windows: qualify ARM64 and 386 MSVC targets (R10; depends on plan9asm#23)

[cpunion]

Depends on xgo-dev/plan9asm#23.

Part of xgo-dev#2325.

This draft qualifies native Windows ARM64 and WoW64 386 MSVC targets while keeping the LLGo compiler, LLVM, LLDB, pkg-config, and other host tools on x64. GOOS=windows GOARCH=... remains authoritative: LLGo selects the canonical target triple and the architecture-matched Visual Studio, vcpkg, compiler-rt, Python, runtime/FFI, language, demo, benchmark, and scheduled GOROOT coverage. Explicit compiler and linker commands remain architecture-checked.

The implementation keeps three concerns separate:

• x64 host tools are built before target activation;
• target SDK and library selection follows GOARCH for amd64, arm64, and 386;
• pure IR generation does not require a physical target linker, while native links do.

Windows/386 execution exposed and fixed target-specific compatibility gaps:

• Go/386 structs and multi-result tuples use four-byte alignment for int64 and float64; Go callers and callees now share the same physical LLVM layout.
• Native MSVC C structs retain their eight-byte alignment. C ABI lowering recursively bridges divergent Go and native aggregate layouts for parameters, results, exports, and callbacks.
• Wide signed and unsigned indexes are checked before narrowing to 32-bit GEP indexes and use Go's standard high/low-word panic helpers.
• Float-to-integer conversion follows the official Go 386 runtime semantics. Wide slice/channel lengths and capacities are range-checked before narrowing, including mixed signed/unsigned operand widths.
• Pointer printing widens 32-bit values for the shared native formatting boundary.
• The syscall bridge is marked SafeSEH-compatible, and function-entry/PC-line records use an explicit Win32 layout without conflating it with Go's 12-byte 386 struct layout.
• Win32 stack walking skips LLGo wrapper frames, preserves exact function-entry records, and keeps Win64 entry slack isolated from 386.
• Windows/386 CPU profiling now captures interrupted PCs and frame chains through the Win32 CONTEXT layout while retaining the existing Go runtime/pprof API.

Go 1.27 additionally moved the Windows ARM64 processor-feature query behind internal/cpu.isProcessorFeaturePresent. LLGo now supplies the same runtime hook as the official implementation through a thin Win32 IsProcessorFeaturePresent call, without adding a C wrapper.

The initial branch replaced five groups of Go 386 assembly with LLGo source patches. Those patches are no longer present. With plan9asm #23 (pinned to its current reviewed head), LLGo compiles the official Go internal/bytealg, internal/chacha8rand, internal/runtime/atomic, internal/runtime/syscall/windows, and math assembly. Only register-ABI metadata for two local internal/bytealg helpers remains LLGo-specific.

CI changes preserve main's current Go 1.20–1.27 coverage and add ARM64/386 rows to the existing Windows matrices. The compiler is always built with the current toolchain before activating a target architecture. Benchmark jobs compare base and head on the same runner; scheduled GOROOT reporting includes Windows MSVC amd64, ARM64, 386, and MinGW amd64 without multiplying old-toolchain Windows lanes.

Local validation on the rebased revision:

• macOS/ARM64: go test ./ssa ./internal/cabi, go test ./internal/plan9asm with the plan9asm workspace, and go test -timeout=20m ./cl (621.6 s)
• macOS/ARM64: all internal/build tests; the workspace-discovery test was also rerun with GOWORK unset as required by that test
• Windows/386 MSVC: complete test/go execution with Go 1.27 and plan9asm feat(coro): add runnable stackless scheduler and panic prototype #23 current head
• Windows/386 MSVC: runtime/pprof CPU start/stop, duplicate-start, goroutine sampling, and fault-recovery tests
• Windows/ARM64 MSVC: runtime line/statement metadata, concurrent first-use function lookup, deferred-panic source line, and logical runtime-tail tests
• Windows/386 MSVC: standard-library hash/crc32, math/bits, math/rand/v2, and full runtime/pprof suites before the final rebase; the affected profiling cases were rerun after it
• Windows/386 and ARM64 outputs were identified as IMAGE_FILE_MACHINE_I386 and IMAGE_FILE_MACHINE_ARM64 and executed in the Windows 11 ARM64 VM

The fresh GitHub Actions matrix, Codecov result, and architecture-specific benchmark comparison are required before this PR is marked ready for review.

[github-actions]

LLGo baseline benchmarks

8f9cd1de92d9 | 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	19560 B	+64 B / +0.3% (worse)	387 B	0 B / +0.0%	347.796 ms	+24.77 ms / +7.7% (worse)	1.263 ms	-13.94 us / -1.1% (better)
Linux	cprintf-lto	19392 B	+64 B / +0.3% (worse)	368 B	0 B / +0.0%	342.562 ms	+9.716 ms / +2.9% (worse)	1.265 ms	+16.19 us / +1.3% (worse)
Linux	fmtprintf	1608368 B	+368 B / +0.02289% (worse)	490042 B	0 B / +0.0%	2.468 s	+23.04 ms / +0.9% (worse)	3.171 ms	+101.6 us / +3.3% (worse)
Linux	fmtprintf-lto	1486096 B	+352 B / +0.02369% (worse)	450888 B	0 B / +0.0%	8.171 s	+29.89 ms / +0.4% (worse)	2.944 ms	-74.32 us / -2.5% (better)
Linux	println	62408 B	+64 B / +0.1% (worse)	15286 B	0 B / +0.0%	335.375 ms	+4.542 ms / +1.4% (worse)	1.573 ms	+33.1 us / +2.1% (worse)
Linux	println-lto	54224 B	+64 B / +0.1% (worse)	12914 B	0 B / +0.0%	507.411 ms	+3.759 ms / +0.7% (worse)	1.574 ms	+13 us / +0.8% (worse)
macOS	cprintf	84480 B	0 B / +0.0%	16797 B	+64 B / +0.4% (worse)	764.998 ms	+248.5 ms / +48.1% (worse)	6.979 ms	+3.278 ms / +88.5% (worse)
macOS	cprintf-lto	100704 B	0 B / +0.0%	16777 B	+64 B / +0.4% (worse)	780.817 ms	+189.5 ms / +32.1% (worse)	4.754 ms	+242.3 us / +5.4% (worse)
macOS	fmtprintf	1470304 B	0 B / +0.0%	867568 B	+368 B / +0.04244% (worse)	2.932 s	+205.6 ms / +7.5% (worse)	7.951 ms	+1.317 ms / +19.8% (worse)
macOS	fmtprintf-lto	1175536 B	0 B / +0.0%	863272 B	+356 B / +0.04126% (worse)	8.947 s	+1.134 s / +14.5% (worse)	8.561 ms	-750 us / -8.1% (better)
macOS	println	114784 B	0 B / +0.0%	35245 B	+64 B / +0.2% (worse)	759.008 ms	+288 ms / +61.1% (worse)	6.021 ms	+716.5 us / +13.5% (worse)
macOS	println-lto	118656 B	0 B / +0.0%	32897 B	+64 B / +0.2% (worse)	949.349 ms	+202.5 ms / +27.1% (worse)	7.648 ms	+3.369 ms / +78.7% (worse)
Windows MinGW	cprintf	20480 B	0 B / +0.0%	4662 B	0 B / +0.0%	845.851 ms	+31.65 ms / +3.9% (worse)	4.654 ms	+1.178 ms / +33.9% (worse)
Windows MinGW	cprintf-lto	18432 B	0 B / +0.0%	4582 B	0 B / +0.0%	858.733 ms	-19.16 ms / -2.2% (better)	3.536 ms	+62.8 us / +1.8% (worse)
Windows MinGW	fmtprintf	1937408 B	+512 B / +0.02643% (worse)	597398 B	0 B / +0.0%	3.530 s	-47.38 ms / -1.3% (better)	8.560 ms	-241.6 us / -2.7% (better)
Windows MinGW	fmtprintf-lto	1991168 B	+512 B / +0.02572% (worse)	590870 B	0 B / +0.0%	9.352 s	-47.22 ms / -0.5% (better)	8.380 ms	-195 us / -2.3% (better)
Windows MinGW	println	74240 B	0 B / +0.0%	25062 B	0 B / +0.0%	847.842 ms	+23.08 ms / +2.8% (worse)	7.052 ms	-348.1 us / -4.7% (better)
Windows MinGW	println-lto	67584 B	0 B / +0.0%	21734 B	0 B / +0.0%	1.057 s	+24.89 ms / +2.4% (worse)	7.005 ms	+748.4 us / +12.0% (worse)
Windows MSVC	cprintf	12288 B	0 B / +0.0%	4438 B	0 B / +0.0%	771.860 ms	-153.9 ms / -16.6% (better)	3.514 ms	-3.225 ms / -47.9% (better)
Windows MSVC	cprintf-lto	11776 B	0 B / +0.0%	4278 B	0 B / +0.0%	939.393 ms	+158.1 ms / +20.2% (worse)	3.543 ms	+40.4 us / +1.2% (worse)
Windows MSVC	fmtprintf	1476096 B	+512 B / +0.0347% (worse)	596982 B	0 B / +0.0%	3.563 s	+37.84 ms / +1.1% (worse)	10.105 ms	+1.312 ms / +14.9% (worse)
Windows MSVC	fmtprintf-lto	1523712 B	+512 B / +0.03361% (worse)	597782 B	0 B / +0.0%	10.832 s	-53.87 ms / -0.5% (better)	10.607 ms	+532.7 us / +5.3% (worse)
Windows MSVC	println	47104 B	0 B / +0.0%	25062 B	0 B / +0.0%	769.249 ms	-5.598 ms / -0.7% (better)	6.976 ms	-82.1 us / -1.2% (better)
Windows MSVC	println-lto	44032 B	0 B / +0.0%	22102 B	0 B / +0.0%	1.030 s	-2.109 ms / -0.2% (better)	6.976 ms	+254.8 us / +3.8% (worse)
Windows MSVC 386	cprintf	9728 B	new	3930 B	new	868.747 ms	new	6.136 ms	new
Windows MSVC 386	cprintf-lto	9216 B	new	3840 B	new	999.460 ms	new	7.744 ms	new
Windows MSVC 386	fmtprintf	1195520 B	new	456560 B	new	3.736 s	new	11.867 ms	new
Windows MSVC 386	fmtprintf-lto	1259520 B	new	454021 B	new	10.311 s	new	11.538 ms	new
Windows MSVC 386	println	35328 B	new	19744 B	new	844.947 ms	new	9.005 ms	new
Windows MSVC 386	println-lto	33280 B	new	17873 B	new	1.094 s	new	9.739 ms	new
Windows MSVC ARM64	cprintf	11264 B	new	3976 B	new	1.825 s	new	8.479 ms	new
Windows MSVC ARM64	cprintf-lto	10752 B	new	3844 B	new	1.846 s	new	8.079 ms	new
Windows MSVC ARM64	fmtprintf	1370112 B	new	510252 B	new	6.083 s	new	17.290 ms	new
Windows MSVC ARM64	fmtprintf-lto	1419776 B	new	508364 B	new	18.853 s	new	18.219 ms	new
Windows MSVC ARM64	println	43008 B	new	22856 B	new	1.777 s	new	14.086 ms	new
Windows MSVC ARM64	println-lto	40960 B	new	20656 B	new	2.253 s	new	14.031 ms	new

Core language and compiler benchmarks

Platform	Benchmark	ns/op	vs base
Linux	BenchmarkLookupPCRandom	13.380 ns/op	+0.06 ns/op / +0.5% (worse)
Linux	BenchmarkMergeCompilerFlags	140.200 ns/op	-0.1 ns/op / -0.1% (better)
Linux	BenchmarkMergeLinkerFlags	88.750 ns/op	-0.08 ns/op / -0.1% (better)
Linux	BenchmarkChannelBuffered	35.200 ns/op	-0.34 ns/op / -1.0% (better)
Linux	BenchmarkChannelHandoff	27410 ns/op	+857 ns/op / +3.2% (worse)
Linux	BenchmarkDefer	46.040 ns/op	-0.54 ns/op / -1.2% (better)
Linux	BenchmarkDirectCall	1.556 ns/op	0 ns/op / +0.0%
Linux	BenchmarkGlobalRead	1.869 ns/op	+0.312 ns/op / +20.0% (worse)
Linux	BenchmarkGlobalWrite	2.488 ns/op	+0.003 ns/op / +0.1% (worse)
Linux	BenchmarkGoroutine	31598 ns/op	-277 ns/op / -0.9% (better)
Linux	BenchmarkInterfaceCall	8.719 ns/op	+0.63 ns/op / +7.8% (worse)
Linux	BenchmarkRuntimeGetG	2.491 ns/op	+0.001 ns/op / +0.04016% (worse)
macOS	BenchmarkLookupPCRandom	15.090 ns/op	+2.15 ns/op / +16.6% (worse)
macOS	BenchmarkMergeCompilerFlags	125.400 ns/op	-12.7 ns/op / -9.2% (better)
macOS	BenchmarkMergeLinkerFlags	78.790 ns/op	+4.7 ns/op / +6.3% (worse)
macOS	BenchmarkChannelBuffered	26.300 ns/op	-9.64 ns/op / -26.8% (better)
macOS	BenchmarkChannelHandoff	8003 ns/op	-2927 ns/op / -26.8% (better)
macOS	BenchmarkDefer	36.970 ns/op	-11.14 ns/op / -23.2% (better)
macOS	BenchmarkDirectCall	1.158 ns/op	-0.252 ns/op / -17.9% (better)
macOS	BenchmarkGlobalRead	1.085 ns/op	-0.221 ns/op / -16.9% (better)
macOS	BenchmarkGlobalWrite	1.060 ns/op	-0.388 ns/op / -26.8% (better)
macOS	BenchmarkGoroutine	45343 ns/op	+11613 ns/op / +34.4% (worse)
macOS	BenchmarkInterfaceCall	4.438 ns/op	-1.089 ns/op / -19.7% (better)
macOS	BenchmarkRuntimeGetG	2.369 ns/op	-0.39 ns/op / -14.1% (better)
Windows MinGW	BenchmarkLookupPCRandom	13.140 ns/op	+0.09 ns/op / +0.7% (worse)
Windows MinGW	BenchmarkMergeCompilerFlags	665.100 ns/op	+49.3 ns/op / +8.0% (worse)
Windows MinGW	BenchmarkMergeLinkerFlags	579.700 ns/op	+21 ns/op / +3.8% (worse)
Windows MinGW	BenchmarkChannelBuffered	34.650 ns/op	+0.17 ns/op / +0.5% (worse)
Windows MinGW	BenchmarkChannelHandoff	910.200 ns/op	+47.6 ns/op / +5.5% (worse)
Windows MinGW	BenchmarkDefer	58.540 ns/op	-0.08 ns/op / -0.1% (better)
Windows MinGW	BenchmarkDirectCall	1.554 ns/op	+0.008 ns/op / +0.5% (worse)
Windows MinGW	BenchmarkGlobalRead	1.549 ns/op	0 ns/op / +0.0%
Windows MinGW	BenchmarkGlobalWrite	2.471 ns/op	+0.002 ns/op / +0.1% (worse)
Windows MinGW	BenchmarkGoroutine	91380 ns/op	+4158 ns/op / +4.8% (worse)
Windows MinGW	BenchmarkInterfaceCall	9.302 ns/op	+0.002 ns/op / +0.02151% (worse)
Windows MinGW	BenchmarkRuntimeGetG	2.501 ns/op	+0.024 ns/op / +1.0% (worse)
Windows MSVC	BenchmarkLookupPCRandom	13.140 ns/op	-0.11 ns/op / -0.8% (better)
Windows MSVC	BenchmarkMergeCompilerFlags	611.800 ns/op	-4.6 ns/op / -0.7% (better)
Windows MSVC	BenchmarkMergeLinkerFlags	532.800 ns/op	-3.5 ns/op / -0.7% (better)
Windows MSVC	BenchmarkChannelBuffered	34.910 ns/op	+0.02 ns/op / +0.1% (worse)
Windows MSVC	BenchmarkChannelHandoff	1030 ns/op	-64 ns/op / -5.9% (better)
Windows MSVC	BenchmarkDefer	55.750 ns/op	-0.64 ns/op / -1.1% (better)
Windows MSVC	BenchmarkDirectCall	1.862 ns/op	+0.001 ns/op / +0.1% (worse)
Windows MSVC	BenchmarkGlobalRead	1.857 ns/op	-0.004 ns/op / -0.2% (better)
Windows MSVC	BenchmarkGlobalWrite	2.452 ns/op	+0.001 ns/op / +0.0408% (worse)
Windows MSVC	BenchmarkGoroutine	86460 ns/op	-2401 ns/op / -2.7% (better)
Windows MSVC	BenchmarkInterfaceCall	8.986 ns/op	-0.011 ns/op / -0.1% (better)
Windows MSVC	BenchmarkRuntimeGetG	2.171 ns/op	+0.002 ns/op / +0.1% (worse)
Windows MSVC 386	BenchmarkLookupPCRandom	26.730 ns/op	new
Windows MSVC 386	BenchmarkMergeCompilerFlags	765.100 ns/op	new
Windows MSVC 386	BenchmarkMergeLinkerFlags	696.500 ns/op	new
Windows MSVC 386	BenchmarkChannelBuffered	43.420 ns/op	new
Windows MSVC 386	BenchmarkChannelHandoff	925.600 ns/op	new
Windows MSVC 386	BenchmarkDefer	49.400 ns/op	new
Windows MSVC 386	BenchmarkDirectCall	1.548 ns/op	new
Windows MSVC 386	BenchmarkGlobalRead	1.548 ns/op	new
Windows MSVC 386	BenchmarkGlobalWrite	7.771 ns/op	new
Windows MSVC 386	BenchmarkGoroutine	93396 ns/op	new
Windows MSVC 386	BenchmarkInterfaceCall	9.946 ns/op	new
Windows MSVC 386	BenchmarkRuntimeGetG	2.168 ns/op	new
Windows MSVC ARM64	BenchmarkLookupPCRandom	12.080 ns/op	new
Windows MSVC ARM64	BenchmarkMergeCompilerFlags	565.200 ns/op	new
Windows MSVC ARM64	BenchmarkMergeLinkerFlags	540.300 ns/op	new
Windows MSVC ARM64	BenchmarkChannelBuffered	43.810 ns/op	new
Windows MSVC ARM64	BenchmarkChannelHandoff	3033 ns/op	new
Windows MSVC ARM64	BenchmarkDefer	70.280 ns/op	new
Windows MSVC ARM64	BenchmarkDirectCall	0.590 ns/op	new
Windows MSVC ARM64	BenchmarkGlobalRead	0.664 ns/op	new
Windows MSVC ARM64	BenchmarkGlobalWrite	3.756 ns/op	new
Windows MSVC ARM64	BenchmarkGoroutine	76819 ns/op	new
Windows MSVC ARM64	BenchmarkInterfaceCall	4.718 ns/op	new
Windows MSVC ARM64	BenchmarkRuntimeGetG	1.801 ns/op	new

Compared with e637682c8772 measured in the same runner job. Platforms without a paired baseline are marked new.

[cpunion]

Replaced by xgo-dev#2454 after rebasing R10 onto current main and integrating the released plan9asm v0.5.1 x87 configuration.