[ia32] Introduce FMA3 instructions on scalar data elements.

src/ia32/disasm-ia32.cc

 // Copyright 2011 the V8 project authors. All rights reserved.
 // Use of this source code is governed by a BSD-style license that can be
 // found in the LICENSE file.
 
 #include <assert.h>
 #include <stdarg.h>
 #include <stdio.h>
 
 #include "src/v8.h"
 
@@ skipping 301 matching lines @@
   }
 
   bool vex_f2() {
     DCHECK(vex_byte0_ == 0xc4 || vex_byte0_ == 0xc5);
     byte checked = vex_byte0_ == 0xc4 ? vex_byte2_ : vex_byte1_;
     return (checked & 3) == 3;
   }
 
   bool vex_w() {
     if (vex_byte0_ == 0xc5) return false;
-    return (vex_byte2_ & 0x80) == 1;
+    return (vex_byte2_ & 0x80) != 0;
   }
 
   bool vex_0f() {
     if (vex_byte0_ == 0xc5) return true;
     return (vex_byte1_ & 3) == 1;
   }
 
   bool vex_0f38() {
     if (vex_byte0_ == 0xc5) return false;
     return (vex_byte1_ & 3) == 2;
@@ skipping 400 matching lines @@
   byte cond = *(data + 1) & 0x0F;
   const char* mnem = conditional_move_mnem[cond];
   int op_size = PrintOperands(mnem, REG_OPER_OP_ORDER, data + 2);
   return 2 + op_size;  // includes 0x0F
 }
 
 
 int DisassemblerIA32::AVXInstruction(byte* data) {
   byte opcode = *data;
   byte* current = data + 1;
-  if (vex_f2() && vex_0f()) {
+  if (vex_66() && vex_0f38()) {
     int mod, regop, rm, vvvv = vex_vreg();
     get_modrm(*current, &mod, &regop, &rm);
     switch (opcode) {
+      case 0x99:
+        AppendToBuffer("vfmadd132s%c %s,%s,", float_size_code(),
+                       NameOfXMMRegister(regop), NameOfXMMRegister(vvvv));
+        current += PrintRightXMMOperand(current);
+        break;
+      case 0xa9:
+        AppendToBuffer("vfmadd213s%c %s,%s,", float_size_code(),
+                       NameOfXMMRegister(regop), NameOfXMMRegister(vvvv));
+        current += PrintRightXMMOperand(current);
+        break;
+      case 0xb9:
+        AppendToBuffer("vfmadd231s%c %s,%s,", float_size_code(),
+                       NameOfXMMRegister(regop), NameOfXMMRegister(vvvv));
+        current += PrintRightXMMOperand(current);
+        break;
+      case 0x9b:
+        AppendToBuffer("vfmsub132s%c %s,%s,", float_size_code(),
+                       NameOfXMMRegister(regop), NameOfXMMRegister(vvvv));
+        current += PrintRightXMMOperand(current);
+        break;
+      case 0xab:
+        AppendToBuffer("vfmsub213s%c %s,%s,", float_size_code(),
+                       NameOfXMMRegister(regop), NameOfXMMRegister(vvvv));
+        current += PrintRightXMMOperand(current);
+        break;
+      case 0xbb:
+        AppendToBuffer("vfmsub231s%c %s,%s,", float_size_code(),
+                       NameOfXMMRegister(regop), NameOfXMMRegister(vvvv));
+        current += PrintRightXMMOperand(current);
+        break;
+      case 0x9d:
+        AppendToBuffer("vfnmadd132s%c %s,%s,", float_size_code(),
+                       NameOfXMMRegister(regop), NameOfXMMRegister(vvvv));
+        current += PrintRightXMMOperand(current);
+        break;
+      case 0xad:
+        AppendToBuffer("vfnmadd213s%c %s,%s,", float_size_code(),
+                       NameOfXMMRegister(regop), NameOfXMMRegister(vvvv));
+        current += PrintRightXMMOperand(current);
+        break;
+      case 0xbd:
+        AppendToBuffer("vfnmadd231s%c %s,%s,", float_size_code(),
+                       NameOfXMMRegister(regop), NameOfXMMRegister(vvvv));
+        current += PrintRightXMMOperand(current);
+        break;
+      case 0x9f:
+        AppendToBuffer("vfnmsub132s%c %s,%s,", float_size_code(),
+                       NameOfXMMRegister(regop), NameOfXMMRegister(vvvv));
+        current += PrintRightXMMOperand(current);
+        break;
+      case 0xaf:
+        AppendToBuffer("vfnmsub213s%c %s,%s,", float_size_code(),
+                       NameOfXMMRegister(regop), NameOfXMMRegister(vvvv));
+        current += PrintRightXMMOperand(current);
+        break;
+      case 0xbf:
+        AppendToBuffer("vfnmsub231s%c %s,%s,", float_size_code(),
+                       NameOfXMMRegister(regop), NameOfXMMRegister(vvvv));
+        current += PrintRightXMMOperand(current);
+        break;
+      default:
+        UnimplementedInstruction();
+    }
+  } else if (vex_f2() && vex_0f()) {
+    int mod, regop, rm, vvvv = vex_vreg();
+    get_modrm(*current, &mod, &regop, &rm);
+    switch (opcode) {
       case 0x58:
         AppendToBuffer("vaddsd %s,%s,", NameOfXMMRegister(regop),
                        NameOfXMMRegister(vvvv));
         current += PrintRightXMMOperand(current);
         break;
       case 0x59:
         AppendToBuffer("vmulsd %s,%s,", NameOfXMMRegister(regop),
                        NameOfXMMRegister(vvvv));
         current += PrintRightXMMOperand(current);
         break;
@@ skipping 395 matching lines @@
             AppendToBuffer("%s", f0mnem);
             data += 2;
           } else if (f0byte == 0x28) {
             data += 2;
             int mod, regop, rm;
             get_modrm(*data, &mod, &regop, &rm);
             AppendToBuffer("movaps %s,%s",
                            NameOfXMMRegister(regop),
                            NameOfXMMRegister(rm));
             data++;
+          } else if (f0byte == 0x2e) {
+            data += 2;
+            int mod, regop, rm;
+            get_modrm(*data, &mod, &regop, &rm);
+            AppendToBuffer("ucomiss %s,", NameOfXMMRegister(regop));
+            data += PrintRightXMMOperand(data);
           } else if (f0byte >= 0x53 && f0byte <= 0x5F) {
             const char* const pseudo_op[] = {
               "rcpps",
               "andps",
               "andnps",
               "orps",
               "xorps",
               "addps",
               "mulps",
               "cvtps2pd",
@@ skipping 550 matching lines @@
             int mod, regop, rm;
             get_modrm(*data, &mod, &regop, &rm);
             AppendToBuffer("movss %s,", NameOfXMMRegister(regop));
             data += PrintRightXMMOperand(data);
           } else if (b2 == 0x2C) {
             data += 3;
             int mod, regop, rm;
             get_modrm(*data, &mod, &regop, &rm);
             AppendToBuffer("cvttss2si %s,", NameOfCPURegister(regop));
             data += PrintRightXMMOperand(data);
+          } else if (b2 == 0x58) {
+            data += 3;
+            int mod, regop, rm;
+            get_modrm(*data, &mod, &regop, &rm);
+            AppendToBuffer("addss %s,", NameOfXMMRegister(regop));
+            data += PrintRightXMMOperand(data);
+          } else if (b2 == 0x59) {
+            data += 3;
+            int mod, regop, rm;
+            get_modrm(*data, &mod, &regop, &rm);
+            AppendToBuffer("mulss %s,", NameOfXMMRegister(regop));
+            data += PrintRightXMMOperand(data);
           } else if (b2 == 0x5A) {
             data += 3;
             int mod, regop, rm;
             get_modrm(*data, &mod, &regop, &rm);
             AppendToBuffer("cvtss2sd %s,", NameOfXMMRegister(regop));
             data += PrintRightXMMOperand(data);
+          } else if (b2 == 0x5c) {
+            data += 3;
+            int mod, regop, rm;
+            get_modrm(*data, &mod, &regop, &rm);
+            AppendToBuffer("subss %s,", NameOfXMMRegister(regop));
+            data += PrintRightXMMOperand(data);
+          } else if (b2 == 0x5e) {
+            data += 3;
+            int mod, regop, rm;
+            get_modrm(*data, &mod, &regop, &rm);
+            AppendToBuffer("divss %s,", NameOfXMMRegister(regop));
+            data += PrintRightXMMOperand(data);
           } else if (b2 == 0x6F) {
             data += 3;
             int mod, regop, rm;
             get_modrm(*data, &mod, &regop, &rm);
             AppendToBuffer("movdqu %s,", NameOfXMMRegister(regop));
             data += PrintRightXMMOperand(data);
           } else if (b2 == 0x7F) {
             AppendToBuffer("movdqu ");
             data += 3;
             int mod, regop, rm;
@@ skipping 143 matching lines @@
       fprintf(f, "  ");
     }
     fprintf(f, "  %s\n", buffer.start());
   }
 }
 
 
 }  // namespace disasm
 
 #endif  // V8_TARGET_ARCH_IA32