[turbofan] Frame elision for code stubs

src/compiler/arm/code-generator-arm.cc

 // Copyright 2014 the V8 project authors. All rights reserved.
 // Use of this source code is governed by a BSD-style license that can be
 // found in the LICENSE file.
 
 #include "src/compiler/code-generator.h"
 
 #include "src/arm/macro-assembler-arm.h"
 #include "src/ast/scopes.h"
 #include "src/compiler/code-generator-impl.h"
 #include "src/compiler/gap-resolver.h"
@@ skipping 226 matching lines @@
       __ JumpIfSmi(value_, exit());
     }
     __ CheckPageFlag(value_, scratch0_,
                      MemoryChunk::kPointersToHereAreInterestingMask, eq,
                      exit());
     RememberedSetAction const remembered_set_action =
         mode_ > RecordWriteMode::kValueIsMap ? EMIT_REMEMBERED_SET
                                              : OMIT_REMEMBERED_SET;
     SaveFPRegsMode const save_fp_mode =
         frame()->DidAllocateDoubleRegisters() ? kSaveFPRegs : kDontSaveFPRegs;
-    if (!frame()->needs_frame()) {
+    if (!frame_access_state()->access_frame_with_fp()) {

[danno, 2016/03/16 18:18:11]

Here and below, and on other ARM-like platforms, I don't think is is correct.
You can set up frame access with the SP even if you have a frame, this is what
happens in tail calls and there is a bot that builds and runs all of the tests
running with sp-based (not fp-based) frame access. You really need a predicate
on the frame_access_state() called has_frame().

[Mircea Trofin, 2016/03/16 20:20:46]

Ah, yes! I started from Intel and ended up cargo-culting the change from
frame()->needs_frame() over. In  my old prototype, I had this done by passing
down the information about whether the frame is constructed at the beginning of
the function or not. I only did it for arm64, it so happens (it was a
prototype).

Now... in this CL, I'm trying to lose any kind of notion of "was frame created
at beginning of function or not". So in that sense, I would probably want to
capture the frame state entering the OOL code. Can there be more than one code
path entering a OOL, btw? One where there is a frame setup, and one where there
isn't. I'll look into it.

       // We need to save and restore lr if the frame was elided.
       __ Push(lr);
     }
     RecordWriteStub stub(isolate(), object_, scratch0_, scratch1_,
                          remembered_set_action, save_fp_mode);
     if (index_.is(no_reg)) {
       __ add(scratch1_, object_, Operand(index_immediate_));
     } else {
       DCHECK_EQ(0, index_immediate_);
       __ add(scratch1_, object_, Operand(index_));
     }
     __ CallStub(&stub);
-    if (!frame()->needs_frame()) {
+    if (!frame_access_state()->access_frame_with_fp()) {
       __ Pop(lr);
     }
   }
 
  private:
   Register const object_;
   Register const index_;
   int32_t const index_immediate_;  // Valid if index_.is(no_reg).
   Register const value_;
   Register const scratch0_;
@@ skipping 108 matching lines @@
     if (instr->InputAt(1)->IsRegister()) {        \
       __ cmp(offset, i.InputRegister(1));         \
     } else {                                      \
       __ cmp(offset, i.InputImmediate(1));        \
     }                                             \
     auto value = i.InputRegister(2);              \
     __ asm_instr(value, i.InputOffset(3), lo);    \
     DCHECK_EQ(LeaveCC, i.OutputSBit());           \
   } while (0)
 
+void CodeGenerator::AssembleDeconstructFrame() {
+  __ LeaveFrame(StackFrame::MANUAL);
+}
 
 void CodeGenerator::AssembleDeconstructActivationRecord(int stack_param_delta) {
   int sp_slot_delta = TailCallFrameStackSlotDelta(stack_param_delta);
   if (sp_slot_delta > 0) {
     __ add(sp, sp, Operand(sp_slot_delta * kPointerSize));
   }
   frame_access_state()->SetFrameAccessToDefault();
 }
 
 
 void CodeGenerator::AssemblePrepareTailCall(int stack_param_delta) {
   int sp_slot_delta = TailCallFrameStackSlotDelta(stack_param_delta);
   if (sp_slot_delta < 0) {
     __ sub(sp, sp, Operand(-sp_slot_delta * kPointerSize));
     frame_access_state()->IncreaseSPDelta(-sp_slot_delta);
   }
-  if (frame()->needs_frame()) {
+  if (frame_access_state()->access_frame_with_fp()) {
     if (FLAG_enable_embedded_constant_pool) {
       __ ldr(cp, MemOperand(fp, StandardFrameConstants::kConstantPoolOffset));
     }
     __ ldr(lr, MemOperand(fp, StandardFrameConstants::kCallerPCOffset));
     __ ldr(fp, MemOperand(fp, StandardFrameConstants::kCallerFPOffset));
   }
   frame_access_state()->SetFrameAccessToSP();
 }
 
 void CodeGenerator::AssemblePopArgumentsAdaptorFrame(Register args_reg,
@@ skipping 157 matching lines @@
       break;
     case kArchStackPointer:
       __ mov(i.OutputRegister(), sp);
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     case kArchFramePointer:
       __ mov(i.OutputRegister(), fp);
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     case kArchParentFramePointer:
-      if (frame_access_state()->frame()->needs_frame()) {
+      if (frame_access_state()->access_frame_with_fp()) {
         __ ldr(i.OutputRegister(), MemOperand(fp, 0));
       } else {
         __ mov(i.OutputRegister(), fp);
       }
       break;
     case kArchTruncateDoubleToI:
       __ TruncateDoubleToI(i.OutputRegister(), i.InputFloat64Register(0));
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     case kArchStoreWithWriteBarrier: {
@@ skipping 598 matching lines @@
   ArmOperandConverter i(this, instr);
   Label* tlabel = branch->true_label;
   Label* flabel = branch->false_label;
   Condition cc = FlagsConditionToCondition(branch->condition);
   __ b(cc, tlabel);
   if (!branch->fallthru) __ b(flabel);  // no fallthru to flabel.
 }
 
 
 void CodeGenerator::AssembleArchJump(RpoNumber target) {
+  AssembleDeconstructFrameBetweenBlocks();
   if (!IsNextInAssemblyOrder(target)) __ b(GetLabel(target));
 }
 
 
 // Assembles boolean materializations after an instruction.
 void CodeGenerator::AssembleArchBoolean(Instruction* instr,
                                         FlagsCondition condition) {
   ArmOperandConverter i(this, instr);
 
   // Materialize a full 32-bit 1 or 0 value. The result register is always the
@@ skipping 40 matching lines @@
   // TODO(turbofan): We should be able to generate better code by sharing the
   // actual final call site and just bl'ing to it here, similar to what we do
   // in the lithium backend.
   __ Call(deopt_entry, RelocInfo::RUNTIME_ENTRY);
   __ CheckConstPool(false, false);
 }
 
 
 void CodeGenerator::AssemblePrologue() {
   CallDescriptor* descriptor = linkage()->GetIncomingDescriptor();
-  if (frame()->needs_frame()) {
+  if (frame_access_state()->access_frame_with_fp()) {
     if (descriptor->IsCFunctionCall()) {
       if (FLAG_enable_embedded_constant_pool) {
         __ Push(lr, fp, pp);
         // Adjust FP to point to saved FP.
         __ sub(fp, sp, Operand(StandardFrameConstants::kConstantPoolOffset));
       } else {
         __ Push(lr, fp);
         __ mov(fp, sp);
       }
     } else if (descriptor->IsJSFunctionCall()) {
       __ Prologue(this->info()->GeneratePreagedPrologue());
     } else {
       __ StubPrologue(info()->GetOutputStackFrameType());
     }
-  } else {
-    frame()->SetElidedFrameSizeInSlots(0);
   }
-  frame_access_state()->SetFrameAccessToDefault();
 
   int stack_shrink_slots = frame()->GetSpillSlotCount();
   if (info()->is_osr()) {
     // TurboFan OSR-compiled functions cannot be entered directly.
     __ Abort(kShouldNotDirectlyEnterOsrFunction);
 
     // Unoptimized code jumps directly to this entrypoint while the unoptimized
     // frame is still on the stack. Optimized code uses OSR values directly from
     // the unoptimized frame. Thus, all that needs to be done is to allocate the
     // remaining stack slots.
@@ skipping 49 matching lines @@
   const RegList saves_fp = descriptor->CalleeSavedFPRegisters();
   if (saves_fp != 0) {
     STATIC_ASSERT(DwVfpRegister::kMaxNumRegisters == 32);
     uint32_t last = base::bits::CountLeadingZeros32(saves_fp) - 1;
     uint32_t first = base::bits::CountTrailingZeros32(saves_fp);
     __ vldm(ia_w, sp, DwVfpRegister::from_code(first),
             DwVfpRegister::from_code(last));
   }
 
   if (descriptor->IsCFunctionCall()) {
-    __ LeaveFrame(StackFrame::MANUAL);
-  } else if (frame()->needs_frame()) {
+    AssembleDeconstructFrameWhenLeaving();
+  } else if (frame_access_state()->access_frame_with_fp()) {
     // Canonicalize JSFunction return sites for now.
     if (return_label_.is_bound()) {
+      AssembleDeconstructFrameBetweenBlocks();
       __ b(&return_label_);
       return;
     } else {
       __ bind(&return_label_);
-      __ LeaveFrame(StackFrame::MANUAL);
+      AssembleDeconstructFrameWhenLeaving();
     }
   }
   __ Ret(pop_count);
 }
 
 
 void CodeGenerator::AssembleMove(InstructionOperand* source,
                                  InstructionOperand* destination) {
   ArmOperandConverter g(this, nullptr);
   // Dispatch on the source and destination operand kinds.  Not all
@@ skipping 198 matching lines @@
       padding_size -= v8::internal::Assembler::kInstrSize;
     }
   }
 }
 
 #undef __
 
 }  // namespace compiler
 }  // namespace internal
 }  // namespace v8