http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48496
--- Comment #19 from Eric Botcazou <ebotcazou at gcc dot gnu.org> 2012-04-09 08:04:40 UTC --- > However, recog.c contains in various places checks that will *accept* -during > reload- a pseudo in places where a memory constraint is required; exactly > because such pseudos will actually get replaced by a MEM: > > case TARGET_MEM_CONSTRAINT: > [snip] > /* During reload, accept a pseudo */ > else if (reload_in_progress && REG_P (op) > && REGNO (op) >= FIRST_PSEUDO_REGISTER) > win = 1; > > Note that those checks were originally added in the same patch that added this > asm validity check ... OK, this makes sense. > So really that validity check shouldn't have failed just because of the > presence of a spilled pseudo. The question is, why doesn't this work for the > ia64 test case as expected? Because the reload insns are of the form: (insn 119 66 120 4 (set (reg:DI 136 f8) (reg:DI 406 [ MEM[(const mp_limb_t *)mip_6(D) + 8B] ])) pr52657.c:27 5 {movdi_internal} and the matching alternative would be (f, Q) with ;; Note that while this accepts mem, it only accepts non-volatile mem, ;; and so cannot be "fixed" by adjusting the address. Thus it cannot ;; and does not use define_memory_constraint. (define_constraint "Q" "Non-volatile memory for FP_REG loads/stores" (and (match_operand 0 "memory_operand") (match_test "!MEM_VOLATILE_P (op)"))) bool insn_extra_memory_constraint (enum constraint_num c) { switch (c) { case CONSTRAINT_S: return true; default: break; } return false; }