On 4/1/2015 10:00 AM, "Marc =?UTF-8?B?U2Now7x0eiI=?= <schue...@gmx.net>" wrote:
On Tuesday, 31 March 2015 at 20:07:19 UTC, Walter Bright wrote:
Consider the code:
struct Foo { C obj; }
@safe void bar(C c, Foo* f) { f.obj = c; }
Under the proposal, for @safe code, this would have to be written as:
@safe void bar(static C c, Foo* f) ...
Either that, or:
@safe void bar(C c return!f, Foo* f) ...
or:
@safe void bar()(C c, Foo* f) ...
I'm concerned this may break an astonishing amount of code.
We could go a step further and infer scope for all @safe functions, except those
with at least one explicit annotation. This would be almost backward compatible,
except for the fact that some arguments then suddenly become scope and therefore
the mangling changes.
The problem with inferring attributes on non-template functions is the source
must be available to the compiler. That's obviously true for templates, but not
so true for non-templates, where only the signature is available.
I'm thinking of a modest step which would be a subset of your proposal:
1. implement 'scope' and 'return' for arrays, classes, and pointers
2. implement inference for templates and lambdas
3. enable it with the -dip25 switch
and see how far that takes us.
