cmd/compile: use MOVBQZX for OpAMD64LoweredHasCPUFeature
In the commit message of CL 212360, I wrote:
> This new intrinsic ... generates MOVB+TESTB+NE.
> (It is possible that MOVBQZX+TESTQ+NE would be better.)
I should have tested. MOVBQZX+TESTQ+NE does in fact appear to be better.
For the benchmark in #36196, on my machine:
name old time/op new time/op delta
FMA-8 0.86ns ± 6% 0.70ns ± 5% -18.79% (p=0.000 n=98+97)
NonFMA-8 0.61ns ± 5% 0.60ns ± 4% -0.74% (p=0.001 n=100+97)
Interestingly, these are both considerably faster than
the measurements I took a couple of months ago (1.4ns/2ns).
It appears that CL 219131 (clearing VZEROUPPER in asyncPreempt) helped a lot.
And FMA is now once again slower than NonFMA, although this change
helps it regain some ground.