[for-7.0,v8,00/31] LoongArch64 port of QEMU TCG

Message

WANG Xuerui Nov. 24, 2021, 6:32 p.m. UTC

Hi all,

This is a port of QEMU TCG to the brand-new CPU architecture LoongArch,
introduced by Loongson with their 3A5000 chips. As the 6.2 merge window
is already over, this work is aimed at 7.0 instead; there's no hurry
though. (I need a QEMU wiki account btw)

Tests (run with `ninja test`) all passed, as usual; I also boot-tested
x86_64 (Debian and Gentoo installation CDs) and install-tested aarch64
(Debian netboot installer), and ran riscv64 linux-user emulation with a
chroot; everything seems fine. Cross build jobs for loongarch64 hosts are
added to CI matrix, and are confirmed working in my GitLab fork.

## About the series

Only the LP64D ABI is supported, as this is the only one fully
implemented and supported by Loongson. 32-bit support is incomplete from
outset, and removed from the very latest upstream submissions, so you
can't even configure for that.

The architecture's documentation is already translated into English;
it can be browsed at https://loongson.github.io/LoongArch-Documentation/.
The LoongArch ELF psABI doc (version 1.00) could be found at [1];
if anything is missing there, it's most likely the same as RISC-V, but
you can always raise an issue over their issue tracker at [2].

[1]: https://loongson.github.io/LoongArch-Documentation/LoongArch-ELF-ABI-EN.html
[2]: https://github.com/loongson/LoongArch-Documentation/issues

In this series I made use of generated instruction encodings and
emitters from https://github.com/loongson-community/loongarch-opcodes
(a community project started by myself, something I must admit), as the
LoongArch encoding is highly irregular even for a fixed 32-bit ISA, and
I want to minimize the maintenance burden for future collaboration.

This series touches some of the same files as Song Gao's previous
submission of LoongArch *target* support, which is a bit unfortunate;
one of us will have to rebase after either series gets in. Actual
conflict should only happen on build system bits and include/elf.h,
though, as we're working on entirely different areas.

## How to build and test this

Upstream support for LoongArch is largely WIP for now, which means you
must apply a lot of patches if you want to even cross-build for this arch.
The main sources I used are as follows:

* binutils: (already upstream as of November 2021)
* gcc: https://github.com/xen0n/gcc/tree/for-gentoo-gcc-12-v2
  based on https://github.com/loongson/gcc/tree/loongarch_upstream
* glibc: https://github.com/xen0n/glibc/tree/for-gentoo-glibc-2.34
  based on https://github.com/loongson/glibc/tree/loongarch_2_34_for_upstream
* Linux: https://github.com/xen0n/linux/tree/loongarch-playground
  based on https://github.com/loongson/linux/tree/loongarch-next
* Gentoo overlay: https://github.com/xen0n/loongson-overlay

I have made ready-to-use Gentoo stage3 tarballs, but they're served with
CDN off my personal cloud account, and I don't want the link to be
exposed so that my bills skyrocket; you can reach me off-list to get the
links if you're interested.

As for the hardware availability, the boards can already be bought in
China on Taobao, and I think some people at Loongson might be able to
arrange for testing environments, if testing on real hardware other than
mine is required before merging; they have their in-house Debian spin-off
from the early days of this architecture. Their kernel is
ABI-incompatible with the version being upstreamed and used by me, but
QEMU should work there regardless.

Lastly, I'm new to QEMU development and this is my first patch series
here; apologizes if I get anything wrong, and any help or suggestion is
certainly appreciated!

## Changelog

v8 -> v7:

- Rebased to latest development branch
- Adjusted component field of Patch 29's subject line to say
  "linux-user" after rebase
- Adjusted Patch 31 slightly according to review (you can check out the
  Patchew diff between revisions for details)

v7 -> v6:

- Rebased to latest development branch
- Added Docker image for cross-compilation and cross-build CI jobs

v6 -> v5:

- Fixed many places using 0/1 to say false/true
- Tweaks to tcg_out_movi
  - Moved variable declarations to top of function, as per QEMU coding
    style
  - Added ASCII art to better explain names like `hi12` `hi32` and `hi52`
- Added example `uname -a` outputs to commit message of Patch 30 to help
  people make sense of the change

v5 -> v4:

- Updated generated instruction definition to latest (added ldx/stx
  family of indexed load/stores)
- Incorporated Richard's suggestion for tcg_out_movi, tested to cover
  the cases it's supposed to improve
- Fixed a "size == MO_64" occurrence to just say "type"
- Used indexed load/stores to optimize qemu_ld/st
- Fixed zero-extension of address register for qemu_ld/st on 32-bit
  targets

v4 -> v3:

- Addressed all review comments from v3
  - Made MAX_CODE_GEN_BUFFER_SIZE to be just SIZE_MAX (but kept
    TCG_TARGET_NB_REGS as macro definition)
  - Updated generated instruction definition, made it clear that the
    whole file is generated
  - Used deposit64 for hand-written relocation code
  - Reworked tcg_out_movi
    - Use pcalau12i + ori for PC-relative values whose offset fits in
      32-bit
    - Every individual insn in the slow path (lu12i.w + ori + cu32i.d +
      cu52i.d) can be suppressed if not needed
  - Fixed constraint of setcond ops, don't allow constant zero for 1st
    operand

v3 -> v2:

- Addressed all review comments from v2
  - Re-organized changes to tcg-target.h so that it's incrementally
    updated in each commit implementing ops
  - Removed support for the eqv op
  - Added support for bswap16_i{32,64} ops
  - Fixed and refactored various places as pointed out during review
- Updated generated instruction definitions to latest

v2 -> v1:

- Addressed all review comments from v1
  - Use "loongarch64" everywhere, tcg directory renamed to "tcg/loongarch64"
  - Removed all redundant TCG_TARGET_REG_BITS conditional
  - Removed support for the neg op
  - Added support for eqv and bswap32_i64 ops
  - Added safe syscall handling for linux-user
  - Fixed everything else I could see
- Updated generated instruction definitions to latest
- Reordered the configure/meson.build changes to come last

v7: https://patchew.org/QEMU/[e-mail]/
v6: https://patchew.org/QEMU/[e-mail]/
v5: https://patchew.org/QEMU/[e-mail]/
v4: https://patchew.org/QEMU/[e-mail]/
v3: https://patchew.org/QEMU/[e-mail]/
v2: https://patchew.org/QEMU/[e-mail]/
v1: https://patchew.org/QEMU/[e-mail]/

WANG Xuerui (31):
  elf: Add machine type value for LoongArch
  MAINTAINERS: Add tcg/loongarch64 entry with myself as maintainer
  tcg/loongarch64: Add the tcg-target.h file
  tcg/loongarch64: Add generated instruction opcodes and encoding
    helpers
  tcg/loongarch64: Add register names, allocation order and input/output
    sets
  tcg/loongarch64: Define the operand constraints
  tcg/loongarch64: Implement necessary relocation operations
  tcg/loongarch64: Implement the memory barrier op
  tcg/loongarch64: Implement tcg_out_mov and tcg_out_movi
  tcg/loongarch64: Implement goto_ptr
  tcg/loongarch64: Implement sign-/zero-extension ops
  tcg/loongarch64: Implement not/and/or/xor/nor/andc/orc ops
  tcg/loongarch64: Implement deposit/extract ops
  tcg/loongarch64: Implement bswap{16,32,64} ops
  tcg/loongarch64: Implement clz/ctz ops
  tcg/loongarch64: Implement shl/shr/sar/rotl/rotr ops
  tcg/loongarch64: Implement add/sub ops
  tcg/loongarch64: Implement mul/mulsh/muluh/div/divu/rem/remu ops
  tcg/loongarch64: Implement br/brcond ops
  tcg/loongarch64: Implement setcond ops
  tcg/loongarch64: Implement tcg_out_call
  tcg/loongarch64: Implement simple load/store ops
  tcg/loongarch64: Add softmmu load/store helpers, implement
    qemu_ld/qemu_st ops
  tcg/loongarch64: Implement tcg_target_qemu_prologue
  tcg/loongarch64: Implement exit_tb/goto_tb
  tcg/loongarch64: Implement tcg_target_init
  tcg/loongarch64: Register the JIT
  linux-user: Add safe syscall handling for loongarch64 hosts
  linux-user: Implement CPU-specific signal handler for loongarch64
    hosts
  configure, meson.build: Mark support for loongarch64 hosts
  tests/docker: Add gentoo-loongarch64-cross image and run cross builds
    in GitLab

 .gitlab-ci.d/container-cross.yml              |   27 +
 .gitlab-ci.d/crossbuilds.yml                  |   19 +
 MAINTAINERS                                   |    7 +
 configure                                     |    7 +-
 include/elf.h                                 |    2 +
 linux-user/host/loongarch64/host-signal.h     |   87 +
 linux-user/host/loongarch64/hostdep.h         |   14 +
 .../host/loongarch64/safe-syscall.inc.S       |   80 +
 meson.build                                   |    2 +-
 tcg/loongarch64/tcg-insn-defs.c.inc           |  979 ++++++++++
 tcg/loongarch64/tcg-target-con-set.h          |   31 +
 tcg/loongarch64/tcg-target-con-str.h          |   28 +
 tcg/loongarch64/tcg-target.c.inc              | 1677 +++++++++++++++++
 tcg/loongarch64/tcg-target.h                  |  180 ++
 tests/docker/Makefile.include                 |   21 +
 .../gentoo-loongarch64-cross.docker           |   19 +
 .../build-toolchain.sh                        |  125 ++
 17 files changed, 3303 insertions(+), 2 deletions(-)
 create mode 100644 linux-user/host/loongarch64/host-signal.h
 create mode 100644 linux-user/host/loongarch64/hostdep.h
 create mode 100644 linux-user/host/loongarch64/safe-syscall.inc.S
 create mode 100644 tcg/loongarch64/tcg-insn-defs.c.inc
 create mode 100644 tcg/loongarch64/tcg-target-con-set.h
 create mode 100644 tcg/loongarch64/tcg-target-con-str.h
 create mode 100644 tcg/loongarch64/tcg-target.c.inc
 create mode 100644 tcg/loongarch64/tcg-target.h
 create mode 100644 tests/docker/dockerfiles/gentoo-loongarch64-cross.docker
 create mode 100755 tests/docker/dockerfiles/gentoo-loongarch64-cross.docker.d/build-toolchain.sh