Re: Another NaT bit propagation bug.

[Richard Henderson <rth at redhat dot com>]

On Wed, Nov 07, 2001 at 06:13:47AM -0800, amacleod@cygnus.com wrote:
> The simplest way to handle this is to look for any registers which
> are live on entry to the function, check to see if their first use is
> in a zero_extract or subreg expression, and if so, set them to 0 immediately
> following the prologue. Since this is specific to ia64, I've put the
> code into ia64_reorg.

I don't think this is really ia64 specific.  Well, the NaT death
is, but not the general problem.  Consider

idefloppy_bcount_reg_t xxx(void)
{
        idefloppy_bcount_reg_t bcount;

        bcount.all = 0;
        bcount.b.high = get_char[0];
        bcount.b.low = get_char[1];

        return bcount;
}

On Alpha, we get

	# $1 has get_char[0]; $3 has get_char[1]; $2 is undefined.
        zapnot $2,12,$2
        sll $1,8,$1
        bis $2,$1,$2
        bis $2,$3,$2
        stl $2,0($16)

I.e. we do see the unnecessary masking.  Contrast this with

        union {
                int x;
                idefloppy_bcount_reg_t bcount;
        } u;

        u.x = 0;
        u.bcount.all = 0;
        u.bcount.b.high = get_char[0];
        u.bcount.b.low = get_char[1];

	# $1 has get_char[0]; $2 has get_char[1].
        sll $1,8,$1
        bis $1,$2,$1
        stl $1,0($16)

It would be nice to handle this in a generic fashion before 
combine, so that combine can simplify the masking.


r~