It's still best practice to put `@setEvalBranchQuota` at the top of
the comptime stack, but as Jimmi notes in #1949, when a function
can be called at comptime and also can be the top of the comptime stack,
this compile error is fundamentally unsound.
So now it's gone.
closes#1949
This deletes some legacy cruft, and produces leaner object files.
Example:
```
var x: i32 = 1234;
export fn entry() i32 {
return x;
}
```
This produces:
```
@x = internal unnamed_addr global i32 1234, align 4
@0 = internal unnamed_addr constant i32* @x, align 8
```
and @0 is never even used. After this commit, @0 is not produced.
This fixes a bug: Zig was creating invalid LLVM IR when one of these
globals that shouldn't exist takes the address of a thread local
variable. In LLVM 8.0.0rc2, it would produce a linker error. But
probably after my bug report is solved it will be caught by the IR
verifier.
https://bugs.llvm.org/show_bug.cgi?id=40652
There's an unfortunate footgun in the current design of error sets.
The debug info type for every error set is the same as the debug info
type of the global error set, which is essentially an enum forward
declaration. The problem is that when we "replace" the forward
declaration with the final value, once we know all the possible errors,
we have to update the pointers of every error set.
So the footgun is that if you ever copy the debug info type of the
global error set, you have to add the address of the pointer to a list
of pointers that need to be updated once we "replace" the forward
declaration. I activated the footgun when I introduced the optimization
that `?anyerror` types are the same size as `anyerror` types (using 0 as
the null value), because I introduced a pointer copy of the global error
set debug info type, but forgot to add it to the list.
I'm sure that there is a better way to code this, which does not have
the footgun, but this commit contains only a fix, not a reworking of the
logic.
closes#1937
this would work if @llvm.sadd.with.overflow supported
vectors, which it does in trunk. but it does not support
them in llvm 7 or even in llvm 8 release branch.
so the next commit after this will have to do a different
strategy, but when llvm 9 comes out it may be worth coming
back to this one.
Previously, std.debug.assert would `@panic` in test builds,
if the assertion failed. Now, it's always `unreachable`.
This makes release mode test builds more accurately test
the actual code that will be run.
However this requires tests to call `std.testing.expect`
rather than `std.debug.assert` to make sure output is correct.
Here is the explanation of when to use either one, copied from
the assert doc comments:
Inside a test block, it is best to use the `std.testing` module
rather than assert, because assert may not detect a test failure
in ReleaseFast and ReleaseSafe mode. Outside of a test block, assert
is the correct function to use.
closes#1304
* zig fmt
* std.mem.join takes a slice of slices instead of var args
* std.mem.join takes a separator slice rather than byte,
and always inserts it. Previously it would not insert the separator
if there already was one, violating the documented behavior.
* std.mem.join calculates exactly the correct amount to allocate
and has no call to allocator.shrink()
* bring back joinWindows and joinPosix and the corresponding tests.
it is intended to be able to call these functions from any OS.
* rename std.os.path.resolveSlice to resolve (now resolve takes
a slice of slices instead of var args)
this should actually improve CI times a bit too
See the description at the top of std/os/startup.zig (deleted in this
commit) for a more detailed understanding of what this commit does.
* rename std.mem.split to std.mem.tokenize
* add future deprecation notice to docs
* (unrelated) add note to std.os.path.resolve docs
* std.mem.separate - assert delimiter.len not zero
* fix implementation of std.mem.separate to respect the delimiter
* separate the two iterators to different structs
`std.mem.Allocator.createOne` is renamed to `std.mem.Allocator.create`.
The problem with the previous API is that even after copy elision,
the initalization value passed as a parameter would always be a copy.
With the new API, once copy elision is done, initialization
functions can directly initialize allocated memory in place.
Related:
* #1872
* #1873
