If the tripcount spill issue is not handled in the pattern, ICE may happen
then.
Here reload is trying to spill pseudo 173, but a memory operand is not allowed
in zero_cost_loop_end pattern.
And this is what I am trying to solve.
pr44023.c:48:1: error: unable to generate reloads for:
}
^
(jump_insn 136 113 116 4 (parallel [
(set (pc)
(if_then_else (ne (reg:SI 173)
(const_int 1 [0x1]))
(label_ref:SI 114)
(pc)))
(set (reg:SI 173)
(plus:SI (reg:SI 173)
(const_int -1 [0xffffffffffffffff])))
(unspec [
(const_int 0 [0])
] 13)
]) pr44023.c:46 48 {zero_cost_loop_end}
(int_list:REG_BR_PROB 9100 (nil))
-> 114)
pr44023.c:48:1: internal compiler error: in find_reloads, at reload.c:3833
0x989383 _fatal_insn(char const*, rtx_def const*, char const*, int, char const*)
../../trunk/gcc/rtl-error.c:110
0x9777cb find_reloads(rtx_insn*, int, int, int, short*)
../../trunk/gcc/reload.c:3833
0x9875f1 calculate_needs_all_insns
../../trunk/gcc/reload1.c:1515
0x9875f1 reload(rtx_insn*, int)
../../trunk/gcc/reload1.c:1003
0x86a8a7 do_reload
../../trunk/gcc/ira.c:5323
0x86acd8 execute
../../trunk/gcc/ira.c:5470
>
> Hi Sterling,
>
> Attached please find the testcase for the spill issue. Try it out with the
> patch :-)
>
>
> >
> > On Wed, Oct 15, 2014 at 7:10 PM, Yangfei (Felix)
> > <felix.y...@huawei.com>
> > wrote:
> > > Hi Sterling,
> > >
> > > Since the patch is delayed for a long time, I'm kind of pushing
> > > it. Sorry for
> > that.
> > > Yeah, you are right. We have some performance issue here as GCC
> > > may
> > use one more general register in some cases with this patch.
> > > Take the following arraysum testcase for example. In doloop
> > > optimization,
> > GCC figures out that the number of iterations is 1024 and creates a
> > new pseudo
> > 79 as the new trip count register.
> > > The pseudo 79 is live throughout the loop, this makes the
> > > register
> > pressure in the loop higher. And it's possible that this new pseudo is
> > spilled by reload when the register pressure is very high.
> > > I know that the xtensa loop instruction copies the trip count
> > > register into
> > the LCOUNT special register. And we need describe this hardware
> > feature in GCC in order to free the trip count register.
> > > But I find it difficult to do. Do you have any good suggestions on
> > > this?
> >
> > There are two issues related to the trip count, one I would like you
> > to solve now, one later.
> >
> > 1. Later: The trip count doesn't need to be updated at all inside
> > these loops, once the loop instruction executes. The code below relates to
> > this
> case.
> >
> > 2. Now: You should be able to use a loop instruction regardless of
> > whether the trip count is spilled. If you have an example where that
> > wouldn't work, I would love to see it.
> >
