else 
             {tmp8u=EX->ub[0]; for (int i=0; i<2; ++i) GX->q[i] >>= tmp8u;}
         break;
-
+    case 0xD4:  /* PADDQ Gx,Ex */
+        nextop = F8;
+        GETEX(0);
+        GETGX;
+        GX->sq[0] += EX->sq[0];
+        GX->sq[1] += EX->sq[1];
+        break;
+    case 0xD5:  /* PMULLW Gx,Ex */
+        nextop = F8;
+        GETEX(0);
+        GETGX;
+        for(int i=0; i<8; ++i) {
+            tmp32s = (int32_t)GX->sw[i] * EX->sw[i];
+            GX->sw[i] = tmp32s&0xffff;
+        }
+        break;
     case 0xD6:                      /* MOVQ Ex,Gx */
         nextop = F8;
         GETEX(0);
diff --git a/src/emu/x64runf20f.c b/src/emu/x64runf20f.c
index bc8b79de..7b8238a8 100644
--- a/src/emu/x64runf20f.c
+++ b/src/emu/x64runf20f.c
@@ -186,6 +186,21 @@ int RunF20F(x64emu_t *emu, rex_t rex)
         }
         break;
 
+    case 0x7C:  /* HADDPS Gx, Ex */
+        nextop = F8;
+        GETEX(0);
+        GETGX;
+        GX->f[0] += GX->f[1];
+        GX->f[1] = GX->f[2] + GX->f[3];
+        if(EX==GX) {
+            GX->f[2] = GX->f[0];
+            GX->f[3] = GX->f[1];
+        } else {
+            GX->f[2] = EX->f[0] + EX->f[1];
+            GX->f[3] = EX->f[2] + EX->f[3];
+        }
+        break;
+
     GOCOND(0x80
         , tmp32s = F32S; CHECK_FLAGS(emu);
         , R_RIP += tmp32s;
diff --git a/src/emu/x64runf30f.c b/src/emu/x64runf30f.c
index c49ef46b..9b25e73a 100644
--- a/src/emu/x64runf30f.c
+++ b/src/emu/x64runf30f.c
@@ -57,6 +57,21 @@ int RunF30F(x64emu_t *emu, rex_t rex)
         GETGX;
         EX->ud[0] = GX->ud[0];
         break;
+    case 0x12:  /* MOVSLDUP Gx, Ex */
+        nextop = F8;
+        GETEX(0);
+        GETGX;
+        GX->ud[1] = GX->ud[0] = EX->ud[0];
+        GX->ud[3] = GX->ud[2] = EX->ud[2];
+        break;
+    
+    case 0x16:  /* MOVSHDUP Gx, Ex */
+        nextop = F8;
+        GETEX(0);
+        GETGX;
+        GX->ud[1] = GX->ud[0] = EX->ud[1];
+        GX->ud[3] = GX->ud[2] = EX->ud[3];
+        break;
 
     case 0x2A:  /* CVTSI2SS Gx, Ed */
         nextop = F8;