On Wed, May 3, 2017 at 11:09 AM, Kyrill Tkachov
<kyrylo.tkac...@foss.arm.com> wrote:
> Hi Bin,
>
>
> On 03/05/17 11:02, Bin.Cheng wrote:
>>
>> On Wed, May 3, 2017 at 9:38 AM, Bin.Cheng <amker.ch...@gmail.com> wrote:
>>>
>>> On Wed, May 3, 2017 at 7:17 AM, Eric Botcazou <ebotca...@adacore.com>
>>> wrote:
>>>>>
>>>>> 2017-04-11 Bin Cheng <bin.ch...@arm.com>
>>>>>
>>>>> * rtlanal.c (rtx_cost): Handle TRUNCATE between tieable modes.
>>>>
>>>> This breaks bootstrap with RTL checking:
>>>>
>>>> /home/eric/build/gcc/native/./gcc/xgcc
>>>> -B/home/eric/build/gcc/native/./gcc/ -
>>>> nostdinc -x c /dev/null -S -o /dev/null -fself-
>>>> test=/home/eric/svn/gcc/gcc/testsuite/selftests
>>>> cc1: internal compiler error: RTL check: expected code 'subreg', have
>>>> 'truncate' in rtx_cost, at rtlanal.c:4169
>>>> 0xbae338 rtl_check_failed_code1(rtx_def const*, rtx_code, char const*,
>>>> int,
>>>> char const*)
>>>> /home/eric/svn/gcc/gcc/rtl.c:829
>>>> 0xbbc9b4 rtx_cost(rtx_def*, machine_mode, rtx_code, int, bool)
>>>> /home/eric/svn/gcc/gcc/rtlanal.c:4169
>>>> 0x8517e6 set_src_cost
>>>> /home/eric/svn/gcc/gcc/rtl.h:2685
>>>> 0x8517e6 init_expmed_one_conv
>>>> /home/eric/svn/gcc/gcc/expmed.c:142
>>>> 0x8517e6 init_expmed_one_mode
>>>> /home/eric/svn/gcc/gcc/expmed.c:209
>>>> 0x853fb2 init_expmed()
>>>> /home/eric/svn/gcc/gcc/expmed.c:270
>>>> 0xc45974 backend_init_target
>>>> /home/eric/svn/gcc/gcc/toplev.c:1665
>>>> 0xc45974 initialize_rtl()
>>>>
>>> Sorry for disturbing, I will revert this if can't fix today.
>>
>> It looks bogus and I couldn't find the motivating case for it, so
>> revert with attached patch. Build on x86 and commit as obvious.
>>
>> Thanks,
>> bin
>> 2017-05-03 Bin Cheng <bin.ch...@arm.com>
>>
>> Revert
>> 2017-05-02 Bin Cheng <bin.ch...@arm.com>
>> * rtlanal.c (rtx_cost): Handle TRUNCATE between tieable modes.
>
>
> Looking at the code in the patch...
>
> + case TRUNCATE:
> + /* If we can tie these modes, make this cheap. */
> + if (MODES_TIEABLE_P (mode, GET_MODE (SUBREG_REG (x))))
>
> 'code' here is GET_CODE (x) and in this case it is TRUNCATE.
> SUBREG_REG asserts (in RTL checking mode) that its argument is a SUBREG, so
> passing it a TRUNCATE rtx would cause
> the checking failure Eric reported. I think you meant to use XEXP (x, 0)
> instead of SUBREG_REG (x) ?
Turned out this is still necessary for a test case pr49781-1.c. I
updated patch as suggested. This is kind of an obvious update.
Thanks,
bin
>
> Thanks,
> Kyrill
>
diff --git a/gcc/rtlanal.c b/gcc/rtlanal.c
index 321363f..d9f57c3 100644
--- a/gcc/rtlanal.c
+++ b/gcc/rtlanal.c
@@ -4164,6 +4164,13 @@ rtx_cost (rtx x, machine_mode mode, enum rtx_code
outer_code,
return COSTS_N_INSNS (2 + factor);
break;
+ case TRUNCATE:
+ if (MODES_TIEABLE_P (mode, GET_MODE (XEXP (x, 0))))
+ {
+ total = 0;
+ break;
+ }
+ /* FALLTHRU */
default:
if (targetm.rtx_costs (x, mode, outer_code, opno, &total, speed))
return total;
