TMS470 C/C++ CODE GENERATION TOOLS
Release 2.58 August 2005


================================================================================
			TMS470 CGT Release History
================================================================================

  v2.1x thread      v2.20 thread         v2.21 thread       v2.30/v2.40 thread
  ------------      ------------         -------------      ------------------

  v2.10             v2.20                v2.21 (CCS 2.20    v2.30
    |                 |                    |    release)      |
    V                 V                    V                  V
  v2.11             v2.201e              v2.22 (CCS 2.20    v2.31beta
    |                 |                    |    release)      |
    V                 V                    V                  V
  v2.12             v2.202e              v2.23 (CCS 2.20    v2.32beta
    |                 |                    |    release)      |
    +-> v2.121e       +----> v2.202.01e    V                  V
    |     |           |        |         v2.24 (CCS 2.20    v2.33beta
    V     V         v2.203e    V           |    release)      |
  v2.13 v2.122e       |      v2.202.02e    V                  V
    |     |           V        |         v2.25 (CCS 2.20    v2.34
    V     V         v2.204e    V                update)       |
  v2.14 v2.123e       |      v2.202.03e                       V
    |                 V        |                            v2.40beta
    V               v2.205e    V                              |
  v2.15               |      v2.202.04e                       V
    |                 V                                     v2.41beta
    V               v2.206e                                   |
  v2.16               |                                       V
    |                 V                                     v2.42
    +-->  v2.161e   v2.207e                                   |
    |       |         |                                       V
    V       V         V                                     v2.43beta
  v2.17   v2.162e   v2.208e                                   |
    |       |         |                                       V
    V       V         V                                     v2.44beta
  v2.18   v2.163e   v2.209e                                   |
    |       |                                                 V
    V       V                                               v2.45
  v2.19   v2.164e                                             |
    |       |                                                 V
    V       V                                               v2.46
  v2.191e v2.165e                                             |
    |       |                                                 V
    V       V                                               v2.47
  v2.192e v2.166e                                             |
    |       |                                                 V
    V       V                                               v2.48beta
  v2.193e v2.167e                                             |
    |                                                         V
    V                                                       v2.49beta
  v2.194e                                                     |
    |                                                         V
    V                                                       v2.50beta
  v2.195e                                                     |
    |                                                         V
    V                                                       v2.51
  v2.196e                                                     |
    |                                                         V
    V                                                       v2.52
  v2.197e                                                     |
    |                                                         V
    V                                                       v2.53
  v2.198e                                                     |
    |                                                         V
    V                                                       v2.54 
  v2.199e                                                     | 
							      V
                                                            v2.55
                                                              | 
							      V
                                                            v2.56
                                                              | 
							      V
                                                            v2.57
                                                              | 
							      V
                                                            v2.58


================================================================================
			TMS470 CGT Defect Repair History
================================================================================

This history of defect repairs is listed from release v1.20, circa April 1997,
to release v2.58.

NOTE:

In general, all bugs that are fixed in a given version are also fixed
in subsequent releases.  For example, all bugs fixed in v2.17 are also
fixed in every version released since v2.17.

--------------------------------------------------------------------------------
   v2.58
--------------------------------------------------------------------------------

-----------------------------
1.  SDSsq41431
-----------------------------

Headline: Assembler causes a segmentation fault.

-----------------------------
RELEASE NOTE

The assembler can cause a segmentation fault with the following code:

cl470 a.asm

Here is a.asm:

        .title  "whatever"
        .state32
QWERTY  .equ    (1 << 12)
        .sect   ".xyz"
        .global _abc
_abc:
        MRC     P15, #0, R0, C1, C0, #0       
        ORR     R0, R0, #QWERTY               
        MCR     P15, #0, R0, C1, C0, #0       
        B       $ + 4
        MOV     PC, R14                       
        .end

-----------------------------
2.  SDSsq42331
-----------------------------

Headline: The rts routine strcmp() may not return expected negative value 
when (string a < string b).

-----------------------------
RELEASE NOTE

The rts routine strcmp() does not generate an expected negative result when 
the first string is lexicographically less than the second string in the
following example:

    char a[5] = { 0x61, 0x62, 0x0 };
    char b[5] = { 0x61, 0xA9, 0x0 };
    int  x    = strcmp(a, b);

-----------------------------
3.  SDSsq42700
-----------------------------

Headline: The compiler generates incorrect results returned for 
'int b=INT_MIN; return b>=INT_MIN;'.

-----------------------------
RELEASE NOTE

When optimization is not used, the TMS470 compiler can generate incorrect 
code for this sample code:

int f1() { int b=INT_MIN; return b>=INT_MIN; }

-----------------------------
4.  SDSsq43499
-----------------------------

Headline: The optimizer may incorrectly assume that an if condition is always 
true in certain code.

-----------------------------
RELEASE NOTE

In certain code, the optimizer will generate code based on an incorrect
assumption that certain comparison expression are always evaluated to true.
The bug only occurs in expressions such as (1 << i) && 11.  The expression 
requires an integer constant shifted by a integer type variable that is ANDED 
with another integer constant.  This second integer constant (11 in this 
example) can not contain a bit pattern of consecutive 1's.  For example, if the 
expression is (1 << i) && 15 the bug does not occur. 


--------------------------------------------------------------------------------
   v2.57
--------------------------------------------------------------------------------

-----------------------------
1.  SDSsq42459
-----------------------------

Headline: Optimizer generates incorrect code for do..while loop.

-----------------------------
RELEASE NOTE

Loops whose termination depends on unsigned-integer wraparound may have
their trip counts computed incorrectly, which may lead to the loop's
incorrect deletion.

There is code that is trying to calculate a loop's trip count, in
general form as "N - I";  when both N and I are ICONs, which the
optimizer computes using the literal values instead of building a
symbolic expression, aiming to have a more precise value for the trip count.

However, when the loop depends on an unsigned integer wrapping
around, I will be 0 or greater and N can be 0, and the simple "N - I"
gives the wrong answer.

-----------------------------
TEST CASE

typedef  unsigned long  u32;
typedef  unsigned short u16;
typedef  unsigned char  u8;

int bug(void)
{
    int test;
    u32 param;

    test = 1;

    param = 0x10000000;
    do
    {
        test++;
        test++;
        param += 0x10000000;
    } while(param != 0);

    return test;
}


volatile int ret;

void c_main(int p)
{
    ret = bug();
}

-----------------------------
2. SDSsq42460
-----------------------------

Headline: Compiler generates incorrect code for certain switch statements.

-----------------------------
RELEASE NOTE

The compiler may sometimes generate incorrect code for certain switch 
statements.

The bug is that a branch to incorrect label is being generated, instead
of branch to default label.

In the failing test case, the control flow looks like this

L0:
	Bcond   SWITCH_BLOC
L1:	...
	...
	B       SOMEWHERE_ELSE
SWITCH_BLOC:
	switch prolog
SW1: ; switch statement

When SW1 is being processed for switch table compression by the code-gen
(happens with optimization in Thumb mode), the compression logic tries
to convert the switch SW1 into a compare branch sequence. At the end of
the compare-branch sequence, the control needs to be transferred to the
default case label. The original logic in switch table compression
assumes that the switch table predecessor branches to the default case.
In the above sequence, the switch table predecessor block L0 branches to
the switch block (SWITCH_BLOC). This causes the switch table
compare-branch sequence to generate a branch to SWITCH_BLOC at the end
of compare-branch sequence causing an infinite loop.

-----------------------------
3. SDSsq42677
-----------------------------

Headline: Optimizer bug with code involving stack objects and logical 
operations.

-----------------------------
RELEASE NOTE

The optimizer may incorrectly simplify A&A or A|A to just A, even if
A has side effects -- eg, printf("x")&printf("x") would become
printf("x").

-----------------------------
TEST CASE

struct Stack
{
    int m_aValues[32];
    int* m_paValues;
};

void
push(Stack* pStack, int value)
{
    *++pStack->m_paValues = value;
}

int
pop(Stack* pStack)
{
    return *pStack->m_paValues--;
}

void
bitwiseAnd(Stack* pStack)
{
    push(pStack, pop(pStack) & pop(pStack));
}

-----------------------------
4. SDSsq41866
-----------------------------

Headline: cg470 crashes with "CONSTANT TABLE ERROR" message.

-----------------------------
RELEASE NOTE
The cg470 can crash with the following message:

   INTERNAL ERROR: CONSTANT TABLE ERROR - CODEGEN ABORTED

   This may be a serious problem.  Please contact customer support with a
   description of this problem and a sample of the source files that caused
   this INTERNAL ERROR message to appear.

-----------------------------
5. SDSsq42192
-----------------------------

Headline: Optimizer generates code with a missing struct variable assignment.

-----------------------------
RELEASE NOTE

The optimizer may remove an assignment to a nested structure bitfield.  Example 
code:

        void missing_assignment (void)  
        {
           structure1.b2_UN.b_ST.bit0 = global2;

           //This assignment is missing in generated ASM
           structure1.ubyte1 = global3; 

           structure1.b1_UN.b_ST.bit2 = global1;

           if (structure1.b1_UN.b_ST.bit3)  
           { 
              structure1.uword1 = global4;  
           } 
        }

During optimisation, we have code that looks like this internally:

        K$1 = &structure1;
        ((unsigned *)K$1)[6] = *((unsigned *)&structure1+24) & 0xffff7fffu| *
                                &global2*32768&0x8000;
        ((unsigned *)K$1)[6] = *((unsigned *)&structure1+24) & 0xffffffdfu| *
                                &global1*32&0x20;

The optimiser does not recognise that "((unsigned *)K$1)[6]" and
"*((unsigned *)&structure1+24)" refer to the same location -- the
casts confuse things -- and thus erroneously concludes that the first
array assignment can be removed as "dead."

-----------------------------
6. SDSsq42769
-----------------------------

Headline: Compiler is missing --v5e switch and support for ARM v5 instructions.

-----------------------------
RELEASE NOTE

In compiler version 2.25, there is assembler support for the new DSP
instructions found in ARM architecture version V5E.  This support is
through a shell switch -v5e or an assembler switch --v5e.  This support
is now included in the 2.5x compilers.  These instruction are supported:

  CLZ
  QADD
  QSUB
  QDADD
  QDSUB
  SMLABB
  SMLABT
  SMLATB
  SMLATT
  SMLAL
  SMLALBB
  SMLALBT
  SMLALTB
  SMLALTT
  SMLAWB
  SMLAWT
  SMULBB
  SMULBT
  SMULTB
  SMULTT
  SMULL
  SMULWB
  SMULWT
  UMLAL
  UMULL


--------------------------------------------------------------------------------
   v2.56
--------------------------------------------------------------------------------

1.  SDSsq40054 

    Headline: Assembler does not recognize CPSR/SPSR new field mask bits.

    --------------------------
    RELEASE NOTE

    The assembler only accepts these fields in the CPSR/SPSR register as 
    part of the MSR instruction:

    MSR CPSR_ALL, reg
    MSR CPSR_FLG, reg

    The latest ARM syntax was changed to:

    MSR CPSR_c, reg
    MSR CPSR_x, reg
    MSR CPSR_s, reg
    MSR CPSR_f, reg

    where each field mask represents specific mask bits in the PSR registers.
    The 2.55 assembler does not accept these newer fields.

2.  SDSsq40212 

    Headline: Compiler generates incorrect code for certain specific functions.

    --------------------------
    RELEASE NOTE

    When the following test case is compiled with "-mt -o2":

    typedef  unsigned long  u32;
    typedef  unsigned short u16;
    typedef  unsigned char  u8;

    u8  global8;
    u32 global32;

    u8 many_paramaters1( u32 r0, u32 r1, u32 r2, u32 r3, u8 stk0, u8 stk1)
    {
        if (r0)
            stk1 = 1;

        if (stk1 == 5)
            global32 = 2;

        return 0;
    }

    u8 many_paramaters2( u32 r0, u32 r1, u32 r2, u32 r3, u16 stk0, u8 stk1)
    {
        if (r0)
            stk1 = 1;

        if (stk1 == 5)
            global32 = 2;

        return 0;
    }

    main()
    {
        global32 = 3;
        global8  = 3;

        global8 = many_paramaters1(0, 1, 2, 3, 4, 5);
        global8 = many_paramaters2(0, 1, 2, 3, 4, 5);
    }

    these assembler warnings are generated:

    "bug.asm", WARNING! at line 99: [W0002] Offset not word aligned
                STR       A2, [V4, #1]          ; |14|

    The bug involves passing a char or short type as an argument to a routine 
    where the number of arguments requires the char or short type to be passed 
    on the stack.  Our compiler will always create a local copy of the same 
    name of these arguments.  The optimizer and the codegen will attempt 
    different optimizations on these copies.  One optimization is that the 
    optimizer will attempt to widen the local to a word size.  This makes 
    sense, particularly in thumb, since you can generate better code with 
    word-size objects than byte, or half-word objects.  A latter codegen pass 
    attempts to remove the copy instruction by using the same stack location as 
    the argument.  The bug occurs at this point.  The copy instruction, after 
    widening, was a move from the 8-bit char argument to a 32-bit local copy.  
    The codegen makes an address change of the 32-bit copy to the original 
    argument location.  Except this is a 8-bit aligned address, which doesn't 
    work on ARM when loading or storing 32-bits.  

    The bug will always create an assembler warning in thumb-mode, since the 
    offset needs to be scaled by 4.  The bug will exist in arm-mode, but no 
    assembler warning is generated since the offset in a [reg, #imm12] indirect 
    operand is not scaled.

3.  SDSsq31999

    Headline: Trailing radix indicator mishandled on command line options.             
    --------------------------
    RELEASE NOTE

    Compiler/linker does not correctly parse command-line hex numbers of 
    the using the trailing indicators, i.e ABCDh for a hex number or 6543Q 
    for an octal number.  Numbers of the form 0xABCD and 06543 are parsed 
    correctly.

4.  SDSsq37922

    Headline: Linker does not add call to trampoline when needed.

    --------------------------
    RELEASE NOTE

    It is possible that the linker will miss adding a call to a trampoline
    for a call that requires one.  This occurs even though the correct 
    trampoline has been generated, resulting in a fatal error from the 
    linker. 

5.  SDSsq39451

    Headline: Parser crash using pragmas with C++ template functions.

6.  SDSsq40969

    Headline: Linker gives no error when call to trampoline is greater than 
    4MB away.  

    --------------------------
    RELEASE NOTE

    Linker should be able to identify situations where call to trampoline 
    could be greater than than 4MB away.

7.  SDSsq41014

    Headline: Linker -f fill option allows 32-bit signed int instead of 
    unsigned int. 

    --------------------------
    RELEASE NOTE

    The compiler/linker accepted only SIGNED values for the -f (linker fill)
    option.  For example, for only values <= 0x7FFF FFFF are accepted.

8.  SDSsq41215

    Headline: Optimizer crash.

9.  SDSsq41311

    Headline: Codegen doesn't account for constant table in the middle of 
    switch table.

    --------------------------
    RELEASE NOTE

    The code generator may not account for a constant table in the middle 
    of a switch table code, therby generating incorrect code.

10. SDSsq41482

    Headline: Optimiser crash with "x += a && b"                                      
    --------------------------
    RELEASE NOTE

    Statements of the form "x += a && b" may cause the optimiser to crash.

11. SDSsq41736

    Headline: Linker memory bugs.

    --------------------------
    RELEASE NOTE

    There are 2 issues in the linker related to trampolines that could cause 
    the linker to crash.

12. SDSsq41271

    Headline: Optimizer (-o2 and -o3) generates incorrect code for nested loops.

    --------------------------
    RELEASE NOTE

    For ARM and MSP only, the optimiser may generate incorrect code for a
    loop nest with an early exit (eg, a return) from the inner loop with an
    expression that uses the outer loop's control variable, when that
    variable is not of type int.

    --------------------------
    TEST CASE
    
    Example C-Code:
    
        unsigned char func1(unsigned char* arg1)
        {
            unsigned char abc;
            unsigned char def;
            unsigned char ghi;

            for(abc = 0; abc < 8; abc++)
            {
                ghi = *(arg1 + 7 - abc);

                if(ghi != 0)
                {
                    for(def = 0; def < 8; def++)
                        if((ghi & (0x80 >> def)) != 0) 
                            return (((7-abc)*8)+(7-def));
                }
             }
             return 0;
         }

    Assembly-Code:

        ...
        ;** ----------------------- def = 0;
        MOV A1, #0
        ...
        ;** 20 ----------------------- return (unsigned char)(63-def);
        MOV A2, #63
        SUB A1, A2, A1 ; compiler uses wrong loop counter
        LSL A1, A1, #24
        LSR A1, A1, #24
        POP {V1, PC} 
        ...


--------------------------------------------------------------------------------
   v2.55
--------------------------------------------------------------------------------

1.  SDSsq37048

    Headline: While loop with default case in a switch stalls the optimizer 
    indefinite.

    --------------------------
    RELEASE NOTE

    While loop with default case in a switch stalls the optimizer 
    indefinitely.  Can get the following test case compiled after removing 
    the default case in the switch.

    #define STATE	0x00
    #define DELAY	15000

    void main()
    {
        unsigned short int           v_state;
        while(1)
        {
            switch (v_state)
            {
                case STATE:
                {
                    break;
                }
		default:
                {
                    break;
                }
            }
        }
    }

2.  SDSsq37207

    Headline: Internal error compiling (E, call()) if call returns a structure.

    --------------------------
    RELEASE NOTE

    The compiler may abort with an internal error on an expression where a 
    comma-operator has an operand which is a function call, as in: "s1 = 
    (E, call());". For this failure to occur, two additional conditions 
    are required: 1) the call returns a structure type, and 2) the result 
    of the expression is used (for example, in an assignment or other 
    expression). The compiler aborts with a message such as: ">> foo.c, 
    line 7: INTERNAL ERROR: no match for SCOMMA".

    --------------------------
    TEST CASE

    /* compile without optimization */

    struct S { int a; int b; };

    int x,y;

    struct S foo()
    {
       return (x=y),foo();
    }

    --------------------------
    WORKAROUND

    Rearrange the expression to not use a comma operator, as in "E; s1 = 
    call();"

3.  SDSsq37212

    Headline: Narrowing cast of volatile expression incorrectly narrows memory 
    operand in little endian mode.

    --------------------------
    RELEASE NOTE

    When an expression which reads a memory object is cast to a narrower
    type, the compiler optimizes the memory reference by narrowing the type
    to the type it will eventually be cast to.  That is, the compiler
    doesn't bother to load the extra bytes that are going to be ignored,
    anyway.  However, this optimization is not legal for volatile objects. 
    Such objects must always be accessed with exactly the type of the object.
    This bug does not affect big-endian targets.

    --------------------------
    TEST CASE

    #include <limits.h>

    #if INT_MAX > CHAR_MAX
    unsigned int foo0(register volatile unsigned int *addr)
    {
        return (char)addr[6];
    }
    #endif

    #if INT_MAX > SHRT_MAX
    unsigned int foo1(register volatile unsigned int *addr)
    {
        return (short)addr[6];
    }
    #endif

    #if !_TMS320C6X
    unsigned long foo2(register volatile unsigned long *addr)
    {
        return (char)addr[6];
    }
    #endif

    #if defined(LLONG_MAX) && (LLONG_MAX > INT_MAX)
    unsigned long long foo3(register volatile unsigned long long *addr)
    {
        return (int)addr[6];
    }
    #endif

4.  SDSsq37224

    Headline: 2.54 compiler missing dead function list with subsections.

    --------------------------
    RELEASE NOTE

    In certain situations, the linker will remove unreferenced functions.
    For example, veneers that are not called or functions compiled with the
    subsection (-ms) switch that are not called.  To obtain a text file of
    those functions removed by the linker, use the --
    generate_dead_funcs_list.

     For example:
       cl470 aa.c -z --generate_dead_funcs_list

    which will create a text file called dead_funcs.txt.  To write the list
    to a different file name use:

       cl470 aa.c -z --generate_dead_funcs_list=fname

    The switch can also be passed directly to the linker:

       lnk470 --generate_dead_funcs_lis

5.  SDSsq37921

    Headline: Compiler incorrectly places certain code in .const sections.

    --------------------------
    RELEASE NOTE

    If a constant C++ array object is initialized with both a integer value
    and a class member object, it is possible that the array object will be
    placed in the .const section but also dynamically initialized at run
    time.  For example,

    class X
    {
       public:
         static const int a;
    };
    static const int arr[] = { 0, X::a };

    In this example, the array is placed in the .const section and its first
    element is initialized as 0.  The second element is then initialized in
    the __sti function that is called at boot time.  The bug here is that the
    .const section should not be written to at run.

    The bug fix is that the codegen will now recognize this situation and
    place these objects in the .bss section.

    --------------------------
    TEST CASE

    void f_test(void);

    class Test
    {
    public:
        typedef void (*fTest)(void);

        enum eTest
        {
            test_1,
            test_2
        };

        static const fTest tTest[];
    };

    /* This structure will be map into memory */
    static const Test::fTest my_test_ok[] = 
    {
        Test::tTest[Test::test_1]
    };

    /* This structure will be map into flash */
    static const Test::fTest my_test_ko[] = 
    {
        Test::tTest[Test::test_1],
        f_test
    };

6.  SDSsq38036

    Headline: Tail merging could merge into an entry block resulting in an 
    assembler failure.

    --------------------------
    RELEASE NOTE

    It is possible with a test case where a loop is merged with an entry
    block, that the loop branch will be to the entry block label.  Since tail
    merging takes place before the prolog code is generated, this would
    create incorrect code.  Here is a sample test case:

    void func()
    {
        int a = 0, b = 0, c = 0;
        int i;

        foo1(a, b, c);
        foo2();

        for ( ; ; )
        {
           a = 0; b = 1000; c = 1;
           foo1(a, b, c);
           foo2();

           blah();

           a = 0; b = 0; c = 0;
           foo1(a, b, c);
           foo2();
        }
    }

    This will only occur if the entire entry block is merged, as in the test
    case above.  The code generator will now split the entry block in the
    above case, which will ensure the label will be placed after the prolog
    code.

    --------------------------
    TEST CASE

    {
        int a = 0, b = 0, c = 0;
        int i;

        foo1(a, b, c);
        foo2();

        for ( ; ; )
        {
           a = 0; b = 1000; c = 1;
           foo1(a, b, c);
           foo2();

           blah();

           a = 0; b = 0; c = 0;
           foo1(a, b, c);
           foo2();
        }
    }

7.  SDSsq38362

    Headline: Linker conditional linking ignores -u symbol inclusion.

    --------------------------
    RELEASE NOTE

    If an input section containing the definition of a -u symbol is also a
    candidate for conditional linking, the linker may erroneously exclude
    that section from the link if the symbol is not referenced by any other
    part of the application.

8.  SDSsq38656

    Headline: Compiler generates incorrect code for shift and division by 
    constant.

    --------------------------
    RELEASE NOTE

    The compiler may generate incorrect code for a shift and division by a 
    constant with 16-bit values. The 8 MSBs of the 16-bit value are lost. 
    The problem appears in arm-mode with -o3 -mf options.

9.  SDSsq38660

    Headline: Linker generates bad error about multiple mappings of 
    sub-sections.

    --------------------------
    RELEASE NOTE

    The linker may generate a bad error message about trying to map
    an input section to multiple output sections if the input section
    is a subsection that has already been allocated when there is a
    reference to the subsection's base name in a subsequent output
    section specification.

    SECTIONS
    {
       OS1: { file.obj(.scn:_sub) } > MEM1
       OS2: { file.obj(.scn)      } > MEM2
    }

    In the above example, the linker will generate an error about trying
    to add ".scn:_sub" to multiple output sections when the reference to
    "file.obj(.scn)" is being processed.

    --------------------------
    TEST CASE

    main.asm:

       .sect   ".text:_main"
       nop
       nop
       nop

       .sect   ".text:_foo"
       nop
       nop
       nop

       .sect   ".text"
       nop
       nop
       nop

       .sect   ".text:_bar"
       nop
       nop
       nop

    bug.cmd:

    main.obj

    -o bug.out
    -m bug.map

    MEMORY
    {
       MEM1: origin = 0x0100, length = 0x400
       MEM2: origin = 0x1100, length = 0x400
    }

    SECTIONS
    {
       OS1: { main.obj(.text:_main) } > MEM1
       OS2: { main.obj(.text)       } > MEM2
    }


    --------------------------
    WORKAROUND

    Place special code in a uniquely named section using CODE_SECTION or
    DATA_SECTION pragma.  For sections that require special placement,
    avoid using sub-sections whose base name matches any other sections 
    in the application.

10.  SDSsq36847

    Headline: Compiler, under optimization, omits add when mixed with 32-bit 
    divides.

    --------------------------
    RELEASE NOTE


    The optimizer attempts to rewrite the expression:

       (((x * 2048) / y) + 8) / 16

    as basically

       ((x * 2048) / (16 * y)) + 8/16

    Since 8/16 is 0 with an integer type, the final expression ends up as

       ((x * 248) / (16 * y))

    which is wrong.  In other words, the optimizer tries to simplify an integer
    division expression but performed an incorrect transformation.

11.  SDSsq37645

    Headline: Optimizer produces code which generates an infinite loop.

    --------------------------
    RELEASE NOTE

    The use of a shift-and-assign sequence ('var<<=expr' or equivalent)
    assigning to a variable used in an expression controlling the loop can
    result in an infinite loop.  Below is a snap shot of the resultant code:

    unsigned long testfunc(unsigned long d)
    {
        unsigned long dadj = d;
        unsigned long cnt = 0;

        while (dadj < (0x08000000)) {
            dadj <<= 5; //dadj *= 32;
            cnt++;
        }
        return cnt;
    }

    With -o2 and -o3, an infinite loop will be generated (for the while
    loop), with -o1, correct code will be generated.

12.  SDSsq38664

    Headline: Shell does not accept --default_order linker option.

    --------------------------
    RELEASE NOTE

    The --default_order option does not work when passed through the
    shell. It works when passed to the linker directly.

    the following does not work:
    cl470 *.c -z --default_order lnk32.cmd -o test.out

    this works:
    lnk470 *.obj --default_order lnk32.cmd -o test.out

13.  SDSsq38712

    Headline: Linker trampoline branches to wrong function.

    --------------------------
    RELEASE NOTE

    If a call is made from one function to another, where the caller and
    callee are in the same source file, but in separate sections, and if
    the caller and callee sections are placed far enough away from each
    other such that a trampoline is needed to help the caller reach the
    callee at run-time, the linker may generate the trampoline to the
    wrong location.

14.  SDSsq38857

    Headline: Comparison of signed long variable with negative number is 
    incorrect.

    --------------------------
    RELEASE NOTE

    The codegen would transform a signed int compare with a negative constant:

    x < > -1

    to

    x + 1 < > 0

    The codegen would also transform a greater than or equal compare to 0,
    or a less than to 0, as follows:

    BLT ==> BMI (minus)
    BGE ==> BPL (plus)

    The bug occurs if the result of the codegen created add expression
    overflows since BMI or BPL will not check for overflow.

15.  SDSsq39093

    Headline: Constant table entries in the middle of a function can create
    incorrect code.

    --------------------------
    RELEASE NOTE

    If the distance between 2 uses of a global variable causes a constant table
    entry for that variable in the middle of the routine, it is possible the
    code generator may create incorrect code.  For example, in this test case:

    static void test(char idx, char b)
    {
        volatile char *ptr;

        ptr = buffer;

        (*ptr)++;
        (*ptr)++;
        (*ptr)++;

        *ptr = idx;
        *ptr = b;
    }

    if the constant table entry for the variable "buffer" occurs in the middle
    of this routine, before the use of the 2 arguments, the code generator 
    will incorrectly determine that the arguments are not used in the routine 
    and the calling function may not set up the code to pass the arguments to 
    the function "test".

16.  SDSsq39401

    Headline: The thumb-mode longjmp RTS routine will not work correctly in 
    cross-mode code.

    --------------------------
    RELEASE NOTE

    The thumb-mode longjmp routine is defined in the setjmp16.asm file as part 
    of the RTS source code.  The return instruction for this routine is 
    MOV PC, LR.  This instruction will not change modes.  Therefore, if setjmp 
    was called in arm-mode, then longjmp called in thumb-mode to return to the 
    arm-mode location, the mode-change will not take place.  For example:

    jmp_buf buffer;
    #pragma CODE_STATE(fun32, 32)
    #pragma CODE_STATE(fun16, 16)

    void fun32();
    void fun16();

    void main()
    {
       fun32();
    }

    void fun32()
    {
       setjmp(buffer);
       fun16();
    }

    void fun16()
    {
       longjmp(buffer, 0);
    }

    the longjmp() call in func16() will not return to arm-mode.  The return 
    instruction needs to be a BX LR instruction.

17.  SDSsq39239

    Headline: Linker trampoline code does not support ARM9 (v5TE) in Thumb mode.

    --------------------------
    RELEASE NOTE

    In version 2.54, a thumb trampoline will make the call through a POP 
    instruction.  On ARM V4, POP { PC } ignores bit[0] in the callee address, 
    and no mode-change can take place.  Therefore, the 2.54 linker will not 
    make any change to the callee address and bit[0] remains 0.  On ARM V5E and 
    higher, POP { PC } will change mode based on bit[0] of the address and thus 
    2.54 thumb trampolines perform an unintended mode change.  The linker is 
    fixed to set bit[0] in this case, since it is ignored in ARM V4 but will 
    behave as expected on the later architectures.

18.  SDSsq39455

    --------------------------
    RELEASE NOTE

    The compiler may sometimes incorrectly removing a compare instruction prior 
    to a conditional branch.  For example, if the code generator generates this 
    sequence of instructions:

       mov reg2, reg1
       any instruction not using reg1, reg2, or reg3, and does not set status
       compare reg2, reg3 conditional branch

    The codegen optimizes this into:

       compare reg1, reg3
       any instruction not using reg1, reg2, or reg3, and does not set status
       conditional branch

    What could happen is if any of the registers in the instruction between the 
    compare and the branch spills, and the spill instruction is turned into a 
    move, specifically a move from a constant to a register, that move will set 
    the status bits. A later code generator pass then removes the compare since 
    it thinks it is dead.

19.  Tool crash fixes

    SDSsq37757: Abnormal termination of the optimizer.
    SDSsq37781: INTERNAL ERROR termination of the code generator.
    SDSsq38067: INTERNAL ERROR termination of the parser.
    SDSsq38276: Segmentation fault of the parser.
    SDSsq39008: Assertion failure in the code generator.


--------------------------------------------------------------------------------
   v2.54.04
--------------------------------------------------------------------------------

1. DDTS #: SDSsq39455

   Description: See version 2.55.


--------------------------------------------------------------------------------
   v2.54.03
--------------------------------------------------------------------------------

This patch contains these 2 fixes only.  It DOES NOT contain a fix for 
SDSsq38656.

1. DDTS #: SDSsq36847

   Description: See version 2.55.

2. DDTS #: SDSsq38857

   Description: See version 2.55.


--------------------------------------------------------------------------------
   v2.54.02
--------------------------------------------------------------------------------

This patch contains this fix only.

1. DDTS #: SDSsq38656

   Description: See version 2.55.


--------------------------------------------------------------------------------
   v2.54.01
--------------------------------------------------------------------------------

Not released.


--------------------------------------------------------------------------------
   v2.54
--------------------------------------------------------------------------------

1. DDTS #: SDSsq36894

   Description: There was an incorrect location for a call to _unlock()
   in the free() routine in the rts library.  This occurs in file memory.c.
   This is the incorrect code:

   /*-----------------------------------------------------------------------*/
   /* NO COLLESCENCE POSSIBLE, JUST INSERT THIS PACKET INTO LIST            */
   /*-----------------------------------------------------------------------*/
   _unlock()
   minsert(current);

   Fix: The rts source code was updated.

2. DDTS #: SDSsq13595

   Description: The old -mw switch, to align structs to a word boundary, 
   was missing.

   Fix:  The --align_structs switch was added.  See CGTNew.txt for details.

3. DDTS #: SDSsq36935 

   Description: assembler crash

   Fix: Updated the assembler to fix this bug.

4. DDTS #: SDSsq36334

   Description: Epilog inlining bug with infinite loops:

   int q190(register int x)
   {   
       if(x)
         for(;;);

       return 0;
   }

   Fix: Update the code generator to fix this bug.

5. DDTS #: SDSsq37022

   Description: The optimizer was incorrectly rewritting certain unsigned
   integer division expressions.  For example, the following expression:

   ((((x-2)*10)*10u)/10)

   would be rewritten to:

   (((x)*10)/10u)/10 + (((-2)*10)*10u)/10
    
   and then to:

   (((x)*10)/10u)/10 + 429496709

   This is incorrect since unsigned division cannot be distributed.  In other
   words, the optimzer was making an incorrect assumption that an unsigned
   division call:

   unsigned_division(a + b, c) is equivalent to:
   
   unsigned_division(a, c) + unsigned_division(b, c).

   Fix: The optimizer was updated to fix this bug.

6. DDTS #: SDSsq37043

   Description: The optimizer would remove the first 'if' block from the 'for'
   loop in the following test case:

      extern unsigned char v, w;

      void func(signed short *ptr)
      {
        unsigned char i;
        unsigned char t = v;
        signed short temp;

        for (i = 0; i < 3; i += 2)
        {
            temp = ptr[i];

	    /* This condition will get removed */
            if (temp < t && temp > (-t))
	         w = 1;

	    if (ptr[i + 1] < 0)
	       ptr[i + 1] *= (-1);
        }
      }

   The problem was caused by the optimizer incorrectly handling a unary minus
   of an unsigned variable.  

   Fix: The optimizer was updated to fix this bug.

7. DDTS #: SDSsq37051

   Description: Linker veneer/trampoline bug.

   Fix: The linker was updated to fix this bug.


--------------------------------------------------------------------------------
   v2.53
--------------------------------------------------------------------------------

1. DDTS #: SDSsq35225

   Description: The compiler did not print any banner information.  A new
   switch, --verbose has been added.  See CGTNew.txt for details.

2. DDTS #: SDSsq33421

   Description: In version 2.52 the linker's default allocation algorithm
   for output sections not listed in a linker command file changed.  A new
   switch, --default_order has been added to reproduce the behavior of the
   2.51 and earlier linkers.  See CGTNew.txt for details.

3. DDTS #: SDSsq34739, SDSsq35382

   Description: There is an optimizer bug involving references to nested
   struct fields.  In the following pseudo-code:

   A.B.C = ...
   ...
         = A.B.C
   ...
   A.B.C = ...
   ...
         = A.B.C
   ...
         = A.B.C

   Two optimizations may take place.  One, the address of A.B will only be
   calculated once and accessed through a pointer: P-C.  Also, the uses of
   A.B.C will be replaced through temporaries.  The bug is that the
   intervening assignment to A.B.C is missed as an alias of P->C:

   P     = &A.B
   temp  = P->C = ...
   ...
         = temp
   ...
   A.B.C = ...   /* bug occurs here */
   ...
         = temp
   ...
         = temp

   Fix: The optimizer was updated to fix this bug.

4. DDTS #: SDSsq34237 

   Description: Compiling with file names, or file name plus directory path, 
   with a string length of greater than 256 characters can cause the optimizer, 
   code generator, or linker to crash.

   Fix: Updated the compiler to handle long file names.

5. DDTS #: SDSsq34759

   Description: With levels -o2 and above, the optimizer can remove assignments 
   to local structures.  This occurs with the use of the local structure as an 
   argument to a function call.

   Sample code:

   extern int ivalue(int);
   extern int iequals(int,int,int);
   extern summary(const char *);

   typedef struct mystruct
   {
      int i;
   }MYSTRUCT;

   MYSTRUCT f_128Y11(MYSTRUCT c)
   {
      c.i *= 2;
      return c;
   }

   int main()
   {
      MYSTRUCT a = { 0 };
      MYSTRUCT b;

      b.i = ivalue(2);
      a = b; /* bug here, a.i stays 0 instead of 2 */
      iequals(__LINE__,f_128Y11(a).i, 4);
      summary("PTR test #SDSsq34759");
   }

   Fix: Updated the optimizer to fix this bug.

6. DDTS #: SDSsq35014

   Description: The optimizer can crash when processing files containing 
   function names greater than 256 characters.  This only occurs with the 
   optimizer interlisting switch -os.

   Fix: Updated the optimizer to fix this bug.

7. DDTS #: SDSsq35356

   Description: There is the possibility of long compile times with extremely
   large files when no optimization is used.  There is a due to a code 
   generation optimization that tracks variables as they are used in registers.  
   Fix: Updated the code generator so that this optimization will now only take 
   place when an optimization level is set.

8. DDTS #: SDSsq35568

   Description: There is a code generator bug with a post-increment, or 
   post-decreemnt, of a 64-bit structure copy.  In the following example,
   the loop in the function will generate incorrect code if compiled with
   optimization and in arm-mode:

   typedef struct _X
   {
      int x;
      int y;
   } X;

   void func(int count, X pts[])
   {
      X*   endP = pts + count - 1;
      X    tmp;

     // THIS LOOP WILL FAIL!
     while (pts < endP)
     {
        tmp     = *pts;
        *pts++  = *endP;
        *endP-- = tmp;
     }
   }

   The code generator will generate an incorrect use of the STMDA instruction.
   The bug can also occur in the copy of an array of 64-bit elements, either
   double floating point type of long long type.  For example:


   double a[]={1.23,2.34,3.45,4.56,5.67,6.78};

   void func()
   {
        double *p,*t;
        double *b,*s;
        int i,j;
        int err_found = 0;

        b = (double*)calloc(sizeof(a),1);
        s = b+sizeof(a)/sizeof(double)-1;
        t = a + sizeof(a)/sizeof(double)-1;
        p = a;

        //THIS LOOP WILL FAIL!
        while (p <= t)
                *s-- = *p++;
   }

   This bug only occurs in arm-mode.

   Fix: The code generator was updated to fix this bug.

9. DDTS #: SDSsq35735

   Description: The code generator may remove C blocks that contain only
   embedded asm statements.  Consider the following C code:

   volatile int noSWI;

   void main(void)
   {
     noSWI = 1;

     if (noSWI == 0)
     {
       asm("  swi #0x80");
     }
   }

   With optimization the comparison of noSWI with 0 is generated but the branch    around the asm statement is removed, resulting in the interrupt being called:

   LDR       V9, [V9, #0]
   CMP       V9, #0
   swi #0x80

   Note, this is not a safe programming practice since the asm statement has
   no effect on the C environment.  Embedded asm statements cannot rely on
   specific instructions being generated by the compiler.  The preferred
   way to program this would be to add a C function call to a hand-coded
   assembly routine.

   Fix: The code generator was updated to avoid removing C blocks of only
   embedded assembly statements.


10. DDTS #: SDSsq35796

    Description: The compiler could generate incorrect code when accessing a
    bitfield member of a word-sized structure in a quest operation.  This
    structure must also be a member of a word-sized union.  Consider the
    following test case:

    #include <stdio.h>

    typedef struct
    {
        unsigned short w;
        unsigned short x;
    }A;

    typedef union
    {
        A x;
        int i;
    }C;

    main()
    {
        C obj;
        int t;

        obj.x.w = 1;

        t = (obj.x.w == 0) ? 0 : 2;

        printf("t = %d (should be 2)\n", t);
    }

    The code generated is as follows, in thumb-mode and level -o0:

    $main:
        PUSH      {LR}
        LSL       A1, A2, #16
        LSR       A2, A1, #16
        MOV       A1, #1
        LSL       A1, A1, #16
        ORR       A2, A1
        MOV       A1, #0
        BQEQ      L1

    which is missing a compare and the load of A1 with 0 causes the branch to
    always be taken.  The bug is that the code generator creates the 16-bit
    field access as first a 32-bit AND operation to mask off the lower 16-bits 
    and then incorrectly treats that result as a 16-bit value without a right
    shift by 16-bits.  The bug can occur in simple if-else statements since
    the optimizer may convert those to a quest operation.  

    Fix: The code generator was updated to fix this bug.


--------------------------------------------------------------------------------
   v2.52
--------------------------------------------------------------------------------

1. DDTS #: SDSsq33291

   Description: There is an optimizer bug in an attempt to combine two
   variables that increment in step in the following loop:


   #include <stdio.h>

   main()
   {
       int i, j, k, l, m;

       i = 0;
       j = 1;
       k = 0;
       l = 0;
       m = 0;

       while (l < 100)
       {
           k += i > 0;
           m += j;

           l++;
           if (l % 35 == 0) { i++; j++; }
       }

       if (k != 65 || m != 195) printf("FAIL (k = %d, m = %d)\n", k, m);
       else                     printf("PASS\n");

       return 0;
   }

   The optimizer incorrectly combines i and j in the above loop.

   Fix: The optimizer was updated to fix this bug.

2. DDTS #: SDSsq32894

   Description: The optimizer would crash when processing certain C++ enumerated
   types within classes.  The bug was in the parser when creating these C++  
   types.

   Fix: Updated the parser to fix this bug.

3. DDTS #: SDSsq32919

   Description: There is bug in the code generator when processing certain 
   64-bit operations.  The following internal error is given:

   >> bn_asm.c, line 6069: INTERNAL ERROR: no match for PLUS

   This may be a serious problem.  Please contact customer support with a
   description of this problem and a sample of the source files that caused
   this INTERNAL ERROR message to appear.

   Fix: The code generator was updated to fix this bug.

4. DDTS $: SDSsq33686

   Description: The linker could crash while using conditional linking, -ms,
   and using the default debug flag, -g.  The linker was not correctly
   removing all debug information when removing an unused function.  This
   would later cause a crash while merging the debug info.

   Fix: The linker was updated to fix this bug.


--------------------------------------------------------------------------------
   v2.51
--------------------------------------------------------------------------------

1. DDTS #: SDSsq32608

   Description: Linker was not defining default output section symbols for 
   .text, .data, or .bss in cases where there was an unresolved reference
   to one of those symbols but the section is empty or does not have
   any constituents.  For example, if the symbol "end" was used the linker 
   would fail to create an executable and exit with an undefined symbol error.

   Fix: The linker was updated to fix this bug.

2. DDTS #: SDSsq29653

   Description: There was a bug in the optimizer when it attempted to group 
   several pointers in a single pointer plus fixed offsets.  For example, code 
   such as:

   *fftrptr++ = arcr - aici; /* line 179 
   *fftiptr++ = aicr + arci;

   was compiled as:

   ;** 179 -----------------------    *W$1 = arcr-aici; 
        MOV       A2, V9 
        SUB       A2, A2, A4            ; |179| 
        STR       A2, [V1, #0]          ; |179| 
   ;** 180 -----------------------    *((++W$1)+Q$2) = aicr+arci; 
        ADD       A2, A1, A3            ; |180| 
        ADD       V1, #4                ; |180| 
        LDR       A1, [SP, #8] 
        STR       A2, [A1, V1]          ; |180| 

   Clearly, W$1 must be post incremented in line 180, not pre-incremented.  The 
   optimizer was incorrectly grouping the 2 pointers together.

   Fix: The optimizer was updated to fix this bug.

3. DDTS #: SDSsq29299 

   Description: It was possible that the code generator could fail with an 
   "OUT OF MEMORY" error when compiling in thumb mode.  The problem could occur 
   with very large functions.

   Fix: Added limits to the code generation pass that tracks register variables 
   in an optimization to reduce code-size.

4. DDTS #: SDSsq32654

   Description: When a multiple assignment of a nested structure is used, if 
   the fields of the structure are not word aligned the assignment may be done 
   incorrectly.  For example, in the following code:

        typedef struct _exp {
                short i;
                short j;
        } exp;
 
        typedef struct _complex_struct {
                int  i;         
                char c;
                exp  a;
                exp  b;
        } complex_struct;
 
        complex_struct x,y;

        void func1( complex_struct* pComp ) 
        {
           /* a is not correctly set */
           pComp->a = pComp->b =  func1a();
        }

   the code generator mistakenly assumes a is word aligned and generates 
   incorrect code.  
   
   Fix: The code generator was updated to set the correct alignment of nested 
   structures.

5. DDTS #: SDSsq32806

   Description: There is bug in the code generator optimization that attempts 
   to remove BL instructions.  This optimization will take place if there are 
   no function calls in the function and the function only saves and restore 
   the return register.  The bug was that the code generator was not check all 
   register references in load/store multiple instructions.

   Fix: The code generator was update to examine all register references in 
   this optimization.

6. DDTS #: SDSsq29398

   Description: There is a bug in the code generator optimization that will 
   attempt to use load multiple instructions to create the frame rather than 
   extra add instructions.  This optimization did not work correctly with 
   interrupt routines.

   Fix: This optimization will not take place in an interrupt routine.


--------------------------------------------------------------------------------
   v2.50
--------------------------------------------------------------------------------

1. DDTS #: SDSsq32465

   Description: The code generator tail merge optimization can generate an
   incorrect branch in a switch table.

   Fix: Updated the code gen to fix this bug.  Tail merging can be disabled
   with --disable:tail_merge.

2. DDTS #: SDSsq32112

   Description: There were 2 memory bugs in the code generator causing the
   executable to run out of memory.

   Fix: Updated the code gen.

3. DDTS #: SDSsq32344

   Description: There was a bug in the code generator's switch table 
   optimization.  This optimization only occurs with optimization and in thumb
   mode.

   Fix: Updated the code gen.

4. DDTS #: SDSsq30723

   Description: The optimizer was removing the "volatileness" from parameters
   of functions defined as inlined.

   Fix: Updated the optimizer.

5. DDTS #: SDSsq32392

   Description: The optimizer could crash with a string parameter that was
   greater than 256 chars.

   Fix: Updated the optimizer.

6. DDTS #: SDSsq32454

   Description: In extremely large functions, it is possible the code 
   generators register allocation pass could run out of memory.

   Fix: Updated the code gen.

7. DDTS #: SDSsq28938

   Description: The code generator could generate incorrect spill code for 
   64-bit integers.

   Fix: Updated the code gen.

8. DDTS #: SDSsq30610

   Description: There was a bug in the optimizer in the simplification of 
   certain sum-of-product expressions.  The following code provides an
   example:

   /*
   Assignment to size should compute '2*nr + nf'.  Code computes
   '2*nr - nf'.

   Changes that result in correct code generation:
     - Change the unsigned literals to signed (3u ==> 3)
     - Delete the 'if (b < nf>' */

   int a, b, nf, nr;

   void report2(void)
   {
    int  size;

    size = 3 + 3u * (nf - 1) + 2u * (nr - nf);

    if (b < nf)
	a = size;
   }


--------------------------------------------------------------------------------
   v2.49
--------------------------------------------------------------------------------

1. DDTS #: SDSsq31240

   Description: The bug is with the optimizer.  It is limited in scope and may
                occur with optimization level 2 in the following situation:

                1. The use of a local variable as an array index.
                2. The optimizer will attempt to optimize array address
                   calculations using that local variable.  The address is
                   calculated once and a temporary pointer is used from that
                   point forward.
                3. Any expression involving that index, say "I", that is not of
                   the form I = I + K, where K is a constant, will prevent this
                   optimization from taking place.  The bug occurs here.  An
                   expression of the form f(&I) was not preventing the
                   optimization as it should have.

                Here is a sample test case:

                static char arr[] = { 1, 2, 3, 4, 5, 6, 7, 8, 9, 10 };
                void func()
                {
                   unsigned short index;
                   index = 2;

                   if (incr(&index)) return;

                   array[index] += 1;
                   index++;

                   if (incr(&index)) return;

                   array[index] += 3;
                }

   Fix : The optimizer was updated to fix this bug

2. DDTS #: SDSsq31477

   Description: There is a bug in the code-size optimization, branch-chaining
                (see CGTNew.txt for a description).  The assembler was 
                generating an incorrect branch target address.  The shell was
                also ignoring the -ab switch.  The assembler and shell were
                updated to fix this bug.  A switch, --disable_branch_chaining,
                was also added to turn off this optimization.


--------------------------------------------------------------------------------
   v2.48
--------------------------------------------------------------------------------

1. SDSsq30789   Load/store peephole optimizations in the code generator did not
                take into account that an operand could be volatile.  The code
                generator was updated to fix this bug.

2. SDSsq30792   When using -o1 as compilation option the "if/else" statement in 
                the code below is not taken into account.

                unsigned char global_variable; 

                void myproc(unsigned char param)
                {

                      unsigned long mask;
                      mask = (param << 24);
                      if (mask & 0x080000000)
                            global_variable = 3;
                      else
                            global_variable = 4;

                }

                The optimizer was updated to fix this bug.

3. SDSsq30745   The optimizer can hang on certain files with -o3 and -mf 
                (compile for speed).  The optimizer was updated to fix this bug.


--------------------------------------------------------------------------------
   v2.47
--------------------------------------------------------------------------------

SDSsq30354      There is a bug in the optimizer when passing an argument by
		reference through a function pointer.  For example, consider 
                the following code:

                typedef void (*weird_func) (unsigned short*);  

                typedef struct {
                    weird_func func_member;
                } my_struct;

                extern my_func2(unsigned short);

                void my_func(my_struct *elem)
                {
                    unsigned short lg =0;
                    elem->func_member(&lg);
                    my_func2(lg);
                }

               When the code is compiled with -mt -o1, it is transformed by the 
               compiler to:
 
               lg = 0u;
               (*elem->func_member)(&lg);
               my_func2(0u);
 
               The parameter lg is not taken into account when calling 
	       func_member function even though it could've been modified.

	       The optimizer was updated to fix this bug.

SDSSq30304     There was a bug in the thumb mode branch chaining code-size 
	       optimization performed by the code generator.  Branch instruction
	       were not being correctly identified in the algorithm in the
	       presence of comments generated with interlisting.

	       The code generator was updated to fix this bug.

--------------------------------------------------------------------------------
   v2.46
--------------------------------------------------------------------------------

SDSsq29165      There is a bug in the optimizer in certain loops.  For example:

                unsigned char X=0;
                unsigned char B[12];
                unsigned char A[12];

                extern func(unsigned int);

                void test(void)
                {
                  unsigned int index;

                  X = 0;

                  for (index = 0; index < 12; index++)
                  {
                    if (A[index] != B[index])
                      X = 1;
                  }   

                  if (X)
                    func(1234);

                  for(;;) ;
                }

                Instead of doing 

	        LDRB      V9, [A1, #0]          ; |20|
                CMP       V9, #0                ; |20|
                BLNE      _func 

                we do an unconditional branch to func :

 	        STMFD     SP!, {LR}
                MOV       V9, #1                ; |12|
                LDR       A1, CON1              ; |12|
                STRB      V9, [A1, #0]          ; |12|
                BL        _func                 ; |18|

                Bug appears with level -o2 and above.

SDSsq29620      Indirect branches not identified as call points as required
                for the Static Stack Depth Analysis Tool.  The assembler was
                updated to fix this bug.

--------------------------------------------------------------------------------
   v2.45
--------------------------------------------------------------------------------

SDSsq28879      The optimizer can increase its run time exponentially on
                certain complicated AND expressions.  The optimizer was 
                updated to avoid certain expression simplifications in these
                complex expressions.

SDSsq28994      There was a bug in a code generator optimization that attempts
                to remove unnecessary shifts.

SDSsq28971      An incorrect code sequence is generated in the following test
                case:

                typedef  unsigned char UBYTE;
                typedef  long int SLONG;

                volatile SLONG global_sl =-2147483648; /* minimum for SLONG 
                (0x80000000) */
                volatile UBYTE BranchCounter1 = 0;
                volatile UBYTE BranchCounter2 = 0;

                void FUN_SIC_WSS_calcRuck_V () 
                {
                  if (global_sl>1457) 
                    BranchCounter1++; 

                  if (global_sl>1456) 
                    BranchCounter2++; 
                }

                The code fragment :

                if (global_sl>1457) 

                generates the following ASM code: 

                LDR       V9, CON1              ; |23| 
                LDR       A1, [V9, #0]          ; |23| 
                SUBS      A1, A1, #178          ; |23| 
                SUBCSS    A1, A1, #1280         ; |23| 

                The SUBS and SUBCSS arm incorrect for the case where global_sl 
                is set to -2147483648. 

SDSsq27673      The RTS floating point multiplication routines would produce
                results of Infinity or NAN instead of 0 when multipling very
                small floats.


--------------------------------------------------------------------------------
   v2.44
--------------------------------------------------------------------------------

SDSsq28399      There was a bug in the code generator during the merging of 
		branch code in switch statements.  The "tail merging"
		optimization, merging equivalent basic blocks into a single
		basic block was incorrectly comparing instructions.  The
		code generator was update to fix this bug.

SDSsq28606      There was a bug in the thumb mode optimization of switch
		tables that could cause the code generator to crash.  Updated
		the code generator to fix this bug.

SDSsq28337      The optimizer was inefficiently creating stack locations
                for a single variable used in a series of "or" and "shift"
		operations.  For example:

                unsigned long func(unsigned long x)
                {
                  x |= (x >> 1);
                  x |= (x >> 2);
                  x |= (x >> 4);
                  x |= (x >> 8);
                  x |= (x >> 16);
                  return x - (x >> 1);
                }

                In this example, several temporary stack locations are 
                created to hold x rather than registers.


--------------------------------------------------------------------------------
   v2.42
--------------------------------------------------------------------------------

SDSsq28019      There was a bug in a thumb-mode peephole optimization.  The
                optimization would attempt to convert 
     
                ADD r1,r2,#x
                LDR r3,[r1,#0]
 
                to
 
                LDR r3,[r2,#x]
                ADD r1,r2,#x

                without ensuring that r2 and r3 were not the same register.
                The code generator was updated to fix this bug.

SDSsq27408      The optimizer was creating an infinite loop in certain nested
                do while loops.

SDSsq27080      The following code generates matcher error when compiled for 
		thumb mode.

                unsigned long long mask;
                int y;

                mask = ((1ULL << y) - 1ULL);

                Also arm mode generates incorrect code.

SDSsq27022      Incorrect code is generated for signed long long compare 
		operation <, <=, > and >=.  


--------------------------------------------------------------------------------
   v2.30
--------------------------------------------------------------------------------

SDSsq25384      DWARF debug information was reporting the wrong starting address
                for some functions and causing breakpoints to be ignored.

SDSsq22792
SDSsq22795      These were requests to reduce the code-size gap between the TI 
		compiler and the ADS compiler.  TI v2.30 code-size vs ADS1.2 in 
		thumb mode on a key wireless benchmark is smaller by 0.1%.

SDSsq22801      Support for ARM 64-bit integers is added with an updated 
		'long long' data type.  The RTS library is also updated to 
		support 'long long'.

SDSsq15939      The v2.30 of the tools includes a new static stack depth 
		analysis tool, sdp470.exe.

SDSsq25325      In some cases, where a switch-case statement does not have a 
                "break" for the "default" case, the code generated by the 
                compiler may be incorrect.

                For example:

                typedef unsigned int  uint8;
                uint8 var1 = 3;
                uint8 var2;
                uint8 var3;
 
                void main (void )
                {
                   uint8 temp=0;
                   switch (var1){
                     {
                       case 12:
                       temp=16;
                       break;
 
                       default:
                       var2=2;
                     }
                   } /* end switch */
                   var3 = temp;
                }

                In this case, the code for "case 12" is executed.

SDSsq27363      The optimizer would crash at level -o2.

SDSsq26352      The linker was not removing all C++ temporary files when 
                using the PC tools.


--------------------------------------------------------------------------------
   v2.25
--------------------------------------------------------------------------------

SDSsq29963      There is a bug in the way indirect calls are handled in Thumb
                mode when -ml (long calls) option is turned on.

                With -ml, all function calls are considered as "far" calls
                i.e., calls to locations that cannot be reached thro PC
                relative addressing. Hence, the address of each callee is
                loaded from memory into registers and the call is made by
                jumping to the contents of register through a BX instruction.

                For eg., if foo() is called and "-ml" option is used (thumb
                mode), then the sequence looks like this

                LDR r1,CON1
                BX  r1
                ...
                CON1: $foo+1 ; The +1 is used to ensure that control stays in
                thumb mode as BX switches mode depending on LSB of the address

                In thumb mode, indirect calls are made by loading the address
                of the callee in r4 and calling IND$CALL. However, with -ml
                turned on, calling IND$CALL is done as follows

                LDR r1, CON2
                BX  r2
                CON2: IND$CALL ;

                Clearly, this is wrong as IND$CALL is a thumb routine and hence
                needs to have its LSB set as follows

                CON2: IND$CALL + 1;

                The code generator was updated to generate the correct address
                for the thumb mode indirect call


SDSsq28673      Multiplication with very small FLOAT/DOUBLE values
                (ie. 10e-25 * 10e-35) does not work as expected. FS$MUL does
                not seem to do a proper checking for underflow/overflow.
                Sometimes the multiplication of very small floating point
                numbers are not changed to zero, but to infinity or NaN. In the
                attached test case there is no branch taken to "unfl", although
                it should go there.

                FS$MUL:
                -------
                $4:     SUB     e1, #0x7E       ; ADJUST THE BIAS
                        BEQ     unfl            ; AND CHECK FOR UNDERFLOW

                        CMP     e1, #0xFF       ; CHECK FOR EXPONENT OVERFLOW
                        BCS     ovfl            ;


               Stepping through the testcase shows that at FS$MUL, label $4 the
               run does not branch to the underflow routine; but that's what
               they would have expected. Instead it branches to the overflow
               routine.

               The RTS library was updated to fix this bug.


SDSsq29398     The compiler will optimize function prologs by attempting to
               push additional registers on the stack to save code-size.  There
               was a bug in this optimization that occurs in interrupt routines
               only.  In certain cases, not all SOE were being saved.  The
               code generator was updated to fix this bug


--------------------------------------------------------------------------------
   v2.24
--------------------------------------------------------------------------------

SDSsq27857      Nesting a linker command file inside of another linker command
                file may corrupt the stream of input commands from the
                enclosing linker command file.

SDSsq27501      A memory bug caused the code generator to segfault in the C2000
                tools. This bug manifested when the code generator was run in
                an MKS korn shell under certain environment. MKS shell did not
                report the segfault so it appeared as the cl2000 was
                exiting without generating the object file and with no
                diagnostics.  It was possible this bug could appear in the
                TMS470 tools.

SDSsq24430      A C++ structure with a nested structure can be incorrectly
                initialized.  This occurs when the nested structure is declared
                as a static const within a function and initialized within that
                function.  The initialization is incorrectly occurring during
                execution of that function.  This does not occur if extra
                braces, { }, are used with the initialization.  For example:

                extern const char arr[];
                extern void bar(char *);

                class X
                {
                   public:

                   typedef struct Y
                   {
                      public:

                      char *p;
                   } y;
                };

                void foo()
                {
                   /* incorrectly initialized */
                   static const X::y y_arr[] = { (char *)&arr[1],
                                                 (char *)&arr[2],
                                                 (char *)&arr[3] };
                   bar((char *)y_arr);

                   /* correctly initialized */
                   static const X::y z_arr[] = { { (char *)&arr[1] },
                                                 { (char *)&arr[2] },
                                                 { (char *)&arr[3] } };
                   bar((char *)z_arr);
                }

SDSsq25609      A faulty peephole optimization rule that removed loads whose
                destination is a dead register. This was causing the removal of
                volatile loads (that weren't being used). According to C
                standards, accesses to volatile variables cannot be optimized
                away (removed). The peephole rule was incorrectly removing
                loads whose destination was a dead register. Such loads should
                be removed by live variable analysis and dead-code elimination,
                not by target dependent phase. This problem occured only in
                thumb mode.

SDSsq25539      The instruction STMDB is not accepted by the 2.2x ARM assembler.
                The previous versions accepted it and it is a legal instruction.
                With v2.2x , the assembler gives warnings. Note: The same
                warning appears for the LDMDB instruction as well.

SDSsq25535      Different behavior existed when using enumerations with and
                without the small-enum compiler option.  The problem was that
                enums were treated as signed when not using small-num and
                treated as unsigned when using small-enum.

SDSsq24805      In the lowlev.c or rts.src file, the tabinit function that
                implements the driver initialization is written in the
                following manner:

                _CODE_ACCESS void tabinit(void)
                {
                   _DEVICE **st;
                   _DEVICE  *dt;

                   static _DATA_ACCESS int init = 0;

                   _lock();
                   if (init) return;

                   /*---------------------------------------------------------*/
                   /* STEP THROUGH THE TABLES SETTING NAME AND POINTER ENTRIES*/
                   /* TO NULL (SKIP PREDEFINED DEVICE AND STREAMS)            */
                   /*---------------------------------------------------------*/
                   for ( st = &_stream[3]; st != &_stream[_NSTREAM];
                         *st++ = NULL );

                   for ( dt = &_device[1]; dt != &_device[_NDEVICE];
                         *(dt++)->name = '\0');

                   init = 1;
                   _unlock();
                }

                If tabinit has already been initialized previously, the second
                iteration of tabinit will have init=1.  The problem is with the
                statement if(init) return; since _lock() has been called
                already.  If init=1, the tabinit will return without a
                corresponding _unlock() call.  Thus deadlock will occur because
                _lock() and _unlock() must be called as sequentially.

SDSsq25323      An abnormal termination (reported for the TMS320C6x optimizer)
                could potentially occur with the TMS470 optimizer.

SDSsq24700      The request for veneer functionality to support conditional
                linking has been added.

SDSsq22793      This is a request to handle large projects.  The compiler does
                not a limit on application size or number of files.

SDSsq23657      The compiler was producing different code on the PC platform
                versus Unix platforms.

SDSsq12627      The optimizer has been updated to allow variables in disjoint
                scopes to share the same stack space.  For example, the stack
                space for disjoint variables within switch statements will be
                optimized.

SDSsq22805      Fixed the archiver, ar470, to avoid truncating long file names.
                The archiver will no longer truncate names.

SDSsq24502      When allocation fails, the linker will now go and find
                information about all of the memory areas that it is trying to
                allocate to a given output section.  The diagnostic message will
                look like this:

                >>   error: can't allocate named_sect, size 00004054
                     (page 0) in P_MEM (avail: 00000300),
                     MEM2 (avail: 00000bf4), D_MEM (avail: 00000500)

                where "named_sect" is a given output section and P_MEM, MEM2,
                and D_MEM are the memory areas that we could allocate named_sect
                into (if we had the available space).

SDSsq25389      The code generator reports an internal error when trying to
                generate STABS debug information for structures and/or unions
                whose data members have C++ only types.

SDSsq25137      This was a bug in the optimizer in which a "volatile" qualifier
                on a type-conversion operator was percolated into operations on
                non-volatile variables after inlining.


--------------------------------------------------------------------------------
   v2.208e
--------------------------------------------------------------------------------

1. DDTS #: SDSsq30835

   Description: There is an optimizer bug with bitfields and the use of that
   structure as an array of bitfields.  It is possible that a write through 
   the array access can overwrite an incorrect part of the bitfield.

   Fix: Update the optimizer to fix this bug


--------------------------------------------------------------------------------
   v2.207e
--------------------------------------------------------------------------------

1. DDTS #: SDSsq28673

   Description: Multiplication with very small FLOAT/DOUBLE values 
                (ie. 10e-25 * 10e-35) does not work as expected. FS$MUL does 
                not seem to do a proper checking for underflow/overflow. 
                Sometimes the multiplication of very small floating point 
                numbers are not changed to zero, but to infinity or NaN. In 
                the floating multiplication routine (shown below) the check
                is for equal to zero rather than less than or equal to zero.

                FS$MUL:
                -------
                $4:     SUB     e1, #0x7E       ; ADJUST THE BIAS
                        BEQ     unfl            ; AND CHECK FOR UNDERFLOW
                        ~~~

                        CMP     e1, #0xFF       ; CHECK FOR EXPONENT OVERFLOW
                        BCS     ovfl            ;

   Fix: The floating point multiplication routines were updated in the RTS
        libraries.


--------------------------------------------------------------------------------
   v2.206e
--------------------------------------------------------------------------------

SDSsq29109      In some functions the LR register was not being saved in the
                prolog when the LR was being used in the function.

SDSsq29147      The following code will generate an CMN instruction with an 
                incorrect BLE instruction in 32-bit mode:

                void foo()
                {
                   if (((((variable_struct.variable_a_s16 >= 0) &&
                       (variable_struct.variable_b_s16 >= CONTSTANT_A1)) ||
                      ((variable_struct.variable_a_s16 < 0) &&
                       (variable_struct.variable_b_s16 <= -CONTSTANT_A1))) ||
                      ((variable_struct.variable_c_s16>>5) < -CONTSTANT_A2) &&
                      (!variable_struct.variable_d_b1)))
                   {
                     variable_struct.variable_e_u8 = CONTSTANT_A3;
                   }

                   else
                   {
                     if (variable_struct.variable_e_u8 > 0)
                     {
                             variable_struct.variable_e_u8 -= 1;
                     }
                   }
                }

SDSsq28971      With the following code:

                void FUN_SIC_WSS_calcRuck_V () 
                {
                    if (global_sl>1457) 
                  {
                    BranchCounter1++; 
                  }
                  if (global_sl>1456) 
                  {
                    BranchCounter2++; 
                  }
                }

                The code fragment 

                if (global_sl>1457) 

                generates the following ASM code: 

                     LDR       V9, CON1              ; |23| 
                     LDR       A1, [V9, #0]          ; |23| 
                     SUBS      A1, A1, #178          ; |23| 
                     SUBCSS    A1, A1, #1280         ; |23| 

                The SUBS and SUBCSS is incorrect for the case where global_sl 
                is set to -2147483648. 

--------------------------------------------------------------------------------
   v2.204e
--------------------------------------------------------------------------------

SDSsq22630      The compiler generates different code with the PC tools
                vs. the HP/Sun tools.  The code is correct in both cases
                however.  The different code only occurs with switch
                tables.

SDSsq24483      The linker can fail with a library listed on the command
                line and included as an input section in a linker command
                file.  

SDSsq24432      Assembler may sometimes fail to detect when a 16-BIS BL
                instruction that refers to an externally defined symbol is 
                out-of-range from the beginning of the section in which it 
                occurs.  This situation will occur only if the BL instruction 
                occurs at an address beyond 0x3ffffc relative to the beginning 
                of the input section in which it is written. This situation 
                will likely be flagged with a relocation overflow error by the 
                linker. 

SDSsq24768      Conditional linking, controlled with the -ms switch, was not
                working correctly.  If dual-state mode is turned off, the -md
                switch, no .clink directives were generated.  With dual-state
                mode, the .clink directives were missing in the veneers.

SDSsq24698      There is a limit of 200 characters on structure type names
                in the compiler with the debug switch.  For each structure
                type, the compiler will generate a .stag directive with
                the type name in the assembly file.  The assembler will
                crash if the .stag directive name is more than 200 characters.

SDSsq24782      The compiler can generate different code with the HP tools
                vs. the Sun and PC tools.  The code is correct in both
                cases however.  The HP code is less efficient.

--------------------------------------------------------------------------------
   v2.203e
--------------------------------------------------------------------------------

 SDSsq24203     Linker does not account for PC offset in post-relocation
                optimization of trampoline code


--------------------------------------------------------------------------------
   v2.202.04e
--------------------------------------------------------------------------------
     
 SDSsq39455     See release 2.55.


--------------------------------------------------------------------------------
   v2.202.03e
--------------------------------------------------------------------------------
     
 SDSsq38857     See release 2.55.


--------------------------------------------------------------------------------
   v2.202.02e (optimizer only)
--------------------------------------------------------------------------------

 SDSsq33291     See release 2.52.


--------------------------------------------------------------------------------
   v2.202.01e (optimizer only)
--------------------------------------------------------------------------------

 SDSsq30610     See release 2.50.


--------------------------------------------------------------------------------
   v2.202e (optimizer only)
--------------------------------------------------------------------------------

 SDSsq24184     The optimizer creates and uses a temporary variable before it 
		is initialized.

 SDSsq23453     Compiler produces a memory access violation during the
                optimization phase

--------------------------------------------------------------------------------
   v2.201e
--------------------------------------------------------------------------------

 SDSsq23918     Linker corrupts debug information for enum type symbol in
                partial link

 SDSsq23860     Linker corrupts fill section names

 SDSsq23845     Compiler produces a constant table overflow internal error
                during code generation

 SDSsq23808     Linker does not detect an illegal attempt to split a GROUP

 SDSsq23800     Linker crashes while allocating far call trampolines

 SDSsq23521     Linker does not encode little-endian far call trampolines
                properly

 SDSsq23453     Compiler produces a memory access violation during the
                optimization phase

 SDSsq23087     Linker executes an infinite loop while trying to split
                an output section under certain circumstances

--------------------------------------------------------------------------------
   v2.20
--------------------------------------------------------------------------------

 SDSsq23121     Fixed an optimizer bug where an incorrect expression was
		generated at level -o2.

 SDSsq22261     Fixed a possible optimizer abnormal termination. (Reported for 
		the TMS320C6x compiler.)

 SDSsq16836     Fixed library-build utility, mk470, from printing mkp470 in
		the banner.

 SDSsq14859     The -mv shell switch was added which causes the generation of
		full veneers of all functions and to avoid using the BX
		instruction to return from any non-veneer function.

 SDSsq14329     Improved code-size in the generation of certain LSL,LSR,STRH,
		LDRH sequences.

 SDSsq14328     Removed generation of a negatable conditional branch followed
		immediately by an unconditional branch.

 SDSsq18309     Changed the way the optimizer lists 16-BIS functions as they
		are being processed.

 SDSsq06756     Dynamic stack overflow checking is added through the new
		-gs shell switch.

 SDSsq23383     The linker was updated to handle relative path names in linker
		command files.

 SDSsq23068     Compiler generates reference to mangled C++ name for memset

 SDSsq23010     Compiler generates incorrect code for "&=" statement

 SDSsq22613     Compiler generates internal error for certain mod 2
                statements

 SDSsq21915     RTS floating-point subtraction yields wrong result for
                zero inputs

 SDSsq21623     Float/Double to unsigned int conversion should retain sign

 SDSsq21461     Linker may overlap the allocation of sections

 SDSsq21157     Problems with handling of infinity and NaNs in floating-
                point functions in RTS

 SDSsq21096     Overflow detection for floating-point addition/subtraction is
                not correct

 SDSsq20812     Assembler fails to detect certain out-of-range immediate
                expressions

 SDSsq20757     RTS function, fcvt(), has rounding error

 SDSsq20710     Trampolines not generated for far calls within the same
                section

 SDSsq19688     Compiler terminates abnormally on particular conditional
                expressions

 SDSsq19590     -i compiler option behaves differently on Solaris OS and HPUX

 SDSsq19550     Returning element of a union produces incorrect code when
                compiled with optimization

 SDSsq19015     Incorrect trampoline callee address displayed in trampoline
                statistics area of the map file

 SDSsq19014     Unnecessary trampolines generated

 SDSsq18689     Linker generates internal error: "missing tramp callee
                symbol"

 SDSsq18556     Implementation of va_arg() macro in stdarg.h fails

 SDSsq18318     Compiler may generate U_DIV for signed integer divide

 SDSsq18214     Preprocessing directives cannot be used within macro
                invocations (in a C/C++ source file)

 SDSsq17506     Compiler incorrectly removes a local structure copy

 SDSsq17338     Assembler marks code section with incorrect section type
                flag

 SDSsq17090     Compiler generates inefficient code for divide by power-
                of-two (Thumb mode)

 SDSsq17079     Assembler encoding for shift by zero is incorrect

 SDSsq17070     Return from interrupt does not correctly restore CPSR
                register (introduced in v2.12)

 SDSsq16972     Compiler generates invalid MUL instruction

 SDSsq16852     Compiler runs out of memory when compiling large files on
                HP-UX

 SDSsq16527     Example linker command files should contain .pinit output
                section

 SDSsq16477     Compiler produces "Bad kind: TYPE::type_pointed_to"
                internal error during optimization phase

 SDSsq15264     Compiler generates incorrect code for union returned
                from a function when union is representable in a register

 SDSsq14643     Compiler generates internal error during optimization

 SDSsq14331     Generate smaller constant tables by using constants more
                efficiently

 SDSsq14297     Performance of RTS integer divide routine can be improved

 SDSsq14007     Library function, getdevice() may try to write into
                constant memory

--------------------------------------------------------------------------------
   v2.199e
--------------------------------------------------------------------------------

SDSsq28769

   Description: The compiler may sometimes perform incorrect sign extension of
   unicode characters.  For example, the following use of unicode characters
   will not initialize correctly:

   unsigned short text[] = L"\x1080\x1040";

   Fix: Updated the parser to fix this bug.

SDSsq28974

   Description: The heuristics for generating switch jump tables was causing
   a significant increase in code-size.

   Fix: Updated the code generator to fix this bug.

--------------------------------------------------------------------------------
   v2.198e
--------------------------------------------------------------------------------

SDSsq28626

   Description: An optimizer bug where a structure initialization is being
   removed. In the test case below, the initialization of h: "h.t = t;", is
   removed.

   typedef struct {
        int f[5];
        struct {
                char b[8];
                char f[8];
                char n;
        } m;
        char t;
        char p[8];
   } H;

   extern char B(short c, char p);
   extern void C(void);
   extern void D(void);

   void A(short c, char t, char p[])
   {
        int i;
        char n;

        h.t = t;
        for (i = 0; i < h.m.n; i++) {
                n = h.m.b[i];
                h.p[n] = B(c, p[n]);
        }
        C();
        D();
    }

    Fix: Updated the optimizer to fix this bug.

--------------------------------------------------------------------------------
   v2.197e
--------------------------------------------------------------------------------

SDSsq25798      In the example below the while loop counter is not read from
                variable a. Instead V4 is loaded with an initial value #0 and
                used as a loop counter.

                while(a > 0)
                {
                   if(foo(arr1[a- 1], &ptr) == b)
                   {
                      if((ptr + c) > d)
                      {
                         e = b;
                         break;
                      }
                    }

                    a--;
                }

                This bug only occurs at optimization level -o2 or higher.  This
                bug does not occur if the -ma switch is also used.

SDSsq25743      Generated code to access a local pointer is done through an
                unitialized stack location.  This only occurs with optimization.

SDSsq25527      The optimizer is marking and removing asserts it considers
                redundant even though they are not.

--------------------------------------------------------------------------------
   v2.196e
--------------------------------------------------------------------------------

SDSsq25005      Linker may perform an illegal optimization of a far
                call trampoline when the generated trampoline is between
                2048 and 4096 bytes from its target.

--------------------------------------------------------------------------------
   v2.195e
--------------------------------------------------------------------------------

SDSsq14287      A performance improvement in the 16-BIS integer divide routine
		to call the 32-BIS integer divide routine.

SDSsq22630      The compiler generates different code with the PC tools
                vs. the HP/Sun tools.  The code is correct in both cases
                however.  The different code only occurs with switch
                tables.
 
SDSsq24483      The linker can fail with a library listed on the command
                line and included as an input section in a linker command
                file.  

SDSsq24698      There is a limit of 200 characters on structure type names
                in the compiler with the debug switch.  For each structure
                type, the compiler will generate a .stag directive with
                the type name in the assembly file.  The assembler will
                crash if the .stag directive name is more than 200 characters.

SDSsq24782      The compiler can generate different code with the HP tools
                vs. the Sun and PC tools.  The code is correct in both
                cases however.  The HP code is less efficient.

--------------------------------------------------------------------------------
   v2.194e
--------------------------------------------------------------------------------

 SDSsq07162     The compiler includes platform support for HP-UX11.

 SDSsq23918     Linker corrupts debug information for enum type symbol in
                partial link (already fixed in v2.201e)

 SDSsq24203     Linker does not account for PC offset in post-relocation
                optimization of trampoline code

--------------------------------------------------------------------------------
   v2.193e
--------------------------------------------------------------------------------

 SDSsq23860     The linker corrupts fill section names.

 SDSsq23845     The compiler produces a constant table overflow internal
                error during code generation.

 SDSsq23808     The linker does not detect an illegal attempt to split a
                GROUP.

 SDSsq23800     The linker crashes while allocating far call trampolines.

 SDSsq23521     The linker does not encode little-endian far call trampolines
                properly.

 SDSsq23087     The linker executes an infinite loop while trying to split
                an output section under certain circumstances.

--------------------------------------------------------------------------------
   v2.192e
--------------------------------------------------------------------------------

 SDSsq23136     The linker crashes when generating far call trampolines
                with -r and -a enabled.

 SDSsq23068     The compiler generates a reference to the mangled C++ name
                for the memset() RTS function.

 SDSsq23010     The compiler generates incorrect code for masking fields
                of a structure that is represented in a register.

--------------------------------------------------------------------------------
   v2.191e
--------------------------------------------------------------------------------

 SDSsq19783
 SDSsq22676     The linker does not detect an invalid allocation attempt
                near the end of the address range.

 SDSsq22613     The compiler generates an internal error during the
                optimization phase.

--------------------------------------------------------------------------------
   v2.19
--------------------------------------------------------------------------------

 SDSsq21461     Linker may overlap the allocation of sections

 SDSsq21157     Problems with handling of infinity and NaNs in floating-
                point functions in RTS

 SDSsq21096     Overflow detection for floating-point addition/subtraction is
                not correct

--------------------------------------------------------------------------------
   v2.18
--------------------------------------------------------------------------------

 SDSsq20812     Assembler fails to detect certain out-of-range immediate
                expressions

 SDSsq20757     RTS function, fcvt(), has rounding error

 SDSsq20710     Trampolines not generated for far calls within the same
                section

--------------------------------------------------------------------------------
   v2.17
--------------------------------------------------------------------------------

 SDSsq24914     The compiler was adding constants to constant tables that
		were never used.

 SDSsq19688     Compiler terminates abnormally on particular conditional
                expressions

 SDSsq19590     -i compiler option behaves differently on Solaris OS and HPUX

 SDSsq19550     Returning element of a union produces incorrect code when
                compiled with optimization

 SDSsq18214     Preprocessing directives cannot be used within macro
                invocations (in a C/C++ source file)


--------------------------------------------------------------------------------
   v2.167e
--------------------------------------------------------------------------------
     
 SDSsq38857     See release 2.55.


--------------------------------------------------------------------------------
   v2.166e
--------------------------------------------------------------------------------

SDSsq33291      See release v2.52.


--------------------------------------------------------------------------------
   v2.165e
--------------------------------------------------------------------------------

SDSsq30610      See release v2.50.


--------------------------------------------------------------------------------
   v2.164e
--------------------------------------------------------------------------------

SDSsq31240      See release v2.49 beta.

--------------------------------------------------------------------------------
   v2.163e
--------------------------------------------------------------------------------

SDSsq28626      In the following function the register V1 is used before
                it is initialized.

                void A(short c, char t, char p[])
                {
                        int i;
                        char n;

                        h.t = t;
                        for (i = 0; i < h.m.n; i++) {
                                n = h.m.b[i];
                                h.p[n] = B(c, p[n]);
                        }
                        C();
                        D();
                }
 

SDSsq25325      In some cases, where a switch-case statement does not have a
                "break" for the "default" case, the code generated by the
                compiler may be incorrect.

                For example:

                typedef unsigned int  uint8;

                uint8 var1 = 3;
                uint8 var2;
                uint8 var3;

                void main (void )
                {
                   uint8 temp=0;
                    switch (var1){
                      {
                        case 12:
                          temp=16;
                          break;

                        default:
                          var2=2;
                      }
                    } /* end switch */
                    var3 = temp;
                }

                In this case the code for case 12 is executed.

--------------------------------------------------------------------------------
   v2.162e
--------------------------------------------------------------------------------

 SDSsq18318     The compiler incorrectly generates a call to the unsigned
                division routine in the rts library rather than to the
                signed division routine.  This only occurs in certain
                instances of complex source lines containing several
                arithmetic operations and casts between signed and unsigned
                integers.

--------------------------------------------------------------------------------
   v2.161e
--------------------------------------------------------------------------------

 SDSsq23010     The compiler generates incorrect code for masking fields
                of a structure that is represented in a register.

--------------------------------------------------------------------------------
   v2.16
--------------------------------------------------------------------------------

 SDSsq19016     Added run-time addresses of trampoline call sites to the linker
		map file.

 SDSsq19015     Incorrect trampoline callee address displayed in trampoline
                statistics area of the map file

 SDSsq19014     Unnecessary trampolines generated

 SDSsq18689     Linker generates internal error: "missing tramp callee
                symbol"

 SDSsq18556     Implementation of va_arg() macro in stdarg.h fails

--------------------------------------------------------------------------------
   v2.15
--------------------------------------------------------------------------------

 <Support for far call trampolines and automatic code overflow introduced
  in the linker>

--------------------------------------------------------------------------------
   v2.14
--------------------------------------------------------------------------------

 SDSsq17506     Compiler incorrectly removes a local structure copy

 SDSsq17338     Assembler marks code section with incorrect section type
                flag

--------------------------------------------------------------------------------
   v2.13
--------------------------------------------------------------------------------

 SDSsq17090     Compiler generates inefficient code for divide by power-
                of-two (Thumb mode)

 SDSsq17079     Assembler encoding for shift by zero is incorrect

 SDSsq17070     Return from interrupt does not correctly restore CPSR
                register (introduced in v2.12)

 SDSsq16852     Compiler runs out of memory when compiling large files on
                HP-UX

 SDSsq16477     Compiler produces "Bad kind: TYPE::type_pointed_to"
                internal error during optimization phase

--------------------------------------------------------------------------------
   v2.121e
--------------------------------------------------------------------------------

 SDSsq23010     The compiler generates incorrect code for masking fields
                of a structure that is represented in a register.

--------------------------------------------------------------------------------
   v2.12
--------------------------------------------------------------------------------

 SDSsq15264     Compiler generates incorrect code for union returned
                from a function when union is representable in a register

 SDSsq14007     Library function, getdevice() may try to write into
                constant memory

--------------------------------------------------------------------------------
   v2.11
--------------------------------------------------------------------------------

 SDSsq14643     Compiler generates internal error during optimization

--------------------------------------------------------------------------------
   v2.10
--------------------------------------------------------------------------------

 SDSsq12511     Fixed an inefficient reuse of a constant offset used to 
		establish a pointer.

 SDSsq14164     Linker crashes when linking several large objects

 SDSsq13859     Compiler generates LDMIA/STMIA instructions with non-aligned
                data

 SDSsq13757     .template section is included in the .out file as a data
                section

 SDSsq13754     Compiler may sometimes optimize function calls away with -o3

 SDSsq13694     Compiler sometimes generates incorrect veneer function code

 SDSsq13218     Compiler may propagate constant improperly with -on1 and -o3

 SDSsq12958     Compiler generates internal error on type information

 SDSsq12288     Compiler generates incorrect branch code when no optimization
                is used

 SDSsq12188     Compiler gives incorrect code when structure passed as
                parameter to func

 SDSsq12150     assert() macro gives error if argument is other than int

 SDSsq12132     Compiler generates internal error: record_list_read

 SDSsq11937     Compiler produces a stack overflow on PC when compiling
                an extremely long function containing only asm() statements
 SDSsq11507     Request for compiler to give warning if a const is not
                initialized.

 SDSsq11502     Printf's may incorrectly print floating point values.

 SDSsq11501     Printf's may incorrectly print floating point values.

 SDSsq11483     Request for tools to use 1 byte of storage space for 1 byte
                variables.

 SDSsq11460     Request for constants to only use 1 byte of ROM instead of 4
                for chars.

 SDSsq11457     Compiler will not initialize constant using pointer
                arithmetic

 SDSsq11074     Request to redefine a linker warning as error instead

 SDSsq10984     Call stack & StepOut broken with Arm/Thumb program

 SDSsq10778     Linker should be able to handle section names larger than 8
                characters

 SDSsq10775     Linker leaves holes in flash and RAM

 SDSsq10773     const unsigned char consumes 4 bytes of ROM when only 1 byte
                is defined

 SDSsq10161     Compiler incorrectly calculates result of floating-point
                multiplication

 SDSsq09808     Compiler sometimes handles structure assignments incorrectly

 SDSsq09557     Linker crashes when both -r and -s switches are specified

 SDSsq09551     Linker does not link when an invalid switch is encountered

 SDSsq09404     C++ virtual member functions won't link

 SDSsq09349     Hex converter gives error reading COFF file - corrupted
                section header

 SDSsq09203     Compiler performs an illegal operation when compiling code

 SDSsq09170     Compiler incorrectly places strings in connection with
                function calls

 SDSsq08602     Compiler does not generate asm code for certain sections of
                C code

 SDSsq08319     Compiler sometimes fails with internal error

 SDSsq08137     strcmp function in runtimes does not check the zero-byte
                condition

 SDSsq07798     Compiler v2.0 generates internal error message when -o1
                option is used.

 SDSsq07533     Request to support the pragma CODE_SECTION

 SDSsq06473     Request to increase the default memory length assumed by the
                linker

 SDSsq06200     Change tools environment variables to support Code Composer
                Studio

 SDSsq05131     User defined sections for CODE segments as well as for DATA
                segment
 SDSsq05130     16/32 bis compilation within same file

 SDSsq04719     Optimizer abnormally terminates if _assert() not a literal
                string

--------------------------------------------------------------------------------
   v2.04e
--------------------------------------------------------------------------------

 SDSsq09808     Compiler sometimes handles structure assignments incorrectly

 SDSsq12958     Compiler generates internal error on type information

--------------------------------------------------------------------------------
   v2.02e, v2.03e
--------------------------------------------------------------------------------

 <not released>

--------------------------------------------------------------------------------
   v2.01
--------------------------------------------------------------------------------

  <no DDTs entry id # available>

- Compilation host based double precision floating point constant
  expression evaluation produces a different result than runtime
  evaluation.

--------------------------------------------------------------------------------
   v2.00
--------------------------------------------------------------------------------

 SDSsq03453     Optimizer terminated abnormally

 SDSsq03657     There are two different definitions of NULL in
                   standard header files

 SDSsq04015     Optimization level 3  produces incorrect code for
                   register stores

 SDSsq04025     Preprocessor does not flag incorrect macro expansion

 SDSsq04035     -o3 generates incorrect asm code

 SDSsq04044     Compiler gives a warning message with warning level
                   -pw2 of statement

 SDSsq04127     Incorrect code generated by compiler v1.23e.

 SDSsq04364     Preprocessor terminates with an internal compiler
                   error:packet error

 SDSsq04365     Compiler generates illegal structure/union member
                   error with -g option

 SDSsq04535     v1.22e compiler terminates abnormally under Windows NT

 SDSsq04562     Reference to unused section in linker command file causes
                   linker to abnormally terminate.

 SDSsq04634     Compiler incorrectly calculates string size  of typedef
                   char arrays in structures during initialization

 SDSsq04636     Ifile linker for -pm mode omitted

 SDSsq04696     Optimizer terminates abnormally

 SDSsq04708     Hex converter crashes

 SDSsq04775     Compiler incorrectly handles code constants in connection
                   with function calls

 SDSsq04813     Archiver string table limit exceeded.

 SDSsq04817     The C++ namespace was polluted.

 SDSsq03756
 SDSsq04922
 SDSsq05129     Shell doesn't generate unique file names for temporary files.

 SDSsq04923     Pragmas are separated across buffer boundaries.

 SDSsq04992     Pragmas are dropped when using -ppa -ppo.

 SDSsq05023     Optimizer terminates abnormally when compiled with -o3

 SDSsq05139     va_start() macro could be implemented more efficiently

 SDSsq05365     Global Register Feature no longer support

 SDSsq05723     SPNU151 - page 2-15: The -pdf definition is incorrect.

 SDSsq05740
 SDSsq05750     returns from software interrupts do not work correctly

 SDSsd05735     Code generator fails with an internal error when
 SDSsq06282     compiled with -o2

 SDSsq05822     Missing right quote of string constant

 SDSsq05962     The -@ compiler option does not allow you to specify
                multiple files

 SDSsq05985     compiler produces internal error when initializing
                structure members

 SDSsq06079     Performance of the library function memset() is poor

 SDSsq06216     -o3 produces bad code for a loop

 SDSsq06248     Compiler generates internal error:TPRW>> internal
                error: tp_read_mdata

 SDSsq06361     linker gives DOS/4GW error or Dr Watson error with -s

--------------------------------------------------------------------------------
   v1.20 <-> v1.28e
--------------------------------------------------------------------------------

 <v1.28e>

DEFECT CORRECTIONS:
===================

- The compiler may incorrectly place constants between the setup of
  the link register and an indirect function call.  This caused the
  called function to return to the wrong address.

--------------------------------------------------------------------------------

<v1.27e>

DEFECT CORRECTIONS:
===================

- The strcmp runtime function does not check for end of string correctly
  when the string is word aligned.

- Highly complex and large programs may cause the assembler to
  fail with an "Block Symbol Stack Overflow - Aborting" message when
  the -g command line flag is supplied.

- Right shifts may result in the upper bits being set to one rather
  than zero.

- Actual function parameter values of the form &"some string" cause
  the optimizer to fail with an internal error.

KNOWN DEFECTS:
==============

- Assignment of the return value of a function call of type integer to
  a variable of type pointer to a structure receives a type mismatch
  error when the return value is cast to the type of the variable.

--------------------------------------------------------------------------------

<v1.26e>

DEFECT CORRECTIONS:
===================

 SDSsq05023     Optimizer fails with "Abnormal termination
                of opt470" when code is compiled with -o3.

--------------------------------------------------------------------------------

<v1.25e>

DEFECT CORRECTIONS:
===================

 SDSsq04775     Compiler fails in generating call with BX.
                Error relates to handling of code constants
                in connection with function calls, using
                function pointers.

--------------------------------------------------------------------------------

<v1.24e>
<v1.23e>

DEFECT CORRECTIONS:
===================

- The compiler may produce different code when run on
  HPUX 9.x than on HPUX 10.2x.

- For loops whose loop increment is greater than 1, if the optimizer
  attempts to convert the loop into a downcounting loop and the
  loop counter is used after the loop exits, the loop counter may
  contain an incorrect value.

- If preprocessing tokens are used inside of a macro invocation, the
  preprocessor may core dump after emitting the error message.

- When enabling -pw2 level warnings, the compiler may incorrectly
  report alignment violations with void pointers.

- When indexing a group of global variables from a single base
  address, the optimizer may incorrectly compute their offsets.

- The code generator may core dump when performing control flow
  optimizations on a switch statement if all cases end in a return
  statement and the switch statement is the last statement in the
  function.

- The optimizer, at level 3 optimization, incorrectly computes
  variable liveness for const auto variables which are given an
  initialization value inside of a loop.

- The optimizer may enter an infinite loop while performing
  interprocedural constant progagation.

--------------------------------------------------------------------------------

<v1.22e>

DEFECT CORRECTIONS:
===================

- Faulty logic in the cross jumping optimization may generate incorrect
  code in rare circumstances.

- Optimizer may illegally attempt to optimize loads of shorts for
  volatile accesses.

--------------------------------------------------------------------------------

<v1.21>
<v1.21b>

DEFECT CORRECTIONS:
===================

- The optimizer may implement subtraction with 2's complement
  addition (adding 251 rather than subtracting 5), but in some cases,
  the result isn't casted back to a byte to perform the wrap around
  effect.

- With -o3 optimization, the optimizer may incorrectly drop a volatile
  type qualifier from a structure member.

- For very large object archive files, the linker may emit an error
  stating that the archive symbol table is too large.

--------------------------------------------------------------------------------

<v1.201e>

  SDSsq24073    (new DDTs entry was entered since original bug was
                reported and fixed (in v1.24e) before current DDTs
                data-base came into existence)

                When indexing a group of global variables from a
                single base address, the optimizer may incorrectly
                compute their offsets.


