I am working toward efficient support of some of the sfixed point
types on avr architecture.
I started out by writing an assembly library to handle all of the
conversions to and from various fixed point, integer, and floating
point types. In many cases the functions are only 1 or 2
instructions, so I would like to implement them in the machine
descriptor language so there isn't a function call needed, and the
registers can be used more efficiently.
I was able to get addition of fixed point rules working, eg:
(define_insn "addqq3"
[(set (match_operand:QQ 0 "register_operand" "=r")
(plus:QQ (match_operand:QQ 1 "register_operand" "%0")
(match_operand:QQ 2 "register_operand" "r")))]
""
"add %0,%2"
[(set_attr "length" "1")
(set_attr "cc" "set_czn")])
This rule works correctly and adds numbers. The problem is I can't
get conversions to work, and I'm wondering if something is missing in
the compiler elsewhere. Here is what I tried:
(define_insn "fractqiqq"
[(set (match_operand:QQ 0 "register_operand" "=r")
(fract_convert:QQ (match_operand:QI 1 "register_operand"
"0")))]
""
"clr %0"
[(set_attr "length" "1")
(set_attr "cc" "set_czn")])
However, it always generates "rcall __fractqiqq" so it isn't using my
rule. I also found that the fixed point conversions specified at
"http://gcc.gnu.org/onlinedocs/gccint/Conversions.html"; are not used
by any backend currently, so I don't have any examples to work with.
Has anyone used any of the following in machine description?
fract_convert
sat_fract
unsigned_fract_convert
unsigned_sat_fract
I would like some examples.
Thanks,
Sean
