v8/test/mjsunit/regress/regress-v8-9394.js
Leszek Swirski e79751b3c1 [parser] Always mark shadowed vars maybe_assigned
If there was an assignment to a maybe-shadowing dynamic variable,
then the shadowing variable would be marked maybe_assigned, but the
maybe-shadowed variable would stay unchanged. This meant that in
non-shadowing cases, the not-actually-shadowed variable would have
the wrong maybe_assigned state, and e.g. would break context
specialization.

This patch pessimistically unconditionally sets maybe_assigned on
variables shadowed by a dynamic variable in a `with` scope. This
marking can cause false positives and sub-optimal optimization for
some functions with 'with' blocks, but it's also the simplest fix
for this issue which doesn't affect performance in the common case
of no 'with' blocks.

Bug: v8:9394
Change-Id: I6924bd7d48dda61232aa9d72c39df1c76c665c67
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/1678365
Commit-Queue: Leszek Swirski <leszeks@chromium.org>
Auto-Submit: Leszek Swirski <leszeks@chromium.org>
Reviewed-by: Georg Neis <neis@chromium.org>
Cr-Commit-Position: refs/heads/master@{#62407}
2019-06-27 08:26:02 +00:00

84 lines
2.0 KiB
JavaScript

// Copyright 2019 the V8 project authors. All rights reserved.
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file.
//
// Flags: --allow-natives-syntax
(function testMaybeAssignedWithShadowing() {
function foo() {
let a = 0;
let g;
with ({}) {
g = function g() {
// Increment a, should set it maybe_assigned but doesn't in the bug.
++a;
}
// Shadowing the outer 'a' with a dynamic one.
a;
}
return function () {
// The access to a would be context specialized (to 2 since it's after the
// second call) if maybe_assigned were incorrectly not set.
g(a);
return a;
}
};
f = foo();
%PrepareFunctionForOptimization(f);
assertEquals(f(), 1);
assertEquals(f(), 2);
%OptimizeFunctionOnNextCall(f);
assertEquals(f(), 3);
})();
// Same test as above, just with more shadowing (including dynamic->dynamic
// shadowing) and skipping over scopes with shadows.
(function testMaybeAssignedWithDeeplyNestedShadowing() {
function foo() {
let a = 0;
let g;
// Increment a, should set it maybe_assigned but doesn't in the bug.
with ({}) {
with ({}) {
with ({}) {
with ({}) {
with ({}) {
g = function g() { ++a; }
// Shadow the second dynamic 'a'.
a;
}
// Shadowing the first dynamic 'a'.
a;
}
// Skip shadowing here
}
// Skip shadowing here
}
// Shadowing the outer 'a' with a dynamic one.
a;
}
return function () {
// The access to a would be context specialized (to 2 since it's after the
// second call) if maybe_assigned were incorrectly not set.
g(a);
return a;
}
};
f = foo();
%PrepareFunctionForOptimization(f);
assertEquals(f(), 1);
assertEquals(f(), 2);
%OptimizeFunctionOnNextCall(f);
assertEquals(f(), 3);
})();