[mlir][EmitC] Expand the MemRefToEmitC pass - Lowering `reinterpret_cast` #152610

Jaddyen · 2025-08-07T23:09:50Z

This patch lowers memref.reinterpret_cast.
From:

func.func @casting(%arg0: memref<999xi32>) {
  %reinterpret_cast_5 = memref.reinterpret_cast %arg0 to offset: [0], sizes: [1, 1, 999], strides: [999, 999, 1] : memref<999xi32> to memref<1x1x999xi32>
  return
}

To:

void casting(int32_t v1[999]) {
  std::size_t v2 = 0;
  int32_t* v3 = &v1[v2];
  int32_t(*v4)[1][1][999]  = reinterpret_cast<int32_t(*)[1][1][999]>(v3);
  return;
}

ilovepi · 2025-08-08T17:23:33Z

In your example, int32_t(*)[1][1][999] v4 is invalid syntax, right? Shouldn't it be: int32_t(*v4)[1][1][999]? Is that a typo in the description? or is that what's generated/translated?

ilovepi · 2025-08-08T18:01:25Z

mlir/lib/Conversion/MemRefToEmitC/MemRefToEmitC.cpp

+    emitc::ConstantOp zeroIndex = rewriter.create<emitc::ConstantOp>(
+        loc, rewriter.getIndexType(), rewriter.getIndexAttr(0));


I think I've seen a zeroIndex crated in several places now. Its not much code, but that may be a good candidate for a helper function (e.g. in the anonymous namespace and marked static).

ilovepi · 2025-08-08T18:02:14Z

mlir/lib/Conversion/MemRefToEmitC/MemRefToEmitC.cpp

+      llvm::SmallVector<mlir::Value> indices;
+      for (int i = 0; i < rank; ++i) {
+        indices.push_back(zeroIndex);
+      }


Suggested change

llvm::SmallVector<mlir::Value> indices;

for (int i = 0; i < rank; ++i) {

indices.push_back(zeroIndex);

}

llvm::SmallVector<mlir::Value> indices(rank, zeroIndex);

ilovepi · 2025-08-08T18:02:34Z

mlir/lib/Conversion/MemRefToEmitC/MemRefToEmitC.cpp

+    auto createPointerFromEmitcArray =
+        [loc, &rewriter, &zeroIndex](
+            mlir::TypedValue<emitc::ArrayType> arrayValue) -> emitc::ApplyOp {
+      int64_t rank = arrayValue.getType().getRank();
+      llvm::SmallVector<mlir::Value> indices;
+      for (int i = 0; i < rank; ++i) {
+        indices.push_back(zeroIndex);
+      }


This closure seems familar from your other patches. I'm guessing that means it should probably be a helper function instead, since I don't see anything in the capture list that couldn't be a parameter.

ilovepi · 2025-08-08T18:13:00Z

mlir/test/Conversion/MemRefToEmitC/memref-to-emitc-reinterpret-cast.mlir

+//CHECK-NEXT:     %1 = "emitc.constant"() <{value = 0 : index}> : () -> index
+//CHECK-NEXT:     %2 = emitc.subscript %0[%1] : (!emitc.array<999xi32>, index) -> !emitc.lvalue<i32>
+//CHECK-NEXT:     %3 = emitc.apply "&"(%2) : (!emitc.lvalue<i32>) -> !emitc.ptr<i32>
+//CHECK-NEXT:     %4 = emitc.call_opaque "reinterpret_cast"(%3) {args = [0 : index], template_args = [!emitc.ptr<!emitc.array<1x1x999xi32>>]} : (!emitc.ptr<i32>) -> !emitc.ptr<!emitc.array<1x1x999xi32>>


an actual reinterpret_cast<T>() is appropriate for C++, but for C, you'd need a C-style cast. given that the C-style cast will work in both languages and what you're doing can't be covered by static or dyn casts, maybe we should use a normal cast operation? Would that work here, do you think?

+1, good to check.

jpienaar · 2025-08-11T03:23:11Z

mlir/lib/Conversion/MemRefToEmitC/MemRefToEmitC.cpp

+        indices.push_back(zeroIndex);
+      }
+
+      emitc::SubscriptOp subPtr = rewriter.create<emitc::SubscriptOp>(


If you switch to the new builder form (emit::SubscriptOp::create(...)) you should no longer need ValueRange here.

jpienaar · 2025-08-11T03:23:42Z

mlir/lib/Conversion/MemRefToEmitC/MemRefToEmitC.cpp

+    auto castCall = rewriter.create<emitc::CastOp>(
+        loc, emitc::PointerType::get(targetInEmitC), srcPtr.getResult());
+
+    rewriter.replaceOp(castOp, castCall);


You can combine this with the preceding by doing rewriter.replaceWithNewOp

jpienaar · 2025-08-11T03:24:32Z

mlir/lib/Target/Cpp/TranslateToCpp.cpp


 LogicalResult CppEmitter::emitVariableDeclaration(Location loc, Type type,
                                                  StringRef name) {
+  if (auto pType = dyn_cast<emitc::PointerType>(type)) {


Could this be a standalone PR? (this feels like it is a general refinement).

jpienaar · 2025-08-11T03:26:02Z

mlir/test/Conversion/MemRefToEmitC/memref-to-emitc-reinterpret-cast.mlir

+//CHECK-NEXT:     %1 = "emitc.constant"() <{value = 0 : index}> : () -> index
+//CHECK-NEXT:     %2 = emitc.subscript %0[%1] : (!emitc.array<999xi32>, index) -> !emitc.lvalue<i32>
+//CHECK-NEXT:     %3 = emitc.apply "&"(%2) : (!emitc.lvalue<i32>) -> !emitc.ptr<i32>
+//CHECK-NEXT:     %4 = emitc.call_opaque "reinterpret_cast"(%3) {args = [0 : index], template_args = [!emitc.ptr<!emitc.array<1x1x999xi32>>]} : (!emitc.ptr<i32>) -> !emitc.ptr<!emitc.array<1x1x999xi32>>


+1, good to check.

jpienaar · 2025-08-11T03:30:35Z

mlir/lib/Conversion/MemRefToEmitC/MemRefToEmitC.cpp

+      return rewriter.notifyMatchFailure(castOp.getLoc(),
+                                         "cannot convert memref type");
+    }
+    Location loc = castOp.getLoc();


Empty lines are for me similar to paragraphs in text, it creates logical separations/groupings which aid reading. Here I can't quite figure those out.

EtoAndruwa · 2025-08-13T09:44:35Z

mlir/test/Conversion/MemRefToEmitC/memref-to-emitc-reinterpret-cast.mlir

@@ -0,0 +1,16 @@
+// RUN: mlir-opt -convert-memref-to-emitc %s -split-input-file | FileCheck %s


Suggested change

// RUN: mlir-opt -convert-memref-to-emitc %s -split-input-file | FileCheck %s

// RUN: mlir-opt -convert-memref-to-emitc %s | FileCheck %s

The -split-input-file option must be dropped since there is no need to use it. Test includes only one function.

EtoAndruwa · 2025-08-13T09:57:11Z

mlir/lib/Conversion/MemRefToEmitC/MemRefToEmitC.cpp

+    auto srcArrayValue =
+        cast<TypedValue<emitc::ArrayType>>(adaptor.getSource());
+
+    emitc::ConstantOp zeroIndex = rewriter.create<emitc::ConstantOp>(


Using emitc::Literal can reduce emitted code by avoiding extra creation of constants, as the literal's value is emitted directly in place.

Jaddyen added 2 commits August 6, 2025 20:18

needs improvment

c6ea995

almost functional option

d09ccc0

Jaddyen requested review from ilovepi and jpienaar August 7, 2025 23:15

ilovepi reviewed Aug 8, 2025

View reviewed changes

cpp output that compiles

19e23d5

jpienaar reviewed Aug 11, 2025

View reviewed changes

EtoAndruwa reviewed Aug 13, 2025

View reviewed changes

		emitc::ConstantOp zeroIndex = rewriter.create<emitc::ConstantOp>(
		loc, rewriter.getIndexType(), rewriter.getIndexAttr(0));

		@@ -0,0 +1,16 @@
		// RUN: mlir-opt -convert-memref-to-emitc %s -split-input-file \| FileCheck %s

[mlir][EmitC] Expand the MemRefToEmitC pass - Lowering reinterpret_cast #152610

Are you sure you want to change the base?

[mlir][EmitC] Expand the MemRefToEmitC pass - Lowering reinterpret_cast #152610

Uh oh!

Conversation

Jaddyen commented Aug 7, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

ilovepi commented Aug 8, 2025

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

[mlir][EmitC] Expand the MemRefToEmitC pass - Lowering `reinterpret_cast` #152610

[mlir][EmitC] Expand the MemRefToEmitC pass - Lowering `reinterpret_cast` #152610

Jaddyen commented Aug 7, 2025 •

edited

Loading