ARM专题（s3c2410a TDMI）

ARM Platform Programming Optimizing

            By Stephen Du

            07-5-16

Part 1 Data Types

•

1.Reason:

    ARM support only 32bits register, so if we don’t choose the proper data type, more instruction will be needed!

2.Detail:

*Use 32bit data types as possible(if mod or % is needed don’t follow this!)

   *Use 32bit data and convert to the type at last

   *Use pointer while not index for array(for short type ldrh don’t support shit addressing)

   *unsigned is better for dividing other the same efficiency

  

*Use as narrow as possible for storing!(armv4 can proccess 8 16 32 the same!)

•*Use explicit convert as possible

•*load and store will convert type automatically!

Part 2 Loop

•1.Fixed count loop should use 0 (--)to end while not ++ to “for” and “do while” is better than “for”(but do while at least once!!)2.To unsigend type use “!=0” is better than “>0”

•3.If possible break up the inloop sentences to request

•4.Reduce the number of variables in loop

 

•For 3

•Int checksum(int *data,unsigned int N){

•unsigned int i;

•int sum = 0;

•for(i=N/4;i!=0;i--){

•sum+=*(data++);

•sum+=*(data++);

•sum+=*(data++);

•sum+=*(data++);

•}

•for(i=N&3;i!=0;i--){

•Sum+=*(data++);

•}

•Return sum

•}

Part 3 Pointer


Alias

•
1.Pointer alias will cause loading the same data from memory twice or more so avoid it by using a local variable

•
2.Avoid using local variable address!!

Part 4 Struct


Padding

1.Proper allocating of struct member will save memory but wasting time,__packed option will force to kick out pads! For cross platform should avoid using __packed and enum type! In struct 8bit should before 16 and so is 16 32 64,the bigger are backer! Avoiding big struct with more levels;Adding pad by hand can do good to cross paltform

Part 5 Inline And


Function Parameters

•
1.Limit the number of parameters to less than 4,because according APCS, the more than 4 or long long alike parameters have to be pushed into  the stack, and these actions make the program cost more time;

•
2.If parameters are too many, make them a structure and pass the structure pointer is a good idea for avoiding pushing and pop;

•
3.Don’t use more than 12 local variable so that the parameters can be allocated register while not stack!

•
3.Put the functions in the same file as possible, and inline the small functions(don’t inline the big ones!!!);

•unsigned int dec16_to_hex(unsigned int dec16){

•If(dec16<10){

•return dec16+’0’;

•}

•return dec16-10+’A’;

•}

•void uint_to_hex(char*out,unsigned int in){

–in = in<<4|in>>28;

–*out++ = (char)dev16_to_hex(in&15);

•}

Part 6 Bit Field

•

1.Avoid using bit-field, use #define or enum instead!

•Bad example:

•void A(); void B(); void C();

•typedef struct{

unsigned int a:1;

unsigned int b:1;

unsigned int c:1;

}P;

•void bit(P* p){

•If(p->a)

•A();

•If(p->b)

•B();

•If(p->c)

•C();

•}

This cause the access of bit field by pointer so that load data more than normal!

•Good ones:

•nsigned int P;

• #define Ba (1u<<0)

•#define Bb (1u<<1)

•#define Bc (1u<<2)

•void bit(unsigned int p){

•If(p&Ba)

•A();

•If(p&Bb)

•B();

•If(p&Bc)

•C();

•}

•
2.Use and or xor is good idear!

Part 7 Edge Aligned And Endian Related 

•1.Be care about __attribute__ ((packed));

•It leads to code uncross usage!Use char type pointer to access the data to avoid the edge aligned problems.

•
2.When data to be processed  is different from our cpu  on endian(cpu is little-endian while data is big endian), just use 8 bit type to access it with 2 times to 16bit origin 4 times to 32 bit origin and also write different sub-code is a good idea!

•

Sample:

 void read_audio_data(short* out,char* in,unsigned int N){

unsigned short* data;

unsigned int sample,next;

switch(in&1){

case 0:

do{

data = (unsigned short*)in;

sample = *(data++);

#ifdef __BIG_ENDIAN

Sample = sample<<8|sample>>8;

#endif

*(out++) =sample;

}while(N--);

break;

case 1:

data = (unsigned short*)(in-1);

Sample = *(data++);

#ifdef __BIG_ENDIAN

Sample = sample&0xff;

#else sample = sample>>8;

#endif

do{

next =*(data++);

#ifdef __BIG_ENDIAN

*(out++)=(short)(next&0xff00)|sample;

Sample =next&0xff;

#else *(out++)=(short)

((next&<<8)|sample);

Sample =next>>8;

#endif

break

}while(--N);

}

}

Part 8 About Divide

And Floating

•
1.ARM doesn’t realize divide, so the armcc will use a C library to simulate the divide function if in need! Avoiding using divide is a good idea! If have to, just use unsigned or const divide, because  they work faster, the signed divide cast 50 cycles ,so much time!

•
2.Sometimes mod or % is better than divide!

•
3.Make full use of the result a=w/c and b=w%c to avoid  divide;

•
4.Other logrithum so hard 

Part 9 Embedded

assemble

•
1.Sometimes embedded assemble works much better, such as the C don’t support types (coprocessor and armv5e extended command!)

•asm(“mcr cp15,00,c7,c5,0”);

Part 10 Cross

Platform

•
1.On ARM platform char is unsigned and if loop is ended with i>=0,it will turned to be infinite, we need to care about it! We can solve it by giving an gcc option or change the I type!

•
2.Old platform int is 16 bits so on arm they will be turned into 32 bits some judgment turned false

•i = 0xf000

•
3.Data  align

•
4.Big and little endian!

•
5.Function types( parameter types)

•
6.Bit-field

•
7.enumeration (use as less as possible)

•
8.Embedded assemble cause uncross

•
9.The port needed to use volatile so that the correct type and command used(ldrsh strsh or ldr str)