Add optimized slide_hash for Power processors

CMakeLists.txt

@@ -7,6 +7,7 @@ set(VERSION "1.2.12.1")
 
 option(ASM686 "Enable building i686 assembly implementation")
 option(AMD64 "Enable building amd64 assembly implementation")
+option(POWER "Enable building power implementation")
 
 set(INSTALL_BIN_DIR "${CMAKE_INSTALL_PREFIX}/bin" CACHE PATH "Installation directory for executables")
 set(INSTALL_LIB_DIR "${CMAKE_INSTALL_PREFIX}/lib" CACHE PATH "Installation directory for libraries")
@@ -140,6 +141,73 @@ if(CMAKE_COMPILER_IS_GNUCC)
 		add_definitions(-DASMV)
 		set_source_files_properties(${ZLIB_ASMS} PROPERTIES LANGUAGE C COMPILE_FLAGS -DNO_UNDERLINE)
 	endif()
+
+    # test to see if we can use a GNU indirect function to detect and load optimized code at runtime
+    CHECK_C_SOURCE_COMPILES("
+    static int test_ifunc_native(void)
+    {
+       return 1;
+    }
+    static int (*(check_ifunc_native(void)))(void)
+    {
+       return test_ifunc_native;
+    }
+    int test_ifunc(void) __attribute__ ((ifunc (\"check_ifunc_native\")));
+    int main(void)
+    {
+       return 0;
+    }
+    " HAS_C_ATTR_IFUNC)
+
+    if(HAS_C_ATTR_IFUNC)
+        add_definitions(-DHAVE_IFUNC)
+        set(ZLIB_PRIVATE_HDRS ${ZLIB_PRIVATE_HDRS} contrib/gcc/zifunc.h)
+    endif()
+
+    if(POWER)
+        # Test to see if we can use the optimizations for Power
+        CHECK_C_SOURCE_COMPILES("
+        #ifndef _ARCH_PPC
+            #error \"Target is not Power\"
+        #endif
+        #ifndef __BUILTIN_CPU_SUPPORTS__
+            #error \"Target doesn't support __builtin_cpu_supports()\"
+        #endif
+        int main() { return 0; }
+        " HAS_POWER_SUPPORT)
+
+        if(HAS_POWER_SUPPORT AND HAS_C_ATTR_IFUNC)
+            add_definitions(-DZ_POWER_OPT)
+
+            set(CMAKE_REQUIRED_FLAGS -mcpu=power8)
+            CHECK_C_SOURCE_COMPILES("int main(void){return 0;}" POWER8)
+
+            if(POWER8)
+                add_definitions(-DZ_POWER8)
+                set(ZLIB_POWER8
+                    contrib/power/slide_hash_power8.c)
+
+                set_source_files_properties(
+                    ${ZLIB_POWER8}
+                    PROPERTIES COMPILE_FLAGS -mcpu=power8)
+            endif()
+
+            set(CMAKE_REQUIRED_FLAGS -mcpu=power9)
+            CHECK_C_SOURCE_COMPILES("int main(void){return 0;}" POWER9)
+
+            if(POWER9)
+                add_definitions(-DZ_POWER9)
+                set(ZLIB_POWER9 )
+
+                set_source_files_properties(
+                    ${ZLIB_POWER9}
+                    PROPERTIES COMPILE_FLAGS -mcpu=power9)
+            endif()
+
+            set(ZLIB_PRIVATE_HDRS ${ZLIB_PRIVATE_HDRS} contrib/power/power.h)
+            set(ZLIB_SRCS ${ZLIB_SRCS} ${ZLIB_POWER8} ${ZLIB_POWER9})
+        endif()
+    endif()
 endif()
 
 if(MSVC)

Makefile.in

@@ -165,6 +165,9 @@ crc32.o: $(SRCDIR)crc32.c
 deflate.o: $(SRCDIR)deflate.c
 	$(CC) $(CFLAGS) $(ZINC) -c -o $@ $(SRCDIR)deflate.c
 
+slide_hash_power8.o: $(SRCDIR)contrib/power/slide_hash_power8.c
+	$(CC) $(CFLAGS) -mcpu=power8 $(ZINC) -c -o $@ $(SRCDIR)contrib/power/slide_hash_power8.c
+
 infback.o: $(SRCDIR)infback.c
 	$(CC) $(CFLAGS) $(ZINC) -c -o $@ $(SRCDIR)infback.c
 
@@ -217,6 +220,11 @@ deflate.lo: $(SRCDIR)deflate.c
 	$(CC) $(SFLAGS) $(ZINC) -DPIC -c -o objs/deflate.o $(SRCDIR)deflate.c
 	-@mv objs/deflate.o $@
 
+slide_hash_power8.lo: $(SRCDIR)contrib/power/slide_hash_power8.c
+	-@mkdir objs 2>/dev/null || test -d objs
+	$(CC) $(SFLAGS) -mcpu=power8 $(ZINC) -DPIC -c -o objs/slide_hash_power8.o $(SRCDIR)contrib/power/slide_hash_power8.c
+	-@mv objs/slide_hash_power8.o $@
+
 infback.lo: $(SRCDIR)infback.c
 	-@mkdir objs 2>/dev/null || test -d objs
 	$(CC) $(SFLAGS) $(ZINC) -DPIC -c -o objs/infback.o $(SRCDIR)infback.c

configure

@@ -836,6 +836,72 @@ EOF
   fi
 fi
 
+# test to see if we can use a gnu indirection function to detect and load optimized code at runtime
+echo >> configure.log
+cat > $test.c <<EOF
+static int test_ifunc_native(void)
+{
+  return 1;
+}
+
+static int (*(check_ifunc_native(void)))(void)
+{
+  return test_ifunc_native;
+}
+
+int test_ifunc(void) __attribute__ ((ifunc ("check_ifunc_native")));
+EOF
+
+if tryboth $CC -c $CFLAGS $test.c; then
+  SFLAGS="${SFLAGS} -DHAVE_IFUNC"
+  CFLAGS="${CFLAGS} -DHAVE_IFUNC"
+  echo "Checking for attribute(ifunc) support... Yes." | tee -a configure.log
+else
+  echo "Checking for attribute(ifunc) support... No." | tee -a configure.log
+fi
+
+# Test to see if we can use the optimizations for Power
+echo >> configure.log
+cat > $test.c <<EOF
+#ifndef _ARCH_PPC
+  #error "Target is not Power"
+#endif
+#ifndef HAVE_IFUNC
+  #error "Target doesn't support ifunc"
+#endif
+#ifndef __BUILTIN_CPU_SUPPORTS__
+  #error "Target doesn't support __builtin_cpu_supports()"
+#endif
+EOF
+
+if tryboth $CC -c $CFLAGS $test.c; then
+  echo "int main(void){return 0;}" > $test.c
+
+  if tryboth $CC -c $CFLAGS -mcpu=power8 $test.c; then
+    POWER8="-DZ_POWER8"
+    PIC_OBJC="${PIC_OBJC} slide_hash_power8.lo"
+    OBJC="${OBJC} slide_hash_power8.o"
+    echo "Checking for -mcpu=power8 support... Yes." | tee -a configure.log
+  else
+    echo "Checking for -mcpu=power8 support... No." | tee -a configure.log
+  fi
+
+  if tryboth $CC -c $CFLAGS -mcpu=power9 $test.c; then
+    POWER9="-DZ_POWER9"
+    PIC_OBJC="${PIC_OBJC}"
+    OBJC="${OBJC}"
+    echo "Checking for -mcpu=power9 support... Yes." | tee -a configure.log
+  else
+    echo "Checking for -mcpu=power9 support... No." | tee -a configure.log
+  fi
+
+  SFLAGS="${SFLAGS} ${POWER8} ${POWER9} -DZ_POWER_OPT"
+  CFLAGS="${CFLAGS} ${POWER8} ${POWER9} -DZ_POWER_OPT"
+  echo "Checking for Power optimizations support... Yes." | tee -a configure.log
+else
+  echo "Checking for Power optimizations support... No." | tee -a configure.log
+fi
+
 # show the results in the log
 echo >> configure.log
 echo ALL = $ALL >> configure.log

contrib/README.contrib

@@ -11,6 +11,10 @@ ada/        by Dmitriy Anisimkov <[email protected]>
 blast/      by Mark Adler <[email protected]>
         Decompressor for output of PKWare Data Compression Library (DCL)
 
+gcc/        by Matheus Castanho <[email protected]>
+            and Rogerio Alves <[email protected]>
+        Optimization helpers using GCC-specific extensions
+
 delphi/     by Cosmin Truta <[email protected]>
         Support for Delphi and C++ Builder
 
@@ -42,6 +46,10 @@ minizip/    by Gilles Vollant <[email protected]>
 pascal/     by Bob Dellaca <[email protected]> et al.
         Support for Pascal
 
+power/      by Matheus Castanho <[email protected]>
+            and Rogerio Alves <[email protected]>
+        Optimized functions for Power processors
+
 puff/       by Mark Adler <[email protected]>
         Small, low memory usage inflate.  Also serves to provide an
         unambiguous description of the deflate format.

contrib/gcc/zifunc.h

@@ -0,0 +1,60 @@
+/* Copyright (C) 2019 Matheus Castanho <[email protected]>, IBM
+ *               2019 Rogerio Alves    <[email protected]>, IBM
+ * For conditions of distribution and use, see copyright notice in zlib.h
+ */
+
+#ifndef Z_IFUNC_H_
+#define Z_IFUNC_H_
+
+/* Helpers for arch optimizations */
+
+#define Z_IFUNC(fname) \
+    typeof(fname) fname __attribute__ ((ifunc (#fname "_resolver"))); \
+    local typeof(fname) *fname##_resolver(void)
+/* This is a helper macro to declare a resolver for an indirect function
+ * (ifunc). Let's say you have function
+ *
+ *    int foo (int a);
+ *
+ * for which you want to provide different implementations, for example:
+ *
+ *    int foo_clever (int a) {
+ *      ... clever things ...
+ *    }
+ *
+ *    int foo_smart (int a) {
+ *      ... smart things ...
+ *    }
+ *
+ * You will have to declare foo() as an indirect function and also provide a
+ * resolver for it, to choose between foo_clever() and foo_smart() based on
+ * some criteria you define (e.g. processor features).
+ *
+ * Since most likely foo() has a default implementation somewhere in zlib, you
+ * may have to rename it so the 'foo' symbol can be used by the ifunc without
+ * conflicts.
+ *
+ *    #define foo foo_default
+ *    int foo (int a) {
+ *      ...
+ *    }
+ *    #undef foo
+ *
+ * Now you just have to provide a resolver function to choose which function
+ * should be used (decided at runtime on the first call to foo()):
+ *
+ *    Z_IFUNC(foo) {
+ *        if (... some condition ...)
+ *          return foo_clever;
+ *
+ *        if (... other condition ...)
+ *          return foo_smart;
+ *
+ *        return foo_default;
+ *    }
+ *
+ * All calls to foo() throughout the code can remain untouched, all the magic
+ * will be done by the linker using the resolver function.
+ */
+
+#endif /* Z_IFUNC_H_ */

contrib/power/power.h

@@ -0,0 +1,8 @@
+/* Copyright (C) 2019 Matheus Castanho <[email protected]>, IBM
+ *               2019 Rogerio Alves    <[email protected]>, IBM
+ * For conditions of distribution and use, see copyright notice in zlib.h
+ */
+
+#include "../../deflate.h"
+
+void _slide_hash_power8(deflate_state *s);

contrib/power/slide_hash_power8.c

@@ -0,0 +1,63 @@
+ /* Copyright (C) 2019 Matheus Castanho <[email protected]>, IBM
+  * For conditions of distribution and use, see copyright notice in zlib.h
+  */
+
+#include <altivec.h>
+#include "../../deflate.h"
+
+local inline void slide_hash_power8_loop OF((deflate_state *s,
+      unsigned n_elems, Posf *table_end)) __attribute__((always_inline));
+
+local void slide_hash_power8_loop(
+    deflate_state *s,
+    unsigned n_elems,
+    Posf *table_end)
+{
+    vector unsigned short vw, vm, *vp;
+    unsigned chunks;
+
+    /* Each vector register (chunk) corresponds to 128 bits == 8 Posf,
+     * so instead of processing each of the n_elems in the hash table
+     * individually, we can do it in chunks of 8 with vector instructions.
+     *
+     * This function is only called from slide_hash_power8(), and both calls
+     * pass n_elems as a power of 2 higher than 2^7, as defined by
+     * deflateInit2_(), so n_elems will always be a multiple of 8. */
+    chunks = n_elems >> 3;
+    Assert(n_elems % 8 == 0, "Weird hash table size!");
+
+    /* This type casting is safe since s->w_size is always <= 64KB
+     * as defined by deflateInit2_() and Posf == unsigned short */
+    vw[0] = (Posf) s->w_size;
+    vw = vec_splat(vw,0);
+
+    vp = (vector unsigned short *) table_end;
+
+    do {
+        /* Processing 8 elements at a time */
+        vp--;
+        vm = *vp;
+
+        /* This is equivalent to: m >= w_size ? m - w_size : 0
+         * Since we are using a saturated unsigned subtraction, any
+         * values that are > w_size will be set to 0, while the others
+         * will be subtracted by w_size. */
+        *vp = vec_subs(vm,vw);
+    } while (--chunks);
+};
+
+void ZLIB_INTERNAL _slide_hash_power8(deflate_state *s)
+{
+    unsigned n;
+    Posf *p;
+
+    n = s->hash_size;
+    p = &s->head[n];
+    slide_hash_power8_loop(s,n,p);
+
+#ifndef FASTEST
+    n = s->w_size;
+    p = &s->prev[n];
+    slide_hash_power8_loop(s,n,p);
+#endif
+}

contrib/power/slide_hash_resolver.c

@@ -0,0 +1,15 @@
+/* Copyright (C) 2019 Matheus Castanho <[email protected]>, IBM
+ * For conditions of distribution and use, see copyright notice in zlib.h
+ */
+
+#include "../gcc/zifunc.h"
+#include "power.h"
+
+Z_IFUNC(slide_hash) {
+#ifdef Z_POWER8
+    if (__builtin_cpu_supports("arch_2_07"))
+        return _slide_hash_power8;
+#endif
+
+    return slide_hash_default;
+}

deflate.c

@@ -196,6 +196,13 @@ local const config configuration_table[10] = {
                  (unsigned)(s->hash_size-1)*sizeof(*s->head)); \
     } while (0)
 
+#ifdef Z_POWER_OPT
+/* Rename function so resolver can use its symbol. The default version will be
+ * returned by the resolver if the host has no support for an optimized version.
+ */
+#define slide_hash slide_hash_default
+#endif /* Z_POWER_OPT */
+
 /* ===========================================================================
  * Slide the hash table when sliding the window down (could be avoided with 32
  * bit values at the expense of memory usage). We slide even when level == 0 to
@@ -227,6 +234,11 @@ local void slide_hash(s)
 #endif
 }
 
+#ifdef Z_POWER_OPT
+#undef slide_hash
+#include "contrib/power/slide_hash_resolver.c"
+#endif /* Z_POWER_OPT */
+
 /* ========================================================================= */
 int ZEXPORT deflateInit_(strm, level, version, stream_size)
     z_streamp strm;