[Based on #143] test/wasm: complete resumable ABI and debug acceptance

[cpunion]

Depends on #143.

Completes the K4 acceptance layer from xgo-dev#2152 for the opt-in LLVM-coroutine wasm backend.

What changes

• Preserve function and parameter DWARF metadata when resumable functions are lowered, and give generated calls a function-scope fallback location when no source location exists.
• Count concrete map range values and exclude function-address intrinsics from suspended-frame GC roots.
• Keep wasm caller-frame instrumentation enabled so runtime.Callers and pclntab lookup still work after suspension and resumption.
• Classify synchronous runtime map, nil-check, and caller-frame helpers as non-suspending boundaries.
• Add the wasm generic math source fallback and a no-libffi wasm compatibility layer.
• Add an end-to-end acceptance workload for indirect variadic suspension, GC roots, reflection metadata and typed method calls, C exports, pclntab, bytes/base64/regexp, and math rounding.
• Run the acceptance workload for J32, J64, and WASI P1, plus a bounded -O0 DWARF fixture for all three ABIs.

The independent K4 diff relative to #143 is 24 files, +581/-16, in five commits.

P1 size fix

Before the runtime-boundary classification, unicode.init acquired 24,772 resume points, including 23,480 RecordPanicLocation calls. Its lowered function reached about 4.25 MiB and Wasmtime rejected it as too large.

After keeping those synchronous helpers outside the suspension set, only real suspension points remain. The largest generated function is about 746 KiB and the P1 module falls from about 6.8 MiB to 3.2 MiB in the local acceptance build.

Compatibility boundaries

The wasm libffi shim allows reflection metadata, typed method values/expressions, and reflection map operations to link and run. Raw P1 still has no compatible libffi calling backend, so dynamic reflect.Value.Call and reflect.MakeFunc are not claimed by this PR.

J32 and J64 debug fixtures pass llvm-dwarfdump --verify. P1 is validated and executed, and its .debug_info and .debug_line sections are asserted. Full post-link P1 range verification remains limited by existing Binaryen/WASI DWARF range rewriting.

Local verification

All commands used GOMAXPROCS=2, GOMEMLIMIT=4GiB, and -p=1 where applicable.

• go test -p=1 ./internal/wasmresume
• go test -p=1 ./ssa
• targeted cl caller-frame and GC-root tests
• go test -p=1 ./internal/build
• go test ./internal/clite/ffi
• default non-resumable P1 -tags=nogc cross-compile/runtime regression
• J32, J64, and P1 acceptance at -O0, -O2, and -O3
• J32/J64 DWARF verification and P1 debug-section/runtime validation
• Ubuntu container: ./internal/wasmresume passed; the broader native Ubuntu matrix is left to CI because the local amd64 image runs through Rosetta under a 7.75 GiB VM cap.

Changed helper coverage is 100%; internal/wasmresume is 95.2% and ssa is 93.3%.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[github-actions]

LLGo baseline benchmarks

ee519ab23efe | workflow run | long-term charts

Program measurements

Platform	Workload	File size	vs main	Build	vs main	Run	vs main
Linux	cprintf	18816 B	+1.5% (worse)	347.917 ms	-15.6% (better)	1.272 ms	-18.1% (better)
Linux	fmtprintf	2219448 B	+0.3% (worse)	3.370 s	-12.6% (better)	2.449 ms	-25.5% (better)
Linux	println	71776 B	-1.2% (better)	343.503 ms	-14.1% (better)	1.612 ms	-3.7% (better)
macOS	cprintf	84672 B	+0.0%	330.599 ms	-20.2% (better)	2.534 ms	-29.4% (better)
macOS	fmtprintf	2361968 B	+0.7% (worse)	2.834 s	-28.1% (better)	17.440 ms	-25.4% (better)
macOS	println	125712 B	-1.0% (better)	320.541 ms	-22.7% (better)	3.173 ms	-56.3% (better)

Core language and compiler benchmarks

Platform	Benchmark	ns/op	vs main
Linux	BenchmarkLookupPCRandom	13.220 ns/op	-1.9% (better)
Linux	BenchmarkMergeCompilerFlags	151.400 ns/op	-2.0% (better)
Linux	BenchmarkMergeLinkerFlags	95.060 ns/op	-6.2% (better)
Linux	BenchmarkChannelBuffered	33.660 ns/op	-3.2% (better)
Linux	BenchmarkChannelHandoff	27261 ns/op	-14.9% (better)
Linux	BenchmarkDefer	43.330 ns/op	-23.1% (better)
Linux	BenchmarkDirectCall	1.556 ns/op	-0.4% (better)
Linux	BenchmarkGlobalRead	1.870 ns/op	+19.9% (worse)
Linux	BenchmarkGlobalWrite	2.486 ns/op	-0.3% (better)
Linux	BenchmarkGoroutine	31204 ns/op	-58.7% (better)
Linux	BenchmarkInterfaceCall	8.092 ns/op	+3.8% (worse)
Linux	BenchmarkRuntimeGetG	1.871 ns/op	-64.7% (better)
macOS	BenchmarkLookupPCRandom	11.160 ns/op	-10.3% (better)
macOS	BenchmarkMergeCompilerFlags	105.700 ns/op	-15.1% (better)
macOS	BenchmarkMergeLinkerFlags	67.180 ns/op	-3.1% (better)
macOS	BenchmarkChannelBuffered	48.390 ns/op	+74.5% (worse)
macOS	BenchmarkChannelHandoff	6690 ns/op	-4.6% (better)
macOS	BenchmarkDefer	78.790 ns/op	+110.5% (worse)
macOS	BenchmarkDirectCall	1.933 ns/op	+33.0% (worse)
macOS	BenchmarkGlobalRead	1.390 ns/op	+14.8% (worse)
macOS	BenchmarkGlobalWrite	1.957 ns/op	+31.6% (worse)
macOS	BenchmarkGoroutine	32241 ns/op	+1.8% (worse)
macOS	BenchmarkInterfaceCall	8.330 ns/op	+54.3% (worse)
macOS	BenchmarkRuntimeGetG	3.654 ns/op	+41.8% (worse)

Compared only with the latest matching platform in the main series.