X87: [turbofan] Frame elision for code stubs.

src/compiler/x87/code-generator-x87.cc

 // Copyright 2013 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/ast/scopes.h"
 #include "src/compiler/code-generator-impl.h"
 #include "src/compiler/gap-resolver.h"
 #include "src/compiler/node-matchers.h"
@@ skipping 319 matching lines @@
     Label done;                                              \
     __ j(above_equal, &done, Label::kNear);                  \
     if (instr->InputAt(2)->IsRegister()) {                   \
       __ asm_instr(i.MemoryOperand(3), i.InputRegister(2));  \
     } else {                                                 \
       __ asm_instr(i.MemoryOperand(3), i.InputImmediate(2)); \
     }                                                        \
     __ bind(&done);                                          \
   } while (false)
 
+void CodeGenerator::AssembleDeconstructFrame() {
+  __ mov(esp, ebp);
+  __ pop(ebp);
+}
+
+// For insert fninit/fld1 instructions after the Prologue
+thread_local bool is_block_0 = false;
+
+void CodeGenerator::AssembleSetupStackPointer() { is_block_0 = true; }
 
 void CodeGenerator::AssembleDeconstructActivationRecord(int stack_param_delta) {
   int sp_slot_delta = TailCallFrameStackSlotDelta(stack_param_delta);
   if (sp_slot_delta > 0) {
     __ add(esp, Immediate(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(esp, Immediate(-sp_slot_delta * kPointerSize));
     frame_access_state()->IncreaseSPDelta(-sp_slot_delta);
   }
-  if (frame()->needs_frame()) {
+  if (frame_access_state()->has_frame()) {
     __ mov(ebp, MemOperand(ebp, 0));
   }
   frame_access_state()->SetFrameAccessToSP();
 }
 
 void CodeGenerator::AssemblePopArgumentsAdaptorFrame(Register args_reg,
                                                      Register, Register,
                                                      Register) {
   // There are not enough temp registers left on ia32 for a call instruction
   // so we pick some scratch registers and save/restore them manually here.
@@ skipping 28 matching lines @@
   __ pop(scratch1);
 
   __ bind(&done);
 }
 
 // Assembles an instruction after register allocation, producing machine code.
 void CodeGenerator::AssembleArchInstruction(Instruction* instr) {
   X87OperandConverter i(this, instr);
   InstructionCode opcode = instr->opcode();
   ArchOpcode arch_opcode = ArchOpcodeField::decode(opcode);
+
+  // Workaround for CL #35139 (https://codereview.chromium.org/1775323002)
+  if (is_block_0) {
+    __ fninit();
+    __ fld1();
+    is_block_0 = false;
+  }
+
   switch (arch_opcode) {
     case kArchCallCodeObject: {
       if (FLAG_debug_code && FLAG_enable_slow_asserts) {
         __ VerifyX87StackDepth(1);
       }
       __ fstp(0);
       EnsureSpaceForLazyDeopt();
       if (HasImmediateInput(instr, 0)) {
         Handle<Code> code = Handle<Code>::cast(i.InputHeapObject(0));
         __ call(code, RelocInfo::CODE_TARGET);
@@ skipping 177 matching lines @@
     case kArchRet:
       AssembleReturn();
       break;
     case kArchFramePointer:
       __ mov(i.OutputRegister(), ebp);
       break;
     case kArchStackPointer:
       __ mov(i.OutputRegister(), esp);
       break;
     case kArchParentFramePointer:
-      if (frame_access_state()->frame()->needs_frame()) {
+      if (frame_access_state()->has_frame()) {
         __ mov(i.OutputRegister(), Operand(ebp, 0));
       } else {
         __ mov(i.OutputRegister(), ebp);
       }
       break;
     case kArchTruncateDoubleToI: {
       if (!instr->InputAt(0)->IsDoubleRegister()) {
         __ fld_d(i.InputOperand(0));
       }
       __ TruncateX87TOSToI(i.OutputRegister());
@@ skipping 1461 matching lines @@
 //                                       | FP | RET | args |  caller frame |
 //                                       ^ esp,ebp
 
 // --{ pop ebp }----------------------------------------------------------------
 //                                            | RET | args |  caller frame |
 //                                            ^ esp                        ^ ebp
 
 
 void CodeGenerator::AssemblePrologue() {
   CallDescriptor* descriptor = linkage()->GetIncomingDescriptor();
-  if (frame()->needs_frame()) {
+  if (frame_access_state()->has_frame()) {
     if (descriptor->IsCFunctionCall()) {
       __ push(ebp);
       __ mov(ebp, esp);
     } else if (descriptor->IsJSFunctionCall()) {
       __ Prologue(this->info()->GeneratePreagedPrologue());
     } else {
       __ StubPrologue(info()->GetOutputStackFrameType());
     }
-  } else {
-    frame()->SetElidedFrameSizeInSlots(kPCOnStackSize / kPointerSize);
   }
-  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.
     if (FLAG_code_comments) __ RecordComment("-- OSR entrypoint --");
     osr_pc_offset_ = __ pc_offset();
     stack_shrink_slots -= OsrHelper(info()).UnoptimizedFrameSlots();
+
+    // Initailize FPU state.
+    __ fninit();
+    __ fld1();
   }
 
   const RegList saves = descriptor->CalleeSavedRegisters();
   if (stack_shrink_slots > 0) {
     __ sub(esp, Immediate(stack_shrink_slots * kPointerSize));
   }
 
   if (saves != 0) {  // Save callee-saved registers.
     DCHECK(!info()->is_osr());
     int pushed = 0;
     for (int i = Register::kNumRegisters - 1; i >= 0; i--) {
       if (!((1 << i) & saves)) continue;
       __ push(Register::from_code(i));
       ++pushed;
     }
     frame()->AllocateSavedCalleeRegisterSlots(pushed);
   }
-
-  // Initailize FPU state.
-  __ fninit();
-  __ fld1();
 }
 
 
 void CodeGenerator::AssembleReturn() {
   CallDescriptor* descriptor = linkage()->GetIncomingDescriptor();
 
   // Clear the FPU stack only if there is no return value in the stack.
   if (FLAG_debug_code && FLAG_enable_slow_asserts) {
     __ VerifyX87StackDepth(1);
   }
@@ skipping 12 matching lines @@
   const RegList saves = descriptor->CalleeSavedRegisters();
   // Restore registers.
   if (saves != 0) {
     for (int i = 0; i < Register::kNumRegisters; i++) {
       if (!((1 << i) & saves)) continue;
       __ pop(Register::from_code(i));
     }
   }
 
   if (descriptor->IsCFunctionCall()) {
-    __ mov(esp, ebp);  // Move stack pointer back to frame pointer.
-    __ pop(ebp);       // Pop caller's frame pointer.
-  } else if (frame()->needs_frame()) {
+    AssembleDeconstructFrame();
+  } else if (frame_access_state()->has_frame()) {
     // Canonicalize JSFunction return sites for now.
     if (return_label_.is_bound()) {
       __ jmp(&return_label_);
       return;
     } else {
       __ bind(&return_label_);
-      __ mov(esp, ebp);  // Move stack pointer back to frame pointer.
-      __ pop(ebp);       // Pop caller's frame pointer.
+      AssembleDeconstructFrame();
     }
   }
   if (pop_count == 0) {
     __ ret(0);
   } else {
     __ Ret(pop_count * kPointerSize, ebx);
   }
 }
 
 
@@ skipping 232 matching lines @@
     int padding_size = last_lazy_deopt_pc_ + space_needed - current_pc;
     __ Nop(padding_size);
   }
 }
 
 #undef __
 
 }  // namespace compiler
 }  // namespace internal
 }  // namespace v8