Analog Devices EE149 Application Notes

Engineer To Engineer Note EE-149

s

a

Technical Notes on using Analog Devices' DSP components and development tools

Contact our technical support by phone: (800) ANALOG-D or e-mail: [e-mail]
Or vi sit ou r on-l ine re sourc es ht tp:// www.analog.com/dsp and http://www.analog.com/dsp/EZAnswer

Tuning C Source Code for the Blackfin® Processor Compiler

Contributed by DSP Tools Compiler Group May 26, 2003

Introduction

This document provides some guidelines for
obtaining the best code execution performance
from the Blackfin® processor family’s C/C++
compiler using VisualDSP++™ release 2.0.

Use the optimizer

There is a vast difference in the performance of
C code that has been compiled optimized and
non-optimized. In some cases optimized code
can run ten or twenty times faster. Optimization
should always be attempted before measuring
performance or shipping code as product. Note
that the default setting is for non-optimized
compilation, the non-optimized default being
there to assist programmers in diagnosing
problems with their initial coding.

The optimizer in the Blackfin processor compiler
is designed to generate efficiently-executing
code from C which has been written in a
straightforward manner. The basic strategy for
tuning a program is to present the algorithm in a
way that gives the optimizer excellent visibility
of the operations and data, hence the greatest
freedom to safely manipulate the code. Note that
future releases will enhance the optimizer, and
expressing algorithms simply will provide the
best path for reaping the benefits of such
enhancements.

Use the Statistical Profiler

Tuning source begins with an understanding of
what areas of the application are the hot spots.
Statistical profiling provided in VisualDSP++ is
an excellent means for finding those hot spots.

If the application is unfamiliar to you, compile it
with diagnostics and run it unoptimized. This
will give you results that connect directly to the
C source. You will obtain a more accurate view
of your application if you build a fully optimized
application and obtain statistics that relate
directly to the assembly code. The only problem
may be in relating assembly lines to the original
source. Do not strip out function names when
linking. If you have the function names then you
can scroll the assembly window to locate the hot
spots. In very complicated code you can locate
the exact source lines by counting the loops –
unless they are unrolled. Look at the line
numbers in the

.s file. Note that the compiler

optimizer may have moved code around.

Data Types

char
unsigned char
short
unsigned short
int
unsigned int
long
unsigned long

8-bit signed integer
8-bit unsigned integer
16-bit signed integer
16-bit unsigned integer
32-bit signed integer
32-bit unsigned integer
32-bit signed integer
32-bit unsigned integer

Table 1: Fixed-Point Data Types (Native Arithmetic)

Copyright 2003, Analog Devices, Inc. All rights reserved. Analog Devices assumes no responsibility for customer product design or the use or application of
customers’ products or for any infringements of patents or rights of others which may result from Analog Devices assistance. All trademarks and logos are property
of their respective holders. Information furnished by Analog Devices Applications and Development Tools Engineers is believed to be accurate and reliable, however
no responsibility is assumed by Analog Devices regarding technical accuracy and topicality of the content provided in Analog Devices’ Engineer-to-Engineer Notes.

a

The compiler directly supports ten scalar data
types as shown in Table 1 and Table 2. double is
equivalent to float on Blackfin processors, since
64-bit values are not supported directly on the
hardware.

float
double

Table 2: Floating-Point Data Types (Emulated
Arithmetic)

32-bit floating point
32-bit floating point

Fractional data types can be represented as either

short or int. Manipulation of these types is best

done by using intrinsics, which will be described
in a subsequent section.

Avoid Float/Double Arithmetic

Floating-point arithmetic operations are
implemented by library routines and,
consequently, are far slower than integer
operations. An arithmetic floating-point
operation inside a loop will prevent the optimizer
from using a hardware loop.

Avoid Integer Division in Loops

The hardware does not provide direct support for
32-bit integer division, so the division and
modulus operations on int variables are multi­cycle operations. The compiler will convert an
integer division by a power of two to a right-shift
operation if the value of the divisor is known.

If the compiler has to issue a full division
operation, it will issue a call to a library function.
In addition to being a multi-cycle operation, this
will prevent the optimizer from using a hardware
loop for any loops around the division.
Whenever possible, do not use divide or modulus
operators inside a loop.

pointer. These two versions of the vector
addition illustrate the two styles:

void va_ind( short a[], short b[],
short out[], int n)
{
int i;
for (i = 0; i < n; ++i)
out[i] = a[i] + b[i];
}

Listing 1: Indexed Arrays

void va_ptr( short a[], short b[],
short out[], int n)
{
int i;
short *pout = out, *pa = a, *pb = b;
for (i = 0; i < n; ++i)
*pout++ = *pa++ + *pb++;
}

Listing 2: Pointers

Common thought might indicate that the chosen
style should not make any difference to the
generated code, but sometimes it does. Often,
one version of an algorithm will generate better
optimized code than the other, but it is not
always the same style that is better; the generated
code is affected by the surrounding code, which
is why there may be differences. The pointer
style introduces additional variables that compete
with the surrounding code for resources during
the optimizer’s analysis. Array accesses, on the
other hand, must be transformed to pointers by
the compiler, and sometimes it does not do the
job as well as you could do by hand.

The best strategy is to start with array notation. If
this looks unsatisfactory try using pointers.
Outside the important loops, use the indexed
style, because it is easier to understand.

Use the -ipa Switch

Indexed Arrays versus Pointers

C allows you to program data accesses from an
array in two ways: either by indexing from an
invariant base pointer or by incrementing a

Tuning C Source Code for the Blackfin® Processor Compiler (EE-149) Page 2 of 10

To ensure the best performance, the optimizer
often needs to know things that can only be
determined by looking outside the routine which
it is working on. In particular, it helps to know
the alignment and value of pointer parameters

a

and the value of loop bounds. The -ipa compiler
switch enables inter-procedural analysis (IPA),
which makes this information available. This
may be switched on from the IDDE by checking
the Interprocedural Optimization box in the Compile
tab of the Project Options dialogue selected from
the Project menu.

When this switch is used the compiler may be
called again from the link phase to recompile the
program using additional information obtained
during previous compilations.

Because it only operates at link time, the

L

effects of

-ipa will not be seen if you

compile with the -S switch. To see the
assembler file put -save-temps in the

Additional Options text box in the Compile

tab of the Project Options dialogue and
look at the .s file produced after your
program has been built.

Much of the following advice assumes that the

-ipa switch is being used.

Initialize Constants Statically

Inter-procedural analysis will also identify
variables that only have one value and replace
them with constants, which can enable better
optimization. For this to happen, a variable must
have a single value throughout the program.

#include <stdio.h>
static int val = 3; // initialized
// once
void init() {
}
void func() {
printf("val %d",val);
}
int main() {
init();
func();
}

Listing 3: Optimal (IPA knows val is 3)

If the variable is statically initialized to zero, as
all global variables are by default, and is
subsequently assigned to some other value at

another point in the program, then the analysis
sees two values and will not consider the variable
to have a constant value.

#include <stdio.h>
static int val; // initialized to zero
void init() {
val = 3; // re-assigned
}
void func() {
printf("val %d",val);
}
int main() {
init();
func();
}

Listing 4: Non-optimal (IPA cannot see that val is a
constant)

Word-align Your Data

To make most efficient use of the hardware, it
must be kept fed with data. In many algorithms,
the balance of data accesses to computations is
such that, to keep the hardware fully utilized,
data must be fetched with 32-bit loads.

Although the Blackfin architecture supports byte
addressing, the hardware requires that references
to memory be naturally aligned. Thus, 16-bit
references must be at even address locations, and
32-bit at word-aligned addresses. So, for the
most efficient code to be generated, you should
ensure that data are word-aligned.

The compiler helps establish the alignment of
array data. The stack frames are kept word­aligned. Top-level arrays are allocated at word­aligned addresses, regardless of their data types.

If you write programs that only pass the address
of the first element of an array as a parameter
and loops that process input arrays an element at
a time, starting at element zero, then inter­procedural analysis should be able to establish
that the alignment is suitable for 32-bit accesses.

Where the inner loop processes a single row of a
multi-dimensional array, be certain that each row

Tuning C Source Code for the Blackfin® Processor Compiler (EE-149) Page 3 of 10

a

begins on a word boundary, possibly inserting
dummy data to do so.

Loop Guidelines

Appendix A gives an overview of how the
optimizer transforms a loop to generate highly
efficient code. It describes the "loop unrolling"
optimization technique.

Do not unroll loops yourself

Not only does loop unrolling make the program
harder to read but it also prevents optimization.
The compiler must be able to unroll the loop
itself in order to use wide loads and both
accumulators automatically.

void va1(short a[], short b[],
short c[], int n)
{
int i;
for (i = 0; i < n; ++i) {
c[i] = b[i] + a[i];
}
}

Avoid loop-carried dependencies

A loop-carried dependency is where
computations in a given iteration of a loop
cannot be completed without knowing the values
calculated in earlier iterations. When a loop has
such dependencies, the compiler cannot overlap
loop iterations.

Some dependencies are caused by scalar
variables that are used before they are defined in
a single iteration.

for (i = 0; i < n; ++i)
x = a[i] - x;

Listing 7: Non-optimal (scalar dependency)

An optimizer can reorder iterations in the
presence of the class of scalar dependencies
known as reductions. These are loops that reduce
a vector of values to a scalar value using an
associative and commutative operator. The most
common example is multiply and accumulate.

for (i = 0; i < n; ++i)
x = x + a[i] * b[i];

Listing 5: Optimal (compiler unrolls and uses both
computational blocks)

void va2(short a[], short b[],
short c[], int n)
{
short xa, xb, xc, ya, yb, yc;
int i;
for (i = 0; i < n; i+=2) {
xb = b[i]; yb = b[i+1];
xa = a[i]; ya = a[i+1];
xc = xa + xb; yc = ya + yb;
c[i] = xc; c[i+1] = yc;
}
}

Listing 6: Non-optimal (compiler leaves 16-bit loads)

In this example, the first version of the loop runs
almost three times faster than the second, in
cases where inter-procedural analysis can
determine that the initial values of
aligned on 32-bit boundaries and

a, b, and c are

n is a multiple

of two.

Listing 8:Optimal (a reduction)

In the first case, the scalar dependency is the
subtraction operation; the variable x changes in a
manner which will give different results if the
iterations are performed out of order. In contrast,
in the second case, the properties of the addition
operator mean that the compiler can perform the
operations in any order and still get the same
result.

Do not rotate loops by hand

Loops in DSP code are often "rotated" by hand,
attempting to do loads and stores from earlier
and future iterations at the same time as
computation from the current iteration. This
technique introduces loop-carried dependencies
that prevent the compiler from rearranging the
code effectively. It is better to give the compiler
a "normalized" version, and leave the rotating to
the compiler.

Tuning C Source Code for the Blackfin® Processor Compiler (EE-149) Page 4 of 10

a

int ss(short *a, short *b, int n) {
short ta, tb;
int sum = 0;
int i = 0;
ta = a[i]; tb = b[i];
for (i = 1; i < n; i++) {
sum += ta + tb;
ta = a[i]; tb = b[i];
}
sum += ta + tb;
return sum;
}

Listing 9: Non-optimal (rotated by hand)

By rotating the loop, the scalar variables ta and

tb have been added, and they have introduced

loop-carried dependencies which prevent the
compiler from issuing iterations in parallel. The
optimizer is capable of doing this kind of loop
rotation itself.

int ss(short *a, short *b, int n) {
short sum = 0;
int i;
for (i = 0; i < n; i++) {
sum += a[i] + b[i];
}
return sum;
}

Listing 10:Optimal (rotated by the compiler)

Avoid array writes in loops

Other dependencies can be caused by writes to
array elements. In the following loop, the
optimizer cannot determine whether the load
from a reads a value defined on a previous
iteration or one that will be overwritten in a
subsequent iteration.

for (i = 0; i < n; ++i)
a[i] = b[i] * a[c[i]];

Listing 11: Non-optimal (array dependency)

for (i = 0; i < n; ++i)
a[i+4] = b[i] * a[i];

The optimizer can resolve access patterns where
the addresses are expressions that vary by a fixed
amount on each iteration. These are known as
"induction variables".

Avoid aliases

void fn(short a[], short b[], int n)
{
for (i = 0; i < n; ++i)
a[i] = b[i];
}

Listing 13: Non-optimal (potential aliasing)

It may seem that a loop that looks like this does
not contain any loop-carried dependencies, but a
and b are both parameters, and, although they are
declared with [ ], they are in fact pointers, which
may point to the same array. When the same data
may be reachable through two pointers, we say
they may alias each other.

If the -ipa switch is enabled, the compiler will be
able to look at the call sites of fn and possibly
determine whether they ever point to the same
array.

Even with the -ipa switch it is quite easy to create
apparent aliases. The interprocedural analysis
works by associating pointers with sets of
variables that they may refer to at some point in
the program. To simplify the analysis no account
is taken of the control flow, and if the sets for
two pointers are found to intersect then both
pointers are assumed to point to the union of the
two sets.

If the function

fn above were to be called in two

places with global arrays as arguments, then for
the following cases the inter-procedural analysis
will have the results shown:

Case 1:

fn(glob1, glob2, N); Sets do not intersect: a and b
fn(glob1, glob2, N); are not aliases (optimal)

Case 2:

Listing 12: Optimal (induction variables)

Tuning C Source Code for the Blackfin® Processor Compiler (EE-149) Page 5 of 10

fn(glob1, glob2, N); Sets do not intersect: a and b
fn(glob3, glob4, N); are not aliases (optimal)

Case 3:

fn(glob1, glob2, N); Sets intersect: a and b
fn(glob3, glob1, N); may be aliases (non-optimal)

The third case shows that IPA considers the
union of all calls, at once, rather than considering
each call individually, when determining whether
there is a risk of aliasing. If each call were
considered individually, IPA would have to take
flow control into account, and the number of
permutations would make compilation time
impracticably long.

Do as much work as possible in the inner loop

The optimizer focuses on the inner loops because
this is where most programs spend the majority
of their time. It is considered a good trade-off for
an optimization to slow down the code before
and after a loop if it is going to make the loop
body run faster. So, make sure that your
algorithm also spends most of its time in the
inner loop. Otherwise, it may actually be made to
run slower by optimization.

A useful technique is "loop switching". If you
have nested loops where the outer loop runs
many times and the inner loop runs a small
number of times, it may be possible to rewrite
the loops so that the outer loop has fewer
iterations.

Avoid conditional code in loops

a

Additionally, they may require that register data
be stored in memory. This will cause a decrease
in code density and execution performance.

Do not place function calls in loops

The compiler will not generate hardware loops if
the loop contains a function call because of the
expense of saving and restoring the context of a
hardware loop. In addition to obvious function
calls, such as printf(), hardware loop generation
will also be prevented by operations such as:
integer division and modulus, floating-point
arithmetic, and conversion between integer and
floating-point data. These operations may require
implicit calls to support routines.

Use integers for loop control variables and array indices

For loop control variables and array indices, it is
always better to use ints rather than shorts. The C
standard states that shorts should be widened to
integer sizes before carrying out computation,
and then truncated back to short size afterwards.

Frequently, the compiler is able to deal with short
loop counters and still detect zero-overhead
loops and pointer induction variables. However,
it does make the compiler’s life harder and may
occasionally result in less-optimized code.

If a loop contains conditional code, there may be

Loop pragmas and aiding vectorization

a large penalty incurred if the decision often has
to branch against the compiler’s prediction. In
some cases, the compiler will be able to convert

if-else and ?: constructs into conditional

moves. In other cases, it will be able to relocate
expression evaluation outside of the loop

void copy(short *a, short *b)
{
int i;
for (i=0; i<100; i++)
a[i] = b[i];
}

entirely. However, for important loops, linear
code should be written.

Listing 14: Non-optimal (without pragma)

If we call the function copy in Listing 14 twice,

copy(x, y) and, later, copy(y, z), then

Keep loops short

say
interprocedural analysis will not be able to tell

For maximum compiler efficiency, loops should
be as small as possible. Large loop bodies are
usually more complex and difficult to optimize.

Tuning C Source Code for the Blackfin® Processor Compiler (EE-149) Page 6 of 10

that a never aliases b, as described above.
Therefore, the loop contains a loop-carried
dependence and cannot be vectorized. A solution

a

in this case is to use the vector_for pragma.
This tells the compiler that the computation in
one iteration of the loop is not dependent on data
computed in the previous iteration.

The following code uses the vector_for
pragma to allow the loop to perform two
iterations in parallel.

void copy(short *a, short *b)
{
int i;
#pragma vector_for
for (i=0; i<100; i++)
a[i] = b[i];
}

Listing 15: Optimal (with pragma)

Note that this pragma does not force the compiler
to vectorize the loop; the optimizer will check
various properties of the loop and will not
vectorize it if it believes it is unsafe or if it
cannot deduce various properties necessary for
the vectorization transformation. The pragma
assures the compiler that there are no loop­carried dependencies, but there may be other
properties of the loop preventing vectorization.

In cases where vectorization is impossible (for
example, if array a were aligned on a word
boundary, but b were not), then the information
given in the assertion made by vector_for may
still be put to good use in aiding other
optimizations.

Const Data is Constant

By default, the compiler will assume that the data
pointed to by a pointer to const data will not
change. Therefore, another way to tell the
compiler that the two arrays a and b do not
overlap is to use the const keyword.

The example in Listing 16 will have a similar
effect to the

const implementation is better since it also

allows the optimizer, after vectorization, to rotate
the loop, which requires knowing that it is not
just adjacent iterations of the loop which have no

vector_for pragma. In fact, the

dependence on each other, but iterations further
apart, too.

void copy(short *a, const short *b)
{
int i;
for (i=0; i<100; i++)
a[i] = b[i];
}

Listing 16: Usage of const keyword

In C, it is legal, though bad programming
practice, to use casts to allow the data pointed at
by pointers to

const data to change. This should

be avoided since, by default, the compiler will
generate code that assumes

const data does not

change. However, if you have a program that
modifies const data through a pointer, you can
generate correct code by using the compile-time
flag -const-read-write.

Fractional Data

Fractional data, represented as 16-bit and 32-bit
integers, can be manipulated in two ways. The
recommended way, giving you the most control
over your data, is by use of intrinsics. Let us
consider the fractional scalar product. This may
be written as:

int sp(short *a, short *b)
{
int i;
int sum=0;
for (i=0; i<100; i++) {
sum += ((a[i]*b[i]) >> 15);
}
return sum;
}

Listing 17: Non-optimal (uses shifts)

However, this presents some problems to the
optimizer. Normally, the code generated here
would be a multiply, followed by a shift,
followed by an accumulation. However, the
Blackfin processor has a fractional multiply
accumulate instruction that performs all these
tasks in one cycle. Moreover, it can do two of
these instructions in parallel.

Tuning C Source Code for the Blackfin® Processor Compiler (EE-149) Page 7 of 10

a

The compiler recognizes this idiom and
acknowledges that, in the DSP world, the
preference is for saturating arithmetic. The
multiply/shift is replaced by a saturating
fractional multiply. The transformation can be
disabled by using the -no_int_to_fract switch in
case saturation is not required.

However, this was a simple case. In more

blocks; if both access the same block, then a stall
will be incurred. Take as an example the dot
product (as shown in the previous section).

Because data is loaded from arrays

a and b in

every cycle, it may be useful to ensure that these
arrays are located in different memory blocks. As
an example, consider defining two banks in the

MEMORY portion of the .LDF file.

complicated cases, where perhaps the multiply is
further separated from the shift, the compiler
may not detect the possibility of using a
fractional multiply. The recommended coding
style is to use the intrinsics. In the following
example, add_fr1x32() and mult_fr1x32 are
used to add and multiply fractional 32-bit data,
respectively.

#include <fract.h>
fract32 sp(fract16 *a, fract16 *b) {
int i;
fract32 sum=0;
for (i=0; i<100; i++) {
sum = add_fr1x32(sum,
mult_fr1x32(a[i],b[i]));
}
return sum;
}

Listing 18: Optimal (uses intrinsics)

The full list of fractional operations is given in
the Blackfin processor C/C++ Compiler manual,
together with descriptions of other intrinsics
available. The intrinsic functions provide
operations that generally operate on single 16- or
32-bit values; the compiler will recognize when a
loop can be vectorized and will generate 2x16
operations in such circumstances. Just as it is

MEMORY {
BANK_A1 {
TYPE(RAM) WIDTH(8)
START(0xFF900000) END(0xFF900FFF)
}
BANK_A2 {
TYPE(RAM) WIDTH(8)
START(0xFF901000) END(0xFF901FFF)
}
}

Listing 19: LDF Memory Layout

Then, configure the SECTIONS portion to tell the
linker to place data sections in specific memory
banks, as shown below.

SECTIONS {
bank_a1 {
INPUT_SECTION_ALIGN(2)
INPUT_SECTIONS( $OBJECTS(bank_a1))
} >BANK_A1
bank_a2 {
INPUT_SECTION_ALIGN(2)
INPUT_SECTIONS( $OBJECTS(bank_a2))
} >BANK_A2
}

Listing 20: LDF Section Assignment

section("bank_a1") short a[100];
section("bank_a2") short b[100];

better to leave loop rotation to the compiler, the
intrinsics leave operation-pairing to the compiler
as well.

If Possible put Arrays into Different Memory Sections

The Blackfin processor can support two memory
operations on a single instruction line. However,
this will only complete in one cycle if the two

Listing 21: Section Assignment in C sources

In the C source code, sections are defined with
the

section("section_name") construct

preceding a buffer declaration.
Note that explicit placement of data in sections

can only be done for global data. Please see the
"VisualDSP++ 2.0 Linker & Utilities Manual for
Blackfin processors" for further details.

addresses are situated in different memory

Tuning C Source Code for the Blackfin® Processor Compiler (EE-149) Page 8 of 10

a

Appendix A: How the optimizer works

We will use the following fractional scalar
product loop to show how the optimizer works.

#include <fract.h>
fract32 sp(fract16 *a, fract16 *b) {
int i;
fract32 sum=0;
for (i=0; i<100; i++) {
sum = add_fr1x32(sum,
mult_fr1x32(a[i],b[i]));
}
return sum;
}

Listing 22: Fractional Dot Product

After code generation and conventional scalar
optimizations, the compiler will have generated a
loop that looks something like this:

P2 = 100;
LSETUP(.P1L3, .P1L4 – 2) LC0 = P2;
.P1L3:
R0 = W[P0++] (X);
R2 = W[P1++] (X);
A0 += R0.L * R2.L;
.P1L4:
R0 = A0.w;

Listing 23: Compiler output

The loop exit test has been moved to the bottom
and the loop counter rewritten to count down to
zero. The sum is being accumulated in A0. P0
and P1 hold pointers that are initialized with the
parameters A and B, respectively, and are
incremented on each iteration. In order to use 32­bit memory accesses, the optimizer unrolls the
loop to run two iterations in parallel.
being accumulated in A0 and A1, which must be
added together after the loop to produce the final
result. In order to use word loads, the compiler
has to know that P0 and P1 have initial values
that are multiples of four bytes. Note also that,
unless the compiler knows that original loop was
executed an even number of times, a
conditionally-executed odd iteration must be
inserted outside the loop.

Sum is now

P2 = 50;
A1 = A0 = 0;
LSETUP(.P1L3, .P1L4 – 4) LC0 = P2;
.P1L3:
R0 = [P0++];
R2 = [P1++];
A1+=R0.H*R2.H, A0+=R0.L*R2.L;
.P1L4:
R0 = (A0+=A1);

Listing 24: Additional Odd Iteration

Finally the optimizer rotates the loop, unrolling
and overlapping iterations to obtain highest
possible use of functional units. The following
code is finally generated:

A1=A0=0 || R0 = [P0++] || NOP;
R2 = [I1++];
P2 = 49;
LSETUP(.P1L3,.P1L4-8) LC0 = P2;
.P1L3:
A1+=R0.H*R2.H, A0+=R0.L*R2.L || R0 =
[P0++] || R2 = [I1++];
.P1L4:
A1+=R0.H*R2.H, A0+=R0.L*R2.L;
R0 = (A0+=A1);

Listing 25: Optimizer output

Appendix B: Compiler switches

The optimization switches supported by the
compiler are:

-O optimize for speed

-Os optimize for size

-Ox

-Ofp

-ipa do inter-procedural optimization

Table 3: Optimizer-related command-line switches

More details of these switches can be found in
the .VisualDSP++ 2.0 C/C++ Compiler and
Library Manual for Blackfin processors..

assume values in shorts remain in
16-bit range
change frame pointer offsets to
enable use of shorter instructions

Tuning C Source Code for the Blackfin® Processor Compiler (EE-149) Page 9 of 10

a

References

[1] VisualDSP++ 2.0 C/C++ Compiler and Library Manual for Blackfin processors.

First Edition, June 2001. Analog Devices, Inc.

Document History

Version Description
May 26, 2003 Updated according to new Blackfin naming conventions and reformatting
December 2001 Initial Release addressing VisualDSP++ version 2.0

Tuning C Source Code for the Blackfin® Processor Compiler (EE-149) Page 10 of 10