On Wed, 11 Feb 2026, Jakub Jelinek wrote:
> Hi!
>
> force_nonfallthru_and_redirect has some extra code to handle asm gotos
> if we are forcing or redirecting the EDGE_FALLTHRU in case. This was
> done for PR51767, PR53589, PR54127.
> It is done for the cases where some asm goto label points to the bb
> after the fallthru edge (i.e. there is edge sharing between the fallthru
> and the label, partly or fully degenerate asm goto) and/or if the label
> points to the TARGET bb (before/after).
> In such case it changes the label from e->dest's label to target's label
> and similarly to the case of 2+ successor e->src a new bb is created
> as new fallthru from e->src and in the asm goto case an additional edge
> is added, so there is
> |
> asm goto -----+ maybe other edges
> | fallthru \
> v v
> jumpblock -> target
> and the jumpblock -> target edge isn't EDGE_FALLTHRU.
> This was done that way to fix various ICEs with the degenerate asm gotos,
> see above for list.
> I believe it will still ICE if we decide to force_nonfallthru_and_redirect
> E bb5->bb6 to l2, because in that case there are already two+ edges, one
> pointing from bb5->l2 and we create another edge bb5->l2:
> bb5:
> asm goto ("" : : : : (l2));
> bb6:
> ...
> l2:
> but maybe nothing tries that in that case. Anyway, the reason why
> the following two (PR116600 and PR123386) testcases are miscompiled is
> that we try to (during shrink-wrapping) emit some insns on the bb3->bb7
> edge, and see bb7 predecessor edge is EDGE_FALLTHRU in:
> bb3:
> __asm__ goto ("" : : : : d);
> if (c)
> bar (a);
> bb6:
> __asm__ goto ("" : : : : d);
> d: // == bb7
> in rtl_split_edge:
> /* We are going to place the new block in front of edge destination.
> Avoid existence of fallthru predecessors. */
> if ((edge_in->flags & EDGE_FALLTHRU) == 0)
> {
> edge e = find_fallthru_edge (edge_in->dest->preds);
>
> if (e)
> force_nonfallthru (e);
> }
> Now, the asm goto in bb6 is degenerate, so shrink-wrapping isn't aware
> that there would be multiple edges from it, there is just single one.
> But we still while splitting the bb3->bb7 edge in order to emit there
> pending insns on that edge turn the single successor asm goto bb into
> one with two successors (initially with two different ways to reach
> the same destination). But unfortunately such change confuses the
> shrink-wrapping logic which isn't aware of such a change and so the
> separate shrink wrapping insns end up just on one of the paths instead
> of being on both, and we then ICE in dwarf2 pass because there is an
> disagreement on the CFI state (of course it is a wrong-code too).
>
> Note, force_nonfallthru calls force_nonfallthru_and_redirect with
> target set to e->dest.
>
> The following patch reworks this.
> Instead of creating that
> |
> asm goto -----+ maybe other edges
> | fallthru \
> v v
> jumpblock -> target
> for initially degenerate asm goto we now create
> |
> asm goto maybe other edges
> | fallthru
> v
> jumpblock -> target
> i.e. the asm goto keeps being degenerate, all we've changed is adding
> a new bb on the fallthru edge and making the edge from that new bb
> non-fallthru. Furthermore, for the case where there would be before
> an asm goto -> target edge before for the e->dest != target case,
> those edges are untouched.
> For the immediate effect after the operation, the asm goto keeps
> the existing behavior, if it falls through, it will reach target
> by hopping through jumpblock, if it jumps to that label, previously
> it jumped directly to target, now it jumps to jumpblock and from there
> to target. But shrink-wrapping etc. then put the right fixups everywhere
> where it belongs.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
OK.
Richard.
> 2026-02-10 Jakub Jelinek <[email protected]>
>
> PR rtl-optimization/116600
> PR middle-end/123386
> * cfgrtl.cc (force_nonfallthru_and_redirect): Don't do any
> asm goto adjustments early, only note in asm_goto_edge if
> any labels point originally to e->dest head. After jumpblock
> creation don't add an extra edge for asm_goto_edge, instead
> adjust those labels pointing to former e->dest head to point
> to jumpblock instead.
>
> * gcc.c-torture/compile/pr116600.c: New test.
> * gcc.c-torture/compile/pr123386.c: New test.
>
> --- gcc/cfgrtl.cc.jj 2026-01-09 21:59:08.701543900 +0100
> +++ gcc/cfgrtl.cc 2026-02-10 11:04:06.811401150 +0100
> @@ -1594,54 +1594,20 @@ force_nonfallthru_and_redirect (edge e,
> }
>
> /* If e->src ends with asm goto, see if any of the ASM_OPERANDS_LABELs
> - don't point to the target or fallthru label. */
> + don't point to the fallthru label. */
> if (JUMP_P (BB_END (e->src))
> && target != EXIT_BLOCK_PTR_FOR_FN (cfun)
> && (e->flags & EDGE_FALLTHRU)
> && (note = extract_asm_operands (PATTERN (BB_END (e->src)))))
> {
> - int i, n = ASM_OPERANDS_LABEL_LENGTH (note);
> - bool adjust_jump_target = false;
> + int n = ASM_OPERANDS_LABEL_LENGTH (note);
>
> - for (i = 0; i < n; ++i)
> - {
> - if (XEXP (ASM_OPERANDS_LABEL (note, i), 0) == BB_HEAD (e->dest))
> - {
> - LABEL_NUSES (XEXP (ASM_OPERANDS_LABEL (note, i), 0))--;
> - XEXP (ASM_OPERANDS_LABEL (note, i), 0) = block_label (target);
> - LABEL_NUSES (XEXP (ASM_OPERANDS_LABEL (note, i), 0))++;
> - adjust_jump_target = true;
> - }
> - if (XEXP (ASM_OPERANDS_LABEL (note, i), 0) == BB_HEAD (target))
> + for (int i = 0; i < n; ++i)
> + if (XEXP (ASM_OPERANDS_LABEL (note, i), 0) == BB_HEAD (e->dest))
> + {
> asm_goto_edge = true;
> - }
> - if (adjust_jump_target)
> - {
> - rtx_insn *insn = BB_END (e->src);
> - rtx note;
> - rtx_insn *old_label = BB_HEAD (e->dest);
> - rtx_insn *new_label = BB_HEAD (target);
> -
> - if (JUMP_LABEL (insn) == old_label)
> - {
> - JUMP_LABEL (insn) = new_label;
> - note = find_reg_note (insn, REG_LABEL_TARGET, new_label);
> - if (note)
> - remove_note (insn, note);
> - }
> - else
> - {
> - note = find_reg_note (insn, REG_LABEL_TARGET, old_label);
> - if (note)
> - remove_note (insn, note);
> - if (JUMP_LABEL (insn) != new_label
> - && !find_reg_note (insn, REG_LABEL_TARGET, new_label))
> - add_reg_note (insn, REG_LABEL_TARGET, new_label);
> - }
> - while ((note = find_reg_note (insn, REG_LABEL_OPERAND, old_label))
> - != NULL_RTX)
> - XEXP (note, 0) = new_label;
> - }
> + break;
> + }
> }
>
> if (EDGE_COUNT (e->src->succs) >= 2 || abnormal_edge_flags ||
> asm_goto_edge)
> @@ -1680,15 +1646,45 @@ force_nonfallthru_and_redirect (edge e,
> and the reg crossing note should be removed. */
> fixup_partition_crossing (new_edge);
>
> - /* If asm goto has any label refs to target's label,
> - add also edge from asm goto bb to target. */
> + /* If asm goto has any label refs to e->dest, change them to point
> + to jump_block instead. */
> if (asm_goto_edge)
> {
> - new_edge->probability /= 2;
> - jump_block->count /= 2;
> - edge new_edge2 = make_edge (new_edge->src, target,
> - e->flags & ~EDGE_FALLTHRU);
> - new_edge2->probability = probability - new_edge->probability;
> + int n = ASM_OPERANDS_LABEL_LENGTH (note);
> +
> + for (int i = 0; i < n; ++i)
> + if (XEXP (ASM_OPERANDS_LABEL (note, i), 0) == BB_HEAD (e->dest))
> + {
> + LABEL_NUSES (XEXP (ASM_OPERANDS_LABEL (note, i), 0))--;
> + XEXP (ASM_OPERANDS_LABEL (note, i), 0)
> + = block_label (jump_block);
> + LABEL_NUSES (XEXP (ASM_OPERANDS_LABEL (note, i), 0))++;
> + }
> +
> + rtx_insn *insn = BB_END (new_edge->src);
> + rtx note;
> + rtx_insn *old_label = BB_HEAD (e->dest);
> + rtx_insn *new_label = BB_HEAD (jump_block);
> +
> + if (JUMP_LABEL (insn) == old_label)
> + {
> + JUMP_LABEL (insn) = new_label;
> + note = find_reg_note (insn, REG_LABEL_TARGET, new_label);
> + if (note)
> + remove_note (insn, note);
> + }
> + else
> + {
> + note = find_reg_note (insn, REG_LABEL_TARGET, old_label);
> + if (note)
> + remove_note (insn, note);
> + if (JUMP_LABEL (insn) != new_label
> + && !find_reg_note (insn, REG_LABEL_TARGET, new_label))
> + add_reg_note (insn, REG_LABEL_TARGET, new_label);
> + }
> + while ((note = find_reg_note (insn, REG_LABEL_OPERAND, old_label))
> + != NULL_RTX)
> + XEXP (note, 0) = new_label;
> }
>
> new_bb = jump_block;
> --- gcc/testsuite/gcc.c-torture/compile/pr116600.c.jj 2026-02-10
> 11:24:22.840629711 +0100
> +++ gcc/testsuite/gcc.c-torture/compile/pr116600.c 2026-02-10
> 11:24:15.386757048 +0100
> @@ -0,0 +1,24 @@
> +/* PR rtl-optimization/116600 */
> +
> +int a, b;
> +int foo ();
> +
> +int
> +bar ()
> +{
> + int c = ({ int d = 0, e = foo ();
> + b = __builtin_expect (e, 1); if (b) ; else d = 4; d; });
> + if (c)
> + return c;
> + foo ();
> + if (a)
> + __asm__ goto ("" : : : : l);
> +l:
> + return 0;
> +}
> +
> +void
> +baz ()
> +{
> + bar ();
> +}
> --- gcc/testsuite/gcc.c-torture/compile/pr123386.c.jj 2026-02-10
> 11:13:55.634344133 +0100
> +++ gcc/testsuite/gcc.c-torture/compile/pr123386.c 2026-02-10
> 11:13:34.528704662 +0100
> @@ -0,0 +1,25 @@
> +/* PR middle-end/123386 */
> +
> +char *a;
> +int b;
> +
> +void
> +foo (char *)
> +{
> + if (b)
> + __builtin_abort ();
> +}
> +
> +void
> +bar (char *c)
> +{
> + __asm__ goto ("" : : : : d);
> + foo (a);
> + __asm__ goto ("" : : : : d);
> + if (c)
> + bar (a);
> + __asm__ goto ("" : : : : d);
> +d:
> + if (b)
> + __builtin_abort ();
> +}
>
> Jakub
>
>
--
Richard Biener <[email protected]>
SUSE Software Solutions Germany GmbH,
Frankenstrasse 146, 90461 Nuernberg, Germany;
GF: Jochen Jaser, Andrew McDonald, Werner Knoblich; (HRB 36809, AG Nuernberg)
