s390x: 3 test failures on Fedora Rawhide

[tuliom]

While updating libdfp on Fedora Rawhide, I noticed the following failures on s390x:

Details of failed tests:
cat ./test-bfp-conversions.out
cat ./test-cast-to-overflow.out
cat ./test-cast-to-underflow.out
FAIL: test-bfp-conversions
FAIL: test-cast-to-overflow
FAIL: test-cast-to-underflow
Summary of test results:
      3 FAIL
    165 PASS

Source: https://koji.fedoraproject.org/koji/taskinfo?taskID=62029024

I still don't have much information about them.
Notice that Rawhide is using GCC 11 now.

[tuliom]

More details:

Details of failed tests:
cat ./test-bfp-conversions.out
12.345 = (double) 12.345DF
1   Success: Expected: "0x1.8b0a3d70a3d710p+3"
             Result:   "0x1.8b0a3d70a3d710p+3"
in: /builddir/build/BUILD/libdfp-1.0.15/tests/test-bfp-conversions.c:168
2.0 = (double) 2.0DF
2   Success: Expected: "0x1.00000000000000p+1"
             Result:   "0x1.00000000000000p+1"
in: /builddir/build/BUILD/libdfp-1.0.15/tests/test-bfp-conversions.c:169
3.402819e+38 = (double) 3.402819e+38DF
3   Success: Expected: "0x1.ffffd1f5422830p+127"
             Result:   "0x1.ffffd1f5422830p+127"
in: /builddir/build/BUILD/libdfp-1.0.15/tests/test-bfp-conversions.c:170
9.999998e96 = (double) 9.999998e96DF
4   Success: Expected: "0x1.2ba091ef1520b0p+322"
             Result:   "0x1.2ba091ef1520b0p+322"
in: /builddir/build/BUILD/libdfp-1.0.15/tests/test-bfp-conversions.c:171
100000 = (double) 100000.0DF
5   Success: Expected: "0x1.86a00000000000p+16"
             Result:   "0x1.86a00000000000p+16"
in: /builddir/build/BUILD/libdfp-1.0.15/tests/test-bfp-conversions.c:172
999999 = (double) 999999.0DF
6   Success: Expected: "0x1.e847e000000000p+19"
             Result:   "0x1.e847e000000000p+19"
in: /builddir/build/BUILD/libdfp-1.0.15/tests/test-bfp-conversions.c:173
1.e-39 = (double) 1.e-39DF
7   Success: Expected: "0x1.5c72fb1552d830p-130"
             Result:   "0x1.5c72fb1552d830p-130"
in: /builddir/build/BUILD/libdfp-1.0.15/tests/test-bfp-conversions.c:174
-1.e-39 = (double) -1.e-39DF
8   Success: Expected: "-0x1.5c72fb1552d830p-130"
             Result:   "-0x1.5c72fb1552d830p-130"
in: /builddir/build/BUILD/libdfp-1.0.15/tests/test-bfp-conversions.c:175
1.2e-38 = (double) 1.2e-38DF
9   Success: Expected: "0x1.05563c4ffe2230p-126"
             Result:   "0x1.05563c4ffe2230p-126"
in: /builddir/build/BUILD/libdfp-1.0.15/tests/test-bfp-conversions.c:176
1.e-40 = (double) 1.e-40DF
10  Success: Expected: "0x1.16c262777579c0p-133"
             Result:   "0x1.16c262777579c0p-133"
in: /builddir/build/BUILD/libdfp-1.0.15/tests/test-bfp-conversions.c:177
8.e-44 = (double) 8.e-44DF
11  Success: Expected: "0x1.c8b82188545670p-144"
             Result:   "0x1.c8b82188545670p-144"
in: /builddir/build/BUILD/libdfp-1.0.15/tests/test-bfp-conversions.c:178
9.e-44 = (double) 9.e-44DF
12  Success: Expected: "0x1.00e792dcaf70a0p-143"
             Result:   "0x1.00e792dcaf70a0p-143"
in: /builddir/build/BUILD/libdfp-1.0.15/tests/test-bfp-conversions.c:179
8.e-46 = (double) 8.e-46DF
13  Success: Expected: "0x1.244ce242c55610p-150"
             Result:   "0x1.244ce242c55610p-150"
in: /builddir/build/BUILD/libdfp-1.0.15/tests/test-bfp-conversions.c:180
-9.e-44 = (double) -9.e-44DF
14  Success: Expected: "-0x1.00e792dcaf70a0p-143"
             Result:   "-0x1.00e792dcaf70a0p-143"
in: /builddir/build/BUILD/libdfp-1.0.15/tests/test-bfp-conversions.c:181
-8.e-46 = (double) -8.e-46DF
15  Success: Expected: "-0x1.244ce242c55610p-150"
             Result:   "-0x1.244ce242c55610p-150"
in: /builddir/build/BUILD/libdfp-1.0.15/tests/test-bfp-conversions.c:182
-7.0 = (double) -7.0DD
16  Success: Expected: "-0x1.c0000000000000p+2"
             Result:   "-0x1.c0000000000000p+2"
in: /builddir/build/BUILD/libdfp-1.0.15/tests/test-bfp-conversions.c:186
3.402819e+38 = (double) 3.402819e+38DD
17  Success: Expected: "0x1.ffffd1f5422830p+127"
             Result:   "0x1.ffffd1f5422830p+127"
in: /builddir/build/BUILD/libdfp-1.0.15/tests/test-bfp-conversions.c:187
1.79768e+308 = (double) 1.79768e+308DD
18  Success: Expected: "0x1.ffff0ad58f4450p+1023"
             Result:   "0x1.ffff0ad58f4450p+1023"
in: /builddir/build/BUILD/libdfp-1.0.15/tests/test-bfp-conversions.c:188
100000000000000 = (double) 100000000000000.0DD
19  Success: Expected: "0x1.6bcc41e9000000p+46"
             Result:   "0x1.6bcc41e9000000p+46"
in: /builddir/build/BUILD/libdfp-1.0.15/tests/test-bfp-conversions.c:189
999999999999999 = (double) 999999999999999.0DD
20  Success: Expected: "0x1.c6bf52633fff80p+49"
             Result:   "0x1.c6bf52633fff80p+49"
in: /builddir/build/BUILD/libdfp-1.0.15/tests/test-bfp-conversions.c:190
1.e290 = (double) 1.e290DD
21  Successcat ./test-cast-to-overflow.out
cat ./test-cast-to-underflow.out
FAIL: test-bfp-conversions
FAIL: test-cast-to-overflow
FAIL: test-cast-to-underflow
Summary of test results:
      3 FAIL
    165 PASS

Source: https://koji.fedoraproject.org/koji/taskinfo?taskID=62087722

[pmur]

@stliibm FYI.

[tuliom]

Notice the same SRPM pass all the tests on F33 (GCC 10): https://koji.fedoraproject.org/koji/taskinfo?taskID=62088777

[tuliom]

I've recently backported 2 patches to Fedora. The code is available at: https://src.fedoraproject.org/fork/tuliom/rpms/libdfp/tree/bug/1923668

[stliibm]

I was able to reproduce it on my system with a self build gcc (commit 78a6d0e30d7950216dc0c5be5d65d0cbed13924c).

All three testcases are segfaulting as they run out of stack due to recursive calls to itself in __dpd_extend/trunc functions while converting _Decimal128|64|32 to or from long double. This is a GCC 11 regression as the pfpo instruction should be used for all bpf <-> dfp conversions as done with previous GCC versions.

I've opened GCC Bugzilla:
"Bug 99134 - S390x: pfpo instructions are not used for dfp[128|64|32] to/from long double conversions"
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99134

• tests/test-bfp-conversions.c:

404: DO_TEST(TF,TD,);
=> converting long double to _Decimal128
in sysdeps/s390/dfpu/extendtftd.c:
     0x3fffdb885b8 <__GI___dpd_extendtftd+40>        brasl       %r14,0x3fffdb86e80 <__dpd_extendtftd@plt>
     => recursive call (instead of pfpo instruction)!

• tests/test_cast_to_overflow.c:

231:       ret = (long double) d128to128[i].set;
=> converting _Decimal128 to long double
in sysdeps/s390/dfpu/trunctdtf.c:
     0x3fffdb89788 <__GI___dpd_trunctdtf+40> brasl	%r14,0x3fffdb86d80 <__dpd_trunctdtf@plt>
     => recursive call (instead of pfpo instruction)!

• tests/test-cast-to-underflow.c:

484:               ret = (long double) d128to128[i].set;
=> converting _Decimal128 to long double
(see tests/test_cast_to_overflow.c)

[stliibm]

I've just restested libdfp with gcc-head:
$ git log --oneline
60b99ee3bc0 (HEAD -> master, origin/master, origin/HEAD) Daily bump.
...
b6e446cb581 IBM Z: Fix long double <-> DFP conversions
a974b8a592e IBM Z: Improve FPRX2 <-> TF conversions
=> See also GCC-Bugzilla:
"Bug 99134 - S390x: pfpo instructions are not used for dfp[128|64|32] to/from long double conversions"
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99134

Now all the long double <-> _Decimal data-type conversions are using the pfpo instruction and the libdfp testsuite passes.

@tuliom Is Fedora automatically syncing its pre "GCC11" to the latest gcc HEAD? Or do we have to enforce this somehow?

[tuliom]

@stliibm No, I don't think it's automatic. But they've been updating it frequently.
The last update was 12 days ago, so I expect a new one soon.
I'll keep an eye on it in order to rebuild libdfp with my patches there.

Thank you!

[tuliom]

gcc-11.0.0-0.20.fc34 has just landed on Fedora 34 and this issue has been fixed.
I created a PR to merge the latest updates: https://src.fedoraproject.org/rpms/libdfp/pull-request/6