Duplicate symbol emitted with SoA and fusion optimizations

[asujeeth]

In OptiML logistic regression, we are getting a duplicated symbol emitted that appears to be a scheduling bug related to some interaction between Multiloop SoA and loop fusion (it does not occur if either one is turned off).

The exact behavior we are seeing is a while loop kernel generated like the following:

def apply(resourceInfo:generated.scala.ResourceInfo,x1231:Int,x1135:Int,x1296:Array[Double],x1229:Array[Double],x628:generated.scala.DenseVectorDouble,x1301:generated.scala.Ref[Double],x1356:generated.scala.Ref[Int],x1355:generated.scala.Ref[generated.scala.DenseVectorDouble]):
Unit = {
val x1880 = while ({
// ...
} {
// ...
val x1135 = x1134.length
// ...
})

Notice that x1135 is re-emitted, despite being an input to the kernel. This also occurs in normal kernel generation, but because x1135 is emitted as a kernel (a separate file), that file is simply overwritten, so the program still compiles and runs correctly.

To reproduce, use the following hashes:
Forge: stanford-ppl/Forge@3de62bb
Delite: c301618
LMS: TiarkRompf/virtualization-lms-core@81ee4b1

and run the program LogRegCompiler from OptiML. When running delite LogRegCompiler, you should see the following walk-time compilation error:

generatedCache/scala/src/kernels/x1880.scala:24: error: not found: value x1134
val x1135 = x1134.length
            ^
one error found

My hypothesis (though unproven) is that the problematic symbol may be being injected into the scope during fusion, which is somehow affected by the split SoA loops. It's possible this could be fixed by Vera's new implementation, or at least made easier to debug.

[asujeeth]

@kjbrown, @TiarkRompf, @vsalvis - let me know if you guys have any ideas. I'm punting on this one for now, I think.