It's me again. The guy who stopped by a year and a half with this thread:
hello AND performance
and the accompanying patches.

Anyway, I dusted off my copy the dosbox code, updated from CVS, and started coding on it a little again. First results: benchmarks again!

Instructions are categorized into types by the following four datums:
1. The first word of the disassembly (ie the characters before whitespce). Usually this is the opcode, but sometimes this is a special prefix like REPNE or somesuch. This is represnted by itself in the type string.
2. Whether or not there was a [] pair in the dissassembly string. Usually this means a memory reference. This is represented in the type string as an "M", for memory.
3. Whether or not there was a comma in the dissassembly string. Only one comma is counted, even if multiple are present.
4. Whether the comma came before or after the [] pair.

Thus, a MOV from memory to a register becomes "mov ,M", whereas a MOV from a register to memory becomes "mov M,", and a MOV from a register to a register becomes "mov ,".

Be aware that the measurement error may be fairly high. I tried a few strategies to reduce error, but all reported identical results. These numbers do not vary from run to run on the same workload. However, they vary between different workloads even when there is no obvious reason for a large discrepancy - for instance, a register to register add took an average of 266 cycles in my benchmark (compiled with MinGW) versus 121 cycles in Daggerfall. All measurements were performed on an Athlon XP 2600+ running windows 2000.

The CPU core used was a modified version of the "normal" core.

Column 1 is the percentage of time used by that instruction type, column 2 is the number of times that instruction type was executed, column 3 is the average numvers of cycles per consumed per execution of that instruction type, and column 4 is the instruction type.

From my simple benchmarks (test1 & test2 combined)

1  15.0324%     8.1M     576  imul ,
2  13.0000%    15.4M     263  mov ,
3   9.2895%    10.9M     266  add ,
4   8.2312%     4.9M     526  mov ,M
5   6.8621%     8.1M     263  jns
6   6.6769%     8.1M     256  dec
7   4.8054%     2.7M     546  movzx ,
8   4.4453%     3.5M     396  mov M,
9   3.4794%     2.7M     395  rol ,
10   3.4747%     3.4M     314  cmp ,
11   3.1752%     3.1M     320  lea ,M
12   3.0938%     2.7M     350  jle
13   2.9471%     2.7M     335  xor ,M
14   2.6834%     2.7M     303  inc
15   2.4157%     2.7M     274  and ,
16   2.2155%     2.7M     250  xor ,
17   1.5693%     1.0M     470  cmp M,
18   0.9937%   710.7K     439  push
19   0.7088%   353.0K     630  call
20   0.6155%   693.0K     279  sar ,

From an actual run of the game "TES2: Daggerfall"

1  20.5124%   167.0M     196  mov ,M
2   7.9630%    68.7M     185  mov M,
3   6.2168%   300.1K   33082  repne
4   6.1500%    81.9M     119  mov ,
5   4.4693%    60.2M     118  pop
6   4.2932%    61.1M     112  push
7   3.9714%    52.0M     121  add ,
8   3.4423%     2.2M    2433  repe
9   2.7341%    23.0M     189  cmp M,
10   2.5600%    33.9M     120  je
11   2.2687%    26.2M     137  and ,
12   2.0686%    24.9M     132  jmp
13   1.9981%    18.7M     169  cmp ,M
14   1.8657%    20.9M     141  shr ,
15   1.7353%    21.0M     131  jne
16   1.4961%    17.0M     140  sub ,
17   1.4107%    16.2M     138  cmp ,
18   1.3851%    18.0M     122  test ,
19   1.3445%     8.6M     247  jle
20   1.1095%     9.0M     196  call