Re: [PATCH 09/11] test/ring: disable problematic tests for RISC-V

[Stephen Hemminger <stephen@networkplumber.org>]

On Thu, 5 May 2022 19:43:43 +0200
Stanisław Kardach <kda@semihalf.com> wrote:

> On Thu, May 5, 2022 at 7:35 PM Stephen Hemminger <stephen@networkplumber.org>
> wrote:
> 
> > On Thu,  5 May 2022 19:30:01 +0200
> > Stanislaw Kardach <kda@semihalf.com> wrote:
> >  
> > > When compiling for RISC-V in debug mode the large amount of inlining in
> > > test_ring_basic_ex() and test_ring_with_exact_size() (in test_ring.c)
> > > leads to large loop bodies. This causes 'goto' and 'for' loop
> > > PC-relative jumps generated by the compiler to go beyond the architecture
> > > limitation of +/-1MB offset (the 'j <offset>' instruction). This
> > > instruction should not be generated by the compiler since C language does
> > > not limit the maximum distance for 'goto' or 'for' loop jumps.
> > >
> > > This only happens in the unit test for ring which tries to perform long
> > > loops with ring enqueue/dequeue and it seems to be caused by excessive
> > > __rte_always_inline usage. ring perf test compiles just fine under
> > > debug.
> > >
> > > To work around this, disable the offending tests in debug mode.
> > >
> > > Signed-off-by: Stanislaw Kardach <kda@semihalf.com>
> > > Sponsored-by: Frank Zhao <Frank.Zhao@starfivetech.com>
> > > Sponsored-by: Sam Grove <sam.grove@sifive.com>
> > > ---  
> >
> > It seems to me that fixing the excessive inlining in the ring code
> > could benefit all architectures, rather than neutering the tests
> > on RISCV.
> >  
>  True. Since this only happened in the tests that I've mentioned, my other
> approach was to introduce a "slow" wrapper over test_ring_dequeue|enqueue()
> which did not force inlining via __rte_always_inline and use it in
> functional test functions. However after talking with Thomas Monjalon we've
> decided to guard the debug build of RISC-V only.
> Another thing is that this is a clear bug in the compiler, the relaxation
> of the jump should not be done since RISC-V has long jump construct for
> arbitrary jumps (auipc+jalr).

There is no good reason the __rte_always_inline should be in the ring code.
The purpose of always inline should be only for code that would break if
not inlined.