What is Profile-Guided Optimization (PGO) in Go, and how do you use it?

PGO (GA in Go 1.21) feeds a CPU profile from production into the compiler so it optimizes hot paths more aggressively. The Go team reports typical gains of 2–14% CPU.

Workflow:

# 1. collect a representative CPU pprof from production
curl -o cpu.pprof 'http://127.0.0.1:6060/debug/pprof/profile?seconds=30'
# 2. place it in the main package directory
mv cpu.pprof ./cmd/server/default.pgo
# 3. build normally — -pgo=auto is the default and picks up default.pgo
go build ./cmd/server
# explicit / merged profiles:
go tool pprof -proto a.pprof b.pprof > merged.pprof
go build -pgo=merged.pprof ./cmd/server

What it does today:

• Hot-call inlining: call sites that are hot in the profile get a much larger inlining budget than the normal ~80 nodes.
• Devirtualization: a hot interface call whose dynamic type is usually *T gets a guarded direct call (if x, ok := i.(*T); ok { x.M() } else { i.M() }), which can then be inlined.
• Layout/register improvements continue to be added.

Gotchas: commit the profile to the repo so builds are reproducible; stale profiles from older code still work (matching is by function name + line offset) but degrade gracefully; build time and binary size grow modestly; it's iterative-stable (building with PGO then re-profiling doesn't oscillate). Profile representative traffic — a profile from a load test that exercises the wrong path helps the wrong code.