runtime,cl: add Go-compatible sampled memory profiling

.github/workflows/benchmark-publish.yml

@@ -14,7 +14,7 @@ permissions:
 jobs:
   publish:
     if: github.event.workflow_run.conclusion == 'success'
-    uses: xgo-dev/setup-benchmark-go-action/.github/workflows/publish.yml@v1.0.4
+    uses: xgo-dev/setup-benchmark-go-action/.github/workflows/publish.yml@v1.0.5
     with:
       run_id: ${{ github.event.workflow_run.id }}
       config_path: .github/llgo-benchmark.yml

.github/workflows/benchmark.yml

@@ -25,7 +25,7 @@ jobs:
             id: macos
             display: macOS
     runs-on: ${{ matrix.os }}
-    timeout-minutes: 20
+    timeout-minutes: 30
     env:
       GOMAXPROCS: "2"
       LLGO_ROOT: ${{ github.workspace }}
@@ -38,7 +38,16 @@ jobs:
         with:
           repository: ${{ github.event.pull_request.base.repo.full_name }}
           ref: ${{ github.event.pull_request.base.sha }}
-          path: .benchmark/source
+          path: .benchmark/base-source
+          persist-credentials: false
+
+      - name: Check out pull request head benchmark source
+        if: github.event_name == 'pull_request'
+        uses: actions/checkout@v7
+        with:
+          repository: ${{ github.event.pull_request.head.repo.full_name }}
+          ref: ${{ github.event.pull_request.head.sha }}
+          path: .benchmark/head-source
           persist-credentials: false
 
       - name: Install dependencies
@@ -49,41 +58,34 @@ jobs:
       - name: Set up Go
         uses: ./.github/actions/setup-go
 
-      - name: Measure pull request base
+      - name: Measure paired pull request revisions
         if: github.event_name == 'pull_request'
         run: |
           benchmark/baseline/run.sh \
-            "$GITHUB_WORKSPACE/.benchmark/source" \
+            "$GITHUB_WORKSPACE/.benchmark/base-source" \
             "$GITHUB_WORKSPACE/.benchmark/base-llgo" \
-            "$GITHUB_WORKSPACE/.benchmark/base-results"
-
-      - name: Check out pull request head benchmark source
-        if: github.event_name == 'pull_request'
-        uses: actions/checkout@v7
-        with:
-          repository: ${{ github.event.pull_request.head.repo.full_name }}
-          ref: ${{ github.event.pull_request.head.sha }}
-          path: .benchmark/source
-          persist-credentials: false
+            "$GITHUB_WORKSPACE/.benchmark/base-results" \
+            "$GITHUB_WORKSPACE/.benchmark/head-source" \
+            "$GITHUB_WORKSPACE/.benchmark/llgo" \
+            "$GITHUB_WORKSPACE/.benchmark/results"
 
       - name: Measure current revision
+        if: github.event_name != 'pull_request'
         run: |
-          source_root="$GITHUB_WORKSPACE"
-          if [[ "$GITHUB_EVENT_NAME" == pull_request ]]; then
-            source_root="$GITHUB_WORKSPACE/.benchmark/source"
-          fi
           benchmark/baseline/run.sh \
-            "$source_root" \
+            "$GITHUB_WORKSPACE" \
             "$GITHUB_WORKSPACE/.benchmark/llgo" \
             "$GITHUB_WORKSPACE/.benchmark/results"
 
       - name: Record benchmark result
-        uses: xgo-dev/setup-benchmark-go-action@v1.0.4
+        uses: xgo-dev/setup-benchmark-go-action@v1.0.5
         with:
           config: .github/llgo-benchmark.yml
           benchmark-file: .benchmark/results/benchmark.txt
           baseline-benchmark-file: >-
             ${{ github.event_name == 'pull_request' &&
                 '.benchmark/base-results/benchmark.txt' || '' }}
+          sample-pairing: >-
+            ${{ github.event_name == 'pull_request' && 'index' || '' }}
           platform-id: ${{ matrix.id }}
           platform-label: ${{ matrix.display }}

benchmark/baseline/README.md

@@ -1,7 +1,7 @@
 # LLGo baseline benchmarks
 
 This suite is the lightweight performance gate for ordinary LLGo changes. It
-uses fixed workloads and short calibrated benchmarks on Linux and macOS so it
+uses fixed workloads and calibrated benchmarks on Linux and macOS so it
 can run on every `main` push and pull request. Branch-only series can be run
 explicitly with `workflow_dispatch`, avoiding duplicate push and pull-request
 jobs for the same commit. The two native jobs record normalized artifacts; a
@@ -15,22 +15,33 @@ The program workloads reuse:
 - `benchmark/binary_size/fmtprintf`: `fmt.Printf`.
 
 For each workload, the collector performs an unmeasured warm build, then records
-median build time, median process time, file size, executable-code bytes,
-allocated non-executable data, and zero-filled data. On ELF, read-only constants
-are included in the data bucket; on Mach-O, `__TEXT` constants are included in
-the text bucket. The Go benchmark stream records five samples of selected
-compiler helpers and LLGo-generated core-language operations: direct/interface
-calls, defer, goroutine creation, channels, `getg`, and global access.
+the median of six builds and eighteen process runs, file size, executable-code
+bytes, allocated non-executable data, and zero-filled data. Workload order is
+rotated between rounds to balance runner drift and cache position. On ELF,
+read-only constants are included in the data bucket; on Mach-O, `__TEXT`
+constants are included in the text bucket. The Go benchmark stream performs one
+unrecorded warmup, then records seven one-second samples of compiler helpers and
+LLGo-generated core-language operations: direct/interface calls, defer,
+channels, `getg`, and global access. Goroutine creation keeps its bounded
+100-iteration samples.
 
-For pull requests, each platform job checks out the recorded base and current
-commits into the same source path, then runs both suites sequentially on the same
-runner. The pull request comment compares that pair, avoiding differences from
-runner machines and embedded source paths. Dependency setup is shared, and Go's
-build cache can be reused by unchanged packages; main pushes still run the suite
-only once. Very small changes can still be scheduler, frequency, or thermal
-noise and should be confirmed by repeated workflow runs. If a workflow does not
-provide a paired result, the publisher falls back to the latest matching `main`
-data.
+The program table also includes standalone memory-profile workloads so the
+whole-program no-consumer path is measurable independently of the retained
+profile paths. Every process disables BDWGC, warms with two million allocations,
+then internally times forty million escaping 16-byte allocations. The reported
+duration is the median of eighteen processes. The workloads are
+`memprofile-no-consumer`, `memprofile-rate0`, and `memprofile-default`.
+
+For pull requests, each platform job builds the recorded base and current
+revisions on one runner, then alternates their measurements within every round.
+Compiler and runtime benchmark binaries are built before sampling, so only their
+execution is interleaved and each matching base/current sample remains adjacent.
+This prevents a phase-wide frequency, thermal, or host load change from being
+attributed entirely to one revision. Dependency setup is shared, and Go's build
+cache can be reused by unchanged packages; main pushes still run the suite only
+once. Very small changes can remain scheduler noise and should be confirmed by
+repeated workflow runs. If a workflow does not provide a paired result, the
+publisher falls back to the latest matching `main` data.
 
 The trusted publisher commits the current result history and generated site to
 the `pages` branch of the configured data repository. Every LLGo repository
@@ -48,39 +59,13 @@ updated summary comment linking to their long-term trend page. If no matching
 `main` history exists yet, the pull-request report is still published and
 marks every metric as `new`.
 
-Local collection:
-
-```sh
-GOMAXPROCS=2 go build -o .benchmark/llgo ./cmd/llgo
-go run ./benchmark/baseline \
-  -llgo .benchmark/llgo \
-  -out .benchmark/results
-```
-
-Write the selected Go benchmark output to `.benchmark/results/go.txt`:
+Run the complete local collection with the same script as CI:
 
 ```sh
-results=.benchmark/results/go.txt
-GOMAXPROCS=1 go test \
-  -run '^$' \
-  -bench '^(BenchmarkMergeCompilerFlags|BenchmarkMergeLinkerFlags|BenchmarkLookupPCRandom)$' \
-  -benchtime=250ms -count=5 -cpu=1 \
-  ./internal/clang ./internal/build/funcinfo | tee "$results"
-GOMAXPROCS=1 .benchmark/llgo test \
-  -run '^$' \
-  -bench '^(BenchmarkRuntimeGetG|BenchmarkGlobal(Read|Write)|Benchmark(DirectCall|InterfaceCall|Defer|ChannelBuffered|ChannelHandoff))$' \
-  -benchtime=250ms -count=5 \
-  ./test/llgoext | tee -a "$results"
-GOMAXPROCS=1 .benchmark/llgo test \
-  -run '^$' -bench '^BenchmarkGoroutine$' -benchtime=100x -count=5 \
-  ./test/llgoext | tee -a "$results"
+benchmark/baseline/run.sh \
+  "$PWD" \
+  "$PWD/.benchmark/llgo" \
+  "$PWD/.benchmark/results"
 ```
 
-Then validate and export the complete artifact in standard Go benchmark format:
-
-```sh
-go run ./benchmark/baseline \
-  -mode export \
-  -out .benchmark/results \
-  -benchmark-output .benchmark/results/benchmark.txt
-```
+The normalized artifact is `.benchmark/results/benchmark.txt`.