mirror of
https://github.com/llvm/llvm-project.git
synced 2025-04-25 23:46:05 +00:00
7e5821bae8
This is the second attempt to bring initial support for [[assume()]] in the Clang Static Analyzer. The first attempt (#116462) was reverted in 2b9abf0db2d106c7208b4372e662ef5df869e6f1 due to some weird failure in a libcxx test involving `#pragma clang loop vectorize(enable) interleave(enable)`. The failure could be reduced into: ```c++ template <class ExecutionPolicy> void transform(ExecutionPolicy) { #pragma clang loop vectorize(enable) interleave(enable) for (int i = 0; 0;) { // The DeclStmt of "i" would be added twice in the ThreadSafety analysis. // empty } } void entrypoint() { transform(1); } ``` As it turns out, the problem with the initial patch was this: ```c++ for (const auto *Attr : AS->getAttrs()) { if (const auto *AssumeAttr = dyn_cast<CXXAssumeAttr>(Attr)) { Expr *AssumeExpr = AssumeAttr->getAssumption(); if (!AssumeExpr->HasSideEffects(Ctx)) { childrenBuf.push_back(AssumeExpr); } } // Visit the actual children AST nodes. // For CXXAssumeAttrs, this is always a NullStmt. llvm::append_range(childrenBuf, AS->children()); // <--- This was not meant to be part of the "for" loop. children = childrenBuf; } return; ``` The solution was simple. Just hoist it from the loop. I also had a closer look at `CFGBuilder::VisitAttributedStmt`, where I also spotted another bug. We would have added the CFG blocks twice if the AttributedStmt would have both the `[[fallthrough]]` and the `[[assume()]]` attributes. With my fix, it will only once add the blocks. Added a regression test for this. Co-authored-by: Vinay Deshmukh <vinay_deshmukh AT outlook DOT com>