PixWin enables to use ProcessVertices which outputs the vs shader outputs.

I looked at the output which would give the IN[4].w

If found that for llvmpipe (used for ProcessVertices) the vs output is between
0 and 1.
However I tried with radeonsi, and... it displays a NaN for that output...

I guess one of the input is NaN and llvmpipe clamps it (the vs output is the
subtraction between two vs inputs).

If that output is really NaN, I guess it can explain the regression. A small
change (x < 0 vs 0 > x) will give different result with NaN.

I tried to change the order for the test in radeonsi's kil_emit, but the
generated isa doesn't change the test order between 0 and x.