Reland "Fix phis for non-sse2 double values"

Remove VerifyX87StackDepth from non-debug code.

BUG=
R=verwaest@chromium.org

Review URL: https://codereview.chromium.org/23621038

git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@16682 ce2b1a6d-e550-0410-aec6-3dcde31c8c00

src/ia32/lithium-codegen-ia32.cc

@@ -382,9 +382,13 @@ bool LCodeGen::GenerateBody() {
 
     instr->CompileToNative(this);
 
-    if (!CpuFeatures::IsSupported(SSE2) &&
-        FLAG_debug_code && FLAG_enable_slow_asserts) {
+    if (!CpuFeatures::IsSupported(SSE2)) {
+      if (instr->IsGoto()) {
+        x87_stack_.LeavingBlock(current_block_, LGoto::cast(instr));
+      } else if (FLAG_debug_code && FLAG_enable_slow_asserts &&
+                 !instr->IsGap() && !instr->IsReturn()) {
         __ VerifyX87StackDepth(x87_stack_.depth());
+      }
     }
   }
   EnsureSpaceForLazyDeopt();
@@ -682,6 +686,21 @@ void LCodeGen::X87Stack::FlushIfNecessary(LInstruction* instr, LCodeGen* cgen) {
       __ fstp(0);
       stack_depth_--;
     }
+    if (FLAG_debug_code && FLAG_enable_slow_asserts) __ VerifyX87StackDepth(0);
+  }
+}
+
+
+void LCodeGen::X87Stack::LeavingBlock(int current_block_id, LGoto* goto_instr) {
+  ASSERT(stack_depth_ <= 1);
+  // If ever used for new stubs producing two pairs of doubles joined into two
+  // phis this assert hits. That situation is not handled, since the two stacks
+  // might have st0 and st1 swapped.
+  if (current_block_id + 1 != goto_instr->block_id()) {
+    // If we have a value on the x87 stack on leaving a block, it must be a
+    // phi input. If the next block we compile is not the join block, we have
+    // to discard the stack state.
+    stack_depth_ = 0;
   }
 }
 
@@ -2486,6 +2505,10 @@ void LCodeGen::EmitGoto(int block) {
 }
 
 
+void LCodeGen::DoClobberDoubles(LClobberDoubles* instr) {
+}
+
+
 void LCodeGen::DoGoto(LGoto* instr) {
   EmitGoto(instr->block_id());
 }

src/ia32/lithium-codegen-ia32.h

@@ -471,6 +471,7 @@ class LCodeGen V8_FINAL BASE_EMBEDDED {
     void PrepareToWrite(X87Register reg);
     void CommitWrite(X87Register reg);
     void FlushIfNecessary(LInstruction* instr, LCodeGen* cgen);
+    void LeavingBlock(int current_block_id, LGoto* goto_instr);
     int depth() const { return stack_depth_; }
     void pop() {
       ASSERT(is_mutable_);

src/ia32/lithium-ia32.cc

@@ -954,6 +954,16 @@ void LChunkBuilder::VisitInstruction(HInstruction* current) {
     if (FLAG_stress_environments && !instr->HasEnvironment()) {
       instr = AssignEnvironment(instr);
     }
+    if (!CpuFeatures::IsSafeForSnapshot(SSE2) && instr->IsGoto() &&
+        LGoto::cast(instr)->jumps_to_join()) {
+      // TODO(olivf) Since phis of spilled values are joined as registers
+      // (not in the stack slot), we need to allow the goto gaps to keep one
+      // x87 register alive. To ensure all other values are still spilled, we
+      // insert a fpu register barrier right before.
+      LClobberDoubles* clobber = new(zone()) LClobberDoubles();
+      clobber->set_hydrogen_value(current);
+      chunk_->AddInstruction(clobber, current_block_);
+    }
     instr->set_hydrogen_value(current);
     chunk_->AddInstruction(instr, current_block_);
   }
@@ -1046,7 +1056,7 @@ LEnvironment* LChunkBuilder::CreateEnvironment(
 
 
 LInstruction* LChunkBuilder::DoGoto(HGoto* instr) {
-  return new(zone()) LGoto(instr->FirstSuccessor()->block_id());
+  return new(zone()) LGoto(instr->FirstSuccessor());
 }
 
 
@@ -1058,7 +1068,7 @@ LInstruction* LChunkBuilder::DoBranch(HBranch* instr) {
     HBasicBlock* successor = HConstant::cast(value)->BooleanValue()
         ? instr->FirstSuccessor()
         : instr->SecondSuccessor();
-    return new(zone()) LGoto(successor->block_id());
+    return new(zone()) LGoto(successor);
   }
 
   ToBooleanStub::Types expected = instr->expected_input_types();

src/ia32/lithium-ia32.h

@@ -73,6 +73,7 @@ class LCodeGen;
   V(ClampTToUint8)                              \
   V(ClampTToUint8NoSSE2)                        \
   V(ClassOfTestAndBranch)                       \
+  V(ClobberDoubles)                             \
   V(CompareNumericAndBranch)                    \
   V(CmpObjectEqAndBranch)                       \
   V(CmpHoleAndBranch)                           \
@@ -406,19 +407,32 @@ class LInstructionGap V8_FINAL : public LGap {
 };
 
 
+class LClobberDoubles V8_FINAL : public LTemplateInstruction<0, 0, 0> {
+ public:
+  LClobberDoubles() { ASSERT(!CpuFeatures::IsSafeForSnapshot(SSE2)); }
+
+  virtual bool ClobbersDoubleRegisters() const { return true; }
+
+  DECLARE_CONCRETE_INSTRUCTION(ClobberDoubles, "clobber-d")
+};
+
+
 class LGoto V8_FINAL : public LTemplateInstruction<0, 0, 0> {
  public:
-  explicit LGoto(int block_id) : block_id_(block_id) { }
+  explicit LGoto(HBasicBlock* block) : block_(block) { }
 
   virtual bool HasInterestingComment(LCodeGen* gen) const V8_OVERRIDE;
   DECLARE_CONCRETE_INSTRUCTION(Goto, "goto")
   virtual void PrintDataTo(StringStream* stream) V8_OVERRIDE;
   virtual bool IsControl() const V8_OVERRIDE { return true; }
 
-  int block_id() const { return block_id_; }
+  int block_id() const { return block_->block_id(); }
+  virtual bool ClobbersDoubleRegisters() const { return false; }
+
+  bool jumps_to_join() const { return block_->predecessors()->length() > 1; }
 
  private:
-  int block_id_;
+  HBasicBlock* block_;
 };
 
 

src/ia32/macro-assembler-ia32.cc

@@ -2798,6 +2798,8 @@ void MacroAssembler::Ret(int bytes_dropped, Register scratch) {
 void MacroAssembler::VerifyX87StackDepth(uint32_t depth) {
   // Make sure the floating point stack is either empty or has depth items.
   ASSERT(depth <= 7);
+  // This is very expensive.
+  ASSERT(FLAG_debug_code && FLAG_enable_slow_asserts);
 
   // The top-of-stack (tos) is 7 if there is one item pushed.
   int tos = (8 - depth) % 8;