https://github.com/marco-milanese-sonarsource created 
https://github.com/llvm/llvm-project/pull/219225

This PR fixes a false positive on `UnreachableCodeChecker` occurring when STU 
exploration exhausts the steps budget and the CTU second phase completes its 
exploration (i.e., has no more work to do).

There are two code pieces to consider:

* When we transition from STU exploration to CTU exploration 
([`CoreEngine.h:173`](https://github.com/llvm/llvm-project/blob/main/clang/lib/StaticAnalyzer/Core/CoreEngine.cpp#L173))
 we move `CTUWList` to `WList`:
 
```c++
  if (CTUWList) {
    NumSTUSteps += STUSteps;
    const unsigned MinCTUSteps =
        this->ExprEng.getAnalysisManager().options.CTUMaxNodesMin;
    const unsigned Pct =
        this->ExprEng.getAnalysisManager().options.CTUMaxNodesPercentage;
    unsigned MaxCTUSteps = std::max(STUSteps * Pct / 100, MinCTUSteps);

    WList = std::move(CTUWList); // remaining STU work gets discarded
    const unsigned CTUSteps = ProcessWList(MaxCTUSteps);
    NumCTUSteps += CTUSteps;
  }

```

* At the end of the analysis `UnreachableCodeChecker` checks for code that was 
not explored. This is corresponds to actual dead code only as long as the 
exploration is complete, thus (correctly) the checker calls `hasWorkRemaining` 
([`CoreEngine.h:149`](https://github.com/llvm/llvm-project/blob/main/clang/include/clang/StaticAnalyzer/Core/PathSensitive/CoreEngine.h#L149-L151)).
 If such predicate is true it immediately bails out as any finding may be due 
to the uncomplete exploration.

* `hasWorkRemaining` relies on `WList` being empty or not to determine if there 
is work remaining (i.e., complete/uncomplete exploration):
```c++
  bool hasWorkRemaining() const { return wasBlocksExhausted() ||
                                         WList->hasWork() ||
                                         wasBlockAborted(); }
```

Now, consider the scenario where:

1. STU exploration was *not* exhaustive (i.e., some code was not analyzed 
because we did not have enough steps budget).
2. CTU exploration was exhaustive.

In this case `hasWorkRemaining` should return `true` as STU exploration had 
more work to do, but because we override it with the CTU worklist (and the 
latter is fully emptied), then `hasWorkRemaining` will return `false`, raising 
the false positive.

This can be simulated in the regression test by setting a very low STU budget 
and enabling CTU mode. Actually we don't 
need to do any CTU work, supply extra TUs; this is because the switch to the 
CTU worklist (empty) is always done when CTU mode is enabled.

>From 59f08b836ae562a43939d8317b5855bafed5e2c0 Mon Sep 17 00:00:00 2001
From: Marco Milanese <[email protected]>
Date: Thu, 27 Aug 2026 16:42:47 +0200
Subject: [PATCH] Add regression test

---
 clang/test/Analysis/ctu/stu-workremaining.cpp | 21 +++++++++++++++++++
 1 file changed, 21 insertions(+)
 create mode 100644 clang/test/Analysis/ctu/stu-workremaining.cpp

diff --git a/clang/test/Analysis/ctu/stu-workremaining.cpp 
b/clang/test/Analysis/ctu/stu-workremaining.cpp
new file mode 100644
index 0000000000000..1524a98ec9a93
--- /dev/null
+++ b/clang/test/Analysis/ctu/stu-workremaining.cpp
@@ -0,0 +1,21 @@
+// RUN: %clang_analyze_cc1 -std=c++20 \
+// RUN:   -analyzer-checker=core,alpha.deadcode.UnreachableCode \
+// RUN:   -analyzer-config experimental-enable-naive-ctu-analysis=true \
+// RUN:   -analyzer-config max-nodes=10 \
+// RUN:   -verify=ctu-on %s
+// ctu-on-no-diagnostics
+
+// RUN: %clang_analyze_cc1 -std=c++20 \
+// RUN:   -analyzer-checker=core,alpha.deadcode.UnreachableCode \
+// RUN:   -analyzer-config experimental-enable-naive-ctu-analysis=false \
+// RUN:   -analyzer-config max-nodes=10 \
+// RUN:   -verify=ctu-off %s
+// ctu-off-no-diagnostics
+
+#define NOP ((void)0)
+
+void entrypoint(int x) {
+  NOP; NOP; NOP; NOP; NOP;
+  NOP; NOP; NOP; NOP; NOP;
+  if (x) NOP;
+}

_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits

Reply via email to