Bug ID: JDK-8362117 C2: compiler/stringopts/TestStackedConcatsAppendUncommonTrap.java fails with a wrong result

Versions (Unresolved/Resolved/Fixed)

The Version table provides details related to the release that this issue/RFE will be addressed.

Unresolved : Release in which this issue/RFE will be addressed.
Resolved: Release in which this issue/RFE has been resolved.
Fixed : Release in which this issue/RFE has been fixed. The release containing this fix may be available for download as an Early Access Release or a General Availability Release.

To download the current JDK release, click here.

JDK 26
26Unresolved

compiler/stringopts/TestStackedConcatsAppendUncommonTrap.java fails with a wrong result after stacked string concatenation optimizations when run with stress flags.

-Xcomp -XX:CompileThreshold=100 -XX:-TieredCompilation -XX:+DeoptimizeALot

These flags inline 'f' into the main method which is OSR compiled in the failure case (and maybe they avoid the uncommon trap that JDK-8357105 was caused by).

This bug is hidden by JDK-8273612 since JDK 18, but to my understanding, that fix should not affect correctness, only inlining decisions for the -Xcomp mode.

===

The test runs fine before JDK-8271341, then starts crashing until JDK-8291775, and after that it produces the wrong result. The failure mode looks similar to JDK-7179138 since in both cases a result with the wrong length is produced.

JDK-8291775 changed the folding of a diamond If-pattern "Region -> (True, False) -> If" by setting the If-test input to a constant zero instead of replacing the Region by the If's input, to prevent a too quick propagation of top() through the graph. One assumption behind that was that the data Phis of the Region would be removed during PhaseRemoveUseless. However, in the attached test case, that does not apply. The data phi is retained until IGVN where the diamond if is first collapsed, then the Phi is transformed as it's replaced by the input from the corresponding "False" branch, which gives the wrong result.
04-08-2025
Reproducer that requires less extra flags: https://github.com/openjdk/jdk/compare/master...danielogh:jdk:stringopts-wrong-result?expand=1 -Xcomp -XX:-TieredCompilation -XX:CompileOnly=compiler.stringopts.TestStackedConcatsAppendUncommonTrap::f -XX:+OptimizeStringConcat === Attached [35_before_stringopts.png] from a run of this test. It looks like the last append is picking up "xx" in the expression (s == "xx") ? s : "aa" and adding it to the string.
24-07-2025
Seems to affect 8 too. It just silently bails out though. EDIT. or maybe we just don't have JDK-8271341 ?
15-07-2025
Thank you!
14-07-2025
I can give it a try.
14-07-2025
[~dskantz] do you plan to look into this or should we assign to someone else? Thanks for the report!
14-07-2025
ILW = Incorrect result of C2 compiled program (string concat), with string concat and stress options, -XX:-OptimizeStringConcat = HLM = P3
14-07-2025
Workaround: -XX:-OptimizeStringConcat
14-07-2025
Another bug with stacked concatenations and a complex merge point between the pair of string builders.
14-07-2025

Causes :	JDK-8291775 - C2: assert(r != __null && r->is_Region()) failed: this phi must have a region
Relates :	JDK-8271341 - Opcode() != Op_If && Opcode() != Op_RangeCheck) \|\| outcnt() == 2 assert failure with Test7179138_1.java
Relates :	JDK-8362118 - C2: add more debug prints to string concatenation optimizations
Relates :	JDK-8357105 - C2: compilation fails with "assert(false) failed: empty program detected during loop optimization"
Relates :	JDK-8273612 - Fix for JDK-8272873 causes timeout in running some tests with -Xcomp