On Wed, May 15, 2019 at 10:20 AM Alexandre Oliva <aol...@redhat.com> wrote:
>
> Gimple jump threading does not duplicate forwarder blocks that might
> be present before or after the second copied block.
Empty forwarder blocks I presume?
> This silently
> drops debug binds and markers that might be present in them. This
> patch attempts to preserve them.
>
> For blocks after either copied block, we attempt to append debug stmts
> to the copied block, if it does not end with a block-ending stmt.
> Failing that, for blocks between both copied blocks, we prepend its
> debug stmts to the copy of the second block.
>
> If everything fails, we still drop debug stmts on the floor, though
> preexisting code consolidates debug binds in the block that threading
> flows into, so only markers are really lost. We can't do much better
> than that without conditional binds and markers, or debug stmts in
> edges, or somesuch.
>
> If we append debug stmts to a reusable template block, we copy it
> after splitting out the debug stmts, and before putting them back.
>
> Regstrapped on x86_64-linux-gnu and i686-linux-gnu. Ok to install?
I wonder if not copying empty forwarders is premature optimization
in that (hopefully) CFG cleanup is run after jump threading anyways?
If we'd copy the blocks the patch would be moot?
Richard.
>
> for gcc/ChangeLog
>
> * tree-ssa-threadupdate.c (struct ssa_local_info_t): Add
> field template_last_to_copy.
> (ssa_create_duplicates): Set it, and use it. Attempt to
> preserve more debug stmts.
> ---
> gcc/tree-ssa-threadupdate.c | 68
> +++++++++++++++++++++++++++++++++++++++++++
> 1 file changed, 68 insertions(+)
>
> diff --git a/gcc/tree-ssa-threadupdate.c b/gcc/tree-ssa-threadupdate.c
> index bccef879db03..d63154317746 100644
> --- a/gcc/tree-ssa-threadupdate.c
> +++ b/gcc/tree-ssa-threadupdate.c
> @@ -235,6 +235,12 @@ struct ssa_local_info_t
> and sharing a template for that block is considerably more difficult.
> */
> basic_block template_block;
>
> + /* If we append debug stmts to the template block after creating it,
> + this iterator won't be the last one in the block, and further
> + copies of the template block shouldn't get debug stmts after
> + it. */
> + gimple_stmt_iterator template_last_to_copy;
> +
> /* Blocks duplicated for the thread. */
> bitmap duplicate_blocks;
>
> @@ -1124,6 +1130,8 @@ ssa_create_duplicates (struct redirection_data **slot,
> create_block_for_threading ((*path)[1]->e->src, rd, 0,
> &local_info->duplicate_blocks);
> local_info->template_block = rd->dup_blocks[0];
> + local_info->template_last_to_copy
> + = gsi_last_bb (local_info->template_block);
>
> /* We do not create any outgoing edges for the template. We will
> take care of that in a later traversal. That way we do not
> @@ -1131,14 +1139,74 @@ ssa_create_duplicates (struct redirection_data **slot,
> }
> else
> {
> + gimple_seq seq = NULL;
> + if (gsi_stmt (local_info->template_last_to_copy)
> + != gsi_stmt (gsi_last_bb (local_info->template_block)))
> + seq = gsi_split_seq_after (local_info->template_last_to_copy);
> create_block_for_threading (local_info->template_block, rd, 0,
> &local_info->duplicate_blocks);
> + if (seq)
> + gsi_insert_seq_after (&local_info->template_last_to_copy,
> + seq, GSI_SAME_STMT);
>
> /* Go ahead and wire up outgoing edges and update PHIs for the
> duplicate
> block. */
> ssa_fix_duplicate_block_edges (rd, local_info);
> }
>
> + if (MAY_HAVE_DEBUG_STMTS)
> + {
> + /* Copy debug stmts from each NO_COPY src block to the block
> + that would have been its predecessor, if we can append to it
> + (we can't add stmts after a block-ending stmt), or prepending
> + to the duplicate of the successor, if there is one. If
> + there's no duplicate successor, we'll mostly drop the blocks
> + on the floor; propagate_threaded_block_debug_into, called
> + elsewhere, will consolidate and preserve the effects of the
> + binds, but none of the markers. */
> + gimple_stmt_iterator copy_to = gsi_last_bb (rd->dup_blocks[0]);
> + if (!gsi_end_p (copy_to))
> + {
> + if (stmt_ends_bb_p (gsi_stmt (copy_to)))
> + {
> + if (rd->dup_blocks[1])
> + copy_to = gsi_after_labels (rd->dup_blocks[1]);
> + else
> + copy_to = gsi_none ();
> + }
> + else
> + gsi_next (©_to);
> + }
> + for (unsigned int i = 2, j = 0; i < path->length (); i++)
> + if ((*path)[i]->type == EDGE_NO_COPY_SRC_BLOCK
> + && gsi_bb (copy_to))
> + {
> + for (gimple_stmt_iterator gsi = gsi_start_bb ((*path)[i]->e->src);
> + !gsi_end_p (gsi); gsi_next (&gsi))
> + {
> + if (!is_gimple_debug (gsi_stmt (gsi)))
> + continue;
> + gimple *stmt = gsi_stmt (gsi);
> + gimple *copy = gimple_copy (stmt);
> + gsi_insert_before (©_to, copy, GSI_SAME_STMT);
> + }
> + }
> + else if ((*path)[i]->type == EDGE_COPY_SRC_BLOCK
> + || (*path)[i]->type == EDGE_COPY_SRC_JOINER_BLOCK)
> + {
> + j++;
> + gcc_assert (j < 2);
> + copy_to = gsi_last_bb (rd->dup_blocks[j]);
> + if (!gsi_end_p (copy_to))
> + {
> + if (stmt_ends_bb_p (gsi_stmt (copy_to)))
> + copy_to = gsi_none ();
> + else
> + gsi_next (©_to);
> + }
> + }
> + }
> +
> /* Keep walking the hash table. */
> return 1;
> }
>
>
> --
> Alexandre Oliva, freedom fighter he/him https://FSFLA.org/blogs/lxo
> Be the change, be Free! FSF Latin America board member
> GNU Toolchain Engineer Free Software Evangelist
> Hay que enGNUrecerse, pero sin perder la terGNUra jamás - Che GNUevara
