cl: keep pcline anchors valid under full LTO

[zhouguangyuan0718]

What changed

• associate ELF llgo_pcline records with their local PC labels instead of source function symbols
• update the IR emission test to require a local SHF_LINK_ORDER anchor
• add a Full LTO build-and-run regression fixture that keeps the affected runtime/pprof path in the link with funcinfo enabled

Why

Full LTO can inline and remove runtime/pprof.writeProfileInternal and writeRuntimeProfile while retaining body-embedded pc-line inline asm. The inline asm previously named the source function as its ELF SHF_LINK_ORDER target, so lld rejected it after that function no longer belonged to an emitted section:

linked-to symbol is not in a section

The local PC label moves with the inline asm and therefore always identifies the final emitted code section, including after Full LTO inlining.

Validation

• go test ./cl -run 'PCLine|PCLines'
• go test ./ssa -run 'PCLine|PCLines'
• go test ./internal/build -run 'PCLine|PCLines'
• go test ./internal/build/funcinfo -run 'PCLine|PCLines'
• go test ./cl -run '^TestRunAndTestFromTestlto$/testpcline$' -count=1
• go test ./cl -run 'TestCompileRuntimeCallerPCLineUsesLocalAnchor|TestRunAndTestFromTestlto/testpcline' -count=1 -coverprofile=/tmp/pr2097-cl.cover (emitPCLineLabel: 88.9%)
• manually built and ran the reproducer with LLGO_FUNCINFO_SITES=1 -lto=full, both with and without the LLGo LTO plugin

The regression fixture intentionally has no out.ll: the test harness disables LLGO_FUNCINFO while capturing an IR golden, which would bypass the metadata link path this PR needs to exercise.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!