Post

Anyone use Intel's 386 processor from 1985, the first 32-bit x86? To improve performance, it preloaded instructions into a 16-byte prefetch queue (highlighted in red). Why does such a small queue take up so much of the chip? Let's look under the microscope at its complicated circuitry... 1/N

How hard can it be to add 1 to a number? The problem is that incrementing a 30-bit number can require carry the 1, carry the 1, repeated 30 times: slow. In the 386's prefetcher, "Manchester carry chain" and "carry skip" circuits make this fast, implemented with complicated "dynamic logic".

Another part of the prefetcher converts an 8-bit value to a 16- or 32-bit value. This is tricky if the value is negative, since the upper bits must be filled with 1's. Here's a closeup of the sign extension circuitry that fills upper bits with the correct value.

This close-up of the 386 processor's die shows the different parts of the prefetch unit. There's a lot going on just to hold 16 bytes of instructions.