Support for multiple register values

runtime/vm/intermediate_language_x64.cc

 // Copyright (c) 2013, the Dart project authors.  Please see the AUTHORS file
 // for details. All rights reserved. Use of this source code is governed by a
 // BSD-style license that can be found in the LICENSE file.
 
 #include "vm/globals.h"  // Needed here to get TARGET_ARCH_X64.
 #if defined(TARGET_ARCH_X64)
 
 #include "vm/intermediate_language.h"
 
 #include "vm/dart_entry.h"
@@ skipping 4734 matching lines @@
     __ Bind(&do_pow);
   }
   __ CallRuntime(TargetFunction(), InputCount());
   __ movaps(locs()->out(0).fpu_reg(), XMM0);
   __ Bind(&skip_call);
   // Restore RSP.
   __ movq(RSP, locs()->temp(kSavedSpTempIndex).reg());
 }
 
 
+LocationSummary* ExtractNthOutputInstr::MakeLocationSummary(bool opt) const {
+  // Only use this instruction in optimized code.
+  ASSERT(opt);
+  const intptr_t kNumInputs = 1;
+  LocationSummary* summary =
+      new LocationSummary(kNumInputs, 0, LocationSummary::kNoCall);
+  if (representation() == kUnboxedDouble) {
+    if (index() == 0) {
+      summary->set_in(0, Location::Pair(Location::RequiresFpuRegister(),
+                                        Location::Any()));
+    } else {
+      ASSERT(index() == 1);
+      summary->set_in(0, Location::Pair(Location::Any(),
+                                        Location::RequiresFpuRegister()));
+    }
+    summary->set_out(0, Location::RequiresFpuRegister());
+  } else {
+    ASSERT(representation() == kTagged);
+    if (index() == 0) {
+      summary->set_in(0, Location::Pair(Location::RequiresRegister(),
+                                        Location::Any()));
+    } else {
+      ASSERT(index() == 1);
+      summary->set_in(0, Location::Pair(Location::Any(),
+                                        Location::RequiresRegister()));
+    }
+    summary->set_out(0, Location::RequiresRegister());
+  }
+  return summary;
+}
+
+
+void ExtractNthOutputInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
+  ASSERT(locs()->in(0).IsPairLocation());
+  PairLocation* pair = locs()->in(0).AsPairLocation();
+  Location in_loc = pair->At(index());
+  if (representation() == kUnboxedDouble) {
+    XmmRegister out = locs()->out(0).fpu_reg();
+    XmmRegister in = in_loc.fpu_reg();
+    __ movaps(out, in);
+  } else {
+    ASSERT(representation() == kTagged);
+    Register out = locs()->out(0).reg();
+    Register in = in_loc.reg();
+    __ movq(out, in);
+  }
+}
+
+
 LocationSummary* MergedMathInstr::MakeLocationSummary(bool opt) const {
   if (kind() == MergedMathInstr::kTruncDivMod) {
     const intptr_t kNumInputs = 2;
     const intptr_t kNumTemps = 1;
     LocationSummary* summary =
         new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
     // Both inputs must be writable because they will be untagged.
     summary->set_in(0, Location::RegisterLocation(RAX));
     summary->set_in(1, Location::WritableRegister());
-    summary->set_out(0, Location::RequiresRegister());
+    summary->set_out(0, Location::Pair(Location::RequiresRegister(),

[Florian Schneider, 2014/04/08 09:48:45]

Same here: It should be possible to have the output in RDX:RAX.

[Cutch, 2014/04/08 15:56:16]

Done.

+                                       Location::RequiresRegister()));
     // Will be used for sign extension and division.
     summary->set_temp(0, Location::RegisterLocation(RDX));
     return summary;
   }
   if (kind() == MergedMathInstr::kSinCos) {
     const intptr_t kNumInputs = 1;
     const intptr_t kNumTemps = 1;
     LocationSummary* summary =
         new LocationSummary(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));
     // R13 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(R13));
-    summary->set_out(0, Location::RegisterLocation(RAX));
+    summary->set_out(0, Location::Pair(Location::FpuRegisterLocation(XMM2),
+                                       Location::FpuRegisterLocation(XMM3)));
     return summary;
   }
   UNIMPLEMENTED();
   return NULL;
 }
 
 
 
 typedef void (*SinCosCFunction) (double x, double* res_sin, double* res_cos);
 
 extern const RuntimeEntry kSinCosRuntimeEntry(
     "libc_sincos", reinterpret_cast<RuntimeFunction>(
         static_cast<SinCosCFunction>(&SinCos)), 1, true, true);
 
 
 void MergedMathInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
   Label* deopt = NULL;
   if (CanDeoptimize()) {
     deopt  = compiler->AddDeoptStub(deopt_id(), kDeoptBinarySmiOp);
   }
   if (kind() == MergedMathInstr::kTruncDivMod) {
     Register left = locs()->in(0).reg();
     Register right = locs()->in(1).reg();
-    Register result = locs()->out(0).reg();
+    ASSERT(locs()->out(0).IsPairLocation());
+    PairLocation* pair = locs()->out(0).AsPairLocation();
+    Register result1 = pair->At(0).reg();
+    Register result2 = pair->At(1).reg();
     Label not_32bit, done;
     Register temp = locs()->temp(0).reg();
     ASSERT(left == RAX);
     ASSERT((right != RDX) && (right != RAX));
     ASSERT(temp == RDX);
-    ASSERT((result != RDX) && (result != RAX));
+    ASSERT((result1 != RDX) && (result1 != RAX));
+    ASSERT((result2 != RDX) && (result2 != RAX));
 
     Range* right_range = InputAt(1)->definition()->range();
     if ((right_range == NULL) || right_range->Overlaps(0, 0)) {
       // Handle divide by zero in runtime.
       __ testq(right, right);
       __ j(ZERO, deopt);
     }
     // Check if both operands fit into 32bits as idiv with 64bit operands
     // requires twice as many cycles and has much higher latency.
     // We are checking this before untagging them to avoid corner case
@@ skipping 49 matching lines @@
       __ Bind(&subtract);
       __ subq(RDX, right);
     } else if (right_range->IsWithin(0, RangeBoundary::kPlusInfinity)) {
       // Right is positive.
       __ addq(RDX, right);
     } else {
       // Right is negative.
       __ subq(RDX, right);
     }
     __ Bind(&all_done);
-    __ SmiTag(result);
 
-    __ LoadObject(result, Array::ZoneHandle(Array::New(2, Heap::kOld)), PP);
-    const intptr_t index_scale = FlowGraphCompiler::ElementSizeFor(kArrayCid);
-    Address trunc_div_address(
-        FlowGraphCompiler::ElementAddressForIntIndex(
-            kArrayCid, index_scale, result,
-            MergedMathInstr::ResultIndexOf(Token::kTRUNCDIV)));
-    Address mod_address(
-        FlowGraphCompiler::ElementAddressForIntIndex(
-            kArrayCid, index_scale, result,
-            MergedMathInstr::ResultIndexOf(Token::kMOD)));
     __ SmiTag(RAX);
     __ SmiTag(RDX);
-    __ StoreIntoObjectNoBarrier(result, trunc_div_address, RAX);
-    __ StoreIntoObjectNoBarrier(result, mod_address, RDX);
+
+    __ movq(result1, RAX);
+    __ movq(result2, RDX);
     // FLAG_throw_on_javascript_int_overflow: not needed.
     // Note that the result of an integer division/modulo of two
     // in-range arguments, cannot create out-of-range result.
     return;
   }
   if (kind() == MergedMathInstr::kSinCos) {
+    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();
+
     // Save RSP.
     __ movq(locs()->temp(0).reg(), RSP);
     // +-------------------------------+
     // | double-argument               |  <- TOS
     // +-------------------------------+
     // | address-cos-result            |  +8
     // +-------------------------------+
     // | address-sin-result            |  +16
     // +-------------------------------+
     // | double-storage-for-cos-result |  +24
     // +-------------------------------+
     // | double-storage-for-sin-result |  +32
     // +-------------------------------+
     // ....
     __ ReserveAlignedFrameSpace(kDoubleSize * 3 + kWordSize * 2);
     __ movsd(Address(RSP, 0), locs()->in(0).fpu_reg());
 
     __ leaq(RDI, Address(RSP, 2 * kWordSize + kDoubleSize));
     __ leaq(RSI, Address(RSP, 2 * kWordSize + 2 * kDoubleSize));
     __ movaps(XMM0, locs()->in(0).fpu_reg());
 
     __ CallRuntime(kSinCosRuntimeEntry, InputCount());
-    __ movsd(XMM0, Address(RSP, 2 * kWordSize + kDoubleSize * 2));  // sin.
-    __ movsd(XMM1, Address(RSP, 2 * kWordSize + kDoubleSize));  // cos.
+    __ movsd(out2, Address(RSP, 2 * kWordSize + kDoubleSize * 2));  // sin.
+    __ movsd(out1, Address(RSP, 2 * kWordSize + kDoubleSize));  // cos.
     // Restore RSP.
     __ movq(RSP, locs()->temp(0).reg());
 
-    Register result = locs()->out(0).reg();
-    const TypedData& res_array = TypedData::ZoneHandle(
-      TypedData::New(kTypedDataFloat64ArrayCid, 2, Heap::kOld));
-    __ LoadObject(result, res_array, PP);
-    const intptr_t index_scale =
-        FlowGraphCompiler::ElementSizeFor(kTypedDataFloat64ArrayCid);
-    Address sin_address(
-        FlowGraphCompiler::ElementAddressForIntIndex(
-            kTypedDataFloat64ArrayCid, index_scale, result,
-            MergedMathInstr::ResultIndexOf(MethodRecognizer::kMathSin)));
-    Address cos_address(
-        FlowGraphCompiler::ElementAddressForIntIndex(
-            kTypedDataFloat64ArrayCid, index_scale, result,
-            MergedMathInstr::ResultIndexOf(MethodRecognizer::kMathCos)));
-    __ movsd(sin_address, XMM0);
-    __ movsd(cos_address, XMM1);
     return;
   }
   UNIMPLEMENTED();
 }
 
 
 LocationSummary* PolymorphicInstanceCallInstr::MakeLocationSummary(
     bool opt) const {
   return MakeCallSummary();
 }
@@ skipping 476 matching lines @@
                          PcDescriptors::kOther,
                          locs());
   __ Drop(ArgumentCount());  // Discard arguments.
 }
 
 }  // namespace dart
 
 #undef __
 
 #endif  // defined TARGET_ARCH_X64