The idea is to separate the notion of const from 'this variable can always be
replaced with its initializer' in the frontend. To do that, I introduced
Declaration::isSameAsInitializer, which is overridden in VarDeclaration to
return false for constants that have a struct literal initializer.
So
{{{
const S s = S(5);
void foo() { auto ps = &s; }
// is no longer replaced by
void foo() { auto ps = &(S(5)); }
}}}
To make taking the address of a struct constant with a struct-initializer
outside of function scope possible, I made sure that AddrExp::optimize doesn't
try to run the argument's optimization with WANTinterpret - that'd again
replace the constant with a struct literal temporary.
is found to see if it can actually happen instead of just assuming it will.
This allows it to catch cases like
{{{
int i;
Foo f;
while (cond(i))
f = new Foo(i++);
return f.value;
}}}
where it previously wouldn't because a phi using the allocation would appear in
the condition block to propagate it to the use after the loop.
allocations too. (A "simple" loop is one where the allocation isn't used in a
subsequent iteration)
This also means it's no longer necessary to run this pass multiple times.
Running it once after inlining should now catch all cases.
`available_externally` linkage. This allows the inliner to inline functions from
other modules while telling the code generator to ignore those functions (treat
them as declarations)
Still generates a few extra `TypeInfo`s and strings...
Disabled when generating debug info because I don't really understand it, and it
doesn't like this.
generating a constant to fill the entire struct with a single `store`.
This is much more efficient at compile time (fixing #320) and vastly reduces
the size of the emitted code. Since LLVM no longer needs to keep the data for
all fields in "registers" until the store happens, it should also be more
efficient at run time in cases where the fields aren't assigned with constants.
There's also some code clean-up by removing duplicated logic.
Applied function type from D1 frontend that got removed in D2, it's critical for member function type to be correct.
Fixed a bunch of type discrepancies in druntime object.di vs. genobj.d .
Disabled (#if 0) some potentally very large type dumps for -vv .
Updated classinfo and typeinfo generation for D2, almost complete now.
Added finer grained checks for vtbl type mismatching, aids debugging.
