[PATCH, MIPS] Add baddu Octeon instruction

[Adam Nemet]

Richard Sandiford wrote:
> (define_predicate "truncate_operator"
>   (ior (and (match_code "subreg")
>             (match_test "subreg_lowpart_p (op)")
>             (match_test "TRULY_NOOP_TRUNCATION
>                           (GET_MODE_BITSIZE (GET_MODE (op)),
>                            GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (op))))"))
>        (match_code "truncate")))
> 
> ?  Then a single pattern:
> 
> (define_insn "*baddu<mode>si3"
>   [(set (match_operand:SI 0 "register_operand" "=d")
> 	(zero_extend:SI
> 	 (match_operator:QI 1 "truncate_operator"
> 	  [(plus:GPR (match_operand:GPR 2 "register_operand" "d")
> 		     (match_operand:GPR 3 "register_operand" "d"))])))]
>   "ISA_HAS_BADDU"
>   "baddu\\t%0,%2,%3"
>   [(set_attr "type" "arith")])

Really cool and it works too!

Surprisingly I ran into a reload assert with this.  Reload wants to
unconditionally reload the argument of a subreg if it is a PLUS and we
fail badly.  I am planning to submit the reload part of the patch
separately after testing it on x86 too.

I ended up not adding the TRULY_NOOP_TRUNCATION part.  I think it's
legitimate to have (subreg:QI (plus:DI ...) <lowpart byte>) and
truncate_operator should match this (unlike my patterns which wouldn't).
 The idea being that if we can prove that the DI addition results in a
valid SI value we can just truncate with subreg.  What do you think?

Adam

-------------- next part --------------
A non-text attachment was scrubbed...
Name: baddu-2.patch
Type: text/x-patch
Size: 3755 bytes
Desc: not available
URL: <http://gcc.gnu.org/pipermail/gcc-patches/attachments/20080828/6cbacf0f/attachment.bin>