Previously, the compiler had special logic to determine whether to
include the startup code, which was in `std/special/start.zig`. Now,
the file is moved to `std/start.zig`, and there is no special logic
in the compiler. Instead, the standard library unconditionally imports
the `start.zig` file, which then has a `comptime` block that does the
logic of determining what, if any, start symbols to export. Instead of
`start.zig` being in its own special package, it is just another normal
file that is part of the standard library.
`std.builtin.TestFn` is now part of the standard library rather than
specially generated by the compiler.
Total bytes used in stage1 std lib tests:
3.418 -> 3.198 GiB (saving 225 MiB)
There's still this from pass 1 not getting freed:
Const: 6909049 items, 72 bytes each, total 474.407 MiB
This is due to 2 things hanging on to references to IrAnalyze pointers:
* ZigVar->owner_exec->analysis
* LazyValue->ira
The LazyValue one could be solved by memoizing the results after the
lazy value is resolved, and then it could unref the IrAnalyze.
ZigVars that are determined to be comptime const, could have their
const_value set to that value, instead of using the mem_slot_index
mechanism. This would prevent an IrAnalyze ref in some cases.
Having ConstGlobalRefs be a pointer in ZigValue was a hack that caused
plenty of bugs. It was used to work around difficulties in type coercing
array values into slices.
However, after #3787 is merged, array values no longer type coerce into
slices, and so this provided an opportunity to clean up the code.
This has the nice effect of reducing stage1 peak RAM usage during the
std lib tests from 3.443 GiB to 3.405 GiB (saving 39 MiB).
There is one behavior test failing in this branch, which I plan to debug
after merging #3787.
