xref: /aosp_15_r20/external/tensorflow/tensorflow/compiler/xla/service/cpu/simple_orc_jit.h (revision b6fb3261f9314811a0f4371741dbb8839866f948)
1 /* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
2 
3 Licensed under the Apache License, Version 2.0 (the "License");
4 you may not use this file except in compliance with the License.
5 You may obtain a copy of the License at
6 
7     http://www.apache.org/licenses/LICENSE-2.0
8 
9 Unless required by applicable law or agreed to in writing, software
10 distributed under the License is distributed on an "AS IS" BASIS,
11 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12 See the License for the specific language governing permissions and
13 limitations under the License.
14 ==============================================================================*/
15 
16 #ifndef TENSORFLOW_COMPILER_XLA_SERVICE_CPU_SIMPLE_ORC_JIT_H_
17 #define TENSORFLOW_COMPILER_XLA_SERVICE_CPU_SIMPLE_ORC_JIT_H_
18 
19 #include <memory>
20 #include <string>
21 #include <vector>
22 
23 #include "llvm/ADT/Triple.h"
24 #include "llvm/ExecutionEngine/JITEventListener.h"
25 #include "llvm/ExecutionEngine/Orc/Core.h"
26 #include "llvm/ExecutionEngine/Orc/ExecutorProcessControl.h"
27 #include "llvm/ExecutionEngine/Orc/IRCompileLayer.h"
28 #include "llvm/ExecutionEngine/Orc/RTDyldObjectLinkingLayer.h"
29 #include "llvm/ExecutionEngine/Orc/SymbolStringPool.h"
30 #include "llvm/IR/Module.h"
31 #include "llvm/Target/TargetMachine.h"
32 #include "tensorflow/compiler/xla/service/cpu/compiler_functor.h"
33 #include "tensorflow/compiler/xla/types.h"
34 
35 namespace xla {
36 namespace cpu {
37 
38 // Simplified LLVM JIT based on the new Orc API.
39 //
40 // This class wraps Orc's functionality into a single interface that only
41 // exposes what we need for XLA.
42 //
43 // Supports JIT-ing multiple modules but without cross-module linking.
44 // Implements eager compilation - the module is lowered to binary as soon as
45 // it's added to the JIT.
46 class SimpleOrcJIT : public llvm::JITEventListener {
47  public:
48   using ObjLayerT = llvm::orc::RTDyldObjectLinkingLayer;
49   using CompileLayerT = llvm::orc::IRCompileLayer;
50 
51   // Create a new JIT, targeting the host architecture.
52   //
53   // {pre,post}_optimization_hook is invoked on the module before/after all
54   // LLVM IR-level optimizations.  post_codegen_hook is invoked after
55   // compiling to machine code.
56   SimpleOrcJIT(
57       std::unique_ptr<llvm::orc::ExecutorProcessControl> target_process_control,
58       std::unique_ptr<llvm::orc::ExecutionSession> execution_session,
59       const llvm::TargetOptions& target_options,
60       llvm::CodeGenOpt::Level opt_level, bool optimize_for_size,
61       bool disable_expensive_passes, llvm::FastMathFlags fast_math_flags,
62       LLVMCompiler::ModuleHook pre_optimization_hook,
63       LLVMCompiler::ModuleHook post_optimization_hook,
64       std::function<void(const llvm::object::ObjectFile&)> post_codegen_hook);
65 
66   static llvm::Expected<std::unique_ptr<SimpleOrcJIT>> Create(
67       const llvm::TargetOptions& target_options,
68       llvm::CodeGenOpt::Level opt_level, bool optimize_for_size,
69       bool disable_expensive_passes, llvm::FastMathFlags fast_math_flags,
70       LLVMCompiler::ModuleHook pre_optimization_hook,
71       LLVMCompiler::ModuleHook post_optimization_hook,
72       std::function<void(const llvm::object::ObjectFile&)> post_codegen_hook);
73 
74   ~SimpleOrcJIT() override;
75 
data_layout()76   const llvm::DataLayout& data_layout() const { return data_layout_; }
77 
target_triple()78   const llvm::Triple& target_triple() const { return target_triple_; }
79 
80   llvm::Error AddModule(llvm::orc::ThreadSafeModule module);
81 
82   // Discards objects we no longer need once we are done compiling.
83   void DoneCompiling();
84 
85   // Get the runtime address of the compiled symbol whose name is given. Returns
86   // nullptr if the symbol cannot be found.
87   llvm::Expected<llvm::JITEvaluatedSymbol> FindCompiledSymbol(
88       const std::string& name);
89 
target_machine()90   llvm::TargetMachine* target_machine() const { return target_machine_.get(); }
91 
92   // Creates an llvm::TargetMachine suitable for JITting code that will run on
93   // the current machine.
94   static std::unique_ptr<llvm::TargetMachine> InferTargetMachineForJIT(
95       const llvm::TargetOptions& target_options,
96       llvm::CodeGenOpt::Level opt_level);
97 
SizeOfGeneratedCodeInBytes()98   int64_t SizeOfGeneratedCodeInBytes() const {
99     return size_of_generated_code_in_bytes_;
100   }
101 
102  private:
103   llvm::JITEvaluatedSymbol ResolveRuntimeSymbol(llvm::StringRef name);
104 
105   void notifyObjectLoaded(
106       llvm::JITEventListener::ObjectKey key,
107       const llvm::object::ObjectFile& object,
108       const llvm::RuntimeDyld::LoadedObjectInfo& object_info) override;
109   void notifyFreeingObject(llvm::JITEventListener::ObjectKey key) override;
110 
111   std::unique_ptr<llvm::TargetMachine> target_machine_;
112   llvm::Triple target_triple_;
113   const llvm::DataLayout data_layout_;
114   std::unique_ptr<llvm::orc::ExecutorProcessControl> target_process_control_;
115   std::unique_ptr<llvm::orc::ExecutionSession> execution_session_;
116   ObjLayerT object_layer_;
117   CompileLayerT compile_layer_;
118   llvm::orc::JITDylib* main_jit_dylib_;
119   int64_t size_of_generated_code_in_bytes_ = 0;
120 
121   // Non owning pointer to a JIT event listener that registers the JIT events
122   // with an attached GDB.
123   //
124   // Note: we get a pointer to this event listener using
125   // `createGDBRegistrationListener` which makes it look like we're supposed to
126   // free this, but the function is poorly named and really just returns a
127   // pointer to a static object.
128   llvm::JITEventListener* gdb_jit_event_listener_;
129 };
130 
131 }  // namespace cpu
132 }  // namespace xla
133 
134 #endif  // TENSORFLOW_COMPILER_XLA_SERVICE_CPU_SIMPLE_ORC_JIT_H_
135