Possible GMP bug

[Shawn Dyer]

I do not know if I have found a bug or not. It may be some built in 
limitation in gmp that I did not notice the documentation for. I was 
attempting to do a computation of pi using a quartically convergent 
algorithm using GMP floating point numbers. Here is the information you 
requested in the bug report:

* GMP version 4.1 compiled locally
* A small test program is attached. (I linked it statically) It works fine 
for 9 iterations which returns a value of pi accurate to 715,319 digits, 
requiring a precision of about 2,376,240 bits. It fails on 10 iterations, 
which should return a value accurate to over 2,800,000 digits, and requires 
about 9,505,060 bits of precision. The error begins to appear between 8 and 
8.5 million bits of precision.
* Also attached is the output from running the program under gdb. I am not 
sure whether it will be useful or not. If you can reproduce the problem, I 
am sure you will have more informative output, but I can attempt to provide 
you with anything else you need.
* I used no configuration options when building gmp.
* I built it using gcc:
gcc -v
Reading specs from /usr/lib/gcc-lib/i386-slackware-linux/2.95.3/specs
gcc version 2.95.3 20010315 (release)
(I had originally observed the problem using Intel's 32 bit compiler:
icc -V
Intel(R) C++ Compiler for 32-bit applications, Version 6.0   Build 020312Z
The gdb output was very similar as for compiled under gcc)
* Linux nonus 2.4.5 #10 Fri Jun 22 02:20:21 PDT 2001 i686 unknown
(I also observed the same problem on a Linux 2.4.18 kernel as well)
* The output of config.guess and configfsf.guess is:
pentium3-pc-linux-gnu    and     i686-pc-linux-gnu

If there is anything other information you need, Please let me know.

gdb_output
Description: Text document

pi-icc-quartic.c
Description: Text document