perf: share wrapper field metadata across variants

codegen/main.go

@@ -112,7 +112,7 @@ type Hooks struct {
 // hooks can be used.
 `, strings.Join(docList, "\n"))
 	g.Printf("func Wrap(w http.ResponseWriter, hooks Hooks) http.ResponseWriter {\n")
-	g.Printf("state := &rwState{w: w}\n")
+	g.Printf("state := &rwState{rwFields: rwFields{w: w}}\n")
 
 	// Precompute hook chains once per Wrap call and
 	// build a uint8 combo index so the switch compiles to a jump table.
@@ -155,10 +155,11 @@ type Hooks struct {
 	g.Printf("panic(\"unreachable\")")
 	g.Printf("}\n\n")
 
-	// rwState holds the underlying writer plus the precomputed hooks.
-	// All variant types are type-definitions over rwState, so a single *rwState
-	// allocation can be reinterpreted as any variant via pointer conversion.
-	g.Printf("type rwState struct {\n")
+	// Nest the fields in a shared type so each variant's runtime type descriptor
+	// contains one embedded field rather than repeating the entire field list.
+	// Value embedding keeps the memory layout and single allocation unchanged.
+	g.Printf("type rwState struct { rwFields }\n\n")
+	g.Printf("type rwFields struct {\n")
 	g.Printf("w http.ResponseWriter\n")
 	for _, iface := range ifaces {
 		for _, fn := range iface.Funcs {

experiments/binary-size/.gitignore

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

experiments/binary-size/README.md

@@ -0,0 +1,103 @@
+# Binary-size experiment
+
+Reproduces the release comparison from [issue #40](https://github.com/felixge/httpsnoop/issues/40).
+The same small HTTP server uses `httpsnoop.CaptureMetrics` to log response metrics
+and is built against each requested release. This exercises the wrapper rather
+than merely importing a package that the linker could discard.
+
+## Run
+
+From the repository root:
+
+```sh
+experiments/binary-size/compare.sh
+
+# Match the toolchain and architecture reported in the issue:
+GOTOOLCHAIN=go1.26.7 GOOS=darwin GOARCH=arm64 \
+  OUT="$PWD/experiments/binary-size/out/go1.26.7" \
+  experiments/binary-size/compare.sh
+
+# Optional release list; deltas are relative to the first release:
+experiments/binary-size/compare.sh v1.0.3 v1.0.4 v1.1.0
+
+# Compare a release with the current working tree via a local module replacement:
+experiments/binary-size/compare.sh v1.1.0 local
+```
+
+Requires Go, Bash, and standard Unix utilities (`awk`, `column`, `wc`). Go downloads
+release dependencies and, if requested, the toolchain.
+
+Each release gets a temporary standalone module with the same module path and
+source. The script does not change the repository's dependencies. Every build
+uses `CGO_ENABLED=0`, `-trimpath`, and `-buildvcs=false`. Inherited `GOFLAGS` are
+cleared, and workspace mode is disabled. Both normal binaries and binaries
+stripped with `-ldflags='-s -w'` are measured using their actual file size in bytes.
+
+`out/` is ignored by Git and contains:
+
+- `environment.txt`: toolchain, target, and build settings.
+- `results.tsv`: sizes and deltas relative to the first release for each mode.
+- `N-module.txt`: resolved httpsnoop version for release number N.
+- `N-default` and `N-stripped`: binaries, retained for further inspection.
+
+Each run overwrites corresponding outputs. Use separate `OUT` directories when
+comparing toolchains. Cross-compiled binaries can be measured without running them.
+
+## Observed results
+
+Measured on darwin/arm64 with CGO disabled:
+
+| Go | httpsnoop | Default bytes | Stripped bytes |
+| --- | --- | ---: | ---: |
+| 1.26.7 | v1.0.4 | 8,084,770 | 5,515,602 |
+| 1.26.7 | v1.1.0 | 9,877,634 | 6,857,186 |
+| 1.27.1 | v1.0.4 | 8,324,050 | 5,615,154 |
+| 1.27.1 | v1.1.0 | 10,115,442 | 6,971,026 |
+
+Upgrading from v1.0.4 to v1.1.0 adds:
+
+| Go | Default increase | Stripped increase |
+| --- | ---: | ---: |
+| 1.26.7 | 1,792,864 bytes (1.71 MiB, 22.18%) | 1,341,584 bytes (1.28 MiB, 24.32%) |
+| 1.27.1 | 1,791,392 bytes (1.71 MiB, 21.52%) | 1,355,872 bytes (1.29 MiB, 24.15%) |
+
+This confirms substantial growth for this application, including after removing
+symbols and debug information. It measures the net effect of upgrading the entire
+release, not the contribution of any individual change. Absolute sizes and deltas
+can vary with the application's reachable code, target, and toolchain. It does
+not measure runtime memory use, compressed size, or build-cache usage.
+
+## Shared-state metadata experiment
+
+The metadata-only implementation nests the underlying writer and hook fields in
+one value-embedded `rwFields` type. The 512 variant types then each describe one
+embedded field instead of repeating the same 14-field list. All method bodies,
+hook initialization, interface combinations, and the allocation layout remain
+unchanged. This does not include method promotion or metrics-only initialization.
+
+Before and after were both built using `local`, with the same repository path
+and build settings, to avoid differences in module replacement metadata:
+
+| Go | Mode | Before bytes | After bytes | Saved bytes |
+| --- | --- | ---: | ---: | ---: |
+| 1.26.7 | Default | 9,877,634 | 9,645,490 | 232,144 (2.35%) |
+| 1.26.7 | Stripped | 6,857,186 | 6,624,994 | 232,192 (3.39%) |
+| 1.27.1 | Default | 10,115,442 | 9,883,282 | 232,160 (2.30%) |
+| 1.27.1 | Stripped | 6,971,026 | 6,738,866 | 232,160 (3.33%) |
+
+The Go 1.26.7 read-only data section shrank by 163,616 bytes; Go 1.27.1's
+`__go_type` section shrank by 159,544 bytes. Machine-code section sizes were
+unchanged on both toolchains. Total file-size savings also reflect alignment and
+other file-format overhead, not just the type-data section.
+
+On Go 1.27.1, allocation benchmarks remained unchanged:
+
+| Benchmark | Bytes/op | Allocs/op |
+| --- | ---: | ---: |
+| Wrap | 128 | 1 |
+| CaptureMetrics | 417 | 13 |
+| CaptureMetricsTwice | 834 | 26 |
+
+Timing samples were noisy and are not evidence of a speed improvement. Tests
+(including all 512 interface combinations) passed on both toolchains, and the
+race tests and `go vet` passed on Go 1.27.1.

experiments/binary-size/compare.sh

@@ -0,0 +1,77 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+# Override GOTOOLCHAIN, GOOS, GOARCH, or OUT to compare other configurations.
+# All builds use the same source, toolchain, target, and CGO_ENABLED=0.
+script_dir=$(cd "$(dirname "$0")" && pwd)
+repo_dir=$(cd "$script_dir/../.." && pwd)
+export GOOS=${GOOS:-$(go env GOOS)}
+export GOARCH=${GOARCH:-$(go env GOARCH)}
+export CGO_ENABLED=0
+# Do not inherit flags that might change stripping or instrumentation.
+unset GOFLAGS
+out=${OUT:-"$script_dir/out"}
+mkdir -p "$out"
+out=$(cd "$out" && pwd)
+work=$(mktemp -d)
+trap 'rm -rf "$work"' EXIT
+export GOWORK=off
+
+if [ "$#" -eq 0 ]; then
+  set -- v1.0.4 v1.1.0
+fi
+
+{
+  go version
+  printf 'GOOS=%s GOARCH=%s CGO_ENABLED=%s\n' "$GOOS" "$GOARCH" "$CGO_ENABLED"
+  printf 'Common flags: -trimpath -buildvcs=false\n'
+  printf 'default: no additional linker flags; stripped: -ldflags="-s -w"\n'
+} | tee "$out/environment.txt"
+printf 'version\tmode\tbytes\tdelta_bytes\tdelta_percent\n' > "$out/results.tsv"
+default_base=0
+stripped_base=0
+index=0
+for version in "$@"; do
+  index=$((index + 1))
+  src="$work/$index"
+  mkdir -p "$src"
+  cp "$script_dir/testdata/main.go" "$src/main.go"
+  # Module path is identical for every build. Dependencies are resolved outside
+  # the repository so its go.mod and go.sum are never modified.
+  printf 'module binary-size-experiment\n\ngo 1.25\n' > "$src/go.mod"
+  (
+    cd "$src"
+    if [ "$version" = local ]; then
+      go mod edit -require=github.com/felixge/httpsnoop@v1.1.0 \
+        "-replace=github.com/felixge/httpsnoop=$repo_dir"
+    else
+      go mod edit "-require=github.com/felixge/httpsnoop@$version"
+    fi
+    go mod tidy
+    go list -m github.com/felixge/httpsnoop > "$out/$index-module.txt"
+    for mode in default stripped; do
+      binary="$out/$index-$mode"
+      if [ "$mode" = stripped ]; then
+        go build -trimpath -buildvcs=false -ldflags='-s -w' -o "$binary" .
+      else
+        go build -trimpath -buildvcs=false -o "$binary" .
+      fi
+    done
+  )
+  for mode in default stripped; do
+    bytes=$(wc -c < "$out/$index-$mode" | tr -d '[:space:]')
+    if [ "$mode" = default ]; then
+      if [ "$index" -eq 1 ]; then default_base=$bytes; fi
+      base=$default_base
+    else
+      if [ "$index" -eq 1 ]; then stripped_base=$bytes; fi
+      base=$stripped_base
+    fi
+    awk -v version="$version" -v mode="$mode" -v bytes="$bytes" -v base="$base" \
+      'BEGIN {printf "%s\t%s\t%d\t%d\t%.2f\n", version, mode, bytes, bytes-base, 100*(bytes-base)/base}' \
+      >> "$out/results.tsv"
+  done
+done
+printf '\n'
+column -t -s $'\t' "$out/results.tsv"
+printf '\nBinaries, resolved versions, and results saved to %s\n' "$out"

experiments/binary-size/testdata/main.go

@@ -0,0 +1,19 @@
+package main
+
+import (
+	"log"
+	"net/http"
+
+	"github.com/felixge/httpsnoop"
+)
+
+func main() {
+	handler := http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		_, _ = w.Write([]byte("hello\n"))
+	})
+	wrapped := http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		m := httpsnoop.CaptureMetrics(handler, w, r)
+		log.Printf("%s %s code=%d duration=%s written=%d", r.Method, r.URL, m.Code, m.Duration, m.Written)
+	})
+	log.Fatal(http.ListenAndServe(":8080", wrapped))
+}

wrap_state_test.go

@@ -0,0 +1,23 @@
+package httpsnoop
+
+import (
+	"reflect"
+	"testing"
+)
+
+func TestSharedStateLayout(t *testing.T) {
+	fields := reflect.TypeOf(rwFields{})
+	for _, value := range []any{rwState{}, rw0{}, rw511{}} {
+		typ := reflect.TypeOf(value)
+		if typ.NumField() != 1 {
+			t.Fatalf("%s has %d fields, want one shared field", typ, typ.NumField())
+		}
+		field := typ.Field(0)
+		if !field.Anonymous || field.Type != fields || field.Offset != 0 {
+			t.Errorf("%s does not embed rwFields at offset zero", typ)
+		}
+		if typ.Size() != fields.Size() || typ.Align() != fields.Align() {
+			t.Errorf("%s changes the shared state's size or alignment", typ)
+		}
+	}
+}