Optimization Techniques in C

Team Emertxe
Optimization Techniques

 Basic Concepts
 Programming Algorithm and Techniques
Optimization Techniques
Basic Concepts

 What is Optimization
 Methods
 Space and Time Optimization
 Evaluation
 When not to Optimize
 Identify Critical Area
 Confusion to Source Level Debugger
Optimization Techniques
Methods

 Choice of Complier
 Compiler setting
 Programming Algorithm and Techniques
 Rewrite Program in Assembly
 Code Profilers
 Disassemble Code Analysis
Optimization Techniques
Programming Algorithm
& Techniques

 Data Handling
 Flow Control
 Other Handling
Optimization Techniques
Data Handling

 RAM Usage
 Data Type Usage
 Avoid Type Conversions
 Unsigned Advantage
 Float and Double
 Constant and Volatile
 Data Alignment – Arrangement and Packing
 Pass by Reference
 Register Store and Restore
Optimization Techniques
Flow Control

 Use of Switch Statement

 Inline Function
 Loop Unrolling
 Loop Hoisting
 Loop Overhead
Optimization Techniques
Other Handling

 Use of Operators
 Inline Assembly
 Fixed Point and Floating Point
Definition

Optimization is a process of improving

efficiency of a program in time (speed) or
space (size)
Basic iIlustration
Time and Space Optimization
//Smaller size
//Faster Speed
main()
{
main()
...
{
function();
...
...
XXXXX
function();
YYYYY
...
ZZZZZ
function();
...
...
XXXXX
}
YYYYY
ZZZZZ
void function()
...
{
XXXXX
XXXXX
YYYYY
YYYYY
ZZZZZ
ZZZZZ
...
}
}
Cont..
 It can be observed that optimization with one method may
affect the other.
 A general phenomenon is faster operating code will have a
bigger code size, whereas smaller code size will have a
slower execution speed. However this may not be always
true.
Evaluation
 A program’s optimization level can be evaluated based on the
measurement of:
 Total code size
 Total number of execution cycles (time taken).
 These are determined by the basic component of a program,
which is the assembly code (Instruction set / Op-code /
Mnemonic).
Cont..
 In the MCU manual, these assembly codes characteristic are
detailed:
 Instruction length -> Determine Code Size
 Number of execution states -> Determine Execution Speed
Do Not Optimize
It is suggested that unless necessary, optimization shall be
omitted. This process should be planned for, and not done at
the end of the development cycle, whereby most scenarios had
been tested. This may cause changes to the initial design and
introduce more wastage of time and resources in debugging.
Data Type Usage
The use of correct data type is important in a recursive
calculation or large array processing. The extra size, which is
not required, is taking up much space and processing time.
RAM Usage
 Shortage of RAM space is a common concern. The
nominal RAM size for most 8-bit MCU is a mere 1 to 4K
bytes size.
 Three main components of RAM are:
• Stack
• Heap
• Global data
 The reduction in one component will enable the increase
in the others. Unlike stack and heap, which are dynamic
in nature, global data is fixed in size.
 Programmers may like to consider the reduction of
global data usage, and place more emphasis on local
variables control in stack and registers.
 I/O considerations
Instead of accessing external bus for two times, it may be
better to read the data as a word (16-bit external data bus),
and process it as a byte.
Use Variables of the
Same Type for Processing

 Programmers should plan to use the same type of variables

for processing. Type conversion must be avoided.
 Otherwise, precious cycles will be waste to convert one type
to another (Unsigned and signed variables are considered as
different types).
Use of Unsigned Type

 All variables must be defined as “unsigned” unless

mathematical calculation for the signed bit is necessary.
 The “signed-bit” may create complication, unwanted failure,
slower processing and extra ROM size.
Float and Double
 Maximum value of Float = 0x7F7F FFFF
 Maximum value of Double = 0x7F7F FFFF FFFF FFFF
 To avoid the unnecessary type conversion or confusion,
programmers can assign the letter “f” following the numeric
value.
 x = y + 0.2f;
 Data Declaration - Constant

 The “const” keyword is to define the data as a constant,

which will allocate it in the ROM space.
 Otherwise a RAM space will also be reserved for this data.
This is unnecessary as the constant data is supposed to be
read-only.
Data Initialization at
Declaration

Data should be initialized at declaration.

Data Definition
Arrangement and Packing

 The declaration of the components in a structure will

determine how the components are being stored.
 Due to the memory alignment, it is possible to have dummy
area within the structure. It is advised to place all similar
size variables in the same group.
Data Arrangement
 Structure Padding
 As the structure is packed, integer b will not be aligned.
 This will improve the RAM size but operational speed will be degraded, as
the access of ‘b’ will take up two cycles.
Global and Local Variables

 Local variable is preferred over the global variable in a function.

 Generally, global variables are stored in the memory, whereas, local
variables are stored in the register.
 Since register access is faster than the memory access, implementing
local variables will improve speed operation.
 Moreover, code portability also encourages the use of local variables.
 However if there are more local variables than the available registers, the
local variables will be temporary stored in the stack.
Passing Reference as Parameters

 Larger numbers of parameters may be costly due to the number of

pushing and popping actions on each function call.
 It is more efficient to pass structure reference as parameters to reduce
this overhead.
Return Value
 The return value of a function will be stored in a register.
 If this return data has no intended usage, time and space are wasted in
storing this information.
 Programmer should define the function as “void” to minimize the extra
handling in the function.
Array & Structure
Initialization

A simple illustration of implementation:

Inline Function
 The technique will cause the compiler to replace all calls to the function,
with a copy of the function’s code.
 This will eliminate the runtime overhead associated with the function
call.
 This is most effective if the function is called frequently, but contains
only a few lines of code.
Inline function
SWITCH & and
Non-contiguous Case Expressions

 Use if-else statements in place of switch statements that have

noncontiguous case expressions.
 If the case expressions are contiguous or nearly contiguous integer values,
most compilers translate the switch statement as a jump table instead of
a comparison chain.
Switch contd...
Jump tables generally improve performance because:
 They reduce the number of branches to a single procedure call.
 The size of the control-flow code is the same no matter how many cases
there are.
 The amount of control-flow code that the processor must execute is the
same for all values of the switch expression.
Switch contd...
 However, if the case expressions are noncontiguous values,
most compilers translate the switch statement as a
comparison chain.
 Comparison chains are undesirable because:
• They use dense sequences of conditional branches, which interfere with
the processor’s ability to successfully perform branch prediction.
• The size of the control-flow code increases with the number of cases.
• The amount of control-flow code that the processor must execute varies
with the value of the switch expression.
Switch – Example 1
A switch statement like this one, provides good
performance:
Switch – Example 2
Because the case expressions in the following switch statement are not
contiguous values, the compiler will likely translate the code into a
comparison chain instead of a jump table:
If-Else statements
To avoid a comparison chain and its undesirable effects on branch prediction,
replace the switch statement with a series of if-else statements, as follows:
Arranging Cases by
Probability of Occurrence
 Arrange switch statement cases by probability of occurrence, from most
probable to least probable.
 Avoid switch statements such as the following, in which the cases are not
arranged by probability of occurrence:
Switch case - Contd...

 Instead, arrange the cases to test for frequently occurring

values first:
Use of Function
Prototypes
 In general, use prototypes for all functions.
 Prototypes can convey additional information to the compiler
that might enable more aggressive optimizations.
Generic Loop
Hoisting
 To improve the performance of inner loops, reduce redundant
constant calculations
 Avoid:
Generic Loop
Hoisting
 Preferred optimization
Contd...
 Un-optimized code
Contd...
 Local Static
Functions
 Declare as static functions that are not used outside the file
where they are defined.
 Declaring a function as static forces internal linkage.
 Functions that are not declared as static default to external
linkage, which may inhibit certain optimizations—for
example, aggressive in-lining—with some compilers.
Explicit Parallelism
in Code
 Where possible, break long dependency chains into several independent
dependency chains that can then be executed in parallel, exploiting the
execution units in each pipeline.
 Avoid:
Contd...

Notice that the four-way unrolling is chosen to exploit the

four-stage fully pipelined floating-point adder. Each stage of
the floating-point adder is occupied on every clock cycle,
ensuring maximum sustained utilization.
Extracting Common
Sub expressions

 Manually extract common sub expressions where C

compilers may be unable to extract them from
floating-point expressions due to the guarantee
against reordering of such expressions in the ANSI
standard.
Contd...
Manually optimize – Example 1
Contd...
Manually optimize – Example 2
Replacing Integer Division
with Multiplication
 Replace integer division with multiplication when there are multiple
divisions in an expression.
 This is possible only if no overflow will occur during the computation of
the product. The possibility of an overflow can be determined by
considering the possible ranges of the divisors.
 Integer division is the slowest of all integer arithmetic operations.
Contd...
Example - Replace division by multiplication
Sign of Integer
Operands
 Where there is a choice of using either a signed or an
unsigned type, take into consideration that some operations
are faster with unsigned types while others are faster for
signed types.
Contd...
Contd...
Signed vs. Unsigned int division
Contd...
 In summary, use unsigned types for:
• Division and remainders
• Loop counters
• Array indexing
 Use signed types for:
• Integer-to-floating-point conversion
Loop Unrolling
Contd...
Loop Invariant
IF Code Floating

Incorrect
Else Clause
Removal

 A jump condition is inevitable in the first case, whereas the

second case gives the higher possibility process (Condition_A)
to have a sequential execution. This technique is possible
only if Undo “DoThat()” process is possible.
Loop Overhead
 The MCU have a conditional branch mechanism that works
well when counting down from positive number to zero.
 It is easier to detect “Zero” (Branch non zero – single
instruction) than a “predefined number” (Compare and
perform jump – two instructions)
Specify Optimization
Type for Each Module
#pragma option can be used to limit and control the
optimization regions.
 Horner’s Rule
of Polynomial Evaluation

The rules state that a polynomial can be rewritten as a nested

factorization. The reduced arithmetic operations will have
better efficiency and execution speed.
Factorization
The compiler may be able to perform better when the formula
Use Finite Differences
to Avoid Multiplies
Modula
Constant in Shift
Operations
 For shift operations, if the shift count is a variable, the
compiler calls a runtime routine to process the operation.
 If the shift count is a constant, the compiler does not call the
routine, which give significant speed improvement.
 Use Formula
 Use formula instead of loops
 What is the formula for sum of naturals numbers?
Contd...
Simplify Condition
Un-optimized – Example 1
Contd..
Optimized
Contd...
Contd...
Contd...
Contd...
Absolute value
Profiling
 What is profiling
 Memory profiling
 Leaks
 Memory usage
 CPU Usage
 Cache
 Stay connected
About us: Emertxe is India’s one of the top IT finishing schools & self learning kits
provider. Our primary focus is on Embedded with diversification focus on Java, Oracle
and Android areas

Emertxe Information Technologies,

No-1, 9th Cross, 5th Main,
Jayamahal Extension,
Bangalore, Karnataka 560046
T: +91 80 6562 9666
E: [email protected]

https://www.facebook.com/Emertxe https://twitter.com/EmertxeTweet https://www.slideshare.net/EmertxeSlides

THANK YOU