On Sun, Feb 2, 2020 at 9:44 PM Anup Patel <anup.pa...@wdc.com> wrote: > > Currently, TIME CSRs are emulated only for user-only mode. This > patch add TIME CSRs emulation for privileged mode. > > For privileged mode, the TIME CSRs will return value provided > by rdtime callback which is registered by QEMU machine/platform > emulation (i.e. CLINT emulation). If rdtime callback is not > available then the monitor (i.e. OpenSBI) will trap-n-emulate > TIME CSRs in software. > > We see 25+% performance improvement in hackbench numbers when > TIME CSRs are not trap-n-emulated. > > Signed-off-by: Anup Patel <anup.pa...@wdc.com> > Reviewed-by: Alistair Francis <alistair.fran...@wdc.com> > --- > target/riscv/cpu.h | 5 +++ > target/riscv/cpu_helper.c | 5 +++ > target/riscv/csr.c | 86 +++++++++++++++++++++++++++++++++++++-- > 3 files changed, 92 insertions(+), 4 deletions(-) >
Jonathan has a patch before: http://patchwork.ozlabs.org/patch/1106480/ The idea in his patch does similar thing according to mcounteren.TM. But in this patch we seem to control the TIME CSR purely by software. Is this behavior spec complaint? Regards, Bin
