memory corruption bug?

[jimclune]

I want to report a possible memory corruption bug in the ocaml bindings for fftw3, version 0.5.2. Here is a small test case:

module FFT = Fftw3.D

let () = let my_array = Array.create 1000 "fish" in Printf.printf "my_array length: %d\n%!" (Array.length myarray); let = FFT.Genarray.create FFT.float Bigarray.c_layout [|10|] in Printf.printf "my_array length: %d\n%!" (Array.length my_array)

If I compile this with ocamlopt (version 4.00.1) and run it, I get:

my_array length: 1000 my_array length: 0

Note the array length is correctly reported as 1000 prior to the call to FFT.Genarray.create, but incorrectly reported as 0 afterwards, which I interpret as an indication of memory corruption. I also observe that if I call Gc.compact () immediately prior to FFT.Genarray.create, then it reports the correct array length afterwards.

[Chris00]

I have not had time to look into it yet — I'll do it ASAP.

[Chris00]

I compiled it to bytecode and native code with OCaml 3.12.1 and the correct result is printed. I'll try with 4.00.1 soon.

[Chris00]

I tried with 4.00.1 and it works fine too. So I cannot reproduce your behavior. Please give more details.

[jimclune]

I am using Mac OS 10.8.2 and have fftw-3.3.3 installed. I am building with: eval opam config -env ocamlbuild -use-ocamlfind junk.native

I can't think of other details that might be relevant, unless I've managed to install something incorrectly (which is certainly possible). If you can think of other details I should provide, please let me know. Thanks.

[Chris00]

I have performed the test with almost the same configuration exept that I am on Linux. Note however that Fftw3.D.Genarray.create does not uses the FFTW3 library. Maybe you want to replace fftw_free and fftw_malloc by the usual free and malloc and tell me whether the bug still occurs?

[jimclune]

The bug still occurs if I replace fftw_malloc and fftw_free with malloc and free. However, I think I see what the problem is. In my environment, STDC_VERSION is 199901L. I suspect you have a version that is lower than this. Can you confirm? See http://caml.inria.fr/mantis/view.php?id=5761 and the conditional preprocessing in ocaml 4.00.1 bigarray_stubs.c. It looks like this bug fix has not yet made it into fftw3_stubs.c. If I replace the two occurrences of "num_dims - 1" in fftw3_stubs.c with "num_dims", then the above test case works for me. I have not yet confirmed that this is sufficient for real fftw usage.

[Chris00]

I have pushed some changes to use the lastest bigarray code. Let me know if it solves the problem.

[jimclune]

Hi, Chris. Thanks for the update. Sorry it took me a while to try it. I still hit the problem. It looks like one spot was missed in fftw3_stubs.c. I think it needs the following change to fftw3_caml_ba_alloc:

147,148c147
<                      sizeof(struct caml_ba_array)
<                      + (num_dims - 1) * sizeof(intnat),
---
>                      SIZEOF_BA_ARRAY + num_dims * sizeof(intnat),

[Chris00]

You are right. I pushed that change.

[Chris00]

Please let me know whether it then works for you. I'll then release a new OPAM package.

[jimclune]

Yes, it works for me. Thanks.

[Chris00]

OPAM package created: https://github.com/OCamlPro/opam-repository/pull/526