clean MSAN warnings

[jianguan210]

1. Changed to "calloc" to create and initialize array.
2. Initialized info before being passed into blas package.

[stevengj]

It's a little frustrating that MSan complains about passing a pointer to uninitialized data, even when it is an output buffer for the routine you are calling, as it is here. I'm worried that using calloc everywhere will potentially hide bugs.

In particular, suppose that you malloc uninitialized memory and then you subsequently read from that memory. A memory sanitizer like valgrind will then (correctly) give you an error or warning at the site of the read, allowing you to inspect the code and fix the logic. If instead you use calloc everywhere, the same code will silently succeed even if zero is not the correct initialization.

Maybe instead define a CHK_MALLOC_ZERO macro that we can call explicitly when we want to initialize the memory to zero, e.g. to silence MSan warnings for LAPACK output buffers.

[jianguan210]

Thanks for the suggestion. Got your point. The explicit initialization is good to me. The only concern is that we have to manually add CHK_MALLOC_ZERO to the places where MSan complains. Hopefully it will not be lot. Let me try it first and reach back to you. Thank you.

[stevengj]

The only concern is that we have to manually add CHK_MALLOC_ZERO to the places where MSan complains.

Yes, that's the point — when a memory sanitizer complains, this is an opportunity to manually check what is the correct initialization at that point in the code. The goal is not to silence the sanitizer, but to find and fix bugs.

[jianguan210]

Hi Prof. Johnson, please take another look, and let me know if this revision is good to you. Thanks.

[stevengj]

Shouldn't this be done in sqmatrix_eigensolve, i.e. won't similar warnings come up on other calls to sqmatrix_eigensolve if Msan doesn't understand LAPACK output variables?

Also should have a comment that that this initialization is not actually necessary, but is just to silence Msan.

[jianguan210]

I agree that this is an incomplete solution to silence the MSan based on the unit tests provided from MEEP. It would give additional warnings if other similar functions are hit, e.g. sqmatrix_eigensolve. It is quite hard to find all the MSan warnings without additional testing cases. Now maybe it is not necessary to silence warnings of use-of-uninitialized-value if you think calloc is not a good solution. Please feel free to close this one and the other one if you agree to leave as it is. Thanks.