[COSCUP 2021] LLVM Project: The Good, The Bad, and The Ugly
1 like410 views
The document provides an overview of Min Hsu's talk on the LLVM Project. It introduces Min Hsu and their background working on LLVM. The talk covers an overview of LLVM's design advantages over traditional compilers, some new features like GlobalISel and ThinLTO that improve compilation speed and optimization, and notes on LLVM's federated community structure and contribution process.
[COSCUP 2021] LLVM Project: The Good, The Bad, and The Ugly
- 1. Min-Yih “Min” Hsu @ COSCUP 2021 LLVM Project: The Good, The Bad, and The Ugly
- 3. “Min” Hsu • Computer Science PhD Candidate in University of California, Irvine • Advised by Prof. Michael Franz about:me 2
- 4. “Min” Hsu • Computer Science PhD Candidate in University of California, Irvine • Advised by Prof. Michael Franz • LLVM contributor since 2015 • Code owner of M68k LLVM backend • Author of book “LLVM Techniques, Tips and Best Practices” (2021) about:me 2
- 5. “Min” Hsu • Computer Science PhD Candidate in University of California, Irvine • Advised by Prof. Michael Franz • LLVM contributor since 2015 • Code owner of M68k LLVM backend • Author of book “LLVM Techniques, Tips and Best Practices” (2021) • Motorcycle rider and co ff ee junkies about:me 2
- 6. *Opinions Are My Own 3
- 7. What is this talk? 4
- 8. What is this talk? 4
- 9. What is this talk? 4
- 10. An Overview 5
- 13. Outline The Good Design Some new features Community The Bad The Ugly 6
- 14. Design
- 15. 8 Frontend Optimizer Backend Traditional compiler design
- 16. 9 Frontend Source Files Optimizer Source Files Backend Source Files Traditional compiler design
- 17. 9 Frontend Source Files Optimizer Source Files Backend Source Files A Single Compiler Executable Traditional compiler design
- 18. 9 Frontend Source Files Optimizer Source Files Backend Source Files A Single Compiler Executable Traditional compiler design Example: GCC
- 19. Scenario 10 We’re adding C++ auto-complete feature
- 20. Scenario 10 We’re adding C++ auto-complete feature Create a new from scratch
- 21. Scenario 10 We’re adding C++ auto-complete feature Create a new from scratch Re-use gcc’s frontend
- 22. Scenario 10 We’re adding C++ auto-complete feature Create a new from scratch Re-use gcc’s frontend Gimme a raise
- 23. Try to re-use gcc’s frontend 11 Frontend Source Files Optimizer Source Files Backend Source Files A single compiler executable
- 24. Try to re-use gcc’s frontend 11 Frontend Source Files Optimizer Source Files Backend Source Files A single compiler executable Option 1: Copy / Extract the frontend source code
- 25. Try to re-use gcc’s frontend 11 Frontend Source Files Optimizer Source Files Backend Source Files A single compiler executable Option 1: Copy / Extract the frontend source code Drawback: Time-consuming
- 26. Try to re-use gcc’s frontend 11 Frontend Source Files Optimizer Source Files Backend Source Files A single compiler executable Option 1: Copy / Extract the frontend source code Option 2: > gcc -fsyntax-only code_snapshot.c Drawback: Time-consuming
- 27. Try to re-use gcc’s frontend 11 Frontend Source Files Optimizer Source Files Backend Source Files A single compiler executable Option 1: Copy / Extract the frontend source code Option 2: > gcc -fsyntax-only code_snapshot.c Drawback: Time-consuming Drawbacks: 1. Potentially low performance 2. Interface to the frontend is weak
- 28. Potential solution: Code sharing 12 Code
- 29. LLVM Project: A collection of compiler libraries 13
- 30. LLVM Project: A collection of compiler libraries 13 Clang libraries LLVM libraries libclangParse libLLVMSupport libLLVMCore libLLVMTransformsUtils libLLVMScalarOpts libLLVMLLParser libLLVMMC libLLVMX86Info libLLVMX86CodeGen libclangAST libclangSema libclangCodeGen
- 31. 14 Clang libraries LLVM libraries libclangParse libLLVMSupport libLLVMCore libLLVMTransformsUtils libLLVMScalarOpts libLLVMLLParser libLLVMMC libLLVMX86Info libLLVMX86CodeGen libclangAST libclangSema libclangCodeGen Tools opt
- 32. 15 Clang libraries LLVM libraries libclangParse libLLVMSupport libLLVMCore libLLVMTransformsUtils libLLVMScalarOpts libLLVMLLParser libLLVMMC libLLVMX86Info libLLVMX86CodeGen libclangAST libclangSema libclangCodeGen clang Tools opt
- 33. 16 Clang libraries LLVM libraries libclangParse libLLVMSupport libLLVMCore libLLVMTransformsUtils libLLVMScalarOpts libLLVMLLParser libLLVMMC libLLVMX86Info libLLVMX86CodeGen libclangAST libclangSema libclangCodeGen Tools My auto-complete tool clang opt
- 34. 17 P…Pa… Pass plugin is not the only “modular way” to use LLVM
- 35. Some (Cool) New Features
- 36. Conventional “performance” measuring on compilers 19 Building Flow Source Code Executable
- 37. Conventional “performance” measuring on compilers 19 Building Flow Source Code Executable
- 38. Conventional “performance” measuring on compilers 19 Building Flow Source Code Executable
- 39. Conventional “performance” measuring on compilers 19 Building Flow Source Code Executable
- 40. Conventional “performance” measuring on compilers 19 Building Flow Source Code Executable
- 41. Conventional “performance” measuring on compilers 19 Building Flow Source Code Executable
- 42. LLVM’s improvement on building speeding 20 Building Flow Source Code Executable CodeGen Linking
- 43. LLVM’s improvement on building speeding 20 Building Flow Source Code Executable CodeGen Linking GlobalISel
- 44. Instruction selection in LLVM: Running time LLVM Passes that have the highest running time 21
- 45. Instruction selection in LLVM: Running time LLVM Passes that have the highest running time 21 LLVM Passes Running Time 0% 10% 20% 30% 40% DAG->DAG Instruction Selection Module Veri fi er Live Variable Analysis Greedy Register Allocator Live Debug Value Analysis Prologue / Epilogue Insertion
- 46. Instruction selection in LLVM: Workflow 22 LLVM IR (Linear) Target Independent
- 47. Instruction selection in LLVM: Workflow 22 LLVM IR (Linear) SelectionDAG (DAG) Target Independent
- 48. Instruction selection in LLVM: Workflow 22 LLVM IR (Linear) SelectionDAG (DAG) MachineInstr (Linear) Target Independent Target Speci fi c
- 49. Instruction selection in LLVM: Workflow 22 LLVM IR (Linear) SelectionDAG (DAG) MachineInstr (Linear) Target Independent Target Speci fi c
- 50. Global Instruction Selection (GlobalISel) 23 LLVM IR (Linear) gMIR (Linear) MachineInstr (Linear) Target Independent Target Speci fi c
- 51. Global Instruction Selection (GlobalISel) 23 LLVM IR (Linear) gMIR (Linear) MachineInstr (Linear) Target Independent Target Speci fi c Hybrid
- 52. Global Instruction Selection (GlobalISel) 24 Credit: Ahmad Bougcha et al. @ LLVM Dev Meeting 2017 • Faster compilation
- 53. Global Instruction Selection (GlobalISel) 24 Credit: Ahmad Bougcha et al. @ LLVM Dev Meeting 2017 • Faster compilation • Better codegen quality
- 54. Global Instruction Selection (GlobalISel) 24 Credit: Ahmad Bougcha et al. @ LLVM Dev Meeting 2017 • Faster compilation • Better codegen quality • Testing GlobalISel is much easier.
- 55. 25 Building Flow Source Code Executable CodeGen Linking
- 56. 25 Building Flow Source Code Executable CodeGen Linking LLD / ThinLTO
- 57. Linker 101 26 A.cpp B.cpp C.cpp Compiler Compiler Compiler A.o B.o C.o Linker Executable
- 58. Linker 101 • Linking speed becomes the bottleneck while building large-scale projects. • We barely can parallelize this process. 26 A.cpp B.cpp C.cpp Compiler Compiler Compiler A.o B.o C.o Linker Executable Thread Thread Thread
- 59. LLD: A (much) faster linker • Gold was the fastest linker, but now LLD is (mostly) faster than gold. 27 Credit: Rui Ueyama @ LLVM Dev Meeting 2017
- 60. LLD: A (much) faster linker • Gold was the fastest linker, but now LLD is (mostly) faster than gold. • Secret sauce 27 Credit: Rui Ueyama @ LLVM Dev Meeting 2017
- 61. LLD: A (much) faster linker • Gold was the fastest linker, but now LLD is (mostly) faster than gold. • Secret sauce • Laziness 27 Credit: Rui Ueyama @ LLVM Dev Meeting 2017
- 62. LLD: A (much) faster linker • Gold was the fastest linker, but now LLD is (mostly) faster than gold. • Secret sauce • Laziness • Certain level of parallelization • Use fl ag -fuse-ld=lld on GCC and Clang to adopt LLD. 27 Credit: Rui Ueyama @ LLVM Dev Meeting 2017
- 63. Conventional Linker 28 A.cpp B.cpp C.cpp Compiler Compiler Compiler A.o B.o C.o Linker Executable
- 64. WMO / LTO 29 A.cpp B.cpp C.cpp Compiler Compiler Compiler A.ir B.ir C.ir Linker Executable Whole Module / Link Time Optimization
- 65. WMO / LTO 29 A.cpp B.cpp C.cpp Compiler Compiler Compiler A.ir B.ir C.ir Linker Executable Whole Module / Link Time Optimization • Perform more aggressive whole program optimization during link time. • Because we can fi nally see the entire program.
- 66. WMO / LTO 29 A.cpp B.cpp C.cpp Compiler Compiler Compiler A.ir B.ir C.ir Linker Executable Whole Module / Link Time Optimization • Perform more aggressive whole program optimization during link time. • Because we can fi nally see the entire program. • (Traditionally) It’s still hard to parallelize WMO / LTO. Thread Thread Thread
- 67. ThinLTO 30 A.cpp B.cpp C.cpp Compiler Compiler Compiler Linker Executable Thread Thread Thread A.o B.o C.o Optimizer Optimizer Optimizer
- 68. ThinLTO 30 A.cpp B.cpp C.cpp Compiler Compiler Compiler Linker Executable • Perform inter-procedural optimizations by exchanging function summary. • “Seeing” other functions via function summary. Thread Thread Thread A.o B.o C.o Optimizer Optimizer Optimizer
- 69. ThinLTO 30 A.cpp B.cpp C.cpp Compiler Compiler Compiler Linker Executable • Perform inter-procedural optimizations by exchanging function summary. • “Seeing” other functions via function summary. • ~3x speed up*† • Using ~10x less memory*† Thread Thread Thread A.o B.o C.o Optimizer Optimizer Optimizer * Teresa Johnson et al. @ LLVM Dev Metting 2016 † Compare to conventional LTO
- 70. Community
- 73. A federated community 33 Code Owner Code Owner Code Owner
- 74. A federated community 33 Code Owner Code Owner Code Owner M68k Backend
- 75. A federated community 33 Code Owner Code Owner Code Owner M68k Backend (LLVM Foundation) Board Members
- 76. “Linux style” contribution process 34
- 77. “Linux style” contribution process 34 Maintainer Maintainer
- 78. “Linux style” contribution process 34 Individual Branch (e.g. fi lesystem) Maintainer Individual Branch (e.g. network) Maintainer
- 79. “Linux style” contribution process 34 Individual Branch (e.g. fi lesystem) Maintainer Individual Branch (e.g. network) Maintainer Patch Patch
- 80. “Linux style” contribution process 34 Individual Branch (e.g. fi lesystem) Maintainer Individual Branch (e.g. network) Maintainer Primary Tree Patch Patch
- 81. “Linux style” contribution process 34 Individual Branch (e.g. fi lesystem) Maintainer Individual Branch (e.g. network) Maintainer Primary Tree Pull Request Pull Request Patch Patch
- 82. LLVM’s contribution process 35 The main branch
- 83. LLVM’s contribution process 35 The main branch Tip of Tree (ToT)
- 84. LLVM’s contribution process 35 The main branch Tip of Tree (ToT) Author Reviewer Reviewer Author
- 85. LLVM’s contribution process 36 The main branch Tip of Tree (ToT) Author Reviewer Reviewer Author
- 86. LLVM’s contribution process 36 The main branch Tip of Tree (ToT) Author Reviewer Reviewer Author Commit Commit
- 87. Outline The Good The Bad (things that LLVM probably should fix) Code Review Learning Resources for Beginners The Ugly 37
- 88. Recap: A federated community 38
- 89. Code review with community members 39 Code Review
- 90. Code review with community members 40 Code Review New Contributor
- 91. Code review with community members 40 Code Review New Contributor ?
- 92. Code review with community members 41 Code Review New Contributor ? Why?
- 93. Code review with community members 41 Code Review New Contributor ? • New member does not have enough “review currency” yet. Why?
- 94. Code review with community members 41 Code Review New Contributor ? • New member does not have enough “review currency” yet. • Patches are less visible. Why?
- 95. Code review with community members 41 Code Review New Contributor ? • New member does not have enough “review currency” yet. • Patches are less visible. • Most of the core contributors are busy. Why?
- 96. Code review with community members 41 Code Review New Contributor ? • New member does not have enough “review currency” yet. • Patches are less visible. • Most of the core contributors are busy. • Nearly all of them are working on LLVM as their day jobs! Why?
- 97. Code review with community members 41 Code Review New Contributor ? • New member does not have enough “review currency” yet. • Patches are less visible. • Most of the core contributors are busy. • Nearly all of them are working on LLVM as their day jobs! • Code owners are obligate to review the patch…but they’re also busy. Why?
- 99. Some attempts… 43 Subscribe to (patch) topics that you’re interested in
- 101. 45 LLVM is hard
- 102. LLVM is Flexible 46
- 103. LLVM is Flexible 47 “Where should I start?”
- 104. 48 Clang libraries LLVM libraries libclangParse libLLVMSupport libLLVMCore libLLVMTransformsUtils libLLVMScalarOpts libLLVMLLParser libLLVMMC libLLVMX86Info libLLVMX86CodeGen libclangAST libclangSema libclangCodeGen clang Tools opt My auto-complete tool
- 105. 48 Clang libraries LLVM libraries libclangParse libLLVMSupport libLLVMCore libLLVMTransformsUtils libLLVMScalarOpts libLLVMLLParser libLLVMMC libLLVMX86Info libLLVMX86CodeGen libclangAST libclangSema libclangCodeGen clang Tools opt My auto-complete tool
- 106. 48 Clang libraries LLVM libraries libclangParse libLLVMSupport libLLVMCore libLLVMTransformsUtils libLLVMScalarOpts libLLVMLLParser libLLVMMC libLLVMX86Info libLLVMX86CodeGen libclangAST libclangSema libclangCodeGen clang Tools opt My auto-complete tool
- 107. 49
- 108. 49
- 109. What do you think? 50
- 110. What do you think? 50 • What’s your story? • Project-oriented tutorial v.s. A broad overview • Online forum / mailing list experiences (for newbie questions)? • Wiki(-style) book?
- 111. Outline The Good The Bad The Ugly (things that are difficult to solve) API & ABI stability 51
- 112. API & ABI Stability
- 113. API stability in a nutshell 53 lib.h int foo(int x, int y); MySrc.cpp int main() { return foo(94, 87); } Version 0.1
- 114. API stability in a nutshell 54 lib.h int foo(Point point); MySrc.cpp int main() { return foo(94, 87); } Version 0.5
- 115. API stability in a nutshell 54 lib.h int foo(Point point); MySrc.cpp int main() { return foo(94, 87); } Version 0.5 Compile Error
- 116. LLVM API stability for downstream users Some common LLVM downstream use cases 55
- 117. LLVM API stability for downstream users Some common LLVM downstream use cases 55 My C++ auto-complete plugin libclangSema libclangParse
- 118. LLVM API stability for downstream users Some common LLVM downstream use cases 55 My C++ auto-complete plugin libclangSema libclangParse My Toy compiler LLVM IR Optimizer LLVM X86 Backend
- 119. LLVM API stability for downstream users 56 LLVM C++ API Stability Policy
- 120. LLVM API stability for downstream users 56 LLVM C++ API Stability Policy NONE
- 121. LLVM API stability for downstream users 56 LLVM C++ API Stability Policy NONE • A pain in the a** for downstream users. • Behaviors of APIs might also change!
- 122. LLVM API stability for downstream users 56 LLVM C++ API Stability Policy NONE • A pain in the a** for downstream users. • Behaviors of APIs might also change! • The C API is (mostly) stable… • …But the provided features are lag (way) behind.
- 123. LLVM API stability for downstream users 56 LLVM C++ API Stability Policy NONE • A pain in the a** for downstream users. • Behaviors of APIs might also change! • The C API is (mostly) stable… • …But the provided features are lag (way) behind. • A design decision made since day 1. • Trade stability for faster feature delivery.
- 124. ABI stability in a nutshell 57 MyExecutable int foo(Point point); int main() { return foo(Point{94, 87}); } libFoo.so Version 0.1
- 125. ABI stability in a nutshell 57 MyExecutable int foo(Point point); int main() { return foo(Point{94, 87}); } libFoo.so Version 0.1 Shared Library
- 126. ABI stability in a nutshell 57 MyExecutable int foo(Point point); int main() { return foo(Point{94, 87}); } libFoo.so Version 0.1 Shared Library Plugin
- 127. ABI stability in a nutshell 58 MyExecutable int foo(Point point); int main() { return foo(Point{94, 87}); } libFoo.so Version 0.5 ? Shared Library Plugin
- 128. ABI stability in a nutshell Name mangling 59 MyExecutable int foo(Point point); libFoo.so Version 0.5 ? Old: _Z3foo5Point
- 129. ABI stability in a nutshell Name mangling 59 MyExecutable int foo(Point point); libFoo.so Version 0.5 ? Old: _Z3foo5Point New: ?foo@@YAHUPoint@@@Z
- 130. ABI stability in a nutshell Type layout 60 MyExecutable int foo(Point point) { return point.X + point.Y; } libFoo.so Version 0.5 ? struct Point { int X; int Y; }; Old
- 131. ABI stability in a nutshell Type layout 60 MyExecutable int foo(Point point) { return point.X + point.Y; } libFoo.so Version 0.5 ? struct Point { int X; int Y; }; Old struct Point { int X; int LMAO; int Y; }; New
- 132. LLVM ABI stability for plugins All available plugin options in Clang & LLVM 61 Clang LLVM Preprocessor: Pragma plugin Sema: Attribute plugin AST: AST plugin LLVM IR: Pass plugin (old / new)
- 133. LLVM ABI stability for plugins If LLVM Pass plugin has a stable ABI… 62 opt -my-pass LLVM version X.Y MyPassPlugin.so
- 134. LLVM ABI stability for plugins If LLVM Pass plugin has a stable ABI… 62 opt -my-pass LLVM version X.Y MyPassPlugin.so
- 135. LLVM ABI stability for plugins If LLVM Pass plugin has a stable ABI… 62 opt -my-pass LLVM version X.Y opt -my-pass LLVM version Y.Z MyPassPlugin.so
- 136. LLVM ABI stability for plugins 63 LLVM Plugin LLVM Plugin “You have stable ABI, right?” “…right?…”
- 137. LLVM ABI stability for plugins 63 LLVM Plugin LLVM Plugin “You have stable ABI, right?” “…right?…” • Again, trading stability for fl exibility • LLVM plugins are pretty powerful.
- 138. LLVM ABI stability for plugins 63 LLVM Plugin LLVM Plugin “You have stable ABI, right?” “…right?…” • Again, trading stability for fl exibility • LLVM plugins are pretty powerful. • Real world use case: Chromium
- 139. LLVM ABI stability for plugins 63 LLVM Plugin LLVM Plugin “You have stable ABI, right?” “…right?…” • Again, trading stability for fl exibility • LLVM plugins are pretty powerful. • Real world use case: Chromium • Using custom clang plugin to enforce coding styles inside Chromium’s code base. • E.g. Catching bad C++ dtor
- 140. LLVM ABI stability for plugins 63 LLVM Plugin LLVM Plugin “You have stable ABI, right?” “…right?…” • Again, trading stability for fl exibility • LLVM plugins are pretty powerful. • Real world use case: Chromium • Using custom clang plugin to enforce coding styles inside Chromium’s code base. • E.g. Catching bad C++ dtor • “Don’t write a clang plugin” @ Chromium Docs - https://bit.ly/3rIUg0f
- 141. Thank You!
- 142. Questions? 65 Email: minyihh@uci.edu Github: mshockwave
- 143. Appendix
- 144. 67
- 145. 67 > ./greeting -say “World” Hello, World
- 146. 67 #include “llvm/Support/CommandLine.h” #include <iostream> #include <string> using namespace llvm; static cl::opt<std::string> GreetingText(“say”, cl::desc(“What to say?”)); int main(int argc, char **argv) { cl::ParseCommandLineOptions(argc, argv); std::cout << “Hello, ” << GreetingText << “n” return 0; } > ./greeting -say “World” Hello, World
- 147. 68 #include “llvm/Support/CommandLine.h” #include <iostream> #include <string> using namespace llvm; static cl::opt<std::string> GreetingText(“say”, cl::desc(“What to say?”)); int main(int argc, char **argv) { cl::ParseCommandLineOptions(argc, argv); std::cout << “Hello, ” << GreetingText << “n” return 0; } > c++ file.cc … -I<LLVM include dir> -L<LLVM library path> -o greeting -lLLVMSupport
- 148. Scenario: Adding a new target (backend) 69 … Assembly syntax Instruction encoding …
- 149. 70 Frontend Source Files Optimizer Source Files Backend Source Files Compiler Assembler Assembler Source Files
- 150. 70 Frontend Source Files Optimizer Source Files Backend Source Files Compiler Assembler Assembler Source Files Know how to print assembly
- 151. 70 Frontend Source Files Optimizer Source Files Backend Source Files Compiler Assembler Assembler Source Files Know how to print assembly Know how to parse assembly
- 152. 70 Frontend Source Files Optimizer Source Files Backend Source Files Compiler Assembler Assembler Source Files Know how to print assembly Know how to parse assembly Di ff erent Codebases!
- 153. 71 Frontend Optimizer Backend Assembler LLVM Solution: Sharing target description library
- 154. 71 Frontend Optimizer Backend Assembler Target Description Library (e.g. libLLVMX86Desc) LLVM Solution: Sharing target description library
- 155. 71 Frontend Optimizer Backend Assembler Target Description Library (e.g. libLLVMX86Desc) Assembly Printer LLVM Solution: Sharing target description library
- 156. 71 Frontend Optimizer Backend Assembler Target Description Library (e.g. libLLVMX86Desc) Assembly Printer Assembly Parser LLVM Solution: Sharing target description library
- 157. LLVM’s binary tooling ecosystem 72 Target Description Library
- 158. LLVM’s binary tooling ecosystem 72 Target Description Library MC Layer Binary Format Library
- 159. LLVM’s binary tooling ecosystem 72 Target Description Library MC Layer Assembler Disassembler Binary Format Library } llvm-mc
- 160. LLVM’s binary tooling ecosystem 72 Target Description Library MC Layer Assembler Disassembler Binary Format Library llvm-objdump llvm-nm } llvm-mc
- 161. LLVM’s binary tooling ecosystem 72 Target Description Library MC Layer Assembler Disassembler Binary Format Library llvm-objdump llvm-nm LLD (Linker) } llvm-mc
- 162. LLVM’s binary tooling ecosystem 73 Target Description Library MC Layer Assembler Disassembler llvm-objdump Binary Format Library llvm-nm LLD (Linker) Why?
- 163. LLVM’s binary tooling ecosystem 73 Target Description Library MC Layer Assembler Disassembler llvm-objdump Binary Format Library llvm-nm LLD (Linker) Why? • Binary tools (e.g. binutils) are essential for many compilation fl ows.
- 164. LLVM’s binary tooling ecosystem 73 Target Description Library MC Layer Assembler Disassembler llvm-objdump Binary Format Library llvm-nm LLD (Linker) Why? • Binary tools (e.g. binutils) are essential for many compilation fl ows. • Faster feature updates.
- 165. LLVM’s binary tooling ecosystem 73 Target Description Library MC Layer Assembler Disassembler llvm-objdump Binary Format Library llvm-nm LLD (Linker) Why? • Binary tools (e.g. binutils) are essential for many compilation fl ows. • Faster feature updates. • Avoid version mismatch.
- 166. LLVM’s binary tooling ecosystem 74 Target Description Library MC Layer Assembler Disassembler llvm-objdump Binary Format Library llvm-nm LLD (Linker)
- 167. 75 Frontend Optimizer Backend Assembler Target Description Library (e.g. libLLVMX86Desc) Recap: Sharing target information
- 168. 76 Frontend Optimizer Backend Compiler Process Assembler Process Assembler Linker Process Linker Recap: Traditional compilation pipeline
- 169. 76 Frontend Optimizer Backend Compiler Process Assembler Process Assembler Linker Process Linker Recap: Traditional compilation pipeline .s fi le .o fi le
- 170. 77 Frontend Optimizer Backend Compiler + Assembler Process Assembler Linker Process Linker ?
- 171. 77 Frontend Optimizer Backend Compiler + Assembler Process Assembler Linker Process Linker ? • An assembly fi le always comes from a single source fi le
- 172. 77 Frontend Optimizer Backend Compiler + Assembler Process Assembler Linker Process Linker ? • An assembly fi le always comes from a single source fi le • Avoid round-trip (i.e. read/write assembly fi le) to disk
- 173. 78 Frontend Optimizer Backend A clang that uses integrated assembler Assembler Clang’s integrated assembler In-memory LLVM IR In-memory LLVM IR Textual assembly code In-memory MCInst objects