[Bug target/90811] [nvptx] ptxas error on OpenMP offloaded code

[kito at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90811

Kito Cheng <kito at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |kito at gcc dot gnu.org,
                   |                            |wilson at gcc dot gnu.org

--- Comment #15 from Kito Cheng <kito at gcc dot gnu.org> ---
RISC-V got fail after g:26d7a5e690169ac04acde90070b0092c41b71c7e for
gfortran.dg/pr45636.f90

It seems like because alignment change cause simplify_builtin_call can't
simplify memset to load/store.

The story for RISC-V here:

- simplify_builtin_call call can_store_by_pieces check it's OK to store by
pieces?
- can_store_by_pieces call targetm.use_by_pieces_infrastructure_p to ask
back-end
- targetm.use_by_pieces_infrastructure_p call by_pieces_ninsns to calculate how
many instruction needed, it was 1 for RV64*, because alignment is changed to 8
bits from 64 bits for char array due to this patch
g:26d7a5e690169ac04acde90070b0092c41b71c7e.
- by_pieces_ninsns got 4 instead of 1, and large than MOVE_RATIO, failed to
simplify memset to load/store.

- ARM (aarch32) seems like same story too.

* Configure option for RV64: --target=riscv64-unknown-linux-gnu
--with-arch=rv64gc --with-abi=lp64


Analysis:

- align_local_variable changed policy, only ask alignment info (SET_DECL_ALIGN)
from back-end only when expanding (expand pass).
- So that's cause cost measurement changed, because some cost measurement are
depend on the alignment.