MIPS:[turbofan] Use Ins, Dins to clear bits instead of And with inverted immediate.

src/compiler/mips64/instruction-selector-mips64.cc

 // Copyright 2014 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 "src/base/adapters.h"
 #include "src/base/bits.h"
 #include "src/compiler/instruction-selector-impl.h"
 #include "src/compiler/node-matchers.h"
 #include "src/compiler/node-properties.h"
 
@@ skipping 277 matching lines @@
         if (lsb + mask_width > 32) mask_width = 32 - lsb;
 
         Emit(kMips64Ext, g.DefineAsRegister(node),
              g.UseRegister(mleft.left().node()), g.TempImmediate(lsb),
              g.TempImmediate(mask_width));
         return;
       }
       // Other cases fall through to the normal And operation.
     }
   }
+  if (m.right().HasValue()) {
+    uint32_t mask = m.right().Value();
+    uint32_t shift = base::bits::CountPopulation32(~mask);
+    uint32_t msb = base::bits::CountLeadingZeros32(~mask);
+    if (shift != 0 && shift != 32 && msb + shift == 32) {
+      // Insert zeros for (x >> K) << K => x & ~(2^K - 1) expression reduction
+      // and remove constant loading of invereted mask.

[paul.l..., 2015/12/01 17:49:34]

nit, inverted. Optional: maybe a comment mentioning DINS instruction is limited
to inserting within low 32-bits only? The code is right, but initial expectation
is that K could be up to 63. (And, if those cases were at all common in
practice, we could add DINSM, DINSU instructions. Might still be less than
loading 64-bit constant.)

[dusan.milosavljevic, 2015/12/01 21:55:54]

Done.

+      Emit(kMips64Ins, g.DefineSameAsFirst(node),
+           g.UseRegister(m.left().node()), g.TempImmediate(0),
+           g.TempImmediate(shift));
+      return;
+    }
+  }
   VisitBinop(this, node, kMips64And);
 }
 
 
 void InstructionSelector::VisitWord64And(Node* node) {
   Mips64OperandGenerator g(this);
   Int64BinopMatcher m(node);
   if (m.left().IsWord64Shr() && CanCover(node, m.left().node()) &&
       m.right().HasValue()) {
     uint64_t mask = m.right().Value();
@@ skipping 17 matching lines @@
         if (lsb + mask_width > 64) mask_width = 64 - lsb;
 
         Emit(kMips64Dext, g.DefineAsRegister(node),
              g.UseRegister(mleft.left().node()), g.TempImmediate(lsb),
              g.TempImmediate(static_cast<int32_t>(mask_width)));
         return;
       }
       // Other cases fall through to the normal And operation.
     }
   }
+  if (m.right().HasValue()) {
+    uint64_t mask = m.right().Value();
+    uint32_t shift = base::bits::CountPopulation64(~mask);
+    uint32_t msb = base::bits::CountLeadingZeros64(~mask);
+    if (shift != 0 && shift < 32 && msb + shift == 64) {
+      // Insert zeros for (x >> K) << K => x & ~(2^K - 1) expression reduction
+      // and remove constant loading of invereted mask.
+      Emit(kMips64Dins, g.DefineSameAsFirst(node),
+           g.UseRegister(m.left().node()), g.TempImmediate(0),
+           g.TempImmediate(shift));
+      return;
+    }
+  }
   VisitBinop(this, node, kMips64And);
 }
 
 
 void InstructionSelector::VisitWord32Or(Node* node) {
   VisitBinop(this, node, kMips64Or);
 }
 
 
 void InstructionSelector::VisitWord64Or(Node* node) {
@@ skipping 1276 matching lines @@
          MachineOperatorBuilder::kFloat32RoundUp |
          MachineOperatorBuilder::kFloat64RoundTruncate |
          MachineOperatorBuilder::kFloat32RoundTruncate |
          MachineOperatorBuilder::kFloat64RoundTiesEven |
          MachineOperatorBuilder::kFloat32RoundTiesEven;
 }
 
 }  // namespace compiler
 }  // namespace internal
 }  // namespace v8