runtime: add memprofile attribution support

[cpunion]

Summary:

• add a lightweight LLGo runtime memprofile snapshot with synthetic frames for runtime.MemProfile
• preserve heap attribution frames by disabling inline/tail-call elimination for packages using runtime.MemProfile/MemProfileRate
• add stable test/go coverage and un-xfail GOROOT heapsampling.go

Dependency / merge order:

• Depends on runtime: add minimal MemProfile allocation records #1902, which provides the minimal MemProfile allocation-record foundation.
• Keep this PR separate and draft until runtime: add minimal MemProfile allocation records #1902 lands; after runtime: add minimal MemProfile allocation records #1902 merges, rebase this branch onto main and resolve the expected conflicts.
• Expected conflict files after runtime: add minimal MemProfile allocation records #1902 lands: runtime/internal/lib/runtime/pprof_runtime_stub_llgo.go, runtime/internal/runtime/memprofile.go, runtime/internal/runtime/z_gc.go, and test/goroot/xfail.yaml.
• Do not merge runtime: add minimal MemProfile allocation records #1902 into this PR before runtime: add minimal MemProfile allocation records #1902 is accepted unless we need to refresh CI/review context; resolving the conflict once after runtime: add minimal MemProfile allocation records #1902 lands keeps the history cleaner.

Tests:

• go test ./test/go -run TestRuntimeMemProfileAttribution -count=1
• go test ./test/goroot -run TestGoRootRunCases -count=1 -args -goroot /opt/homebrew/Cellar/go@1.24/1.24.11/libexec -dirs . -case '^heapsampling.go$' -directive-mode runlike
• go test ./cl ./ssa
• cd runtime && go test ./internal/lib/runtime ./internal/runtime
• git diff --cached --check

[cpunion]

Verification update for commit 622e1e5:\n\n- Root cause: Linux CI could not reliably recover LLGo Go function names from libunwind for memory profile stack attribution; the profile saw frames like n / __libc_start_main, so TestRuntimeMemProfileAttribution could not find main.profiledAlloc.\n- Fix: for packages that use runtime.MemProfile, the compiler now emits lightweight MemProfileEnter/Exit calls and the runtime uses that Go call stack before falling back to libunwind.\n- Local tests:\n - go test ./test/go -run TestRuntimeMemProfileAttribution -count=1 -v\n - go test ./test/go -count=1\n - go test ./ssa -count=1\n - go test ./cl -count=1\n\nPushed only to cpunion/llgo:codex/goroot-heapsampling-coverage. Monitoring the new CI run now.

[cpunion]

CI update for 622e1e5a0e3eeccd097599b6f4a6abb8639740f9:

• Previously failing test/go memprofile attribution failure is fixed by keeping a lightweight Go call stack for runtime.MemProfile test packages.
• Local verification passed:
  • go test ./test/go -run TestRuntimeMemProfileAttribution -count=1 -v
  • go test ./test/go -count=1
  • go test ./ssa -count=1
  • go test ./cl -count=1
• New CI: the three previously failing jobs now pass:
  • test (ubuntu-latest, 19)
  • test (ubuntu-latest, 19, 1.24.2, 2)
  • test (ubuntu-latest, 19, 1.26.0, 2)

At the time of this update, the remaining visible checks are queued macOS jobs; no new failure is showing on the PR.

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[cpunion]

Resolved the latest merge conflict with xgo-dev/main (51d665e). The merge keeps main's generic/linkonce and BDWGC MemStats changes while preserving this PR's memprofile attribution instrumentation and heapsampling xfail removal.\n\nLocal verification:\n- go test -timeout 20m ./test/go/memprofile -count=1\n- go test -timeout 20m ./test/go -run 'Test(RuntimeMemProfileAttribution|RuntimeReadMemStatsAfterInitGC)' -count=1\n- (runtime module) go test -timeout 20m ./internal/runtime ./internal/lib/runtime -count=1\n- git diff --check

[cpunion]

Added focused coverage for the memprofile attribution instrumentation paths.

Changes:

• Added cl tests for runtime.MemProfile / runtime.MemProfileRate AST detection, command-line-arguments frame-name normalization, and generated MemProfileEnter / MemProfileExit plus noinline / disable-tail-calls IR attributes.
• Added ssa tests for Builder.MemProfileEnter, Builder.MemProfileExit, and Function.DisableTailCalls IR emission.

Local verification:

• go test ./cl -run 'TestPackageUsesRuntimeMemProfile|TestMemProfileFunctionName|TestCompileRuntimeMemProfileInstrumentation' -count=1
• go test -coverprofile=/tmp/llgo-cl-focused.cover ./cl -run 'TestPackageUsesRuntimeMemProfile|TestMemProfileFunctionName|TestCompileRuntimeMemProfileInstrumentation' -count=1
  • covered applyNoInline, memProfileFunctionName, and packageUsesRuntimeMemProfile at 100% in the focused profile.
• go test ./ssa -run TestMemProfileBuilderAndTailCallAttribute -count=1
• go test -coverprofile=/tmp/llgo-ssa-focused.cover ./ssa -run TestMemProfileBuilderAndTailCallAttribute -count=1
  • covered DisableTailCalls, MemProfileEnter, and MemProfileExit at 100% in the focused profile.
• go test ./test/go -run TestRuntimeMemProfileAttribution -count=1
• go test ./test/go/memprofile -count=1
• go test ./ssa -count=1
• go test ./cl -count=1
• git diff --check

CI note: the previous LLGo/download-model failure is still the known workflow-side HuggingFace download issue tracked by #1945. The job failed while downloading stories15M.bin from HuggingFace with 429 Too Many Requests; I did not change workflow/model download behavior in this PR.