Update ProGuard default shrinking rules #2448

sgjesse · 2023-07-26T13:01:56Z

This change deals correctly with classes without a no-args constructor. If a no-args constructor was not present the conditional keep rule would not trigger.

See https://issuetracker.google.com/150189783#comment11.

…s without a no-args constructor

eamonnmcmanus

Looks like ShrinkingIT needs to be updated?

sgjesse · 2023-07-26T13:16:12Z

Yes, and that makes sense. However, I cannot reproduce locally. If I run mvn clean verify I get 7 failures, but not the one on ShrinkingIT. What am I missing?

eamonnmcmanus · 2023-07-26T18:07:24Z

Yes, and that makes sense. However, I cannot reproduce locally. If I run mvn clean verify I get 7 failures, but not the one on ShrinkingIT. What am I missing?

I was seeing the same thing with JDK 11 on Google's internal Linux. #2450 should fix it. With that PR, I am able to reproduce the test failure in that environment.

Marcono1234 · 2023-07-26T19:18:56Z

This case of using <init>() vs. <init>(...) was what we were discussing in #2420 (comment) I think.

Out of curiosity, could you please explain a bit why you think it is needed?

In the end it is still Éamonn's decission whether this is included, but I am a bit afraid that with this change we contribute to people unwittingly relying on Unsafe, while we would have the chance here to indicate to them that their class is misconfigured (missing a no-args constructor).

Marcono1234 · 2023-07-26T19:37:10Z

If I run mvn clean verify I get 7 failures, but not the one on ShrinkingIT. What am I missing?

Just to be sure, you run this in the parent directory, right? ShrinkingIT is part of a separate Maven module, so it won't be executed if you run inside /gson/gson.

Edit: Nevermind, sorry the reason is the other test failures both of you mentioned. Probably due to that the Maven build already fails for the gson module and doesn't reach the shrinker-test module which contains ShrinkingIT.

sgjesse · 2023-07-27T13:18:32Z

After testing with the new configuration I realized that with just <init>() the conditional -if rule would not actually match classes without a no-args constructor - which was also what we discussed and decided on. However, for common code patterns data classes will not have a no-args constructor, e.g.:

Java

public class Data {
  @SerializedName("s")
  private final String s;
  public Data(String s) { this.s = s; }
}

Kotlin

data class Data(
  @field:SerializedName("s") val s: String
)

To get no-args constructors requires effectively assigning default values to all fields for these patterns (either directly for the field or assigning in an explicit no-args constructor), which I don't think is feasible to require, and then the provided rules will not work for many developers.

I is still an option for developers to add no-args constructors to not use sun.misc.Unsafe, and if present they will still be kept.

sgjesse · 2023-07-27T13:19:17Z

Tests updated, please review.

sfreilich

Three other issues stand out to me here:

There are two rules about classes with @SerializedName fields, which I think could be consolidated into one -keepclasseswithmembers rule?
More of these -keep/-keepclassmembers declarations could use allowobfuscation, which should work fine with reflection that refers to things by symbol instead of string name (including these annotations).
Possibly this should have -keep,allowobfuscation class rules for the annotations used here. I think that can matter when there are multiple optimization passes? Not 100% sure, basing this on other examples that follow that pattern. Not sure if that would be best to do with a rule that matches com.google.gson.annotations.* or just JsonAdapter and SerializedName specifically.

sfreilich · 2023-07-27T13:52:28Z

gson/src/main/resources/META-INF/proguard/gson.pro

@@ -65,6 +65,6 @@
 # See also https://github.com/google/gson/pull/2420#discussion_r1241813541 for a more detailed explanation
 -if class *


I think this if class *, -classeswithmembers... class <1> doesn't do anything relative to -classeswithmembers... class * , after reading some of the follow-up threads on that change.

At least for R8 it does seem to make a difference: -keepclasseswithmembers ... class * always keeps the class, even if it isn't used. Whereas -if class * ... -keepclasseswithmembers ... class <1> only keeps the class if it is actually used.

Unfortunately this is not covered by the tests yet; have created #2455 to add a test for this.

Thanks also for the other points you raised, but unless sgjesse wants to address them here as well could you please create a separate GitHub issue or pull request? Otherwise the discussion on this pull request here might drift too far away from the original topic.

Changed the rules to be two -if rules

Keep all @SerializedName fields for classes which are present after shrinking

If a class with a @SerializedName field is present keep its no-args constructor

sfreilich · 2023-07-27T14:10:49Z

Possibly this should have -keep,allowobfuscation class rules for the annotations used here.

Realized this bit probably doesn't matter, I was confused by examples for annotations that are just referenced by proguard config rules, and not actually otherwise referenced in the code.

Marcono1234

@sgjesse, hmmm ok, let's hope then that those users don't encounter the pitfalls of JDK Unsafe (see documentation of GsonBuilder.disableJdkUnsafe()) (and also let's hope that neither the JDK nor Android remove the needed JDK Unsafe functionality any time soon).

Personal opinion on this (click to show)

Personally I am a bit indecisive on these changes; on one hand I can definitely see that it would be convenient for users if it just worked automatically. On the other hand, the problem is that JDK Unsafe usage is opt-out and I assume most users are not aware of it being used at all, nor what the implications of it are (if it was opt-in I wouldn't have a problem with that). If as user you want to use your classes with Gson, then you should (in my opinion) adjust them to follow Gson's requirements (or if necessary differentiate between application logic classes and DTO classes, or however you want to call them), and from that perspective I would appreciate any changes which make users less dependent on JDK Unsafe.

Though I am not a direct member of this project, so it is not up to me to decide this.

Thanks for the test changes, could you please adjust the comments which are now partially obsolete?

Also, ideally there would be a new test in shrinker-test's com.example.Main which covers the successful case of a class without no-args constructor but with fields annotated with @SerializedName, which is now supported with your changes.
If you don't want to do that I can do it in a follow-up PR.

Technically Troubleshooting.md is now also a bit outdated (e.g. it says "has a no-args constructor"), but I would leave it that way for now to encourage usage of a no-args constructor to avoid relying on JDK Unsafe.

Marcono1234 · 2023-07-29T11:35:49Z

gson/src/main/resources/META-INF/proguard/gson.pro

@@ -65,6 +65,6 @@
 # See also https://github.com/google/gson/pull/2420#discussion_r1241813541 for a more detailed explanation
 -if class *
 -keepclasseswithmembers,allowobfuscation,allowoptimization class <1> {
-  <init>();
+  <init>(...);


Please adjust the comment in lines 62 and 63 (GitHub does not let me comment there); for example reword it to:

# If a class is used in some way by the application and has fields annotated with @SerializedName, # keep those fields and the constructors of the class # For convenience this also matches classes without no-args constructor, in which case Gson uses JDK Unsafe

Marcono1234 · 2023-07-29T11:42:40Z

shrinker-test/src/main/java/com/example/DefaultConstructorMain.java

  // making class abstract); other constructors are ignored to suggest to user adding default
  // constructor instead of implicitly relying on JDK Unsafe


Please adjust this comment (starting in line 19), for example:

// Specify explicit constructor with args to remove implicit no-args default constructor

The current comment mentioning ProGuard rules and JDK Unsafe is now obsolete.

Marcono1234 · 2023-07-29T11:44:48Z

shrinker-test/src/main/java/com/example/DefaultConstructorMain.java

@@ -20,7 +20,7 @@ static class TestClassNotAbstract {
  // making class abstract); other constructors are ignored to suggest to user adding default
  // constructor instead of implicitly relying on JDK Unsafe
  static class TestClassWithoutDefaultConstructor {
-    @SerializedName("s")
+    // No @SerializedName annotation to avoid matching the consumer rules from gson.pro.


This is a good addition; could you please copy it to the other classes TestClass and TestClassNotAbstract above as well?

Though I think the trailing period should be omitted, it just looks a bit weird after the file name:

Suggested change

// No @SerializedName annotation to avoid matching the consumer rules from gson.pro.

// No @SerializedName annotation to avoid matching the consumer rules from gson.pro

sfreilich · 2023-08-01T16:53:14Z

I'll file a PR for whatever @sgjesse doesn't want to handle here. We were just discussing that since I was trying to solve some obstacles to adopting these rules downstream, and we're colleagues, we share a downstream.

Marcono1234 · 2023-08-05T12:34:52Z

@sgjesse, I have tried to address my review comments in sgjesse#1 now; I hope that is easier for you. That pull request targets your fix-proguard-rules branch, so in case you merge it, the changes should show up here as well.

…-changes Adjust shrinker tests

Marcono1234 · 2023-08-13T00:02:14Z

shrinker-test/proguard.pro

+#-keep class com.example.ClassWithSerializedName {
+#  <init>(...);
+#}


Is this needed?

Good catch, no removed the rules in comments. Also removed the rule above, as that was also not needed. I probably added it while testing different rules.

Marcono1234 · 2023-08-13T00:04:11Z

gson/src/main/resources/META-INF/proguard/gson.pro

+# Keep fields annotated with @SerializedName for classes which are present.
+# If classes with fields annotated with @SerializedName have a no-args
+# constructor keep that as well.


Why did you remove the link to the pull request discussion?

At least in my opinion this -if class * ... is pretty cryptic and its intended behavior is not directly obvious, so it would be good to have a reference explaining what this actually does.

Good point, added those references back.

Marcono1234 · 2023-08-13T00:04:49Z

gson/src/main/resources/META-INF/proguard/gson.pro

 -if class *
-keepclasseswithmembers,allowobfuscation,allowoptimization class <1> {


Why did you remove allowoptimization?

Having allowoptimization can allow R8 to propagate field values, so if you have an explicit construction of a serialized class always setting a field to a specific value, then that single value can become the default value for the created instances instead of 0/null.

Marcono1234 · 2023-08-13T00:06:22Z

gson/src/main/resources/META-INF/proguard/gson.pro

 -if class *
-keepclasseswithmembers,allowobfuscation,allowoptimization class <1> {
-  <init>(...);


Is it intentional that you removed <init>(...) again and used <init>() below? My personal opinion in #2448 (review) was not blocking the merge of this pull request (sorry if I did not make that clear enough).

~~This change to use <init>() again might render some or all of the shrinker test changes of this pull request obsolete.~~
Edit: Maybe not, see comment below.

Or at least I am a bit surprised that no exception occurs for ClassWithoutDefaultConstructor; I assume it might be related to your newly added -keepclasseswithmembers.

So, do I understand it correctly that your new approach of using -keepclasseswithmembers achieves the same thing as your initial <init>(...) approach (i.e. Gson can instantiate classes without no-args constructor using Unsafe; instead of failing because R8 made the class abstract). But this new approach is better than the <init>(...) rule because that would have unnecessarily kept the constructors with args around, even though they are not actually used by Gson?

Marcono1234 · 2023-08-13T00:09:18Z

gson/src/main/resources/META-INF/proguard/gson.pro

-# Keep fields with @SerializedName annotation, but allow obfuscation of their names
-keepclassmembers,allowobfuscation class * {
-  @com.google.gson.annotations.SerializedName <fields>;
-}


Out of curiosity, why did you remove this here and add it below guarded by a -if class *? Does that make a difference for the end result?

Yes, in R8 full mode we have a distinction between referenced and instantiated classes. For classes which are not seen by R8 as instantiated (e.g. if only referenced statically using .class as one will do in fromJson then -keepclassmembers will only keep static members and not instance members. For ProGuard and R8 in compatability mode the two will do the same.

Hmm thanks, ok that sounds like the previous rules would not have worked properly then because -keepclassmembers would not have kepts the non-static @SerializedName fields. Though I think the shrinker-test showed that it did work as expected?
I am bit confused, or am I misunderstanding your comment or are the shrinker-tests flawed and not close enough to real code?

One way to see the difference is to have an instance field annotated with @SerializedName which is not used in the code at all, then the field will be removed with the previous rule in R8 full mode (which is default from AGP 8.0.0). Then in a pass through from fromJson to toJson the field value will be lost. Of course it is a matter of taste if unused fields should be kept this way.

sgjesse

Addressed all comments.

sgjesse · 2023-08-10T11:33:13Z

gson/src/main/resources/META-INF/proguard/gson.pro

@@ -65,6 +65,6 @@
 # See also https://github.com/google/gson/pull/2420#discussion_r1241813541 for a more detailed explanation
 -if class *


Changed the rules to be two -if rules

Keep all @SerializedName fields for classes which are present after shrinking

If a class with a @SerializedName field is present keep its no-args constructor

sgjesse · 2023-08-15T08:38:36Z

gson/src/main/resources/META-INF/proguard/gson.pro

-# Keep fields with @SerializedName annotation, but allow obfuscation of their names
-keepclassmembers,allowobfuscation class * {
-  @com.google.gson.annotations.SerializedName <fields>;
-}


Yes, in R8 full mode we have a distinction between referenced and instantiated classes. For classes which are not seen by R8 as instantiated (e.g. if only referenced statically using .class as one will do in fromJson then -keepclassmembers will only keep static members and not instance members. For ProGuard and R8 in compatability mode the two will do the same.

sgjesse · 2023-08-15T08:41:23Z

gson/src/main/resources/META-INF/proguard/gson.pro

+# Keep fields annotated with @SerializedName for classes which are present.
+# If classes with fields annotated with @SerializedName have a no-args
+# constructor keep that as well.


Good point, added those references back.

sgjesse · 2023-08-15T08:46:24Z

gson/src/main/resources/META-INF/proguard/gson.pro

 -if class *
-keepclasseswithmembers,allowobfuscation,allowoptimization class <1> {


Having allowoptimization can allow R8 to propagate field values, so if you have an explicit construction of a serialized class always setting a field to a specific value, then that single value can become the default value for the created instances instead of 0/null.

sgjesse · 2023-08-15T08:51:58Z

shrinker-test/proguard.pro

+#-keep class com.example.ClassWithSerializedName {
+#  <init>(...);
+#}


Good catch, no removed the rules in comments. Also removed the rule above, as that was also not needed. I probably added it while testing different rules.

Marcono1234

Thanks for your update and also for your work on this!

Though it looks like some of the changes in your last commit might not have been intentional (see comments)?

Also, I think #2448 (comment) is still open; just to clarify that nothing was overlooked.

See also #2448 (comment) (unfortunately GitHub does not show that comment down here as well, so it might be easy to miss).

Marcono1234 · 2023-08-15T20:47:15Z

gson/src/main/resources/META-INF/proguard/gson.pro

@@ -22,6 +22,7 @@
 # Keep class TypeToken (respectively its generic signature) if present
 -if class com.google.gson.reflect.TypeToken
 -keep,allowobfuscation class com.google.gson.reflect.TypeToken
+#-keep class com.google.gson.reflect.TypeToken { *; }


Is this commented line intentional?

No removed.

Marcono1234 · 2023-08-15T20:50:29Z

gson/src/main/resources/META-INF/proguard/gson.pro

 -if class *
 -keepclasseswithmembers,allowobfuscation class <1> {
  @com.google.gson.annotations.SerializedName <fields>;
 }
 -if class * {
  @com.google.gson.annotations.SerializedName <fields>;
 }
-keepclassmembers,allowobfuscation class <1> {
+-keepclassmembers,allowobfuscation,allowoptimization class <1> {


I thought in #2448 (comment) you just explained that you intentionally removed allowoptimization, but now you are adding it back? Is that intentional?

(Or am I misunderstanding this?)

This rule is only covering the no-args constructor (<init>()), so here there is no issue with allowoptimization. For the rule above for the annotated fields it should not be there as otherwise R8 can do value propagation.

Marcono1234 · 2023-08-15T20:52:40Z

shrinker-test/common.pro

@@ -38,3 +38,5 @@
 -keepclassmembernames class com.example.ClassWithVersionAnnotations {
  <fields>;
 }
+
+-dontobfuscate


Is this intentional? I think this should not be necessary and there were explicit rules which preserved names where needed for the tests.
Though it looks like you (accidentally?) removed one in proguard.pro:

- -keepclassmembernames class com.example.NoSerializedNameMain$TestClassWithoutDefaultConstructor { - <fields>; - }

No, this was all for some testing. These changes should not have been there.

sgjesse · 2023-08-21T12:24:19Z

@Marcono1234 Thank you for being such an observant reviewer! Addressed my mistakes in my previous upload. Please take another look.

eamonnmcmanus

I think this is close enough now. If there are further changes needed, we can make them in a follow-up PR.

Marcono1234 · 2023-08-22T20:56:17Z

@Marcono1234 Thank you for being such an observant reviewer!

Thank you for all the work you have put into this, and all the detailed explanations! That was helpful for me to understand at least a bit better how R8 works. And hopefully it will also be useful for future reference.

The changes look good, so I don't think there is any direct follow-up PR needed.

sfreilich · 2023-08-22T22:15:33Z

@Marcono1234, @sgjesse, @eamonnmcmanus: Thank you all for your work on this, I really appreciate it!

Update ProGuard default shrinking rules to correctly deal with classe…

7fa9ef5

…s without a no-args constructor

eamonnmcmanus approved these changes Jul 26, 2023

View reviewed changes

eamonnmcmanus requested changes Jul 26, 2023

View reviewed changes

eamonnmcmanus added a commit to eamonnmcmanus/gson that referenced this pull request Jul 26, 2023

Copy of google#2448 atop text fixes.

f13322c

sgjesse added 2 commits July 27, 2023 13:37

Merge branch 'google:main' into fix-proguard-rules

e384d4d

Update test after changing default shrinking rules

f60e17d

sfreilich suggested changes Jul 27, 2023

View reviewed changes

Marcono1234 mentioned this pull request Jul 29, 2023

Copy of #2448 atop test fixes #2451

Closed

9 tasks

Marcono1234 suggested changes Jul 29, 2023

View reviewed changes

Marcono1234 mentioned this pull request Jul 29, 2023

Add test for unused class to shrinker-test #2455

Merged

9 tasks

Adjust shrinker tests

eea08d0

Marcono1234 mentioned this pull request Aug 5, 2023

Adjust shrinker tests sgjesse/gson#1

Merged

sgjesse added 3 commits August 7, 2023 10:22

Merge pull request #1 from Marcono1234/marcono1234/fix-proguard-rules…

4c40a91

…-changes Adjust shrinker tests

Update rules

c34126b

Merge branch 'main' into fix-proguard-rules

af98d32

Marcono1234 reviewed Aug 13, 2023

View reviewed changes

Addressed review comments

ac6e590

sgjesse commented Aug 15, 2023

View reviewed changes

Marcono1234 reviewed Aug 15, 2023

View reviewed changes

sgjesse added 2 commits August 21, 2023 14:12

Addressed more review comments

8d14f1a

Addressed more review comments

743d529

Merge branch 'main' into fix-proguard-rules

3468700

sgjesse requested a review from eamonnmcmanus August 21, 2023 13:30

eamonnmcmanus approved these changes Aug 22, 2023

View reviewed changes

eamonnmcmanus merged commit 7b5629b into google:main Aug 22, 2023
9 checks passed

Marcono1234 added the proguard-r8 Issues relating to the use of ProGuard and/or R8, such as problems due to obfuscation label Mar 29, 2024

		@@ -65,6 +65,6 @@
		# See also https://github.com/google/gson/pull/2420#discussion_r1241813541 for a more detailed explanation
		-if class *

		// making class abstract); other constructors are ignored to suggest to user adding default
		// constructor instead of implicitly relying on JDK Unsafe

	// No @SerializedName annotation to avoid matching the consumer rules from gson.pro.
	// No @SerializedName annotation to avoid matching the consumer rules from gson.pro

		-if class *
		-keepclasseswithmembers,allowobfuscation,allowoptimization class <1> {

Update ProGuard default shrinking rules #2448

Update ProGuard default shrinking rules #2448

Conversation

sgjesse commented Jul 26, 2023

eamonnmcmanus left a comment

Choose a reason for hiding this comment

sgjesse commented Jul 26, 2023

eamonnmcmanus commented Jul 26, 2023

Marcono1234 commented Jul 26, 2023

Marcono1234 commented Jul 26, 2023 • edited

sgjesse commented Jul 27, 2023

sgjesse commented Jul 27, 2023

sfreilich left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Marcono1234 Jul 29, 2023 • edited

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

sfreilich commented Jul 27, 2023

Marcono1234 left a comment • edited

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

sfreilich commented Aug 1, 2023

Marcono1234 commented Aug 5, 2023

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Marcono1234 Aug 13, 2023 • edited

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Marcono1234 Aug 15, 2023 • edited

Choose a reason for hiding this comment

Choose a reason for hiding this comment

sgjesse left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Marcono1234 left a comment • edited

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Marcono1234 Aug 15, 2023 • edited

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

sgjesse commented Aug 21, 2023

eamonnmcmanus left a comment

Choose a reason for hiding this comment

Marcono1234 commented Aug 22, 2023

sfreilich commented Aug 22, 2023

Marcono1234 commented Jul 26, 2023 •

edited

Marcono1234 Jul 29, 2023 •

edited

Marcono1234 left a comment •

edited

Marcono1234 Aug 13, 2023 •

edited

Marcono1234 Aug 15, 2023 •

edited

Marcono1234 left a comment •

edited

Marcono1234 Aug 15, 2023 •

edited