Goffi

#goffi — Zero-CGO FFI for Go

Pure Go Foreign Function Interface for calling C libraries without CGO. Designed for WebGPU and GPU computing — zero C dependencies, zero per-call allocations, 88–114 ns overhead.

Deep dive: How We Call C Libraries Without a C Compiler — architecture, assembly, callbacks, and ecosystem.

// Load library, prepare once, call many times — no CGO required
handle, _ := ffi.LoadLibrary("wgpu_native.dll")
sym, _ := ffi.GetSymbol(handle, "wgpuCreateInstance")

cif := &types.CallInterface{}
ffi.PrepareCallInterface(cif, types.DefaultCall, returnType, argTypes)
_, _ = ffi.CallFunction(cif, sym, unsafe.Pointer(&result), args)

#Features

#Quick Start

Android arm64/API 29+ is a preview candidate in both CGO modes. Cross-build, ABI, and ELF probes pass, but physical-device startup proof is still pending; see docs/ANDROID.md for the runtime ABI, NDK probe, and the intentional callback limitation.

#Installation

go get github.com/go-webgpu/goffi

#Requirements

goffi builds under both CGO_ENABLED=0 and CGO_ENABLED=1. CGO_ENABLED=0 needs no C compiler and is the recommended mode for the gogpu ecosystem and most users. The Go toolchain enables cgo by default when it finds a C compiler (check go env CGO_ENABLED), so set CGO_ENABLED=0 explicitly if you want that mode.

# CGO_ENABLED=0 (recommended, no C compiler needed)
CGO_ENABLED=0 go build ./...

# CGO_ENABLED=1 (needs a C compiler; e.g. for go test -race, which requires cgo on Linux)
CGO_ENABLED=1 go build ./...

# FreeBSD with CGO_ENABLED=0: fakecgo must export environ/__progname for libc
CGO_ENABLED=0 go build -gcflags=github.com/go-webgpu/goffi/internal/fakecgo=-std ./...

Known issue — Linux and FreeBSD, CGO_ENABLED=1: a binary that contains goffi and other cgo code (any package with import "C", including purego), or that is linked with -linkmode=external, -buildmode=c-shared or c-archive, currently fails to link with unhandled relocation … (SDYNIMPORT …) (#72; the fix is discussed in #104). macOS and Windows are not affected. Workarounds are listed under Known Limitations.

How? goffi uses Go's cgo_import_dynamic for dynamic library loading. Under CGO_ENABLED=0 the cgo runtime is supplied by internal/fakecgo; under CGO_ENABLED=1 the standard runtime/cgo is linked in. Both modes share the same FFI fast path and ABIs.

#Linking modes (Linux)

CGO_ENABLED=0 does not imply a fully static ELF when goffi is imported. //go:cgo_import_dynamic for dlopen / libc still records PT_INTERP and DT_NEEDED (libdl.so.2, libc.so.6, libpthread.so.0). This matches purego and is required for host dlopen — the kernel only maps ld.so when PT_INTERP is present. See goffi#74 and gogpu#474.

# Fully static Linux amd64/arm64 binary (FFI unavailable)
CGO_ENABLED=0 GOOS=linux GOARCH=amd64 go build -tags goffi_static -o app .
file app   # statically linked
# Verify: no INTERP / NEEDED
scripts/check-elf-linking.sh --static ./app

Under -tags goffi_static, errno capture is unavailable (always returns 0): ErrnoFnAddr() is a no-op so the assembly trampoline skips __errno_location / __error, which need dynamic libc.

FROM scratch + Vulkan/Wayland/libX11 via host dlopen is not possible without either ld.so or a userspace ELF loader (see docs/ADR-001-userspace-elf-loader.md). Windows is unaffected (LoadLibraryW via ntdll).

#Example: Calling strlen

package main

import (
	"fmt"
	"runtime"
	"unsafe"

	"github.com/go-webgpu/goffi/ffi"
	"github.com/go-webgpu/goffi/types"
)

func main() {
	// Load platform-specific C library
	libName := "libc.so.6"
	if runtime.GOOS == "windows" {
		libName = "msvcrt.dll"
	}

	handle, err := ffi.LoadLibrary(libName)
	if err != nil {
		panic(err)
	}
	defer ffi.FreeLibrary(handle)

	strlen, err := ffi.GetSymbol(handle, "strlen")
	if err != nil {
		panic(err)
	}

	// Prepare call interface once — reuse for all subsequent calls
	cif := &types.CallInterface{}
	err = ffi.PrepareCallInterface(
		cif,
		types.DefaultCall,                                     // auto-detects platform ABI
		types.UInt64TypeDescriptor,                            // return: size_t
		[]*types.TypeDescriptor{types.PointerTypeDescriptor},  // arg: const char*
	)
	if err != nil {
		panic(err)
	}

	// Call strlen — avalue elements are pointers TO argument values.
	// Keep pointer arguments in unsafe.Pointer variables, never uintptr —
	// in your binding wrappers' parameters too (see SECURITY.md).
	testStr := "Hello, goffi!\x00"
	strPtr := unsafe.Pointer(unsafe.StringData(testStr))
	var length uint64

	_, err = ffi.CallFunction(cif, strlen, unsafe.Pointer(&length), []unsafe.Pointer{unsafe.Pointer(&strPtr)})
	if err != nil {
		panic(err)
	}

	fmt.Printf("strlen(%q) = %d\n", testStr[:len(testStr)-1], length)
	// Output: strlen("Hello, goffi!") = 13
}

#Example: Calling a Variadic C Function

Use PrepareVariadicCallInterface instead of PrepareCallInterface for C variadic functions (printf, sprintf, custom variadic APIs). Specify nfixedargs — the count of fixed parameters before ... in the C prototype. goffi automatically applies Apple's ARM64 stack-force rule on darwin/arm64 so the same Go code works correctly on all platforms.

// C prototype: int64_t sum_variadic(int64_t count, ...)
// Call as: sum_variadic(3, 10, 20, 30) → 60

var cif types.CallInterface
err := ffi.PrepareVariadicCallInterface(
    &cif,
    types.DefaultCall,
    1, // nfixedargs: only 'count' is fixed; 10/20/30 are variadic
    types.SInt64TypeDescriptor,
    []*types.TypeDescriptor{
        types.SInt64TypeDescriptor, // count (fixed)
        types.SInt64TypeDescriptor, // arg1 (variadic)
        types.SInt64TypeDescriptor, // arg2 (variadic)
        types.SInt64TypeDescriptor, // arg3 (variadic)
    },
)

count := int64(3)
a1, a2, a3 := int64(10), int64(20), int64(30)
var result int64
_, _ = ffi.CallFunction(&cif, sym, unsafe.Pointer(&result), []unsafe.Pointer{
    unsafe.Pointer(&count),
    unsafe.Pointer(&a1),
    unsafe.Pointer(&a2),
    unsafe.Pointer(&a3),
})
// result == 60

A new CIF must be prepared for each unique combination of variadic argument types. The fixed-arg portion of the CIF can be reused by re-calling PrepareVariadicCallInterface with different variadic arg type slices.

#Example: Passing and Returning Structs

goffi handles C struct pass-by-value across all ABI size classes. Describe the struct layout with a TypeDescriptor, then pass struct values directly via unsafe.Pointer(&s).

// C struct: typedef struct { int64_t x; int64_t y; } Point;
pointType := &types.TypeDescriptor{
    Kind:      types.StructType,
    Size:      16,          // must match C sizeof(Point)
    Alignment: 8,           // must match C alignof(Point)
    Members: []*types.TypeDescriptor{
        types.SInt64TypeDescriptor, // x
        types.SInt64TypeDescriptor, // y
    },
}

var cif types.CallInterface
ffi.PrepareCallInterface(&cif, types.DefaultCall,
    pointType,                                            // return: Point
    []*types.TypeDescriptor{types.SInt64TypeDescriptor, types.SInt64TypeDescriptor},
)

x, y := int64(3), int64(4)
var result Point
_, _ = ffi.CallFunction(&cif, makePointFn,
    unsafe.Pointer(&result),                              // buffer for struct return value
    []unsafe.Pointer{unsafe.Pointer(&x), unsafe.Pointer(&y)},
)
// result.X == 3, result.Y == 4

// Pass struct as argument:
var distCif types.CallInterface
ffi.PrepareCallInterface(&distCif, types.DefaultCall,
    types.SInt64TypeDescriptor,
    []*types.TypeDescriptor{pointType, pointType},        // two Point args
)

a := Point{X: 0, Y: 0}
b := Point{X: 3, Y: 4}
var dist int64
_, _ = ffi.CallFunction(&distCif, distFn,
    unsafe.Pointer(&dist),
    []unsafe.Pointer{unsafe.Pointer(&a), unsafe.Pointer(&b)}, // pointer to struct data
)
// dist == 25

Structs >16 bytes (on Windows AMD64: every struct whose size is not 1, 2, 4 or 8 bytes) are returned via a hidden pointer (sret) — goffi handles this transparently. Windows AMD64 passes struct arguments of other sizes as a pointer to a copy, so the callee never modifies your value.

See examples/struct/ for a complete working example with compile-and-run.

#avalue Convention: Pointer-to-Value

Every avalue[i] must be a pointer TO the argument value. GoFFI dereferences it to read the actual value passed to C. This follows the libffi convention.

For most cases this is straightforward — unsafe.Pointer(&x) for scalars, unsafe.Pointer(&buf) for pointers. The non-obvious case is out-pointer parameters (where C writes a result through a pointer):

// C: int get_device(uint32_t index, void **device)
// The function writes *device — an out-pointer parameter.

// WRONG — C receives NULL (GoFFI reads the VALUE of device, which is nil):
var device unsafe.Pointer
_, _ = ffi.CallFunction(&cif, fn, unsafe.Pointer(&ret),
    []unsafe.Pointer{unsafe.Pointer(&index), unsafe.Pointer(&device)})

// CORRECT — C receives &device (GoFFI reads the VALUE of devicePtr, which is &device):
var device unsafe.Pointer
devicePtr := unsafe.Pointer(&device)
_, _ = ffi.CallFunction(&cif, fn, unsafe.Pointer(&ret),
    []unsafe.Pointer{unsafe.Pointer(&index), unsafe.Pointer(&devicePtr)})
// After call: device holds the handle written by C

Rule of thumb: if the C parameter is T *out (C writes through it), you need an intermediate outPtr := unsafe.Pointer(&out) and pass unsafe.Pointer(&outPtr) in avalue.

#Performance

FFI overhead: 88–114 ns/op (Windows AMD64, Intel i7-1255U)

syscallArgs is heap-allocated via sync.Pool for callback safety (goroutine stack may move during C→Go callbacks). Pool reuse gives 0 allocs/op in steady state.

The avalue slice escapes to the heap (its pointers are handed to C), so a slice literal built on every call — []unsafe.Pointer{unsafe.Pointer(&x)} — costs one allocation per call (8 B per argument). Build the slice once and reuse it to keep calls at 0 allocs/op.

At 60 FPS with ~50 FFI calls per frame, overhead is 5 µs per frame — 0.03% of the 16.6 ms budget. Unmeasurable in profiling.

See docs/PERFORMANCE.md for detailed analysis, optimization strategies, and when NOT to use goffi.

#Architecture

goffi transitions from Go's managed runtime to C code through three layers:

Go Code
  │  ffi.CallFunction()
  ▼
runtime.cgocall               ← Go runtime: system stack switch, GC coordination
  │
  ▼
Assembly Wrapper              ← Hand-written: load GP/SSE registers per ABI
  │  CALL target_function
  ▼
C Function                    ← External library

Three ABIs — hand-written assembly for System V and AAPCS64; on Windows amd64 the Go runtime's syscall.SyscallN builds the Win64 frame:

See docs/ARCHITECTURE.md for the full technical deep dive.

#Callbacks (C → Go)

WebGPU fires async callbacks from internal Metal/Vulkan threads. These threads have no goroutine — calling Go directly would crash.

goffi uses crosscall2 (Linux, macOS, FreeBSD) and syscall.NewCallback (Windows) for safe C→Go transitions from any thread:

// uintptr-sized arguments and one uintptr result work on every platform;
// Windows callbacks accept only that shape.
cb := ffi.NewCallback(func(status, adapter, msg, ud uintptr) uintptr {
    // Safe even when called from a C thread
    result.handle = adapter
    close(done)
    return 0
})

_, _ = ffi.CallFunction(cif, wgpuRequestAdapter, nil, args)
<-done // Wait for GPU driver callback

On Unix: 2000 pre-compiled trampoline entries per process (AMD64: 5 bytes/entry, ARM64: 8 bytes/entry); float32/float64 arguments and results are supported there. Windows callbacks take uintptr-sized arguments only (syscall.NewCallback rules).

#Error Handling

Five typed error types for precise diagnostics:

handle, err := ffi.LoadLibrary("nonexistent.dll")
if err != nil {
	var libErr *ffi.LibraryError
	if errors.As(err, &libErr) {
		fmt.Printf("Failed to %s %q: %v\n", libErr.Operation, libErr.Name, libErr.Err)
	}
}

#Comparison: goffi vs purego vs CGO

Choose goffi for GPU/real-time workloads: struct passing, zero per-call overhead, callback float returns, typed errors.

Choose purego for general-purpose bindings: string auto-marshaling, broad architecture support, less boilerplate.

See also:

• JupiterRider/ffi — pure Go binding for libffi via purego. Supports struct pass/return and variadic functions; requires libffi at runtime.
• unxed/pureffi — A 1:1 API-compatible drop-in replacement for purego implemented entirely on top of goffi. It resolves fakecgo linker conflicts without build tags, correctly handles macOS ARM64 variadic stack-packing ABI requirements under the hood, and provides a zero-code-change migration path.

#Known Limitations

Linux and FreeBSD: CGO_ENABLED=1 together with other cgo code does not link (#72, fix discussed in #104)

• Affects binaries that also contain another import "C" package (purego, database drivers, …) or are linked with -linkmode=external, -buildmode=c-shared or c-archive. goffi alone with CGO_ENABLED=1 links (internal linker). macOS and Windows are not affected.
• goffi + a plain cgo package: add -ldflags=-linkmode=internal. This does not help with purego, c-shared or c-archive.
• goffi + purego: build with CGO_ENABLED=0 -tags nofakecgo (see the purego entry below).

Linux: default builds are dynamically linked (#74)

• Importing goffi records libdl/libc via cgo_import_dynamic even with CGO_ENABLED=0. Use -tags goffi_static for a fully static ELF (no runtime .so loading), or build against musl for Alpine. See Linking modes.

Windows: C++ exceptions may crash the program (#12516)

• Go runtime limitation, not goffi-specific. Go 1.22+ added partial SEH support (#58542), but edge cases remain.
• Workaround: build native libraries with panic=abort.

Apple ARM64: variadic args always go on stack

• Per Apple's AAPCS64 extension, variadic arguments must be passed on the stack even when GP/FP registers are available. Use PrepareVariadicCallInterface (not PrepareCallInterface) for variadic C functions on all platforms — goffi handles the Darwin-specific register flush automatically.

Struct packing follows System V ABI only

• Windows #pragma pack not honored. Manually specify Size/Alignment in TypeDescriptor.

No bitfields in struct types.

Unix: duplicate symbol conflict with purego (#22)

• When using goffi and purego in the same binary with CGO_ENABLED=0, the linker reports duplicated definition of symbol _cgo_init. Both libraries include internal/fakecgo which defines identical runtime symbols.
• Workaround A: Build with -tags nofakecgo to disable goffi's fakecgo, relying on purego's copy:

  CGO_ENABLED=0 go build -tags nofakecgo ./...
  # FreeBSD: also -gcflags=github.com/ebitengine/purego/internal/fakecgo=-std
  

• Workaround B (macOS only for now): Build with CGO_ENABLED=1. In cgo mode the real runtime/cgo supplies the symbols and both libraries' fakecgo packages are gated out by //go:build !cgo. On Linux and FreeBSD this currently fails to link (#72, see above).
• Workaround C (Recommended for complete integration): Use pureffi. pureffi is a 1:1, drop-in API-compatible wrapper over goffi that acts as a transparent replacement for purego via Go's replace directive. It completely eliminates the need for compilation tags, resolves the duplicate symbol conflict, and natively fixes Apple Silicon variadic ABI stack-packing.

#Platform Support

#Roadmap

Planned work: ROADMAP.md and the milestones. What each release changed: CHANGELOG.md.

#Testing

go test ./...                          # all tests (C-backed e2e tests need a C compiler, else they skip)
go test -cover ./...                   # with coverage (CI gate: 70%)
CGO_ENABLED=1 go test -race ./...      # race detector (needs cgo and a C compiler)
go test -bench=. -benchmem ./ffi       # benchmarks
go test -v ./ffi                       # verbose, auto-detects platform

#Documentation

#Contributing

See CONTRIBUTING.md for guidelines.

1. Fork → branch from main → tests (CI gate: 70% coverage) → lint → PR to main
2. Conventional commits: feat:, fix:, docs:, test:

#Acknowledgments

• purego — proved that pure Go FFI is possible. The crosscall2 callback mechanism, fakecgo approach, and assembly trampoline patterns were pioneered by purego. goffi exists because purego cleared the path.
• libffi — reference for FFI architecture patterns and CIF design.
• Go runtime — runtime.cgocall for GC-safe stack switching, crosscall2 for C→Go transitions.

#Ecosystem

goffi powers an ecosystem of pure Go GPU libraries:

#Star History

#License

MIT — see LICENSE.

GitHub | pkg.go.dev | Dev.to