[PATCH] D157417: [RISCV][SelectionDAG] Lower shuffles as bitrotates with vror.vi when possible

[Craig Topper via Phabricator via llvm-commits]

craig.topper added a comment.

In D157417#4595013 <https://reviews.llvm.org/D157417#4595013>, @luke wrote:

> In D157417#4593411 <https://reviews.llvm.org/D157417#4593411>, @reames wrote:
>
>> Broader point is that maybe we should be doing this even without zbb.
>
> For reference, this is the new sequence we would be generating without zvbb:
>
>    define <8 x i16> @shuffle_v8i16_as_i32(<8 x i16> %v) {
>    ; CHECK-LABEL: shuffle_v8i16_as_i32:
>    ; CHECK:       # %bb.0:
>   -; CHECK-NEXT:    lui a0, %hi(.LCPI18_0)
>   -; CHECK-NEXT:    addi a0, a0, %lo(.LCPI18_0)
>   -; CHECK-NEXT:    vsetivli zero, 8, e16, m1, ta, ma
>   -; CHECK-NEXT:    vle16.v v10, (a0)
>   -; CHECK-NEXT:    vrgather.vv v9, v8, v10
>   -; CHECK-NEXT:    vmv.v.v v8, v9
>   +; CHECK-NEXT:    vsetivli zero, 4, e16, mf2, ta, ma
>   +; CHECK-NEXT:    vmv.v.i v9, 0
>   +; CHECK-NEXT:    li a0, 16
>   +; CHECK-NEXT:    vwsubu.vx v10, v9, a0
>   +; CHECK-NEXT:    li a1, 31
>   +; CHECK-NEXT:    vsetvli zero, zero, e32, m1, ta, ma
>   +; CHECK-NEXT:    vand.vx v9, v10, a1
>   +; CHECK-NEXT:    vsrl.vv v9, v8, v9
>   +; CHECK-NEXT:    vmv.v.x v10, a0
>   +; CHECK-NEXT:    vand.vx v10, v10, a1
>   +; CHECK-NEXT:    vsll.vv v8, v8, v10
>   +; CHECK-NEXT:    vor.vv v8, v8, v9
>    ; CHECK-NEXT:    ret
>    ;
>    ; ZVBB_V-LABEL: shuffle_v8i16_as_i32:
>    ; ZVBB_V:       # %bb.0:
>    ; ZVBB_V-NEXT:    vsetivli zero, 4, e32, m1, ta, ma
>    ; ZVBB_V-NEXT:    vror.vi v8, v8, 16
>    ; ZVBB_V-NEXT:    ret
>    ;
>    ; ZVBB_ZVE32X-LABEL: shuffle_v8i16_as_i32:
>    ; ZVBB_ZVE32X:       # %bb.0:
>    ; ZVBB_ZVE32X-NEXT:    vsetivli zero, 4, e32, m4, ta, ma
>    ; ZVBB_ZVE32X-NEXT:    vror.vi v8, v8, 16
>    ; ZVBB_ZVE32X-NEXT:    ret
>      %shuffle = shufflevector <8 x i16> %v, <8 x i16> poison, <8 x i32> <i32 1, i32 0, i32 3, i32 2, i32 5, i32 4, i32 7, i32 6>
>      ret <8 x i16> %shuffle
>    }

Why isn't this constant folded

  +; CHECK-NEXT:    vmv.v.i v9, 0
  +; CHECK-NEXT:    li a0, 16
  +; CHECK-NEXT:    vwsubu.vx v10, v9, a0

This and feels unnecessary. This shift only uses the lower 5 bits

  +; CHECK-NEXT:    li a1, 31
  +; CHECK-NEXT:    vsetvli zero, zero, e32, m1, ta, ma
  +; CHECK-NEXT:    vand.vx v9, v10, a1
  +; CHECK-NEXT:    vsrl.vv v9, v8, v9


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D157417/new/

https://reviews.llvm.org/D157417