Porting question: define_expand Vs define_split

Porting question: define_expand Vs define_split

[Balaji S]

Hi all,
I am porting GCC to a new architecture using stormy16 as base.

In my architecture there is no instruction for /neg/ operation. So i planned to handle it using /sub/. I am able to accomplish this job by using either define_expand or define_split without knowing its internals as follows:

(define_expand "neghi2"
  [(set (match_operand:HI 0 "register_operand" "")
	(const_int 0))
   (set (match_dup 0)
	(minus:HI (match_dup 0)
		  (match_operand:HI 1 "register_operand" "")))]
  ""
  "")

(define_insn_and_split "neghi2"
  [(set (match_operand:HI 0 "register_operand" "=r")
	(neg:HI (match_operand:HI 1 "register_operand" "0")))]
  ""
  "#"
  "reload_completed"
  [(set (match_dup 0)
	(const_int 0))
   (set (match_dup 0)
	(minus:HI (match_dup 0)
		  (match_dup 1)))]
  "")

I do not know whether the above m/c descriptions are correct. Please correct me if any mistakes.

Moreover, i would like to know, when to use /define_expand/ and when to use /define_split/.


Thanks in advance,
Balaji S

Re: Porting question: define_expand Vs define_split

[DJ Delorie]

Choosing between define_expand and define_insn_and_split IMHO depends
on whether gcc can optimize better with the split-up insns or with the
combined insn.  In cases with linear math, it's probably better to
define_expand.  In cases where you're splitting up a wide op into
pairs of narrow ops (like movdi->movsi), it's probably better to
define_insn_and_split (and defer until after reload) so gcc can see
the wider ops during optimization.

In the case of a missing neg, though, you can use a not/inc pair if
your machine is 2s-compliment and avoid the need for a scratch
register to hold the zero or op1 (it can be done in place).  You'd
still use a define_expand:

(define_expand "neghi2"
  [(set (match_dup 2)
	(not:HI (match_operand:HI 1 "" "")))
   (set (match_operand:HI 0 "" "")
	(plus:HI (match_dup 2)
                 (const_int 1)))]
  ""
  "operands[2] = gen_reg_rtx (HImode);"
  )

Add "register_operand" if your not/inc ops need registers.

Re: Porting question: define_expand Vs define_split

[Daniel Jacobowitz]

On Wed, Nov 03, 2004 at 12:11:13PM +0530, Balaji S wrote:
> Hi all,
> I am porting GCC to a new architecture using stormy16 as base.
> 
> In my architecture there is no instruction for /neg/ operation. So i 
> planned to handle it using /sub/. I am able to accomplish this job by using 
> either define_expand or define_split without knowing its internals as 
> follows:
> 
> (define_expand "neghi2"
>  [(set (match_operand:HI 0 "register_operand" "")
> 	(const_int 0))
>   (set (match_dup 0)
> 	(minus:HI (match_dup 0)
> 		  (match_operand:HI 1 "register_operand" "")))]
>  ""
>  "")
> 
> (define_insn_and_split "neghi2"
>  [(set (match_operand:HI 0 "register_operand" "=r")
> 	(neg:HI (match_operand:HI 1 "register_operand" "0")))]
>  ""
>  "#"
>  "reload_completed"
>  [(set (match_dup 0)
> 	(const_int 0))
>   (set (match_dup 0)
> 	(minus:HI (match_dup 0)
> 		  (match_dup 1)))]
>  "")
> 
> I do not know whether the above m/c descriptions are correct. Please 
> correct me if any mistakes.

The splitter is not correct.  You have constrained the input and output
to be in the same register, and are overwriting that register with 0. 
The constraints would have to be "=&r" and "r".

-- 
Daniel Jacobowitz