R8 removes enum class values, despite usage from generated code [339085669]

WAI

Bug

Status Update

No update yet.

Description

cc...@google.com

created issue #1

May 6, 2024 10:04PM

In tweaking proguard rules in androidx microbenchmarks, I noticed that some of our generated moshi code references to a values array are failing to resolve, and crashing at runtime:

> ./gradlew benchmark:benchmark-benchmark:installRelease && adb shell am instrument -w -r -e debug false -e newRunListenerMode true -e class 'androidx.benchmark.benchmark.TrivialJavaBenchmark' -e androidx.benchmark.output.enable true -e listener androidx.benchmark.junit4.InstrumentationResultsRunListener androidx.benchmark.benchmark.test/androidx.benchmark.junit4.AndroidBenchmarkRunner

Time: 9.563
There were 2 failures:
1) increment(androidx.benchmark.benchmark.TrivialJavaBenchmark)
java.lang.ExceptionInInitializerError
	at androidx.benchmark.junit4.BenchmarkRule$applyInternal$$inlined$Statement$1.evaluate(SourceFile:123)
	at androidx.test.rule.GrantPermissionRule$RequestPermissionStatement.evaluate(Unknown Source:11)
	at org.junit.rules.RunRules.evaluate(Unknown Source:2)
...
	at androidx.test.runner.AndroidJUnitRunner.onStart(SourceFile:134)
	at android.app.Instrumentation$InstrumentationThread.run(Instrumentation.java:2402)
Caused by: java.lang.RuntimeException: java.lang.NoSuchMethodException: androidx.benchmark.json.BenchmarkData$TestResult$ProfilerOutput$Type.values []
	at java.lang.Enum.enumValues(Enum.java:315)
...
	at androidx.benchmark.ResultWriter.<clinit>(SourceFile:20)
	... 31 more
Caused by: java.lang.NoSuchMethodException: androidx.benchmark.json.BenchmarkData$TestResult$ProfilerOutput$Type.values []
	at java.lang.Class.getMethod(Class.java:2937)
	at java.lang.Class.getDeclaredMethod(Class.java:2914)
	at java.lang.Enum.enumValues(Enum.java:312)
	... 64 more

# also repros with
#    ./gradlew benchmark:benchmark-benchmark:cC -P android.testInstrumentationRunnerArguments.class=androidx.benchmark.benchmark.TrivialJavaBenchmark
# but above makes crash stack clearer

Repro AndroidX CL, which just enables R8 in a trivial sample benchmark, and tweaks proguard rules to get it (almost) running: https://android-review.git.corp.google.com/c/platform/frameworks/support/+/3073538

Comments

cc...@google.com <cc...@google.com> #2May 6, 2024 10:22PM

Sergio shared a WIP CL to calculate more correct bounds. You can see in the screenshot attached here that it does a better job of tightening the bounds around the object (the yellow box), but that it misses some of the curves which expand beyond these tighter bounds. In particular, note the two triangles and the two blob shapes (rotated round rects).
I've attached another 2 screenshots from the ShapeEditor view of one of the blobs, showing the shape and controls/anchors. You can see that the tighter bounds (yellow rectangle) are constrained to the anchors, but the curves at the end go beyond those points and exceed the bounds.

ap...@google.com <ap...@google.com> #3May 6, 2024 10:53PM

Part of the problem is that the min/max values are assigned the anchors to begin with. That would cause the problem we're seeing in #2, because these new bounds are constrained to the rectangle created by the anchor points, even if the control points cause curves to go outside of those bounds.
But that's not all of the problem here. When I changed those min/max values to use the control points instead, the bounds expanded to be essentially the same as the estimated bounds from before. So there's more to it.

ch...@google.com <ch...@google.com> #4May 7, 2024 07:34AM

Reassigned to ch...@google.com.

Some of #3 is incorrect - the bounds can be expanded from the initial min/max values, so it is fine to have them start at the anchor points.

Meanwhile, I thought part of the problem might be calculating derivatives/etc for zero-length curves (of which there are many, given the way our shape rounding works). But adding in a check for this (setting the bounds of a curve to the same point when the anchors are equal within the distance epsilon) did not fix the problem; we still end up with bounds that are inside some of the curves (see attached screenshot)

cc...@google.com <cc...@google.com> #5May 7, 2024 10:44PM

Project: platform/frameworks/support
Branch: androidx-main

commit 958577b15712233aa487003deb0ab37de3af7ab7
Author: Chet Haase <chet@google.com>
Date: Wed Dec 20 13:21:53 2023

Added more bounds calculation utilities

Previously, calculateBounds() returned the estimated bounds
for a RoundedPolygon. This is reasonable in most cases, but sometimes
you want the actual/minimal bounds instead. Also, sometimes you want
the max bounds for an object that might be rotated within its container,
so you need to size the container appropriately.

This CL introduces options for bounds calculations to handle these
new cases, along with some minor cleanups along the way.

Bug: 317286450
Test: Existing tests pass, added new bounds tests
Relnote: Now more options for retrieving exact and max bounds

Change-Id: I6d49f468a28c1f360000e8370f02a50841f744e4

M graphics/graphics-shapes/api/current.txt
M graphics/graphics-shapes/api/restricted_current.txt
M graphics/graphics-shapes/src/androidInstrumentedTest/kotlin/androidx/graphics/shapes/PolygonTest.kt
M graphics/graphics-shapes/src/commonMain/kotlin/androidx/graphics/shapes/Cubic.kt
M graphics/graphics-shapes/src/commonMain/kotlin/androidx/graphics/shapes/RoundedPolygon.kt
M graphics/graphics-shapes/src/commonMain/kotlin/androidx/graphics/shapes/Utils.kt
M graphics/integration-tests/testapp-compose/src/main/java/androidx/graphics/shapes/testcompose/DebugDraw.kt
M graphics/integration-tests/testapp-compose/src/main/java/androidx/graphics/shapes/testcompose/MainActivity.kt

https://android-review.googlesource.com/2889768

ze...@google.com <ze...@google.com> #6May 8, 2024 04:44AM

Problem fixed - we now offer the option to request actual bounds (or faster/approximated, based on a parameter to calculateBounds()).
We also offer a new API, calculateMaxBounds(), which returns the square bounds in which the object can be rotated without ever exceeding the bounds.

cc...@google.com <cc...@google.com> #7May 9, 2024 04:58PM

Is it possible for R8 to figure out which enum classes are potentially fed to such stl methods? If so, and the stl behavior is consistent (and stl using values to implement class.enumConstants seems reasonable), it would be nice to avoid this necessary configuration (even if studio gives it to you by default in the very stale proguard-android-optimize list).

Either that, or could make enum value/valueOf minification opt-in.

ze...@google.com <ze...@google.com> #8May 13, 2024 06:12AM

Unfortunately, we have not been able to come up with a safe/sound analysis that can determine when the rules are not needed. Conversely, we have no way of safely keeping the enums that is less conservative than the rules. There are clients that do not have these rules as they manually ensure that they have no usages of enums that would break. Thus we can't make the rules part of the compiler and we can't safely determine when they are needed. We are not happy with that situation, but we currently have no better alternatives.

cc...@google.com <cc...@google.com> #9May 13, 2024 11:40PM

Fair enough, feel free to close as WAI. Understand that the cost to migrate the default is non-zero, with minimal benefit.

ch...@google.com <ch...@google.com> May 14, 2024 07:31AM

Status: Won't Fix (Intended Behavior)

pr...@google.com <pr...@google.com> #10May 14, 2024 05:22PM

The following release(s) address this bug.It is possible this bug has only been partially addressed:

androidx.benchmark:benchmark-junit4:1.3.0-alpha05