Skip to content
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.

Sign up for GitHub

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

[mlir] ArithToLLVM: fix memref bitcast lowering #125148

Merged
merged 3 commits into from
Feb 12, 2025
Merged

[mlir] ArithToLLVM: fix memref bitcast lowering #125148

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
bdeccd9
[mlir] ArithToLLVM: fix memref bitcast lowering
Hardcode84 Jan 31, 2025
8c3d5c7
add comments
Hardcode84 Jan 31, 2025
01d05df
pointer support
Hardcode84 Feb 12, 2025
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
22 changes: 22 additions & 0 deletions mlir/lib/Conversion/ArithToLLVM/ArithToLLVM.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -54,6 +54,25 @@ struct ConstrainedVectorConvertToLLVMPattern
}
};

/// No-op bitcast. Propagate type input arg if converted source and dest types
/// are the same.
struct IdentityBitcastLowering final
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, to get the behavior you want here where identity bitcast on memref gets folded away before the general pattern for arith.bitcast kicks in, you want to set a PatternBenefit

: public OpConversionPattern<arith::BitcastOp> {
using OpConversionPattern::OpConversionPattern;

LogicalResult
matchAndRewrite(arith::BitcastOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const final {
Value src = adaptor.getIn();
Type resultType = getTypeConverter()->convertType(op.getType());
if (src.getType() != resultType)
return rewriter.notifyMatchFailure(op, "Types are different");

rewriter.replaceOp(op, src);
return success();
}
};

//===----------------------------------------------------------------------===//
// Straightforward Op Lowerings
//===----------------------------------------------------------------------===//
Expand Down Expand Up @@ -524,6 +543,9 @@ void mlir::arith::registerConvertArithToLLVMInterface(

void mlir::arith::populateArithToLLVMConversionPatterns(
const LLVMTypeConverter &converter, RewritePatternSet &patterns) {

patterns.add<IdentityBitcastLowering>(converter, patterns.getContext());

// clang-format off
patterns.add<
AddFOpLowering,
Expand Down
10 changes: 9 additions & 1 deletion mlir/lib/Conversion/LLVMCommon/VectorPattern.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -103,6 +103,14 @@ LogicalResult LLVM::detail::handleMultidimensionalVectors(
return success();
}

static bool isVectorCompatibleType(Type type) {
Copy link
Member

Choose a reason for hiding this comment

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

Can just LLVM::isCompatibleType be used here? It already checks for LLVMArrayType, VectorType, etc. Alternatively, there is also LLVM::isCompatibleVectorType, which may be useful here.

Copy link
Contributor Author

Choose a reason for hiding this comment

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

It was using LLVM::isCompatibleType before, but it's too broad, I specifically want to limit this transform to scalar and vector types.

Copy link
Contributor

Choose a reason for hiding this comment

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

Instead of fiddling, can you just set PatternBenefit on the bitcast pattern?

Copy link
Contributor Author

Choose a reason for hiding this comment

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

While PatternBenefit will probably work in this specific case, I still think there is a potential problem in vectorOneToOneRewrite as it can generate llvm bitcasts for unsupported types like structs.

// Limit `vectorOneToOneRewrite` to scalar and vector types (and to
// `LLVM::LLVMArrayType` which have a special handling).
return isa<LLVM::LLVMArrayType, LLVM::LLVMPointerType, VectorType,
IntegerType, FloatType>(type) &&
LLVM::isCompatibleType(type);
}

LogicalResult LLVM::detail::vectorOneToOneRewrite(
Operation *op, StringRef targetOp, ValueRange operands,
ArrayRef<NamedAttribute> targetAttrs,
Expand All @@ -111,7 +119,7 @@ LogicalResult LLVM::detail::vectorOneToOneRewrite(
assert(!operands.empty());

// Cannot convert ops if their operands are not of LLVM type.
if (!llvm::all_of(operands.getTypes(), isCompatibleType))
if (!llvm::all_of(operands.getTypes(), isVectorCompatibleType))
return failure();

auto llvmNDVectorTy = operands[0].getType();
Expand Down
12 changes: 12 additions & 0 deletions mlir/test/Conversion/ArithToLLVM/arith-to-llvm.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -727,3 +727,15 @@ func.func @ops_supporting_overflow(%arg0: i64, %arg1: i64) {
%3 = arith.shli %arg0, %arg1 overflow<nsw, nuw> : i64
return
}

// -----

// CHECK-LABEL: func @memref_bitcast
// CHECK-SAME: (%[[ARG:.*]]: memref<?xi16>)
// CHECK: %[[V1:.*]] = builtin.unrealized_conversion_cast %[[ARG]] : memref<?xi16> to !llvm.struct<(ptr, ptr, i64, array<1 x i64>, array<1 x i64>)>
// CHECK: %[[V2:.*]] = builtin.unrealized_conversion_cast %[[V1]] : !llvm.struct<(ptr, ptr, i64, array<1 x i64>, array<1 x i64>)> to memref<?xbf16>
// CHECK: return %[[V2]]
func.func @memref_bitcast(%1: memref<?xi16>) -> memref<?xbf16> {
%2 = arith.bitcast %1 : memref<?xi16> to memref<?xbf16>
func.return %2 : memref<?xbf16>
}