Bug 822369

Summary:	sys-libs/libomp-13.0.0 fails to build with USE="offload abi_x86_32" LLVM_TARGETS=NVPTX
Product:	Gentoo Linux	Reporter:	Fat-Zer <fatzer2>
Component:	Current packages	Assignee:	Bernard Cafarelli <voyageur>
Status:	RESOLVED FIXED
Severity:	normal	CC:	ionen, llvm, mgorny
Priority:	Normal	Keywords:	PATCH
Version:	unspecified
Hardware:	All
OS:	Linux
Whiteboard:
Package list:		Runtime testing required:	---
Attachments:	build.log emerge --info '=sys-libs/libomp-13.0.0::gentoo' environment proposed patch

Description Fat-Zer 2021-11-08 00:01:37 UTC

subj;

It looks like the assumption in the ebuild comments is not true anymore:

 97         # this is non-fatal and libomp checks for CUDA conditionally
 98         # to ABI, so we can just ignore passing the wrong value
 99         # on non-amd64 ABIs

So the offloading should be enabled only for the amd64 ABI.

Reproducible: Always

Steps to Reproduce:
1. USE="cuda offload abi_x86_32" emerge --info '=sys-libs/libomp-13.0.0::gentoo'
Actual Results:  
Configuration error

Expected Results:  
Compilation

Comment 1 Fat-Zer 2021-11-08 00:04:24 UTC

Created attachment 749469 [details]
build.log

Comment 2 Fat-Zer 2021-11-08 00:05:10 UTC

Created attachment 749472 [details]
emerge --info '=sys-libs/libomp-13.0.0::gentoo'

Comment 3 Fat-Zer 2021-11-08 00:07:06 UTC

Created attachment 749475 [details]
environment

Comment 4 Fat-Zer 2021-11-08 04:13:26 UTC

Created attachment 749481 [details, diff]
proposed patch

Comment 5 Michał Górny archtester

2021-11-08 08:31:09 UTC

What nvidia-cuda-toolkit version are you using?  I'm surprised you have /usr/lib/libcuda.so.

Comment 6 Ionen Wolkens gentoo-dev

2021-11-08 08:49:08 UTC

(In reply to Michał Górny from comment #5)
> What nvidia-cuda-toolkit version are you using?  I'm surprised you have
> /usr/lib/libcuda.so.
/usr/lib/libcuda.so is installed by nvidia-drivers[abi_x86_32], not nvidia-cuda-toolkit

Comment 7 Fat-Zer 2021-11-08 09:31:13 UTC

(In reply to Michał Górny from comment #5)
> What nvidia-cuda-toolkit version are you using?

   $ emerge -pqv nvidia-cuda-toolkit
   [ebuild   R   ] dev-util/nvidia-cuda-toolkit-11.5.0  USE="-debugger -nsight -profiler -sanitizer -vis-profiler"

> I'm surprised you have /usr/lib/libcuda.so.

As Ionen Wolkens noticed the libcuda.so is installed by nvidia-drivers:

  $ equery b /usr/lib/libcuda.so
  * Searching for /usr/lib/libcuda.so ...
  x11-drivers/nvidia-drivers-495.44 (/usr/lib/libcuda.so.495.44)
  x11-drivers/nvidia-drivers-495.44 (/usr/lib/libcuda.so -> libcuda.so.495.44)

Comment 8 Michał Górny archtester

2021-11-08 09:57:49 UTC

Ok, it turned out I couldn't reproduce because I had ccache enabled... and the stupid package assumes that llvm-link must be present in the same directory as the compiler.  Since it wasn't present in the ccache shadow directory, the check for llvm-link failed and it disabled nvidia stuff.

Comment 9 Larry the Git Cow gentoo-dev

2021-11-08 10:34:33 UTC

The bug has been closed via the following commit(s):

https://gitweb.gentoo.org/repo/gentoo.git/commit/?id=49fe1591b8692a955419f9bd944aa51b5e4f19b4

commit 49fe1591b8692a955419f9bd944aa51b5e4f19b4
Author:     Michał Górny <mgorny@gentoo.org>
AuthorDate: 2021-11-08 10:18:39 +0000
Commit:     Michał Górny <mgorny@gentoo.org>
CommitDate: 2021-11-08 10:34:28 +0000

    sys-libs/libomp: Selectively enable deviceRTLs based on CHOST
    
    deviceRTLs are only supported on amd64, aarch64 and ppc64le, and trying
    to build them elsewhere may cause CMake failures.  Check CHOST
    to determine whether to enable them.
    
    Closes: https://bugs.gentoo.org/822369
    Signed-off-by: Michał Górny <mgorny@gentoo.org>

 sys-libs/libomp/libomp-13.0.0-r1.ebuild   | 32 +++++++++++++++++--------------
 sys-libs/libomp/libomp-13.0.0.9999.ebuild | 32 +++++++++++++++++--------------
 sys-libs/libomp/libomp-14.0.0.9999.ebuild | 32 +++++++++++++++++--------------
 3 files changed, 54 insertions(+), 42 deletions(-)