Go Assembly: A Semi-Abstract Instruction Set

Why it exists

Go's compiler suite was designed without a traditional, separate assembler pass. Instead, the compiler generates code in a semi-abstract instruction set. Go assembly is the human-readable syntax for this intermediate form, which is then processed by the linker to create the final machine code.

The mental model

Think of Go assembly not as writing machine code, but as providing high-level hints to the Go compiler's final code generation stage. You write MOV, but the compiler decides the most efficient way to move that data, which might not be a literal MOV instruction on the CPU. It's an abstraction layer above the metal, based on the Plan 9 assembler style.

How it works

The Go assembler parses files written in its specific syntax. It understands abstract operations like MOV (move data) and TEXT (define a function), as well as machine-specific instructions. The key is that the final instruction selection happens later in the toolchain. This allows the compiler to inject runtime metadata, like the FUNCDATA and PCDATA directives used by the garbage collector, and make late-stage optimization decisions.

When to use it

Use Go assembly for two main purposes. First, to inspect the compiler's output with go tool compile -S to understand performance and see what your Go code becomes. Second, for hand-tuning performance-critical, low-level routines that require specific CPU features, as seen in standard library packages like runtime and math/big.

When not to use it

Avoid writing Go assembly for general application logic. It is not a direct representation of the underlying machine, making it a poor choice for precise hardware control. It's also non-portable and bypasses many of Go's safety guarantees. Reserve it for deep optimization after profiling proves it's necessary.

One canonical example

To see the compiler's assembly output for a simple program, run go tool compile -S my_program.go. The output will show abstract instructions like TEXT "".main(SB), $16-0 and MOVQ (TLS), CX, mixed with runtime directives like FUNCDATA that are meaningless to the CPU but vital for the Go runtime.