[patch, avr] Fix ICE PR64452 pushing eliminated rtxes

[Georg-Johann Lay]

Am 02/17/2015 um 03:34 PM schrieb Denis Chertykov:
> 2015-02-17 14:12 GMT+03:00 Georg-Johann Lay <avr@gjlay.de>:
>> Byte-wise pushing virtual regs like arg pointer migth result in patterns
>> like
>>
>>   (set (mem:QI (post_dec:HI (reg:HI 32 SP)))
>>        (subreg:QI (plus:HI (reg:HI 28)
>>                            (const_int 17)) 0))
>>
>> after elimination.
>>
>> Attached patch uses new pushhi1_insn to push virtuals in HImode so that
>> expressions like in subreg_reg from above can be reloaded.
>>
>> Ok to commit ?
>>
>> Johann
>>
>>          PR target/64452
>>
>>          * config/avr/avr.md (pushhi_insn): New insn.
>>          (push<mode>1): Push virtual regs in one chunk using pushhi1_insn.
>
> Approved.
> (But I'm worry about this because it's reload related problem and it
> can have a side effect)
>
> Denis.

So you have a superior solution in mind?

What side effects specifically?

Currently the side effect is that reload gets simpler expressions and hence 
does not ICE.  There isn't even an insn that can push complex (plus rtx in this 
case) expressions or subregs thereof.  Even if there were such insns I don't 
think reload is supposed to handle them.

The current implementation of push<mode>1 assumes that all RTXes which ever 
appear in a push can be decomposed into subregs and these can be simplified to 
some of the push insns, i.e. the push operand simplifies to REG or CONST0_RTX. 
  The subreg above, however, cannot be simplified to anything reload can handle 
and does not match an insn.  And supplying such an insn is pointless because 
that insn would need a scratch and hence require secondary reloads...

plus rtxes are special as they might be produced by reload (R28 above is 
(hard_)frame_pointer).  For similar reason there are two addhi3 insns (one 
without scratch to accommodate reload and one generic with scratch for better 
performance.)


Johann