The previous instruction selection was too loose, it only required
registers for the inputs. The codegen also used Unpcklps(dst, mask), and
failed to use src at all. The test case was accidentally passing
because dst == src (xmm0) by chance.

We fix this bug requiring that for AVX, any register is fine, but for
SSE, require dst == src. Also redefine Unpcklps to check dst == src in
the no AVX case.

Bug: v8:11265
Change-Id: I1988b2d2da8263512bf6e675e6297c50f55663f7
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2668918Reviewed-by: Deepti Gandluri <gdeepti@chromium.org>
Commit-Queue: Zhi An Ng <zhin@chromium.org>
Cr-Commit-Position: refs/heads/master@{#72536}