[PATCH][PR21138] Lower <u, u, u, u, idx1, idx2, idx3, idx4> using 256bit shuffle instead of splitting

[Quentin Colombet]

Hi Chandler,

Attached is a tentative patch to fix PR21138 http://llvm.org/bugs/show_bug.cgi?id=21138

I have not tested it against the test suite, I wanted to know if the direction matches what you expected first.
In particular, the check that guards the transformation is simple than I would have expected at first, but I did not find cases where the transformation is actually not beneficial (though, I haven’t tried hard).

Anyhow, let me know what you think on the patch before I make the measurements.

Thanks,
-Quentin