[mlir][gpu] `gpu.rotate` verifier shouldn't check for constant operands

[kuhar]

Op verifiers are not supposed to check non-local properties in ::verify(): https://mlir.llvm.org/getting_started/DeveloperGuide/#ir-verifier

We should remove this code from the verifier for gpu.rotate:

llvm-project/mlir/lib/Dialect/GPU/IR/GPUDialect.cpp

Lines 1360 to 1384 in 0f48baf

	auto offsetConstOp = getOffset().getDefiningOp<arith::ConstantOp>();
	if (!offsetConstOp)
	return emitOpError() << "offset is not a constant value";
	auto offsetIntAttr =
	llvm::dyn_cast<mlir::IntegerAttr>(offsetConstOp.getValue());
	auto widthConstOp = getWidth().getDefiningOp<arith::ConstantOp>();
	if (!widthConstOp)
	return emitOpError() << "width is not a constant value";
	auto widthIntAttr =
	llvm::dyn_cast<mlir::IntegerAttr>(widthConstOp.getValue());
	llvm::APInt offsetValue = offsetIntAttr.getValue();
	llvm::APInt widthValue = widthIntAttr.getValue();
	if (!widthValue.isPowerOf2())
	return emitOpError() << "width must be a power of two";
	if (offsetValue.sge(widthValue) || offsetValue.slt(0)) {
	int64_t widthValueInt = widthValue.getSExtValue();
	return emitOpError() << "offset must be in the range [0, " << widthValueInt
	<< ")";
	}

An alternative would be to take width as an attribute, if we want to have it verified.

[llvmbot]

@llvm/issue-subscribers-mlir-gpu

Author: Jakub Kuderski (kuhar)

Op verifiers are not supposed to check non-local properties in `::verify()`: https://mlir.llvm.org/getting_started/DeveloperGuide/#ir-verifier

We should remove this code from the verifier for gpu.rotate:

llvm-project/mlir/lib/Dialect/GPU/IR/GPUDialect.cpp

Lines 1360 to 1384 in 0f48baf

	auto offsetConstOp = getOffset().getDefiningOp<arith::ConstantOp>();
	if (!offsetConstOp)
	return emitOpError() << "offset is not a constant value";
	auto offsetIntAttr =
	llvm::dyn_cast<mlir::IntegerAttr>(offsetConstOp.getValue());
	auto widthConstOp = getWidth().getDefiningOp<arith::ConstantOp>();
	if (!widthConstOp)
	return emitOpError() << "width is not a constant value";
	auto widthIntAttr =
	llvm::dyn_cast<mlir::IntegerAttr>(widthConstOp.getValue());
	llvm::APInt offsetValue = offsetIntAttr.getValue();
	llvm::APInt widthValue = widthIntAttr.getValue();
	if (!widthValue.isPowerOf2())
	return emitOpError() << "width must be a power of two";
	if (offsetValue.sge(widthValue) || offsetValue.slt(0)) {
	int64_t widthValueInt = widthValue.getSExtValue();
	return emitOpError() << "offset must be in the range [0, " << widthValueInt
	<< ")";
	}

An alternative would be to take width as an attribute, if we want to have it verified.

[kuhar]

@Hsiangkai I missed this during the code review for #142796