[mlir][sparse] Factoring out Transforms/CodegenUtils.{cpp,h}

[wrengr]

Summary

This moves a bunch of helper functions from Transforms/SparseTensorConversion.cpp into Transforms/CodegenUtils.{cpp,h} so that they can be reused by Transforms/Sparsification.cpp, etc.

See also the dependent D115010 which cleans up some corner cases in this change.

[wrengr]

Updating the two workaround comments with the differential ID that removes them

[aartbik]

mlir/include/mlir/Dialect/SparseTensor/Utils/CodegenUtils.h

22–24 ↗

(On Diff #391510)

This is a good explanation, but sort of a very well known technique.
I think a simple

// Forward references.
....

suffices ;-)

mlir/lib/Dialect/SparseTensor/Utils/CodegenUtils.cpp

16–18

// Forward.

utils/bazel/llvm-project-overlay/mlir/BUILD.bazel

1810–1815

Is this autogen? Shouldn't we just add the right deps, instead of adding outside headers to this lib?

[wrengr]

addressing nits

[wrengr]

utils/bazel/llvm-project-overlay/mlir/BUILD.bazel

1810–1815

I tried using build_cleaner (which made the changes to the deps below), but it couldn't figure out how to fulfill these headers without pulling in a huge plethora of other stuff. I'll try taking another whack at it

[wrengr]

cleaning up BUILD.bazel

utils/bazel/llvm-project-overlay/mlir/BUILD.bazel

1810–1815

Okay, I managed to get a decently clean version. The new version still has "include/mlir/ExecutionEngine/SparseTensorUtils.h" in the hdrs (which should be fine(?) since that's what :SparseTensorTransforms does as well). Alternatively we could add ":mlir_c_runner_utils" to the deps (which will also bring that header in, though at the cost of building all of :mlir_c_runner_utils).

[aartbik]

utils/bazel/llvm-project-overlay/mlir/BUILD.bazel

1810–1815

Yeah, this look okay to me, since now we only include what we "own" in a sense.
Let's see if others chime in, but I am okay with this.

[rriddle]

mlir/lib/Dialect/SparseTensor/Utils/CodegenUtils.cpp

14–18

namespaces should only really have classes, please remove these and use static methods and full namespace resolution instead.

[rriddle]

mlir/lib/Dialect/SparseTensor/Utils/CodegenUtils.cpp

14–18

Some relevant docs:

https://llvm.org/docs/CodingStandards.html#use-namespace-qualifiers-to-implement-previously-declared-functions

https://llvm.org/docs/CodingStandards.html#anonymous-namespaces

[wrengr]

adjusting namespaces

[rriddle]

mlir/include/mlir/Dialect/SparseTensor/Utils/CodegenUtils.h

22–24 ↗

(On Diff #391510)

I wouldn't even add a comment, we don't do that anywhere else.

1–11 ↗

(On Diff #392508)

It isn't clear to me that we should be exposing this kind of API in the external API, can this just be local to the lib/ directory?

96–113 ↗

(On Diff #392508)

I'm not really sure about this kind of API honestly, how much is it really saving? This comment isn't entirely blocking assuming this is local to lib/, but just a general comment.

100–103 ↗

(On Diff #392508)

This looks unused.

[wrengr]

addressing nit

mlir/include/mlir/Dialect/SparseTensor/Utils/CodegenUtils.h

1–11 ↗

(On Diff #392508)

I have no idea where the external API is specified. This code is intended to be internal to the sparse_tensor dialect, but for general reuse by that dialect (i.e., across both lib/Dialect/SparseTensor/Utils/*.cpp and lib/Dialect/SparseTensor/Transforms/*.cpp)

96–113 ↗

(On Diff #392508)

It saves quite a lot actually. Saving 16–20 characters (46–54%) every single time we want to emit a constant adds up very quickly. Not to mention ensuring uniformity re which of the many different constant ops gets used. Let alone if we want to change that choice (as we did just recently). Not to mention the costs of onboarding new developers. It may have been some time since you saw MLIR with fresh eyes, but I assure you that for any newcomer to the codebase the constant drag from wading through the wall of text without these macros greatly impedes their velocity. And mere velocity is not the issue, but rather the fact that having such low initial velocity generally causes new developers to lose heart and decide the project isn't for them. If you want to discuss this further, we can set up a meeting to do so.

100–103 ↗

(On Diff #392508)

At present, yes; though I have no reason to believe it will remain so, and it strikes me as unnecessarily baroque to punch a hole in the regularity of the API just to eliminate a one-liner.

[rriddle]

mlir/include/mlir/Dialect/SparseTensor/Utils/CodegenUtils.h

1–11 ↗

(On Diff #392508)

Anything in include/mlir is considered an "external" API. This should be moved to the lib/ folder.

96–113 ↗

(On Diff #392508)

It may have been some time since you saw MLIR with fresh eyes,

I'll start off by saying that I really don't appreciate the tone of this message

Saving 16–20 characters (46–54%) every single time we want to emit a constant adds up very quickly.

Some of these functions are unused, and some are used only a handful of times. If you want to save code, writing local utilities like this can be fine but I am somewhat strongly opposed to exposing stuff like in any kind of external API.

100–103 ↗

(On Diff #392508)

I don't see a reason to add code that is untested, seems better to add it when it is actually necessary.

[mehdi_amini]

mlir/include/mlir/Dialect/SparseTensor/Utils/CodegenUtils.h

96–113 ↗

(On Diff #392508)

It saves quite a lot actually. Saving 16–20 characters (46–54%) every single time we want to emit a constant adds up very quickly. [...]

Thanks for noticing this.

The usual process in MLIR is to raise this on Discourse, write an RFC, and gain consensus.
We should definitely look into this kind of areas of improvement, and each newcomer to the project has indeed a "fresh pair of eyes" that is unique and very valuable! (after a few months, most people get "used to things" even when they are suboptimal).

That said the things we've been consistent about is to keep all the common infrastructure shared and unified in the project: if you have a good case for helpers method to build various constant, it's unlikely that these have anything specific to the SparseTensor domain and so they really don't belong here.

[aartbik]

mlir/include/mlir/Dialect/SparseTensor/Utils/CodegenUtils.h

96–113 ↗

(On Diff #392508)

It may have been some time since you saw MLIR with fresh eyes,

I'll start off by saying that I really don't appreciate the tone of this message

Wren, please don't say stuff like this. River is our top contributor with a great passion for keeping our source and organization clean and consistent.

River, I probably set the wrong example here by using include/mlir/Dialect/SparseTensor/Utils/ as the place to define the headers of utilities that are used in lib/mlir/Dialect/SparseTensor/* files, with the current (and only) example of Merger.h. I got that inspiration from several other dialects, e.g. include/mlir/Dialect/Linalg/Utils/Utils.h. I was not aware that defining headers in the lib directory is preferred in such places. Just for my understanding, do you think Merger.h was the right decision (since it it more elaborate and has its own testing) and is the objection purely based on having very tiny utils in the header, or do you prefer we move in that direction for everything?

[wrengr]

rebase (over D115004 and D115005)

[rriddle]

mlir/include/mlir/Dialect/SparseTensor/Utils/CodegenUtils.h

96–113 ↗

(On Diff #392508)

The distinction that has generally been drawn between include/ vs lib/ is what is considered an "implementation detail" of a library, and what is meant to be exposed as a general utility to external users of this library. If we just want to have a place to share code between different passes within the same library then lib/ is generally a better choice, given that the stipulations and expectations of internal API are much more relaxed than external API.

Merger could have also been defined in lib/ if it's not meant to be used by external clients, not sure what the expectation there is. Some of the GPU conversion utils are like this: https://github.com/llvm/llvm-project/tree/main/mlir/lib/Conversion/GPUCommon

I will note that you can still write targeted tests for things only defined in lib/, the test directory can use relative include for these types of things: https://github.com/llvm/llvm-project/blob/f638c4d6e4a28fbd5508f853a8715cc92bb66d48/mlir/unittests/Conversion/PDLToPDLInterp/RootOrderingTest.cpp#L9

[aartbik]

mlir/include/mlir/Dialect/SparseTensor/Utils/CodegenUtils.h

96–113 ↗

(On Diff #392508)

Merger could have also been defined in lib/ if it's not meant to be used by external clients, not sure what the expectation there is.

Thanks River for the background. That makes me feel better, since the expectation indeed was that other clients could start using this (since lattice operations are central to the TACO way of generating code), for example, when external contributor start writing improvements on sparse codegen, but want to use the same lattice theory.

Wren, I think we should do the following next
(1) take a look at our utils, see which one could be beneficial for all (a) and which one are specific to sparse
(2a) move (a) to a general place, with River's blessing
(2b) move (b) to a header inside SparseDialects lib

[rriddle]

mlir/include/mlir/Dialect/SparseTensor/Utils/CodegenUtils.h

96–113 ↗

(On Diff #392508)

Right, exactly what Mehdi says here. The "external API" of MLIR (and to a lesser extent the "internal API") need to be consistent. As newcomers engage with the project there can be a tendency to claim "ahah, this is definitely the way it should be because it's easy for me". In many cases this can help the project grow, but it has to be something agreed upon by the community lest we develop pockets of consistency (as different parts of the project develop differently). For this specific API, we've had discussions in the past that never really reached maturity (https://llvm.discourse.group/t/evolving-builder-apis-based-on-lessons-learned-from-edsc/879), and maybe there can be some renewed interest in starting a discussion there. Until then though, we need to ensure the API of the project is consistent.

[wrengr]

mlir/include/mlir/Dialect/SparseTensor/Utils/CodegenUtils.h

96–113 ↗

(On Diff #392508)

Aha! I'd thought the include/ vs lib/ split was just for *.h vs *.cpp file types. I'll move this over to lib/ then, since the intention wasn't to make it public API. Sorry for the confusion

96–113 ↗

(On Diff #392508)

Yes of course, consensus is crucial for any major style/API changes. Again, my intent with this differential was never to change anything outside of the SparseTensor codebase, hence why I did not expect such pushback against this internal cleanup.

Both getOneAttr and genIsNonzero are of general utility, though the latter needs renaming (afaik the "gen" prefix is our own thing rather than mlir standard). I can write up RFCs for both of these, but I'd rather not have the current differential (or rather D115010, D115012) depend on those being resolved.

The constant generators are also of general utility, but they're far trickier to upstream (especially to do so without losing their brevity). There are several different versions of constant ops for different dialects, so we can't just make them into OpBuilder methods. I suppose we could make an ArithOpBuilder subclass, but since that's not an established pattern it strikes me as a terrible precedent. So there's considerable design work to be done just to determine what it would even mean to make these public, let alone reaching consensus. It would be great to do that work, but that's a considerably larger undertaking than what this differential aims to accomplish.

That said, I'm getting the impression that I'm forbidden to do any sort of code cleanup that could be construed as general purpose. I fully appreciate the importance of avoiding design islands, believe me. Yet every other open-source project I have worked on has had it as part of their RFC process to demonstrate the utility of a change, namely by implementing it and using it on a small scale. If such small-scale implementations are forbidden, then I have difficulty seeing how any progress can be made.

[wrengr]

Moved CodegenUtils.h into lib/. Dunno if the "../Utils/CodegenUtils.h" is acceptable style, or if there's a better way. Also rebased.

[wrengr]

added todos for filing RFCs to moving things upstream

[mehdi_amini]

mlir/include/mlir/Dialect/SparseTensor/Utils/CodegenUtils.h

96–113 ↗

(On Diff #392508)

The difficulty is how to manage these "islands" and avoid these to be creeping up. So in general I'm fairly conservative in trying to prevent them from existing in the first place (there are enough out-of-tree users to find such patterns through experiment) and instead be more proactive in refactoring our core infra.

We saw with EDSC that a small island quickly becomes its own continent and is really hard to reconcile with the rest of the codebase later, I think it took more than a year of cleanups to purge it (I don't even know if we succeeded entirely).

Another aspect is that letting islands emerge removes the incentive for folks to take the time to abstract away and generalize the concepts, so pushing back on a new "island" early is a way to keep a stronger incentive to creation motion around things like https://llvm.discourse.group/t/evolving-builder-apis-based-on-lessons-learned-from-edsc/879

In the end I'm not sure there is a perfect answer but it ends up being about various tradeoffs to make...

[aartbik]

mlir/lib/Dialect/SparseTensor/Transforms/SparseTensorConversion.cpp

17

This looks strange. Also, because CodeGenUtils.* is now a private facing API that supports only our codegen in Ttransforms, it seems to me all the the files should simply live in lib/Dialect/SparseTensor/Transforms, i.e. local to our codegen, since Utils should be reserved for implementation of our more public facing support.

@rriddle @mehdi_amini can you please advise?

[mehdi_amini]

mlir/lib/Dialect/SparseTensor/Transforms/SparseTensorConversion.cpp

17

Either way is fine with me.

[aartbik]

mlir/lib/Dialect/SparseTensor/Transforms/SparseTensorConversion.cpp

17

If either is fine with the masters, then I have a slight preference moving all the CodeGenUtils.* into Transforms, Wren. For now, I think you can even keep the "general" constant getters there, and go through the RFC of making them more widely available if you feel like going that route. But having these, now private, convenience methods inside the sparse codegen for now seems totally acceptable for the time being.

[wrengr]

Moving CodegenUtils.{h,cpp} into Transforms/ instead of Utils/.
Also rebasing

mlir/lib/Dialect/SparseTensor/Transforms/SparseTensorConversion.cpp

17

The only issue I can think of re moving CodegenUtils into Transforms/ is that Utils/Merger.cpp also uses it. It's only a very minor issue since, iirc, Merger.cpp only uses one of the simpler functions once; so I'm totally cool with undoing that change for the sake of cleaning up the structure of things in Transforms/*. But it is something to bear in mind in the future depending on how Merger.* and CodegenUtils.* evolve.

[aartbik]

mlir/lib/Dialect/SparseTensor/Transforms/CMakeLists.txt

2 ↗

(On Diff #394691)

I believe you have to list the header file here too now (others please chime in if that is wrong)

[wrengr]

mlir/lib/Dialect/SparseTensor/Transforms/CMakeLists.txt

2 ↗

(On Diff #394691)

It does build without listing the header... And none of the other lib/Dialect/**/CMakeLists.txt list their internal headers (albeit, all their internal headers are named PassDetail.h except for two instances of TypeDetail.h, so I don't know if those are considered special)... So I don't think it should be listed, but then I know very little about cmake so I'd rather get input from someone who does

[aartbik]

utils/bazel/llvm-project-overlay/mlir/BUILD.bazel

1821

Why does the sparse tensor utils depend on sparse tensors now and not the other way around. like it used to?
Is this still required with the new structure? It feels a bit the wrong way to me.

[wrengr]

utils/bazel/llvm-project-overlay/mlir/BUILD.bazel

1821

Imo, it makes the most sense to have the definition of a language be a priori to and independent of any utilities/transformations on that language. The naturalness of this layering is evidenced by the fact that the IR library doesn't actually depend on the utils at all (cf., line 1807 above), as well as the utils needing to pull in the IncGen dependencies if it can't get them from the IR library (which strikes me as a failure/violation of cohesion/encapsulation).

But I can revert the change if you feel strongly

[aartbik]

utils/bazel/llvm-project-overlay/mlir/BUILD.bazel

1821

I suppose that is true. I was thinking mainly of the new utils that support IR building, but there are also utils that indeed use the sparse encoding. Note that part of this is also because the sparse tensor build rule is too large, I really would like to see the sparse tensor attribute being independent of the dialect, especially since this may be used outside the sparse tensor dialect without having to pull in everything (see issue https://github.com/llvm/llvm-project/issues/52748 I filed recently).

[wrengr]

utils/bazel/llvm-project-overlay/mlir/BUILD.bazel

1821

Making the attributes independent of the rest of the dialect makes sense to me. Nicely separates the type definition from the operations, avoiding the analogue of the forward class declaration hack (and opening the way to split up the sparse_tensor dialect or define alternatives to it, if we needed either for some reason).

So, should I revert the BUILD.bazel change or no?

[aartbik]

I am okay with BUILD file as is right now. I think we may bikeshed about this later when we extract encoding out ;-)

[rriddle]

I don't have any concerns after moving to lib/, deferring the approval for this dir to Aart.

[wrengr]

rebase

mlir/lib/Dialect/SparseTensor/Transforms/SparseTensorConversion.cpp

 //===- SparseTensorConversion.cpp - Sparse tensor primitives conversion ---===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 //
 // Convert sparse tensor primitives to calls into a runtime support library.
 // Note that this is a current implementation choice to keep the conversion
 // simple. In principle, these primitives could also be converted to actual
 // elaborate IR code that implements the primitives on the selected sparse
 // tensor storage schemes.
 //
 //===----------------------------------------------------------------------===//
 
+#include "../Utils/CodegenUtils.h"
 #include "mlir/Dialect/Bufferization/IR/Bufferization.h"
 #include "mlir/Dialect/LLVMIR/LLVMDialect.h"
 #include "mlir/Dialect/Linalg/Utils/Utils.h"
 #include "mlir/Dialect/MemRef/IR/MemRef.h"
 #include "mlir/Dialect/SCF/SCF.h"
 #include "mlir/Dialect/SparseTensor/IR/SparseTensor.h"
 #include "mlir/Dialect/SparseTensor/Transforms/Passes.h"
 #include "mlir/Dialect/StandardOps/IR/Ops.h"

 /// Shorthand aliases for the `emitCInterface` argument to `getFunc()`,
 /// `createFuncCall()`, and `replaceOpWithFuncCall()`.
 enum class EmitCInterface : bool { Off = false, On = true };
 
 //===----------------------------------------------------------------------===//
 // Helper methods.
 //===----------------------------------------------------------------------===//
 
-/// Generates a constant zero of the given type.
-inline static Value constantZero(ConversionPatternRewriter &rewriter,
-                                 Location loc, Type t) {
-  return rewriter.create<arith::ConstantOp>(loc, t, rewriter.getZeroAttr(t));
-}
-
-/// Generates a constant of `index` type.
-inline static Value constantIndex(ConversionPatternRewriter &rewriter,
-                                  Location loc, int64_t i) {
-  return rewriter.create<arith::ConstantIndexOp>(loc, i);
-}
-
-/// Generates a constant of `i32` type.
-inline static Value constantI32(ConversionPatternRewriter &rewriter,
-                                Location loc, int32_t i) {
-  return rewriter.create<arith::ConstantIntOp>(loc, i, 32);
-}
-
-/// Generates a constant of `i8` type.
-inline static Value constantI8(ConversionPatternRewriter &rewriter,
-                               Location loc, int8_t i) {
-  return rewriter.create<arith::ConstantIntOp>(loc, i, 8);
-}
-
-/// Generates a constant of the given `Action`.
-static Value constantAction(ConversionPatternRewriter &rewriter, Location loc,
-                            Action action) {
-  return constantI32(rewriter, loc, static_cast<uint32_t>(action));
-}
-
-/// Generates a constant of the internal type encoding for overhead storage.
-static Value constantOverheadTypeEncoding(ConversionPatternRewriter &rewriter,
-                                          Location loc, unsigned width) {
-  OverheadType sec;
-  switch (width) {
-  default:
-    sec = OverheadType::kU64;
-    break;
-  case 32:
-    sec = OverheadType::kU32;
-    break;
-  case 16:
-    sec = OverheadType::kU16;
-    break;
-  case 8:
-    sec = OverheadType::kU8;
-    break;
-  }
-  return constantI32(rewriter, loc, static_cast<uint32_t>(sec));
-}
-
-/// Generates a constant of the internal type encoding for pointer
-/// overhead storage.
-static Value constantPointerTypeEncoding(ConversionPatternRewriter &rewriter,
-                                         Location loc,
-                                         SparseTensorEncodingAttr &enc) {
-  return constantOverheadTypeEncoding(rewriter, loc, enc.getPointerBitWidth());
-}
-
-/// Generates a constant of the internal type encoding for index overhead
-/// storage.
-static Value constantIndexTypeEncoding(ConversionPatternRewriter &rewriter,
-                                       Location loc,
-                                       SparseTensorEncodingAttr &enc) {
-  return constantOverheadTypeEncoding(rewriter, loc, enc.getIndexBitWidth());
-}
-
-/// Generates a constant of the internal type encoding for primary storage.
-static Value constantPrimaryTypeEncoding(ConversionPatternRewriter &rewriter,
-                                         Location loc, Type tp) {
-  PrimaryType primary;
-  if (tp.isF64())
-    primary = PrimaryType::kF64;
-  else if (tp.isF32())
-    primary = PrimaryType::kF32;
-  else if (tp.isInteger(64))
-    primary = PrimaryType::kI64;
-  else if (tp.isInteger(32))
-    primary = PrimaryType::kI32;
-  else if (tp.isInteger(16))
-    primary = PrimaryType::kI16;
-  else if (tp.isInteger(8))
-    primary = PrimaryType::kI8;
-  else
-    llvm_unreachable("Unknown element type");
-  return constantI32(rewriter, loc, static_cast<uint32_t>(primary));
-}
-
-/// Generates a constant of the internal dimension level type encoding.
-static Value
-constantDimLevelTypeEncoding(ConversionPatternRewriter &rewriter, Location loc,
-                             SparseTensorEncodingAttr::DimLevelType dlt) {
-  DimLevelType dlt2;
-  switch (dlt) {
-  case SparseTensorEncodingAttr::DimLevelType::Dense:
-    dlt2 = DimLevelType::kDense;
-    break;
-  case SparseTensorEncodingAttr::DimLevelType::Compressed:
-    dlt2 = DimLevelType::kCompressed;
-    break;
-  case SparseTensorEncodingAttr::DimLevelType::Singleton:
-    dlt2 = DimLevelType::kSingleton;
-    break;
-  }
-  return constantI8(rewriter, loc, static_cast<uint8_t>(dlt2));
-}
-
 /// Returns the equivalent of `void*` for opaque arguments to the
 /// execution engine.
 static Type getOpaquePointerType(PatternRewriter &rewriter) {
   return LLVM::LLVMPointerType::get(rewriter.getI8Type());
 }
 
 /// Returns a function reference (first hit also inserts into module). Sets
 /// the "_emit_c_interface" on the function declaration when requested,
static void newParams(ConversionPatternRewriter &rewriter,
   // User action.
   params.push_back(constantAction(rewriter, loc, action));
   // Payload pointer.
   if (!ptr)
     ptr = rewriter.create<LLVM::NullOp>(loc, getOpaquePointerType(rewriter));
   params.push_back(ptr);
 }
 
-/// Generates the comparison `v != 0` where `v` is of numeric type `t`.
-/// For floating types, we use the "unordered" comparator (i.e., returns
-/// true if `v` is NaN).
-static Value genIsNonzero(ConversionPatternRewriter &rewriter, Location loc,
-                          Value v) {
-  Type t = v.getType();
-  Value zero = constantZero(rewriter, loc, t);
-  if (t.isa<FloatType>())
-    return rewriter.create<arith::CmpFOp>(loc, arith::CmpFPredicate::UNE, v,
-                                          zero);
-  if (t.isIntOrIndex())
-    return rewriter.create<arith::CmpIOp>(loc, arith::CmpIPredicate::ne, v,
-                                          zero);
-  llvm_unreachable("Unknown element type");
-}
-
 /// Generates the code to read the value from tensor[ivs], and conditionally
 /// stores the indices ivs to the memory in ind. The generated code looks like
 /// the following and the insertion point after this routine is inside the
 /// if-then branch behind the assignment to ind. This is to ensure that the
 /// addEltX call generated after is inside the if-then branch.
 ///    if (tensor[ivs]!=0) {
 ///      ind = ivs
 static Value genIndexAndValueForDense(ConversionPatternRewriter &rewriter,

mlir/lib/Dialect/SparseTensor/Transforms/Sparsification.cpp

 //===- Sparsification.cpp - Implementation of sparsification --------------===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 //
 // This file implements converting sparse tensor types to actual sparse code.
 //
 //===----------------------------------------------------------------------===//
 
+#include "../Utils/CodegenUtils.h"
 #include "mlir/Dialect/Affine/IR/AffineOps.h"
 #include "mlir/Dialect/Arithmetic/IR/Arithmetic.h"
 #include "mlir/Dialect/Bufferization/IR/Bufferization.h"
 #include "mlir/Dialect/Linalg/ComprehensiveBufferize/BufferizableOpInterface.h"
 #include "mlir/Dialect/Linalg/IR/LinalgOps.h"
 #include "mlir/Dialect/Linalg/Utils/Utils.h"
 #include "mlir/Dialect/MemRef/IR/MemRef.h"
 #include "mlir/Dialect/SCF/SCF.h"

 /// and a straightforward horizontal reduction will complete the operation.
 static Value genVectorReducInit(CodeGen &codegen, PatternRewriter &rewriter,
                                 Location loc, VectorType vtp) {
   Value r = codegen.redVal;
   switch (codegen.redKind) {
   case kNoReduc:
     break;
   case kSum:
-  case kXor: {
+  case kXor:
     // Initialize reduction vector to: | 0 | .. | 0 | r |
-    Attribute zero = rewriter.getZeroAttr(vtp);
-    Value vec = rewriter.create<arith::ConstantOp>(loc, vtp, zero);
     return rewriter.create<vector::InsertElementOp>(
-        loc, r, vec, rewriter.create<arith::ConstantIndexOp>(loc, 0));
-  }
-  case kProduct: {
+        loc, r, constantZero(rewriter, loc, vtp),
+        constantIndex(rewriter, loc, 0));
+  case kProduct:
     // Initialize reduction vector to: | 1 | .. | 1 | r |
-    Type etp = vtp.getElementType();
-    Attribute one;
-    if (etp.isa<FloatType>())
-      one = rewriter.getFloatAttr(etp, 1.0);
-    else
-      one = rewriter.getIntegerAttr(etp, 1);
-    Value vec = rewriter.create<arith::ConstantOp>(
-        loc, vtp, DenseElementsAttr::get(vtp, one));
     return rewriter.create<vector::InsertElementOp>(
-        loc, r, vec, rewriter.create<arith::ConstantIndexOp>(loc, 0));
-  }
+        loc, r, constantOne(rewriter, loc, vtp),
+        constantIndex(rewriter, loc, 0));
   case kAnd:
   case kOr:
     // Initialize reduction vector to: | r | .. | r | r |
     return rewriter.create<vector::BroadcastOp>(loc, vtp, r);
   }
   llvm_unreachable("unknown reduction kind");
 }
 
static void updateReduc(Merger &merger, CodeGen &codegen, Value reduc) {
   assert(codegen.redKind != kNoReduc);
   codegen.redVal = merger.exp(codegen.redExp).val = reduc;
 }
 
 //===----------------------------------------------------------------------===//
 // Sparse compiler synthesis methods (statements and expressions).
 //===----------------------------------------------------------------------===//
 
-/// Maps sparse integer option to actual integral storage type.
-static Type genIntType(PatternRewriter &rewriter, unsigned width) {
-  if (width == 0)
-    return rewriter.getIndexType();
-  return rewriter.getIntegerType(width);
-}
-
 /// Generates buffer for the output tensor. Note that all sparse kernels
 /// assume that when all elements are written to (viz. x(i) = y(i) * z(i)),
 /// the output buffer is already initialized to all zeroes and only nonzeroes
 /// values are computed and written out. For updates (viz. x(i) += y(i) * z(i)),
 /// only nonzeroes values are used for the updates and no assumption on the
 /// original contents of the output buffer is necessary..
 static Value genOutputBuffer(CodeGen &codegen, PatternRewriter &rewriter,
                              linalg::GenericOp op, MemRefType denseTp,
                              ArrayRef<Value> args) {
   Location loc = op.getLoc();
   Value tensor = op.getOutputOperand(0)->get();
   // The output tensor simply could materialize from the buffer that will
   // be generated for the tensor present in the outs() clause. This has
   // the major advantage that the sparse kernel only updates the nonzero
   // positions for the output tensor.
   if (isInPlace(tensor))
     return rewriter.create<bufferization::ToMemrefOp>(loc, denseTp, tensor);
   // By default, a new buffer is allocated which is initialized to the
   // tensor defined in the outs() clause. This is always correct but
   // introduces a dense initialization component that may negatively
   // impact the running complexity of the sparse kernel. If the tensor
   // materializes into the computation, we need to preserve the zero
   // initialization assumption of all sparse output buffers.
   if (isMaterializing(tensor)) {
-    Type tp = denseTp.getElementType();
     Value alloc = rewriter.create<memref::AllocOp>(loc, denseTp, args);
-    Value zero =
-        rewriter.create<arith::ConstantOp>(loc, tp, rewriter.getZeroAttr(tp));
+    Value zero = constantZero(rewriter, loc, denseTp.getElementType());
     rewriter.create<linalg::FillOp>(loc, zero, alloc);
     return alloc;
   }
   Value init = rewriter.create<bufferization::ToMemrefOp>(loc, denseTp, tensor);
   Value alloc = rewriter.create<memref::AllocOp>(loc, denseTp, args);
   rewriter.create<memref::CopyOp>(loc, init, alloc);
   return alloc;
 }
for (OpOperand *t : op.getInputAndOutputOperands()) {
     for (unsigned d = 0, rank = map.getNumResults(); d < rank; d++) {
       AffineExpr a = map.getResult(perm(enc, d));
       if (a.getKind() != AffineExprKind::DimId)
         continue; // compound
       unsigned idx = a.cast<AffineDimExpr>().getPosition();
       // Handle sparse storage schemes.
       if (merger.isDim(tensor, idx, Dim::kSparse)) {
         auto dynShape = {ShapedType::kDynamicSize};
-        auto ptrTp = MemRefType::get(
-            dynShape, genIntType(rewriter, enc.getPointerBitWidth()));
-        auto indTp = MemRefType::get(
-            dynShape, genIntType(rewriter, enc.getIndexBitWidth()));
-        Value dim = rewriter.create<arith::ConstantIndexOp>(loc, d);
+        auto ptrTp =
+            MemRefType::get(dynShape, getPointerOverheadType(rewriter, enc));
+        auto indTp =
+            MemRefType::get(dynShape, getIndexOverheadType(rewriter, enc));
+        Value dim = constantIndex(rewriter, loc, d);
         // Generate sparse primitives to obtains pointer and indices.
         codegen.pointers[tensor][idx] =
             rewriter.create<ToPointersOp>(loc, ptrTp, t->get(), dim);
         codegen.indices[tensor][idx] =
             rewriter.create<ToIndicesOp>(loc, indTp, t->get(), dim);
       }
       // Find upper bound in current dimension.
       unsigned p = perm(enc, d);
if (!enc) {
       if (tensor < op.getNumInputs())
         codegen.buffers[tensor] =
             rewriter.create<bufferization::ToMemrefOp>(loc, denseTp, t->get());
       else
         codegen.buffers[tensor] =
             genOutputBuffer(codegen, rewriter, op, denseTp, args);
     } else if (t == codegen.sparseOut) {
       // True sparse output needs a lexIdx array.
-      Value rank = rewriter.create<arith::ConstantIndexOp>(loc, op.getRank(t));
+      Value rank = constantIndex(rewriter, loc, op.getRank(t));
       auto dynShape = {ShapedType::kDynamicSize};
       auto memTp = MemRefType::get(dynShape, rewriter.getIndexType());
       codegen.lexIdx = rewriter.create<memref::AllocaOp>(loc, memTp, rank);
     } else {
       // Annotated sparse tensors.
       auto dynShape = {ShapedType::kDynamicSize};
       auto sparseTp = MemRefType::get(dynShape, elementType);
       codegen.buffers[tensor] =

 static VectorType vectorType(CodeGen &codegen, Value ptr) {
   return vectorType(codegen, ptr.getType().cast<MemRefType>().getElementType());
 }
 
 /// Constructs vector iteration mask.
 static Value genVectorMask(CodeGen &codegen, PatternRewriter &rewriter,
                            Value iv, Value lo, Value hi, Value step) {
   Location loc = iv.getLoc();
-  VectorType mtp = vectorType(codegen, genIntType(rewriter, 1));
+  VectorType mtp = vectorType(codegen, rewriter.getI1Type());
   // Special case if the vector length evenly divides the trip count (for
   // example, "for i = 0, 128, 16"). A constant all-true mask is generated
   // so that all subsequent masked memory operations are immediately folded
   // into unconditional memory operations.
   IntegerAttr loInt, hiInt, stepInt;
   if (matchPattern(lo, m_Constant(&loInt)) &&
       matchPattern(hi, m_Constant(&hiInt)) &&
       matchPattern(step, m_Constant(&stepInt))) {
     if (((hiInt.getInt() - loInt.getInt()) % stepInt.getInt()) == 0)
       return rewriter.create<vector::BroadcastOp>(
-          loc, mtp, rewriter.create<arith::ConstantIntOp>(loc, 1, 1));
+          loc, mtp, constantI1(rewriter, loc, true));
   }
   // Otherwise, generate a vector mask that avoids overrunning the upperbound
   // during vector execution. Here we rely on subsequent loop optimizations to
   // avoid executing the mask in all iterations, for example, by splitting the
   // loop into an unconditional vector loop and a scalar cleanup loop.
   auto minMap = AffineMap::get(
       /*dimCount=*/2, /*symbolCount=*/1,
       {rewriter.getAffineSymbolExpr(0),
        rewriter.getAffineDimExpr(0) - rewriter.getAffineDimExpr(1)},
       rewriter.getContext());
   Value end =
       rewriter.createOrFold<AffineMinOp>(loc, minMap, ValueRange{hi, iv, step});
   return rewriter.create<vector::CreateMaskOp>(loc, mtp, end);
 }
 
 /// Generates a vectorized load lhs = a[ind[lo:hi]] or lhs = a[lo:hi].
 static Value genVectorLoad(CodeGen &codegen, PatternRewriter &rewriter,
                            Value ptr, ArrayRef<Value> args) {
   Location loc = ptr.getLoc();
   VectorType vtp = vectorType(codegen, ptr);
-  Value pass =
-      rewriter.create<arith::ConstantOp>(loc, vtp, rewriter.getZeroAttr(vtp));
+  Value pass = constantZero(rewriter, loc, vtp);
   if (args.back().getType().isa<VectorType>()) {
     SmallVector<Value, 4> scalarArgs(args.begin(), args.end());
     Value indexVec = args.back();
-    scalarArgs.back() = rewriter.create<arith::ConstantIndexOp>(loc, 0);
+    scalarArgs.back() = constantIndex(rewriter, loc, 0);
     return rewriter.create<vector::GatherOp>(
         loc, vtp, ptr, scalarArgs, indexVec, codegen.curVecMask, pass);
   }
   return rewriter.create<vector::MaskedLoadOp>(loc, vtp, ptr, args,
                                                codegen.curVecMask, pass);
 }
 
 /// Generates a vectorized store a[ind[lo:hi]] = rhs or a[lo:hi] = rhs.
 static void genVectorStore(CodeGen &codegen, PatternRewriter &rewriter,
                            Value rhs, Value ptr, ArrayRef<Value> args) {
   Location loc = ptr.getLoc();
   if (args.back().getType().isa<VectorType>()) {
     SmallVector<Value, 4> scalarArgs(args.begin(), args.end());
     Value indexVec = args.back();
-    scalarArgs.back() = rewriter.create<arith::ConstantIndexOp>(loc, 0);
+    scalarArgs.back() = constantIndex(rewriter, loc, 0);
     rewriter.create<vector::ScatterOp>(loc, ptr, scalarArgs, indexVec,
                                        codegen.curVecMask, rhs);
     return;
   }
   rewriter.create<vector::MaskedStoreOp>(loc, ptr, args, codegen.curVecMask,
                                          rhs);
 }
 
static Value genAffine(CodeGen &codegen, PatternRewriter &rewriter,
   case AffineExprKind::Mul: {
     auto binOp = a.cast<AffineBinaryOpExpr>();
     return rewriter.create<arith::MulIOp>(
         loc, genAffine(codegen, rewriter, binOp.getLHS(), loc),
         genAffine(codegen, rewriter, binOp.getRHS(), loc));
   }
   case AffineExprKind::Constant: {
     int64_t c = a.cast<AffineConstantExpr>().getValue();
-    return rewriter.create<arith::ConstantIndexOp>(loc, c);
+    return constantIndex(rewriter, loc, c);
   }
   default:
     llvm_unreachable("unexpected affine subscript");
   }
 }
 
 /// Generates index for load/store on sparse tensor.
 static Value genIndex(CodeGen &codegen, linalg::GenericOp op, OpOperand *t) {

 
 /// Generates insertion code to implement dynamic tensor load.
 static Value genInsertionLoad(CodeGen &codegen, PatternRewriter &rewriter,
                               linalg::GenericOp op, OpOperand *t) {
   Location loc = op.getLoc();
   // Direct lexicographic index order, tensor loads as zero.
   if (!codegen.expValues) {
     Type tp = getElementTypeOrSelf(t->get().getType());
-    return rewriter.create<arith::ConstantOp>(loc, tp,
-                                              rewriter.getZeroAttr(tp));
+    return constantZero(rewriter, loc, tp);
   }
   // Load from expanded access pattern.
   Value index = genIndex(codegen, op, t);
   return rewriter.create<memref::LoadOp>(loc, codegen.expValues, index);
 }
 
 /// Generates insertion code to implement dynamic tensor store.
 static void genInsertionStore(CodeGen &codegen, PatternRewriter &rewriter,
                               linalg::GenericOp op, OpOperand *t, Value rhs) {
   Location loc = op.getLoc();
   // Direct insertion in lexicographic index order.
   if (!codegen.expValues) {
     rewriter.create<LexInsertOp>(loc, t->get(), codegen.lexIdx, rhs);
     return;
   }
   // Generates insertion code along expanded access pattern.
   //   if (!expFilled[i]) then
   //     expFilled[i] = true
   //     expAdded[inserts++] = i
   //   endif
   //   values[i] = rhs
   Value index = genIndex(codegen, op, t);
-  Value fval = rewriter.create<arith::ConstantIntOp>(loc, 0, 1); // false
-  Value tval = rewriter.create<arith::ConstantIntOp>(loc, 1, 1); // true
+  Value fval = constantI1(rewriter, loc, false);
+  Value tval = constantI1(rewriter, loc, true);
   // If statement.
   Value filled = rewriter.create<memref::LoadOp>(loc, codegen.expFilled, index);
   Value cond = rewriter.create<arith::CmpIOp>(loc, arith::CmpIPredicate::eq,
                                               filled, fval);
   scf::IfOp ifOp = rewriter.create<scf::IfOp>(loc, rewriter.getIndexType(),
                                               cond, /*else=*/true);
   // True branch.
   rewriter.setInsertionPointToStart(&ifOp.thenRegion().front());
   rewriter.create<memref::StoreOp>(loc, tval, codegen.expFilled, index);
   rewriter.create<memref::StoreOp>(loc, index, codegen.expAdded,
                                    codegen.expCount);
-  Value one = rewriter.create<arith::ConstantIndexOp>(loc, 1);
+  Value one = constantIndex(rewriter, loc, 1);
   Value add = rewriter.create<arith::AddIOp>(loc, codegen.expCount, one);
   rewriter.create<scf::YieldOp>(loc, add);
   // False branch.
   rewriter.setInsertionPointToStart(&ifOp.elseRegion().front());
   rewriter.create<scf::YieldOp>(loc, codegen.expCount);
   rewriter.setInsertionPointAfter(ifOp);
   // Value assignment.
   codegen.expCount = ifOp.getResult(0);
if (codegen.curVecLength > 1) {
     // to state that the indices are unsigned, with creates the potential of
     // incorrect address calculations in the unlikely case we need such
     // extremely large offsets.
     Type etp = ptr.getType().cast<MemRefType>().getElementType();
     Value vload = genVectorLoad(codegen, rewriter, ptr, {s});
     if (!etp.isa<IndexType>()) {
       if (etp.getIntOrFloatBitWidth() < 32)
         vload = rewriter.create<arith::ExtUIOp>(
-            loc, vload, vectorType(codegen, genIntType(rewriter, 32)));
+            loc, vload, vectorType(codegen, rewriter.getI32Type()));
       else if (etp.getIntOrFloatBitWidth() < 64 &&
                !codegen.options.enableSIMDIndex32)
         vload = rewriter.create<arith::ExtUIOp>(
-            loc, vload, vectorType(codegen, genIntType(rewriter, 64)));
+            loc, vload, vectorType(codegen, rewriter.getI64Type()));
     }
     return vload;
   }
   // For the scalar case, we simply zero extend narrower indices into 64-bit
   // values before casting to index without a performance penalty. Here too,
   // however, indices that already are 64-bit, in theory, cannot express the
   // full range as explained above.
   Value load = rewriter.create<memref::LoadOp>(loc, ptr, s);
   if (!load.getType().isa<IndexType>()) {
     if (load.getType().getIntOrFloatBitWidth() < 64)
-      load =
-          rewriter.create<arith::ExtUIOp>(loc, load, genIntType(rewriter, 64));
+      load = rewriter.create<arith::ExtUIOp>(loc, load, rewriter.getI64Type());
     load =
         rewriter.create<arith::IndexCastOp>(loc, load, rewriter.getIndexType());
   }
   return load;
 }
 
 /// Generates an invariant value.
 static Value genInvariantValue(Merger &merger, CodeGen &codegen,
if (!lhs || codegen.outerParNest != op.getRank(lhs) - 1 ||
     return; // not needed at this level
   // Generate start or end of an expanded access pattern.
   Value tensor = lhs->get();
   Location loc = op.getLoc();
   if (atStart) {
     auto dynShape = {ShapedType::kDynamicSize};
     Type etp = tensor.getType().cast<ShapedType>().getElementType();
     Type t1 = MemRefType::get(dynShape, etp);
-    Type t2 = MemRefType::get(dynShape, genIntType(rewriter, 1));
-    Type t3 = MemRefType::get(dynShape, genIntType(rewriter, 0));
+    Type t2 = MemRefType::get(dynShape, rewriter.getI1Type());
+    Type t3 = MemRefType::get(dynShape, rewriter.getIndexType());
     Type t4 = rewriter.getIndexType();
     auto res =
         rewriter.create<ExpandOp>(loc, TypeRange({t1, t2, t3, t4}), tensor);
     assert(res.getNumResults() == 4);
     assert(!codegen.expValues);
     codegen.expValues = res.getResult(0);
     codegen.expFilled = res.getResult(1);
     codegen.expAdded = res.getResult(2);
if (inits[b]) {
       if (merger.isDim(b, Dim::kSparse)) {
         // Initialize sparse index.
         unsigned pat = at;
         for (; pat != 0; pat--) {
           if (codegen.pidxs[tensor][topSort[pat - 1]])
             break;
         }
         Value ptr = codegen.pointers[tensor][idx];
-        Value one = rewriter.create<arith::ConstantIndexOp>(loc, 1);
-        Value p0 = (pat == 0) ? rewriter.create<arith::ConstantIndexOp>(loc, 0)
+        Value one = constantIndex(rewriter, loc, 1);
+        Value p0 = (pat == 0) ? constantIndex(rewriter, loc, 0)
                               : codegen.pidxs[tensor][topSort[pat - 1]];
         codegen.pidxs[tensor][idx] = genLoad(codegen, rewriter, loc, ptr, p0);
         Value p1 = rewriter.create<arith::AddIOp>(loc, p0, one);
         codegen.highs[tensor][idx] = genLoad(codegen, rewriter, loc, ptr, p1);
       } else {
         // Dense index still in play.
         needsUniv = true;
       }
     }
   }
 
   // Initialize the universal dense index.
-  codegen.loops[idx] = rewriter.create<arith::ConstantIndexOp>(loc, 0);
+  codegen.loops[idx] = constantIndex(rewriter, loc, 0);
   return needsUniv;
 }
 
 /// Returns vectorization strategy. Any implicit inner loop in the Linalg
 /// operation is a candidate. Whether it is actually converted to SIMD code
 /// depends on the requested strategy.
 static bool isVectorFor(CodeGen &codegen, bool isInner, bool isSparse) {
   switch (codegen.options.vectorizationStrategy) {
static Operation *genFor(Merger &merger, CodeGen &codegen,
   // Prepare vector length.
   if (isVector)
     codegen.curVecLength = codegen.options.vectorLength;
 
   // Loop bounds and increment.
   Location loc = op.getLoc();
   Value lo = isSparse ? codegen.pidxs[tensor][idx] : codegen.loops[idx];
   Value hi = isSparse ? codegen.highs[tensor][idx] : codegen.sizes[idx];
-  Value step =
-      rewriter.create<arith::ConstantIndexOp>(loc, codegen.curVecLength);
+  Value step = constantIndex(rewriter, loc, codegen.curVecLength);
 
   // Emit a parallel loop.
   if (isParallel) {
     assert(!isVector);
     scf::ParallelOp parOp = rewriter.create<scf::ParallelOp>(loc, lo, hi, step);
     if (isSparse)
       codegen.pidxs[tensor][idx] = parOp.getInductionVars()[0];
     else
for (unsigned b = 0, be = locals.size(); b < be; b++) {
     if ((locals[b] || merger.isOutTensor(b, idx)) &&
         merger.isDim(b, Dim::kDense)) {
       unsigned tensor = merger.tensor(b);
       assert(idx == merger.index(b));
       unsigned pat = at;
       for (; pat != 0; pat--)
         if (codegen.pidxs[tensor][topSort[pat - 1]])
           break;
-      Value p = (pat == 0) ? rewriter.create<arith::ConstantIndexOp>(loc, 0)
+      Value p = (pat == 0) ? constantIndex(rewriter, loc, 0)
                            : codegen.pidxs[tensor][topSort[pat - 1]];
       codegen.pidxs[tensor][idx] = genAddress(
           codegen, rewriter, loc, codegen.sizes[idx], p, codegen.loops[idx]);
     }
   }
 
   // Move the insertion indices in lexicographic index order. During access
   // pattern expansion, we can skip setting the innermost dimension.
   if (codegen.sparseOut && !codegen.expValues) {
-    Value pos = rewriter.create<arith::ConstantIndexOp>(loc, at);
+    Value pos = constantIndex(rewriter, loc, at);
     rewriter.create<memref::StoreOp>(loc, codegen.loops[idx], codegen.lexIdx,
                                      pos);
   }
 }
 
 /// Generates the induction structure for a while-loop.
 static void genWhileInduction(Merger &merger, CodeGen &codegen,
                               PatternRewriter &rewriter, linalg::GenericOp op,
static void genWhileInduction(Merger &merger, CodeGen &codegen,
   rewriter.setInsertionPointToEnd(&whileOp.after().front());
   // Finalize the induction. Note that the induction could be performed
   // in the individual if-branches to avoid re-evaluating the conditions.
   // However, that would result in a rather elaborate forest of yield
   // instructions during code generation. Moreover, performing the induction
   // after the if-statements more closely resembles code generated by TACO.
   unsigned o = 0;
   SmallVector<Value, 4> operands;
-  Value one = rewriter.create<arith::ConstantIndexOp>(loc, 1);
+  Value one = constantIndex(rewriter, loc, 1);
   for (unsigned b = 0, be = induction.size(); b < be; b++) {
     if (induction[b] && merger.isDim(b, Dim::kSparse)) {
       unsigned tensor = merger.tensor(b);
       assert(idx == merger.index(b));
       Value op1 = codegen.idxs[tensor][idx];
       Value op2 = codegen.loops[idx];
       Value op3 = codegen.pidxs[tensor][idx];
       Value cmp = rewriter.create<arith::CmpIOp>(loc, arith::CmpIPredicate::eq,
if (conditions[b]) {
       assert(idx == merger.index(b));
       Value clause;
       if (merger.isDim(b, Dim::kSparse)) {
         Value op1 = codegen.idxs[tensor][idx];
         Value op2 = codegen.loops[idx];
         clause = rewriter.create<arith::CmpIOp>(loc, arith::CmpIPredicate::eq,
                                                 op1, op2);
       } else {
-        clause = rewriter.create<arith::ConstantIntOp>(loc, 1, 1); // true
+        clause = constantI1(rewriter, loc, true);
       }
       cond = cond ? rewriter.create<arith::AndIOp>(loc, cond, clause) : clause;
     }
   }
   if (codegen.redVal)
     types.push_back(codegen.redVal.getType());
   if (codegen.expValues)
     types.push_back(rewriter.getIndexType());

mlir/lib/Dialect/SparseTensor/Utils/CMakeLists.txt

 add_mlir_dialect_library(MLIRSparseTensorUtils
   Merger.cpp
+  CodegenUtils.cpp
 
   ADDITIONAL_HEADER_DIRS
   ${MLIR_MAIN_INCLUDE_DIR}/mlir/Dialect/SparseTensor
 
   LINK_LIBS PUBLIC
   MLIRArithmetic
   MLIRIR
   MLIRLinalg
 )

mlir/lib/Dialect/SparseTensor/Utils/CodegenUtils.h

+//===- CodegenUtils.h - Utilities for generating MLIR -----------*- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This header file defines utilities for generating MLIR.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef MLIR_DIALECT_SPARSETENSOR_UTILS_CODEGENUTILS_H_
+#define MLIR_DIALECT_SPARSETENSOR_UTILS_CODEGENUTILS_H_
+
+#include "mlir/Dialect/Arithmetic/IR/Arithmetic.h"
+#include "mlir/Dialect/SparseTensor/IR/SparseTensor.h"
+#include "mlir/ExecutionEngine/SparseTensorUtils.h"
+#include "mlir/IR/Builders.h"
+
+namespace mlir {
+class Location;
+class Type;
+class Value;
+
+namespace sparse_tensor {
+
+//===----------------------------------------------------------------------===//
+// ExecutionEngine/SparseTensorUtils helper functions.
+//===----------------------------------------------------------------------===//
+
+/// Converts an overhead storage bitwidth to its internal type-encoding.
+OverheadType overheadTypeEncoding(unsigned width);
+
+/// Converts the internal type-encoding for overhead storage to an mlir::Type.
+Type getOverheadType(Builder &builder, OverheadType ot);
+
+/// Returns the mlir::Type for pointer overhead storage.
+Type getPointerOverheadType(Builder &builder,
+                            const SparseTensorEncodingAttr &enc);
+
+/// Returns the mlir::Type for index overhead storage.
+Type getIndexOverheadType(Builder &builder,
+                          const SparseTensorEncodingAttr &enc);
+
+/// Converts a primary storage type to its internal type-encoding.
+PrimaryType primaryTypeEncoding(Type elemTp);
+
+/// Converts the IR's dimension level type to its internal type-encoding.
+DimLevelType dimLevelTypeEncoding(SparseTensorEncodingAttr::DimLevelType dlt);
+
+//===----------------------------------------------------------------------===//
+// Misc code generators.
+//===----------------------------------------------------------------------===//
+
+/// Generates a 1-valued attribute of the given type.  This supports
+/// all the same types as `getZeroAttr`; however, unlike `getZeroAttr`,
+/// for unsupported types we raise `llvm_unreachable` rather than
+/// returning a null attribute.
+Attribute getOneAttr(Builder &builder, Type tp);
+
+/// Generates the comparison `v != 0` where `v` is of numeric type.
+/// For floating types, we use the "unordered" comparator (i.e., returns
+/// true if `v` is NaN).
+Value genIsNonzero(OpBuilder &builder, Location loc, Value v);
+
+//===----------------------------------------------------------------------===//
+// Constant generators.
+//
+// All these functions are just wrappers to improve code legibility;
+// therefore, we mark them as `inline` to avoid introducing any additional
+// overhead due to the legibility.
+//===----------------------------------------------------------------------===//
+
+/// Generates a 0-valued constant of the given type.  In addition to
+/// the scalar types (`FloatType`, `IndexType`, `IntegerType`), this also
+/// works for `RankedTensorType` and `VectorType` (for which it generates
+/// a constant `DenseElementsAttr` of zeros).
+inline Value constantZero(OpBuilder &builder, Location loc, Type tp) {
+  return builder.create<arith::ConstantOp>(loc, tp, builder.getZeroAttr(tp));
+}
+
+/// Generates a 1-valued constant of the given type.  This supports all
+/// the same types as `constantZero`.
+inline Value constantOne(OpBuilder &builder, Location loc, Type tp) {
+  return builder.create<arith::ConstantOp>(loc, tp, getOneAttr(builder, tp));
+}
+
+/// Generates a constant of `index` type.
+inline Value constantIndex(OpBuilder &builder, Location loc, int64_t i) {
+  return builder.create<arith::ConstantIndexOp>(loc, i);
+}
+
+/// Generates a constant of `i32` type.
+inline Value constantI32(OpBuilder &builder, Location loc, int32_t i) {
+  return builder.create<arith::ConstantIntOp>(loc, i, 32);
+}
+
+/// Generates a constant of `i16` type.
+inline Value constantI16(OpBuilder &builder, Location loc, int16_t i) {
+  return builder.create<arith::ConstantIntOp>(loc, i, 16);
+}
+
+/// Generates a constant of `i8` type.
+inline Value constantI8(OpBuilder &builder, Location loc, int8_t i) {
+  return builder.create<arith::ConstantIntOp>(loc, i, 8);
+}
+
+/// Generates a constant of `i1` type.
+inline Value constantI1(OpBuilder &builder, Location loc, bool b) {
+  return builder.create<arith::ConstantIntOp>(loc, b, 1);
+}
+
+/// Generates a constant of the given `Action`.
+inline Value constantAction(OpBuilder &builder, Location loc, Action action) {
+  return constantI32(builder, loc, static_cast<uint32_t>(action));
+}
+
+/// Generates a constant of the internal type-encoding for overhead storage.
+inline Value constantOverheadTypeEncoding(OpBuilder &builder, Location loc,
+                                          unsigned width) {
+  return constantI32(builder, loc,
+                     static_cast<uint32_t>(overheadTypeEncoding(width)));
+}
+
+/// Generates a constant of the internal type-encoding for pointer
+/// overhead storage.
+inline Value constantPointerTypeEncoding(OpBuilder &builder, Location loc,
+                                         const SparseTensorEncodingAttr &enc) {
+  return constantOverheadTypeEncoding(builder, loc, enc.getPointerBitWidth());
+}
+
+/// Generates a constant of the internal type-encoding for index overhead
+/// storage.
+inline Value constantIndexTypeEncoding(OpBuilder &builder, Location loc,
+                                       const SparseTensorEncodingAttr &enc) {
+  return constantOverheadTypeEncoding(builder, loc, enc.getIndexBitWidth());
+}
+
+/// Generates a constant of the internal type-encoding for primary storage.
+inline Value constantPrimaryTypeEncoding(OpBuilder &builder, Location loc,
+                                         Type elemTp) {
+  return constantI32(builder, loc,
+                     static_cast<uint32_t>(primaryTypeEncoding(elemTp)));
+}
+
+/// Generates a constant of the internal dimension level type encoding.
+inline Value
+constantDimLevelTypeEncoding(OpBuilder &builder, Location loc,
+                             SparseTensorEncodingAttr::DimLevelType dlt) {
+  return constantI8(builder, loc,
+                    static_cast<uint8_t>(dimLevelTypeEncoding(dlt)));
+}
+
+} // namespace sparse_tensor
+} // namespace mlir
+
+#endif // MLIR_DIALECT_SPARSETENSOR_UTILS_CODEGENUTILS_H_

mlir/lib/Dialect/SparseTensor/Utils/CodegenUtils.cpp

+//===- CodegenUtils.cpp - Utilities for generating MLIR -------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#include "CodegenUtils.h"
+
+#include "mlir/IR/Types.h"
+#include "mlir/IR/Value.h"
+
+using namespace mlir::sparse_tensor;
+
+//===----------------------------------------------------------------------===//
+// ExecutionEngine/SparseTensorUtils helper functions.
+//===----------------------------------------------------------------------===//
+
+OverheadType mlir::sparse_tensor::overheadTypeEncoding(unsigned width) {
+  switch (width) {
+  default:
+    return OverheadType::kU64;
+  case 32:
+    return OverheadType::kU32;
+  case 16:
+    return OverheadType::kU16;
+  case 8:
+    return OverheadType::kU8;
+  }
+}
+
+mlir::Type mlir::sparse_tensor::getOverheadType(mlir::Builder &builder,
+                                                OverheadType ot) {
+  switch (ot) {
+  case OverheadType::kU64:
+    return builder.getIntegerType(64);
+  case OverheadType::kU32:
+    return builder.getIntegerType(32);
+  case OverheadType::kU16:
+    return builder.getIntegerType(16);
+  case OverheadType::kU8:
+    return builder.getIntegerType(8);
+  }
+  llvm_unreachable("Unknown OverheadType");
+}
+
+mlir::Type mlir::sparse_tensor::getPointerOverheadType(
+    mlir::Builder &builder, const SparseTensorEncodingAttr &enc) {
+  // NOTE(wrengr): This workaround will be fixed in D115010.
+  unsigned width = enc.getPointerBitWidth();
+  if (width == 0)
+    return builder.getIndexType();
+  return getOverheadType(builder, overheadTypeEncoding(width));
+}
+
+mlir::Type
+mlir::sparse_tensor::getIndexOverheadType(mlir::Builder &builder,
+                                          const SparseTensorEncodingAttr &enc) {
+  // NOTE(wrengr): This workaround will be fixed in D115010.
+  unsigned width = enc.getIndexBitWidth();
+  if (width == 0)
+    return builder.getIndexType();
+  return getOverheadType(builder, overheadTypeEncoding(width));
+}
+
+PrimaryType mlir::sparse_tensor::primaryTypeEncoding(mlir::Type elemTp) {
+  if (elemTp.isF64())
+    return PrimaryType::kF64;
+  if (elemTp.isF32())
+    return PrimaryType::kF32;
+  if (elemTp.isInteger(64))
+    return PrimaryType::kI64;
+  if (elemTp.isInteger(32))
+    return PrimaryType::kI32;
+  if (elemTp.isInteger(16))
+    return PrimaryType::kI16;
+  if (elemTp.isInteger(8))
+    return PrimaryType::kI8;
+  llvm_unreachable("Unknown primary type");
+}
+
+DimLevelType mlir::sparse_tensor::dimLevelTypeEncoding(
+    SparseTensorEncodingAttr::DimLevelType dlt) {
+  switch (dlt) {
+  case SparseTensorEncodingAttr::DimLevelType::Dense:
+    return DimLevelType::kDense;
+  case SparseTensorEncodingAttr::DimLevelType::Compressed:
+    return DimLevelType::kCompressed;
+  case SparseTensorEncodingAttr::DimLevelType::Singleton:
+    return DimLevelType::kSingleton;
+  }
+  llvm_unreachable("Unknown SparseTensorEncodingAttr::DimLevelType");
+}
+
+//===----------------------------------------------------------------------===//
+// Misc code generators.
+//===----------------------------------------------------------------------===//
+
+mlir::Attribute mlir::sparse_tensor::getOneAttr(mlir::Builder &builder,
+                                                mlir::Type tp) {
+  if (tp.isa<FloatType>())
+    return builder.getFloatAttr(tp, 1.0);
+  if (tp.isa<IndexType>())
+    return builder.getIndexAttr(1);
+  if (auto intTp = tp.dyn_cast<IntegerType>())
+    return builder.getIntegerAttr(tp, APInt(intTp.getWidth(), 1));
+  if (tp.isa<RankedTensorType, VectorType>()) {
+    auto shapedTp = tp.cast<ShapedType>();
+    if (auto one = getOneAttr(builder, shapedTp.getElementType()))
+      return DenseElementsAttr::get(shapedTp, one);
+  }
+  llvm_unreachable("Unsupported attribute type");
+}
+
+mlir::Value mlir::sparse_tensor::genIsNonzero(mlir::OpBuilder &builder,
+                                              mlir::Location loc,
+                                              mlir::Value v) {
+  mlir::Type tp = v.getType();
+  mlir::Value zero = constantZero(builder, loc, tp);
+  if (tp.isa<FloatType>())
+    return builder.create<arith::CmpFOp>(loc, arith::CmpFPredicate::UNE, v,
+                                         zero);
+  if (tp.isIntOrIndex())
+    return builder.create<arith::CmpIOp>(loc, arith::CmpIPredicate::ne, v,
+                                         zero);
+  llvm_unreachable("Non-numeric type");
+}

mlir/lib/Dialect/SparseTensor/Utils/Merger.cpp

 //===- Merger.cpp - Implementation of iteration lattices ------------------===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 
 #include "mlir/Dialect/SparseTensor/Utils/Merger.h"
+#include "CodegenUtils.h"
 #include "mlir/Dialect/Arithmetic/IR/Arithmetic.h"
 
 #include "mlir/IR/Operation.h"
 #include "llvm/Support/Debug.h"
 
 namespace mlir {
 namespace sparse_tensor {
 
Value Merger::buildExp(PatternRewriter &rewriter, Location loc, unsigned e,
   case kCeilF:
     return rewriter.create<math::CeilOp>(loc, v0);
   case kFloorF:
     return rewriter.create<math::FloorOp>(loc, v0);
   case kNegF:
     return rewriter.create<arith::NegFOp>(loc, v0);
   case kNegI: // no negi in std
     return rewriter.create<arith::SubIOp>(
-        loc,
-        rewriter.create<arith::ConstantOp>(loc, v0.getType(),
-                                           rewriter.getZeroAttr(v0.getType())),
-        v0);
+        loc, constantZero(rewriter, loc, v0.getType()), v0);
   case kTruncF:
     return rewriter.create<arith::TruncFOp>(loc, v0, inferType(e, v0));
   case kExtF:
     return rewriter.create<arith::ExtFOp>(loc, v0, inferType(e, v0));
   case kCastFS:
     return rewriter.create<arith::FPToSIOp>(loc, v0, inferType(e, v0));
   case kCastFU:
     return rewriter.create<arith::FPToUIOp>(loc, v0, inferType(e, v0));

utils/bazel/llvm-project-overlay/mlir/BUILD.bazel

cc_library(
     includes = ["include"],
     deps = [
         ":ArithmeticDialect",
         ":IR",
         ":InferTypeOpInterface",
         ":SideEffectInterfaces",
         ":SparseTensorAttrDefsIncGen",
         ":SparseTensorOpsIncGen",
-        ":SparseTensorUtils",
         ":StandardOps",
         "//llvm:Support",
     ],
 )
 
 cc_library(
     name = "SparseTensorUtils",
-    srcs = glob(["lib/Dialect/SparseTensor/Utils/*.cpp"]),
-    hdrs = glob(["include/mlir/Dialect/SparseTensor/Utils/*.h"]),
+    srcs = glob([
+        "lib/Dialect/SparseTensor/Utils/*.cpp",
+        "lib/Dialect/SparseTensor/Utils/*.h",
+    ]),
+    hdrs = glob([
+        "include/mlir/Dialect/SparseTensor/Utils/*.h",
+    ]) + [
+        "include/mlir/ExecutionEngine/SparseTensorUtils.h",
+    ],
     includes = ["include"],
     deps = [
         ":ArithmeticDialect",
         ":IR",
         ":LinalgOps",
-        ":SideEffectInterfaces",
-        ":SparseTensorAttrDefsIncGen",
-        ":SparseTensorOpsIncGen",
-        ":StandardOps",
+        ":SparseTensor",
         "//llvm:Support",
     ],
 )
 
 cc_library(
     name = "SparseTensorTransforms",
-    srcs = glob(["lib/Dialect/SparseTensor/Transforms/*.cpp"]),
+    srcs = glob([
+        "lib/Dialect/SparseTensor/Transforms/*.cpp",
+    ]) + [
+        "lib/Dialect/SparseTensor/Utils/CodegenUtils.h",
+    ],
     hdrs = [
         "include/mlir/Dialect/SparseTensor/Transforms/Passes.h",
         "include/mlir/ExecutionEngine/SparseTensorUtils.h",
     ],
     includes = ["include"],
     deps = [
         ":Affine",
         ":ArithmeticDialect",