mirror of
https://github.com/hanzoai/wasm2go.git
synced 2026-10-30 15:06:48 +00:00
* wasm2go: a WebAssembly-to-Go ahead-of-time transpiler wasm2go translates a WebAssembly binary into standalone Go source. The generated package runs the module natively — no interpreter and no embedded wasm runtime — so it builds and links like ordinary Go code. This makes it practical to ship a wasm-compiled library (for example a large C/C++ codebase built with the WASI SDK) inside a Go program without the startup and memory cost of a wasm engine. Components: - internal/wasm — the WebAssembly binary parser. - internal/codegen — translates a parsed module into Go source through the SSA lowering + optimization + emission pipeline. - internal/ssa — a lightweight SSA IR plus optimization passes (const-propagation, DCE, CSE, branch folding, memory promotion). - transpile — the public library API: Parse, Translate, and the Transpile one-shot convenience wrapper. - cmd/wasm2go — the command-line interface. - internal/testfixture — shared helper that compiles testdata/*.wat to wasm on demand via wat2wasm. Library API (Options): - Required: Package, OutputImportPath. - Optional: BulkExportPrefix (per-export Inv_<svc>_<mt> dispatch for modules with uniformly-named exports), EntryExports (narrow DCE root set), KeepDeadFuncs (disable DCE for diff/debug), PromotionReportPath (write SSA memory-promotion JSON). Everything else is auto-derived from the input: the SSA pipeline runs unconditionally; data segments emit as //go:embed sidecars; wasi imports translate to Go-native syscalls; modules whose total function-body bytes exceed an internal threshold are emitted as a chain of //go:linkname-split sub-packages so the Go compiler stays inside a bounded per-package memory budget. WASI runtime: full wasi_snapshot_preview1 coverage. Args/environ, clocks, fd_*, path_* (open/create_directory/unlink_file/ remove_directory/rename/filestat_get/filestat_set_times/link/ symlink/readlink), readdir, pread/pwrite, sync/datasync, poll_oneoff (clock + fd-event branches), sched_yield, random_get, proc_exit, proc_raise, sock_accept/recv/send/shutdown — every call routes to the Go stdlib or golang.org/x/sys/unix equivalent. Per-platform companion files (unix/linux/darwin/windows) carry the bits that differ across OSes; the translated module is a normal Go program with the host's ordinary filesystem and network reach. Release engineering for the public repository: - GitHub Actions CI: build, vet, race-enabled tests behind an 85% coverage gate, golangci-lint — all driven through make targets. - Release workflow runs GoReleaser on version tags, including deb/rpm/apk packages, with build-provenance attestations recorded for every release asset. - A release-test workflow runs `make release/check` (GoReleaser config validation + a publish-free dry-run build) on every PR and main push so a broken release is caught before tag time. - golangci-lint and goreleaser are pinned in tools/go.mod and run via `go tool`, so local and CI invocations use the same versions. - `make install/wat2wasm` provisions the wat2wasm CLI the test suite needs, dispatching to the host's package manager (Homebrew on macOS, apt/dnf/pacman/apk on Linux). - Makefile targets for local development. Test coverage is 85.2% of statements overall. * README: document the WASI runtime, performance, and library entry points - Add a "WASI support" subsection explaining the single-file, stdlib-only `wasi_snapshot_preview1` implementation and the resulting cross-platform reach. - Add a "Performance" section with the wazero-vs-wasm2go corpus numbers (cold start 8.4×, short queries up to 81×, TPC-H scans within ±10%). Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]> * ci: align release-test with release.yml — use stable Go The release-test workflow exercises the same `make release/check` dry-run that the publish job runs at tag time, so it should pull the same toolchain. Switch it from `go-version-file: tools/go.mod` to `go-version: stable` to match release.yml. Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]> --------- Co-authored-by: Claude Opus 4.7 (1M context) <[email protected]>
95 lines
3.0 KiB
Go
95 lines
3.0 KiB
Go
package codegen
|
|
|
|
import (
|
|
"github.com/goccy/wasm2go/internal/wasm"
|
|
)
|
|
|
|
// computeReachableFuncs returns the set of DEFINED functions — keyed by
|
|
// local (defined-relative) index — reachable from the module's entry
|
|
// points through the direct-call graph.
|
|
//
|
|
// A wasm module compiled from a large C++ program carries thousands of
|
|
// functions that the program can never actually invoke. The Go linker
|
|
// cannot drop them because the generated InvokeExport dispatch
|
|
// references every export and the element segment takes the address of
|
|
// every table entry. Computing reachability here, at transpile time,
|
|
// lets wasm2go simply not emit the dead functions.
|
|
//
|
|
// Roots:
|
|
// - the start function;
|
|
// - every function in an element segment — call_indirect can target
|
|
// any table slot, so all table entries are conservatively live;
|
|
// - the function exports. When entryExports is nil EVERY export is a
|
|
// root, so only functions dead in the wasm itself are dropped.
|
|
// A non-nil entryExports restricts the export roots to the named
|
|
// subset (used by whole-program pruning when the caller knows the
|
|
// real entry set).
|
|
//
|
|
// The returned map is keyed by local index: function localIdx is kept
|
|
// iff result[localIdx] is true.
|
|
//
|
|
// callees, when non-nil, must be the precomputed call-graph adjacency
|
|
// list (one entry per defined function, with each entry listing the
|
|
// local indices it directly calls). Passing nil falls back to running
|
|
// buildCallGraph here — both PlanMultiPackage and reachability use
|
|
// the same graph, so the caller pre-computes it once and threads the
|
|
// result through to avoid scanning the bytecode three times.
|
|
func computeReachableFuncs(mod *wasm.Module, entryExports map[string]bool, callees [][]uint32) (map[uint32]bool, error) {
|
|
if callees == nil {
|
|
var err error
|
|
callees, err = buildCallGraph(mod) // callees[localIdx] = []localIdx
|
|
if err != nil {
|
|
return nil, err
|
|
}
|
|
}
|
|
nImp := mod.NumImportedFuncs
|
|
nDef := uint32(len(mod.Functions))
|
|
|
|
reachable := make(map[uint32]bool, nDef)
|
|
var work []uint32
|
|
addLocal := func(d uint32) {
|
|
if d >= nDef || reachable[d] {
|
|
return
|
|
}
|
|
reachable[d] = true
|
|
work = append(work, d)
|
|
}
|
|
addGlobal := func(fi uint32) {
|
|
if fi < nImp { // imported function — no body to emit or walk
|
|
return
|
|
}
|
|
addLocal(fi - nImp)
|
|
}
|
|
|
|
if mod.Start != nil {
|
|
addGlobal(*mod.Start)
|
|
}
|
|
for _, seg := range mod.Elements {
|
|
for _, fi := range seg.FuncIdxs {
|
|
addGlobal(fi)
|
|
}
|
|
}
|
|
for _, e := range mod.Exports {
|
|
if e.Kind != wasm.ExportFunc {
|
|
continue
|
|
}
|
|
// When entryExports is non-nil it lists the exact set of export
|
|
// names to keep as reachability roots. Any export not in the
|
|
// set is dropped from the root set (its function can still
|
|
// become reachable through the table or a direct call).
|
|
if entryExports != nil && !entryExports[e.Name] {
|
|
continue
|
|
}
|
|
addGlobal(e.Index)
|
|
}
|
|
|
|
for len(work) > 0 {
|
|
d := work[len(work)-1]
|
|
work = work[:len(work)-1]
|
|
for _, c := range callees[d] {
|
|
addLocal(c)
|
|
}
|
|
}
|
|
return reachable, nil
|
|
}
|