Re: [PATCH bpf-next] bpf: abstract loop unrolling pragmas in BPF selftests

[Eduard Zingerman <eddyz87@gmail.com>]

On Thu, 2024-02-08 at 16:05 +0100, Jose E. Marchesi wrote:
> > On Thu, 2024-02-08 at 13:55 +0100, Jose E. Marchesi wrote:
> > [...]
> > 
> > > However, it would be good if some clang wizard could confirm what
> > > impact, if any, #pragma unroll (aka #pragma clang loop unroll(enabled))
> > > has over -O2, before ditching these pragmas from the selftests.
> > 
> > I compiled sefltests both with and without this patch,
> > there are no differences in disassembly of generated BPF object files.
> > (using current clang main).
> > 
> > [...]
> 
> Hmm, wouldn't that mean that the loops in profiler.inc.h never get
> unrolled regardless of optimization level or pragma? (profiler2.c)
> 

No, the generated code is different between profiler{1,2,3}, e.g.:

$ llvm-objdump -d before/profiler1.bpf.o | wc -l
5356
$ llvm-objdump -d before/profiler2.bpf.o | wc -l
2329
$ llvm-objdump -d before/profiler3.bpf.o | wc -l
1915

What I meant, is that generated code for before/profiler1.bpf.o
and after/profiler1.bpf.o is identical, etc.