JDK-8186780: clang fastdebug assertion failure in os_linux_x86:os::verify_stack_alignment()

[Martin Buchholz]

Let's try to get the last straggler clang linux failure in OpenJDK resolved.
https://bugs.openjdk.java.net/browse/JDK-8186780
This is not a RFR because we need to get consensus on strategy.
I'm hoping a hotspot developer takes ownership.

---
We should stop trying to deduce stack alignment using non-portable
assembly.  Instead, we should use more portable constructs that ensure that
stack objects are actually correctly aligned.  We can get rid
of os::verify_stack_alignment() and replace it with a single portable
function.  Here's a proof of concept:
(I was surprised to see no pre-existing use of alignof)
(we will soon be using C++14, where presence of alignof is guaranteed)


// Are C++11 alignof, alignas and std::max_align_t available?
// This could instead be implemented via a configure check.
// The __cplusplus macro is not reliable on Microsoft Visual C++:
//
https://devblogs.microsoft.com/cppblog/msvc-now-correctly-reports-__cplusplus/
#if __cplusplus >= 201103L || (defined(_MSC_VER) && _MSC_VER >= 1900)
#define HAVE_ALIGNOF 1
#endif

#ifndef PRODUCT
void os::verify_stack_alignment() {
#ifdef HAVE_ALIGNOF
  STATIC_ASSERT((StackAlignmentInBytes & (alignof(std::max_align_t) - 1))
== 0);

  {
    char jiggle_alignment;
    std::max_align_t object;
    assert(((uintptr_t)&object & (alignof(std::max_align_t) - 1)) == 0,
           "incorrect stack alignment");
  }

  {
    char jiggle_alignment;
    alignas(StackAlignmentInBytes) char object;
    assert(((uintptr_t)&object & (StackAlignmentInBytes - 1)) == 0,
           "incorrect stack alignment");
  }
#endif
}
#endif

---
If we really really want to check the alignment of the stack frame itself
(why?!) then we should not be checking the stack pointer (esp) but instead
using the gcc/clang API __builtin_frame_address(0)

+// TODO: this should be a configure check
+#if defined(__clang__) || defined(__GNUC__)
+#define HAVE_BUILTIN_FRAME_ADDRESS 1
+#endif
+
 #ifndef PRODUCT
 void os::verify_stack_alignment() {
 #ifdef AMD64
-  assert(((intptr_t)os::current_stack_pointer() &
(StackAlignmentInBytes-1)) == 0, "incorrect stack alignment");
+#if defined(HAVE_BUILTIN_FRAME_ADDRESS)
+  void *address = __builtin_frame_address(0);
+#else
+  void *address = os::current_stack_pointer();
+#endif
+  assert(((intptr_t)address & (StackAlignmentInBytes-1)) == 0, "incorrect
stack alignment");
 #endif
 }
 #endif

---
Many other solutions exist - we can simply delete the check, or we can
prevent inlining of  verify_stack_alignment using NOINLINE.