Examining the collected profile is the same for any address profiler, no matter if it is really a data profiler, a branch recorder, or something else. For the examples, we use the branch recorder from section 13.1.
Run a few million instructions or so with the branch profiler attached to the processor to get some data to look at, and then type:
simics> cpu0_branch_recorder.address-profile-data
View 0 of cpu0_branch_recorder: execution count
64-bit physical addresses, profiler granularity 4 bytes
Each cell covers 36 address bits (64 Gigabytes).
column offsets:
0x1000000000* 0x0 0x1 0x2 0x3 0x4 0x5 0x6 0x7
---------------------------------------------------------------------------
0x0000000000000000: 17000k . . . . . . .
0x0000008000000000: . . . . . . . .
0x0000010000000000: . . . . . . . .
0x0000018000000000: . . . . . . . .
0x0000020000000000: . . . . . . . .
0x0000028000000000: . . . . . . . .
0x0000030000000000: . . . . . . . .
0x0000038000000000: . . . . . . . .
0x0000040000000000: . . . . . . . .
0x0000048000000000: . . . . . . . .
0x0000050000000000: . . . . . . . .
0x0000058000000000: . . . . . . . .
0x0000060000000000: . . . . . . . .
0x0000068000000000: . . . . . . . .
0x0000070000000000: . . . . . . . .
0x0000078000000000: . . . . . . . 141
16999781 (17000k) counts shown. 0 not shown.
(View 0 is the default, so we did not have to specify it explicitly.) This gives us an overview of the address space. Since we did not specify what address interval we wanted to see, we got the smallest interval that contained all counts. By giving arguments to address-profile-data, we can zoom in on the interesting part where almost all the action is. A few orders of magnitude closer, it looks like this:
simics> cpu0_branch_recorder.address-profile-data address = 0x1f800000 table-bits = 19
View 0 of cpu0_branch_recorder: execution count
64-bit physical addresses, profiler granularity 4 bytes
Each cell covers 12 address bits (4 kilobytes).
column offsets:
0x1000* 0x0 0x1 0x2 0x3 0x4 0x5 0x6 0x7
---------------------------------------------------------------------------
0x000000001f800000: 31 760 . . . . 9472k 204633
0x000000001f808000: 14380 39728 420 1 2751 38226 119215 560668
0x000000001f810000: 375 301 5979k 1036 . . . .
0x000000001f818000: . . 390 . . . . .
0x000000001f820000: . . . . . . . .
0x000000001f828000: . . . . . . . .
0x000000001f830000: . . . . . . . .
0x000000001f838000: 16 . . . . . . 129960
0x000000001f840000: 16048 194576 12930 . . . . .
0x000000001f848000: . 1264 152 . . 18930 12338 174868
0x000000001f850000: 5736 . . . . . . 6
0x000000001f858000: 52 . . . . . . .
0x000000001f860000: . . . . . . . .
0x000000001f868000: . . . . . . . .
0x000000001f870000: . . . . . . . .
0x000000001f878000: . . . . . . . .
16999640 (17000k) counts shown. 141 not shown.
Here, we have zoomed in on the 19 bits of address space containing the address 0x1f800000. It is also possible to specify exactly what address interval you are interested in; see the Simics Reference Manual or the online help for address-profile-data.
If you want to see the current numbers every time you disassemble code, use the command aprof-views to select the views you want:
simics> cpu0.aprof-views add = cpu0_branch_recorder view = 0
simics> si 5
[cpu0] <v:0x00000000f00066b4> <p:0x000000001f8066b4> 1183160 sllx %o1, 63, %o1
[cpu0] <v:0x00000000f00066b8> <p:0x000000001f8066b8> 1183160 jmpl [%o7 + 8], %g0
[cpu0] <v:0x00000000f00066bc> <p:0x000000001f8066bc> 1183160 andn %o0, %o1, %o0
[cpu0] <v:0x00000000f0012358> <p:0x000000001f812358> 1181028 cmp %o0, %i0
[cpu0] <v:0x00000000f001235c> <p:0x000000001f81235c> 1181028 bcs,pt %xcc, 0xf0012350
Note that since the disassembly printed by commands such as c and si are the instruction just about to execute, the statistics reflect the way things were immediately before that instruction executed.
For most real-world profiling tasks, the above methods of looking at profile data are inadequate. Unfortunately, Simics does not currently offer a way to display large sets of profiling data in a user-friendly way; what it does offer is a few primitive operations (such as accessing individual counters and summing all counters in an interval) that users can call from their own scripts. See the Profiling API section in the Simics Reference Manual for details.