https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98190
--- Comment #16 from rguenther at suse dot de <rguenther at suse dot de> --- On Thu, 10 Dec 2020, jakub at gcc dot gnu.org wrote: > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98190 > > --- Comment #15 from Jakub Jelinek <jakub at gcc dot gnu.org> --- > Created attachment 49727 > --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=49727&action=edit > gcc11-pr98190.patch > > So, I have bootstrapped/regtested this patch last night on x86_64, i686, > aarch64, armv7hl, powerpc64le (and s390x still pending) linux. > Unfortunately, on aarch64 it regresses: > gcc.c-torture/execute/pr93213.c > and on powerpc64le that test plus: > g++.dg/warn/Wstrict-aliasing-bogus-char-1.C > gcc.dg/pr87273.c > gcc.dg/torture/pr91656-1.c > gcc.dg/tree-ssa/pr92085-2.c > gcc.dg/tree-ssa/pr94703.c > > Seems the assumption that for promoted SUBREG to_rtx the store is always to > all > the bits is incorrect, e.g. on pr93213.c the memcpy is copying just half of > the bits. So, shall we check the bitpos 0 bitsize all to_rtx bits for the > store_rtx case and otherwise check depending on endianity if the most > significant bit of to_rtx is overwritten and extend in that case? in foo() you mean? For __builtin_memmove (&u16_1, &u128_1, 1); ? So that's a parameter destination - does it at least have correctly DECL_NOT_GIMPLE_REG_P set? Did expansion really do sth different when we had it TREE_ADDRESSABLE?
