test/goroot: retire stale expectations and classify full-shard outcomes

[cpunion]

Summary

This PR retires the fully audited stale hosted expectations and classifies the outcomes found by the complete GOROOT matrix.

• keeps all 175 previously validated expectation removals;
• makes notapplicable.yaml global by construction: all 121 version and 9 platform selectors are removed, duplicate rangegen.go/maymorestack.go entries are collapsed, and stack.go is added once globally;
• accepts either success or an ordinary failure for a not-applicable case while keeping resource-guard failures fatal;
• globally host-skips fixedbugs/issue16016.go, chan/goroutines.go, and chanlinear.go, which can exhaust LLGo's one-thread-per-goroutine host model;
• adds ordinary xfail coverage for index0.go, fixedbugs/issue34123.go, fixedbugs/issue52612.go, and Linux Go 1.26 convert5.go, and consolidates heapsampling.go globally;
• makes the observed typeparam/chans.go flake global.

The repository now has 133 unique global not-applicable entries, with no version/platform selectors. stack.go is no longer present in flakes.

Full-shard findings

The unmodified 16-shard Darwin/Linux × Go 1.25.0/1.26.5 run on the prior head received all 16 reports and observed all 4,170 selected cases:

• 4,170 selected / 4,170 observed;
• 4,125 classified passes, 41 failures, 4 skips;
• 16 unique failing paths;
• zero failure tuple intersections with the 175 validated removals.

Those 16 paths are now handled as follows:

• 8 design-level cases are global not-applicable entries;
• 5 semantic/runtime error cases have ordinary xfail coverage;
• 2 unbounded thread-creation cases are global host skips;
• fixedbugs/bug449.go remains owned by ssa: lower large array comparisons to runtime equality algorithms #2416 and is not duplicated here.

The older full run also exposed chanlinear.go; it is now a global host skip. typeparam/chans.go was independently reproduced as flaky and is globally classified.

Relationship to other PRs

• The branch is based directly on main and contains none of ssa: lower large array comparisons to runtime equality algorithms #2416 or cl: preserve diagnostics for invalid method receivers #2417.
• fixedbugs/bug449.go remains owned by ssa: lower large array comparisons to runtime equality algorithms #2416.
• fixedbugs/issue5089.go and the cl diagnostics remain owned by cl: preserve diagnostics for invalid method receivers #2417.
• runtime: stabilize goroutine hooks and GOROOT classifications #2421 remains the implementation follow-up: after this PR, its runtime fixes can remove the corresponding fixedbugs/issue52612.go and goprint.go expectations without duplicating the runner/not-applicable classifications moved here.

Validation

• GOMAXPROCS=2 go test -p=2 -parallel=2 -count=1 ./test/goroot;
• YAML parse, global-selector, uniqueness, and cross-file separation tests;
• exact audit of all 16 shard artifacts from https://github.com/cpunion/llgo/actions/runs/32972501873;
• git diff --check;
• current unmodified 16-shard validation: https://github.com/cpunion/llgo/actions/runs/32980102573.

The current integration head cb9de75e5 is exact PR head 9501a7ba6 plus #2416 head 78a9f63bc; .github/workflows/goroot.yml is byte-identical to main. Go 1.24 is not a compatibility target; supported runtime validation is Go 1.25/1.26.

[cpunion]

Full GOROOT validation run (16 configured OS/Go-version/shard jobs):

https://github.com/cpunion/llgo/actions/runs/32942085570

This run is pinned to head a94c25f7f25a0127046022ae447380db477f12b6.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[fennoai]

Review Summary

This is a clean, well-scoped change. The bulk of the diff is data (retiring 187 validated stability expectations from xfail.yaml), and the code changes are small and targeted:

• The procPin relocation from runtime/internal/lib/runtime/sync_runtime_llgo.go into the new core file runtime/internal/runtime/procpin.go is a pure relocation — the mutex/Once logic is byte-for-byte identical to the pre-existing code, and the darwin || linux build tag is preserved. The complementary !darwin && !linux no-op in runtime/_patch/sync/atomic/value_other.go keeps the linkname coverage exhaustive with no overlap. No behavior change.
• runtime.NumGoroutine now reports live LLGo runtime contexts; calling getg() first to register a lazily-initialized main/foreign thread before reading the count is correct (getg -> initRuntimeContext -> retainG).
• TestRuntimeNumGoroutineTracksWorkers is correctly designed around the worker-release race, using a deadline-bounded Gosched poll rather than an immediate assertion.
• Forcing GOTOOLCHAIN=local and running stdlib discovery from an isolated temp dir is a sensible hardening/isolation improvement, and the stack.go global not-applicable reason ("recursively grows through 8,000 large frames") matches upstream test/stack.go (recur(8000)).

No blocking issues found. A few low-severity maintainability/documentation notes are inline.

[fennoai]

Review notes (follow-up)

The two low-severity maintainability/documentation notes from my summary review are attached inline below. No blocking issues.

[cpunion]

The first dispatched run was cancelled after all started Linux shards exposed one deterministic runner setup error: the temporary GOPATH still linked the checkout under the pre-migration github.com/goplus/llgo path, while the module/runtime imports use github.com/xgo-dev/llgo. The failures happened before individual expectation validity could be evaluated.

Fixed in f9eeed831 with a path-level regression test and a bounded real compile/run of 235.go on Go 1.26.5.

Replacement full 16-shard validation run (exact updated head):
https://github.com/cpunion/llgo/actions/runs/32942797439

[github-actions]

LLGo baseline benchmarks

9501a7ba60aa | workflow run | long-term charts

Program measurements

Platform	Workload	File size	vs base	Text size	vs base	Build	vs base	Run	vs base
Linux	cprintf	19256 B	0 B / +0.0%	387 B	0 B / +0.0%	315.243 ms	+71.65 ms / +29.4% (worse)	925.272 us	+78.26 us / +9.2% (worse)
Linux	cprintf-lto	19088 B	0 B / +0.0%	368 B	0 B / +0.0%	270.670 ms	+26.2 ms / +10.7% (worse)	895.976 us	+117.1 us / +15.0% (worse)
Linux	fmtprintf	1653232 B	0 B / +0.0%	501083 B	0 B / +0.0%	2.305 s	+7.763 ms / +0.3% (worse)	2.256 ms	-13.6 us / -0.6% (better)
Linux	fmtprintf-lto	1525824 B	0 B / +0.0%	459700 B	0 B / +0.0%	6.852 s	+11.09 ms / +0.2% (worse)	2.249 ms	-75.59 us / -3.3% (better)
Linux	println	62032 B	0 B / +0.0%	15261 B	0 B / +0.0%	285.753 ms	+38.52 ms / +15.6% (worse)	1.102 ms	+75.17 us / +7.3% (worse)
Linux	println-lto	53840 B	0 B / +0.0%	12882 B	0 B / +0.0%	383.959 ms	+3.505 ms / +0.9% (worse)	1.070 ms	+34.26 us / +3.3% (worse)
macOS	cprintf	84480 B	0 B / +0.0%	16493 B	0 B / +0.0%	615.907 ms	+167.2 ms / +37.3% (worse)	3.717 ms	+808.2 us / +27.8% (worse)
macOS	cprintf-lto	100704 B	0 B / +0.0%	16473 B	0 B / +0.0%	776.654 ms	+370.7 ms / +91.3% (worse)	10.016 ms	+6.938 ms / +225.4% (worse)
macOS	fmtprintf	1498256 B	0 B / +0.0%	882133 B	0 B / +0.0%	4.613 s	+1.692 s / +57.9% (worse)	7.460 ms	+939.3 us / +14.4% (worse)
macOS	fmtprintf-lto	1208976 B	0 B / +0.0%	875129 B	0 B / +0.0%	7.517 s	+264.2 ms / +3.6% (worse)	7.971 ms	+1.596 ms / +25.0% (worse)
macOS	println	114832 B	0 B / +0.0%	34849 B	0 B / +0.0%	712.881 ms	+344.2 ms / +93.4% (worse)	4.348 ms	+738.2 us / +20.5% (worse)
macOS	println-lto	118656 B	0 B / +0.0%	32489 B	0 B / +0.0%	788.978 ms	+232.5 ms / +41.8% (worse)	5.115 ms	+2.236 ms / +77.7% (worse)

Core language and compiler benchmarks

Platform	Benchmark	ns/op	vs base
Linux	BenchmarkLookupPCRandom	8.291 ns/op	+0.027 ns/op / +0.3% (worse)
Linux	BenchmarkMergeCompilerFlags	103.100 ns/op	+1.8 ns/op / +1.8% (worse)
Linux	BenchmarkMergeLinkerFlags	65.880 ns/op	+0.8 ns/op / +1.2% (worse)
Linux	BenchmarkChannelBuffered	41.090 ns/op	+0.97 ns/op / +2.4% (worse)
Linux	BenchmarkChannelHandoff	23954 ns/op	+967 ns/op / +4.2% (worse)
Linux	BenchmarkDefer	43.690 ns/op	+2.56 ns/op / +6.2% (worse)
Linux	BenchmarkDirectCall	0.259 ns/op	+0.0025 ns/op / +1.0% (worse)
Linux	BenchmarkGlobalRead	0.373 ns/op	+0.0114 ns/op / +3.2% (worse)
Linux	BenchmarkGlobalWrite	7.154 ns/op	+0.142 ns/op / +2.0% (worse)
Linux	BenchmarkGoroutine	40242 ns/op	+6343 ns/op / +18.7% (worse)
Linux	BenchmarkInterfaceCall	4.336 ns/op	-0.021 ns/op / -0.5% (better)
Linux	BenchmarkRuntimeGetG	0.754 ns/op	+0.0265 ns/op / +3.6% (worse)
macOS	BenchmarkLookupPCRandom	15.500 ns/op	+0.83 ns/op / +5.7% (worse)
macOS	BenchmarkMergeCompilerFlags	147.900 ns/op	+22.6 ns/op / +18.0% (worse)
macOS	BenchmarkMergeLinkerFlags	99.880 ns/op	+17.64 ns/op / +21.4% (worse)
macOS	BenchmarkChannelBuffered	29.440 ns/op	-2.54 ns/op / -7.9% (better)
macOS	BenchmarkChannelHandoff	8842 ns/op	+1459 ns/op / +19.8% (worse)
macOS	BenchmarkDefer	42.160 ns/op	-8.69 ns/op / -17.1% (better)
macOS	BenchmarkDirectCall	1.111 ns/op	-0.312 ns/op / -21.9% (better)
macOS	BenchmarkGlobalRead	1.051 ns/op	-0.274 ns/op / -20.7% (better)
macOS	BenchmarkGlobalWrite	1.054 ns/op	-0.51 ns/op / -32.6% (better)
macOS	BenchmarkGoroutine	51952 ns/op	+14836 ns/op / +40.0% (worse)
macOS	BenchmarkInterfaceCall	5.218 ns/op	-1.948 ns/op / -27.2% (better)
macOS	BenchmarkRuntimeGetG	2.488 ns/op	-0.152 ns/op / -5.8% (better)

Compared with 6fe8babd468c measured in the same runner job.

[cpunion]

Fresh deletion-only GOROOT validation run:

https://github.com/cpunion/llgo/actions/runs/32954622318

The integration ref is fa08f98c8: exact #2419 head 6f1a070f6 plus only #2416's workspace-path commit 8ba212556, which is required for the fork/module rename. It does not include #2421 or #2416's compiler fixes. The workflow runs all 16 Darwin/Linux × Go 1.25.0/1.26.5 shards with the existing build/run timeouts and 3 GiB RSS guard.

[cpunion]

Regular CI retry note: the only failed check on 6f1a070f6 was the macOS Intel release-artifact job failing to download a 272 MB artifact after five retries, before testing it. The contribution commit was amended without content changes to trigger a fresh CI run; 6f1a070f6 and current head 4e5c75d90 have the identical tree 5d56480e3735d7d718a62b56ee068e16c974f554.

The full GOROOT run on the temporary integration ref remains an exact tree validation of the current PR content plus the separately owned workspace-path commit.

[cpunion]

CI retry: the macOS Go job on 4e5c75d9 failed because GitHub returned HTTP 502 while downloading goplus/compiler-rt; a later attempt in the same job downloaded it successfully. I refreshed the fork contribution commit to retrigger checks without changing source content. New head: 408e13e1444e8da5673a24681b204153a32f1204; tree: 5d56480e3735d7d718a62b56ee068e16c974f554 (identical to 4e5c75d9).

[cpunion]

Final deletion-only GOROOT result: https://github.com/cpunion/llgo/actions/runs/32954622318

• all 16 Darwin/Linux × Go 1.25.0/1.26.5 shard artifacts were produced;
• selected = 4,170 and observed = 4,170, so no selected case was omitted;
• the workflow is red because main currently has 41 unclassified failures across 17 case paths;
• conservative path-only comparison against all 176 removed entries (100 unique paths) found zero overlap; exact platform/version/directive matching therefore also has zero overlap;
• all 4 skips are selectors deliberately retained in this deletion-only PR.

The integration commit fa08f98c8 contains the PR tree 5d56480e3735d7d718a62b56ee068e16c974f554 plus only #2416 workspace-path commit 8ba212556; it contains none of #2421 and none of #2416 compiler fixes. No deleted expectation needs to be restored.

[cpunion]

Focused GOROOT pass run for the deletion set only: https://github.com/cpunion/llgo/actions/runs/32961976670

The temporary fork ref keeps the exact #2419 tree, plus #2416 workspace-path setup and a workflow-only filter commit. It runs every removed selector rather than the unrelated full GOROOT corpus:

• Darwin/Linux × Go 1.24.0/1.26.5 × 2 shards (8 jobs);
• 99/61/10/9 removed entries in the four platform/version lanes, covering 97/61/10/9 unique directive/path executions;
• versionless selectors run in both Go lanes;
• GOMAXPROCS=1, 3-minute build timeout, 1-minute run timeout, 3 GiB RSS cap, and 90-minute job cap.

Other PR checks are intentionally not being used as the gate until this focused GOROOT run passes.

[cpunion]

Focused GOROOT replacement run: https://github.com/cpunion/llgo/actions/runs/32962217541

The first focused run was cancelled after the Go 1.24 lanes failed before case selection: the repository go.mod requires Go 1.25+, so Go 1.24 cannot compile the test runner. The replacement keeps Go 1.24 as the target GOROOT/baseline but uses Go 1.26.5 to compile the repository runner. The candidate set and all resource bounds are unchanged.

[cpunion]

Supported-toolchain focused GOROOT run: https://github.com/cpunion/llgo/actions/runs/32962609618

Per maintainer direction, Go 1.24 is not a compatibility target. This final matrix runs only Go 1.25.0 and 1.26.5 on Darwin/Linux. It covers every removed selector applicable to those supported toolchains (2/61/1/9 selectors by platform/version lane); the 106 Go1.24-only entries are deletion of unsupported compatibility metadata, not runtime claims. Resource bounds remain GOMAXPROCS=1, 3-minute build, 1-minute run, 3 GiB RSS, and 90-minute job cap.

[cpunion]

Focused run finding: https://github.com/cpunion/llgo/actions/runs/32962609618 completed 72/73 applicable selectors successfully. Darwin Go 1.26 rangegen.go crossed the strict RSS guard (3076.7 MiB > 3072 MiB) while baseline Go compiled the generated program. Because it failed a bounded validation run, its Go1.26 20-minute timeout entry will be retained; the unsupported Go1.24 entry remains removed.

Replacement validation for the resulting 175-entry deletion set (60 Darwin Go1.26 selectors instead of 61): https://github.com/cpunion/llgo/actions/runs/32963761626. The PR branch itself is intentionally not updated until this GOROOT run passes, so other checks are not retriggered early.

[cpunion]

Final focused GOROOT gate passed: https://github.com/cpunion/llgo/actions/runs/32963761626

All four supported platform/toolchain jobs are green and their artifacts report:

• Darwin Go 1.25: 2 selected / 2 observed / 2 passed;
• Darwin Go 1.26: 60 / 60 / 60;
• Linux Go 1.25: 1 / 1 / 1;
• Linux Go 1.26: 9 / 9 / 9;
• total: 72 selected / 72 observed / 72 passed, 0 failed, 0 skipped.

Final PR head 47647ffc8 retains the Darwin Go1.26 rangegen.go timeout found necessary by the strict RSS run. The final semantic diff is 175 deletions and zero additions. Other PR checks have now been retriggered from this final head.

[cpunion]

Final head 47647ffc8 is fully green: 42 checks passed, 1 release job was expectedly skipped, 0 failed, and 0 remain pending. The PR is mergeable with a clean merge state. Both review conversations are resolved. The diff remains one file (test/goroot/xfail.yaml) with 175 semantic deletions and zero additions.

[cpunion]

Correction to the focused-run evidence: run 32963761626 used a temporary modified goroot.yml, so it is diagnostic evidence only and is not the final GOROOT gate. That temporary remote branch has been deleted.

The final full-shard run uses the repository workflow unchanged: https://github.com/cpunion/llgo/actions/runs/32972501873

Integration head c5668df9d is exact #2419 head 47647ffc8 plus only workspace-path commit 8ba212556. .github/workflows/goroot.yml is byte-identical to the PR/base version. This run executes the original 16 Darwin/Linux × Go1.25/1.26 × 4-shard jobs.

[cpunion]

Scope updated from deletion-only based on the complete 16-shard result.

• notapplicable.yaml is now genuinely global: no version/platform selectors, duplicates collapsed, and stack.go added once.
• not-applicable success and ordinary failure are both classified; resource-guard failures remain fatal.
• the remaining observed semantic/runtime errors have ordinary xfail coverage; unbounded thread-creation cases use host skips instead of being executed as xfails.
• fixedbugs/bug449.go is still left entirely to ssa: lower large array comparisons to runtime equality algorithms #2416.

New head: 9501a7ba6. The replacement unmodified 16-shard run is https://github.com/cpunion/llgo/actions/runs/32980102573.