Decl objects need to know whether they are the owner of the Type/Value
associated with them, in order to decide whether to destroy the
associated Namespace, Fn, or Var when cleaning up.
This commit takes advantage of the new "NameStrategy" that is exposed
in the ZIR in order to name Decls after their parent when asked to. This
makes the next failing test case pass.
which can be either parent, func, or anon. Here's the enum reproduced in
the commit message for convenience:
```zig
pub const NameStrategy = enum(u2) {
/// Use the same name as the parent declaration name.
/// e.g. `const Foo = struct {...};`.
parent,
/// Use the name of the currently executing comptime function call,
/// with the current parameters. e.g. `ArrayList(i32)`.
func,
/// Create an anonymous name for this declaration.
/// Like this: "ParentDeclName_struct_69"
anon,
};
```
With this information in the ZIR, a future commit can improve the
names of structs, unions, enums, and opaques.
In order to accomplish this, the following ZIR instruction forms were
removed and replaced with Extended op codes:
* struct_decl
* struct_decl_packed
* struct_decl_extern
* union_decl
* union_decl_packed
* union_decl_extern
* enum_decl
* enum_decl_nonexhaustive
By being extended opcodes, one more u32 is needed, however we more than
make up for it by repurposing the 16 "small" bits to provide shorter
encodings for when decls_len == 0, fields_len == 0, a source node is not
provided, etc. There tends to be no downside, and in fact sometimes
upsides, to using an extended op code when there is a need for flag
bits, which is the case for all three of these. Likewise, the container
layout can be encoded in these bits rather than into the opcode.
The following 4 ZIR instructions were added, netting a total of 4 freed
up ZIR enum tags for future use:
* opaque_decl_anon
* opaque_decl_func
* error_set_decl_anon
* error_set_decl_func
This is so that opaques and error sets can have the same name hint as
structs, enums, and unions.
`std.builtin.ContainerLayout` gets an explicit integer tag type so that
it can be used inside packed structs.
This commit also makes `Module.Namespace` use a separate set for
anonymous decls, thus allowing anonymous decls to share the same
`Decl.name` as their owner `Decl` objects.
Some of the reworkings in this branch put us over the limit, on Linux,
where the kernel disregards the fact that we ask for 16 MiB in the ELF
file. So we ask for more stack space in `main`.
* Sema: implement global variables
- Improved global constants to stop needlessly creating a Var
structure; they can just store the value directly.
- This required making memory management a bit more sophisticated to
detect when a Decl owns the Namespace associated with it, for the
purposes of deinitialization.
* Decl.name and Namespace decl table keys no longer directly
reference ZIR; instead they have heap-duped names, so that deleted
decls, which no longer have any ZIR to reference for their names, can
be removed from the parent Namespace table.
- In the future I would like to explore going a different direction
with this, where the strings would still point to the ZIR however
they would be removed from their owner Namespace objects during the
update detection. The design principle here is that the existence
of incremental compilation as a feature should not incur any cost
for the use case when it is not used. In this example Decl names
could simply point to ZIR string table memory, and it is only
because of incremental compilation that we duplicate their names.
* AstGen: implement threadlocal variables
* CLI: call cleanExit after building a compilation so that in release
modes we don't bother freeing memory or closing file descriptors,
allowing the OS to do it more efficiently.
* Avoid calling `freeDecl` in the linker for unreferenced Decl objects.
* Fix CBE test case expecting the compile error to point to the wrong
column.
After this commit, `pub export fn main() c_int { ... }` will be
correctly detected as the intended entry point, and therefore start code
will not try to export its own conflicting `main` function.
* Implement basic union support
- lots of stuff is still TODO, including runtime field access
- also TODO: resolving the union tag type
- comptime field access is implemented
* DRY up some code by using the `Zir.DeclIterator` for skipping over
decls in structs and unions.
* Start to clean up Sema with regards to calling `.value()` to find out
a const value. Instead, Sema code should call one of these two:
- `resolvePossiblyUndefinedValue` (followed by logic dealing with
undefined values)
- `resolveDefinedValue` (a compile error will be emitted if the value
is undefined)
* An exported function with an unspecified calling convention gets the
C calling convention.
* Implement comptime field access for structs.
* Add another implementation of "type has one possible value" in Sema.
This is a bit unfortunate since the logic is duplicated, but the one
in Type asserts that the types are resolved already, and is
appropriate to call from codegen, while the one in Sema performs
type resolution if necessary, reporting any compile errors that occur
in the process.
The goal is to get start code to be able to inspect the calling
convention of `main` in order to determine whether to export a main for
libc to call, or to allow the root source file to do it.
Now that they are lazy, they need to get analyzed in the correct
context, when requested.
This commit also hooks up std.builtin type values being resolved
properly. This is needed, for example, with the `@export` builtin
function, which occurs in start.zig, for `std.builtin.ExportOptions`.
The ZIR code uses the special `Ref.export_options` value, and semantic
analysis has to map this to the corresponding type from `std.builtin`.
AstGen is now completely independent from the rest of the compiler. It
ingests an AST tree and produces ZIR code as the output, without
depending on any of the glue code of the compiler.
instead of node indexes.
* AstGen: dbg_stmt instructions now have line and column indexes,
relative to the parent declaration. This allows codegen to emit debug
info without having the source bytes, tokens, or AST nodes loaded
in memory.
* ZIR: each decl has the absolute line number. This allows computing
line numbers from offsets without consulting source code bytes.
Memory management: creating a function definition does not prematurely
set the Decl arena. Instead the function is allocated with the general
purpose allocator.
Codegen no longer looks at source code bytes for any reason. They can
remain unloaded from disk.
* AstGen: LocalVal and LocalPtr use string table indexes for their
names. This is more efficient because local variable declarations do
need to include the variable names so that semantic analysis can emit
a compile error if a declaration is shadowed. So we take advantage of
this fact by comparing string table indexes when resolving names.
* The arg ZIR instructions are needed for the above reasoning, as well
as to emit equivalent AIR instructions for debug info.
Now that we have these arg instructions, get rid of the special
`Zir.Inst.Ref` range for parameters. ZIR instructions now refer
to the arg instructions for parameters.
* Move identAsString and strLitAsString from Module.GenZir to AstGen
where they belong.
Before there was this "top_decl" and "tmp_namespace" stack values that
were kludgy and buggy. Now Sema is slightly reworked so that files which
are structs are analyzed with their own Decl and Namespace already set
up.
After this commit there are no memory leaks for a successful build-obj.
* AstGen: add missing `break_inline` for comptime blocks.
* Module: call getTree() in byteOffset(). This generates the AST when
using cached ZIR and compile errors need to be reported.
* Scope.File: distinguish between successful ZIR generation and AIR
generation (when Decls in scope have been scanned).
- `semaFile` correctly avoids doing work twice.
* Implement first pass at `lookupInNamespace`. It has various TODOs
left, such as `usingnamespace`, and setting up Decl dependencies.
* Remove some unused imports in AstGen.zig. I think it would make sense
to start decoupling AstGen from the rest of the compiler code,
similar to how the tokenizer and parser are decoupled.
* AstGen: For decls, move the block_inline instructions to the top of
the function so that they get lower ZIR instruction indexes. With
this, the block_inline instruction index combined with its corresponding
break_inline instruction index can be used to form a ZIR instruction
range. This is useful for allocating an array to map ZIR instructions
to semantically analyzed instructions.
* Module: extract emit-h functionality into a struct, and only allocate
it when emit-h is activated.
* Module: remove the `decl_table` field. This previously was a table of
all Decls in the entire Module. A "name hash" strategy was used to
find decls within a given namespace, using this global table. Now,
each Namespace has its own map of name to children Decls.
- Additionally, there were 3 places that relied on iterating over
decl_table in order to function:
- C backend and SPIR-V backend. These now have their own decl_table
that they keep populated when `updateDecl` and `removeDecl` are
called.
- emit-h. A `decl_table` field has been added to the new GlobalEmitH
struct which is only allocated when emit-h is activated.
* Module: fix ZIR serialization/deserialization bug in debug mode having
to do with the secret safety tag for untagged unions. There is still an
open TODO to investigate a friendlier solution to this problem with
the language.
* Module: improve deserialization of ZIR to allocate only exactly as
much capacity as length in the instructions array so as to not waste
space.
* Module: move `srcHashEql` to `std.zig` to live next to the definition
of `SrcHash` itself.
* Module: re-introduce the logic for scanning top level declarations
within a namespace.
* Compilation: add an `analyze_pkg` Job which is used to kick off the
start of semantic analysis by doing the equivalent of
`_ = @import("std");`. The `analyze_pkg` job is unconditionally added
to the work queue on every update(), with pkg set to the std lib pkg.
* Rename TZIR to AIR in a few places. A more comprehensive rename will
come later.
Inside a nosuspend block, emit function calls as nosuspend calls.
Also inside a comptime block, emit function calls as comptime calls.
Also emit `async foo()` calls as async calls.
Remove compile error for `nosuspend` block inside `suspend` block.
Instead of implicitly treating every `suspend` block also as a
`nosuspend` block (which would make sense), we leave suspension points
as compile errors, to hint to the programmer about accidents. Of course
they may then assert `nosuspend` by introducing a block within their
suspend block.
To make room in `Zir.Inst.Tag` I moved `typeof_peer` and `compile_log`
to `Extended`.
* AstGen: implement `anyframe_literal` and `anyframe_type`.
* Introduce `makeSubBlock` to avoid redundant AstGen code for GenZir
scopes. Allows adding/removing a field without possibility of
accidentally introducing a bug of forgetting to set the new field.
* Add to GenZir `nosuspend_node` and `suspend_node` in preparation for
implementing `suspend` blocks and `nosuspend` blocks.
* AstGen: fix assembly to support clobbers, multiple outputs, and
outputs without `->` syntax.
- `asm` and `asm_volatile` move to `Extended` enum with `small` being
repurposed for a few things. This frees up 2 ZIR tags, 1 of which
is used in this commit and 1 is leftover.
* AstGen: fix `simple_types` incorrectly having multiple conflicting
values for "undefined" and "null".
- Also add "anyframe" to `simple_types`.
* Add `anyframe_type` to type.zig, value.zig and `Zir.Inst.Ref`.
- Also add i128 and u128 types to `Zir.Inst.Ref` and `simple_types`.
* Sema/Zir: Fix incorrect math causing the function body to be messed
up for Extended-encoded functions.
* Zir: support `i32` fields for "extra" payloads.
