Return the NRVO variable by reference

[ibuclaw]

This should fix the broken builds after updating to latest GCC snapshot. The test in sdtor has always been flaky because of this anyway. With this the compiler should now consistently generate the same code regardless of optimization level.

[ibuclaw]

Ah, I hadn't updated my local copy of gcc to the latest. I can make it work, but only with http://bugzilla.gdcproject.org/show_bug.cgi?id=42 fixed. And that breaks a lot of other stuff with it.

I'll get back in a while when I've done some triaging.

[ibuclaw]

And it is (more or less) working! At least the build failure is because of an error in the codegen, rather than an ICE. :-)

[ibuclaw]

Split out the more complex logic in #221. Now the change is rather trivial.

[ibuclaw]

I might finally be at the drum-roll moment...

[ibuclaw]

Three failed unittests, testsuite passes. This may be a sign that there may still be more cases that cause silent corruption of codegen as a result of this.

[ibuclaw]

That aren't caught by testsuite nor unittester, that is.

[ibuclaw]

Now passing the unittester. :-o

[ibuclaw]

It turned out that the frontend returns false from isPod for nested structs. However unlike structs with this(this) or ~this(), we can't pass them around by reference. But at the same time, I can't see a nice way to force them in memory.

So I'm adjusting the condition in TypeVisitor(TypeStruct) to only set TREE_ADDRESSABLE if sym>postblit || sym->dtor.

[ibuclaw]

Passes!

[jpf91]

Nice work! Now I'll just have to port this to all other branches ;-)

[ibuclaw]

Looks like there's more to be done. I just updated to the latest snapshot, and get more ICE's relating to setting TREE_ADDRESSABLE. This may need some ironing yet.

[jpf91]

Maybe this is related:

Failing test for GCC 5:
https://semaphoreci.com/jpf91/gdc/branches/gdc-5/builds/45

ICE while compiling phobos: GCC 4.9/4.8
https://semaphoreci.com/jpf91/gdc/branches/gdc-4-9/builds/41
https://semaphoreci.com/jpf91/gdc/branches/gdc-4-8/builds/43

[ibuclaw]

Nope, I get something other. I'd imagine that some things may need to be adjusted for older versions of GCC.