Fix unannounced modification of input operand 8 (lda4) in Haswell GEMVN microkernel

[martin-frbg]

Fixes miscompilation with gcc9 -ftree-vectorize (related to issue #2009 and PR #2018)

[bartoldeman]

Another way to do this is to rename. Swapping %2 (x) with %8 and then adding the + for 2 gives a minimal diff, but you could also rename 2 to 3, 3 to 4, .. 7 to 8 and then 8 to 2.

[bartoldeman]

That looks good to me. An alternative that avoids this awkward reordering is to use "asmSymbolicName"s, as documented here: https://gcc.gnu.org/onlinedocs/gcc/Extended-Asm.html
Very old GCCs did not have those: I was digging a bit and found those were added in GCC 3.1 (from 2002).

[bartoldeman]

In any case this becomes a moot point once it's rewritten to use intrinsics like the skylake-x code, so for now I understand it's just about fixing bugs.

[martin-frbg]

Thanks again for the review and additional pointers. I'll probably go with the reordering for now, although it is a bit uglier with the next file (the dtrsm_kernel_RN_haswell mentioned in #2009).
Similar bugs exist in SGEMVN for Nehalem and Sandybridge, and in some GEMVN and TRSM
kernels for Bulldozer and Piledriver, I'll see if I can get them fixed over the weekend.