Introduction to LLVM

At the BCS Central London Offices, I spoke to the BCS Open Source Specialist Group, about LLVM as an alternative to GCC, how being a newer compile changes the way it feels to compiler developers and how ultimately having multiple high-quality open source compilers is very good for users.

Finally I conclude with a comparison of compilation and execution times, and describe how compilers are also useful in the hardware design space.

Transcript

1. LLVM Simon Cook, Embecosm
2. Copyright © 2014 Embecosm. Freely available under a Creative Commons license What is LLVM? • Open Source Compiler – Taking code written in various languages, translating to machine code for a variety of languages. • Set of Compilation Libraries – For e.g. including JIT functionality in a program.
3. Copyright © 2014 Embecosm. Freely available under a Creative Commons license Languages and Targets Supported • Languages – C/C++ – D – Hydra – Java Bitcode (IcedTea JRE) – Rust – Scheme – etc. • Targets – AArch32 (ARM) – AArch64 – MIPS – PowerPC – R600 – Sparc – X86 – etc.
4. Copyright © 2014 Embecosm. Freely available under a Creative Commons license How does LLVM differ from GCC? • UIUC vs GPL • Designed as a set of clean C++ libraries – Easier to extend • One binary for many architectures – No need to recompile all common libraries to support a second architecture. • Clang is command line compatible with both gcc and cl.exe – Allows you to try a different compiler without changing your build system. • Warnings
5. Copyright © 2014 Embecosm. Freely available under a Creative Commons license Clang Warnings • Take the following C program: #include <stdio.h> #define MAX(x,y) (x > y ? x : y) int (int argv, char **argc) { printf("Max is %i\n", MAX(argv,argc[0])); return 0; }
6. Copyright © 2014 Embecosm. Freely available under a Creative Commons license Clang Warnings • Historically GCC will have produced: test.c: In function ‘main’: test.c:6:25: warning: comparison between pointer and integer [enabled by default] test.c:6:25: warning: pointer/integer type mismatch in conditional expression [enabled by default] test.c:6:3: warning: format ‘%i’ expects argument of type ‘int’, but argument 2 has type ‘char *’ [-Wformat]
7. Copyright © 2014 Embecosm. Freely available under a Creative Commons license Clang Warnings • Clang however produces: test.c:6:25: warning: ordered comparison between pointer and integer ('int' and 'char *') printf("Max is %i\n", MAX(argv,argc[0])); ^ ~~~~ ~~~~~~~ test.c:3:21: note: expanded from macro 'MAX' #define MAX(x,y) (x > y ? x : y) ^ test.c:6:25: warning: pointer/integer type mismatch in conditional expression ('int' and 'char *') [-Wconditional-type-mismatch] printf("Max is %i\n", MAX(argv,argc[0])); ^ ~~~~ ~~~~~~~ test.c:3:25: note: expanded from macro 'MAX' #define MAX(x,y) (x > y ? x : y) ^ test.c:6:25: warning: format specifies type 'int' but the argument has type 'char *' [- Wformat] printf("Max is %i\n", MAX(argv,argc[0])); ~~ ^~~~~~~~~~~~~~~~~ %s test.c:3:18: note: expanded from macro 'MAX' #define MAX(x,y) (x > y ? x : y) ^~~~~~~~~~~~~~~ 3 warnings generated.
8. Copyright © 2014 Embecosm. Freely available under a Creative Commons license Clang Warnings • This has affected the warnings that GCC provide: test.c: In function ‘main’: test.c:3:21: warning: comparison between pointer and integer [enabled by default] #define MAX(x,y) (x > y ? x : y) ^ test.c:6:25: note: in expansion of macro ‘MAX’ printf("Max is %i\n", MAX(argv,argc[0])); ^ test.c:3:29: warning: pointer/integer type mismatch in conditional expression [enabled by default] #define MAX(x,y) (x > y ? x : y) ^ test.c:6:25: note: in expansion of macro ‘MAX’ printf("Max is %i\n", MAX(argv,argc[0])); ^ test.c:6:3: warning: format ‘%i’ expects argument of type ‘int’, but argument 2 has type ‘char *’ [-Wformat=] printf("Max is %i\n", MAX(argv,argc[0])); ^
9. Copyright © 2014 Embecosm. Freely available under a Creative Commons license TableGen • The ease of porting to a new Target comes from TableGen. • TableGen is a language used to describe features of an architecture. – e.g register and instruction definitions • When the compiler is built, TableGen definitions are processed by tblgen, generating target-specific classes used during the instruction selection and generation process.
10. Copyright © 2014 Embecosm. Freely available under a Creative Commons license TableGen Example class ALU_RR<bits<4> subOp, string asmstr, list<dag> pattern> : InstRR<0x8, (outs GPR:$rD), (ins GPR: $rA, GPR:$rB), !strconcat(asmstr, "\t$rD, $rA, $rB"), pattern> { bits<5> rD; bits<5> rA; bits<5> rB; bits<4> op; let Inst{25-21} = rD; let Inst{20-16} = rA; let Inst{15-11} = rB; let Inst{3-0} = op; let op = subOp; } class ALU1_RR<bits<4> subOp, string asmstr, SDNode OpNode> : ALU_RR<subOp, asmstr, [(set GPR:$rD, (OpNode (i32 GPR:$rA), (i32 GPR:$rB)))]>; let isAsCheapAsAMove = 1 in { def SUB : ALU1_RR<0x2, "l.sub", sub>; let isCommutable=1 in { def ADD : ALU1_RR<0x0, "l.add", add>; def AND : ALU1_RR<0x3, "l.and", and>; def OR : ALU1_RR<0x4, "l.or", or>; def XOR : ALU1_RR<0x5, "l.xor", xor>; } } class Ri<bits<5> num, string n> : Register<n> { let Num = num; } def R0 : Ri< 0, "r0">, DwarfRegNum<[0]>; def R1 : Ri< 1, "r1">, DwarfRegNum<[1]>; def R2 : Ri< 2, "r2">, DwarfRegNum<[2]>;
11. Copyright © 2014 Embecosm. Freely available under a Creative Commons license Comparison of Compilation Time Data from http://www.phoronix.com/scan.php?page=article&item=llvm_clang33_3
12. Copyright © 2014 Embecosm. Freely available under a Creative Commons license Comparison of Execution Time Data from http://www.phoronix.com/scan.php?page=article&item=llvm_clang33_3
13. Thank You www.embecosm.com