Cyclic dependency checking broken between version 2015-11 to 2014.12

[gdcarver]

This relates to issue 151.

A cyclic dependency exists in some versions of the IFS code (see issue above). Earlier versions of FCM correctly catch this and report it (see report below). However, all versions from 2014.12 to 2015-11 fail to report the cyclic dependency, will appear to work with no fail messages but not produce executable.

I've found that the change happened between the releases 2014.11.0 and 2014.12.0.

Example output from 2014.11.0:

fcm --version
FCM 2014.11.0

fcm make -v -j4 -f oifs.cfg

[FAIL] make oifs           # 9.4s
[FAIL] make                # 9.5s
[FAIL] mpl_message_mod.mod: target depends on itself
[FAIL]     required by: mpl_buffer_method_mod.o
[FAIL]     required by: mpl_buffer_method_mod.mod
[FAIL]     required by: mpl_init_mod.o
[FAIL]     required by: mpl_init_mod.mod
[FAIL]     required by: dr_hook_util.intfb.h
[FAIL]     required by: yomhook.o
[FAIL]     required by: yomhook.mod
[FAIL]     required by: sdl_mod.o
[FAIL]     required by: sdl_mod.mod
[FAIL]     required by: mpl_abort_mod.o
[FAIL]     required by: mpl_abort_mod.mod
[FAIL]     required by: mpl_message_mod.o
[FAIL]     required by: mpl_message_mod.mod
[FAIL]     required by: mpl_open_mod.o
[FAIL]     required by: mpl_open_mod.mod
[FAIL]     required by: mpl_module.o
[FAIL]     required by: mpl_module.mod
[FAIL]     required by: vod2uv.o
[FAIL]     required by: vod2uv.exe

Example output from 2014.12.0:

fcm --version
FCM 2014.12.0

[info] compile+   0.0 M tpm_flt.mod          <- trans/module/tpm_flt.F90
[info] compile   targets: modified=52, unchanged=0, failed=0, total-time=2.3s
[info] compile+  targets: modified=51, unchanged=0, failed=0, total-time=0.1s
[info] ext-iface targets: modified=2, unchanged=0, failed=0, total-time=0.0s
[info] TOTAL     targets: modified=105, unchanged=0, failed=0, elapsed-time=2.9s
[done] make oifs           # 11.2s
[done] make                # 11.3s

[matthewrmshin]

@gdcarver This is puzzling. I have just done a diff between 2014.11.0 and 2014.12.0. I cannot see any change related to fcm make between these versions. I'll try and see if I can repeat your problem in current master.

[matthewrmshin]

@gdcarver I am unable to repeat this problem in several simple test cases. fcm make correctly identifies the problem in all my little test cases. Do you mind reducing this problem into a simple repeatable test case, and send us a tar ball?

[gdcarver]

@matthewrmshin I was also a bit puzzled as I looked at the changes doc and couldn't see anything that I thought might affect it. The tests are repeatable but I will check again to make sure.

I'll work on reducing the code and send a test case in a day or two.

[gdcarver]

@matthewrmshin Oddly, more testing show 2014-12 behaving correctly not but 2015 releases. But I am getting inconsistent results which might be related to the number of threads I'm using.
Can I check - all I should need to do is correctly set my PATH env var to change between installed versions of FCM? There are no possibilities for perl to get confused about which perl modules fcm is using?

[matthewrmshin]

Can I check - all I should need to do is correctly set my PATH env var to change between installed versions of FCM?

This is the normal expectation.

[gdcarver]

Then I have some odd behaviour that I do not understand at present.

Installed multiple versions of FCM, I build in separate directories for each version (i.e. mode=new) and delete previous build dir to be sure. I use the same source dir and 1 thread for each test.
In each test, PATH is changed to point to the version of FCM under test (making sure it is the only FCM binary in PATH). fcm --version confirms correct fcm is used in the script log.
Each test run has identical config file and source.

In these tests, FCM should not be able to build the final executable but should correctly report a FAIL and a cyclic dependency and not appear to finish successfully.

I get:

version        FAIL reported (out of 4 runs)
2014-06      3 out of 4
2014-09      4 / 4
2014-11      4 / 4
2014-12      2 / 4
2015-02      2 / 4  (n.b. not the same 2 runs as 2014-12)
2015-03      3 / 4
2015-05,-07,-08,-11 : 0 / 4 
(i.e. never report a FAIL and appear to work though no executable is produced)

Another point to note is that if the FAIL is reported, it is not always the same dependency list between versions.

So, for some versions, they may or may not correctly report the FAIL. I do not understand why. Maybe there is some caching or a race condition going on somewhere but I don't understand where with only 1 thread e.g. fcm make -j1. Have you ever seen anything like this before?

I am working on reducing the code to send you a small test.

[gdcarver]

Matt, apologies for not getting back to you about this yet. Creating the test case was not easy and then I had to work on something else. Will get back to this ASAP.

[matthewrmshin]

@gdcarver any further progress in creating the test case?