[TBAA] Refine pointer-tbaa for void pointers by pointer depth #126047

brunodf-snps · 2025-02-06T11:01:24Z

Commit 77d3f8a avoids distinct tags for any pointers where the ultimate pointee type is void, to solve breakage in real-world code that uses (indirections to) void* for polymorphism over different pointer types.

While this matches the TBAA implementation in GCC, this patch implements a refinement that distinguishes void pointers by pointer depth, as described in the "strict aliasing" documentation included in the aforementioned commit:

void* is permitted to alias any pointer type, void** is permitted
to alias any pointer to pointer type, and so on.

For example, void** is no longer considered to alias int* in this refinement, but it remains possible to use void** for polymorphism over pointers to pointers.

Commit 77d3f8a avoids distinct tags for any pointers where the ultimate pointee type is `void`, to solve breakage in real-world code that uses (indirections to) `void*` for polymorphism over different pointer types. While this matches the TBAA implementation in GCC, this patch implements a refinement that distinguishes void pointers by pointer depth, as described in the "strict aliasing" documentation included in the aforementioned commit: > `void*` is permitted to alias any pointer type, `void**` is permitted > to alias any pointer to pointer type, and so on. For example, `void**` is no longer considered to alias `int*` in this refinement, but it remains possible to use `void**` for polymorphism over pointers to pointers.

llvmbot · 2025-02-06T11:01:56Z

@llvm/pr-subscribers-clang-codegen

@llvm/pr-subscribers-clang

Author: Bruno De Fraine (brunodf-snps)

Changes

Commit 77d3f8a avoids distinct tags for any pointers where the ultimate pointee type is void, to solve breakage in real-world code that uses (indirections to) void* for polymorphism over different pointer types.

While this matches the TBAA implementation in GCC, this patch implements a refinement that distinguishes void pointers by pointer depth, as described in the "strict aliasing" documentation included in the aforementioned commit:
> void* is permitted to alias any pointer type, void** is permitted
> to alias any pointer to pointer type, and so on.

For example, void** is no longer considered to alias int* in this refinement, but it remains possible to use void** for polymorphism over pointers to pointers.

Full diff: https://github.com/llvm/llvm-project/pull/126047.diff

4 Files Affected:

(modified) clang/lib/CodeGen/CodeGenTBAA.cpp (+24-6)
(modified) clang/lib/CodeGen/CodeGenTBAA.h (+5)
(modified) clang/test/CXX/drs/cwg158.cpp (+2-1)
(modified) clang/test/CodeGen/tbaa-pointers.c (+14-8)

diff --git a/clang/lib/CodeGen/CodeGenTBAA.cpp b/clang/lib/CodeGen/CodeGenTBAA.cpp
index 3f1a24791ddd816..3a294fd5980472a 100644
--- a/clang/lib/CodeGen/CodeGenTBAA.cpp
+++ b/clang/lib/CodeGen/CodeGenTBAA.cpp
@@ -80,6 +80,25 @@ llvm::MDNode *CodeGenTBAA::getChar() {
   return Char;
 }
 
+llvm::MDNode *CodeGenTBAA::getAnyPtr(unsigned PtrDepth) {
+  assert(PtrDepth >= 1 && "Pointer must have some depth");
+
+  if (AnyPtrs.size() < PtrDepth) {
+    AnyPtrs.reserve(PtrDepth);
+    auto Size = Module.getDataLayout().getPointerSize();
+    // Populate first element.
+    if (AnyPtrs.empty())
+      AnyPtrs.push_back(createScalarTypeNode("any pointer", getChar(), Size));
+    // Populate further elements.
+    for (size_t Idx = AnyPtrs.size(); Idx < PtrDepth; ++Idx) {
+      auto Name = ("any p" + llvm::Twine(Idx + 1) + " pointer").str();
+      AnyPtrs.push_back(createScalarTypeNode(Name, AnyPtrs[Idx - 1], Size));
+    }
+  }
+
+  return AnyPtrs[PtrDepth - 1];
+}
+
 static bool TypeHasMayAlias(QualType QTy) {
   // Tagged types have declarations, and therefore may have attributes.
   if (auto *TD = QTy->getAsTagDecl())
@@ -202,9 +221,8 @@ llvm::MDNode *CodeGenTBAA::getTypeInfoHelper(const Type *Ty) {
   // they involve a significant representation difference.  We don't
   // currently do so, however.
   if (Ty->isPointerType() || Ty->isReferenceType()) {
-    llvm::MDNode *AnyPtr = createScalarTypeNode("any pointer", getChar(), Size);
     if (!CodeGenOpts.PointerTBAA)
-      return AnyPtr;
+      return getAnyPtr();
     // C++ [basic.lval]p11 permits objects to accessed through an l-value of
     // similar type. Two types are similar under C++ [conv.qual]p2 if the
     // decomposition of the types into pointers, member pointers, and arrays has
@@ -232,7 +250,7 @@ llvm::MDNode *CodeGenTBAA::getTypeInfoHelper(const Type *Ty) {
     // common idioms and there is no good alternative to re-write the code
     // without strict-aliasing violations.
     if (Ty->isVoidType())
-      return AnyPtr;
+      return getAnyPtr(PtrDepth);
 
     assert(!isa<VariableArrayType>(Ty));
     // When the underlying type is a builtin type, we compute the pointee type
@@ -256,7 +274,7 @@ llvm::MDNode *CodeGenTBAA::getTypeInfoHelper(const Type *Ty) {
       // similar-types rule.
       const auto *RT = Ty->getAs<RecordType>();
       if (!RT)
-        return AnyPtr;
+        return getAnyPtr();
 
       // For unnamed structs or unions C's compatible types rule applies. Two
       // compatible types in different compilation units can have different
@@ -270,7 +288,7 @@ llvm::MDNode *CodeGenTBAA::getTypeInfoHelper(const Type *Ty) {
       // compatibility rule, but it doesn't matter because you can never have a
       // pointer to an anonymous struct or union.
       if (!RT->getDecl()->getDeclName())
-        return AnyPtr;
+        return getAnyPtr();
 
       // For non-builtin types use the mangled name of the canonical type.
       llvm::raw_svector_ostream TyOut(TyName);
@@ -281,7 +299,7 @@ llvm::MDNode *CodeGenTBAA::getTypeInfoHelper(const Type *Ty) {
     OutName += std::to_string(PtrDepth);
     OutName += " ";
     OutName += TyName;
-    return createScalarTypeNode(OutName, AnyPtr, Size);
+    return createScalarTypeNode(OutName, getAnyPtr(PtrDepth), Size);
   }
 
   // Accesses to arrays are accesses to objects of their element types.
diff --git a/clang/lib/CodeGen/CodeGenTBAA.h b/clang/lib/CodeGen/CodeGenTBAA.h
index ab3b05df7713ba6..0aae171d168df2a 100644
--- a/clang/lib/CodeGen/CodeGenTBAA.h
+++ b/clang/lib/CodeGen/CodeGenTBAA.h
@@ -139,6 +139,7 @@ class CodeGenTBAA {
 
   llvm::MDNode *Root;
   llvm::MDNode *Char;
+  llvm::SmallVector<llvm::MDNode *, 4> AnyPtrs;
 
   /// getRoot - This is the mdnode for the root of the metadata type graph
   /// for this translation unit.
@@ -148,6 +149,10 @@ class CodeGenTBAA {
   /// considered to be equivalent to it.
   llvm::MDNode *getChar();
 
+  /// getAnyPtr - This is the mdnode for any pointer type of (at least) the
+  /// given pointer depth.
+  llvm::MDNode *getAnyPtr(unsigned PtrDepth = 1);
+
   /// CollectFields - Collect information about the fields of a type for
   /// !tbaa.struct metadata formation. Return false for an unsupported type.
   bool CollectFields(uint64_t BaseOffset,
diff --git a/clang/test/CXX/drs/cwg158.cpp b/clang/test/CXX/drs/cwg158.cpp
index 1f5c319e6bd25fc..a083ded82bb2e06 100644
--- a/clang/test/CXX/drs/cwg158.cpp
+++ b/clang/test/CXX/drs/cwg158.cpp
@@ -42,5 +42,6 @@ const int * h(const int * (*p)[10],  int *(*q)[9]) {
 }
 
 // POINTER-TBAA: [[PTRARRAY_TBAA]] = !{[[PTRARRAY_TY:!.+]], [[PTRARRAY_TY]], i64 0}
-// POINTER-TBAA: [[PTRARRAY_TY]] = !{!"p2 int", [[ANYPTR:!.+]], i64 0}
+// POINTER-TBAA: [[PTRARRAY_TY]] = !{!"p2 int", [[ANYP2PTR:!.+]], i64 0}
+// POINTER-TBAA: [[ANYP2PTR]] = !{!"any p2 pointer", [[ANYPTR:!.+]],
 // POINTER-TBAA: [[ANYPTR]] = !{!"any pointer"
diff --git a/clang/test/CodeGen/tbaa-pointers.c b/clang/test/CodeGen/tbaa-pointers.c
index 48adac503357f00..9cfaa0a47af6e75 100644
--- a/clang/test/CodeGen/tbaa-pointers.c
+++ b/clang/test/CodeGen/tbaa-pointers.c
@@ -208,8 +208,10 @@ int void_ptrs(void **ptr) {
 // COMMON-LABEL: define i32 @void_ptrs(
 // COMMON-SAME: ptr noundef [[PTRA:%.+]])
 // COMMON:        [[PTR_ADDR:%.+]]  = alloca ptr, align 8
-// COMMON-NEXT:   store ptr [[PTRA]], ptr [[PTR_ADDR]], align 8, !tbaa [[ANYPTR]]
-// COMMON-NEXT:   [[L0:%.+]] = load ptr, ptr  [[PTR_ADDR]], align 8, !tbaa  [[ANYPTR]]
+// DISABLE-NEXT:  store ptr [[PTRA]], ptr [[PTR_ADDR]], align 8, !tbaa [[ANYPTR]]
+// DEFAULT-NEXT:  store ptr [[PTRA]], ptr [[PTR_ADDR]], align 8, !tbaa [[ANYP2:!.+]]
+// DISABLE-NEXT:  [[L0:%.+]] = load ptr, ptr [[PTR_ADDR]], align 8, !tbaa [[ANYPTR]]
+// DEFAULT-NEXT:  [[L0:%.+]] = load ptr, ptr [[PTR_ADDR]], align 8, !tbaa [[ANYP2]]
 // COMMON-NEXT:   [[L1:%.+]] = load ptr, ptr [[L0]], align 8, !tbaa  [[ANYPTR]]
 // COMMON-NEXT:   [[BOOL:%.+]] = icmp ne ptr [[L1]], null
 // COMMON-NEXT:   [[BOOL_EXT:%.+]] = zext i1 [[BOOL]] to i64
@@ -220,7 +222,8 @@ int void_ptrs(void **ptr) {
 }
 
 // DEFAULT: [[P2INT_0]] = !{[[P2INT:!.+]], [[P2INT]], i64 0}
-// DEFAULT: [[P2INT]] = !{!"p2 int", [[ANY_POINTER:!.+]], i64 0}
+// DEFAULT: [[P2INT]] = !{!"p2 int", [[ANY_P2_POINTER:!.+]], i64 0}
+// DEFAULT: [[ANY_P2_POINTER]] = !{!"any p2 pointer", [[ANY_POINTER:!.+]], i64 0}
 // DISABLE: [[ANYPTR]] = !{[[ANY_POINTER:!.+]], [[ANY_POINTER]], i64 0}
 // COMMON: [[ANY_POINTER]] = !{!"any pointer", [[CHAR:!.+]], i64 0}
 // COMMON: [[CHAR]] = !{!"omnipotent char", [[TBAA_ROOT:!.+]], i64 0}
@@ -228,17 +231,19 @@ int void_ptrs(void **ptr) {
 // DEFAULT: [[P1INT_0]] = !{[[P1INT:!.+]], [[P1INT]], i64 0}
 // DEFAULT: [[P1INT]] = !{!"p1 int", [[ANY_POINTER]], i64 0}
 // DEFAULT: [[P3INT_0]] = !{[[P3INT:!.+]], [[P3INT]], i64 0}
-// DEFAULT: [[P3INT]] = !{!"p3 int", [[ANY_POINTER]], i64 0}
+// DEFAULT: [[P3INT]] = !{!"p3 int", [[ANY_P3_POINTER:!.+]], i64 0}
+// DEFAULT: [[ANY_P3_POINTER]] = !{!"any p3 pointer", [[ANY_P2_POINTER]], i64 0}
 // DEFAULT: [[P4CHAR_0]] = !{[[P4CHAR:!.+]], [[P4CHAR]], i64 0}
-// DEFAULT: [[P4CHAR]] = !{!"p4 omnipotent char", [[ANY_POINTER]], i64 0}
+// DEFAULT: [[P4CHAR]] = !{!"p4 omnipotent char", [[ANY_P4_POINTER:!.*]], i64 0}
+// DEFAULT: [[ANY_P4_POINTER]] = !{!"any p4 pointer", [[ANY_P3_POINTER]], i64 0}
 // DEFAULT: [[P3CHAR_0]] = !{[[P3CHAR:!.+]], [[P3CHAR]], i64 0}
-// DEFAULT: [[P3CHAR]] = !{!"p3 omnipotent char", [[ANY_POINTER]], i64 0}
+// DEFAULT: [[P3CHAR]] = !{!"p3 omnipotent char", [[ANY_P3_POINTER]], i64 0}
 // DEFAULT: [[P2CHAR_0]] = !{[[P2CHAR:!.+]], [[P2CHAR]], i64 0}
-// DEFAULT: [[P2CHAR]] = !{!"p2 omnipotent char", [[ANY_POINTER]], i64 0}
+// DEFAULT: [[P2CHAR]] = !{!"p2 omnipotent char", [[ANY_P2_POINTER]], i64 0}
 // DEFAULT: [[P1CHAR_0]] = !{[[P1CHAR:!.+]], [[P1CHAR]], i64 0}
 // DEFAULT: [[P1CHAR]] = !{!"p1 omnipotent char", [[ANY_POINTER]], i64 0}
 // DEFAULT: [[P2S1_TAG]] = !{[[P2S1:!.+]], [[P2S1]], i64 0}
-// DEFAULT: [[P2S1]] = !{!"p2 _ZTS2S1", [[ANY_POINTER]], i64 0}
+// DEFAULT: [[P2S1]] = !{!"p2 _ZTS2S1", [[ANY_P2_POINTER]], i64 0}
 // DEFAULT: [[P1S1_TAG:!.+]] = !{[[P1S1:!.+]], [[P1S1]], i64 0}
 // DEFAULT: [[P1S1]] = !{!"p1 _ZTS2S1", [[ANY_POINTER]], i64 0}
 // DEFAULT: [[P1S2_TAG]] = !{[[P1S2:!.+]], [[P1S2]], i64 0}
@@ -251,3 +256,4 @@ int void_ptrs(void **ptr) {
 // COMMON:  [[INT_TAG]] = !{[[INT_TY:!.+]], [[INT_TY]], i64 0}
 // COMMON:  [[INT_TY]] = !{!"int", [[CHAR]], i64 0}
 // DEFAULT: [[ANYPTR]] = !{[[ANY_POINTER]],  [[ANY_POINTER]], i64 0}
+// DEFAULT: [[ANYP2]] = !{[[ANY_P2_POINTER]],  [[ANY_P2_POINTER]], i64 0}

Endilll

I'm not qualified to review this, so please make sure you understand changes to CWG158 test and they are what you expect. Specifically, I'd like to avoid updating the test just to make FileCheck happy.

brunodf-snps · 2025-02-06T21:39:52Z

See discussion with @rjmccall in #122116 for context.

clang/lib/CodeGen/CodeGenTBAA.cpp

…e cases

rjmccall

Functionality LGTM; just some suggested tweaks to the comments.

clang/lib/CodeGen/CodeGenTBAA.cpp

…n suggestion

brunodf-snps · 2025-02-11T14:33:09Z

I'm not qualified to review this, so please make sure you understand changes to CWG158 test and they are what you expect. Specifically, I'd like to avoid updating the test just to make FileCheck happy.

All the updates to FileCheck patterns match what I expected. The affected test case of cwg158.cpp was added by @fhahn in commit decb874, I hope he will still chime in.

rjmccall

Thanks, the new comment looks great. LGTM.

brunodf-snps · 2025-02-17T13:49:18Z

If no other remarks get raised, I think this can be merged in a day or two? (Will have to be done by someone with commit access.)

Endilll · 2025-02-17T14:31:55Z

@brunodf-snps A reminder at the moment when you think this should be merged would be appreciated.

brunodf-snps · 2025-02-20T09:07:24Z

@Endilll @rjmccall OK to merge.

…s-O0-O2-opt-disabled.c` The test failed because of upstream changes to TBAA ``` commit 458b1e9 error: cannot run gpg: No such file or directory Author: Bruno De Fraine <[email protected]> Date: Thu Feb 20 13:06:56 2025 +0100 [TBAA] Refine pointer-tbaa for void pointers by pointer depth (llvm#126047) ``` rdar://145309621

brunodf-snps requested a review from Endilll as a code owner February 6, 2025 11:01

llvmbot added clang Clang issues not falling into any other category clang:codegen IR generation bugs: mangling, exceptions, etc. labels Feb 6, 2025

brunodf-snps mentioned this pull request Feb 6, 2025

[TBAA] Don't emit pointer-tbaa for void pointers. #122116

Merged

Endilll reviewed Feb 6, 2025

View reviewed changes

fhahn requested review from fhahn, rjmccall, AaronBallman and thesamesam February 6, 2025 20:21

CodeGenTBAA: getAnyPtr: add extra comment describing AnyPtrs

aa4ccd6

rjmccall reviewed Feb 7, 2025

View reviewed changes

clang/lib/CodeGen/CodeGenTBAA.cpp Outdated Show resolved Hide resolved

CodeGenTBAA: getTypeInfoHelper: also use pointer depth in conservativ…

8068c5b

…e cases

rjmccall reviewed Feb 10, 2025

View reviewed changes

clang/lib/CodeGen/CodeGenTBAA.cpp Show resolved Hide resolved

CodeGenTBAA: getAnyPtr: further extend description in comment based o…

ffa0edc

…n suggestion

rjmccall approved these changes Feb 11, 2025

View reviewed changes

Endilll merged commit 458b1e9 into llvm:main Feb 20, 2025
8 checks passed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[TBAA] Refine pointer-tbaa for void pointers by pointer depth #126047

[TBAA] Refine pointer-tbaa for void pointers by pointer depth #126047

Uh oh!

brunodf-snps commented Feb 6, 2025

Uh oh!

llvmbot commented Feb 6, 2025 •

edited

Loading

Uh oh!

Endilll left a comment

Uh oh!

brunodf-snps commented Feb 6, 2025

Uh oh!

Uh oh!

rjmccall left a comment

Uh oh!

Uh oh!

brunodf-snps commented Feb 11, 2025

Uh oh!

rjmccall left a comment

Uh oh!

brunodf-snps commented Feb 17, 2025

Uh oh!

Endilll commented Feb 17, 2025

Uh oh!

brunodf-snps commented Feb 20, 2025

Uh oh!

Uh oh!

Uh oh!

[TBAA] Refine pointer-tbaa for void pointers by pointer depth #126047

[TBAA] Refine pointer-tbaa for void pointers by pointer depth #126047

Uh oh!

Conversation

brunodf-snps commented Feb 6, 2025

Uh oh!

llvmbot commented Feb 6, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Endilll left a comment

Choose a reason for hiding this comment

Uh oh!

brunodf-snps commented Feb 6, 2025

Uh oh!

Uh oh!

rjmccall left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

brunodf-snps commented Feb 11, 2025

Uh oh!

rjmccall left a comment

Choose a reason for hiding this comment

Uh oh!

brunodf-snps commented Feb 17, 2025

Uh oh!

Endilll commented Feb 17, 2025

Uh oh!

brunodf-snps commented Feb 20, 2025

Uh oh!

Uh oh!

Uh oh!

llvmbot commented Feb 6, 2025 •

edited

Loading