<!DOCTYPE html>
<html>
<head>
<meta http-equiv="Content-Type" content="text/xhtml; charset=utf-8">
</head>
<body>
<div style="font-family:sans-serif"><div style="white-space:normal">
<p dir="auto">On 9 Mar 2020, at 15:05, Chris Tetreault via llvm-dev wrote:</p>
</div>
<div style="white-space:normal"><blockquote style="border-left:2px solid #777; color:#777; margin:0 0 5px; padding-left:5px"><p dir="auto">Hi,<br>
I am helping with the effort to implement scalable vectors in the codebase in order to add support for generating SVE code in the Arm backend. I would like to propose a refactor of the Type class hierarchy in order to eliminate issues related to the misuse of SequentialType::getNumElements(). I would like to introduce a new class FixedVectorType that inherits from SequentialType and VectorType. VectorType would no longer inherit from SequentialType, instead directly inheriting from Type. After this change, it will be statically impossible to accidentally call SequentialType::getNumElements() via a VectorType pointer.</p>
</blockquote></div>
<div style="white-space:normal">
<p dir="auto">I’m sorry that I missed this thread when you posted it. I’m very much in favor of changing the type hierarchy to statically distinguish fixed from scalable vector types, but I think that making VectorType the common base type is unnecessarily disruptive. Practically speaking, this is going to break every out-of-tree frontend, backend, or optimization pass that supports SIMD types. Relatively little LLVM code will just naturally support scalable vector types without any adjustment. Following the principle of iterative development, as well as just good conservative coding practice, it’s much better for code that does support both to explicitly opt in by checking for and handling the more general type, rather than being implicitly “volunteered” to support both by having the <code style="background-color:#F7F7F7; border-radius:3px; margin:0; padding:0 0.4em" bgcolor="#F7F7F7">VectorType</code> type semantically repurposed out from under them.</p>
<p dir="auto">I understand the argument that <code style="background-color:#F7F7F7; border-radius:3px; margin:0; padding:0 0.4em" bgcolor="#F7F7F7">VectorType</code> is a better name for the abstract base type, but in this case I don’t think that consideration justifies the disruption for the vast majority of LLVM developers. There are plenty of names you could give the abstract base type that adequately express that’s a more general type, and the historical baggage of <code style="background-color:#F7F7F7; border-radius:3px; margin:0; padding:0 0.4em" bgcolor="#F7F7F7">VectorType</code> being slightly misleadingly named if you’re aware of a particular largely-vendor-specific extension does not seem overbearing.</p>
<p dir="auto">John.</p>
</div>
<div style="white-space:normal"><blockquote style="border-left:2px solid #777; color:#777; margin:0 0 5px; padding-left:5px"><p dir="auto">Background:<br>
Recently, scalable vectors have been introduced into the codebase. Previously, vectors have been written <n x ty> in IR, where n is a fixed number of elements known at compile time, and ty is some type. Scalable vectors are written <vscale x n x ty> where vscale is a runtime constant value. A new function has been added to VectorType (defined in llvm/IR/DerivedTypes.h), getElementCount(), that returns an ElementCount, which is defined as such in llvm/Support/TypeSize.h:<br>
class ElementCount {<br>
public:<br>
unsigned Min;<br>
bool Scalable;<br>
...<br>
}<br>
Min is the minimum number of elements in the vector (the "n" in <vscale x n x ty>), and Scalable is true if the vector is scalable (true for <vscale x n x ty>, false for <n x ty>) The idea is that if a vector is not scalable, then Min is exactly equal to the number of vector elements, but if the vector is scalable, then the number of vector elements is equal to some runtime-constant multiple of Min. The key takeaway here is that scalable vectors and fixed length vectors need to be treated differently by the compiler. For a fixed length vector, it is valid to iterate over the vector elements, but this is impossible for a scalable vector.<br>
Discussion:<br>
The trouble is that all instances of VectorType have getNumElements() inherited from SequentialType. Prior to the introduction of scalable vectors, this function was guaranteed to return the number of elements in a vector or array. Today, there is a comment that documents the fact that this returns only the minimum number of elements for scalable vectors, however there exists a ton of code in the codebase that is now misusing getNumElements(). Some examples:<br>
Auto *V = VectorType::get(Ty, SomeOtherVec->getNumElements());<br>
This code was previously perfectly fine but is incorrect for scalable vectors. When scalable vectors were introduced VectorType::get() was refactored to take a bool to tell if the vector is scalable. This bool has a default value of false. In this example, get() is returning a non-scalable vector even if SomeOtherVec was scalable. This will manifest later in some unrelated code as a type mismatch between a scalable and fixed length vector.<br>
for (unsigned I = 0; I < SomeVec->getNumElements(); ++I) { ... }<br>
Previously, since there was no notion of scalable vectors, this was perfectly reasonable code. However, for scalable vectors, this is always a bug.<br>
With vigilance in code review, and good test coverage we will eventually find and squash most of these bugs. Unfortunately, code review is hard, and test coverage isn't perfect. Bugs will continue to slip through as long as it's easier to do the wrong thing.<br>
One other factor to consider, is that there is a great deal of code which deals exclusively with fixed length vectors. Any backend for which there are no scalable vectors should not need to care about their existence. Even in Arm, if Neon code is being generated, then the vectors will never be scalable. In this code, the current status quo is perfectly fine, and any code related to checking if the vector is scalable is just noise.<br>
Proposal:<br>
In order to support users who only need fixed width vectors, and to ensure that nobody can accidentally call getNumElements() on a scalable vector, I am proposing the introduction of a new FixedVectorType which inherits from both VectorType and SequentialType. In turn, VectorType will no longer inherit from SequentialType. An example of what this will look like, with some misc. functions omitted for clarity:<br>
class VectorType : public Type {<br>
public:<br>
static VectorType *get(Type *ElementType, ElementCount EC);<br>
<br>
Type *getElementType() const;<br>
ElementCount getElementCount() const;<br>
bool isScalable() const;<br>
};<br>
<br>
class FixedVectorType : public VectorType, public SequentialType {<br>
public:<br>
static FixedVectorType *get(Type *ElementType, unsigned NumElts);<br>
};<br>
<br>
class SequentialType : public CompositeType {<br>
public:<br>
uint64_t getNumElements() const { return NumElements; }<br>
};<br>
In this proposed architecture, VectorType does not have a getNumElements() function because it does not inherit from SequentialType. In generic code, users will call VectorType::get() to obtain a new instance of VectorType just as they always have. VectorType implements the safe subset of functionality of fixed and scalable vectors that is suitable for use anywhere. If the user passes false to the scalable parameter of get(), they will get an instance of FixedVectorType back. Users can then inspect its type and cast it to FixedVectorType using the usual mechanisms. In code that deals exclusively in fixed length vectors, the user can call FixedVectorType::get() to directly get an instance of FixedVectorType, and their code can remain largely unchanged from how it was prior to the introduction of scalable vectors. At this time, there exists no use case that is only valid for scalable vectors, so no ScalableVectorType is being added.<br>
With this change, in generic code it is now impossible to accidentally call getNumElements() on a scalable vector. If a user tries to pass a scalable vector to a function that expects a fixed length vector, they will encounter a compilation failure at the site of the bug, rather than a runtime error in some unrelated code. If a user attempts to cast a scalable vector to FixedVectorType, the cast will fail at the call site. This will make it easier to track down all the places that are currently incorrect, and will prevent future developers from introducing bugs by misusing getNumElements().<br>
Outstanding design choice:<br>
One issue with this architecture as proposed is the fact that SequentialType (by way of CompositeType) inherits from Type. This introduces a diamond inheritance in FixedVectorType. Unfortunately, llvm::cast uses a c-style cast internally, so we cannot use virtual inheritance to resolve this issue. Thus, we have a few options:<br>
<br>
1. Break CompositeType's inheritance on Type and introduce functions to convert from a Type to a CompositeType and vice versa. The conversion from CompositeType is always safe because all instances of CompositeType (StructType, ArrayType, and FixedVectorType) are instances of Type. A CompositeType can be cast to the most derived class, then back to Type. The other way is not always safe, so a function will need to be added to check if a Type is an instance of CompositeType. This change is not that big, and I have a prototype implementation up at <a href="https://reviews.llvm.org/D75486" style="color:#777">https://reviews.llvm.org/D75486</a> ([SVE] Make CompositeType not inherit from Type)<br>
* Pros: this approach would result in minimal changes to the codebase. If the llvm casts can be made to work for the conversion functions, then it would touch very few files.<br>
* Cons: There are those who think that CompositeType adds little value and should be removed. Now would be an ideal time to do this. Additionally, the conversion functions would be more complicated if we left CompositeType in.<br>
2. Remove CompositeType and break SequentialType's inheritance of Type. Add functions to convert a SequentialType to and from Type. The conversion functions would work the same as those in option 1 above. Currently, there exists only one class that derives directly from CompositeType: StructType. The functionality of CompositeType can be directly moved into StructType, and APIs that use CompositeType can directly use Type and cast appropriately. We feel that this would be a fairly simple change, and we have a prototype implementation up at <a href="https://reviews.llvm.org/D75660" style="color:#777">https://reviews.llvm.org/D75660</a> (Remove CompositeType class)<br>
* Pros: Removing CompositeType would simplify the type hierarchy. Leaving SequentialType in would simplify some code and be more typesafe than having a getSequentialNumElements on Type.<br>
* Cons: The value of SequentialType has also been called into question. If we wanted to remove it, now would be a good time. Conversion functions add complexity to the design. Introduces additional casting from Type.<br>
3. Remove CompositeType and SequentialType. Roll the functions directly into the most derived classes. A helper function can be added to Type to handle choosing from FixedVectorType and ArrayType and calling getNumElements():<br>
static unsigned getSequentialNumElements() {<br>
assert(isSequentialType()); // This already exists and does the<br>
// right thing<br>
if (auto *AT = dyn_cast<ArrayType>(this))<br>
return AT->getNumElements();<br>
return cast<FixedVectorType>(this)->getNumElements();<br>
}<br>
A prototype implementation of this strategy can be found at <a href="https://reviews.llvm.org/D75661" style="color:#777">https://reviews.llvm.org/D75661</a> (Remove SequentialType from the type heirarchy.)<br>
<br>
* Pros: By removing the multiple inheritance completely, we greatly simplify the design and eliminate the need for any conversion functions. The value of CompositeType and SequentialType has been called into question, and removing them now might be of benefit to the codebase<br>
* Cons: getSequentialNumElements() has similar issues to those that we are trying to solve in the first place and potentially subverts the whole design. Omitting getSequentialNumElements() would add lots of code duplication. Introduces additional casting from Type.<br>
I believe that all three of these options are reasonable. My personal preference is currently option 2. I think that option 3's getSequentialNumElements() subverts the design because every Type has getSequentialNumElements(), it is tempting to just call it. However, the cast will fail at the call site in debug, and in release it will return a garbage value rather than a value that works most of the time. For option 1, the existence of CompositeType complicates the conversion logic for little benefit.<br>
Conclusion:<br>
Thank you for your time in reviewing this RFC. Your feedback on my work is greatly appreciated.<br>
<br>
Thank you,<br>
Christopher Tetreault</p>
</blockquote><br><blockquote style="border-left:2px solid #777; color:#777; margin:0 0 5px; padding-left:5px"><p dir="auto">_______________________________________________<br>
LLVM Developers mailing list<br>
llvm-dev@lists.llvm.org<br>
<a href="https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev" style="color:#777">https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev</a></p>
</blockquote></div>
<div style="white-space:normal">
</div>
</div>
</body>
</html>