[plabondutta]

Root cause:

caml_array_make_far always built a generic array (tag 0) and copied init into every slot. For a float, init is the address of a boxed float, so Array.make ~far:true n 2.0 stored addresses that read back as ~4.7e-310, and Array.fill took the generic path for the same reason.

Fix:

When init is a boxed float (FLAT_FLOAT_ARRAY), use caml_array_create_float_far to get a proper Double_array_tag array and store Double_val(init) in each slot, matching the stock Array.make.

Test:

testsuite/tests/lib-array/far_float_array.ml

[Binkersss]

This looks good to me. Not sure how I made this oversight haha.

[khale]

Looks like there are some checks failing?

[Binkersss]

Their ci pipeline configuration does not pull in libnuma as a dependency. We can probably fix that if needed though, I haven't looked at their workflow files so unsure of how easy that will be.

[plabondutta]

I have created this PR to fix the CI issue. Once that one gets approved and merged, I will work on this one and update this PR.