test: support target-matrix and post-ABI IR checks by zhouguangyuan0718 · Pull Request #2319 · xgo-dev/llgo

zhouguangyuan0718 · 2026-08-13T17:21:24Z

Summary

Add explicit, source-embedded FileCheck support for target matrices and opt-in post-ABI IR while keeping focused handwritten checks as the default.

// LITTEST <GOOS/GOARCH...> cross-generates the existing default IR stage for every listed target.
// LITTEST: POST-ABI <GOOS/GOARCH...> explicitly selects target-ABI-lowered, pre-optimization IR.
The fixture's current effective target is always checked; listed targets add cross-compilation coverage and a matching current target is deduplicated.
An unlisted CI platform still exercises portable CHECK and applicable GOARCH assertions. Add it to the matrix when it has a distinct IR contract.
Runtime output still runs once on the host; target-matrix IR generation is compile-only.
FileCheck enables a layered prefix set such as CHECK,ARM64,DARWIN-ARM64: portable assertions use CHECK, same-architecture assertions use ARM64/AMD64, and only true OS differences use the exact target prefix.
litgen supports automatic target-matrix maintenance at the default and post-ABI stages. It first merges all-target output, then safely merges identical gaps among two or more targets with the same GOARCH.
The float conversion post-ABI matrix is automatically maintained; its directive count drops from 101 to 73 by sharing identical Darwin/Linux arm64 lowering checks without weakening coverage.
Existing platform alternatives such as closure nest|swiftself, jmpbuf sizes, setjmp names, and stderr symbols are replaced by explicit target assertions.
44 non-cgo default-stage fixtures now cross-check Darwin/arm64 and Linux/amd64 on either host. Four direct or transitive cgo fixtures retain plain // LITTEST, because cross-cgo additionally requires a target C compiler, headers, and sysroot; their explicit platform checks run on matching platform CI.
Obsolete literal out.ll IR-check modes remain removed; source-embedded FileCheck is the only IR golden path.

Validation

go fmt ./...
git diff --check
go test ./internal/littest ./internal/filecheck ./internal/llgen ./chore/litgen ./cl/cltest
go test ./chore/litgen -coverprofile=... (80.7% statement coverage)
go run ./chore/litgen -u --check cl
go test ./cl -count=1 on Darwin/arm64 (includes current and additional matrix targets; 307.038s)
Three-target POST-ABI integration coverage verifies ARM64 folding while retaining LINUX-AMD64 differences
Focused unit and fixture suites on goallc-ubuntu, proving Linux current-target checks are deduplicated while Darwin/arm64 and Linux/arm64 share the generated architecture contract

codecov · 2026-08-13T17:44:07Z

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

github-actions · 2026-08-15T02:37:52Z

LLGo baseline benchmarks

8b5e93fc852d | workflow run | long-term charts

Program measurements

Platform	Workload	File size	vs base	Build	vs base	Run	vs base
Linux	cprintf	19384 B	+0.0%	341.795 ms	-1.6% (better)	1.261 ms	-4.9% (better)
Linux	fmtprintf	1880776 B	+0.0%	2.635 s	-1.2% (better)	3.291 ms	+0.2% (worse)
Linux	println	68880 B	+0.0%	338.839 ms	-2.8% (better)	1.541 ms	-8.4% (better)
macOS	cprintf	84672 B	+0.0%	448.508 ms	+10.7% (worse)	2.783 ms	+7.9% (worse)
macOS	fmtprintf	1892208 B	+0.0%	2.440 s	-34.5% (better)	12.662 ms	-17.7% (better)
macOS	println	121376 B	+0.0%	381.447 ms	-2.7% (better)	4.011 ms	-9.0% (better)

Core language and compiler benchmarks

Platform	Benchmark	ns/op	vs base
Linux	BenchmarkLookupPCRandom	13.230 ns/op	-0.5% (better)
Linux	BenchmarkMergeCompilerFlags	150.900 ns/op	-0.1% (better)
Linux	BenchmarkMergeLinkerFlags	94.840 ns/op	+0.1% (worse)
Linux	BenchmarkChannelBuffered	33.690 ns/op	+0.1% (worse)
Linux	BenchmarkChannelHandoff	26133 ns/op	-10.9% (better)
Linux	BenchmarkDefer	44.840 ns/op	+0.9% (worse)
Linux	BenchmarkDirectCall	1.557 ns/op	-0.1% (better)
Linux	BenchmarkGlobalRead	1.557 ns/op	+0.1% (worse)
Linux	BenchmarkGlobalWrite	2.487 ns/op	+0.1% (worse)
Linux	BenchmarkGoroutine	32651 ns/op	-0.7% (better)
Linux	BenchmarkInterfaceCall	8.106 ns/op	+0.1% (worse)
Linux	BenchmarkRuntimeGetG	1.868 ns/op	+0.0%
macOS	BenchmarkLookupPCRandom	10.970 ns/op	-14.3% (better)
macOS	BenchmarkMergeCompilerFlags	97.230 ns/op	-9.0% (better)
macOS	BenchmarkMergeLinkerFlags	61.110 ns/op	-23.3% (better)
macOS	BenchmarkChannelBuffered	20.470 ns/op	-1.3% (better)
macOS	BenchmarkChannelHandoff	6483 ns/op	+0.7% (worse)
macOS	BenchmarkDefer	25.620 ns/op	-1.5% (better)
macOS	BenchmarkDirectCall	0.941 ns/op	-6.6% (better)
macOS	BenchmarkGlobalRead	0.957 ns/op	-5.9% (better)
macOS	BenchmarkGlobalWrite	0.941 ns/op	-5.9% (better)
macOS	BenchmarkGoroutine	27189 ns/op	-26.3% (better)
macOS	BenchmarkInterfaceCall	5.356 ns/op	-1.3% (better)
macOS	BenchmarkRuntimeGetG	2.322 ns/op	+6.4% (worse)

Compared with 0f480253029a measured in the same runner job.

zhouguangyuan0718 · 2026-08-15T10:17:19Z

/review

zhouguangyuan0718 · 2026-08-15T10:29:21Z

/review

zhouguangyuan0718 · 2026-08-15T10:39:33Z

/review

zhouguangyuan0718 · 2026-08-16T14:50:09Z

/review

zhouguangyuan0718 · 2026-08-16T14:57:55Z

/review

zhouguangyuan0718 · 2026-08-17T00:01:20Z

/review

zhouguangyuan0718 · 2026-08-17T02:53:54Z

/review

fennoai

Review: target-matrix and post-ABI IR checks

Solid, well-structured PR. The Spec-based refactor of littest is a clear improvement over the old Mode enum, error handling is consistent, dead code (generalizeClosureEnvAttrs, genZip/inCompilerDir, out.ll special-casing) is cleanly removed, and the multi-target merge logic (LCS + per-GOARCH grouping) is thoughtfully separated. Empty-slice guards are present in the merge/rewrite functions, and the panic→error recovery in genTargetVariants correctly re-panics on unexpected payloads.

Findings below are non-blocking; the first is the one most worth confirming. No blocking correctness bugs found.

Verification: four review dimensions (quality, performance, security, docs) run against the fixed diff range 7f954d6...17d1c86.

Additional notes (no inline location)

Performance (informational): per-target compiles in testFrom/testIRTargets and genTargetVariants scale linearly with the target matrix, and ForceRebuild=true (required for correctness — a cache hit skips the module text) means each compile pays full build cost. Fine for test tooling; just keep LITTEST matrices deliberately small.
Docs (cosmetic): CLAUDE.md:80-83 shows both // LITTEST ... and // LITTEST: POST-ABI ... markers in one fenced block. They are mutually exclusive first-line markers; splitting into two blocks (as dev/README.md does) would avoid reading them as one header.

fennoai · 2026-08-17T03:04:20Z

+func additionalIRTargets(targets []littest.Target, currentPrefix string) []littest.Target {
+	additional := make([]littest.Target, 0, len(targets))
+	for _, target := range targets {
+		prefixes := filecheck.TargetPrefixes(target.GOOS, target.GOARCH, "")


[P2] additionalIRTargets dedup uses inconsistent prefix derivation

currentPrefix is computed by the caller from the actual generated base target via filecheck.TargetPrefixes(goos, goarch, generated.Target), so if the base build has a non-empty named Target it becomes TARGET-<X>. But here each matrix entry's prefix is recomputed with an empty target string (TargetPrefixes(target.GOOS, target.GOARCH, "")), yielding <GOOS>-<GOARCH>. When the base target is named, the equality check never matches, so the entry equal to the base is not skipped and gets recompiled/rechecked redundantly. In the common golden-test flow Target is empty so this is latent, but the two sides derive the prefix from different inputs. Consider deduplicating on GOOS/GOARCH directly, or feeding the same TargetPrefixes inputs on both sides.

fennoai · 2026-08-17T03:04:20Z

+	default:
+		return Spec{}, false, nil


[P3] Malformed LITTEST marker silently disables the spec file

A first line that starts with the marker but doesn't parse — e.g. // LITTEST: POST-ABI-EXTRA or // LITTESTfoo — fails the exact-match switch and both HasPrefix(marker, X+" ") cases, so it falls into default and returns found=false, treating the file as having no marker at all. Given the fail-loud handling elsewhere (parseTargets errors on bad targets), a mistyped LITTEST/LITTEST: POST-ABI directive vanishing silently is inconsistent and could mask an author error. Consider returning an error when the line begins with the marker prefix but doesn't parse.

fennoai · 2026-08-17T03:04:20Z

+	seen := make(map[string]struct{}, len(fields))
+	for _, field := range fields {
+		goos, goarch, ok := strings.Cut(field, "/")
+		if !ok || goos == "" || goarch == "" || strings.Contains(goarch, "/") {


[P3] parseTargets: whitelist GOOS/GOARCH characters

parseTargets rejects empty components and a goarch containing /, but not ., .., backslashes, or other path-meaningful characters. These values later flow into a result.<GOOS>-<GOARCH>.txt filename (cltest.go) and into a FileCheck --check-prefix= arg. Practical path traversal is blocked (only one / is guaranteed and it's stripped; reaching the write requires the target to actually compile), and prefixes are passed as separate argv entries so there's no injection — but a defensive whitelist like ^[a-z0-9_]+$ on GOOS/GOARCH would remove the ambiguity cheaply. Also note the strings.Contains(goarch, "/") guard is what rejects a/b/c (since Cut splits on the first /); a brief comment there would clarify intent.