feat: estimate cycles

.github/actions/build-capstone/action.yml

@@ -0,0 +1,25 @@
+name: Build Capstone (cycle-estimation decoder)
+description: >-
+  Build the static Capstone that Callgrind cycle estimation links against, and
+  export its install prefix as CAPSTONE_DIR for later steps.
+
+runs:
+  using: composite
+  steps:
+    - shell: bash
+      run: |
+        CS_SRC="$GITHUB_WORKSPACE/third_party/capstone"
+        CS_PREFIX="$GITHUB_WORKSPACE/.capstone"
+        cmake -S "$CS_SRC" -B "$CS_SRC/build" \
+          -DCMAKE_BUILD_TYPE=Release \
+          -DCAPSTONE_ARCHITECTURE_DEFAULT=OFF \
+          -DCAPSTONE_X86_SUPPORT=ON \
+          -DCAPSTONE_ARM64_SUPPORT=ON \
+          -DCAPSTONE_BUILD_SHARED_LIBS=OFF \
+          -DCAPSTONE_BUILD_CSTOOL=OFF \
+          -DCMAKE_INSTALL_LIBDIR=lib \
+          -DCMAKE_INSTALL_PREFIX="$CS_PREFIX" \
+          -DCMAKE_C_FLAGS="-fno-stack-protector -U_FORTIFY_SOURCE -D_FORTIFY_SOURCE=0 -fPIC"
+        cmake --build "$CS_SRC/build" -j"$(nproc)"
+        cmake --install "$CS_SRC/build"
+        echo "CAPSTONE_DIR=$CS_PREFIX" >> "$GITHUB_ENV"

.github/workflows/ci.yml

@@ -21,6 +21,8 @@ jobs:
 
     steps:
       - uses: actions/checkout@v4
+        with:
+          submodules: recursive
 
       # Skip installing package docs to avoid wasting time when installing valgrind
       # See: https://github.com/actions/runner-images/issues/10977#issuecomment-2810713336
@@ -42,20 +44,25 @@ jobs:
             build-essential \
             automake \
             autoconf \
+            cmake \
             libc6-dev \
-            gcc-multilib \
-            libc6-dev-i386 \
             gdb \
             docbook \
             docbook-xsl \
             docbook-xml \
             xsltproc
 
+      - name: Build Capstone (cycle-estimation decoder)
+        uses: ./.github/actions/build-capstone
+
       - name: Run autogen
         run: ./autogen.sh
 
+      # CodSpeed only runs the 64-bit tool, and cycle estimation needs a
+      # 64-bit Capstone, so skip the 32-bit secondary build entirely. This
+      # halves build time and avoids compiling cycledecode.c without Capstone.
       - name: Configure
-        run: ./configure
+        run: ./configure --enable-only64bit
 
       - name: Build Valgrind
         run: make -j$(nproc)

.github/workflows/codspeed.yml

@@ -22,6 +22,7 @@ jobs:
       - uses: actions/checkout@v4
         with:
           lfs: true
+          submodules: recursive
       - uses: extractions/setup-just@v3
 
       # Skip installing package docs to avoid wasting time when installing build dependencies
@@ -58,12 +59,17 @@ jobs:
             build-essential \
             automake \
             autoconf \
+            cmake \
             gdb \
             docbook \
             docbook-xsl \
             docbook-xml \
             xsltproc
 
+      - name: Build Capstone (cycle-estimation decoder)
+        if: steps.valgrind-cache.outputs.cache-hit != 'true' && matrix.valgrind == 'local'
+        uses: ./.github/actions/build-capstone
+
       - name: Build Valgrind (${{ matrix.valgrind }})
         if: steps.valgrind-cache.outputs.cache-hit != 'true'
         run: just build ${{ matrix.valgrind }}
@@ -97,3 +103,4 @@ jobs:
         with:
           mode: walltime
           config: bench/codspeed.yml
+          runner-version: 4.17.7-alpha.1

.github/workflows/release.yml

@@ -16,11 +16,9 @@ jobs:
           - architecture: amd64
             platform: ubuntu-22.04
             ubuntu-version: 22.04
-            additional-deps: gcc-multilib libc6-dev-i386
           - architecture: amd64
             platform: ubuntu-24.04
             ubuntu-version: 24.04
-            additional-deps: gcc-multilib libc6-dev-i386
           - architecture: arm64
             platform: codspeedhq-arm64-ubuntu-22.04
             ubuntu-version: 22.04
@@ -31,20 +29,27 @@ jobs:
     runs-on: ${{ matrix.runner.platform }}
     steps:
       - uses: actions/checkout@v3
+        with:
+          submodules: recursive
       - name: Update apt-get cache
         run: sudo apt-get update
       - name: Install packaging deps
         run: sudo apt-get install -y build-essential devscripts debhelper dh-make
       - name: Install build deps
-        run: sudo apt-get install -y debhelper-compat gdb mpi-default-dev pkgconf docbook docbook-xsl docbook-xml xsltproc ${{ matrix.runner.additional-deps }}
+        run: sudo apt-get install -y debhelper-compat gdb mpi-default-dev pkgconf cmake docbook docbook-xsl docbook-xml xsltproc
       - name: Configure GPG Key
         run: echo -n "$GPG_SIGNING_KEY" | base64 --decode | gpg --import
         env:
           GPG_SIGNING_KEY: ${{ secrets.GPG_SIGNING_KEY }}
+
+      # debian/rules forwards $CAPSTONE_DIR to configure via --with-capstone.
+      - name: Build Capstone (cycle-estimation decoder)
+        uses: ./.github/actions/build-capstone
+
       - name: Build the deb package
         id: build_deb
         run: |
-          debuild --no-tgz-check -nc
+          debuild -e CAPSTONE_DIR --no-tgz-check -nc
           echo "asset-path=$(find .. -name 'valgrind_*.deb')" >> "$GITHUB_OUTPUT"
         env:
           DEBEMAIL: ${{ vars.MAINTAINER_EMAIL }}

.gitignore

@@ -2565,3 +2565,6 @@ test-suite.log
 
 # autoconf backup
 /configure~
+
+# Capstone build install prefix (built from the third_party/capstone submodule)
+/.capstone

.gitmodules

@@ -0,0 +1,4 @@
+[submodule "third_party/capstone"]
+	path = third_party/capstone
+	url = https://github.com/capstone-engine/capstone.git
+	shallow = true

Justfile

@@ -30,7 +30,10 @@ build-in dir:
         ./autogen.sh
     fi
 
-    ./configure
+    # 64-bit only: CodSpeed runs the 64-bit tool, and cycle estimation needs a
+    # 64-bit Capstone, so the 32-bit secondary build (which has no Capstone) is
+    # skipped.
+    ./configure --enable-only64bit
     make include/vgversion.h
     make -j$(nproc) -C VEX
     make -j$(nproc) -C coregrind

bench/generate_config.py

@@ -26,11 +26,14 @@
     "stress-ng --cpu 4 --cpu-ops 10",
 ]
 
-# Callgrind configurations: (extra args, config name). The config name is the
-# last segment of the benchmark id, e.g. `test_valgrind[<version>, <cmd>, no-inline]`.
+# Callgrind configurations: (extra args, config name, requires_codspeed). The
+# config name is the last segment of the benchmark id, e.g.
+# `test_valgrind[<version>, <cmd>, no-inline]`. `requires_codspeed` marks configs
+# that rely on CodSpeed-only options (e.g. `--cycle-estimation`); they are skipped
+# for upstream Valgrind builds, which would otherwise abort with "Unknown option".
 CONFIGS = [
-    (["--read-inline-info=no"], "no-inline"),
-    (["--read-inline-info=yes"], "inline"),
+    (["--read-inline-info=no"], "no-inline", False),
+    (["--read-inline-info=yes"], "inline", False),
     (
         [
             "--trace-children=yes",
@@ -45,6 +48,7 @@
             "--read-inline-info=yes",
         ],
         "full-with-inline",
+        False,
     ),
     (
         [
@@ -59,9 +63,31 @@
             "--dump-line=no",
         ],
         "full-no-inline",
+        False,
     ),
+    (
+        [
+            "--trace-children=yes",
+            "--cache-sim=yes",
+            "--I1=32768,8,64",
+            "--D1=32768,8,64",
+            "--LL=8388608,16,64",
+            "--collect-systime=nsec",
+            "--compress-strings=no",
+            "--combine-dumps=yes",
+            "--dump-line=no",
+            "--read-inline-info=yes",
+            "--cycle-estimation=yes"
+        ],
+        "full-with-inline-with-cycle-estimation",
+        True,
+    ),
+    (["--cycle-estimation=yes"], "cycle-estimation", True),
 ]
 
+# Label produced by `valgrind_version` for CodSpeed's custom build.
+CODSPEED_VERSION = "valgrind.codspeed"
+
 
 def valgrind_version(valgrind_path: str) -> str:
     """Return the normalized version label used in benchmark ids.
@@ -80,7 +106,7 @@ def valgrind_version(valgrind_path: str) -> str:
 
     version = result.stdout.strip()
     if "codspeed" in version:
-        return "valgrind.codspeed"
+        return CODSPEED_VERSION
     return version
 
 
@@ -89,8 +115,12 @@ def build_config(valgrind_paths: list) -> dict:
     benchmarks = []
     for valgrind_path in valgrind_paths:
         version = valgrind_version(valgrind_path)
+        is_codspeed = version == CODSPEED_VERSION
         for cmd in COMMANDS:
-            for args, config_name in CONFIGS:
+            for args, config_name, requires_codspeed in CONFIGS:
+                should_skip = requires_codspeed and not is_codspeed
+                if should_skip:
+                    continue
                 name = f"test_valgrind[{version}, {cmd}, {config_name}]"
                 exec_cmd = " ".join(
                     [valgrind_path, "--tool=callgrind", "--log-file=/dev/null", *args, cmd]

callgrind/Makefile.am

@@ -18,9 +18,16 @@ bin_SCRIPTS = \
 
 noinst_HEADERS = \
 	costs.h \
+	cycledecode.h \
+	cycledecode_capstone.h \
+	sigkey.h \
 	events.h \
 	global.h
 
+# Generated cost tables (#included by cycledecode.c under CLG_WITH_CAPSTONE;
+# the arch is selected at compile time).
+EXTRA_DIST += x86_caps_lut.inc arm64_caps_lut.inc
+
 #----------------------------------------------------------------------------
 # callgrind-<platform>
 #----------------------------------------------------------------------------
@@ -37,6 +44,8 @@ CALLGRIND_SOURCES_COMMON = \
 	clo.c \
 	context.c \
 	costs.c \
+	cycledecode.c \
+	cycledecode_capstone.c \
 	debug.c \
 	dump.c \
 	events.c \
@@ -54,11 +63,12 @@ callgrind_@VGCONF_ARCH_PRI@_@VGCONF_OS@_SOURCES      = \
 callgrind_@VGCONF_ARCH_PRI@_@VGCONF_OS@_CPPFLAGS     = \
 	$(AM_CPPFLAGS_@VGCONF_PLATFORM_PRI_CAPS@)
 callgrind_@VGCONF_ARCH_PRI@_@VGCONF_OS@_CFLAGS       = $(LTO_CFLAGS) \
-	$(AM_CFLAGS_@VGCONF_PLATFORM_PRI_CAPS@) $(CALLGRIND_CFLAGS_COMMON)
+	$(AM_CFLAGS_@VGCONF_PLATFORM_PRI_CAPS@) $(CALLGRIND_CFLAGS_COMMON) \
+	@CAPSTONE_CFLAGS@
 callgrind_@VGCONF_ARCH_PRI@_@VGCONF_OS@_DEPENDENCIES = \
 	$(TOOL_DEPENDENCIES_@VGCONF_PLATFORM_PRI_CAPS@)
 callgrind_@VGCONF_ARCH_PRI@_@VGCONF_OS@_LDADD        = \
-	$(TOOL_LDADD_@VGCONF_PLATFORM_PRI_CAPS@)
+	$(TOOL_LDADD_@VGCONF_PLATFORM_PRI_CAPS@) @CAPSTONE_LIBS@
 callgrind_@VGCONF_ARCH_PRI@_@VGCONF_OS@_LDFLAGS      = \
 	$(TOOL_LDFLAGS_@VGCONF_PLATFORM_PRI_CAPS@)
 callgrind_@VGCONF_ARCH_PRI@_@VGCONF_OS@_LINK = \