This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [lra] patch to revert a code from previous patch.
- From: Richard Sandiford <rdsandiford at googlemail dot com>
- To: Vladimir Makarov <vmakarov at redhat dot com>
- Cc: GCC Patches <gcc-patches at gcc dot gnu dot org>
- Date: Mon, 15 Oct 2012 21:25:19 +0100
- Subject: Re: [lra] patch to revert a code from previous patch.
- References: <507C4A38.3000005@redhat.com> <87bog3lfew.fsf@talisman.home>
Richard Sandiford <rdsandiford@googlemail.com> writes:
> Vladimir Makarov <vmakarov@redhat.com> writes:
>> After committing a patch yesterday to implement proposals from a
>> review, I found that GCC crashes on SPEC2000 gap. LRA is trying to find
>> a mode of operand (const_int 1) in *lea_general_1 insn and can not find
>> it as the operand and insn template operand has VOIDmode.
>>
>> There are still cases when context lookup is necessary to find a mode of
>> the operand. So I am reversing the change I did yesterday.
>>
>> The patch is committed as rev. 192462.
>>
>> 2012-10-15 Vladimir Makarov <vmakarov@redhat.com>
>>
>> * lra-int.h (lra_get_mode): Remove.
>> * lra-constraints.c (find_mode, get_op_mode): New functions.
>> (match_reload): Use get_op_mode instead of lra_get_mode.
>> (process_alt_operands, curr_insn_transform): Ditto.
>
> But my objection to this code still stands. It's wrong to assume
> that an operand to an rtx has the same mode as the containing rtx.
>
> Please add a testcase that shows the problem.
(...because I was hoping to have a look myself). But if that's too
difficult to reduce, then which operand to *lea_general_1 was the problem?
The pattern looks like:
(define_insn_and_split "*lea_general_1"
[(set (match_operand 0 "register_operand" "=r")
(plus (plus (match_operand 1 "index_register_operand" "l")
(match_operand 2 "register_operand" "r"))
(match_operand 3 "immediate_operand" "i")))]
So operands 0, 1 and 2 should have been registers. Operand 3 never
needs reloading, so its mode shouldn't matter.
Richard