How To Retarget the GNU Toolchain in 21 Patches

atgreen.github.com

70 points·atgreen··4 comments

Thank you, I thoroughly enjoyed this guide. It seems however, that your initial assumption is somewhat invalid: you wanted to build an architecture that a compiler writer would design.

But then you make all design decisions by looking at what the toolchain does well, which is inevitably the feature set of the most popular processors. Which are, in turn, designed by hardware engineers, per your original assumption.

In the preamble I freely admit that this project was based on naive assumptions and ideas. On the other hand, I believe that I was successful in building an architecture that a compiler writer would design, free from the knowledge of how actual hardware works, simply because I only implemented things that were obvious and easy for GCC.

Note the lack of any kind of instruction scheduling, unnatural addressing modes, weird arithmetic, etc.

The ggx processor's successor, moxie (which is part of the upstream GNU toolchain), is slowly developing characteristics shared with real architectures as the FPGA implementation develops. For instance, I just introduced a pipeline hazard with my new multi-port register file implementation that will benefit from some instruction scheduling in GCC.

Great tutorial! I wish there was a similar article for LLVM/Clang.

http://llvm.org/docs/WritingAnLLVMBackend.html

Reply on news.ycombinator.com