cachegrind/cg_branchpred.c

/*--------------------------------------------------------------------*/
/*--- Branch predictor simulation                  cg_branchpred.c ---*/
/*--------------------------------------------------------------------*/

/*
   This file is part of Cachegrind, a Valgrind tool for cache
   profiling programs.

   Copyright (C) 2002-2013 Nicholas Nethercote
      njn@valgrind.org

   This program is free software; you can redistribute it and/or
   modify it under the terms of the GNU General Public License as
   published by the Free Software Foundation; either version 2 of the
   License, or (at your option) any later version.

   This program is distributed in the hope that it will be useful, but
   WITHOUT ANY WARRANTY; without even the implied warranty of
   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
   General Public License for more details.

   You should have received a copy of the GNU General Public License
   along with this program; if not, write to the Free Software
   Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
   02111-1307, USA.

   The GNU General Public License is contained in the file COPYING.
*/


/* This file contains the actual branch predictor simulator and its
   associated state.  As with cg_sim.c it is #included directly into
   cg_main.c.  It provides:

   - a taken/not-taken predictor for conditional branches
   - a branch target address predictor for indirect branches

   Function return-address prediction is not modelled, on the basis
   that return stack predictors almost always predict correctly, and
   also that it is difficult for Valgrind to robustly identify
   function calls and returns.
*/

/* How many bits at the bottom of an instruction address are
   guaranteed to be zero? */
#if defined(VGA_ppc32) || defined(VGA_ppc64) \
    || defined(VGA_mips32) || defined(VGA_mips64) || defined(VGA_arm64)
#  define N_IADDR_LO_ZERO_BITS 2
#elif defined(VGA_x86) || defined(VGA_amd64)
#  define N_IADDR_LO_ZERO_BITS 0
#elif defined(VGA_s390x) || defined(VGA_arm)
#  define N_IADDR_LO_ZERO_BITS 1
#else
#  error "Unsupported architecture"
#endif


/* Get a taken/not-taken prediction for the instruction (presumably a
   conditional branch) at instr_addr.  Once that's done, update the
   predictor state based on whether or not it was actually taken, as
   indicated by 'taken'.  Finally, return 1 for a mispredict and 0 for
   a successful predict.

   The predictor is an array of 16k (== 2^14) 2-bit saturating
   counters.  Given the address of the branch instruction, the array
   index to use is computed both from the low order bits of the branch
   instruction's address, and the global history - that is, from the
   taken/not-taken behaviour of the most recent few branches.  This
   makes the predictor able to correlate this branch's behaviour with
   that of other branches. 

   TODO: use predictor written by someone who understands this stuff.
   Perhaps it would be better to move to a standard GShare predictor
   and/or tournament predictor.
*/
/* The index is composed of N_HIST bits at the top and N_IADD bits at
   the bottom.  These numbers chosen somewhat arbitrarily, but note
   that making N_IADD_BITS too small (eg 4) can cause large amounts of
   aliasing, and hence misprediction, particularly if the history bits
   are mostly unchanging. */
#define N_HIST_BITS 7
#define N_IADD_BITS 7

#define N_BITS     (N_HIST_BITS + N_IADD_BITS)
#define N_COUNTERS (1 << N_BITS)

static UWord shift_register = 0;   /* Contains global history */
static UChar counters[N_COUNTERS]; /* Counter array; presumably auto-zeroed */


static ULong do_cond_branch_predict ( Addr instr_addr, Word takenW )
{
   UWord indx;
   Bool  predicted_taken, actually_taken, mispredict;

   const UWord hist_mask = (1 << N_HIST_BITS) - 1;
   const UWord iadd_mask = (1 << N_IADD_BITS) - 1;
         UWord hist_bits = shift_register & hist_mask;
         UWord iadd_bits = (instr_addr >> N_IADDR_LO_ZERO_BITS)
                           & iadd_mask;

   tl_assert(hist_bits <= hist_mask);
   tl_assert(iadd_bits <= iadd_mask);
   indx = (hist_bits << N_IADD_BITS) | iadd_bits;
   tl_assert(indx < N_COUNTERS);
   if (0) VG_(printf)("index = %d\n", (Int)indx);

   tl_assert(takenW <= 1);
   predicted_taken = counters[ indx ] >= 2;
   actually_taken  = takenW > 0;

   mispredict = (actually_taken && (!predicted_taken))
                || ((!actually_taken) && predicted_taken);

   shift_register <<= 1;
   shift_register |= (actually_taken ? 1 : 0);

   if (actually_taken) {
      if (counters[indx] < 3)
         counters[indx]++;
   } else {
      if (counters[indx] > 0)
         counters[indx]--;
   }

   tl_assert(counters[indx] <= 3);

   return mispredict ? 1 : 0;
}


/* A very simple indirect branch predictor.  Use the branch's address
   to index a table which records the previous target address for this
   branch (or whatever aliased with it) and use that as the
   prediction. */
#define N_BTAC_BITS 9
#define N_BTAC      (1 << N_BTAC_BITS)
static Addr btac[N_BTAC]; /* BTAC; presumably auto-zeroed */

static ULong do_ind_branch_predict ( Addr instr_addr, Addr actual )
{
   Bool mispredict;
   const UWord mask = (1 << N_BTAC_BITS) - 1;
         UWord indx = (instr_addr >> N_IADDR_LO_ZERO_BITS) 
                      & mask;
   tl_assert(indx < N_BTAC);
   mispredict = btac[indx] != actual;
   btac[indx] = actual;
   return mispredict ? 1 : 0;
}


/*--------------------------------------------------------------------*/
/*--- end                                          cg_branchpred.c ---*/
/*--------------------------------------------------------------------*/