[Turbofan] Add ARM NEON instructions for implementing SIMD.

src/arm/disasm-arm.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.
 
 // A Disassembler object is used to disassemble a block of code instruction by
 // instruction. The default implementation of the NameConverter object can be
 // overriden to modify register names or to do symbol lookup on addresses.
 //
 // The example below will disassemble a block of code and print it to stdout.
 //
@@ skipping 1401 matching lines @@
 // Dd = vmls(Dn, Dm)
 // Sd = vmls(Sn, Sm)
 // Dd = vdiv(Dn, Dm)
 // Sd = vdiv(Sn, Sm)
 // vcmp(Dd, Dm)
 // vcmp(Sd, Sm)
 // Dd = vsqrt(Dm)
 // Sd = vsqrt(Sm)
 // vmrs
 // vmsr
+// Qd = vdup.size(Qd, Rt)
 void Decoder::DecodeTypeVFP(Instruction* instr) {
   VERIFY((instr->TypeValue() == 7) && (instr->Bit(24) == 0x0) );
   VERIFY(instr->Bits(11, 9) == 0x5);
 
   if (instr->Bit(4) == 0) {
     if (instr->Opc1Value() == 0x7) {
       // Other data processing instructions
       if ((instr->Opc2Value() == 0x0) && (instr->Opc3Value() == 0x1)) {
         // vmov register to register.
         if (instr->SzValue() == 0x1) {
@@ skipping 92 matching lines @@
       } else {
         Format(instr, "vdiv'cond.f32 'Sd, 'Sn, 'Sm");
       }
     } else {
       Unknown(instr);  // Not used by V8.
     }
   } else {
     if ((instr->VCValue() == 0x0) &&
         (instr->VAValue() == 0x0)) {
       DecodeVMOVBetweenCoreAndSinglePrecisionRegisters(instr);
-    } else if ((instr->VLValue() == 0x0) &&
-               (instr->VCValue() == 0x1) &&
-               (instr->Bit(23) == 0x0)) {
-      if (instr->Bit(21) == 0x0) {
-        Format(instr, "vmov'cond.32 'Dd[0], 'rt");
+    } else if ((instr->VLValue() == 0x0) && (instr->VCValue() == 0x1)) {
+      if (instr->Bit(23) == 0) {
+        if (instr->Bit(21) == 0x0) {
+          Format(instr, "vmov'cond.32 'Dd[0], 'rt");
+        } else {
+          Format(instr, "vmov'cond.32 'Dd[1], 'rt");
+        }
       } else {
-        Format(instr, "vmov'cond.32 'Dd[1], 'rt");
+        int size = 32;
+        if (instr->Bit(5) != 0)
+          size = 16;
+        else if (instr->Bit(22) != 0)
+          size = 8;
+        int Vd = (instr->Bit(7) << 3) | (instr->VnValue() >> 1);
+        int Rt = instr->RtValue();
+        out_buffer_pos_ += SNPrintF(out_buffer_ + out_buffer_pos_,
+                                    "vdup.%i q%d, r%d", size, Vd, Rt);
       }
     } else if ((instr->VLValue() == 0x1) &&
                (instr->VCValue() == 0x1) &&
                (instr->Bit(23) == 0x0)) {
       if (instr->Bit(21) == 0x0) {
         Format(instr, "vmov'cond.32 'rt, 'Dd[0]");
       } else {
         Format(instr, "vmov'cond.32 'rt, 'Dd[1]");
       }
     } else if ((instr->VCValue() == 0x0) &&
                (instr->VAValue() == 0x7) &&
                (instr->Bits(19, 16) == 0x1)) {
       if (instr->VLValue() == 0) {
         if (instr->Bits(15, 12) == 0xF) {
           Format(instr, "vmsr'cond FPSCR, APSR");
         } else {
           Format(instr, "vmsr'cond FPSCR, 'rt");
         }
       } else {
         if (instr->Bits(15, 12) == 0xF) {
           Format(instr, "vmrs'cond APSR, FPSCR");
         } else {
           Format(instr, "vmrs'cond 'rt, FPSCR");
         }
       }
+    } else {
+      Unknown(instr);  // Not used by V8.
     }
   }
 }
 
 void Decoder::DecodeTypeCP15(Instruction* instr) {
   VERIFY((instr->TypeValue() == 7) && (instr->Bit(24) == 0x0));
   VERIFY(instr->CoprocessorValue() == 15);
 
   if (instr->Bit(4) == 1) {
     int crn = instr->Bits(19, 16);
@@ skipping 226 matching lines @@
 void Decoder::DecodeSpecialCondition(Instruction* instr) {
   switch (instr->SpecialValue()) {
     case 4:
       if (instr->Bits(21, 20) == 2 && instr->Bits(11, 8) == 1 &&
           instr->Bit(4) == 1) {
         // vmov Qd, Qm
         int Vd = instr->VFPDRegValue(kSimd128Precision);
         int Vm = instr->VFPMRegValue(kSimd128Precision);
         out_buffer_pos_ +=
             SNPrintF(out_buffer_ + out_buffer_pos_, "vmov q%d, q%d", Vd, Vm);
+      } else if (instr->Bits(11, 8) == 8) {
+        const char* op = (instr->Bit(4) == 0) ? "vadd" : "vtst";
+        int size = kBitsPerByte * (1 << instr->Bits(21, 20));
+        int Vd = instr->VFPDRegValue(kSimd128Precision);
+        int Vm = instr->VFPMRegValue(kSimd128Precision);
+        int Vn = instr->VFPNRegValue(kSimd128Precision);
+        // vadd/vtst.I<size> Qd, Qm, Qn.
+        out_buffer_pos_ +=
+            SNPrintF(out_buffer_ + out_buffer_pos_, "%s.I%d q%d, q%d, q%d", op,
+                     size, Vd, Vn, Vm);
+      } else if (instr->Bits(11, 8) == 0xd && instr->Bit(4) == 0) {
+        const char* op = (instr->Bits(21, 20) == 0) ? "vadd" : "vsub";
+        int size = kBitsPerByte * (1 << instr->Bits(21, 20));
+        int Vd = instr->VFPDRegValue(kSimd128Precision);
+        int Vm = instr->VFPMRegValue(kSimd128Precision);
+        int Vn = instr->VFPNRegValue(kSimd128Precision);
+        // vadd/vsub.F32 Qd, Qm, Qn.
+        out_buffer_pos_ += SNPrintF(out_buffer_ + out_buffer_pos_,
+                                    "%s.F32 q%d, q%d, q%d", op, Vd, Vn, Vm);

[Rodolph Perfetta (ARM), 2016/12/07 15:59:00]

use lower case for the type.

[bbudge, 2016/12/08 03:07:31]

Done.

       } else {
         Unknown(instr);
       }
       break;
     case 5:
       if ((instr->Bits(18, 16) == 0) && (instr->Bits(11, 6) == 0x28) &&
           (instr->Bit(4) == 1)) {
         // vmovl signed
         if ((instr->VdValue() & 1) != 0) Unknown(instr);
         int Vd = (instr->Bit(22) << 3) | (instr->VdValue() >> 1);
         int Vm = (instr->Bit(5) << 4) | instr->VmValue();
         int imm3 = instr->Bits(21, 19);
         out_buffer_pos_ += SNPrintF(out_buffer_ + out_buffer_pos_,
                                     "vmovl.s%d q%d, d%d", imm3*8, Vd, Vm);
       } else {
         Unknown(instr);
       }
       break;
     case 6:
-      if (instr->Bits(21, 20) == 0 && instr->Bits(11, 8) == 1 &&
-          instr->Bit(4) == 1) {
+      if (instr->Bits(11, 8) == 8) {
+        int size = kBitsPerByte * (1 << instr->Bits(21, 20));
+        int Vd = instr->VFPDRegValue(kSimd128Precision);
+        int Vm = instr->VFPMRegValue(kSimd128Precision);
+        int Vn = instr->VFPNRegValue(kSimd128Precision);
+        if (instr->Bit(4) == 0) {
+          out_buffer_pos_ +=
+              SNPrintF(out_buffer_ + out_buffer_pos_, "vsub.I%d q%d, q%d, q%d",

[Rodolph Perfetta (ARM), 2016/12/07 15:59:00]

lower case

[bbudge, 2016/12/08 03:07:31]

Done.

+                       size, Vd, Vn, Vm);
+        } else {
+          out_buffer_pos_ +=
+              SNPrintF(out_buffer_ + out_buffer_pos_, "vceq.I%d q%d, q%d, q%d",

[Rodolph Perfetta (ARM), 2016/12/07 15:59:00]

lower case.

[bbudge, 2016/12/08 03:07:31]

Done.

+                       size, Vd, Vn, Vm);
+        }
+      } else if (instr->Bits(21, 20) == 1 && instr->Bits(11, 8) == 1 &&
+                 instr->Bit(4) == 1) {
+        int Vd = instr->VFPDRegValue(kSimd128Precision);
+        int Vm = instr->VFPMRegValue(kSimd128Precision);
+        int Vn = instr->VFPNRegValue(kSimd128Precision);
+        out_buffer_pos_ += SNPrintF(out_buffer_ + out_buffer_pos_,
+                                    "vbsl q%d, q%d, q%d", Vd, Vn, Vm);
+      } else if (instr->Bits(21, 20) == 0 && instr->Bits(11, 8) == 1 &&
+                 instr->Bit(4) == 1) {
         if (instr->Bit(6) == 0) {
           // veor Dd, Dn, Dm
           int Vd = instr->VFPDRegValue(kDoublePrecision);
           int Vn = instr->VFPNRegValue(kDoublePrecision);
           int Vm = instr->VFPMRegValue(kDoublePrecision);
           out_buffer_pos_ += SNPrintF(out_buffer_ + out_buffer_pos_,
                                       "veor d%d, d%d, d%d", Vd, Vn, Vm);
 
         } else {
           // veor Qd, Qn, Qm
@@ skipping 10 matching lines @@
     case 7:
       if ((instr->Bits(18, 16) == 0) && (instr->Bits(11, 6) == 0x28) &&
           (instr->Bit(4) == 1)) {
         // vmovl unsigned
         if ((instr->VdValue() & 1) != 0) Unknown(instr);
         int Vd = (instr->Bit(22) << 3) | (instr->VdValue() >> 1);
         int Vm = (instr->Bit(5) << 4) | instr->VmValue();
         int imm3 = instr->Bits(21, 19);
         out_buffer_pos_ += SNPrintF(out_buffer_ + out_buffer_pos_,
                                     "vmovl.u%d q%d, d%d", imm3*8, Vd, Vm);
+      } else if (instr->Opc1Value() == 7 && instr->Bits(19, 16) == 0 &&
+                 instr->Bits(11, 6) == 0x17 && instr->Bit(4) == 0) {
+        int Vd = instr->VFPDRegValue(kSimd128Precision);
+        int Vm = instr->VFPMRegValue(kSimd128Precision);
+        out_buffer_pos_ +=
+            SNPrintF(out_buffer_ + out_buffer_pos_, "vmvn q%d, q%d", Vd, Vm);
+      } else if (instr->Opc1Value() == 7 && instr->Bits(19, 16) == 0xB &&
+                 instr->Bits(11, 9) == 0x3 && instr->Bit(4) == 0) {
+        int Vd = instr->VFPDRegValue(kSimd128Precision);
+        int Vm = instr->VFPMRegValue(kSimd128Precision);
+        const char* suffix = nullptr;
+        int op = instr->Bits(8, 7);
+        switch (op) {
+          case 0:

[Rodolph Perfetta (ARM), 2016/12/07 15:59:00]

same as the assembler this is incorrect, it should be "f32.s32"

[bbudge, 2016/12/08 03:07:31]

Done.

+            suffix = "S32.F32";
+            break;
+          case 1:
+            suffix = "U32.F32";
+            break;
+          case 2:
+            suffix = "F32.S32";
+            break;
+          case 3:
+            suffix = "F32.U32";
+            break;
+        }
+        out_buffer_pos_ += SNPrintF(out_buffer_ + out_buffer_pos_,
+                                    "vcvt.%s q%d, q%d", suffix, Vd, Vm);
       } else if ((instr->Bits(21, 16) == 0x32) && (instr->Bits(11, 7) == 0) &&
                  (instr->Bit(4) == 0)) {
         if (instr->Bit(6) == 0) {
           int Vd = instr->VFPDRegValue(kDoublePrecision);
           int Vm = instr->VFPMRegValue(kDoublePrecision);
           out_buffer_pos_ +=
               SNPrintF(out_buffer_ + out_buffer_pos_, "vswp d%d, d%d", Vd, Vm);
         } else {
           int Vd = instr->VFPDRegValue(kSimd128Precision);
           int Vm = instr->VFPMRegValue(kSimd128Precision);
           out_buffer_pos_ +=
               SNPrintF(out_buffer_ + out_buffer_pos_, "vswp q%d, q%d", Vd, Vm);
         }
+      } else if (instr->Opc1Value() == 0x7 && instr->Bits(11, 7) == 0x18 &&
+                 instr->Bit(4) == 0x0) {
+        int Vd = instr->VFPDRegValue(kSimd128Precision);
+        int Vm = instr->VFPMRegValue(kDoublePrecision);
+        int index = instr->Bit(19);
+        out_buffer_pos_ += SNPrintF(out_buffer_ + out_buffer_pos_,
+                                    "vdup q%d, d%d[%d]", Vd, Vm, index);
+      } else if (instr->Opc1Value() == 0x7 && instr->Bits(11, 10) == 0x2 &&
+                 instr->Bit(4) == 0x0) {
+        int Vd = instr->VFPDRegValue(kDoublePrecision);
+        int Vn = instr->VFPNRegValue(kDoublePrecision);
+        int Vm = instr->VFPMRegValue(kDoublePrecision);
+        int len = instr->Bits(9, 8);
+        NeonListOperand list(DwVfpRegister::from_code(Vn), len + 1);
+        out_buffer_pos_ += SNPrintF(out_buffer_ + out_buffer_pos_, "%s d%d, ",
+                                    instr->Bit(6) == 0 ? "vtbl" : "vtbx", Vd);
+        FormatNeonList(Vn, list.type());
+        Print(", ");
+        PrintDRegister(Vm);
       } else {
         Unknown(instr);
       }
       break;
     case 8:
       if (instr->Bits(21, 20) == 0) {
         // vst1
         int Vd = (instr->Bit(22) << 4) | instr->VdValue();
         int Rn = instr->VnValue();
         int type = instr->Bits(11, 8);
@@ skipping 325 matching lines @@
     pc += d.InstructionDecode(buffer, pc);
     v8::internal::PrintF(f, "%p    %08x      %s\n", static_cast<void*>(prev_pc),
                          *reinterpret_cast<int32_t*>(prev_pc), buffer.start());
   }
 }
 
 
 }  // namespace disasm
 
 #endif  // V8_TARGET_ARCH_ARM