Skip to content

[XLS JIT] Improve LLVM IR codegen for partitions, temporaries, literals, and selects #5164

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
copybara-service wants to merge 1 commit into from
Closed

[XLS JIT] Improve LLVM IR codegen for partitions, temporaries, literals, and selects #5164

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
55 changes: 39 additions & 16 deletions xls/jit/function_base_jit.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -267,12 +267,8 @@ class LlvmFunctionWrapper final : public JitCompilationMetadata {
// `temp_buffer` argument to the jitted function.
llvm::Value* GetOffsetIntoTempBuffer(int64_t offset,
llvm::IRBuilder<>& builder) {
llvm::Type* i64 = llvm::Type::getInt64Ty(function()->getContext());
llvm::Value* start_buffer = builder.CreatePtrToInt(GetTempBufferArg(), i64);
llvm::Value* start_plus_offset =
builder.CreateAdd(start_buffer, llvm::ConstantInt::get(i64, offset));
return builder.CreateIntToPtr(
start_plus_offset, llvm::PointerType::get(function()->getContext(), 0));
return builder.CreateInBoundsGEP(builder.getInt8Ty(), GetTempBufferArg(),
{builder.getInt64(offset)});
}

private:
Expand Down Expand Up @@ -378,7 +374,7 @@ class BufferAllocator {
absl::flat_hash_map<Node*, AllocationKind> allocation_kinds_;
};

// The maximum number of xls::Nodes in a partition.
// The maximum number of non-materialized xls::Nodes in a partition.
static constexpr int64_t kMaxPartitionSize = 100;

// Abstraction representing a point (partition function) at which an early exit
Expand Down Expand Up @@ -464,10 +460,12 @@ absl::StatusOr<std::vector<Partition>> PartitionFunctionBase(FunctionBase* f) {
continue;
}
partition_map[node] = current_partition;
++partition_size;
if (partition_size > kMaxPartitionSize) {
++current_partition;
partition_size = 0;
if (!ShouldMaterializeAtUse(node)) {
++partition_size;
if (partition_size > kMaxPartitionSize) {
++current_partition;
partition_size = 0;
}
}
}

Expand All @@ -491,7 +489,6 @@ absl::StatusOr<std::vector<Partition>> PartitionFunctionBase(FunctionBase* f) {
// Assemble nodes into partitions. `current_partition` is the maximum number
// of any partition.
std::vector<Partition> partitions(current_partition + 1);
XLS_ASSIGN_OR_RETURN(topo_sort_nodes, TopoSort(f));
for (Node* node : topo_sort_nodes) {
int64_t partition = partition_map.at(node);
partitions.at(partition).nodes.push_back(node);
Expand Down Expand Up @@ -587,6 +584,9 @@ absl::StatusOr<llvm::Function*> BuildPartitionFunction(
LlvmFunctionWrapper wrapper = LlvmFunctionWrapper::Create(
name, global_input_nodes, global_output_nodes,
llvm::Type::getInt1Ty(jit_context.context()), jit_context);
// Mark partition functions private so LLVM can discard them after inlining
// into the top-level function.
wrapper.function()->setLinkage(llvm::GlobalValue::PrivateLinkage);
llvm::IRBuilder<>& b = wrapper.entry_builder();

// Whether to interrupt execution of the FunctionBase. Only used for
Expand Down Expand Up @@ -925,11 +925,34 @@ absl::StatusOr<PartitionedFunction> BuildFunctionInternal(
partition_functions.push_back(partition_function);
}

// When execution cannot exit early (i.e. no blocking receives/sends that
// require state in `temp_buffer` to persist across calls) and the temporary
// buffer fits comfortably on the stack, allocate it locally with `alloca`
// instead of using the external `temp_buffer` pointer. Once partition
// functions are inlined, LLVM's SROA and dead-store elimination promote
// cross-partition temporaries to SSA registers and eliminate dead stores.
static constexpr int64_t kMaxLocalTempBufferBytes = 32 * 1024;
bool has_early_exit = absl::c_any_of(partitions, [](const Partition& p) {
return p.early_exit_point.has_value() || p.resume_point.has_value();
});
llvm::Value* temp_buffer_arg = wrapper.GetTempBufferArg();
if (!has_early_exit && allocator.size() > 0 &&
allocator.size() <= kMaxLocalTempBufferBytes) {
llvm::AllocaInst* local_temp = wrapper.entry_builder().CreateAlloca(
llvm::ArrayType::get(llvm::Type::getInt8Ty(jit_context.context()),
allocator.size()),
nullptr, "local_temp_buffer");
local_temp->setAlignment(llvm::Align(allocator.alignment()));
temp_buffer_arg = local_temp;
}

// Args passed to each partition function.
std::vector<llvm::Value*> args = {
wrapper.GetInputsArg(), wrapper.GetOutputsArg(),
wrapper.GetTempBufferArg(), wrapper.GetInterpreterEventsArg(),
wrapper.GetInstanceContextArg(), wrapper.GetJitRuntimeArg()};
std::vector<llvm::Value*> args = {wrapper.GetInputsArg(),
wrapper.GetOutputsArg(),
temp_buffer_arg,
wrapper.GetInterpreterEventsArg(),
wrapper.GetInstanceContextArg(),
wrapper.GetJitRuntimeArg()};

// To handle continuation points, sequential basic blocks are created in the
// body.
Expand Down
Loading
Loading