http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57351
--- Comment #6 from ktkachov at gcc dot gnu.org --- (In reply to chrbr from comment #5) > > If I print out the regno argument to arm_dbx_register_number, it's 272 which > > seem to me like it could be reg number in DWARF numbering? > > yes it is. strange, I can't see this new failure, with a full newlib build > for > > --enable-languages=c,c++ \ > --with-fpu=neon-vfpv4 --with-float=hard --with-arch=armv7-a > > could you please attach the preprocessed file ? thanks, My apologies, the attachment I added in the bug report is reduced from newlib (I initially said it was from libgfortran). I'm getting the ICE in arm_dbx_register_number on the initial file I attached. On a side note, in your patch you say: - t = one_reg_loc_descriptor (REGNO (XVECEXP (regs, 0, i)), + reg = REGNO (XVECEXP (regs, 0, i)); + t = one_reg_loc_descriptor (DBX_REGISTER_NUMBER (reg), but reg is undeclared in this scope (changing it to "unsigned reg = ..." fixes it) which suggests to me there's some other changes? In your original patch you also moved the #ifdef LEAF_REG_MAP and associated definitions inside the "if (regs == NULL_RTX)" block...