As a follow-up to this discussion (http://gcc.gnu.org/ml/gcc/2012-11/msg00307.html), I have a patch that makes dwarf2out.c emit correct location information for a FP register spanning multiple hard regs.
Before the patch, the DW_AT_location for a local variable on the AVR target (whose FP spans registers R29 and R28) looks like this <4f> DW_AT_location : 2 byte block: 8c 1 (DW_OP_breg28 (r28): 1) After the patch, it looks like this <4f> DW_AT_location : 11 byte block: 6c 93 1 6d 93 1 11 1 22 94 2 (DW_OP_reg28 (r28); DW_OP_piece: 1; DW_OP_reg29 (r29); DW_OP_piece: 1; DW_OP_consts: 1; DW_OP_plus; DW_OP_deref_size: 2) What do you guys think? Regards Senthil diff --git gcc/dwarf2out.c gcc/dwarf2out.c index f0256ae..53a4e56 100644 --- gcc/dwarf2out.c +++ gcc/dwarf2out.c @@ -10875,11 +10875,25 @@ based_loc_descr (rtx reg, HOST_WIDE_INT offset, return new_loc_descr (DW_OP_fbreg, offset, 0); } - if (regno <= 31) - result = new_loc_descr ((enum dwarf_location_atom) (DW_OP_breg0 + regno), - offset, 0); + rtx regs = targetm.dwarf_register_span (reg); + + if (hard_regno_nregs[REGNO (reg)][GET_MODE (reg)] > 1 || regs) + { + result = multiple_reg_loc_descriptor (reg, regs, initialized); + add_loc_descr (&result, new_loc_descr (DW_OP_consts, offset, 0)); + add_loc_descr (&result, new_loc_descr (DW_OP_plus, 0, 0)); + add_loc_descr (&result, new_loc_descr (DW_OP_deref_size, + GET_MODE_SIZE (GET_MODE (reg)), + 0)); + } else - result = new_loc_descr (DW_OP_bregx, regno, offset); + { + if (regno <= 31) + result = new_loc_descr ((enum dwarf_location_atom) (DW_OP_breg0 + regno), + offset, 0); + else + result = new_loc_descr (DW_OP_bregx, regno, offset); + } if (initialized == VAR_INIT_STATUS_UNINITIALIZED) add_loc_descr (&result, new_loc_descr (DW_OP_GNU_uninit, 0, 0));