Skip to content

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

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Draft
wants to merge 3 commits into
base: main
Choose a base branch
from
Draft
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
62 changes: 61 additions & 1 deletion mlir/lib/Conversion/MemRefToEmitC/MemRefToEmitC.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -20,8 +20,11 @@
#include "mlir/IR/PatternMatch.h"
#include "mlir/IR/TypeRange.h"
#include "mlir/IR/Value.h"
#include "mlir/IR/ValueRange.h"
#include "mlir/Transforms/DialectConversion.h"
#include "llvm/Support/FormatVariadic.h"
#include <cstdint>
#include <string>

using namespace mlir;

Expand Down Expand Up @@ -269,6 +272,62 @@ struct ConvertLoad final : public OpConversionPattern<memref::LoadOp> {
}
};

struct ConvertReinterpretCastOp final
: public OpConversionPattern<memref::ReinterpretCastOp> {
using OpConversionPattern::OpConversionPattern;

LogicalResult
matchAndRewrite(memref::ReinterpretCastOp castOp, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {

MemRefType srcType = cast<MemRefType>(castOp.getSource().getType());

MemRefType targetMemRefType =
cast<MemRefType>(castOp.getResult().getType());

auto srcInEmitC = convertMemRefType(srcType, getTypeConverter());
auto targetInEmitC =
convertMemRefType(targetMemRefType, getTypeConverter());
if (!srcInEmitC || !targetInEmitC) {
return rewriter.notifyMatchFailure(castOp.getLoc(),
"cannot convert memref type");
}
Location loc = castOp.getLoc();
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

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.


auto srcArrayValue =
cast<TypedValue<emitc::ArrayType>>(adaptor.getSource());

emitc::ConstantOp zeroIndex = rewriter.create<emitc::ConstantOp>(
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

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

loc, rewriter.getIndexType(), rewriter.getIndexAttr(0));
Comment on lines +300 to +301
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

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).


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);
}
Comment on lines +307 to +310
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

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);

Comment on lines +303 to +310
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

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.


emitc::SubscriptOp subPtr = rewriter.create<emitc::SubscriptOp>(
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

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

loc, arrayValue, mlir::ValueRange(indices));
emitc::ApplyOp ptr = rewriter.create<emitc::ApplyOp>(
loc, emitc::PointerType::get(arrayValue.getType().getElementType()),
rewriter.getStringAttr("&"), subPtr);

return ptr;
};

auto srcPtr = createPointerFromEmitcArray(srcArrayValue);

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

rewriter.replaceOp(castOp, castCall);
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

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

return success();
}
};

struct ConvertStore final : public OpConversionPattern<memref::StoreOp> {
using OpConversionPattern::OpConversionPattern;

Expand Down Expand Up @@ -321,5 +380,6 @@ void mlir::populateMemRefToEmitCTypeConversion(TypeConverter &typeConverter) {
void mlir::populateMemRefToEmitCConversionPatterns(
RewritePatternSet &patterns, const TypeConverter &converter) {
patterns.add<ConvertAlloca, ConvertAlloc, ConvertGlobal, ConvertGetGlobal,
ConvertLoad, ConvertStore>(converter, patterns.getContext());
ConvertLoad, ConvertReinterpretCastOp, ConvertStore>(
converter, patterns.getContext());
}
27 changes: 25 additions & 2 deletions mlir/lib/Target/Cpp/TranslateToCpp.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -1756,6 +1756,20 @@ LogicalResult CppEmitter::emitOperation(Operation &op, bool trailingSemicolon) {

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

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

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

if (auto aType = dyn_cast<emitc::ArrayType>(pType.getPointee())) {
if (failed(emitType(loc, aType.getElementType())))
return failure();
os << " (*" << name << ")";
for (auto dim : aType.getShape())
os << "[" << dim << "]";
return success();
}
if (failed(emitType(loc, pType.getPointee())))
return failure();
os << " *" << name;
return success();
}
if (auto arrType = dyn_cast<emitc::ArrayType>(type)) {
if (failed(emitType(loc, arrType.getElementType())))
return failure();
Expand Down Expand Up @@ -1848,8 +1862,17 @@ LogicalResult CppEmitter::emitType(Location loc, Type type) {
if (auto lType = dyn_cast<emitc::LValueType>(type))
return emitType(loc, lType.getValueType());
if (auto pType = dyn_cast<emitc::PointerType>(type)) {
if (isa<ArrayType>(pType.getPointee()))
return emitError(loc, "cannot emit pointer to array type ") << type;
// Check if the pointee is an array type.
if (auto aType = dyn_cast<emitc::ArrayType>(pType.getPointee())) {
// Handle pointer to array: `element_type (*)[dim]`.
if (failed(emitType(loc, aType.getElementType())))
return failure();
os << "(*)";
for (auto dim : aType.getShape())
os << "[" << dim << "]";
return success();
}
// Handle standard pointer: `element_type*`.
if (failed(emitType(loc, pType.getPointee())))
return failure();
os << "*";
Expand Down
Original file line number Diff line number Diff line change
@@ -0,0 +1,16 @@
// RUN: mlir-opt -convert-memref-to-emitc %s -split-input-file | FileCheck %s
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

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.


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
}

//CHECK: module {
//CHECK-NEXT: func.func @casting(%arg0: memref<999xi32>) {
//CHECK-NEXT: %0 = builtin.unrealized_conversion_cast %arg0 : memref<999xi32> to !emitc.array<999xi32>
//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>>
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

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?

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

+1, good to check.

//CHECK-NEXT: return

Loading