test/ken/convert.go fails for sparc64 SSA backend

[binarycrusader]

# go run run.go -- ken/convert.go
incorrect output
78 +1.000000e+001 +5.500000e+000 8 9
139 +1.750000e+002 +1.740000e+002 2 8
148 +1.750000e+002 +1.740000e+002 3 8
157 +1.750000e+002 +1.740000e+002 4 8
166 +1.750000e+002 +1.740000e+002 5 8
175 +1.750000e+002 +1.740000e+002 6 8
184 +1.750000e+002 +1.740000e+002 7 8
193 +1.750000e+002 +1.740000e+002 8 8
194 +1.750000e+002 +9.500000e+001 8 9
195 +1.750000e+002 +1.740000e+002 8 10
202 +1.750000e+002 +1.740000e+002 9 8
211 +1.750000e+002 +1.740000e+002 10 8
217 +4.125900e+004 +4.125800e+004 4 8
224 +4.125900e+004 +4.125800e+004 5 8
231 +4.125900e+004 +4.125800e+004 6 8
238 +4.125900e+004 +4.125800e+004 7 8
245 +4.125900e+004 +4.125800e+004 8 8
246 +4.125900e+004 +2.267700e+004 8 9
247 +4.125900e+004 +4.125800e+004 8 10
252 +4.125900e+004 +4.125800e+004 9 8
259 +4.125900e+004 +4.125800e+004 10 8
275 +3.758096e+009 +2.013266e+009 8 9
287 +1.671736e+019 +1.729382e+018 8 8
288 +1.671736e+019 +9.367487e+017 8 9
289 +1.671736e+019 +1.729382e+018 8 10
290 +1.671736e+019 +9.223372e+018 9 8
293 +1.671736e+019 +1.729382e+018 10 8

FAIL    ken/convert.go  0.496s

[4ad]

Interesting.

[binarycrusader]

FSTOX needs the same aglorithm applied as FDTOX (use native conversion if value less than 2^63; otherwise subtract 2^63, convert, and add back).

FXTOS/FXTOD were not handling the unsigned case correctly, not only did they need to copy the input value to a temporary register for adjustments, but they also had the 2^63 comparison reversed.

FXTOS was wrongly using AFMULD (instead of AFMULS) to double the value after conversion for the unsigned > 2^63 case.

test/ken/convert.go now passes; as of this moment, all tests in test that can be cross-compiled and executed work just as well as they did with the old compiler backend. There's only one known failure at the moment, as described in #33.