[TableGen] Add a general-purpose JSON backend.

[simon_tatham]

Summary

The aim of this backend is to output everything TableGen knows about
the record set, similarly to the default -print-records backend. But
where -print-records produces output in TableGen's input syntax
(convenient for humans to read), this backend produces it as
structured JSON data, which is convenient for loading into standard
scripting languages such as Python, in order to extract information
from the data set in an automated way.

The output data contains a JSON representation of the variable
definitions in output 'def' records, and a few pieces of metadata such
as which of those definitions are tagged with the 'field' prefix and
which defs are derived from which classes. It doesn't dump out
absolutely every piece of knowledge it _could_ produce, such as type
information and complicated arithmetic operator nodes in abstract
superclasses; the main aim is to allow consumers of this JSON dump to
essentially act as new backends, and backends don't generally need to
depend on that kind of data.

The new backend is implemented as an EmitJSON() function similar to
all of llvm-tblgen's other EmitFoo functions, except that it lives in
lib/TableGen instead of utils/TableGen on the basis that I'm expecting
to add it to clang-tblgen too in a future patch.

To test it, I've written a Python script that loads the JSON output
and tests properties of it based on comments in the .td source - more
or less like FileCheck, except that the CHECK: lines have Python
expressions after them instead of textual pattern matches.

[labath]

utils/TableGen/JSONEmitter.cpp

1 ↗

(On Diff #143909)

You should take a look at <D45753>, which is about to add a JSON library to llvm. It would be a shame to add two of them in the same week. :)

[simon_tatham]

utils/TableGen/JSONEmitter.cpp

1 ↗

(On Diff #143909)

Ha! You're right, I hadn't noticed that. I'll replace my ad-hockery with calls to that code with great pleasure as soon as it lands – even without looking I'm sure it will be better than the 'only just enough' code I have here. Thanks for pointing it out!

[simon_tatham]

@nhaehnle : while this change is blocked on that one, do you have any opinions on my design questions? Particularly the one about whether I ought to move most of the code into some sort of getAsJSONObject method in the various classes in Record.h / Record.cpp, because if I'm going to do that then I should do it before making any other detailed changes :-)

[nhaehnle]

Consider what to do about integer values that don't fit exactly into a 'double'. This code will simply emit them as decimal integer literals, which JSON parsers are within their rights to round to the nearest double precision float, losing data. Some JSON readers (e.g. Python json.load) will deliver accurate integer values anyway, but it might be better not to rely on that, and instead output very large integers in some other form, such as a JSON object containing an identifying type field and two doubles whose sum is the desired integer, or a string representation of the integer, or both.

Well. The one thing I do have a strong opinion on is that the representation of very large integers should not be different from the representation of small integers, because consumers would almost certainly get that wrong. So that leaves two options: string representations, or sending JavaScript to the hell it deserves and going with integers.

Consider adding the new -dump-json option to clang-tblgen as well as llvm-tblgen. (As I understand it they wouldn't do anything differently, but it seems asymmetric not to have both of them support it. They both have -print-records, after all.)

Sure, that seems like a good idea.

Consider providing a cut-down version, enabled by another option such as '-dump-simple-json', in which all the complicated parametric expression nodes like !add and !foreach and !foldl are simply not emitted, and replaced by some kind of small object indicating that a complex expression was elided. The motivation is that I expect a lot of uses for this system would only be interested in the output fields that consist of final well-defined values of primitive type, so constructing the complicated parts is a waste of both TableGen's time and the consumer's. But I'm not sure where the line should be drawn - DAG arguments might well still need to be output in full, for example, and type information might be omittable. There may be no one good answer.

I believe there is a good answer :)

Take a look at TGParser.cpp, checkConcrete. All valid final and needed records should pass that check. There are unfortunately still some exceptions, although it'd be nice if we could get rid of those actually.

So I would argue that you should only bother emitting what fits this definition of "concrete", i.e. don't have a "complex" JSON option at all. If you do encounter something that doesn't fit the definition of concrete, just output a JSON object { "kind": "complex", "code": <getAsString()> } in its place.

On a related note, I'm not convinced you really want to print out class definitions. The benefit of printing class definitions with -print-records is that it can help you understand what's going on while you're writing class definitions, but the existing TableGen backends really don't care about class definitions. Since the idea here is to basically allow writing TableGen backends in a scripting language like Python, providing the class definitions is unlikely to be useful. It doesn't hurt, but I don't think it's a good motivation for building all this infrastructure for printing ! operations, for example.

Decide where all this code should live. It might be better to move a lot of it into Record.cpp in the form of getAsJSONObject() methods or something like that. That would remove the risk of forgetting to update the JSON back end if a new node type is introduced - anyone forgetting to implement that method in any new subclass of Init or RecTy would be reminded by a compile error.

I'd rather keep it separate for orthogonality. Adding a new fundamental concrete data type is a rare enough occurence, and the worst that would happen with my suggestion above is that you get an unexpected "complex" object, which is not too difficult to track down.

[simon_tatham]

Well. The one thing I do have a strong opinion on is that the representation of very large integers should not be different from the representation of small integers, because consumers would almost certainly get that wrong.

A good point – now you put it that way, I suddenly agree strongly!

So that leaves two options: string representations, or sending JavaScript to the hell it deserves and going with integers.

Well, since my personal use cases all involve Python, and Python copes fine with arbitrary integers, I'm happy with the latter if you are :-) And I'd definitely prefer not to have to put a decode-from-string operation in every single consumer of this JSON output.

I suppose a cl::opt<bool> to switch to a more cumbersome representation of integers could always be added later, if anyone turns out to really need one.

Take a look at TGParser.cpp, checkConcrete. All valid final and needed records should pass that check. There are unfortunately still some exceptions, although it'd be nice if we could get rid of those actually.

Ah! Yes, that seems nice. And if I'm not dumping the class definitions, then perhaps it's not worth dumping the details of all the types either, for the same reason (a 'back end' consuming this format will already know what type to expect from any field it cares about), in which case I can simplify the output representation a great deal by removing the extra level of dereference where you have to suffix ['value'] all the time.

I agree that none of my own use cases will care about any of the things that this redesign throws away – and it makes the output JSON a great deal smaller and simpler. Thanks for the suggestions!

[simon_tatham]

OK, here's my second draft. Changes since last time:

• thrown out the ad-hoc JSON emitter in favour of the new JSON library in D45753 (also requires the integer-handling followup patch D46209)
• moved the new source file into lib/TableGen where clang-tblgen will be able to get at it more easily (but I haven't actually added it to clang-tblgen yet)
• removed all the type and abstract class information, leaving only the concrete records and a couple of pieces of metadata that I know backends do actually want (list of field keywords, list of superclasses, list of instances of each class). Exotic subclasses of Init are now rendered as kind="complex" with only a printable representation.
• flattened the JSON structure by several layers to make it more convenient to consume
• added documentation of the format.

I think from my perspective this is no longer an unfinished draft; I'd be happy to commit it in this state, subject to code review approval and its dependencies landing.

[nhaehnle]

Thanks, this already looks very good. I do have some suggestions though.

docs/TableGen/BackEnds.rst

516–518

This is a minor point, but I suspect it would be slightly more convenient for consumers if this were instead represented as [[arg, name], [arg, name], ...]. So the example below would have args be [[22, null], ["hello", "foo"]]. What do you think?

lib/TableGen/JSONBackend.cpp

46–49

C-style comments are rather uncommon in LLVM; best to be consistent and use C++-style comments here (and below).

93

I think this should be "var" instead of "variable", for consistency.

172–179

I think it would be slightly nicer to merge this into the loop above, to avoid iterating over the same data twice.

test/TableGen/JSON-check.py

49

Cool, I didn't know this was possible.

[simon_tatham]

docs/TableGen/BackEnds.rst

516–518

That's actually more like how I had it in the first version of the patch, and I had second thoughts and changed it to this :-) so I'm already on the fence and could easily be persuaded to change it back again.

My thought was that some use cases wouldn't care about the name at all (e.g. an entire backend might use dag-typed data for some purpose that never gives a name to any argument), and those users would find it more convenient if they could get the actual arguments by just saying node['args'][n] instead of having to say node['args'][n][0] (in your version) or node['args'][n]['value'] (as I originally had it). So I moved the names out into a separate array that you'd only have to look at if you cared about names at all.

On the other hand, I can certainly see the counterargument – if you do care about names, it's nicer to have a single array to iterate over. I'm happy to change it to your style.

lib/TableGen/JSONBackend.cpp

93

Another thing I was trying to do (but forgot to actually mention anywhere) was to arrange for all the field names that depend on "kind" not to be the same as each other, as a means of error detection – it would stop a user accidentally mistaking a var for a varbit, by absentmindedly retrieving node['var'] and forgetting there was another field to look at too. But that's quite a marginal consideration in the first place, and also I admit this particular choice of two different names was terrible :-) so yes, I'm happy to change over to being consistent.

test/TableGen/JSON-check.py

49

You mean passing a string to sys.exit? That was new to me as well quite recently. It's one of those functions where as a C programmer I automatically assumed I already knew how its API would work, so it took me years to find a reason to read its docs!

[simon_tatham]

Updated with all those review comments.

[nhaehnle]

Great, LGTM!

docs/TableGen/BackEnds.rst

516–518

Yeah, I admit it was really a minor thing. Thanks for changing it though!

[simon_tatham]

Thanks for the review!

Of course, having got to this point, I can't actually commit it until D45753 is ready. And I'm about to be on holiday for several weeks, so if that happens soon then I probably won't notice for a while. But I won't actually forget about this patch, I promise :-)

[simon_tatham]

@nhaehnle, following up discussion about NAME in D47430 (and hopefully posted in the right place this time):

I was going to change this patch so that it uses !name instead of NAME for the key inside each JSON def object that gives the def's own name. (Mostly the reason I think it's useful to have such a key at all is so that client code that's consuming the JSON can pass those dictionaries to its own subfunctions without having to pass the name alongside it, so it makes sense to me to use a key that indicates that it's a JSON-specific convenience.)

But it's just struck me that there might also be a use case for knowing whether the record is anonymous or not, in the sense of whether its name is something that was deliberately specified in the TableGen input or whether it was some anonymous_123 value made up by tblgen itself.

I was thinking of adding a boolean field !anonymous, or alternatively perhaps having !key and !name (where !key is always the key under which this record is stored in the JSON root object, and !name is either the same as !key or null). Any particular preference, before I make the changes? Or is the entire idea not worth bothering with?

(Also, I noticed in passing that the IsAnonymous field isn't set reliably: the records defined by an anonymous defm have it false rather than true. That looks easy to fix, and I could fold that into this change or make it a separate one.)

[nhaehnle]

There are arguments both for !key + !name and for !name + !anonymous, although thinking about it for a minute or two I weakly prefer !name + !anonymous because it matches the representation in C++. It makes it easier for people to move between JSON and C++.

P.S.: No worries about the confusion with the other review :)

[nhaehnle]

And by the way, I do agree with your rationale for why !name is very useful to have in JSON. The C++ backends can (and do) use Record::getName() for the same functionality.

[simon_tatham]

Renamed the NAME attribute to !name in line with D47430.

Added the !anonymous attribute, a set of tests for it, and a fix in TGParser.cpp to set it correctly for anonymous defms.

[simon_tatham]

OK, !name + !anonymous it is. (That was how I'd drafted it too, so I think we had the same mild preference.)

I'm afraid this patch now has a very tiny conflict with D47430, in that we've both added braces to the same if statement in TGParser::ParseDefm.

[simon_tatham]

Rebase to current trunk, and revert change to 'anonymous' handling.

In line with the discussion in D47431, I've removed my previous tweak
in TGParser.cpp that sets the !anonymous flag if any part of a def's
final name was derived from an anonymous def or defm. Now the
!anonymous field in the JSON output is consistent with the existing
behavior of the isAnonymous() query function.

[simon_tatham]

@nhaehnle , are you still happy for me to commit this, now that its dependencies have landed and I've tweaked its handing of !anonymous?

[simon_tatham]

Nearly forgot to mention the new option in the tblgen man page!

[simon_tatham]

Actually, on second thoughts, I'm going to assume it was overcautious to ask for a re-approval, since this version of the patch introduces no new controversy and in fact removes the only previous tweak in the Tablegen core (in that I'm not trying to change the semantics of anonymous any more). So I'll commit this as is, based on the previous review approval.

[stella.stamenova]

The test json.td fails on Windows when there's a space in the path to python. python should be specified in the test file as '%python' or "%python" rather than just '%python. I can submit a fix for review if you're not able. Thanks!

[simon_tatham]

Oops, sorry about that :-(

(I admit I didn't consider the possibility of spacey filenames at all, but now I have, it's a mild surprise to me that %python doesn't expand to an already correctly-quoted string.)

The fix sounds easy, but if you have a setup where you can actually check it works, there might be less chance of a typo if you do it rather than me...

docs/TableGen/BackEnds.rst

 Generate ARM NEON tests for clang.
 
 AttrDocs
 --------
 
 **Purpose**: Creates ``AttributeReference.rst`` from ``AttrDocs.td``, and is
 used for documenting user-facing attributes.
 
+General BackEnds
+================
+
+JSON
+----
+
+**Purpose**: Output all the values in every ``def``, as a JSON data
+structure that can be easily parsed by a variety of languages. Useful
+for writing custom backends without having to modify TableGen itself,
+or for performing auxiliary analysis on the same TableGen data passed
+to a built-in backend.
+
+**Output**:
+
+The root of the output file is a JSON object (i.e. dictionary),
+containing the following fixed keys:
+
+* ``!tablegen_json_version``: a numeric version field that will
+  increase if an incompatible change is ever made to the structure of
+  this data. The format described here corresponds to version 1.
+
+* ``!instanceof``: a dictionary whose keys are the class names defined
+  in the TableGen input. For each key, the corresponding value is an
+  array of strings giving the names of ``def`` records that derive
+  from that class. So ``root["!instanceof"]["Instruction"]``, for
+  example, would list the names of all the records deriving from the
+  class ``Instruction``.
+
+For each ``def`` record, the root object also has a key for the record
+name. The corresponding value is a subsidiary object containing the
+following fixed keys:
+
+* ``!superclasses``: an array of strings giving the names of all the
+  classes that this record derives from.
+
+* ``!fields``: an array of strings giving the names of all the variables
+  in this record that were defined with the ``field`` keyword.
+
+For each variable defined in a record (including the special implicit
+variable ``NAME``), the ``def`` object for that record also has a key
+for the variable name. The corresponding value is a translation into
+JSON of the variable's value, using the conventions described below.
+
+Some TableGen data types are translated directly into the
+corresponding JSON type:
+
+* A completely undefined value (e.g. for a variable declared without
+  initializer in some superclass of this record, and never initialized
+  by the record itself or any other superclass) is emitted as the JSON
+  ``null`` value.
+
+* ``int`` and ``bit`` values are emitted as numbers. Note that
+  TableGen ``int`` values are capable of holding integers too large to
+  be exactly representable in IEEE double precision. The integer
+  literal in the JSON output will show the full exact integer value.
+  So if you need to retrieve large integers with full precision, you
+  should use a JSON reader capable of translating such literals back
+  into 64-bit integers without losing precision, such as Python's
+  standard ``json`` module.
+
+* ``string`` and ``code`` values are emitted as JSON strings.
+
+* ``list<T>`` values, for any element type ``T``, are emitted as JSON
+  arrays. Each element of the array is represented in turn using these
+  same conventions.
+
+* ``bits`` values are also emitted as arrays. A ``bits`` array is
+  ordered from least-significant bit to most-significant. So the
+  element with index ``i`` corresponds to the bit described as
+  ``x{i}`` in TableGen source. However, note that this means that
+  scripting languages are likely to *display* the array in the
+  opposite order from the way it appears in the TableGen source or in
+  the diagnostic ``-print-records`` output.
+
+All other TableGen value types are emitted as a JSON object,
+containing two standard fields: ``kind`` is a discriminator describing
+which kind of value the object represents, and ``printable`` is a
+string giving the same representation of the value that would appear
+in ``-print-records``.
+
+* A reference to a ``def`` object has ``kind=="def"``, and has an
+  extra field ``def`` giving the name of the object referred to.
+
+* A reference to another variable in the same record has
+  ``kind=="var"``, and has an extra field ``var`` giving the name of
+  the variable referred to.
+
+* A reference to a specific bit of a ``bits``-typed variable in the
+  same record has ``kind=="varbit"``, and has two extra fields:
+  ``variable`` gives the name of the variable referred to, and
+  ``index`` gives the index of the bit.
+
+* A value of type ``dag`` has ``kind=="dag"``, and has three extra
+  fields. ``operator`` gives the initial value after the opening
+  parenthesis of the dag initializer; ``args`` is an array of the
+  following arguments; and ``argnames`` is a list giving the
+  colon-suffixed name of each argument, if any. For example, in the
+  JSON representation of the dag value ``(Op 22, "hello":$foo)``
+  (assuming that ``Op`` is the name of a record defined elsewhere with
+  a ``def`` statement):
+
+  * ``operator`` will be an object in which ``kind=="def"`` and
+    ``def=="Op"``
+
+  * ``args`` will be the array ``[22, "hello"]``
+
+  * ``argnames`` will be the array ``[null, "foo"]``.
+
+* If any other kind of value or complicated expression appears in the
+  output, it will have ``kind=="complex"``, and no additional fields.
+  These values are not expected to be needed by backends. The standard
+  ``printable`` field can be used to extract a representation of them
+  in TableGen source syntax if necessary.
+
 How to write a back-end
 =======================
 
 TODO.
 
 Until we get a step-by-step HowTo for writing TableGen backends, you can at
 least grab the boilerplate (build system, new files, etc.) from Clang's
 r173931.
 
 TODO: How they work, how to write one.  This section should not contain details
 about any particular backend, except maybe ``-print-enums`` as an example.  This
 should highlight the APIs in ``TableGen/Record.h``.
 

docs/TableGen/index.rst

.. code-block:: bash
 
   $ llvm-tblgen X86.td -print-enums -class=Instruction 
   ABS_F, ABS_Fp32, ABS_Fp64, ABS_Fp80, ADC32mi, ADC32mi8, ADC32mr, ADC32ri,
   ADC32ri8, ADC32rm, ADC32rr, ADC64mi32, ADC64mi8, ADC64mr, ADC64ri32, ADC64ri8,
   ADC64rm, ADC64rr, ADD16mi, ADD16mi8, ADD16mr, ADD16ri, ADD16ri8, ADD16rm,
   ADD16rr, ADD32mi, ADD32mi8, ADD32mr, ADD32ri, ADD32ri8, ADD32rm, ADD32rr,
   ADD64mi32, ADD64mi8, ADD64mr, ADD64ri32, ...
 
-The default backend prints out all of the records.
+The default backend prints out all of the records. There is also a general
+backend which outputs all the records as a JSON data structure, enabled using
+the `-dump-json` option.
 
 If you plan to use TableGen, you will most likely have to write a `backend`_
 that extracts the information specific to what you need and formats it in the
-appropriate way.
+appropriate way. You can do this by extending TableGen itself in C++, or by
+writing a script in any language that can consume the JSON output.
 
 Example
 -------
 
 With no other arguments, `llvm-tblgen` parses the specified file and prints out all
 of the classes, then all of the definitions.  This is a good way to see what the
 various definitions expand to fully.  Running this on the ``X86.td`` file prints
 this (at the time of this writing):

include/llvm/TableGen/Record.h

public:
   explicit TrackUnresolvedResolver(Resolver *R = nullptr)
       : Resolver(R ? R->getCurrentRecord() : nullptr), R(R) {}
 
   bool foundUnresolved() const { return FoundUnresolved; }
 
   Init *resolve(Init *VarName) override;
 };
 
+void EmitJSON(RecordKeeper &RK, raw_ostream &OS);
+
 } // end namespace llvm
 
 #endif // LLVM_TABLEGEN_RECORD_H

lib/TableGen/CMakeLists.txt

 add_llvm_library(LLVMTableGen
   Error.cpp
+  JSONBackend.cpp
   Main.cpp
   Record.cpp
   SetTheory.cpp
   StringMatcher.cpp
   TableGenBackend.cpp
   TGLexer.cpp
   TGParser.cpp
 
   ADDITIONAL_HEADER_DIRS
   ${LLVM_MAIN_INCLUDE_DIR}/llvm/TableGen
   )

lib/TableGen/JSONBackend.cpp

+//===- JSONBackend.cpp - Generate a JSON dump of all records. -*- C++ -*-=====//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This TableGen back end generates a machine-readable representation
+// of all the classes and records defined by the input, in JSON format.
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/ADT/BitVector.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/TableGen/Error.h"
+#include "llvm/TableGen/Record.h"
+#include "llvm/TableGen/TableGenBackend.h"
+#include "llvm/Support/JSON.h"
+
+#define DEBUG_TYPE "json-emitter"
+
+using namespace llvm;
+
+namespace {
+
+class JSONEmitter {
+private:
+  RecordKeeper &Records;
+
+  json::Value translateInit(const Init &I);
+  json::Array listSuperclasses(const Record &R);
+
+public:
+  JSONEmitter(RecordKeeper &R);
+
+  void run(raw_ostream &OS);
+};
+
+} // end anonymous namespace
+
+JSONEmitter::JSONEmitter(RecordKeeper &R) : Records(R) {}
+
+json::Value JSONEmitter::translateInit(const Init &I) {
+  /*
+   * Init subclasses that we return as JSON primitive values of one
+   * kind or another.
+   */
+  if (isa<UnsetInit>(&I)) {
+    return nullptr;
+  } else if (auto *Bit = dyn_cast<BitInit>(&I)) {
+    return Bit->getValue() ? 1 : 0;
+  } else if (auto *Bits = dyn_cast<BitsInit>(&I)) {
+    json::Array array;
+    for (unsigned i = 0, limit = Bits->getNumBits(); i < limit; i++)
+      array.push_back(translateInit(*Bits->getBit(i)));
+    return array;
+  } else if (auto *Int = dyn_cast<IntInit>(&I)) {
+    return Int->getValue();
+  } else if (auto *Str = dyn_cast<StringInit>(&I)) {
+    return Str->getValue();
+  } else if (auto *Code = dyn_cast<CodeInit>(&I)) {
+    return Code->getValue();
+  } else if (auto *List = dyn_cast<ListInit>(&I)) {
+    json::Array array;
+    for (auto val : *List)
+      array.push_back(translateInit(*val));
+    return array;
+  }
+
+  /*
+   * Init subclasses that we return as JSON objects containing a
+   * 'kind' discriminator. For these, we also provide the same
+   * translation back into TableGen input syntax that -print-records
+   * would give.
+   */
+
+  json::Object obj;
+  obj["printable"] = I.getAsString();
+
+  if (auto *Def = dyn_cast<DefInit>(&I)) {
+    obj["kind"] = "def";
+    obj["def"] = Def->getDef()->getName();
+    return obj;
+  } else if (auto *Var = dyn_cast<VarInit>(&I)) {
+    obj["kind"] = "var";
+    obj["var"] = Var->getName();
+    return obj;
+  } else if (auto *VarBit = dyn_cast<VarBitInit>(&I)) {
+    if (auto *Var = dyn_cast<VarInit>(VarBit->getBitVar())) {
+      obj["kind"] = "varbit";
+      obj["variable"] = Var->getName();
+      obj["index"] = VarBit->getBitNum();
+      return obj;
+    }
+  } else if (auto *Dag = dyn_cast<DagInit>(&I)) {
+    obj["kind"] = "dag";
+    obj["operator"] = translateInit(*Dag->getOperator());
+    if (auto name = Dag->getName())
+      obj["name"] = name->getAsUnquotedString();
+    json::Array args, argnames;
+    for (unsigned i = 0, limit = Dag->getNumArgs(); i < limit; ++i) {
+      json::Object arg;
+      args.push_back(translateInit(*Dag->getArg(i)));
+      if (auto argname = Dag->getArgName(i))
+        argnames.push_back(argname->getAsUnquotedString());
+      else
+        argnames.push_back(nullptr);
+    }
+    obj["args"] = std::move(args);
+    obj["argnames"] = std::move(argnames);
+    return obj;
+  }
+
+  /*
+   * Final fallback: anything that gets past here is simply given a
+   * kind field of 'complex', and the only other field is the standard
+   * 'printable' representation.
+   */
+  assert(!I.isConcrete());
+  obj["kind"] = "complex";
+  return obj;
+}
+
+void JSONEmitter::run(raw_ostream &OS) {
+  json::Object root;
+
+  root["!tablegen_json_version"] = 1;
+
+  for (const auto &D : Records.getDefs()) {
+    auto &Name = D.second->getNameInitAsString();
+    auto &Def = *D.second;
+
+    json::Object obj;
+    json::Array fields;
+
+    for (const RecordVal &RV : Def.getValues()) {
+      if (!Def.isTemplateArg(RV.getNameInit())) {
+        auto Name = RV.getNameInitAsString();
+        if (RV.getPrefix())
+          fields.push_back(Name);
+        obj[Name] = translateInit(*RV.getValue());
+      }
+    }
+
+    obj["!fields"] = std::move(fields);
+
+    json::Array superclasses;
+    for (const auto &SuperPair : Def.getSuperClasses())
+      superclasses.push_back(SuperPair.first->getNameInitAsString());
+    obj["!superclasses"] = std::move(superclasses);
+
+    // The special "NAME" field, documented as giving the record's
+    // symbolic name as a string, is not reliably defined in the
+    // actual Record structure itself - the '-print-records' output
+    // lists it as ? most of the time. Scripts processing this JSON
+    // output will likely find it convenient to be able to pass around
+    // def objects without having to separately remember their name,
+    // so we ensure here that NAME is properly filled in in the JSON
+    // output.
+    obj["NAME"] = Name;
+
+    root[Name] = std::move(obj);
+  }
+
+  std::map<std::string, json::Array> instance_lists;
+  for (const auto &C : Records.getClasses()) {
+    auto &Name = C.second->getNameInitAsString();
+    (void)instance_lists[Name]; // ensure even empty arrays exist
+  }
+  for (const auto &D : Records.getDefs()) {
+    auto &Name = D.second->getNameInitAsString();
+    auto &Def = *D.second;
+    for (const auto &SuperPair : Def.getSuperClasses()) {
+      auto SuperName = SuperPair.first->getNameInitAsString();
+      instance_lists[SuperName].push_back(Name);
+    }
+  }
+  json::Object instanceof;
+  for (auto kv: instance_lists)
+    instanceof[kv.first] = std::move(kv.second);
+  root["!instanceof"] = std::move(instanceof);
+
+  OS << json::Value(std::move(root)) << "\n";
+}
+
+namespace llvm {
+
+void EmitJSON(RecordKeeper &RK, raw_ostream &OS) { JSONEmitter(RK).run(OS); }
+} // end namespace llvm

test/TableGen/JSON-check.py

+#!/usr/bin/env python
+
+import sys
+import subprocess
+import traceback
+import json
+
+data = json.load(sys.stdin)
+testfile = sys.argv[1]
+
+prefix = "CHECK: "
+
+fails = 0
+passes = 0
+with open(testfile) as testfh:
+    lineno = 0
+    for line in iter(testfh.readline, ""):
+        lineno += 1
+        line = line.rstrip("\r\n")
+        try:
+            prefix_pos = line.index(prefix)
+        except ValueError:
+            continue
+        check_expr = line[prefix_pos + len(prefix):]
+
+        try:
+            exception = None
+            result = eval(check_expr, {"data":data})
+        except Exception:
+            result = False
+            exception = traceback.format_exc().splitlines()[-1]
+
+        if exception is not None:
+            sys.stderr.write(
+                "{file}:{line:d}: check threw exception: {expr}\n"
+                "{file}:{line:d}: exception was: {exception}\n".format(
+                    file=testfile, line=lineno,
+                    expr=check_expr, exception=exception))
+            fails += 1
+        elif not result:
+            sys.stderr.write(
+                "{file}:{line:d}: check returned False: {expr}\n".format(
+                    file=testfile, line=lineno, expr=check_expr))
+            fails += 1
+        else:
+            passes += 1
+
+if fails != 0:
+    sys.exit("{} checks failed".format(fails))
+else:
+    sys.stdout.write("{} checks passed\n".format(passes))

test/TableGen/JSON.td

+// RUN: llvm-tblgen -dump-json %s | %python %S/JSON-check.py %s
+
+// CHECK: data['!tablegen_json_version'] == 1
+
+// CHECK: all(data[s]['NAME'] == s for s in data if not s.startswith("!"))
+
+class Base {}
+class Intermediate : Base {}
+class Derived : Intermediate {}
+
+def D : Intermediate {}
+// CHECK: 'D' in data['!instanceof']['Base']
+// CHECK: 'D' in data['!instanceof']['Intermediate']
+// CHECK: 'D' not in data['!instanceof']['Derived']
+// CHECK: 'Base' in data['D']['!superclasses']
+// CHECK: 'Intermediate' in data['D']['!superclasses']
+// CHECK: 'Derived' not in data['D']['!superclasses']
+
+def ExampleDagOp;
+
+def FieldKeywordTest {
+    int a;
+    field int b;
+    // CHECK: 'a' not in data['FieldKeywordTest']['!fields']
+    // CHECK: 'b' in data['FieldKeywordTest']['!fields']
+}
+
+class Variables {
+    int i;
+    string s;
+    bit b;
+    bits<8> bs;
+    code c;
+    list<int> li;
+    Base base;
+    dag d;
+}
+def VarNull : Variables {
+    // A variable not filled in at all has its value set to JSON
+    // 'null', which translates to Python None
+    // CHECK: data['VarNull']['i'] is None
+}
+def VarPrim : Variables {
+    // Test initializers that map to primitive JSON types
+
+    int i = 3;
+    // CHECK: data['VarPrim']['i'] == 3
+
+    // Integer literals should be emitted in the JSON at full 64-bit
+    // precision, for the benefit of JSON readers that preserve that
+    // much information. Python's is one such.
+    int enormous_pos = 9123456789123456789;
+    int enormous_neg = -9123456789123456789;
+    // CHECK: data['VarPrim']['enormous_pos'] == 9123456789123456789
+    // CHECK: data['VarPrim']['enormous_neg'] == -9123456789123456789
+
+    string s = "hello, world";
+    // CHECK: data['VarPrim']['s'] == 'hello, world'
+
+    bit b = 0;
+    // CHECK: data['VarPrim']['b'] == 0
+
+    // bits<> arrays are stored in logical order (array[i] is the same
+    // bit identified in .td files as bs{i}), which means the _visual_
+    // order of the list (in default rendering) is reversed.
+    bits<8> bs = { 0,0,0,1,0,1,1,1 };
+    // CHECK: data['VarPrim']['bs'] == [ 1,1,1,0,1,0,0,0 ]
+
+    code c = [{ \"  }];
+    // CHECK: data['VarPrim']['c'] == r' \"  '
+
+    list<int> li = [ 1, 2, 3, 4 ];
+    // CHECK: data['VarPrim']['li'] == [ 1, 2, 3, 4 ]
+}
+def VarObj : Variables {
+    // Test initializers that map to JSON objects containing a 'kind'
+    // discriminator
+
+    Base base = D;
+    // CHECK: data['VarObj']['base']['kind'] == 'def'
+    // CHECK: data['VarObj']['base']['def'] == 'D'
+    // CHECK: data['VarObj']['base']['printable'] == 'D'
+
+    dag d = (ExampleDagOp 22, "hello":$foo);
+    // CHECK: data['VarObj']['d']['kind'] == 'dag'
+    // CHECK: data['VarObj']['d']['operator']['kind'] == 'def'
+    // CHECK: data['VarObj']['d']['operator']['def'] == 'ExampleDagOp'
+    // CHECK: data['VarObj']['d']['operator']['printable'] == 'ExampleDagOp'
+    // CHECK: data['VarObj']['d']['args'] == [22, "hello"]
+    // CHECK: data['VarObj']['d']['argnames'] == [None, 'foo']
+    // CHECK: data['VarObj']['d']['printable'] == '(ExampleDagOp 22, "hello":$foo)'
+
+    int undef_int;
+    field int ref_int = undef_int;
+    // CHECK: data['VarObj']['ref_int']['kind'] == 'var'
+    // CHECK: data['VarObj']['ref_int']['var'] == 'undef_int'
+    // CHECK: data['VarObj']['ref_int']['printable'] == 'undef_int'
+
+    bits<2> undef_bits;
+    bits<4> ref_bits;
+    let ref_bits{3-2} = 0b10;
+    let ref_bits{1-0} = undef_bits{1-0};
+    // CHECK: data['VarObj']['ref_bits'][3] == 1
+    // CHECK: data['VarObj']['ref_bits'][2] == 0
+    // CHECK: data['VarObj']['ref_bits'][1]['kind'] == 'varbit'
+    // CHECK: data['VarObj']['ref_bits'][1]['variable'] == 'undef_bits'
+    // CHECK: data['VarObj']['ref_bits'][1]['index'] == 1
+    // CHECK: data['VarObj']['ref_bits'][1]['printable'] == 'undef_bits{1}'
+    // CHECK: data['VarObj']['ref_bits'][0]['kind'] == 'varbit'
+    // CHECK: data['VarObj']['ref_bits'][0]['variable'] == 'undef_bits'
+    // CHECK: data['VarObj']['ref_bits'][0]['index'] == 0
+    // CHECK: data['VarObj']['ref_bits'][0]['printable'] == 'undef_bits{0}'
+
+    field int complex_ref_int = !add(undef_int, 2);
+    // CHECK: data['VarObj']['complex_ref_int']['kind'] == 'complex'
+    // CHECK: data['VarObj']['complex_ref_int']['printable'] == '!add(undef_int, 2)'
+}

utils/TableGen/TableGen.cpp

 #include "llvm/TableGen/Main.h"
 #include "llvm/TableGen/Record.h"
 #include "llvm/TableGen/SetTheory.h"
 
 using namespace llvm;
 
 enum ActionType {
   PrintRecords,
+  DumpJSON,
   GenEmitter,
   GenRegisterInfo,
   GenInstrInfo,
   GenInstrDocs,
   GenAsmWriter,
   GenAsmMatcher,
   GenDisassembler,
   GenPseudoLowering,
enum ActionType {
   GenRegisterBank,
 };
 
 namespace {
   cl::opt<ActionType>
   Action(cl::desc("Action to perform:"),
          cl::values(clEnumValN(PrintRecords, "print-records",
                                "Print all records to stdout (default)"),
+                    clEnumValN(DumpJSON, "dump-json",
+                               "Dump all records as machine-readable JSON"),
                     clEnumValN(GenEmitter, "gen-emitter",
                                "Generate machine code emitter"),
                     clEnumValN(GenRegisterInfo, "gen-register-info",
                                "Generate registers and register classes info"),
                     clEnumValN(GenInstrInfo, "gen-instr-info",
                                "Generate instruction descriptions"),
                     clEnumValN(GenInstrDocs, "gen-instr-docs",
                                "Generate instruction documentation"),
namespace {
   Class("class", cl::desc("Print Enum list for this class"),
         cl::value_desc("class name"), cl::cat(PrintEnumsCat));
 
 bool LLVMTableGenMain(raw_ostream &OS, RecordKeeper &Records) {
   switch (Action) {
   case PrintRecords:
     OS << Records;           // No argument, dump all contents
     break;
+  case DumpJSON:
+    EmitJSON(Records, OS);
+    break;
   case GenEmitter:
     EmitCodeEmitter(Records, OS);
     break;
   case GenRegisterInfo:
     EmitRegisterInfo(Records, OS);
     break;
   case GenInstrInfo:
     EmitInstrInfo(Records, OS);