Skip to content

Commit 4334f31

Browse files
fhahnfhahn
authored
[TBAA] Extend pointer TBAA to pointers of non-builtin types. (#110569)
Extend the logic added in 123c036 (#76612) to support pointers to non-builtin types by using the mangled name of the canonical type. PR: #110569
1 parent 10f0c1a commit 4334f31

File tree

4 files changed

+104
-36
lines changed

4 files changed

+104
-36
lines changed

‎clang/lib/CodeGen/CodeGenTBAA.cpp‎

Lines changed: 24 additions & 18 deletions
Original file line numberDiff line numberDiff line change
@@ -39,8 +39,9 @@ CodeGenTBAA::CodeGenTBAA(ASTContext &Ctx, CodeGenTypes &CGTypes,
3939
llvm::Module &M, const CodeGenOptions &CGO,
4040
const LangOptions &Features)
4141
: Context(Ctx), CGTypes(CGTypes), Module(M), CodeGenOpts(CGO),
42-
Features(Features), MDHelper(M.getContext()), Root(nullptr),
43-
Char(nullptr) {}
42+
Features(Features),
43+
MangleCtx(ItaniumMangleContext::create(Ctx, Ctx.getDiagnostics())),
44+
MDHelper(M.getContext()), Root(nullptr), Char(nullptr) {}
4445

4546
CodeGenTBAA::~CodeGenTBAA() {
4647
}
@@ -202,14 +203,6 @@ llvm::MDNode *CodeGenTBAA::getTypeInfoHelper(const Type *Ty) {
202203
// Other qualifiers could theoretically be distinguished, especially if
203204
// they involve a significant representation difference. We don't
204205
// currently do so, however.
205-
//
206-
// Computing the pointee type string recursively is implicitly more
207-
// forgiving than the standards require. Effectively, we are turning
208-
// the question "are these types compatible/similar" into "are
209-
// accesses to these types allowed to alias". In both C and C++,
210-
// the latter question has special carve-outs for signedness
211-
// mismatches that only apply at the top level. As a result, we are
212-
// allowing e.g. `int *` l-values to access `unsigned *` objects.
213206
if (Ty->isPointerType() || Ty->isReferenceType()) {
214207
llvm::MDNode *AnyPtr = createScalarTypeNode("any pointer", getChar(), Size);
215208
if (!CodeGenOpts.PointerTBAA)
@@ -221,21 +214,34 @@ llvm::MDNode *CodeGenTBAA::getTypeInfoHelper(const Type *Ty) {
221214
PtrDepth++;
222215
Ty = Ty->getPointeeType().getTypePtr();
223216
} while (Ty->isPointerType());
224-
// TODO: Implement C++'s type "similarity" and consider dis-"similar"
225-
// pointers distinct for non-builtin types.
217+
Ty = Context.getBaseElementType(QualType(Ty, 0)).getTypePtr();
218+
assert(!isa<VariableArrayType>(Ty));
219+
// When the underlying type is a builtin type, we compute the pointee type
220+
// string recursively, which is implicitly more forgiving than the standards
221+
// require. Effectively, we are turning the question "are these types
222+
// compatible/similar" into "are accesses to these types allowed to alias".
223+
// In both C and C++, the latter question has special carve-outs for
224+
// signedness mismatches that only apply at the top level. As a result, we
225+
// are allowing e.g. `int *` l-values to access `unsigned *` objects.
226+
SmallString<256> TyName;
226227
if (isa<BuiltinType>(Ty)) {
227228
llvm::MDNode *ScalarMD = getTypeInfoHelper(Ty);
228229
StringRef Name =
229230
cast<llvm::MDString>(
230231
ScalarMD->getOperand(CodeGenOpts.NewStructPathTBAA ? 2 : 0))
231232
->getString();
232-
SmallString<256> OutName("p");
233-
OutName += std::to_string(PtrDepth);
234-
OutName += " ";
235-
OutName += Name;
236-
return createScalarTypeNode(OutName, AnyPtr, Size);
233+
TyName = Name;
234+
} else {
235+
// For non-builtin types use the mangled name of the canonical type.
236+
llvm::raw_svector_ostream TyOut(TyName);
237+
MangleCtx->mangleCanonicalTypeName(QualType(Ty, 0), TyOut);
237238
}
238-
return AnyPtr;
239+
240+
SmallString<256> OutName("p");
241+
OutName += std::to_string(PtrDepth);
242+
OutName += " ";
243+
OutName += TyName;
244+
return createScalarTypeNode(OutName, AnyPtr, Size);
239245
}
240246

241247
// Accesses to arrays are accesses to objects of their element types.

‎clang/lib/CodeGen/CodeGenTBAA.h‎

Lines changed: 2 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -24,6 +24,7 @@ namespace clang {
2424
class ASTContext;
2525
class CodeGenOptions;
2626
class LangOptions;
27+
class MangleContext;
2728
class QualType;
2829
class Type;
2930

@@ -119,6 +120,7 @@ class CodeGenTBAA {
119120
llvm::Module &Module;
120121
const CodeGenOptions &CodeGenOpts;
121122
const LangOptions &Features;
123+
std::unique_ptr<MangleContext> MangleCtx;
122124

123125
// MDHelper - Helper for creating metadata.
124126
llvm::MDBuilder MDHelper;

‎clang/test/CodeGen/tbaa-pointers.c‎

Lines changed: 55 additions & 11 deletions
Original file line numberDiff line numberDiff line change
@@ -116,10 +116,12 @@ void p2struct(struct S1 **ptr) {
116116
// COMMON-LABEL: define void @p2struct(
117117
// COMMON-SAME: ptr noundef [[PTR:%.+]])
118118
// COMMON: [[PTR_ADDR:%.+]] = alloca ptr, align 8
119-
// ENABLED-NEXT: store ptr [[PTR]], ptr [[PTR_ADDR]], align 8, !tbaa [[ANYPTR:!.+]]
119+
// ENABLED-NEXT: store ptr [[PTR]], ptr [[PTR_ADDR]], align 8, !tbaa [[P2S1_TAG:!.+]]
120120
// DEFAULT-NEXT: store ptr [[PTR]], ptr [[PTR_ADDR]], align 8, !tbaa [[ANYPTR]]
121-
// COMMON-NEXT: [[BASE:%.+]] = load ptr, ptr [[PTR_ADDR]], align 8, !tbaa [[ANYPTR]]
122-
// COMMON-NEXT: store ptr null, ptr [[BASE]], align 8, !tbaa [[ANYPTR]]
121+
// ENABLED-NEXT: [[BASE:%.+]] = load ptr, ptr [[PTR_ADDR]], align 8, !tbaa [[P2S1_TAG]]
122+
// DEFAULT-NEXT: [[BASE:%.+]] = load ptr, ptr [[PTR_ADDR]], align 8, !tbaa [[ANYPTR]]
123+
// ENABLED-NEXT: store ptr null, ptr [[BASE]], align 8, !tbaa [[P1S1_TAG:!.+]]
124+
// DEFAULT-NEXT: store ptr null, ptr [[BASE]], align 8, !tbaa [[ANYPTR]]
123125
// COMMON-NEXT: ret void
124126
//
125127
*ptr = 0;
@@ -129,9 +131,10 @@ void p2struct_const(struct S1 const **ptr) {
129131
// COMMON-LABEL: define void @p2struct_const(
130132
// COMMON-SAME: ptr noundef [[PTR:%.+]])
131133
// COMMON: [[PTR_ADDR:%.+]] = alloca ptr, align 8
132-
// COMMON-NEXT: store ptr [[PTR]], ptr [[PTR_ADDR]], align 8, !tbaa [[ANYPTR]]
134+
// COMMON-NEXT: store ptr [[PTR]], ptr [[PTR_ADDR]], align 8, !tbaa [[ANYPTR:!.+]]
133135
// COMMON-NEXT: [[BASE:%.+]] = load ptr, ptr [[PTR_ADDR]], align 8, !tbaa [[ANYPTR]]
134-
// COMMON-NEXT: store ptr null, ptr [[BASE]], align 8, !tbaa [[ANYPTR]]
136+
// ENABLED-NEXT: store ptr null, ptr [[BASE]], align 8, !tbaa [[P1S1_TAG]]
137+
// DEFAULT-NEXT: store ptr null, ptr [[BASE]], align 8, !tbaa [[ANYPTR]]
135138
// COMMON-NEXT: ret void
136139
//
137140
*ptr = 0;
@@ -145,14 +148,44 @@ void p2struct2(struct S2 *ptr) {
145148
// COMMON-LABEL: define void @p2struct2(
146149
// COMMON-SAME: ptr noundef [[PTR:%.+]])
147150
// COMMON: [[PTR_ADDR:%.+]] = alloca ptr, align 8
148-
// COMMON-NEXT: store ptr [[PTR]], ptr [[PTR_ADDR]], align 8, !tbaa [[ANYPTR]]
149-
// COMMON-NEXT: [[BASE:%.+]] = load ptr, ptr [[PTR_ADDR]], align 8, !tbaa [[ANYPTR]]
150-
// COMMON-NEXT: [[S:%.+]] = getelementptr inbounds nuw %struct.S2, ptr [[BASE]], i32 0, i32 0
151-
// COMMON-NEXT: store ptr null, ptr [[S]], align 8, !tbaa [[S2_S_TAG:!.+]]
151+
// ENABLED-NEXT: store ptr [[PTR]], ptr [[PTR_ADDR]], align 8, !tbaa [[P1S2_TAG:!.+]]
152+
// ENABLED-NEXT: [[BASE:%.+]] = load ptr, ptr [[PTR_ADDR]], align 8, !tbaa [[P1S2_TAG]]
153+
// ENABLED-NEXT: [[S:%.+]] = getelementptr inbounds nuw %struct.S2, ptr [[BASE]], i32 0, i32 0
154+
// ENABLED-NEXT: store ptr null, ptr [[S]], align 8, !tbaa [[S2_S_TAG:!.+]]
155+
// DEFAULT-NEXT: store ptr [[PTR]], ptr [[PTR_ADDR]], align 8, !tbaa [[ANYPTR]]
156+
// DEFAULT-NEXT: [[BASE:%.+]] = load ptr, ptr [[PTR_ADDR]], align 8, !tbaa [[ANYPTR]]
157+
// DEFAULT-NEXT: [[S:%.+]] = getelementptr inbounds nuw %struct.S2, ptr [[BASE]], i32 0, i32 0
158+
// DEFAULT-NEXT: store ptr null, ptr [[S]], align 8, !tbaa [[S2_S_TAG:!.+]]
152159
// COMMON-NEXT: ret void
153160
ptr->s = 0;
154161
}
155162

163+
164+
void vla1(int n, int ptr[][n], int idx) {
165+
// COMMON-LABEL: define void @vla1(
166+
// COMMON-SAME: i32 noundef [[N:%.+]], ptr noundef [[PTR:%.+]], i32 noundef [[IDX:%.+]])
167+
// COMMON: [[N_ADDR:%.+]] = alloca i32, align 4
168+
// COMMON-NEXT: [[PTR_ADDR:%.+]] = alloca ptr, align 8
169+
// COMMON-NEXT: [[IDX_ADDR:%.+]] = alloca i32, align 4
170+
// COMMON-NEXT: store i32 [[N]], ptr [[N_ADDR]], align 4, !tbaa [[INT_TY:!.+]]
171+
// ENABLED-NEXT: store ptr [[PTR]], ptr [[PTR_ADDR]], align 8, !tbaa [[P1INT0:!.+]]
172+
// DEFAULT-NEXT: store ptr [[PTR]], ptr [[PTR_ADDR]], align 8, !tbaa [[ANYPTR]]
173+
// COMMON-NEXT: store i32 [[IDX]], ptr [[IDX_ADDR]], align 4, !tbaa [[INT_TY]]
174+
// COMMON-NEXT: [[L:%.+]] = load i32, ptr [[N_ADDR]], align 4, !tbaa [[INT_TY]]
175+
// COMMON-NEXT: [[L_EXT:%.+]] = zext i32 [[L]] to i64
176+
// ENABLED-NEXT: [[L_PTR:%.+]] = load ptr, ptr [[PTR_ADDR]], align 8, !tbaa [[P1INT0]]
177+
// DEFAULT-NEXT: [[L_PTR:%.+]] = load ptr, ptr [[PTR_ADDR]], align 8, !tbaa [[ANYPTR]]
178+
// COMMON-NEXT: [[L_IDX:%.+]] = load i32, ptr [[IDX_ADDR]], align 4, !tbaa [[INT_TY]]
179+
// COMMON-NEXT: [[IDX_EXT:%.+]] = sext i32 [[L_IDX]] to i64
180+
// COMMON-NEXT: [[MUL:%.+]] = mul nsw i64 [[IDX_EXT]], [[L_EXT]]
181+
// COMMON-NEXT: [[GEP1:%.+]] = getelementptr inbounds i32, ptr [[L_PTR]], i64 [[MUL]]
182+
// COMMON-NEXT: [[GEP2:%.+]] = getelementptr inbounds i32, ptr [[GEP1]], i64 0
183+
// COMMON-NEXT: store i32 0, ptr [[GEP2]], align 4, !tbaa [[INT_TAG:!.+]]
184+
// ENABLED-NEXT: ret void
185+
186+
ptr[idx][0] = 0;
187+
}
188+
156189
// ENABLED: [[P2INT_0]] = !{[[P2INT:!.+]], [[P2INT]], i64 0}
157190
// ENABLED: [[P2INT]] = !{!"p2 int", [[ANY_POINTER:!.+]], i64 0}
158191
// DEFAULT: [[ANYPTR]] = !{[[ANY_POINTER:!.+]], [[ANY_POINTER]], i64 0}
@@ -171,5 +204,16 @@ void p2struct2(struct S2 *ptr) {
171204
// ENABLED: [[P2CHAR]] = !{!"p2 omnipotent char", [[ANY_POINTER]], i64 0}
172205
// ENABLED: [[P1CHAR_0]] = !{[[P1CHAR:!.+]], [[P1CHAR]], i64 0}
173206
// ENABLED: [[P1CHAR]] = !{!"p1 omnipotent char", [[ANY_POINTER]], i64 0}
174-
// COMMON: [[S2_S_TAG]] = !{[[S2_TY:!.+]], [[ANY_POINTER]], i64 0}
175-
// COMMON: [[S2_TY]] = !{!"S2", [[ANY_POINTER]], i64 0}
207+
// ENABLED: [[P2S1_TAG]] = !{[[P2S1:!.+]], [[P2S1]], i64 0}
208+
// ENABLED: [[P2S1]] = !{!"p2 _ZTS2S1", [[ANY_POINTER]], i64 0}
209+
// ENABLED: [[P1S1_TAG:!.+]] = !{[[P1S1:!.+]], [[P1S1]], i64 0}
210+
// ENABLED: [[P1S1]] = !{!"p1 _ZTS2S1", [[ANY_POINTER]], i64 0}
211+
// ENABLED: [[P1S2_TAG]] = !{[[P1S2:!.+]], [[P1S2]], i64 0}
212+
// ENABLED: [[P1S2]] = !{!"p1 _ZTS2S2", [[ANY_POINTER]], i64 0}
213+
214+
// ENABLED: [[S2_S_TAG]] = !{[[S2_TY:!.+]], [[P1S1]], i64 0}
215+
// ENABLED: [[S2_TY]] = !{!"S2", [[P1S1]], i64 0}
216+
// DEFAULT: [[S2_S_TAG]] = !{[[S2_TY:!.+]], [[ANY_POINTER]], i64 0}
217+
// DEFAULT: [[S2_TY]] = !{!"S2", [[ANY_POINTER]], i64 0}
218+
// COMMON: [[INT_TAG]] = !{[[INT_TY:!.+]], [[INT_TY]], i64 0}
219+
// COMMON: [[INT_TY]] = !{!"int", [[CHAR]], i64 0}

‎clang/test/CodeGen/tbaa-reference.cpp‎

Lines changed: 23 additions & 7 deletions
Original file line numberDiff line numberDiff line change
@@ -1,7 +1,7 @@
11
// RUN: %clang_cc1 -triple x86_64-linux -O1 -disable-llvm-passes %s -emit-llvm -o - | FileCheck %s -check-prefixes=CHECK,OLD-PATH
2-
// RUN: %clang_cc1 -triple x86_64-linux -O1 -disable-llvm-passes -pointer-tbaa %s -emit-llvm -o - | FileCheck %s -check-prefixes=CHECK,OLD-PATH
2+
// RUN: %clang_cc1 -triple x86_64-linux -O1 -disable-llvm-passes -pointer-tbaa %s -emit-llvm -o - | FileCheck %s -check-prefixes=CHECK,OLD-PATH-POINTER
33
// RUN: %clang_cc1 -triple x86_64-linux -O1 -disable-llvm-passes %s -emit-llvm -new-struct-path-tbaa -o - | FileCheck %s -check-prefixes=CHECK,NEW-PATH
4-
// RUN: %clang_cc1 -triple x86_64-linux -O1 -disable-llvm-passes %s -pointer-tbaa -emit-llvm -new-struct-path-tbaa -o - | FileCheck %s -check-prefixes=CHECK,NEW-PATH
4+
// RUN: %clang_cc1 -triple x86_64-linux -O1 -disable-llvm-passes %s -pointer-tbaa -emit-llvm -new-struct-path-tbaa -o - | FileCheck %s -check-prefixes=CHECK,NEW-PATH-POINTER
55
//
66
// Check that we generate correct TBAA information for reference accesses.
77

@@ -16,13 +16,13 @@ struct B {
1616
B::B(S &s) : s(s) {
1717
// CHECK-LABEL: _ZN1BC2ER1S
1818
// Check initialization of the reference parameter.
19-
// CHECK: store ptr {{.*}}, ptr {{.*}}, !tbaa [[TAG_pointer:!.*]]
19+
// CHECK: store ptr {{.*}}, ptr %s.addr, align 8, !tbaa [[TAG_S_PTR:!.*]]
2020

2121
// Check loading of the reference parameter.
22-
// CHECK: load ptr, ptr {{.*}}, !tbaa [[TAG_pointer]]
22+
// CHECK: load ptr, ptr {{.*}}, !tbaa [[TAG_S_PTR:!.*]]
2323

2424
// Check initialization of the reference member.
25-
// CHECK: store ptr {{.*}}, ptr {{.*}}, !tbaa [[TAG_pointer]]
25+
// CHECK: store ptr {{.*}}, ptr {{.*}}, !tbaa [[TAG_S_PTR]]
2626
}
2727

2828
S &B::get() {
@@ -32,16 +32,32 @@ S &B::get() {
3232
return s;
3333
}
3434

35-
// OLD-PATH-DAG: [[TAG_pointer]] = !{[[TYPE_pointer:!.*]], [[TYPE_pointer]], i64 0}
35+
// OLD-PATH-DAG: [[TAG_S_PTR]] = !{[[TYPE_pointer:!.*]], [[TYPE_pointer]], i64 0}
3636
// OLD-PATH-DAG: [[TAG_B_s]] = !{[[TYPE_B:!.*]], [[TYPE_pointer]], i64 0}
3737
//
3838
// OLD-PATH-DAG: [[TYPE_B]] = !{!"_ZTS1B", [[TYPE_pointer]], i64 0}
3939
// OLD-PATH-DAG: [[TYPE_pointer]] = !{!"any pointer", [[TYPE_char:!.*]], i64 0}
4040
// OLD-PATH-DAG: [[TYPE_char]] = !{!"omnipotent char", {{!.*}}, i64 0}
4141

42-
// NEW-PATH-DAG: [[TAG_pointer]] = !{[[TYPE_pointer:!.*]], [[TYPE_pointer]], i64 0, i64 8}
42+
// OLD-PATH-POINTER-DAG: [[TAG_S_PTR]] = !{[[TYPE_S_PTR:!.*]], [[TYPE_S_PTR]], i64 0}
43+
// OLD-PATH-POINTER-DAG: [[TAG_B_s]] = !{[[TYPE_B:!.*]], [[TYPE_S_PTR:!.*]], i64 0}
44+
//
45+
// OLD-PATH-POINTER-DAG: [[TYPE_B]] = !{!"_ZTS1B", [[TYPE_S_PTR:!.*]], i64 0}
46+
// OLD-PATH-POINTER-DAG: [[TYPE_pointer:!.*]] = !{!"any pointer", [[TYPE_char:!.*]], i64 0}
47+
// OLD-PATH-POINTER-DAG: [[TYPE_char]] = !{!"omnipotent char", {{!.*}}, i64 0}
48+
// OLD-PATH-POINTER-DAG: [[TYPE_S_PTR]] = !{!"p1 _ZTS1S", [[TYPE_pointer]], i64 0}
49+
50+
// NEW-PATH-DAG: [[TAG_S_PTR]] = !{[[TYPE_pointer:!.*]], [[TYPE_pointer]], i64 0, i64 8}
4351
// NEW-PATH-DAG: [[TAG_B_s]] = !{[[TYPE_B:!.*]], [[TYPE_pointer]], i64 0, i64 8}
4452
//
4553
// NEW-PATH-DAG: [[TYPE_B]] = !{[[TYPE_char:!.*]], i64 8, !"_ZTS1B", [[TYPE_pointer]], i64 0, i64 8}
4654
// NEW-PATH-DAG: [[TYPE_pointer]] = !{[[TYPE_char:!.*]], i64 8, !"any pointer"}
4755
// NEW-PATH-DAG: [[TYPE_char]] = !{{{!.*}}, i64 1, !"omnipotent char"}
56+
57+
// NEW-PATH-POINTER-DAG: [[TAG_S_PTR]] = !{[[TYPE_S_PTR:!.*]], [[TYPE_S_PTR]], i64 0, i64 8}
58+
// NEW-PATH-POINTER-DAG: [[TAG_B_s]] = !{[[TYPE_B:!.*]], [[TYPE_S_PTR]], i64 0, i64 8}
59+
//
60+
// NEW-PATH-POINTER-DAG: [[TYPE_B]] = !{[[TYPE_char:!.*]], i64 8, !"_ZTS1B", [[TYPE_S_PTR]], i64 0, i64 8}
61+
// NEW-PATH-POINTER-DAG: [[TYPE_S_PTR]] = !{[[TYPE_pointer:!.+]], i64 8, !"p1 _ZTS1S"}
62+
// NEW-PATH-POINTER-DAG: [[TYPE_pointer]] = !{[[TYPE_char:!.*]], i64 8, !"any pointer"}
63+
// NEW-PATH-POINTER-DAG: [[TYPE_char]] = !{{{!.*}}, i64 1, !"omnipotent char"}

0 commit comments

Comments
 (0)