Bug #17787: Four AIX build issues with xlc compiler and ruby-3.0.1

• Status: Closed
• Assignee: -
• Target version: -
• ruby -v: ruby 3.0.1p64 (2021-04-05 revision 0fb782ee38) [powerpc-aix7.1.5.0]
• Backport: 2.6: REQUIRED, 2.7: REQUIRED, 3.0: DONE

[lamont (Lamont Granquist)]

Yesterday found 4 small issues trying to use xlc to build ruby 3.0 on AIX that prevented building. My solutions are ugly fixes just to get our own builds going which will not be acceptable, but I'll include links to them.

#1: internal/bits.h uses the gcc extension __builtin_unreachable() directly, this does not exist on xlc, I think it should be replaced by "UNREACHABLE" but I just used an empty statement:

https://github.com/chef/omnibus-software/blob/3120a25111a29503a10866cbd4ee9fedcb438cee/config/patches/ruby/ruby-3.0.1-aix-builtin-unreachable.patch

the ./configure script is correctly finding rb_cv_func___builtin_unreachable=no

#2: With xlc to get cpp output you must use '>' and not '-o' which differs from COUTFLAG and this line in Makefile.in is incorrect and really needs better detection of this feature and not just assuming CPPOUTFLAG = COUTFLAG. I simply hardcoded the right value of CPPOUTFLAG:

https://github.com/chef/omnibus-software/blob/3120a25111a29503a10866cbd4ee9fedcb438cee/config/patches/ruby/ruby-3.0.1-aix-cppoutflag.patch

#3: The xlc compiler actually supports gcc's sync builtins. See the docs here:

https://www.ibm.com/docs/en/SSGH3R_16.1.0/com.ibm.compilers.aix.doc/compiler.pdf#BIFS_GCC_ATOMIC

The checks for this in the configure.ac script are guarded by a check for intel architecture which would need to be properly expanded properly in some way so that those checks would run on aix+xlc as well. I just forced it on:

https://github.com/chef/omnibus-software/blob/3120a25111a29503a10866cbd4ee9fedcb438cee/config/patches/ruby/ruby-3.0.1-aix-force-gcc-sync-builtins.patch

#4: The configure script incorrectly identifies the compiler as supporting -fdeclspec. It might be good to just omit these checks on the xlc compiler. I also tested the patches to fdeclspec which are currently unreleased in master since we needed to backport those to 3.0.1 for M1 mac builds. This one I just removed the fdeclspec checks entirely:

https://github.com/chef/omnibus-software/blob/3120a25111a29503a10866cbd4ee9fedcb438cee/config/patches/ruby/ruby-3.0.1-aix-nofdeclspec.patch

[xtkoba (Tee KOBAYASHI)]

An advice from me is that you should not edit configure directly, as that file is auto-generated from configure.ac which should be edited instead. I am aware that these patches are not intended to be merged as they are.

[lamont (Lamont Granquist)]

Yeah that road requires having all the build tooling installed on our builders, which it turns out we don't (chaotic versions of autoconf and lack of bison on many of them). Went down that road with some ruby-3.0.0-to-3.0-head patches that were obsoleted by the 3.0.1 release and it became painful to try to produce distributions from the sources and then diff them and sort the diffs to produce patches that applied in correct order. It was ultimately easier to hit ./configure with a hammer just to get some builds.

[nobu (Nobuyoshi Nakada)]

• Description updated (diff)
• Status changed from Open to Feedback

lamont (Lamont Granquist) wrote:

#1: internal/bits.h uses the gcc extension __builtin_unreachable() directly, this does not exist on xlc, I think it should be replaced by "UNREACHABLE" but I just used an empty statement:

Probably UNREACHABLE_RETURN is preferable.

#2: With xlc to get cpp output you must use '>' and not '-o' which differs from COUTFLAG and this line in Makefile.in is incorrect and really needs better detection of this feature and not just assuming CPPOUTFLAG = COUTFLAG. I simply hardcoded the right value of CPPOUTFLAG:

Seems fine.

#3: The xlc compiler actually supports gcc's sync builtins. See the docs here:

It isn't considered as gcc?
What is CONFIG["GCC"] in rbconfig.rb file?
If it is "yes", could you show the part for rb_cv_gcc_sync_builtins in config.log file?

#4: The configure script incorrectly identifies the compiler as supporting -fdeclspec. It might be good to just omit these checks on the xlc compiler. I also tested the patches to fdeclspec which are currently unreleased in master since we needed to backport those to 3.0.1 for M1 mac builds. This one I just removed the fdeclspec checks entirely:

Could you show the part for -fdecspec in config.log file?

[lamont (Lamont Granquist)]

fdeclspec section is:

configure:9228: result: yes
configure:9348: checking whether -fdeclspec is accepted as CFLAGS
configure:9375: xlc_r -q64 -c -q64 -I/opt/chef/embedded/include -D_LARGE_FILES -O  -fdeclspec -q64 -I/opt/chef/embedded/include -D_LARGE_FILES -O conftest.c >&5
configure:9375: $? = 0
configure:9400: result: yes
configure:9425: checking whether -fdeclspec is accepted as CXXFLAGS
configure:9457: xlC_r -q64 -o conftest -q64 -I/opt/chef/embedded/include -D_LARGE_FILES -O -fdeclspec -q64 -I/opt/chef/embedded/include -D_LARGE_FILES -O -q64 -L/opt/chef/embedded/lib -Wl,-blibpath:/opt/chef/e
mbedded/lib:/usr/lib:/lib conftest.cpp  >&5
ld: 0706-005 Cannot find or open file: declspec
        ld:fopen(): No such file or directory

later this trainwrecks the configure script with:

configure:14675: checking size of long long
configure:14680: xlc_r -q64 -fdeclspec -o conftest -q64 -I/opt/chef/embedded/include -D_LARGE_FILES -O -q64 -I/opt/chef/embedded/include -D_LARGE_FILES -O -q64 -L/opt/chef/embedded/lib -Wl,-blibpath:/opt/chef/embedded/lib:/usr/lib:/lib conftest.c -lm  >&5
ld: 0706-005 Cannot find or open file: declspec
        ld:fopen(): No such file or directory
configure:14680: $? = 255
configure: program exited with status 255

[lamont (Lamont Granquist)]

xlc does not identify as GCC:

# egrep GCC ./3.0.0/powerpc-aix7.1.5.0/rbconfig.rb
  CONFIG["GCC"] = ""

# egrep gnu config.log
ac_cv_c_compiler_gnu=no
ac_cv_cxx_compiler_gnu=no
rb_cv_prog_gnu_ld=no

IF I REMOVE this check: https://github.com/ruby/ruby/blob/ruby_3_0/configure.ac#L1482 then it works fine:

checking for __atomic builtins... no
checking for __sync builtins... yes
checking for __builtin_unreachable... no

[nobu (Nobuyoshi Nakada)]

Do these patches work?
https://github.com/ruby/ruby/compare/master...nobu:bug/17787-aix-workarounds

[lamont (Lamont Granquist)]

Those work fine, applied here:

https://github.com/chef/omnibus-software/pull/1328/files#diff-a57a53cdb786ca02b297de88fce6855e7928fc0ff4f7e022d0f22185955da1ea

I ditched all my old patches, combined yours with the existing fdeclspec patch in ruby-head (that one would be useful to backport BTW) and applied them not just to AIX but all our build systems and they passed on windows/freebsd/solaris/aix/ubuntu/debian.

[nobu (Nobuyoshi Nakada)]

• Status changed from Feedback to Closed

Applied in changeset git|3a3033c138c0951ad5edbf02b5c233b58a190d2f.

---

get rid of using __builtin_unreachable directly [Bug #17787]

As it is an independent feature from clz, ctz and popcount,
it might be unavailable even if the latters are built in.

[nobu (Nobuyoshi Nakada)]

• Backport changed from 2.5: UNKNOWN, 2.6: UNKNOWN, 2.7: UNKNOWN, 3.0: UNKNOWN to 2.6: REQUIRED, 2.7: REQUIRED, 3.0: REQUIRED

[nagachika (Tomoyuki Chikanaga)]

• Backport changed from 2.6: REQUIRED, 2.7: REQUIRED, 3.0: REQUIRED to 2.6: REQUIRED, 2.7: REQUIRED, 3.0: DONE

ruby_3_0 3e07a4edc92f18dec4806961c1b724ac5cf4d6b2 merged revision(s) 3a3033c138c0951ad5edbf02b5c233b58a190d2f.