Update nloptr c

[aadler]

While reading through WRE, I noticed that Rdefines.h is no longer maintained, and that is what @jyypma used in nloptr.c, so I figured we should bring nlopt.c more in line with the API. While going through the file I made some other changes (see the commit log for exact details). They include:

• Import Rinternals.h instead of Rdefines.h and manually define the necessary coercion shortcuts.
• Bugfix: ranseed was processed through the AS_INTEGER macro, which casts to int although nlopt_srand requires an unsigned long. It is now coerced to a double and then cast to the unsigned long retaining the full range of random seeds.
• Convert loop variables to size_t where appropriate and declare them inside the for loops for safety.
• Use more R-centric accessor functions where appropriate (e.g. asReal, asInteger).
• Convert the huge nested if-else ladder in "find algorithm" to a switch statement based on an algorithm lookup table.
  • This should be more efficient as strcmp is now only called once. The table has the algorithms in lexicographical order, so bsearch can be used to find the proper case. If a new algorithm is added, it should be added in alphabetical order and the table updated.
• Some minor efficiency in the gradient nested loops by calculating values which will be constants in the inner loop prior to entering the loop instead of each time through the inner loop.
• Creating pointer variables outside of loops instead of continually calling the REAL or INTEGER macros inside the loops. See the section titled Accessing vector data by Hadley.
• Minor cleanup of DESCRIPTION (Biarch is no longer needed, but NeedsCompilation and UseLTO should be set).
• Declare and initialize variables on the same line where appropriate.
• Use unsigned integers/size_t in structures where numbers < 0 make no sense.
• Copy-edit comments.
• block out some C code in nocov which cannot be tested either because they depend on NLopt functions which are not exposed or because they are completely guarded by is.nloptr.

[aadler]

Correction. Strcmp can be called more than once, but since it’s a bisection search on an ordered list, it should be called fewer times than the if-else ladder.

[aadler]

Team, any thoughts?

[aadler]

@astamm Any objections?

[aadler]

Once the pull requests fixing the packages are completed, this PR may need to be retested.

[astamm]

I merged #176. I will run reverse dependency checks today. Then, I'll merge the fixes in this PR and see if it breaks or not.

[astamm]

Nice changes, thank you @aadler! I merged the master and slightly edited nloptr.c. It seems not to break anything now. Can you test quickly on your side? My plan is to submit to CRAN today or tomorrow at the latest. Thanks again.

[aadler]

Unfortunately I can’t right now. I just built a new PC and am having trouble building R 😞