Optimize (the very inefficient) RawRepresentable comparison function call to a simple compare of enum tags.
For example,
```
enum E: String {
case a, b, c
}
```
is compared by getting the raw values of both operands and doing a string compare.
This peephole optimizations replaces the call to such a comparison function with a direct compare of the enum tags, which boils down to a single integer comparison instruction.
rdar://151788987
Introduce a new pass MandatoryTempRValueElimination, which works as the original TempRValueElimination, except that it does not remove any alloc_stack instruction which are associated with source variables.
Running this pass at Onone helps to reduce copies of large structs, e.g. InlineArrays or structs containing InlineArrays.
Copying large structs can be a performance problem, even at Onone.
rdar://151629149
This is a workaround for a bug in the move-only checker: rdar://151841926.
The move-only checker sometimes inserts destroy_addr within read-only static access scopes.
Therefore don't consider static access scopes as immutable scopes.
* Move the mutating APIs into Context.swift, because SIL can only be mutated through a MutatingContext
* move the `baseOperand` and `base` properties from the instruction classes to the `MarkDependenceInstruction` protocol
* add `valueOrAddressOperand` and `valueOrAddress` in the `MarkDependenceInstruction` protocol
This prevents simplification and SILCombine passes to remove (alive) `mark_dependence_addr`.
The instruction is conceptually equivalent to
```
%v = load %addr
%d = mark_dependence %v on %base
store %d to %addr
```
Therefore the address operand has to be defined as writing to the address.
Instead of looking for a single store to the global in a global-init function, build a GlobalInitValue tree.
This is a data structure representing the init value of the global. It can handle complex InlineArray initializations,
like `init(repeating:)`.
rdar://150859232
It is like `zeroInitializer`, but does not actually initialize the memory.
It only indicates to mandatory passes that the memory is going to be initialized.
Diagnostics only work with `SourceLoc` which is basically a pointer into a buffer of the loaded source file.
But when debug info is de-serialized, the SIL `Location` consists of a filename+line+column.
To "convert" this to a `SourceLoc`, the file must be loaded.
This change adds `DiagnosticEngine.getLocationFromExternalSource` for this purpose.
Also, the new protocol `ProvidingSourceLocation` - to which `SourceLoc` and `Location` conform - help to generalize the helper struct `Diagnostic` and make this "conversion" happen automatically.
Treat `load`-`store` pairs as if they were `copy_addr` instructions.
This can catch more cases. For example it can eliminate unnecessary copies of InlineArray when subscripting it.
rdar://149250346
Add a boolean parameter `salvageDebugInfo` to `Context.erase(instruction:)`.
Sometimes it needs to be turned off because the caller might require that after erasing the original instruction the operands no users anymore.
Reimplement the simplification in swift and add a new transformation:
```
%1 = unchecked_addr_cast %0 : $*Builtin.FixedArray<N, Element> to $*Element
```
->
```
%1 = vector_base_addr %0 : $*Builtin.FixedArray<N, Element>
```
It derives the address of the first element of a vector, i.e. a `Builtin.FixedArray`, from the address of the vector itself.
Addresses of other vector elements can then be derived with `index_addr`.
Beside cleaning up the source code, the motivation for the translation into Swift is to make it easier to improve the pass for some InlineArray specific optimizations (though I'm not sure, yet if we really need those).
Also, the new implementation doesn't contain the optimize-store-into-temp optimization anymore, because this is covered by redundant load elimination.
These APIs are quite convoluted. The checks for var_decl need to be performed in
just the right order. The is a consequence of complexity in the SIL
representation itself, not a problem with the APIs.
It is common for code to accidentally call a less-complete form of the API. It
is essential that they be defined in a central location, and the we get the same
answer whether we start with an Instruction, Argument, or Value. The primary
public interface should always check for debug_value users. The varDecl property
is actually an implementation detail.
It is questionable whether a function like findVarDecl() that returns a basic
property of SIL and does not require arguments should be a property instead. It
is a function to hint that it may scan the use-list, which is not something
we normally want SIL properties to do. Use-lists can grow linearly in function
size. But, again, this is a natural result of the SIL representation and needs
to be considered an implementation detail.
As SourceKit explicitly disables WMO, silence the diagnostic in this case (but leave it enabled for explicit non-WMO builds otherwise).
rdar://150596807
If the block is passed to another function - either as closure argument or as closure capture - it's "converted" to a swift closure with the help of a thunk. The thunk just calls the block.
If this is done with a non-escaping partial_apply, the block does not escape.
rdar://149095630
This utility is used by DependentAddressUseDefWalker which now conservatively
follows all possible uses. This could result in the same address being reached
multiple times during a def-use walk. Ensure that we don't infinitely recurse.
There is no small test case for this, but the fix is trivial and standard
practice for such walkers, and this is hit quickly in real usage, so there is no
danger of it regressing.
Fixes rdar://150403948 ([nonescapable] Infinite recursion compiler crash in
lifetime dependence checking)
Correctly generate dependency tracking for functions that return a non-Escapable
existential, such as:
func getMutableSpanWithOpaqueReturn(_ array: inout [Int]) -> any PAny & ~Copyable & ~Escapable
Previously, dependency insertion assumed that @out storage always initialized an
alloc_stack. But existentials are always boxed.
First, add a diagnostic to catch any missing dependency insertions now that
we're past the bootstrapping phase.
Then, generalize handling of dependency insertion to handle any access base as
long as it has a recognizable address source.
Fixes rdar://150388126 (Missing mark_dependence for opaque lifetime dependent
value)
Array/ArraySlice/ContiguousArray have support for mutableSpan property.
These types are "optimized COW" types, we use compiler builtins begin_cow_mutation/end_cow_mutation to
optimize uniqueness checks. Since mutableSpan is a property and not a coroutine there is
no way to schedule the end_cow_mutaton operation at the end of the access.
This can lead to miscompiles in rare cases where we can end up using a stale storage buffer after a cow.
This PR inserts end_cow_mutation_addr to avoid this issue.
Note: We can end up with unnecessary end_cow_mutation_addr. But it is just a barrier to prevent invalid optimizations
and has no impact.
When a coroutine is extended because of a dependent lifetime, extend all scopes
that enclose that coroutine even if the coroutine itself has no lifetime
dependencies.
Fixes rdar://150275147 (Invalid SIL after lifetime dependence fixup involving
coroutines)
Add a note explaining that dependence on closure captures is not
supported. Otherwise, the diagnostics are very confusing:
"it depends on a closure capture; this is not yet supported"
Ensure that we always issue a diagnostic on error, but avoid emitting any notes that don't have source locations.
With implicit accessors and thunks, report the correct line number and indicate which accessor generates the error.
Always check for debug_value users.
Consistently handle access scopes across diagnostic analysis and diagnostic messages.
Fix a simple typo that results in infinite recursion on invalid code.
Fixes rdar://147470493 ([nonescapable] LifetimeDependenceInsertion: infinite
recursion in VariableUseDefWalker.walkup with immortal setter)
* rename `ScopedInstruction.endOperands` -> `scopeEndingOperands`
* let them behave the same way. For `load_borrow` there was a difference because `endOperands` didn't consider branches to re-borrow phis.
Briefly (April 2025), RawSpan._extracting, Span._extracting, and UTF8Span.span
returned a borrowed value that depended on a copied argument. Continue to
support those interfaces. The implementations were correct but needed an
explicit _overrideLifetime.
