This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
[PATCH, AArch64 00/11] LSE atomics out-of-line
- From: rth7680 at gmail dot com
- To: gcc-patches at gcc dot gnu dot org
- Cc: ramana dot radhakrishnan at arm dot com, agraf at suse dot de, matz at suse dot de, Richard Henderson <richard dot henderson at linaro dot org>
- Date: Tue, 25 Sep 2018 22:03:44 -0700
- Subject: [PATCH, AArch64 00/11] LSE atomics out-of-line
From: Richard Henderson <richard.henderson@linaro.org>
ARMv8.1 adds an (mandatory) Atomics extension, also known as the
Large System Extension. Deploying this extension at the OS level
has proved challenging.
The following is the result of a conversation between myself,
Alex Graf of SuSE, and Ramana Radhakrishnan of ARM, at last week's
Linaro Connect in Vancouver.
The current state of the world is that one could distribute two
different copies of a given shared library and place the LSE-enabled
version in /lib64/atomics/ and it will be selected over the /lib64/
version by ld.so when HWCAP_ATOMICS is present.
Alex's main concern with this is that (1) he doesn't want to
distribute two copies of every library, or determine what a
resonable subset would be and (2) this solution does not work
for executables, e.g. mysql.
Ramana's main concern was to avoid the overhead of an indirect jump,
especially in how that would affect the (non-)branch-prediction of
the smallest implementations.
Therefore, I've created small out-of-line helpers that are directly
linked into every library or executable that requires them. There
will be two direct branches, both of which will be well-predicted.
In the process, I discovered a number of places within the code
where the existing implementation could be improved. In particular:
- the LSE patterns didn't use predicates or constraints that
match the actual instructions, requiring unnecessary splitting.
- the non-LSE compare-and-swap can use an extending compare to
avoid requiring the input to have been previously extended.
- TImode compare-and-swap was missing entirely. This brings
aarch64 to parity with x86_64 wrt __sync_val_compare_and_swap.
There is a final patch that enables the new option by default.
I am not necessarily expecting this to be merged upstream, but
for the operating system to decide what the default should be.
It might be that this should be a configure option, so as to
make that OS choice easier, but I've just now thought of that. ;-)
I'm going to have to rely on Alex and/or Ramana to perform
testing on a system that supports LSE.
r~
Richard Henderson (11):
aarch64: Simplify LSE cas generation
aarch64: Improve cas generation
aarch64: Improve swp generation
aarch64: Improve atomic-op lse generation
aarch64: Emit LSE st<op> instructions
Add visibility to libfunc constructors
Link static libgcc after shared libgcc for -shared-libgcc
aarch64: Add out-of-line functions for LSE atomics
aarch64: Implement -matomic-ool
aarch64: Implement TImode compare-and-swap
Enable -matomic-ool by default
gcc/config/aarch64/aarch64-protos.h | 20 +-
gcc/optabs-libfuncs.h | 2 +
gcc/common/config/aarch64/aarch64-common.c | 6 +-
gcc/config/aarch64/aarch64.c | 480 ++++++--------
gcc/gcc.c | 9 +-
gcc/optabs-libfuncs.c | 26 +-
.../atomic-comp-swap-release-acquire.c | 2 +-
.../gcc.target/aarch64/atomic-inst-ldadd.c | 18 +-
.../gcc.target/aarch64/atomic-inst-ldlogic.c | 54 +-
.../gcc.target/aarch64/atomic-op-acq_rel.c | 2 +-
.../gcc.target/aarch64/atomic-op-acquire.c | 2 +-
.../gcc.target/aarch64/atomic-op-char.c | 2 +-
.../gcc.target/aarch64/atomic-op-consume.c | 2 +-
.../gcc.target/aarch64/atomic-op-imm.c | 2 +-
.../gcc.target/aarch64/atomic-op-int.c | 2 +-
.../gcc.target/aarch64/atomic-op-long.c | 2 +-
.../gcc.target/aarch64/atomic-op-relaxed.c | 2 +-
.../gcc.target/aarch64/atomic-op-release.c | 2 +-
.../gcc.target/aarch64/atomic-op-seq_cst.c | 2 +-
.../gcc.target/aarch64/atomic-op-short.c | 2 +-
.../aarch64/atomic_cmp_exchange_zero_reg_1.c | 2 +-
.../atomic_cmp_exchange_zero_strong_1.c | 2 +-
.../gcc.target/aarch64/sync-comp-swap.c | 2 +-
.../gcc.target/aarch64/sync-op-acquire.c | 2 +-
.../gcc.target/aarch64/sync-op-full.c | 2 +-
libgcc/config/aarch64/lse.c | 280 ++++++++
gcc/config/aarch64/aarch64.opt | 4 +
gcc/config/aarch64/atomics.md | 608 ++++++++++--------
gcc/config/aarch64/iterators.md | 8 +-
gcc/config/aarch64/predicates.md | 12 +
gcc/doc/invoke.texi | 14 +-
libgcc/config.host | 4 +
libgcc/config/aarch64/t-lse | 48 ++
33 files changed, 1050 insertions(+), 577 deletions(-)
create mode 100644 libgcc/config/aarch64/lse.c
create mode 100644 libgcc/config/aarch64/t-lse
--
2.17.1
- Follow-Ups:
- [PATCH, AArch64 01/11] aarch64: Simplify LSE cas generation
- [PATCH, AArch64 03/11] aarch64: Improve swp generation
- [PATCH, AArch64 02/11] aarch64: Improve cas generation
- [PATCH, AArch64 05/11] aarch64: Emit LSE st<op> instructions
- [PATCH, AArch64 06/11] Add visibility to libfunc constructors
- [PATCH, AArch64 07/11] Link static libgcc after shared libgcc for -shared-libgcc
- [PATCH, AArch64 04/11] aarch64: Improve atomic-op lse generation
- [PATCH, AArch64 08/11] aarch64: Add out-of-line functions for LSE atomics
- [PATCH, AArch64 11/11] Enable -matomic-ool by default
- [PATCH, AArch64 09/11] aarch64: Implement -matomic-ool
- [PATCH, AArch64 10/11] aarch64: Implement TImode compare-and-swap
- Re: [PATCH, AArch64 00/11] LSE atomics out-of-line
- Re: [PATCH, AArch64 00/11] LSE atomics out-of-line
- From: Ramana Radhakrishnan