diff options
author | jeanPerier <jperier@nvidia.com> | 2025-10-08 10:36:52 +0200 |
---|---|---|
committer | GitHub <noreply@github.com> | 2025-10-08 10:36:52 +0200 |
commit | f59736758113c568908c6b9293622c530d6ab827 (patch) | |
tree | 503b6831d8235affca466667fbeb5dcb477458a5 /llvm/lib/Bitcode/Reader/BitcodeReader.cpp | |
parent | 1ab69d7004fe8b404c1f83d306838442f549e1ee (diff) | |
download | llvm-f59736758113c568908c6b9293622c530d6ab827.zip llvm-f59736758113c568908c6b9293622c530d6ab827.tar.gz llvm-f59736758113c568908c6b9293622c530d6ab827.tar.bz2 |
[flang][HLFIR] add skip_rebox option to hlfir.declare (#162305)
hlfir.declare with a fir.box input always lead to a fir.rebox being
created to ensure the lower bounds and attributes are set correctly in
the local descriptor for the entity. This is really needed for
hlfir.declare using fir.box function argument that do not come with any
guarantees with regards to the lower bounds.
Sometimes however, this fir.rebox just adds a lot of noise in the SSA
chain, especially at the LLVM level and it is known that the input
descriptor is already correct.
I am making this patch in the context of OpenACC where I want to remap
the variables inside the compute region, creating a fir.rebox on the
way. This fir.rebox cannot be optimized away by FIR because of the
OpenACC ops in the SSA chain.
This patch adds a flag to indicate the the fir.box is known to have the
correct lower bounds and attributes so that it can have a simpler code
generation to FIR.
Diffstat (limited to 'llvm/lib/Bitcode/Reader/BitcodeReader.cpp')
0 files changed, 0 insertions, 0 deletions