hi, Richard,
This version of patch showed no regressions on mipsel-unknown-linux-gnu.
                         Thanks, Dinar.


On Wed, Sep 21, 2011 at 4:59 PM, Richard Sandiford
<rdsandiford@googlemail.com> wrote:
> Dinar Temirbulatov <dtemirbulatov@gmail.com> writes:
>> @@ -14696,7 +14696,11 @@ mips_avoid_hazard (rtx after, rtx insn,
>>       clobber hi and lo.  */
>>    if (*hilo_delay < 2 && reg_set_p (lo_reg, pattern))
>>      nops = 2 - *hilo_delay;
>> -  else if (*delayed_reg != 0 && reg_referenced_p (*delayed_reg, pattern))
>> +  else if ((*delayed_reg != 0 && reg_referenced_p (*delayed_reg, pattern))
>> +         && !((GET_MODE (*delayed_reg) == DFmode && set_after != 0
>> +            && (set = single_set (insn)) != NULL_RTX && GET_MODE (SET_DEST(set)) == DFmode
>> +            && XINT((XEXP (pattern, 1)), 1) == UNSPEC_LOAD_HIGH
>> +            && XINT((XEXP (*set_after, 1)), 1) == UNSPEC_LOAD_LOW)))
>
> This isn't safe because the patterns might not be UNSPECs (so XINT (...)
> would be meaningless).  It's better to check the insn code instead.
> Something like:
>
>  else if (*delayed_reg != 0
>           && reg_referenced_p (*delayed_reg, pattern)
>           && !(recog_memoized (insn) == CODE_FOR_load_highdf
>                && recog_memoized (*set_after) == CODE_FOR_load_lowdf))
>
> (untested).  Note that *set_after should always be nonnull if
> *delayed_reg is.
>
> Looks good otherwise.
>
> Richard
>