GCC

This is a study report from my time on the Compiler team at Marvell. It covers how the GCC back-end generates and optimizes machine code.

Transcript

1. GCC Marvell Taiwa n Intern Wei-Sheng Chou  1
2. Outline Chapter1 - Compile r – Complication … P 3 – GCC compiler … P1 2 Chapter2 - GCC Source Code – Configuration & Building … P1 7 – Cross Compile GCC compiler … P2 1 Chapter3 - Gray box probing of GC C – Gray box probing of GCC … P2 9 – Passes … P3 3 – Gimple Optimization … P4 6 Reference … P57  2
3. Chapter1
 
 Complication  3
4. Binding conceptualizing Programming Compile linking loading Executing Problem Algorithm, Data Structure, Strategy Function, Variable, Type、Control Flow Instruction, Register Function addressing, library Code, Data addressing Value of Variable unbind object time  4
5.  5 Problem Algorithm, Data Structure, Strategy Function, Variable, Type、Control Flow Instruction, Register Function addressing, library Code, Data addressing Value of Variable unbind object time Binding conceptualizing Programming Compile linking loading Executing
6. Interpreter VS Compiler Source Program Inpu t Data  6 Interpreter Machine Develo p Flexibility
7. Interpreter VS Compiler Compiler Targe t Program Machine Source Program Inpu t Data  7 Executin g Speed
8. Model Front-End Optimizer Code Generator AST Ind. IR Targe t Program Front-End Expander Optimizer Recognizer AST Register Transfer Register Transfer Aho Ullman Model Davidson Fraser Model Machine & Optimizer Independent  8 Machine & Optimizer Dependent
9. Structure of Compiler Front-End Back-End Parser Scanner Semantic Analyzer Symbol table Handler Source Program Instruction Selector Assembly Emitter Register Allocator Assembl y Program AST  9
10. Typical Front-End Parser Scanner Semantic Analyzer Symbol table Handler Source Program AST Parse Tree Error Handler Token AST / IR + Symbol Table  10
11. Typical Back-End Ind. Optimizer Code Generator Register Allocator Assembl y Program Instruction Scheduler Peephole Optimizer Ind. IR Dep. IR AST / IR + Symbol Table  11 Assembly Code Generator
12. Chapter1
 
 GCC
 GNU Compiler Collection
 Great Compiler Challenge  12
13. GCC GCC compiler GCC Source Program Targe t Program  13 CC1 CPP AS LD LIB compiler assembler linker C Preprocessor GCC is a collection that invokes compiler, assembler and linker… Static dynamic
14.  14 Parser Source Program Assembl y Program 14 Architecture Gimplifer TreeSS A Optimizer Expander Optimizer Recoginer Input language Target name Language & Machine Independent Generic Code Machine Dependent Generator Code Machine Descriptions Language Specific Code
15.  15 Parser Source Program Assembl y Program 15 Different Time Gimplifer TreeSS A Optimizer Expander Optimizer Recoginer Input language Target name Language & Machine Independent Generic Code Machine Dependent Generator Code Machine Descriptions Language Specific Code Develop GCC Time Build GCC time Use GCC time Select Copy Copy Generate Generate
16.  16 Parser Source Program Assembl y Program 16 Different View Gimplifer TreeSS A Optimizer Expander Optimizer Recoginer Input language Target name Language & Machine Independent Generic Code Machin e Dependent Generato r Code Machin e Descriptions Language Specific Code C or C++? Arm or x86?
17. Chapter2
 
 GCC Source Code 
 Configuration & Building  17
18. Pre-requisites • ISO C9 0 • GC C • GNU Bas h • Aw k • bzip, gzip, unta r • GNU Make  18 Mpfr librar y Mpc librar y Pp l C LooG-PP L Ja r Libel f GMP https://gcc.gnu.org/install/prerequisites.html
19. Directory • GCC Source – source cod e – $(SOURCE_D ) • GCC Build – make source cod e – $(BUILD ) • GCC Install – install binary fil e – $(INSTALL)  19 *GCC will generate file in build time
20. Step 1. Build pre-requisite s 2. --prefix = /usr/loca l 3. ldconfi g 4. Build gc c cd $(BUILD ) $(SOURCE_D)/configure make; make install  20 => #create makefile => #install path => #link library
21. Chapter3
 Gray box probing of GCC  21
22. What is Gray box?  22 Black Box Input Output
23. What is Gray box?  23 Black Box Input Output White Box Input Output Obj Obj Obj Obj Obj Obj Obj Obj
24. What is Gray box?  24 Black Box Input Output White Box Input Output Gray Box Input Output Obj Obj Obj Obj Obj Obj Obj Obj Obj Obj Obj
25. Chapter3
 
 Passes
 Examining Dumps  25
26. Passes  26 Parse Gimplify TreeSS A Optimize Generat e RTL RT L Optimize Generat e ASM Target Ind. Target dep. Gimple Passes RTL -> ASM RTL Passes Gimple -> RTL
27. Command • gcc -fdump-<stage>-<passname> <file > – ex. gcc -fdump-tree-original test. c – ex. gcc -fdump-tree-cfg-raw test. c – ex. gcc -fdump-ipa-all test. c – Stage : • tre e • ip a • rtl  27
28. Important passes  28 Parse Gimplify Source Program CFG Grammer RTL Generator Reg Allocator pro_epilogue generator Pattern Matcher AS M Program AST: 003t.original GIMPLE: 004t.gimple CFG: 013t.cfg RTL expand: 144r.expand IRA: 191r.ira 198r.prologue-and- epilogue
29. Command Result  29
30. 1. gcc -fdump-tree-original-raw test.c Examples: AST dumps test. c  30 int a ; main( ) { a = 55 ; }
31. test.c.004t.gimpl e 2. gcc -fdump-tree-gimple test.c test. c Examples: GIMPLE dumps  31 int main( ) { int a[3], x ; a[1] = a[2] = 10 ; x = a[1] + a[2] ; a[0] = a[1] + a[1]*x ; } main ( ) { int D.1589 ; int D.1590 ; int D.1591 ; int D.1592 ; int D.1593 ; int D.1594 ; int a[3] ; int x; a[2] = 10 ; D.1589 = a[2] ; a[1] = D.1589 ; D.1590 = a[1] ; D.1591 = a[2] ; x = D.1590 + D.1591 ; D.1592 = x + 1 ; D.1593 = a[1] ; D.1594 = D.1592 * D.1593 ; a[0] = D.1594 ; }
32. Examples: CFG dumps 3. gcc -fdump-tree-cfg test.c  32 test.c.004t.gimple (part ) if (a<=12) goto <D.1200 > else goto <D.1201 > <D.1200> : D.1199 = a + b ; a = D.1199 + c ; <D.1201>: test.c (part ) If (a<=12 ) a = a+b+c;
33. Examples: RTL dumps 4. gcc -fdump-rtl-expand test.c  33 test.c.144r.expand (part) test. c int a ; main( ) { a = a+1 ; } (insn 5 4 6 3 (set (reg:SI 59 [ a.0 ] ) (mem/c/i:SI (symbol_ref:DI ("a") <var_decl 0x7f13bdac3000 a>) [0 a+0 S 4 A32])) test.c:4 - 1 (nil) ) (insn 6 5 7 3 (parallel [ (set (reg:SI 60 [ a.1 ] ) (plus:SI (reg:SI 59 [ a.0 ] ) (const_int 1 [0x1])) ) (clobber (reg:CC 17 flags) ) ]) test.c:4 - 1 (nil) ) (insn 7 6 13 3 (set (mem/c/i:SI (symbol_ref:DI ("a") <var_decl 0x7f13bdac3000 a>) [0 a+0 S4 A32] ) (reg:SI 60 [ a.1 ])) test.c:4 - 1 (nil))
34.  34 (insn 5 4 6 3 (set (reg:SI 59 [ a.0 ] ) (mem/c/i:SI (symbol_ref:DI ("a") <var_decl 0x7f13bdac3000 a>) [0 a+0 S 4 A32])) test.c:4 - 1 (nil) ) (insn 6 5 7 3 (parallel [ (set (reg:SI 60 [ a.1 ] ) (plus:SI (reg:SI 59 [ a.0 ] ) (const_int 1 [0x1])) ) (clobber (reg:CC 17 flags) ) ]) test.c:4 - 1 (nil) ) (insn 7 6 13 3 (set (mem/c/i:SI (symbol_ref:DI ("a") <var_decl 0x7f13bdac3000 a>) [0 a+0 S4 A32] ) (reg:SI 60 [ a.1 ])) test.c:4 - 1 (nil)) r59 = a; r60 = r59 + 1 a = r60
35.  35 (insn 5 4 6 3 (set (reg:SI 59 [ a.0 ] ) (mem/c/i:SI (symbol_ref:DI ("a") <var_decl 0x7f13bdac3000 a>) [0 a+0 S 4 A32])) test.c:4 - 1 (nil) ) (insn 6 5 7 3 (parallel [ (set (reg:SI 60 [ a.1 ] ) (plus:SI (reg:SI 59 [ a.0 ] ) (const_int 1 [0x1])) ) (clobber (reg:CC 17 flags) ) ]) test.c:4 - 1 (nil) ) (insn 7 6 13 3 (set (mem/c/i:SI (symbol_ref:DI ("a") <var_decl 0x7f13bdac3000 a>) [0 a+0 S4 A32] ) (reg:SI 60 [ a.1 ])) test.c:4 - 1 (nil)) Current Inst Previous Inst Next Inst CFG: Basic Block Single Integar Register Integer 1 Scalar Memory reference
36. Examples: Assembly dumps 5. gcc -S test.c | | objdump -d a.out  36 test.s (part ) test.c main : .LFB0 : .cfi_startpro c pushq %rb p .cfi_def_cfa_offset 1 6 .cfi_offset 6, -1 6 movq %rsp, %rb p .cfi_def_cfa_register 6 movl $1, -4(%rbp ) popq %rb p .cfi_def_cfa 7, 8 re t .cfi_endproc int main( ) { int a ; a=1 ; } test.c.144r.expand (part ) (insn 5 4 11 3 (set (mem/c/i:SI (plus:DI (reg/f:DI 54 virtual-stack-vars ) (const_int -4 [0xfffffffffffffffc])) [0 a+0 S4 A32] ) (const_int 1 [0x1])) test.c:4 - 1 (nil))
37. Chapter3 
 
 Gimple Optimization  37
38. Passes  38 Parse Gimplify TreeSS A Optimize Generat e RTL RT L Optimize Generat e ASM Target Ind. Target dep. Gimple Passes RTL -> ASM RTL Passes Gimple -> RTL
39. Brief 0. Pre-procedure (017t.ssa) (022t.copyrename1 ) 1. constant propagation (023t.ccp1) (059t.ccp2 ) 2. copy propagation (027t.copyprop1 ) 3. loop unrolling (058t.cunrolli) 4. dead code elimination (029t.cddce1) Command : gcc -fdump-tree-all -O2 test.c  39
40. Source Code int main( ) { int a=1 ; int b=2 ; int c=3 ; int n=c*2 ; while (a<=n ) a = a+1 ; if (a<12 ) a = a+b+c ; return a ; }  40
41. cfg (Control Flow Graph)  41 a=1 b=2 c=3 n=c*2 B2 If a<=n B4 If a<12 B5 If a<12 B3 D.1200 = a+ b a=D.1200+c B6 D.1201=a return D.1201 B7 T T F F
42. cfg -> ssa (017t.ssa)  42 a_3=1 b_4=2 c_5=3 n_6=c_5*2 B2 a_1=phi(a_3, a_7) If a_1<=n_6 B4 If a_1<12 B5 If a_7<12 B3 D.1200_8 = a_1+b_4 a_9=D.1200_8+c_5 B6 a_2 = phi(a_1, a_9) D.1201_10=a_2 return D.1201_10 B7 T T F F
43. ssa -> copyrename (022t) <bb 7> : #a_2 = PHI<a_1(5),a_9(6) > D.1201_10=a_2 ; return D.1201_10 ; <bb 7> : #a_2 = PHI<a_1(5),a_9(6) > a_10=a_2 ; return a_10;  43 017t.ssa 022t.copyrename1
44. copyrename -> ccp (023t) <bb 2 > a_3=1 ; b_4=2 ; c_5=3 ; n_6=c_5*2 ; goto <bb 4 > <bb 2 > a_3=1 ; b_4=2 ; c_5=3 ; n_6=6 ; goto <bb 4>  44 022t.copyrename1 023t.ccp1
45. ccp -> copyprop (027t) <bb 7> : #a_2 = PHI<a_1(5),a_9(6) > a_10=a_2 ; return a_10 ; <bb 7> : #a_2 = PHI<a_1(5),a_9(6) > return a_2;  45 023t.ccp1 027t.copyprop1
46. copyprop -> cddc (029t)  46 029t.cddc <bb 2> : a_3 = 1 ; b_4 = 2 ; c_5 = 3 ; n_6 = 6 ; goto <bb 4> ; <bb 3> : a_7 = a_1 + 1 ; <bb 4> : … <bb 2> : goto <bb 4> ; <bb 3> : a_7 = a_1 + 1 ; <bb 4> : … 029t.cddc
47. cddc ->cunrolli (058t)  47 <bb 2> : goto <bb 4> ; <bb 3> : a_7 = a_1 + 1 ; <bb 4> : # a_1 = PHI <1(2), a 7(3) > if (a_1 <= 6) goto <bb 3> ; else goto <bb 5> ; <bb 5> : if (a_1 <= 11) goto <bb 6> ; else goto <bb 7> ; <bb 6> : a_9 = a_1 + 5 ; <bb 7> : # a_2 = PHI <a_1(5), a_9(6) > return a_2; a= 1 a=a+ 1 a=a+ 1 a=a+ 1 a=a+ 1 a=a+ 1 a=a+1 <bb 2> : a_12 = 2 ; a_14 = a_12 + 1 ; a_16 = a_14 + 1 ; a_18 = a_16 + 1 ; a_20 = a_18 + 1 ; a_22 = a_20 + 1 ; if (a_22 <= 11) goto <bb 3> ; else goto <bb 4> ; <bb 3> : a_9 = a_22 + 5 ; <bb 4> : # a_2 = PHI <a_22(2), a_9(3) > return a_2; 029t.cddc 058t.cunrolli
48. cunrolli -> ccp2 (059t) main( ) { <bb 2> : return 12 ; }  48 <bb 2> : a_12 = 2 ; a_14 = a_12 + 1 ; a_16 = a_14 + 1 ; a_18 = a_16 + 1 ; a_20 = a_18 + 1 ; a_22 = a_20 + 1 ; if (a_22 <= 11) goto <bb 3> ; else goto <bb 4> ; <bb 3> : a_9 = a_22 + 5 ; <bb 4> : # a_2 = PHI <a_22(2), a_9(3) > return a_2; 058t.cunrolli 059t.ccp2
49. Reference • GCC Source Cod e – https://github.com/mirrors/gc c • IITB GCC workshop OC W – http://www.cse.iitb.ac.in/grc/index.php? page=gcc-pldi14-tut  49