On Thu, Oct 1, 2026 at 6:22 PM <[email protected]> wrote:
>
> From: Kyrylo Tkachov <[email protected]>
>
> bbcb0c056be (PR15242) stopped try_crossjump_bb from sharing the tails of
> the predecessors of a computed jump unless optimizing for size.
> pass_duplicate_computed_gotos copies the jump back into each predecessor
> and a shared tail defeats that. efd8f7507b3 replaced !optimize_size with
> optimize_bb_for_size_p (bb), which inverted the test: tails are now shared
> in code optimized for speed and kept apart in code optimized for size.
>
> Test optimize_function_for_speed_p (cfun) instead. PR42621 fixed the
> bb-reorder half of the same change in the same way, in the gate of
> pass_duplicate_computed_gotos. As in bbcb0c056be, the guard then applies
> to the whole function, also to a block that is too large to be unfactored.
> try_head_merge_bb has a copy of the inverted test that cannot change code,
> because every successor edge of a computed jump is abnormal. It stays as
> it is.
>
> On the PR127246 testcase with identical handler tails, jump2 no longer
> crossjumps the handlers at -O2 only for compgotos to copy the tails back.
> At -Os the handler tails are shared again, which makes the aarch64 code of
> the PR testcase 54% smaller.
>
> Bootstrapped and regression tested on aarch64-unknown-linux-gnu and
> x86_64-pc-linux-gnu.
>
> Ok for trunk?
> Thanks,
> Kyrill
>
> gcc/ChangeLog:
>
> PR rtl-optimization/127246
> * cfgcleanup.cc (try_crossjump_bb): Keep predecessors of a computed
> jump apart when the function, not the block, is optimized for speed.
>
> gcc/testsuite/ChangeLog:
>
> PR rtl-optimization/127246
> * gcc.dg/pr127246-1.c: New test.
> * gcc.dg/pr127246-2.c: New test.
>
> Signed-off-by: Kyrylo Tkachov <[email protected]>
> ---
> gcc/cfgcleanup.cc | 2 +-
> gcc/testsuite/gcc.dg/pr127246-1.c | 48 +++++++++++++++++++++++++++++++
> gcc/testsuite/gcc.dg/pr127246-2.c | 12 ++++++++
> 3 files changed, 61 insertions(+), 1 deletion(-)
> create mode 100644 gcc/testsuite/gcc.dg/pr127246-1.c
> create mode 100644 gcc/testsuite/gcc.dg/pr127246-2.c
>
> diff --git a/gcc/cfgcleanup.cc b/gcc/cfgcleanup.cc
> index 1d9ec908dab..7e5a0775e22 100644
> --- a/gcc/cfgcleanup.cc
> +++ b/gcc/cfgcleanup.cc
> @@ -2182,7 +2182,7 @@ try_crossjump_bb (int mode, basic_block bb)
>
> /* Don't crossjump if this block ends in a computed jump,
> unless we are optimizing for size. */
> - if (optimize_bb_for_size_p (bb)
> + if (optimize_function_for_speed_p (cfun)
pass_duplicate_computed_gotos::gate has
return (optimize > 0
&& flag_expensive_optimizations
&& ! optimize_function_for_size_p (fun));
so please use ! optimize_function_for_size_p (cfun) &&
flag_expensive_optimizations
and adjust the comment to mention the reason, duplicating computed gotos.
OK with that change.
Richard.
> && bb != EXIT_BLOCK_PTR_FOR_FN (cfun)
> && computed_jump_p (BB_END (bb)))
> return false;
> diff --git a/gcc/testsuite/gcc.dg/pr127246-1.c
> b/gcc/testsuite/gcc.dg/pr127246-1.c
> new file mode 100644
> index 00000000000..cfea56d631d
> --- /dev/null
> +++ b/gcc/testsuite/gcc.dg/pr127246-1.c
> @@ -0,0 +1,48 @@
> +/* PR rtl-optimization/127246 */
> +/* { dg-do compile } */
> +/* { dg-require-effective-target label_values } */
> +/* { dg-require-effective-target indirect_jumps } */
> +/* { dg-options "-O2 -fno-tree-slp-vectorize -fdump-rtl-jump2-details" } */
> +
> +/* pass_duplicate_computed_gotos copies the dispatch back into every handler,
> + so tail merging must not share the identical handler tails in front of
> + it. */
> +
> +struct frame
> +{
> + const unsigned short *instr_ptr;
> + long *stackpointer;
> +};
> +
> +#define OPS(X) \
> + X (0) X (1) X (2) X (3) X (4) X (5) X (6) X (7) X (8) X (9) X (10) \
> + X (11) X (12) X (13) X (14) X (15) X (16) X (17) X (18) X (19)
> +
> +extern long op (long);
> +
> +#define DISPATCH() goto *targets[*next_instr & 0xff]
> +
> +#define HANDLER(N) \
> + L_##N: \
> + frame->instr_ptr = next_instr; \
> + next_instr += 1; \
> + frame->stackpointer = stack_pointer; \
> + stack_pointer[-1] = op (N); \
> + stack_pointer -= 1; \
> + DISPATCH ();
> +
> +#define ENTRY(N) &&L_##N,
> +
> +void
> +interp (struct frame *frame)
> +{
> + static const void *const targets[] = { OPS (ENTRY) };
> + const unsigned short *next_instr = frame->instr_ptr;
> + long *stack_pointer = frame->stackpointer;
> +
> + DISPATCH ();
> +
> + OPS (HANDLER)
> +}
> +
> +/* { dg-final { scan-rtl-dump-not "Cross jumping" "jump2" } } */
> diff --git a/gcc/testsuite/gcc.dg/pr127246-2.c
> b/gcc/testsuite/gcc.dg/pr127246-2.c
> new file mode 100644
> index 00000000000..02328f04eab
> --- /dev/null
> +++ b/gcc/testsuite/gcc.dg/pr127246-2.c
> @@ -0,0 +1,12 @@
> +/* PR rtl-optimization/127246 */
> +/* { dg-do compile } */
> +/* { dg-require-effective-target label_values } */
> +/* { dg-require-effective-target indirect_jumps } */
> +/* { dg-options "-Os -fdump-rtl-jump2-details" } */
> +
> +/* When optimizing for size the computed goto is not unfactored, so tail
> + merging the handler tails is what we want. */
> +
> +#include "pr127246-1.c"
> +
> +/* { dg-final { scan-rtl-dump "Cross jumping" "jump2" } } */
> --
> 2.50.1 (Apple Git-155)
>
