|
From: | John Regehr |
Subject: | RE: [avr-gcc-list] Re: AVR LLVM backend? |
Date: | Thu, 13 Nov 2008 21:15:04 -0700 (MST) |
User-agent: | Alpine 1.00 (DEB 882 2007-12-20) |
My experience with -fwhole-program -fcombine has been less than stellar. I have seen a 25% code reduction, which is great! But I have also seen up to 25% code *increase*, which is really bad. I would have thought that, at the very least, they would ensure that the code size would not increase. But because of this wide range of results, I can't always rely on those switches, which is a real shame.
I think that ad hoc and embedded-inappropriate inlining heuristics are entirely to blame here. Both llvm and gcc use speed-focused inlining where code bloat is an afterthought. For a target like AVR it won't be hard to improve this.
John
[Prev in Thread] | Current Thread | [Next in Thread] |