test: replace cl out.ll files with LIT checks

[cpunion]

Summary

• replace 25 placeholder out.ll cases with source-embedded LIT checks covering their complete relevant functions and control flow
• remove the other seven obsolete out.ll files: six empty files already covered by LIT checks and one unused nested artifact
• preserve all expect.txt runtime-output comparisons and leave llgen, gentests, and the test framework unchanged

Portability

• keep full instruction ordering with CHECK-LABEL, CHECK-NEXT, and CHECK-EMPTY
• use patterns only for unstable identifiers and observed platform ABI differences, including nest/swiftself, cgo hashes and C types, libc symbols, pthread layout, and reflect method slots
• match the defer save-area allocation as alloca i8 without fixing its platform-defined size
• retain explicit debug checks for function !dbg, DICompileUnit, and DILocalVariable metadata

Companion documentation update: #2279

Validation

• all 25 affected cases passed on macOS arm64
• all 25 affected cases passed on Linux amd64 through dev/docker.sh
• go test ./internal/littest ./internal/llgen ./chore/litgen ./chore/llgen ./chore/gentests -count=1
• git diff --check

[fennoai]

Review summary

This is a clean, well-executed migration from golden out.ll snapshots to source-embedded // LITTEST + // CHECK FileCheck directives. Dead code removal is thorough (no dangling references to SmartDoFile, DoFile, genZip, inCompilerDir, cltest.Pkg, decodeLinkFile, or the Mode enum remain), the internal/littest simplification and new ErrSpecNotFound sentinel are genuine improvements, and the removed TestCgofullGeneratesC2func is now covered by CHECK-DAG in cgofull.go. Moving defer_ir_regression_test.go off a go run ./chore/llgen subprocess to an in-process llgen.GenFrom call is also a correctness/speed win. internal/littest vets clean.

The findings below are all in files the PR did not touch, so they cannot be placed inline.

Findings

1. (High) Stale documentation references the deleted chore/gentests / chore/llgen tooling and the removed out.ll workflow. This PR deletes both chore/gentests/ and chore/llgen/ entirely and removes all out.ll files, but the docs still instruct contributors to use them — these commands will now fail and the described workflow no longer exists:

• CLAUDE.md:60-102 — the "Understanding out.ll files" / "Required steps after modifying IR generation logic" section still describes out.ll as "generated by llgen" and instructs go install -v ./chore/gentests, go install -v ./chore/llgen, and running gentests / llgen. CLAUDE.md:16 also lists llgen as a chore/ tool.
• README.md:425-426 — the tools list still documents [llgen](chore/llgen) and [gentests](chore/gentests) (now broken in-repo links) with out.ll/expect.txt regeneration descriptions.
• dev/README.md:77-97 — the "Refresh test goldens" section documents a ### gentests flow with go run ./chore/gentests for regenerating out.ll.

Suggest updating these to describe the // LITTEST + // CHECK FileCheck workflow (and the surviving litgen tool, go run ./chore/litgen <path>) before merge. Note the internal/llgen library package still exists and is fine — only the chore/llgen command was removed.

2. (Minor) cltest.CaptureMeta is now dead exported code. cl/cltest/cltest.go:370 — CaptureMeta was only ever called from the deleted chore/gentests/gentests.go; it now has zero callers (its sibling RunAndCapture is still used by cl/compile_test.go:449). Consider removing it (and any orphaned meta-capture plumbing), or adding a comment noting an intended future caller.

3. (Minor / optional, pre-existing) IR checks now shell out per test package. With the ModeSkip short-circuit and the in-process ModeLiteral string compare both removed, every // LITTEST package now routes through filecheck.Match → llvm.New("") (internal/filecheck/filecheck.go:36), which re-execs llvm-config --bindir and then FileCheck on every call. Across the ~200 marked packages that is a meaningful number of redundant llvm-config subprocess spawns per run. This is pre-existing behavior in internal/filecheck (not introduced here), but this PR increases the number of packages routed through it. A process-level sync.Once around the llvm.New("") bindir lookup would remove the redundant execs at zero correctness cost. Not blocking.

No security concerns: removing genZip (a zip shell-out, previously dead behind if false) and decodeLinkFile (an archive/zip decode of committed fixtures) is a net reduction in surface area.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[cpunion]

Thanks for the review. The review was generated against the initial broader version of this PR. Commit d9b118f has since narrowed the final diff to cl/_test*/** only.

1. chore/llgen and chore/gentests are restored unchanged, so the documentation links and commands are no longer broken by this PR. The companion documentation cleanup remains in docs: add contributor guide and validation wrappers #2279.
2. cltest.CaptureMeta is restored to its previous state and is again called by chore/gentests.
3. ModeSkip, ModeLiteral, and the existing FileCheck plumbing are restored unchanged. Only the 25 newly migrated fixture cases now use source-embedded LIT checks.

The current PR diff contains only fixture sources and out.ll deletions; no tooling or test-framework code remains changed.

[github-actions]

LLGo baseline benchmarks

275b202668d4 | workflow run | long-term charts

Program measurements

Platform	Workload	File size	vs base	Build	vs base	Run	vs base
Linux	cprintf	18456 B	+0.0%	316.176 ms	-1.4% (better)	1.314 ms	-4.8% (better)
Linux	fmtprintf	1829968 B	+0.0%	3.196 s	+2.2% (worse)	3.708 ms	+4.5% (worse)
Linux	println	68008 B	+0.0%	306.229 ms	+0.6% (worse)	1.703 ms	-0.9% (better)
macOS	cprintf	84672 B	+0.0%	438.352 ms	-10.5% (better)	3.723 ms	+21.7% (worse)
macOS	fmtprintf	1869328 B	+0.0%	6.376 s	+84.8% (worse)	20.222 ms	+20.4% (worse)
macOS	println	121200 B	+0.0%	409.030 ms	-11.3% (better)	5.417 ms	+18.7% (worse)

Core language and compiler benchmarks

Platform	Benchmark	ns/op	vs base
Linux	BenchmarkLookupPCRandom	12.250 ns/op	-1.5% (better)
Linux	BenchmarkMergeCompilerFlags	146.400 ns/op	+0.3% (worse)
Linux	BenchmarkMergeLinkerFlags	94.830 ns/op	-0.5% (better)
Linux	BenchmarkChannelBuffered	37.020 ns/op	+2.0% (worse)
Linux	BenchmarkChannelHandoff	24299 ns/op	-4.1% (better)
Linux	BenchmarkDefer	52.550 ns/op	+1.6% (worse)
Linux	BenchmarkDirectCall	1.758 ns/op	-0.1% (better)
Linux	BenchmarkGlobalRead	1.757 ns/op	-0.1% (better)
Linux	BenchmarkGlobalWrite	2.809 ns/op	+0.0% (worse)
Linux	BenchmarkGoroutine	30847 ns/op	+1.0% (worse)
Linux	BenchmarkInterfaceCall	9.143 ns/op	+0.0% (worse)
Linux	BenchmarkRuntimeGetG	2.111 ns/op	-0.1% (better)
macOS	BenchmarkLookupPCRandom	18.530 ns/op	+0.4% (worse)
macOS	BenchmarkMergeCompilerFlags	220.400 ns/op	+28.5% (worse)
macOS	BenchmarkMergeLinkerFlags	109.100 ns/op	-9.7% (better)
macOS	BenchmarkChannelBuffered	24.540 ns/op	-22.3% (better)
macOS	BenchmarkChannelHandoff	7615 ns/op	-31.2% (better)
macOS	BenchmarkDefer	33.750 ns/op	-42.1% (better)
macOS	BenchmarkDirectCall	1.202 ns/op	-9.2% (better)
macOS	BenchmarkGlobalRead	1.133 ns/op	-24.6% (better)
macOS	BenchmarkGlobalWrite	1.237 ns/op	-23.8% (better)
macOS	BenchmarkGoroutine	56589 ns/op	+5.2% (worse)
macOS	BenchmarkInterfaceCall	6.229 ns/op	-5.1% (better)
macOS	BenchmarkRuntimeGetG	2.171 ns/op	-17.9% (better)

Compared with d37108c85377 measured in the same runner job.

[cpunion]

Expanded the migrated checks in 275b20266.

Each of the 25 cases now checks its complete relevant function bodies and control-flow blocks instead of only a few symbols. Platform-specific matching is limited to observed ABI differences: nest/swiftself, cgo hashes and C types, sigsetjmp/stderr names, pthread once layout, reflect method slots, and the defer save-area allocation. The debug case also retains explicit function !dbg, DICompileUnit, and DILocalVariable coverage.

All affected cases passed on macOS arm64 and Linux amd64. No out.ll or non-fixture tooling changes were reintroduced.