The previous PR only updated simple assignment expressions (where the lvalue is
an identifier), this PR extends the same idea to all assignment variants. Note
that there is one case that doesn't work yet, which is complex destructuring
assignment as a value:
```javascript
let x = makeObject();
x.foo(([[x]] = makeObject()));
```
What happens here is that we lower the destructuring to a series of steps:
```
tmp1: Destructure Const [ tmp0 ] = makeObject();
tmp2: Destructure Reassign [ x ] = tmp0;
PropertyCall x, 'foo', [ tmp1 ]
```
Thankfully we can detect this case: if we have a const/let declaration with an
lvalue, that's invalid. See the new error test case which shows we correctly
detect & reject this case for now.
This PR subtly changes how we represent assignment expressions in order to
accurately model them _as expressions_. Specifically, the result of lowering an
assignment is now the temporary created for the assignment's lvalue. This allows
us to restore the assignment as a value (expression) during codegen. Note how
this fixes a bug and cleans up some output.
Updates ConstantPropagation so that each instruction is responsible for whether
to replace its `.value` with the resolved constant value (if found).
Specifically, for `StoreLocal` we don't want to replace the value — we want to
keep the assignment — but we do want to propagate the _result_ of the assignment
downstream. This more accurately models the semantics of assignment expressions,
and helps with subsequent PRs.
Rewrite ArrowFunctionExpression to FunctionDeclaration and compile it. This lets
us reuse all the export gating logic, rather than writing separate, specific
logic for ArrowFunctionExpression.
Cleans up duplicated code for processing call/constructor arguments. As a side
benefit, we now support spread elements for constructor arguments (and if we
want to change how we represent that, we can do it in one place).
We currently are not lowering property calls in evaluation order.
I wonder if the following lowering for a PropertyCall (with static or computed
property) is semantically equivalent to source:
1. eval + resolve receiver (store in t0)
2. eval computed property (if present)
3. resolve t0.property (binding the call to receiver and storing in t1)
4. eval args
5. eval t1(args)
Although codegen might then generate something like this (if args are named
temporaries)
```js
const tmp = receiver.property.bind(receiver);
// lower args to temporaries
tmp(arg1, arg2);
```
Adds a `DeclareLocal` instruction which represents declaring a named variable
without initializing it. Currently declarations without an initializer (`let x`)
are transformed into a declaration to undefined (`let x = undefined`) which
changes the semantics due to hoisting and TDZ (temporary dead zone). The correct
thing is to represent declaration without initialization.
These examples previously errored all the way in codegen, when we detected that
a value block (eg a `while` test expression) was declaring a new variable. We
now detect this in LeaveSSA and error. The actual fix is a bit tricky, we'd need
to add a new declaration in the nearest block scope (or selectively not DCE the
declaration if its reassigned in just this way).
If a logical or conditional expression is unused, then a phi node isn't created
for the identifier it assigns to. Then when we leave SSA form the two branches
will assign to separate values, and we aren't sure which identifier to use as
the lvalue of the resulting ReactiveInstruction (remember that
logicals/conditionals decompose into control flow in HIR, but are a single
compound instruction in ReactiveFunction). If the two sides don't assign to the
same location, it could be because of a bug in the compiler or because the value
wasn't used. Ideally we'd represent this explicitly, but for now i'm just making
this a TODO since most logicals/conditionals should have their value used.
Enables support for assignment expressions in value blocks (which includes in
loop init/test/update blocks). This was pretty straightforward, the main changes
are:
* During PropagateScopeDependencies, we currently record scope reassignments
based on `Identifier` object identity. In the case where a variable is
reassigned in multiple control-flow paths of a value block, however, there can
be multiple object identities. So we now de-dupe reassignments based on
identifier id.
* MergeOverlappingScopes now treats value blocks as regular blocks, allowing it
to correctly merge scopes from the value with other scopes from the outer block.
Otherwise this is mostly just lots of tests. Note that there is an outstanding
todo, which is that we currently error for ternaries and logicals whose value is
unused (eg `cond ? (x = 1) : null`). I'll address that in a follow-up.
This PR clarifies the logic for adjust mutable ranges of phis and their operands
during LeaveSSA. Previously we had logic in several places to determine
whether/how to extend the ranges of each phi and its operands: this occurred
while traversing reassignmentPhis (in 2+ places) and rewritePhis, as well as in
rewritePlace().
This was kind of a band-aid to make things work, but the logic was imprecise.
The actual rules are as follows:
If there is a back-edge, or the phi id is unnamed, then were extend the ranges
of the phi and its operands to min(starts) and max(ends). This ensures that the
operands are computed as one unit, ie put into a single reactive scope. For
loops this is necessary because...looping! For unnamed values this is necessary
because of the way we collapse logical and ternary expressions back to a
hierarchical ReactiveFunction — we need to make sure the final mutable range
extends from the start of the final instruction up to the end of the
logical/ternaries value blocks.
Otherwise this is a phi where operands come from predecessors and are named. If
the phi is mutated later, then we have to extend the end of each operand's range
to account for the fact that they can be mutated later. Else, we leave the
operands alone.
Behavior doesn't change, but we consolidate all of the mutable range logic in
one place.
Makes JSX memoized by default again, but adds an option to disable memoization
of JSX. Also adds a new test and fixtures directory to test the opt-in
no-jsx-memoization behavior.
---
Currently, we run type inference passes early in the pipeline and do not check
inference output in any tests, test fixtures, or verifier passes. In fact, the
only ways to view inferred types are (1) locally add a test fixture with`@only`
and inspect console logs or (2) scroll to the relevant section on a playground
example.
However, inferred types and effects significantly affect the output of later
passes (Alias / MutableRange analysis, InferReactiveIdentifiers, etc), and we
have already found some bugs due to incorrect inference (e.g. #1274).
This PR add the `typer-tests` fixture with the following goals
1. Record relevant current compiler type + effect inference output.
2. Have relatively stable output (with respect to changes in HIR and PrintHIR).
- we try to achieve this by annotating the source code.
---
Currently, we run type inference passes early in the pipeline and do not check
inference output in any tests, test fixtures, or verifier passes. In fact, the
only ways to view inferred types are (1) locally add a test fixture with`@only`
and inspect console logs or (2) scroll to the relevant section on a playground
example.
However, inferred types and effects significantly affect the output of later
passes (Alias / MutableRange analysis, InferReactiveIdentifiers, etc), and we
have already found some bugs due to incorrect inference (e.g. #1274).
This PR add the `typer-tests` fixture with the following goals
1. Record relevant current compiler type + effect inference output.
2. Have relatively stable output (with respect to changes in HIR and PrintHIR).
- we try to achieve this by annotating the source code.
We were treating Destructuring as if it could never allocate and therefore
didn't have to be memoized. That's only true if there are no rest spreads
though. This PR teaches the compiler to treat rest spreads differently for
scoping and memoization
purposes, fixing the newly added test case and some existing bugs.
Adds a new pass that uses escape analysis and React-specific heuristics to tune
the amount of memoization applied. Specifically, the pass ensures that we only
memoize:
* Values which escape (are directly returned or transitively aliased by a
returned value)
* ...and that are not JSX elements
* OR values which are _dependencies_ of scopes that produce an escaping value.
The latter case is necessary to avoid breaking memoization of an escaping value
bc a scope happened to have a non-escaping dependency.
## Algorithm
1. First we build up a graph, a mapping of IdentifierId to a node describing all
the scopes and inputs involved in creating that identifier. Individual nodes are
marked as definitely aliased, conditionally aliased, or unaliased:
a. Arrays, objects, function calls all produce a new value and are always marked
as aliased
b. Conditional and logical expressions (and a few others) are conditinally
aliased, depending on whether their result value is aliased.
c. JSX is always unaliased (though its props children may be)
2. The same pass which builds the graph also stores the set of returned
identifiers
3. We traverse the graph starting from the returned identifiers and mark
reachable dependencies as escaping, based on the combination of the parent
node's type and its children (eg a conditional node with an aliased dep promotes
to aliased).
4. Finally we prune scopes whose outputs weren't marked.
This was the actual bug. When EliminateRedundantPhis eliminates a phi, it has to
rewrite downstream usages of the phi id to the single operand id. We were
correctly doing that in all but one place. When we iterate _downstream phis_, we
were looking up the operands against the rewrite table, but not updating the phi
operands themselves to the rewritten value.
This fixes the bug, and incidentally fixes a test that has been broken for a
while and nagging at me.
Found while debugging the previous issue: `mapInstructionOperands()` should not
look at lvalues. The previous version was causing us to create extra phi nodes,
which interestingly weren't the actual problem behind the "SSA" bug, but sure
looked like it at first.
Refactors the representation of ObjectExpression properties from a Map to an
`Array<ObjectProperty>` to prepare for the next diff which adds spread element
support.