Loop metadata intrinsics

[p0nce]

I'd like to be able to tell LDC to unroll a loop, to assume that this loop has no loop dependencies, etc... like with the Intel compiler. Does something similar exist?

[dnadlinger]

LLVM knows the concept of loop metadata: http://llvm.org/docs/LangRef.html#llvm-loop

However, I don't think there is currently a way to make LDC emit these from user code. Do you have a design in mind?

[p0nce]

If I had a wish-list :) I'd like :

• From clang loop unrolling http://clang.llvm.org/docs/LanguageExtensions.html#loop-unrolling
  The most useful in my experience would be

#pragma clang loop unroll_count(8)

Of course it would be especially awesome to be like pragma(inline, bool) and allow a statically known integer. eg:

pragma(ldc.attr.loop_unroll, 4) 
for(int i = 0; i < n; ++i)
{
    // body
}

• Often auto-vectorization doesn't happen because of a loop dependency, real or perceived. A way to suppress those worries would be most useful.
  Alternatively, the Intel compiler can also emit information about why loops were not vectorized as diagnostics:

Diagnostic 15543: loop was not vectorized: reason XXX

I tend to prefer this solution which is less blunt, doesn't break code and explain things.

In years of optimization I rarely forced the width of auto-vectorization, however the autovectorization was an essential tool especially to know when it does not happen and why.
Loop unrolling can be useful when the loop inner is memory-bound, as it often hide latency around accesses.

[kinke]

Alternatively, the Intel compiler can also emit information about why loops were not vectorized as diagnostics.

Johan added support for outputting a file with detailed LLVM optimization info in 1.3, see #2089 and #2091.