Memory disambiguation is a technique to accelerate the execution of memory-related instructions.
All Intel CPUs since Pentium Pro have an out-of-order engine, which allows the CPU to execute non-dependant instructions in any order. What happens is that memory-related instructions are traditionally executed in the same order they appear on the program, otherwise data inconsistency could appear. For example, if the original program has an instruction like ”store 10 at address 5555“ and then a ”load data stored at 5555“, they cannot be reversed (i.e., executed out of order) or the second instruction would get wrong data, as the data of address 5555 was changed by the first instruction.
What the memory disambiguation engine does is locate and execute memory-related instructions that can be executed out of order, accelerating the execution of the program (we will explain how this is accomplished in a minute).
In Figure 3 you have an example of a CPU without memory disambiguation (i.e., all CPUs not based on Core microarchitecture). As you can see, the CPU has to execute the instructions as they appear on the original program. For example, ”Load4“ isn’t related to any other memory-related instruction and could be executed first, however it has to wait all other instructions to be executed first.
click to enlarge
Figure 3: CPU without memory disambiguation.
In Figure 4 you see how the program shown in Figure 3 works on a CPU based on Core microarchitecture. It ”knows“ that ”Load4“ isn’t related to the other instructions and can be executed first.
click to enlarge
Figure 4: CPU with memory disambiguation.
This improves the CPU performance because now that ”Load4“ is executed, the CPU has the data required for executing other instructions that need the value of ”X“ to be executed.
On a regular CPU, if after this ”Load4“ we had an ”Add 50“, this ”Add 50“ (and all other instructions that depend on that result) would have to wait all other instructions shown in Figure 3 to be executed. With memory disambiguation, these instructions can be executed early, since the CPU will now have the value of ”X“ early.
