[Bug rtl-optimization/66614] LRA might fail to eliminate dead code

[bernd.edlinger at hotmail dot de]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66614

--- Comment #2 from Bernd Edlinger <bernd.edlinger at hotmail dot de> ---
this is what I understand from lra in my own words:

lra () consists of a sequence of


  for (;;) {
     ...
          lra_eliminate (false, false);
          /* Do inheritance only for regular algorithms.  */
          if (! lra_simple_p)
            {
              if (flag_ipa_ra)
                {
                  if (live_p)
                    lra_clear_live_ranges ();
                  /* As a side-effect of lra_create_live_ranges, we calculate
                     actual_call_used_reg_set,  which is needed during
                     lra_inheritance.  */
                  lra_create_live_ranges (true, true);
                  live_p = true;
                }
              lra_inheritance ();
            }
          if (live_p)
            lra_clear_live_ranges ();
     ...
    }

and it is finally fixated  by

lra_eliminate (true, false);

which actually replaces the base register, but keeps the offsets as they are

the RTL transformations are done by lra_eliminate_regs_1 ():

first it is called with subst_p=false, update_p=false, full_p=true,
this replaces the offset relative to from_rtx to to_rtx, but keeps the
previous base register.

then it is repeatedly called with subst_p=false, update_p=true, full_p=false,
which adds some minor corrections to the offset, and again keeps the to be
eliminated base register.

finally it is called with subst_p=true, update_p=false, full_p=false,
this time only the base register is replaced but the offsets are already
relative to the target base register.

lra_create_live_ranges calls process_bb_lives repeatedly
which uses this as a pre-condition of a dead insn which and can be eliminated:

      if (dead_insn_p && set != NULL_RTX
          && REG_P (SET_DEST (set)) && REGNO (SET_DEST (set)) >=
FIRST_PSEUDO_REGISTER
          && find_reg_note (curr_insn, REG_EH_REGION, NULL_RTX) == NULL_RTX
          && ! may_trap_p (PATTERN (curr_insn))
          /* Don't do premature remove of pic offset pseudo as we can
             start to use it after some reload generation.  */
          && (pic_offset_table_rtx == NULL_RTX
              || pic_offset_table_rtx != SET_DEST (set)))

now may_trap_p sees the offsets which are not consistent with the base.