CPUNegativeOffsetState is a struct placed immediately before
CPUArchState in the ArchCPU struct.  Its purpose is to ensure that
certain fields (CPUTLBDescFast, IcountDecr) lay within a small negative
offset of CPUArchState in memory.  This is desired for better
code-generation on arm[32|64] and riscv hosts which has addressing
modes with 12- and 11 bits of displacement respectively.

This patchset removes CPUNegativeOffsetState, moves its fields to
CPUState, and statically asserts that the offset to the fields above
is expressable in 11 bits of displacement ( >= -(1 << 11) ). In order
to achieve this the TARGET_PAGE_ENTRY_EXTRA macro in CPUTLBEntryFull
had to be replaced with a union to make CPUTLB target independent.

The motivation for this patchset is twofold:

    1. Parts of the codebase that previously depended on CPUArchState
       to access either CPUTLB or IcountDecr now only depend on the
       target-agnostic CPUState.  This is a step towards building
       accel/ once for system- and once for user-mode.

    2. Targets no longer have to define a CPUNegativeOffsetState
       member of ArchCPU, and QEMU will fail to compile if CPUTLB
       and IcountDecr drift too far from CPUArchState.

Patches will follow that convert accel/tcg/cputlb.c and
accel/tcg/user-exec.c away from CPUArchState.

Anton Johansson (9):
  target/arm: Replace TARGET_PAGE_ENTRY_EXTRA
  include: Move MMUAccessType to tlb-common.h
  include/exec: Move CPUTLB and friends to tlb-common.h
  include/hw: introduce CPU_MAX_NEGATIVE_ENV_OFFSET
  accel: Move CPUTLB to CPUState and assert offset
  Move IcountDecr to CPUState and assert offset
  include/exec: Remove [cpu|env]_neg() functions
  target: Remove CPUNegativeOffsetState field from ArchCPU
  include/exec: Remove CPUNegativeOffsetState

 include/exec/cpu-all.h           |  28 +-----
 include/exec/cpu-defs.h          | 141 ----------------------------
 include/exec/exec-all.h          |   2 +-
 include/exec/tlb-common.h        | 153 +++++++++++++++++++++++++++++++
 include/hw/core/cpu.h            |  27 ++++--
 target/alpha/cpu.h               |   1 -
 target/arm/cpu-param.h           |  12 ---
 target/arm/cpu.h                 |   1 -
 target/avr/cpu.h                 |   1 -
 target/cris/cpu.h                |   1 -
 target/hexagon/cpu.h             |   1 -
 target/hppa/cpu.h                |   1 -
 target/i386/cpu.h                |   1 -
 target/loongarch/cpu.h           |   1 -
 target/m68k/cpu.h                |   1 -
 target/microblaze/cpu.h          |   1 -
 target/mips/cpu.h                |   3 +-
 target/nios2/cpu.h               |   1 -
 target/openrisc/cpu.h            |   1 -
 target/ppc/cpu.h                 |   1 -
 target/riscv/cpu.h               |   1 -
 target/rx/cpu.h                  |   1 -
 target/s390x/cpu.h               |   1 -
 target/sh4/cpu.h                 |   1 -
 target/sparc/cpu.h               |   1 -
 target/tricore/cpu.h             |   1 -
 target/xtensa/cpu.h              |   3 +-
 accel/tcg/cpu-exec.c             |  14 +--
 accel/tcg/tcg-accel-ops-icount.c |   6 +-
 accel/tcg/tcg-accel-ops.c        |   2 +-
 accel/tcg/translate-all.c        |  19 +++-
 accel/tcg/translator.c           |  15 ++-
 softmmu/icount.c                 |   2 +-
 target/arm/ptw.c                 |   4 +-
 target/arm/tcg/mte_helper.c      |   2 +-
 target/arm/tcg/sve_helper.c      |   2 +-
 target/arm/tcg/tlb_helper.c      |   4 +-
 target/arm/tcg/translate-a64.c   |   2 +-
 38 files changed, 222 insertions(+), 238 deletions(-)

--
2.41.0

On 6/30/23 14:25, Anton Johansson via wrote:
> CPUNegativeOffsetState is a struct placed immediately before
> CPUArchState in the ArchCPU struct.  Its purpose is to ensure that
> certain fields (CPUTLBDescFast, IcountDecr) lay within a small negative
> offset of CPUArchState in memory.  This is desired for better
> code-generation on arm[32|64] and riscv hosts which has addressing
> modes with 12- and 11 bits of displacement respectively.

The purpose is also to ensure that general purpose registers stay close 
to the beginning of the CPUArchState and thus can also be accessed with 
a small displacement.

Can you check if this becomes worse for any architecture?  On some 
64-bit targets, 8 bytes * 32 registers is 512 bytes and it's a 
substantial part of the 11 bits that are available.

Paolo

On 7/1/23 11:21, Paolo Bonzini wrote:
> On 6/30/23 14:25, Anton Johansson via wrote:
>> CPUNegativeOffsetState is a struct placed immediately before
>> CPUArchState in the ArchCPU struct.  Its purpose is to ensure that
>> certain fields (CPUTLBDescFast, IcountDecr) lay within a small negative
>> offset of CPUArchState in memory.  This is desired for better
>> code-generation on arm[32|64] and riscv hosts which has addressing
>> modes with 12- and 11 bits of displacement respectively.
>
> The purpose is also to ensure that general purpose registers stay 
> close to the beginning of the CPUArchState and thus can also be 
> accessed with a small displacement.
>
> Can you check if this becomes worse for any architecture?  On some 
> 64-bit targets, 8 bytes * 32 registers is 512 bytes and it's a 
> substantial part of the 11 bits that are available.
>
> Paolo
>
I'm dropping the CPUNegativeOffsetState changes, but the fields would 
still have had
a negative displacement from the beginning of env, so the positive 
offset of the gprs from env
wouldn't be affected.

-- 
Anton Johansson,
rev.ng Labs Srl.