Enable correct optimized double modulo operation. (TODO: enable remainder optimization).

runtime/vm/intermediate_language_x64.cc

Index: runtime/vm/intermediate_language_x64.cc
===================================================================
--- runtime/vm/intermediate_language_x64.cc	(revision 17815)
+++ runtime/vm/intermediate_language_x64.cc	(working copy)
@@ -2421,6 +2421,18 @@
   // currently we can't specify these registers because ParallelMoveResolver
   // assumes that XMM0 is free at all times.
   // TODO(vegorov): allow XMM0 to be used.
+  if (recognized_kind() == MethodRecognizer::kDoubleMod) {
+    ASSERT(InputCount() == 2);
+    const intptr_t kNumTemps = 2;
+    LocationSummary* result =
+        new LocationSummary(InputCount(), kNumTemps, LocationSummary::kCall);
+    result->set_in(0, Location::FpuRegisterLocation(XMM1, Location::kDouble));
+    result->set_in(1, Location::FpuRegisterLocation(XMM2, Location::kDouble));
+    result->set_out(Location::FpuRegisterLocation(XMM1, Location::kDouble));
+    result->set_temp(0, Location::FpuRegisterLocation(XMM2, Location::kDouble));
+    result->set_temp(1, Location::FpuRegisterLocation(XMM3, Location::kDouble));
+    return result;
+  }
   ASSERT((InputCount() == 1) || (InputCount() == 2));
   const intptr_t kNumTemps = 0;
   LocationSummary* result =
@@ -2435,6 +2447,12 @@
 
 
 void InvokeMathCFunctionInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
+  if (recognized_kind() == MethodRecognizer::kDoubleMod) {
+    // Preserve right argument.
+    __ pushq(RAX);
+    __ movsd(Address(RSP, 0), locs()->in(1).fpu_reg());
+  }
+
   ASSERT(locs()->in(0).fpu_reg() == XMM1);
   __ EnterFrame(0);
   __ ReserveAlignedFrameSpace(0);
@@ -2446,6 +2464,39 @@
   __ CallRuntime(TargetFunction());
   __ movaps(locs()->out().fpu_reg(), XMM0);
   __ leave();
+
+  if (recognized_kind() == MethodRecognizer::kDoubleMod) {
+    // Result of C call is remainder, convert it to modulo.
+    Label done, equal_zero, right_greater_equal_zero;
+    XmmRegister remainder = locs()->out().fpu_reg();
+    XmmRegister zero_temp = locs()->temp(0).fpu_reg();
+    XmmRegister right_temp = locs()->temp(1).fpu_reg();
+
+    __ xorpd(zero_temp, zero_temp);  // 0.0 -> 'temp'.
+    __ comisd(remainder, zero_temp);
+    __ j(PARITY_EVEN, &done, Assembler::kNearJump);  // NaN -> false;
+    __ j(EQUAL, &equal_zero, Assembler::kNearJump);
+    __ j(ABOVE, &done, Assembler::kNearJump);  // (remainder) > 0 -> done.
+    // remainder < 0.
+    // Load preserved right argument.
+    __ movsd(right_temp, Address(RSP, 0));
+    __ comisd(right_temp, zero_temp);
+    __ j(ABOVE_EQUAL, &right_greater_equal_zero, Assembler::kNearJump);
+    // right < 0.
+    __ subsd(remainder, right_temp);
+    __ jmp(&done, Assembler::kNearJump);
+
+    __ Bind(&right_greater_equal_zero);
+    __ addsd(remainder, right_temp);
+    __ jmp(&done);
+
+    __ Bind(&equal_zero);
+    // Switch to the positive 0.0 (just in case it was negative).
+    __ movsd(remainder, zero_temp);
+
+    __ Bind(&done);
+    __ Drop(1);  // Remove preserved right argument.
+  }
 }