[LLVMdev] llvm-ld optimization options

Chris Brody chris.brody at gmail.com
Fri Apr 18 02:09:58 PDT 2008

> What would be great is to compute an optimal set of passes and -O
> levels for llvm-gcc and llvm-ld.  It makes sense to start with llvm-gcc.
> Presumably what you would do is take a large body of code (spec etc),
> and compile each compilation unit down to bitcode using llvm-gcc at
> -O0.  These bitcode files would be input to COLE.  What COLE would
> have to do is run opt on the bitcode with varying selections of passes,
> and came up with optimal sets of passes based on various criteria
> (execution speed, bitcode size [this one is the easiest to do, maybe
> you should start here], etc), which can then be turned into -O levels
> for llvm-gcc.  I think you can assume that llvm-ld will be run at the

I would guess that an upcoming COLE optimizer should target the raw,
unoptimized bitcode output of any frontend, such as gcc, clang, or any
other languages, right?

http://repo.or.cz/w/hlafsm.git - use for a macro-extensible language

More information about the llvm-dev mailing list