Re: [dpdk-dev] [PATCH] Minor fixes in rte_common.h file.

[Ravi Kerur <rkerur@gmail.com>]

On Tue, Dec 16, 2014 at 9:23 AM, Ananyev, Konstantin <
konstantin.ananyev@intel.com> wrote:
>
>
>
> > -----Original Message-----
> > From: dev [mailto:dev-bounces@dpdk.org] On Behalf Of Ravi Kerur
> > Sent: Tuesday, December 16, 2014 4:47 PM
> > To: Neil Horman
> > Cc: dev@dpdk.org
> > Subject: Re: [dpdk-dev] [PATCH] Minor fixes in rte_common.h file.
> >
> > On Sat, Dec 13, 2014 at 2:39 AM, Neil Horman <nhorman@tuxdriver.com>
> wrote:
> > >
> > > On Fri, Dec 12, 2014 at 03:04:34PM -0800, r k wrote:
> > > > Subject: [PATCH] Minor fixes in rte_common.h file.
> > > >
> > > > Fix rte_is_power_of_2 since 0 is not.
> > > > Avoid branching instructions in RTE_MAX and RTE_MIN.
> > > >
> > > > Signed-off-by: Ravi Kerur <rkerur@gmail.com>
> > > > ---
> > > >  lib/librte_eal/common/include/rte_common.h | 6 +++---
> > > >  lib/librte_pmd_e1000/igb_pf.c              | 4 ++--
> > > >  lib/librte_pmd_ixgbe/ixgbe_pf.c            | 4 ++--
> > > >  3 files changed, 7 insertions(+), 7 deletions(-)
> > > >
> > > > diff --git a/lib/librte_eal/common/include/rte_common.h
> > > > b/lib/librte_eal/common/include/rte_common.h
> > > > index 921b91f..e163f35 100644
> > > > --- a/lib/librte_eal/common/include/rte_common.h
> > > > +++ b/lib/librte_eal/common/include/rte_common.h
> > > > @@ -203,7 +203,7 @@ extern int RTE_BUILD_BUG_ON_detected_error;
> static
> > > > inline int  rte_is_power_of_2(uint32_t n)  {
> > > > -       return ((n-1) & n) == 0;
> > > > +       return n && !(n & (n - 1));
> > > >  }
> > > >
> > > >  /**
> > > > @@ -259,7 +259,7 @@ rte_align64pow2(uint64_t v)  #define RTE_MIN(a,
> b)
> > > ({ \
> > > >                 typeof (a) _a = (a); \
> > > >                 typeof (b) _b = (b); \
> > > > -               _a < _b ? _a : _b; \
> > > > +                _b ^ ((_a ^ _b) & -(_a < _b)); \
> > > Are you sure this is actually faster than the branch version?  What
> about
> > > using
> > > a cmov instead?
> > >
> > >
> > <rk> i am pretty sure modified code is faster than branching. I remember
> > cmov had performance issues esp. on Pentuim-4 not sure how new intel
> cpu's
> > perform.
>
> I also think most modern compilers are smart enough to avoid any branching
> here and will use cmov instead.
> And we are way ahead of Pentium 4 times these days.

Konstantin
>

<rk>Konstantin,  Can you please elaborate, is it something done
automatically with Intel's icc compiler? My understanding is branch
prediction can be influenced only by using compiler builtin i.e.
__builtin_expect() , without this compiler will generate regular
instructions(cmp/jump instructions). I wrote small program and compiled
with gcc -02/-03, don't see cmov instruction.


> >
> > >         })
> > > >
> > > >  /**
> > > > @@ -268,7 +268,7 @@ rte_align64pow2(uint64_t v)  #define RTE_MAX(a,
> b)
> > > ({ \
> > > >                 typeof (a) _a = (a); \
> > > >                 typeof (b) _b = (b); \
> > > > -               _a > _b ? _a : _b; \
> > > > +               _a ^ ((_a ^ _b) & -(_a < _b)); \
> > > Same as above
> > >
> > > <rk> Same as above.
> >
> > > >         })
> > > >
> > > >  /*********** Other general functions / macros ********/ diff --git
> > > > a/lib/librte_pmd_e1000/igb_pf.c b/lib/librte_pmd_e1000/igb_pf.c index
> > > > bc3816a..546499c 100644
> > > > --- a/lib/librte_pmd_e1000/igb_pf.c
> > > > +++ b/lib/librte_pmd_e1000/igb_pf.c
> > > > @@ -321,11 +321,11 @@ igb_vf_set_mac_addr(struct rte_eth_dev *dev,
> > > uint32_t
> > > > vf, uint32_t *msgbuf)  static int  igb_vf_set_multicast(struct
> > > rte_eth_dev
> > > > *dev, __rte_unused uint32_t vf, uint32_t *msgbuf)  {
> > > > -       int i;
> > > > +       int16_t i;
> > > >         uint32_t vector_bit;
> > > >         uint32_t vector_reg;
> > > >         uint32_t mta_reg;
> > > > -       int entries = (msgbuf[0] & E1000_VT_MSGINFO_MASK) >>
> > > > +       int32_t entries = (msgbuf[0] & E1000_VT_MSGINFO_MASK) >>
> > > >                 E1000_VT_MSGINFO_SHIFT;
> > > NAK, this has nothing to do with the included changelog
> > >
> >
> > <rk> It does, it causes compilation errors such as
> >
> > /root/dpdk-new/dpdk/lib/librte_pmd_e1000/igb_pf.c: In function
> > \u2018igb_pf_mbx_process\u2019:
> > /root/dpdk-new/dpdk/lib/librte_pmd_e1000/igb_pf.c:350:23: error: array
> > subscript is above array bounds [-Werror=array-bounds]
> >    vfinfo->vf_mc_hashes[i] = hash_list[i];
> >                        ^
> > cc1: all warnings being treated as errors
> >
> > Also it is always better to use explicit int definitions esp. for 64bit
> > systems.
> >
> >
> >
> > >
> > > >         uint16_t *hash_list = (uint16_t *)&msgbuf[1];
> > > >         struct e1000_hw *hw =
> > > > E1000_DEV_PRIVATE_TO_HW(dev->data->dev_private);
> > > > diff --git a/lib/librte_pmd_ixgbe/ixgbe_pf.c
> > > > b/lib/librte_pmd_ixgbe/ixgbe_pf.c index 51da1fd..426caf9 100644
> > > > --- a/lib/librte_pmd_ixgbe/ixgbe_pf.c
> > > > +++ b/lib/librte_pmd_ixgbe/ixgbe_pf.c
> > > > @@ -390,7 +390,7 @@ ixgbe_vf_set_multicast(struct rte_eth_dev *dev,
> > > > __rte_unused uint32_t vf, uint32
> > > >         struct ixgbe_hw *hw =
> > > > IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
> > > >         struct ixgbe_vf_info *vfinfo =
> > > >
>  *(IXGBE_DEV_PRIVATE_TO_P_VFDATA(dev->data->dev_private));
> > > > -       int nb_entries = (msgbuf[0] & IXGBE_VT_MSGINFO_MASK) >>
> > > > +       int32_t nb_entries = (msgbuf[0] & IXGBE_VT_MSGINFO_MASK) >>
> > > >                 IXGBE_VT_MSGINFO_SHIFT;
> > > ditto
> > > >         uint16_t *hash_list = (uint16_t *)&msgbuf[1];
> > > >         uint32_t mta_idx;
> > > > @@ -399,7 +399,7 @@ ixgbe_vf_set_multicast(struct rte_eth_dev *dev,
> > > > __rte_unused uint32_t vf, uint32
> > > >         const uint32_t IXGBE_MTA_BIT_SHIFT = 5;
> > > >         const uint32_t IXGBE_MTA_BIT_MASK = (0x1 <<
> IXGBE_MTA_BIT_SHIFT)
> > > -
> > > > 1;
> > > >         uint32_t reg_val;
> > > > -       int i;
> > > > +       int16_t i;
> > > ditto
> > >
> > > <rk> Same as above.
> >
> > > >
> > > >         /* only so many hash values supported */
> > > >         nb_entries = RTE_MIN(nb_entries, IXGBE_MAX_VF_MC_ENTRIES);
> > > > --
> > > > 1.9.1
> > > >
> > >
>