IQMath function IQNdiv blows up when den = 0

Eddie3909

IQMath lib 2.1.4.0

C64xplus

IQNdiv

According to the header on this function in IQmath_inline.h

;; On Exit: A4 = Num/Den result in IQ format
;; = 0x7FFFFFFF if +ve overflow
;; = 0x80000000 if -ve overflow
;; = 0x7FFFFFFF if Den == 0

However, if den = 0, the offset in the code below becomes 0xffffff81 and is a really bad offset to use in IQDivTable[] and will give you a memory exception.

t1 = qfmt + sbits;
var2 = (59U - t1);
offset = (U32_IQ)Dnorm >> 23U;
offset = offset - 127U;

Dm=IQdivTable[offset];

Cheers2u

Eddie

over 9 years ago

0 Sivaraj Kuppuraj over 9 years ago

TI__Mastermind 35645 points

Hi Eddie,

Thanks for your post.

In my opinion, the IQNdiv module divide two IQN number and provide the result in 32-bit aligned IQN format as per newton raphson technique and with this context and as per your code snippet "offset = (U32_IQ)Dnorm >> 23U", the left side operand "offset" and the right side operand "Dnorm" variable rightshifted by 23 should both be 32-bit aligned and typecasted with IQN format. Please ensure both the operands are 32-bit aligned and typecasted with IQN format appropriately, so that, you could avoid memory exception issues.

For more details, please refer section IQNdiv module C64x+ IQmath library user guide as below:

http://www.ti.com/lit/ug/sprugg9/sprugg9.pdf (page no 39)

Thanks & regards,

Sivaraj K

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

Please click the Verify Answer button on this post if it answers your question.

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

0 Eddie3909 over 9 years ago in reply to Sivaraj Kuppuraj

Guru 10385 points

Thanks for the reply

I've stepped through the code and this is what I see happening:

If den = 0, then hi1 = 0, and consequently Dnorm also equals zero. Then offset then computes to

offset = (U32_IQ)Dnorm >> 23U
= 0 >> 23
= 0

Then offset is adjusted

offset = offset - 127U;
= 0 - 127
= -127

Using the negative number -127 to index the table results in an invalid memory access.

The IQN format for zero is 0. True?

Cheers

Eddie

    sbits = _norm(den);   
    hi1 = den;     
    Dnorm = (I32_IQ)(hi1 << sbits);

    /*==========================================================================*/
    /* Step 2)  Obtain initial estimate of Dm = 1/Dnorm from                    */
    /*          "_IQdivTable" using the upper 8-bits of the                     */
    /*          normalized value.                                               */
    /* =========================================================================*/

    t1 = qfmt + sbits;
    var2 = (59U - t1);
    offset = (U32_IQ)Dnorm >> 23U; 
    offset = offset - 127U;    

    Dm=IQdivTable[offset];

Processors

Processors forum

IQMath function IQNdiv blows up when den = 0