runtime: LLGo-owned frame-pointer unwinder (Stage 5)

.github/workflows/go.yml

@@ -69,12 +69,10 @@ jobs:
       - name: Build
         run: go build -v ./...
 
-      - name: Test
-        if: ${{!startsWith(matrix.os, 'macos')}}
-        run: go test -timeout 30m ./...
-
+      # Both platforms upload coverage: OS-specific paths (ELF vs Mach-O
+      # emission, per-OS runtime shims) are otherwise invisible to
+      # codecov/patch and fail it on lines only the other OS executes.
       - name: Test with coverage
-        if: startsWith(matrix.os, 'macos')
         run: go test -timeout 30m -coverprofile="coverage.txt" -covermode=atomic ./...
 
       - name: Test with embedded emulator env

benchmark/runtime_funcinfo/.gitignore

@@ -0,0 +1,2 @@
+out/
+out-*

benchmark/runtime_funcinfo/README.md

@@ -0,0 +1,62 @@
+# Runtime Funcinfo Benchmark
+
+This benchmark keeps runtime funcinfo measurements comparable across branches by
+generating the same probe programs and rebuilding them with each compiler/root
+pair in one run.
+
+It covers:
+
+- hot runtime metadata calls: `Caller`, `Callers`, `CallersFrames`,
+  `FuncForPC`, and `Func.FileLine`.
+- deep stacks through direct calls, interface calls, and closures.
+- many packages and methods, generated from configurable package/method counts.
+- cold first-use runtime metadata paths, including lazy table initialization.
+- a stdlib-heavy program with `encoding/json`, `text/template`, `regexp`,
+  `go/parser`, `go/token`, and `net/netip` imports.
+- ordinary code (`plain`): pure-compute probes (recursive `fib`, JSON
+  round-trip, `sort.Ints`, map churn) with no runtime introspection at all,
+  measuring what the funcinfo machinery costs code that never asks for it.
+
+Generated modules use `example.com/llgo-bench/...` import paths. This is
+intentional: LLGo does not enable caller-frame tracking for stdlib-shaped paths
+without a dot, and that would benchmark the fallback path instead of normal
+third-party package behavior.
+
+Example:
+
+```sh
+go run ./benchmark/runtime_funcinfo \
+  -runs=11 \
+  -llgo-opt=2 \
+  -variant go=go \
+  -variant main=llgo,/path/to/llgo-main,/path/to/llgo-main-root \
+  -variant 2002=llgo,/path/to/llgo-2002,/path/to/llgo-2002-root \
+  -variant 2009=llgo,/path/to/llgo-2009,/path/to/llgo-2009-root \
+  -variant 2010=llgo,/path/to/llgo-2010,/path/to/llgo-2010-root
+```
+
+Add `-include-lto` to build an additional `+lto` variant for every LLGo
+compiler. LLGo builds use `-O2` by default; pass `-llgo-opt=` to omit the
+optimization flag. Add `-scales=6x6,12x12,24x24` to generate separate
+`multipkg_*` and `cold_*` scenarios for several package/function counts in one
+run. Output is written to `benchmark/runtime_funcinfo/out` by default:
+
+- `summary.md`: markdown performance and size tables.
+- `results.json`: raw build and run data.
+- `work/`: generated probe modules.
+- `bin/`: generated executables.
+
+Performance cells are `best/trimmed avg` from process-level runs. The trimmed
+average drops one minimum and one maximum when at least three runs are present.
+`-iters` is a base iteration count: `hot` uses the full count, `deep` uses a
+quarter, `multipkg`/`stdlib` use one twentieth, and `plain` uses 1/2000
+because each operation does substantially more work.
+
+`multipkg.FuncForPCMany` and `multipkg.FileLineMany` are batch metrics over all
+generated target functions (`-packages * -methods`, 144 targets with the default
+settings), not single-lookup timings.
+
+`cold.First*` metrics are single measurements from a fresh process and include
+lazy runtime initialization that has not already happened in that process.
+`cold.WarmFuncForPCMany` and `cold.WarmFileLineMany` use the same batch target
+count as `multipkg`.