It used to be done with a library intrinsic which returns the array and an element address pointer.
A `mark_dependence` was used to "connect" the two results.
But this resulted in completely wrong OSSA after inlining.
It just didn't get caught be the verifier because the SIL was obfuscated by address-pointer conversions.
Now we don't use the second result (= the element address) of the intrinsic but generate a correct borrow scope from the first (= array) result. Within that borrow scope we project out the element address with a `ref_tail_addr` instruction.
This relies on knowledge of the internal Array data structures. However those data structures are baked into the ABI since a long time and will not change.
Array literals only need to be finalized, if the array is really allocated.
In case of zero elements, no allocation is done, but the empty-array singleton is used.
"Finalization" means to emit an end_cow_mutation instruction on the array.
As the empty-array singleton is a read-only and shared object, it's not legal to do a end_cow_mutation on it.
VarargExpansionExpr shows up in call argument lists in synthesized
initializers and modify accessors when we need to forward arguments
to a call taking varargs.
Previously we would say that the type of VarargExpansionExpr is
$T when its subexpression type is [$T]. matchCallArguments() would
then 'collect' the single VarargExpansionExpr into a variadic
argument list with a single element, and build an ArgumentShuffleExpr
for the argument list.
In turn, SILGen would peephole vararg emission of a variadic
argument list with a single entry that happens to be a
VarargExpansionExpr, by returning the subexpression's value,
which happened to be an array of the right element type,
instead of building a new array containing the elements of the
variadic argument list.
This was all too complicated. Instead, let's say that the type of
a VarargExpansionExpr is [$T], except that when it appears in a
TupleExpr, the variadic bit of the corresponding element is set.
Then, matchCallArguments() needs to support a case where both
the parameter and argument list have a matching vararg element.
In this case, instead of collecting multiple arguments into a
single variadic argument list, we treat the variadic argument like
an ordinary parameter, bypassing construction of the
ArgumentShuffleExpr altogether.
Finally, SILGen now needs to be able to emit a VarargExpansionExpr
in ordinary rvalue position, since it now appears as a child of a
TupleExpr; it can do this by simply emitting the sub-expression
to produce an array value.
I needed this for materializeForSet remission, but it makes inherited
variadic initializers work, too.
I tried to make this a reasonable starting point for a real language
feature. Here's what's still missing:
- syntax
- semantic restrictions to ensure that the expression isn't written in
invalid places or arbitrarily converted
- SILGen support for expansions that aren't the only variadic argument
rdar://16331406
the call instead of during the formal evaluation of the argument.
This is the last major chunk of the semantic changes proposed
in the accessors document. It has two purposes, both related
to the fact that it shortens the duration of the formal access.
First, the change isolates later evaluations (as long as they
precede the call) from the formal access, preventing them from
spuriously seeing unspecified behavior. For example::
foo(&array[0], bar(array))
Here the value passed to bar is a proper copy of 'array',
and if bar() decides to stash it aside, any modifications
to 'array[0]' made by foo() will not spontaneously appear
in the copy. (In contrast, if something caused a copy of
'array' during foo()'s execution, that copy would violate
our formal access rules and would therefore be allowed to
have an arbitrary value at index 0.)
Second, when a mutating access uses a pinning addressor, the
change limits the amount of arbitrary code that falls between
the pin and unpin. For example::
array[0] += countNodes(subtree)
Previously, we would begin the access to array[0] before the
call to countNodes(). To eliminate the pin and unpin, the
optimizer would have needed to prove that countNodes didn't
access the same array. With this change, the call is evaluated
first, and the access instead begins immediately before the call
to +=. Since that operator is easily inlined, it becomes
straightforward to eliminate the pin/unpin.
A number of other changes got bundled up with this in ways that
are hard to tease apart. In particular:
- RValueSource is now ArgumentSource and can now store LValues.
- It is now illegal to use emitRValue to emit an l-value.
- Call argument emission is now smart enough to emit tuple
shuffles itself, applying abstraction patterns in reverse
through the shuffle. It also evaluates varargs elements
directly into the array.
- AllowPlusZero has been split in two. AllowImmediatePlusZero
is useful when you are going to immediately consume the value;
this is good enough to avoid copies/retains when reading a 'var'.
AllowGuaranteedPlusZero is useful when you need a stronger
guarantee, e.g. when arbitrary code might intervene between
evaluation and use; it's still good enough to avoid copies
from a 'let'. The upshot is that we're now a lot smarter
about generally avoiding retains on lets, but we've also
gotten properly paranoid about calling non-mutating methods
on vars.
(Note that you can't necessarily avoid a copy when passing
something in a var to an @in_guaranteed parameter! You
first have to prove that nothing can assign to the var during
the call. That should be easy as long as the var hasn't
escaped, but that does need to be proven first, so we can't
do it in SILGen.)
Swift SVN r24709
