internal/packages: eliminate //go:linkname and decouple from x/tools internals

[cpunion]

Summary

Eliminate all //go:linkname directives and private struct mirroring in internal/packages/load.go, replacing them with standard public golang.org/x/tools/go/packages.Load and a self-contained, post-order go/types.NewChecker typechecker.

This resolves fragility across golang.org/x/tools upgrades (e.g., #2368) and achieves a cleaner import pipeline with reduced binary sizes for compiled programs.

Key Changes

1. Decouple from x/tools Private Internals:

   • Removed 5 //go:linkname hooks (defaultDriver, newLoader, loadFromExportData, parseFiles, typesinternalSetUsesCgo).
   • Removed mirrored private types (loader, loaderPackage, loadRecursiveEx, refineEx).
   • Standard public packages.Load handles go list driver invocation, module metadata, embed discovery, and file resolution.

2. Self-Contained Bottom-Up Typechecking:

   • Traverses the dependency DAG post-order via packages.Visit.
   • Uses standard go/types.NewChecker with a custom importerFunc ensuring complete type information across dependencies.
   • Accurately computes target architecture pointer sizes (supporting cross-compilation such as WASM 32-bit via cfg.Env target resolution).
   • Fully preserves:
     • Deduper caching (safely guarded on pkg.Types.Complete()) and preload
     • Custom types.Sizes overrides
     • llgoFiles patch file injection
     • Per-module / global GoVersion propagation (requesting NeedModule)
     • Authoritative syntax error detection vs parser secondary errors
     • Embed error normalization and proper *os.PathError propagation

3. Binary Size & Dead Code Elimination (DCE) Improvement:

   • Pure source-based typechecking avoids synthetic export data stubs and unneeded export metadata.
   • Yields clean types.Info records, allowing LLGo's backend SSA dead-code elimination and symbol pruning to be more effective.
   • Benchmark measurements show a modest binary size reduction on stdlib-dependent binaries (e.g., fmtprintf: -832 B on Linux, -272 B on macOS) with zero regression in performance.

4. Code Quality & Maintenance:

   • Net reduction of -355 lines of code.
   • Full GoDoc comments preserved on all exported public APIs (LoadMode, Config, Visit, LoadEx).

Verification

• go test -v ./internal/packages (PASS)
• go test -v ./internal/build -run TestWasm (PASS)
• bash test/buildcache/test.sh (18/18 Native, WASM, ESP32-C3 cache tests PASS)
• Full GitHub Actions CI Matrix (41/41 jobs PASS)

[github-actions]

LLGo baseline benchmarks

adb0bec9da4f | workflow run | long-term charts

Program measurements

Platform	Workload	File size	vs base	Build	vs base	Run	vs base
Linux	cprintf	19288 B	0 B / +0.0%	330.910 ms	-4.327 ms / -1.3% (better)	1.273 ms	-12.61 us / -1.0% (better)
Linux	fmtprintf	1860152 B	0 B / +0.0%	2.851 s	+130.5 ms / +4.8% (worse)	3.321 ms	+88.78 us / +2.7% (worse)
Linux	println	68776 B	0 B / +0.0%	338.743 ms	+10.36 ms / +3.2% (worse)	1.590 ms	+32.49 us / +2.1% (worse)
macOS	cprintf	84672 B	0 B / +0.0%	581.063 ms	+144.7 ms / +33.1% (worse)	3.730 ms	+622.3 us / +20.0% (worse)
macOS	fmtprintf	1892720 B	0 B / +0.0%	3.648 s	+467.3 ms / +14.7% (worse)	17.114 ms	+4.577 ms / +36.5% (worse)
macOS	println	121360 B	0 B / +0.0%	489.399 ms	+115.7 ms / +31.0% (worse)	4.704 ms	+621.5 us / +15.2% (worse)

Core language and compiler benchmarks

Platform	Benchmark	ns/op	vs base
Linux	BenchmarkLookupPCRandom	13.260 ns/op	+0.05 ns/op / +0.4% (worse)
Linux	BenchmarkMergeCompilerFlags	150.900 ns/op	+0.1 ns/op / +0.1% (worse)
Linux	BenchmarkMergeLinkerFlags	94.840 ns/op	-1.71 ns/op / -1.8% (better)
Linux	BenchmarkChannelBuffered	33.990 ns/op	+0.01 ns/op / +0.02943% (worse)
Linux	BenchmarkChannelHandoff	27868 ns/op	+813 ns/op / +3.0% (worse)
Linux	BenchmarkDefer	46.250 ns/op	-0.96 ns/op / -2.0% (better)
Linux	BenchmarkDirectCall	1.557 ns/op	0 ns/op / +0.0%
Linux	BenchmarkGlobalRead	1.557 ns/op	0 ns/op / +0.0%
Linux	BenchmarkGlobalWrite	2.482 ns/op	+0.001 ns/op / +0.04031% (worse)
Linux	BenchmarkGoroutine	32260 ns/op	-232 ns/op / -0.7% (better)
Linux	BenchmarkInterfaceCall	8.098 ns/op	-0.006 ns/op / -0.1% (better)
Linux	BenchmarkRuntimeGetG	2.180 ns/op	+0.001 ns/op / +0.04589% (worse)
macOS	BenchmarkLookupPCRandom	14.050 ns/op	-0.97 ns/op / -6.5% (better)
macOS	BenchmarkMergeCompilerFlags	187.600 ns/op	+32.1 ns/op / +20.6% (worse)
macOS	BenchmarkMergeLinkerFlags	111.400 ns/op	+21.57 ns/op / +24.0% (worse)
macOS	BenchmarkChannelBuffered	27.820 ns/op	-3.69 ns/op / -11.7% (better)
macOS	BenchmarkChannelHandoff	5873 ns/op	-2780 ns/op / -32.1% (better)
macOS	BenchmarkDefer	40.400 ns/op	-2.87 ns/op / -6.6% (better)
macOS	BenchmarkDirectCall	1.324 ns/op	-0.002 ns/op / -0.2% (better)
macOS	BenchmarkGlobalRead	1.157 ns/op	+0.011 ns/op / +1.0% (worse)
macOS	BenchmarkGlobalWrite	1.187 ns/op	-0.015 ns/op / -1.2% (better)
macOS	BenchmarkGoroutine	28182 ns/op	-5001 ns/op / -15.1% (better)
macOS	BenchmarkInterfaceCall	5.857 ns/op	-0.062 ns/op / -1.0% (better)
macOS	BenchmarkRuntimeGetG	2.363 ns/op	-0.298 ns/op / -11.2% (better)

Compared with bed9d7c4d4c0 measured in the same runner job.

[fennoai]

Review: refactor internal/packages to drop //go:linkname internals

This is a solid, well-motivated refactor. It removes the fragile //go:linkname hooks into golang.org/x/tools internals (defaultDriver, newLoader, loadFromExportData, parseFiles, typesinternalSetUsesCgo) and the private loader/loaderPackage types, replacing them with a self-contained driver: call the public packages.Load with an adjusted Mode, then parse + typecheck bottom-up via packages.Visit. This is much less likely to break across x/tools upgrades. internal/packages tests pass locally.

I verified the main worry — that dropping the export-data fallback would break NeedImports-without-NeedDeps loads — and it does not reproduce: the go list driver still returns dependency packages with populated CompiledGoFiles, so packages.Visit reaches and typechecks them from source (confirmed internal/bytealg loads with complete import types, IllTyped=false, 0 errors). So that path is fine.

A few findings below, ordered by impact. The wasm-sizes one (P1) is the only likely functional regression.

[fennoai]

Review (cont.): one additional finding

Follow-up to the review above — this P1 finding was omitted from the first submission due to a title-length limit. It is the most impactful item.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!