macro-fusion

Demo - Macro-op Fusion

Overview

Macro-op fusion is an optimisation where two consecutive instructions combine into one micro-instruction for execution within the CPU. This reduces resource usage in all stages of the pipeline after decoding, possibly increasing execution throughput.

Requirements

CPU:

• Intel x86-64 since Core
• AMD x86-64 since Bulldozer

Software:

• Linux
• GCC
• Make
• perf

Other microarchitectural assumptions:

• 2-8 scalar integer execution units/ports

Tutorial

This demonstration has two assembly programs, with-fusion.s and without-fusion.s. Both consist of a loop with more arithmetic instructions than the CPU can execute simultaneously. The only difference is that in with-fusion, macro-op fusion can occur.

To build the programs, run in this directory:

make

Then run and check the number of instructions, cycles, and micro-instructions issued:

perf stat -e instructions,cycles,uops_issued.any ./with-fusion
perf stat -e instructions,cycles,uops_issued.any ./without-fusion

If your CPU supports macro-op fusion, you should see that with-fusion completes with fewer micro-instructions issued and cycles, for the same number of instructions overall.

The reason is macro-op fusion. Certain pairs of consecutive instructions can macro-fuse; the most common being inc/dec/add/sub/test/cmp with jCC (CC being some conditional code). In with-fusion, the dec rax and jnz can be fused, whereas in without-fusion they are not consecutive and cannot fuse.

As a result, each loop iteration, with-fusion has one fewer issued (i.e scheduled for execution) micro-instruction over without-fusion, which is reflected by the uops_issued.any counter. When the bottleneck is parallel instruction execution, as in these programs, more cycles are required to execute the extra micro-instruction.