Bug report is
https://lore.kernel.org/qemu-devel/33748BB7-E617-4661-BDE3-5D29780FC9FF@wdc.com

Jørgen properly identified the problem, though calling set_can_do_io
in the middle of translator_access is incorrect.

It forced me to do something that's been in the back of my mind for
a while and track the insns being emitted so that we can go back and
insert the proper store to can_do_io after code for the final insn
has been emitted.

If we decide to take this for 9.0, most of the target changes could
technically be left for 9.1.  With the exception of i386 and its funky
case of "pretend we never started translating the current insn".
But I think the target changes are minor enough to take them anyway.


r~


Richard Henderson (9):
  tcg: Add TCGContext.emit_before_op
  accel/tcg: Add insn_start to DisasContextBase
  target/arm: Use insn_start from DisasContextBase
  target/hppa: Use insn_start from DisasContextBase
  target/i386: Preserve DisasContextBase.insn_start across rewind
  target/microblaze: Use insn_start from DisasContextBase
  target/riscv: Use insn_start from DisasContextBase
  target/s390x: Use insn_start from DisasContextBase
  accel/tcg: Improve can_do_io management

 include/exec/translator.h      |  4 ++-
 include/tcg/tcg.h              |  6 +++++
 target/arm/tcg/translate.h     | 12 ++++-----
 accel/tcg/translator.c         | 47 ++++++++++++++++++----------------
 target/arm/tcg/translate-a64.c |  2 +-
 target/arm/tcg/translate.c     |  2 +-
 target/hppa/translate.c        | 10 ++++----
 target/i386/tcg/translate.c    |  3 +++
 target/microblaze/translate.c  |  8 ++----
 target/riscv/translate.c       | 11 ++++----
 target/s390x/tcg/translate.c   |  4 +--
 tcg/tcg.c                      | 14 ++++++++--
 12 files changed, 70 insertions(+), 53 deletions(-)

-- 
2.34.1