Control Flow Integrity Design Documentation

This page documents the design of the Control Flow Integrity schemes supported by Clang.

Forward-Edge CFI for Virtual Calls

This scheme works by allocating, for each static type used to make a virtual call, a region of read-only storage in the object file holding a bit vector that maps onto to the region of storage used for those virtual tables. Each set bit in the bit vector corresponds to the address point for a virtual table compatible with the static type for which the bit vector is being built.

For example, consider the following three C++ classes:

struct A {
  virtual void f1();
  virtual void f2();
  virtual void f3();

struct B : A {
  virtual void f1();
  virtual void f2();
  virtual void f3();

struct C : A {
  virtual void f1();
  virtual void f2();
  virtual void f3();

The scheme will cause the virtual tables for A, B and C to be laid out consecutively:

Virtual Table Layout for A, B, C
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14
A::offset-to-top &A::rtti &A::f1 &A::f2 &A::f3 B::offset-to-top &B::rtti &B::f1 &B::f2 &B::f3 C::offset-to-top &C::rtti &C::f1 &C::f2 &C::f3

The bit vector for static types A, B and C will look like this:

Bit Vectors for A, B, C
Class 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14
A 0 0 1 0 0 0 0 1 0 0 0 0 1 0 0
B 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0
C 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0

Bit vectors are represented in the object file as byte arrays. By loading from indexed offsets into the byte array and applying a mask, a program can test bits from the bit set with a relatively short instruction sequence. Bit vectors may overlap so long as they use different bits. For the full details, see the ByteArrayBuilder class.

In this case, assuming A is laid out at offset 0 in bit 0, B at offset 0 in bit 1 and C at offset 0 in bit 2, the byte array would look like this:

char bits[] = { 0, 0, 1, 0, 0, 0, 3, 0, 0, 0, 0, 5, 0, 0 };

To emit a virtual call, the compiler will assemble code that checks that the object’s virtual table pointer is in-bounds and aligned and that the relevant bit is set in the bit vector.

For example on x86 a typical virtual call may look like this:

ca7fbb:       48 8b 0f                mov    (%rdi),%rcx
ca7fbe:       48 8d 15 c3 42 fb 07    lea    0x7fb42c3(%rip),%rdx
ca7fc5:       48 89 c8                mov    %rcx,%rax
ca7fc8:       48 29 d0                sub    %rdx,%rax
ca7fcb:       48 c1 c0 3d             rol    $0x3d,%rax
ca7fcf:       48 3d 7f 01 00 00       cmp    $0x17f,%rax
ca7fd5:       0f 87 36 05 00 00       ja     ca8511
ca7fdb:       48 8d 15 c0 0b f7 06    lea    0x6f70bc0(%rip),%rdx
ca7fe2:       f6 04 10 10             testb  $0x10,(%rax,%rdx,1)
ca7fe6:       0f 84 25 05 00 00       je     ca8511
ca7fec:       ff 91 98 00 00 00       callq  *0x98(%rcx)
ca8511:       0f 0b                   ud2

The compiler relies on co-operation from the linker in order to assemble the bit vectors for the whole program. It currently does this using LLVM’s type metadata mechanism together with link-time optimization.


The scheme as described above is the fully general variant of the scheme. Most of the time we are able to apply one or more of the following optimizations to improve binary size or performance.

In fact, if you try the above example with the current version of the compiler, you will probably find that it will not use the described virtual table layout or machine instructions. Some of the optimizations we are about to introduce cause the compiler to use a different layout or a different sequence of machine instructions.

Stripping Leading/Trailing Zeros in Bit Vectors

If a bit vector contains leading or trailing zeros, we can strip them from the vector. The compiler will emit code to check if the pointer is in range of the region covered by ones, and perform the bit vector check using a truncated version of the bit vector. For example, the bit vectors for our example class hierarchy will be emitted like this:

Bit Vectors for A, B, C
Class 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14
A     1 0 0 0 0 1 0 0 0 0 1    
B               1              
C                         1    

Short Inline Bit Vectors

If the vector is sufficiently short, we can represent it as an inline constant on x86. This saves us a few instructions when reading the correct element of the bit vector.

If the bit vector fits in 32 bits, the code looks like this:

 dc2:       48 8b 03                mov    (%rbx),%rax
 dc5:       48 8d 15 14 1e 00 00    lea    0x1e14(%rip),%rdx
 dcc:       48 89 c1                mov    %rax,%rcx
 dcf:       48 29 d1                sub    %rdx,%rcx
 dd2:       48 c1 c1 3d             rol    $0x3d,%rcx
 dd6:       48 83 f9 03             cmp    $0x3,%rcx
 dda:       77 2f                   ja     e0b <main+0x9b>
 ddc:       ba 09 00 00 00          mov    $0x9,%edx
 de1:       0f a3 ca                bt     %ecx,%edx
 de4:       73 25                   jae    e0b <main+0x9b>
 de6:       48 89 df                mov    %rbx,%rdi
 de9:       ff 10                   callq  *(%rax)
 e0b:       0f 0b                   ud2

Or if the bit vector fits in 64 bits:

11a6:       48 8b 03                mov    (%rbx),%rax
11a9:       48 8d 15 d0 28 00 00    lea    0x28d0(%rip),%rdx
11b0:       48 89 c1                mov    %rax,%rcx
11b3:       48 29 d1                sub    %rdx,%rcx
11b6:       48 c1 c1 3d             rol    $0x3d,%rcx
11ba:       48 83 f9 2a             cmp    $0x2a,%rcx
11be:       77 35                   ja     11f5 <main+0xb5>
11c0:       48 ba 09 00 00 00 00    movabs $0x40000000009,%rdx
11c7:       04 00 00
11ca:       48 0f a3 ca             bt     %rcx,%rdx
11ce:       73 25                   jae    11f5 <main+0xb5>
11d0:       48 89 df                mov    %rbx,%rdi
11d3:       ff 10                   callq  *(%rax)
11f5:       0f 0b                   ud2

If the bit vector consists of a single bit, there is only one possible virtual table, and the check can consist of a single equality comparison:

9a2:   48 8b 03                mov    (%rbx),%rax
9a5:   48 8d 0d a4 13 00 00    lea    0x13a4(%rip),%rcx
9ac:   48 39 c8                cmp    %rcx,%rax
9af:   75 25                   jne    9d6 <main+0x86>
9b1:   48 89 df                mov    %rbx,%rdi
9b4:   ff 10                   callq  *(%rax)
9d6:   0f 0b                   ud2

Virtual Table Layout

The compiler lays out classes of disjoint hierarchies in separate regions of the object file. At worst, bit vectors in disjoint hierarchies only need to cover their disjoint hierarchy. But the closer that classes in sub-hierarchies are laid out to each other, the smaller the bit vectors for those sub-hierarchies need to be (see “Stripping Leading/Trailing Zeros in Bit Vectors” above). The GlobalLayoutBuilder class is responsible for laying out the globals efficiently to minimize the sizes of the underlying bitsets.


If all gaps between address points in a particular bit vector are multiples of powers of 2, the compiler can compress the bit vector by strengthening the alignment requirements of the virtual table pointer. For example, given this class hierarchy:

struct A {
  virtual void f1();
  virtual void f2();

struct B : A {
  virtual void f1();
  virtual void f2();
  virtual void f3();
  virtual void f4();
  virtual void f5();
  virtual void f6();

struct C : A {
  virtual void f1();
  virtual void f2();

The virtual tables will be laid out like this:

Virtual Table Layout for A, B, C
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
A::offset-to-top &A::rtti &A::f1 &A::f2 B::offset-to-top &B::rtti &B::f1 &B::f2 &B::f3 &B::f4 &B::f5 &B::f6 C::offset-to-top &C::rtti &C::f1 &C::f2

Notice that each address point for A is separated by 4 words. This lets us emit a compressed bit vector for A that looks like this:

2 6 10 14
1 1 0 1

At call sites, the compiler will strengthen the alignment requirements by using a different rotate count. For example, on a 64-bit machine where the address points are 4-word aligned (as in A from our example), the rol instruction may look like this:

dd2:       48 c1 c1 3b             rol    $0x3b,%rcx

Padding to Powers of 2

Of course, this alignment scheme works best if the address points are in fact aligned correctly. To make this more likely to happen, we insert padding between virtual tables that in many cases aligns address points to a power of 2. Specifically, our padding aligns virtual tables to the next highest power of 2 bytes; because address points for specific base classes normally appear at fixed offsets within the virtual table, this normally has the effect of aligning the address points as well.

This scheme introduces tradeoffs between decreased space overhead for instructions and bit vectors and increased overhead in the form of padding. We therefore limit the amount of padding so that we align to no more than 128 bytes. This number was found experimentally to provide a good tradeoff.

Eliminating Bit Vector Checks for All-Ones Bit Vectors

If the bit vector is all ones, the bit vector check is redundant; we simply need to check that the address is in range and well aligned. This is more likely to occur if the virtual tables are padded.

Forward-Edge CFI for Indirect Function Calls

Under forward-edge CFI for indirect function calls, each unique function type has its own bit vector, and at each call site we need to check that the function pointer is a member of the function type’s bit vector. This scheme works in a similar way to forward-edge CFI for virtual calls, the distinction being that we need to build bit vectors of function entry points rather than of virtual tables.

Unlike when re-arranging global variables, we cannot re-arrange functions in a particular order and base our calculations on the layout of the functions’ entry points, as we have no idea how large a particular function will end up being (the function sizes could even depend on how we arrange the functions). Instead, we build a jump table, which is a block of code consisting of one branch instruction for each of the functions in the bit set that branches to the target function, and redirect any taken function addresses to the corresponding jump table entry. In this way, the distance between function entry points is predictable and controllable. In the object file’s symbol table, the symbols for the target functions also refer to the jump table entries, so that addresses taken outside the module will pass any verification done inside the module.

In more concrete terms, suppose we have three functions f, g, h which are all of the same type, and a function foo that returns their addresses:

mov 0, %eax

mov 1, %eax

mov 2, %eax

mov f, %eax
mov g, %edx
mov h, %ecx

Our jump table will (conceptually) look like this:

jmp .Ltmp0 ; 5 bytes
int3       ; 1 byte
int3       ; 1 byte
int3       ; 1 byte

jmp .Ltmp1 ; 5 bytes
int3       ; 1 byte
int3       ; 1 byte
int3       ; 1 byte

jmp .Ltmp2 ; 5 bytes
int3       ; 1 byte
int3       ; 1 byte
int3       ; 1 byte

mov 0, %eax

mov 1, %eax

mov 2, %eax

mov f, %eax
mov g, %edx
mov h, %ecx

Because the addresses of f, g, h are evenly spaced at a power of 2, and function types do not overlap (unlike class types with base classes), we can normally apply the Alignment and Eliminating Bit Vector Checks for All-Ones Bit Vectors optimizations thus simplifying the check at each call site to a range and alignment check.

Shared library support


The basic CFI mode described above assumes that the application is a monolithic binary; at least that all possible virtual/indirect call targets and the entire class hierarchy are known at link time. The cross-DSO mode, enabled with -f[no-]sanitize-cfi-cross-dso relaxes this requirement by allowing virtual and indirect calls to cross the DSO boundary.

Assuming the following setup: the binary consists of several instrumented and several uninstrumented DSOs. Some of them may be dlopen-ed/dlclose-d periodically, even frequently.

  • Calls made from uninstrumented DSOs are not checked and just work.
  • Calls inside any instrumented DSO are fully protected.
  • Calls between different instrumented DSOs are also protected, with
    a performance penalty (in addition to the monolithic CFI overhead).
  • Calls from an instrumented DSO to an uninstrumented one are
    unchecked and just work, with performance penalty.
  • Calls from an instrumented DSO outside of any known DSO are
    detected as CFI violations.

In the monolithic scheme a call site is instrumented as

if (!InlinedFastCheck(f))
call *f

In the cross-DSO scheme it becomes

if (!InlinedFastCheck(f))
  __cfi_slowpath(CallSiteTypeId, f);
call *f


CallSiteTypeId is a stable process-wide identifier of the call-site type. For a virtual call site, the type in question is the class type; for an indirect function call it is the function signature. The mapping from a type to an identifier is an ABI detail. In the current, experimental, implementation the identifier of type T is calculated as follows:

  • Obtain the mangled name for “typeinfo name for T”.
  • Calculate MD5 hash of the name as a string.
  • Reinterpret the first 8 bytes of the hash as a little-endian 64-bit integer.

It is possible, but unlikely, that collisions in the CallSiteTypeId hashing will result in weaker CFI checks that would still be conservatively correct.


In the general case, only the target DSO knows whether the call to function f with type CallSiteTypeId is valid or not. To export this information, every DSO implements

void __cfi_check(uint64 CallSiteTypeId, void *TargetAddr)

This function provides external modules with access to CFI checks for the targets inside this DSO. For each known CallSiteTypeId, this function performs an llvm.type.test with the corresponding type identifier. It aborts if the type is unknown, or if the check fails.

The basic implementation is a large switch statement over all values of CallSiteTypeId supported by this DSO, and each case is similar to the InlinedFastCheck() in the basic CFI mode.

CFI Shadow

To route CFI checks to the target DSO’s __cfi_check function, a mapping from possible virtual / indirect call targets to the corresponding __cfi_check functions is maintained. This mapping is implemented as a sparse array of 2 bytes for every possible page (4096 bytes) of memory. The table is kept readonly (FIXME: not yet) most of the time.

There are 3 types of shadow values:

  • Address in a CFI-instrumented DSO.
  • Unchecked address (a “trusted” non-instrumented DSO). Encoded as value 0xFFFF.
  • Invalid address (everything else). Encoded as value 0.

For a CFI-instrumented DSO, a shadow value encodes the address of the __cfi_check function for all call targets in the corresponding memory page. If Addr is the target address, and V is the shadow value, then the address of __cfi_check is calculated as

__cfi_check = AlignUpTo(Addr, 4096) - (V + 1) * 4096

This works as long as __cfi_check is aligned by 4096 bytes and located below any call targets in its DSO, but not more than 256MB apart from them.


The slow path check is implemented in compiler-rt library as

void __cfi_slowpath(uint64 CallSiteTypeId, void *TargetAddr)

This functions loads a shadow value for TargetAddr, finds the address of __cfi_check as described above and calls that.

Position-independent executable requirement

Cross-DSO CFI mode requires that the main executable is built as PIE. In non-PIE executables the address of an external function (taken from the main executable) is the address of that function’s PLT record in the main executable. This would break the CFI checks.

Backward-edge CFI for return statements (RCFI)

This section is a proposal. As of March 2017 it is not implemented.

Backward-edge control flow (RET instructions) can be hijacked via overwriting the return address (RA) on stack. Various mitigation techniques (e.g. SafeStack, RFG, Intel CET) try to detect or prevent RA corruption on stack.

RCFI enforces the expected control flow in several different ways described below. RCFI heavily relies on LTO.

Leaf Functions

If f() is a leaf function (i.e. it has no calls except maybe no-return calls) it can be called using a special calling convention that stores RA in a dedicated register R before the CALL instruction. f() does not spill R and does not use the RET instruction, instead it uses the value in R to JMP to RA.

This flavour of CFI is precise, i.e. the function is guaranteed to return to the point exactly following the call.

An alternative approach is to copy RA from stack to R in the first instruction of f(), then JMP to R. This approach is simpler to implement (does not require changing the caller) but weaker (there is a small window when RA is actually stored on stack).

Functions called once

Suppose f() is called in just one place in the program (assuming we can verify this in LTO mode). In this case we can replace the RET instruction with a JMP instruction with the immediate constant for RA. This will precisely enforce the return control flow no matter what is stored on stack.

Another variant is to compare RA on stack with the known constant and abort if they don’t match; then JMP to the known constant address.

Functions called in a small number of call sites

We may extend the above approach to cases where f() is called more than once (but still a small number of times). With LTO we know all possible values of RA and we check them one-by-one (or using binary search) against the value on stack. If the match is found, we JMP to the known constant address, otherwise abort.

This protection is near-precise, i.e. it guarantees that the control flow will be transferred to one of the valid return addresses for this function, but not necessary to the point of the most recent CALL.

General case

For functions called multiple times a return jump table is constructed in the same manner as jump tables for indirect function calls (see above). The correct jump table entry (or it’s index) is passed by CALL to f() (as an extra argument) and then spilled to stack. The RET instruction is replaced with a load of the jump table entry, jump table range check, and JMP to the jump table entry.

This protection is also near-precise.

Returns from functions called indirectly

If a function is called indirectly, the return jump table is constructed for the equivalence class of functions instead of a single function.

Cross-DSO calls

Consider two instrumented DSOs, A and B. A defines f() and B calls it.

This case will be handled similarly to the cross-DSO scheme using the slow path callback.


RCFI does not protect RET instructions:
  • in non-instrumented DSOs,
  • in instrumented DSOs for functions that are called from non-instrumented DSOs,
  • embedded into other instructions (e.g. 0f4fc3 cmovg %ebx,%eax).

Hardware support

We believe that the above design can be efficiently implemented in hardware. A single new instruction added to an ISA would allow to perform the forward-edge CFI check with fewer bytes per check (smaller code size overhead) and potentially more efficiently. The current software-only instrumentation requires at least 32-bytes per check (on x86_64). A hardware instruction may probably be less than ~ 12 bytes. Such instruction would check that the argument pointer is in-bounds, and is properly aligned, and if the checks fail it will either trap (in monolithic scheme) or call the slow path function (cross-DSO scheme). The bit vector lookup is probably too complex for a hardware implementation.

//  This instruction checks that 'Ptr'
//   * is aligned by (1 << kAlignment) and
//   * is inside [kRangeBeg, kRangeBeg+(kRangeSize<<kAlignment))
//  and if the check fails it jumps to the given target (slow path).
// 'Ptr' is a register, pointing to the virtual function table
//    or to the function which we need to check. We may require an explicit
//    fixed register to be used.
// 'kAlignment' is a 4-bit constant.
// 'kRangeSize' is a ~20-bit constant.
// 'kRangeBeg' is a PC-relative constant (~28 bits)
//    pointing to the beginning of the allowed range for 'Ptr'.
// 'kFailedCheckTarget': is a PC-relative constant (~28 bits)
//    representing the target to branch to when the check fails.
//    If kFailedCheckTarget==0, the process will trap
//    (monolithic binary scheme).
//    Otherwise it will jump to a handler that implements `CFI_SlowPath`
//    (cross-DSO scheme).
CFI_Check(Ptr, kAlignment, kRangeSize, kRangeBeg, kFailedCheckTarget) {
   if (Ptr < kRangeBeg ||
       Ptr >= kRangeBeg + (kRangeSize << kAlignment) ||
       Ptr & ((1 << kAlignment) - 1))

An alternative and more compact encoding would not use kFailedCheckTarget, and will trap on check failure instead. This will allow us to fit the instruction into 8-9 bytes. The cross-DSO checks will be performed by a trap handler and performance-critical ones will have to be black-listed and checked using the software-only scheme.

Note that such hardware extension would be complementary to checks at the callee side, such as e.g. Intel ENDBRANCH. Moreover, CFI would have two benefits over ENDBRANCH: a) precision and b) ability to protect against invalid casts between polymorphic types.