[Bug tree-optimization/63986] [5 Regression][SH] gcc.target/sh/pr51244-15.c failures

["rguenth at gcc dot gnu.org" <gcc-bugzilla at gcc dot gnu dot org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63986

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |NEW
                 CC|                            |ppalka at gcc dot gnu.org,
                   |                            |rguenth at gcc dot gnu.org
           Assignee|rguenth at gcc dot gnu.org         |unassigned at gcc dot gnu.org

--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
Ok, now already existing forwprop code gets fed with

  <bb 2>:
  _3 = a_2(D) == 0;
  x_4 = (char) _3;
  _7 = ~_3;
  _8 = (int) _7;
  MEM[(int *)d_5(D) + 8B] = _8;
  if (x_4 != 0)

where we now in the first forwprop pass identified the opportunity
to use ~_3 instead of x_4 == 0 thus x_4 is now no longer multi-use.
This makes us optimize if (x_4 != 0) to if (_3 != 0) which we
re-optimize in fold_gimple_cond now to '_3' and then of course
if (_3 != 0) (err, and we return "changed"....) which means we
now propagate _again_ via forward_propagate_into_gimple_cond
which now specifically allows aggressive forwarding of compares,
bypassing single-use restrictions.  See

2014-11-16  Patrick Palka  <ppalka@gcc.gnu.org>

        PR middle-end/63790
        * tree-ssa-forwprop.c (forward_propagate_into_comparison_1):
        Always combine comparisons or conversions from booleans.

thus me fixing my "mistake" does not help anymore.

I suppose RTL CSE cannot CSE flag register sets...?

Btw, my previous comment was incorrect - the code is what is now produced
on trunk while on the 4.9 branch we create

test_0:
.LFB0:
        .cfi_startproc
        testl   %edi, %edi
        movl    %edx, %eax
        sete    %r8b
        movl    %r8d, %edi
        xorl    $1, %edi
        testb   %r8b, %r8b
        movzbl  %dil, %edi
        cmovne  %esi, %eax
        movl    %edi, 8(%rcx)
        ret

which means code generation improved for x86...