no codegen for parts of failed template instantiations

[rainers]

do not write failed template instantiations or parts of these to the object file.

This solves the duplicate COMDAT error in dlang/phobos#1411

[yebblies]

Usual style in dmd is if (

[rainers]

Fixed spacing.

[rainers]

It seems b69eb50 fixed the error propagation, so I have reduced it to a simple function in Dsymbol. The function name isInFailedInstantiation now is too specific, but hasErrors sounds too generic. The patch #2480 uses isInErrorTree but it's ugly. Any better suggestion?

[yebblies]

Just use a loop in TemplateInstance::toObjFile.

[rainers]

Just use a loop in TemplateInstance::toObjFile

To avoid finding a good name? ;-)

The PR #2480 calls it from other places aswell (e.g. when TypeInfo for erronuous types has been generated), so a function would be nice.

[yebblies]

The problem is that we shouldn't be getting here in the first place - at this point all instances with errors are speculative and shouldn't be outputted anyway. Maybe we should keep a list of all child template instances and recursively mark them as non-speculative?

https://d.puremagic.com/issues/show_bug.cgi?id=11724

[rainers]

Following Kenjis' suggestion, I moved the parent check to isError.

[9rnsr]

s is already != NULL, so s && is redundant.

[rainers]

Hmmm, I'd rather say the test is missing in the check for s->errors. If we can assume isErrors is always a Dsymbol, we should pass it as that.

[rainers]

BTW: is there any reason why code isn't written as "if (Dsymbol* s = isDsymbol(o)) {}"? I'm very much tempted to change it to this style if I stumble over it.

[9rnsr]

o is neither Type, Expression, nor Tuple. so it is always Dsymbol.

[9rnsr]

Additionally, s->errors at line 105 already assumes that s != NULL. Therefore, testing it here looks weird to me.

[9rnsr]

Anyway, this change does not have serious problem, and I also confirmed that the it would fix the issue in #1411 in my local.
LGTM.

[rainers]

Thanks. The win64 tests now pass.
I've created a small followup pull request #3079 for the suggested cleanup.