[turbofan, arm64] Fix native stack parameters on arm64.

I added a flag to the CallDescriptor which indicates that the native
stack should be used for a CallObject instead of the js stack on arm64.

Additionally I removed the use of EmitPrepareArguments because the
current implementation does not work when float and int parameters are
mixed. I plan to fix it in a future CL, because currently I have a
problem figuring out the type of a parameter.

R=titzer@chromium.org, v8-arm-ports@googlegroups.com

Review URL: https://codereview.chromium.org/1494123002

Cr-Commit-Position: refs/heads/master@{#32577}

src/arm64/macro-assembler-arm64.cc

@@ -2634,14 +2634,13 @@ void MacroAssembler::TruncateHeapNumberToI(Register result,
 
 
 void MacroAssembler::StubPrologue() {
-  DCHECK(StackPointer().Is(jssp));
   UseScratchRegisterScope temps(this);
   Register temp = temps.AcquireX();
   __ Mov(temp, Smi::FromInt(StackFrame::STUB));
   // Compiled stubs don't age, and so they don't need the predictable code
   // ageing sequence.
   __ Push(lr, fp, cp, temp);
-  __ Add(fp, jssp, StandardFrameConstants::kFixedFrameSizeFromFp);
+  __ Add(fp, StackPointer(), StandardFrameConstants::kFixedFrameSizeFromFp);
 }
 
 

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

@@ -877,12 +877,22 @@ void CodeGenerator::AssembleArchInstruction(Instruction* instr) {
     }
     case kArm64Poke: {
       Operand operand(i.InputInt32(1) * kPointerSize);
+      if (instr->InputAt(0)->IsDoubleRegister()) {
+        __ Poke(i.InputFloat64Register(0), operand);
+      } else {
         __ Poke(i.InputRegister(0), operand);
+      }
       break;
     }
     case kArm64PokePair: {
       int slot = i.InputInt32(2) - 1;
-      __ PokePair(i.InputRegister(1), i.InputRegister(0), slot * kPointerSize);
+      if (instr->InputAt(0)->IsDoubleRegister()) {
+        __ PokePair(i.InputFloat64Register(1), i.InputFloat64Register(0),
+                    slot * kPointerSize);
+      } else {
+        __ PokePair(i.InputRegister(1), i.InputRegister(0),
+                    slot * kPointerSize);
+      }
       break;
     }
     case kArm64Clz:
@@ -1292,9 +1302,18 @@ void CodeGenerator::AssemblePrologue() {
     __ SetStackPointer(jssp);
     __ Prologue(info->IsCodePreAgingActive());
   } else if (frame()->needs_frame()) {
+    if (descriptor->UseNativeStack()) {
+      __ SetStackPointer(csp);
+    } else {
       __ SetStackPointer(jssp);
+    }
     __ StubPrologue();
   } else {
+    if (descriptor->UseNativeStack()) {
+      __ SetStackPointer(csp);
+    } else {
+      __ SetStackPointer(jssp);
+    }
     frame()->SetElidedFrameSizeInSlots(0);
   }
   frame_access_state()->SetFrameAccessToDefault();
@@ -1315,9 +1334,12 @@ void CodeGenerator::AssemblePrologue() {
     stack_shrink_slots -= OsrHelper(info()).UnoptimizedFrameSlots();
   }
 
-  if (csp.Is(masm()->StackPointer())) {
+  // If frame()->needs_frame() is false, then
+  // frame()->AlignSavedCalleeRegisterSlots() is guaranteed to return 0.
+  if (csp.Is(masm()->StackPointer()) && frame()->needs_frame()) {
     // The system stack pointer requires 16-byte alignment at function call
     // boundaries.
+
     stack_shrink_slots += frame()->AlignSavedCalleeRegisterSlots();
   }
   __ Claim(stack_shrink_slots);
@@ -1374,9 +1396,15 @@ void CodeGenerator::AssembleReturn() {
       return;
     } else {
       __ Bind(&return_label_);
+      if (descriptor->UseNativeStack()) {
+        __ Mov(csp, fp);
+      } else {
         __ Mov(jssp, fp);
+      }
       __ Pop(fp, lr);
     }
+  } else if (descriptor->UseNativeStack()) {
+    pop_count += (pop_count & 1);
   }
   __ Drop(pop_count);
   __ Ret();

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

@@ -1535,7 +1535,14 @@ void InstructionSelector::EmitPrepareArguments(NodeVector* arguments,
 
   // Push the arguments to the stack.
   int aligned_push_count = static_cast<int>(arguments->size());
+
   bool pushed_count_uneven = aligned_push_count & 1;
+  int claim_count = aligned_push_count;
+  if (pushed_count_uneven && descriptor->UseNativeStack()) {
+    // We can only claim for an even number of call arguments when we use the
+    // native stack.
+    claim_count++;
+  }
   // TODO(dcarney): claim and poke probably take small immediates,
   //                loop here or whatever.
   // Bump the stack pointer(s).
@@ -1543,23 +1550,20 @@ void InstructionSelector::EmitPrepareArguments(NodeVector* arguments,
     // TODO(dcarney): it would be better to bump the csp here only
     //                and emit paired stores with increment for non c frames.
     Emit(kArm64ClaimForCallArguments, g.NoOutput(),
-         g.TempImmediate(aligned_push_count));
+         g.TempImmediate(claim_count));
   }
+
   // Move arguments to the stack.
-  {
   int slot = aligned_push_count - 1;
-    // Emit the uneven pushes.
-    if (pushed_count_uneven) {
-      Node* input = (*arguments)[slot];
-      Emit(kArm64Poke, g.NoOutput(), g.UseRegister(input),
+  while (slot >= 0) {
+    Emit(kArm64Poke, g.NoOutput(), g.UseRegister((*arguments)[slot]),
          g.TempImmediate(slot));
     slot--;
-    }
-    // Now all pushes can be done in pairs.
-    for (; slot >= 0; slot -= 2) {
-      Emit(kArm64PokePair, g.NoOutput(), g.UseRegister((*arguments)[slot]),
-           g.UseRegister((*arguments)[slot - 1]), g.TempImmediate(slot));
-    }
+    // TODO(ahaas): Poke arguments in pairs if two subsequent arguments have the
+    //              same type.
+    // Emit(kArm64PokePair, g.NoOutput(), g.UseRegister((*arguments)[slot]),
+    //      g.UseRegister((*arguments)[slot - 1]), g.TempImmediate(slot));
+    // slot -= 2;
   }
 }
 

src/compiler/linkage.h

@@ -153,6 +153,9 @@ class CallDescriptor final : public ZoneObject {
     kHasLocalCatchHandler = 1u << 4,
     kSupportsTailCalls = 1u << 5,
     kCanUseRoots = 1u << 6,
+    // Indicates that the native stack should be used for a code object. This
+    // information is important for native calls on arm64.
+    kUseNativeStack = 1u << 7,
     kPatchableCallSiteWithNop = kPatchableCallSite | kNeedsNopAfterCall
   };
   typedef base::Flags<Flag> Flags;
@@ -218,6 +221,7 @@ class CallDescriptor final : public ZoneObject {
 
   bool NeedsFrameState() const { return flags() & kNeedsFrameState; }
   bool SupportsTailCalls() const { return flags() & kSupportsTailCalls; }
+  bool UseNativeStack() const { return flags() & kUseNativeStack; }
 
   LinkageLocation GetReturnLocation(size_t index) const {
     return location_sig_->GetReturn(index);

test/cctest/compiler/test-run-native-calls.cc

@@ -21,13 +21,6 @@ namespace v8 {
 namespace internal {
 namespace compiler {
 
-#if V8_TARGET_ARCH_ARM64
-// TODO(titzer): fix native stack parameters on arm64
-#define DISABLE_NATIVE_STACK_PARAMS true
-#else
-#define DISABLE_NATIVE_STACK_PARAMS false
-#endif
-
 namespace {
 typedef float float32;
 typedef double float64;
@@ -221,7 +214,7 @@ class RegisterConfig {
         compiler::Operator::kNoProperties,  // properties
         kCalleeSaveRegisters,               // callee-saved registers
         kCalleeSaveFPRegisters,             // callee-saved fp regs
-        CallDescriptor::kNoFlags,           // flags
+        CallDescriptor::kUseNativeStack,    // flags
         "c-call");
   }
 
@@ -559,8 +552,6 @@ static void TestInt32Sub(CallDescriptor* desc) {
 
 
 static void CopyTwentyInt32(CallDescriptor* desc) {
-  if (DISABLE_NATIVE_STACK_PARAMS) return;
-
   const int kNumParams = 20;
   int32_t input[kNumParams];
   int32_t output[kNumParams];
@@ -672,7 +663,6 @@ TEST_INT32_SUB_WITH_RET(19)
 
 
 TEST(Run_Int32Sub_all_allocatable_single) {
-  if (DISABLE_NATIVE_STACK_PARAMS) return;
   Int32Signature sig(2);
   RegisterPairs pairs;
   while (pairs.More()) {
@@ -690,7 +680,6 @@ TEST(Run_Int32Sub_all_allocatable_single) {
 
 
 TEST(Run_CopyTwentyInt32_all_allocatable_pairs) {
-  if (DISABLE_NATIVE_STACK_PARAMS) return;
   Int32Signature sig(20);
   RegisterPairs pairs;
   while (pairs.More()) {
@@ -743,7 +732,6 @@ static int32_t Compute_Int32_WeightedSum(CallDescriptor* desc, int32_t* input) {
 
 
 static void Test_Int32_WeightedSum_of_size(int count) {
-  if (DISABLE_NATIVE_STACK_PARAMS) return;
   Int32Signature sig(count);
   for (int p0 = 0; p0 < Register::kNumRegisters; p0++) {
     if (Register::from_code(p0).IsAllocatable()) {
@@ -805,8 +793,6 @@ static void RunSelect(CallDescriptor* desc) {
 
 template <int which>
 void Test_Int32_Select() {
-  if (DISABLE_NATIVE_STACK_PARAMS) return;
-
   int parray[] = {
       RegisterConfiguration::ArchDefault(RegisterConfiguration::TURBOFAN)
           ->GetAllocatableGeneralCode(0)};
@@ -930,7 +916,6 @@ TEST(Float64Select_registers) {
 
 
 TEST(Float32Select_stack_params_return_reg) {
-  if (DISABLE_NATIVE_STACK_PARAMS) return;
   int rarray[] = {
       RegisterConfiguration::ArchDefault(RegisterConfiguration::TURBOFAN)
           ->GetAllocatableDoubleCode(0)};
@@ -953,7 +938,6 @@ TEST(Float32Select_stack_params_return_reg) {
 
 
 TEST(Float64Select_stack_params_return_reg) {
-  if (DISABLE_NATIVE_STACK_PARAMS) return;
   int rarray[] = {
       RegisterConfiguration::ArchDefault(RegisterConfiguration::TURBOFAN)
           ->GetAllocatableDoubleCode(0)};
@@ -1008,8 +992,6 @@ static void Build_Select_With_Call(CallDescriptor* desc,
 
 
 TEST(Float64StackParamsToStackParams) {
-  if (DISABLE_NATIVE_STACK_PARAMS) return;
-
   int rarray[] = {
       RegisterConfiguration::ArchDefault(RegisterConfiguration::TURBOFAN)
           ->GetAllocatableDoubleCode(0)};
@@ -1030,7 +1012,6 @@ TEST(Float64StackParamsToStackParams) {
 
 
 void MixedParamTest(int start) {
-  if (DISABLE_NATIVE_STACK_PARAMS) return;
   if (RegisterConfiguration::ArchDefault(RegisterConfiguration::TURBOFAN)
           ->num_double_registers() < 2)
     return;
@@ -1045,7 +1026,7 @@ void MixedParamTest(int start) {
 #else
   static MachineType types[] = {
       kMachInt32,   kMachInt64,   kMachFloat32, kMachFloat64, kMachInt32,
-      kMachFloat64, kMachFloat32, kMachInt64,   kMachFloat64, kMachInt32,
+      kMachFloat64, kMachFloat32, kMachInt64,   kMachInt64,   kMachFloat32,
       kMachFloat32, kMachInt32,   kMachFloat64, kMachFloat64, kMachInt64,
       kMachInt32,   kMachFloat64, kMachInt32,   kMachFloat32};
 #endif