https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65979
--- Comment #15 from Oleg Endo <olegendo at gcc dot gnu.org> ---
(In reply to Oleg Endo from comment #14)
>
> I think the check operands[1] / operands[2] check should go into the
> preparation statement. operands[0] is dying after this peephole, so I guess
> this should work:
>
I have confirmed that the following fixes the test case in attachment 35572:
Index: gcc/config/sh/sh.md
===================================================================
--- gcc/config/sh/sh.md (revision 223416)
+++ gcc/config/sh/sh.md (working copy)
@@ -14721,6 +14721,9 @@
|| REGNO (operands[2]) == REGNO (operands[5]))"
[(const_int 0)]
{
+ if (reg_overlap_mentioned_p (operands[1], operands[2]))
+ std::swap (operands[0], operands[2]);
+
sh_check_add_incdec_notes (emit_move_insn (operands[2], operands[3]));
emit_insn (gen_tstsi_t (operands[2],
gen_rtx_REG (SImode, (REGNO (operands[1])))));
However, I think that the emit_move_insn could also be a source of hidden
problems. For instance, if the captured insn
(set (match_operand:SI 2 "arith_reg_dest")
(match_operand:SI 3))
is not a move insn, but some computation like (set (reg) (plus:SI ... )). I'm
not sure what emit_move_insn is going to do in this case. Maybe a better way
would be to copy/re-emit the captured original insn in a different way.
