cppgc: Remove redzone from stack scan

The trampoline should never inline the actual iteration method and
always use a full call. As a result, there can be no interesting data
in the redzone.

This allows simplifying as the existence and size of a red zone is platform
dependent.

Bug: chromium:1056170
Change-Id: I38d686b0e60fdcc383c40a45e7a81ec0e91f4d62
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2132788Reviewed-by: Anton Bikineev <bikineev@chromium.org>
Commit-Queue: Michael Lippautz <mlippautz@chromium.org>
Cr-Commit-Position: refs/heads/master@{#66955}

src/heap/cppgc/stack.cc

@@ -111,12 +111,7 @@ void Stack::IteratePointersImpl(StackVisitor* visitor,
   // All supported platforms should have their stack aligned to at least
   // sizeof(void*).
   constexpr size_t kMinStackAlignment = sizeof(void*);
-  // Redzone should not contain any pointers as the iteration is always called
-  // from the assembly trampoline. If inline assembly is ever inlined through
-  // LTO this may become necessary.
-  constexpr size_t kRedZoneBytes = 128;
-  void** current = reinterpret_cast<void**>(
-      reinterpret_cast<uintptr_t>(stack_end - kRedZoneBytes));
+  void** current = reinterpret_cast<void**>(stack_end);
   CHECK_EQ(0u, reinterpret_cast<uintptr_t>(current) & (kMinStackAlignment - 1));
   for (; current < stack_start_; ++current) {
     // MSAN: Instead of unpoisoning the whole stack, the slot's value is copied

src/heap/cppgc/stack.h

@@ -44,7 +44,11 @@ class V8_EXPORT_PRIVATE Stack final {
 #endif  // CPPGC_SUPPORTS_CONSERVATIVE_STACK_SCAN
 
  private:
-  void IteratePointersImpl(StackVisitor* visitor, intptr_t* stack_end) const;
+  // Called by the trampoline that pushes registers on the stack. This method
+  // should never be inlined to ensure that a possible redzone cannot contain
+  // any data that needs to be scanned.
+  V8_NOINLINE void IteratePointersImpl(StackVisitor* visitor,
+                                       intptr_t* stack_end) const;
 
   const void* stack_start_;
 };