Inline binary Float32x4 ops.

runtime/vm/intermediate_language_ia32.cc

Index: runtime/vm/intermediate_language_ia32.cc
diff --git a/runtime/vm/intermediate_language_ia32.cc b/runtime/vm/intermediate_language_ia32.cc
index 689f965d55ebccbdc62c85076b7393f045b6224b..caf321681e11680f5f38d0fb542bee17d18027bf 100644
--- a/runtime/vm/intermediate_language_ia32.cc
+++ b/runtime/vm/intermediate_language_ia32.cc
@@ -2768,10 +2768,15 @@ void BoxFloat32x4Instr::EmitNativeCode(FlowGraphCompiler* compiler) {
 
 LocationSummary* UnboxFloat32x4Instr::MakeLocationSummary() const {
   const intptr_t kNumInputs = 1;
-  const intptr_t kNumTemps = 0;
+  const intptr_t value_cid = value()->Type()->ToCid();
+  const intptr_t kNumTemps = value_cid != kFloat32x4Cid ? 1 : 0;
   LocationSummary* summary =
       new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
   summary->set_in(0, Location::RequiresRegister());
+  if (kNumTemps > 0) {
+    ASSERT(kNumTemps == 1);
+    summary->set_temp(0, Location::RequiresRegister());
+  }
   summary->set_out(Location::RequiresFpuRegister());
   return summary;
 }
@@ -2782,7 +2787,14 @@ void UnboxFloat32x4Instr::EmitNativeCode(FlowGraphCompiler* compiler) {
   const Register value = locs()->in(0).reg();
   const XmmRegister result = locs()->out().fpu_reg();
 
-  ASSERT(value_cid == kFloat32x4Cid);
+  if (value_cid != kFloat32x4Cid) {

[Vyacheslav Egorov (Google), 2013/04/15 11:09:35]

See the comment in the optimizer. Emit CheckClass for operands of the binary
operation and restore assert here. 

It should never trigger.

[Cutch, 2013/04/15 13:34:56]

Done.

+    Label* deopt = compiler->AddDeoptStub(deopt_id_, kDeoptCheckClass);
+    Register temp = locs()->temp(0).reg();
+    __ testl(value, Immediate(kSmiTagMask));
+    __ j(ZERO, deopt);
+    __ CompareClassId(value, kFloat32x4Cid, temp);
+    __ j(NOT_EQUAL, deopt);
+  }
   __ movups(result, FieldAddress(value, Float32x4::value_offset()));
 }
 
@@ -2815,6 +2827,33 @@ void BinaryDoubleOpInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
 }
 
 
+LocationSummary* BinaryFloat32x4OpInstr::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 BinaryFloat32x4OpInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
+  XmmRegister left = locs()->in(0).fpu_reg();
+  XmmRegister right = locs()->in(1).fpu_reg();
+
+  ASSERT(locs()->out().fpu_reg() == left);
+
+  switch (op_kind()) {
+    case Token::kADD: __ addps(left, right); break;
+    case Token::kSUB: __ subps(left, right); break;
+    case Token::kMUL: __ mulps(left, right); break;
+    case Token::kDIV: __ divps(left, right); break;
+    default: UNREACHABLE();
+  }
+}
+
 LocationSummary* MathSqrtInstr::MakeLocationSummary() const {
   const intptr_t kNumInputs = 1;
   const intptr_t kNumTemps = 0;