Re: [PATCH] Fix PR33642, unrecognizable insn for -frtl-abstract-sequences

["Andrew Pinski" <pinskia@gmail.com>]

On Tue, Apr 8, 2008 at 3:00 AM, Gabor Loki <loki@gcc.gnu.org> wrote:
>  Well, you are probably right. The sequence abstraction algorithm
>  cannot live without insn_*_length functions. I will add HAVE_ATTR_length
>  check.
>
>  BTW, why don't IA-64 have insn_default_length or something similar?
>  I am not familiar with IA64 port, but I guessed insn_*_length functions
>  are not negligible for code size optimization passes.

Well ia64 is not usually a target which you care much about code size.
 Also insn_*_length was only used for branch relaxation so it was not
needed for ia64 or some other targets.  And in some cases it can be
wrong for some targets.

>  Some years ago I used rtx_cost instead of insn_default_length.
>  Unfortunately many ports do not involve optimize_size check to
>  produce size related numbers in their PORT_rtx_cost* functions.
>  And the result wasn't optimal to decrease code size.

I think those targets should be considered broken and fix the issue
there instead of changing to use insn_default_length.

Thanks,
Andrew Pinski