Make asm use more selectable

- Make the SIMD ASM code off by default. Use configure --enable-simd-asm
  to enable.
- Allow MD5 ASM code to be requested even when OpenSSL is handling MD4
  checksums. Use configure --enable-md5-asm to enable.

Author: WayneD

Makefile.in

@@ -30,8 +30,9 @@ SHELL=/bin/sh
 .SUFFIXES:
 .SUFFIXES: .c .o
 
-SIMD_x86_64=simd-checksum-x86_64.o simd-checksum-avx2.o
-ASM_x86_64=lib/md5-asm-x86_64.o
+ROLL_SIMD_x86_64=simd-checksum-x86_64.o
+ROLL_ASM_x86_64=simd-checksum-avx2.o
+MD5_ASM_x86_64=lib/md5-asm-x86_64.o
 
 GENFILES=configure.sh aclocal.m4 config.h.in rsync.1 rsync.1.html \
 	 rsync-ssl.1 rsync-ssl.1.html rsyncd.conf.5 rsyncd.conf.5.html \
@@ -46,7 +47,7 @@ OBJS1=flist.o rsync.o generator.o receiver.o cleanup.o sender.o exclude.o \
 	util1.o util2.o main.o checksum.o match.o syscall.o log.o backup.o delete.o
 OBJS2=options.o io.o compat.o hlink.o token.o uidlist.o socket.o hashtable.o \
 	usage.o fileio.o batch.o clientname.o chmod.o acls.o xattrs.o
-OBJS3=progress.o pipe.o @ASM@ @SIMD@
+OBJS3=progress.o pipe.o @MD5_ASM@ @ROLL_SIMD@ @ROLL_ASM@
 DAEMON_OBJ = params.o loadparm.o clientserver.o access.o connection.o authenticate.o
 popt_OBJS=popt/findme.o  popt/popt.o  popt/poptconfig.o \
 	popt/popthelp.o popt/poptparse.o
@@ -147,13 +148,13 @@ git-version.h: ALWAYS_RUN
 ALWAYS_RUN:
 
 simd-checksum-x86_64.o: simd-checksum-x86_64.cpp
-	@$(srcdir)/cmd-or-msg disable-simd $(CXX) -I. $(CXXFLAGS) $(CPPFLAGS) -c -o $@ $(srcdir)/simd-checksum-x86_64.cpp
+	@$(srcdir)/cmd-or-msg disable-roll-simd $(CXX) -I. $(CXXFLAGS) $(CPPFLAGS) -c -o $@ $(srcdir)/simd-checksum-x86_64.cpp
 
 simd-checksum-avx2.o: simd-checksum-avx2.S
-	@$(srcdir)/cmd-or-msg disable-asm $(CC) $(CFLAGS) --include=$(srcdir)/rsync.h -DAVX2_ASM -I. @NOEXECSTACK@ -c -o $@ $(srcdir)/simd-checksum-avx2.S
+	@$(srcdir)/cmd-or-msg disable-roll-asm $(CC) $(CFLAGS) -I. @NOEXECSTACK@ -c -o $@ $(srcdir)/simd-checksum-avx2.S
 
-lib/md5-asm-x86_64.o: lib/md5-asm-x86_64.S config.h lib/md-defines.h
-	@$(srcdir)/cmd-or-msg disable-asm $(CC) -I. @NOEXECSTACK@ -c -o $@ $(srcdir)/lib/md5-asm-x86_64.S
+lib/md5-asm-x86_64.o: lib/md5-asm-x86_64.S lib/md-defines.h
+	@$(srcdir)/cmd-or-msg disable-md5-asm $(CC) -I. @NOEXECSTACK@ -c -o $@ $(srcdir)/lib/md5-asm-x86_64.S
 
 tls$(EXEEXT): $(TLS_OBJ)
 	$(CC) $(CFLAGS) $(LDFLAGS) -o $@ $(TLS_OBJ) $(LIBS)

NEWS.md

@@ -136,7 +136,8 @@
    (keeping the behavior the same as before), so specifying `--info=nonreg0`
    can be used to turn the warnings off.
 
- - More ASM optimizations from Shark64.
+ - An optional asm optimization for the rolling checksum from Shark64. Enable
+   it with `./configure --enable-roll-asm`.
 
  - Using `--debug=FILTER` now outputs a caution message if a filter rule
    has trailing whitespace.
@@ -192,14 +193,24 @@
    using the output of `git describe` when building inside a non-shallow git
    checkout, though.)
 
- - Improved the IPv6 determination in configure.
+ - Renamed configure's `--enable-simd` option to `--enable-roll-simd` and added
+   the option `--enable-roll-asm` to use the new asm version of the code.  Both
+   are x86_64/amd64 only.
+
+ - Renamed configure's `--enable-asm` option to `--enable-md5-asm` to avoid
+   confusion with the asm option for the rolling checksum.  It is also honored
+   even when openssl crypto is in use.  This allows: normal MD4 & MD5, normal
+   MD4 + asm MD5, openssl MD4 & MD5, or openssl MD4 + asm MD5.
 
- - Made SIMD & ASM configure default to "no" on non-Linux hosts due to various
-   reports of problems on NetBSD & macOS hosts.  These tests were also tweaked
-   to allow enabling the feature on a host_cpu of amd64 (was only x86_64).
+ - Made SIMD & asm configure checks default to "no" on non-Linux hosts due to
+   various reports of problems on NetBSD & macOS hosts.  These were also
+   tweaked to allow enabling the feature on a host_cpu of amd64 (was only
+   allowed on x86_64 before).
 
  - Fixed configure to not fail at the SIMD check when cross-compiling.
 
+ - Improved the IPv6 determination in configure.
+
  - Compile the C files with `-pedantic-errors` (when possible) so that we will
    get warned if a static initialization overflows in the future (among other
    things).

checksum.c

@@ -179,7 +179,7 @@ int canonical_checksum(int csum_type)
 	return 0;
 }
 
-#ifndef HAVE_SIMD /* See simd-checksum-*.cpp. */
+#ifndef USE_ROLL_SIMD /* See simd-checksum-*.cpp. */
 /*
   a simple 32 bit checksum that can be updated from either end
   (inspired by Mark Adler's Adler-32 checksum)
@@ -222,23 +222,23 @@ void get_checksum2(char *buf, int32 len, char *sum)
 	  }
 #endif
 	  case CSUM_MD5: {
-		MD5_CTX m5;
+		md5_context m5;
 		uchar seedbuf[4];
-		MD5_Init(&m5);
+		md5_begin(&m5);
 		if (proper_seed_order) {
 			if (checksum_seed) {
 				SIVALu(seedbuf, 0, checksum_seed);
-				MD5_Update(&m5, seedbuf, 4);
+				md5_update(&m5, seedbuf, 4);
 			}
-			MD5_Update(&m5, (uchar *)buf, len);
+			md5_update(&m5, (uchar *)buf, len);
 		} else {
-			MD5_Update(&m5, (uchar *)buf, len);
+			md5_update(&m5, (uchar *)buf, len);
 			if (checksum_seed) {
 				SIVALu(seedbuf, 0, checksum_seed);
-				MD5_Update(&m5, seedbuf, 4);
+				md5_update(&m5, seedbuf, 4);
 			}
 		}
-		MD5_Final((uchar *)sum, &m5);
+		md5_result(&m5, (uchar *)sum);
 		break;
 	  }
 	  case CSUM_MD4:
@@ -374,18 +374,18 @@ void file_checksum(const char *fname, const STRUCT_STAT *st_p, char *sum)
 	  }
 #endif
 	  case CSUM_MD5: {
-		MD5_CTX m5;
+		md5_context m5;
 
-		MD5_Init(&m5);
+		md5_begin(&m5);
 
 		for (i = 0; i + CHUNK_SIZE <= len; i += CHUNK_SIZE)
-			MD5_Update(&m5, (uchar *)map_ptr(buf, i, CHUNK_SIZE), CHUNK_SIZE);
+			md5_update(&m5, (uchar *)map_ptr(buf, i, CHUNK_SIZE), CHUNK_SIZE);
 
 		remainder = (int32)(len - i);
 		if (remainder > 0)
-			MD5_Update(&m5, (uchar *)map_ptr(buf, i, remainder), remainder);
+			md5_update(&m5, (uchar *)map_ptr(buf, i, remainder), remainder);
 
-		MD5_Final((uchar *)sum, &m5);
+		md5_result(&m5, (uchar *)sum);
 		break;
 	  }
 	  case CSUM_MD4:
@@ -443,7 +443,7 @@ static union {
 #ifdef USE_OPENSSL
 	MD4_CTX m4;
 #endif
-	MD5_CTX m5;
+	md5_context m5;
 } ctx;
 #ifdef SUPPORT_XXHASH
 static XXH64_state_t* xxh64_state;
@@ -482,7 +482,7 @@ void sum_init(int csum_type, int seed)
 		break;
 #endif
 	  case CSUM_MD5:
-		MD5_Init(&ctx.m5);
+		md5_begin(&ctx.m5);
 		break;
 	  case CSUM_MD4:
 #ifdef USE_OPENSSL
@@ -532,7 +532,7 @@ void sum_update(const char *p, int32 len)
 		break;
 #endif
 	  case CSUM_MD5:
-		MD5_Update(&ctx.m5, (uchar *)p, len);
+		md5_update(&ctx.m5, (uchar *)p, len);
 		break;
 	  case CSUM_MD4:
 #ifdef USE_OPENSSL
@@ -597,7 +597,7 @@ int sum_end(char *sum)
 	  }
 #endif
 	  case CSUM_MD5:
-		MD5_Final((uchar *)sum, &ctx.m5);
+		md5_result(&ctx.m5, (uchar *)sum);
 		break;
 	  case CSUM_MD4:
 #ifdef USE_OPENSSL

configure.ac

@@ -229,12 +229,12 @@ fi
 AC_DEFINE_UNQUOTED(NOBODY_USER, "$NOBODY_USER", [unprivileged user--e.g. nobody])
 AC_DEFINE_UNQUOTED(NOBODY_GROUP, "$NOBODY_GROUP", [unprivileged group for unprivileged user])
 
-# SIMD optimizations
-SIMD=
+# rolling-checksum SIMD optimizations
+ROLL_SIMD=
 
-AC_MSG_CHECKING([whether to enable SIMD optimizations])
-AC_ARG_ENABLE(simd,
-    AS_HELP_STRING([--enable-simd],[enable/disable to control SIMD optimizations (requires c++)]))
+AC_MSG_CHECKING([whether to enable rolling-checksum SIMD optimizations])
+AC_ARG_ENABLE(roll-simd,
+    AS_HELP_STRING([--enable-roll-simd],[enable/disable to control rolling-checksum SIMD optimizations (requires c++)]))
 
 # Clag is crashing with -g -O2, so we'll get rid of -g for now.
 CXXFLAGS=`echo "$CXXFLAGS" | sed 's/-g //'`
@@ -263,14 +263,14 @@ __attribute__ ((target("ssse3"))) void more_testing(char* buf, int len)
 }
 ]])
 
-if test x"$enable_simd" = x""; then
+if test x"$enable_roll_simd" = x""; then
     case "$host_os" in
 	*linux*) ;;
-	*) enable_simd=no ;;
+	*) enable_roll_simd=no ;;
     esac
 fi
 
-if test x"$enable_simd" != x"no"; then
+if test x"$enable_roll_simd" != x"no"; then
     # For x86-64 SIMD, g++ >=5 or clang++ >=7 is required
     if test x"$host_cpu" = x"x86_64" || test x"$host_cpu" = x"amd64"; then
 	AC_LANG(C++)
@@ -283,23 +283,23 @@ if test x"$enable_simd" != x"no"; then
 	AC_LANG(C)
 	if test x"$CXX_OK" = x"yes"; then
 	    # AC_MSG_RESULT() is called below.
-	    SIMD="$host_cpu"
-	elif test x"$enable_simd" = x"yes"; then
+	    ROLL_SIMD="$host_cpu"
+	elif test x"$enable_roll_simd" = x"yes"; then
 	    AC_MSG_RESULT(error)
-	    AC_MSG_ERROR(The SIMD compilation test failed.
-Omit --enable-simd to continue without it.)
+	    AC_MSG_ERROR(The rolling-checksum SIMD compilation test failed.
+Omit --enable-roll-simd to continue without it.)
 	fi
-    elif test x"$enable_simd" = x"yes"; then
+    elif test x"$enable_roll_simd" = x"yes"; then
         AC_MSG_RESULT(unavailable)
-        AC_MSG_ERROR(The SIMD optimizations are currently x86_64|amd64 only.
-Omit --enable-simd to continue without it.)
+        AC_MSG_ERROR(The rolling-checksum SIMD optimizations are currently x86_64|amd64 only.
+Omit --enable-roll-simd to continue without it.)
     fi
 fi
 
-if test x"$SIMD" != x""; then
-    AC_MSG_RESULT([yes ($SIMD)])
-    AC_DEFINE(HAVE_SIMD, 1, [Define to 1 to enable SIMD optimizations])
-    SIMD='$(SIMD_'"$SIMD)"
+if test x"$ROLL_SIMD" != x""; then
+    AC_MSG_RESULT([yes ($ROLL_SIMD)])
+    AC_DEFINE(USE_ROLL_SIMD, 1, [Define to 1 to enable rolling-checksum SIMD optimizations])
+    ROLL_SIMD='$(ROLL_SIMD_'"$ROLL_SIMD)"
     # We only use c++ for its target attribute dispatching, disable unneeded bulky features
     CXXFLAGS="$CXXFLAGS -fno-exceptions -fno-rtti"
     # Apple often has "g++" as a symlink for clang. Try to find out the truth.
@@ -311,7 +311,7 @@ else
     AC_MSG_RESULT(no)
 fi
 
-AC_SUBST(SIMD)
+AC_SUBST(ROLL_SIMD)
 
 AC_MSG_CHECKING([if assembler accepts noexecstack])
 OLD_CFLAGS="$CFLAGS"
@@ -433,45 +433,66 @@ if test x"$enable_openssl" != x"no"; then
 	err_msg="$err_msg$nl- Failed to find openssl/md4.h and openssl/md5.h for openssl crypto lib support."
 	no_lib="$no_lib openssl"
     fi
-    if test x"$enable_asm" != x"yes"; then
-	enable_asm=no
+    if test x"$enable_md5_asm" != x"yes"; then
+	enable_md5_asm=no
     fi
 else
     AC_MSG_RESULT(no)
 fi
 
-ASM=
+MD5_ASM=
 
-AC_MSG_CHECKING([whether to enable ASM optimizations])
-AC_ARG_ENABLE(asm,
-    AS_HELP_STRING([--enable-asm],[enable/disable to control ASM optimizations]))
+AC_MSG_CHECKING([whether to enable MD5 ASM optimizations])
+AC_ARG_ENABLE(md5-asm,
+    AS_HELP_STRING([--enable-md5-asm],[enable/disable to control MD5 ASM optimizations]))
 
-if test x"$enable_asm" = x""; then
+if test x"$enable_md5_asm" = x""; then
     case "$host_os" in
 	*linux*) ;;
-	*) enable_asm=no ;;
+	*) enable_md5_asm=no ;;
     esac
 fi
 
-if test x"$enable_asm" != x"no"; then
+if test x"$enable_md5_asm" != x"no"; then
     if test x"$host_cpu" = x"x86_64" || test x"$host_cpu" = x"amd64"; then
-	ASM="$host_cpu"
-    elif test x"$enable_asm" = x"yes"; then
+	MD5_ASM="$host_cpu"
+    elif test x"$enable_md5_asm" = x"yes"; then
         AC_MSG_RESULT(unavailable)
         AC_MSG_ERROR(The ASM optimizations are currently x86_64|amd64 only.
-Omit --enable-asm to continue without it.)
+Omit --enable-md5-asm to continue without it.)
     fi
 fi
 
-if test x"$ASM" != x""; then
-    AC_MSG_RESULT([yes ($ASM)])
-    AC_DEFINE(HAVE_ASM, 1, [Define to 1 to enable ASM optimizations])
-    ASM='$(ASM_'"$ASM)"
+if test x"$MD5_ASM" != x""; then
+    AC_MSG_RESULT([yes ($MD5_ASM)])
+    AC_DEFINE(USE_MD5_ASM, 1, [Define to 1 to enable MD5 ASM optimizations])
+    MD5_ASM='$(MD5_ASM_'"$MD5_ASM)"
 else
     AC_MSG_RESULT(no)
 fi
 
-AC_SUBST(ASM)
+AC_SUBST(MD5_ASM)
+
+ROLL_ASM=
+
+AC_MSG_CHECKING([whether to enable rolling-checksum ASM optimizations])
+AC_ARG_ENABLE(roll-asm,
+    AS_HELP_STRING([--enable-roll-asm],[enable/disable to control rolling-checksum ASM optimizations (requires --enable-roll-simd)]))
+
+if test x"$ROLL_SIMD" = x""; then
+    enable_roll_asm=no
+fi
+
+if test x"$enable_roll_asm" = x"yes"; then
+    ROLL_ASM="$host_cpu"
+    AC_MSG_RESULT([yes ($ROLL_ASM)])
+    AC_DEFINE(USE_ROLL_ASM, 1, [Define to 1 to enable rolling-checksum ASM optimizations (requires --enable-roll-simd)])
+    ROLL_ASM='$(ROLL_ASM_'"$ROLL_ASM)"
+else
+    AC_MSG_RESULT(no)
+fi
+
+AC_SUBST(ROLL_ASM)
 
 AC_MSG_CHECKING([whether to enable xxhash checksum support])
 AC_ARG_ENABLE([xxhash],
@@ -1421,10 +1442,6 @@ esac
 AC_CONFIG_FILES([Makefile lib/dummy zlib/dummy popt/dummy shconfig])
 AC_OUTPUT
 
-if test "$enable_openssl" = yes && test "$enable_asm" = yes; then
-    echo "*** Ignoring --enable-asm option -- using openssl for MD5 checksums ***"
-fi
-
 AC_MSG_RESULT()
 AC_MSG_RESULT([    rsync $PACKAGE_VERSION configuration successful])
 AC_MSG_RESULT()

lib/md5-asm-x86_64.S

@@ -27,7 +27,7 @@
 #include "config.h"
 #include "md-defines.h"
 
-#if !defined USE_OPENSSL && CSUM_CHUNK == 64
+#ifdef USE_MD5_ASM /* { */
 
 #ifdef __APPLE__
 #define md5_process_asm _md5_process_asm
@@ -698,4 +698,4 @@ md5_process_asm:
 	pop	%rbp
 	ret
 
-#endif /* !USE_OPENSSL ... */
+#endif /* } USE_MD5_ASM */