[PATCH] ARM NEON: Handle v16i8 and v8i16 reverse shuffles

[Nadav Rotem]

+/// If the mask is a reverse operation on an v16i8/v8i16 operation we can
+/// implement it in terms of a vrev64 and vext operation.
+/// \return true if this is a reverse operation on an v16i8/v8i16 vector.
+static bool isReverse_v16i8_or_v8i16_Mask(ArrayRef<int> M, EVT VT) {
+  if (VT != MVT::v16i8 && VT != MVT::v8i16)
+      return false;
+
+  unsigned NumElts = VT.getVectorNumElements();
+  // Make sure the mask has the right size.
+  if (NumElts != M.size())
+      return false;
+
+  // Look for <15, ..., 3, -1, 1, 0>.
+  for (unsigned i = 0; i != NumElts; ++i)
+    if (M[i] >= 0 && M[i] != (int) (NumElts - 1 - i))
+      return false;
+
+  return true;
+}

In the future may want to add support for additional types (such as floats). Can we make this function generic and check for the types outside the function ? 

+  if (isReverse_v16i8_or_v8i16_Mask(ShuffleMask, VT))
+    return LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(Op, DAG);
+

Can become:

+  if (isReverseMask(ShuffleMask, VT)  && (VT == MVT::v16i8 || VT == MVT::v8i16 ))
+    return LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(Op, DAG);
+


Thanks,
Nadav

On Feb 9, 2013, at 11:02 AM, Arnold Schwaighofer <aschwaighofer at apple.com> wrote:

> Lower reverse shuffles to a vrev64 and a vext instruction instead of the default
> legalization of storing and loading to the stack. This is important because we
> generate reverse shuffles in the loop vectorizer when we reverse store to an
> array.
> 
>  uint8_t Arr[N];
>  for (i = 0; i < N; ++i)
>    Arr[N - i - 1] = …
> 
> For v8i16 we now generate something like:
> 
>  vrev64.16       q9, q9
>  vext.16 q9, q9, q9, #4
> 
> instead of:
> 
>  orr r1, r0, #14
>  vst1.16 {d16[0]}, [r1, :16]
>  orr r1, r0, #12
>  vst1.16 {d16[1]}, [r1, :16]
>  orr r1, r0, #10
>  vst1.16 {d16[2]}, [r1, :16]
>  orr r1, r0, #8
>  vst1.16 {d16[3]}, [r1, :16]
>  orr r1, r0, #6
>  vst1.16 {d17[0]}, [r1, :16]
>  orr r1, r0, #4
>  vst1.16 {d17[1]}, [r1, :16]
>  orr r1, r0, #2
>  vst1.16 {d17[2]}, [r1, :16]
>  vst1.16 {d17[3]}, [r0, :16]
>  vld1.64 {d16, d17}, [r0, :128]
> 
> 
> For v16i8 we now generate something like:
> 
>  vrev64.8        q8, q8
>  vext.8  q8, q8, q8, #8
> 
> <0001-ARM-NEON-Handle-v16i8-and-v8i16-reverse-shuffles.patch>
>