Two or three initial observations on the fp prining patch

[Paolo Carlini <pcarlini at suse dot de>]

Hi,

finally I'm beginning to play a bit with the actual patch.

At build time I got a warning about the __mod parameter of _M_insert_float
being unused: Jerry, it looks like can be removed.

The first, good indeed!, numbers on my P4-2400 linux are the
following (always -O2)

For this testcase:

 for (double i = 0.0; i < 3000000; i++)
   std::cout << i << '\n';

2.96-99 RH
----------
4.050u 0.000s 0:04.07 99.5%     0+0k 0+0io 143pf+0w

3.4
---
7.760u 0.000s 0:07.80 99.4%     0+0k 0+0io 248pf+0w

3.4 + fp-patch
--------------
3.210u 0.000s 0:03.23 99.3%     0+0k 0+0io 204pf+0w

Overall, more than a 2x speedup, therefore, and /better/ than v2!

For this testcase, as expected, slightly worse numbers:

 for (int i = 0; i < 3000000; i++) {
   std::cout << std::setprecision(30)
        << (double)i + 
9.09923452563465256292345256346525629234525634652562e+18 << '\n';
 }

2.96-99 RH
----------
8.780u 0.030s 0:08.85 99.5%     0+0k 0+0io 143pf+0w

3.4
---
12.320u 0.030s 0:12.46 99.1%    0+0k 0+0io 248pf+0w

3.4 + fp-patch
--------------
10.980u 0.020s 0:11.11 99.0%    0+0k 0+0io 204pf+0w

All in all, the results seem definitely worth the trouble, especially
considering that we can also remove the __convert_from_v mess...

The next step is dealing with the various configury bits:

* uint32_t and int32_t types
This seems easy: probably we can do something very similar to
the the LFS work: an autoconf test that checks for the presence
of the C99 types with a fall back to unsigned long and long.

* IEEE, IBM, or VAX
For this, one of the most difficult, we should probably start from
the macros present in v2 code and generalize it to the architectures
currently supported.

* big or little endian
This seems easy, in principle. Probably, instead of the awful test
present in v2 we can use FLOAT_WORDS_BIG_ENDIAN, as suggested by
Jerry

* Check_FLT_ROUNDS - i'm not sure what this really does
This seems minor, already more or less ok.

More about all those things in the next days...
Paolo.