The `comptime_args` field of Fn has a clarified purpose:
For generic function instantiations, there is a `TypedValue` here
for each parameter of the function:
* Non-comptime parameters are marked with a `generic_poison` for the value.
* Non-anytype parameters are marked with a `generic_poison` for the type.
Sema now has a `fn_ret_ty` field. Doc comments reproduced here:
> When semantic analysis needs to know the return type of the function whose body
> is being analyzed, this `Type` should be used instead of going through `func`.
> This will correctly handle the case of a comptime/inline function call of a
> generic function which uses a type expression for the return type.
> The type will be `void` in the case that `func` is `null`.
Various places in Sema are modified in accordance with this guidance.
Fixed `resolveMaybeUndefVal` not returning `error.GenericPoison` when
Value Tag of `generic_poison` is encountered.
Fixed generic function memoization incorrect equality checking. The
logic now clearly deals properly with any combination of anytype and
comptime parameters.
Fixed not removing generic function instantiation from the table in case
a compile errors in the rest of `call` semantic analysis. This required
introduction of yet another adapter which I have called
`GenericRemoveAdapter`. This one is nice and simple - it's the same hash
function (the same precomputed hash is passed in) but the equality
function checks pointers rather than doing any logic.
Inline/comptime function calls coerce each argument in accordance with
the function parameter type expressions. Likewise the return type
expression is evaluated and provided (see `fn_ret_ty` above).
There's a new compile error "unable to monomorphize function". It's
pretty unhelpful and will need to get improved in the future. It happens
when a type expression in a generic function did not end up getting
resolved at a callsite. This can happen, for example, if a runtime
parameter is attempted to be used where it needed to be comptime known:
```zig
fn foo(x: anytype) [x]u8 { _ = x; }
```
In this example, even if we pass a number such as `10` for `x`, it is
not marked `comptime`, so `x` will have a runtime known value, making
the return type unable to resolve.
In the LLVM backend I implement cmp instructions for float types to pass
some behavior tests that used floats.
For now these errors are handled via `@panic` rather than `unreachable`.
These are relatively likely bugs to occur at this early stage of
development, and handling them as panics lets us ship release builds
of the compiler without worrying about undefined behavior.
Furthermore, in stage1, `@panic` is implemented to include an error
return trace, while `unreachable` is not. In this case, the error return
traces are extremely helpful in debugging the compiler.
* ZIR encoding for function instructions have a body for the return
type. This lets Sema for generic functions do the same thing it does
for parameters, handling `error.GenericPoison` in the evaluation of
the return type by marking the function as generic.
* Sema: fix missing block around the new Decl arena finalization. This
led to a memory corruption.
* Added some floating point support to the LLVM backend but didn't get
far enough to pass any new tests.
Module has a new field `monomorphed_funcs` which stores the set of
`*Module.Fn` objects which are generic function instantiations.
The hash is based on hashes of comptime values of parameters known to be
comptime based on an explicit comptime keyword or must-be-comptime
type expressions that can be evaluated without performing monomorphization.
This allows function calls to be semantically analyzed cheaply for
generic functions which are already instantiated.
The table is updated with a single `getOrPutAdapted` in the semantic
analysis of `call` instructions, by pre-allocating the `Fn` object and
passing it to the child `Sema`.
AstGen result locations now have a `coerced_ty` tag which is the same as
`ty` except it assumes that Sema will do a coercion, so it does not
redundantly add an `as` instruction into the ZIR code. This results in
cleaner ZIR and about a 14% reduction of ZIR bytes.
param and param_comptime ZIR instructions now have a block body for
their type expressions. This allows Sema to skip evaluation of the
block in the case that the parameter is comptime-provided. It also
allows a new mechanism to function: when evaluating type expressions of
generic functions, if it would depend on another parameter, it returns
`error.GenericPoison` which bubbles up and then is caught by the
param/param_comptime instruction and then handled.
This allows parameters to be evaluated independently so that the type
info for functions which have comptime or anytype parameters will still
have types populated for parameters that do not depend on values of
previous parameters (because evaluation of their param blocks will return
successfully instead of `error.GenericPoison`).
It also makes iteration over the block that contains function parameters
slightly more efficient since it now only contains the param
instructions.
Finally, it fixes the case where a generic function type expression contains
a function prototype. Formerly, this situation would cause shared state
to clobber each other; now it is in a proper tree structure so that
can't happen. This fix also required adding a field to Sema
`comptime_args_fn_inst` to make sure that the `comptime_args` field
passed into Sema is applied to the correct `func` instruction.
Source location for `node_offset_asm_ret_ty` is fixed; it was pointing at
the asm output name rather than the return type as intended.
Generic function instantiation is fixed, notably with respect to
parameter type expressions that depend on previous parameters, and with
respect to types which must be always comptime-known. This involves
passing all the comptime arguments at a callsite of a generic function,
and allowing the generic function semantic analysis to coerce the values
to the proper types (since it has access to the evaluated parameter type
expressions) and then decide based on the type whether the parameter is
runtime known or not. In the case of explicitly marked `comptime`
parameters, there is a check at the semantic analysis of the `call`
instruction.
Semantic analysis of `call` instructions does type coercion on the
arguments, which is needed both for generic functions and to make up for
using `coerced_ty` result locations (mentioned above).
Tasks left in this branch:
* Implement the memoization table.
* Add test coverage.
* Improve error reporting and source locations for compile errors.
The general strategy is that Sema will pre-map comptime arguments into
the inst_map, and then re-run the block body that contains the `param`
and `func` instructions. This re-runs all the parameter type expressions
except with comptime values populated.
In Sema, param instructions are now handled specially: they detect
whether they are comptime-elided or not. If so, they skip putting a
value in the inst_map, since it is already pre-populated. If not, then
they append to the `fields` field of `Sema` for use with the `func`
instruction.
So when the block body is re-run, a new function is generated with
all the comptime arguments elided, and the new function type has only
runtime parameters in it. TODO: give the generated Decls better names
than "foo__anon_x".
The new function is then added to the work queue to have its body
analyzed and a runtime call AIR instruction to the new function is
emitted.
When the new function gets semantically analyzed, comptime parameters are
pre-mapped to the corresponding `comptime_args` values rather than
mapped to an `arg` AIR instruction. `comptime_args` is a new field that
`Fn` has which is a `TypedValue` for each parameter. This field is non-null
for generic function instantiations only. The values are the comptime
arguments. For non-comptime parameters, a sentinel value is used. This is
because we need to know the information of which parameters are
comptime-known.
Additionally:
* AstGen: align and section expressions are evaluated in the scope that
has comptime parameters in it.
There are still some TODO items left; see the BRANCH_TODO file.
* ZIR function instructions encode the index of the block that
contains the function instruction. This allows Zig to later scan the
block and find the parameter instructions, which is needed for
semantically analyzing function bodies.
* Runtime function calls insert AIR arg instructions and then inserts
Sema inst_map entries mapping the ZIR param instructions to them.
* comptime/inline function call inserts Sema inst_map entries mapping
the ZIR param instructions to the AIR callsite arguments.
With this commit we are back to the tests passing.
ZIR encoding for functions is changed in preparation for generic
function support. As an example:
```zig
const std = @import("std");
const expect = std.testing.expect;
test "example" {
var x: usize = 0;
x += checkSize(i32, 1);
x += checkSize(bool, true);
try expect(x == 5);
}
fn checkSize(comptime T: type, x: T) usize {
_ = x;
return @sizeOf(T);
}
```
Previous ZIR for the `checkSize` function:
```zir
[165] checkSize line(10) hash(0226f62e189fd0b1c5fca02cf4617562): %55 = block_inline({
%56 = decl_val("T") token_offset:11:35
%57 = as_node(@Ref.type_type, %56) node_offset:11:35
%69 = extended(func([comptime @Ref.type_type, %57], @Ref.usize_type, {
%58 = arg("T") token_offset:11:23
%59 = as_node(@Ref.type_type, %58) node_offset:11:35
%60 = arg("x") token_offset:11:32
%61 = dbg_stmt(11, 4)
```
ZIR for the `checkSize` function after this commit:
```zir
[157] checkSize line(10) hash(0226f62e189fd0b1c5fca02cf4617562): %55 = block_inline({
%56 = param_comptime("T", @Ref.type_type) token_offset:11:23
%57 = as_node(@Ref.type_type, %56) node_offset:11:35
%58 = param("x", %57) token_offset:11:32
%67 = func(@Ref.usize_type, {
%59 = dbg_stmt(11, 4)
```
Noted differences:
* Previously the type expression was redundantly repeated.
* Previously the parameter names were redundantly stored in the ZIR
extra array.
* Instead of `arg` ZIR instructions as the first instructions within a
function body, they are now outside the function body, in the same
block as the `func` instruction. There are variants:
- param
- param_comptime
- param_anytype
- param_anytype_comptime
* The param instructions additionally encode the type.
* Because of the param instructions, the `func` instruction no longer
encodes the list of parameter types or the comptime bits.
It's implied that Sema will collect the parameters so that when a `func`
instruction is encountered, they will be implicitly used to construct
the function's type. This is so that we can satisfy all 3 ways of
performing semantic analysis on a function:
1. runtime: Sema will insert AIR arg instructions for each parameter,
and insert into the Sema inst_map ZIR param => AIR arg.
2. comptime/inline: Sema will insert into the inst_map ZIR param =>
callsite arguments.
3. generic: Sema will map *only the comptime* ZIR param instructions to
the AIR instructions for the comptime arguments at the callsite, and
then re-run Sema for the function's Decl. This will produce a new
function which is the monomorphized function.
Additionally:
* AstGen: Update usage of deprecated `ensureCapacity` to
`ensureUnusedCapacity` or `ensureTotalCapacity`.
* Introduce `Type.fnInfo` for getting a bunch of data about a function
type at once, and use it in `analyzeCall`.
This commit starts a branch to implement generic functions in stage2.
Test regressions have not been addressed yet.
* `Value.toType` accepts a buffer parameter instead of an allocator
parameter and can no longer fail.
* Module: remove the unused `mod: *Module` parameter from various
functions.
* `Value.compare` now accepts a `Type` parameter which indicates the
type of both operands. There is also a `Value.compareHetero` which
accepts only Value parameters and supports comparing mixed types.
Likewise, `Value.eql` requires a `Type` parameter.
* `Value.hash` is removed; instead the hash map context structs now
have a `ty: Type` field, and the hash function lives there, where it
has access to a Value's Type when it computes a hash.
- This allowed the hash function to be greatly simplified and sound
in the sense that the same Values, even with different
representations, always hash to the same thing.
* Sema: Fix source location of zirCmp when an operand is runtime known
but needs to be comptime known.
* Remove unused target parameter from `Value.floatCast`.
After this change, the frontend and backend cooperate to keep track of
which Decls are actually emitted into the machine code. When any backend
sees a `decl_ref` Value, it must mark the corresponding Decl `alive`
field to true.
This prevents unused comptime data from spilling into the output object
files. For example, if you do an `inline for` loop, previously, any
intermediate value calculations would have gone into the object file.
Now they are garbage collected immediately after the owner Decl has its
machine code generated.
In the frontend, when it is time to send a Decl to the linker, if it has
not been marked "alive" then it is deleted instead.
Additional improvements:
* Resolve type ABI layouts after successful semantic analysis of a
Decl. This is needed so that the backend has access to struct fields.
* Sema: fix incorrect logic in resolveMaybeUndefVal. It should return
"not comptime known" instead of a compile error for global variables.
* `Value.pointerDeref` now returns `null` in the case that the pointer
deref cannot happen at compile-time. This is true for global
variables, for example. Another example is if a comptime known
pointer has a hard coded address value.
* Binary arithmetic sets the requireRuntimeBlock source location to the
lhs_src or rhs_src as appropriate instead of on the operator node.
* Fix LLVM codegen for slice_elem_val which had the wrong logic for
when the operand was not a pointer.
As noted in the comment in the implementation of deleteUnusedDecl, a
future improvement will be to rework the frontend/linker interface to
remove the frontend's responsibility of calling allocateDeclIndexes.
I discovered some issues with the plan9 linker backend that are related
to this, and worked around them for now.
* AIR no longer has a `variables` array. Instead of the `varptr`
instruction, Sema emits a constant with a `decl_ref`.
* AIR no longer has a `ref` instruction. There is no longer any
instruction that takes a value and returns a pointer to it. If this
is desired, Sema must either create an anynomous Decl and return a
constant `decl_ref`, or in the case of a runtime value, emit an
`alloc` instruction, `store` the value to it, and then return the
`alloc`.
* The `ref_val` Value Tag is eliminated. `decl_ref` should be used
instead. Also added is `eu_payload_ptr` which points to the payload
of an error union, given an error union pointer.
In general, Sema should avoid calling `analyzeRef` if it can be helped.
For example in the case of field_val and elem_val, there should never be
a reason to create a temporary (alloc or decl). Recent previous commits
made progress along that front.
There is a new abstraction in Sema, which looks like this:
var anon_decl = try block.startAnonDecl();
defer anon_decl.deinit();
// here 'anon_decl.arena()` may be used
const decl = try anon_decl.finish(ty, val);
// decl is typically now used with `decl_ref`.
This pattern is used to upgrade `ref_val` usages to `decl_ref` usages.
Additional improvements:
* Sema: fix source location resolution for calling convention
expression.
* Sema: properly report "unable to resolve comptime value" for loads of
global variables. There is now a set of functions which can be
called if the callee wants to obtain the Value even if the tag is
`variable` (indicating comptime-known address but runtime-known value).
* Sema: `coerce` resolves builtin types before checking equality.
* Sema: fix `u1_type` missing from `addType`, making this type have a
slightly more efficient representation in AIR.
* LLVM backend: fix `genTypedValue` for tags `decl_ref` and `variable`
to properly do an LLVMConstBitCast.
* Remove unused parameter from `Value.toEnum`.
After this commit, some test cases are no longer passing. This is due to
the more principled approach to comptime references causing more
anonymous decls to get sent to the linker for codegen. However, in all
these cases the decls are not actually referenced by the runtime machine
code. A future commit in this branch will implement garbage collection
of decls so that unused decls do not get sent to the linker for codegen.
This will make the tests go back to passing.
Frontend improvements:
* When compiling in `zig test` mode, put a task on the work queue to
analyze the main package root file. Normally, start code does
`_ = import("root");` to make Zig analyze the user's code, however in
the case of `zig test`, the root source file is the test runner.
Without this change, no tests are picked up.
* In the main pipeline, once semantic analysis is finished, if there
are no compile errors, populate the `test_functions` Decl with the
set of test functions picked up from semantic analysis.
* Value: add `array` and `slice` Tags.
LLVM backend improvements:
* Fix incremental updates of globals. Previously the
value of a global would not get replaced with a new value.
* Fix LLVM type of arrays. They were incorrectly sending
the ABI size as the element count.
* Remove the FuncGen parameter from genTypedValue. This function is for
generating global constants and there is no function available when
it is being called.
- The `ref_val` case is now commented out. I'd like to eliminate
`ref_val` as one of the possible Value Tags. Instead it should
always be done via `decl_ref`.
* Implement constant value generation for slices, arrays, and structs.
* Constant value generation for functions supports the `decl_ref` tag.
* Add AIR instruction: struct_field_val
- This is part of an effort to eliminate the AIR instruction `ref`.
- It's implemented for C backend and LLVM backend so far.
* Rename `resolvePossiblyUndefinedValue` to `resolveMaybeUndefVal` just
to save some columns on long lines.
* Sema: add `fieldVal` alongside `fieldPtr` (renamed from
`namedFieldPtr`). This is part of an effort to eliminate the AIR
instruction `ref`. The idea is to avoid unnecessary loads, stores,
stack usage, and IR instructions, by paying a DRY cost.
LLVM backend improvements:
* internal linkage vs exported linkage is implemented, along with
aliases. There is an issue with incremental updates due to missing
LLVM API for deleting aliases; see the relevant comment in this commit.
- `updateDeclExports` is hooked up to the LLVM backend now.
* Fix usage of `Type.tag() == .noreturn` rather than calling `isNoReturn()`.
* Properly mark global variables as mutable/constant.
* Fix llvm type generation of function pointers
* Fix codegen for calls of function pointers
* Implement llvm type generation of error unions and error sets.
* Implement AIR instructions: addwrap, subwrap, mul, mulwrap, div,
bit_and, bool_and, bit_or, bool_or, xor, struct_field_ptr,
struct_field_val, unwrap_errunion_err, add for floats, sub for
floats.
After this commit, `zig test` on a file with `test "example" {}`
correctly generates and executes a test binary. However the
`test_functions` slice is undefined and just happens to be going into
the .bss section, causing the length to be 0. The next step towards
`zig test` will be replacing the `test_functions` Decl Value with the
set of test function pointers, before it is sent to linker/codegen.
* There is now a main_pkg in addition to root_pkg. They are usually the
same. When using `zig test`, main_pkg is the user's source file and
root_pkg has the test runner.
* scanDecl no longer looks for test decls outside the package being
tested. honoring `--test-filter` is still TODO.
* test runner main function has a void return value rather than
`anyerror!void`
* Sema is improved to generate better AIR for for loops on slices.
* Sema: fix incorrect capacity calculation in zirBoolBr
* Sema: add compile errors for trying to use slice fields as an lvalue.
* Sema: fix type coercion for error unions
* Sema: fix analyzeVarRef generating garbage AIR
* C codegen: fix renderValue for error unions with 0 bit payload
* C codegen: implement function pointer calls
* CLI: fix usage text
Adds 4 new AIR instructions:
* slice_len, slice_ptr: to get the ptr and len fields of a slice.
* slice_elem_val, ptr_slice_elem_val: to get the element value of
a slice, and a pointer to a slice.
AstGen gains a new functionality:
* One of the unused flags of struct decls is now used to indicate
structs that are known to have non-zero size based on the AST alone.
Previously we had codegen_decl for both constant values as well as
function bodies. A recent commit updated the linker backends to add
updateFunc as a separate function than updateDecl, and now this commit
does the same with work queue tasks.
The frontend now distinguishes between function pointers and function
bodies.
Now supports multiple items pointing to the same body. This is a common
pattern even when using a jump table, with multiple cases pointing to
the same block of code.
In the case of a range specified, the items are moved to branches in the
else body. A future improvement may make it possible to have jump table
items as well as ranges pointing to the same block of code.
* some instructions are not implemented yet
* fix off-by-1 in Air.getMainBody
* Compilation: use `@import("builtin")` rather than `std.builtin`
for the values that are different for different build configurations.
* Sema: avoid calling `addType` in between
air_instructions.ensureUnusedCapacity and corresponding
appendAssumeCapacity because it can possibly add an instruction.
* Value: functions print their names
Additionally: ZIR encoding for floats now supports float literals up to
f64, not only f32. This is because we no longer need a source location
for this instruction.
Now you can pass `.unneeded` for a `LazySrcLoc` and if there ended up
being a compile error that needed it, you'll get
`error.NeededSourceLocation`.
Callsites can now exploit this error to do the expensive computation
to produce a source location object and then repeat the operation.
Now the branch is compiling again, provided that one uses
`-Dskip-non-native`, but many code paths are disabled. The code paths
can now be re-enabled one at a time and updated to conform to the new
AIR memory layout.
to the link infrastructure, instead of being stored with Module.Fn. This
moves towards a strategy to make more efficient use of memory by not
storing Air or Liveness data in the Fn struct, but computing it on
demand, immediately sending it to the backend, and then immediately
freeing it.
Backends which want to defer codegen until flush() such as SPIR-V
must move the Air/Liveness data upon `updateFunc` being called and keep
track of that data in the backend implementation itself.
This commit changes the AIR file and the documentation of the memory
layout. The actual work of modifying the surrounding code (in Sema and
codegen) is not yet done.
* Inferred error sets are stored in the return Type of the function,
owned by the Module.Fn. So it cleans up that memory in deinit().
* Sema: update the inferred error set in zirRetErrValue
- Update relevant code in wrapErrorUnion
* C backend: improve some some instructions to take advantage of
liveness analysis to avoid being emitted when unused.
* C backend: when an error union has a payload type with no runtime
bits, emit the error union as the same type as the error set.
* ZIR: add two instructions:
- ret_err_value_code
- ret_err_value
* AstGen: add countDefers and utilize it to emit more efficient ZIR for
return expressions in the presence of defers.
* AstGen: implement |err| payloads for `errdefer` syntax.
- There is not an "unused capture" error for it yet.
* AstGen: `return error.Foo` syntax gets a hot path in return
expressions, using the new ZIR instructions. This also is part of
implementing inferred error sets, since we need to tell Sema to add
an error value to the inferred error set before it gets coerced.
* Sema: implement `@setCold`.
- Implement `@setCold` support for C backend.
* `@panic` and regular safety panics such as `unreachable` now properly
invoke `std.builtin.panic`.
* C backend: improve pointer and function value rendering.
* C linker: fix redundant typedefs.
* Add Type.error_set_inferred.
* Fix Value.format for enum_literal, enum_field_index, bytes.
* Remove the C backend test that checks for identical text
I measured a 14% reduction in Total ZIR Bytes from master branch
for std/os.zig.
It now displays the byte with proper printability handling. This makes
the relevant compile error test case no longer a regression in quality
from stage1 to stage2.
In order to not regress the quality of compile errors, some improvements
had to be made.
* std.zig.parseCharLiteral is improved to return more detailed parse
failure information.
* tokenizer is improved to handle null bytes in the middle of strings,
character literals, and line comments.
* validating how many unicode escape digits in string literals is moved
to std.zig.parseStringLiteral rather than handled in the tokenizer.
* when a tokenizer error occurs, if the reported token is the 'invalid'
tag, an error note is added to point to the invalid byte location.
Further improvements would be:
- Mention the expected set of allowed bytes at this location.
- Display the invalid byte (if printable, print it, otherwise
escape-print it).
There is now a distinction between `@import` with a .zig extension and
without. Without a .zig extension it assumes it is a package name, and
returns error.PackageNotFound if not mapped into the package table.
- hash/eql functions moved into a Context object
- *Context functions pass an explicit context
- *Adapted functions pass specialized keys and contexts
- new getPtr() function returns a pointer to value
- remove functions renamed to fetchRemove
- new remove functions return bool
- removeAssertDiscard deleted, use assert(remove(...)) instead
- Keys and values are stored in separate arrays
- Entry is now {*K, *V}, the new KV is {K, V}
- BufSet/BufMap functions renamed to match other set/map types
- fixed iterating-while-modifying bug in src/link/C.zig
