Denis Chertykov schrieb:
> 2011/4/14 Georg-Johann Lay <a...@gjlay.de>:
>> The "rotl<mode>3" expanders (mode \in {HI,SI,DI}) violates synopsis by
>> using 4 operands instead of 3. This runs in ICE in top of
>> optabs.c:maybe_gen_insn.
>>
>> The right way to do this is to use match_dup, not match_operand. So
>> the fix is obvious.
>>
>> Regenerated avr-gcc and run against avr testsuite:
>>
>> gcc.target/avr/torture/pr41885.c generates these patterns
>>
>> Johann
>>
>> 2011-04-14 Georg-Johann Lay <a...@gjlay.de>
>>
>> * config/avr/avr.md ("rotlhi3", "rotlsi3", "rotldi3"): Fix
>> expanders operand 3 from match_operand to match_dup.
>
> May be better to use (clobber (match_scratch ...)) here and in
> `*rotw<mode>' and `*rotb<mode>'.
These are splitters that might split before reload, producing strange,
non-matching insns like
(set (scratch:QI) ...
or crashing already in avr_rotate_bytes becase the split condition reads
"&& (reload_completed || <MODE>mode == DImode)"
By the way: this and other patterns implement DImode insns, but there
is *no* movdi in avr backend. That is strongly discouraged, i.e.
implementing insns for FOOmode if there in no movfoo insn that can do
reloading for that mode. It's tempting to omit movdi (because movdi
would be the most complicated DI insn) but it's unhealthy.
IMO the DI stuff should be removed / deactivated as long as there is
no proper support of DImode.
>
> Denis.
>
