build: preserve main package identity for test builds

[visualfc]

Summary

• Preserve the main package identity during test builds instead of renaming it to main.test.
• Add an internal @fortest marker to distinguish test-main packages when resolving ABI paths.
• Keep c-shared test exports pointed at main.init and main.main.
• Add linkname coverage for same-package, external-test, and main-package references.
• Update the stdlib build-mode test runner for test binaries that retain the main package identity.
• Keep generated test library names package-specific, such as libtar.test.
• Link runners against main.init and main.main rather than package-path-derived test symbols.

Testing

• llgo test -v ./test/go/linkname
• LLGO=./dev/llgo.sh dev/test_std_buildmodes.sh ./test/std/archive/tar

[github-actions]

LLGo baseline benchmarks

c8bdb683fc11 | workflow run | long-term charts

Program measurements

Platform	Workload	File size	vs base	Build	vs base	Run	vs base
Linux	cprintf	19288 B	0 B / +0.0%	325.824 ms	+1.377 ms / +0.4% (worse)	1.351 ms	-13 us / -1.0% (better)
Linux	fmtprintf	1860168 B	-8 B / -0.0004301% (better)	2.696 s	+12.67 ms / +0.5% (worse)	3.558 ms	-739 ns / -0.02077% (better)
Linux	println	68776 B	0 B / +0.0%	330.526 ms	+2.662 ms / +0.8% (worse)	1.710 ms	-33.78 us / -1.9% (better)
macOS	cprintf	84672 B	0 B / +0.0%	686.106 ms	+153.6 ms / +28.8% (worse)	6.716 ms	+3.067 ms / +84.0% (worse)
macOS	fmtprintf	1892720 B	0 B / +0.0%	3.180 s	+209.9 ms / +7.1% (worse)	16.400 ms	+2.121 ms / +14.9% (worse)
macOS	println	121360 B	0 B / +0.0%	584.248 ms	+184.3 ms / +46.1% (worse)	5.179 ms	+1.626 ms / +45.8% (worse)

Core language and compiler benchmarks

Platform	Benchmark	ns/op	vs base
Linux	BenchmarkLookupPCRandom	12.260 ns/op	-0.08 ns/op / -0.6% (better)
Linux	BenchmarkMergeCompilerFlags	144.500 ns/op	0 ns/op / +0.0%
Linux	BenchmarkMergeLinkerFlags	94.200 ns/op	-0.05 ns/op / -0.1% (better)
Linux	BenchmarkChannelBuffered	36.320 ns/op	+0.02 ns/op / +0.1% (worse)
Linux	BenchmarkChannelHandoff	23833 ns/op	+823 ns/op / +3.6% (worse)
Linux	BenchmarkDefer	44.390 ns/op	+0.36 ns/op / +0.8% (worse)
Linux	BenchmarkDirectCall	1.758 ns/op	0 ns/op / +0.0%
Linux	BenchmarkGlobalRead	1.758 ns/op	-0.001 ns/op / -0.1% (better)
Linux	BenchmarkGlobalWrite	2.806 ns/op	0 ns/op / +0.0%
Linux	BenchmarkGoroutine	29868 ns/op	+482 ns/op / +1.6% (worse)
Linux	BenchmarkInterfaceCall	8.651 ns/op	+0.03 ns/op / +0.3% (worse)
Linux	BenchmarkRuntimeGetG	2.112 ns/op	0 ns/op / +0.0%
macOS	BenchmarkLookupPCRandom	16.560 ns/op	-0.17 ns/op / -1.0% (better)
macOS	BenchmarkMergeCompilerFlags	168.100 ns/op	+11.9 ns/op / +7.6% (worse)
macOS	BenchmarkMergeLinkerFlags	107.600 ns/op	+5.3 ns/op / +5.2% (worse)
macOS	BenchmarkChannelBuffered	27.470 ns/op	+2.5 ns/op / +10.0% (worse)
macOS	BenchmarkChannelHandoff	11267 ns/op	+1347 ns/op / +13.6% (worse)
macOS	BenchmarkDefer	34.580 ns/op	+4.75 ns/op / +15.9% (worse)
macOS	BenchmarkDirectCall	1.214 ns/op	+0.006 ns/op / +0.5% (worse)
macOS	BenchmarkGlobalRead	1.068 ns/op	-0.155 ns/op / -12.7% (better)
macOS	BenchmarkGlobalWrite	1.161 ns/op	-0.211 ns/op / -15.4% (better)
macOS	BenchmarkGoroutine	32112 ns/op	-20614 ns/op / -39.1% (better)
macOS	BenchmarkInterfaceCall	5.589 ns/op	-0.323 ns/op / -5.5% (better)
macOS	BenchmarkRuntimeGetG	2.529 ns/op	-0.178 ns/op / -6.6% (better)

Compared with 0314faeba69d measured in the same runner job.

[fennoai]

Review Summary

This PR reworks how //go:linkname resolves into test-compiled main packages. It replaces the previous approach (renaming the package to "main.test", which mutated the real package name) with a scope sentinel: an @ForTest marker constant is inserted into the package scope when a main package is built for test, and ssa/abi/abi.go PathOf() consults that marker to return the real import path instead of collapsing to "main".

The approach is sound and a genuine improvement over mutating Types.Name():

• The @ prefix is deliberately un-spellable as a Go identifier, so the sentinel cannot collide with, or be spoofed by, user code. No security issue.
• The added Scope().Lookup in PathOf is O(1) and guarded by Name() == "main" short-circuit, so it only runs for main packages. No performance concern.
• Removing the obsolete main.test rename (and its test) is correct hygiene; the surviving .test suffix checks operate on pkg.ID/pkg.PkgPath, not the removed rename, so they are unaffected.

Findings below are maintainability/robustness improvements, not blockers. The new test/linkname/ regression tests are a good addition.

[codecov]

Codecov Report

❌ Patch coverage is 87.50000% with 1 line in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
internal/build/build.go	85.71%	0 Missing and 1 partial ⚠️

📢 Thoughts on this report? Let us know!