Compiling lots of 'junk' helps in catching bugs. I added a new tool (located under llvm/tools/llvm-stress) for generating random LL files. The tool can be used to test different llvm components using various compilation flags. Until now, I only found bugs in the codegen, and not in general llvm optimizations. This probably means that the generated tests are currently too simple for the higher-level optimizations.
The command line below generates a random ll file, and llc compiles this file. It often crashes.
The "-seed" flag sets the initial seed to be used by the random function. I implemented a simple portable 'random' function so that the result should be identical on all platforms. The initial seed also appears in the name of the generated function. The "-size" parameter sets the size of the generated random file.
Wow, nifty tool! I’ve already found a couple crashes!
It is also really easy to pinpoint what is causing the error. Whenever you trigger a bug, run llvm-stress with the same seed but a really small size that doesn’t trigger the bug (e.g. like 10). Then do binary search on the size. Eventually you find exactly the cutoff of size that triggers the bug (e.g. 539 runs fine, but 540 crashes), and then you can diff the crashing and non-crashing .ll files and there should only be a tiny difference.
Thanks for looking at this. Knowing that the last instruction triggered the bug is often not enough. I use bugpoint to reduce the failing test. The reason is that some of the bugs may be caused by the interaction between several instruction. Having said that, I think that the change that you proposed is a good one. Can you send a patch ?