[arm] Improve Float(32|64)(Max|Min).

ARMv8 can use vminnm and vmaxnm to handle most inputs. Other platforms
use an implementation similar to what was there before, except that
out-of-line code is used for the uncommon cases.

BUG=

Review-Url: https://codereview.chromium.org/2313863003
Cr-Commit-Position: refs/heads/master@{#39202}

src/arm/macro-assembler-arm.cc

@@ -250,15 +250,17 @@ void MacroAssembler::Move(Register dst, Register src, Condition cond) {
   }
 }
 
-void MacroAssembler::Move(SwVfpRegister dst, SwVfpRegister src) {
+void MacroAssembler::Move(SwVfpRegister dst, SwVfpRegister src,
+                          Condition cond) {
   if (!dst.is(src)) {
-    vmov(dst, src);
+    vmov(dst, src, cond);
   }
 }
 
-void MacroAssembler::Move(DwVfpRegister dst, DwVfpRegister src) {
+void MacroAssembler::Move(DwVfpRegister dst, DwVfpRegister src,
+                          Condition cond) {
   if (!dst.is(src)) {
-    vmov(dst, src);
+    vmov(dst, src, cond);
   }
 }
 
@@ -3428,6 +3430,144 @@ void MacroAssembler::RestoreFPRegs(Register location, Register scratch) {
   add(location, location, Operand(16 * kDoubleSize), LeaveCC, eq);
 }
 
+template <typename T>
+void MacroAssembler::FloatMaxHelper(T result, T left, T right,
+                                    Label* out_of_line) {
+  // This trivial case is caught sooner, so that the out-of-line code can be
+  // completely avoided.
+  DCHECK(!left.is(right));
+
+  if (CpuFeatures::IsSupported(ARMv8)) {
+    CpuFeatureScope scope(this, ARMv8);
+    VFPCompareAndSetFlags(left, right);
+    b(vs, out_of_line);
+    vmaxnm(result, left, right);
+  } else {
+    Label done;
+    VFPCompareAndSetFlags(left, right);
+    b(vs, out_of_line);
+    // Avoid a conditional instruction if the result register is unique.
+    bool aliased_result_reg = result.is(left) || result.is(right);
+    Move(result, right, aliased_result_reg ? mi : al);
+    Move(result, left, gt);
+    b(ne, &done);
+    // Left and right are equal, but check for +/-0.
+    VFPCompareAndSetFlags(left, 0.0);
+    b(eq, out_of_line);
+    // The arguments are equal and not zero, so it doesn't matter which input we
+    // pick. We have already moved one input into the result (if it didn't
+    // already alias) so there's nothing more to do.
+    bind(&done);
+  }
+}
+
+template <typename T>
+void MacroAssembler::FloatMaxOutOfLineHelper(T result, T left, T right) {
+  DCHECK(!left.is(right));
+
+  // ARMv8: At least one of left and right is a NaN.
+  // Anything else: At least one of left and right is a NaN, or both left and
+  // right are zeroes with unknown sign.
+
+  // If left and right are +/-0, select the one with the most positive sign.
+  // If left or right are NaN, vadd propagates the appropriate one.
+  vadd(result, left, right);
+}
+
+template <typename T>
+void MacroAssembler::FloatMinHelper(T result, T left, T right,
+                                    Label* out_of_line) {
+  // This trivial case is caught sooner, so that the out-of-line code can be
+  // completely avoided.
+  DCHECK(!left.is(right));
+
+  if (CpuFeatures::IsSupported(ARMv8)) {
+    CpuFeatureScope scope(this, ARMv8);
+    VFPCompareAndSetFlags(left, right);
+    b(vs, out_of_line);
+    vminnm(result, left, right);
+  } else {
+    Label done;
+    VFPCompareAndSetFlags(left, right);
+    b(vs, out_of_line);
+    // Avoid a conditional instruction if the result register is unique.
+    bool aliased_result_reg = result.is(left) || result.is(right);
+    Move(result, left, aliased_result_reg ? mi : al);
+    Move(result, right, gt);
+    b(ne, &done);
+    // Left and right are equal, but check for +/-0.
+    VFPCompareAndSetFlags(left, 0.0);
+    // If the arguments are equal and not zero, it doesn't matter which input we
+    // pick. We have already moved one input into the result (if it didn't
+    // already alias) so there's nothing more to do.
+    b(ne, &done);
+    // At this point, both left and right are either 0 or -0.
+    // We could use a single 'vorr' instruction here if we had NEON support.
+    // The algorithm used is -((-L) + (-R)), which is most efficiently expressed
+    // as -((-L) - R).
+    if (left.is(result)) {
+      DCHECK(!right.is(result));
+      vneg(result, left);
+      vsub(result, result, right);
+      vneg(result, result);
+    } else {
+      DCHECK(!left.is(result));
+      vneg(result, right);
+      vsub(result, result, left);
+      vneg(result, result);
+    }
+    bind(&done);
+  }
+}
+
+template <typename T>
+void MacroAssembler::FloatMinOutOfLineHelper(T result, T left, T right) {
+  DCHECK(!left.is(right));
+
+  // At least one of left and right is a NaN. Use vadd to propagate the NaN
+  // appropriately. +/-0 is handled inline.
+  vadd(result, left, right);
+}
+
+void MacroAssembler::FloatMax(SwVfpRegister result, SwVfpRegister left,
+                              SwVfpRegister right, Label* out_of_line) {
+  FloatMaxHelper(result, left, right, out_of_line);
+}
+
+void MacroAssembler::FloatMin(SwVfpRegister result, SwVfpRegister left,
+                              SwVfpRegister right, Label* out_of_line) {
+  FloatMinHelper(result, left, right, out_of_line);
+}
+
+void MacroAssembler::FloatMax(DwVfpRegister result, DwVfpRegister left,
+                              DwVfpRegister right, Label* out_of_line) {
+  FloatMaxHelper(result, left, right, out_of_line);
+}
+
+void MacroAssembler::FloatMin(DwVfpRegister result, DwVfpRegister left,
+                              DwVfpRegister right, Label* out_of_line) {
+  FloatMinHelper(result, left, right, out_of_line);
+}
+
+void MacroAssembler::FloatMaxOutOfLine(SwVfpRegister result, SwVfpRegister left,
+                                       SwVfpRegister right) {
+  FloatMaxOutOfLineHelper(result, left, right);
+}
+
+void MacroAssembler::FloatMinOutOfLine(SwVfpRegister result, SwVfpRegister left,
+                                       SwVfpRegister right) {
+  FloatMinOutOfLineHelper(result, left, right);
+}
+
+void MacroAssembler::FloatMaxOutOfLine(DwVfpRegister result, DwVfpRegister left,
+                                       DwVfpRegister right) {
+  FloatMaxOutOfLineHelper(result, left, right);
+}
+
+void MacroAssembler::FloatMinOutOfLine(DwVfpRegister result, DwVfpRegister left,
+                                       DwVfpRegister right) {
+  FloatMinOutOfLineHelper(result, left, right);
+}
 
 void MacroAssembler::JumpIfBothInstanceTypesAreNotSequentialOneByte(
     Register first, Register second, Register scratch1, Register scratch2,

src/arm/macro-assembler-arm.h

@@ -170,8 +170,8 @@ class MacroAssembler: public Assembler {
       mov(dst, src, sbit, cond);
     }
   }
-  void Move(SwVfpRegister dst, SwVfpRegister src);
-  void Move(DwVfpRegister dst, DwVfpRegister src);
+  void Move(SwVfpRegister dst, SwVfpRegister src, Condition cond = al);
+  void Move(DwVfpRegister dst, DwVfpRegister src, Condition cond = al);
 
   void Load(Register dst, const MemOperand& src, Representation r);
   void Store(Register src, const MemOperand& dst, Representation r);
@@ -1082,6 +1082,32 @@ class MacroAssembler: public Assembler {
   // values to location, restoring [d0..(d15|d31)].
   void RestoreFPRegs(Register location, Register scratch);
 
+  // Perform a floating-point min or max operation with the
+  // (IEEE-754-compatible) semantics of ARM64's fmin/fmax. Some cases, typically
+  // NaNs or +/-0.0, are expected to be rare and are handled in out-of-line
+  // code. The specific behaviour depends on supported instructions.
+  //
+  // These functions assume (and assert) that !left.is(right). It is permitted
+  // for the result to alias either input register.
+  void FloatMax(SwVfpRegister result, SwVfpRegister left, SwVfpRegister right,
+                Label* out_of_line);
+  void FloatMin(SwVfpRegister result, SwVfpRegister left, SwVfpRegister right,
+                Label* out_of_line);
+  void FloatMax(DwVfpRegister result, DwVfpRegister left, DwVfpRegister right,
+                Label* out_of_line);
+  void FloatMin(DwVfpRegister result, DwVfpRegister left, DwVfpRegister right,
+                Label* out_of_line);
+
+  // Generate out-of-line cases for the macros above.
+  void FloatMaxOutOfLine(SwVfpRegister result, SwVfpRegister left,
+                         SwVfpRegister right);
+  void FloatMinOutOfLine(SwVfpRegister result, SwVfpRegister left,
+                         SwVfpRegister right);
+  void FloatMaxOutOfLine(DwVfpRegister result, DwVfpRegister left,
+                         DwVfpRegister right);
+  void FloatMinOutOfLine(DwVfpRegister result, DwVfpRegister left,
+                         DwVfpRegister right);
+
   // ---------------------------------------------------------------------------
   // Runtime calls
 
@@ -1513,6 +1539,16 @@ class MacroAssembler: public Assembler {
   MemOperand SafepointRegisterSlot(Register reg);
   MemOperand SafepointRegistersAndDoublesSlot(Register reg);
 
+  // Implementation helpers for FloatMin and FloatMax.
+  template <typename T>
+  void FloatMaxHelper(T result, T left, T right, Label* out_of_line);
+  template <typename T>
+  void FloatMinHelper(T result, T left, T right, Label* out_of_line);
+  template <typename T>
+  void FloatMaxOutOfLineHelper(T result, T left, T right);
+  template <typename T>
+  void FloatMinOutOfLineHelper(T result, T left, T right);
+
   bool generating_stub_;
   bool has_frame_;
   // This handle will be patched with the code object on installation.

src/compiler/arm/code-generator-arm.cc

@@ -271,6 +271,37 @@ class OutOfLineRecordWrite final : public OutOfLineCode {
   UnwindingInfoWriter* const unwinding_info_writer_;
 };
 
+template <typename T>
+class OutOfLineFloatMin final : public OutOfLineCode {
+ public:
+  OutOfLineFloatMin(CodeGenerator* gen, T result, T left, T right)
+      : OutOfLineCode(gen), result_(result), left_(left), right_(right) {}
+
+  void Generate() final { __ FloatMinOutOfLine(result_, left_, right_); }
+
+ private:
+  T const result_;
+  T const left_;
+  T const right_;
+};
+typedef OutOfLineFloatMin<SwVfpRegister> OutOfLineFloat32Min;
+typedef OutOfLineFloatMin<DwVfpRegister> OutOfLineFloat64Min;
+
+template <typename T>
+class OutOfLineFloatMax final : public OutOfLineCode {
+ public:
+  OutOfLineFloatMax(CodeGenerator* gen, T result, T left, T right)
+      : OutOfLineCode(gen), result_(result), left_(left), right_(right) {}
+
+  void Generate() final { __ FloatMaxOutOfLine(result_, left_, right_); }
+
+ private:
+  T const result_;
+  T const left_;
+  T const right_;
+};
+typedef OutOfLineFloatMax<SwVfpRegister> OutOfLineFloat32Max;
+typedef OutOfLineFloatMax<DwVfpRegister> OutOfLineFloat64Max;
 
 Condition FlagsConditionToCondition(FlagsCondition condition) {
   switch (condition) {
@@ -1377,145 +1408,59 @@ CodeGenerator::CodeGenResult CodeGenerator::AssembleArchInstruction(
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     case kArmFloat32Max: {
-      FloatRegister left_reg = i.InputFloat32Register(0);
-      FloatRegister right_reg = i.InputFloat32Register(1);
-      FloatRegister result_reg = i.OutputFloat32Register();
-      Label result_is_nan, return_left, return_right, check_zero, done;
-      __ VFPCompareAndSetFlags(left_reg, right_reg);
-      __ b(mi, &return_right);
-      __ b(gt, &return_left);
-      __ b(vs, &result_is_nan);
-      // Left equals right => check for -0.
-      __ VFPCompareAndSetFlags(left_reg, 0.0);
-      if (left_reg.is(result_reg) || right_reg.is(result_reg)) {
-        __ b(ne, &done);  // left == right != 0.
+      SwVfpRegister result = i.OutputFloat32Register();
+      SwVfpRegister left = i.InputFloat32Register(0);
+      SwVfpRegister right = i.InputFloat32Register(1);
+      if (left.is(right)) {
+        __ Move(result, left);
       } else {
-        __ b(ne, &return_left);  // left == right != 0.
+        auto ool = new (zone()) OutOfLineFloat32Max(this, result, left, right);
+        __ FloatMax(result, left, right, ool->entry());
+        __ bind(ool->exit());
       }
-      // At this point, both left and right are either 0 or -0.
-      // Since we operate on +0 and/or -0, vadd and vand have the same effect;
-      // the decision for vadd is easy because vand is a NEON instruction.
-      __ vadd(result_reg, left_reg, right_reg);
-      __ b(&done);
-      __ bind(&result_is_nan);
-      __ vadd(result_reg, left_reg, right_reg);
-      __ b(&done);
-      __ bind(&return_right);
-      __ Move(result_reg, right_reg);
-      if (!left_reg.is(result_reg)) __ b(&done);
-      __ bind(&return_left);
-      __ Move(result_reg, left_reg);
-      __ bind(&done);
+      DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     }
     case kArmFloat64Max: {
-      DwVfpRegister left_reg = i.InputDoubleRegister(0);
-      DwVfpRegister right_reg = i.InputDoubleRegister(1);
-      DwVfpRegister result_reg = i.OutputDoubleRegister();
-      Label result_is_nan, return_left, return_right, check_zero, done;
-      __ VFPCompareAndSetFlags(left_reg, right_reg);
-      __ b(mi, &return_right);
-      __ b(gt, &return_left);
-      __ b(vs, &result_is_nan);
-      // Left equals right => check for -0.
-      __ VFPCompareAndSetFlags(left_reg, 0.0);
-      if (left_reg.is(result_reg) || right_reg.is(result_reg)) {
-        __ b(ne, &done);  // left == right != 0.
+      DwVfpRegister result = i.OutputDoubleRegister();
+      DwVfpRegister left = i.InputDoubleRegister(0);
+      DwVfpRegister right = i.InputDoubleRegister(1);
+      if (left.is(right)) {
+        __ Move(result, left);
       } else {
-        __ b(ne, &return_left);  // left == right != 0.
+        auto ool = new (zone()) OutOfLineFloat64Max(this, result, left, right);
+        __ FloatMax(result, left, right, ool->entry());
+        __ bind(ool->exit());
       }
-      // At this point, both left and right are either 0 or -0.
-      // Since we operate on +0 and/or -0, vadd and vand have the same effect;
-      // the decision for vadd is easy because vand is a NEON instruction.
-      __ vadd(result_reg, left_reg, right_reg);
-      __ b(&done);
-      __ bind(&result_is_nan);
-      __ vadd(result_reg, left_reg, right_reg);
-      __ b(&done);
-      __ bind(&return_right);
-      __ Move(result_reg, right_reg);
-      if (!left_reg.is(result_reg)) __ b(&done);
-      __ bind(&return_left);
-      __ Move(result_reg, left_reg);
-      __ bind(&done);
+      DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     }
     case kArmFloat32Min: {
-      FloatRegister left_reg = i.InputFloat32Register(0);
-      FloatRegister right_reg = i.InputFloat32Register(1);
-      FloatRegister result_reg = i.OutputFloat32Register();
-      Label result_is_nan, return_left, return_right, check_zero, done;
-      __ VFPCompareAndSetFlags(left_reg, right_reg);
-      __ b(mi, &return_left);
-      __ b(gt, &return_right);
-      __ b(vs, &result_is_nan);
-      // Left equals right => check for -0.
-      __ VFPCompareAndSetFlags(left_reg, 0.0);
-      if (left_reg.is(result_reg) || right_reg.is(result_reg)) {
-        __ b(ne, &done);  // left == right != 0.
-      } else {
-        __ b(ne, &return_left);  // left == right != 0.
-      }
-      // At this point, both left and right are either 0 or -0.
-      // We could use a single 'vorr' instruction here if we had NEON support.
-      // The algorithm is: -((-L) + (-R)), which in case of L and R being
-      // different registers is most efficiently expressed as -((-L) - R).
-      __ vneg(left_reg, left_reg);
-      if (left_reg.is(right_reg)) {
-        __ vadd(result_reg, left_reg, right_reg);
+      SwVfpRegister result = i.OutputFloat32Register();
+      SwVfpRegister left = i.InputFloat32Register(0);
+      SwVfpRegister right = i.InputFloat32Register(1);
+      if (left.is(right)) {
+        __ Move(result, left);
       } else {
-        __ vsub(result_reg, left_reg, right_reg);
+        auto ool = new (zone()) OutOfLineFloat32Min(this, result, left, right);
+        __ FloatMin(result, left, right, ool->entry());
+        __ bind(ool->exit());
       }
-      __ vneg(result_reg, result_reg);
-      __ b(&done);
-      __ bind(&result_is_nan);
-      __ vadd(result_reg, left_reg, right_reg);
-      __ b(&done);
-      __ bind(&return_right);
-      __ Move(result_reg, right_reg);
-      if (!left_reg.is(result_reg)) __ b(&done);
-      __ bind(&return_left);
-      __ Move(result_reg, left_reg);
-      __ bind(&done);
+      DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     }
     case kArmFloat64Min: {
-      DwVfpRegister left_reg = i.InputDoubleRegister(0);
-      DwVfpRegister right_reg = i.InputDoubleRegister(1);
-      DwVfpRegister result_reg = i.OutputDoubleRegister();
-      Label result_is_nan, return_left, return_right, check_zero, done;
-      __ VFPCompareAndSetFlags(left_reg, right_reg);
-      __ b(mi, &return_left);
-      __ b(gt, &return_right);
-      __ b(vs, &result_is_nan);
-      // Left equals right => check for -0.
-      __ VFPCompareAndSetFlags(left_reg, 0.0);
-      if (left_reg.is(result_reg) || right_reg.is(result_reg)) {
-        __ b(ne, &done);  // left == right != 0.
-      } else {
-        __ b(ne, &return_left);  // left == right != 0.
-      }
-      // At this point, both left and right are either 0 or -0.
-      // We could use a single 'vorr' instruction here if we had NEON support.
-      // The algorithm is: -((-L) + (-R)), which in case of L and R being
-      // different registers is most efficiently expressed as -((-L) - R).
-      __ vneg(left_reg, left_reg);
-      if (left_reg.is(right_reg)) {
-        __ vadd(result_reg, left_reg, right_reg);
+      DwVfpRegister result = i.OutputDoubleRegister();
+      DwVfpRegister left = i.InputDoubleRegister(0);
+      DwVfpRegister right = i.InputDoubleRegister(1);
+      if (left.is(right)) {
+        __ Move(result, left);
       } else {
-        __ vsub(result_reg, left_reg, right_reg);
+        auto ool = new (zone()) OutOfLineFloat64Min(this, result, left, right);
+        __ FloatMin(result, left, right, ool->entry());
+        __ bind(ool->exit());
       }
-      __ vneg(result_reg, result_reg);
-      __ b(&done);
-      __ bind(&result_is_nan);
-      __ vadd(result_reg, left_reg, right_reg);
-      __ b(&done);
-      __ bind(&return_right);
-      __ Move(result_reg, right_reg);
-      if (!left_reg.is(result_reg)) __ b(&done);
-      __ bind(&return_left);
-      __ Move(result_reg, left_reg);
-      __ bind(&done);
+      DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     }
     case kArmFloat64SilenceNaN: {

test/cctest/test-assembler-arm.cc

@@ -30,11 +30,11 @@
 #include "src/v8.h"
 #include "test/cctest/cctest.h"
 
-#include "src/arm/assembler-arm-inl.h"
 #include "src/arm/simulator-arm.h"
 #include "src/base/utils/random-number-generator.h"
 #include "src/disassembler.h"
 #include "src/factory.h"
+#include "src/macro-assembler.h"
 #include "src/ostreams.h"
 
 using namespace v8::base;
@@ -2382,7 +2382,7 @@ TEST(ARMv8_vsel) {
 }
 
 TEST(ARMv8_vminmax_f64) {
-  // Test the vsel floating point instructions.
+  // Test the vminnm and vmaxnm floating point instructions.
   CcTest::InitializeVM();
   Isolate* isolate = CcTest::i_isolate();
   HandleScope scope(isolate);
@@ -2464,7 +2464,7 @@ TEST(ARMv8_vminmax_f64) {
 }
 
 TEST(ARMv8_vminmax_f32) {
-  // Test the vsel floating point instructions.
+  // Test the vminnm and vmaxnm floating point instructions.
   CcTest::InitializeVM();
   Isolate* isolate = CcTest::i_isolate();
   HandleScope scope(isolate);
@@ -2545,6 +2545,236 @@ TEST(ARMv8_vminmax_f32) {
   }
 }
 
+template <typename T, typename Inputs, typename Results>
+static F4 GenerateMacroFloatMinMax(MacroAssembler& assm) {
+  T a = T::from_code(0);  // d0/s0
+  T b = T::from_code(1);  // d1/s1
+  T c = T::from_code(2);  // d2/s2
+
+  // Create a helper function:
+  //  void TestFloatMinMax(const Inputs* inputs,
+  //                       Results* results);
+  Label ool_min_abc, ool_min_aab, ool_min_aba;
+  Label ool_max_abc, ool_max_aab, ool_max_aba;
+
+  Label done_min_abc, done_min_aab, done_min_aba;
+  Label done_max_abc, done_max_aab, done_max_aba;
+
+  // a = min(b, c);
+  __ vldr(b, r0, offsetof(Inputs, left_));
+  __ vldr(c, r0, offsetof(Inputs, right_));
+  __ FloatMin(a, b, c, &ool_min_abc);
+  __ bind(&done_min_abc);
+  __ vstr(a, r1, offsetof(Results, min_abc_));
+
+  // a = min(a, b);
+  __ vldr(a, r0, offsetof(Inputs, left_));
+  __ vldr(b, r0, offsetof(Inputs, right_));
+  __ FloatMin(a, a, b, &ool_min_aab);
+  __ bind(&done_min_aab);
+  __ vstr(a, r1, offsetof(Results, min_aab_));
+
+  // a = min(b, a);
+  __ vldr(b, r0, offsetof(Inputs, left_));
+  __ vldr(a, r0, offsetof(Inputs, right_));
+  __ FloatMin(a, b, a, &ool_min_aba);
+  __ bind(&done_min_aba);
+  __ vstr(a, r1, offsetof(Results, min_aba_));
+
+  // a = max(b, c);
+  __ vldr(b, r0, offsetof(Inputs, left_));
+  __ vldr(c, r0, offsetof(Inputs, right_));
+  __ FloatMax(a, b, c, &ool_max_abc);
+  __ bind(&done_max_abc);
+  __ vstr(a, r1, offsetof(Results, max_abc_));
+
+  // a = max(a, b);
+  __ vldr(a, r0, offsetof(Inputs, left_));
+  __ vldr(b, r0, offsetof(Inputs, right_));
+  __ FloatMax(a, a, b, &ool_max_aab);
+  __ bind(&done_max_aab);
+  __ vstr(a, r1, offsetof(Results, max_aab_));
+
+  // a = max(b, a);
+  __ vldr(b, r0, offsetof(Inputs, left_));
+  __ vldr(a, r0, offsetof(Inputs, right_));
+  __ FloatMax(a, b, a, &ool_max_aba);
+  __ bind(&done_max_aba);
+  __ vstr(a, r1, offsetof(Results, max_aba_));
+
+  __ bx(lr);
+
+  // Generate out-of-line cases.
+  __ bind(&ool_min_abc);
+  __ FloatMinOutOfLine(a, b, c);
+  __ b(&done_min_abc);
+
+  __ bind(&ool_min_aab);
+  __ FloatMinOutOfLine(a, a, b);
+  __ b(&done_min_aab);
+
+  __ bind(&ool_min_aba);
+  __ FloatMinOutOfLine(a, b, a);
+  __ b(&done_min_aba);
+
+  __ bind(&ool_max_abc);
+  __ FloatMaxOutOfLine(a, b, c);
+  __ b(&done_max_abc);
+
+  __ bind(&ool_max_aab);
+  __ FloatMaxOutOfLine(a, a, b);
+  __ b(&done_max_aab);
+
+  __ bind(&ool_max_aba);
+  __ FloatMaxOutOfLine(a, b, a);
+  __ b(&done_max_aba);
+
+  CodeDesc desc;
+  assm.GetCode(&desc);
+  Handle<Code> code = assm.isolate()->factory()->NewCode(
+      desc, Code::ComputeFlags(Code::STUB), Handle<Code>());
+#ifdef DEBUG
+  OFStream os(stdout);
+  code->Print(os);
+#endif
+  return FUNCTION_CAST<F4>(code->entry());
+}
+
+TEST(macro_float_minmax_f64) {
+  // Test the FloatMin and FloatMax macros.
+  CcTest::InitializeVM();
+  Isolate* isolate = CcTest::i_isolate();
+  HandleScope scope(isolate);
+
+  MacroAssembler assm(isolate, NULL, 0, CodeObjectRequired::kYes);
+
+  struct Inputs {
+    double left_;
+    double right_;
+  };
+
+  struct Results {
+    // Check all register aliasing possibilities in order to exercise all
+    // code-paths in the macro assembler.
+    double min_abc_;
+    double min_aab_;
+    double min_aba_;
+    double max_abc_;
+    double max_aab_;
+    double max_aba_;
+  };
+
+  F4 f = GenerateMacroFloatMinMax<DwVfpRegister, Inputs, Results>(assm);
+
+  Object* dummy = nullptr;
+  USE(dummy);
+
+#define CHECK_MINMAX(left, right, min, max)                                  \
+  do {                                                                       \
+    Inputs inputs = {left, right};                                           \
+    Results results;                                                         \
+    dummy = CALL_GENERATED_CODE(isolate, f, &inputs, &results, 0, 0, 0);     \
+    /* Use a bit_cast to correctly identify -0.0 and NaNs. */                \
+    CHECK_EQ(bit_cast<uint64_t>(min), bit_cast<uint64_t>(results.min_abc_)); \
+    CHECK_EQ(bit_cast<uint64_t>(min), bit_cast<uint64_t>(results.min_aab_)); \
+    CHECK_EQ(bit_cast<uint64_t>(min), bit_cast<uint64_t>(results.min_aba_)); \
+    CHECK_EQ(bit_cast<uint64_t>(max), bit_cast<uint64_t>(results.max_abc_)); \
+    CHECK_EQ(bit_cast<uint64_t>(max), bit_cast<uint64_t>(results.max_aab_)); \
+    CHECK_EQ(bit_cast<uint64_t>(max), bit_cast<uint64_t>(results.max_aba_)); \
+  } while (0)
+
+  double nan_a = bit_cast<double>(UINT64_C(0x7ff8000000000001));
+  double nan_b = bit_cast<double>(UINT64_C(0x7ff8000000000002));
+
+  CHECK_MINMAX(1.0, -1.0, -1.0, 1.0);
+  CHECK_MINMAX(-1.0, 1.0, -1.0, 1.0);
+  CHECK_MINMAX(0.0, -1.0, -1.0, 0.0);
+  CHECK_MINMAX(-1.0, 0.0, -1.0, 0.0);
+  CHECK_MINMAX(-0.0, -1.0, -1.0, -0.0);
+  CHECK_MINMAX(-1.0, -0.0, -1.0, -0.0);
+  CHECK_MINMAX(0.0, 1.0, 0.0, 1.0);
+  CHECK_MINMAX(1.0, 0.0, 0.0, 1.0);
+
+  CHECK_MINMAX(0.0, 0.0, 0.0, 0.0);
+  CHECK_MINMAX(-0.0, -0.0, -0.0, -0.0);
+  CHECK_MINMAX(-0.0, 0.0, -0.0, 0.0);
+  CHECK_MINMAX(0.0, -0.0, -0.0, 0.0);
+
+  CHECK_MINMAX(0.0, nan_a, nan_a, nan_a);
+  CHECK_MINMAX(nan_a, 0.0, nan_a, nan_a);
+  CHECK_MINMAX(nan_a, nan_b, nan_a, nan_a);
+  CHECK_MINMAX(nan_b, nan_a, nan_b, nan_b);
+
+#undef CHECK_MINMAX
+}
+
+TEST(macro_float_minmax_f32) {
+  // Test the FloatMin and FloatMax macros.
+  CcTest::InitializeVM();
+  Isolate* isolate = CcTest::i_isolate();
+  HandleScope scope(isolate);
+
+  MacroAssembler assm(isolate, NULL, 0, CodeObjectRequired::kYes);
+
+  struct Inputs {
+    float left_;
+    float right_;
+  };
+
+  struct Results {
+    // Check all register aliasing possibilities in order to exercise all
+    // code-paths in the macro assembler.
+    float min_abc_;
+    float min_aab_;
+    float min_aba_;
+    float max_abc_;
+    float max_aab_;
+    float max_aba_;
+  };
+
+  F4 f = GenerateMacroFloatMinMax<SwVfpRegister, Inputs, Results>(assm);
+  Object* dummy = nullptr;
+  USE(dummy);
+
+#define CHECK_MINMAX(left, right, min, max)                                  \
+  do {                                                                       \
+    Inputs inputs = {left, right};                                           \
+    Results results;                                                         \
+    dummy = CALL_GENERATED_CODE(isolate, f, &inputs, &results, 0, 0, 0);     \
+    /* Use a bit_cast to correctly identify -0.0 and NaNs. */                \
+    CHECK_EQ(bit_cast<uint32_t>(min), bit_cast<uint32_t>(results.min_abc_)); \
+    CHECK_EQ(bit_cast<uint32_t>(min), bit_cast<uint32_t>(results.min_aab_)); \
+    CHECK_EQ(bit_cast<uint32_t>(min), bit_cast<uint32_t>(results.min_aba_)); \
+    CHECK_EQ(bit_cast<uint32_t>(max), bit_cast<uint32_t>(results.max_abc_)); \
+    CHECK_EQ(bit_cast<uint32_t>(max), bit_cast<uint32_t>(results.max_aab_)); \
+    CHECK_EQ(bit_cast<uint32_t>(max), bit_cast<uint32_t>(results.max_aba_)); \
+  } while (0)
+
+  float nan_a = bit_cast<float>(UINT32_C(0x7fc00001));
+  float nan_b = bit_cast<float>(UINT32_C(0x7fc00002));
+
+  CHECK_MINMAX(1.0f, -1.0f, -1.0f, 1.0f);
+  CHECK_MINMAX(-1.0f, 1.0f, -1.0f, 1.0f);
+  CHECK_MINMAX(0.0f, -1.0f, -1.0f, 0.0f);
+  CHECK_MINMAX(-1.0f, 0.0f, -1.0f, 0.0f);
+  CHECK_MINMAX(-0.0f, -1.0f, -1.0f, -0.0f);
+  CHECK_MINMAX(-1.0f, -0.0f, -1.0f, -0.0f);
+  CHECK_MINMAX(0.0f, 1.0f, 0.0f, 1.0f);
+  CHECK_MINMAX(1.0f, 0.0f, 0.0f, 1.0f);
+
+  CHECK_MINMAX(0.0f, 0.0f, 0.0f, 0.0f);
+  CHECK_MINMAX(-0.0f, -0.0f, -0.0f, -0.0f);
+  CHECK_MINMAX(-0.0f, 0.0f, -0.0f, 0.0f);
+  CHECK_MINMAX(0.0f, -0.0f, -0.0f, 0.0f);
+
+  CHECK_MINMAX(0.0f, nan_a, nan_a, nan_a);
+  CHECK_MINMAX(nan_a, 0.0f, nan_a, nan_a);
+  CHECK_MINMAX(nan_a, nan_b, nan_a, nan_a);
+  CHECK_MINMAX(nan_b, nan_a, nan_b, nan_b);
+
+#undef CHECK_MINMAX
+}
+
 TEST(unaligned_loads) {
   // All supported ARM targets allow unaligned accesses.
   CcTest::InitializeVM();