split-init interaction with warning about local-from-distributed

[mppf]

How should split init interact with the warning about initializing a non-distributed domain from a distributed one?
Previously this warning relied on the difference between initialization and assignment. The warning was added in PR #10791.

For example:

// copy to a default-rectangular array, for writeln()
proc copyToDF(A:[]) {
  var D: domain(A.rank, A.domain.idxType, A.domain.stridable);
  D = A.domain; // warning here if this is now initialization
  var Res: [D] A.eltType = A;
  return Res;
}

warning: initializing a non-distributed domain from a distributed domain. If you didn't mean to do that, add a dmapped clause to the type expression or remove the type expression altogether

When this is resolved, update these tests, as PR #14691 adds a workaround to them:

• test/distributions/dm/s7.chpl
• test/distributions/dm/s8.chpl
• test/distributions/dm/s9.chpl
• test/distributions/ferguson/init-local-from-distributed-warning.chpl
• test/studies/hpcc/HPL/bradc/hpl-blc-noreindex.chpl
• distributions/robust/arithmetic/basics/test_domain_align.chpl

[mppf]

Brainstorming some ideas to resolve this:

1. make the error only apply to non-split init by adjusting the compiler to avoid calling chpl_checkCopyInit for split init
2. accept the warning for the pattern demonstrated above
   a. tests use the workaround of mentioning the variable before initializing it
   b. tests include the warning in .good files
   c. tests are updated to call a new library routine to copy to default-rectangular
   d. turn it off with fine-grained-warning control generalized warning control #7355
3. extend the warning to apply also to assignment

[bradcray]

I'd probably do 2b myself. Could you save me some time to reproduce myself by pointing to where the warning would come from in test/studies/hpcc/HPL/bradc/hpl-blc-noreindex.chpl and distributions/robust/arithmetic/basics/test_domain_align.chpl (the two files I'm most familiar with in your list).

[Adding more rationale for this choice:

• doesn't require much effort
• unlikely to happen much in practice, in my opinion
• treats initialization consistently
• could always change later if it ended up being a real hassle for someone, by:
  • going to case 1
  • providing a means of disabling a warning for a specific chunk of code

My gut reaction is not to extend the warning to assignment (because it seems like it would make many domain-to-domain assignments annoying), but maybe that's just because I understand the language well enough that I'm not a huge fan of the warning to begin with and think people shouldn't be declaring their domain variable types so much? Though I understand that many new users have gotten pretty flummoxed by this behavior which is why we have it...]

[mppf]

Sure.

chapel/test/studies/hpcc/HPL/bradc/hpl-blc-noreindex.chpl

Lines 186 to 189 in 921f541

	var replAD: domain(2, indexType),
	replBD: domain(2, indexType);
	replAD = AD;
	replBD = BD;

chapel/test/distributions/robust/arithmetic/basics/test_domain_align.chpl

Lines 42 to 44 in 921f541

	D = rangeTuple(D.rank, 1..10);
	var R : domain(D.rank, D.idxType, D.stridable);
	R = D;

[bradcray]

Thanks for pointing to the code segments. I don't think either of those change my answer (and I provided a bit more rationale for my answer in an edit just above, capturing some of our offline discussion).

[mppf]

I'll plan to do 2b for the moment.

[mppf]

Addressed (for now at least) by PR #14802 which does 2b.