Unexpected numerical results with LLVM Native but not with the Interpreter

[tsoernes]

I am submitting a...

• bug report
• feature request
• support request => you might also like to ask your question on the mailing list or gitter chat.

Description

Using LLVM as backend gives different results than using the Interpreter.

From a high-level view, ignoring everything thought irrelevant, I have a simulator executing an iterated function x' = f(x), where f adds or subtracts at most 1 from a subset of elements of x. Domain rules dictate that no element will ever leave the inclusive integer bound [0, 70], which holds true when running with the interpreter. When using LLVM, this does not hold true.

I am inclined to believe that this might be a bug outside of my program for the following reasons:

• The simulator checks every iteration (given run-time flags below) if any element of x exceeds 70, and if so,
  prints out previous x (where no element should exceed the bounds, lest the simulator would have quit). x here refers to the slice of elements with depth 70 (which is maximum depth) for any spatial position in the 3D array freps, that is, in numpy notation, freps[:, :, -1].
• Elements of x are only ever changed by f in increments or decrements of 1
• Example observed value 283480585120

Steps to reproduce (for bugs)

1. Clone
   https://github.com/tsoernes/haskelldca/tree/070ee164250fe1668392d18075de864107f3fde5
2. Build
   stack build

Compare
stack exec dca-exe -- --backend interp --log_iter 1
against
stack exec dca-exe -- --backend cpu --log_iter 1

The interpreter runs without error and yields expected results.

The second, which uses LLVM, always quits due to previously mentioned check on "out of bounds" values, at a non-deterministic time (usually 200-1000 iterations which takes a minute or 2 on a laptop) despite there being no inherent randomness in the program (random seed is fixed). On exit, it will print (amongst other things), x and x', named Prev frep and Current frep respectively. "Prev" will be the last input without out-of-bounds values, while "Current" will have large values.

Your environment

• Version used:
• Backend(s) used and version:
• accelerate-1.2.0.0
• accelerate-llvm-1.2.0.0
• accelerate-llvm-native-1.2.0.0
• accelerate-blas-0.2.0.0
• llvm-hs-6.0.0
• llvm-hs-pure-6.0.0
• GHC version: 8.4.3
• Operating system and version: Linux Fedora 28
• Link to your project: https://github.com/tsoernes/haskelldca/tree/overflow

Tell me if you need more information about or explanation of the code.

[tsoernes]

As I just found out, both LLVM and the Interpreter gives the correct results when run with a different stack-yaml and .stack.work, that is, switching from this stack-yaml to this stack-yaml, and building with:
stack build --stack-yaml stack-release.yaml.

Do note that the two yamls uses different .stack-work directory, which might point to the problem being corruption of temporary files, however I have tried both stack clean and ./myprog +ACC -fflush-cache -fforce-recomp -ACC without effect.

[tmcdonell]

Uh, that is very strange. I never use ghc-options in the stack.yaml files, so not sure if those are doing something unexpected, though I doubt it? That looks like the only difference between the two files, right?

I wonder if this bug is still lurking somewhere, or if this really was a strange incremental build problem... |: