This is the mail archive of the mailing list for the GCC project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [Patch, ira] Invalid assert in reload1.c::finish_spills?

On 8/7/19 8:15 AM, Vladimir Makarov wrote:
> On 8/7/19 7:36 AM, wrote:
>> Hi,
>>    gcc/testsuite/c-c++-common/pr60101.c fails with an ICE for the
>>    avr target, because of a gcc_assert firing at reload1.c:4233
>>    The assert (in the patch below) looks bogus to me, as it's in
>>    the if block of
>>      if (! ira_conflicts_p || reg_renumber[i] >= 0)
>>    For this testcase and for the avr target, ira_conflicts_p is
>>    false because build_conflict_bit_table bailed out early
>>    (conflict table too big).
>>    If reg_renumber[i] is now negative, the assert fires and causes
>>    the ICE.
>>    Getting rid of the assert (patch below) makes the ICE go away,
>>    not sure if that's the right fix though.
>>    Comments?
> Mike Matz is right.  Removing the assertion will make the bug even worse
> (changing memory beyond pseudo_previous_regs).
> I did some investigation.  This bug occurred from a 10 years old patch
> avoiding building big conflict tables in IRA.  And the assert was in
> reload before IRA.
> I think the solution should be
[ ... ]
That seems to be working on the embedded targets which probably covers
most of the remaining reload targets.

m68k will bootstrap with it overnight, that's probably the most
interesting test we'll get for this change.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]