Binary Exploitation 1

Buffer Overflows

(return-to-libc, ROP, Canaries, W^X, ASLR)

Chester Rebeiro

Indian Institute of Technology Madras

 Parts of Malware
• Two parts
Subvert execution:
change the normal execution behavior of the
program

Payload:
the code which the attacker wants to execute

2
 Subvert Execution
• In application software
– SQL Injection

• In system software
– Buffers overflows and overreads
– Heap: double free, use after free
– Integer overflows
– Format string
– Control Flow

• In peripherials
– USB drives; Printers

• In Hardware
– Hardware Trojans
These do not really subvert execution,
• Covert Channels but can lead to confidentiality attacks.
– Can exist in hardware or software

3
 Buffer Overflows in the Stack
• We need to first know how a stack is managed

http://insecure.org/stf/smashstack.html 4
 Stack in a Program
(when function is executing)
Stack
ESP
ESP Parameters
for function
ESP
ESP return Address
ESP prev frame pointer EBP

Locals of function
ESP
In main In function
push $3 push %ebp
push $2 movl %esp, %ebp %ebp: Frame Pointer
push $1 sub $20, %esp %esp : Stack Pointer
call function

5
 Stack Usage (example)
Stack (top to bottom):
address stored data
1000 to 997 3

996 to 993 2

992 to 989 1
Parameters
for function 988 to 985 return address

984 to 981 %ebp (stored

Return Address frame pointer)
frame pointer prev frame pointer (%ebp)980 to 976 buffer1

Locals of function 975 to 966 buffer2

(%sp) 964
stack pointer
6
 Stack Usage Contd.
Stack (top to bottom):
address stored data
1000 to 997 3

996 to 993 2

992 to 989 1
What is the output of the following?
988 to 985 return address
• printf(“%x”, buffer2) : 966
• printf(“%x”, &buffer2[10]) 984 to 981 %ebp (stored
frame pointer)
976 à buffer1
(%ebp)980 to 976 buffer1
Therefore buffer2[10] = buffer1[0]
A BUFFER OVERFLOW 975 to 966 buffer2
(%sp) 964

7
 Modifying the Return Address
Stack (top to bottom):
buffer2[19] =
address stored data
&arbitrary memory location
1000 to 997 3

This causes execution of an 996 to 993 2

arbitrary memory location
instead of the standard return 992 to 989 1

988 to 985 Arbitrary

Return Address
Location

19 984 to 981 %ebp (stored

frame pointer)
(%ebp)980 to 976 buffer1

976 to 966 buffer2

(%sp) 964

8
 Stack (top to bottom):
address stored data
1000 to 997 3 Now that we seen how buffer
overflows can skip an instruction,
996 to 993 2

992 to 989 1 We will see how an attacker can use

it to execute his own code (exploit
988 to 985 ATTACKER’S code)
code pointer
984 to 981 %ebp (stored
frame pointer)
(%ebp)980 to 976 buffer1

976 to 966 buffer2

(%sp) 964

9
 Big Picture of the exploit
Fill the stack as follows
(where BA is buffer address) BA
BA
Parameters BA
for function BA
BA
Return Address
BA
frame pointer prev frame pointer
BA
buffer BA

stack pointer Exploit code BA

buffer Address

10
 Payload
• Lets say the attacker wants to spawn a shell
• ie. do as follows:

• How does he put this code onto the stack?

11
Step 1 : Get machine codes

• objdump –disassemble-all shellcode.o

• Get machine code : “eb 1e 5e 89 76 08 c6
46 07 00 c7 46 0c 00 00 00 00 b8 0b 00 00
00 89 f3 8d 4e 08 8d 56 0c cd 80 cd 80”
• If there are 00s replace it with other
instructions

12
Step 2: Find Buffer overflow in an
application

Defined on stack

O
O
O
O
o

13
 Step 3 :
Put Machine Code in Large String

large_string
shellcode

14
 Step 3 (contd) :
Fill up Large String with BA

Address of buffer is BA

large_string
shellcode BA BA BA BA BA BA BA BA

15
 Final state of Stack
BA
BA
• Copy large string into buffer
BA
BA
BA
• When strcpy returns the BA
exploit code would be executed BA
BA
buffer
shellcode

large_string
shellcode BA BA BA BA BA BA BA BA BA
buffer Address
BA

16
Putting it all together

bash$ gcc overflow1.c

bash$ ./a.out
$sh

17
 Buffer overflow in the Wild
• Worm CODERED … released on 13th July 2001
• Infected 3,59,000 computers by 19th July.

18
 Defenses
• Eliminate program flaws that could lead to subverting of execution
Safer programming languages; Safer libraries; hardware enhancements;
static analysis
• If can’t eliminate, make it more difficult for malware to subvert
execution
W^X , ASLR, canaries
• If malware still manages to execute, try to detect its execution at
runtime
malware run-time detection techniques using learning techniques, ANN and malware signatures
• If can’t detect at runtime, try to restrict what the malware can do..
– Sandbox system
so that malware affects only part of the system; access control; virtualization; trustzone; SGX
– Track information flow
DIFT; ensure malware does not steal sensitive information

19
Preventing Buffer Overflows
with Canaries and W^X

20
 Canaries
• Known (pseudo random) values placed Stack (top to bottom):
on stack to monitor buffer overflows. stored data
• A change in the value of the canary
3
indicates a buffer overflow.
• Will cause a ‘stack smashing’ to be 2
detected
1
ret addr

Insert a canary here

sfp (%ebp)
Insert canary here
buffer1
check if the canary value
has got modified buffer2

21
 Canaries and gcc
• As on gcc 4.4.5, canaries are not added to functions by default
o Could cause overheads as they are executed for every function
that gets executed
• Canaries can be added into the code by –fstack-protector option
o If -fstack-protector is specified, canaries will get added based on
a gcc heuristic
• For example, buffer of size at-least 8 bytes is allocated
• Use of string operations such as strcpy, scanf, etc.

o Canaries can be evaded quite easily by not altering the contents of

the canary

22
 Canaries Example
Without canaries, the return address on stack gets overwritten resulting in a
segmentation fault. With canaries, the program gets aborted due to stack smashing.

23
 Canaries Example
Without canaries, the return address on stack gets overwritten resulting in a
segmentation fault. With canaries, the program gets aborted due to stack smashing.

24
 Canary Internals

Store canary onto stack

Verify if the canary has Without canaries

changed

With canaries

gs is a segment that shows thread local data; in this case it is used

for picking out canaries

25
 Non Executable Stacks (W^X)
• In Intel/AMD processors, ND/NX bit present to mark non code
regions as non-executable.
– Exception raised when code in a page marked W^X executes
• Works for most programs
– Supported by Linux kernel from 2004
– Supported by Windows XP service pack 1 and Windows Server 2003
• Called DEP – Data Execution Prevention
• Does not work for some programs that NEED to execute from the
stack.
– Eg. JIT Compiler, constructs assembly code from external data and then
executes it.
(Need to disable the W^X bit, to get this to work)

26
26
 Will non executable
stack prevent buffer
overflow attacks ?

Return – to – LibC Attacks

(Bypassing non-executable stack

during exploitation using return-
to-libc attacks)

https://css.csail.mit.edu/6.858/2010/readings/return-to-libc.pdf 27
27
 Return to Libc
(big picture)
BA
BA
BA
BA
Return Address BA
This will not work if ND bit is set
BA
BA
BA
buffer
Exploit code

28
28
 Return to Libc
(replace return address to point to a function within libc)

F1 Addr
F1 Addr Stack

F1 Addr
F1 Addr
Heap
Return Address F1 Addr
F1 Addr
F1 Addr Data
F1 Addr
buffer
F1 Addr

Text

Bypasses W^X since F1 is in the code segment,

And can be legally executed.
29
29
 F1 = system()
• One option is function system present in libc
system(“/bin/bash”);
would create a bash shell

(there could be other options as well)

So we need to
1. Find the address of system in the program
(does not have to be a user specified function, could be a function
present in one of the linked libraries)
2. Supply an address that points to the string
/bin/sh

30
30
 The return-to-libc attack
F1ptr
F1 ptr /bin/bash
Shell ptr
F1ptr
Return Address F1ptr
F1ptr
F1ptr system()
F1ptr In libc
buffer
F1ptr

31
31
Find address of system in the
executable

32
32
 Find address of /bin/sh
• Every process stores the enviroment variables at
the bottom of the stack
• We need to find this and extract the string
/bin/sh from it

33
33
Finding the address of the string
/bin/sh

34
 The final Exploit Stack
xxx
xxx /bin/sh
0xbfbffe25
dead
Return Address 0x28085260
xxx
xxx system()
xxx In libc
buffer
xxx

35
 A clean exit
xxx
xxx /bin/bash
0xbfbffe25
0x281130d0 exit()
Return Address 0x28085260 In libc
xxx
xxx system()
xxx In libc
buffer
xxx

36
 Limitation of ret2libc

Limitation on what the attacker can do

(only restricted to certain functions in the library)

These functions could be removed from the library

37
37
Return Oriented Programming
(ROP)

38
 Return Oriented Programming Attacks

• Discovered by Hovav Shacham of Stanford University

• Subverts execution to libc
– As with the regular ret-2-libc, can be used with non executable stacks since the
instructions can be legally execute
– Unlike ret-2-libc does not require to execute functions in libc (can execute any
arbitrary code)

The Geometry of Innocent Flesh on the Bone: Return-into-libc without Function Calls
(on the x86
39
 Target Payload
Lets say this is the payload needed to be executed by an attacker.

Suppose there is a function in libc, which has exactly this sequence of

instructions … then we are done.. we just need to subvert execution
to the function

What if such a function does not exist?

If you can’t find it then build it

40
 Step 1: Find Gadgets
• Find gadgets
• A gadget is a short sequence of instructions followed by a return
useful instruction(s)
ret
• Useful instructions : should not transfer control outside the gadget

• This is a pre-processing step by statically analyzing the libc library

41
 Step 2: Stitching
• Stitch gadgets so that the payload is built

movb $0x0, 0x7(%esi)

ret G2

movl $0xb, %eax

G4
ret

movb $0x0, 0xc(%esi)

ret G3

Ret instruction has 2 steps: movl %esi, 0x8(%esi)

G1
• Pops the contents pointed to by ESP into EIP ret
• Increment ESP by 4 (32bit machine)
Program Binary 42
 Step 3: Construct the Stack

xxx
AG4 movb $0x0, 0x7(%esi)
ret G2
AG3
movl $0xb, %eax
AG2 G4
ret
Return Address AG1 movb $0x0, 0xc(%esi)
ret G3
xxx
xxx
xxx movl %esi, 0x8(%esi)
buffer G1
ret
xxx

Program Binary

Program Stack
AGi: Address of Gadget i 43
 Finding Gadgets
• Static analysis of libc
• To find
1. A set of instructions that end in a ret (0xc3)
The instructions can be intended (put in by the compiler) or unintended
2. Besides ret, none of the instructions transfer control out of the
gadget

44
 Intended vs Unintended Instructions
• Intended : machine code intentionally put in by the compiler
• Unintended : interpret machine code differently in order to build new
instructions
Machine Code : F7 C7 07 00 00 00 0F 95 45 C3
What the compiler intended..

What was not ntended

Highly likely to find many diverse instructions of this form in x86; not so likely to
have such diverse instructions in RISC processors 45
 Geometry
• Given an arbitrary string of machine code, what is the
probability that the code can be interpreted as useful
instructions.
– x86 code is highly dense
– RISC processors like (SPARC, ARM, etc.) have low geometry
• Thus finding gadgets in x86 code is considerably more easier
than that of ARM or SPARC
• Fixed length instruction set reduces geometry

46
 Finding Gadgets
• Static analysis of libc
• Find any memory location with 0xc3 (RETurn instruction)
• Build a trie data structure with 0xc3 as a root
• Every path (starting from any node, not just the leaf) to the root is a
possible gadget
C
3
child of

00 46

24 89
24 43

94
37 16

47
 Finding Gadgets

33 b2 23 12 a0 31 a5 67 22 ab ba 4a 3c c3 ff ee ab 31 11 09

• Scan libc from the beginning toward the end

• If 0xc3 is found
– Start scanning backward
– With each byte, ask the question if the subsequence forms a valid
instruction
– If yes, add as child
– If no, go backwards until we reach the maximum instruction length (20
bytes)
– Repeat this till (a predefined) length W, which is the max instructions in
the gadget

48
Finding Gadgets Algorithm

49
Finding Gadgets Algorithm

Found 15,121 nodes in

~1MB of libc binary

is this sequence of instructions valid x86 instruction?

Boring: not interesting to look further;

Eg. pop %ebp; ret;;;; leave; ret (these are boring if we want to ignore intended instructions)
Jump out of the gadget instructions
50
 More about Gadgets
• Example Gadgets
– Loading a constant into a register (edx ß deadbeef)

pop %edx
deadbeef ret
esp GadgetAdd

• A previous return will pop the gadget address int %eip

• %esp will also be incremented to point to deadbeef
stack (4 bytes on 32 bit platform)
• The pop %edx will pop deadbeef onto the stack and increment
%esp to point to the next 4 bytes on the stack

51
 Stitch
G1
pop %edx
G2
ret
addr G2
esp G1 mov 64(%edx), %eax
ret
+64

Load arbitrary data into eax register using

Gadgets G1 and G2
deadbeef

stack

52
 Store Gadget
• Store the contents of a register to a memory location in the
stack
mov %eax, 24(%edx)
ret
GadgetAddr 2
0
esp GadgetAddr 1 pop %edx
ret
24

stack

53
 Gadget for addition
Add the memory pointed
to by %edx to %eax.
The result is stored in %eax

addl (%edx), %eax

push %edi pushes %edi.. onto the stack
Modified
GadgetAddr2 why is this present?
ret …. This is unnecessary, but
esp GadgetAddr
this is best gadget that we can
find for addition
But can create problems!!

stack We need work arounds!

54
 Gadget for addition
(put 0xc3 into %edi)
1. First put gadget ptr for 0xC3 into
%edi
2. 0xC3 corresponds to NOP in
addl (%edx), %eax ROP
GadgetAddr3 push %edi 3. Push %edi in gadget 2 just pushes
Gadget_RET
ret 0xc3 back into the stack
GadgetAddr2
Therefore not disturbing the stack
Gadget_RET
0xc3 contents
esp GadgetAddr1
4. Gadget 3 executes as planned
stack pop %edi
ret

0xc3 is ret ; in ROP ret is equivalent to NOP v

55
 Unconditional Branch
in ROP
• Changing the %esp causes unconditional
jumps

pop %esp
ret
esp GA

stack

56
 Conditional Branches
In x86 instructions conditional branches have 2 parts

1. An instruction which modifies a condition flag (eg CF, OF, ZF)

eg. CMP %eax, %ebx (will set ZF if %eax = %ebx)
2. A branch instruction (eg. JZ, JCC, JNZ, etc)
which internally checks the conditional flag and
changes the EIP accordingly

In ROP, we need flags to modify %esp register instead of EIP

Needs to be explicitly handled

In ROP conditional branches have 3 parts

1. An ROP which modifies a condition flag (eg CF, OF, ZF)

eg. CMP %eax, %ebx (will set ZF if %eax = %ebx)
2. Transfer flags to a register or memory
3. Perturb %esp based on flags stored in memory

57
 Step 1 : Set the flags
Find suitable ROPs that set appropriate flags
CMP %eax, %ebx subtraction
RET Affects flags CF, OF, SF, ZF, AF, PF

NEG %eax 2s complement negation

RET Affects flags CF

58
 Step 2: Transfer flags to
memory or register
• Using lahf instruction
stores 5 flags (ZF, SF, AF, PF, CF) in the %ah register

• Using pushf instruction where would one

pushes the eflags into the stack use this
instruction?

ROPs for these two not easily found.

A third way – perform an operation whose result depends on the flag
contents.

59
Step 2: Indirect way to transfer flags to
memory
Several instructions operate using the contents of the flags

ADC %eax, %ebx : add with carry; performs eax <- eax + ebx + CF

(if eax and ebx are 0 initially, then the result will be either 1 or 0 depending on the CF)

RCL : rotate left with carry;

RCL %eax, 1
(if eax = 0. then the result is either 0 or 1 depending on CF)

60
Gadget to transfer flags to memory

%edx will have value A

%ecx will contain 0x0

61
 Step 3: Perturb %esp depending
on flag
What we hope to achieve
If (CF is set){
perturb %esp What we have One way of achieving …
}else{ CF stored in a memory location (say X) negate X
leave %esp as it is Current %esp offset = delta & X
} delta, how much to perturb %esp %esp = %esp + offset

1. Negate X (eg. Using instruction negl)

finds the 2’s complement of X
if (X = 1) 2’s complement is 111111111…
if (X = 0) 2’s complement is 000000000...
2. offset = delta if X = 1
offset = 0 if X = 0
3. %esp = %esp + offset if X = 1
%esp = %esp if X = 0

62
 Turing Complete
• Gadgets can do much more…
invoke libc functions,
invoke system calls, ...
• For x86, gadgets are said to be turning complete
– Can program just about anything with gadgets
• For RISC processors, more difficult to find gadgets
– Instructions are fixed width
– Therefore can’t find unintentional instructions
• Tools available to find gadgets automatically
Eg. ROPGadget (https://github.com/JonathanSalwan/ROPgadget)
Ropper (https://github.com/sashs/Ropper)

63
Address Space Layout Randomization
(ASLR)

64
 The Attacker’s Plan
• Find the bug in the source code (for eg. Kernel) that can be
exploited
– Eyeballing
– Noticing something in the patches
– Following CVE
• Use that bug to insert malicious code to perform something
nefarious
– Such as getting root privileges in the kernel

Attacker depends upon knowning where these functions reside in

memory. Assumes that many systems use the same address mapping.
Therefore one exploit may spread easily

65
 Address Space Randomization
• Address space layout
randomization (ASLR)
randomizes the address space
layout of the process
• Each execution would have a
different memory map, thus
making it difficult for the attacker
to run exploits
• Initiated by Linux PaX project in
2001
• Now a default in many operating
systems

Memory layout across boots for a Windows box

66
 ASLR in the Linux Kernel
• Locations of the base, libraries, heap, and stack can be randomized in a
process’ address space

• Built into the Linux kernel and controlled by

/proc/sys/kernel/randomize_va_space

• randomize_va_space can take 3 values

0 : disable ASLR
1 : positions of stack, VDSO, shared memory regions are randomized
the data segment is immediately after the executable code
2 : (default setting) setting 1 as well as the data segment location is
randomized

67
ASLR in Action

First Run

Another Run

68
 ASLR in the Linux Kernel
• Permanent changes can be made by editing the /etc/sysctl.conf file

/etc/sysctl.conf, for example:

kernel.randomize_va_space = value
sysctl -p

69
Internals : Making code relocatable
• Load time relocatable
– where the loader modifies a program executable so
that all addresses are adjusted properly
– Relocatable code
• Slow load time since executable code needs to be modified.
• Requires a writeable code segment, which could pose
problems
• PIE : position independent executable
– a.k.a PIC (position independent code)
– code that executes properly irrespective of its absolute address
– Used extensively in shared libraries
• Easy to find a location where to load them without overlapping with
other modules

70
 Load Time Relocatable
1

71
Load Time Relocatable
2

note the 0x0 here…

the actual address of mylib_int is not filled in

72
 Load Time Relocatable

Relocatable table present in the executable

that contains all references of mylib_int
3

73
 Load Time Relocatable

The loader fills in the actual address of mylib_int

at run time.
4

74
 Load Time Relocatable
Limitations
• Slow load time since executable code needs to be modified

• Requires a writeable code segment, which could pose problems.

• Since executable code of each program needs to be customized, it

would prevent sharing of code sections

75
 PIC Internals
• An additional level of indirection for all global data and
function references
• Uses a lot of relative addressing schemes and a global offset
table (GOT)
• For relative addressing,
– data loads and stores should not be at absolute addresses but must be
relative

Details about PIC and GOT taken from …

http://eli.thegreenplace.net/2011/11/03/position-independent-code-pic-in-shared-libraries/ 76
 Global Offset Table (GOT)
• Table at a fixed (known) location in memory
space and known to the linker
• Has the location of the absolute address of
variables and functions
Without GOT

With GOT

77
Enforcing Relative Addressing
(example)
With load time relocatable

With PIC

78
 Enforcing Relative Addressing
(example)
With load time relocatable

With PIC
Get address of next instruction
to achieve relativeness

Index into GOT and get the

actual address of mylib_int into
eax

Now work with the actual

address.

79
 Advantage of the GOT
• With load time relocatable code, every variable reference would need to
be changed
– Requires writeable code segments
– Huge overheads during load time
– Code pages cannot be shared
• With GOT, the GOT table needs to be constructed just once during the
execution
– GOT is in the data segment, which is writeable
– Data pages are not shared anyway
– Drawback : runtime overheads due to multiple loads

80
 An Example of working with GOT

$gcc –m32 –shared –fpic –S got.c

Besides a.out, this compilation also generates got.s

The assembly code for the program

81
Data section

Text section

The macro for the GOT is known by the linker.

%ecx will now contain the offset to GOT

Load the absolute address of myglob from the

GOT into %eax

Fills %ecx with the eip of the next

instruction.
Why do we need this indirect way of doing this?
In this case what will %ecx contain?
82
More

offset of myglob
in GOT

GOT it!

83
 Deep Within the Kernel
loading the executable
(randomizing the data section)

Check if randomize_va_space
is > 1 (it can be 1 or 2)

Compute the end of the data

segment (m->brk + 0x20)

Finally Randomize

84
 Function Calls in PIC
• Theoretically could be done similar with the data…
– call instruction gets location from GOT entry that is filled in during load
time (this process is called binding)
– In practice, this is time consuming. Much more functions than global
variables. Most functions in libraries are unused
• Lazy binding scheme
– Delay binding till invocation of the function
– Uses a double indirection – PLT – procedure linkage table in addition
to GOT

85
 The PLT
• Instead of directly calling func, invoke an offset in the
PLT instead.
• PLT is part of the executable text section, and consists
of one entry for each external function the shared
library calls.
1 • Each PLT entry has
a jump location to a specific GOT entry
Preparation of arguments for a ‘resolver’
Call to resolver function

86
 First Invocation of Func
First Invocation of fun (steps 2 and 3)
On first invocation of func, PLT[n]
jumps to GOT[n], which simply jumps
back to PLT[n]
1

87
 First Invocation of Func
(step 4). Invoke resolver, which resolves
the actual of func,
places this actual address into GOT
and then invokes func

The arguments passed to resolver, that

1 helps to do symbol resolution

Note that the contents of GOT is now

changed to point to the actual address
of func

4 3

88
Example of PLT

Compiler converts the call to set_mylib_int

into set_mylib_int@plt

89
Example of PLT
ebx points to the GOT table
ebx + 0x10 is the offset
corresponding
to set_mylib_int

Offset of set_mylib_int in the

GOT (+0x10).
It contains the address of the
next instruction (ie. 0x3c2)

90
Example of PLT
Jump to the resolver, which
resolves the actual address
of set_mylib_int and fills it
into the GOT

Push arguments for the

resolver.

Jump to the first entry of the PLT

Ie. PLT[0]

91
Subsequent invocations of Func

3
2

92
 Advantages
• Functions are relocatable, therefore good for ASLR
• Functions resolved only on need, therefore saves
time during the load phase

93
 Bypassing ASLR
• Brute force
• Return-to-PLT
• Overwriting the GOT
• Timing Attacks

94
Safer Programming Languages,
and Compiler Techniques

95
 Other Precautions for buffer overflows
• Enforce memory safety in programming language
– Example java, C# (slow and not feasible for system programming)
• Cannot replace C and C++.
(Too much software already developed in C / C++)

– Newer languages like Rust seem promising

• Use securer libraries. For example C11 annex K, gets_s, strcpy_s,

strncpy_s, etc.
(_s is for secure)

96
 Compile Bounds Checking
• Check accesses to each buffer so that it cannot be beyond the
bounds
• In C and C++, bound checking performed at pointer calculation time
or dereference time.
• Requires run-time bound information for each allocated block.
• Two methodologies
– Object based techniques
– Pointer based techniques

Softbound : Highly Compatible and Complete Spatial Memory Safety for C

Santosh Nagarakatte, Jianzhou Zhao, Milo M. K. Martin, and Steve Zdancewic 97
 Softbound
• Every pointer in the program is associated with a base and bound
• Before every pointer dereference to verify to verify if the dereference is
legally permitted

These checks are automatically inserted at compile time for all pointer
variables. For non-pointers, this check is not required.

98
 Softbound – more details
• pointer arithmetic and assignment
The new pointer inherits the base and bound of the original
pointer

No specific checks are required, until dereferencing is done

99
 Storing Metadata
• Table maintained for metadata

100
 Softbound – more details
• Pointers passed to functions
– If pointers are passed by the stack
no issues. The compiler can put information related to metadata onto
the stack
– If pointers passed by registers.
Compiler modifies every function declaration to
add more arguments related to metadata
For each function parameter that is a pointer, the corresponding base
and bound values are also sent to the function

101