[PATCH] Update documentation for unroll pragmas on loops with runtime trip counts

[Mark Heffernan]

REPOSITORY
  rL LLVM

================
Comment at: include/clang/Basic/AttrDocs.td:1330
@@ -1328,1 +1329,3 @@
+the loop unroller tries to unroll the loop by a power-of-two factor
+subject to a limit on unrolled code size.
 
----------------
broune wrote:
> Is #pragma unroll different from #pragma unroll(full) ? If not, why not make the two doc strings the same?
In AttrDocs.td there is not documentation for "#pragma clang loop unroll".  Instead it just mentions "#pragma clang loop" and points to LanguageExtensions for details.  Regarding LanguageExtensions documentation of "#pragma clang loop unroll" and AttrDocs documentation for "#pragma unroll", I believe they do both say the same thing (or at least aren't conflicting with each other) with the LanguageExtensions documentation providing more context and describing the limits on unrolling.

Little bit of background here.  There are two forms of the loop unroll pragmas:

#pragma unroll
#pragma unroll N

and:

#pragma clang loop unroll(full)
#pragma clang loop unroll_count(N)

The "#pragma unroll" forms are supported for compatibilty with other compilers (such as nvcc CUDA compiler) where this form is common while "#pragma clang .." is a more cannonical form specific to clang.  "#pragma unroll" and "#pragma clang loop unroll(full)" are equivalent.  What I propose in the related optimization patch (http://reviews.llvm.org/D10854) is that "#pragma unroll" should enable loop unrolling with a runtime trip count of the loop.  So if you have:

void foo(int n) {
#pragma unroll
  for (int i=0; i<n; i++) { ... }
}

Then the optimizer will unroll the loop with some reasonable power-of-two factor.  One potentially funny bit about this approach is that "#pragma clang loop unroll(full)" will also enable unrolling of a loop with a runtime trip count.  This may be surprising... or maybe not. Full unrolling is clearly not possible in this case so just being more aggressive with unrolling may be the sensible thing to do.

http://reviews.llvm.org/D10857

EMAIL PREFERENCES
  http://reviews.llvm.org/settings/panel/emailpreferences/