Alex Coplan <alex.cop...@arm.com> writes:
>> > + writeback_pats[i] = orig_rtl[i];
>> > +
>> > + // Now that we've characterized the defs involved, go through the
>> > + // debug uses and determine how to update them (if needed).
>> > + for (auto use : set->debug_insn_uses ())
>> > + {
>> > + if (*pair_dst < *use->insn () && defs[1])
>> > + // We're re-ordering defs[1] above a previous use of the
>> > + // same resource.
>> > + update_debug_use (use, defs[1], writeback_pats[1]);
>> > + else if (*pair_dst >= *use->insn ())
>> > + // We're re-ordering defs[0] below its use.
>> > + update_debug_use (use, defs[0], writeback_pats[0]);
>> > + }
>> > + }
>> > +
>> > + // Now let's look at registers which are def'd by the second insn
>> > + // but not by the first insn, there may still be debug uses of a
>> > + // previous def which can be affected by moving the second insn up.
>> > + for (auto def : insns[1]->defs ())
>> > + {
>> > + // This should be M log N where N is the number of defs in
>> > + // insns[0] and M is the number of defs in insns[1].
>> > + if (def->is_mem () || find_access (insns[0]->defs (), def->regno
>> > ()))
>> > + continue;
>> > +
>> > + auto prev_set = safe_dyn_cast<set_info *> (def->prev_def ());
>> > + if (!prev_set)
>> > + continue;
>> > +
>> > + rtx writeback_pat = NULL_RTX;
>> > + if (def->regno () == base_regno && (writeback & 2))
>> > + writeback_pat = orig_rtl[1];
>> > +
>> > + // We have a def in insns[1] which isn't def'd by the first insn.
>> > + // Look to the previous def and see if it has any debug uses.
>> > + for (auto use : prev_set->debug_insn_uses ())
>> > + if (*pair_dst < *use->insn ())
>> > + // We're ordering DEF above a previous use of the same register.
>> > + update_debug_use (use, def, writeback_pat);
>>
>> This might be more efficient as a reverse walk that breaks as soon as
>> *pair_dst >= *use->insn (), since the previous def could be arbitrarily
>> far away and have arbitrarily many leading uses. But it probably doesn't
>> matter much in practice.
>
> Agreed, provided it's easy to get at the last debug use in constant time
Yeah, sorry, I should have checked whether that was possible. Like you
pointed out off-list, it isn't possible as things stand, so never mind. :)
Richard
> (if so I guess 1/3 might need updating to add a last_debug_insn_use
> accessor). I'll look into that tomorrow.
>
>>
>> > + }
>> > +
>> > + if ((writeback & 2) && !writeback_effect)
>> > + {
>> > + // If the second insn initially had writeback but the final
>> > + // pair does not, then there may be trailing debug uses of the
>> > + // second writeback def which need re-parenting: do that.
>> > + auto def = find_access (insns[1]->defs (), base_regno);
>> > + gcc_assert (def);
>> > + for (auto use : as_a<set_info *> (def)->debug_insn_uses ())
>> > + {
>> > + insn_change change (use->insn ());
>> > + change.new_uses = check_remove_regno_access (attempt,
>> > + change.new_uses,
>> > + base_regno);
>> > + auto new_use = find_access (insns[0]->uses (), base_regno);
>> > +
>> > + // N.B. insns must have already shared a common base due to writeback.
>> > + gcc_assert (new_use);
>> > +
>> > + if (dump_file)
>> > + fprintf (dump_file,
>> > + " i%d: cancelling wb, re-parenting trailing debug use\n",
>> > + use->insn ()->uid ());
>> > +
>> > + change.new_uses = insert_access (attempt, new_use, change.new_uses);
>> > + crtl->ssa->change_insn (change);
>> > + }
>> > + }
>> > + else if (trailing_add)
>> > + fixup_debug_uses_trailing_add (attempt, pair_dst, trailing_add,
>> > + writeback_effect);
>> > +}
>> > +
>> > // Try and actually fuse the pair given by insns I1 and I2.
>> > //
>> > // Here we've done enough analysis to know this is safe, we only
>> > @@ -1378,6 +1681,9 @@ ldp_bb_info::fuse_pair (bool load_p,
>> > insn_info *first = (*i1 < *i2) ? i1 : i2;
>> > insn_info *second = (first == i1) ? i2 : i1;
>> >
>> > + insn_info *pair_dst = move_range.singleton ();
>> > + gcc_assert (pair_dst);
>> > +
>> > insn_info *insns[2] = { first, second };
>> >
>> > auto_vec<insn_change *> changes;
>> > @@ -1388,6 +1694,13 @@ ldp_bb_info::fuse_pair (bool load_p,
>> > PATTERN (second->rtl ())
>> > };
>> >
>> > + // Make copies of the patterns as we might need to refer to the
>> > original RTL
>> > + // later, for example when updating debug uses (which is after we've
>> > updated
>> > + // one or both of the patterns in the candidate insns).
>> > + rtx orig_rtl[2];
>> > + for (int i = 0; i < 2; i++)
>> > + orig_rtl[i] = copy_rtx (pats[i]);
>> > +
>>
>> FWIW, an alternative (that avoids RTL allocation) would be to use
>> temporarily_undo_changes and redo_changes. But this is fine too.
>
> Presumably using temporarily_undo_changes would require some bookkeping
> around the change numbers that might end up making things more
> complicated?
>
> I thought about making the copies conditional on MAY_HAVE_DEBUG_INSNS as
> that at least saves copying in the common case with no debug insns (and
> perhaps value-initializing orig_rtl to avoid it being used uninitialized
> by potential future patches). WDYT about that?
>
>>
>> The patch is OK from my POV with the as_a change above, but I'd be
>> interested in the answer to the writeback question above.
>
> Great, thanks a lot for the reviews!
>
> Alex
>
>>
>> Thanks,
>> Richard
>>
>> > use_array input_uses[2] = { first->uses (), second->uses () };
>> > def_array input_defs[2] = { first->defs (), second->defs () };
>> >
>> > @@ -1604,9 +1917,7 @@ ldp_bb_info::fuse_pair (bool load_p,
>> > using Action = stp_change_builder::action;
>> > insn_info *store_to_change = try_repurpose_store (first, second,
>> > move_range);
>> > - insn_info *stp_dest = move_range.singleton ();
>> > - gcc_assert (stp_dest);
>> > - stp_change_builder builder (insns, store_to_change, stp_dest);
>> > + stp_change_builder builder (insns, store_to_change, pair_dst);
>> > insn_change *change;
>> > set_info *new_set = nullptr;
>> > for (; !builder.done (); builder.advance ())
>> > @@ -1677,7 +1988,7 @@ ldp_bb_info::fuse_pair (bool load_p,
>> > fprintf (dump_file,
>> > " stp: changing i%d to use mem from new stp "
>> > "(after i%d)\n",
>> > - action.insn->uid (), stp_dest->uid ());
>> > + action.insn->uid (), pair_dst->uid ());
>> > change->new_uses = drop_memory_access (change->new_uses);
>> > gcc_assert (new_set);
>> > auto new_use = crtl->ssa->create_use (attempt, action.insn,
>> > @@ -1741,6 +2052,11 @@ ldp_bb_info::fuse_pair (bool load_p,
>> >
>> > gcc_assert (crtl->ssa->verify_insn_changes (changes));
>> >
>> > + // Fix up any debug uses that will be affected by the changes.
>> > + if (MAY_HAVE_DEBUG_INSNS)
>> > + fixup_debug_uses (attempt, insns, orig_rtl, pair_dst, trailing_add,
>> > + load_p, writeback, writeback_effect, base_regno);
>> > +
>> > confirm_change_group ();
>> > crtl->ssa->change_insns (changes);
>> >
>> > @@ -2807,7 +3123,7 @@ try_promote_writeback (insn_info *insn)
>> >
>> > rtx wb_effect = NULL_RTX;
>> > def_info *add_def;
>> > - const insn_range_info pair_range (insn->prev_nondebug_insn ());
>> > + const insn_range_info pair_range (insn);
>> > insn_info *insns[2] = { nullptr, insn };
>> > insn_info *trailing_add = find_trailing_add (insns, pair_range, 0,
>> > &wb_effect,
>> > &add_def, base_def, offset,
>> > @@ -2830,14 +3146,16 @@ try_promote_writeback (insn_info *insn)
>> > pair_change.new_defs);
>> > gcc_assert (pair_change.new_defs.is_valid ());
>> >
>> > - pair_change.move_range = insn_range_info (insn->prev_nondebug_insn ());
>> > -
>> > auto is_changing = insn_is_changing (changes);
>> > for (unsigned i = 0; i < ARRAY_SIZE (changes); i++)
>> > gcc_assert (rtl_ssa::restrict_movement_ignoring (*changes[i],
>> > is_changing));
>> >
>> > gcc_assert (rtl_ssa::recog_ignoring (attempt, pair_change,
>> > is_changing));
>> > gcc_assert (crtl->ssa->verify_insn_changes (changes));
>> > +
>> > + if (MAY_HAVE_DEBUG_INSNS)
>> > + fixup_debug_uses_trailing_add (attempt, insn, trailing_add,
>> > wb_effect);
>> > +
>> > confirm_change_group ();
>> > crtl->ssa->change_insns (changes);
>> > }
>> > diff --git a/gcc/testsuite/gcc.c-torture/compile/pr113089.c
>> > b/gcc/testsuite/gcc.c-torture/compile/pr113089.c
>> > new file mode 100644
>> > index 00000000000..70c71f23f1c
>> > --- /dev/null
>> > +++ b/gcc/testsuite/gcc.c-torture/compile/pr113089.c
>> > @@ -0,0 +1,26 @@
>> > +/* { dg-do compile } */
>> > +/* { dg-options "-g -funroll-loops" } */
>> > +
>> > +typedef unsigned short uint16;
>> > +
>> > +void intrapred_chroma_plane(uint16 ***mb_preds, int* max_imgpel_values,
>> > int crx, int cry, int px) {
>> > + for (int uv = 0; uv < 2; uv++) {
>> > + uint16 **mb_pred = mb_preds[uv + 1];
>> > + uint16 **predU2 = &mb_pred[px - 2];
>> > + uint16 *upPred = &mb_pred[px][px];
>> > + int max_imgpel_value = max_imgpel_values[uv];
>> > +
>> > + int ih = upPred[crx - 1];
>> > + for (int i = 0; i < crx*3; ++i)
>> > + ih += upPred[crx*3];
>> > +
>> > + int iv = (mb_pred[cry - 1][px+1]);
>> > + for (int i = 0; i < cry - 1; ++i) {
>> > + iv += (i + 1) * (*(mb_preds[uv][0]) - *(*predU2--));
>> > + }
>> > +
>> > + for (int j = 0; j < cry; ++j)
>> > + for (int i = 0; i < crx; ++i)
>> > + mb_pred[j][i] = (uint16) (max_imgpel_value * ((i * ih + iv)));
>> > + }
>> > +}
