Fix codegen for comparisons to be able to handle refs

[e-kayrakli]

Without this patch codegen seems to omit dereferences for a in the following snippet:

proc foo() {
  var a = 10;
  on Locales[0] {
    writeln(a>10);
    writeln(a>=10);
    writeln(a<10);
    writeln(a<=10);
  }
}

foo();

Such code compiled fine with --no-denormalize, since the dereferencing was handled by PRIM_MOVE. However, since denormalize removes that PRIM_MOVE this code was compiled without derefrenceing causing pointer vs integer comparison.

It seems that ref intent for a is inserted post denormalize, so there doesn't seem to be a direct way of handling this at denormalize time. Moreover, such issues are handled at codegen time for arithmetic operations. I followed the same strategy for comparisons.

This PR also adds a test that generates faulty code on master.

(I can't easily reproduce this issue without on statements. I tried to do it with ref variables but it seems to work fine for those)

Tests:

• release/examples/hello*chpl passed with GASNet
• users/engin/refComparison.chpl passed with GASNet
• Full GASNet
• Full LLVM (?)

[e-kayrakli]

This was encountered by @LouisJenkinsCS.

Also, I can't run paratest myself.

[benharsh]

I'll start some paratesting:

• full local
• full gasnet
• gasnet llvm examples/

[benharsh]

Does this test really need a skipif?

[e-kayrakli]

Hmm.. not that it fails with CHPL_COMM!=none. Just that the on statement is optimized away (I think) with that environment, thus it doesn't test anything new. In other words, with CHPL_COMM=none this test passes without this PR, too.

I can remove it if you want.

[benharsh]

Should it really be "CHPL_COMM==none" , which would only run the test in multilocale configurations?

[e-kayrakli]

Oh boy :( You are right, sorry about the confusion. Do you want me to fix it or remove it altogether?

[benharsh]

Let's just remove it

[benharsh]

Other than the skipif, these changes look good to me.

Testing also looks good.