eliminate redundant inline scanning

[WalterBright]

While working on a fix to do better delegate inlining, I discovered that the inliner is constantly rescanning functions for inline possibilities. Added a flag FUNCFLAGinlineScanned to put a stop to that, and a flag again to enable rescanning in case of delegate parameters.

The latter isn't complete yet, but this PR won't hurt.

[yebblies]

Indentation

[andralex]

s/inline possibilities/inlining opportunities/ I guess but no matter

[9rnsr]

Looks good to me.

[andralex]

Auto-merge toggled on

[9rnsr]

This PR introduced a regression:
https://issues.dlang.org/show_bug.cgi?id=15253

[9rnsr]

I have a question about this change.

Q1. To see the inline possibility of a function, InlineScanVisitor.visit(FuncDeclaration) will be called:

1. from inlineScanModule - it's invoked in the global "inlining" stage. In this case, InlineScanVisitor.visit(FuncDeclaration) will be called only once for each functions.

2. from canInline - it's called from each expression/statement inlining, so it could be called during expandInline process. However in it, fd.inlineNest is already tested and rejecting recursive scanning, so InlineScanVisitor.visit(FuncDeclaration) won't be called recursively.
   And also, the scanning result is cached into fd.inlineStatus(Stmt|Exp) field, so repeated scanning is not also there.

   Therefore, I cannot see why we need to have a new flag FUNCFLAGinlineScanned. To me it's just a duplication of fd.inlineNest.

Q2. As I already specified in Q1, recursive expansion is tested and prevented by canInline. If we do repeated scanning for more inline possibility, the again flag seems to be set in canInline function, IMO.

[9rnsr]

Ah, if InlineScanVisitor.visit(FuncDeclaration) is called from inlineScanModule, it would ignore the cached result fd.inlineStatusStmt or fd.inlineStatusExp.
And if InlineScanVisitor.visit(FuncDeclaration) is called from canInline, it would ignore the previous scan result that is calculated via inlineScanModule.

In other words, a function will be scanned at most three times - 1. as function declaration, 2. as a statement, and 3. as an expression. Before this change, the 2. result is cached to fd.inlineStatusStmt, and 3. is cached to fd.inlineStatusExp. After this change the 1. result is cached to fd.flags & FUNCFLAGinlineScanned.

But, as I already said in Q1, the No.1 case == InlineScanVisitor.visit(FuncDeclaration) call from inlineScanModule would happen at most once. I still think caching its result would not be necessary.