On Friday, 17 October 2014 at 10:30:14 UTC, eles wrote:
On Friday, 17 October 2014 at 09:46:49 UTC, Ola Fosheim Grøstad
wrote:
The second thing I would change is to make whole program
analysis mandatory so that you can deduce and constrain value
ranges.
Nice idea, but how to persuade libraries to play that game?
1. Provide a meta-language for writing propositions that
describes what libraries do if they are foreign (pre/post
conditions). Could be used for "asserts" too.
2. Provide a C compiler that compiles to the same internal
representation as the new language, so you can run the same
analysis on C code.
3. Remove int so that you have to specify the range and make
typedefs local to the library
4. Provide the ability to specify additional constraints on
library functions you use in your project or even probabilistic
information.
Essentially it is a cultural thing, so the standard library has
to be very well written.
Point 4 above could let you specify properties on the input to a
sort function on the call site and let the compiler use that
information for optimization. E.g. if one million values are
evenly distributed over a range of 0..100000 then a quick sort
could break it down without using pivots. If the range is 0..1000
then it could switch to an array of counters. If the input is 99%
sorted then it could switch to some insertion-sort based scheme.
If you allow both absolute and probabilistic meta-information
then the probabilistic information can be captured on a corpus of
representative test-data. You could run the algorithm within the
"measured probable range" and switch to a slower algorithm when
you detect values outside it.
Lots of opportunities for improving "state-of-the-art".
