[Bug tree-optimization/80155] [7 regression] Performance regression with code hoisting enabled

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80155

--- Comment #18 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Thomas Preud'homme from comment #17)
> (In reply to rguenther@suse.de from comment #16)
> > On Thu, 23 Mar 2017, thopre01 at gcc dot gnu.org wrote:
> > > 
> > > Funnily this led back to the Cortex-M0+ reduced testcase. With the patch in
> > > comment #13 applied we can still see a difference in the push (one register
> > > pushed Vs 0).
> > 
> > I can't reproduce zero pushes here I get three with/without 
> > -fno-code-hoisting.  code hoisting hoists the two loads inside
> > the switch before it so we have
> 
> Ooops my apologize, it needs more flags indeed. -O2 -funroll-all-loops shows
> 2 registers pushed Vs 1 when -fno-code-hoisting is added.

Still can't reproduce.  I've configured with

/space/rguenther/src/svn/gcc-7-branch/configure --target=arm-suse-linux-gnueabi
--disable-libstdcxx-pch --enable-languages=c,c++

and am using

./cc1 -quiet cortex-m0plus_reproducer.c -O2 -mcpu=cortex-m0plus -mthumb -I
include  -fdump-tree-pre-details -fdump-tree-crited  -fdump-tree-all
-funroll-all-loops -fno-code-hoisting

seeing 3 pushes:

fn1:
        @ args = 0, pretend = 0, frame = 0
        @ frame_needed = 0, uses_anonymous_args = 0
        movs    r3, #0
        push    {r4, r5, lr}

I see softfp being used, not sure how I could change that (if that's what I'm
missing).  Btw, -funroll-all-loops disqualifies this testcase ;)