Add a few GDB pretty printers for MLIR.

[csigg]

Summary

GDB pretty printer script to improve printing for a few MLIR types.

Most printers just remove one layer of indirection when printing an MLIR type that wraps another.

Directly print (only) the first member for the following types:

• mlir::AbstractOperation
• mlir::Location
• mlir::Attribute
• mlir::NamedAttributeList
• mlir::Value

Directly print (only) the first base class for the following types:

• mlir::DictionaryAttr
• mlir::ElementsAttr
• mlir::EnumAttr
• mlir::EnumCaseAttr
• mlir::LocationAttr
• mlir::StructAttr

Custom printers for the following types:

• mlir::Identifier (print identifier string)
• mlir::OperationName (print name string)
• mlir::detail::StorageUserBase (cast and print *impl)

[merge_guards_bot]

Unit tests: pass. 62221 tests passed, 0 failed and 816 were skipped.

clang-tidy: pass.

clang-format: pass.

Build artifacts: diff.json, clang-tidy.txt, clang-format.patch, CMakeCache.txt, console-log.txt, test-results.xml

Pre-merge checks is in beta. Report issue. Please join beta or enable it for your project.

[nicolasvasilache]

Sorry I am a bit lost here since there is no commit message :)
What does this do concretely, how is it used, why do we want this in python etc?
Atm the only thing I see is some python file that prints some stuff.

[csigg]

GDB has a python API to customize how values are printed. See here for an introduction. The API documentation is here.

You can load the script from within gdb with source path/to/prettyprinters.py or put that command into ~/.gdbinit.

The current pretty printers are quite simple, they primarily forward to the interesting member or base class for MLIR types that are mere wrappers.

LLVM already provides pretty printers for some of it's common types.

[csigg]

No change, but I updated the revision summary.

[merge_guards_bot]

Unit tests: pass. 62221 tests passed, 0 failed and 816 were skipped.

clang-tidy: pass.

clang-format: pass.

Build artifacts: diff.json, clang-tidy.txt, clang-format.patch, CMakeCache.txt, console-log.txt, test-results.xml

Pre-merge checks is in beta. Report issue. Please join beta or enable it for your project.

[mehdi_amini]

That's really nice!

Maybe ping @dblaikie ; I think they worked on the LLVM ones and can maybe review here.

[nicolasvasilache]

Very cool!

I am absolutely not qualified to review this but I am happy to rubberstamp to unblock you if no one chimes in by mid-next week.
I am not sure whether this is a step towards using gdb with MLIR or whether you already have it working but if you have a way of GDB MLIR please please please give a presentation at the open design meeting (and internally if you get an earlier spot :) ).

Assuming you can also GDB MLIR somehow, would it also work with JIT'ed MLIR?

[rriddle]

In D73546#1852321, @nicolasvasilache wrote:

Assuming you can also GDB MLIR somehow, would it also work with JIT'ed MLIR?

You mean debug the generated code? That requires some changes to execution engine, as well as emission of line information to LLVM IR. I had a patch to add enough support for line stepping JIT'ed code, but it needs to be updated.

[csigg]

@nicolasvasilache, it's currently blocked on D73609 and D73726. Then I will add tests and send it to @dblaikie for review.

[dblaikie]

mlir/utils/gdb-scripts/prettyprinters.py

21–22

What types would not have a default visualizer/what case is this intended to handle?

53–54

When does this code path happen (when does a type's field have a None type?

64–74

Looks easy/simple enough & could be committed separately.

mlir/utils/gdb-scripts/prettyprinters.py

+"""GDB pretty printers for MLIR types."""
+
+import gdb.printing
+
+class StructPrinter:
+  """Prints (optionally) bases of a struct and it's fields."""
+
+  def __init__(self, val, print_base = True):
+    self.val = val
+    self.print_base = print_base
+
+  def children(self):
+    for field in self.val.type.values():
+      if field.is_base_class:
+        if self.print_base:
+          yield (field.name, self.val.cast(field.type))
+      else:
+        yield (field.name, self.val[field.name])
+
+def get_default_or_struct_printer(val, print_base = True):
+  """Returns gdb.default_visualizer(val) or fall back to StructPrinter."""
+  default_printer = gdb.default_visualizer(val)
+  if default_printer:
+    return default_printer
+  return StructPrinter(val, print_base)
+
+class FieldPrinterFactory:
+  """Returns printer of one of the fields."""
+
+  def __init__(self, name):
+    self.name = name
+
+  def __call__(self, val):
+    field = val[self.name]
+    if field.type.code == gdb.TYPE_CODE_PTR:
+      field = field.dereference()
+    return get_default_or_struct_printer(field)
+
+def get_first_member_printer(val):
+  """Returns printer for the first member of val."""
+  for field in val.type.fields():
+    if not field.is_base_class:
+      return FieldPrinterFactory(field.name)(val)
+  return None
+
+class CastPrinterFactory:
+  """Returns printer after casting the value."""
+
+  def __init__(self, type):
+    self.type = type
+
+  def __call__(self, val):
+    if not self.type:
+      return None
+    return get_default_or_struct_printer(val.cast(self.type))
+
+def get_first_base_printer(val):
+  """Returns printer for the first base class of val."""
+  for field in val.type.fields():
+    if field.is_base_class:
+      return CastPrinterFactory(field.type)(val)
+  return None
+
+class IdentifierPrinter:
+  """Prints an mlir::Identifier instance."""
+
+  def __init__(self, val):
+    self.val = val
+
+  def to_string(self):
+    return self.val['pointer'].string()
+
+  def display_hint(self):
+    return 'string'
+
+def get_operation_name_printer(val):
+  """Returns printer for an mlir::OperationName instance."""
+  rep_printer = gdb.default_visualizer(val['representation'])
+  if not rep_printer:
+    return None
+  if not hasattr(rep_printer, 'pointer'):
+    return rep_printer
+  pointer = rep_printer.pointer
+  if pointer.type.code == gdb.TYPE_CODE_PTR:
+    pointer = pointer.dereference()
+  return get_default_or_struct_printer(pointer)
+
+def get_storage_user_base_printer(val):
+  """Returns printer for an mlir::detail::StorageUserBase instance."""
+  if not val['impl']:
+    return None
+  storage_type = val.type.template_argument(2)
+  impl = val['impl'].dereference().cast(storage_type)
+  return get_default_or_struct_printer(impl)
+
+pp = gdb.printing.RegexpCollectionPrettyPrinter('MLIRSupport')
+
+# Forwarding printers to only relevant field.
+pp.add_printer('mlir::AbstractOperation', '^mlir::AbstractOperation$', FieldPrinterFactory('name'))
+for name in ['Location', 'Attribute', 'NamedAttributeList', 'Value']:
+  pp.add_printer('mlir::%s' % name, '^mlir::%s' % name, get_first_member_printer)
+
+# Forwarding printers to the only relevant base class.
+for name in ['DictionaryAttr', 'ElementsAttr', 'EnumAttr', 'EnumCaseAttr', 'LocationAttr', 'StructAttr']:
+  pp.add_printer('mlir::%s' % name, '^mlir::%s' % name, get_first_base_printer)
+
+pp.add_printer('mlir::Identifier', '^mlir::Identifier$', IdentifierPrinter)
+pp.add_printer('mlir::OperationName', '^mlir::OperationName$', get_operation_name_printer)
+pp.add_printer('mlir::detail::StorageUserBase', '^mlir::detail::StorageUserBase<.*>', get_storage_user_base_printer)
+
+gdb.printing.register_pretty_printer(gdb.current_objfile(), pp, True)