• d-bugmail puremagic.com (34/34) Mar 27 2018 https://issues.dlang.org/show_bug.cgi?id=18671

d-bugmail puremagic.com writes:

https://issues.dlang.org/show_bug.cgi?id=18671

          Issue ID: 18671
           Summary: Implement loop unrolling in dmd's optimizer
           Product: D
           Version: D2
          Hardware: All
                OS: All
            Status: NEW
          Severity: enhancement
          Priority: P1
         Component: dmd
          Assignee: nobody puremagic.com
          Reporter: hsteoh quickfur.ath.cx

It has been known for a while that dmd's optimizer tends to lag behind
gdc/ldc's when it comes to the performance of the generated code.  While this
may give the appearance of general poor quality of the optimizer, closer
inspection reveals that actually dmd's optimizer does perform many standard
optimization operations, and arguably pretty well at it (it even performs code
hoisting from loops, for example). One big reason it fails to generate code
with performance competitive with ldc/gdc is because, as Walter himself said,
it doesn't unroll loops.

When it comes to optimization, there is often a domino effect where one
optimization leads to opportunities for further optimizations, whereas if that
first optimization wasn't performed, it also closes the door to subsequent
optimizations.  Loops being generally the locus of most performance-sensitive
code means that loop optimizations are especially sensitive to this effect, and
loop unrolling is an important step that could often lead to further
optimization opportunities.  Missing this step is a big reason why dmd's
optimizer often doesn't produce competitive code over ldc/gdc.

This has been talked about often, but it's about time it's logged as a real
request so that it won't be forgotten (and, one hopes, the chances of it
actually being implemented might increase just a tad further away from zero).

tl;dr: please implement loop unrolling in dmd. ;-)

--