runtime: preserve type metadata in assertion panics

[cpunion]

Type assertion failures currently lower the asserted type to a string before calling the runtime. That loses package identity, the source interface type, and instantiated generic type information.

Pass ABI type descriptors for the source interface and asserted type instead, and construct the existing Go-compatible TypeAssertionError directly. Empty source interfaces remain nil so ordinary interface-conversion messages keep their Go spelling.

This removes the xfails for:

• fixedbugs/issue18911.go
• interface/embed3.go
• typeparam/mdempsky/16.go
• fixedbugs/issue26094.go

The scoped same-name type regression is covered in TestInterfaceAssertRejectsSameNameTypesFromDifferentScopes; it checks both local-to-local and local-to-package assertions.

Validation:

• go test ./ssa ./test/goroot -count=1 -timeout=15m
• affected 14 cl runtime/IR fixtures
• focused llgo test assertion regressions
• the first three GOROOT cases with Go 1.24.11, 1.25.0, and 1.26.5 on darwin/arm64
• fixedbugs/issue26094.go with empty xfail on all three Go versions on darwin/arm64

[github-actions]

LLGo baseline benchmarks

6477f8e80fbb | workflow run | long-term charts

Program measurements

Platform	Workload	File size	vs main	Build	vs main	Run	vs main
Linux	cprintf	18512 B	new	372.222 ms	new	1.340 ms	new
Linux	fmtprintf	2215624 B	new	3.537 s	new	2.537 ms	new
Linux	println	71472 B	new	370.604 ms	new	1.597 ms	new
macOS	cprintf	84672 B	new	509.855 ms	new	4.079 ms	new
macOS	fmtprintf	2344880 B	new	5.793 s	new	32.580 ms	new
macOS	println	125712 B	new	377.698 ms	new	4.136 ms	new

Core language and compiler benchmarks

Platform	Benchmark	ns/op	vs main
Linux	BenchmarkLookupPCRandom	13.310 ns/op	new
Linux	BenchmarkMergeCompilerFlags	154.700 ns/op	new
Linux	BenchmarkMergeLinkerFlags	94.960 ns/op	new
Linux	BenchmarkChannelBuffered	34.550 ns/op	new
Linux	BenchmarkChannelHandoff	25694 ns/op	new
Linux	BenchmarkDefer	46.440 ns/op	new
Linux	BenchmarkDirectCall	1.556 ns/op	new
Linux	BenchmarkGlobalRead	1.559 ns/op	new
Linux	BenchmarkGlobalWrite	2.486 ns/op	new
Linux	BenchmarkGoroutine	33071 ns/op	new
Linux	BenchmarkInterfaceCall	7.787 ns/op	new
Linux	BenchmarkRuntimeGetG	2.809 ns/op	new
macOS	BenchmarkLookupPCRandom	17.010 ns/op	new
macOS	BenchmarkMergeCompilerFlags	183.200 ns/op	new
macOS	BenchmarkMergeLinkerFlags	118.900 ns/op	new
macOS	BenchmarkChannelBuffered	32.030 ns/op	new
macOS	BenchmarkChannelHandoff	7869 ns/op	new
macOS	BenchmarkDefer	50.960 ns/op	new
macOS	BenchmarkDirectCall	1.495 ns/op	new
macOS	BenchmarkGlobalRead	1.358 ns/op	new
macOS	BenchmarkGlobalWrite	2.151 ns/op	new
macOS	BenchmarkGoroutine	43872 ns/op	new
macOS	BenchmarkInterfaceCall	7.314 ns/op	new
macOS	BenchmarkRuntimeGetG	3.016 ns/op	new

No main baseline exists yet; all metrics are marked new.

Warning

• The rendered benchmark data could not be pushed.

[fennoai]

Review summary

Reviewed the type-assertion panic changes against the PR's actual diff (head 959307217 vs its parent). Four review passes (code quality, performance, security, docs) were run; findings that turned out to target unrelated code already on main (the RewriteMainPrefix removal) were discarded as out of scope for this PR.

No blocking issues found. This is a clean, well-scoped change.

What the change does and why it holds up:

• runtime/internal/runtime/errors.go: PanicTypeAssert now takes (source, concrete, asserted *_type, missingMethod) and constructs a TypeAssertionError, delegating all message formatting to the existing Error() method instead of building strings inline. The dead helpers sameTypeAssertName/typeNameSuffix are removed with no remaining references (hasPrefix is still used elsewhere). Error() already distinguishes "different packages" vs "different scopes" and handles the nil-concrete and missing-method cases, so behavior is preserved and made more gc-accurate.
• ssa/interface.go: the failure branch now passes the source interface's abi type — real pointer for non-empty interfaces (NumMethods() > 0), nil for empty interfaces. The source computation is emitted only into the panic block (blks[1]), so the successful-assertion hot path is unaffected; the single TypeAssertionError heap allocation is confined to the cold, terminal panic path.
• The regenerated FileCheck goldens are consistent with the new call shape: any assertions emit ptr null for source, non-empty-interface assertions emit a real type pointer (e.g. cl/_testrt/any/in.go, cl/_testgo/tpinst/main.go).
• test/go/interface_type_assert_panic_test.go: TestGenericNilInterfaceAssertMentionsSourceType correctly exercises the new source-type-in-message behavior for a nil non-empty interface.

Non-blocking observations (no action required):

• TypeAssertionError.Error() dereferences e.asserted.String() without a nil guard, unlike the _interface/concrete fields. This line is pre-existing (not changed by this PR) and every construction site (ssa/interface.go, z_face.go) passes a non-nil asserted, so it is safe today — noting only as a latent invariant.
• The removed test/goroot/xfail.yaml entries (typeparam/mdempsky/16.go, fixedbugs/issue18911.go, interface/embed3.go) match what this change fixes. The embed3.go case depends on _type.String()/pkgpath() output, which this PR does not modify; worth confirming those three goroot cases are green in CI before merge.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!