Status Update
Comments
je...@google.com <je...@google.com>
hu...@google.com <hu...@google.com> #2
Thank you for the report and reproduction. I can reproduce on a Android TV emulator.
an...@google.com <an...@google.com> #3
Status of current findings.
This is related to the SplitReturnRewriter
added in
Not running this code pass on java.lang.Object androidx.collection.LruCache.remove(java.lang.Object)
works around the issue. I used the following patch:
diff --git a/src/main/java/com/android/tools/r8/ir/conversion/passes/SplitReturnRewriter.java b/src/main/java/com/android/tools/r8/ir/conversion/passes/SplitReturnRewriter.java
index 157d26c66a..63e0f76064 100644
--- a/src/main/java/com/android/tools/r8/ir/conversion/passes/SplitReturnRewriter.java
+++ b/src/main/java/com/android/tools/r8/ir/conversion/passes/SplitReturnRewriter.java
@@ -43,6 +43,10 @@ public class SplitReturnRewriter extends CodeRewriterPass<AppInfo> {
@Override
protected CodeRewriterResult rewriteCode(IRCode code) {
+ if (code.context().toSourceString().equals(
+ "java.lang.Object androidx.collection.LruCache.remove(java.lang.Object)")) {
+ return CodeRewriterResult.NO_CHANGE;
+ }
int color = colorExceptionHandlers(code);
AffectedValues affectedValues = new AffectedValues();
boolean changed = false;
Comparing the DEX with and without this patch shows one instruction diff in void androidx.leanback.widget.GridLayoutManager$2.addItem(java.lang.Object, int, int, int, int)
where java.lang.Object androidx.collection.LruCache.remove(java.lang.Object)
is inlined into.
Tracing the IR for void androidx.leanback.widget.GridLayoutManager$2.addItem(java.lang.Object, int, int, int, int)
for the two versions shows that they start to deviate in second pass of IR after inserting assume instructions (SSA)
with
block 29, pred-counts: 1, succ-count: 2, filled: true, sealed: false
predecessors: 28
successors: 31 30 (no try/catch successors)
no phis
#143:loadView;1762:addItem: -1: InstanceGet v67 <- v136; field: androidx.collection.LruCache androidx.leanback.widget.ViewsStateBundle.mChildStates
: -1: Assume v140 <- v67; lower bound: @Nullable androidx.collection.LruCache {}
: -1: SafeCheckCast v131 <- v65; java.lang.String
: -1: Invoke-Virtual v132 <- v140, v131; method: java.lang.Object androidx.collection.LruCache.remove(java.lang.String)
: -1: CheckCast v69 <- v132; android.util.SparseArray
#144:loadView;1762:addItem: -1: If v69 EQZ block 31 (fallthrough 30)
vs.
block 29, pred-counts: 1, succ-count: 2, filled: true, sealed: false
predecessors: 28
successors: 31 30 (no try/catch successors)
no phis
#143:loadView;1762:addItem: -1: InstanceGet v67 <- v136; field: androidx.collection.LruCache androidx.leanback.widget.ViewsStateBundle.mChildStates
: -1: Assume v140 <- v67; lower bound: @Nullable androidx.collection.LruCache {}
: -1: SafeCheckCast v131 <- v65; java.lang.String
: -1: Invoke-Virtual v132 <- v140, v131; method: java.lang.Object androidx.collection.LruCache.remove(java.lang.String)
: -1: Assume v141 <- v132; not null
: -1: CheckCast v69 <- v141; android.util.SparseArray
#144:loadView;1762:addItem: -1: If v69 EQZ block 31 (fallthrough 30)
This is from computeDynamicReturnType
concluding a dynamic return type of DynamicTypeWithUpperBound(@NotNull java.lang.Object {})
.
Description
Steps to reproduce:
whyareyounotinlining
option in proguard configgradle.properties
:Sample project is attached.
Actual behavior
When R8 runs inside the main Gradle Daemon process, the system property works correctly and build finishes successfully. When R8 runs outside the process, the build failes with:
As I understand, the option
systemProp.com.android.tools.r8.experimental.enablewhyareyounotinlining=1
wasn't passed to forked process.Environment:
Gradle version: 8.10.2 AGP version: 8.7.2