- Fix sin-cos merge, add a flag to turn it on/off. Using fsincos is much slower than calling out to…

runtime/vm/intermediate_language_ia32.cc

Index: runtime/vm/intermediate_language_ia32.cc
===================================================================
--- runtime/vm/intermediate_language_ia32.cc	(revision 43881)
+++ runtime/vm/intermediate_language_ia32.cc	(working copy)
@@ -5541,12 +5541,18 @@
   }
   if (kind() == MergedMathInstr::kSinCos) {
     const intptr_t kNumInputs = 1;
-    const intptr_t kNumTemps = 0;
+    const intptr_t kNumTemps = 2;
     LocationSummary* summary = new(zone) LocationSummary(
-        zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
-    summary->set_in(0, Location::RequiresFpuRegister());
-    summary->set_out(0, Location::Pair(Location::RequiresFpuRegister(),
-                                       Location::RequiresFpuRegister()));
+        zone, kNumInputs, kNumTemps, LocationSummary::kCall);
+    // Because we always call into the runtime (LocationSummary::kCall) we
+    // must specify each input, temp, and output register explicitly.
+    summary->set_in(0, Location::FpuRegisterLocation(XMM1));
+    // EDI is chosen because it is callee saved so we do not need to back it
+    // up before calling into the runtime.
+    summary->set_temp(0, Location::RegisterLocation(EDI));
+    summary->set_temp(1, Location::RegisterLocation(EBX));
+    summary->set_out(0, Location::Pair(Location::FpuRegisterLocation(XMM2),
+                                       Location::FpuRegisterLocation(XMM3)));
     return summary;
   }
   UNIMPLEMENTED();
@@ -5631,27 +5637,47 @@
   }
 
   if (kind() == MergedMathInstr::kSinCos) {
-    XmmRegister in = locs()->in(0).fpu_reg();
     ASSERT(locs()->out(0).IsPairLocation());
     PairLocation* pair = locs()->out(0).AsPairLocation();
     XmmRegister out1 = pair->At(0).fpu_reg();
     XmmRegister out2 = pair->At(1).fpu_reg();
 
-    // Do x87 sincos, since the ia32 compilers may not fuse sin/cos into
-    // sincos.
-    __ pushl(EAX);
-    __ pushl(EAX);
-    __ movsd(Address(ESP, 0), in);
-    __ fldl(Address(ESP, 0));
-    __ fsincos();
-    __ fstpl(Address(ESP, 0));
-    __ movsd(out1, Address(ESP, 0));
-    __ fstpl(Address(ESP, 0));
-    __ movsd(out2, Address(ESP, 0));
-    __ addl(ESP, Immediate(2 * kWordSize));
+    // Save ESP.
+    __ movl(locs()->temp(0).reg(), ESP);
+    // +-------------------------------+
+    // | double-argument               |  <- TOS
+    // +-------------------------------+
+    // | address-cos-result            |  +8
+    // +-------------------------------+
+    // | address-sin-result            |  +12
+    // +-------------------------------+
+    // | double-storage-for-cos-result |  +16
+    // +-------------------------------+
+    // | double-storage-for-sin-result |  +24
+    // +-------------------------------+
+    // ....
+    __ ReserveAlignedFrameSpace(kDoubleSize * 3 + kWordSize * 2);
+    __ movsd(Address(ESP, 0), locs()->in(0).fpu_reg());
+
+    Address cos_result(ESP, 2 * kWordSize + kDoubleSize);
+    Address sin_result(ESP, 2 * kWordSize + 2 * kDoubleSize);
+
+    // 'cos' result storage address.
+    __ leal(locs()->temp(1).reg(), cos_result);
+    __ movl(Address(ESP, kDoubleSize), locs()->temp(1).reg());
+
+    // 'sin' result storage address.
+    __ leal(locs()->temp(1).reg(), sin_result);
+    __ movl(Address(ESP, kDoubleSize + kWordSize), locs()->temp(1).reg());
+
+    __ CallRuntime(kSinCosRuntimeEntry, InputCount());
+    __ movsd(out2, sin_result);  // sin.
+    __ movsd(out1, cos_result);  // cos.
+    // Restore RSP.
+    __ movl(ESP, locs()->temp(0).reg());
+
     return;
   }
-
   UNIMPLEMENTED();
 }