[clang] Document llvm options controlling pragma unroll

[yaxunl]

Summary

Add documentation about LLVM options for controlling pragma unroll.

[tra]

clang/docs/LanguageExtensions.rst

3412

Nit: I think that "loop unrolling increases code size and, therefore, compilation time" is kind of obvious and not particularly helpful. My preference would be to remove it and just leave the details on how to control the unrolling.

3413

Considering that we're talking about clang extensions, we should probably use -mllvm -unroll-count=n.
I'd also rephrase it a bit so that "what" comes before "how". I.e. Loop unroll parameters can be controlled via -mllvm foo and -mllvm bar.

[yaxunl]

revised by Artem's comments.

clang/docs/LanguageExtensions.rst

.. code-block:: c++
 
   #pragma clang loop unroll_count(8)
   for(...) {
     ...
   }
 
 Unrolling of a loop can be prevented by specifying ``unroll(disable)``.
 
+Fully unrolling loops can cause excessive compilation time. LLVM options
+`-unroll-count=n` and `-pragma-unroll-threshold=n` can be used to control loop
+unrolling specified by pragmas, including `#pragma unroll` directives.
+
 Loop Distribution
 -----------------
 
 Loop Distribution allows splitting a loop into multiple loops.  This is
 beneficial for example when the entire loop cannot be vectorized but some of the
 resulting loops can.
 
 If ``distribute(enable))`` is specified and the loop has memory dependencies