add dexts operation

[rodonisi]

• the supported types are integers, vectors and signals
• the verifier checks for type conversions and result length not being larger than the input

[maerhart]

You might want to consider to add additional class declarations to extract the slice size easily when working with this operation e.g. in passes. Something like this:

let extraClassDeclaration = [{
        unsigned getSliceWidth() {
            if (result().getType().isSignlessInteger()) {
                return result().getType().getIntOrFloatBitWidth();
            } else if (auto sigRes = result().getType().dyn_cast<llhd::SigType>()) {
               return sigRes.getUnderlyingType().getIntOrFloatBitWidth(); 
            } else if (auto vecRes = result().getType().dyn_cast<VectorType>()) {
                vecRes.getNumElements();
            }
            return 0;
        }

        unsigned getTargetWidth() {
            if (target().getType().isSignlessInteger()) {
                return target().getType().getIntOrFloatBitWidth();
            } else if (auto sigRes = target().getType().dyn_cast<llhd::SigType>()) {
               return sigRes.getUnderlyingType().getIntOrFloatBitWidth(); 
            } else if (auto vecRes = target().getType().dyn_cast<VectorType>()) {
                vecRes.getNumElements();
            }
            return 0;
        }
    }];

Then you can also use these methods to get the widths in the verify method or maybe consider traits like this:

def LLHD_DextsOp : LLHD_Op<"dexts",
        [NoSideEffect,
         PredOpTrait<"the result width cannot be larger than the target width",
            CPred<"this->getTargetWidth() <= this->getSliceWidth()">>,
         PredOpTrait<"the target and result kinds have to match",
            CPred<"$target.getType().getKind() == $result.getType().getKind()">>,
         PredOpTrait<"vector element types do not match",
            Or<[CPred<"!$result.getType().isa<VectorType>()">, TCresVTEtIsSameAsOp<0,0>]>>
        ]> { ... }

[jmgorius]

You might want to consider to add additional class declarations to extract the slice size easily when working with this operation e.g. in passes. Something like this:

let extraClassDeclaration = [{
        unsigned getSliceWidth() {
            if (result().getType().isSignlessInteger()) {
                return result().getType().getIntOrFloatBitWidth();
            } else if (auto sigRes = result().getType().dyn_cast<llhd::SigType>()) {
               return sigRes.getUnderlyingType().getIntOrFloatBitWidth(); 
            } else if (auto vecRes = result().getType().dyn_cast<VectorType>()) {
                vecRes.getNumElements();
            }
            return 0;
        }

        unsigned getTargetWidth() {
            if (target().getType().isSignlessInteger()) {
                return target().getType().getIntOrFloatBitWidth();
            } else if (auto sigRes = target().getType().dyn_cast<llhd::SigType>()) {
               return sigRes.getUnderlyingType().getIntOrFloatBitWidth(); 
            } else if (auto vecRes = target().getType().dyn_cast<VectorType>()) {
                vecRes.getNumElements();
            }
            return 0;
        }
    }];

Then you can also use these methods to get the widths in the verify method or maybe consider traits like this:

def LLHD_DextsOp : LLHD_Op<"dexts",
        [NoSideEffect,
         PredOpTrait<"the result width cannot be larger than the target width",
            CPred<"this->getTargetWidth() <= this->getSliceWidth()">>,
         PredOpTrait<"the target and result kinds have to match",
            CPred<"$target.getType().getKind() == $result.getType().getKind()">>,
         PredOpTrait<"vector element types do not match",
            Or<[CPred<"!$result.getType().isa<VectorType>()">, TCresVTEtIsSameAsOp<0,0>]>>
        ]> { ... }

Just nitpicking here: if you go for such an approach, I would strongly recommend you assign result.getType() and target.getType() to a local to avoid potentially redundant and deep function calls.

[maerhart]

Right, that would be a nice optimization.

[rodonisi]

Thanks for the feedback!
I moved the verification to predicate traits as suggested. The TCresVTEtIsSameAsOp though seems only to work if the checked types are explicitly VectorTypes, and I couldn't see a way have general Types casted for that check (correct me if I'm wrong). Instead I added two additional class declarations that return the element type if the type is a vector, or return null. I think that this could also come handy in other situations.

[maerhart]

Thanks for the feedback!
I moved the verification to predicate traits as suggested. The TCresVTEtIsSameAsOp though seems only to work if the checked types are explicitly VectorTypes, and I couldn't see a way have general Types casted for that check (correct me if I'm wrong). Instead I added two additional class declarations that return the element type if the type is a vector, or return null. I think that this could also come handy in other situations.

Thanks for adding the extra class declarations!

TCresVTEtIsSameAsOp uses the getElementTypeOrSelf method defined in TypeUtilities.h. You need to add #include "mlir/IR/TypeUtilities.h" to LLHDOps.cpp to make it work.

• I added a trait that checks the types in one go and can also be used for the llhd.inss instruction. The overhead should be smaller than in the current implementation.
• I deleted the vectorElementTypeOrNull methods because I think they are not of much use outside of getting the trait to work. Because if you just want to get the element type if it's a vector you can call target().getType().cast<VectorType>() directly if you already know it's a vector and otherwise you need to either check that it is a vector type or check that the extra class declaration didn't return NULL which makes no difference. But if you have a use-case for it, I can add it again.

[rodonisi]

Thanks for the feedback!
I moved the verification to predicate traits as suggested. The TCresVTEtIsSameAsOp though seems only to work if the checked types are explicitly VectorTypes, and I couldn't see a way have general Types casted for that check (correct me if I'm wrong). Instead I added two additional class declarations that return the element type if the type is a vector, or return null. I think that this could also come handy in other situations.

Thanks for adding the extra class declarations!

TCresVTEtIsSameAsOp uses the getElementTypeOrSelf method defined in TypeUtilities.h. You need to add #include "mlir/IR/TypeUtilities.h" to LLHDOps.cpp to make it work.

• I added a trait that checks the types in one go and can also be used for the llhd.inss instruction. The overhead should be smaller than in the current implementation.
• I deleted the vectorElementTypeOrNull methods because I think they are not of much use outside of getting the trait to work. Because if you just want to get the element type if it's a vector you can call target().getType().cast<VectorType>() directly if you already know it's a vector and otherwise you need to either check that it is a vector type or check that the extra class declaration didn't return NULL which makes no difference. But if you have a use-case for it, I can add it again.

Nice addition, thank you!
I was thinking at use cases where you would do something specific with the element type if the extraction is on vectors, I have no specific example for that though. The idea was to enable something like

if (auto elemTy = dexts. getSliceVectorElementTypeOrNull()) {
    ...
}

which would be simpler than checking the cast though not by much.
All in all it was mostly a solution to make it work and that could maybe become handy later, but not necessarily. I think we can leave it out.

[rodonisi]

Also thank you for clarifying about the getElementTypeOrSelf, at first I thought that to be specific to VectorType