[turbofan][ia32] Remove call to code object via slot.

R=jgruber@chromium.org
BUG=v8:6666,v8:9810

Change-Id: I972983d8e86729843f4a1bbe050e3b37a3c0c61c
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/1881147Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Commit-Queue: Michael Starzinger <mstarzinger@chromium.org>
Cr-Commit-Position: refs/heads/master@{#64578}

src/builtins/builtins-definitions.h

@@ -1283,6 +1283,7 @@ namespace internal {
   V(WasmToNumber)                        \
   V(WasmThrow)                           \
   V(WasmRethrow)                         \
+  V(ArgumentsAdaptorTrampoline)          \
   V(DoubleToI)                           \
   V(WasmI64ToBigInt)                     \
   V(WasmI32PairToBigInt)                 \

src/compiler/backend/ia32/code-generator-ia32.cc

@@ -670,7 +670,7 @@ CodeGenerator::CodeGenResult CodeGenerator::AssembleArchInstruction(
       if (op->IsImmediate()) {
         Handle<Code> code = i.InputCode(0);
         __ Call(code, RelocInfo::CODE_TARGET);
-      } else if (op->IsRegister()) {
+      } else {
         Register reg = i.InputRegister(0);
         DCHECK_IMPLIES(
             HasCallDescriptorFlag(instr, CallDescriptor::kFixedTargetRegister),
@@ -681,23 +681,6 @@ CodeGenerator::CodeGenResult CodeGenerator::AssembleArchInstruction(
         } else {
           __ call(reg);
         }
-      } else {
-        CHECK(tasm()->root_array_available());
-        // This is used to allow calls to the arguments adaptor trampoline from
-        // code that only has 5 gp registers available and cannot call through
-        // an immediate. This happens when the arguments adaptor trampoline is
-        // not an embedded builtin.
-        // TODO(v8:6666): Remove once only embedded builtins are supported.
-        __ push(eax);
-        frame_access_state()->IncreaseSPDelta(1);
-        Operand virtual_call_target_register(
-            kRootRegister, IsolateData::virtual_call_target_register_offset());
-        __ mov(eax, i.InputOperand(0));
-        __ LoadCodeObjectEntry(eax, eax);
-        __ mov(virtual_call_target_register, eax);
-        __ pop(eax);
-        frame_access_state()->IncreaseSPDelta(-1);
-        __ call(virtual_call_target_register);
       }
       RecordCallPosition(instr);
       frame_access_state()->ClearSPDelta();

src/compiler/backend/instruction-selector.cc

@@ -905,7 +905,6 @@ void InstructionSelector::InitializeCallBuffer(Node* call, CallBuffer* buffer,
   bool call_code_immediate = (flags & kCallCodeImmediate) != 0;
   bool call_address_immediate = (flags & kCallAddressImmediate) != 0;
   bool call_use_fixed_target_reg = (flags & kCallFixedTargetRegister) != 0;
-  bool call_through_slot = (flags & kAllowCallThroughSlot) != 0;
   switch (buffer->descriptor->kind()) {
     case CallDescriptor::kCallCodeObject:
       // TODO(jgruber, v8:7449): The below is a hack to support tail-calls from
@@ -919,8 +918,7 @@ void InstructionSelector::InitializeCallBuffer(Node* call, CallBuffer* buffer,
               : call_use_fixed_target_reg
                     ? g.UseFixed(callee, kJavaScriptCallCodeStartRegister)
                     : is_tail_call ? g.UseUniqueRegister(callee)
-                                   : call_through_slot ? g.UseUniqueSlot(callee)
-                                                       : g.UseRegister(callee));
+                                   : g.UseRegister(callee));
       break;
     case CallDescriptor::kCallAddress:
       buffer->instruction_args.push_back(
@@ -2802,15 +2800,6 @@ void InstructionSelector::VisitCall(Node* node, BasicBlock* handler) {
   // Improve constant pool and the heuristics in the register allocator
   // for where to emit constants.
   CallBufferFlags call_buffer_flags(kCallCodeImmediate | kCallAddressImmediate);
-  if (flags & CallDescriptor::kAllowCallThroughSlot) {
-    // TODO(v8:6666): Remove kAllowCallThroughSlot and use a pc-relative call
-    // instead once builtins are embedded in every build configuration.
-    call_buffer_flags |= kAllowCallThroughSlot;
-#ifndef V8_TARGET_ARCH_32_BIT
-    // kAllowCallThroughSlot is only supported on ia32.
-    UNREACHABLE();
-#endif
-  }
   InitializeCallBuffer(node, &buffer, call_buffer_flags, false);
 
   EmitPrepareArguments(&(buffer.pushed_nodes), call_descriptor, node);
@@ -2892,7 +2881,6 @@ void InstructionSelector::VisitTailCall(Node* node) {
   if (callee->flags() & CallDescriptor::kFixedTargetRegister) {
     flags |= kCallFixedTargetRegister;
   }
-  DCHECK_EQ(callee->flags() & CallDescriptor::kAllowCallThroughSlot, 0);
   InitializeCallBuffer(node, &buffer, flags, true, stack_param_delta);
 
   // Select the appropriate opcode based on the call type.

src/compiler/backend/instruction-selector.h

@@ -559,8 +559,7 @@ class V8_EXPORT_PRIVATE InstructionSelector final {
     kCallCodeImmediate = 1u << 0,
     kCallAddressImmediate = 1u << 1,
     kCallTail = 1u << 2,
-    kCallFixedTargetRegister = 1u << 3,
-    kAllowCallThroughSlot = 1u << 4
+    kCallFixedTargetRegister = 1u << 3
   };
   using CallBufferFlags = base::Flags<CallBufferFlag>;
 

src/compiler/linkage.h

@@ -210,14 +210,13 @@ class V8_EXPORT_PRIVATE CallDescriptor final
     // Use the kJavaScriptCallCodeStartRegister (fixed) register for the
     // indirect target address when calling.
     kFixedTargetRegister = 1u << 7,
-    kAllowCallThroughSlot = 1u << 8,
-    kCallerSavedRegisters = 1u << 9,
+    kCallerSavedRegisters = 1u << 8,
     // The kCallerSavedFPRegisters only matters (and set) when the more general
     // flag for kCallerSavedRegisters above is also set.
-    kCallerSavedFPRegisters = 1u << 10,
+    kCallerSavedFPRegisters = 1u << 9,
     // AIX has a function descriptor by default but it can be disabled for a
     // certain CFunction call (only used for Kind::kCallAddress).
-    kNoFunctionDescriptor = 1u << 11,
+    kNoFunctionDescriptor = 1u << 10,
   };
   using Flags = base::Flags<Flag>;
 

src/compiler/wasm-compiler.cc

@@ -5872,8 +5872,9 @@ class WasmWrapperGraphBuilder : public WasmGraphBuilder {
             LOAD_RAW(callable_node,
                      wasm::ObjectAccess::ContextOffsetInTaggedJSFunction(),
                      MachineType::TypeCompressedTaggedPointer());
-        args[pos++] = BuildLoadBuiltinFromIsolateRoot(
-            Builtins::kArgumentsAdaptorTrampoline);
+        args[pos++] = mcgraph()->RelocatableIntPtrConstant(
+            wasm::WasmCode::kArgumentsAdaptorTrampoline,
+            RelocInfo::WASM_STUB_CALL);
         args[pos++] = callable_node;                         // target callable
         args[pos++] = undefined_node;                        // new target
         args[pos++] = mcgraph()->Int32Constant(wasm_count);  // argument count
@@ -5901,16 +5902,10 @@ class WasmWrapperGraphBuilder : public WasmGraphBuilder {
           args[pos++] = undefined_node;
         }
 
-#ifdef V8_TARGET_ARCH_IA32
-        // TODO(v8:6666): Remove kAllowCallThroughSlot and use a pc-relative
-        // call instead once builtins are embedded in every build configuration.
-        CallDescriptor::Flags flags = CallDescriptor::kAllowCallThroughSlot;
-#else
-        CallDescriptor::Flags flags = CallDescriptor::kNoFlags;
-#endif
         auto call_descriptor = Linkage::GetStubCallDescriptor(
             mcgraph()->zone(), ArgumentsAdaptorDescriptor{}, 1 + wasm_count,
-            flags, Operator::kNoProperties);
+            CallDescriptor::kNoFlags, Operator::kNoProperties,
+            StubCallMode::kCallWasmRuntimeStub);
 
         // Convert wasm numbers to JS values.
         pos = AddArgumentNodes(VectorOf(args), pos, wasm_count, sig_);

src/execution/isolate-data.h

@@ -74,11 +74,6 @@ class IsolateData final {
     return builtins_table_offset() + id * kSystemPointerSize;
   }
 
-  // Root-register-relative offset of the virtual call target register value.
-  static constexpr int virtual_call_target_register_offset() {
-    return kVirtualCallTargetRegisterOffset - kIsolateRootBias;
-  }
-
   // The FP and PC that are saved right before TurboAssembler::CallCFunction.
   Address* fast_c_call_caller_fp_address() { return &fast_c_call_caller_fp_; }
   Address* fast_c_call_caller_pc_address() { return &fast_c_call_caller_pc_; }
@@ -131,7 +126,6 @@ class IsolateData final {
   V(kThreadLocalTopOffset, ThreadLocalTop::kSizeInBytes)                      \
   V(kBuiltinEntryTableOffset, Builtins::builtin_count* kSystemPointerSize)    \
   V(kBuiltinsTableOffset, Builtins::builtin_count* kSystemPointerSize)        \
-  V(kVirtualCallTargetRegisterOffset, kSystemPointerSize)                     \
   V(kStackIsIterableOffset, kUInt8Size)                                       \
   /* This padding aligns IsolateData size by 8 bytes. */                      \
   V(kPaddingOffset,                                                           \
@@ -184,11 +178,6 @@ class IsolateData final {
   // The entries in this array are tagged pointers to Code objects.
   Address builtins_[Builtins::builtin_count] = {};
 
-  // For isolate-independent calls on ia32.
-  // TODO(v8:6666): Remove once wasm supports pc-relative jumps to builtins on
-  // ia32 (otherwise the arguments adaptor call runs out of registers).
-  void* virtual_call_target_register_ = nullptr;
-
   // Whether the SafeStackFrameIterator can successfully iterate the current
   // stack. Only valid values are 0 or 1.
   uint8_t stack_is_iterable_ = 1;
@@ -227,8 +216,6 @@ void IsolateData::AssertPredictableLayout() {
   STATIC_ASSERT(offsetof(IsolateData, thread_local_top_) ==
                 kThreadLocalTopOffset);
   STATIC_ASSERT(offsetof(IsolateData, builtins_) == kBuiltinsTableOffset);
-  STATIC_ASSERT(offsetof(IsolateData, virtual_call_target_register_) ==
-                kVirtualCallTargetRegisterOffset);
   STATIC_ASSERT(offsetof(IsolateData, external_memory_) ==
                 kExternalMemoryOffset);
   STATIC_ASSERT(offsetof(IsolateData, external_memory_limit_) ==

src/snapshot/embedded/embedded-data.cc

@@ -200,16 +200,6 @@ EmbeddedData EmbeddedData::FromIsolate(Isolate* isolate) {
         saw_unsafe_builtin = true;
         fprintf(stderr, "%s is not isolate-independent.\n", Builtins::name(i));
       }
-      if (Builtins::IsWasmRuntimeStub(i) &&
-          RelocInfo::RequiresRelocation(code)) {
-        // Wasm additionally requires that its runtime stubs must be
-        // individually PIC (i.e. we must be able to copy each stub outside the
-        // embedded area without relocations). In particular, that means
-        // pc-relative calls to other builtins are disallowed.
-        saw_unsafe_builtin = true;
-        fprintf(stderr, "%s is a wasm runtime stub but needs relocation.\n",
-                Builtins::name(i));
-      }
       if (BuiltinAliasesOffHeapTrampolineRegister(isolate, code)) {
         saw_unsafe_builtin = true;
         fprintf(stderr, "%s aliases the off-heap trampoline register.\n",