https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107389
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Keywords| |missed-optimization
CC| |rguenth at gcc dot gnu.org
--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
Yes, that's definitely a no-go since pointer types have no semantic meanings.
I don't understand s390 assembly much but I see calls to __atomic_load_16@PLT
for both -O0 and -O2 so I wonder if it is the testcase you are seeing a
difference for?
If __atomic_load_n (and others) imply certain alignment of the pointer
arguments then their expanders have to explicitely add that. But I guess
there's no requirement that __builtin_atomic_load (p, 16) has p 16 byte
aligned?
There was talks to "rectify" specified type alignment for pointers in
function arguments but IIRC that has never materialized. See duplicate
bugreports about memcpy involving properly align-typed function parameters.
That said, iff __atomic_* constraints are not exploited we should do that.
