[9fans] 'fp: stack overflow'

[9fans] 'fp: stack overflow'

[David Tolpin]

Hi,

I am getting 'fp: stack overfow' on a pretty simple code from graphviz
building Bezier curves. The code literally just computes a mean of two
numbers t=(high+low)/2.0 , and when t is about to become 0.9375, it
dies with fp: stack overflow. Of course, it is not reproducable on simpler
tests.

Any thoughts?

David

Re: [9fans] 'fp: stack overflow'

[David Presotto]

Are you using notes that might be interrupting in the middle of a
floating point expression and then doing more floating point in the
note handler?  Other than that, it just sounds like a bug.  Staring
at the assembly language (acid it and asm() the routine) would
probably tell you what's wrong.

Re: [9fans] 'fp: stack overflow'

[David Tolpin]

> Are you using notes that might be interrupting in the middle of a
> floating point expression and then doing more floating point in the
> note handler?  Other than that, it just sounds like a bug.  Staring
> at the assembly language (acid it and asm() the routine) would
> probably tell you what's wrong.

It's a working code from graphviz, used on many platforms; the rest
works (that is, I can just comment this part out and get a graph
with straight lines instead of splines), and the floating point
numbers are only introduced inside the function.

I'll try to debug the assembly code.

Re: [9fans] 'fp: stack overflow'

[Charles Forsyth]

the compiler might be miscalculating the depth of the FP
stack, allowing it to overflow at run-time.
(alternatively, it might just be missing a pop variant in a particular case.)
either way it would be helpful to see if you can isolate the expression(s)
that cause the trouble.

Re: [9fans] 'fp: stack overflow'

[David Tolpin]

>
> the compiler might be miscalculating the depth of the FP
> stack, allowing it to overflow at run-time.
> (alternatively, it might just be missing a pop variant in a particular case.)
> either way it would be helpful to see if you can isolate the expression(s)
> that cause the trouble.
>

The trouble is moving all around the floating point code if I am trying
to isolate it. The expressions are moderately complex, such as

  Vtemp[i][j].x = (1.0 - t) * Vtemp[i-1][j].x + t * Vtemp[i-1][j+1].x

but the problem moves to another expression in another function or
another part of the library (or just to a later invocation of the same
function) if I simpify them.

I must admit that I do not remember the x86 assembler well enough
to debug. But it looks to me like FP stack is unbalanced, and 
gets overflowed due to missing pops (or are they called folds?)

David

Re: [9fans] 'fp: stack overflow'

[David Tolpin]

I am getting stack overflow on eigth call to Bezier on any input file.
I'm recalling that there are eight FP registers, so if it is off by one
in the FP code generator, then it should be filled in eight calls.

However, my knowledge of the assembler is very bad, and to actually find
where one register is lost I will have to re-learn it. 

If anyone is willing to help me, I am ready to provide all I have.

David

Re: [9fans] 'fp: stack overflow'

[David Presotto]

[-- Attachment #1: Type: text/plain, Size: 65 bytes --]

When I said bug, I meant in our compiler not the graphviz code...

[-- Attachment #2: Type: message/rfc822, Size: 2685 bytes --]

From: David Tolpin <dvd@davidashen.net>
To: 9fans@cse.psu.edu
Subject: Re: [9fans] 'fp: stack overflow'
Date: Thu, 4 Mar 2004 22:35:25 +0400 (AMT)
Message-ID: <200403041835.i24IZPTZ058629@adat.davidashen.net>

> Are you using notes that might be interrupting in the middle of a
> floating point expression and then doing more floating point in the
> note handler?  Other than that, it just sounds like a bug.  Staring
> at the assembly language (acid it and asm() the routine) would
> probably tell you what's wrong.

It's a working code from graphviz, used on many platforms; the rest
works (that is, I can just comment this part out and get a graph
with straight lines instead of splines), and the floating point
numbers are only introduced inside the function.

I'll try to debug the assembly code.

Re: [9fans] 'fp: stack overflow'

[boyd, rounin]

> Any thoughts?

diddums