https://gcc.gnu.org/bugzilla/show_bug.cgi?id=122844
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Target|riscv aarch64 |riscv aarch64 x86_64-*-*
--- Comment #5 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Andrew Pinski from comment #3)
> Also reproducible on aarch64 with `-O3 -march=armv9-a -fno-vect-cost-model` .
And with -O3 -march=znver4 (aka AVX512 plus a masked epilog).
The testcase requires both inlining into main and unswitching to miscompile.
When enabling ranger I wondered whether we eventually mess up "context" when
transforming multiple loops since we do not fully analyze all loops in a
function, but analyze + transform one loop after each other. We are also
not removing "dead" IVs and their value ranges might no longer be correct
(but I fail to see how this could affect later code).
