fix(pclntab): preserve Darwin line sites with DWARF

[cpunion]

Problem

Darwin builds with DWARF currently disable every compiler-emitted funcinfo site record. That avoids an LLDB lexical-scope regression from inline assembly at function entry, but also removes final-PC anchors used for precise runtime line tables. runtime.Caller, runtime.CallersFrames, logging prefixes, and panic locations can then fall back to an earlier address.

Function-entry address records and PC-line anchors have different debugger effects: entry records alter the initial line-table row, while a PC-line anchor is safe when its inline-assembly call has no debug location.

Changes

• add ssa.Builder.InlineAsmNoDebug, clearing only the emitted call location while preserving the builder location;
• emit PC-line anchors whenever funcinfo sites are enabled, including Darwin DWARF builds;
• split runtime policy between PC-line and function-entry address records;
• suppress only function-entry address records for embedded pclntab + Darwin + DWARF;
• retain address records for Darwin without DWARF, external pclntab, and ELF targets;
• add explicit -ldflags=-w=false execution coverage for runtime.Caller, runtime.CallersFrames, and log.Lshortfile;
• clarify the global site gate and use explicit PC-line naming.

This does not change pclntab mode selection, default -w, runtime table format, or LLDB language behavior.

Fixes #2115.

Dependency update

#2143 and #2215 are merged. This branch is rebased directly onto current main.

Verification

• go test ./internal/build ./ssa ./cl;
• go test ./test/go -run 'TestDWARF.*PCLN|TestPCLN.*DWARF' -count=1;
• the complete go test ./... passes on the dependent [Based on #2157] build: match cmd/link DWARF defaults #2142 stack;
• git diff --check.

[gemini-code-assist]

Caution

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

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[fennoai]

FennoAI Review — PR #2157 (DWARF/pcln runtime line sites)

Overall a careful, well-tested change. The moveInstr → moveInstrsAfter rewrite preserves stable relative order, DebugRefs correctly move with the values they describe (and are excluded from the moveWouldBreakSSA value set since *ssa.DebugRef is not an ssa.Value), the nonDebugReferrers conversion faithfully preserves the prior nil/length/identity checks, and the comment rewrites match the new gating. No correctness, security, or memory-safety defects found. A few items worth confirming, noted inline plus below.

Behavioral change to confirm (highest-impact item in this PR):

• internal/build/build.go:532 — passOpt is now shouldRunLLVMPasses(mode) (mode != ModeGen). The previous if emitDebugInfo || mode == ModeGen { passOpt = false } guard is gone, so debug/DWARF builds now run the full LLVM pass pipeline (RunPasses + VerifyModule) where they previously ran none. Because emitDebugInfo and OptLevel are independent, a normal DWARF-producing llgo build still defaults to O2, so the common debug path moves from no passes to full default<O2> per package. This looks intentional (it pairs with the new SetDebugInfoOptimized(passOpt && OptLevel != optlevel.O0) at build.go:540 and the "precise DWARF under optimization" goal), but it is a real per-package compile-time increase on the default path and is not obvious from the diff — please confirm it is the intended tradeoff. If a fast no-pass debug path at -O0 is still desired, note default<O0> is still invoked (verify + minimal passes) rather than skipped.

Stale doc reference (adjacent to edited code):

• internal/build/build.go:1195 — the rewritePrebuiltFuncTab comment (whose body you edit just below) references doc/design/pclntab-linkphase.md, which does not exist in the repo. The same dangling reference also appears in chore/pclnpost/main.go:19, internal/pclnpost/binary.go:18, and doc/design/stage5-handoff.md. Consider restoring the doc or dropping the references.

Remaining inline notes are minor/latent.

[cpunion]

Review follow-up after rebasing onto merged #2143:

• the DWARF LLVM pass pipeline is intentional and now belongs to main;
• doc/design/pclntab-linkphase.md is present on current main, so the previously dangling references are valid;
• PC-line/address-site naming and the global gate wording were updated;
• the moveInstrsAfter precondition guard and tests are inherited from final debug: keep optimization and SSA rewrites DWARF-safe #2143.

Focused tests and the dependent-stack full test suite pass.

[cpunion]

CI follow-up: the Go workflow and LLGo compatibility shard had the same compile failure: the new acceptance test referenced the obsolete private helper findStringConversionRepoRoot. Commit b63aae8e1 now reuses the shared findRepoRoot helper.

Verified with:

• go test -vet=off ./test/go -run '^TestDWARFPCLNLineSites$' -count=1
• complete internal/build coverage run (including the branches previously shown as missing by Codecov).

[github-actions]

LLGo baseline benchmarks

b63aae8e1408 | workflow run | long-term charts

Program measurements

Platform	Workload	File size	vs base	Build	vs base	Run	vs base
Linux	cprintf	18456 B	+0.0%	298.838 ms	-1.9% (better)	1.381 ms	+2.3% (worse)
Linux	fmtprintf	1829968 B	+0.0%	3.001 s	+0.1% (worse)	3.539 ms	+2.3% (worse)
Linux	println	68008 B	+0.0%	294.275 ms	-1.2% (better)	1.766 ms	+2.0% (worse)
macOS	cprintf	84672 B	+0.0%	558.540 ms	+12.4% (worse)	5.614 ms	-3.7% (better)
macOS	fmtprintf	1869328 B	+0.0%	3.523 s	-27.8% (better)	14.911 ms	-1.7% (better)
macOS	println	121200 B	+0.0%	483.365 ms	+39.1% (worse)	4.999 ms	+28.6% (worse)

Core language and compiler benchmarks

Platform	Benchmark	ns/op	vs base
Linux	BenchmarkLookupPCRandom	12.250 ns/op	+0.1% (worse)
Linux	BenchmarkMergeCompilerFlags	143.600 ns/op	-0.5% (better)
Linux	BenchmarkMergeLinkerFlags	94.420 ns/op	+0.4% (worse)
Linux	BenchmarkChannelBuffered	37.080 ns/op	-0.0% (better)
Linux	BenchmarkChannelHandoff	23723 ns/op	-0.1% (better)
Linux	BenchmarkDefer	49.450 ns/op	-0.9% (better)
Linux	BenchmarkDirectCall	1.758 ns/op	+0.1% (worse)
Linux	BenchmarkGlobalRead	1.758 ns/op	+0.0%
Linux	BenchmarkGlobalWrite	2.810 ns/op	+0.1% (worse)
Linux	BenchmarkGoroutine	30226 ns/op	+1.5% (worse)
Linux	BenchmarkInterfaceCall	9.154 ns/op	+0.1% (worse)
Linux	BenchmarkRuntimeGetG	2.111 ns/op	-0.0% (better)
macOS	BenchmarkLookupPCRandom	13.430 ns/op	-10.9% (better)
macOS	BenchmarkMergeCompilerFlags	128.200 ns/op	-17.3% (better)
macOS	BenchmarkMergeLinkerFlags	77.970 ns/op	-19.8% (better)
macOS	BenchmarkChannelBuffered	28.460 ns/op	+20.4% (worse)
macOS	BenchmarkChannelHandoff	9038 ns/op	+27.2% (worse)
macOS	BenchmarkDefer	40.150 ns/op	+36.3% (worse)
macOS	BenchmarkDirectCall	1.201 ns/op	+15.1% (worse)
macOS	BenchmarkGlobalRead	1.204 ns/op	+15.4% (worse)
macOS	BenchmarkGlobalWrite	1.199 ns/op	+12.6% (worse)
macOS	BenchmarkGoroutine	61170 ns/op	+26.3% (worse)
macOS	BenchmarkInterfaceCall	5.389 ns/op	+20.0% (worse)
macOS	BenchmarkRuntimeGetG	2.767 ns/op	+9.6% (worse)

Compared with 6670dae3884d measured in the same runner job.