This commit is mostly refactoring.
*) Introduce a new OptimizationMode enum and use that in SILOptions and IRGenOptions
*) Allow the optimization mode also be specified for specific SILFunctions. This is not used in this commit yet and thus still a NFC.
Also, fixes a minor bug: we didn’t run mandatory IRGen passes for functions with @_semantics("optimize.sil.never")
Switch most general endpoint to be `flags, parameters, parameterFlags, result`,
instead of opaque `void **`, more specialized ones to use follow argument scheme:
`flags, param0, [flags0], ..., paramN, [flagsN], result` and store parameter/flags
information separately in `FunctionCacheEntry::{Key, Data}` as well.
Currently only single 'inout' flag has been encoded into function
metadata, these changes extend function metadata to support up to
32 flags per parameter.
Switch most general endpoint to be `flags, parameters, parameterFlags, result`,
instead of opaque `void **`, more specialized ones to use follow argument scheme:
`flags, param0, [flags0], ..., paramN, [flagsN], result` and store parameter/flags
information separately in `FunctionCacheEntry::{Key, Data}` as well.
Currently only single 'inout' flag has been encoded into function
metadata, these changes extend function metadata to support up to
32 flags per parameter.
This is a small code size win, and also gives us some abstraction so that future cooperative ObjC compilers/runtimes might be able to interoperate ObjC class objects with Swift type metadata efficiently than they currently are in the fragile Swift runtime.
While I'm here, I also noticed that swift_getObjCClassMetadata was unnecessarily getting exposed in non-ObjC-interop runtime builds, so I fixed that as well.
This changes layout of the parameter metadata from single tuple record
(in case of materializable type) to N records each corresponding to
invididual function parameter, where functions with no parameters
`() -> Void` get 0 records allocated.
Alter the value metadata layout to use an absolute pointer for the
nominal type descriptor rather than a relative offset relative to the
complete type metadata. Although this is slightly less efficient in
terms of load times, this is more portable across different
environments. For example, PE/COFF does not provide a cross-section
relative offset relocation. Other platform ports are unable to provide
a 64-bit relative offset encoding.
Given that the value witness table reference in the value metadata is
currently an absolute pointer, this page is most likely going to be
dirtied by the loader.
Finally, remove the parent type metadata argument from type
constructors.
Now that type constructors don't take a parent metadata pointer,
we can hit some asserts concerning type constructors that do not
have any parameters. This happens when you define a concrete type
in a fully-constrained extension of a generic type.
A more efficient ABI would use concrete type metadata for these
cases, but that would be a bigger change that we can do later, so
for now just relax these assertions.
This resolves a runtime crasher since a circular metadata case is
no longer circular. I renamed the crasher to reference the more
specific radar since the more general issue of circular metadata
is still unresolved.
We no longer need this for anything, so remove it from metadata
altogether. This simplifies logic for emitting type metadata and
makes type metadata smaller.
We still pass the parent metadata pointer to type constructors;
removing that is a separate change.
This change makes it so that type metadata stores arguments
from each nesting depth. This means that parent metadata is
no longer needed to fulfill generic arguments, but we still
store a parent pointer in metadata and pass it to type
constructor functions at this point.
Once generic type metadata includes arguments from all outer contexts,
we need to know how many arguments there are at each nesting depth in
order to properly reconstruct the type name from metadata.
We were inlining the size and alignment, which was not correct.
This was the cause of a long-standing ASAN failure in some
library evolution tests.
Fixes <rdar://problem/24540778>.
- Always include an array of requirement descriptors in the protocol
descriptor. For now, this doesn't contain anything except a general
requirement kind and an optional default implementation, but eventually
this can be augmented with type / name metadata. This array is always
emitted as constant.
- Guarantee the value of the extent fields in a protocol descriptor and
slightly tweak their meaning.
- Move the private-data field out of line in a generic witness table
descriptor so that the main descriptor can be emitted as constant.
- Rely on IRGen's notion of the witness-table layout instead of assuming
that SILWitnessTable and SILDefaultWitnessTable match the actual
physical layout.
This version of the patch uses a hack in which we assign internal rather
than private linkage to certain symbols in order to work around a Darwin
linker bug.
Previous patches changed the runtime to copy the vtable from the
superclass rather than IRGen emitting it statically for generic
and resilient classes.
However for generic classes we would still copy the vtable entries
for methods defined in the immediate class from the template.
Instead, store them in the nominal type descriptor, where they use
less space since we can use relative pointers, and copy them out of
there.
This will allow us to 'slim down' generic class templates eventually.
Instead of emitting the vtable statically, copy it from the
superclass, and fill in method overrides, if any.
For now this is only done for classes initialized by
swift_initClassMetadata_UniversalStrategy(); I need to
add a new entry point for classes with static field
layout but a dynamic vtable.
The emitInitializeFieldOffsetVector() function used for
class metadata is now almost identical to the logic used
for structs. Refactor the code used for classes and also
use it to layout structs.
Both swift_init{Struct,Class}Metadata_UniversalStrategy() wish to
avoid instantiating type metadata for field types if possible.
The struct version took an array of the more general TypeLayout objects,
whereas the class version was implemented earlier and took an array
of size/alignment pairs.
Since we can emit static TypeLayouts for all fixed-size types,
the class version can use the more general TypeLayout type also.
A single-field struct with an @_alignment attribute does not
have the same layout as its field.
This caused the runtime to compute incorrect layouts for
generic structs containing SIMD types.
Fixes <rdar://problem/33794343>.
