[CodeExtractor] Resolving the Inconsistency of Compiled Binary Files

[dongjianqiang2]

When the compiler enables ALSR by default, binary inconsistency occurs when the extractCodeRegion function is called. The reason is that the sequence of traversing ExitBlocks containers of the SmallPtrSet type is changed due to randomization of BasicBlock pointers.
This fixes #86427

[llvmbot]

@llvm/pr-subscribers-llvm-transforms

Author: dong jianqiang (dongjianqiang2)

Changes

When the compiler enables ALSR by default, binary inconsistency occurs when the extractCodeRegion function is called. The reason is that the sequence of traversing ExitBlocks containers of the SmallPtrSet type is changed due to randomization of BasicBlock pointers.
This fixes #86427

---

Full diff: https://github.com/llvm/llvm-project/pull/86497.diff

2 Files Affected:

• (modified) llvm/include/llvm/Transforms/Utils/CodeExtractor.h (+1-1)
• (modified) llvm/lib/Transforms/Utils/CodeExtractor.cpp (+2-2)

diff --git a/llvm/include/llvm/Transforms/Utils/CodeExtractor.h b/llvm/include/llvm/Transforms/Utils/CodeExtractor.h
index 27b34ef023db72..333ed6774d6c7e 100644
--- a/llvm/include/llvm/Transforms/Utils/CodeExtractor.h
+++ b/llvm/include/llvm/Transforms/Utils/CodeExtractor.h
@@ -249,7 +249,7 @@ class CodeExtractorAnalysisCache {
                        Instruction *Addr, BasicBlock *ExitBlock) const;
 
     void severSplitPHINodesOfEntry(BasicBlock *&Header);
-    void severSplitPHINodesOfExits(const SmallPtrSetImpl<BasicBlock *> &Exits);
+    void severSplitPHINodesOfExits(const SetVector<BasicBlock *> &Exits);
     void splitReturnBlocks();
 
     Function *constructFunction(const ValueSet &inputs,
diff --git a/llvm/lib/Transforms/Utils/CodeExtractor.cpp b/llvm/lib/Transforms/Utils/CodeExtractor.cpp
index 3191751d92e176..6988292ac71561 100644
--- a/llvm/lib/Transforms/Utils/CodeExtractor.cpp
+++ b/llvm/lib/Transforms/Utils/CodeExtractor.cpp
@@ -745,7 +745,7 @@ void CodeExtractor::severSplitPHINodesOfEntry(BasicBlock *&Header) {
 /// and other with remaining incoming blocks; then first PHIs are placed in
 /// outlined region.
 void CodeExtractor::severSplitPHINodesOfExits(
-    const SmallPtrSetImpl<BasicBlock *> &Exits) {
+    const SetVector<BasicBlock *> &Exits) {
   for (BasicBlock *ExitBB : Exits) {
     BasicBlock *NewBB = nullptr;
 
@@ -1751,7 +1751,7 @@ CodeExtractor::extractCodeRegion(const CodeExtractorAnalysisCache &CEAC,
   // Calculate the exit blocks for the extracted region and the total exit
   // weights for each of those blocks.
   DenseMap<BasicBlock *, BlockFrequency> ExitWeights;
-  SmallPtrSet<BasicBlock *, 1> ExitBlocks;
+  SetVector<BasicBlock *> ExitBlocks;
   for (BasicBlock *Block : Blocks) {
     for (BasicBlock *Succ : successors(Block)) {
       if (!Blocks.count(Succ)) {

[github-actions]

✅ With the latest revision this PR passed the Python code formatter.

[github-actions]

✅ With the latest revision this PR passed the C/C++ code formatter.

[adrian-prantl]

Makes sense, we can't iterate over the elements if it affects the output. And this shouldn't even be slower.

[hstk30-hw]

https://llvm.org/docs/CodingStandards.html#beware-of-non-determinism-due-to-ordering-of-pointers

Say beware of non-determinism!

I see we use SmallPtrSet in many pass,

https://github.com/search?q=repo%3Allvm%2Fllvm-project+SmallPtrSet+path%3A%2F%5Ellvm%5C%2Flib%5C%2FTransforms%5C%2F%2F&type=code

maybe have same problem.