Bug 94972

Using vector selects (cdf7c6b83dad7eb6a7600af61403315b02dcf13f) caused some regressions when using llvm 3.7 (large number of blend tests in deqp, also mentioned in bug 94957, piglit gl-1.0-blend-func, and dolphin was mentioned on IRC).

The problem is actually entirely llvm's fault after some digging (initially wasn't sure if we weren't relying on some undefined behavior, hence tracked this down), which very obviously miscompiles rather simple vector selects (I suspect this only affects per-byte selects, possibly only when constant values are involved).
It only seems to affect 3.7, I've tried 3.3 through 3.8 and everything else worked.

Mostly filing a bug so I've got something to refer to when working around it in mesa.

define <16 x i8> @novsel(<16 x i8> %val1, <16 x i8> %val2) {
entry:
  %val1a = and <16 x i8> %val1, <i8 0, i8 0, i8 0, i8 -1, i8 0, i8 0, i8 0, i8 -1, i8 0, i8 0, i8 0, i8 -1, i8 0, i8 0, i8 0, i8 -1>
  %val2rgb = and <16 x i8> %val2, <i8 -1, i8 -1, i8 -1, i8 0, i8 -1, i8 -1, i8 -1, i8 0, i8 -1, i8 -1, i8 -1, i8 0, i8 -1, i8 -1, i8 -1, i8 0>
  %res = or <16 x i8> %val1a, %val2rgb
  ret <16 x i8> %res
}

define <16 x i8> @vsel(<16 x i8> %val1, <16 x i8> %val2) {
entry:
  %res = select <16 x i1> <i1 false, i1 false, i1 false, i1 true, i1 false, i1 false, i1 false, i1 true, i1 false, i1 false, i1 false, i1 true, i1 false, i1 false, i1 false, i1 true>, <16 x i8> %val1, <16 x i8> %val2
   ret <16 x i8> %res
}

The former gets compiled to:
        vmovdqa .LCPI0_0(%rip), %xmm2   # xmm2 = [255,255,255,0,255,255,255,0,255,255,255,0,255,255,255,0]
        vpblendvb       %xmm2, %xmm1, %xmm0, %xmm0
        retq

But the latter to:
        vmovdqa .LCPI1_0(%rip), %xmm2   # xmm2 = [0,0,0,255,0,0,0,255,255,255,255,255,255,255,255,255]
        vpblendvb       %xmm2, %xmm0, %xmm1, %xmm0
        retq

So only the first 8 of the 16 values are correct (and yes the .LCPI1_0 byte values look the same as indicated in the comment).