ssa: use defer selectors on all targets (superseded by #2208)

[cpunion]

Superseded by #2208.

This PR explored replacing defer continuation block addresses with selectors on every target. After reviewing Native lowering and performance risk, the final design keeps Native blockaddress/indirectbr code generation unchanged and applies selector dispatch only to Wasm.

The complete Rund and Reth implementation, Native/Wasm lit coverage, runtime execution coverage, and Wasm build validation have been moved to #2208. Closing this PR keeps review and CI focused on the target-specific design.

[fennoai]

Review: generic static defer dispatch

Clean, well-motivated refactor. Replacing blockaddress/indirectbr defer dispatch with dense integer continuation selectors dispatched via switch is a net codegen improvement (enables jump-table lowering and unblocks CFG passes), the reserved-selector scheme is documented at both the definition and consumption sites, and the unreachable default is the correct pattern. I traced the rethChain/rethTargets/rundTargets index math and the selectors stay dense and consistent with the emitted switch cases — no off-by-one in the selector math. Good coverage of the tricky RunDefers-before-Defer and i32/i64 switch-width cases in the new test.

A few findings below (2 inline).

Dead code no longer reachable but left in tree

• ssa/stmt_builder.go:212 — Builder.IndirectJump (emits CreateIndirectBr) has no remaining callers after this PR removes its only uses from ssa/eh.go. Its companion aBasicBlock.Addr() (ssa/stmt_builder.go:52, emits blockaddress) also appears to have no live callers now (all other .Addr() hits resolve to net/reflect/netip). Since the PR's stated goal is to eliminate defer-generated blockaddress/indirectbr, leaving the helpers that emit them invites accidental reuse — consider removing them here (after confirming no external/generated caller depends on Addr()).

Loop-defer drainer not converted to switch dispatch (optional)

• ssa/eh.go:452-473 — the reth/rund continuations now use switch, but loopDeferDrainer still dispatches on nodeID with a hand-rolled linear chain of equality tests, emitting 2 extra basic blocks per case and an O(k) compare chain per drained node in the drain loop. The c.id values are dense integer constants, so this is exactly the shape CreateSwitch/jumpDeferTarget handles. Not a regression (this path is unchanged by the PR), but aligning it with the new dispatch strategy would be more consistent and lower IR size. Optional follow-up.

_Reviewed ssa/eh.go, runtime/internal/runtime/z_rt.go, ssa/eh_defer_test.go, and the cl/testgo CHECK updates.

[codecov]

Codecov Report

❌ Patch coverage is 73.01587% with 17 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
ssa/eh.go	73.01%	14 Missing and 3 partials ⚠️

📢 Thoughts on this report? Let us know!

[cpunion]

Follow-up on the two optional/general review notes:

• Confirmed there are no remaining in-repository callers of BasicBlock.Addr or Builder.IndirectJump. I am keeping them because they are exported general SSA-builder APIs; removing them would be a separate public API compatibility decision. This PR removes every defer-generated use.
• The loop nodeID chain is existing payload-tag dispatch rather than a continuation selector. Its unknown-ID behavior also exits the drainer rather than being unreachable. Converting that path to an LLVM switch can be evaluated separately without expanding this correctness/portability change.

[github-actions]

LLGo baseline benchmarks

2353dec74d1f | workflow run | long-term charts

Program measurements

Platform	Workload	File size	vs main	Build	vs main	Run	vs main
Linux	cprintf	18344 B	+0.0%	239.146 ms	-19.0% (better)	1.003 ms	-25.0% (better)
Linux	fmtprintf	1833064 B	+0.1% (worse)	2.588 s	-13.6% (better)	2.540 ms	-18.0% (better)
Linux	println	67720 B	+0.0%	233.233 ms	-21.0% (better)	1.149 ms	-33.7% (better)
macOS	cprintf	84672 B	+0.0%	376.967 ms	+14.2% (worse)	4.028 ms	+58.8% (worse)
macOS	fmtprintf	1861456 B	-0.4% (better)	5.024 s	+96.9% (worse)	16.572 ms	+42.4% (worse)
macOS	println	121200 B	+0.0%	445.398 ms	+41.4% (worse)	4.630 ms	+33.0% (worse)

Core language and compiler benchmarks

Platform	Benchmark	ns/op	vs main
Linux	BenchmarkLookupPCRandom	10.180 ns/op	-24.2% (better)
Linux	BenchmarkMergeCompilerFlags	123.600 ns/op	-17.8% (better)
Linux	BenchmarkMergeLinkerFlags	85.860 ns/op	-9.0% (better)
Linux	BenchmarkChannelBuffered	54.730 ns/op	+35.9% (worse)
Linux	BenchmarkChannelHandoff	27833 ns/op	+7.7% (worse)
Linux	BenchmarkDefer	39.810 ns/op	-6.5% (better)
Linux	BenchmarkDirectCall	0.971 ns/op	-37.6% (better)
Linux	BenchmarkGlobalRead	1.331 ns/op	-14.6% (better)
Linux	BenchmarkGlobalWrite	7.914 ns/op	+218.2% (worse)
Linux	BenchmarkGoroutine	35633 ns/op	+4.8% (worse)
Linux	BenchmarkInterfaceCall	6.389 ns/op	-17.9% (better)
Linux	BenchmarkRuntimeGetG	1.220 ns/op	-36.9% (better)
macOS	BenchmarkLookupPCRandom	17.450 ns/op	+62.2% (worse)
macOS	BenchmarkMergeCompilerFlags	166.400 ns/op	+48.4% (worse)
macOS	BenchmarkMergeLinkerFlags	134.700 ns/op	+98.8% (worse)
macOS	BenchmarkChannelBuffered	32.820 ns/op	-6.7% (better)
macOS	BenchmarkChannelHandoff	10041 ns/op	+116.7% (worse)
macOS	BenchmarkDefer	51.290 ns/op	+22.8% (worse)
macOS	BenchmarkDirectCall	1.408 ns/op	+16.4% (worse)
macOS	BenchmarkGlobalRead	1.200 ns/op	-9.3% (better)
macOS	BenchmarkGlobalWrite	1.442 ns/op	+2.4% (worse)
macOS	BenchmarkGoroutine	50210 ns/op	+123.1% (worse)
macOS	BenchmarkInterfaceCall	6.557 ns/op	+1.1% (worse)
macOS	BenchmarkRuntimeGetG	3.042 ns/op	+4.5% (worse)

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

[cpunion]

Superseded by #2208, which now contains the tested Wasm-only Rund/Reth selector design while preserving Native defer dispatch.