Re: [PATCH v3 18/37] target/ppc: implement vgnb

[Richard Henderson]

On 2/10/22 23:34, matheus.ferst@eldorado.org.br wrote:
> +    for (int dw = 1; dw >= 0; dw--) {
> +        get_avr64(vrb, a->vrb, dw);
> +        for (; in >= 0; in -= a->n, out--) {
> +            if (in > out) {
> +                tcg_gen_shri_i64(tmp, vrb, in - out);
> +            } else {
> +                tcg_gen_shli_i64(tmp, vrb, out - in);
> +            }
> +            tcg_gen_andi_i64(tmp, tmp, 1ULL << out);
> +            tcg_gen_or_i64(rt, rt, tmp);
> +        }
> +        in += 64;
> +    }

This is going to produce up to 3*64 operations (n=2).

You can produce more than one output pairing per shift,
and produce the same result in 3*lg2(64) operations.

I've given an example like this on the list before, recently.
I think it was in the context of some riscv bit manipulation.

> N = 2
>
> AxBxCxDxExFxGxHxIxJxKxLxMxNxOxPxQxRxSxTxUxVxWxXxYxZx0x1x2x3x4x5x
>   & rep(0b10)
> A.B.C.D.E.F.G.H.I.J.K.L.M.N.O.P.Q.R.S.T.U.V.W.X.Y.Z.0.1.2.3.4.5.
>   << 1
> .B.C.D.E.F.G.H.I.J.K.L.M.N.O.P.Q.R.S.T.U.V.W.X.Y.Z.0.1.2.3.4.5..
>   |
> ABBCCDDEEFFGGHHIIJJKKLLMMNNOOPPQQRRSSTTUUVVWWXXYYZZ001122334455.
>   & rep(0b1100)
> AB..CD..EF..GH..IJ..KL..MN..OP..QR..ST..UV..WX..YZ..01..23..45..
>   << 2
> ..CD..EF..GH..IJ..KL..MN..OP..QR..ST..UV..WX..YZ..01..23..45....
>   |
> ABCDCDEFEFGHGHIJIJKLKLMNMNOPOPWQQRSTSTUVUVWXWXYZYZ010123234545..
>   & rep(0xf0)
> ABCD....EFGH....IJKL....MNOP....QRST....UVWX....YZ01....2345....
>   << 4
> ....EFGH....IJKL....MNOP....QRST....UVWX....YZ01....2345........
>   |
> ABCDEFGHEFGHIJKLIJKLMNOPMNOPQRSTQRSTUVWXUVWXYZ01YZ0123452345....
>   & rep(0xff00)
> ABCDEFGH........IJKLMNOP........QRSTUVWX........YZ012345........
>   << 8
> ........IJKLMNOP........QRSTUVWX........YZ012345................
>   |
> ABCDEFGHIJKLMNOPIJKLMNOPQRSTUVWXQRSTUVWXYZ012345YZ012345........
>   & rep(0xffff0000)
> ABCDEFGHIJKLMNOP................QRSTUVWXYZ012345................
>   deposit(t, 32, 16)
> ABCDEFGHIJKLMNOPQRSTUVWXYZ012346................................

and similarly for larger N.  For N >= 4, I believe that half of the masking may be elided, 
because there are already zeros in which to place bits.

> N = 5
>
> AxxxxBxxxxCxxxxDxxxxExxxxFxxxxGxxxxHxxxxIxxxxJxxxxKxxxxLxxxxMxxx
>   & rep(0b10000)
> A....B....C....D....E....F....G....H....I....J....K....L....M...
>   << (5 - 1)
> .B....C....D....E....F....G....H....I....J....K....L....M.......
>   |
> AB...BC...CD...DE...EF...FG...GH...HI...IJ...JK...KL...LM...M...
>   << (10 - 2)
> ..CD...DE...EF...FG...GH...HI...IJ...JK...KL...LM...M...
>   |
> ABCD.BCDE.CDEF.DEFG.EFGH.FGHI.GHIJ.HIJK.IJKL.JKLM.KLM..LM...M...
>   & rep(0xf0000)
> ABCD................EFGH................IJKL................M...
>   << (20 - 4)
> ....EFGH................IJKL................M...................
>   |
> ABCDEFGH............EFGHIJKL............IJKLM...............M...
>   << (40 - 8)
> ........IJKLM...............M...................................
>   |
> ABCDEFGHIJKLM.......EFGHIJKLM...........IJKLM...............M...
>   & 0xfff8_0000_0000_0000
> ABCDEFGHIJKLM...................................................

It's probably worth working through the various N to make sure you know which masking is 
required.


r~