https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113358
Bug ID: 113358
Summary: OpenMP inhibits vectorization
Product: gcc
Version: 13.2.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c
Assignee: unassigned at gcc dot gnu.org
Reporter: thomas.koopman at ru dot nl
Target Milestone: ---
Created attachment 57054
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=57054&action=edit
The three different versions as well as preprocessed output.
I attached three programs that compute an array accel of points in R^3 from a
3d array of positions in R^3. accel[i] is sum_j ||positions[i] -
positions[j]||^2.
These are seq.c which is the most basic, omp.c which parallelises the outer
loop with OpenMP and block.c which uses the blocking optimisation. The first
version vectorizes as expected, but the other two do not. objdump -d omp.o |
grep ymm shows up empty.
They are compiled with
gcc -c seq.c -Ofast -mavx2 -mfma -save-temps -Wall -Wextra -o seq.o -lm
gcc -c omp.c -Ofast -mavx2 -mfma -save-temps -Wall -Wextra -o omp.o -lm
-fopenmp
gcc -c block.c -Ofast -mavx2 -mfma -save-temps -Wall -Wextra -o block.o -lm
and
gcc -v gives the following.
Using built-in specs.
COLLECT_GCC=gcc
COLLECT_LTO_WRAPPER=/home/thomas/.local/libexec/gcc/x86_64-pc-linux-gnu/13.2.0/lto-wrapper
Target: x86_64-pc-linux-gnu
Configured with: ./configure --prefix=/home/thomas/.local
Thread model: posix
Supported LTO compression algorithms: zlib zstd
gcc version 13.2.0 (GCC)
