https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118401
Bug ID: 118401
Summary: [Offload][OpenMP][OpenACC] Compilation failure with
-finstrument-functions active
Product: gcc
Version: 14.2.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c
Assignee: unassigned at gcc dot gnu.org
Reporter: j.reu...@fz-juelich.de
Target Milestone: ---
Created attachment 60091
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=60091&action=edit
Source files & Makefile to reproduce the issue
Some applications, e.g. performance tools, are interested in instrumenting
other applications to collect information and present them to users. For this,
tool might use compiler plugins, or rely on existing flags.
Many compilers support function instrumentation via -finstrument-functions or
similar flags. This allows tools to implement '__cyg_profile_func_enter' and
'__cyg_profile_func_exit' to get information about the called function and its
callsite.
However, this can cause issues in the case where users try to offload code via
OpenACC or OpenMP. With GCC 13.3.0 and 14.2.0, different error messages can be
observed depending on the optimization level.
A reproducer for this is attached. With a simple 'make' the error should be
reproducible.
I've checked this on two different systems (Arch Linux, GCC 14.2.0 built via
EasyBuild, RTX 3070 & Rocky Linux 9, GCC 13.3.0 built via EasyBuild, A100 )
with the same results.
----
The reproducer includes three different C files with basically the same code.
One variant adds OpenMP target offloading, another OpenACC. The third variant
doesn't do any offloading. The last C file just implements
'__cyg_profile_func_enter' and '__cyg_profile_func_exit' and prints the
pointers.
The Makefile first compiles all files to object files. For the actual "user
code", instrumentation via '-finstrument-functions' is used. This is not active
for the "profiling interface". After that, the interface is linked together
with the user code.
This represents the behavior tools might use to instrument applications.
Here, different errors occur:
With -O0:
gcc -O0 -g -Wall -Wextra -Werror -save-temps -fno-strict-aliasing -fwrapv
-fopenacc -foffload=nvptx-none instrumented-application-openacc.o
profile-interface.o -o instrumented-application-openacc
unresolved symbol __cyg_profile_func_exit
collect2: error: ld returned 1 exit status
nvptx mkoffload: fatal error: x86_64-pc-linux-gnu-accel-nvptx-none-gcc returned
1 exit status
compilation terminated.
lto-wrapper: fatal error:
/data/EasyBuild-develop/software/GCCcore/13.3.0/libexec/gcc/x86_64-pc-linux-gnu/13.3.0//accel/nvptx-none/mkoffload
returned 1 exit status
compilation terminated.
/data/EasyBuild-develop/software/binutils/2.42-GCCcore-13.3.0/bin/ld: error:
lto-wrapper failed
collect2: error: ld returned 1 exit status
make: *** [Makefile:32: instrumented-application-openacc] Error 1
This makes sense, as '__cyg_profile_func_exit' is not implemented for any
accelerator. However, many tools are not interested in doing such thing, and
only want to rely on the host side. A similar issue was encountered with LLVM
some time ago: https://github.com/llvm/llvm-project/issues/59799. Here, the
solution was to allow specifying the flag for the host-side only (via
'-Xarch_host'). I'm not aware of a similar flag for GCC.
With -O1:
gcc -O1 -g -Wall -Wextra -Werror -save-temps -fno-strict-aliasing -fwrapv
-fopenacc -foffload=nvptx-none instrumented-application-openacc.o
profile-interface.o -o instrumented-application-openacc
ptxas ./instrumented-application-openacc.xnvptx-none.mkoffload.o, line 33;
error : Illegal operand type to instruction 'ld'
ptxas ./instrumented-application-openacc.xnvptx-none.mkoffload.o, line 42;
error : Illegal operand type to instruction 'ld'
ptxas ./instrumented-application-openacc.xnvptx-none.mkoffload.o, line 33;
error : Unknown symbol '%frame'
ptxas ./instrumented-application-openacc.xnvptx-none.mkoffload.o, line 42;
error : Unknown symbol '%frame'
ptxas ./instrumented-application-openacc.xnvptx-none.mkoffload.o, line 66;
error : Illegal operand type to instruction 'ld'
ptxas ./instrumented-application-openacc.xnvptx-none.mkoffload.o, line 75;
error : Illegal operand type to instruction 'ld'
ptxas ./instrumented-application-openacc.xnvptx-none.mkoffload.o, line 66;
error : Unknown symbol '%frame'
ptxas ./instrumented-application-openacc.xnvptx-none.mkoffload.o, line 75;
error : Unknown symbol '%frame'
ptxas fatal : Ptx assembly aborted due to errors
nvptx-as: ptxas returned 255 exit status
nvptx mkoffload: fatal error: x86_64-pc-linux-gnu-accel-nvptx-none-gcc returned
1 exit status
compilation terminated.
lto-wrapper: fatal error:
/data/EasyBuild-develop/software/GCCcore/13.3.0/libexec/gcc/x86_64-pc-linux-gnu/13.3.0//accel/nvptx-none/mkoffload
returned 1 exit status
compilation terminated.
/data/EasyBuild-develop/software/binutils/2.42-GCCcore-13.3.0/bin/ld: error:
lto-wrapper failed
collect2: error: ld returned 1 exit status
make: *** [Makefile:32: instrumented-application-openacc] Error 1
This error message is more confusing. I guess something trips here by trying to
instrument with '-finstrument-functions', but this is just pure guessing from
my side.
----
GCC information:
Using built-in specs.
COLLECT_GCC=gcc
COLLECT_LTO_WRAPPER=/data/EasyBuild-develop/software/GCCcore/14.2.0/libexec/gcc/x86_64-pc-linux-gnu/14.2.0/lto-wrapper
OFFLOAD_TARGET_NAMES=nvptx-none
Target: x86_64-pc-linux-gnu
Configured with: ../configure --enable-languages=c,c++,fortran
--without-cuda-driver --enable-offload-targets=nvptx-none --enable-lto
--enable-checking=release --disable-multilib --enable-shared=yes
--enable-static=yes --enable-threads=posix --enable-plugins --enable-gold
--enable-ld=default --prefix=/data/EasyBuild-develop/software/GCCcore/14.2.0
--with-local-prefix=/data/EasyBuild-develop/software/GCCcore/14.2.0
--enable-bootstrap
--with-isl=/data/EasyBuild-develop/build/GCCcore/14.2.0/system-system/gcc-14.2.0/stage2_stuff
--build=x86_64-pc-linux-gnu --host=x86_64-pc-linux-gnu
Thread model: posix
Supported LTO compression algorithms: zlib zstd
gcc version 14.2.0 (GCC)
