차세대컴파일러, VM의미래: 애플 오픈소스 LLVM
•
33 likes•4,607 views
JCO 컨퍼런스에서 발표했던 자료입니다. 애플이 만드는 오픈소스 LLVM 프로젝트가 왜 차세대 컴파일러인지 소개합니다. Java와 여러 다른 언어 사이에서 밀회를 즐길 수 있는 LLVM 활용 사례들을 공개합니다. LLVM과 함께 새로운 경험을 해보실 수 있습니다.
Recommended
More Related Content
What's hot
What's hot (20)
Similar to 차세대컴파일러, VM의미래: 애플 오픈소스 LLVM
Similar to 차세대컴파일러, VM의미래: 애플 오픈소스 LLVM (20)
More from Jung Kim
More from Jung Kim (15)
Recently uploaded
Recently uploaded (20)
차세대컴파일러, VM의미래: 애플 오픈소스 LLVM
- 1. 차세대 컴파일러 VM의 미래 애플 오픈소스 프로젝트 LLVM twitter.com/@godrm facebook.com/godrm
- 2. JCO
- 3. Mac OS X MacOS X iOS의 아버지
- 4. Father of OS X
- 5. Java vs. Objective-C “ ! ! ” - Patrick Naughton
- 6. JVM vs Objective-C Runtime Java Java Obj-C Obj-C App. App. App. App. Runtime JVM 운영체제 하드웨어
- 9. LLVM
- 10. 아이디어의 시작은… • JVMs do all optimizations online at JIT time: - Hugely redundant across runs - Applications launch slowly - What if we could do heavy lifting at install time? ! • Problem: Java bytecode is too limiting! - Memory safety prevents some optzns (eg. bounds checks) - JVM type system doesn’t lend itself to machine optzns
- 11. “ With some sort of low level virtual machine, we could optimize better and a JIT compiler would have to do less work online! ”
- 12. Introduction • LLVM - Low-Level Virtual Machine • An Infrastructure for Multi-stage Optimization - by Chris Arthur Lattner @2002 • Design and Implementation of a compiler infrastructure - support a unique multi-stage optimization system - support inter-procedural and profile-driven optimizations • LLVM virtual instruction (IR) - with high-level type information • Sponsored by APPLE
- 13. LLVM Vision and Approach • Primary mission: Build a set of modular compiler components: - Reduces the time & cost to construct a particular compiler - Components are shared across different compilers - Allows choice of the right component for the job • Secondary mission: Build compilers out of these components X86 PPC CBE clang GCC LTO DWARF Code Target gen JIT Optzn linker IPO BC IO LL IO System Core Support xforms analysis GC ...
- 14. Authors • Vikram Adve - At University of Illinois • Chris Lattner - LLVM: A Compilation Framework for Lifelong Program Analysis & Transformation @2004 - Macroscopic Data Structure Analysis and Optimization @2005 Ph.D Thesis - work for Apple from 2007 • Related Publications - 15 < @2007, 30 ~ @2008, 50 ~ @2009, 30 ~ @2010
- 15. LLVM Timeline LLVM 1.4 @Dec. 2004 WWDC @June. 2008 Clang C @Oct. 2006 LLVM 2.6 @Oct. 2009 LLVM 3.2 @Dec. 2012 LLVM 3.0 @Dec. 2011 LLVM 2.8 @Oct. 2010 LLVM 2.0 @May. 2007 LLVM 2.4 @Nov. 2008 LLVM 3.4 @Jan. 2014 LLVM 1.0 @Oct. 2003 - LTO/LTA Open @July. 2007 osxdev.org
- 16. Compiler Architectures C" C++" Objec)ve+C" GCC#4.2# 프론트엔드# Tree*SSA# 최적화기# 코드 생성기# 실행파일" LLVM# 코드 생성기# 실행파일" LLVM# 코드 생성기# 실행파일" GCC"4.2" C" C++" Objec)ve+C" GCC#4.2# 프론트엔드# LLVM# 최적화기# LLVM+GCC"4.2" C" C++" Objec)ve+C" Clang# 프론트엔드# LLVM# 최적화기# LLVM" osxdev.org
- 17. The Architecture Libraries Compiler FE 1 . . Compiler FE N LLVM .o files LLVM LLVM Native CodeGen Offline Reoptimizer exe & LLVM exe Linker IPO / IPA LLVM exe LLVM CPU exe Link Time Runtime Optimizer LLVM JIT Compile Time Profile & Trace Info Profile Info LLVM Run-Time
- 19. Virtual Instruction Set for ( i=0; i<N; ++i) sum(&A[i], &P); loop: %i.1 = phi int [ 0, %bb0 ], [ %i.2, %loop ] %AiAddr = getelementptr float* %A, int %i.1 call void %sum(float %AiAddr, %pair* %P) %i.2 = add int %i.1, 1 %tmp.4 = setlt int %i.1, %N br bool %tmp.4, label %loop, label %outloop http://llvm.org/docs/LangRef.html
- 20. LLVM Demo #include <stdio.h> #include <stdlib.h> ! int factorial(int X) { if (X == 0) return 1; return X*factorial(X-1); } ! int main(int argc, char **argv) { printf("%dn", factorial(atoi(argv[1]))); } http://llvm.org/demo
- 22. Summary of the optimization • Analysis Passes (~50) - Basic Alias Analysis - Basic CallGraph construction - Count alias analysis query response - Dominator Tree Construction - Counts the various types of Instructions - Loop Dependence Analysis - Track pointer bounds • Transform Passes (~70) - Remove redundant conditional branches - Aggressive Dead Code Elimination - Dead Code Elimination - Deduce function attributes - Unroll loops - Optimize use of memcpy and friends - Strip debug info for unused symbols • Utility Passes (~10) - Dead Argument Hacking - View CFG of function http://llvm.org/docs/Passes.html
- 23. Optimization Example - IPO
- 24. Optimization Example - IPO
- 25. Compile Time Performance • Problem: System headers are huge! - Carbon.h contains: ✴ 558 files ✴ 12.3 megabytes ✴ 10,000 function declarations ✴ 2000 structure definitions, 8000 fields ✴ 3000 enum definitions, 20000 enum constants ✴ 5000 typedefs ✴ 2000 file scoped variables ✴ 6000 macros
- 26. AppKit Framework Compile AppKit Framework Compile Times Time Compile Time (minutes) GCC 4.2 LLVM-GCC 24.5 16.0 35% Faster 5% Faster 4.2 -Os 3.9 -O0 -g
- 27. Full Build Speedup (Debug Option) Full Build Speedup (Debug Config) 3x 2.5x 2.3x 2.2x 2.2x 1.9x 1.8x 2.0x 1.8x 1.9x BlastApp 2x 2.1x Growl 2.1x 1.6x 1.4x 1x Sketch Adium Xcode iCal AppleScript Edit Preview Message (Mail) Interface Builder Dashcode Automator AddressBook 0x Full Build is 2x Faster!
- 28. Clang Front-end C, Objective-C and C++ front-end • Aggressive project with many goals... • - Compatibility with GCC - Fast compilation - Expressive error messages (static analysis) ! t.c:6:49: error: invalid operands to binary expression (‘int’ and ‘struct A’)! ! return intArg + func(intArg ? ((someA.X+40) + someA) / 42 : someA.X));! ~~~~~~~~~~~~~ ^ ~~~~~ ! • Host for a broad range of source-level tools http://llvm.org/devmtg/2008-08/Kremenek_StaticAnalyzer.pdf
- 29. Clang Front-end Analyzer Xcode Integration C, Objective-C and C++ front-end • Aggressive project with many goals... • - Compatibility with GCC - Fast compilation - Expressive error messages (static analysis) ! t.c:6:49: error: invalid operands to binary expression (‘int’ and ‘struct A’)! ! return intArg + func(intArg ? ((someA.X+40) + someA) / 42 : someA.X));! ~~~~~~~~~~~~~ ^ ~~~~~ ! • Host for a broad range of source-level tools Improving Your Application with the Xcode Static Analyzer North Beach Tuesday 5:00PM http://llvm.org/devmtg/2008-08/Kremenek_StaticAnalyzer.pdf
- 30. Better Diagnosis of the Problem Better Diagnosis of the Problem GCC 4.2 Clang test.m:4:1: error: unknown type name 'NSString' NSString *P = @"good stuff"; ^
- 31. libclang • Clang is not just a great compiler... - also a library for processing source code ✴ ✴ Resolves identifiers and symbols ✴ Expands macros ✴ • Translates text into AST Makes implicit information explicit Features - Parsing - Indexing and cross-referencing - Syntax highlighting - Code completion NEW
- 32. libclang • Clang is not just a great compiler... - also a library for processing source code ✴ ✴ Resolves identifiers and symbols ✴ Expands macros ✴ • Translates text into AST Makes implicit information explicit Features - Parsing - Indexing and cross-referencing - Syntax highlighting - Code completion NEW
- 33. ARC • NEW Automatic Reference Coun Automatic Reference Counting - Automatic memory management of Objective-C objects - Just Compile-time, Not Run-time Not Garbage-Collector • Migration Tool in Xcode 4.2 • - with LLVM 3.0 - build-settings : -fobjc-arc (cf. -fno-objc-arc) • New Rules - remove dealloc, retain/release/autorelease ✴ can still use CFRetain / CFRelease in CF - Can’t use NSAllocateObject / NSDeallocateObject - Can’t use object pointer in C Structures - no casual casting id -> void* - Can’t use NSAutoreleasePool -> @autoreleasepool - Can’t use memory zone (NSZone) - Can’t give a property name with new- osxdev.org
- 34. libc++ • Another C++ standard library? - http://libcxx.llvm.org • The C++0x spec introduces several fundamentally new ideas - Move semantics - Perfect forwarding - Variadic templates • New language features - C++03 implementation from the beginning - driven several low-level design decisions. NEW
- 35. DragonEgg • gcc plugin - replaces gcc’s optimizers and code generators - reimplementation of llvm-gcc. with gcc-4.5 or later • Current Status (v2.8) - C works well, C++ works fairly well - can compile a small amount of Obj-C/C++ - Limited debug info - Requires patching gcc - Only supports x86-32/64 - Only supports linux/darwin ex) gcc hello.c -S -O1 -o -fplugin=./dragonegg.so NEW
- 37. LLDB NEW • Next-generation! • & High-performance Debugger! • a set of reusable components in LLVM! • Clang expression parser! • LLVM disassembler! • C/C++, Objective-C! • Efficient Multi-threading, symbol manager! • Extension - Python script! • Support Remote protocol/debug server
- 38. Reusability Reusability Extensibility Reusability Extensibility Reusability Extensibility Reusability Extensibility lldb lldblldb lldb Xcode Xcode 44 4 4 Xcode Xcode Python Python Python Python LLDB.framework LLDB.framework LLDB.framework LLDB.framework LLDB Core LLDB Core LLDB Core LLDB Core Process Dynamic Dynamic Object FIles Object Symbols Disassembly Process Process Dynamic Object FIles Object Symbols Disassembly Process Dynamic Object FIles FIlesObject Dynamic Object FIles Object Object Symbols Symbols Disassembly Disassembly Process Object Symbols Disassembly Loader Containers Loader Loader Containers Containers Loader Containers Loader Containers Mac OS X GDB Remote Darwin Mach-O Universal DWARF ELF BSD Archive ObjectFile LLVM
- 39. Introduction GDB % gdb a.out (gdb) break main Breakpoint 1 at 0x100000f33:file main.c line4 (gdb) run LLDB % lldb a.out (lldb) breakpoint set --name main Breakpoint created:1:name=‘main’, locations=1 (lldb) process launch
- 40. Introduction GDB (gdb) info args argc = 1 argv = (const char **) 0x7fff5fbff550 (gdb) info locals i = 32767 LLDB (lldb) frame variable argc = 1 argv = 0x00007fff5fbfff68 i=0
- 41. Expression in LLDB LLDB (lldb) expression x+y->getCount() (int) $0 = 2 (lldb) expression pt (struct point_tag) $1 = { (int) x = 2 (int) y = 3 } (lldb) expression $1.x (int) $2 = 2
- 42. LLDB Command Syntax Command Syntax <noun> <verb> [-options [option-value]] [argument [argument...]] Uses standard getopt_long() for predicate behavior (lldb) process launch a.out --stop-at-entry (lldb) process launch a.out -- --arg0 --arg1 (lldb) process launch a.out -st Options know which other options they are compatible with (lldb) process attach --pid 123 --name a.out
- 43. Common Commands GDB (gdb) ^C (gdb) signal 2 (gdb) info break (gdb) continue (gdb) step (gdb) stepi (gdb) next (gdb) nexti (gdb) finish (gdb) info threads (gdb) backtrace LLDB (lldb) process interrupt (lldb) process signal SIGINT (lldb) breakpoint list (lldb) process continue (lldb) thread step-in (lldb) thread step-inst (lldb) thread step-over (lldb) thread step-over-inst (lldb) thread step-out (lldb) thread list (lldb) thread backtrace
- 44. Common Commands GDB LLDB (gdb) ^C (gdb) signal 2 (gdb) in br (gdb) c (gdb) s (gdb) si (gdb) n (gdb) ni (gdb) f (gdb) info threads (gdb) bt (lldb) pro int (lldb) pro s SIGINT (lldb) br l (lldb) c (lldb) s (lldb) si (lldb) n (lldb) ni (lldb) f (lldb) th l (lldb) bt
- 45. Java LLVM
- 46. Call for help! Call for help OSS community needs to unite work on various scripting languages – Common module to represent/type infer an arbitrary dynamic language – Who will provide this? pypy? parrot? llvm itself someday (“hlvm”)? – Ruby Python Perl Javascript ... Common Dynamic Language Representation + Type Inference C, C++, Ada, ... GLSL, ARB VP, ... llvm-gcc OpenGL What Next? LLVM LTO JIT Install Time Cross Lang Debugger Codegen Optzn Support IPO http://llvm.org/
- 47. LLVM sub-projects Clang & LLDB • OpenMP • Compiler-rt • - provides highly tuned implementations of the low-level code generator. • cf. libgcc VMKit - is an implementation of the Java and .NET Virtual Machines • KLEE - implements a "symbolic virtual machine" - which uses a theorem prover to try to evaluate all dynamic paths through a program
- 48. LLVM sub-projects • Polly - implements a suite of cache-locality optimizations as well as auto-parallelism and vectorization using a polyhedral model. • libclc - implement the OpenCL standard library. • SAFECode - is a memory safety compiler for C/C++ programs. - It instruments code with run-time checks to detect memory safety errors (ex. buffer overflows) at run-time. LLD (Linker) • ... to be continue ... •
- 49. VMKit: a substrate for virtual machines MMTk Classpath Mono Pnetlib LLVM JIT POSIX 운영체제 하드웨어 http://vmkit.llvm.org/tuto/VMKit_pres_eng.pdf
- 51. LLVM
- 52. LLVM use in Open Source OSes Minix moved to Clang as default compiler • FreeBSD is working on ClangBSD • LLVM a hard dependency for Gallium3D • Building Debian with Clang • Unsupported GCC Flags / C Extensions • 출처 : http://llvm.org/devmtg/2012-04-12/Slides/Mark_Charlebois.pdf
- 53. Use-case #1 - New Compiler • Cling - CERN • CtoVerilog - Haifa University • OpenCL - AMD • Click - Ericsson • EDGtoLLVM - ARM • Delphi XE- Embarcadero • Jaguar - Cray
- 54. OpenCL
- 55. OpenCL
- 56. OpenGL – • Very fragile, hard to understand and change (hex opcodes) OpenGL Interpreter: – JIT didn’t support all OpenGL features: fallback to interpreter – Interpreter was very slow, 100x or worse than JIT LLVM 이전까지는… GFX Card GLSL Text OpenGL Parser Custom JIT OpenGL AST Interpreter htt
- 57. 1.Translate OpenGL AST into LLVM call instructions: one per operation – Very fragile, hard to understand and change (hex 2.Use the LLVM inliner to inline opcodes from precompiled bytecode opcodes) 3.Optimize/codegen as before OpenGL • GLSL OpenGL Interpreter: – JIT didn’t support all OpenGL features: fallback to interpreter – Interpreter was very slow, 100xLLVM OpenGL to or worse thanLLVM JIT OpenGL LLVM 이전까지는… Parser LLVM Optimizer JIT LLVM IR OpenGL AST OpenGL to LLVM ... vec3 viewVec = normalize(-ecPosition); GLSL float diffuse = max(dot(lightVec, tnorm), 0.0); Text ... OpenGL Parser OpenGL Optimize, AST Codegen PPC X86 LLVM IR GFX Card ... %tmp1 = call opengl_negate(ecPosition) %viewVec = call opengl_normalize(%tmp1); %tmp2 = call opengl_dot(%lightVec, %tnorm) %diffuse = call opengl_max(%tmp2, 0.0); ... Custom JIT Interpreter LLVM Inliner ... %tmp1 = sub <4 x float> <0,0,0,0>, %ecPosition %tmp3 = shuffle <4 x float> %tmp1, ...; %tmp4 = mul <4 x float> %tmp3, %tmp3 ... http://llvm.org/ htt
- 58. Use-case #2 - Optimization Flash ActionScript (.as3) ➔ ActionScript Bytecode (.abc) ➔ LLVM Bytecode (.bc) ➔.abc ➔ .swf
- 59. ActionScript3 + LLVM? Alchemy Project Simple AS3 function function CopyMatrix(B:Array, A:Array):void { var M:uint = A.length; var N:uint = A[0].length; var remainder:uint = N & 3; for (var i:uint=0; i<M; i++) { var Bi:Array = B[i]; var Ai:Array = A[i]; for (var j:uint=0; j<remainder; j++) Bi[j] = Ai[j]; for (j=remainder; j<N; j+=4) { Bi[j] = Ai[j]; Bi[j+1] = Ai[j+1]; Bi[j+2] = Ai[j+2]; Bi[j+3] = Ai[j+3]; } } } // N mod 4;
- 60. As ABC function CopyMatrix(Array,Array):void /* disp_id=45 method_id=0 */ { // local_count=10 max_scope=1 max_stack=5 code_len=210 0 getlocal0 1 pushscope 2 pushnull 3 coerce Array 5 setlocal 7 7 pushnull 8 coerce Array 10 setlocal 8 12 pushbyte 0 14 convert_u 15 setlocal 9 17 getlocal2 18 getproperty length 20 convert_u 21 setlocal3 22 getlocal2 23 pushbyte 0 25 getproperty null 27 getproperty length 29 convert_u 30 setlocal 4 32 getlocal 4 34 pushbyte 3 36 bitand 37 convert_u 38 setlocal 5 40 pushbyte 0 42 convert_u 43 setlocal 6 45 jump L1
- 61. ActionScript3 + LLVM? As BC ; ModuleID = 'SparseCompRow' declare avm2val @avm2_getproperty(...) readonly declare void @avm2_setproperty(...) declare avm2val @avm2_coerce(...) readnone define avm2val @GO_m6_CopyMatrix(avm2val, avm2val, avm2val) { bb_m6_b0_0_: %i = call avm2val (...)* @avm2_getproperty( avm2val %2, avm2ref bitcast (i32 24 to avm2ref) ) ; <avm2val> [#uses=1] %i41 = add i32 0, 0 ; <i32> [#uses=1] %i42 = call avm2val @avm2box_i32( i32 %i41 ) ; <avm2val> [#uses=1] %i1 = call avm2val (...)* @avm2_pushbyte( i32 0 ) ; <avm2val> [#uses=0] %i2 = call avm2val (...)* @avm2_getproperty( avm2val %2, avm2val %i42, avm2ref bitcast (i32 5 to avm2ref) ) ; <avm2val> [#uses=1] %i3 = call avm2val (...)* @avm2_getproperty( avm2val %i2, avm2ref bitcast (i32 58 to avm2ref) ) ; <avm2val> [#uses=1] %i4 = call avm2val (...)* @avm2_convert_u( avm2val %i3 ) ; <avm2val> [#uses=3] %i43 = add i32 3, 0 ; <i32> [#uses=2] %i44 = call avm2val @avm2box_i32( i32 %i43 ) ; <avm2val> [#uses=6] %i88 = call double @avm2unbox_double( avm2val %i44 ) ; <double> [#uses=1] %i84 = call double @avm2unbox_double( avm2val %i44 ) ; <double> [#uses=1] %i5 = call avm2val (...)* @avm2_pushbyte( i32 3 ) ; <avm2val> [#uses=0] %i45 = call i32 @avm2unbox_i32( avm2val %i4 ) ; <i32> [#uses=1] %i46 = call i32 @avm2unbox_i32( avm2val %i44 ) ; <i32> [#uses=0] %i47 = and i32 %i45, %i43 ; <i32> [#uses=1] %i48 = call avm2val @avm2box_i32( i32 %i47 ) ; <avm2val> [#uses=1] %i6 = call avm2val (...)* @avm2_bitand( avm2val %i4, avm2val %i44 ) ; <avm2val> [#uses=0] %i7 = call avm2val (...)* @avm2_convert_u( avm2val %i48 ) ; <avm2val> [#uses=3] %i63 = call i32 @avm2unbox_i32( avm2val %i7 ) ; <i32> [#uses=1] %i8 = call avm2val (...)* @avm2_pushuint( i32 0 ) ; <avm2val> [#uses=4] %i53 = call i32 @avm2unbox_i32( avm2val %i8 ) ; <i32> [#uses=1] %i49 = call i32 @avm2unbox_i32( avm2val %i8 ) ; <i32> [#uses=1] br label %bb_m6_b1_0_ ; ... ®
- 62. Use-case #3 - Cross Language Crack Scripting Language ➔ C/C++/Java-like Scripting Language ➔ Speed of a compiled language, ease of use of a scripting language ➔ Unladen Sparrow(Python), Rubinius(Ruby), V8(JS) cf. PNaCl (Portable Native Client)
- 63. Use-case #3 - Emscripten compiles LLVM bytecode into JavaScript ➔ C/C++ to JavaScript ➔ can be run on the web ➔ Python, the Bullet physics engine, eSpeak (TTS) https://github.com/kripken/emscripten/wiki
- 64. Everything compiles into LLVM bitcode Use-case #3 - Emscripten compiles LLVM bytecode into JavaScript ➔ C/C++ to JavaScript ➔ The web is everywhere, and runs JavaScript can be run on the web ➔ Python, the Bullet physics engine, eSpeak (TTS) Compiling LLVM bitcode to JavaScript lets us run ~ everything, everywhere https://github.com/kripken/emscripten/wiki
- 67. asmjs.org
- 69. References • Official Site - http://llvm.org - Online Demo : http://llvm.org/demo • Developer Meetings - Sponsored by Apple, Google, Adobe, Qualcomm. , AMD, CERN, Sony Pictures, Intel • LLVM User - Adobe, Apple, Cray, Google, Electronic Arts, NVIDIA, Siemens, Sun Microsystems, XMOS ... - Objective Modula-2, IcedTea, PyPy, iPhone tool chain, IOQuake3, LLVM D, Mono, MacRuby, Pure, LLVM-Lua - CMU, ETH Zurich, NYU, Stanford ...
- 70. 놀라운 사실들 • • • • • • • First production JIT compiler for C-based languages Clang/LLVM have fully replaced GCC in XCode 5 Used on both major mobile platforms: iOS and Android Most GPU compute languages (OpenCL, CUDA, Renderscript) use LLVM First complete C++-11x: language + library First ARM64 compiler in production (iPhone 5s) 2012 ACM Software System Award
- 71. Hazard in LLVM 25 Crash Errors 18.75 12.5 6.25 0 1.9 2.0 2.1 2.2 2.3 2.4 LLVM version 2.5 2.6 2.7 2.8
- 72. Bug fixed… From March, 2008 to present: 170 bugs fixed + 3 reported but not yet fixed ( 57 wrong code bugs, 116 crash bugs) • 2.4 revision simplifies “( a>13 ) & ( a==15)” into “( a>13 )” • 2.8 revision folds “((x==x) | (y&z))==1” into 0 • 2.8 revision reduces this loop into “i=1” for (i=0; i<5; i++) { if (i) continue; if (i) break; }
- 73. More to come… • • • • • • More complete Windows support More effective profile-guided optimization Improved usability, parallelization for LTO Improved auto-vectorization Improved debugging support State-of-the-art pointer analysis
- 74. Conclusion • LLVM is - still continue evolution... - not omnia(all mighty) - maybe a shortcut to the new compiler - new strategy for mobile - alternative solution for HW emulation or VM - can mix with the other languages - synergy with OpenCL, OpenGL - new chance with LLVM projects
- 75.