Optimize double min/max by reducing the code size (reusing left register as result removes the need…

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 "lib/error.h"
@@ skipping 3337 matching lines @@
 
 
 LocationSummary* MathMinMaxInstr::MakeLocationSummary() const {
   if (result_cid() == kDoubleCid) {
     const intptr_t kNumInputs = 2;
     const intptr_t kNumTemps = 1;
     LocationSummary* summary =
         new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
     summary->set_in(0, Location::RequiresFpuRegister());
     summary->set_in(1, Location::RequiresFpuRegister());
+    // Reuse the left register so that code can be made shorter.
+    summary->set_out(Location::SameAsFirstInput());
     summary->set_temp(0, Location::RequiresRegister());
-    summary->set_out(Location::RequiresFpuRegister());
     return summary;
   }
   ASSERT(result_cid() == kSmiCid);
-  UNIMPLEMENTED();
-  return NULL;
+  const intptr_t kNumInputs = 2;
+  const intptr_t kNumTemps = 0;
+  LocationSummary* summary =
+      new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
+  summary->set_in(0, Location::RequiresRegister());
+  summary->set_in(1, Location::RequiresRegister());
+  // Reuse the left register so that code can be made shorter.
+  summary->set_out(Location::SameAsFirstInput());
+  return summary;
 }
 
 
 void MathMinMaxInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
   ASSERT((op_kind() == MethodRecognizer::kMathMin) ||
          (op_kind() == MethodRecognizer::kMathMax));
+  const intptr_t is_min = (op_kind() == MethodRecognizer::kMathMin);
   if (result_cid() == kDoubleCid) {
     Label done, returns_nan, are_equal;
     DRegister left = EvenDRegisterOf(locs()->in(0).fpu_reg());
     DRegister right = EvenDRegisterOf(locs()->in(1).fpu_reg());
     DRegister result = EvenDRegisterOf(locs()->out().fpu_reg());
     Register temp = locs()->temp(0).reg();
     __ vcmpd(left, right);
     __ vmstat();
     __ b(&returns_nan, VS);
     __ b(&are_equal, EQ);
-    const intptr_t is_min = (op_kind() == MethodRecognizer::kMathMin);
-    const Condition double_condition =
-        is_min ? TokenKindToDoubleCondition(Token::kLT)
-               : TokenKindToDoubleCondition(Token::kGT);
     const Condition neg_double_condition =
         is_min ? TokenKindToDoubleCondition(Token::kGTE)
                : TokenKindToDoubleCondition(Token::kLTE);
-    __ vmovd(result, left, double_condition);
+    ASSERT(left == result);
     __ vmovd(result, right, neg_double_condition);
     __ b(&done);
 
     __ Bind(&returns_nan);
     __ LoadDImmediate(result, NAN, temp);
     __ b(&done);
 
     __ Bind(&are_equal);
     // Check for negative zero: -0.0 is equal 0.0 but min or max must return
     // -0.0 or 0.0 respectively.
     // Check for negative left value (get the sign bit):
     // - min -> left is negative ? left : right.
     // - max -> left is negative ? right : left
     // Check the sign bit.
     __ vmovrrd(IP, temp, left);  // Sign bit is in bit 31 of temp.
     __ cmp(temp, ShifterOperand(0));
     if (is_min) {
-      __ vmovd(result, left, LT);
+      ASSERT(left == result);
       __ vmovd(result, right, GE);
     } else {
       __ vmovd(result, right, LT);
-      __ vmovd(result, left, GE);
+      ASSERT(left == result);
     }
     __ Bind(&done);
     return;
   }
+
   ASSERT(result_cid() == kSmiCid);
-  UNIMPLEMENTED();
+  Register left = locs()->in(0).reg();
+  Register right = locs()->in(1).reg();
+  Register result = locs()->out().reg();
+  __ cmp(left, ShifterOperand(right));
+  ASSERT(result == left);
+  if (is_min) {
+    __ mov(result, ShifterOperand(right), GE);

[regis, 2013/07/23 19:45:47]

GT would be more intuitive than GE, but it does not make any difference in the
number of cycles or in correctness.

[srdjan, 2013/07/23 20:34:01]

Done.

+  } else {
+    __ mov(result, ShifterOperand(right), LT);
+  }
 }
 
 
 LocationSummary* UnarySmiOpInstr::MakeLocationSummary() const {
   const intptr_t kNumInputs = 1;
   const intptr_t kNumTemps = 0;
   LocationSummary* summary =
       new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
   summary->set_in(0, Location::RequiresRegister());
   // We make use of 3-operand instructions by not requiring result register
@@ skipping 681 matching lines @@
   compiler->GenerateCall(token_pos(),
                          &label,
                          PcDescriptors::kOther,
                          locs());
   __ Drop(2);  // Discard type arguments and receiver.
 }
 
 }  // namespace dart
 
 #endif  // defined TARGET_ARCH_ARM