Re: [PATCH v5 1/4] Remove stale legacy code from rs6000-aix-nat.c file.

[Aditya Kamath <Aditya.Kamath1@ibm.com>]

[-- Attachment #1: Type: text/plain, Size: 4923 bytes --]

Hi Ulrich and community members,

>>                          32-bit
>>                                    before   after    after + your
patches
>>  expected passes             41687   41541    41640
>>  unexpected failures         2302    2294     2308
>>  unexpected successes           1       1        1
>>  expected failures                     15      15       15
>>  known failures                   30      30       30
>>  unresolved testcases            11      86       11
>>  untested testcases           107     107      108
>>  unsupported tests             179     179      178

>That looks better, thanks!   It is still surprising to see
>the differences in untested vs. unsupported tests - do you
>know what this is about?

>Also it is slightly concerning to see the increase in
>unexpected failures - is this caused by different sets
>of tests being run, or just noise?

I git fetched and git merged my repository this morning since there were commits made like here<https://sourceware.org/git/?p=binutils-gdb.git;a=commitdiff;h=bf842a2e8d6e9eb7429444904b1b31c8b85d5fb2> that can change the numbers.

Here are the new numbers.  "before" is without the v5 aix-thread patches,
"after" is with them.  Both columns include the fixes below (wchar,
shell.exp, bigcore, filename-completion).

                          64-bit            32-bit
                          before  after     before  after
  expected passes          41820  41828      41629  41628
  unexpected failures       2341   2338       2294   2295
  unexpected successes         1      1          1      1
  expected failures           17     17         15     15
  known failures              30     30         30     30
  unresolved testcases        16     16         11     11
  untested testcases         105    105        108    108
  unsupported tests          179    179        179    179

What changed since the last numbers:

 1: wchar.exp crashed GDB in 64-bit.  The AIX target used UTF-16 as the
    wide charset while wchar_t is 4 bytes in 64-bit mode, so
    wide strings were decoded wrongly and GDB hit a SIGSEGV on the next
    print.  64-bit now uses UTF-32.  This removed 15 of the
    unresolved results.
 2: shell.exp: accepts ksh88's exit (130) for the "kill -2 $$" tests, which
    fixed the 4 interrupt failures.
 3: bigcore.exp and filename-completion.exp are skipped on AIX.

I will be sending a patch for all 3 soon. These 3 patches are the reason the numbers are as above.

 Regarding the UNTESTED:  Of the 105 (64-bit) / 108 (32-bit), 83 / 84 are build or   link failures on AIX, mostly shared-library builds (ld: 0706-005 and "cannot open
 as COFF file").  7 are "memory at address 0 is possibly executable”, 2 are "can't generate a core file", and the others are missing compiler or tool features (__float128, _FloatN, stap probes, gdb index).

Regarding the UNSUPPORTED: the test needs something AIX target does not have.  Of the 179, 71 are "require failed" (Linux-only features, Python, hardware watchpoints, displaced stepping), 20 are "non-stop mode not supported", and 8 are "address zero is readable".

Both kinds reflect AIX limits.  The 32-bit run has 3 more untested than the 64-bit run: longest-types and the two ptype-offsets tests, which only run on lp64.

Regarding the unexpected failures -  The count is noise from timing-dependent tests, not a regression which is summarized as below :

 - 32-bit: 2294 -> 2295.  No test name that appears in both runs changes
   status.  The extra failure is in vfork-follow-parent, whose failing
   names change from run to run (suffixes such as "(timeout)" and "(the
   program is no longer running)").
 - 64-bit: 2341 -> 2338.  The only changes are in vfork-follow-parent
   (47 -> 44 FAILs) and gdb-sigterm (1 -> 0), plus a multi-forks swap
   (detach 4 and detach 5 swap PASS/FAIL, 50 FAILs in both runs).
 - The FAIL counts per test file are identical, except for those two files in
   64-bit.  The attach, foll-fork, sigstep and bg-exec families have the same
   FAIL counts before and after the patches.
 - About 100 FAILs in each run are "(timeout)": 107/106 in 64-bit,
   98/97 in 32-bit.  They are in the fork, vfork and attach tests.
 - The largest sources of failures are the same in every run and
   are unrelated to the patches: whatis-ptype-typedefs (1062),
   max-depth-c++ (91), nodebug (89), align-c++ (70), sigstep (60).


I hope this answers the questions. Based on these results, none of the observed differences appear to be related to this patch series. Please let me know if you would like me to investigate any of the remaining differences in more detail. I will update the numbers in patch 1/4 before the next revision if there are no further concerns.

Have a nice day ahead.

Thanks and regards,
Aditya.



[-- Attachment #2: Type: text/html, Size: 14608 bytes --]