This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: Ping shrink wrap patches
On Thu, Oct 13, 2011 at 07:04:59PM +0200, Bernd Schmidt wrote:
> On 10/13/11 18:50, Bernd Schmidt wrote:
> > On 10/13/11 14:27, Alan Modra wrote:
> >> Without the ifcvt
> >> optimization for a function "int foo (int x)" we might have something
> >> like
> >>
> >> r29 = r3; // save r3 in callee saved reg
> >> if (some test) goto exit_label
> >> // main body of foo, calling other functions
> >> r3 = 0;
> >> return;
> >> exit_label:
> >> r3 = 1;
> >> return;
> >>
> >> Bernd's http://gcc.gnu.org/ml/gcc-patches/2011-10/msg00380.html quite
> >> happily rearranges the r29 assignment to be after the "if", and shrink
> >> wrapping occurs. With the ifcvt optimization we get
> >>
> >> r29 = r3; // save r3 in callee saved reg
> >> r3 = 1;
> >> if (some test) goto exit_label
> >> // main body of foo, calling other functions
> >> r3 = 0;
> >> exit_label:
> >> return;
> >
> > I wonder if this can't be described as another case for moving an insn
> > downwards in prepare_shrink_wrap, rather than stopping ifcvt?
Actually, I don't stop ifcvt completely, just on the ifcvt passes
before function prologue/epilogue is emitted and only on a very
restricted set of blocks. Also, x86 and any other target that don't
define both FUNCTION_VALUE_REGNO_P and FUNCTION_ARG_REGNO_P, won't be
affected at all.
> I.e. something like this? Minimally tested by inspecting some generated
> assembly. I haven't found a case where it enables extra shrink-wrapping
> on i686, but maybe it's different on ppc?
It certainly is different on ppc. :) x86 doesn't even run into this
issue, since x86 function args are passed on the stack they won't be
trashed by ifcvt setting up function return values early.
I'll try out your patch, but I think it likely won't work.
prepare_shrink_wrap pushes insns down into successor blocks only when
there's just one live edge for the reg, and that's quite unlikely for
an input arg like r3.
> Index: /local/src/egcs/scratch-trunk/gcc/function.c
> ===================================================================
> --- /local/src/egcs/scratch-trunk/gcc/function.c (revision 179848)
> +++ /local/src/egcs/scratch-trunk/gcc/function.c (working copy)
> @@ -5369,13 +5369,13 @@ static void
> prepare_shrink_wrap (basic_block entry_block)
> {
> rtx insn, curr;
> - FOR_BB_INSNS_SAFE (entry_block, insn, curr)
> + FOR_BB_INSNS_REVERSE_SAFE (entry_block, insn, curr)
> {
> basic_block next_bb;
> edge e, live_edge;
> edge_iterator ei;
> - rtx set, scan;
> - unsigned destreg, srcreg;
> + rtx set, src, dst, scan;
> + unsigned destreg;
>
> if (!NONDEBUG_INSN_P (insn))
> continue;
> @@ -5383,12 +5383,14 @@ prepare_shrink_wrap (basic_block entry_b
> if (!set)
> continue;
>
> - if (!REG_P (SET_SRC (set)) || !REG_P (SET_DEST (set)))
> + src = SET_SRC (set);
> + dst = SET_DEST (set);
> + if (!(REG_P (src) || CONSTANT_P (src)) || !REG_P (dst))
> continue;
> - srcreg = REGNO (SET_SRC (set));
> - destreg = REGNO (SET_DEST (set));
> - if (hard_regno_nregs[srcreg][GET_MODE (SET_SRC (set))] > 1
> - || hard_regno_nregs[destreg][GET_MODE (SET_DEST (set))] > 1)
> + destreg = REGNO (dst);
> + if (hard_regno_nregs[destreg][GET_MODE (dst)] > 1)
> + continue;
> + if (REG_P (src) && hard_regno_nregs[REGNO (src)][GET_MODE (src)] > 1)
> continue;
>
> next_bb = entry_block;
> @@ -5436,7 +5438,8 @@ prepare_shrink_wrap (basic_block entry_b
> if (REG_NOTE_KIND (link) == REG_INC)
> record_hard_reg_sets (XEXP (link, 0), NULL, &set_regs);
>
> - if (TEST_HARD_REG_BIT (set_regs, srcreg)
> + if ((REG_P (src)
> + && TEST_HARD_REG_BIT (set_regs, REGNO (src)))
> || reg_referenced_p (SET_DEST (set),
> PATTERN (scan)))
> {
--
Alan Modra
Australia Development Lab, IBM