[PATCH] [X86][SSE] Vectorized i8 and i16 shift operators

[Fiona Glaser]

Is it possible to do an i8 variable shift as follows? pseudocode (in0 = value being shifted, in1 = shift amounts):

vec8_multiplies: [1, 2, 4, 8, 16, 32, 64, 128]

pshufb tmp1, [vec8_multiplies], in1
pmovzxbw tmp2, in0
pmovzxbw tmp3, tmp1
pmullw tmp2, tmp3

or something like that? would it be faster, and could something of the sort maybe be doable for right shifts (multiply, then shift back down?) i16 might also be doable with appropriate hacks...


REPOSITORY
  rL LLVM

http://reviews.llvm.org/D9474

EMAIL PREFERENCES
  http://reviews.llvm.org/settings/panel/emailpreferences/