[turbofan] Implement rounding of floats on x64 and ia32 without sse4.1.

src/x64/disasm-x64.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>
 
 #if V8_TARGET_ARCH_X64
 
@@ skipping 976 matching lines @@
       case 0x2a:
         AppendToBuffer("%s %s,%s,", vex_w() ? "vcvtqsi2ss" : "vcvtlsi2ss",
                        NameOfXMMRegister(regop), NameOfXMMRegister(vvvv));
         current += PrintRightOperand(current);
         break;
       case 0x2c:
         AppendToBuffer("vcvttss2si%s %s,", vex_w() ? "q" : "",
                        NameOfCPURegister(regop));
         current += PrintRightXMMOperand(current);
         break;
+      case 0x2d:
+        AppendToBuffer("vcvtss2si%s %s,", vex_w() ? "q" : "",
+                       NameOfCPURegister(regop));
+        current += PrintRightXMMOperand(current);
+        break;
       case 0x58:
         AppendToBuffer("vaddss %s,%s,", NameOfXMMRegister(regop),
                        NameOfXMMRegister(vvvv));
         current += PrintRightXMMOperand(current);
         break;
       case 0x59:
         AppendToBuffer("vmulss %s,%s,", NameOfXMMRegister(regop),
                        NameOfXMMRegister(vvvv));
         current += PrintRightXMMOperand(current);
         break;
@@ skipping 225 matching lines @@
         break;
       case 0x2e:
         AppendToBuffer("vucomiss %s,", NameOfXMMRegister(regop));
         current += PrintRightXMMOperand(current);
         break;
       case 0x54:
         AppendToBuffer("vandps %s,%s,", NameOfXMMRegister(regop),
                        NameOfXMMRegister(vvvv));
         current += PrintRightXMMOperand(current);
         break;
+      case 0x56:
+        AppendToBuffer("vorps %s,%s,", NameOfXMMRegister(regop),
+                       NameOfXMMRegister(vvvv));
+        current += PrintRightXMMOperand(current);
+        break;
       case 0x57:
         AppendToBuffer("vxorps %s,%s,", NameOfXMMRegister(regop),
                        NameOfXMMRegister(vvvv));
         current += PrintRightXMMOperand(current);
         break;
       default:
         UnimplementedInstruction();
     }
   } else if (vex_66() && vex_0f()) {
     int mod, regop, rm, vvvv = vex_vreg();
@@ skipping 29 matching lines @@
       case 0x57:
         AppendToBuffer("vxorpd %s,%s,", NameOfXMMRegister(regop),
                        NameOfXMMRegister(vvvv));
         current += PrintRightXMMOperand(current);
         break;
       case 0x6e:
         AppendToBuffer("vmov%c %s,", vex_w() ? 'q' : 'd',
                        NameOfXMMRegister(regop));
         current += PrintRightOperand(current);
         break;
+      case 0x72:
+        AppendToBuffer("%s %s,", regop == 6 ? "vpslld" : "vpsrld",
+                       NameOfXMMRegister(vvvv));
+        current += PrintRightXMMOperand(current);
+        AppendToBuffer(",%u", *current++);
+        break;
       case 0x73:
         AppendToBuffer("%s %s,", regop == 6 ? "vpsllq" : "vpsrlq",
                        NameOfXMMRegister(vvvv));
         current += PrintRightXMMOperand(current);
         AppendToBuffer(",%u", *current++);
         break;
       case 0x76:
         AppendToBuffer("vpcmpeqd %s,%s,", NameOfXMMRegister(regop),
                        NameOfXMMRegister(vvvv));
         current += PrintRightXMMOperand(current);
@@ skipping 403 matching lines @@
       AppendToBuffer("%s %s,", mnemonic, NameOfXMMRegister(regop));
       current += PrintRightOperand(current);
     } else if (opcode == 0x2C) {
       // CVTTSS2SI:
       // Convert with truncation scalar single-precision FP to dword integer.
       int mod, regop, rm;
       get_modrm(*current, &mod, &regop, &rm);
       AppendToBuffer("cvttss2si%c %s,",
           operand_size_code(), NameOfCPURegister(regop));
       current += PrintRightXMMOperand(current);
+    } else if (opcode == 0x2D) {
+      // CVTSS2SI:
+      // Convert with rounded scalar single-precision FP to dword integer.
+      int mod, regop, rm;
+      get_modrm(*current, &mod, &regop, &rm);
+      AppendToBuffer("cvtss2si%c %s,", operand_size_code(),
+                     NameOfCPURegister(regop));
+      current += PrintRightXMMOperand(current);
     } else if (opcode == 0x7E) {
       int mod, regop, rm;
       get_modrm(*current, &mod, &regop, &rm);
       AppendToBuffer("movq %s,", NameOfXMMRegister(regop));
       current += PrintRightXMMOperand(current);
     } else if ((opcode & 0xF8) == 0x58 || opcode == 0x51) {
       // XMM arithmetic. Mnemonic was retrieved at the start of this function.
       int mod, regop, rm;
       get_modrm(*current, &mod, &regop, &rm);
       AppendToBuffer("%s %s,", mnemonic, NameOfXMMRegister(regop));
@@ skipping 140 matching lines @@
     }
   } else if (opcode == 0xB8 || opcode == 0xBC || opcode == 0xBD) {
     // POPCNT, CTZ, CLZ.
     AppendToBuffer("%s%c ", mnemonic, operand_size_code());
     int mod, regop, rm;
     get_modrm(*current, &mod, &regop, &rm);
     AppendToBuffer("%s,", NameOfCPURegister(regop));
     current += PrintRightOperand(current);
   } else if (opcode == 0x0B) {
     AppendToBuffer("ud2");
+  } else if (opcode == 0xAE) {
+    byte modrm = *(data + 2);
+    int mod, regop, rm;
+    get_modrm(modrm, &mod, &regop, &rm);
+    regop &= 0x7;  // The REX.R bit does not affect the operation.
+    const char* mnem = NULL;
+    switch (regop) {
+      case 2:
+        mnem = "ldmxcsr";
+        break;
+      case 3:
+        mnem = "stmxcsr";
+        break;
+      default:
+        UnimplementedInstruction();
+        return 2;
+    }
+    DCHECK_NOT_NULL(mnem);
+    AppendToBuffer("%s ", mnem);
+    current +=
+        PrintRightOperandHelper(current, &DisassemblerX64::NameOfCPURegister);
   } else {
     UnimplementedInstruction();
   }
   return static_cast<int>(current - data);
 }
 
 
 // Mnemonics for two-byte opcode instructions starting with 0x0F.
 // The argument is the second byte of the two-byte opcode.
 // Returns NULL if the instruction is not handled here.
@@ skipping 606 matching lines @@
     for (int i = 6 - static_cast<int>(pc - prev_pc); i >= 0; i--) {
       fprintf(f, "  ");
     }
     fprintf(f, "  %s\n", buffer.start());
   }
 }
 
 }  // namespace disasm
 
 #endif  // V8_TARGET_ARCH_X64