Add flags to find libatomic in build-tree testing, fixing a catastrophic
libgomp testsuite failure with targets such as `riscv-linux-gnu' that
imply `-latomic' with the `-pthread' GCC option, implied in turn by the
`-fopenacc' and `-fopenmp' options, removing failures like:
.../bin/riscv64-linux-gnu-ld: cannot find -latomic
collect2: error: ld returned 1 exit status
compiler exited with status 1
FAIL: libgomp.c/../libgomp.c-c++-common/atomic-18.c (test for excess errors)
Excess errors:
.../bin/riscv64-linux-gnu-ld: cannot find -latomic
UNRESOLVED: libgomp.c/../libgomp.c-c++-common/atomic-18.c compilation failed to
produce executable
and bringing overall test results for the said target (here with the
`x86_64-linux-gnu' host and RISC-V QEMU in the Linux user emulation mode
as the target board) from:
=== libgomp Summary ===
# of expected passes90
# of unexpected failures3267
# of expected failures 2
# of unresolved testcases 3247
# of unsupported tests 548
to:
=== libgomp Summary ===
# of expected passes6834
# of unexpected failures4
# of expected failures 4
# of unsupported tests 518
libgomp/
* testsuite/lib/libgomp.exp (libgomp_init): Add flags to find
libatomic in build-tree testing.
---
Hi Jakub,
Thank you for your review.
On Mon, 18 Nov 2019, Jakub Jelinek wrote:
> > --- gcc.orig/libgomp/testsuite/lib/libgomp.exp
> > +++ gcc/libgomp/testsuite/lib/libgomp.exp
> > @@ -174,6 +174,16 @@ proc libgomp_init { args } {
> > # For build-tree testing, also consider the library paths used for
> > builing.
> > # For installed testing, we assume all that to be provided in the
> > sysroot.
> > if { $blddir != "" } {
> > + # Offload options imply `-pthread', and that implies `-latomic'
> > + # on some targets, so wire in libatomic build directories.
>
> -fopenmp is not an option I'd like to call Offload option, OpenMP is
> much more than just offloading, and this isn't on some targets, but only
> one, riscv*.
> So, I think it should be done only for that target and talk about
> -fopenmp/-fopenacc options instead of Offload options.
Thanks for straightening me out on OpenMP; I have adjusted the comment
and the change description accordingly.
I think it makes no sense to require people to update this test harness
if a future target also requires access to libatomic; the `-L' option and
the dynamic loader path are harmless if not required, as libatomic won't
actually be pulled unless otherwise requested and there is no other
library in libatomic's build directory that could be used unintentionally.
Whereas in the absence of these settings a random version of libatomic
already present somewhere on the system may accidentally be used and
affect test results somehow.
Why do you think it is important that it is only the relevant (i.e.
`riscv*') targets that the directory providing newly-built libatomic is
pointed at ?
Maciej
Changes from v1:
- Replaced references to "offload options" with "`-fopenacc' and
`-fopenmp' options".
---
libgomp/testsuite/lib/libgomp.exp | 11 +++
1 file changed, 11 insertions(+)
gcc-test-libgomp-atomic-lib-path.diff
Index: gcc/libgomp/testsuite/lib/libgomp.exp
===
--- gcc.orig/libgomp/testsuite/lib/libgomp.exp
+++ gcc/libgomp/testsuite/lib/libgomp.exp
@@ -174,6 +174,17 @@ proc libgomp_init { args } {
# For build-tree testing, also consider the library paths used for builing.
# For installed testing, we assume all that to be provided in the sysroot.
if { $blddir != "" } {
+ # The `-fopenacc' and `-fopenmp' options imply `-pthread', and
+ # that implies `-latomic' on some targets, so wire in libatomic
+ # build directories.
+ set shlib_ext [get_shlib_extension]
+ set atomic_library_path "${blddir}/../libatomic/.libs"
+ if { [file exists "${atomic_library_path}/libatomic.a"]
+|| [file exists \
+"${atomic_library_path}/libatomic.${shlib_ext}"] } {
+ lappend ALWAYS_CFLAGS "additional_flags=-L${atomic_library_path}"
+ append always_ld_library_path ":${atomic_library_path}"
+ }
global cuda_driver_include
global cuda_driver_lib
if { $cuda_driver_include != "" } {
