dgemv_kernel_4x4(Haswell): add missing clobbers for xmm0,xmm1,xmm2,xmm3

[bartoldeman]

This fixes a crash in dblat2 when OpenBLAS is compiled using
-march=znver1 -ftree-vectorize -O2

See also:
easybuilders/easybuild-easyconfigs#7180

[martin-frbg]

Thanks. Yet another error pattern to look for ... but at first glance it appears to be a unique oversight in #561. gcc >=8 is really bringing out the worst traits of the code it seems...

[martin-frbg]

FYI compiling with -ftree-vectorize -O2 (on Kaby Lake, i.e. Haswell kernels) leads to a "less than half accurate" warning for SGEMV so there appears to be more such niceness. (Possibly related to my misgivings about input arguments beyond %0,%1 getting modified #2009)

[martin-frbg]

The SGEMV mishandling does indeed appear to come from abusing one of the input parameters (%8 a.k.a. "lda4") of sgemv_kernel_4x8 in sgemv_n_microk_haswell-4.c . I did not manage to come up with a proper constraint, so my current workaround is to "movq" the value from %8 into the otherwise unused %xmm10 at the start and restore it from there - not sure if that is a sane solution...