[PATCH][AArch64] PR 68102: Check that operand is REG before checking the REGNO in mov-immediate splitters

[Kyrill Tkachov]

Hi James,

On 27/10/15 18:26, James Greenhalgh wrote:
> On Tue, Oct 27, 2015 at 06:06:12PM +0000, Kyrill Tkachov wrote:
>> Hi all,
>>
>> This is another RTL checking error occuring in the splitting condition of the
>> mov-immediate patterns.  We take a REGNO of operands[0] which is a
>> nonimmediate_operand.  Since the immediate splitting code only makes sense
>> when the destination is a register, we should be guarding that condition on
>> REG_P (operands[0]).
>>
>> The reported error occurs on the *movdi_aarch64 pattern but I see the same
>> vulnerability in the *movsi_aarch64 pattern, although I wasn't able to get it
>> to trigger an ICE.
>>
>> This patch adds a REG_P check on the splitting condition of both.  The
>> testcase (taken from the BZ for PR 68102 and with an #if 1 removed)now
>> compiles fine on an aarch64 compiler with RTL checking enabled.
>> Bootstrapped and tested on aarch64-linux with RTL checking enabled.
>>
>> Ok for trunk?
> OK.
>
>> Thanks,
>> Kyrill
>>
>> The BZ says this occurs on the GCC 5 branch but I don't have a checking
>> compiler from that branch yet. I'll be investigating whether to backport this
>> patch there in the meantime.
> Sounds good to me.

So I reproduced the checking ICE on the GCC 5 and the patch applies
cleanly there and fixes it.

So ok to commit it there after a bootstrap and test on that branch?

Thanks,
Kyrill

> Thanks,
> James
>
>> 2015-10-27  Kyrylo Tkachov  <kyrylo.tkachov@arm.com>
>>
>>      PR target/68102
>>      * config/aarch64/aarch64.md (*movsi_aarch64): Check that
>>      operands[0] is a reg before taking its REGNO in split condition.
>>      (*movdi_aarch64): Likewise.
>>
>> 2015-10-27  Kyrylo Tkachov  <kyrylo.tkachov@arm.com>
>>
>>      PR target/68102
>>      * gcc.target/aarch64/pr68102_1.c: New test.