This document discusses various techniques for optimizing code on ARM processors, including using conditional instructions, benchmarking with cycle counts, utilizing hardware features like multiplication and DMA, choosing optimal data structures and algorithms, and using mutexes and exclusive monitors for thread synchronization. Some key points covered are using bitwise operations instead of shifts/masks when possible, structs for packing data efficiently in memory, and preferring to reuse existing libraries over reimplementing functionality.

1. 1
Embedded System - ARM
Development and Optimization Sharing Session
 Thumb2 Conditional Code
 Benchmark using cycle count
 Optimization Technique
 Mutex & ARM Exclusive Monitor
竹內宏輝 hirokiht <hiroki04030@yahoo.com>

2. 2
Thumb2 Conditional Code
Thumb Thumb2
ISA 16bit 16bit/32bit
Conditional
Only supported in
Code
branch
Use IT-block
Bitwise - BFC/BFI/SBFX
/UBFX
Table-branch
TBB/TBH
List of Condition Code:
 Eq, hs, mi, vs, hi, ge, gt,
al, ne, lo, pl, vc, ls, lt, le
 Example:
ITETT EQ
MOVEQ r0, #1
MOVNE r0, #0
MOVEQ r1, #0
MOVEQ r2, #0

3. 3
Thumb2 Conditional Code

4. 4
Benchmark using cycle count
Naive Approach:
 Using cycle
count/systick/system time
difference, less is better.
 QEMU doesn't have
systick, so can't be
emulated. Have to use real
hardware.
Commercial Approach
 Use SDK/Development
Kit/Workbench, highly
emulated mcu. Very
expensive....

5. 5
Benchmark using cycle count
My Thoughts:
 For low-end hardware, hardware is cheaper
than the commercial development tool, just
buy and plug it in. The best benchmark is to do
a REAL benchmark on hardware.

6. 6
Optimization Technique
Choose the best compilation option,
fully utilize the hardware (if
possible)
 Hardware Multiplication and Division
 DSP
 FPU
 DMA
 Saturation
 Thumb2 ISA
 Hardware controller - Ethernet, Graphic,
SDIO, RTC, Encoder/Decoder, etc

7. 7
Optimization Technique
 If the operation involve 2n, think again,
maybe bitwise operator can help
 Multiplication x<<n = x*2^n
 Division x >> n = x/2^n (Note, lose
precision)
 Modulus x&(n-1) = x%2^n
 Check is 2^n or not: x & (x-1) == 0
 Swap (without using tmp):
– x ^= y; y ^= x; x ^= y;

8. 8
Optimization Technique (for C)
 Pointer is your friend, learn it, use it! (Esp. in memory limited
environment)
 Struct/Union is the best tool to pack data
– C99 support bitwise struct. (Older version can use bitwise shifting and
masking)
– Variable order is important, in memory and in files.
 If many boolean variable is needed, consider using "flag"
implementation
 #define can help to make better code readability, also can be used
to define macro

9. 9
Optimization Technique
Best Code Density != Best Code Performance
(most of the time)
 Decide before start to develop
 Again use the compilation option accordingly

10. 10
Optimization Technique
Algorithm is important, choose the one that
best fit your application.
 Searching, Sorting, Random, Hashing, etc
Data Structure is more important!
 But don't overkill, no need to use an indexed list, to
store 10 single-attribute item, a simple array will do.

11. 11
Optimization Technique
But sometimes overkill is good:
 If the software need 128MB RAM, no harm giving
512MB (Memory is cheap, Time is expensive to debug).
Breathing room for further software modification.
 A well structured program may not be needed for small
project, but program only grows bigger and bigger.

12. 12
Optimization Technique
Standards and Protocols are complicated, but
they're there for some reason.....
 Well designed, implemented, and tested.
 Compatible with other software/hardware.

13. 13
Optimization Technique
Don't reinvent the wheel, make the wheel
better.
 Google/Stack Overflow is your friend, see how other
people implement and why.
 If there's a (open source) library, why not use it?

14. 14
Mutex & ARM Exclusive Monitor
What is Mutex?
 Mutual Exclusion - Make sure no more than one
concurrent process is in their critical section, so that
race condition will not occur. Mutex only have two
states -- "locked" and "unlocked".

15. 15
Mutex & ARM Exclusive Monitor
Mutex vs Binary Semaphore
 Semaphore is to control access by multiple process to a common resource. Typical
producer-consumer problem can be solved using semaphore.
 Binary semaphore is semaphore variable with only 0 or 1. Means only one process can
access the resource at the time. Then how is it different with mutex?
 Mutex have a owner concept, only the process who locks it can unlocks it. But
semaphore doesn't, process A may be using the resource, and process B contain a
bug which "release" the resource it doesn't own, and process C will attempt to use it,
and causing concurrent access with process A.

16. 16
Mutex & ARM Exclusive Monitor
ARM Exclusive Monitor
 Each processor that supports
exclusive accesses has a local
monitor. As the name indicate, "local
monitor" is to monitor the memory
which is local (shared and non-shared).
Whereas, "global monitor" is
to monitor shared memory. Access to
shared memory will be checked for
both local and global monitor.

17. 17
Mutex & ARM Exclusive Monitor
 Since during "exclusive" state, access to the specific memory is
only done by one process/thread.
 To enter the "exclusive" state, we must use LDREX, if we decided
to update the value, we use STREX to update the value and return
to the "open access" state; otherwise it is advisable that we use
CLREX to return to "open access" state without change any value.
 If we attempt to use STREX during "open access" state, it will fail.
For safety purpose, it is necessary to check if an operation has
been succeeded.