Revert "[arm64] Switch jssp to csp"

This reverts commit 50baf934.

Reason for revert: breaks arm64 nosnap debug tests:

https://build.chromium.org/p/client.v8.ports/builders/V8%20Linux%20-%20arm64%20-%20sim%20-%20nosnap%20-%20debug/builds/8418

Original change's description:
> [arm64] Switch jssp to csp
> 
> Switch stack pointer to using csp directly, making jssp redundant.
> 
> Bug: v8:6644
> Change-Id: I8e38eda50d56a25161b187c0a033608dd9f90239
> Reviewed-on: https://chromium-review.googlesource.com/860097
> Reviewed-by: Benedikt Meurer <bmeurer@chromium.org>
> Commit-Queue: Martyn Capewell <martyn.capewell@arm.com>
> Cr-Commit-Position: refs/heads/master@{#50487}

TBR=martyn.capewell@arm.com,bmeurer@chromium.org

Change-Id: I20015885e6029271ee6558509cdb92ff1a106e5f
No-Presubmit: true
No-Tree-Checks: true
No-Try: true
Bug: v8:6644
Reviewed-on: https://chromium-review.googlesource.com/860319Reviewed-by: Adam Klein <adamk@chromium.org>
Commit-Queue: Adam Klein <adamk@chromium.org>
Cr-Commit-Position: refs/heads/master@{#50488}

src/arm64/code-stubs-arm64.cc

@@ -42,6 +42,7 @@ void DoubleToIStub::Generate(MacroAssembler* masm) {
   Register result = destination();
 
   DCHECK(result.Is64Bits());
+  DCHECK(jssp.Is(masm->StackPointer()));
 
   UseScratchRegisterScope temps(masm);
   Register scratch1 = temps.AcquireX();
@@ -283,6 +284,7 @@ void CEntryStub::Generate(MacroAssembler* masm) {
   //
   // The arguments are in reverse order, so that arg[argc-2] is actually the
   // first argument to the target function and arg[0] is the last.
+  DCHECK(jssp.Is(__ StackPointer()));
   const Register& argc_input = x0;
   const Register& target_input = x1;
 
@@ -414,6 +416,7 @@ void CEntryStub::Generate(MacroAssembler* masm) {
   __ Peek(target, 3 * kPointerSize);
 
   __ LeaveExitFrame(save_doubles(), x10, x9);
+  DCHECK(jssp.Is(__ StackPointer()));
   if (!argv_in_register()) {
     // Drop the remaining stack slots and return from the stub.
     __ DropArguments(x11);
@@ -450,6 +453,12 @@ void CEntryStub::Generate(MacroAssembler* masm) {
     __ CallCFunction(find_handler, 3);
   }
 
+  // We didn't execute a return case, so the stack frame hasn't been updated
+  // (except for the return address slot). However, we don't need to initialize
+  // jssp because the throw method will immediately overwrite it when it
+  // unwinds the stack.
+  __ SetStackPointer(jssp);
+
   // Retrieve the handler context, SP and FP.
   __ Mov(cp, Operand(pending_handler_context_address));
   __ Ldr(cp, MemOperand(cp));
@@ -487,6 +496,7 @@ void CEntryStub::Generate(MacroAssembler* masm) {
 // Output:
 //   x0: result.
 void JSEntryStub::Generate(MacroAssembler* masm) {
+  DCHECK(jssp.Is(__ StackPointer()));
   Register code_entry = x0;
 
   // Enable instruction instrumentation. This only works on the simulator, and
@@ -503,6 +513,7 @@ void JSEntryStub::Generate(MacroAssembler* masm) {
   __ SetStackPointer(csp);
   __ PushCalleeSavedRegisters();
   __ Mov(jssp, csp);
+  __ SetStackPointer(jssp);
 
   ProfileEntryHookStub::MaybeCallEntryHook(masm);
 
@@ -579,6 +590,7 @@ void JSEntryStub::Generate(MacroAssembler* masm) {
   __ Bind(&invoke);
 
   // Push new stack handler.
+  DCHECK(jssp.Is(__ StackPointer()));
   static_assert(StackHandlerConstants::kSize == 2 * kPointerSize,
                 "Unexpected offset for StackHandlerConstants::kSize");
   static_assert(StackHandlerConstants::kNextOffset == 0 * kPointerSize,
@@ -655,6 +667,8 @@ void JSEntryStub::Generate(MacroAssembler* masm) {
                 "Size of entry frame is not a multiple of 16 bytes");
   __ Drop(EntryFrameConstants::kFixedFrameSize / kPointerSize);
   // Restore the callee-saved registers and return.
+  DCHECK(jssp.Is(__ StackPointer()));
+  __ Mov(csp, jssp);
   __ SetStackPointer(csp);
   __ PopCalleeSavedRegisters();
   // After this point, we must not modify jssp because it is a callee-saved

src/arm64/macro-assembler-arm64-inl.h

@@ -1230,10 +1230,7 @@ void TurboAssembler::Push(Handle<HeapObject> handle) {
   UseScratchRegisterScope temps(this);
   Register tmp = temps.AcquireX();
   Mov(tmp, Operand(handle));
-  // This is only used in test-heap.cc, for generating code that is not
-  // executed. Push a padding slot together with the handle here, to
-  // satisfy the alignment requirement.
-  Push(padreg, tmp);
+  Push(tmp);
 }
 
 void TurboAssembler::Push(Smi* smi) {

src/arm64/macro-assembler-arm64.cc

@@ -44,7 +44,7 @@ TurboAssembler::TurboAssembler(Isolate* isolate, void* buffer, int buffer_size,
 #endif
       tmp_list_(DefaultTmpList()),
       fptmp_list_(DefaultFPTmpList()),
-      sp_(csp),
+      sp_(jssp),
       use_real_aborts_(true) {
   if (create_code_object == CodeObjectRequired::kYes) {
     code_object_ =
@@ -2145,6 +2145,7 @@ void TurboAssembler::PrepareForTailCall(const ParameterCount& callee_args_count,
   Add(dst_reg, dst_reg, 15);
   Bic(dst_reg, dst_reg, 15);
 
+  DCHECK(jssp.Is(StackPointer()));
   Register src_reg = caller_args_count_reg;
   // Calculate the end of source area. +kPointerSize is for the receiver.
   if (callee_args_count.is_reg()) {
@@ -2188,6 +2189,7 @@ void TurboAssembler::PrepareForTailCall(const ParameterCount& callee_args_count,
 
   // Leave current frame.
   Mov(StackPointer(), dst_reg);
+  SetStackPointer(jssp);
   AssertStackConsistency();
 }
 
@@ -2431,6 +2433,7 @@ void TurboAssembler::TruncateDoubleToIDelayed(Zone* zone, Register result,
     // it should use.
     Push(jssp, xzr);  // Push xzr to maintain csp required 16-bytes alignment.
     Mov(jssp, csp);
+    SetStackPointer(jssp);
   }
 
   // If we fell through then inline version didn't succeed - call stub instead.
@@ -2456,6 +2459,7 @@ void TurboAssembler::TruncateDoubleToIDelayed(Zone* zone, Register result,
 }
 
 void TurboAssembler::Prologue() {
+  DCHECK(jssp.Is(StackPointer()));
   Push(lr, fp, cp, x1);
   Add(fp, StackPointer(), StandardFrameConstants::kFixedFrameSizeFromFp);
 }
@@ -2466,6 +2470,7 @@ void TurboAssembler::EnterFrame(StackFrame::Type type) {
   Register code_reg = temps.AcquireX();
 
   if (type == StackFrame::INTERNAL) {
+    DCHECK(jssp.Is(StackPointer()));
     Mov(type_reg, StackFrame::TypeToMarker(type));
     Mov(code_reg, Operand(CodeObject()));
     Push(lr, fp, type_reg, code_reg);
@@ -2486,6 +2491,7 @@ void TurboAssembler::EnterFrame(StackFrame::Type type) {
     // csp[0] : for alignment
   } else {
     DCHECK_EQ(type, StackFrame::CONSTRUCT);
+    DCHECK(jssp.Is(StackPointer()));
     Mov(type_reg, StackFrame::TypeToMarker(type));
 
     // Users of this frame type push a context pointer after the type field,
@@ -2510,6 +2516,7 @@ void TurboAssembler::LeaveFrame(StackFrame::Type type) {
     AssertStackConsistency();
     Pop(fp, lr);
   } else {
+    DCHECK(jssp.Is(StackPointer()));
     // Drop the execution stack down to the frame pointer and restore
     // the caller frame pointer and return address.
     Mov(StackPointer(), fp);
@@ -2543,6 +2550,7 @@ void MacroAssembler::ExitFrameRestoreFPRegs() {
 void MacroAssembler::EnterExitFrame(bool save_doubles, const Register& scratch,
                                     int extra_space,
                                     StackFrame::Type frame_type) {
+  DCHECK(jssp.Is(StackPointer()));
   DCHECK(frame_type == StackFrame::EXIT ||
          frame_type == StackFrame::BUILTIN_EXIT);
 
@@ -2595,8 +2603,19 @@ void MacroAssembler::EnterExitFrame(bool save_doubles, const Register& scratch,
   //         sp[8]: Extra space reserved for caller (if extra_space != 0).
   //   sp -> sp[0]: Space reserved for the return address.
 
+  // Align and synchronize the system stack pointer with jssp.
+  AlignAndSetCSPForFrame();
   DCHECK(csp.Is(StackPointer()));
 
+  //         fp[8]: CallerPC (lr)
+  //   fp -> fp[0]: CallerFP (old fp)
+  //         fp[-8]: STUB marker
+  //         fp[-16]: Space reserved for SPOffset.
+  //         fp[-24]: CodeObject()
+  //         fp[-24 - fp_size]: Saved doubles (if save_doubles is true).
+  //         csp[8]: Memory reserved for the caller if extra_space != 0.
+  //  csp -> csp[0]: Space reserved for the return address.
+
   // ExitFrame::GetStateForFramePointer expects to find the return address at
   // the memory address immediately below the pointer stored in SPOffset.
   // It is not safe to derive much else from SPOffset, because the size of the
@@ -2637,7 +2656,8 @@ void MacroAssembler::LeaveExitFrame(bool restore_doubles,
   //         fp[8]: CallerPC (lr)
   //   fp -> fp[0]: CallerFP (old fp)
   //         fp[...]: The rest of the frame.
-  Mov(csp, fp);
+  Mov(jssp, fp);
+  SetStackPointer(jssp);
   AssertStackConsistency();
   Pop(fp, lr);
 }
@@ -3059,6 +3079,7 @@ void TurboAssembler::Abort(AbortReason reason) {
   // simplify the CallRuntime code, make sure that jssp is the stack pointer.
   // There is no risk of register corruption here because Abort doesn't return.
   Register old_stack_pointer = StackPointer();
+  SetStackPointer(jssp);
   Mov(jssp, old_stack_pointer);
 
   // We need some scratch registers for the MacroAssembler, so make sure we have

src/builtins/arm64/builtins-arm64.cc

@@ -653,6 +653,8 @@ void Builtins::Generate_ResumeGeneratorTrampoline(MacroAssembler* masm) {
 
 static void Generate_StackOverflowCheck(MacroAssembler* masm, Register num_args,
                                         Label* stack_overflow) {
+  DCHECK(masm->StackPointer().Is(jssp));
+
   UseScratchRegisterScope temps(masm);
   Register scratch = temps.AcquireX();
 
@@ -1008,6 +1010,7 @@ void Builtins::Generate_InterpreterEntryTrampoline(MacroAssembler* masm) {
   // Open a frame scope to indicate that there is a frame on the stack.  The
   // MANUAL indicates that the scope shouldn't actually generate code to set up
   // the frame (that is done below).
+  DCHECK(jssp.Is(__ StackPointer()));
   FrameScope frame_scope(masm, StackFrame::MANUAL);
   __ Push(lr, fp, cp, closure);
   __ Add(fp, __ StackPointer(), StandardFrameConstants::kFixedFrameSizeFromFp);
@@ -1061,6 +1064,7 @@ void Builtins::Generate_InterpreterEntryTrampoline(MacroAssembler* masm) {
 
     // Do a stack check to ensure we don't go over the limit.
     Label ok;
+    DCHECK(jssp.Is(__ StackPointer()));
     __ Sub(x10, __ StackPointer(), Operand(x11));
     __ CompareRoot(x10, Heap::kRealStackLimitRootIndex);
     __ B(hs, &ok);
@@ -1646,6 +1650,7 @@ void Generate_ContinueToBuiltinHelper(MacroAssembler* masm,
                         allocatable_register_count)) *
                        kPointerSize;
 
+  DCHECK(jssp.Is(__ StackPointer()));
   // Set up frame pointer.
   __ Add(fp, __ StackPointer(), frame_size);
 
@@ -2956,6 +2961,10 @@ void Builtins::Generate_ArgumentsAdaptorTrampoline(MacroAssembler* masm) {
 }
 
 void Builtins::Generate_WasmCompileLazy(MacroAssembler* masm) {
+  // Wasm code uses the csp. This builtin excepts to use the jssp.
+  // Thus, move csp to jssp when entering this builtin (called from wasm).
+  DCHECK(masm->StackPointer().is(jssp));
+  __ Move(jssp, csp);
   {
     FrameScope scope(masm, StackFrame::INTERNAL);
 
@@ -2980,6 +2989,9 @@ void Builtins::Generate_WasmCompileLazy(MacroAssembler* masm) {
     __ PopDRegList(fp_regs);
     __ PopXRegList(gp_regs);
   }
+  // Move back to csp land. jssp now has the same value as when entering this
+  // function, but csp might have changed in the runtime call.
+  __ Move(csp, jssp);
   // Now jump to the instructions of the returned code object.
   __ Jump(x8);
 }

src/compiler/arm64/code-generator-arm64.cc

@@ -455,7 +455,12 @@ Condition FlagsConditionToCondition(FlagsCondition condition) {
   } while (0)
 
 void CodeGenerator::AssembleDeconstructFrame() {
-  __ Mov(csp, fp);
+  const CallDescriptor* descriptor = linkage()->GetIncomingDescriptor();
+  if (descriptor->IsCFunctionCall() || descriptor->UseNativeStack()) {
+    __ Mov(csp, fp);
+  } else {
+    __ Mov(jssp, fp);
+  }
   __ Pop(fp, lr);
 
   unwinding_info_writer_.MarkFrameDeconstructed(__ pc_offset());
@@ -2323,7 +2328,11 @@ void CodeGenerator::FinishFrame(Frame* frame) {
   frame->AlignFrame(16);
   CallDescriptor* descriptor = linkage()->GetIncomingDescriptor();
 
-  __ SetStackPointer(csp);
+  if (descriptor->UseNativeStack() || descriptor->IsCFunctionCall()) {
+    __ SetStackPointer(csp);
+  } else {
+    __ SetStackPointer(jssp);
+  }
 
   // Save FP registers.
   CPURegList saves_fp = CPURegList(CPURegister::kVRegister, kDRegSizeInBits,
@@ -2413,6 +2422,7 @@ void CodeGenerator::AssembleConstructFrame() {
         __ EnterFrame(StackFrame::WASM_COMPILED);
       }
       DCHECK(__ StackPointer().Is(csp));
+      __ SetStackPointer(jssp);
       __ AssertStackConsistency();
       // Initialize the jssp because it is required for the runtime call.
       __ Mov(jssp, csp);

src/compiler/arm64/instruction-selector-arm64.cc

@@ -1682,21 +1682,28 @@ void InstructionSelector::EmitPrepareArguments(
     Node* node) {
   Arm64OperandGenerator g(this);
 
+  bool from_native_stack = linkage()->GetIncomingDescriptor()->UseNativeStack();
+  bool to_native_stack = descriptor->UseNativeStack();
+
+  bool always_claim = to_native_stack != from_native_stack;
+
   // `arguments` includes alignment "holes". This means that slots bigger than
   // kPointerSize, e.g. Simd128, will span across multiple arguments.
   int claim_count = static_cast<int>(arguments->size());
   int slot = claim_count - 1;
   claim_count = RoundUp(claim_count, 2);
   // Bump the stack pointer(s).
-  if (claim_count > 0) {
+  if (claim_count > 0 || always_claim) {
     // TODO(titzer): claim and poke probably take small immediates.
     // TODO(titzer): it would be better to bump the csp here only
     //                and emit paired stores with increment for non c frames.
-    ArchOpcode claim = kArm64ClaimCSP;
+    ArchOpcode claim = to_native_stack ? kArm64ClaimCSP : kArm64ClaimJSSP;
+    // ClaimJSSP(0) or ClaimCSP(0) isn't a nop if there is a mismatch between
+    // CSP and JSSP.
     Emit(claim, g.NoOutput(), g.TempImmediate(claim_count));
   }
 
-  ArchOpcode poke = kArm64PokeCSP;
+  ArchOpcode poke = to_native_stack ? kArm64PokeCSP : kArm64PokeJSSP;
   if (claim_count > 0) {
     // Store padding, which might be overwritten.
     Emit(poke, g.NoOutput(), g.UseImmediate(0),

src/compiler/instruction-selector.cc

@@ -2387,6 +2387,15 @@ void InstructionSelector::VisitCall(Node* node, BasicBlock* handler) {
     buffer.instruction_args.push_back(g.Label(handler));
   }
 
+  bool from_native_stack = linkage()->GetIncomingDescriptor()->UseNativeStack();
+  bool to_native_stack = descriptor->UseNativeStack();
+  if (from_native_stack != to_native_stack) {
+    // (arm64 only) Mismatch in the use of stack pointers. One or the other
+    // has to be restored manually by the code generator.
+    flags |= to_native_stack ? CallDescriptor::kRestoreJSSP
+                             : CallDescriptor::kRestoreCSP;
+  }
+
   // Select the appropriate opcode based on the call type.
   InstructionCode opcode = kArchNop;
   switch (descriptor->kind()) {

test/cctest/heap/test-heap.cc

@@ -3943,8 +3943,7 @@ static Handle<Code> DummyOptimizedCode(Isolate* isolate) {
                       v8::internal::CodeObjectRequired::kYes);
   CodeDesc desc;
   masm.Push(isolate->factory()->undefined_value());
-  masm.Push(isolate->factory()->undefined_value());
-  masm.Drop(2);
+  masm.Drop(1);
   masm.GetCode(isolate, &desc);
   Handle<Object> undefined(isolate->heap()->undefined_value(), isolate);
   Handle<Code> code =