Lower saturating_cast in bounds inference

[rootjalex]

This should revert bounds inference on saturating_cast to the behavior of pre - #6900 .

@steven-johnson could you test to see if this fixes things?

The issues detected were not a direct result of #6900 - before that PR, there was still signed_integer_overflow inside of Bounds.cpp, but it was being hidden by complex Cast handling inside the bounds visitor. @abadams and I are discussing a fix to the underlying issue, but it is incredibly involved and requires possibly using infinite-precision integers in some parts of the compiler.

[steven-johnson]

Should this be applied in addition to #6961, or instead of?

[rootjalex]

Sorry for the lack of clarification - this is instead of #6961

[steven-johnson]

OK, looks like this one is passing -- Let's just insert that fix for OpenGLCompute too; I monkey-patched it for my testing

[steven-johnson]

LGTM pending the fix to OpenGLCompute as well

[abadams]

A proper fix seems intractable without getting some sort of repro for the harder failure cases. AJ's earlier attempt at a fix really should have worked.

[abadams]

Maybe we need a test-case reducer like bugpoint

[steven-johnson]

A proper fix seems intractable without getting some sort of repro for the harder failure cases

I can try to devote more time to the other repro case, but don't rely on it, it may be hard to disentangle

[rootjalex]

LGTM pending the fix to OpenGLCompute as well

Just added.

AJ's earlier attempt at a fix really should have worked.

To clarify for @steven-johnson , Andrew and I both think that 129b464 really should have worked, and cannot figure out what could have changed by #6900 that was not solved by that commit. It's very understandable if you can't disentangle that, but I think looking at that change would probably be easier than looking at the entire #6961 change.

[rootjalex]

The only failure on the buildbots appears unrelated, I think we should merge this in as a temporary fix for now. I've opened #6974 to track the fundamental issue.