glibc has always filtered out -O3 CFLAG and replaced it with -O2, relevant portion of the ebuild file: # Lock glibc at -O2; we want to be conservative here. # -fno-strict-aliasing is to work around #155906. filter-flags '-O?' append-flags -O2 -fno-strict-aliasing On the other hand, Clear Linux is compiling glibc with -O3: https://github.com/clearlinux-pkgs/glibc/blob/master/glibc.spec export CFLAGS="-O3 .... Isn't it time to relax the -O flag restriction in glibc? The bug #155906 and its workaround -fno-strict-aliasing seems to be obsolete too. It is now 13 years old anyway.
(In reply to Jan Fikar from comment #0) > glibc has always filtered out -O3 CFLAG and replaced it with -O2, relevant > portion of the ebuild file: > > # Lock glibc at -O2; we want to be conservative here. > # -fno-strict-aliasing is to work around #155906. > filter-flags '-O?' > append-flags -O2 -fno-strict-aliasing > > On the other hand, Clear Linux is compiling glibc with -O3: > > https://github.com/clearlinux-pkgs/glibc/blob/master/glibc.spec > > export CFLAGS="-O3 .... > > > Isn't it time to relax the -O flag restriction in glibc? -O? filter is mostly against -O0/-O1 which tends to break glibc. > The bug #155906 and its workaround -fno-strict-aliasing seems to be obsolete > too. It is now 13 years old anyway. Yeah, probably. That flags is a big hammer and should be an upstream setting if set at all. Can you run glibc testsuite as 'USE=test FEATURES=test emerge -v1 glibc' with and without proposed changes to see if it introduces no regressions?
I have tried 'USE=test FEATURES=test emerge -v1 glibc' with the original ebuild and I'm getting some fails: XPASS: elf/tst-protected1a XPASS: elf/tst-protected1b UNSUPPORTED: io/tst-getcwd-abspath FAIL: misc/check-installed-headers-c FAIL: misc/tst-bz21269 UNSUPPORTED: misc/tst-pkey UNSUPPORTED: nptl/test-cond-printers UNSUPPORTED: nptl/test-condattr-printers UNSUPPORTED: nptl/test-mutex-printers UNSUPPORTED: nptl/test-mutexattr-printers UNSUPPORTED: nptl/test-rwlock-printers UNSUPPORTED: nptl/test-rwlockattr-printers UNSUPPORTED: nss/tst-nss-files-alias-leak UNSUPPORTED: nss/tst-nss-files-hosts-erange UNSUPPORTED: nss/tst-nss-files-hosts-getent UNSUPPORTED: nss/tst-nss-test3 UNSUPPORTED: posix/tst-sysconf-empty-chroot FAIL: resolv/tst-resolv-ai_idn FAIL: resolv/tst-resolv-ai_idn-latin1 UNSUPPORTED: resolv/tst-resolv-threads FAIL: stdio-common/tst-printfsz-islongdouble UNSUPPORTED: sunrpc/tst-svc_register UNSUPPORTED: time/tst-y2039 Summary of test results: 5 FAIL 5857 PASS 16 UNSUPPORTED 17 XFAIL 2 XPASS This is glibc-2.29-r2 with gcc-8.3.0-r1. I guess these fails are expected, aren't they? So now I'll try with -O3 and see, if there are more fails or not, right?
(In reply to Jan Fikar from comment #2) > I have tried 'USE=test FEATURES=test emerge -v1 glibc' with the original > ebuild and I'm getting some fails: > > XPASS: elf/tst-protected1a > XPASS: elf/tst-protected1b > UNSUPPORTED: io/tst-getcwd-abspath > FAIL: misc/check-installed-headers-c > FAIL: misc/tst-bz21269 > UNSUPPORTED: misc/tst-pkey > UNSUPPORTED: nptl/test-cond-printers > UNSUPPORTED: nptl/test-condattr-printers > UNSUPPORTED: nptl/test-mutex-printers > UNSUPPORTED: nptl/test-mutexattr-printers > UNSUPPORTED: nptl/test-rwlock-printers > UNSUPPORTED: nptl/test-rwlockattr-printers > UNSUPPORTED: nss/tst-nss-files-alias-leak > UNSUPPORTED: nss/tst-nss-files-hosts-erange > UNSUPPORTED: nss/tst-nss-files-hosts-getent > UNSUPPORTED: nss/tst-nss-test3 > UNSUPPORTED: posix/tst-sysconf-empty-chroot > FAIL: resolv/tst-resolv-ai_idn > FAIL: resolv/tst-resolv-ai_idn-latin1 > UNSUPPORTED: resolv/tst-resolv-threads > FAIL: stdio-common/tst-printfsz-islongdouble > UNSUPPORTED: sunrpc/tst-svc_register > UNSUPPORTED: time/tst-y2039 > Summary of test results: > 5 FAIL > 5857 PASS > 16 UNSUPPORTED > 17 XFAIL > 2 XPASS > > This is glibc-2.29-r2 with gcc-8.3.0-r1. I guess these fails are expected, > aren't they? Yes, they look benign and probably are locale-dependent. Should be fixed separately. > So now I'll try with -O3 and see, if there are more fails or not, right? Correct. Please do!
It seems the -O0 and -O3 flags still get converted to -O2 by 'strip-flags' (line 351). When I comment strip-flags and the other two lines 375-376 (filter-flags, append-flags) the -Ox seems to be transmitted to glibc correctly. Here are the test results: -O0 glibc itself fails with: 'glibc cannot be compiled without optimization' -O1 I get 3 extra failed tests when compared to baseline -O2: FAIL: gmon/tst-gmon-gprof FAIL: gmon/tst-gmon-pie-gprof FAIL: gmon/tst-gmon-static-gprof the files are smaller compared to baseline (-609,692 bytes as reported by 'equery s glibc | cat') -O2 but without -fno-strict-aliasing produces the same test results as baseline, the files are just slightly bigger (+1846 bytes) -O3 I get 4 extra failed tests when compared to baseline -O2: FAIL: elf/tst-audit1 FAIL: elf/tst-audit2 FAIL: elf/tst-audit3 FAIL: elf/tst-audit8 the files are bigger compared to baseline (+1,857,314 bytes) -Os I get 4 extra failed tests when compared to baseline -O2: FAIL: elf/tst-execstack FAIL: elf/tst-execstack-needed FAIL: elf/tst-execstack-prog FAIL: nptl/tst-execstack the files are smaller compared to baseline (-3,001,954 bytes) Not sure if the extra fails are important or not. Seems we do not need -fno-strict-aliasing flag any more. To stop filtering -Ox flags we need to comment line 351 (strip-flags) and lines 375-376 (filter-flags, append-flags).
(In reply to Jan Fikar from comment #4) > It seems the -O0 and -O3 flags still get converted to -O2 by 'strip-flags' > (line 351). When I comment strip-flags and the other two lines 375-376 > (filter-flags, append-flags) the -Ox seems to be transmitted to glibc > correctly. Here are the test results: > > -O0 glibc itself fails with: 'glibc cannot be compiled without optimization' > > -O1 I get 3 extra failed tests when compared to baseline -O2: > FAIL: gmon/tst-gmon-gprof > FAIL: gmon/tst-gmon-pie-gprof > FAIL: gmon/tst-gmon-static-gprof > the files are smaller compared to baseline (-609,692 bytes as reported by > 'equery s glibc | cat') > > -O2 but without -fno-strict-aliasing produces the same test results as > baseline, the files are just slightly bigger (+1846 bytes) > > -O3 I get 4 extra failed tests when compared to baseline -O2: > FAIL: elf/tst-audit1 > FAIL: elf/tst-audit2 > FAIL: elf/tst-audit3 > FAIL: elf/tst-audit8 > the files are bigger compared to baseline (+1,857,314 bytes) > > -Os I get 4 extra failed tests when compared to baseline -O2: > FAIL: elf/tst-execstack > FAIL: elf/tst-execstack-needed > FAIL: elf/tst-execstack-prog > FAIL: nptl/tst-execstack > the files are smaller compared to baseline (-3,001,954 bytes) > > Not sure if the extra fails are important or not. Seems we do not need > -fno-strict-aliasing flag any more. To stop filtering -Ox flags we need to > comment line 351 (strip-flags) and lines 375-376 (filter-flags, > append-flags). Thank you for the extensive test! Looks reasonable. We can leave only -O0 mangling into -O1.
The bug has been referenced in the following commit(s): https://gitweb.gentoo.org/repo/gentoo.git/commit/?id=5a00ffa513c9cc38af9a6c573005521fada0b6d7 commit 5a00ffa513c9cc38af9a6c573005521fada0b6d7 Author: Sergei Trofimovich <slyfox@gentoo.org> AuthorDate: 2019-05-19 08:33:27 +0000 Commit: Sergei Trofimovich <slyfox@gentoo.org> CommitDate: 2019-05-19 08:39:48 +0000 sys-libs/glibc: drop obsolete -fno-strict-aliasing, bug #686018 Reported-and-tested-by: Jan Fikar Bug: https://bugs.gentoo.org/686018 Package-Manager: Portage-2.3.66, Repoman-2.3.12 Signed-off-by: Sergei Trofimovich <slyfox@gentoo.org> sys-libs/glibc/glibc-9999.ebuild | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-)
The bug has been referenced in the following commit(s): https://gitweb.gentoo.org/repo/gentoo.git/commit/?id=9ecb3b089082f39a15c4c1a315ed91b388a80ade commit 9ecb3b089082f39a15c4c1a315ed91b388a80ade Author: Sergei Trofimovich <slyfox@gentoo.org> AuthorDate: 2019-05-19 09:15:02 +0000 Commit: Sergei Trofimovich <slyfox@gentoo.org> CommitDate: 2019-05-19 09:15:02 +0000 sys-libs/glibc: add USE=custom-cflags, bug #686018 Allow users to pass in more (or less) aggressive CFLAGS for USE=custom-cflags case (disabled by default). Reported-and-tested-by:Jan Fikar Bug: https://bugs.gentoo.org/686018 Package-Manager: Portage-2.3.66, Repoman-2.3.12 Signed-off-by: Sergei Trofimovich <slyfox@gentoo.org> sys-libs/glibc/glibc-9999.ebuild | 25 ++++++++++++++++--------- 1 file changed, 16 insertions(+), 9 deletions(-)
I've added USE=custom-cflags to allow bypassing 'strip-flags' and dropped -O2 forcing. That should allow you to pass more exotic flags. Will be available in next glibc revbump.
