Enable stub generation using Hydrogen/Lithium (again)

src/ia32/lithium-codegen-ia32.cc

Index: src/ia32/lithium-codegen-ia32.cc
diff --git a/src/ia32/lithium-codegen-ia32.cc b/src/ia32/lithium-codegen-ia32.cc
index 388d4969dd67ea085861780be1e2f346e2872353..00da461d0004e3e04bc5adddbade358c26bc10ee 100644
--- a/src/ia32/lithium-codegen-ia32.cc
+++ b/src/ia32/lithium-codegen-ia32.cc
@@ -30,6 +30,7 @@
 #if defined(V8_TARGET_ARCH_IA32)
 
 #include "ia32/lithium-codegen-ia32.h"
+#include "ic.h"
 #include "code-stubs.h"
 #include "deoptimizer.h"
 #include "stub-cache.h"
@@ -70,7 +71,6 @@ bool LCodeGen::GenerateCode() {
   HPhase phase("Z_Code generation", chunk());
   ASSERT(is_unused());
   status_ = GENERATING;
-  CpuFeatures::Scope scope(SSE2);
 
   CodeStub::GenerateFPStubs();
 
@@ -79,13 +79,15 @@ bool LCodeGen::GenerateCode() {
   // the frame (that is done in GeneratePrologue).
   FrameScope frame_scope(masm_, StackFrame::MANUAL);
 
-  dynamic_frame_alignment_ = (chunk()->num_double_slots() > 2 &&
-                              !chunk()->graph()->is_recursive()) ||
-                             !info()->osr_ast_id().IsNone();
+  dynamic_frame_alignment_ = info()->IsOptimizing() &&
+      ((chunk()->num_double_slots() > 2 &&
+        !chunk()->graph()->is_recursive()) ||
+       !info()->osr_ast_id().IsNone());
 
   return GeneratePrologue() &&
       GenerateBody() &&
       GenerateDeferredCode() &&
+      GenerateJumpTable() &&
       GenerateSafepointTable();
 }
 
@@ -95,7 +97,9 @@ void LCodeGen::FinishCode(Handle<Code> code) {
   code->set_stack_slots(GetStackSlotCount());
   code->set_safepoint_table_offset(safepoints_.GetCodeOffset());
   PopulateDeoptimizationData(code);
-  Deoptimizer::EnsureRelocSpaceForLazyDeoptimization(code);
+  if (!info()->IsStub()) {
+    Deoptimizer::EnsureRelocSpaceForLazyDeoptimization(code);
+  }
 }
 
 
@@ -126,35 +130,36 @@ void LCodeGen::Comment(const char* format, ...) {
 bool LCodeGen::GeneratePrologue() {
   ASSERT(is_generating());
 
-  ProfileEntryHookStub::MaybeCallEntryHook(masm_);
+  if (info()->IsOptimizing()) {
+    ProfileEntryHookStub::MaybeCallEntryHook(masm_);
 
 #ifdef DEBUG
-  if (strlen(FLAG_stop_at) > 0 &&
-      info_->function()->name()->IsEqualTo(CStrVector(FLAG_stop_at))) {
-    __ int3();
-  }
+    if (strlen(FLAG_stop_at) > 0 &&
+        info_->function()->name()->IsEqualTo(CStrVector(FLAG_stop_at))) {
+      __ int3();
+    }
 #endif
 
-  // Strict mode functions and builtins need to replace the receiver
-  // with undefined when called as functions (without an explicit
-  // receiver object). ecx is zero for method calls and non-zero for
-  // function calls.
-  if (!info_->is_classic_mode() || info_->is_native()) {
-    Label begin;
-    __ bind(&begin);
-    Label ok;
-    __ test(ecx, Operand(ecx));
-    __ j(zero, &ok, Label::kNear);
-    // +1 for return address.
-    int receiver_offset = (scope()->num_parameters() + 1) * kPointerSize;
-    __ mov(Operand(esp, receiver_offset),
-           Immediate(isolate()->factory()->undefined_value()));
-    __ bind(&ok);
-    ASSERT(!FLAG_age_code ||
-           (kSizeOfOptimizedStrictModePrologue == ok.pos() - begin.pos()));
+    // Strict mode functions and builtins need to replace the receiver
+    // with undefined when called as functions (without an explicit
+    // receiver object). ecx is zero for method calls and non-zero for
+    // function calls.
+    if (!info_->is_classic_mode() || info_->is_native()) {
+      Label begin;
+      __ bind(&begin);
+      Label ok;
+      __ test(ecx, Operand(ecx));
+      __ j(zero, &ok, Label::kNear);
+      // +1 for return address.
+      int receiver_offset = (scope()->num_parameters() + 1) * kPointerSize;
+      __ mov(Operand(esp, receiver_offset),
+             Immediate(isolate()->factory()->undefined_value()));
+      __ bind(&ok);
+      ASSERT(!FLAG_age_code ||
+             (kSizeOfOptimizedStrictModePrologue == ok.pos() - begin.pos()));
+    }
   }
 
-
   if (dynamic_frame_alignment_) {
     Label begin;
     __ bind(&begin);
@@ -185,10 +190,19 @@ bool LCodeGen::GeneratePrologue() {
             do_not_pad.pos() - begin.pos()));
   }
 
-  __ push(ebp);  // Caller's frame pointer.
-  __ mov(ebp, esp);
-  __ push(esi);  // Callee's context.
-  __ push(edi);  // Callee's JS function.
+  if (NeedsEagerFrame()) {
+    ASSERT(!frame_is_built_);
+    frame_is_built_ = true;
+    __ push(ebp);  // Caller's frame pointer.
+    __ mov(ebp, esp);
+    if (info()->IsStub()) {
+      __ push(esi);

[Jakob Kummerow, 2012/11/28 16:28:22]

nit: could hoist this out of the if/else blocks.

[danno, 2012/11/30 16:23:24]

Done.

[danno, 2012/11/30 16:23:24]

Done.

+      __ push(Immediate(Smi::FromInt(StackFrame::STUB)));
+    } else {
+      __ push(esi);  // Callee's context.
+      __ push(edi);  // Callee's JS function.
+    }
+  }
 
   if (dynamic_frame_alignment_ && FLAG_debug_code) {
     __ test(esp, Immediate(kPointerSize));
@@ -197,50 +211,52 @@ bool LCodeGen::GeneratePrologue() {
 
   // Reserve space for the stack slots needed by the code.
   int slots = GetStackSlotCount();
-  ASSERT_GE(slots, 1);
-  if (slots == 1) {
-    if (dynamic_frame_alignment_) {
-      __ push(edx);
-    } else {
-      __ push(Immediate(kNoAlignmentPadding));
-    }
-  } else {
-    if (FLAG_debug_code) {
-      __ mov(Operand(eax), Immediate(slots));
-      Label loop;
-      __ bind(&loop);
-      __ push(Immediate(kSlotsZapValue));
-      __ dec(eax);
-      __ j(not_zero, &loop);
+  ASSERT(slots != 0 || !info()->IsOptimizing());
+  if (slots > 0) {
+    if (slots == 1) {
+      if (dynamic_frame_alignment_) {
+        __ push(edx);
+      } else {
+        __ push(Immediate(kNoAlignmentPadding));
+      }
     } else {
-      __ sub(Operand(esp), Immediate(slots * kPointerSize));
-  #ifdef _MSC_VER
-      // On windows, you may not access the stack more than one page below
-      // the most recently mapped page. To make the allocated area randomly
-      // accessible, we write to each page in turn (the value is irrelevant).
-      const int kPageSize = 4 * KB;
-      for (int offset = slots * kPointerSize - kPageSize;
-           offset > 0;
-           offset -= kPageSize) {
-        __ mov(Operand(esp, offset), eax);
+      if (FLAG_debug_code) {
+        __ mov(Operand(eax), Immediate(slots));
+        Label loop;
+        __ bind(&loop);
+        __ push(Immediate(kSlotsZapValue));
+        __ dec(eax);
+        __ j(not_zero, &loop);
+      } else {
+        __ sub(Operand(esp), Immediate(slots * kPointerSize));
+#ifdef _MSC_VER
+        // On windows, you may not access the stack more than one page below
+        // the most recently mapped page. To make the allocated area randomly
+        // accessible, we write to each page in turn (the value is irrelevant).
+        const int kPageSize = 4 * KB;
+        for (int offset = slots * kPointerSize - kPageSize;
+             offset > 0;
+             offset -= kPageSize) {
+          __ mov(Operand(esp, offset), eax);
+        }
+#endif
       }
-  #endif
-    }
 
-    // Store dynamic frame alignment state in the first local.
-    if (dynamic_frame_alignment_) {
-      __ mov(Operand(ebp,
-                     JavaScriptFrameConstants::kDynamicAlignmentStateOffset),
-             edx);
-    } else {
-      __ mov(Operand(ebp,
-                     JavaScriptFrameConstants::kDynamicAlignmentStateOffset),
-             Immediate(kNoAlignmentPadding));
+      // Store dynamic frame alignment state in the first local.
+      if (dynamic_frame_alignment_) {
+        __ mov(Operand(ebp,
+                       JavaScriptFrameConstants::kDynamicAlignmentStateOffset),
+               edx);
+      } else {
+        __ mov(Operand(ebp,
+                       JavaScriptFrameConstants::kDynamicAlignmentStateOffset),
+               Immediate(kNoAlignmentPadding));
+      }
     }
   }
 
   // Possibly allocate a local context.
-  int heap_slots = scope()->num_heap_slots() - Context::MIN_CONTEXT_SLOTS;
+  int heap_slots = info_->num_heap_slots() - Context::MIN_CONTEXT_SLOTS;
   if (heap_slots > 0) {
     Comment(";;; Allocate local context");
     // Argument to NewContext is the function, which is still in edi.
@@ -280,7 +296,7 @@ bool LCodeGen::GeneratePrologue() {
   }
 
   // Trace the call.
-  if (FLAG_trace) {
+  if (FLAG_trace && info()->IsOptimizing()) {
     // We have not executed any compiled code yet, so esi still holds the
     // incoming context.
     __ CallRuntime(Runtime::kTraceEnter, 0);
@@ -334,16 +350,104 @@ bool LCodeGen::GenerateBody() {
 }
 
 
+bool LCodeGen::GenerateJumpTable() {
+  Label needs_frame_not_call;
+  bool has_generated_needs_frame_not_call = false;
+  Label needs_frame_is_call;
+  bool has_generated_needs_frame_is_call = false;
+  for (int i = 0; i < jump_table_.length(); i++) {
+    __ bind(&jump_table_[i].label);
+    Address entry = jump_table_[i].address;
+    if (jump_table_[i].needs_frame) {
+      __ push(Immediate(ExternalReference::ForDeoptEntry(entry)));
+      if (jump_table_[i].is_lazy_deopt) {
+        if (!has_generated_needs_frame_is_call) {

[Jakob Kummerow, 2012/11/28 16:28:22]

I think you can replace the check for !has_generated_needs_frame_is_call with
!needs_frame_is_call.is_bound(), eliminating the need to have an extra bool for
each label.

Also, I think readability would improve a bit (fewer double negations) if you
reordered the two blocks:
if (label.is_bound()) {
  __ jmp(&label);
} else {
  __ bind(&label);
  // ...
}

[danno, 2012/11/30 16:23:24]

Done.

+          has_generated_needs_frame_is_call = true;
+          __ bind(&needs_frame_is_call);
+          __ push(esi);
+          // If there is not frame, we don't have access to the JSFunction that

[Jakob Kummerow, 2012/11/28 16:28:22]

"there is not frame" has not grammar.
Aside from that, of course there is no frame, we're just building one. Did you
mean "When we're compiling a stub, ..."?

[danno, 2012/11/30 16:23:24]

Done.

+          // needs to be put into the frame.
+          ASSERT(info()->IsStub());
+          __ push(Immediate(Smi::FromInt(StackFrame::STUB)));
+          // Push a PC inside the function so that the deopt code can find where
+          // the deopt comes from. It doesn't have to be the precise return
+          // address of a "calling" LAZY deopt, it only has to be somewhere
+          // inside the code body.
+          Label push_approx_pc;
+          __ call(&push_approx_pc);
+          __ bind(&push_approx_pc);
+          // Push the continuation which was stashed were the ebp should
+          // be. Replace it with the saved ebp.
+          __ push(MemOperand(esp, 3 * kPointerSize));
+          __ mov(MemOperand(esp, 4 * kPointerSize), ebp);
+          __ lea(ebp, MemOperand(esp, 4 * kPointerSize));
+          __ ret(0);  // Call the continuation without clobbering registers.
+        } else {
+          __ jmp(&needs_frame_is_call);
+        }
+      } else {
+        if (!has_generated_needs_frame_not_call) {

[Jakob Kummerow, 2012/11/28 16:28:22]

same here

[danno, 2012/11/30 16:23:24]

Done.

+          has_generated_needs_frame_not_call = true;
+          __ bind(&needs_frame_not_call);
+          __ push(esi);
+          // If there is not frame, we don't have access to the JSFunction that

[Jakob Kummerow, 2012/11/28 16:28:22]

see above.

[danno, 2012/11/30 16:23:24]

Done.

+          // needs to be put into the frame.
+          ASSERT(info()->IsStub());
+          __ push(Immediate(Smi::FromInt(StackFrame::STUB)));
+          // Push the continuation which was stashed were the ebp should
+          // be. Replace it with the saved ebp.
+          __ push(MemOperand(esp, 2 * kPointerSize));
+          __ mov(MemOperand(esp, 3 * kPointerSize), ebp);
+          __ lea(ebp, MemOperand(esp, 3 * kPointerSize));
+          __ ret(0);  // Call the continuation without clobbering registers.
+        } else {
+          __ jmp(&needs_frame_not_call);
+        }
+      }
+    } else {
+      if (jump_table_[i].is_lazy_deopt) {
+        __ call(entry, RelocInfo::RUNTIME_ENTRY);
+      } else {
+        __ jmp(entry, RelocInfo::RUNTIME_ENTRY);
+      }
+    }
+  }
+  return !is_aborted();
+}
+
+
 bool LCodeGen::GenerateDeferredCode() {
   ASSERT(is_generating());
   if (deferred_.length() > 0) {
     for (int i = 0; !is_aborted() && i < deferred_.length(); i++) {
       LDeferredCode* code = deferred_[i];
       __ bind(code->entry());
+      if (NeedsDeferredFrame()) {
+        Comment(";;; Deferred build frame",
+                code->instruction_index(),
+                code->instr()->Mnemonic());
+        ASSERT(!frame_is_built_);
+        ASSERT(info()->IsStub());
+        frame_is_built_ = true;
+        // Build the frame in such a way that esi isn't trashed.
+        __ push(ebp);  // Caller's frame pointer.
+        __ push(Operand(ebp, StandardFrameConstants::kContextOffset));
+        __ push(Immediate(Smi::FromInt(StackFrame::STUB)));
+        __ lea(ebp, Operand(esp, 2 * kPointerSize));
+      }
       Comment(";;; Deferred code @%d: %s.",
               code->instruction_index(),
               code->instr()->Mnemonic());
       code->Generate();
+      if (NeedsDeferredFrame()) {
+        Comment(";;; Deferred destory frame",

[Jakob Kummerow, 2012/11/28 16:28:22]

nit: "destroy"

[danno, 2012/11/30 16:23:24]

Done.

+                code->instruction_index(),
+                code->instr()->Mnemonic());
+        ASSERT(frame_is_built_);
+        frame_is_built_ = false;
+        __ mov(esp, ebp);
+        __ pop(ebp);
+      }
       __ jmp(code->exit());
     }
   }
@@ -357,6 +461,15 @@ bool LCodeGen::GenerateDeferredCode() {
 
 bool LCodeGen::GenerateSafepointTable() {
   ASSERT(is_done());
+  if (!info()->IsStub()) {
+    // For lazy deoptimization we need space to patch a call after every call.
+    // Ensure there is always space for such patching, even if the code ends
+    // in a call.
+    int target_offset = masm()->pc_offset() + Deoptimizer::patch_size();
+    while (masm()->pc_offset() < target_offset) {
+      masm()->nop();
+    }
+  }
   safepoints_.Emit(masm(), GetStackSlotCount());
   return !is_aborted();
 }
@@ -372,6 +485,11 @@ XMMRegister LCodeGen::ToDoubleRegister(int index) const {
 }
 
 
+bool LCodeGen::IsX87TopOfStack(LOperand* op) const {
+  return op->IsDoubleRegister();
+}
+
+
 Register LCodeGen::ToRegister(LOperand* op) const {
   ASSERT(op->IsRegister());
   return ToRegister(op->index());
@@ -457,7 +575,9 @@ void LCodeGen::WriteTranslation(LEnvironment* environment,
                    translation,
                    arguments_index,
                    arguments_count);
-  int closure_id = *info()->closure() != *environment->closure()
+  bool has_closure_id = !info()->closure().is_null() &&
+      *info()->closure() != *environment->closure();
+  int closure_id = has_closure_id
       ? DefineDeoptimizationLiteral(environment->closure())
       : Translation::kSelfLiteralId;
   switch (environment->frame_type()) {
@@ -480,6 +600,11 @@ void LCodeGen::WriteTranslation(LEnvironment* environment,
     case ARGUMENTS_ADAPTOR:
       translation->BeginArgumentsAdaptorFrame(closure_id, translation_size);
       break;
+    case STUB:
+      translation->BeginCompiledStubFrame();
+      break;
+    default:
+      UNREACHABLE();
   }
 
   // Inlined frames which push their arguments cause the index to be
@@ -614,6 +739,8 @@ void LCodeGen::CallRuntime(const Runtime::Function* fun,
   __ CallRuntime(fun, argc);
 
   RecordSafepointWithLazyDeopt(instr, RECORD_SIMPLE_SAFEPOINT);
+
+  ASSERT(info()->is_calling());
 }
 
 
@@ -638,6 +765,8 @@ void LCodeGen::CallRuntimeFromDeferred(Runtime::FunctionId id,
   __ CallRuntimeSaveDoubles(id);
   RecordSafepointWithRegisters(
       instr->pointer_map(), argc, Safepoint::kNoLazyDeopt);
+
+  ASSERT(info()->is_calling());
 }
 
 
@@ -683,7 +812,11 @@ void LCodeGen::DeoptimizeIf(Condition cc, LEnvironment* environment) {
   RegisterEnvironmentForDeoptimization(environment, Safepoint::kNoLazyDeopt);
   ASSERT(environment->HasBeenRegistered());
   int id = environment->deoptimization_index();
-  Address entry = Deoptimizer::GetDeoptimizationEntry(id, Deoptimizer::EAGER);
+  ASSERT(info()->IsOptimizing() || info()->IsStub());
+  Deoptimizer::BailoutType bailout_type = frame_is_built_
+      ? Deoptimizer::EAGER
+      : Deoptimizer::LAZY;
+  Address entry = Deoptimizer::GetDeoptimizationEntry(id, bailout_type);
   if (entry == NULL) {
     Abort("bailout was not prepared");
     return;
@@ -717,19 +850,44 @@ void LCodeGen::DeoptimizeIf(Condition cc, LEnvironment* environment) {
     __ popfd();
   }
 
+  ASSERT(info()->IsStub() || frame_is_built_);
+  bool lazy_deopt_needed = info()->IsStub();
   if (cc == no_condition) {
     if (FLAG_trap_on_deopt) __ int3();
-    __ jmp(entry, RelocInfo::RUNTIME_ENTRY);
+    if (lazy_deopt_needed) {
+      __ call(entry, RelocInfo::RUNTIME_ENTRY);
+    } else {
+      __ jmp(entry, RelocInfo::RUNTIME_ENTRY);
+    }
   } else {
+    Label done;
     if (FLAG_trap_on_deopt) {
-      Label done;
       __ j(NegateCondition(cc), &done, Label::kNear);
       __ int3();
-      __ jmp(entry, RelocInfo::RUNTIME_ENTRY);
-      __ bind(&done);
+    }
+    if (!lazy_deopt_needed && frame_is_built_) {
+      if (FLAG_trap_on_deopt) {
+        __ jmp(entry, RelocInfo::RUNTIME_ENTRY);
+      } else {
+        __ j(cc, entry, RelocInfo::RUNTIME_ENTRY);
+      }
     } else {
-      __ j(cc, entry, RelocInfo::RUNTIME_ENTRY);
+      // We often have several deopts to the same entry, reuse the last
+      // jump entry if this is the case.
+      if (jump_table_.is_empty() ||
+          jump_table_.last().address != entry ||
+          jump_table_.last().needs_frame != !frame_is_built_ ||
+          jump_table_.last().is_lazy_deopt != lazy_deopt_needed) {
+        JumpTableEntry table_entry(entry, !frame_is_built_, lazy_deopt_needed);
+        jump_table_.Add(table_entry, zone());
+      }
+      if (FLAG_trap_on_deopt) {
+        __ jmp(&jump_table_.last().label);
+      } else {
+        __ j(cc, &jump_table_.last().label);
+      }
     }
+    __ bind(&done);
   }
 }
 
@@ -1430,7 +1588,8 @@ void LCodeGen::DoConstantD(LConstantD* instr) {
     int32_t lower = static_cast<int32_t>(int_val);
     int32_t upper = static_cast<int32_t>(int_val >> (kBitsPerInt));
     if (CpuFeatures::IsSupported(SSE4_1)) {
-      CpuFeatures::Scope scope(SSE4_1);
+      CpuFeatures::Scope scope1(SSE2);
+      CpuFeatures::Scope scope2(SSE4_1);
       if (lower != 0) {
         __ Set(temp, Immediate(lower));
         __ movd(res, Operand(temp));
@@ -1442,6 +1601,8 @@ void LCodeGen::DoConstantD(LConstantD* instr) {
         __ pinsrd(res, Operand(temp), 1);
       }
     } else {
+      ASSERT(CpuFeatures::IsSupported(SSE2));

[Jakob Kummerow, 2012/11/28 16:28:22]

nit: not needed. The Scope c'tor right below contains the same ASSERT.

[danno, 2012/11/30 16:23:24]

Done.

+      CpuFeatures::Scope scope(SSE2);
       __ Set(temp, Immediate(upper));
       __ movd(res, Operand(temp));
       __ psllq(res, 32);
@@ -1595,6 +1756,7 @@ void LCodeGen::DoAddI(LAddI* instr) {
 
 
 void LCodeGen::DoMathMinMax(LMathMinMax* instr) {
+  CpuFeatures::Scope scope(SSE2);
   LOperand* left = instr->left();
   LOperand* right = instr->right();
   ASSERT(left->Equals(instr->result()));
@@ -1656,6 +1818,8 @@ void LCodeGen::DoMathMinMax(LMathMinMax* instr) {
 
 
 void LCodeGen::DoArithmeticD(LArithmeticD* instr) {
+  ASSERT(CpuFeatures::IsSupported(SSE2));

[Jakob Kummerow, 2012/11/28 16:28:22]

nit: not needed

[danno, 2012/11/30 16:23:24]

Done.

+  CpuFeatures::Scope scope(SSE2);
   XMMRegister left = ToDoubleRegister(instr->left());
   XMMRegister right = ToDoubleRegister(instr->right());
   XMMRegister result = ToDoubleRegister(instr->result());
@@ -1666,8 +1830,8 @@ void LCodeGen::DoArithmeticD(LArithmeticD* instr) {
       __ addsd(left, right);
       break;
     case Token::SUB:
-       __ subsd(left, right);
-       break;
+      __ subsd(left, right);
+      break;
     case Token::MUL:
       __ mulsd(left, right);
       break;
@@ -1740,6 +1904,7 @@ void LCodeGen::EmitBranch(int left_block, int right_block, Condition cc) {
 void LCodeGen::DoBranch(LBranch* instr) {
   int true_block = chunk_->LookupDestination(instr->true_block_id());
   int false_block = chunk_->LookupDestination(instr->false_block_id());
+  CpuFeatures::Scope scope(SSE2);
 
   Representation r = instr->hydrogen()->value()->representation();
   if (r.IsInteger32()) {
@@ -1899,6 +2064,7 @@ void LCodeGen::DoCmpIDAndBranch(LCmpIDAndBranch* instr) {
   int false_block = chunk_->LookupDestination(instr->false_block_id());
   int true_block = chunk_->LookupDestination(instr->true_block_id());
   Condition cc = TokenToCondition(instr->op(), instr->is_double());
+  CpuFeatures::Scope scope(SSE2);
 
   if (left->IsConstantOperand() && right->IsConstantOperand()) {
     // We can statically evaluate the comparison.
@@ -2408,7 +2574,7 @@ void LCodeGen::DoCmpT(LCmpT* instr) {
 
 
 void LCodeGen::DoReturn(LReturn* instr) {
-  if (FLAG_trace) {
+  if (FLAG_trace && info()->IsOptimizing()) {
     // Preserve the return value on the stack and rely on the runtime call
     // to return the value in the same register.  We're leaving the code
     // managed by the register allocator and tearing down the frame, it's
@@ -2422,8 +2588,10 @@ void LCodeGen::DoReturn(LReturn* instr) {
     __ mov(edx, Operand(ebp,
       JavaScriptFrameConstants::kDynamicAlignmentStateOffset));
   }
-  __ mov(esp, ebp);
-  __ pop(ebp);
+  if (NeedsEagerFrame()) {
+    __ mov(esp, ebp);
+    __ pop(ebp);
+  }
   if (dynamic_frame_alignment_) {
     Label no_padding;
     __ cmp(edx, Immediate(kNoAlignmentPadding));
@@ -2436,7 +2604,12 @@ void LCodeGen::DoReturn(LReturn* instr) {
     __ Ret((GetParameterCount() + 2) * kPointerSize, ecx);
     __ bind(&no_padding);
   }
-  __ Ret((GetParameterCount() + 1) * kPointerSize, ecx);
+  if (info()->IsStub()) {
+    __ mov(esi, Operand(ebp, StandardFrameConstants::kContextOffset));
+    __ Ret();
+  } else {
+    __ Ret((GetParameterCount() + 1) * kPointerSize, ecx);
+  }
 }
 
 
@@ -2812,11 +2985,23 @@ void LCodeGen::DoLoadKeyedExternalArray(LLoadKeyed* instr) {
       0,
       instr->additional_index()));
   if (elements_kind == EXTERNAL_FLOAT_ELEMENTS) {
-    XMMRegister result(ToDoubleRegister(instr->result()));
-    __ movss(result, operand);
-    __ cvtss2sd(result, result);
+    if (CpuFeatures::IsSupported(SSE2)) {
+      CpuFeatures::Scope scope(SSE2);
+      XMMRegister result(ToDoubleRegister(instr->result()));
+      __ movss(result, operand);
+      __ cvtss2sd(result, result);
+    } else {
+      __ fld_s(operand);
+      HandleX87FPReturnValue(instr);
+    }
   } else if (elements_kind == EXTERNAL_DOUBLE_ELEMENTS) {
-    __ movdbl(ToDoubleRegister(instr->result()), operand);
+    if (CpuFeatures::IsSupported(SSE2)) {
+      CpuFeatures::Scope scope(SSE2);
+      __ movdbl(ToDoubleRegister(instr->result()), operand);
+    } else {
+      __ fld_d(operand);
+      HandleX87FPReturnValue(instr);
+    }
   } else {
     Register result(ToRegister(instr->result()));
     switch (elements_kind) {
@@ -2860,9 +3045,30 @@ void LCodeGen::DoLoadKeyedExternalArray(LLoadKeyed* instr) {
 }
 
 
-void LCodeGen::DoLoadKeyedFixedDoubleArray(LLoadKeyed* instr) {
-  XMMRegister result = ToDoubleRegister(instr->result());
+void LCodeGen::HandleX87FPReturnValue(LInstruction* instr) {
+  if (IsX87TopOfStack(instr->result())) {
+    // Return value is already on stack. If the value has no uses, then
+    // pop it off the FP stack. Otherwise, make sure that there are enough
+    // copies of the value on the stack to feed all of the usages, e.g.
+    // when the following instruction uses the return value in multiple
+    // inputs.
+    int count = instr->hydrogen_value()->UseCount();
+    if (count == 0) {
+      __ fstp(0);
+    } else {
+      count--;
+      ASSERT(count <= 7);
+      while (count-- > 0) {
+        __ fld(0);
+      }
+    }
+  } else {
+    __ fstp_d(ToOperand(instr->result()));
+  }
+}
+
 
+void LCodeGen::DoLoadKeyedFixedDoubleArray(LLoadKeyed* instr) {
   if (instr->hydrogen()->RequiresHoleCheck()) {
     int offset = FixedDoubleArray::kHeaderSize - kHeapObjectTag +
         sizeof(kHoleNanLower32);
@@ -2883,7 +3089,14 @@ void LCodeGen::DoLoadKeyedFixedDoubleArray(LLoadKeyed* instr) {
       FAST_DOUBLE_ELEMENTS,
       FixedDoubleArray::kHeaderSize - kHeapObjectTag,
       instr->additional_index());
-  __ movdbl(result, double_load_operand);
+  if (CpuFeatures::IsSupported(SSE2)) {
+    CpuFeatures::Scope scope(SSE2);
+    XMMRegister result = ToDoubleRegister(instr->result());
+    __ movdbl(result, double_load_operand);
+  } else {
+    __ fld_d(double_load_operand);
+    HandleX87FPReturnValue(instr);
+  }
 }
 
 
@@ -3299,6 +3512,7 @@ void LCodeGen::DoMathAbs(LUnaryMathOperation* instr) {
   ASSERT(instr->value()->Equals(instr->result()));
   Representation r = instr->hydrogen()->value()->representation();
 
+  CpuFeatures::Scope scope(SSE2);
   if (r.IsDouble()) {
     XMMRegister  scratch = xmm0;
     XMMRegister input_reg = ToDoubleRegister(instr->value());
@@ -3320,6 +3534,7 @@ void LCodeGen::DoMathAbs(LUnaryMathOperation* instr) {
 
 
 void LCodeGen::DoMathFloor(LUnaryMathOperation* instr) {
+  CpuFeatures::Scope scope(SSE2);
   XMMRegister xmm_scratch = xmm0;
   Register output_reg = ToRegister(instr->result());
   XMMRegister input_reg = ToDoubleRegister(instr->value());
@@ -3384,6 +3599,7 @@ void LCodeGen::DoMathFloor(LUnaryMathOperation* instr) {
 }
 
 void LCodeGen::DoMathRound(LUnaryMathOperation* instr) {
+  CpuFeatures::Scope scope(SSE2);
   XMMRegister xmm_scratch = xmm0;
   Register output_reg = ToRegister(instr->result());
   XMMRegister input_reg = ToDoubleRegister(instr->value());
@@ -3429,6 +3645,7 @@ void LCodeGen::DoMathRound(LUnaryMathOperation* instr) {
 
 
 void LCodeGen::DoMathSqrt(LUnaryMathOperation* instr) {
+  CpuFeatures::Scope scope(SSE2);
   XMMRegister input_reg = ToDoubleRegister(instr->value());
   ASSERT(ToDoubleRegister(instr->result()).is(input_reg));
   __ sqrtsd(input_reg, input_reg);
@@ -3436,6 +3653,7 @@ void LCodeGen::DoMathSqrt(LUnaryMathOperation* instr) {
 
 
 void LCodeGen::DoMathPowHalf(LMathPowHalf* instr) {
+  CpuFeatures::Scope scope(SSE2);
   XMMRegister xmm_scratch = xmm0;
   XMMRegister input_reg = ToDoubleRegister(instr->value());
   Register scratch = ToRegister(instr->temp());
@@ -3512,6 +3730,7 @@ void LCodeGen::DoRandom(LRandom* instr) {
 
   DeferredDoRandom* deferred = new(zone()) DeferredDoRandom(this, instr);
 
+  CpuFeatures::Scope scope(SSE2);
   // Having marked this instruction as a call we can use any
   // registers.
   ASSERT(ToDoubleRegister(instr->result()).is(xmm1));
@@ -3579,6 +3798,7 @@ void LCodeGen::DoDeferredRandom(LRandom* instr) {
 
 
 void LCodeGen::DoMathLog(LUnaryMathOperation* instr) {
+  CpuFeatures::Scope scope(SSE2);
   ASSERT(instr->value()->Equals(instr->result()));
   XMMRegister input_reg = ToDoubleRegister(instr->value());
   Label positive, done, zero;
@@ -3610,6 +3830,8 @@ void LCodeGen::DoMathLog(LUnaryMathOperation* instr) {
 
 
 void LCodeGen::DoMathExp(LMathExp* instr) {
+  ASSERT(CpuFeatures::IsSupported(SSE2));

[Jakob Kummerow, 2012/11/28 16:28:22]

nit: not needed

[danno, 2012/11/30 16:23:24]

Done.

+  CpuFeatures::Scope scope(SSE2);
   XMMRegister input = ToDoubleRegister(instr->value());
   XMMRegister result = ToDoubleRegister(instr->result());
   Register temp1 = ToRegister(instr->temp1());
@@ -3878,6 +4100,11 @@ void LCodeGen::DoBoundsCheck(LBoundsCheck* instr) {
     }
     DeoptimizeIf(below_equal, instr->environment());
   } else {
+    if (instr->hydrogen()->index()->representation().IsTagged() &&
+        !instr->hydrogen()->index()->type().IsSmi()) {
+      __ test(ToRegister(instr->index()), Immediate(kSmiTagMask));
+      DeoptimizeIf(not_zero, instr->environment());
+    }
     __ cmp(ToRegister(instr->index()), ToOperand(instr->length()));
     DeoptimizeIf(above_equal, instr->environment());
   }
@@ -3900,9 +4127,11 @@ void LCodeGen::DoStoreKeyedExternalArray(LStoreKeyed* instr) {
       0,
       instr->additional_index()));
   if (elements_kind == EXTERNAL_FLOAT_ELEMENTS) {
+    CpuFeatures::Scope scope(SSE2);
     __ cvtsd2ss(xmm0, ToDoubleRegister(instr->value()));
     __ movss(operand, xmm0);
   } else if (elements_kind == EXTERNAL_DOUBLE_ELEMENTS) {
+    CpuFeatures::Scope scope(SSE2);
     __ movdbl(operand, ToDoubleRegister(instr->value()));
   } else {
     Register value = ToRegister(instr->value());
@@ -3938,6 +4167,7 @@ void LCodeGen::DoStoreKeyedExternalArray(LStoreKeyed* instr) {
 
 
 void LCodeGen::DoStoreKeyedFixedDoubleArray(LStoreKeyed* instr) {
+  CpuFeatures::Scope scope(SSE2);
   XMMRegister value = ToDoubleRegister(instr->value());
 
   if (instr->NeedsCanonicalization()) {
@@ -4188,15 +4418,21 @@ void LCodeGen::DoStringAdd(LStringAdd* instr) {
 
 
 void LCodeGen::DoInteger32ToDouble(LInteger32ToDouble* instr) {
-  LOperand* input = instr->value();
-  ASSERT(input->IsRegister() || input->IsStackSlot());
-  LOperand* output = instr->result();
-  ASSERT(output->IsDoubleRegister());
-  __ cvtsi2sd(ToDoubleRegister(output), ToOperand(input));
+  if (CpuFeatures::IsSupported(SSE2)) {
+    CpuFeatures::Scope scope(SSE2);
+    LOperand* input = instr->value();
+    ASSERT(input->IsRegister() || input->IsStackSlot());
+    LOperand* output = instr->result();
+    ASSERT(output->IsDoubleRegister());
+    __ cvtsi2sd(ToDoubleRegister(output), ToOperand(input));
+  } else {
+    UNREACHABLE();
+  }
 }
 
 
 void LCodeGen::DoUint32ToDouble(LUint32ToDouble* instr) {
+  CpuFeatures::Scope scope(SSE2);
   LOperand* input = instr->value();
   LOperand* output = instr->result();
   LOperand* temp = instr->temp();
@@ -4274,9 +4510,21 @@ void LCodeGen::DoDeferredNumberTagI(LInstruction* instr,
     // the value in there. If that fails, call the runtime system.
     __ SmiUntag(reg);
     __ xor_(reg, 0x80000000);
-    __ cvtsi2sd(xmm0, Operand(reg));
+    if (CpuFeatures::IsSupported(SSE2)) {
+      CpuFeatures::Scope feature_scope(SSE2);
+      __ cvtsi2sd(xmm0, Operand(reg));
+    } else {
+      __ push(reg);
+      __ fild_s(Operand(esp, 0));
+      __ pop(reg);
+    }
   } else {
-    __ LoadUint32(xmm0, reg, xmm1);
+    if (CpuFeatures::IsSupported(SSE2)) {
+      CpuFeatures::Scope feature_scope(SSE2);
+      __ LoadUint32(xmm0, reg, xmm1);
+    } else {
+      UNREACHABLE();
+    }
   }
 
   if (FLAG_inline_new) {
@@ -4305,7 +4553,12 @@ void LCodeGen::DoDeferredNumberTagI(LInstruction* instr,
   // Done. Put the value in xmm0 into the value of the allocated heap
   // number.
   __ bind(&done);
-  __ movdbl(FieldOperand(reg, HeapNumber::kValueOffset), xmm0);
+  if (CpuFeatures::IsSupported(SSE2)) {
+    CpuFeatures::Scope feature_scope(SSE2);
+    __ movdbl(FieldOperand(reg, HeapNumber::kValueOffset), xmm0);
+  } else {
+    __ fstp_d(FieldOperand(reg, HeapNumber::kValueOffset));
+  }
   __ StoreToSafepointRegisterSlot(reg, reg);
 }
 
@@ -4321,7 +4574,6 @@ void LCodeGen::DoNumberTagD(LNumberTagD* instr) {
     LNumberTagD* instr_;
   };
 
-  XMMRegister input_reg = ToDoubleRegister(instr->value());
   Register reg = ToRegister(instr->result());
   Register tmp = ToRegister(instr->temp());
 
@@ -4332,7 +4584,16 @@ void LCodeGen::DoNumberTagD(LNumberTagD* instr) {
     __ jmp(deferred->entry());
   }
   __ bind(deferred->exit());
-  __ movdbl(FieldOperand(reg, HeapNumber::kValueOffset), input_reg);
+  if (CpuFeatures::IsSupported(SSE2)) {
+    CpuFeatures::Scope scope(SSE2);
+    XMMRegister input_reg = ToDoubleRegister(instr->value());
+    __ movdbl(FieldOperand(reg, HeapNumber::kValueOffset), input_reg);
+  } else {
+    if (!IsX87TopOfStack(instr->value())) {
+    __ fld_d(ToOperand(instr->value()));
+    }
+    __ fstp_d(FieldOperand(reg, HeapNumber::kValueOffset));
+  }
 }
 
 
@@ -4489,7 +4750,8 @@ void LCodeGen::DoDeferredTaggedToI(LTaggedToI* instr) {
       DeoptimizeIf(not_equal, instr->environment());
       DeoptimizeIf(parity_even, instr->environment());  // NaN.
     }
-  } else {
+  } else if (CpuFeatures::IsSupported(SSE2)) {
+    CpuFeatures::Scope scope(SSE2);
     // Deoptimize if we don't have a heap number.
     __ RecordComment("Deferred TaggedToI: not a heap number");
     DeoptimizeIf(not_equal, instr->environment());
@@ -4511,6 +4773,8 @@ void LCodeGen::DoDeferredTaggedToI(LTaggedToI* instr) {
       __ RecordComment("Deferred TaggedToI: minus zero");
       DeoptimizeIf(not_zero, instr->environment());
     }
+  } else {
+    UNREACHABLE();
   }
   __ bind(&done);
 }
@@ -4553,19 +4817,24 @@ void LCodeGen::DoNumberUntagD(LNumberUntagD* instr) {
   LOperand* result = instr->result();
   ASSERT(result->IsDoubleRegister());
 
-  Register input_reg = ToRegister(input);
-  XMMRegister result_reg = ToDoubleRegister(result);
-
-  bool deoptimize_on_minus_zero =
-      instr->hydrogen()->deoptimize_on_minus_zero();
-  Register temp_reg = deoptimize_on_minus_zero ? ToRegister(temp) : no_reg;
-
-  EmitNumberUntagD(input_reg,
-                   temp_reg,
-                   result_reg,
-                   instr->hydrogen()->deoptimize_on_undefined(),
-                   deoptimize_on_minus_zero,
-                   instr->environment());
+  if (CpuFeatures::IsSupported(SSE2)) {
+    CpuFeatures::Scope scope(SSE2);
+    Register input_reg = ToRegister(input);
+    XMMRegister result_reg = ToDoubleRegister(result);
+
+    bool deoptimize_on_minus_zero =
+        instr->hydrogen()->deoptimize_on_minus_zero();
+    Register temp_reg = deoptimize_on_minus_zero ? ToRegister(temp) : no_reg;
+
+    EmitNumberUntagD(input_reg,
+                     temp_reg,
+                     result_reg,
+                     instr->hydrogen()->deoptimize_on_undefined(),
+                     deoptimize_on_minus_zero,
+                     instr->environment());
+  } else {
+    UNIMPLEMENTED();
+  }
 }
 
 
@@ -4574,6 +4843,7 @@ void LCodeGen::DoDoubleToI(LDoubleToI* instr) {
   ASSERT(input->IsDoubleRegister());
   LOperand* result = instr->result();
   ASSERT(result->IsRegister());
+  CpuFeatures::Scope scope(SSE2);
 
   XMMRegister input_reg = ToDoubleRegister(input);
   Register result_reg = ToRegister(result);
@@ -4763,10 +5033,10 @@ void LCodeGen::DoCheckFunction(LCheckFunction* instr) {
 void LCodeGen::DoCheckMapCommon(Register reg,
                                 Handle<Map> map,
                                 CompareMapMode mode,
-                                LEnvironment* env) {
+                                LInstruction* instr) {
   Label success;
   __ CompareMap(reg, map, &success, mode);
-  DeoptimizeIf(not_equal, env);
+  DeoptimizeIf(not_equal, instr->environment());
   __ bind(&success);
 }
 
@@ -4784,12 +5054,14 @@ void LCodeGen::DoCheckMaps(LCheckMaps* instr) {
     __ j(equal, &success);
   }
   Handle<Map> map = map_set->last();
-  DoCheckMapCommon(reg, map, REQUIRE_EXACT_MAP, instr->environment());
+  DoCheckMapCommon(reg, map, REQUIRE_EXACT_MAP, instr);
   __ bind(&success);
 }
 
 
 void LCodeGen::DoClampDToUint8(LClampDToUint8* instr) {
+  ASSERT(CpuFeatures::IsSupported(SSE2));

[Jakob Kummerow, 2012/11/28 16:28:22]

nit: not needed

+  CpuFeatures::Scope scope(SSE2);
   XMMRegister value_reg = ToDoubleRegister(instr->unclamped());
   Register result_reg = ToRegister(instr->result());
   __ ClampDoubleToUint8(value_reg, xmm0, result_reg);
@@ -4804,6 +5076,9 @@ void LCodeGen::DoClampIToUint8(LClampIToUint8* instr) {
 
 
 void LCodeGen::DoClampTToUint8(LClampTToUint8* instr) {
+  ASSERT(CpuFeatures::IsSupported(SSE2));

[Jakob Kummerow, 2012/11/28 16:28:22]

nit: not needed

[danno, 2012/11/30 16:23:24]

Done.

+  CpuFeatures::Scope scope(SSE2);
+
   ASSERT(instr->unclamped()->Equals(instr->result()));
   Register input_reg = ToRegister(instr->unclamped());
   Label is_smi, done, heap_number;
@@ -4850,7 +5125,7 @@ void LCodeGen::DoCheckPrototypeMaps(LCheckPrototypeMaps* instr) {
   // Check prototype maps up to the holder.
   while (!current_prototype.is_identical_to(holder)) {
     DoCheckMapCommon(reg, Handle<Map>(current_prototype->map()),
-                     ALLOW_ELEMENT_TRANSITION_MAPS, instr->environment());
+                     ALLOW_ELEMENT_TRANSITION_MAPS, instr);
 
     current_prototype =
         Handle<JSObject>(JSObject::cast(current_prototype->GetPrototype()));
@@ -4860,7 +5135,7 @@ void LCodeGen::DoCheckPrototypeMaps(LCheckPrototypeMaps* instr) {
 
   // Check the holder map.
   DoCheckMapCommon(reg, Handle<Map>(current_prototype->map()),
-                   ALLOW_ELEMENT_TRANSITION_MAPS, instr->environment());
+                   ALLOW_ELEMENT_TRANSITION_MAPS, instr);
 }
 
 
@@ -5397,13 +5672,15 @@ void LCodeGen::EmitIsConstructCall(Register temp) {
 
 
 void LCodeGen::EnsureSpaceForLazyDeopt() {
-  // Ensure that we have enough space after the previous lazy-bailout
-  // instruction for patching the code here.
-  int current_pc = masm()->pc_offset();
-  int patch_size = Deoptimizer::patch_size();
-  if (current_pc < last_lazy_deopt_pc_ + patch_size) {
-    int padding_size = last_lazy_deopt_pc_ + patch_size - current_pc;
-    __ Nop(padding_size);
+  if (!info()->IsStub()) {
+    // Ensure that we have enough space after the previous lazy-bailout
+    // instruction for patching the code here.
+    int current_pc = masm()->pc_offset();
+    int patch_size = Deoptimizer::patch_size();
+    if (current_pc < last_lazy_deopt_pc_ + patch_size) {
+      int padding_size = last_lazy_deopt_pc_ + patch_size - current_pc;
+      __ Nop(padding_size);
+    }
   }
   last_lazy_deopt_pc_ = masm()->pc_offset();
 }