build: build the vendored Capstone from the autotools build

.github/workflows/ci.yml

@@ -44,17 +44,13 @@ jobs:
             build-essential \
             automake \
             autoconf \
-            cmake \
             libc6-dev \
             gdb \
             docbook \
             docbook-xsl \
             docbook-xml \
             xsltproc
 
-      - name: Build Capstone (cycle-estimation decoder)
-        uses: ./.github/actions/build-capstone
-
       - name: Run autogen
         run: ./autogen.sh
 

.github/workflows/codspeed.yml

@@ -72,17 +72,12 @@ 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 }}

.github/workflows/release.yml

@@ -36,20 +36,16 @@ jobs:
       - 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 cmake docbook docbook-xsl docbook-xml xsltproc
+        run: sudo apt-get install -y debhelper-compat gdb mpi-default-dev pkgconf 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 $CAPSTONE_DIR to configure via --with-capstone.
-      - name: Build Capstone (cycle-estimation decoder)
-        uses: ./.github/actions/build-capstone
-
       - name: Build the deb package
         id: build_deb
         run: |
-          debuild -e CAPSTONE_DIR --no-tgz-check -nc
+          debuild --no-tgz-check -nc
           echo "asset-path=$(find .. -name 'valgrind_*.deb')" >> "$GITHUB_OUTPUT"
         env:
           DEBEMAIL: ${{ vars.MAINTAINER_EMAIL }}

.gitignore

@@ -2566,7 +2566,14 @@ test-suite.log
 # autoconf backup
 /configure~
 
-# Capstone build install prefix (built from the third_party/capstone submodule)
+# /third_party/ (vendored Capstone, built by third_party/Makefile.am)
+/third_party/Makefile
+/third_party/Makefile.in
+/third_party/libcapstone.a
+/third_party/capstone-stubs/.deps
+/third_party/capstone-stubs/.dirstamp
+
+# install prefix of a Capstone built by hand for --with-capstone
 /.capstone
 
 # fake CodSpeed benchmark fixture binary (compiled from testdata/llsc_tzconvert_bench.c)

Makefile.am

@@ -22,6 +22,7 @@ EXP_TOOLS = \
 # Put docs last because building the HTML is slow and we want to get
 # everything else working before we try it.
 SUBDIRS = \
+	third_party \
 	include \
 	VEX \
 	coregrind \

autogen.sh

@@ -20,6 +20,16 @@ run autoconf
 if git rev-parse --is-inside-work-tree > /dev/null 2>&1 ; then
     echo "running: git configuration"
     git config blame.ignoreRevsFile .git-blame-ignore-revs
+    # CodSpeed: the vendored Capstone decoder that Callgrind's cycle estimation
+    # links against. A clone without --recurse-submodules leaves it empty.
+    # Not fatal: a build using --with-capstone=PATH (or CAPSTONE_DIR) needs no
+    # submodule, and must still work in a sandbox with no network. configure
+    # reports the empty submodule if neither is available.
+    echo "running: git submodule update --init third_party/capstone"
+    if ! git submodule update --init third_party/capstone ; then
+	echo "warning: could not check out third_party/capstone."
+	echo "warning: pass --with-capstone=PATH to configure to use a prebuilt Capstone."
+    fi
 else
     echo "skipping: git configuration"
 fi

callgrind/Makefile.am

@@ -67,7 +67,7 @@ callgrind_@VGCONF_ARCH_PRI@_@VGCONF_OS@_CFLAGS       = $(LTO_CFLAGS) \
 	$(AM_CFLAGS_@VGCONF_PLATFORM_PRI_CAPS@) $(CALLGRIND_CFLAGS_COMMON) \
 	@CAPSTONE_CFLAGS@
 callgrind_@VGCONF_ARCH_PRI@_@VGCONF_OS@_DEPENDENCIES = \
-	$(TOOL_DEPENDENCIES_@VGCONF_PLATFORM_PRI_CAPS@)
+	$(TOOL_DEPENDENCIES_@VGCONF_PLATFORM_PRI_CAPS@) @CAPSTONE_LIBS@
 callgrind_@VGCONF_ARCH_PRI@_@VGCONF_OS@_LDADD        = \
 	$(TOOL_LDADD_@VGCONF_PLATFORM_PRI_CAPS@) @CAPSTONE_LIBS@
 callgrind_@VGCONF_ARCH_PRI@_@VGCONF_OS@_LDFLAGS      = \

callgrind/cycledecode_capstone.c

@@ -55,6 +55,7 @@ extern char* vgPlain_strncpy(char* d, const char* s, unsigned long n);
 extern char* vgPlain_strchr(const char* s, char c);
 extern char* vgPlain_strrchr(const char* s, char c);
 extern char* vgPlain_strstr(const char* h, const char* n);
+extern char* vgPlain_strcat(char* d, const char* s);
 
 static const char* const CLG_CD_CC = "clg.cycledecode";
 
@@ -91,10 +92,16 @@ int printf(const char* fmt, ...)
 }
 int puts(const char* s) { return printf("%s\n", s); }
 
-/* Capstone's SStream references stderr/fwrite on a buffer-overflow guard in the
- * op_str text path (which this code never reads). Valgrind has no FILE* layer,
- * so stderr is a sentinel and fwrite routes the bytes to the Valgrind log fd,
- * making such an overflow visible rather than swallowed. */
+/* Capstone's SStream references stderr/fprintf on a buffer-overflow guard in
+ * the op_str text path (which this code never reads). Valgrind has no FILE*
+ * layer, so stderr is a sentinel and these route the bytes to the Valgrind log
+ * fd, making such an overflow visible rather than swallowed.
+ *
+ * Both fprintf and fwrite are needed: the guard calls fprintf, which GCC folds
+ * into fwrite for a format string with no conversions, but only when builtins
+ * are enabled. Valgrind's tool CFLAGS pass -fno-builtin, so whether the object
+ * ends up referencing fprintf or fwrite depends on how Capstone was compiled.
+ * Defining both keeps the tool linkable either way. */
 extern int vgPlain_write(int fd, const void* buf, int count);
 FILE*      stderr = 0;
 size_t     fwrite(const void* p, size_t size, size_t nmemb, FILE* f)
@@ -103,6 +110,15 @@ size_t     fwrite(const void* p, size_t size, size_t nmemb, FILE* f)
    vgPlain_write(2, p, (int)(size * nmemb));
    return nmemb;
 }
+int fprintf(FILE* f, const char* fmt, ...)
+{
+   (void)f;
+   va_list ap;
+   va_start(ap, fmt);
+   unsigned int r = vgPlain_vprintf(fmt, ap);
+   va_end(ap);
+   return (int)r;
+}
 
 size_t strlen(const char* s) { return vgPlain_strlen(s); }
 int    strcmp(const char* a, const char* b) { return vgPlain_strcmp(a, b); }
@@ -118,6 +134,7 @@ char* strncpy(char* d, const char* s, size_t n)
 char* strchr(const char* s, int c) { return vgPlain_strchr(s, (char)c); }
 char* strrchr(const char* s, int c) { return vgPlain_strrchr(s, (char)c); }
 char* strstr(const char* h, const char* n) { return vgPlain_strstr(h, n); }
+char* strcat(char* d, const char* s) { return vgPlain_strcat(d, s); }
 
 /*------------------------------------------------------------*/
 /*--- Capstone handle: open / decode                        -*/

configure.ac

@@ -63,6 +63,9 @@ AC_PROG_CC
 AS_IF([test "x$ac_cv_prog_cc_c11" = "xno"],
       [AC_MSG_ERROR([Valgrind relies on a C compiler supporting C11])])
 
+# CodSpeed: pin the C dialect. See m4/codspeed_capstone.m4.
+CODSPEED_C_STD_GNU17
+
 AC_PROG_CPP
 AC_PROG_CXX
 AC_PROG_RANLIB
@@ -1122,36 +1125,8 @@ AC_ARG_WITH(tmpdir,
 AC_DEFINE_UNQUOTED(VG_TMPDIR, "$tmpdir", [Temporary files directory])
 AC_SUBST(VG_TMPDIR, [$tmpdir])
 
-#----------------------------------------------------------------------------
-# CodSpeed: mandatory Capstone for Callgrind per-instruction cycle estimation
-#----------------------------------------------------------------------------
-# Callgrind's primary (native) tool is always built with the Capstone decoder
-# and a generated cost table (x86 or arm64, selected at compile time), enabling
-# --cycle-estimation=yes. The decoder location comes from --with-capstone=PATH
-# or, when omitted, the CAPSTONE_DIR environment variable (`nix develop` sets
-# it). A build without Capstone is not supported and fails here.
-AC_ARG_WITH(capstone,
-   [  --with-capstone=PATH    Path to the Capstone decoder install used for
-                          Callgrind cycle estimation (Cy/Cl). Defaults to the
-                          CAPSTONE_DIR environment variable (amd64/arm64)],
-   [capstone_dir="$withval"],
-   [capstone_dir="$CAPSTONE_DIR"])
-
-if test -z "$capstone_dir"; then
-   AC_MSG_ERROR([Capstone is required for Callgrind cycle estimation. Pass --with-capstone=PATH or set CAPSTONE_DIR (`nix develop` provides it).])
-fi
-if test ! -f "$capstone_dir/lib/libcapstone.a" \
-     -o ! -f "$capstone_dir/include/capstone/capstone.h"; then
-   AC_MSG_ERROR([--with-capstone=$capstone_dir: libcapstone.a or capstone.h not found])
-fi
-# Fortify off: the tool links -nodefaultlibs, so glibc's __*_chk fortify
-# wrappers are unavailable, and our libc shims (sprintf/snprintf/...) must
-# be real definitions, not fortify macro-expansions.
-CAPSTONE_CFLAGS="-DCLG_WITH_CAPSTONE -U_FORTIFY_SOURCE -D_FORTIFY_SOURCE=0 -I$capstone_dir/include"
-CAPSTONE_LIBS="$capstone_dir/lib/libcapstone.a"
-AC_MSG_NOTICE([Callgrind cycle estimation enabled with Capstone at $capstone_dir])
-AC_SUBST(CAPSTONE_CFLAGS)
-AC_SUBST(CAPSTONE_LIBS)
+# CodSpeed: Capstone decoder for Callgrind cycle estimation. See m4/codspeed_capstone.m4.
+CODSPEED_CAPSTONE
 
 #----------------------------------------------------------------------------
 # Detect xcode path
@@ -5790,6 +5765,7 @@ CFLAGS=$safe_CFLAGS
 AC_CONFIG_FILES([
    Makefile
    VEX/Makefile:Makefile.vex.in
+   third_party/Makefile
    valgrind.spec
    valgrind.pc
    glibc-2.X.supp

debian/rules

@@ -28,8 +28,9 @@ CONFARGS = \
 	--enable-only64bit \
 	--with-gdbscripts-dir=/usr/share/gdb/auto-load
 
-# Callgrind cycle estimation requires Capstone; the release workflow builds a
-# static decoder and exports its prefix as CAPSTONE_DIR.
+# Callgrind cycle estimation needs Capstone. The build compiles the vendored
+# third_party/capstone submodule by default; CAPSTONE_DIR, when set, points at a
+# prebuilt decoder to use instead.
 ifneq ($(CAPSTONE_DIR),)
   CONFARGS += --with-capstone=$(CAPSTONE_DIR)
 endif

flake.nix

@@ -36,8 +36,9 @@
         });
       in
       {
-        # Expose the pinned Capstone so the autotools build and scripts can find
-        # it via `nix build .#capstone` or the CAPSTONE_DIR env var below.
+        # Expose the pinned Capstone for the builds that want a prebuilt decoder
+        # rather than the vendored submodule: `nix build .#capstone`, then pass it
+        # to configure as --with-capstone=PATH (or CAPSTONE_DIR).
         packages.capstone = capstone;
 
         devShells.default = pkgs.mkShell {
@@ -62,10 +63,6 @@
             pkgs.gcc
             pkgs.pkg-config
           ];
-
-          # Consumed by configure (--with-capstone), the LUT generator, and the
-          # standalone cycledecode test. Point them at the hardening-free build.
-          CAPSTONE_DIR = "${capstone}";
         };
       }
     );

m4/codspeed_capstone.m4

@@ -0,0 +1,73 @@
+# codspeed_capstone.m4 -- CodSpeed additions to Valgrind's configure.
+
+# CODSPEED_CAPSTONE
+# -----------------
+# Locate the Capstone decoder that Callgrind's per-instruction cycle
+# estimation (--cycle-estimation=yes) links against, and export
+# CAPSTONE_CFLAGS / CAPSTONE_LIBS for callgrind/Makefile.am.
+#
+# By default the vendored submodule in third_party/capstone is compiled by
+# third_party/Makefile.am as part of `make`, so a plain
+# `./autogen.sh && ./configure && make && make install` is all that is needed.
+# --with-capstone=PATH (or the CAPSTONE_DIR environment variable, which `nix
+# develop` sets) selects a prebuilt Capstone install instead.
+AC_DEFUN([CODSPEED_CAPSTONE], [
+AC_ARG_WITH([capstone],
+   [AS_HELP_STRING([--with-capstone=PATH],
+      [use a prebuilt Capstone install for Callgrind cycle estimation instead
+       of the vendored third_party/capstone submodule. Defaults to the
+       CAPSTONE_DIR environment variable])],
+   [capstone_dir="$withval"],
+   [capstone_dir="$CAPSTONE_DIR"])
+
+AM_CONDITIONAL([BUILD_VENDORED_CAPSTONE], [test -z "$capstone_dir"])
+
+if test -z "$capstone_dir"; then
+   # Built by third_party/Makefile.am. Note there is deliberately no check
+   # for libcapstone.a here: it does not exist yet at configure time.
+   if test ! -f "$srcdir/third_party/capstone/cs.c"; then
+      AC_MSG_ERROR([third_party/capstone is empty. Run:
+   git submodule update --init third_party/capstone
+or pass --with-capstone=PATH to use a prebuilt Capstone.])
+   fi
+   CAPSTONE_INCLUDES='-I$(top_srcdir)/third_party/capstone/include'
+   CAPSTONE_LIBS='$(top_builddir)/third_party/libcapstone.a'
+   AC_MSG_NOTICE([Callgrind cycle estimation enabled with the vendored Capstone])
+else
+   if test ! -f "$capstone_dir/lib/libcapstone.a" \
+        -o ! -f "$capstone_dir/include/capstone/capstone.h"; then
+      AC_MSG_ERROR([--with-capstone=$capstone_dir: libcapstone.a or capstone.h not found])
+   fi
+   CAPSTONE_INCLUDES="-I$capstone_dir/include"
+   CAPSTONE_LIBS="$capstone_dir/lib/libcapstone.a"
+   AC_MSG_NOTICE([Callgrind cycle estimation enabled with Capstone at $capstone_dir])
+fi
+
+# Fortify off: the tool links -nodefaultlibs, so glibc's __*_chk fortify
+# wrappers are unavailable, and our libc shims (sprintf/snprintf/...) must
+# be real definitions, not fortify macro-expansions.
+CAPSTONE_CFLAGS="-DCLG_WITH_CAPSTONE -U_FORTIFY_SOURCE -D_FORTIFY_SOURCE=0 $CAPSTONE_INCLUDES"
+AC_SUBST([CAPSTONE_CFLAGS])
+AC_SUBST([CAPSTONE_LIBS])
+])
+
+# CODSPEED_C_STD_GNU17
+# --------------------
+# Pin the C dialect to gnu17.
+#
+# AC_PROG_CC under Autoconf 2.70+ selects the newest dialect the compiler
+# supports, which is -std=gnu23 on GCC 15+. Under C23, glibc 2.42+ defines
+# strchr/strrchr/strstr as _Generic macros, which clash with Callgrind's own
+# definitions of those symbols. Appending to CFLAGS is enough to win: Autoconf
+# puts its own -std= into $CC, and CFLAGS comes later on the command line.
+#
+# Must be called after AC_PROG_CC.
+AC_DEFUN([CODSPEED_C_STD_GNU17], [
+AC_MSG_CHECKING([whether $CC accepts -std=gnu17])
+codspeed_save_CFLAGS="$CFLAGS"
+CFLAGS="$CFLAGS -std=gnu17"
+AC_COMPILE_IFELSE([AC_LANG_PROGRAM([], [])],
+   [AC_MSG_RESULT([yes])],
+   [AC_MSG_RESULT([no])
+    CFLAGS="$codspeed_save_CFLAGS"])
+])