Fedora 25

I filed bug against GCC7 on Fedora here.

I've compiled MariaDB with gcc 7.0.1-0.16

$ gcc --version

gcc (GCC) 7.0.1 20170425 (Red Hat 7.0.1-0.16)

And I am not getting any failures on my end on a debug build:
MariaDB Version 10.1.22-MariaDB-debug

On the other hand, switching to a release build, everything changes.

Taking a simple test case main.range, I get a crash in xtradb only (consistent with your experience). I've attached a GDB to it. Code crashes in:

storage/xtradb/btr/btr0cur.cc:1562

	if (UNIV_UNLIKELY(thr && thr_get_trx(thr)->fake_changes)) {

		/* skip CHANGE, LOG */

		*big_rec = big_rec_vec;

		return(err); /* == DB_SUCCESS */

	}

With a dereference of a nullptr thr. This should not happen, as the first part of the condition guards against it.

So in this case it is definitely a compiler bug. Assembly debugging confirms:

# No relevant asm instructions before this.

mov    0xe0(%rsp),%rax   # $rsp + 0xe0 is the address of thr

mov    0x48(%rax),%rax   # Here we attempt to load thr->graph and crash

Turns out that it's not a compiler bug. _attribute_(nonnull) can affect what optimizations compiler does.

Dirty fix is to remove all MY_ATTRIBUTES from xtradb.

The file storage/xtradb/include/univ.i

@@ -261,7 +261,7 @@

 that are only referenced from within InnoDB, not from MySQL. We disable the

 GCC visibility directive on all Sun operating systems because there is no

 easy way to get it to work. See http://bugs.mysql.com/bug.php?id=52263. */

-#define MY_ATTRIBUTE __attribute__

+#define MY_ATTRIBUTE(X) /* empty */

 #if defined(__GNUC__) && (__GNUC__ >= 4) && !defined(sun) || defined(__INTEL_COMPILER)

 # define UNIV_INTERN MY_ATTRIBUTE((visibility ("hidden")))

 #else

CC marko This may be something for you to consider fixing later.

I enabled '--big-test' option for the testsuite
I made some builds with the patch for the file storage/xtradb/include/univ.i

So far, it looks like the issue has been solved.
Tests fails by hundreds no more and I haven't encountered any other similar issue.
I'm going to propose builds with the patch.

I want to point out, that even though the 'storage/xtradb/include/univ.i' file has been patched, the same line of code appears on one more place:

# grep -R "MY_ATTRIBUTE __attribute__" mariadb-10.1.23/

mariadb-10.1.23/storage/xtradb/include/univ.i:#define MY_ATTRIBUTE __attribute__

mariadb-10.1.23/storage/innobase/include/univ.i:#define MY_ATTRIBUTE __attribute__

However the compilation seems to work fine even if only xtradb file is patched.

I made group of builds and unfortunatelly, it seems, many tests keep failing against the 10.1.21 version.

Here is example build log, with '--big-test' disabled

Again, there is huge difference in amount of tests failing in GCC7 builds (up to 100 failing tetsts) and GCC6 builds (up to 12 failing tests)

I think I tried 7.1 yesterday against 10.2. And didn't have significant troubles - still a few too many warning however generally isolated to mgrooma and connect.

I tried again with gcc-7.1.0 and gcc-7.1.1(from branch 7 head) and had no troubles on 10.2 however quickly hit the test 40 failures (my limit) on 10.1 branch. This is on Ubuntu-16.04 x86_64 and ppc64le

I will make new analysis, when I catch some time.
I'll post the results here later.

I fixed some GCC 7 warnings, but did not yet attack the MY_ATTRIBUTE problem that cvicentiu mentioned.
Most notably, -Wimplicit-fallthrough does identify genuine bugs. I plan to silence bogus -Wimplicit-fallthrough warnings in MariaDB 10.0 and later.

Thanks for attacking -Wimplicit-fallthough. I picked up on https://github.com/codership/galera/issues/464 and https://github.com/codership/galera/issues/465 doing the same fixes in galera which looks much more serious.

The remaining attribute fixes are the ones provided by Google in patches in MDEV-9998 over a year ago

The header changes in:

https://github.com/MariaDB/server/pull/341/files#diff-db850e092e18ea7b5b4fb9872ee2c3ce

I pushed the silencing of bogus -Wimplicit-fallthrough warnings to 10.0, and filed MDEV-12824 for the remaining ones (some of which definitely look like bugs).
I pushed the following to bb-10.0-marko and intend to push to 10.0:
739b47e220a MDEV-12358 Work around what looks like a bug in GCC 7.1.0
e22d86a3eb8 fil_create_new_single_table_tablespace(): Correct a bogus nonnull attribute
956d2540c45 Remove redundant UT_LIST_INIT() calls
The first and last are working around what I think are GCC 7 -O3 code generation bugs (Debian 7.1.0-5).
The second one is fixing a bug that was reported by a GCC 7 warning.

ok to push above ones ^

To get something good out of GCC 7.1.0, I compiled the debug version with -O3 -fsanitize=undefined, ran the InnoDB tests and fixed all InnoDB errors:
cbb6db8c0d7 Fix some _attribute_((nonnull)) misuse
e42a002c2fe ibuf_get_volume_buffered_hash(): Use a proper type cast
5bc1ed8ba74 Simplify trx_undo_report_row_operation()
8b34aabf86b Follow-up to MDEV-12534: Align srv_sys

ol to push above ones ^ (one could argue about 5bc simplify claim but I'm ok with it).

I pushed the above-mentioned commits to 10.0 (replacing the word "Simplify" with "Refactor" in one commit comment).
I would expect these to be merged to 10.1 or 10.2 within a week or two.

I pushed a tentative merge to bb-10.1-marko and compiled and tested it as follows. No UBSAN warnings were reported for InnoDB or XtraDB:

mkdir build

cd build

CC=gcc-7 CXX=g++-7 cmake -DWITH_EMBEDDED_SERVER=OFF -DWITH_UNIT_TESTS=OFF -DCMAKE_BUILD_TYPE=Debug -DCMAKE_{C,CXX}_FLAGS_DEBUG='-g -O3 -fsanitize=undefined' -DPLUGIN_TOKUDB=NO -DPLUGIN_MROONGA=NO -DPLUGIN_OQGRAPH=NO -DPLUGIN_ROCKSDB=NO -DWITH_SSL=bundled ..

make -kj4

cd mysql-test

./mtr --mem --force --parallel=auto --suite=innodb,innodb_fts,innodb_zip,encryption

grep 'storage/[ix].*runtime' var/*/log/mysqld*err*

Note: at least one 10.1 target will fail to build with GCC 7.1.0 in a Debian unstable/experimental chroot. In 10.0, storage/connect would fail to build with a similar message:

In file included from /mariadb/10.1-gcc7/zlib/zconf.h:12:0,

                 from /mariadb/10.1-gcc7/zlib/zlib.h:34,

                 from /mariadb/10.1-gcc7/extra/mariabackup/crc/crc_glue.c:25:

/mariadb/10.1-gcc7/build/include/config.h:690:2: error: #error <my_config.h> MUST be included first!

Can you point me to some specific GIT point, at which I can download the tarball and try your fixes on Fedora builders?
Beacause you all looks like it more or less works for you, so I would give you some feedback.
... But I'm kinda lost at your GitHub.

mschorm, I intend to push the merge to 10.1 soon. Meanwhile, you can experiment with commit ae04f2968ea6 which I linked in my previous comment.

I also performed a tentative merge to 10.2, but pushing that one could take a week. For 10.2, there were a few follow-up fixes necessary which I hope jplindst can review:
850c39f1cce Remove dict_index_t::is_ngram
bbe7eb370a2 Remove bogus _attribute_((nonnull))
7896861d385 Fix a compilation warning
4fb9a820e51 Fix a compilation warning introduced by MDEV-11782
0debfe196ef Remove unnecessary conversions from integer to double
78df4a758d7 Fix some -Wimplicit-fallthrough warnings in InnoDB

ok to push bbe73, 78968, 4fb9a, 0debf, 78df4. 850c3 requires some explanation to commit message before I can review.

mschorm, I pushed a merge to 10.1. It does not fix the my_config.h issue yet; that you should be able to work around by adding

#include "my_config.h"

to the start of the affected files.

I pushed a merge to 10.2 as well.

I tried to build MariaDB on Fedora from the commit in this comment Marko.
And it looks like the issue with hundreds test failing has been solved.

Great work!

MariaDB 10.1.24 is now in stable repos in Fedora.
The GCC issue looks resolved - good job!

However, I'm still not able to make a debug build. But that is different issue, so I should probabbly create a new JIRA ticket - shouldn't I?

Yeh, I'd do a new ticket. You seem comfortable that all gcc-7 issues raised here been resolved.

Regarding debug builds look at MDEV-4646 as this introduces a CMake based RPM build of debug symbol packages.