The other day Emilio noticed that we weren't setting CONFIG_ATOMIC128 for x86_64, despite forcing -mcx16 on the command-line. It seems that gcc has changed behaviour with version 7.
There's a rather long discussion about this: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80878 It does not appear that gcc will ever quite align with our needs here. I have added a second configure test, and have split the test to be used by the rest of QEMU into two compile-time constants. This lets us refactor the code a bit and remove some conditional compilation. In addition, no version of gcc supports __sync_compare_and_swap_16 for aarch64. Inline some asm for that case. I've also checked power8, which does support CONFIG_ATOMIC128. Changes since v1: * Fixed some errors in the lesser used paths (Emilio). * Dropped the target/s390x changes for now, as they conflict with other patches in-flight on list. * Raise EXCP_ATOMIC at translate time for AArch64, if needed. r~ Richard Henderson (5): tcg: Split CONFIG_ATOMIC128 target/i386: Convert to HAVE_CMPXCHG128 target/arm: Convert to HAVE_CMPXCHG128 target/arm: Check HAVE_CMPXCHG128 at transate time target/ppc: Convert to HAVE_CMPXCHG128 and HAVE_ATOMIC128 accel/tcg/atomic_template.h | 20 ++- include/qemu/atomic128.h | 155 ++++++++++++++++++++++ target/ppc/helper.h | 2 +- tcg/tcg.h | 16 ++- accel/tcg/cputlb.c | 3 +- accel/tcg/user-exec.c | 5 +- target/arm/helper-a64.c | 251 ++++++++++++++++++------------------ target/arm/translate-a64.c | 38 +++--- target/i386/mem_helper.c | 9 +- target/ppc/mem_helper.c | 33 ++++- target/ppc/translate.c | 115 +++++++++-------- configure | 19 +++ 12 files changed, 443 insertions(+), 223 deletions(-) create mode 100644 include/qemu/atomic128.h -- 2.17.1
