Support for multiple register values

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 dd7cf89227fee0f72b81ab85a7df2297aae343d6..7699adb41eb78a550ab919e445885d0206b50f01 100644
--- a/runtime/vm/intermediate_language_ia32.cc
+++ b/runtime/vm/intermediate_language_ia32.cc
@@ -4844,8 +4844,57 @@ void InvokeMathCFunctionInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
 }
 
 
-LocationSummary* MergedMathInstr::MakeLocationSummary(bool opt) const {
-  if (kind() == MergedMathInstr::kTruncDivMod) {
+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();
+    __ movl(out, in);
+  }
+}
+
+
+LocationSummary* MergedMath2Instr::MakeLocationSummary(bool opt) const {
+  if (kind() == MergedMath2Instr::kTruncDivMod) {
     const intptr_t kNumInputs = 2;
     const intptr_t kNumTemps = 1;
     LocationSummary* summary =
@@ -4853,18 +4902,21 @@ LocationSummary* MergedMathInstr::MakeLocationSummary(bool opt) const {
     // Both inputs must be writable because they will be untagged.
     summary->set_in(0, Location::RegisterLocation(EAX));
     summary->set_in(1, Location::WritableRegister());
-    summary->set_out(0, Location::RequiresRegister());
+    // Output is a pair of registers.
+    summary->set_out(0, Location::Pair(Location::RequiresRegister(),
+                                       Location::RequiresRegister()));
     // Will be used for sign extension and division.
     summary->set_temp(0, Location::RegisterLocation(EDX));
     return summary;
   }
-  if (kind() == MergedMathInstr::kSinCos) {
+  if (kind() == MergedMath2Instr::kSinCos) {
     const intptr_t kNumInputs = 1;
     const intptr_t kNumTemps = 0;
     LocationSummary* summary =
-        new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kCall);
-    summary->set_in(0, Location::FpuRegisterLocation(XMM1));
-    summary->set_out(0, Location::RegisterLocation(EAX));
+        new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
+    summary->set_in(0, Location::RequiresFpuRegister());
+    summary->set_out(0, Location::Pair(Location::RequiresFpuRegister(),
+                                       Location::RequiresFpuRegister()));
     return summary;
   }
   UNIMPLEMENTED();
@@ -4879,6 +4931,132 @@ extern const RuntimeEntry kSinCosRuntimeEntry(
         static_cast<SinCosCFunction>(&SinCos)), 1, true, true);
 
 
+void MergedMath2Instr::EmitNativeCode(FlowGraphCompiler* compiler) {
+  Label* deopt = NULL;
+  if (CanDeoptimize()) {
+    deopt  = compiler->AddDeoptStub(deopt_id(), kDeoptBinarySmiOp);
+  }
+
+  if (kind() == MergedMath2Instr::kTruncDivMod) {
+    Register left = locs()->in(0).reg();
+    Register right = locs()->in(1).reg();
+    ASSERT(locs()->out(0).IsPairLocation());
+    PairLocation* pair = locs()->out(0).AsPairLocation();
+    Register result1 = pair->At(0).reg();
+    Register result2 = pair->At(1).reg();
+    Range* right_range = InputAt(1)->definition()->range();
+    if ((right_range == NULL) || right_range->Overlaps(0, 0)) {
+      // Handle divide by zero in runtime.
+      __ testl(right, right);
+      __ j(ZERO, deopt);
+    }
+    ASSERT(left == EAX);
+    ASSERT((right != EDX) && (right != EAX));
+    ASSERT(locs()->temp(0).reg() == EDX);
+    ASSERT((result1 != EDX) && (result1 != EAX));
+    ASSERT((result2 != EDX) && (result2 != EAX));
+    __ SmiUntag(left);
+    __ SmiUntag(right);
+    __ cdq();  // Sign extend EAX -> EDX:EAX.
+    __ idivl(right);  //  EAX: quotient, EDX: remainder.
+    // Check the corner case of dividing the 'MIN_SMI' with -1, in which
+    // case we cannot tag the result.
+    // TODO(srdjan): We could store instead untagged intermediate results in a
+    // typed array, but then the load indexed instructions would need to be
+    // able to deoptimize.
+    __ cmpl(EAX, Immediate(0x40000000));
+    __ j(EQUAL, deopt);
+    // Modulo result (EDX) correction:
+    //  res = left % right;
+    //  if (res < 0) {
+    //    if (right < 0) {
+    //      res = res - right;
+    //    } else {
+    //      res = res + right;
+    //    }
+    //  }
+    Label done;
+    __ cmpl(EDX, Immediate(0));
+    __ j(GREATER_EQUAL, &done, Assembler::kNearJump);
+    // Result is negative, adjust it.
+    if ((right_range == NULL) || right_range->Overlaps(-1, 1)) {
+      Label subtract;
+      __ cmpl(right, Immediate(0));
+      __ j(LESS, &subtract, Assembler::kNearJump);
+      __ addl(EDX, right);
+      __ jmp(&done, Assembler::kNearJump);
+      __ Bind(&subtract);
+      __ subl(EDX, right);
+    } else if (right_range->IsWithin(0, RangeBoundary::kPlusInfinity)) {
+      // Right is positive.
+      __ addl(EDX, right);
+    } else {
+      // Right is negative.
+      __ subl(EDX, right);
+    }
+    __ Bind(&done);
+
+    __ SmiTag(EAX);
+    __ SmiTag(EDX);
+    __ movl(result1, EAX);
+    __ movl(result2, EDX);
+    return;
+  }
+
+  if (kind() == MergedMath2Instr::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));
+    return;
+  }
+
+  UNIMPLEMENTED();
+}
+
+
+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(EAX));
+    summary->set_in(1, Location::WritableRegister());
+    summary->set_out(0, Location::RequiresRegister());
+    // Will be used for sign extension and division.
+    summary->set_temp(0, Location::RegisterLocation(EDX));
+    return summary;
+  }
+  if (kind() == MergedMathInstr::kSinCos) {

[Florian Schneider, 2014/04/04 11:52:46]

I wonder how much code is actually shared between SinCos and DivMod. I think
it's too little (only little boilerplate in intermediate_language.h) to have one
class for both. I'd split them up -  maybe inherit from a common
MergedInstruction super-class if at all.

[Cutch, 2014/04/04 16:34:46]

Note that this is a clone of MergedMathInstr that will go away before I land it.
I'd rather cleanup MergedMathInstr in a separate CL.

+    const intptr_t kNumInputs = 1;
+    const intptr_t kNumTemps = 0;
+    LocationSummary* summary =
+        new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kCall);
+    summary->set_in(0, Location::FpuRegisterLocation(XMM1));
+    summary->set_out(0, Location::RegisterLocation(EAX));
+    return summary;
+  }
+  UNIMPLEMENTED();
+  return NULL;
+}
+
+
 void MergedMathInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
   Label* deopt = NULL;
   if (CanDeoptimize()) {