Home Original page

[Fwd: Re: [PATCH] MIPS32 DSP instructions again]

Chao-ying Fu fu@mips.com
Mon Jun 13 22:08:00 GMT 2005 
Hello,

1. I think I put extra tab, so that the table is aligned
for all cases.

2. Someone complained about that I put some magic number
inside my code.  So, I generated the magic number using
the existing mask.  I can change it back to use the number
(ex: 64) directly.

3. I reuse the 'c' format, so that I don't need to create
a new format for the same shift and the same mask and the
same immediate.  I can create a new mask and the new shift,
if you want.

4. Some DSP instructions write or read accumulators which
may be $ac0, $ac1, $ac2, $ac3.  Because $ac0 is HI/LO,
these DSP instructions need the attritubes (WR_HILO/RD_HILO)
to inform the assembler to maintain correct HI/LO dependences.
Or, if the dependences of accumulators and HI/LO are not necessary
at all, I can remove all WR_a and RD_a.

Regards,
Chao-ying

----- Original Message ----- 
From: "Eric Christopher" <echristo@redhat.com>
To: "Chao-ying Fu" <fu@mips.com>
Cc: "Thiemo Seufer" <ths@networkno.de>; <dom@mips.com>; <ian@airs.com>;
"Paul Koning" <pkoning@equallogic.com>; <nigel@mips.com>; "Richard
Sandiford" <rsandifo@nildram.co.uk>; <radhika@mips.com>;
<binutils@sourceware.org>
Sent: Monday, June 13, 2005 2:52 PM
Subject: Re: [Fwd: Re: [PATCH] MIPS32 DSP instructions again]


> On Mon, 2005-06-13 at 14:47 -0700, Chao-ying Fu wrote:
> > Hello,
> >
> > 1. I didn't see any new line after case '5' from my patch.
>
> The tabs in the table.
>
> > 3. For checking the immediate range, my original code is
> > as follows.
> >
> > case ':': /* dsp 7-bit signed immediate in bit 19 */
> >   my_getExpression (&imm_expr, s);
> >   check_absolute_expr (ip, &imm_expr);
> >   if ((unsigned long) (imm_expr.X_add_number + 64) > 127)
> >     {
> >       as_warn (_("DSP immediate not in range -64..63 (%ld)"),
> >                (long) imm_expr.X_add_number);
> >     }
> >
> >   But, I changed the magic number 64 to
> > ((OP_MASK_DSPSFT_7 + 1) >> 1).
> > And, I changed the comparison of 127 to "AND" the inverse
> > of the mask.  Do you prefer my original code or may I put
> > some comments to explain the code?
> >
>
> Both ways are pretty foul. I suppose as long as the warning message
> contains the actual number range I'm ok with it. But why don't you just
> check that the number is in the range, as opposed to the bizarre
> addition and shifts?
>
> > 4. For case 'c', I change the warning message, because
> > our DSP instructions reuse this 'c' format.  The old
> > warning message said "Illegal break code" which is not suitable
> > for DSP instructions.
> >
>
> Why are you reusing?
>
> > 6. WR_a/RD_a reuse WR_HILO/RD_HILO, because DSP accumulators
> > may map to the old HI/LO and we want to maintain correct
> > dependences of reading and writing HI/LO.
>
> What do you mean "map to the old HI/LO"?
>
> -eric
>
>
More information about the Binutils mailing list