On Tue, Nov 29, 2016 at 5:07 PM, Jeff Law <l...@redhat.com> wrote:
> On 11/29/2016 03:23 AM, Richard Biener wrote:
>> On Mon, Nov 28, 2016 at 10:23 PM, Jeff Law <l...@redhat.com> wrote:
>>> I was digging into issues around the patches for 78120 when I stumbled
>>> undesirable bb copying in bb-reorder.c on the m68k.
>>> The core issue is that the m68k does not define a length attribute and
>>> therefore generic code assumes that the length of all insns is 0 bytes.
>> What other targets behave like this?
> ft32, nvptx, mmix, mn10300, m68k, c6x, rl78, vax, ia64, m32c
> cris has a hack to define a length, even though no attempt is made to make
> it accurate. The hack specifically calls out that it's to make bb-reorder
>>> That in turn makes bb-reorder think it is infinitely cheap to copy basic
>>> blocks. In the two codebases I looked at (GCC's runtime libraries and
>>> newlib) this leads to a 10% and 15% undesirable increase in code size.
>>> I've taken a slight variant of this patch and bootstrapped/regression
>>> it on x86_64-linux-gnu to verify sanity as well as built the m68k target
>>> libraries noted above.
>>> OK for the trunk?
>> I wonder if it isn't better to default to a length of 1 instead of zero
>> there is no length attribute. There are more users of the length
>> in bb-reorder.c (and elsewhere as well I suppose).
> I pondered that as well, but felt it was riskier given we've had a default
> length of 0 for ports that don't define lengths since the early 90s. It's
> certainly easy enough to change that default if you'd prefer. I don't have
> a strong preference either way.
Thinking about this again maybe targets w/o insn-length should simply
always use the 'simple' algorithm instead of the STV one? At least that
might be what your change effectively does in some way?