llvm-project/mlir/docs/Bufferization.md

Bufferization on MLIR

The general task of bufferization is to move SSA values (like tensors) into allocated memory buffers that have to be freed when they are no longer needed. This also involves the placement of copies to clone contents of allocated memory buffers at specific locations (similar to register allocation). On the one hand, these copies are needed to ensure the right behavior of a program, on the other hand, introducing several aliases for a certain buffer could lead to a wrong freeing of buffers. Therefore, we have to take care of them and the program structure. The introduction of copies solves this problem. Several unnecessary introduced copies during this process can be eliminated afterwards.

func @func_on_tensors(%source: tensor<4xf32>) -> tensor<4xf32> {
  %0 = mhlo.exp %source : (tensor<4xf32>) -> (tensor<4xf32>)
  return %0 : tensor<4xf32>
}

Will be transformed to:

func @func_on_buffers(%source: memref<4xf32>, %target: memref<4xf32>) {
  %0 = alloc() : memref<4xf32>
  lmhlo.exp %source, %0 : (memref<4xf32>, memref<4xf32>) -> ()
  lmhlo.copy %0, %target : (memref<4xf32>, memref<4xf32>) -> ()
  dealloc %0 : memref<4xf32>
  return
}

In general, Bufferization is split into three separate phases: a preparation phase, a bufferization phase and a post-processing phase. The assignment process happens during dialect conversion and allocates buffers for each value that should be moved into a memory buffer. This has to be implemented by each dialect using the following tools and patterns. Thereby, all operations on memory buffers have to be changed to memref<T> types (see Preparation Phase). Afterwards, the placement transformation (see BufferDeallocation) adds all required deallocation operations, temporary buffers and copy operations automatically.

Preparation Phase

In order to apply the BufferDeallocation code transformation, the input MLIR program needs to leverage allocations (buffers in general) and memref<T> types(as outlined above). If your input program does not work on buffers, you need to perform this preparation step in order to port it to the “buffer world”. This is a user-defined preparation step that is intended to be applied during dialect conversion. The user has to take care for the right conversion by providing conversion patterns relying on a type converter to assign buffers.

A necessary step is to apply type and function signature conversion. Furthermore, after changing all function signatures, the associated return and call operations must comply with the new corresponding function signatures. For this purpose, three operation conversion patterns are introduced:

• BufferizeFuncOpConverter
• BufferizeReturnOpConverter
• BufferizeCallOpConverter

In order to use these conversion patterns, the user needs to define a custom BufferizeTypeConverter implementation.

BufferizeTypeConverter

The BufferizeTypeConverter is an extension to the TypeConverter class that provides additional functionality for dialect developers to decide on the signature of a function. The extra features include:

• setResultConversionKind to decide on a function result after the conversion with a specific type to be appended to the function argument list as an output argument or remains as a function result.
• define their own callback function for type or value unwrapping.

ResultConversionKind

ResultConversionKind is an enum with two values

• AppendToArgumentList
• KeepAsFunctionResult

that defines how BufferizeFuncOpConverter should handle function results in order to convert the signature of the function. The other two operation conversion patterns also use ResultConversionKind to adapt themselves with the new function signature.

ResultConversionKind can be set using setResultConversionKind, which needs two template parameters that correspond to the types before and after type conversion. This mapping specifies whether the resulting type should stay as a function result or should be appended to the arguments list after the conversion is done. Note that the default value for unspecified mappings is KeepAsFunctionResult. For instance, the following command updates the BufferizeTypeConverter instance that defines all MemRefType function results (converted from RankedTensorTypes). These results should be appended to the function argument list in BufferizeFuncOpConverter:

converter.setResultConversionKind<RankedTensorType, MemRefType>(
        BufferizeTypeConverter::AppendToArgumentsList);

Callbacks for Unpacking Types

func @func_on_tensors(%arg0: tuple<i1,f32>) -> (tuple<tensor<10xf32>, tensor<5xf32>>) {
  ...
}

Will be transformed to:

func @func_on_buffers(%arg0: i1, %arg1: f32, %target0: memref<10xf32>, %target1: memref<5xf32>) {
  ...
}

BufferizeFuncOpConverter is also able to unpack the types of arguments and results of a function during function signature conversion. In the example above, it unwraps the tuple type and converts the type of each constituent.

For this purpose, users can provide custom callbacks by using addDecomposeTypeConversion for the BufferizeTypeConverter instance to show how a specific type (i.e. TupleType) can be unpacked. However, when a type decomposition is provided, there are two additional callbacks that have to be defined as well. They specify how to pack or unpack values of that particular type. These two callbacks can be provided by the addArgumentMaterialization (packing) and addDecomposeValueConversion (unpacking) functions:

The following piece of code demonstrates this functionality by flattening a TupleType.

converter.addDecomposeTypeConversion(
        [](TupleType tupleType, SmallVectorImpl<Type> &types) {
          tupleType.getFlattenedTypes(types);
          return success();
        });

converter.addArgumentMaterialization(
        [](OpBuilder &builder, TupleType resultType, ValueRange inputs,
           Location loc) -> Optional<Value> {
          if (inputs.size() == 1)
            return llvm::None;
          TypeRange TypeRange = inputs.getTypes();
          SmallVector<Type, 2> types(TypeRange.begin(), TypeRange.end());
          TupleType tuple = TupleType::get(types, builder.getContext());
          mlir::Value value = builder.create<MakeTupleOp>(loc, tuple, inputs);
          return value;
        });

converter.addDecomposeValueConversion([](OpBuilder &builder, Location loc,
                                         TupleType resultType, Value value,
                                         SmallVectorImpl<Value> &values) {
      for (unsigned i = 0, e = resultType.size(); i < e; ++i) {
        Value res = builder.create<GetTupleElementOp>(
            loc, resultType.getType(i), value, builder.getI32IntegerAttr(i));
        values.push_back(res);
      }
      return success();
 });

In the scope of these callback functions, the elements of a tuple value can be decomposed using GetTupleElementOp. Conversely, MakeTupleOp is used to pack a list of values as a single tuple type.

Bufferization Operation Conversion Patterns

The following conversion patterns can be used to conveniently transform the signature of a function, the return and call operations:

• BufferizeFuncOpConverter
• BufferizeReturnOpConverter
• BufferizeCallOpConverter

Any combination of these conversion patterns can be specified by the user. If you need to apply all of these operation converters, you can use populateWithBufferizeOpConversionPatterns which sets up all converters.

BufferizeFuncOpConverter

The BufferizeFuncOpConverter is the actual function operation converter that applies signature conversion by using a previously defined BufferizeTypeConverter.

In the following example, we configure a BufferizeTypeConverter instance such that

• all RankedTensorTypes should be converted to MemRefTypes.
• all function results that are results of type conversion from RankedTensorTypes to MemRefTypes should be appended to the function argument list.
• all TupleTypes should be flattened and decomposed to its constituents.

converter.addConversion([](RankedTensorType type) {
    return (Type)MemRefType::get(type.getShape(), type.getElementType());
  });
converter.setResultConversionKind<RankedTensorType, MemRefType>(
        BufferizeTypeConverter::AppendToArgumentsList);
converter.addDecomposeTypeConversion(
        [](TupleType tupleType, SmallVectorImpl<Type> &types) {
          tupleType.getFlattenedTypes(types);
          return success();
        });

Consider the following signature conversion:

func @on_tensors(%arg1: tuple<i1,f32>) -> (tuple<memref<10xf32>, tensor<5xf32>>){
 ...
}

Will be transformed to:

func @on_buffers(%arg0: i1, %arg1: f32, %out: memref<5xf32>) -> memref<10xf32> {
 ...
}

Using the presented converter setup, all TupleType arguments and results are decomposed first. The tensor<5xf32> result is converted to a memref<5xf32> type and appended to the argument list. There is no conversion for the types memref, i1, and f32. Therefore, the memref<10xf32> result is kept as it is and will also be kept as a function result since there is no ResultConversionKind mapping from a MemRefType to a MemRefType. However, if we change the result-conversion behavior via

converter.setResultConversionKind<RankedTensorType, MemRefType>(
        BufferizeTypeConverter::KeepAsFunctionResult);

the output will be:

func @on_buffers(%arg0: i1, %arg1: f32) -> (memref<10xf32>, memref<5xf32>) {
 ...
}

BufferizeReturnOpConverter

When changing the signature of a function, the return operands must match with the results of the corresponding function if buffer-typed-results have been configured to be appended to the function arguments list. This matching consists of two separate steps. First, we have to remove the operands that have been appended to the argument list as output arguments. Second, we have to introduce additional copies for each operand. However, since each dialect has its own dialect-dependent return and copy operations, this conversion pattern comes with three template parameters which are the original return operation, target return operation, and copy operation kinds.

In the following example, two conversion patterns are inserted into the pattern list. The BufferizeReturnOpConverter is set to replace a standard return operation with the same operation type.

patterns->insert<
  BufferizeFuncOpConverter,
  BufferizeReturnOpConverter
    <mlir::ReturnOp, mlir::ReturnOp, linalg::CopyOp>
                >(...)

Consider the following input/output program using a single return:

func @on_tensors(%arg0: tensor<5xf32>, %arg1: i1) -> (tensor<5xf32>, i1) {
  return %arg0, %arg1 : tensor<5xf32>, i1
}

Will be transformed to:

func @on_buffers(%arg0: memref<5xf32>, %arg1: i1, %out: memref<5xf32>) -> i1 {
  linalg.copy(%arg0, %out) : memref<5xf32>, memref<5xf32>
  return %arg1 : i1
}

Based on our previously configured BufferizeTypeConverter instance which requires buffer-typed-function-results to be appended to the function argument list, the new on_buffers function signature is created. The operands of the return operation must be adapted with the new function signature. Therefore, the buffer-typed operand is removed from the operand list of the new return operation. Instead, a copy operation is inserted right before the return operation to copy the content of the operand buffer to the target buffer and yields the output as shown above.

BufferizeCallOpConverter

The BufferizeCallOpConverter is a call operation converter that transforms and matches the operands and results of a call operation with the arguments and results of the callee. Besides converting operand and result types, it allocates a buffer for each buffer-based result of the called function that is appended to the argument list (if buffer typed results have been configured to be appended to the function arguments list).

The following piece of code shows a sample call site, based on our previously configured BufferizeTypeConversion:

func @callee(%arg0: tensor<5xf32>) -> (tensor<5xf32>) {
  return %arg0 : tensor<5xf32>
}

func @caller(%arg0: tensor<5xf32>) -> tensor<5xf32> {
  %x = call @callee(%arg0) : (tensor<5xf32>) -> tensor<5xf32>
  return %x : tensor<5xf32>
}

Will be transformed to:

func @callee(%arg0: memref<5xf32>, %out: memref<5xf32>) {
  linalg.copy(%arg0, %out) : memref<5xf32>, memref<5xf32>
  return
}

func @caller(%arg0: memref<5xf32>, %out: memref<5xf32>) {
  %0 = alloc() : memref<5xf32>
  call @callee(%arg0, %0) : (memref<5xf32>, memref<5xf32>) -> ()
  linalg.copy(%0, %out) : memref<5xf32>, memref<5xf32>
  return
}

Summarizing Example

To summarize all preparation converters, the following sample is a complete listing of an input IR program and its output after applying all converters:

func @callee(%arg0: tuple<tensor<5xf32>,i1>) -> tuple<tensor<5xf32>,i1> {
  return %arg0 : tuple<tensor<5xf32>,i1>
}

func @caller(%arg0: tuple<tensor<5xf32>,i1>) -> tuple<tensor<5xf32>,i1> {
  %x = call @callee(%arg0) : (tuple<tensor<5xf32>,i1>) -> tuple<tensor<5xf32>,i1>
  return %x : tuple<tensor<5xf32>,i1>
}

Will be transformed to:

func @callee(%arg0: memref<5xf32>, %arg1: i1, %arg2: memref<5xf32>) -> i1 {
  %0 = "test.make_tuple"(%arg0, %arg1) : (memref<5xf32>, i1) -> tuple<memref<5xf32>, i1>
  %1 = "test.get_tuple_element"(%0) {index = 0 : i32} : (tuple<memref<5xf32>, i1>) -> memref<5xf32>
  %2 = "test.get_tuple_element"(%0) {index = 1 : i32} : (tuple<memref<5xf32>, i1>) -> i1
  linalg.copy(%1, %arg2) : memref<5xf32>, memref<5xf32>
  return %2 : i1
}
func @caller(%arg0: memref<5xf32>, %arg1: i1, %arg2: memref<5xf32>) -> i1 {
  %0 = "test.make_tuple"(%arg0, %arg1) : (memref<5xf32>, i1) -> tuple<memref<5xf32>, i1>
  %1 = "test.get_tuple_element"(%0) {index = 0 : i32} : (tuple<memref<5xf32>, i1>) -> memref<5xf32>
  %2 = "test.get_tuple_element"(%0) {index = 1 : i32} : (tuple<memref<5xf32>, i1>) -> i1
  %3 = alloc() : memref<5xf32>
  %4 = call @callee(%1, %2, %3) : (memref<5xf32>, i1, memref<5xf32>) -> i1
  %5 = "test.make_tuple"(%3, %4) : (memref<5xf32>, i1) -> tuple<memref<5xf32>, i1>
  %6 = "test.get_tuple_element"(%5) {index = 0 : i32} : (tuple<memref<5xf32>, i1>) -> memref<5xf32>
  %7 = "test.get_tuple_element"(%5) {index = 1 : i32} : (tuple<memref<5xf32>, i1>) -> i1
  linalg.copy(%6, %arg2) : memref<5xf32>, memref<5xf32>
  return %7 : i1
}

Buffer Deallocation - Internal Functionality

This section covers the internal functionality of the BufferDeallocation transformation. The transformation consists of several passes. The main pass called BufferDeallocation can be applied via “-buffer-deallocation” on MLIR programs. Currently, there are three optimization passes, that move allocs and convert AllocOps to AllocaOps, if possible. The first and second pass can be applied using “-buffer-hoisting” or “-buffer-loop-hoisting”, the third one using “-promote-buffers-to-stack”. However, these optimizations must be applied before using the BufferDeallocation pass.

Requirements

In order to use BufferDeallocation on an arbitrary dialect, several control-flow interfaces have to be implemented when using custom operations. This is particularly important to understand the implicit control-flow dependencies between different parts of the input program. Without implementing the following interfaces, control-flow relations cannot be discovered properly and the resulting program can become invalid:

• Branch-like terminators should implement the BranchOpInterface to query and manipulate associated operands.
• Operations involving structured control flow have to implement the RegionBranchOpInterface to model inter-region control flow.
• Terminators yielding values to their parent operation (in particular in the scope of nested regions within RegionBranchOpInterface-based operations), should implement the ReturnLike trait to represent logical “value returns”.

Example dialects that are fully compatible are the “std” and “scf” dialects with respect to all implemented interfaces.

Detection of Buffer Allocations

The first step of the BufferDeallocation transformation is to identify manageable allocation operations that implement the SideEffects interface. Furthermore, these ops need to apply the effect MemoryEffects::Allocate to a particular result value while not using the resource SideEffects::AutomaticAllocationScopeResource (since it is currently reserved for allocations, like Alloca that will be automatically deallocated by a parent scope). Allocations that have not been detected in this phase will not be tracked internally, and thus, not deallocated automatically. However, BufferDeallocation is fully compatible with “hybrid” setups in which tracked and untracked allocations are mixed:

func @mixedAllocation(%arg0: i1) {
   %0 = alloca() : memref<2xf32>  // aliases: %2
   %1 = alloc() : memref<2xf32>  // aliases: %2
   cond_br %arg0, ^bb1, ^bb2
^bb1:
  use(%0)
  br ^bb3(%0 : memref<2xf32>)
^bb2:
  use(%1)
  br ^bb3(%1 : memref<2xf32>)
^bb3(%2: memref<2xf32>):
  ...
}

Example of using a conditional branch with alloc and alloca. BufferDeallocation can detect and handle the different allocation types that might be intermixed.

Note: the current version does not support allocation operations returning multiple result buffers.

Conversion from AllocOp to AllocaOp

The PromoteBuffersToStack-pass converts AllocOps to AllocaOps, if possible. In some cases, it can be useful to use such stack-based buffers instead of heap-based buffers. The conversion is restricted to several constraints like:

• Control flow
• Buffer Size
• Dynamic Size

If a buffer is leaving a block, we are not allowed to convert it into an alloca. If the size of the buffer is large, we could convert it, but regarding stack overflow, it makes sense to limit the size of these buffers and only convert small ones. The size can be set via a pass option. The current default value is 1KB. Furthermore, we can not convert buffers with dynamic size, since the dimension is not known a priori.

Movement and Placement of Allocations

Using the buffer hoisting pass, all buffer allocations are moved as far upwards as possible in order to group them and make upcoming optimizations easier by limiting the search space. Such a movement is shown in the following graphs. In addition, we are able to statically free an alloc, if we move it into a dominator of all of its uses. This simplifies further optimizations (e.g. buffer fusion) in the future. However, movement of allocations is limited by external data dependencies (in particular in the case of allocations of dynamically shaped types). Furthermore, allocations can be moved out of nested regions, if necessary. In order to move allocations to valid locations with respect to their uses only, we leverage Liveness information.

The following code snippets shows a conditional branch before running the BufferHoisting pass:

func @condBranch(%arg0: i1, %arg1: memref<2xf32>, %arg2: memref<2xf32>) {
  cond_br %arg0, ^bb1, ^bb2
^bb1:
  br ^bb3(%arg1 : memref<2xf32>)
^bb2:
  %0 = alloc() : memref<2xf32>  // aliases: %1
  use(%0)
  br ^bb3(%0 : memref<2xf32>)
^bb3(%1: memref<2xf32>):  // %1 could be %0 or %arg1
  "linalg.copy"(%1, %arg2) : (memref<2xf32>, memref<2xf32>) -> ()
  return
}

Applying the BufferHoisting pass on this program results in the following piece of code:

func @condBranch(%arg0: i1, %arg1: memref<2xf32>, %arg2: memref<2xf32>) {
  %0 = alloc() : memref<2xf32>  // moved to bb0
  cond_br %arg0, ^bb1, ^bb2
^bb1:
  br ^bb3(%arg1 : memref<2xf32>)
^bb2:
   use(%0)
   br ^bb3(%0 : memref<2xf32>)
^bb3(%1: memref<2xf32>):
  "linalg.copy"(%1, %arg2) : (memref<2xf32>, memref<2xf32>) -> ()
  return
}

The alloc is moved from bb2 to the beginning and it is passed as an argument to bb3.

The following example demonstrates an allocation using dynamically shaped types. Due to the data dependency of the allocation to %0, we cannot move the allocation out of bb2 in this case:

func @condBranchDynamicType(
  %arg0: i1,
  %arg1: memref<?xf32>,
  %arg2: memref<?xf32>,
  %arg3: index) {
  cond_br %arg0, ^bb1, ^bb2(%arg3: index)
^bb1:
  br ^bb3(%arg1 : memref<?xf32>)
^bb2(%0: index):
  %1 = alloc(%0) : memref<?xf32>   // cannot be moved upwards to the data
                                   // dependency to %0
  use(%1)
  br ^bb3(%1 : memref<?xf32>)
^bb3(%2: memref<?xf32>):
  "linalg.copy"(%2, %arg2) : (memref<?xf32>, memref<?xf32>) -> ()
  return
}

Introduction of Copies

In order to guarantee that all allocated buffers are freed properly, we have to pay attention to the control flow and all potential aliases a buffer allocation can have. Since not all allocations can be safely freed with respect to their aliases (see the following code snippet), it is often required to introduce copies to eliminate them. Consider the following example in which the allocations have already been placed:

func @branch(%arg0: i1) {
  %0 = alloc() : memref<2xf32>  // aliases: %2
  cond_br %arg0, ^bb1, ^bb2
^bb1:
  %1 = alloc() : memref<2xf32>  // resides here for demonstration purposes
                                // aliases: %2
  br ^bb3(%1 : memref<2xf32>)
^bb2:
  use(%0)
  br ^bb3(%0 : memref<2xf32>)
^bb3(%2: memref<2xf32>):
  …
  return
}

The first alloc can be safely freed after the live range of its post-dominator block (bb3). The alloc in bb1 has an alias %2 in bb3 that also keeps this buffer alive until the end of bb3. Since we cannot determine the actual branches that will be taken at runtime, we have to ensure that all buffers are freed correctly in bb3 regardless of the branches we will take to reach the exit block. This makes it necessary to introduce a copy for %2, which allows us to free %alloc0 in bb0 and %alloc1 in bb1. Afterwards, we can continue processing all aliases of %2 (none in this case) and we can safely free %2 at the end of the sample program. This sample demonstrates that not all allocations can be safely freed in their associated post-dominator blocks. Instead, we have to pay attention to all of their aliases.

Applying the BufferDeallocation pass to the program above yields the following result:

func @branch(%arg0: i1) {
  %0 = alloc() : memref<2xf32>
  cond_br %arg0, ^bb1, ^bb2
^bb1:
  %1 = alloc() : memref<2xf32>
  %3 = alloc() : memref<2xf32>  // temp copy for %1
  "linalg.copy"(%1, %3) : (memref<2xf32>, memref<2xf32>) -> ()
  dealloc %1 : memref<2xf32> // %1 can be safely freed here
  br ^bb3(%3 : memref<2xf32>)
^bb2:
  use(%0)
  %4 = alloc() : memref<2xf32>  // temp copy for %0
  "linalg.copy"(%0, %4) : (memref<2xf32>, memref<2xf32>) -> ()
  br ^bb3(%4 : memref<2xf32>)
^bb3(%2: memref<2xf32>):
  …
  dealloc %2 : memref<2xf32> // free temp buffer %2
  dealloc %0 : memref<2xf32> // %0 can be safely freed here
  return
}

Note that a temporary buffer for %2 was introduced to free all allocations properly. Note further that the unnecessary allocation of %3 can be easily removed using one of the post-pass transformations.

Reconsider the previously introduced sample demonstrating dynamically shaped types:

func @condBranchDynamicType(
  %arg0: i1,
  %arg1: memref<?xf32>,
  %arg2: memref<?xf32>,
  %arg3: index) {
  cond_br %arg0, ^bb1, ^bb2(%arg3: index)
^bb1:
  br ^bb3(%arg1 : memref<?xf32>)
^bb2(%0: index):
  %1 = alloc(%0) : memref<?xf32>  // aliases: %2
  use(%1)
  br ^bb3(%1 : memref<?xf32>)
^bb3(%2: memref<?xf32>):
  "linalg.copy"(%2, %arg2) : (memref<?xf32>, memref<?xf32>) -> ()
  return
}

In the presence of DSTs, we have to parameterize the allocations with additional dimension information of the source buffers, we want to copy from. BufferDeallocation automatically introduces all required operations to extract dimension specifications and wires them with the associated allocations:

func @condBranchDynamicType(
  %arg0: i1,
  %arg1: memref<?xf32>,
  %arg2: memref<?xf32>,
  %arg3: index) {
  cond_br %arg0, ^bb1, ^bb2(%arg3 : index)
^bb1:
  %c0 = constant 0 : index
  %0 = dim %arg1, %c0 : memref<?xf32>   // dimension operation to parameterize
                                        // the following temp allocation
  %1 = alloc(%0) : memref<?xf32>
  "linalg.copy"(%arg1, %1) : (memref<?xf32>, memref<?xf32>) -> ()
  br ^bb3(%1 : memref<?xf32>)
^bb2(%2: index):
  %3 = alloc(%2) : memref<?xf32>
  use(%3)
  %c0_0 = constant 0 : index
  %4 = dim %3, %c0_0 : memref<?xf32>  // dimension operation to parameterize
                                      // the following temp allocation
  %5 = alloc(%4) : memref<?xf32>
  "linalg.copy"(%3, %5) : (memref<?xf32>, memref<?xf32>) -> ()
  dealloc %3 : memref<?xf32>  // %3 can be safely freed here
  br ^bb3(%5 : memref<?xf32>)
^bb3(%6: memref<?xf32>):
  "linalg.copy"(%6, %arg2) : (memref<?xf32>, memref<?xf32>) -> ()
  dealloc %6 : memref<?xf32>  // %6 can be safely freed here
  return
}

BufferDeallocation performs a fix-point iteration taking all aliases of all tracked allocations into account. We initialize the general iteration process using all tracked allocations and their associated aliases. As soon as we encounter an alias that is not properly dominated by our allocation, we mark this alias as critical (needs to be freed and tracked by the internal fix-point iteration). The following sample demonstrates the presence of critical and non-critical aliases:

func @condBranchDynamicTypeNested(
  %arg0: i1,
  %arg1: memref<?xf32>,  // aliases: %3, %4
  %arg2: memref<?xf32>,
  %arg3: index) {
  cond_br %arg0, ^bb1, ^bb2(%arg3: index)
^bb1:
  br ^bb6(%arg1 : memref<?xf32>)
^bb2(%0: index):
  %1 = alloc(%0) : memref<?xf32>   // cannot be moved upwards due to the data
                                   // dependency to %0
                                   // aliases: %2, %3, %4
  use(%1)
  cond_br %arg0, ^bb3, ^bb4
^bb3:
  br ^bb5(%1 : memref<?xf32>)
^bb4:
  br ^bb5(%1 : memref<?xf32>)
^bb5(%2: memref<?xf32>):  // non-crit. alias of %1, since %1 dominates %2
  br ^bb6(%2 : memref<?xf32>)
^bb6(%3: memref<?xf32>):  // crit. alias of %arg1 and %2 (in other words %1)
  br ^bb7(%3 : memref<?xf32>)
^bb7(%4: memref<?xf32>):  // non-crit. alias of %3, since %3 dominates %4
  "linalg.copy"(%4, %arg2) : (memref<?xf32>, memref<?xf32>) -> ()
  return
}

Applying BufferDeallocation yields the following output:

func @condBranchDynamicTypeNested(
  %arg0: i1,
  %arg1: memref<?xf32>,
  %arg2: memref<?xf32>,
  %arg3: index) {
  cond_br %arg0, ^bb1, ^bb2(%arg3 : index)
^bb1:
  %c0 = constant 0 : index
  %d0 = dim %arg1, %c0 : memref<?xf32>
  %5 = alloc(%d0) : memref<?xf32>  // temp buffer required due to alias %3
  "linalg.copy"(%arg1, %5) : (memref<?xf32>, memref<?xf32>) -> ()
  br ^bb6(%5 : memref<?xf32>)
^bb2(%0: index):
  %1 = alloc(%0) : memref<?xf32>
  use(%1)
  cond_br %arg0, ^bb3, ^bb4
^bb3:
  br ^bb5(%1 : memref<?xf32>)
^bb4:
  br ^bb5(%1 : memref<?xf32>)
^bb5(%2: memref<?xf32>):
  %c0_0 = constant 0 : index
  %d1 = dim %2, %c0_0 : memref<?xf32>
  %6 = alloc(%d1) : memref<?xf32>  // temp buffer required due to alias %3
  "linalg.copy"(%1, %6) : (memref<?xf32>, memref<?xf32>) -> ()
  dealloc %1 : memref<?xf32>
  br ^bb6(%6 : memref<?xf32>)
^bb6(%3: memref<?xf32>):
  br ^bb7(%3 : memref<?xf32>)
^bb7(%4: memref<?xf32>):
  "linalg.copy"(%4, %arg2) : (memref<?xf32>, memref<?xf32>) -> ()
  dealloc %3 : memref<?xf32>  // free %3, since %4 is a non-crit. alias of %3
  return
}

Since %3 is a critical alias, BufferDeallocation introduces an additional temporary copy in all predecessor blocks. %3 has an additional (non-critical) alias %4 that extends the live range until the end of bb7. Therefore, we can free %3 after its last use, while taking all aliases into account. Note that %4 does not need to be freed, since we did not introduce a copy for it.

The actual introduction of buffer copies is done after the fix-point iteration has been terminated and all critical aliases have been detected. A critical alias can be either a block argument or another value that is returned by an operation. Copies for block arguments are handled by analyzing all predecessor blocks. This is primarily done by querying the BranchOpInterface of the associated branch terminators that can jump to the current block. Consider the following example which involves a simple branch and the critical block argument %2:

  custom.br ^bb1(..., %0, : ...)
  ...
  custom.br ^bb1(..., %1, : ...)
  ...
^bb1(%2: memref<2xf32>):
  ...

The BranchOpInterface allows us to determine the actual values that will be passed to block bb1 and its argument %2 by analyzing its predecessor blocks. Once we have resolved the values %0 and %1 (that are associated with %2 in this sample), we can introduce a temporary buffer and clone its contents into the new buffer. Afterwards, we rewire the branch operands to use the newly allocated buffer instead. However, blocks can have implicitly defined predecessors by parent ops that implement the RegionBranchOpInterface. This can be the case if this block argument belongs to the entry block of a region. In this setting, we have to identify all predecessor regions defined by the parent operation. For every region, we need to get all terminator operations implementing the ReturnLike trait, indicating that they can branch to our current block. Finally, we can use a similar functionality as described above to add the temporary copy. This time, we can modify the terminator operands directly without touching a high-level interface.

Consider the following inner-region control-flow sample that uses an imaginary “custom.region_if” operation. It either executes the “then” or “else” region and always continues to the “join” region. The “custom.region_if_yield” operation returns a result to the parent operation. This sample demonstrates the use of the RegionBranchOpInterface to determine predecessors in order to infer the high-level control flow:

func @inner_region_control_flow(
  %arg0 : index,
  %arg1 : index) -> memref<?x?xf32> {
  %0 = alloc(%arg0, %arg0) : memref<?x?xf32>
  %1 = custom.region_if %0 : memref<?x?xf32> -> (memref<?x?xf32>)
   then(%arg2 : memref<?x?xf32>) {  // aliases: %arg4, %1
    custom.region_if_yield %arg2 : memref<?x?xf32>
   } else(%arg3 : memref<?x?xf32>) {  // aliases: %arg4, %1
    custom.region_if_yield %arg3 : memref<?x?xf32>
   } join(%arg4 : memref<?x?xf32>) {  // aliases: %1
    custom.region_if_yield %arg4 : memref<?x?xf32>
   }
  return %1 : memref<?x?xf32>
}

Non-block arguments (other values) can become aliases when they are returned by dialect-specific operations. BufferDeallocation supports this behavior via the RegionBranchOpInterface. Consider the following example that uses an “scf.if” operation to determine the value of %2 at runtime which creates an alias:

func @nested_region_control_flow(%arg0 : index, %arg1 : index) -> memref<?x?xf32> {
  %0 = cmpi "eq", %arg0, %arg1 : index
  %1 = alloc(%arg0, %arg0) : memref<?x?xf32>
  %2 = scf.if %0 -> (memref<?x?xf32>) {
    scf.yield %1 : memref<?x?xf32>   // %2 will be an alias of %1
  } else {
    %3 = alloc(%arg0, %arg1) : memref<?x?xf32>  // nested allocation in a div.
                                                // branch
    use(%3)
    scf.yield %1 : memref<?x?xf32>   // %2 will be an alias of %1
  }
  return %2 : memref<?x?xf32>
}

In this example, a dealloc is inserted to release the buffer within the else block since it cannot be accessed by the remainder of the program. Accessing the RegionBranchOpInterface, allows us to infer that %2 is a non-critical alias of %1 which does not need to be tracked.

func @nested_region_control_flow(%arg0: index, %arg1: index) -> memref<?x?xf32> {
    %0 = cmpi "eq", %arg0, %arg1 : index
    %1 = alloc(%arg0, %arg0) : memref<?x?xf32>
    %2 = scf.if %0 -> (memref<?x?xf32>) {
      scf.yield %1 : memref<?x?xf32>
    } else {
      %3 = alloc(%arg0, %arg1) : memref<?x?xf32>
      use(%3)
      dealloc %3 : memref<?x?xf32>  // %3 can be safely freed here
      scf.yield %1 : memref<?x?xf32>
    }
    return %2 : memref<?x?xf32>
}

Analogous to the previous case, we have to detect all terminator operations in all attached regions of “scf.if” that provides a value to its parent operation (in this sample via scf.yield). Querying the RegionBranchOpInterface allows us to determine the regions that “return” a result to their parent operation. Like before, we have to update all ReturnLike terminators as described above. Reconsider a slightly adapted version of the “custom.region_if” example from above that uses a nested allocation:

func @inner_region_control_flow_div(
  %arg0 : index,
  %arg1 : index) -> memref<?x?xf32> {
  %0 = alloc(%arg0, %arg0) : memref<?x?xf32>
  %1 = custom.region_if %0 : memref<?x?xf32> -> (memref<?x?xf32>)
   then(%arg2 : memref<?x?xf32>) {  // aliases: %arg4, %1
    custom.region_if_yield %arg2 : memref<?x?xf32>
   } else(%arg3 : memref<?x?xf32>) {
    %2 = alloc(%arg0, %arg1) : memref<?x?xf32>  // aliases: %arg4, %1
    custom.region_if_yield %2 : memref<?x?xf32>
   } join(%arg4 : memref<?x?xf32>) {  // aliases: %1
    custom.region_if_yield %arg4 : memref<?x?xf32>
   }
  return %1 : memref<?x?xf32>
}

Since the allocation %2 happens in a divergent branch and cannot be safely deallocated in a post-dominator, %arg4 will be considered a critical alias. Furthermore, %arg4 is returned to its parent operation and has an alias %1. This causes BufferDeallocation to introduce additional copies:

func @inner_region_control_flow_div(
  %arg0 : index,
  %arg1 : index) -> memref<?x?xf32> {
  %0 = alloc(%arg0, %arg0) : memref<?x?xf32>
  %1 = custom.region_if %0 : memref<?x?xf32> -> (memref<?x?xf32>)
   then(%arg2 : memref<?x?xf32>) {
    %c0 = constant 0 : index  // determine dimension extents for temp allocation
    %2 = dim %arg2, %c0 : memref<?x?xf32>
    %c1 = constant 1 : index
    %3 = dim %arg2, %c1 : memref<?x?xf32>
    %4 = alloc(%2, %3) : memref<?x?xf32>  // temp buffer required due to critic.
                                          // alias %arg4
    linalg.copy(%arg2, %4) : memref<?x?xf32>, memref<?x?xf32>
    custom.region_if_yield %4 : memref<?x?xf32>
   } else(%arg3 : memref<?x?xf32>) {
    %2 = alloc(%arg0, %arg1) : memref<?x?xf32>
    %c0 = constant 0 : index  // determine dimension extents for temp allocation
    %3 = dim %2, %c0 : memref<?x?xf32>
    %c1 = constant 1 : index
    %4 = dim %2, %c1 : memref<?x?xf32>
    %5 = alloc(%3, %4) : memref<?x?xf32>  // temp buffer required due to critic.
                                          // alias %arg4
    linalg.copy(%2, %5) : memref<?x?xf32>, memref<?x?xf32>
    dealloc %2 : memref<?x?xf32>
    custom.region_if_yield %5 : memref<?x?xf32>
   } join(%arg4: memref<?x?xf32>) {
    %c0 = constant 0 : index  // determine dimension extents for temp allocation
    %2 = dim %arg4, %c0 : memref<?x?xf32>
    %c1 = constant 1 : index
    %3 = dim %arg4, %c1 : memref<?x?xf32>
    %4 = alloc(%2, %3) : memref<?x?xf32>  // this allocation will be removed by
                                          // applying the copy removal pass
    linalg.copy(%arg4, %4) : memref<?x?xf32>, memref<?x?xf32>
    dealloc %arg4 : memref<?x?xf32>
    custom.region_if_yield %4 : memref<?x?xf32>
   }
  dealloc %0 : memref<?x?xf32>  // %0 can be safely freed here
  return %1 : memref<?x?xf32>
}

Placement of Deallocs

After introducing allocs and copies, deallocs have to be placed to free allocated memory and avoid memory leaks. The deallocation needs to take place after the last use of the given value. The position can be determined by calculating the common post-dominator of all values using their remaining non-critical aliases. A special-case is the presence of back edges: since such edges can cause memory leaks when a newly allocated buffer flows back to another part of the program. In these cases, we need to free the associated buffer instances from the previous iteration by inserting additional deallocs.

Consider the following “scf.for” use case containing a nested structured control-flow if:

func @loop_nested_if(
  %lb: index,
  %ub: index,
  %step: index,
  %buf: memref<2xf32>,
  %res: memref<2xf32>) {
  %0 = scf.for %i = %lb to %ub step %step
    iter_args(%iterBuf = %buf) -> memref<2xf32> {
    %1 = cmpi "eq", %i, %ub : index
    %2 = scf.if %1 -> (memref<2xf32>) {
      %3 = alloc() : memref<2xf32>  // makes %2 a critical alias due to a
                                    // divergent allocation
      use(%3)
      scf.yield %3 : memref<2xf32>
    } else {
      scf.yield %iterBuf : memref<2xf32>
    }
    scf.yield %2 : memref<2xf32>
  }
  "linalg.copy"(%0, %res) : (memref<2xf32>, memref<2xf32>) -> ()
  return
}

In this example, the then branch of the nested “scf.if” operation returns a newly allocated buffer.

Since this allocation happens in the scope of a divergent branch, %2 becomes a critical alias that needs to be handled. As before, we have to insert additional copies to eliminate this alias using copies of %3 and %iterBuf. This guarantees that %2 will be a newly allocated buffer that is returned in each iteration. However, “returning” %2 to its alias %iterBuf turns %iterBuf into a critical alias as well. In other words, we have to create a copy of %2 to pass it to %iterBuf. Since this jump represents a back edge, and %2 will always be a new buffer, we have to free the buffer from the previous iteration to avoid memory leaks:

func @loop_nested_if(
  %lb: index,
  %ub: index,
  %step: index,
  %buf: memref<2xf32>,
  %res: memref<2xf32>) {
  %4 = alloc() : memref<2xf32>
  "linalg.copy"(%buf, %4) : (memref<2xf32>, memref<2xf32>) -> ()
  %0 = scf.for %i = %lb to %ub step %step
    iter_args(%iterBuf = %4) -> memref<2xf32> {
    %1 = cmpi "eq", %i, %ub : index
    %2 = scf.if %1 -> (memref<2xf32>) {
      %3 = alloc() : memref<2xf32> // makes %2 a critical alias
      use(%3)
      %5 = alloc() : memref<2xf32> // temp copy due to crit. alias %2
      "linalg.copy"(%3, %5) : memref<2xf32>, memref<2xf32>
      dealloc %3 : memref<2xf32>
      scf.yield %5 : memref<2xf32>
    } else {
      %6 = alloc() : memref<2xf32> // temp copy due to crit. alias %2
      "linalg.copy"(%iterBuf, %6) : memref<2xf32>, memref<2xf32>
      scf.yield %6 : memref<2xf32>
    }
    %7 = alloc() : memref<2xf32> // temp copy due to crit. alias %iterBuf
    "linalg.copy"(%2, %7) : memref<2xf32>, memref<2xf32>
    dealloc %2 : memref<2xf32>
    dealloc %iterBuf : memref<2xf32> // free backedge iteration variable
    scf.yield %7 : memref<2xf32>
  }
  "linalg.copy"(%0, %res) : (memref<2xf32>, memref<2xf32>) -> ()
  dealloc %0 : memref<2xf32> // free temp copy %0
  return
}

Example for loop-like control flow. The CFG contains back edges that have to be handled to avoid memory leaks. The bufferization is able to free the backedge iteration variable %iterBuf.

Private Analyses Implementations

The BufferDeallocation transformation relies on one primary control-flow analysis: BufferPlacementAliasAnalysis. Furthermore, we also use dominance and liveness to place and move nodes. The liveness analysis determines the live range of a given value. Within this range, a value is alive and can or will be used in the course of the program. After this range, the value is dead and can be discarded - in our case, the buffer can be freed. To place the allocs, we need to know from which position a value will be alive. The allocs have to be placed in front of this position. However, the most important analysis is the alias analysis that is needed to introduce copies and to place all deallocations.

Post Phase

In order to limit the complexity of the BufferDeallocation transformation, some tiny code-polishing/optimization transformations are not applied on-the-fly during placement. Currently, there is only the CopyRemoval transformation to remove unnecessary copy and allocation operations.

Note: further transformations might be added to the post-pass phase in the future.

CopyRemoval Pass

A common pattern that arises during placement is the introduction of unnecessary temporary copies that are used instead of the original source buffer. For this reason, there is a post-pass transformation that removes these allocations and copies via -copy-removal. This pass, besides removing unnecessary copy operations, will also remove the dead allocations and their corresponding deallocation operations. The CopyRemoval pass can currently be applied to operations that implement the CopyOpInterface in any of these two situations which are

• reusing the source buffer of the copy operation.
• reusing the target buffer of the copy operation.

Reusing the Source Buffer of the Copy Operation

In this case, the source of the copy operation can be used instead of target. The unused allocation and deallocation operations that are defined for this copy operation are also removed. Here is a working example generated by the BufferDeallocation pass that allocates a buffer with dynamic size. A deeper analysis of this sample reveals that the highlighted operations are redundant and can be removed.

func @dynamic_allocation(%arg0: index, %arg1: index) -> memref<?x?xf32> {
  %7 = alloc(%arg0, %arg1) : memref<?x?xf32>
  %c0_0 = constant 0 : index
  %8 = dim %7, %c0_0 : memref<?x?xf32>
  %c1_1 = constant 1 : index
  %9 = dim %7, %c1_1 : memref<?x?xf32>
  %10 = alloc(%8, %9) : memref<?x?xf32>
  linalg.copy(%7, %10) : memref<?x?xf32>, memref<?x?xf32>
  dealloc %7 : memref<?x?xf32>
  return %10 : memref<?x?xf32>
}

Will be transformed to:

func @dynamic_allocation(%arg0: index, %arg1: index) -> memref<?x?xf32> {
  %7 = alloc(%arg0, %arg1) : memref<?x?xf32>
  %c0_0 = constant 0 : index
  %8 = dim %7, %c0_0 : memref<?x?xf32>
  %c1_1 = constant 1 : index
  %9 = dim %7, %c1_1 : memref<?x?xf32>
  return %7 : memref<?x?xf32>
}

In this case, the additional copy %10 can be replaced with its original source buffer %7. This also applies to the associated dealloc operation of %7.

To limit the complexity of this transformation, it only removes copy operations when the following constraints are met:

• The copy operation, the defining operation for the target value, and the deallocation of the source value lie in the same block.
• There are no users/aliases of the target value between the defining operation of the target value and its copy operation.
• There are no users/aliases of the source value between its associated copy operation and the deallocation of the source value.

Reusing the Target Buffer of the Copy Operation

In this case, the target buffer of the copy operation can be used instead of its source. The unused allocation and deallocation operations that are defined for this copy operation are also removed.

Consider the following example where a generic linalg operation writes the result to %temp and then copies %temp to %result. However, these two operations can be merged into a single step. Copy removal removes the copy operation and %temp, and replaces the uses of %temp with %result:

func @reuseTarget(%arg0: memref<2xf32>, %result: memref<2xf32>){
  %temp = alloc() : memref<2xf32>
  linalg.generic {
    args_in = 1 : i64,
    args_out = 1 : i64,
    indexing_maps = [#map0, #map0],
    iterator_types = ["parallel"]} %arg0, %temp {
  ^bb0(%gen2_arg0: f32, %gen2_arg1: f32):
    %tmp2 = exp %gen2_arg0 : f32
    linalg.yield %tmp2 : f32
  }: memref<2xf32>, memref<2xf32>
  "linalg.copy"(%temp, %result) : (memref<2xf32>, memref<2xf32>) -> ()
  dealloc %temp : memref<2xf32>
  return
}

Will be transformed to:

func @reuseTarget(%arg0: memref<2xf32>, %result: memref<2xf32>){
  linalg.generic {
    args_in = 1 : i64,
    args_out = 1 : i64,
    indexing_maps = [#map0, #map0],
    iterator_types = ["parallel"]} %arg0, %result {
  ^bb0(%gen2_arg0: f32, %gen2_arg1: f32):
    %tmp2 = exp %gen2_arg0 : f32
    linalg.yield %tmp2 : f32
  }: memref<2xf32>, memref<2xf32>
  return
}

Like before, several constraints to use the transformation apply:

• The copy operation, the defining operation of the source value, and the deallocation of the source value lie in the same block.
• There are no users/aliases of the target value between the defining operation of the source value and the copy operation.
• There are no users/aliases of the source value between the copy operation and the deallocation of the source value.

Known Limitations

BufferDeallocation introduces additional copies using allocations from the “std” dialect (“std.alloc”). Analogous, all deallocations use the “std” dialect-free operation “std.dealloc”. The actual copy process is realized using “linalg.copy”. Furthermore, buffers are essentially immutable after their creation in a block. Another limitations are known in the case using unstructered control flow.