https://gcc.gnu.org/bugzilla/show_bug.cgi?id=124907

Christoph Müllner <cmuellner at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |ice-on-valid-code

--- Comment #1 from Christoph Müllner <cmuellner at gcc dot gnu.org> ---
The underlying issue seems to be that dwarf2cfi assumes every path entering the
same code region has the same set of currently saved/restored registers, but
separate shrink-wrapping can violate that assumption:
* shrink-wrap can place save/restore code on only some CFG paths
* when those paths later merge, one predecessor can arrive with a register
still marked as saved, while another arrives with it restored
* in this bug, that happened for AArch64 FP callee-save v15
* dwarf2cfi treated that difference as impossible and aborted in
maybe_record_trace_start

I believe this affects other architectures as well.

Reply via email to