On 7/15/20 9:13 AM, Roger Sayle wrote:
>
> This patch provides standard vec_extract and vec_set patterns to the
> nvptx backend, to extract an element from a PTX vector and set an
> element of a PTX vector respectively. PTX vectors (I hesitate to
> call them SIMD vectors) may contain up to four elements, so vector
> modes up to size four are supported by this patch even though the
> nvptx backend currently only allows V2SI and V2DI, i.e. two out
> of the ten possible vector modes.
>
> As an example of the improvement, the following C function:
>
> typedef int __v2si __attribute__((__vector_size__(8)));
> int foo (__v2si arg) { return arg[0]+arg[1]; }
>
> previously generated this code using a shift:
>
> mov.u64 %r25, %ar0;
> ld.v2.u32 %r26, [%r25];
> mov.b64 %r28, %r26;
> shr.s64 %r30, %r28, 32;
> cvt.u32.u32 %r31, %r26.x;
> cvt.u32.u64 %r32, %r30;
> add.u32 %value, %r31, %r32;
>
> but with this patch now generates:
>
> mov.u64 %r25, %ar0;
> ld.v2.u32 %r26, [%r25];
> mov.u32 %r28, %r26.x;
> mov.u32 %r29, %r26.y;
> add.u32 %value, %r28, %r29;
>
> I've implemented these getters and setters as their own instructions
> instead of attempting the much more intrusive patch of changing the
> backend's definition of register_operand. Given the limited utility
> of PTX vectors, I'm not convinced that attempting to support them as
> operands in every instruction would be worth the effort involved.
>
> This patch has been tested on nvptx-none hosted on x86_64-pc-linux-gnu
> with "make" and "make check" with no new regressions.
> Ok for mainline?
>
>
> 2020-07-15 Roger Sayle <ro...@nextmovesoftware.com>
>
> gcc/ChangeLog:
> * config/nvptx/nvptx.md (nvptx_vector_index_operand): New predicate.
> (VECELEM): New mode attribute for a vector's uppercase element mode.
> (Vecelem): New mode attribute for a vector's lowercase element mode.
> (*vec_set<mode>_0, *vec_set<mode>_1, *vec_set<mode>_2,
> *vec_set<mode>_3): New instructions.
> (vec_set<mode>): New expander to generate one of the above insns.
> (vec_extract<mode><Vecelem>): New instruction.
Added test-case, fixed some nits, pushed (not reposting).
Thanks,
- Tom
