Inline fsin/fcos. Huge speedup on Box2D.

runtime/vm/intermediate_language_x64.cc

Index: runtime/vm/intermediate_language_x64.cc
===================================================================
--- runtime/vm/intermediate_language_x64.cc	(revision 25800)
+++ runtime/vm/intermediate_language_x64.cc	(working copy)
@@ -3724,7 +3724,7 @@
 }
 
 
-LocationSummary* MathSqrtInstr::MakeLocationSummary() const {
+LocationSummary* MathUnaryInstr::MakeLocationSummary() const {
   const intptr_t kNumInputs = 1;
   const intptr_t kNumTemps = 0;
   LocationSummary* summary =
@@ -3735,8 +3735,29 @@
 }
 
 
-void MathSqrtInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
-  __ sqrtsd(locs()->out().fpu_reg(), locs()->in(0).fpu_reg());
+void MathUnaryInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
+  if (kind() == MethodRecognizer::kMathSqrt) {

[Ivan Posva, 2013/08/05 23:32:16]

ditto

[srdjan, 2013/08/06 15:43:05]

Done.

+    __ sqrtsd(locs()->out().fpu_reg(), locs()->in(0).fpu_reg());
+    return;
+  }
+
+  if ((kind() == MethodRecognizer::kMathCos) ||
+      (kind() == MethodRecognizer::kMathSin)) {
+    __ pushq(RAX);
+    __ movsd(Address(RSP, 0), locs()->in(0).fpu_reg());
+    __ fldl(Address(RSP, 0));
+    if (kind() == MethodRecognizer::kMathSin) {
+      __ fsin();
+    } else {
+      ASSERT(kind() == MethodRecognizer::kMathCos);
+      __ fcos();
+    }
+    __ fstpl(Address(RSP, 0));
+    __ movsd(locs()->out().fpu_reg(), Address(RSP, 0));
+    __ addq(RSP, Immediate(kWordSize));
+    return;
+  }
+  UNREACHABLE();
 }