1817565 - gcc-12 fails to inline a couple of always_inline functions in mozilla/mfbt/lz4/xxhash.h when -Og is specified

Status: NEW

(Core :: MFBT, enhancement)

Description

[ISHIKAWA, Chiaki]

Attachment: gcc-12 fails to inline two always_inline funcitons with -Og

I compile C-C Thunderbird under Debian GNU/Linux locally.

$ gcc-12 --version
gcc-12 (Debian 12.2.0-14) 12.2.0
Copyright (C) 2022 Free Software Foundation, Inc.
This is free software; see the source for copying conditions. There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

With the option "-Og", gcc-12 fails to inline two functions that are declared with
XXX_FORCED_INLINE:
XXH3_accumulate_512_sse2(), and
XXH3_scrambleAcc_sse2()

I am attaching the error log.

Without "-Og", compilation succeeds.
gcc-10 and gcc-11 did not have the problem.

If I comment out XXX_FORCE_INLINE for these two functions, gcc-12 compilation succeeds with -Og.

Comment 1

[ISHIKAWA, Chiaki]

Looking at https://github.com/Perl/perl5/issues/19776,
I thought the issue may be triggered due to "-Og" AND "-g" on the same command line.
But no, even if I remove all "-g" options from the command line, gcc-12 still fails to inline the two always_inline functions when -Og is specified.

Comment 2

[ISHIKAWA, Chiaki]

I think this problem is related to the issue discussed in the following GCC bugzilla.
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107931

So I reported my problem there.
It sound as if there is no remedy.
We may have to disable |always_inline| when the compiler is GCC 12 and -Og is specified on the command line.

Comment 3

[Robert Longson [:longsonr]]

maybe report it to https://github.com/Cyan4973/xxHash

Comment 4

[ISHIKAWA, Chiaki]

(In reply to Robert Longson [:longsonr] from comment #3)

maybe report it to https://github.com/Cyan4973/xxHash

Thank you. Will do.

Comment 5

[ISHIKAWA, Chiaki]

(In reply to ISHIKAWA, Chiaki from comment #4)

(In reply to Robert Longson [:longsonr] from comment #3)

maybe report it to https://github.com/Cyan4973/xxHash

Thank you. Will do.

Reported to the github.
https://github.com/Cyan4973/xxHash/issues/800

I am now sure now if GCC-10 and GCC-11 really inlined the functions because the gcc bugzilla mentions that sometimes errors were/are not reported in GCC. Oh well.

But for my local debugging purpose, getting TB compiled with -Og is more important than inlining these functions.

Comment 6

[ISHIKAWA, Chiaki]

It seems that the issue was taken care of upstream.
Or I am using gcc-14 now and gcc-14 is smarter or more lenient than gcc-12.

I don't see the error any more (I think so. I build two versions of TB locally. One for valgrind testing using gcc, which unfortunately doesn't work very well, the other for ASAN test using clang. "-Og" is passed to valgrind version.)

GCC bug discussion at https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107931
confirmed the existence of the issue (it is NOT a bug. It is a feature documented in GCC manual.), and it is a subtle issue.
It was made clear that gcc-10 and gcc-11 did not have this problem in the gcc bug report discussion.

I am not closing this yet since there will be someone starting to use gcc-12 thinking it is really mature and may get bitten by this issue.

According to GCC site ( https://gcc.gnu.org/ ) the following versions are officially supported versions.:
Supported Releases

GCC 14.2 (changes)
Status: 2024-07-11 (regression fixes & docs only).
GCC 13.3 (changes)
Status: 2024-04-30 (regression fixes & docs only).
GCC 12.4 (changes)
Status: 2024-06-20 (regression fixes & docs only).

Like I said I am using gcc-14.2 now.
I needed to apply some patches to compile the source code due to strange warnings/errors when I pass -Werror to catch as many dubious constructs as possible at compile-time in OTHER parts of the source code, but that is another story.