[RFC PATCH 00/25] exec: Add load/store API for aligned pointers

["Philippe Mathieu-Daudé" <philmd@redhat.com>]

Hi,

This series is the result of a chat with Stefan after looking at
Peter response to Bibo on a problem with unoptimized memcpy()
leading to atomic issues:
https://www.mail-archive.com/qemu-devel@nongnu.org/msg808209.html

MIPS R6 (32/64-bit) does support unaligned accesses, but for the
older releases the compiler can't optimize the __builtin_memcpy().

When a pointer is known to be aligned, we can skip the checks for
misalignment and directly access the data. The pair of virtio vring
fields happen to be aligned.

To be able to call the aligned functions from the virtio layer, we
have to fill the gap in multiple API layers.

The series is decomposed as:
- cleanups (1-6)
- clean ldst API using macros (7-13)
- add aligned ldst methods (14)
- add aligned memory methods (15-16)
- similar changes in virtio (17-24)
- use the new methods on vring aligned values (25)

There are some checkpatch errors related to the macro used.

Regards,

Phil.

Philippe Mathieu-Daudé (25):
  exec/memory_ldst_cached: Sort declarations
  exec/memory_ldst_phys: Sort declarations
  exec/memory_ldst: Use correct type sizes
  exec/memory_ldst_phys: Use correct type sizes
  exec/memory_ldst_cached: Use correct type size
  exec/memory: Use correct type size
  qemu/bswap: Introduce ST_CONVERT() macro
  qemu/bswap: Use ST_CONVERT() macro to emit 16-bit load/store functions
  qemu/bswap: Introduce LD_CONVERT() macro
  qemu/bswap: Use LD_CONVERT macro to emit 16-bit signed load/store code
  qemu/bswap: Use LD_CONVERT macro to emit 16-bit unsigned load/store
    code
  qemu/bswap: Use LDST_CONVERT macro to emit 32-bit load/store functions
  qemu/bswap: Use LDST_CONVERT macro to emit 64-bit load/store functions
  qemu/bswap: Introduce load/store for aligned pointer
  exec/memory: Add methods for aligned pointer access (address space)
  exec/memory: Add methods for aligned pointer access (physical memory)
  hw/virtio: Use correct type sizes
  hw/virtio: Introduce VIRTIO_LD_CONVERT() macro
  hw/virtio: Use LD_CONVERT macro to emit 16-bit unsigned load/store
    code
  hw/virtio: Introduce VIRTIO_ST_CONVERT() macro
  hw/virtio: Use ST_CONVERT() macro to emit 16-bit load/store functions
  hw/virtio: Use LDST_CONVERT macro to emit 32-bit load/store functions
  hw/virtio: Use LDST_CONVERT macro to emit 64-bit load/store functions
  hw/virtio: Add methods for aligned pointer access
  hw/virtio: Optimize accesses on vring aligned pointers

 docs/devel/loads-stores.rst           |  27 +--
 include/exec/memory.h                 |  18 +-
 include/hw/virtio/virtio-access.h     | 239 +++++++++-----------------
 include/qemu/bswap.h                  | 149 ++++++----------
 include/exec/memory_ldst.h.inc        |  16 +-
 include/exec/memory_ldst_cached.h.inc |  66 ++++---
 include/exec/memory_ldst_phys.h.inc   |  72 ++++----
 hw/virtio/virtio.c                    |  13 +-
 memory_ldst.c.inc                     |  20 +--
 9 files changed, 270 insertions(+), 350 deletions(-)

-- 
2.26.3