[Bug]: internal error when assigning array slices with `--debug`

[jabraham17]

The following code results in an internal error when compiled with --debug

use BlockDist;
proc copy(x : [?d] ?t) {
  var retVal = blockDist.createArray(d, t);
  coforall loc in Locales do on loc {
    var xLD = x.localSubdomain() ;
    retVal[xLD] = x[xLD];
  }
  return retVal;
}

var arr = blockDist.createArray(1..100, int);

var arrCopy = copy(arr);
writeln(arrCopy);

chpl --debug bug.chpl

$CHPL_HOME/modules/standard/CTypes.chpl:99: internal error: no field 'eltType' in class 'c_ptr(_array(unmanaged BlockArr(1,int(64),one,int(64),unmanaged DefaultDist)))' in getField() [codegen/cg-type.cpp:777]

[jabraham17]

Actually, the reproducer is even simpler

var arr: [1..100] int;
var arrCopy: [1..100] int;
arrCopy[1..10] = arr[1..10];
writeln(arrCopy);

chpl --debug bar.chpl

$CHPL_HOME/modules/standard/CTypes.chpl:99: internal error: no field 'eltType' in class 'c_ptr(_array(unmanaged DefaultRectangularArr(1,int(64),one,int(64),int(64))))' in getField() [codegen/cg-type.cpp:777]

[jabraham17]

The array slices are a symptom of the biger problem, I can replicate the issue without them

use CTypes;
var arr: [1..100] int;
var ptr = c_addrOf(arr);
writeln(ptr.eltType:string);
// writeln(ptr.deref());

This is compiled with --no-dead-code-elimination and results in the same error. Any access of ptr.eltType results in the issue. So for example, ptr.deref() hits the issue because it has a check like the following

proc deref() {
  if this.eltType == void then compilerError(...);
  ...
}

During resolution, the compile time conditional is folded away. But before that happens, the compiler inserts code to build a runtime type record for eltType (since its an array type). So while the check is removed, the call to eltType does not get removed.

Similarly in the cast to string case. During resolution, ptr.eltType:string is compile time folded to the proper string value. But because of the runtime type, the call to eltType does no get properly removed. This is dead code, so with --dead-code-elimination this gets removed before codegen. But without that, the compiler tries to codegen a field access of eltType (which is a fake field) and blows up.

I think this is a problem that needs to be fixed before codegen. Ideally, during resolution when these compile-time type calls are removed the runtime type values also get removed