i965/fs: Only unroll high-accuracy dFdy() from SIMD16 to SIMD8 on gen4 and IVB.

i965/fs: Only unroll high-accuracy dFdy() from SIMD16 to SIMD8 on gen4 and IVB.

In commit 800610f (i965/fs: Improve accuracy of dFdy() to match
dFdx()) I unrolled the high-accuracy dFdy() computation from a single
SIMD16 instruction to two SIMD8 instructions because of text I found
in the i965 (gen4) PRM saying that instruction compression could not
be used in align16 mode.  I couldn't find similar text in later
hardware docs, and I observed problems trying to use instruction
compression on align16 mode on Ivy Bridge, so I assumed that the
restriction still applied and the associated documentation had simply
been lost.

After consultation with the hardware engineers, it turns out this is
not the case.  In point of fact, the restriction was dropped in gen5,
re-introduced in Ivy Bridge, and dropped again in Haswell.  The reason
I didn't notice this is that in the Ivy Bridge documentation, the
restriction was in a different section, and described using different
language.

Now that we know that the restriction only applies to Gen4 and Ivy
Bridge, we can limit the unrolling to those platforms.

Tested on gen5, gen6, and gen7 (both Ivy Bridge and Haswell).

Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Eric Anholt <eric@anholt.net>