Lots of places rely on reference stability of ContainerInfo objects
(CodeGen's deduplication, Container nodes' containerInfo_ member).
In the key capture work, we need to be able to append to this list,
which would invalidate references before this change.
Dummy and DummyAllocator nodes had been changed to use NodeIds, but
were still printed out in full when visited for a second time.
[[nodiscard]] prevents future bugs of this type by turning them into
compilation errors.
Example of the now-fixed bug:
[1] Container: std::map<int32_t, int32_t, DummySizedOperator<0, 0, 8>, std::allocator<std::pair<int32_t const, int32_t>>>
Param
Primitive: int32_t
Param
Primitive: int32_t
Param
[2] Dummy [less<int>]
Param
...
[3] Container: std::map<int32_t, int32_t, DummySizedOperator<0, 0, 8>, std::allocator<std::pair<int32_t const, int32_t>>>
Param
Primitive: int32_t
Param
Primitive: int32_t
Param
[2]
Dummy [less<int>]
Param
...
With this patch, the second "Dummy" line will not be printed.
We only want to do the extra work if it's explicitly requested.
chaseRawPointers is already explicitly requested whenever it's needed
and readEnumValues currently isn't needed at all.
Summary:
Previously on large types OIL would have problems with corrupting the `std::stack<exporter::inst::Inst>` that is passed to the processors. This change hides the implementation of the stack from the processors by wrapping the call to emplace in a `std::function` written by the non-generated code, which solves the test case I've seen for this crashing. It also allows us to easily change the stack implementation in future - I plan to change it to a `std::stack<T, std::vector<T>>` in a follow up.
Reviewed By: tyroguru
Differential Revision: D49273116
Summary:
Update to clang-15 compiler and libraries as clang-12 is ancient.
The changes to oilgen are necessary because the new internal toolchain is being more picky about linking PIC to PIC. In certain modes we build with PIC, but try to link a non-PIC oilgen artifact. Add the ability to build the oilgen artifacts with PIC which sorts this.
Reviewed By: ttreyer
Differential Revision: D46220858
Summary:
The iterator was incremented without checking it in the JSON exporter. This caused an assertion to trigger on the last run in debug mode (weirdly no crashes). This change should fix that by checking the iterator at the increment site and not just when the loop rolls around.
Differential Revision: D49151482
Summary:
Update `OIGenerator` and out BUCK stuff for compile time OIL with OIL v2. Main things:
- Switch `OIGenerator` from the `getObjectSize` call to the new `introspect` call.
- Switch from looking at template parameters to looking at function parameters, as this was exposing a bug in our elfutils/drgn and this way it's the same as OID.
- Migrate `OIGenerator` to CodeGen v2 and update CodeGen v2 to accept a linkage name.
- Update the compile time example to be the same as the JIT example, using the new interface and the JSON exporter.
- Clean up the `ObjectIntrospection.h` header.
Differential Revision: D48687728
fbshipit-source-id: 2c3c041fd1b6499c5e02eb5e2082a977bfa529d7
Types within containers were previously named TODO. This sorts it out so
they're named as their most resolved type. The current implementation
skips Typedef names.
The TypeGraph class should only be responsible for storing Type nodes.
Traversing the graph and tracking which nodes have been visited should
not be included there.
Passes now take a NodeTrackerHolder as an input parameter, which
provides access to a zeroed-out NodeTracker.
Type Graph deduplicates and modifies names to better fit the generated
code, for example `int32_t[4]` becomes `OIArray<int32_t, 4>` and `struct
MyStruct` might become `struct MyStruct_0`.
Add an `inputName` which better represents the original input code which
can be used when building the tree.
This removes Printer's legacy behaviour of generating an ID for each
node as it gets printed. This old method meant that if new nodes were
added to or removed from a graph, every ID after the new/removed node
would change.
Now IDs are stable so it is easier to follow specific nodes through
multiple transformation passes in CodeGen.
Names which were generated on-demand are now stored in member variables,
which are set during the ctor and can be regenerated when required (by
NameGen).
We previously only marked as packed if there was no tail padding, which
was not a sufficient condition.
The new AlignmentCalcTest.PackedMembers test case is an example which
would previously not have been marked as packed.
CodeGen v1 does not record anything for pointers to incomplete types.
Not even the address, as is done for other pointers.
Introduce a new Primitive type "Incomplete". This behaves identically to
"Void", but allows us to tell whether a type was defined as void or if
it ended up like that because of incomplete DWARF information.
This extracts the compatibility logic from AddPadding, which allows for it to be
simplified and will make it easier to extend and eventually remove in the
future. No functional changes.
This lets us remove fields from types when they are no longer needed,
speeding up later passes.
A secondary benefit of pruning unused types means that we sometimes
remove types for which we can't generate correct C++ code. This can
allow us to CodeGen for complex types which reference these broken types
without actually requiring them (e.g. as template parameters).
Add a new feature flag "prune-type-graph" to control this pass. It makes
sense to prune most of the time, but for testing CodeGen functionality
on a wider range of types, it will be useful to have the option to not
prune.
In general, we can't tell which member is active in a union so it is not
safe to try and measure any of them.
Explicitly set the alignment of unions (and structs/classes) in CodeGen
if it is available, as the C++ compiler can no longer infer it from the
members.
Also reshuffle CodeGen's passes to fix an alignment bug with removed
members.
Change RemoveMembers to actually remove members instead of replacing
them with padding. AddPadding must be run afterwards to fill in the
gaps.
The underlying type of bitfield is important to the size of a struct:
struct Foo { int64_t bitfield : 1; };
struct Bar { int8_t bitfield : 1; };
sizeof(Foo) = 8;
sizeof(Bar) = 1;
Previously this code would not have removed all members which it was
supposed to.
Also remove some now-redundant code from TypeIdentifier. RemoveIgnored
will take over the responsibility of removing members from classes.
TypeGraphParser parses a textual type graph, as emitted by Printer.
It also doubles as a way of ensuring that Printer displays all
information about a type graph, to aid with debugging.
Convert Flattener unit tests over to this new framework as a first step.
Summary:
Stubbing tuple without alignment info can cause failures. I added a test
case for this which failed before this fix but works now.
Test Plan:
Ran the test.
Summary:
If a container has 0 template params (e.g. IOBuf, IOBufQueue), it wasn't
being added to containerTypeMap. This causes the deserialization to go wrong
as TreeBuilder doesn't treat the types as container
Test Plan:
make test
Use fully qualified names to determine if a class is really the child of
our type. It may be that it is the child of another type with an
identical name in another namespace.
visit(Type&) and visit(Type*) were helper functions than did cycle
detection and provided nicer syntax for the type.accept(*this) calls.
However, because they overloaded the visit() function, it was easy to
accidentally call a concrete visit method (e.g. visit(Class&)) instead
of the virtual-dispatching visit(Type&).
By changing the name of this wrapper, it will make it much more obvious
when code is introduced which bypasses the cycle detection.
Pass-through-types represent classes to be turned into containers. We
don't want these to turn these containers into Dummy's on a second run
of TypeIdentifier.
This will eventually be used to enable running with Tree Builder v2.
For now, when it is disabled it puts CodeGen v2 into compatibility mode,
disabling features which weren't present in CodeGen v1 so that its
output can be understood by Tree Builder v1.
These aren't used for anything yet, but should be useful for stable IDs
when printing nodes before and after passes and for faster cycle
detection than the current map of pointers.
As we now store ContainerInfo objects in OICodeGen::Config, we can not
copy it any more. Change all places that took copies to take const
references instead.
The copy in OICodeGen modified membersToStub, the contents of which form
part of OICache's hash. However, as OICache also previously had its own
copy, it would not have been OICodeGen's modifications.
- Change member and parent offsets to work in bits, not bytes
- Printer still displays offsets in bytes, with decimals when using
bitfields
- AddPadding: Don't pad bitfields
- CodeGen: Emit code for bitfields
For std::vector and std::list, template parameters are not required to
be defined before they can be used. Delay sorting them until the end.
Also fix a CodeGen bug where we were defining typedefs in the middle of
the forward declarations. They only need to be defined when other types
are defined.
