[DYNAREC][ENV] Disabled SEP by default

[ksco]

Now that we have callret enabled by default, I noticed SEP causes the branch misprediction by an order of magnitude in coremark, while the memory-saving effect is almost imperceptible.

ksco@steel [ coremark@main ] $ BOX64_LOG=0 BOX64_NOBANNER=1 BOX64_DYNACACHE=0 BOX64_DYNAREC_SEP=0 perf stat -e task-clock,cycles,instructions,branches,branch-misses -- taskset -c 2 box64 ./coremark.exe_x64 0 0 0x66 200000 7 1 >/dev/null

 Performance counter stats for 'taskset -c 2 box64 ./coremark.exe_x64 0 0 0x66 200000 7 1':

         16,339.65 msec task-clock:u
    35,828,022,082      cycles:u
   118,528,644,489      instructions:u
    12,994,274,457      branches:u
        36,068,961      branch-misses:u

      16.348269274 seconds time elapsed

      16.331188000 seconds user
       0.009995000 seconds sys

ksco@steel [ coremark@main ] $ BOX64_LOG=0 BOX64_NOBANNER=1 BOX64_DYNACACHE=0 BOX64_DYNAREC_SEP=1 perf stat -e task-clock,cycles,instructions,branches,branch-misses -- taskset -c 2 box64 ./coremark.exe_x64 0 0 0x66 200000 7 1 >/dev/null

 Performance counter stats for 'taskset -c 2 box64 ./coremark.exe_x64 0 0 0x66 200000 7 1':

         18,571.57 msec task-clock:u
    40,721,222,650      cycles:u
   118,523,943,716      instructions:u
    12,993,208,645      branches:u
       383,363,320      branch-misses:u

      18.581470637 seconds time elapsed

      18.564061000 seconds user
       0.008996000 seconds sys

ksco@steel [ coremark@main ] $ perf stat -e task-clock,cycles,instructions,branches,branch-misses -- taskset -c 2 ./coremark.exe 0 0 0x66 200000 7 1 >/dev/null

 Performance counter stats for 'taskset -c 2 ./coremark.exe 0 0 0x66 200000 7 1':

          9,757.79 msec task-clock:u
    21,407,080,976      cycles:u
    62,785,977,365      instructions:u
    11,852,476,157      branches:u
        18,851,352      branch-misses:u

       9.762550998 seconds time elapsed

       9.757495000 seconds user
       0.000999000 seconds sys

[ptitSeb]

Are you sure about the memory saving being minimmum? (how about it's enable by default on SAVE_MEM configuration?)

[ksco]

Are you sure about the memory saving being minimmum? (how about it's enable by default on SAVE_MEM configuration?)

Do you have any workload where SEP reduces memory consumption noticeably? Because I can't find one.

[ksco]

Anyway, I think "enable by default on SAVE_MEM" is a good idea indeed.

[ksco]

Done