SIMD shuffle API changes

runtime/vm/intermediate_language_arm.cc

 // Copyright (c) 2013, the Dart project authors.  Please see the AUTHORS file
 // for details. All rights reserved. Use of this source code is governed by a
 // BSD-style license that can be found in the LICENSE file.
 
 #include "vm/globals.h"  // Needed here to get TARGET_ARCH_ARM.
 #if defined(TARGET_ARCH_ARM)
 
 #include "vm/intermediate_language.h"
 
 #include "vm/dart_entry.h"
@@ skipping 3013 matching lines @@
   switch (op_kind()) {
     case Token::kADD: __ vaddqs(result, left, right); break;
     case Token::kSUB: __ vsubqs(result, left, right); break;
     case Token::kMUL: __ vmulqs(result, left, right); break;
     case Token::kDIV: __ Vdivqs(result, left, right); break;
     default: UNREACHABLE();
   }
 }
 
 
-LocationSummary* Float32x4ShuffleInstr::MakeLocationSummary() const {
+LocationSummary* Simd32x4ShuffleInstr::MakeLocationSummary() const {
   const intptr_t kNumInputs = 1;
   const intptr_t kNumTemps = 0;
   LocationSummary* summary =
       new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
   // Low (< Q7) Q registers are needed for the vcvtds and vmovs instructions.
   summary->set_in(0, Location::FpuRegisterLocation(Q5));
   summary->set_out(Location::FpuRegisterLocation(Q6));
   return summary;
 }
 
 
-void Float32x4ShuffleInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
+void Simd32x4ShuffleInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
   QRegister value = locs()->in(0).fpu_reg();
   QRegister result = locs()->out().fpu_reg();
   DRegister dresult0 = EvenDRegisterOf(result);
   DRegister dresult1 = OddDRegisterOf(result);
   SRegister sresult0 = EvenSRegisterOf(dresult0);
   SRegister sresult1 = OddSRegisterOf(dresult0);
   SRegister sresult2 = EvenSRegisterOf(dresult1);
   SRegister sresult3 = OddSRegisterOf(dresult1);
 
   DRegister dvalue0 = EvenDRegisterOf(value);
@@ skipping 14 matching lines @@
       __ vcvtds(dresult0, sresult0);
       break;
     case MethodRecognizer::kFloat32x4ShuffleZ:
       __ vdup(kWord, result, dvalue1, 0);
       __ vcvtds(dresult0, sresult0);
       break;
     case MethodRecognizer::kFloat32x4ShuffleW:
       __ vdup(kWord, result, dvalue1, 1);
       __ vcvtds(dresult0, sresult0);
       break;
+    case MethodRecognizer::kUint32x4Shuffle:
     case MethodRecognizer::kFloat32x4Shuffle:
       if (mask_ == 0x00) {
         __ vdup(kWord, result, dvalue0, 0);
       } else if (mask_ == 0x55) {
         __ vdup(kWord, result, dvalue0, 1);
       } else if (mask_ == 0xAA) {
         __ vdup(kWord, result, dvalue1, 0);
       } else  if (mask_ == 0xFF) {
         __ vdup(kWord, result, dvalue1, 1);
       } else {
         SRegister svalues[4];
 
         svalues[0] = EvenSRegisterOf(dtemp0);
         svalues[1] = OddSRegisterOf(dtemp0);
         svalues[2] = EvenSRegisterOf(dtemp1);
         svalues[3] = OddSRegisterOf(dtemp1);
 
         __ vmovq(QTMP, value);
         __ vmovs(sresult0, svalues[mask_ & 0x3]);
         __ vmovs(sresult1, svalues[(mask_ >> 2) & 0x3]);
         __ vmovs(sresult2, svalues[(mask_ >> 4) & 0x3]);
         __ vmovs(sresult3, svalues[(mask_ >> 6) & 0x3]);
       }
       break;
     default: UNREACHABLE();
   }
 }
 
 
+LocationSummary* Simd32x4ShuffleMixInstr::MakeLocationSummary() const {
+  const intptr_t kNumInputs = 2;
+  const intptr_t kNumTemps = 0;
+  LocationSummary* summary =
+      new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
+  // Low (< Q7) Q registers are needed for the vcvtds and vmovs instructions.
+  summary->set_in(0, Location::FpuRegisterLocation(Q4));
+  summary->set_in(1, Location::FpuRegisterLocation(Q5));
+  summary->set_out(Location::FpuRegisterLocation(Q6));
+  return summary;
+}
+
+
+void Simd32x4ShuffleMixInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
+  QRegister left = locs()->in(0).fpu_reg();
+  QRegister right = locs()->in(1).fpu_reg();
+  QRegister result = locs()->out().fpu_reg();
+
+  DRegister dresult0 = EvenDRegisterOf(result);
+  DRegister dresult1 = OddDRegisterOf(result);
+  SRegister sresult0 = EvenSRegisterOf(dresult0);
+  SRegister sresult1 = OddSRegisterOf(dresult0);
+  SRegister sresult2 = EvenSRegisterOf(dresult1);
+  SRegister sresult3 = OddSRegisterOf(dresult1);
+
+  DRegister dleft0 = EvenDRegisterOf(left);
+  DRegister dleft1 = OddDRegisterOf(left);
+  DRegister dright0 = EvenDRegisterOf(right);
+  DRegister dright1 = OddDRegisterOf(right);
+
+  switch (op_kind()) {
+    case MethodRecognizer::kFloat32x4ShuffleMix:
+    case MethodRecognizer::kUint32x4ShuffleMix:

[zra, 2013/10/30 17:39:27]

Is it possible to have special cases where vdup can be used?

[Cutch, 2013/10/31 01:43:55]

I'm not familiar enough with the NEON shuffling instructions, but, I expect all
shuffle and shuffleMix masks have better instruction sequences than the __ vmovs
code.

[zra, 2013/10/31 15:10:20]

Please leave a TODO here for me, and I'll look into it.

+      SRegister left_svalues[4];
+      SRegister right_svalues[4];
+
+      left_svalues[0] = EvenSRegisterOf(dleft0);
+      left_svalues[1] = OddSRegisterOf(dleft0);
+      left_svalues[2] = EvenSRegisterOf(dleft1);
+      left_svalues[3] = OddSRegisterOf(dleft1);
+      right_svalues[0] = EvenSRegisterOf(dright0);
+      right_svalues[1] = OddSRegisterOf(dright0);
+      right_svalues[2] = EvenSRegisterOf(dright1);
+      right_svalues[3] = OddSRegisterOf(dright1);
+
+      __ vmovs(sresult0, left_svalues[mask_ & 0x3]);
+      __ vmovs(sresult1, left_svalues[(mask_ >> 2) & 0x3]);
+      __ vmovs(sresult2, right_svalues[(mask_ >> 4) & 0x3]);
+      __ vmovs(sresult3, right_svalues[(mask_ >> 6) & 0x3]);
+      break;
+    default: UNREACHABLE();
+  }
+}
+
+
 LocationSummary* Simd32x4GetSignMaskInstr::MakeLocationSummary() const {
   const intptr_t kNumInputs = 1;
   const intptr_t kNumTemps = 1;
   LocationSummary* summary =
       new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
   summary->set_in(0, Location::FpuRegisterLocation(Q5));
   summary->set_temp(0, Location::RequiresRegister());
   summary->set_out(Location::RequiresRegister());
   return summary;
 }
@@ skipping 346 matching lines @@
 void Float32x4ToUint32x4Instr::EmitNativeCode(FlowGraphCompiler* compiler) {
   QRegister value = locs()->in(0).fpu_reg();
   QRegister result = locs()->out().fpu_reg();
 
   if (value != result) {
     __ vmovq(result, value);
   }
 }
 
 
-LocationSummary* Float32x4TwoArgShuffleInstr::MakeLocationSummary() const {
-  const intptr_t kNumInputs = 2;
-  const intptr_t kNumTemps = 0;
-  LocationSummary* summary =
-      new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
-  summary->set_in(0, Location::RequiresFpuRegister());
-  summary->set_in(1, Location::RequiresFpuRegister());
-  summary->set_out(Location::SameAsFirstInput());
-  return summary;
-}
-
-
-void Float32x4TwoArgShuffleInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
-  QRegister left = locs()->in(0).fpu_reg();
-  QRegister right = locs()->in(1).fpu_reg();
-  QRegister result = locs()->out().fpu_reg();
-
-  ASSERT(result == left);
-
-  DRegister dleft0 = EvenDRegisterOf(left);
-  DRegister dleft1 = OddDRegisterOf(left);
-  DRegister dright0 = EvenDRegisterOf(right);
-  DRegister dright1 = OddDRegisterOf(right);
-
-  switch (op_kind()) {
-    case MethodRecognizer::kFloat32x4WithZWInXY:
-      __ vmovd(dleft0, dright1);
-      break;
-    case MethodRecognizer::kFloat32x4InterleaveXY:
-      __ vmovq(QTMP, right);
-      __ vzipqw(left, QTMP);
-      break;
-    case MethodRecognizer::kFloat32x4InterleaveZW:
-      __ vmovq(QTMP, right);
-      __ vzipqw(left, QTMP);
-      __ vmovq(left, QTMP);
-      break;
-    case MethodRecognizer::kFloat32x4InterleaveXYPairs:
-      __ vmovd(dleft1, dright0);
-      break;
-    case MethodRecognizer::kFloat32x4InterleaveZWPairs:
-      __ vmovq(QTMP, right);
-      __ vmovd(EvenDRegisterOf(QTMP), dleft1);
-      __ vmovq(result, QTMP);
-      break;
-    default: UNREACHABLE();
-  }
-}
-
-
 LocationSummary* Uint32x4BoolConstructorInstr::MakeLocationSummary() const {
   const intptr_t kNumInputs = 4;
   const intptr_t kNumTemps = 1;
   LocationSummary* summary =
       new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
   summary->set_in(0, Location::RequiresRegister());
   summary->set_in(1, Location::RequiresRegister());
   summary->set_in(2, Location::RequiresRegister());
   summary->set_in(3, Location::RequiresRegister());
   summary->set_temp(0, Location::RequiresRegister());
@@ skipping 1115 matching lines @@
   compiler->GenerateCall(token_pos(),
                          &label,
                          PcDescriptors::kOther,
                          locs());
   __ Drop(2);  // Discard type arguments and receiver.
 }
 
 }  // namespace dart
 
 #endif  // defined TARGET_ARCH_ARM