On Saturday 20 September 2008 12:37, you wrote:
>> The md file states:
>>
>> (define_insn "abssf2"
>> [(set (match_operand:SF 0 "register_operand" "=d,r")
>> (abs:SF (match_operand:SF 1 "register_operand" "0,0")))]
>> ""
>> "@
>> andi %D0,0x7f
>> clt\;bld %D0,7"
>> [(set_attr "length" "1,2")
>> (set_attr "cc" "set_n,clobber")])
>>
>> I don't know how to read it, but loosely interpreted i
>> understand the following:
>> abssf2 seems to be involved in the calculation of the
>> absolute value of a single precesion float. There are
>> two ways to clean the msb of register %DO:
>> - one is and immideate with 0x7f and the
>> - second is clear the T bit and load that zero on msb
>> Now since the first only works on registers r16 to r31
>> sometimes the compiler chooses the second option?
>
> Yes that is exactly correct. "d" means r16..31 and "r" is r0..r31
Nice, would it then not be better to replace
clt\;bld %D0,7"
by:
rol %D0 \; lsr %D0"
This does the same thing, are not more instructions, and leave
the Transfer bit in tact (it conserves carry two, like andi)
Since this is the only place where the T bit is used, it seems
a sensible thing to do and it gives control of that bit back
to the assembly programmers.
Ruud.
_______________________________________________
AVR-GCC-list mailing list
AVR-GCC-list@nongnu.org
http://lists.nongnu.org/mailman/listinfo/avr-gcc-list
