GCC An Architectural Overview

Transcription

1 GCC An Architectural Overview Diego Novillo Red Hat Canada OSDL Japan - Linux Symposium Tokyo, September 2006

2 Topics 1. Overview 2. Development model 3. Compiler infrastructure 4. Current status and future work 05/09/06 2

3 Overview Key strengths Widely popular Freely available almost everywhere Open development model However Large code base (2.2 MLOC) and aging (~15 years) Difficult to maintain and enhance Technically demanding Recent architectural changes bring hope 05/09/06 3

4 Development Model 05/09/06 4

5 Development Model Project organization Steering Committee Administrative, political Release Manager Release coordination Maintainers Design, implementation Three main stages (~2 months each) Stage 1 Big disruptive changes. Stage 2 Stabilization, minor features. Stage 3 Bug fixes only (driven by bugzilla, mostly). 05/09/06 5

6 Development Model Major development is done in branches Design/implementation discussion on public lists Frequent merges from mainline Final contribution into mainline only at stage 1 and approved by maintainers Anyone with SVN access may create a development branch Vendors create own branches from FSF release branches 05/09/06 6

7 Development Model All contributors must sign FSF copyright release Even if only working on branches Three levels of access Snapshots (weekly) Anonymous SVN Read/write SVN 05/09/06 7

8 Compiler Infrastructure 05/09/06 8

9 Source code <src> gcc libcpp libada libstdc++-v3 libgfortran libobjc boehm-gc libffi libjava zlib libiberty libgomp libssp libmudflap libdecnumber Front/middle/back end Pre-processor Ada runtime C++ runtime Fortran runtime Objective-C runtime Java runtime Utility functions and generic data structures OpenMP runtime Stack Smash Protection runtime Pointer/memory check runtime Decimal arithmetic library 05/09/06 9

10 Compiler pipeline C C++ Java Fortran Front End Middle End Back End GENERIC GIMPLE RTL Assembly Inter Procedural Optimizer RTL Optimizer SSA Optimizer Final Code Generation Call Graph Manager Pass Manager 05/09/06 10

11 GENERIC and GIMPLE GENERIC High GIMPLE Low GIMPLE if (foo (a + b,c)) c = b++ / a endif return c t1 = a + b t2 = foo (t1, c) if (t2!= 0) t3 = b b = b + 1 c = t3 / a endif return c t1 = a + b t2 = foo (t1, c) if (t2!= 0) <L1,L2> L1: t3 = b b = b + 1 c = t3 / a goto L3 L2: L3: return c 05/09/06 11

12 SSA Form Static Single Assignment (SSA) Versioning representation to expose data flow explicitly Assignments generate new versions of symbols Convergence of multiple versions generates new one (Φ functions) a 1 = 3 b 2 = 9 if (i 3 > 20) a 3 = a 1 2 b 4 = b 2 + a 3 a 5 = a a 6 = Φ(a 3, a 5 ) b 7 = Φ(b 4, b 2 ) c 8 = a 6 + b 7 05/09/06 12

13 SSA Optimizers Operate on GIMPLE IL Around 100 passes Vectorization Various loop optimizations Traditional scalar optimizations: CCP, DCE, DSE, FRE, PRE, VRP, SRA, jump threading, forward propagation Field-sensitive, points-to alias analysis Pointer checking instrumentation for C/C++ 05/09/06 13

14 RTL Register Transfer Language Assembler for abstract machine with infinite registers b = a - 1 (set (reg/v:si 59 [ b ]) (plus:si (reg/v:si 60 [ a ] (const_int -1 [0xffffffff])))) 05/09/06 14

15 RTL Abstracts Register classes Memory addressing modes Word sizes and types Compare-and-branch instructions Calling conventions Bitfield operations Type and sign conversions 05/09/06 15

16 RTL Optimizers Operate closer to the hardware Register allocation Scheduling Software pipelining Common subexpression elimination Instruction recombination Mode switching reduction Peephole optimizations Machine specific reorganization 05/09/06 16

17 Current Status and Future Work 05/09/06 17

18 Current Status New Intermediate Representations decouple Front End and Back End Increased internal modularity Lots of new features Fortran 95, mudflap, vectorizer, OpenMP, inter/intra procedural optimizers, stack protection, profiling, etc. Easier to modify 05/09/06 18

19 Future Work Static analysis support Extensibility mechanism to allow 3 rd party tools Link time optimizations Write intermediate representation Read and combine multiple compilation units Dynamic compilation Emit bytecodes Implement virtual machine with optimizing JIT 05/09/06 19

20 Contacts Home page Wiki Mailing lists IRC irc.oft.net/#gcc 05/09/06 20