Strict aliasing violation in scan_with_buffer

[jlebar]

https://github.com/thrust/thrust/blob/25669cd425a8f4d896e4a631683899e55086f733/thrust/system/cuda/detail/bulk/algorithm/scan.hpp#L357

If input_type and intermediate_type are not the same, a C++ compiler can assume that a pointer of type input_type* does not alias a pointer of type intermediate_type*. Yet here they clearly do alias when we invoke scan() below.

This can lead to all sorts of difficult-to-track-down bugs.

[brycelelbach]

We've switched to a new implementation for the CUDA backend. Amusingly, there is another union problem in the new backend, but I'm pretty sure this is no longer an issue. Leaving it open until I can confirm. Also, hi Justin.

[jlebar]

Hi, Bryce. :)

We've switched to a new implementation for the CUDA backend.

Hm, I'm not sure what you mean.

I was referring to the fact that this code can miscompile if you compile with clang.

[brycelelbach]

Sorry: what I meant is that this is likely no longer an issue, because the code in question has been rewritten in the CUDA toolkit version of Thrust.

[andrewcorrigan]

Is the CUDA toolkit version of Thrust open-source? Are you going to merge the github repository with the CUDA toolkit version?

[brycelelbach]

Yes, the version of CUDA that ships with Thrust is open source.

I don't have an answer for the second question right now.

[brycelelbach]

Closing; this is no longer an issue, although there are other places where unions may be abused.