On Thu, Feb 25, 2021 at 2:29 PM Richard Biener <rguent...@suse.de> wrote:
>
> Most callers of connect_infinite_loops_to_exit already do this but
> the few that do not end up with extra exit edges. The following
> makes that consistent, also matching the post-dominance DFS walk code.
>
> Boostrapped and tested on x86_64-unknown-linux-gnu, queued for stage1.
g:58ad6b2802592f1448eed48e8ad4e0e87985cecb
> 2021-02-25 Richard Biener <rguent...@suse.de>
>
> * cfganal.c (connect_infinite_loops_to_exit): First call
> add_noreturn_fake_exit_edges.
> * ipa-sra.c (process_scan_results): Do not call the now redundant
> add_noreturn_fake_exit_edges.
> * predict.c (tree_estimate_probability): Likewise.
> (rebuild_frequencies): Likewise.
> * store-motion.c (one_store_motion_pass): Likewise.
> ---
> gcc/cfganal.c | 10 +++++++---
> gcc/ipa-sra.c | 1 -
> gcc/predict.c | 2 --
> gcc/store-motion.c | 1 -
> 4 files changed, 7 insertions(+), 7 deletions(-)
>
> diff --git a/gcc/cfganal.c b/gcc/cfganal.c
> index 2627c2ff457..cec5abe30f9 100644
> --- a/gcc/cfganal.c
> +++ b/gcc/cfganal.c
> @@ -582,9 +582,9 @@ add_noreturn_fake_exit_edges (void)
> make_single_succ_edge (bb, EXIT_BLOCK_PTR_FOR_FN (cfun), EDGE_FAKE);
> }
>
> -/* This function adds a fake edge between any infinite loops to the
> - exit block. Some optimizations require a path from each node to
> - the exit node.
> +/* This function adds a fake edge between any noreturn block and
> + infinite loops to the exit block. Some optimizations require a path
> + from each node to the exit node.
>
> See also Morgan, Figure 3.10, pp. 82-83.
>
> @@ -596,6 +596,10 @@ add_noreturn_fake_exit_edges (void)
> void
> connect_infinite_loops_to_exit (void)
> {
> + /* First add fake exits to noreturn blocks, this is required to
> + discover only truly infinite loops below. */
> + add_noreturn_fake_exit_edges ();
> +
> /* Perform depth-first search in the reverse graph to find nodes
> reachable from the exit block. */
> depth_first_search dfs;
> diff --git a/gcc/ipa-sra.c b/gcc/ipa-sra.c
> index 1571921cb48..7a89906cee6 100644
> --- a/gcc/ipa-sra.c
> +++ b/gcc/ipa-sra.c
> @@ -2394,7 +2394,6 @@ process_scan_results (cgraph_node *node, struct
> function *fun,
> if (!pdoms_calculated)
> {
> gcc_checking_assert (cfun);
> - add_noreturn_fake_exit_edges ();
> connect_infinite_loops_to_exit ();
> calculate_dominance_info (CDI_POST_DOMINATORS);
> pdoms_calculated = true;
> diff --git a/gcc/predict.c b/gcc/predict.c
> index d0a8e5f8e04..0bf1748ffa8 100644
> --- a/gcc/predict.c
> +++ b/gcc/predict.c
> @@ -3106,7 +3106,6 @@ tree_estimate_probability (bool dry_run)
> {
> basic_block bb;
>
> - add_noreturn_fake_exit_edges ();
> connect_infinite_loops_to_exit ();
> /* We use loop_niter_by_eval, which requires that the loops have
> preheaders. */
> @@ -4291,7 +4290,6 @@ rebuild_frequencies (void)
> if (profile_status_for_fn (cfun) == PROFILE_GUESSED)
> {
> loop_optimizer_init (0);
> - add_noreturn_fake_exit_edges ();
> mark_irreducible_loops ();
> connect_infinite_loops_to_exit ();
> estimate_bb_frequencies (true);
> diff --git a/gcc/store-motion.c b/gcc/store-motion.c
> index f0401cae272..3f6e003219d 100644
> --- a/gcc/store-motion.c
> +++ b/gcc/store-motion.c
> @@ -1152,7 +1152,6 @@ one_store_motion_pass (void)
>
> /* Now compute kill & transp vectors. */
> build_store_vectors ();
> - add_noreturn_fake_exit_edges ();
> connect_infinite_loops_to_exit ();
>
> edge_list = pre_edge_rev_lcm (num_stores, st_transp, st_avloc,
> --
> 2.26.2
