This was the most glaring case, and would result in LRA crashing if this code snippet was actually hit on big-endian, since simplify_gen_subreg() will return NULL in such a case and then we try to blindly emit a move to 'subreg'.
There is code in match_reload which seems to have a similar problem, specifically these sequences: if (SCALAR_INT_MODE_P (inmode)) new_out_reg = gen_lowpart_SUBREG (outmode, reg); else new_out_reg = gen_rtx_SUBREG (outmode, reg, 0); ... if (SCALAR_INT_MODE_P (outmode)) new_in_reg = gen_lowpart_SUBREG (inmode, reg); else new_in_reg = gen_rtx_SUBREG (inmode, reg, 0); But I have not tried to address those cases in this patch. Vlad, is this OK to commit? 2015-09-14 David S. Miller <da...@davemloft.net> * lra-constraints.c (simplify_operand_subreg): Do not assume that lowpart of a SUBREG has offset zero. diff --git a/gcc/lra-constraints.c b/gcc/lra-constraints.c index cdb2695..fc8e43d 100644 --- a/gcc/lra-constraints.c +++ b/gcc/lra-constraints.c @@ -1545,7 +1545,7 @@ simplify_operand_subreg (int nop, machine_mode reg_mode) bool insert_before, insert_after; PUT_MODE (new_reg, mode); - subreg = simplify_gen_subreg (innermode, new_reg, mode, 0); + subreg = gen_lowpart_SUBREG (innermode, new_reg); bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg)); insert_before = (type != OP_OUT);