nir: Optimize fpow with small constant exponents
They would be turned into exp(log(a)*b) instead, which is slow.
Totals from 2146 (2.52% of 85071) affected shaders:
MaxWaves: 35769 -> 35779 (+0.03%); split: +0.03%, -0.01%
Instrs: 6476835 -> 6465494 (-0.18%); split: -0.18%, +0.00%
CodeSize: 35382288 -> 35347092 (-0.10%); split: -0.10%, +0.00%
SpillSGPRs: 1055 -> 1017 (-3.60%)
Latency: 75211743 -> 75063623 (-0.20%); split: -0.20%, +0.00%
InvThroughput: 17525115 -> 17501745 (-0.13%); split: -0.14%, +0.00%
VClause: 200089 -> 200077 (-0.01%); split: -0.01%, +0.01%
SClause: 293566 -> 293480 (-0.03%); split: -0.03%, +0.00%
Copies: 649631 -> 640516 (-1.40%); split: -1.44%, +0.03%
Branches: 268441 -> 268325 (-0.04%)
PreSGPRs: 146868 -> 146045 (-0.56%)
PreVGPRs: 134125 -> 134128 (+0.00%); split: -0.00%, +0.01%
Reviewed-by: Georg Lehmann <dadschoorse@gmail.com>
Reviewed-by: Faith Ekstrand <faith.ekstrand@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26727>
b88ac6b381