[mlir] Added oilist primitive

[shraiysh]

Summary

This patch attempts to add the oilist primitive proposed in the RFC: Extending Declarative Assembly Format to support order-independent variadic segments.

This element supports optional order-independent variadic segments for operations. This will allow OpenACC and OpenMP Dialects to have similar and relaxed requirements while encouraging the use of Declarative Assembly Format and avoiding code duplication.

An oilist element parses grammar of the form:

clause-list := clause clause-list | empty
clause := `keyword` clause1 | `otherKeyword` clause2
clause1 := <assembly-format element>
clause2 := <assembly-format element>

AssemblyFormat specification:

let assemblyFormat = [{
  oilist( `keyword` clause1
        | `otherkeyword` clause2
        ...
        )
}];

Example:

oilist( `private` `(` $arg0 `:` type($arg0) `)`
      | `nowait`
      | `reduction` custom<ReductionClause>($arg1, type($arg1)))

oilist( `private` `=` $arg0 `:` type($arg0)
      | `reduction` `=` $arg1 `:` type($arg1)
      | `firstprivate` `=` $arg3 `:` type($arg2))

[shraiysh]

Remove unrelated change

[Mogball]

Nice!

mlir/test/mlir-tblgen/op-format-spec.td

368

Can you add an op to the test dialect that uses this new directive and then add some list tests to test round-tripping the new format?

mlir/tools/mlir-tblgen/FormatGen.h

45

[ and ] are called l_square and r_square in other parts of the codebase.

mlir/tools/mlir-tblgen/OpFormatGen.cpp

71

Because oilist is a keyword, I would consider it more of a "directive" and wouldn't really group it with optional.

370–371

377

Can you encapsulate these members and document them?

857

Can you be more descriptive here?

859–862

Can you also reference oilist somewhere in the error message? Otherwise, it might be a bit vague to someone who isn't aware of the construct.

957–961

1351

Use cast if you're immediately going to dereference.

1352–1353

If you're generating an int for every element whose only purpose is to index into a BitVector, why not just generate a bool flag for each element instead and ditch the bit vector?

1357

You could instead do

::llvm::StringRef oilistKey;
while (succeeded(parser.parseOptionalKeyword(&oilistKey))) {
  if (oilistKey == "literal0") { ... } // or StringSwitch
}

1362

2105

Handling the spaces can be tricky. I think you should delete this line and call genLiteralPrinter below (instead of directly generating the literal print yourself), which has the correct behaviour for printing spaces and updating the flags.

2110

2110–2111

hasAttrOfType?

2113–2118

Let genElementPrinter handle the spaces.

2122–2123

2919–2926

Can you add a comment explaining what this is meant to guard against? (I like the error message btw)

3302–3303

3306–3310

Leftover debugging?

3314

3322–3324

3326–3327

A ] is required to exit the loop, so this branch is unreachable.

[shraiysh]

Thanks for the detailed review @Mogball. I have tried to address all your comments. Please let me know if I missed something!

mlir/tools/mlir-tblgen/OpFormatGen.cpp

1357

I tried doing that earlier, but that was causing issues when this was the last element in the assembly format.

For example,
Operation Definition:

let assemblyFormat = [{ oilist [ `keyword` | `otherKeyword` ] }];

MLIR file:

func @testing() {
  test.op keyword
  return
}

It tends to accept return as a keyword and enters the while loop. Once it is parsed, I do not know how to un-parse it for the next operation to take over. Please let me know if you have an alternative solution for this.

[shraiysh]

mlir/tools/mlir-tblgen/OpFormatGen.cpp

377

I have documented them. I am not too sure if I can encapsulate literalElements as OIListElement is itself encapsulating it, but I guess we can do something about parsingElements. Please let me know if you want me to make a new class ElementListElement for one parsing element - it is a element that is a list of elements (along with its own parsing and printing). I know that the name is a bit weird, looking for better suggestions for the same.

3322–3324

I cannot put it in the while condition, because then this testcase will be accepted with no errors:

oilist [ `something` | ]

This is because after getting a pipe, continue will check the condition and it will break the loop as curToken is now ]. So, now this will get accepted without any error reporting.

[rriddle]

Thanks, will take a look within the next day or two.

mlir/test/lib/Dialect/Test/TestDialect.cpp

531–533

Drop trivial braces.

[Mogball]

mlir/tools/mlir-tblgen/OpFormatGen.cpp

377

I meant to make them private members and provide accessors:

auto getLiteralElements() const {
  return llvm::make_pointee_range(llvm::map_range(literalElements, [](auto &el) { return cast<LiteralElement>(el.get()); }));
}

861

1348–1351

Use a for-each loop (and the provided accessor)

1357

Ah, right, because there's no delimiter to end the list. This is fine as-is then.

1357

You can use llvm::zip to iterate over literalElements and parsingElements at the same time

for (auto it : llvm::zip(oilist->getLiteralElements(), oilist->getParsingElements())) {
  LiteralElement &literalElement = std::get<0>(it);
  auto &parsingElement = std::get<1>(it);
  ...
}

You can move printing else to the end of the loop without checking the index. Then after the loop print

body << " {\n"

3322–3324

Right, makes sense.

[rriddle]

Can you add documentation for this in docs/OpDefinitions.md?

Also, how does this interact with the detection of parse ambiguities? We should still detect ambiguities inside of (and adjacent to) oilist.

mlir/test/lib/Dialect/Test/TestOps.td

613–615

Why surround with brackets instead of using something like ()(like all of the other directives)?

mlir/tools/mlir-tblgen/FormatGen.cpp

177–178

Keep this sorted.

mlir/tools/mlir-tblgen/FormatGen.h

69–70

Keep this sorted.

mlir/tools/mlir-tblgen/OpFormatGen.cpp

53–54

Please keep these sorted.

369

857

typo: coressponding

861–862

Drop the reference to the current operation, leave that to the parser to handle infra to callout.

The message is also a bit confusing, this is calling out a duplicate entry for the clause but the message seems to state that only one clause of the list is valid.

1346–1347

1348

Cache the end iterator unless it changes within the loop

2920–2926

Can you flip the if so that error happens first? We generally prefer early exit.

3297

[shraiysh]

Thanks for the review @Mogball and @rriddle. I have tried to address all your comments.

@rriddle can you please elaborate on the parse ambiguities. Are you talking about something like this?

oilist ( `keyword` | `otherkeyword` ) `keyword` attr-dict

I realise this translates to no valid syntax with the current implementation of the parser. Would you suggest not allowing oilist-literals right after an oilist element? (Looking for ideas on this)

If there are any other ambiguities which should be discussed, we can discuss them before moving ahead.

[Mogball]

LG enough for me once the issues I've commented on below are resolved. You should wait for river to sign off too.

mlir/tools/mlir-tblgen/OpFormatGen.cpp

364

Constructors and public methods should come before private methods and class members

393–395

make_pointee_range will work on any range of types that implement operator*, including unique_ptr. You don't need to map them to get(). Also, I would just inline this map function into your accessor method below.

412–413

You don't need to store this in a std::function

1356

Should this be a reference?

2109

Should this be a reference?

[shraiysh]

Thanks for the review @Mogball. Addressed Comments.

mlir/tools/mlir-tblgen/OpFormatGen.cpp

364

I wanted the vectorConverter method to be private, but because its return type is auto, I was forced to put it before its use in getParsingElements(). So, if I try to put the other public methods above, the structure will be something like this. Please let me know if I should change it to this:

public:
  OIListElement(...) { ... }
  static bool classof(...) { ... }
  auto getLiteralElements() const {...}
private:
  static auto vectorConverter(...) { ... }
  std::vector<...> literalElements;
  std::vector<...> parsingElements;
public:
  auto getParsingElements() const {...}

393–395

Thanks for the unique_ptr comment. Addressed it.

About the inlining, I can make it an std::function there, but the declaration will not be pretty as it returns a very complex templated llvm::iterator_range. (can't use it directly or use auto with lambda, because lambdas cannot be copied and llvm::zip will try to copy lambda).

std::function<llvm::iterator_range<llvm::pointee_iterator<decltype((
    std::begin(std::declval<std::vector<std::unique_ptr<Element>>>())))>>(
    const std::vector<std::unique_ptr<Element>> &vec)>
    vectorConverter =
        [](auto &vec) { return llvm::make_pointee_range(vec); };

Let me know if you prefer this, and I will change it accordingly.

412–413

I cannot use auto as copy constructor for lambda functions is deleted but it is not deleted for std::function. The function is stored as is in a llvm::map_range iterator and it is copied while using llvm::zip. Using auto will make it a lambda and lead to errors.

[rriddle]

In D115215#3190124, @shraiysh wrote:

Thanks for the review @Mogball and @rriddle. I have tried to address all your comments.

@rriddle can you please elaborate on the parse ambiguities. Are you talking about something like this?

oilist ( `keyword` | `otherkeyword` ) `keyword` attr-dict

I realise this translates to no valid syntax with the current implementation of the parser. Would you suggest not allowing oilist-literals right after an oilist element? (Looking for ideas on this)

If there are any other ambiguities which should be discussed, we can discuss them before moving ahead.

One case in particular is https://github.com/llvm/llvm-project/blob/1652871473a70e9908d794e858388edc89cd6e88/mlir/tools/mlir-tblgen/OpFormatGen.cpp#L2435
https://github.com/llvm/llvm-project/blob/1652871473a70e9908d794e858388edc89cd6e88/mlir/test/mlir-tblgen/op-format-spec.td#L463

mlir/docs/OpDefinitions.md

626

I would avoid the use of "string" here.

628

Something non-obvious here is the treatment of non-optional constructs, should those be allowed? What happens in those cases? Should we error out?

mlir/tools/mlir-tblgen/OpFormatGen.cpp

364

That method doesn't look like it pays for itself, I'd just inline it and keep proper public -> private ordering.

367

Documentation should use /// comments.

388–394

This doesn't appear super useful, I'd just inline it wherever this gets called.

407–408

408

412–413

The use of std::function is quite unfortunate, can you add a comment here capturing this?

417–418

1356

Can you spell out auto here?

1360

nit: Make sure to use camelCase for variable names.

1363

nit: Spell out auto here.

[shraiysh]

Thanks for the review @rriddle. I have tried to address all the comments. Please let me know if I have missed something.

[shraiysh]

Ping!

[rriddle]

Thanks for the ping, appreciated!

mlir/tools/mlir-tblgen/OpFormatGen.cpp

2336

I think at head we can drop the llvm:: on these.

2730

nit: Drop the trivial braces here.

3338

Can we drop the ::mlir:: on these? (I'm not sure why they are on the others in this file).

[clementval]

mlir/docs/OpDefinitions.md

627

[shraiysh]

Thanks for accepting this @rriddle, and thanks for the review @clementval. Addressed comments.

@Mogball, I have rebased this patch with main and there were some substantial changes by you in these files. Can you please recheck if everything is where it is supposed to be in this patch. Thank you.

[Mogball]

Overall LGTM.

mlir/tools/mlir-tblgen/FormatGen.h

146

Fill this in? :P

147

Let's move this into OpFormatGen.cpp. There is an expectation that elements defined in this file are available to both Attribute/Type and Operation formats.

152–153

158

166

avoid returning a copy

mlir/tools/mlir-tblgen/OpFormatGen.cpp

958

avoid copying the vector

1354

2109

3361

Unneeded?

3367

Same here

[shraiysh]

Rebase with main, addressed comments. Thanks for the review @Mogball.

mlir/tools/mlir-tblgen/FormatGen.h

158

Sadly, I have to use std::function here because lambdas can't be copied (copy constructor is required in llvm::zip). I have added a comment about this.

mlir/docs/OpDefinitions.md

 
 *   `functional-type` ( inputs , results )
 
     -   Formats the `inputs` and `results` arguments as a
         [function type](Dialects/Builtin.md/#functiontype).
     -   The constraints on `inputs` and `results` are the same as the `input` of
         the `type` directive.
 
+*   `oilist` ( \`keyword\` elements | \`otherKeyword\` elements ...)
+
+    -   Represents an optional order-independent list of clauses. Each clause
+        has a keyword associated with it and a string of other assembly format
+        elements.
+    -   Each clause can appear 0 or 1 times (in any order).
+
 *   `operands`
 
     -   Represents all of the operands of an operation.
 
 *   `ref` ( input )
 
     -   Represents a reference to the a variable or directive, that must have
         already been resolved, that may be used as a parameter to a `custom`

mlir/test/IR/traits.mlir

 func @succeededResultSizeAttr() {
   // CHECK: test.attr_sized_results
   %0:4 = "test.attr_sized_results"() {result_segment_sizes = dense<[0, 2, 1, 1]>: vector<4xi32>} : () -> (i32, i32, i32, i32)
   return
 }
 
 // -----
 
+// CHECK-LABEL: @succeededOilistTrivial
+func @succeededOilistTrivial() {
+  // CHECK: test.oilist_with_keywords_only keyword
+  test.oilist_with_keywords_only keyword
+  // CHECK: test.oilist_with_keywords_only otherKeyword
+  test.oilist_with_keywords_only otherKeyword
+  // CHECK: test.oilist_with_keywords_only keyword otherKeyword
+  test.oilist_with_keywords_only keyword otherKeyword
+  // CHECK: test.oilist_with_keywords_only keyword otherKeyword
+  test.oilist_with_keywords_only otherKeyword keyword
+  return
+}
+
+// -----
+
+// CHECK-LABEL: @succeededOilistSimple
+func @succeededOilistSimple(%arg0 : i32, %arg1 : i32, %arg2 : i32) {
+  // CHECK: test.oilist_with_simple_args keyword %{{.*}} : i32
+  test.oilist_with_simple_args keyword %arg0 : i32
+  // CHECK: test.oilist_with_simple_args otherKeyword %{{.*}} : i32
+  test.oilist_with_simple_args otherKeyword %arg0 : i32
+  // CHECK: test.oilist_with_simple_args thirdKeyword %{{.*}} : i32
+  test.oilist_with_simple_args thirdKeyword %arg0 : i32
+
+  // CHECK: test.oilist_with_simple_args keyword %{{.*}} : i32 otherKeyword %{{.*}} : i32
+  test.oilist_with_simple_args keyword %arg0 : i32 otherKeyword %arg1 : i32
+  // CHECK: test.oilist_with_simple_args keyword %{{.*}} : i32 thirdKeyword %{{.*}} : i32
+  test.oilist_with_simple_args keyword %arg0 : i32 thirdKeyword %arg1 : i32
+  // CHECK: test.oilist_with_simple_args otherKeyword %{{.*}} : i32 thirdKeyword %{{.*}} : i32
+  test.oilist_with_simple_args thirdKeyword %arg0 : i32 otherKeyword %arg1 : i32
+
+  // CHECK: test.oilist_with_simple_args keyword %{{.*}} : i32 otherKeyword %{{.*}} : i32 thirdKeyword %{{.*}} : i32
+  test.oilist_with_simple_args keyword %arg0 : i32 otherKeyword %arg1 : i32 thirdKeyword %arg2 : i32
+  // CHECK: test.oilist_with_simple_args keyword %{{.*}} : i32 otherKeyword %{{.*}} : i32 thirdKeyword %{{.*}} : i32
+  test.oilist_with_simple_args otherKeyword %arg0 : i32 keyword %arg1 : i32 thirdKeyword %arg2 : i32
+  // CHECK: test.oilist_with_simple_args keyword %{{.*}} : i32 otherKeyword %{{.*}} : i32 thirdKeyword %{{.*}} : i32
+  test.oilist_with_simple_args otherKeyword %arg0 : i32 thirdKeyword %arg1 : i32 keyword %arg2 : i32
+  return
+}
+
+// -----
+
+// CHECK-LABEL: @succeededOilistVariadic
+// CHECK-SAME: (%[[ARG0:.*]]: i32, %[[ARG1:.*]]: i32, %[[ARG2:.*]]: i32)
+func @succeededOilistVariadic(%arg0: i32, %arg1: i32, %arg2: i32) {
+  // CHECK: test.oilist_variadic_with_parens keyword(%[[ARG0]], %[[ARG1]] : i32, i32)
+  test.oilist_variadic_with_parens keyword (%arg0, %arg1 : i32, i32)
+  // CHECK: test.oilist_variadic_with_parens keyword(%[[ARG0]], %[[ARG1]] : i32, i32) otherKeyword(%[[ARG2]], %[[ARG1]] : i32, i32)
+  test.oilist_variadic_with_parens otherKeyword (%arg2, %arg1 : i32, i32) keyword (%arg0, %arg1 : i32, i32)
+  // CHECK: test.oilist_variadic_with_parens keyword(%[[ARG0]], %[[ARG1]] : i32, i32) otherKeyword(%[[ARG0]], %[[ARG1]] : i32, i32) thirdKeyword(%[[ARG2]], %[[ARG0]], %[[ARG1]] : i32, i32, i32)
+  test.oilist_variadic_with_parens thirdKeyword (%arg2, %arg0, %arg1 : i32, i32, i32) keyword (%arg0, %arg1 : i32, i32) otherKeyword (%arg0, %arg1 : i32, i32)
+  return
+}
+
+// -----
+// CHECK-LABEL: succeededOilistCustom
+// CHECK-SAME: (%[[ARG0:.*]]: i32, %[[ARG1:.*]]: i32, %[[ARG2:.*]]: i32)
+func @succeededOilistCustom(%arg0: i32, %arg1: i32, %arg2: i32) {
+  // CHECK: test.oilist_custom private(%[[ARG0]], %[[ARG1]] : i32, i32)
+  test.oilist_custom private (%arg0, %arg1 : i32, i32)
+  // CHECK: test.oilist_custom private(%[[ARG0]], %[[ARG1]] : i32, i32) nowait
+  test.oilist_custom private (%arg0, %arg1 : i32, i32) nowait
+  // CHECK: test.oilist_custom private(%arg0, %arg1 : i32, i32) nowait reduction (%arg1)
+  test.oilist_custom nowait reduction (%arg1) private (%arg0, %arg1 : i32, i32)
+  return
+}
+
+// -----
+
 func @failedHasDominanceScopeOutsideDominanceFreeScope() -> () {
   "test.ssacfg_region"() ({
     test.graph_region {
       // expected-error @+1 {{operand #0 does not dominate this use}}
       %2:3 = "bar"(%1) : (i64) -> (i1,i1,i1)
     }
     // expected-note @+1 {{operand defined here}}
     %1 = "baz"() : () -> (i64)

 // -----
 
 // Check that we can query traits in attributes
 func @failed_attr_traits() {
   // expected-error@+1 {{'attr' attribute should have trait 'TestAttrTrait'}}
   "test.attr_with_trait"() {attr = 42 : i32} : () -> ()
   return
 }
 No newline at end of file

mlir/test/lib/Dialect/Test/TestDialect.cpp

 
 //===----------------------------------------------------------------------===//
 // Test Format* operations
 //===----------------------------------------------------------------------===//
 
 //===----------------------------------------------------------------------===//
 // Parsing
 
+static ParseResult
+parseCustomOptionalOperand(OpAsmParser &parser,
+                           Optional<OpAsmParser::OperandType> &optOperand) {
+  if (succeeded(parser.parseOptionalLParen())) {
+    optOperand.emplace();
+    if (parser.parseOperand(*optOperand) || parser.parseRParen())
+      return failure();
+  }
+  return success();
+}
 static ParseResult parseCustomDirectiveOperands(
     OpAsmParser &parser, OpAsmParser::OperandType &operand,
     Optional<OpAsmParser::OperandType> &optOperand,
     SmallVectorImpl<OpAsmParser::OperandType> &varOperands) {
   if (parser.parseOperand(operand))
     return failure();
   if (succeeded(parser.parseOptionalComma())) {
     optOperand.emplace();
if (parser.parseInteger(operandCount))
     return failure();
   bool expectedOptionalOperand = operandCount == 0;
   return success(expectedOptionalOperand != optOperand.hasValue());
 }
 
 //===----------------------------------------------------------------------===//
 // Printing
 
+static void printCustomOptionalOperand(OpAsmPrinter &printer, Operation *,
+                                       Value optOperand) {
+  if (optOperand)
+    printer << "(" << optOperand << ") ";
+}
 static void printCustomDirectiveOperands(OpAsmPrinter &printer, Operation *,
                                          Value operand, Value optOperand,
                                          OperandRange varOperands) {
   printer << operand;
   if (optOperand)
     printer << ", " << optOperand;
   printer << " -> (" << varOperands << ")";
 }

mlir/test/lib/Dialect/Test/TestOps.td

let results = (outs
     Variadic<I32>:$d
   );
 }
 
 // This is used to test that the fallback for a custom op's parser and printer
 // is the dialect parser and printer hooks.
 def CustomFormatFallbackOp : TEST_Op<"dialect_custom_format_fallback">;
 
+// Ops related to OIList primitive
+def OIListTrivial : TEST_Op<"oilist_with_keywords_only"> {
+  let assemblyFormat = [{
+    oilist( `keyword`
+          | `otherKeyword`) attr-dict
+  }];
+}
+
+def OIListSimple : TEST_Op<"oilist_with_simple_args", [AttrSizedOperandSegments]> {
+  let arguments = (ins Optional<AnyType>:$arg0,
+                       Optional<AnyType>:$arg1,
+                       Optional<AnyType>:$arg2);
+  let assemblyFormat = [{
+    oilist( `keyword` $arg0 `:` type($arg0)
+          | `otherKeyword` $arg1 `:` type($arg1)
+          | `thirdKeyword` $arg2 `:` type($arg2) ) attr-dict
+  }];
+}
+
+def OIListVariadic : TEST_Op<"oilist_variadic_with_parens", [AttrSizedOperandSegments]> {
+  let arguments = (ins Variadic<AnyType>:$arg0,
+                       Variadic<AnyType>:$arg1,
+                       Variadic<AnyType>:$arg2);
+  let assemblyFormat = [{
+    oilist( `keyword` `(` $arg0 `:` type($arg0) `)`
+          | `otherKeyword` `(` $arg1 `:` type($arg1) `)`
+          | `thirdKeyword` `(` $arg2 `:` type($arg2) `)`) attr-dict
+  }];
+}
+
+def OIListCustom : TEST_Op<"oilist_custom", [AttrSizedOperandSegments]> {
+  let arguments = (ins Variadic<AnyType>:$arg0,
+                       Optional<I32>:$optOperand,
+                       UnitAttr:$nowait);
+  let assemblyFormat = [{
+    oilist( `private` `(` $arg0 `:` type($arg0) `)`
+          | `nowait`
+          | `reduction` custom<CustomOptionalOperand>($optOperand)
+    ) attr-dict
+  }];
+}
+
 // This is used to test encoding of a string attribute into an SSA name of a
 // pretty printed value name.
 def StringAttrPrettyNameOp
  : TEST_Op<"string_attr_pretty_name",
            [DeclareOpInterfaceMethods<OpAsmOpInterface, ["getAsmResultNames"]>]> {
   let arguments = (ins StrArrayAttr:$names);
   let results = (outs Variadic<I32>:$r);
 

mlir/test/mlir-tblgen/op-format-spec.td

 }]>;
 // CHECK-NOT: error
 def LiteralValid : TestFormat_Op<[{
   `_` `:` `,` `=` `<` `>` `(` `)` `[` `]` `?` `+` `*` ` ` `` `->` `\n` `abc$._`
   attr-dict
 }]>;
 
 //===----------------------------------------------------------------------===//
+// oilist
+
+// CHECK: error: expected literal, but got ')'
+def OIListErrorExpectedLiteral : TestFormat_Op<[{
+  oilist( `keyword` | ) attr-dict
+}]>;
+// CHECK: error: expected literal, but got ')'
+def OIListErrorExpectedEmpty : TestFormat_Op<[{
+  oilist() attr-dict
+}]>;
+// CHECK: error: expected literal, but got '$arg0'
+def OIListErrorNoLiteral : TestFormat_Op<[{
+  oilist( $arg0 `:` type($arg0) | $arg1 `:` type($arg1) ) attr-dict
+}], [AttrSizedOperandSegments]>, Arguments<(ins Optional<AnyType>:$arg0, Optional<AnyType>:$arg1)>;
+// CHECK: error: expected '(' before oilist argument list
+def OIListStartingToken : TestFormat_Op<[{
+  oilist `wrong` attr-dict
+}]>;
+
+// CHECK-NOT: error
+def OIListTrivial : TestFormat_Op<[{
+  oilist(`keyword` `(` `)` | `otherkeyword` `(` `)`) attr-dict
+}]>;
+def OIListSimple : TestFormat_Op<[{
+  oilist( `keyword` $arg0 `:` type($arg0)
+        | `otherkeyword` $arg1 `:` type($arg1)
+        | `thirdkeyword` $arg2 `:` type($arg2) )
+  attr-dict
+}], [AttrSizedOperandSegments]>, Arguments<(ins Optional<AnyType>:$arg0, Optional<AnyType>:$arg1, Optional<AnyType>:$arg2)>;
+def OIListVariadic : TestFormat_Op<[{
+  oilist( `keyword` `(` $args0 `:` type($args0) `)`
+        | `otherkeyword` `(` $args1 `:` type($args1) `)`
+        | `thirdkeyword` `(` $args2 `:` type($args2) `)`)
+  attr-dict
+}], [AttrSizedOperandSegments]>, Arguments<(ins Variadic<AnyType>:$args0, Variadic<AnyType>:$args1, Variadic<AnyType>:$args2)>;
+def OIListCustom : TestFormat_Op<[{
+  oilist( `private` `(` $arg0 `:` type($arg0) `)`
+        | `nowait`
+        | `reduction` custom<ReductionClause>($arg1, type($arg1))) attr-dict
+}], [AttrSizedOperandSegments]>, Arguments<(ins Optional<AnyType>:$arg0, Optional<AnyType>:$arg1)>;
+
+//===----------------------------------------------------------------------===//
 // Optional Groups
 //===----------------------------------------------------------------------===//
 
 // CHECK: error: optional groups can only be used as top-level elements
 def OptionalInvalidA : TestFormat_Op<[{
   type(($attr^)?) attr-dict
 }]>, Arguments<(ins OptionalAttr<I64Attr>:$attr)>;
 // CHECK: error: expected directive, literal, variable, or optional group

mlir/tools/mlir-tblgen/FormatGen.h

public:
   enum Kind {
     // Markers.
     eof,
     error,
 
     // Tokens with no info.
     l_paren,
     r_paren,
     caret,
     colon,
     comma,
     equal,
     less,
     greater,
     question,
     star,
+    pipe,
 
     // Keywords.
     keyword_start,
     kw_attr_dict,
     kw_attr_dict_w_keyword,
     kw_custom,
     kw_functional_type,
+    kw_oilist,
     kw_operands,
     kw_params,
     kw_ref,
     kw_regions,
     kw_results,
     kw_struct,
     kw_successors,
     kw_type,
     keyword_end,
 
     // String valued tokens.
     identifier,
     literal,
     variable,
   };
 
   FormatToken(Kind kind, StringRef spelling) : kind(kind), spelling(spelling) {}
private:
   /// Location of the format string.
   llvm::SMLoc loc;
   /// Buffer containing the format string.
   StringRef curBuffer;
   /// Current pointer in the buffer.
   const char *curPtr;
 };
 
 /// Whether a space needs to be emitted before a literal. E.g., two keywords
 /// back-to-back require a space separator, but a keyword followed by '<' does
 /// not require a space.
 bool shouldEmitSpaceBefore(StringRef value, bool lastWasPunctuation);
 
 /// Returns true if the given string can be formatted as a keyword.
 bool canFormatStringAsKeyword(StringRef value,
                               function_ref<void(Twine)> emitError = nullptr);
 
 /// Returns true if the given string is valid format literal element.
 /// If `emitError` is provided, it is invoked with the reason for the failure.
 bool isValidLiteral(StringRef value,
                     function_ref<void(Twine)> emitError = nullptr);
 
 /// Whether a failure in parsing the assembly format should be a fatal error.
 extern llvm::cl::opt<bool> formatErrorIsFatal;
 
 } // namespace tblgen
 } // namespace mlir
 
 #endif // MLIR_TOOLS_MLIRTBLGEN_FORMATGEN_H_

mlir/tools/mlir-tblgen/FormatGen.cpp

FormatToken FormatLexer::lexToken() {
   case '?':
     return formToken(FormatToken::question, tokStart);
   case '(':
     return formToken(FormatToken::l_paren, tokStart);
   case ')':
     return formToken(FormatToken::r_paren, tokStart);
   case '*':
     return formToken(FormatToken::star, tokStart);
+  case '|':
+    return formToken(FormatToken::pipe, tokStart);
 
   // Ignore whitespace characters.
   case 0:
   case ' ':
   case '\t':
   case '\n':
     return lexToken();
 
FormatToken FormatLexer::lexIdentifier(const char *tokStart) {
   // Check to see if this identifier is a keyword.
   StringRef str(tokStart, curPtr - tokStart);
   auto kind =
       StringSwitch<FormatToken::Kind>(str)
           .Case("attr-dict", FormatToken::kw_attr_dict)
           .Case("attr-dict-with-keyword", FormatToken::kw_attr_dict_w_keyword)
           .Case("custom", FormatToken::kw_custom)
           .Case("functional-type", FormatToken::kw_functional_type)
+          .Case("oilist", FormatToken::kw_oilist)
           .Case("operands", FormatToken::kw_operands)
           .Case("params", FormatToken::kw_params)
           .Case("ref", FormatToken::kw_ref)
           .Case("regions", FormatToken::kw_regions)
           .Case("results", FormatToken::kw_results)
           .Case("struct", FormatToken::kw_struct)
           .Case("successors", FormatToken::kw_successors)
           .Case("type", FormatToken::kw_type)
           .Default(FormatToken::identifier);
   return FormatToken(kind, str);
 }
 
 //===----------------------------------------------------------------------===//
 // Utility Functions
 //===----------------------------------------------------------------------===//
 
 bool mlir::tblgen::shouldEmitSpaceBefore(StringRef value,

mlir/tools/mlir-tblgen/OpFormatGen.cpp

 /// This class represents a single format element.
 class Element {
 public:
   enum class Kind {
     /// This element is a directive.
     AttrDictDirective,
     CustomDirective,
     FunctionalTypeDirective,
+    OIListDirective,
     OperandsDirective,
     RefDirective,
     RegionsDirective,
     ResultsDirective,
     SuccessorsDirective,
     TypeDirective,
 
     /// This element is a literal.
     Literal,
 
     /// This element is a whitespace.
     Newline,
     Space,
 
     /// This element is an variable value.
     AttributeVariable,
     OperandVariable,
     RegionVariable,
     ResultVariable,
     SuccessorVariable,
 
     /// This element is an optional element.
     Optional,
   };
   Element(Kind kind) : kind(kind) {}
   virtual ~Element() = default;
 
   /// Return the kind of this element.
   Kind getKind() const { return kind; }
 
 private:
   /// The kind of this element.
private:
   unsigned anchor;
   /// The index of the first element that is parsed (is not a
   /// WhitespaceElement).
   unsigned parseStart;
 };
 } // namespace
 
 //===----------------------------------------------------------------------===//
+// OIListElement
+
+namespace {
+class OIListElement : public Element {
+
+private:
+  // A vector of `LiteralElement` objects. Each element stores *the keyword* for
+  // one case of oilist element. For example, an oilist element along with the
+  // literalElements vector:
+  // ```
+  //  oilist [ `keyword` `=` `(` $arg0 `)` | `otherKeyword` `<` $arg1 `>`]
+  //  literalElements = { `keyword`, `otherKeyword` }
+  // ```
+  std::vector<std::unique_ptr<Element>> literalElements;
+
+  // A vector of valid declarative assembly format vectors. Each object in
+  // parsing elements is a vector of elements in assembly format syntax.
+  // For example, an oilist element along with the parsingElements vector:
+  // ```
+  //  oilist [ `keyword` `=` `(` $arg0 `)` | `otherKeyword` `<` $arg1 `>`]
+  //  parsingElements = {
+  //    { `=`, `(`, $arg0, `)` },
+  //    { `<`, $arg1, `>` }
+  //  }
+  // ```
+  std::vector<std::vector<std::unique_ptr<Element>>> parsingElements;
+
+  // A helper function to take a "vector of unique pointers to Element" and
+  // return an iterator for a "vector to Elements". This function is used in
+  // `getParsingElements`.
+  static auto
+  vectorConverter(const std::vector<std::unique_ptr<Element>> &vec) {
+    return llvm::make_pointee_range(vec);
+  }
+
+public:
+  OIListElement(
+      std::vector<std::unique_ptr<Element>> &&literalElements,
+      std::vector<std::vector<std::unique_ptr<Element>>> &&parsingElements)
+      : Element(Kind::OIListDirective),
+        literalElements(std::move(literalElements)),
+        parsingElements(std::move(parsingElements)) {}
+
+  static bool classof(const Element *element) {
+    return element->getKind() == Kind::OIListDirective;
+  }
+
+  // Returns an llvm::iterator_range to iterate over the LiteralElements.
+  auto getLiteralElements() const {
+    std::function<LiteralElement *(const std::unique_ptr<Element> &)>
+        literalElementCastConverter =
+            [](auto &el) { return cast<LiteralElement>(el.get()); };
+    return llvm::make_pointee_range(
+        llvm::map_range(literalElements, literalElementCastConverter));
+  }
+
+  // Returns an llvm::iterator_range to iterate over the parsing elements
+  // coressponding to the clauses.
+  auto getParsingElements() const {
+    return llvm::map_range(parsingElements, OIListElement::vectorConverter);
+  }
+};
+} // end anonymous namespace
+
+//===----------------------------------------------------------------------===//
 // OperationFormat
 //===----------------------------------------------------------------------===//
 
 namespace {
 
 using ConstArgument =
     llvm::PointerUnion<const NamedAttribute *, const NamedTypeConstraint *>;
 

 /// The code snippet used to generate a parser call for a successor.
 ///
 /// {0}: The name of the successor.
 const char *successorParserCode = R"(
   if (parser.parseSuccessor({0}Successor))
     return ::mlir::failure();
 )";
 
+/// The code snippet used to generate a parser for OIList
+///
+/// {0}: literal keyword corresponding to a case for oilist
+const char *oilistParserCode = R"(
+  if ({0}Clause) {
+    return parser.emitError(parser.getNameLoc())
+          << "`{0}` clause can appear at most once in the expansion of the "
+             "oilist directive";
+  }
+  {0}Clause = true;
+  result.addAttribute("{0}", UnitAttr::get(parser.getContext()));
+)";
+
 namespace {
 /// The type of length for a given parse argument.
 enum class ArgumentLengthKind {
   /// The argument is a variadic of a variadic, and may contain 0->N range
   /// elements.
   VariadicOfVariadic,
   /// The argument is variadic, and may contain 0->N elements.
   Variadic,
if (auto *optional = dyn_cast<OptionalElement>(element)) {
     if (anchor && anchor != &*elements.begin() && anchor->isUnitAttr())
       elidedAnchorElement = anchor;
     for (auto &childElement : elements)
       if (&childElement != elidedAnchorElement)
         genElementParserStorage(&childElement, op, body);
     for (auto &childElement : optional->getElseElements())
       genElementParserStorage(&childElement, op, body);
 
+  } else if (auto *oilist = dyn_cast<OIListElement>(element)) {
+    for (auto pelement : oilist->getParsingElements())
+      for (auto &element : pelement)
+        genElementParserStorage(&element, op, body);
+
   } else if (auto *custom = dyn_cast<CustomDirective>(element)) {
     for (auto &paramElement : custom->getArguments())
       genElementParserStorage(&paramElement, op, body);
 
   } else if (isa<OperandsDirective>(element)) {
     body << "  ::mlir::SmallVector<::mlir::OpAsmParser::OperandType, 4> "
             "allOperands;\n";
 
if (auto *optional = dyn_cast<OptionalElement>(element)) {
     if (!elseElements.empty()) {
       body << " else {\n";
       for (Element &childElement : elseElements)
         genElementParser(&childElement, body, attrTypeCtx);
       body << "  }";
     }
     body << "\n";
 
+    // OIList Directive
+  } else if (OIListElement *oilist = dyn_cast<OIListElement>(element)) {
+    for (LiteralElement &le : oilist->getLiteralElements()) {
+      body << "  bool " << le.getLiteral() << "Clause = false;\n";
+    }
+
+    // Generate the parsing loop
+    body << "  while(true) {\n";
+    for (auto it : llvm::zip(oilist->getLiteralElements(),
+                             oilist->getParsingElements())) {
+      LiteralElement &lelement = std::get<0>(it);
+      auto &pelement = std::get<1>(it);
+      body << "if (succeeded(parser.parseOptional";
+      genLiteralParser(lelement.getLiteral(), body);
+      body << ")) {\n";
+      StringRef attrname = lelement.getLiteral();
+      body << formatv(oilistParserCode, attrname);
+      inferredAttributes.insert(attrname);
+      for (auto &el : pelement)
+        genElementParser(&el, body, attrTypeCtx);
+      body << "    } else ";
+    }
+    body << " {\n";
+    body << "    break;\n";
+    body << "  }\n";
+    body << "}\n";
+
     /// Literals.
   } else if (LiteralElement *literal = dyn_cast<LiteralElement>(element)) {
     body << "  if (parser.parse";
     genLiteralParser(literal->getLiteral(), body);
     body << ")\n    return ::mlir::failure();\n";
 
     /// Whitespaces.
   } else if (isa<WhitespaceElement>(element)) {
if (!elseElements.empty()) {
       }
       body << "  }";
     }
 
     body << "\n";
     return;
   }
 
+  if (auto *oilist = dyn_cast<OIListElement>(element)) {
+    genLiteralPrinter(" ", body, shouldEmitSpace, lastWasPunctuation);
+    for (auto it : llvm::zip(oilist->getLiteralElements(),
+                             oilist->getParsingElements())) {
+      LiteralElement &lelement = std::get<0>(it);
+      auto &pelement = std::get<1>(it);
+
+      body << "  if ((*this)->hasAttrOfType<UnitAttr>(\""
+           << lelement.getLiteral() << "\")) {\n";
+      genLiteralPrinter(lelement.getLiteral(), body, shouldEmitSpace,
+                        lastWasPunctuation);
+      for (auto &element : pelement) {
+        genElementPrinter(&element, body, op, shouldEmitSpace,
+                          lastWasPunctuation);
+      }
+      body << "  }\n";
+    }
+    return;
+  }
+
   // Emit the attribute dictionary.
   if (auto *attrDict = dyn_cast<AttrDictDirective>(element)) {
     genAttrDictPrinter(*this, op, body, attrDict->isWithKeyword());
     lastWasPunctuation = false;
     return;
   }
 
   // Optionally insert a space before the next element. The AttrDict printer
private:
   /// Verify the state of operation results within the format.
   LogicalResult
   verifyResults(llvm::SMLoc loc,
                 llvm::StringMap<TypeResolutionInstance> &variableTyResolver);
 
   /// Verify the state of operation successors within the format.
   LogicalResult verifySuccessors(llvm::SMLoc loc);
 
   /// Given the values of an `AllTypesMatch` trait, check for inferable type
   /// resolution.
   void handleAllTypesMatchConstraint(
       ArrayRef<StringRef> values,
       llvm::StringMap<TypeResolutionInstance> &variableTyResolver);
   /// Check for inferable type resolution given all operands, and or results,
   /// have the same type. If 'includeResults' is true, the results also have the
   /// same type as all of the operands.
   void handleSameTypesConstraint(
LogicalResult parseRegionsDirective(std::unique_ptr<Element> &element,
                                       llvm::SMLoc loc, ParserContext context);
   LogicalResult parseResultsDirective(std::unique_ptr<Element> &element,
                                       llvm::SMLoc loc, ParserContext context);
   LogicalResult parseSuccessorsDirective(std::unique_ptr<Element> &element,
                                          llvm::SMLoc loc,
                                          ParserContext context);
   LogicalResult parseTypeDirective(std::unique_ptr<Element> &element,
                                    FormatToken tok, ParserContext context);
+  LogicalResult parseOIListDirective(std::unique_ptr<Element> &element,
+                                     FormatToken tok, ParserContext context);
   LogicalResult parseTypeDirectiveOperand(std::unique_ptr<Element> &element,
                                           bool isRefChild = false);
 
   //===--------------------------------------------------------------------===//
   // Lexer Utilities
   //===--------------------------------------------------------------------===//
 
   /// Advance the current lexer onto the next token.
for (unsigned i = 0, e = values.size(); i != e; ++i) {
     // Mark this value as the type resolver for the other variables.
     for (unsigned j = 0; j != i; ++j)
       variableTyResolver[values[j]] = {arg, llvm::None};
     for (unsigned j = i + 1; j != e; ++j)
       variableTyResolver[values[j]] = {arg, llvm::None};
   }
 }
 
 void FormatParser::handleSameTypesConstraint(
     llvm::StringMap<TypeResolutionInstance> &variableTyResolver,
     bool includeResults) {
   const NamedTypeConstraint *resolver = nullptr;
   int resolvedIt = -1;
 
   // Check to see if there is an operand or result to use for the resolution.
   if ((resolvedIt = seenOperandTypes.find_first()) != -1)
     resolver = &op.getOperand(resolvedIt);
LogicalResult FormatParser::parseDirective(std::unique_ptr<Element> &element,
   case FormatToken::kw_results:
     return parseResultsDirective(element, dirTok.getLoc(), context);
   case FormatToken::kw_successors:
     return parseSuccessorsDirective(element, dirTok.getLoc(), context);
   case FormatToken::kw_ref:
     return parseReferenceDirective(element, dirTok.getLoc(), context);
   case FormatToken::kw_type:
     return parseTypeDirective(element, dirTok, context);
+  case FormatToken::kw_oilist:
+    return parseOIListDirective(element, dirTok, context);
 
   default:
     llvm_unreachable("unknown directive token");
   }
 }
 
 LogicalResult FormatParser::parseLiteral(std::unique_ptr<Element> &element,
                                          ParserContext context) {
   FormatToken literalTok = curToken;
   if (context != TopLevelContext) {
     return emitError(
         literalTok.getLoc(),
         "literals may only be used in a top-level section of the format");
   }
   consumeToken();
 
-  StringRef value = literalTok.getSpelling().drop_front().drop_back();
+  StringRef value = literalTok.getSpelling();
+  // Prevents things like `$arg0` or empty literals (when a literal is expected
+  // but not found) from getting segmentation faults.
+  if (value.size() < 2 || value[0] != '`' || value[value.size() - 1] != '`')
+    return emitError(literalTok.getLoc(),
+                     "expected literal, but got '" + value + "'");
+  value = value.drop_front().drop_back();
 
   // The parsed literal is a space element (`` or ` `).
   if (value.empty() || (value.size() == 1 && value.front() == ' ')) {
     element = std::make_unique<SpaceElement>(!value.empty());
     return ::mlir::success();
   }
   // The parsed literal is a newline element.
   if (value == "\\n") {
     element = std::make_unique<NewlineElement>();
if (hasAllSuccessors || !seenSuccessors.empty())
       return emitError(loc, "'successors' directive creates overlap in format");
     hasAllSuccessors = true;
   }
   element = std::make_unique<SuccessorsDirective>();
   return ::mlir::success();
 }
 
 LogicalResult
+FormatParser::parseOIListDirective(std::unique_ptr<Element> &element,
+                                   FormatToken tok, ParserContext context) {
+  if (failed(parseToken(FormatToken::l_paren,
+                        "expected '(' before oilist argument list")))
+    return failure();
+  std::vector<std::unique_ptr<Element>> literalElements;
+  std::vector<std::vector<std::unique_ptr<Element>>> parsingElements;
+  do {
+    literalElements.emplace_back();
+    parsingElements.emplace_back();
+    if (failed(parseLiteral(literalElements.back(), context)))
+      return failure();
+    auto &currParsingElements = parsingElements.back();
+    while (curToken.getKind() != FormatToken::pipe &&
+           curToken.getKind() != FormatToken::r_paren) {
+      currParsingElements.emplace_back();
+      if (failed(parseElement(currParsingElements.back(), context)))
+        return failure();
+    }
+    if (curToken.getKind() == FormatToken::pipe) {
+      consumeToken();
+      continue;
+    }
+    if (curToken.getKind() == FormatToken::r_paren) {
+      consumeToken();
+      break;
+    }
+  } while (true);
+
+  element = std::make_unique<OIListElement>(std::move(literalElements),
+                                            std::move(parsingElements));
+  return success();
+}
+
+LogicalResult
 FormatParser::parseTypeDirective(std::unique_ptr<Element> &element,
                                  FormatToken tok, ParserContext context) {
   llvm::SMLoc loc = tok.getLoc();
   if (context == TypeDirectiveContext)
     return emitError(loc, "'type' cannot be used as a child of another `type`");
 
   bool isRefChild = context == RefDirectiveContext;
   std::unique_ptr<Element> operand;
   if (failed(parseToken(FormatToken::l_paren,
                         "expected '(' before argument list")) ||
       failed(parseTypeDirectiveOperand(operand, isRefChild)) ||
       failed(
           parseToken(FormatToken::r_paren, "expected ')' after argument list")))
     return ::mlir::failure();
 
   element = std::make_unique<TypeDirective>(std::move(operand));
   return ::mlir::success();
 }
 
 LogicalResult
 FormatParser::parseTypeDirectiveOperand(std::unique_ptr<Element> &element,
                                         bool isRefChild) {
   llvm::SMLoc loc = curToken.getLoc();
   if (failed(parseElement(element, TypeDirectiveContext)))
     return ::mlir::failure();
   if (isa<LiteralElement>(element.get()))
     return emitError(
         loc, "'type' directive operand expects variable or directive operand");
 
   if (auto *var = dyn_cast<OperandVariable>(element.get())) {
     unsigned opIdx = var->getVar() - op.operand_begin();
     if (!isRefChild && (fmt.allOperandTypes || seenOperandTypes.test(opIdx)))
       return emitError(loc, "'type' of '" + var->getVar()->name +
                                 "' is already bound");
     if (isRefChild && !(fmt.allOperandTypes || seenOperandTypes.test(opIdx)))
       return emitError(loc, "'ref' of 'type($" + var->getVar()->name +
                                 ")' is not bound by a prior 'type' directive");
     seenOperandTypes.set(opIdx);
   } else if (auto *var = dyn_cast<ResultVariable>(element.get())) {
     unsigned resIdx = var->getVar() - op.result_begin();
     if (!isRefChild && (fmt.allResultTypes || seenResultTypes.test(resIdx)))
       return emitError(loc, "'type' of '" + var->getVar()->name +
                                 "' is already bound");
     if (isRefChild && !(fmt.allResultTypes || seenResultTypes.test(resIdx)))
       return emitError(loc, "'ref' of 'type($" + var->getVar()->name +
                                 ")' is not bound by a prior 'type' directive");
     seenResultTypes.set(resIdx);