Re: [RFC PATCH v1 03/43] accel/tcg: Add gvec size changing operations

[Richard Henderson]

On 12/3/24 14:15, Anton Johansson wrote:
> The point is that we have a lot of Hexagon instructions where size
> changes are probably unavoidable, another example is V6_vshuffh which
> takes in a <16 x i16> vector and shuffles the upper <8xi16> into the upper
> 16-bits of a <8 x i32> vector
>
>      void emit_V6_vshuffh(intptr_t vec3, intptr_t vec7) {
>          VectorMem mem = {0};
>          intptr_t vec2 = temp_new_gvec(&mem, 128);
>          tcg_gen_gvec_zext(MO_32, MO_16, vec2, vec7, 128, 64, 128);
>
>          intptr_t vec0 = temp_new_gvec(&mem, 128);
>          tcg_gen_gvec_zext(MO_32, MO_16, vec0, (vec7 + 64ull), 128, 64, 128);
>
>          intptr_t vec1 = temp_new_gvec(&mem, 128);
>          tcg_gen_gvec_shli(MO_32, vec1, vec0, 16, 128, 128);
>          tcg_gen_gvec_or(MO_32, vec3, vec1, vec2, 128, 128);
>      }
>
> Not to bloat the email too much with examples, you can see 3 more here
>
>    https://pad.rev.ng/11IvAKhiRy2cPwC7MX9nXA
>
> Maybe we rely on the target defining size-changing operations if they
> are needed?

Perhaps.

I'll note that emit_V6_vpackwh_sat in particular should probably not use vectors at all. 
I'm sure it would be shorter to simply expand directly to integer code.

I'll also note that tcg's vector support isn't really designed for the way you're using 
it.  It leads to the creation of many on-stack temporaries that would not otherwise be 
required.

When targets are emitting their own complex patterns, the expected method is to use the 
GVecGen* structures and the callbacks therein.  This allows the JIT to select different 
expansions depending on the host cpu vector support (or lack thereof).

For a simple example, see gen_gvec_xar() in target/arm/tcg/gengvec64.c, which simply 
combines a rotate and an xor.  For a more complex example, see gen_gvec_usqadd_qc() later 
in that same file, where in the worst case we call an out-of-line helper.


r~