Richard Henderson <r...@redhat.com> writes: > On 02/15/2012 11:53 AM, Richard Sandiford wrote: >> We then trip: >> >> /* Don't let us unwind past the handler context. */ >> gcc_assert (!match_handler); >> >> in _Unwind_RaiseException_Phase2. What's the right thing to do here? >> > > Ug. The Right Thing is to fix the unwinder so that it identifies > frames by pc+cfa, like gdb does, rather than by cfa alone. The > assumption has been that a function without a frame can't perform > a call, and can only by unwound to via signals. > > As a workaround for 4.7, you can try this hack: > > .cfi_startproc simple > .cfi_def_cfa 29, -1 # fake cfa one byte below sp > .cfi_register 29, 29 # "save" sp in itself so we don't use > the fake cfa > move $18,$31 > .cfi_register 31, 18 > ...
Ooh, nice (if that's the word). It certainly fixes the testcase, although I had to use -4 rather than -1 in order to defeat DWARF2_CIE_DATA_ALIGNMENT. That should still be OK, since the stack is 8-byte aligned. GDB doesn't seem to be able to backtrace through this, but that has to come second to correctness. I'll aim to get a tested fix in this weekend. Thanks, Richard
