Reapply "[TBAA] Add verifier for tbaa.struct metadata (llvm#86709)"

This reverts commit b9cd48f.

-------------------------------------------------------------
Original commit message:

Adds logic to the IR verifier that checks whether !tbaa.struct nodes are
well-formed. That is, it checks that the operands of !tbaa.struct nodes
are in groups of three, that each group of three operands consists of
two integers and a valid tbaa node, and that the regions described by
the offset and size operands are non-overlapping.

PR: llvm#86709

llvm/include/llvm/IR/Verifier.h

@@ -77,6 +77,7 @@ class TBAAVerifier {
   /// Visit an instruction and return true if it is valid, return false if an
   /// invalid TBAA is attached.
   bool visitTBAAMetadata(Instruction &I, const MDNode *MD);
+  bool visitTBAAStructMetadata(Instruction &I, const MDNode *MD);
 };
 
 /// Check a function for errors, useful for use when debugging a

llvm/lib/IR/Verifier.cpp

@@ -5122,6 +5122,9 @@ void Verifier::visitInstruction(Instruction &I) {
   if (MDNode *TBAA = I.getMetadata(LLVMContext::MD_tbaa))
     TBAAVerifyHelper.visitTBAAMetadata(I, TBAA);
 
+  if (MDNode *TBAA = I.getMetadata(LLVMContext::MD_tbaa_struct))
+    TBAAVerifyHelper.visitTBAAStructMetadata(I, TBAA);
+
   if (MDNode *MD = I.getMetadata(LLVMContext::MD_noalias))
     visitAliasScopeListMetadata(MD);
   if (MDNode *MD = I.getMetadata(LLVMContext::MD_alias_scope))
@@ -7453,6 +7456,35 @@ bool TBAAVerifier::visitTBAAMetadata(Instruction &I, const MDNode *MD) {
   return true;
 }
 
+bool TBAAVerifier::visitTBAAStructMetadata(Instruction &I, const MDNode *MD) {
+  CheckTBAA(MD->getNumOperands() % 3 == 0,
+            "tbaa.struct operands must occur in groups of three", &I, MD);
+
+  // Each group of three operands must consist of two integers and a
+  // tbaa node. Moreover, the regions described by the offset and size
+  // operands must be non-overlapping.
+  std::optional<APInt> NextFree;
+  for (unsigned int Idx = 0; Idx < MD->getNumOperands(); Idx += 3) {
+    auto *OffsetCI =
+        mdconst::dyn_extract_or_null<ConstantInt>(MD->getOperand(Idx));
+    CheckTBAA(OffsetCI, "Offset must be a constant integer", &I, MD);
+
+    auto *SizeCI =
+        mdconst::dyn_extract_or_null<ConstantInt>(MD->getOperand(Idx + 1));
+    CheckTBAA(SizeCI, "Size must be a constant integer", &I, MD);
+
+    MDNode *TBAA = dyn_cast_or_null<MDNode>(MD->getOperand(Idx + 2));
+    CheckTBAA(TBAA, "TBAA tag missing", &I, MD);
+    visitTBAAMetadata(I, TBAA);
+
+    bool NonOverlapping = !NextFree || NextFree->ule(OffsetCI->getValue());
+    CheckTBAA(NonOverlapping, "Overlapping tbaa.struct regions", &I, MD);
+
+    NextFree = OffsetCI->getValue() + SizeCI->getValue();
+  }
+  return true;
+}
+
 char VerifierLegacyPass::ID = 0;
 INITIALIZE_PASS(VerifierLegacyPass, "verify", "Module Verifier", false, false)
 

llvm/test/CodeGen/AArch64/arm64-abi_align.ll

@@ -518,4 +518,6 @@ attributes #5 = { nobuiltin }
 !1 = !{!"omnipotent char", !2}
 !2 = !{!"Simple C/C++ TBAA"}
 !3 = !{!"short", !1}
-!4 = !{i64 0, i64 4, !0, i64 4, i64 2, !3, i64 8, i64 4, !0, i64 12, i64 2, !3, i64 16, i64 4, !0, i64 20, i64 2, !3}
+!4 = !{i64 0, i64 4, !5, i64 4, i64 2, !6, i64 8, i64 4, !5, i64 12, i64 2, !6, i64 16, i64 4, !5, i64 20, i64 2, !6}
+!5 = !{!0, !0, i64 0}
+!6 = !{!3, !3, i64 0}

llvm/test/Transforms/InferAddressSpaces/AMDGPU/mem-intrinsics.ll

@@ -141,4 +141,4 @@ attributes #1 = { argmemonly nounwind }
 !5 = distinct !{!5, !"some domain"}
 !6 = !{!7}
 !7 = distinct !{!7, !5, !"some scope 2"}
-!8 = !{i64 0, i64 8, null}
+!8 = !{i64 0, i64 8, !0}

llvm/test/Transforms/InstCombine/struct-assign-tbaa.ll

@@ -75,7 +75,7 @@ entry:
 !1 = !{!"omnipotent char", !0}
 !2 = !{!5, !5, i64 0}
 !3 = !{i64 0, i64 4, !2}
-!4 = !{i64 0, i64 8, null}
+!4 = !{i64 0, i64 8, !2}
 !5 = !{!"float", !0}
 !6 = !{i64 0, i64 4, !2, i64 4, i64 4, !2}
 !7 = !{i64 0, i64 2, !2, i64 4, i64 6, !2}

llvm/test/Transforms/SROA/tbaa-struct3.ll

@@ -539,7 +539,7 @@ declare void @llvm.memcpy.p0.p0.i64(ptr noalias nocapture writeonly, ptr noalias
 !6 = !{!5, !5, i64 0}
 !7 = !{i64 0, i64 8, !6, i64 8, i64 4, !1}
 !8 = !{i64 0, i64 4, !1, i64 4, i64 8, !6}
-!9 = !{i64 0, i64 8, !6, i64 4, i64 8, !1}
+!9 = !{i64 0, i64 8, !6, i64 8, i64 8, !1}
 !10 = !{i64 0, i64 2, !1, i64 2, i64 2, !1}
 !11 = !{i64 0, i64 1, !1, i64 1, i64 3, !1}
 !12 = !{i64 0, i64 2, !1, i64 2, i64 6, !1}

llvm/test/Transforms/Scalarizer/basic-inseltpoison.ll

@@ -836,5 +836,6 @@ define <2 x i32> @f23_crash(<2 x i32> %srcvec, i32 %v1) {
 !2 = !{ !"set2", !0 }
 !3 = !{ !3, !{!"llvm.loop.parallel_accesses", !13} }
 !4 = !{ float 4.0 }
-!5 = !{ i64 0, i64 8, null }
+!5 = !{ i64 0, i64 8, !6 }
+!6 = !{ !1, !1, i64 0 }
 !13 = distinct !{}

llvm/test/Transforms/Scalarizer/basic.ll

@@ -870,5 +870,6 @@ define <2 x float> @f25(<2 x float> %src) {
 !2 = !{ !"set2", !0 }
 !3 = !{ !3, !{!"llvm.loop.parallel_accesses", !13} }
 !4 = !{ float 4.0 }
-!5 = !{ i64 0, i64 8, null }
+!5 = !{ i64 0, i64 8, !6 }
+!6 = !{ !1, !1, i64 0 }
 !13 = distinct !{}

llvm/test/Verifier/tbaa-struct.ll

@@ -1,28 +1,36 @@
-; RUN: llvm-as < %s 2>&1
-
-; FIXME: The verifer should reject the invalid !tbaa.struct nodes below.
+; RUN: not llvm-as < %s 2>&1 | FileCheck %s
 
 define void @test_overlapping_regions(ptr %a1) {
+; CHECK: Overlapping tbaa.struct regions
+; CHECK-NEXT:  %ld = load i8, ptr %a1, align 1, !tbaa.struct !0
   %ld = load i8, ptr %a1, align 1, !tbaa.struct !0
   ret void
 }
 
 define void @test_size_not_integer(ptr %a1) {
+; CHECK: Size must be a constant integer
+; CHECK-NEXT:  store i8 1, ptr %a1, align 1, !tbaa.struct !5
   store i8 1, ptr %a1, align 1, !tbaa.struct !5
   ret void
 }
 
 define void @test_offset_not_integer(ptr %a1, ptr %a2) {
+; CHECK: Offset must be a constant integer
+; CHECK-NEXT:  tail call void @llvm.memcpy.p0.p0.i64(ptr align 8 %a1, ptr align 8 %a2, i64 16, i1 false), !tbaa.struct !6
   tail call void @llvm.memcpy.p0.p0.i64(ptr align 8 %a1, ptr align 8 %a2, i64 16, i1 false), !tbaa.struct !6
   ret void
 }
 
 define void @test_tbaa_missing(ptr %a1, ptr %a2) {
+; CHECK: TBAA tag missing
+; CHECK-NEXT:  tail call void @llvm.memcpy.p0.p0.i64(ptr align 8 %a1, ptr align 8 %a2, i64 16, i1 false), !tbaa.struct !7
   tail call void @llvm.memcpy.p0.p0.i64(ptr align 8 %a1, ptr align 8 %a2, i64 16, i1 false), !tbaa.struct !7
   ret void
 }
 
 define void @test_tbaa_invalid(ptr %a1) {
+; CHECK: Old-style TBAA is no longer allowed, use struct-path TBAA instead
+; CHECK-NEXT:  store i8 1, ptr %a1, align 1, !tbaa.struct !8
   store i8 1, ptr %a1, align 1, !tbaa.struct !8
   ret void
 }