The Linux kernel and the following testcase distilled from it is
miscompiled, because tree-outof-ssa.cc (eliminate_phi) emits some
fixups on some of the edges (but doesn't commit edge insertions).
Later expand_asm_stmt emits further instructions on the same edge.
Now the problem is that expand_asm_stmt uses insert_insn_on_edge
to add its own fixups, but that function appends to the existing
sequence on the edge if any.  And the bug triggers when the
fixup sequence emitted by eliminate_phi uses a pseudo which the
fixup sequence emitted by expand_asm_stmt later on sets.
So, we end up with
  (set (reg A) (asm_operands ...))
and on one of the edges queued sequence
  (set (reg C) (reg B)) // added by eliminate_phi
  (set (reg B) (reg A)) // added by expand_asm_stmt
That is wrong, what we emit by expand_asm_stmt needs to be as close
to the asm_operands as possible (they aren't known until expand_asm_stmt
is called, the PHI fixup code assumes it is reg B which holds the right
value) and the PHI adjustments need to be done after it.

So, the following patch introduces a prepend_insn_to_edge function and
uses it from expand_asm_stmt, so that we queue
  (set (reg B) (reg A)) // added by expand_asm_stmt
  (set (reg C) (reg B)) // added by eliminate_phi
instead and so the value from the asm_operands output propagates correctly
to the PHI result.

2024-02-15  Jakub Jelinek  <jakub@redhat.com>

	PR middle-end/113921
	* cfgrtl.h (prepend_insn_to_edge): New declaration.
	* cfgrtl.cc (insert_insn_on_edge): Clarify behavior in function
	comment.
	(prepend_insn_to_edge): New function.
	* cfgexpand.cc (expand_asm_stmt): Use prepend_insn_to_edge instead of
	insert_insn_on_edge.

	* gcc.target/i386/pr113921.c: New test.