Recognize larger regions of no-op control flow. Ideally we would be able to remove these regions from the CFG. Until then, pattern matching larger regions leaves fewer ghost conditions in the generated code.
With pattern matching, we are seeing more examples where this is necessary.
Bug: #29475
Change-Id: Ib0981400883631f56940288f8a32fbc5f3985dfb
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/340067
Reviewed-by: Sigmund Cherem <sigmund@google.com>
Commit-Queue: Stephen Adams <sra@google.com>
Treat the loop update marker more like an ordinary symbolic value. The old symbolic marker attempted to do widening 'on the fly', which could lead to incorrect results when the update could reset the value to a constant lower than the starting value. The new version moves the widening to a single place, at the loop update.
Fix#53355 by caching intermediate results so that long chains of diamond control flow are not explored exponentially.
There are very few changes in apps. There is one change in a Flutter app that is like the changed codegen/value_range_test where the bounds check can be eliminated because the loop index may be decremented, but not more than the increment, so is still weakly monotonic. There is one change in a large ACX app where a lower bounds check is no longer removed but I *think* it was previously removed incorrectly, though it is hard to tell since it is in a huge function.
Issue: #48465
Issue: #53078
Issue: #53355
Change-Id: Ib125cd6bb30cef52f8dfcd53eaa13e439f26316c
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/322594
Reviewed-by: Sigmund Cherem <sigmund@google.com>
Commit-Queue: Stephen Adams <sra@google.com>
The fix is to enable type inference.
The old `List(n)` constructor was replaced by `List.filled`. `List(n)`
could be recognized as fixed-length purely by syntax. The test failed
after the replacement because `List.filled` is not recognized to be
fixed-length purely syntactically in the SSA builder, but rather by
type inference.
With type inference enabled, `List.filled` is recognized, but other
parts of the test need to be more realistic, as type inference also
detects the element type and length.
Bug: #52139
Change-Id: Icfaeb3e341b448bbc263efe43f48a2b76d05bf5b
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/298020
Reviewed-by: Nate Biggs <natebiggs@google.com>
Commit-Queue: Stephen Adams <sra@google.com>
This CL replaces https://dart-review.googlesource.com/c/sdk/+/296900
The `List` constructor is removed in Dart 3.0.
Some of the `@patch` implementations were not removed.
This is *high priority*. It seems the left-over `@patch factory List` constructor did not cause any errors, instead it *added* a constructor to `List` that can be used in web compiled code. Even if `List` doesn't have such a constructor in the SDK code proper.
The VM and analyzer will say the invocation is an error, but dart2js happily compiles it and runs.
(It used to be that patches couldn't add public members, that security seems to have been removed.)
Also removes code which tries to detect "the unnamed List constructor",
which is no longer a thing, and a number of invocations of the constructor, where it's not clear that the test is aware that the constructor no longer exists, and is not marked as `@dart=2.x` with x < 12.
TEST=ci
Change-Id: I4ffaf3ae2c4e75ca06e7ba0bf19187b6376f3888
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/297100
Reviewed-by: Brian Quinlan <bquinlan@google.com>
Reviewed-by: Nate Bosch <nbosch@google.com>
Reviewed-by: Sigmund Cherem <sigmund@google.com>
Commit-Queue: Stephen Adams <sra@google.com>
`o is T` where `T` is an interface implemented by a class hierarchy
can be implemented as `o instanceof R` where `R` is the root of the
hierarchy. The common case is that `R` is a single implementation
class.
`is Record` and `is Type` now generally use `instanceof`.
Fixes: #51366
Change-Id: I1943533bc53024f3199f71910034d49cf6661b22
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/294320
Commit-Queue: Stephen Adams <sra@google.com>
Reviewed-by: Mayank Patke <fishythefish@google.com>
Suppose we have code like
if (isDevMode) {
// do foo
} else {
// do bar
}
where isDevMode is const (or can be evaluated as const). In particular,
isDevMode can be controlled by compile-time options, like
bool.fromEnvironment.
We currently eliminate the dead branch during SSA, but that means we do
the work of compiling dead code in the previous phases. Instead, we can
recognize that the condition is const (or effectively const) and
eliminate the dead AST subtree directly.
Change-Id: Ia91da5ebc7fa496a1b963308c6e02d572cab936e
Bug: b/254543452
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/270281
Reviewed-by: Sigmund Cherem <sigmund@google.com>
The library wasn't taken into account when handling computing class
members in the K/J-model and not handled correctly when computing
applicable selectors. This made dart2js unable to handle members with
a name private to a different library than the enclosing library.
Fixes#33732Fixes#49226
Change-Id: I5ba143d87662bbd42e0ff02355054e4a937be8f8
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/252665
Commit-Queue: Johnni Winther <johnniwinther@google.com>
Reviewed-by: Stephen Adams <sra@google.com>
This annotation prevents inlining at call sites within the method that
is annotated.
The main expected use is to control inlining in unit tests.
Change-Id: Ib6652f0b4cb7a34c6bf829b1be5c09e241ffbde2
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/248345
Commit-Queue: Stephen Adams <sra@google.com>
Reviewed-by: Sigmund Cherem <sigmund@google.com>
Late fields start out holding a sentinel value and after they are initialized or assigned, never hold a sentinel value again.
This CL implements some removal of checks and loads based on this monotonic behaviour.
On two large Dart Angular apps, this reduces the number of late-instance-field checks by ~25%.
One app is 0.44% smaller, the other is 0.98% smaller.
The optimization_test.dart no longer forces `--disable-inlining` because that prevents the inlining that exposes the optimization opportunities.
Most tests already use `@pragma('dart2js:noInline')` to prevent inlining where relevant.
Change-Id: I2327a96e69191313ed6c07eabe94aa2962fb2648
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/246881
Reviewed-by: Joshua Litt <joshualitt@google.com>
Commit-Queue: Stephen Adams <sra@google.com>
Reviewed-by: Sigmund Cherem <sigmund@google.com>
Always import 'lib/compiler_api.dart' with prefix 'api'.
Remove import of 'lib/src/compiler.dart' with prefix 'api'.
Some of the test helper files under 'compiler/test' are a little tedious
with lots of 'api.' prefixes and I could be convinced to change them
back but I think we should stick with always using a prefix under
'compiler/lib/'.
Change-Id: I3f82d7d31f1c90a860c9811fce031a64cea04ddb
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/241867
Reviewed-by: Joshua Litt <joshualitt@google.com>
Commit-Queue: Stephen Adams <sra@google.com>
`int` variables can attain NaN values because web int arithmetic
implemented by JavaScript numbers (doubles) is not closed under many
operations. It is possible get NaN using only addition:
int a = 1, b = -1;
while (a + a != a) { a += a; b += b; }
int nan = a + b;
On the VM, a, b and nan are all zero.
On the web, a, b and nan are Infinity, -Infinity and NaN, respectively.
Since NaN can leak into int arithmetic, is it helpful if bounds checks
catch NaN indexes. NaN compares false in any comparison, so a test
of the form
if (index < 0 || index >= a.length) throw ioore(a, index);
fails to detect a NaN value of `index`.
This is fixed by negating the comparisons, and applying De Morgan's law:
if (!(index >= 0 && index < a.length)) throw ioore(a, index);
These changes have been applied to JSArray.[], JSArray.[]= and String.[]
For dart2js the change is a little more involved. Primitive indexing is
lowered to code with a HBoundsCheck check instruction. The code generated
for the instruction now uses, e.g. `!(i>=0)` instead of `i<0`.
This leads to a small code size regression.
There is no regression at -O4 since bounds checks are omitted at -O4.
At -O3 (where the regression is largest) the regression is
0.01% for cm
0.06% for flutter gallery -- array-heavy diff and layout
0.21% for Meteor -- array-heavy code
0.30% for Box2DOctane -- array-heavy code
I believe the regression can be largely alleviated by determining if
NaN is impossible at the index check, and if so, reverting to the smaller
code pattern. The analysis could be global, incorporating NaN into the
global abstract value domain, or a much simpler a local dataflow
analysis. Many indexes are loop driven and cannot reach infinity because
they are incremented by a small bump and eventually (even without a loop
guard) the index would stop growing when the increment falls below the
rounding error in O(2^53) iterations.
Change-Id: I23ab1eb779f1d0c9c6655e13d69f65d453db9284
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/210321
Commit-Queue: Stephen Adams <sra@google.com>
Reviewed-by: Mayank Patke <fishythefish@google.com>
Mayank Patke