commit 06402ab1ca284cfd394a8bfb7fddbcbe242fc86e
author Jens Johansen <jensj@google.com> Wed Jul 30 07:05:12 2025 -0700
committer Commit Queue <dart-scoped@luci-project-accounts.iam.gserviceaccount.com> Wed Jul 30 07:05:12 2025 -0700
tree 7393f1ba5f7daa4e9db17b0f67ec4cf8ac7970fb
parent 9c5eea2b1af09ded9c9ab0d7aaa8e91e80e824d7

[parser] Optimize mayFollowTypeArgs

Doing benchmarks before and after I get the below.
Notice that this is on tokens/ms, i.e. higher is better.

JIT:

pkg/analyzer/lib/src/dart/ast/ast.dart:

```
Difference at 95.0% confidence
        1.27922 +/- 0.393002
        6.47159% +/- 1.9882%
        (Student's t, pooled s = 0.418268)
```

pkg/front_end/lib/src/type_inference/inference_visitor.dart:

```
Difference at 95.0% confidence
        0.968609 +/- 0.287367
        6.8824% +/- 2.04187%
        (Student's t, pooled s = 0.305841)
```

In both cases it processes a little over 6% more tokens per ms.

AOT:

pkg/analyzer/lib/src/dart/ast/ast.dart:

```
Difference at 95.0% confidence
        2.39988 +/- 0.429312
        9.50898% +/- 1.70105%
        (Student's t, pooled s = 0.456911)
```

pkg/front_end/lib/src/type_inference/inference_visitor.dart:

```
Difference at 95.0% confidence
        2.28604 +/- 0.279366
        12.9521% +/- 1.58282%
        (Student's t, pooled s = 0.297326)
```

It processes between 9-13% more tokens per ms.


Additionally, for AOT, I ran it through the benchmarker and got this:

pkg/analyzer/lib/src/dart/ast/ast.dart:

```
msec task-clock:u: -8.5488% +/- 2.1840% (-275.52 +/- 70.39) (3222.95 -> 2947.42)
cycles:u: -8.6402% +/- 2.1619% (-1210657004.60 +/- 302921394.06) (14011942063.40 -> 12801285058.80)
instructions:u: -9.0870% +/- 0.0000% (-2746030853.20 +/- 876.33) (30219334757.30 -> 27473303904.10)
branch-misses:u: -13.3846% +/- 10.1850% (-5542357.40 +/- 4217459.09) (41408381.20 -> 35866023.80)
seconds time elapsed: -8.5442% +/- 2.1816% (-0.28 +/- 0.07) (3.22 -> 2.95)
seconds user: -8.5614% +/- 2.2098% (-0.27 +/- 0.07) (3.20 -> 2.93)
```

pkg/front_end/lib/src/type_inference/inference_visitor.dart

```
msec task-clock:u: -11.6033% +/- 0.9079% (-402.51 +/- 31.50) (3468.96 -> 3066.45)
cycles:u: -11.6826% +/- 0.9053% (-1765194611.60 +/- 136793475.00) (15109578869.80 -> 13344384258.20)
instructions:u: -11.4435% +/- 0.0000% (-3905886611.70 +/- 1150.33) (34131775222.10 -> 30225888610.40)
branch-misses:u: -17.5244% +/- 4.6453% (-8196278.70 +/- 2172645.25) (46770630.60 -> 38574351.90)
seconds time elapsed: -11.6004% +/- 0.9056% (-0.40 +/- 0.03) (3.47 -> 3.07)
seconds user: -11.6215% +/- 0.9879% (-0.40 +/- 0.03) (3.45 -> 3.05)
```

I.e. in both cases it's 8-11% less work (instructions, time etc).

Compiling the CFE with the CFE before and after (AOT) I get this change:

```
msec task-clock:u: -0.4356% +/- 0.3110% (-23.07 +/- 16.47) (5295.49 -> 5272.42)
page-faults:u: 0.1135% +/- 0.0534% (123.64 +/- 58.22) (108956.82 -> 109080.46)
cycles:u: -0.5004% +/- 0.3175% (-109702416.52 +/- 69596415.67) (21921571730.18 -> 21811869313.66)
instructions:u: -0.3994% +/- 0.0005% (-104979706.80 +/- 138268.02) (26284569470.80 -> 26179589764.00)
branch-misses:u: -1.6117% +/- 1.4644% (-1285239.90 +/- 1167813.84) (79746730.52 -> 78461490.62)
seconds time elapsed: -0.4291% +/- 0.3090% (-0.02 +/- 0.02) (5.30 -> 5.28)
seconds user: -0.3992% +/- 0.3801% (-0.02 +/- 0.02) (5.04 -> 5.02)
```

So this improvement to the parser reduces the runtime of compiling by ~0.4%.

Change-Id: Ia89d02e8cce2655fb23ffe18ecbab63521ce75e3
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/439721
Commit-Queue: Jens Johansen <jensj@google.com>
Reviewed-by: Johnni Winther <johnniwinther@google.com>