[cperciva]

A few months ago I was doing FFTs on arrays larger than 4GB. Amusingly, this uncovered a bug in the LLVM optimizer: It was looking at stride lengths to figure out if accesses were independent, and truncated a 4GB stride down to 0.

[ksml]

I would be extremely interested to hear how you found this bug. Sounds like a difficult bug to track down, and I always learn from good debugging stories.

[cperciva]

It was pretty easy to track down: clang38 was exiting with

    Assertion failed: (Distance > 0 && "The distance must be non-zero"),
    function areStridedAccessesIndependent, file /wrkdirs/usr/ports/devel/
    llvm38/work/llvm-3.8.0.src/lib/Analysis/LoopAccessAnalysis.cpp, line 1004.

Looking at the file it was easy to see what was being asserted, and to see that the type was a 32-bit integer; since I knew I was dealing with huge FFTs, the problem was obvious.

Let this be a lesson: Asserting that impossible things don't happen makes debugging much easier when they do happen!

[flamedoge]

Well good thing it was ReleaseWithAsserts build!