[Bug tree-optimization/98775] New: missing optimization opportunity on nbody

[Bug tree-optimization/98775] New: missing optimization opportunity on nbody

[vanyacpp at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98775

            Bug ID: 98775
           Summary: missing optimization opportunity on nbody
           Product: gcc
           Version: 10.2.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: vanyacpp at gmail dot com
  Target Milestone: ---

Created attachment 50015
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=50015&action=edit
nbody.cpp

On the attached sample (208 LOC), clang 11.0 generates the code that is almost
twice as fast as the one generated by GCC 10.2 (-O3 -ffast-math -flto).

$ ./nbody 50000000
4.0s for clang vs 7.5s for GCC.

A quick look at the generated code shows that clang aggressively unrolled all
inner loops. If I unroll all inner loops manually I get:

$ ./nbody-unrolled 50000000
3.7s for clang vs 6.3s for GCC.
17.6B instructions for clang vs 29.6B instructions for GCC.

While the first sample is a subject to unrolling heuristic, the second is about
optimizing the completely linear chunk of code with many floating point
multiplications and additions.

I tried reducing the sample further, but I only came up with PR98774.

[Bug tree-optimization/98775] missing optimization opportunity on nbody

[vanyacpp at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98775

--- Comment #1 from Ivan Sorokin <vanyacpp at gmail dot com> ---
Created attachment 50016
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=50016&action=edit
nbody-unrolled.cpp

[Bug tree-optimization/98775] missing optimization opportunity on nbody

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98775

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
A quick check comparing GCC 10 to GCC 11 shows an improvement from 3.5s to 2.7s
on a Zen2 machine (with -march=native).