Custom Alias Analysis in LLVM

At Codeplay I currently work on a compiler backend for an embedded accelerator. The backend is the part of the compiler which takes some representation of the source code and translates it to machine code. In my case I’m working with LLVM, so this representation is LLVM IR.

It’s very common for accelerators like GPUs to have multiple regions of addressable memory – each with distinct properties. One important optimisation I’ve implemented recently is extending LLVM’s alias analysis functionality to handle the different address spaces for our target architecture.

This article will give an overview of the aim of alias analysis – along with an example based around address spaces – and show how custom alias analyses can be expressed in LLVM. You should be able to understand this article even if you don’t work with compilers or LLVM; hopefully it gives some insight into what compiler developers do to make the tools you use generate better code. LLVM developers may find the implementation section helpful, but may want to read the documentation or examples linked at the bottom for more details.

What is alias analysis

Alias analysis is the process of determining whether two pointers can point to the same object (alias) or not. This is very important for some valuable optimisations.

Take this C function as an example:

int foo (int __attribute__((address_space(0)))* a,
         int __attribute__((address_space(1)))* b) {
    *a = 42;
    *b = 20;
    return *a;
}

Those __attribute__s specify that a points to an int in address space 0, b points to an int in address space 1. An important detail of the target architecture for this code is that address spaces 0 and 1 are completely distinct: modifying memory in address space 0 can never affect memory in address space 1. Here’s some LLVM IR which could be generated from this function:

define i32 @foo(i32 addrspace(0)* %a, i32 addrspace(1)* %b) #0 {
entry:
  store i32 42, i32 addrspace(0)* %a, align 4
  store i32 20, i32 addrspace(1)* %b, align 4
  %0 = load i32, i32* %a, align 4
  ret i32 %0
}

For those unfamiliar with LLVM IR, the first store is storing 42 into *a, the second storing 20 into *b. The %0 = ... line is like loading *a into a temporary variable, which is then returned in the final line.

Optimising foo

Now we want foo to be optimised. Can you see an optimisation which could be made?

What we really want is for that load from a (the line beginning %0 = ...) to be removed and for the final statement to instead return 42. We want the optimised code to look like this:

define i32 @foo(i32 addrspace(0)* %a, i32 addrspace(1)* %b) #0 {
entry:
  store i32 42, i32 addrspace(0)* %a, align 4
  store i32 20, i32 addrspace(1)* %b, align 4
  ret i32 42
}

However, we have to be very careful, because this optimisation is only valid if a and b do not alias, i.e. they must not point at the same object. Forgetting about the address spaces for a second, consider this call to foo where we pass pointers which do alias:

int i = 0;
int result = foo(&i, &i);

Inside the unoptimised version of foo, i will be set to 42, then to 20, then 20 will be returned. However, if we carry out desired optimisation then the two stores will occur, but 42 will be returned instead of 20. We’ve just broken the behaviour of our function.

The only way that a compiler can reasonably carry out the above optimisation is if it can prove that the two pointers cannot possibly alias. This reasoning is carried out through alias analysis.

Custom alias analysis in LLVM

As I mentioned above, address spaces 0 and 1 for our target architecture are distinct. However, this may not hold for some systems, so LLVM cannot assume that it holds in general: we need to make it explicit.

One way to achieve this is llvm::AAResultBase. If our target is called TAR then we can create a class called TARAAResult which inherits from AAResultBase<TARAAResult>1:

class TARAAResult : public AAResultBase<TARAAResult> {
public:
  explicit TARAAResult() : AAResultBase() {}
  TARAAResult(TARAAResult &&Arg) : AAResultBase(std::move(Arg)) {}

  AliasResult alias(const MemoryLocation &LocA, const MemoryLocation &LocB);
};

The magic happens in the alias member function, which takes two MemoryLocations and returns an AliasResult. The result indicates whether the locations can never alias, may alias, partially alias, or precisely alias. We want our analysis to say “If the address spaces for the two memory locations are different, then they can never alias”. The resulting code is surprisingly close to this English description:

AliasResult TARAAResult::alias(const MemoryLocation &LocA,
                               const MemoryLocation &LocB) {
  auto AsA = LocA.Ptr->getType()->getPointerAddressSpace();
  auto AsB = LocB.Ptr->getType()->getPointerAddressSpace();

  if (AsA != AsB) {
    return NoAlias;
  }

  // Forward the query to the next analysis.
  return AAResultBase::alias(LocA, LocB);
}

Alongside this you need a bunch of boilerplate for creating a pass out of this analysis (I’ll link to a full example at the end), but after that’s done you just register the pass and ensure that the results of it are tracked:

void TARPassConfig::addIRPasses() {
  addPass(createTARAAWrapperPass());
  auto AnalysisCallback = [](Pass &P, Function &, AAResults &AAR) {
    if (auto *WrapperPass = P.getAnalysisIfAvailable<TARAAWrapper>()) {
      AAR.addAAResult(WrapperPass->getResult());
    }
  }; 
  addPass(createExternalAAWrapperPass(AliasCallback));
  TargetPassConfig::addIRPasses();
}

We also want to ensure that there is an optimisation pass which will remove unnecessary loads and stores which our new analysis will find. One example is Global Value Numbering.

  addPass(createNewGVNPass());

After all this is done, running the optimiser on the LLVM IR from the start of the post will eliminate the unnecessary load, and the generated code will be faster as a result.

You can see an example with all of the necessary boilerplate in LLVM’s test suite. The AMDGPU target has a full implementation which is essentially what I presented above extended for more address spaces.

Hopefully you have got a taste of what alias analysis does and the kinds of work involved in writing compilers, even if I have not gone into a whole lot of detail. I may write some more short posts on other areas of compiler development if readers find this interesting. Let me know on Twitter or in the comments!


  1. This pattern of inheriting from a class and passing the derived class as a template argument is known as the Curiously Recurring Template Pattern

Libc++ v5 PPA for Ubuntu now available

I am please to announce that I have made v5.0 of libc++ available for Ubuntu Linux via a PPA. Initially, the binaries are only available for Ubuntu 16.04 Xenial.

I was frustrated at the lack of an up-to-date build of libc++ for linux, especially with the v5.0 release of clang. Even though the LLVM project provide Ubuntu packages for clang, they don't provide one for libc++, and the version in the Ubuntu repositories is woefully out of date (v3.7). I therefore decided to package it myself, in this PPA.

This is especially good, since in combination with clang v5.0, it provides support for the Coroutines TS!

Enjoy!

Posted by Anthony Williams
[/ news /] permanent link
Tags: , , , ,

| Stumble It! stumbleupon logo | Submit to Reddit reddit logo | Submit to DZone dzone logo

Comment on this post

Follow me on Twitter