---
Changes in `@enableOptimizeFunctionExpressions` caused a bug in the last Forget
sync to VR Store. The repro can be summarized to something like this:
```js
function foo() {
const x = true; // some constant or global
// Add some branching for type inference
// This can be a Logical expression as well (e.g. `4 || 5`)
if (...) { }
// In this HIR block, SSA inserts a `x$2 = phi(x$1, x$1)`.
// EliminateRedundantPhiNodes needs to rewrite all references of `x$2` to `x$1`
const accessXInLambda = () => x;
return accessXInLambda;
}
---
Revives e2e test infra from #587.
- All React component-like functions are compiled.
- `yarn jest` runs each e2e test twice (forget and no forget)
Github Actions is already running `yarn test`, which includes all jest tests
```
Run yarn test
yarn run v1.22.19
$ yarn workspaces run test
> babel-plugin-react-forget
$ yarn jest && yarn snap:build && yarn snap
$ tsc && jest
PASS main src/__tests__/Result-test.ts
PASS main src/__tests__/DisjointSet-test.ts
PASS e2e with forget src/__tests__/e2e/hello.e2e.js
PASS e2e no forget src/__tests__/e2e/hello.e2e.js
Test Suites:
[4](https://github.com/facebook/react-forget/actions/runs/5732016200/job/15534129231?pr=1881#step:8:5)
passed, 4 total
Tests: 23 passed, 23 total
Snapshots: 11 passed, 11 total
Time:
6.1[5](https://github.com/facebook/react-forget/actions/runs/5732016200/job/15534129231?pr=1881#step:8:6)3
s
```
The current heuristic to check if setState is called in render is based on
whether the lambda containing the call to setState has a mutable range that got
extended. This doesn't seem to work in all cases so this validation needs a bit
more work before we can turn it on by default
Console warnings when running tests: ``` ts-jest[versions] (WARN) Version 5.1.3
of typescript installed has not been tested with ts-jest. ```
Our typescript version is ahead of what's supported in ts-jest. This PR updates
ts-jest to work with the latest typescript compiler.
This PR adds a Hole kind that can be present in both ArrayExpression and
ArrayPatterns.
This Hole type is not interesting for our inference passes and is skipped over
for all of the pipeline.
Forget doesn't understand the React namespace object and generates incorrect
code when compiling code that loads props from this namespace object.
This PR makes Forget bailout when we see a property load from React namespace
object.
Upgrades every dependency on `hermes-parser` and `prettier` to the versions that
we're using in the rest of our other codebases. Notably, these package versions
are necessary for our other codebases to make use of mapped types in Flow.
Updates eliminate_redundant_phis and constant_propagation to recurse into
function expressions. I also realized there was a bug in EliminateRedundantPhis
in which we wouldn't traverse into function expressions encountered after
finding a back edge, so i fixed that logic in both versions.
Makes the same improvement to constant propagation in both the JS and Rust
versions. The core algorithm only populates phi variables if all operands have a
known value (no back edges) and all those values are the same: this allows us to
propagate constants in most cases and simply punts on handling propagating
values that are affected by loops. However, since we collapse if statements into
gotos when the test condition is a constant, there can be cases where a phi that
originally existed will be pruned out:
```
// bb0
let x1;
if (true) {
// bb1
x2 = 1;
} else {
// bb2
x3 = 2; // this block becomes unreachable
}
// bb3
x4 = phi(bb1: x2, bb2: x3); // this phi will get pruned s.t. x4 = x2 = 1
return x4;
```
However, the algorithm doesn't prune phis until _after_ applying constants, so
currently we would see this phi node w different inputs and not propagate a
constant for the final usage of x, even though it will clearly be `1`.
The change is to make constant propagation use fixpoint iteration, iterating so
long as terminals changed on the previous iteration. If no terminals change the
algorithm completes in a single pass, but if terminals do change then we update
phis and continue. As you can see from the new test case this allows us to find
arbitrary length sequences of values and terminals that can be pruned.
This mode improves compilation and makes optimizations easier, let's make it the
default. I previously confirmed that enabling this mode didn't affect output
when synced internally, and I'll do that again before landing the PR.
I added this as a quick workaround, since we didn't support unused
logical/conditional expression statements. Now that we handle them we don't need
InstructionValue::ExpressionStatement anymore. I found this when porting our
lowering to Rust.
lowerIdentifierForAssignment does extra checks such as checking if globals are
lvalues.
UpdateExpression lowers into an assignment and previously we missed out on such
checks.
- Made most static methods on CompilerError take a single options object as an
argument. With the exception of invariant which takes a condition and an options
object. - Added a new `suggestions` field on CompilerErrorDetail, which we'll
use to provide eslint suggestions - Updated eslint-plugin-react-forget to handle
suggestions
Most of these errors were incorrectly using InvalidInput as a catchall for
rejecting code. I went through each one and manually updated them to be more
accurate
Found this while running Forget on the React tests.
This isn't a high priority because the ESLint plugin would've caught this. But
it'd be nice if either our validation rules caught this or if our compiler did
correctly eliminate the dep array.
This allows us to mimic the `invariant` api, which means you can just assert
that something holds true after execution proceeds to the next line without
throwing
These were effectively unused since we almost always passed in null when
creating errors, so remove them and instead pass in a `loc` explicitly. The
downside is that we no longer will see Babel codeframes in our test fixtures,
which doesn't seem like a huge loss
Fixes the mutable range validation in PrintHIR: when we checked the identifier
scope's range we failed to allow end=start=0 cases which are also valid.
While I was here, I created a separate assertion pass to check all ranges, that
way we aren't relying on the printer to validate them. The pass is on for
playground and tests, but disabled by default so it can't break internal apps.
Distinguishes between two types of "validation" passes:
- Passes which assert the validity of the HIR. These remain in HIR/ but are
renamed "assertFoo".
- Passes which validate that the user code is correct. These move to
Validation/.
Fixes#1751. Function ids can be plain strings, and we can refer to them as
globals rather than via a local identifier. In addition to the bug from $1751
this also cleans up an existing todo.
This is a fix specifically for the `enableOptimizeFunctionExpression` feature
(disabled by default). I tried running all of our fixtures with that flag on
everywhere, and this was the only issue. It's actually extracted from another
fixture which is more complicated, this is a distilled version. There were two
bugs:
* DCE was running after LeaveSSA when it needs to run before. Fixing the order
means code inside the function expression stops getting removed.
* In the feature flag, context variables share Identifier instances with the
surrounding code, whereas they are distinct instances without the feature. This
meant that mutable ranges from inside the function propagated outside the
function, throwing off our inference. The fix was to reset the ranges of context
variables after inferring the function expression's effects.
inference
Integrates the new inference to detect definitive mutations of context
variables. This allows us to more precisely infer mutable function expressions
and validate that they aren't passed where frozen values are expected.
New pass to infer context variables which are definitively mutated,
distinguishing from context variables which _may_ have been mutated. See the
code comment on the new pass for more details.
Small tweak necessary for the subsequent PRs, refs and ref values take
precedence over other mutations when computing the effect type of context
variables. We always want to record ref access within a function as capture
(since we have later validation) rather than a mutation. For now this has no
impact, either order records a Capture. But it allows later diffs to make
non-ref cases have other effects.
The original version of the code wasn't checking return values. I missed this
since my examples were passing functions _into_ the return value, as opposed to
return the functions directly. This revealed some existing test fixtures that
were technically invalid, but easy to fix by changing the return value.
In InferReferenceEffects, locations that are ConditionallyMutate are either
recorded as a Mutate or Read, which means we lose the distinction btw
conditional/unconditional mutation in later passes. This PR changes to remember
that these places were conditionally mutable, used in later analysis.
Defaults to false, ie it runs the codegen pass. When enabled it will simply run
all passes up to codegen and then skip over it. Naming of this option is copied
from [TypeScript](https://www.typescriptlang.org/tsconfig#noEmit) which has the
same named option that makes the compiler only perform typechecking.
I'm adding this option primarily to get around some issues running the eslint
plugin on Meta code. The plugin would error because Forget would report
duplicate Babel AST nodes, which I presume would only occur during codegen. It
should also make it a tiny bit faster to not run codegen, which is a nice plus.
The existing errors thrown were marked as InvalidInput, which is now considered
critical. This was causing an error in the sync since we had 2 occurrences of
the errors being thrown. These are really todos and not invalid code.