This module serves the purpose of generating debug symbols. We use LLVM's source level debugging features for generating the debug information. The general principle is this:
Given the right metadata in the LLVM IR, the LLVM code generator is able to create DWARF debug symbols for the given code. The metadata is structured much like DWARF debugging information entries (DIE), representing type information such as datatype layout, function signatures, block layout, variable location and scope information, etc. It is the purpose of this module to generate correct metadata and insert it into the LLVM IR.
As the exact format of metadata trees may change between different LLVM versions, we now use LLVM DIBuilder to create metadata where possible. This will hopefully ease the adaption of this module to future LLVM versions.
The public API of the module is a set of functions that will insert the
correct metadata into the LLVM IR when called with the right parameters.
The module is thus driven from an outside client with functions like
debuginfo::create_local_var_metadata(bx: block, local: &ast::local).
Internally the module will try to reuse already created metadata by utilizing a cache. The way to get a shared metadata node when needed is thus to just call the corresponding function in this module:
let file_metadata = file_metadata(cx, file);
The function will take care of probing the cache for an existing node for that exact file path.
All private state used by the module is stored within either the CrateDebugContext struct (owned by the CodegenCx) or the FunctionDebugContext (owned by the FunctionCx).
This file consists of three conceptual sections:
- The public interface of the module
- Module-internal metadata creation functions
- Minor utility functions
Some kinds of types, such as structs and enums can be recursive. That means that the type definition of some type X refers to some other type which in turn (transitively) refers to X. This introduces cycles into the type referral graph. A naive algorithm doing an on-demand, depth-first traversal of this graph when describing types, can get trapped in an endless loop when it reaches such a cycle.
For example, the following simple type for a singly-linked list...
struct List {
value: i32,
tail: Option<Box<List>>,
}
will generate the following callstack with a naive DFS algorithm:
describe(t = List)
describe(t = i32)
describe(t = Option<Box<List>>)
describe(t = Box<List>)
describe(t = List) // at the beginning again...
...
To break cycles like these, we use "forward declarations". That is, when the algorithm encounters a possibly recursive type (any struct or enum), it immediately creates a type description node and inserts it into the cache before describing the members of the type. This type description is just a stub (as type members are not described and added to it yet) but it allows the algorithm to already refer to the type. After the stub is inserted into the cache, the algorithm continues as before. If it now encounters a recursive reference, it will hit the cache and does not try to describe the type anew.
This behavior is encapsulated in the 'RecursiveTypeDescription' enum, which represents a kind of continuation, storing all state needed to continue traversal at the type members after the type has been registered with the cache. (This implementation approach might be a tad over- engineered and may change in the future)
In addition to data type descriptions the debugging information must also allow to map machine code locations back to source code locations in order to be useful. This functionality is also handled in this module. The following functions allow to control source mappings:
set_source_location()clear_source_location()start_emitting_source_locations()
set_source_location() allows to set the current source location. All IR
instructions created after a call to this function will be linked to the
given source location, until another location is specified with
set_source_location() or the source location is cleared with
clear_source_location(). In the later case, subsequent IR instruction
will not be linked to any source location. As you can see, this is a
stateful API (mimicking the one in LLVM), so be careful with source
locations set by previous calls. It's probably best to not rely on any
specific state being present at a given point in code.
One topic that deserves some extra attention is function prologues. At
the beginning of a function's machine code there are typically a few
instructions for loading argument values into allocas and checking if
there's enough stack space for the function to execute. This prologue is
not visible in the source code and LLVM puts a special PROLOGUE END marker
into the line table at the first non-prologue instruction of the function.
In order to find out where the prologue ends, LLVM looks for the first
instruction in the function body that is linked to a source location. So,
when generating prologue instructions we have to make sure that we don't
emit source location information until the 'real' function body begins. For
this reason, source location emission is disabled by default for any new
function being codegened and is only activated after a call to the third
function from the list above, start_emitting_source_locations(). This
function should be called right before regularly starting to codegen the
top-level block of the given function.
There is one exception to the above rule: llvm.dbg.declare instruction
must be linked to the source location of the variable being declared. For
function parameters these llvm.dbg.declare instructions typically occur
in the middle of the prologue, however, they are ignored by LLVM's prologue
detection. The create_argument_metadata() and related functions take care
of linking the llvm.dbg.declare instructions to the correct source
locations even while source location emission is still disabled, so there
is no need to do anything special with source location handling here.
In order for link-time optimization to work properly, LLVM needs a unique
type identifier that tells it across compilation units which types are the
same as others. This type identifier is created by
TypeMap::get_unique_type_id_of_type() using the following algorithm:
-
Primitive types have their name as ID
-
Structs, enums and traits have a multipart identifier
-
The first part is the SVH (strict version hash) of the crate they were originally defined in
-
The second part is the ast::NodeId of the definition in their original crate
-
The final part is a concatenation of the type IDs of their concrete type arguments if they are generic types.
-
Tuple-, pointer-, and function types are structurally identified, which means that they are equivalent if their component types are equivalent (i.e.,
(i32, i32)is the same regardless in which crate it is used).
This algorithm also provides a stable ID for types that are defined in one
crate but instantiated from metadata within another crate. We just have to
take care to always map crate and NodeIds back to the original crate
context.
As a side-effect these unique type IDs also help to solve a problem arising
from lifetime parameters. Since lifetime parameters are completely omitted
in debuginfo, more than one Ty instance may map to the same debuginfo
type metadata, that is, some struct Struct<'a> may have N instantiations
with different concrete substitutions for 'a, and thus there will be N
Ty instances for the type Struct<'a> even though it is not generic
otherwise. Unfortunately this means that we cannot use ty::type_id() as
cheap identifier for type metadata -- we have done this in the past, but it
led to unnecessary metadata duplication in the best case and LLVM
assertions in the worst. However, the unique type ID as described above
can be used as identifier. Since it is comparatively expensive to
construct, though, ty::type_id() is still used additionally as an
optimization for cases where the exact same type has been seen before
(which is most of the time).