Re: [RFC, PATCH][LRA, MIPS] ICE: in decompose_normal_address, at rtlanal.c:5817

[Richard Sandiford <rdsandiford at googlemail dot com>]

Jeff Law <law@redhat.com> writes:
> On 01/09/15 04:32, Robert Suchanek wrote:
>> Hi Steven/Vladimir,
>>
>>> It's hard to say what the correct fix should be, but it sounds like
>>> the address you get after the substitutions should be simplified
>>> (folded).
>>
>> Coming back to the original testcase and re-analyzing the problem, it appears
>> that there is, indeed, a missing case for simplification of LO_SUM/HIGH pair.
>> The patch attached resolves the issue.
>>
>> Although the testcase is not reproducible on the trunk, I think it is still
>> worth to include it in case the ICE reoccurred.
>>
>> The patch has been bootstrapped and regtested on
>> x86_64-unknown-linux-gnu target
>> and similarly tested against SVN revision r212763 where it can be reproduced.
>>
>> Regards,
>> Robert
>>
>> 2015-01-08  Robert Suchanek  <robert.suchanek@imgtec.com>
>>
>> gcc/
>> 	* simplify-rtx.c (simplify_replace_fn_rtx): Simplify (lo_sum (high x)
>> 	(const (plus x offset))) to (const (plus x offset)).
> You have to be careful here.  Whether or not this transformation is 
> valid depends on the size of the offset and whether or not the port has 
> an overlap between its sethi and losum insns and whether or not any 
> rounding occurs when applying the relocations for sethi/losum as well as 
> potentially other factors.
>
> We don't currently have a way for ports to indicate what offsets would 
> make this kind of simplification valid.   In fact, there's at least one 
> port (PA) where the validity of this kind of simplification can't be 
> determined until after LRA/reload when you know the full context of how 
> the result is going to be used.

I agree this is the kind of thing we'd need to consider if we were
deciding whether it's valid to connect a (lo_sum (high x+N) x+N) to an
existing (high x).  But this code is handling cases where the connection
has already been made and we're trying to simplify the result.  Would it
be valid RTL to use:

  (lo_sum (high x) (const (plus x offset)))

to mean anything other than x+offset?

Hmm, admittedly the documentation doesn't guarantee that...

If we do go for the change, I think we should generalise it to handle
(lo_sum (high x+N) x+N') and (lo_sum (high x-N) x) too.  Things like
get_related_value or split_const could help there.

Thanks,
Richard