Sub-integer compares under select nodes give wrong result on x86-64

[aviansie-ben]

As part of refactoring the *select evaluators on Power in #5497, new Tril tests were added to exercise previously untested codegen optimizations for when the first child is a compare. When running these tests on x86, numerous incorrect results were observed for select nodes with a sub-integer compare as their first child, such as:

[2020-08-27T01:03:20.027Z] 47: /home/jenkins/workspace/Build/fvtest/compilertriltest/SelectTest.cpp:1382: Failure
[2020-08-27T01:03:20.027Z] 47:       Expected: xselectOracle(xcmpOracle(param.cmp, param.c1, param.c2), param.v1, param.v2)
[2020-08-27T01:03:20.027Z] 47:       Which is: 1
[2020-08-27T01:03:20.027Z] 47: To be equal to: entry_point(param.c1, param.c2, param.v1, param.v2)
[2020-08-27T01:03:20.027Z] 47:       Which is: 0
[2020-08-27T01:03:20.027Z] 47: [  FAILED  ] SelectCompareTest/Int32SelectInt8CompareTest.UsingLoadParam/50, where GetParam() = (Comparison::lt, ('\x80' (-128), '\x7F' (127)), (1, 0)) (2 ms)