Changes since v1:
* Emit SWP/SWPB for pre-armv6, non-linux (i.e. no external __sync routines).
The assumption here is that we must be on some sort of embedded OS, and
further that the system is not SMP. I'm more or less assuming that all
of the embedded ARM multi-core are at least v6. If we don't do this, then
the C++ atomic_bool implementation will fall back to truely non-atomic
store and load sequences. I can't believe that SWP is worse than that.
* Patch 3 converts the compare-and-swap routines to generate the success
boolean inside the flags register. This is similar to an optimization
that we made for powerpc recently. I believe that this does in fact
generate better code inside libgomp now.
* Patch 4 marks some more instructions as predicable. While examining
the output of the previous patch, we encounter things like
while (!__atomic_compare_exchange_n (sem, &count,
(count + SEM_INC) & ~SEM_WAIT, true,
MEMMODEL_RELEASE, MEMMODEL_RELAXED))
continue;
which can be optmized (in part) to
b8: f57ff05f dmb sy
bc: e1901f9f ldrex r1, [r0]
c0: e151000c cmp r1, ip
c4: 01804f92 strexeq r4, r2, [r0]
c8: 03340001 teqeq r4, #1
cc: e5831000 str r1, [r3]
d0: 1a000004 bne e8 <gomp_ordered_first+0xe8>
but only if strex and teq are both predicable. Otherwise we get a
branch after c0 to the store before the branch at cc.
While exploring how predication actually happens in the arm backend,
I went through and marked the CMP, CMN, and TST insns as predicable,
since ifcvt.c does a better job than arm_final_prescan_insn here.
* Patch 5 has been updated to deal with hwcap-y issues, as well as to
fix instruction errors under -mthumb.
It's been under testing for 36 hours now on a armv7 hw; I was really
hoping that it would be completed before eob today but no such luck.
I decided to send the patch series for review anyway.
r~
Richard Henderson (5):
arm: Convert to atomic optabs.
arm: Emit swp for pre-armv6.
arm: Use CC_REGNUM as success output from compare-and-swap.
arm: Set predicable on more instructions.
arm-linux: Add libitm support.
gcc/config/arm/arm-protos.h | 7 +-
gcc/config/arm/arm.c | 818 +++++++++++++---------------------
gcc/config/arm/arm.h | 24 +-
gcc/config/arm/arm.md | 66 ++--
gcc/config/arm/constraints.md | 5 +
gcc/config/arm/predicates.md | 4 +
gcc/config/arm/sync.md | 727 ++++++++++++++-----------------
libitm/Makefile.am | 3 +
libitm/Makefile.in | 20 +-
libitm/config/arm/hwcap.cc | 67 +++
libitm/config/arm/hwcap.h | 41 ++
libitm/config/arm/sjlj.S | 135 ++++++
libitm/config/arm/target.h | 62 +++
libitm/config/generic/asmcfi.h | 13 +-
libitm/config/linux/arm/futex_bits.h | 48 ++
libitm/configure | 18 +-
libitm/configure.ac | 1 +
libitm/configure.tgt | 2 +
18 files changed, 1077 insertions(+), 984 deletions(-)
create mode 100644 libitm/config/arm/hwcap.cc
create mode 100644 libitm/config/arm/hwcap.h
create mode 100644 libitm/config/arm/sjlj.S
create mode 100644 libitm/config/arm/target.h
create mode 100644 libitm/config/linux/arm/futex_bits.h
--
1.7.6.4
