https://github.com/AZero13 updated https://github.com/llvm/llvm-project/pull/207540
>From de8a23d5fd6e2f681d82974ab27bc31e2959364c Mon Sep 17 00:00:00 2001 From: AZero13 <[email protected]> Date: Sat, 4 Jul 2026 16:41:32 -0400 Subject: [PATCH] [Clang] Allow devirtualization involving array subscripts with constant indices when the pointee type is known By C++17 [expr.add]/6, if the array element type and the pointee type are not similar, behavior is undefined. --- clang/docs/ReleaseNotes.md | 6 +++ clang/lib/AST/DeclCXX.cpp | 32 ++++++++++++++ clang/test/CXX/drs/cwg15xx.cpp | 19 ++++++++ .../devirtualize-virtual-function-calls.cpp | 43 ++++++++++++++++++- clang/www/cxx_dr_status.html | 2 +- 5 files changed, 99 insertions(+), 3 deletions(-) diff --git a/clang/docs/ReleaseNotes.md b/clang/docs/ReleaseNotes.md index bbd42848a98c2..fedaeebf76520 100644 --- a/clang/docs/ReleaseNotes.md +++ b/clang/docs/ReleaseNotes.md @@ -275,6 +275,9 @@ latest release, please see the [Clang Web Site](https://clang.llvm.org) or the #### Resolutions to C++ Defect Reports +- Clang now implements [CWG1504](https://cplusplus.github.io/CWG/issues/1504.html) + which makes pointer arithmetic after derived-base conversion undefined behavior, + allowing devirtualization of calls through non-zero array subscripts. - Implemented [CWG1780 Explicit instantiation/specialization of generic lambda operator()](https://cplusplus.github.io/CWG/issues/1780.html) - Clang now allows omitting `typename` before a template name in a @@ -323,6 +326,9 @@ latest release, please see the [Clang Web Site](https://clang.llvm.org) or the ### Non-comprehensive list of changes in this release +- Clang can now devirtualize virtual function calls on objects accessed through + array subscript expressions with non-zero constant indices, based on the + resolution of [CWG1504](https://cplusplus.github.io/CWG/issues/1504.html). - Added support for floating point and pointer values in most `__atomic_` builtins. - Added `__builtin_stdc_rotate_left` and `__builtin_stdc_rotate_right` diff --git a/clang/lib/AST/DeclCXX.cpp b/clang/lib/AST/DeclCXX.cpp index 0573cdf95952a..b78ab4393cc7a 100644 --- a/clang/lib/AST/DeclCXX.cpp +++ b/clang/lib/AST/DeclCXX.cpp @@ -2600,6 +2600,38 @@ CXXMethodDecl *CXXMethodDecl::getDevirtualizedMethod(const Expr *Base, } } + // By CWG1504 / [expr.add]/6, pointer arithmetic on a base pointer into an + // array of derived objects is undefined behavior when the element type and + // pointee type are not similar. This means we can devirtualize calls on + // objects accessed through array subscripts or pointer arithmetic with + // non-zero offsets, since the dynamic type must match the static type. + // + // A single object is considered to be an array of one element, so p[0] + // could still be a derived object, but p[N] for N != 0 cannot. + const Expr *Inner = Base->IgnoreParenImpCasts(); + + // Handle p[N].f() (dot syntax with array subscript). + if (const auto *ASE = dyn_cast<ArraySubscriptExpr>(Inner)) { + Expr::EvalResult Result; + if (ASE->getIdx()->EvaluateAsInt(Result, getASTContext()) && + !Result.Val.getInt().isZero()) + return DevirtualizedMethod; + } + + // Handle (p + N)->f() (arrow syntax with pointer arithmetic). + if (const auto *BO = dyn_cast<BinaryOperator>(Inner)) { + if (BO->getOpcode() == BO_Add || BO->getOpcode() == BO_Sub) { + // Identify the integer operand (the offset). + const Expr *IdxExpr = BO->getLHS()->getType()->isPointerType() + ? BO->getRHS() + : BO->getLHS(); + Expr::EvalResult Result; + if (IdxExpr->EvaluateAsInt(Result, getASTContext()) && + !Result.Val.getInt().isZero()) + return DevirtualizedMethod; + } + } + // We can't devirtualize the call. return nullptr; } diff --git a/clang/test/CXX/drs/cwg15xx.cpp b/clang/test/CXX/drs/cwg15xx.cpp index 5a9b80ed028c4..53f262740eae2 100644 --- a/clang/test/CXX/drs/cwg15xx.cpp +++ b/clang/test/CXX/drs/cwg15xx.cpp @@ -11,6 +11,25 @@ // cxx98-error@-1 {{variadic macros are a C99 feature}} #endif +namespace cwg1504 { // cwg1504: 23 +#if __cplusplus >= 201103L + // CWG1504: Pointer arithmetic after derived-base conversion + struct Base { int x; }; + struct Derived : Base { int y; }; + constexpr Derived arr[2] = {}; + + // Pointer arithmetic on a base pointer into a derived array is UB, + // and the constexpr evaluator must diagnose it. + constexpr int test(int n) { + return ((const Base*)arr)[n].x; // #cwg1504-x + } + constexpr int bad = test(1); + // since-cxx11-error@-1 {{constexpr variable 'bad' must be initialized by a constant expression}} + // since-cxx11-note@#cwg1504-x {{cannot access field of pointer past the end of object}} + // since-cxx11-note@-3 {{in call to 'test(1)'}} +#endif +} // namespace cwg1504 + namespace cwg1512 { // cwg1512: 4 void f(char *p) { if (p > 0) {} diff --git a/clang/test/CodeGenCXX/devirtualize-virtual-function-calls.cpp b/clang/test/CodeGenCXX/devirtualize-virtual-function-calls.cpp index b50881db63e05..bfee72464176d 100644 --- a/clang/test/CodeGenCXX/devirtualize-virtual-function-calls.cpp +++ b/clang/test/CodeGenCXX/devirtualize-virtual-function-calls.cpp @@ -92,10 +92,49 @@ void fd(D d, XD xd, D *p) { // CHECK: call void % p[0].f(); - // FIXME: We can devirtualize this, by C++1z [expr.add]/6 (if the array + // We can devirtualize this, by CWG1504 / [expr.add]/6 (if the array // element type and the pointee type are not similar, behavior is undefined). - // CHECK: call void % + // CHECK: call void @_ZN1A1fEv p[1].f(); + + // Negative indices are also UB for the same reason. + // CHECK: call void @_ZN1A1fEv + p[-1].f(); + + // Pointer arithmetic with arrow syntax: (p + N)->f() + // CHECK: call void @_ZN1A1fEv + (p + 1)->f(); + + // Pointer subtraction with arrow syntax: (p - N)->f() + // CHECK: call void @_ZN1A1fEv + (p - 1)->f(); + + // Can't devirtualize with non-constant index; we can't prove N != 0. + int n = 1; + // CHECK: call void % + p[n].f(); + + // Zero through expression: p[1-1] evaluates to 0, can't devirtualize. + // CHECK: call void % + p[1-1].f(); + + // (p + 0)->f() also can't be devirtualized (same as *p). + // CHECK: call void % + (p + 0)->f(); + + // 1 + p is legal pointer arithmetic too. + // CHECK: call void @_ZN1A1fEv + (1 + p)->f(); + + // Constant variables are evaluated. + const int N = 1; + // CHECK: call void @_ZN1A1fEv + p[N].f(); + + // Pointer arithmetic with deref: *(p + 1) + // This is a UnaryOperator, which our current AST matching doesn't catch. + // CHECK: call void % + (*(p + 1)).f(); } struct B { diff --git a/clang/www/cxx_dr_status.html b/clang/www/cxx_dr_status.html index af91ac559d274..73c8be98e7848 100755 --- a/clang/www/cxx_dr_status.html +++ b/clang/www/cxx_dr_status.html @@ -10309,7 +10309,7 @@ <h2 id="cxxdr">C++ defect report implementation status</h2> <td>[<a href="https://wg21.link/expr.add";>expr.add</a>]</td> <td>CD3</td> <td>Pointer arithmetic after derived-base conversion</td> - <td class="unknown" align="center">Unknown</td> + <td class="full" align="center">Clang 23</td> </tr> <tr id="1505"> <td><a href="https://cplusplus.github.io/CWG/issues/1505.html";>1505</a></td> _______________________________________________ cfe-commits mailing list [email protected] https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
