[llvm] [LLVM] Fix an ASAN error in `lookupLLVMIntrinsicByName` (PR #147444)

[Rahul Joshi via llvm-commits]

================
@@ -80,6 +80,28 @@ TEST(IntrinsicNameLookup, Basic) {
   EXPECT_EQ(memcpy_inline, lookupIntrinsicID("llvm.memcpy.inline.p0.p0.i1024"));
 }
 
+TEST(IntrinsicNameLookup, NonNullterminatedStringRef) {
+  using namespace Intrinsic;
+  // This reproduces an issue where lookupIntrinsicID() can access memory beyond
+  // the bounds of the passed in StringRef. For ASAN to catch this as an error,
+  // create a StringRef using heap allocated memory and make it not null
+  // terminated.
+
+  // ASAN will report a "AddressSanitizer: heap-buffer-overflow" error in
+  // `lookupLLVMIntrinsicByName` when LLVM is built with these options:
+  //  -DCMAKE_BUILD_TYPE=Debug
+  //  -DLLVM_USE_SANITIZER=Address
+  //  -DLLVM_OPTIMIZE_SANITIZED_BUILDS=OFF
+
+  // Make an intrinsic name "llvm.memcpy.inline" on the heap.
+  std::string Name = "llvm.memcpy.inline";
+  assert(Name.size() == 18);
+  std::unique_ptr<char[]> Data = std::make_unique<char[]>(Name.size());
+  std::strncpy(Data.get(), Name.data(), Name.size());
----------------
jurahul wrote:

I want access to Data[18] to be flagged by asan as a bad access. With SmallVector, I don't know or want to rely on its internal implementation detail of whether it will allocate extra capacity, in which case asan will not flag the bad access. With this, I am directly allocating data on the heap so nothing unexpected might happen and "break" the unit test.

https://github.com/llvm/llvm-project/pull/147444