[cpunion]

Purpose

This draft is the complete native-Windows integration snapshot rebased onto xgo-dev/llgo main at daef162f1a64c460d0f3681c562f645612a388e8. It supersedes the closed #169 and is retained only as a validation/reference branch while the implementation is reconstructed into reviewable PRs. It is not intended to be merged as one PR.

Related proposal: xgo-dev#2325.

The new PRs will use the final integrated behavior rather than blindly cherry-picking historical commits. In particular, commits that mix concerns or revise an earlier approach will not be picked whole. Each PR will be reconstructed from the smallest coherent final file/function/hunk set so that superseded implementations and unrelated changes do not leak across review boundaries.

Planned dependency order

1. windows: scope LLVM 19 host flags to the binding package xgo-dev/llvm#49: keep Windows LLVM 19 host flags package-local (existing external prerequisite).
2. LLGo target and portable host-tool foundation: MSVC triples, typed architecture features/cache inputs, Windows-safe metadata mapping, file locking, paths, and tool discovery.
3. PE/COFF and MSVC compiler/build layer: COMDAT and imports, MSVC C ABI, Plan 9 assembly consumption, linker/archive handling, and executable/C archive/C shared build modes.
4. Two branches can then proceed from the compiler/build layer:
   • hosted runtime substrate: Win32 threads, TLS/FLS, synchronization, stdio/OS primitives, lifecycle, and BDWGC registration;
   • Windows-host cross-compilation and embedded tooling: target/host separation, response files, native archive extraction, SDK/sysroot inputs, and ESP/WASI tooling.
5. Runtime execution ABI: SyscallN, FFI and callbacks, foreign-thread entry, faults, SEH/unwind, PC metadata, traceback, and profiling.
6. Windows standard-library integration: files/processes, signals, testing hooks, IOCP/networking, users, and OS behavior. The required goplus/lib Windows CRT/WinSock and pthread/synchronization changes will be upstreamed as their own prerequisite PRs before this layer.
7. Final validation: shared Linux/macOS/Windows CI matrices, stdlib coverage, GOROOT, demos, build cache, benchmarks, and release qualification.

Generic receiver-metadata and concurrent-output-parser fixes, if still needed, will be independent non-blocking PRs. Linux poller changes, macOS timing/xfail changes, generic converthash, broad timing relaxations, and the old large-aggregate lowering are excluded from the Windows series unless independently justified and tested.

Only the foundation prerequisite is advanced for review first. Later layers will not be submitted until their direct base has been reviewed and merged.

Rebase notes

• Replayed the integration history onto the latest upstream main and resolved the ESP download/file-lock and historical Plan 9 dependency overlaps.
• Kept the released github.com/xgo-dev/plan9asm v0.5.0; no temporary plan9asm replacement remains.
• Removed the old generic large-aggregate lowering from this snapshot because it regressed the latest main static-initialization contract and is not required by the first Windows milestone.
• Retained the existing temporary cpunion/lib and cpunion/llvm replacements only in this integration snapshot. Reviewable LLGo PRs must consume upstreamed/released dependencies at the appropriate layer.

Validation

• go test ./internal/xtool/llvm ./internal/goarch ./internal/meta ./internal/crosscompile ./internal/cabi ./internal/build
• go test ./cl -run '^TestStaticGlobalInitSkipsLargeByteSlice$' -count=1
• go test ./cl -run '^TestRunAndTestFromTestrt/any$' -count=1 -timeout=3m -v
• The first full go test ./cl exposed the stale large-aggregate interaction; after removing it, a 20-minute run reached only six LTO output mismatches caused by the local LLVM 19 warning that RISC-V +zcm/+zcz are unrecognized. The targeted LTO command produces the same warnings and failures on a clean checkout of the exact main base commit, so they are not introduced by this branch.
• Fresh CI and three-platform GOROOT validation are required for the rebased head. The earlier all-green CI and GOROOT run remain evidence for the pre-rebase integration snapshot only.

[github-actions]

LLGo baseline benchmarks

cf155a435101 | workflow run | long-term charts

Program measurements

Platform	Workload	File size	vs base	Build	vs base	Run	vs base
Linux	cprintf	19592 B	0 B / +0.0%	362.596 ms	-12.05 ms / -3.2% (better)	1.307 ms	+7.935 us / +0.6% (worse)
Linux	cprintf-lto	19424 B	0 B / +0.0%	358.651 ms	+5.696 ms / +1.6% (worse)	1.301 ms	+7.503 us / +0.6% (worse)
Linux	fmtprintf	1871960 B	0 B / +0.0%	3.013 s	-10.59 ms / -0.4% (better)	3.292 ms	-219.4 us / -6.2% (better)
Linux	fmtprintf-lto	1784488 B	0 B / +0.0%	14.840 s	-101.3 ms / -0.7% (better)	3.271 ms	+64.58 us / +2.0% (worse)
Linux	println	69192 B	0 B / +0.0%	353.034 ms	-8.51 ms / -2.4% (better)	1.632 ms	-33.44 us / -2.0% (better)
Linux	println-lto	62912 B	0 B / +0.0%	579.845 ms	+9.738 ms / +1.7% (worse)	1.674 ms	-3.628 us / -0.2% (better)
macOS	cprintf	84672 B	0 B / +0.0%	546.870 ms	-106.8 ms / -16.3% (better)	4.474 ms	-820.4 us / -15.5% (better)
macOS	cprintf-lto	100912 B	0 B / +0.0%	552.090 ms	+32.49 ms / +6.3% (worse)	3.930 ms	-525.2 us / -11.8% (better)
macOS	fmtprintf	1909424 B	0 B / +0.0%	2.433 s	-1.339 s / -35.5% (better)	11.759 ms	-2.365 ms / -16.7% (better)
macOS	fmtprintf-lto	1626272 B	0 B / +0.0%	19.760 s	-7.046 s / -26.3% (better)	5.854 ms	-4.801 ms / -45.1% (better)
macOS	println	121360 B	0 B / +0.0%	512.853 ms	-162.3 ms / -24.0% (better)	5.152 ms	-456.4 us / -8.1% (better)
macOS	println-lto	128528 B	0 B / +0.0%	604.681 ms	-415.1 ms / -40.7% (better)	4.334 ms	-2.801 ms / -39.3% (better)
Windows	cprintf	119808 B	0 B / +0.0%	798.840 ms	-65.37 ms / -7.6% (better)	3.345 ms	-106.4 us / -3.1% (better)
Windows	cprintf-lto	119296 B	0 B / +0.0%	834.330 ms	-45.45 ms / -5.2% (better)	3.360 ms	-749.2 us / -18.2% (better)
Windows	fmtprintf	1827328 B	+6144 B / +0.3% (worse)	5.400 s	+134 ms / +2.5% (worse)	8.662 ms	+266.1 us / +3.2% (worse)
Windows	fmtprintf-lto	1926144 B	+5120 B / +0.3% (worse)	17.122 s	+349.4 ms / +2.1% (worse)	8.446 ms	-572.5 us / -6.3% (better)
Windows	println	202240 B	+3584 B / +1.8% (worse)	839.672 ms	-38.95 ms / -4.4% (better)	6.767 ms	-1.238 ms / -15.5% (better)
Windows	println-lto	200704 B	+3584 B / +1.8% (worse)	1.042 s	-71.77 ms / -6.4% (better)	6.734 ms	-1.202 ms / -15.1% (better)

Core language and compiler benchmarks

Platform	Benchmark	ns/op	vs base
Linux	BenchmarkLookupPCRandom	13.140 ns/op	-0.23 ns/op / -1.7% (better)
Linux	BenchmarkMergeCompilerFlags	154.900 ns/op	+2.6 ns/op / +1.7% (worse)
Linux	BenchmarkMergeLinkerFlags	96.090 ns/op	+0.05 ns/op / +0.1% (worse)
Linux	BenchmarkChannelBuffered	34.930 ns/op	+0.02 ns/op / +0.1% (worse)
Linux	BenchmarkChannelHandoff	28261 ns/op	+786 ns/op / +2.9% (worse)
Linux	BenchmarkDefer	45.150 ns/op	+0.04 ns/op / +0.1% (worse)
Linux	BenchmarkDirectCall	1.557 ns/op	0 ns/op / +0.0%
Linux	BenchmarkGlobalRead	1.557 ns/op	-0.002 ns/op / -0.1% (better)
Linux	BenchmarkGlobalWrite	2.488 ns/op	0 ns/op / +0.0%
Linux	BenchmarkGoroutine	41122 ns/op	+8057 ns/op / +24.4% (worse)
Linux	BenchmarkInterfaceCall	7.789 ns/op	+0.001 ns/op / +0.01284% (worse)
Linux	BenchmarkRuntimeGetG	2.495 ns/op	+0.004 ns/op / +0.2% (worse)
macOS	BenchmarkLookupPCRandom	11.410 ns/op	-2.99 ns/op / -20.8% (better)
macOS	BenchmarkMergeCompilerFlags	105 ns/op	-60.5 ns/op / -36.6% (better)
macOS	BenchmarkMergeLinkerFlags	70.540 ns/op	-30.26 ns/op / -30.0% (better)
macOS	BenchmarkChannelBuffered	22.530 ns/op	-2.9 ns/op / -11.4% (better)
macOS	BenchmarkChannelHandoff	6941 ns/op	-1116 ns/op / -13.9% (better)
macOS	BenchmarkDefer	31.220 ns/op	+2.71 ns/op / +9.5% (worse)
macOS	BenchmarkDirectCall	1.087 ns/op	+0.023 ns/op / +2.2% (worse)
macOS	BenchmarkGlobalRead	1.090 ns/op	-0.374 ns/op / -25.5% (better)
macOS	BenchmarkGlobalWrite	1.229 ns/op	+0.047 ns/op / +4.0% (worse)
macOS	BenchmarkGoroutine	28098 ns/op	-23258 ns/op / -45.3% (better)
macOS	BenchmarkInterfaceCall	5.148 ns/op	+0.262 ns/op / +5.4% (worse)
macOS	BenchmarkRuntimeGetG	2.048 ns/op	-0.249 ns/op / -10.8% (better)
Windows	BenchmarkLookupPCRandom	15.970 ns/op	-0.05 ns/op / -0.3% (better)
Windows	BenchmarkMergeCompilerFlags	603.200 ns/op	-20.2 ns/op / -3.2% (better)
Windows	BenchmarkMergeLinkerFlags	523.200 ns/op	-4.1 ns/op / -0.8% (better)
Windows	BenchmarkChannelBuffered	33.640 ns/op	-0.12 ns/op / -0.4% (better)
Windows	BenchmarkChannelHandoff	789.200 ns/op	-81.3 ns/op / -9.3% (better)
Windows	BenchmarkDefer	53.350 ns/op	-0.66 ns/op / -1.2% (better)
Windows	BenchmarkDirectCall	1.550 ns/op	+0.001 ns/op / +0.1% (worse)
Windows	BenchmarkGlobalRead	1.548 ns/op	-0.015 ns/op / -1.0% (better)
Windows	BenchmarkGlobalWrite	2.469 ns/op	-0.006 ns/op / -0.2% (better)
Windows	BenchmarkGoroutine	89582 ns/op	-7125 ns/op / -7.4% (better)
Windows	BenchmarkInterfaceCall	9.037 ns/op	-0.045 ns/op / -0.5% (better)
Windows	BenchmarkRuntimeGetG	1.863 ns/op	-0.312 ns/op / -14.3% (better)

Compared with 7e11371113c8 measured in the same runner job.