Updated CRC32 Power Optimization Changes

Summary:
Support for PowerPC Architecture
Detecting AltiVec Support
Closes https://github.com/facebook/rocksdb/pull/2716

Differential Revision: D5606836

Pulled By: siying

fbshipit-source-id: 720262453b1546e5fdbbc668eff56848164113f3

Makefile

@@ -96,6 +96,18 @@ OPT += -momit-leaf-frame-pointer
 endif
 endif
 
+ifeq (,$(shell $(CXX) -fsyntax-only -maltivec -xc /dev/null 2>&1))
+CXXFLAGS += -DHAS_ALTIVEC
+CFLAGS += -DHAS_ALTIVEC
+HAS_ALTIVEC=1
+endif
+
+ifeq (,$(shell $(CXX) -fsyntax-only -mcpu=power8 -xc /dev/null 2>&1))
+CXXFLAGS += -DHAVE_POWER8
+CFLAGS +=  -DHAVE_POWER8
+HAVE_POWER8=1
+endif
+
 # if we're compiling for release, compile without debug code (-DNDEBUG) and
 # don't treat warnings as errors
 ifeq ($(DEBUG_LEVEL),0)
@@ -322,6 +334,12 @@ util/build_version.cc: FORCE
 endif
 
 LIBOBJECTS = $(LIB_SOURCES:.cc=.o)
+ifeq ($(HAVE_POWER8),1)
+LIB_CC_OBJECTS = $(LIB_SOURCES:.cc=.o)
+LIBOBJECTS += $(LIB_SOURCES_C:.c=.o)
+LIBOBJECTS += $(LIB_SOURCES_ASM:.S=.o)
+endif
+
 LIBOBJECTS += $(TOOL_LIB_SOURCES:.cc=.o)
 MOCKOBJECTS = $(MOCK_LIB_SOURCES:.cc=.o)
 
@@ -573,14 +591,35 @@ $(SHARED2): $(SHARED4)
 $(SHARED3): $(SHARED4)
 	ln -fs $(SHARED4) $(SHARED3)
 endif
-
+ifeq ($(HAVE_POWER8),1)
+SHARED_C_OBJECTS = $(LIB_SOURCES_C:.c=.o)
+SHARED_ASM_OBJECTS = $(LIB_SOURCES_ASM:.S=.o)
+SHARED_C_LIBOBJECTS = $(patsubst %.o,shared-objects/%.o,$(SHARED_C_OBJECTS))
+SHARED_ASM_LIBOBJECTS = $(patsubst %.o,shared-objects/%.o,$(SHARED_ASM_OBJECTS))
+shared_libobjects = $(patsubst %,shared-objects/%,$(LIB_CC_OBJECTS))
+else
 shared_libobjects = $(patsubst %,shared-objects/%,$(LIBOBJECTS))
+endif
+
 CLEAN_FILES += shared-objects
+shared_all_libobjects = $(shared_libobjects)
 
+ifeq ($(HAVE_POWER8),1)
+shared-ppc-objects = $(SHARED_C_LIBOBJECTS) $(SHARED_ASM_LIBOBJECTS)
+
+shared-objects/util/crc32c_ppc.o: util/crc32c_ppc.c
+	$(AM_V_CC)$(CC) $(CFLAGS) -c $< -o $@
+
+shared-objects/util/crc32c_ppc_asm.o: util/crc32c_ppc_asm.S
+	$(AM_V_CC)$(CC) $(CFLAGS) -c $< -o $@
+endif
 $(shared_libobjects): shared-objects/%.o: %.cc
 	$(AM_V_CC)mkdir -p $(@D) && $(CXX) $(CXXFLAGS) $(PLATFORM_SHARED_CFLAGS) -c $< -o $@
 
-$(SHARED4): $(shared_libobjects)
+ifeq ($(HAVE_POWER8),1)
+shared_all_libobjects = $(shared_libobjects) $(shared-ppc-objects)
+endif
+$(SHARED4): $(shared_all_libobjects)
 	$(CXX) $(PLATFORM_SHARED_LDFLAGS)$(SHARED3) $(CXXFLAGS) $(PLATFORM_SHARED_CFLAGS) $(shared_libobjects) $(LDFLAGS) -o $@
 
 endif  # PLATFORM_SHARED_EXT
@@ -1665,13 +1704,36 @@ rocksdbjavastaticpublishcentral:
 	mvn gpg:sign-and-deploy-file -Durl=https://oss.sonatype.org/service/local/staging/deploy/maven2/ -DrepositoryId=sonatype-nexus-staging -DpomFile=java/rocksjni.pom -Dfile=java/target/rocksdbjni-$(ROCKSDB_MAJOR).$(ROCKSDB_MINOR).$(ROCKSDB_PATCH).jar
 
 # A version of each $(LIBOBJECTS) compiled with -fPIC
+ifeq ($(HAVE_POWER8),1)
+JAVA_CC_OBJECTS = $(SHARED_CC_OBJECTS)
+JAVA_C_OBJECTS = $(SHARED_C_OBJECTS)
+JAVA_ASM_OBJECTS = $(SHARED_ASM_OBJECTS)
+
+JAVA_C_LIBOBJECTS = $(patsubst %.c.o,jl/%.c.o,$(JAVA_C_OBJECTS))
+JAVA_ASM_LIBOBJECTS = $(patsubst %.S.o,jl/%.S.o,$(JAVA_ASM_OBJECTS))
+endif
+
 java_libobjects = $(patsubst %,jl/%,$(LIBOBJECTS))
 CLEAN_FILES += jl
+java_all_libobjects = $(java_libobjects)
+
+ifeq ($(HAVE_POWER8),1)
+java_ppc_libobjects = $(JAVA_C_LIBOBJECTS) $(JAVA_ASM_LIBOBJECTS)
+
+jl/crc32c_ppc.o: util/crc32c_ppc.c
+	$(AM_V_CC)$(CC) $(CFLAGS) -c $< -o $@
+
+jl/crc32c_ppc_asm.o: util/crc32c_ppc_asm.S
+	$(AM_V_CC)$(CC) $(CFLAGS) -c $< -o $@
+java_all_libobjects += $(java_ppc_libobjects)
+endif
 
 $(java_libobjects): jl/%.o: %.cc
 	$(AM_V_CC)mkdir -p $(@D) && $(CXX) $(CXXFLAGS) -fPIC -c $< -o $@ $(COVERAGEFLAGS)
 
-rocksdbjava: $(java_libobjects)
+
+
+rocksdbjava: $(java_all_libobjects)
 	$(AM_V_GEN)cd java;$(MAKE) javalib;
 	$(AM_V_at)rm -f ./java/target/$(ROCKSDBJNILIB)
 	$(AM_V_at)$(CXX) $(CXXFLAGS) -I./java/. $(JAVA_INCLUDE) -shared -fPIC -o ./java/target/$(ROCKSDBJNILIB) $(JNI_NATIVE_SOURCES) $(java_libobjects) $(JAVA_LDFLAGS) $(COVERAGEFLAGS)
@@ -1726,30 +1788,54 @@ IOSVERSION=$(shell defaults read $(PLATFORMSROOT)/iPhoneOS.platform/version CFBu
 	lipo ios-x86/$@ ios-arm/$@ -create -output $@
 
 else
+ifeq ($(HAVE_POWER8),1)
+util/crc32c_ppc.o: util/crc32c_ppc.c
+	$(AM_V_CC)$(CC) $(CFLAGS) -c $< -o $@
+
+util/crc32c_ppc_asm.o: util/crc32c_ppc_asm.S
+	$(AM_V_CC)$(CC) $(CFLAGS) -c $< -o $@
+endif
 .cc.o:
 	$(AM_V_CC)$(CXX) $(CXXFLAGS) -c $< -o $@ $(COVERAGEFLAGS)
 
 .c.o:
 	$(AM_V_CC)$(CC) $(CFLAGS) -c $< -o $@
 endif
-
 # ---------------------------------------------------------------------------
 #  	Source files dependencies detection
 # ---------------------------------------------------------------------------
 
 all_sources = $(LIB_SOURCES) $(MAIN_SOURCES) $(MOCK_LIB_SOURCES) $(TOOL_LIB_SOURCES) $(BENCH_LIB_SOURCES) $(TEST_LIB_SOURCES) $(EXP_LIB_SOURCES)
-DEPFILES = $(all_sources:.cc=.d)
+DEPFILES = $(all_sources:.cc=.cc.d)
 
 # Add proper dependency support so changing a .h file forces a .cc file to
 # rebuild.
 
 # The .d file indicates .cc file's dependencies on .h files. We generate such
 # dependency by g++'s -MM option, whose output is a make dependency rule.
-$(DEPFILES): %.d: %.cc
+%.cc.d: %.cc
 	@$(CXX) $(CXXFLAGS) $(PLATFORM_SHARED_CFLAGS) \
 	  -MM -MT'$@' -MT'$(<:.cc=.o)' "$<" -o '$@'
 
+ifeq ($(HAVE_POWER8),1)
+DEPFILES_C = $(LIB_SOURCES_C:.c=.c.d)
+DEPFILES_ASM = $(LIB_SOURCES_ASM:.S=.S.d)
+
+%.c.d: %.c
+	@$(CXX) $(CXXFLAGS) $(PLATFORM_SHARED_CFLAGS) \
+	  -MM -MT'$@' -MT'$(<:.c=.o)' "$<" -o '$@'
+
+%.S.d: %.S
+	@$(CXX) $(CXXFLAGS) $(PLATFORM_SHARED_CFLAGS) \
+	  -MM -MT'$@' -MT'$(<:.S=.o)' "$<" -o '$@'
+
+$(DEPFILES_C): %.c.d
+
+$(DEPFILES_ASM): %.S.d
+depend: $(DEPFILES) $(DEPFILES_C) $(DEPFILES_ASM)
+else
 depend: $(DEPFILES)
+endif
 
 # if the make goal is either "clean" or "format", we shouldn't
 # try to import the *.d files.

db/db_impl.cc

@@ -133,8 +133,8 @@ void DumpSupportInfo(Logger* logger) {
   ROCKS_LOG_HEADER(logger, "\tZSTDNotFinal supported: %d",
                    ZSTDNotFinal_Supported());
   ROCKS_LOG_HEADER(logger, "\tZSTD supported: %d", ZSTD_Supported());
-  ROCKS_LOG_HEADER(logger, "Fast CRC32 supported: %d",
-                   crc32c::IsFastCrc32Supported());
+  ROCKS_LOG_HEADER(logger, "Fast CRC32 supported: %s",
+                   crc32c::IsFastCrc32Supported().c_str());
 }
 
 int64_t kDefaultLowPriThrottledRate = 2 * 1024 * 1024;

db/db_test_util.cc

@@ -288,12 +288,11 @@ Options DBTestBase::GetOptions(
   Options options = default_options;
   BlockBasedTableOptions table_options;
   bool set_block_based_table_factory = true;
-#if !defined(OS_MACOSX) && !defined(OS_WIN) && !defined(OS_SOLARIS) &&  \
-  !defined(OS_AIX)
+#if !defined(OS_MACOSX) && !defined(OS_WIN) && !defined(OS_SOLARIS) && \
+    !defined(OS_AIX)
   rocksdb::SyncPoint::GetInstance()->ClearCallBack(
       "NewRandomAccessFile:O_DIRECT");
-  rocksdb::SyncPoint::GetInstance()->ClearCallBack(
-      "NewWritableFile:O_DIRECT");
+  rocksdb::SyncPoint::GetInstance()->ClearCallBack("NewWritableFile:O_DIRECT");
 #endif
 
   bool can_allow_mmap = IsMemoryMappedAccessSupported();

src.mk

@@ -205,6 +205,16 @@ LIB_SOURCES =                                                   \
   utilities/write_batch_with_index/write_batch_with_index.cc    \
   utilities/write_batch_with_index/write_batch_with_index_internal.cc    \
 
+ifeq (,$(shell $(CXX) -fsyntax-only -maltivec -xc /dev/null 2>&1))
+LIB_SOURCES_ASM =\
+  util/crc32c_ppc_asm.S
+LIB_SOURCES_C = \
+  util/crc32c_ppc.c
+else
+LIB_SOURCES_ASM =
+LIB_SOURCES_C =
+endif
+
 TOOL_LIB_SOURCES = \
   tools/ldb_cmd.cc                                               \
   tools/ldb_tool.cc                                              \

tools/db_stress.cc

@@ -2422,15 +2422,15 @@ int main(int argc, char** argv) {
 #if !defined(NDEBUG) && !defined(OS_MACOSX) && !defined(OS_WIN) && \
   !defined(OS_SOLARIS) && !defined(OS_AIX)
   rocksdb::SyncPoint::GetInstance()->SetCallBack(
-    "NewWritableFile:O_DIRECT", [&](void* arg) {
-      int* val = static_cast<int*>(arg);
-      *val &= ~O_DIRECT;
-    });
+      "NewWritableFile:O_DIRECT", [&](void* arg) {
+        int* val = static_cast<int*>(arg);
+        *val &= ~O_DIRECT;
+      });
   rocksdb::SyncPoint::GetInstance()->SetCallBack(
-    "NewRandomAccessFile:O_DIRECT", [&](void* arg) {
-      int* val = static_cast<int*>(arg);
-      *val &= ~O_DIRECT;
-    });
+      "NewRandomAccessFile:O_DIRECT", [&](void* arg) {
+        int* val = static_cast<int*>(arg);
+        *val &= ~O_DIRECT;
+      });
   rocksdb::SyncPoint::GetInstance()->EnableProcessing();
 #endif
 

util/crc32c.cc

@@ -18,9 +18,32 @@
 #endif
 #include "util/coding.h"
 
+#ifdef __powerpc64__
+#include "util/crc32c_ppc.h"
+#include "util/crc32c_ppc_constants.h"
+
+#if __linux__
+#include <sys/auxv.h>
+
+#ifndef PPC_FEATURE2_VEC_CRYPTO
+#define PPC_FEATURE2_VEC_CRYPTO 0x02000000
+#endif
+
+#ifndef AT_HWCAP2
+#define AT_HWCAP2 26
+#endif
+
+#endif /* __linux__ */
+
+#endif
+
 namespace rocksdb {
 namespace crc32c {
 
+#ifdef __powerpc64__
+static int arch_ppc_crc32 = 0;
+#endif
+
 static const uint32_t table0_[256] = {
   0x00000000, 0xf26b8303, 0xe13b70f7, 0x1350f3f4,
   0xc79a971f, 0x35f1141c, 0x26a1e7e8, 0xd4ca64eb,
@@ -371,6 +394,7 @@ uint32_t ExtendImpl(uint32_t crc, const char* buf, size_t size) {
 }
 
 // Detect if SS42 or not.
+#ifndef HAVE_POWER8
 static bool isSSE42() {
 #ifndef HAVE_SSE42
   return false;
@@ -387,15 +411,67 @@ static bool isSSE42() {
   return false;
 #endif
 }
+#endif
 
 typedef uint32_t (*Function)(uint32_t, const char*, size_t);
 
+#if defined(HAVE_POWER8) && defined(HAS_ALTIVEC)
+uint32_t ExtendPPCImpl(uint32_t crc, const char *buf, size_t size) {
+  return crc32c_ppc(crc, (const unsigned char *)buf, size);
+}
+
+#if __linux__
+static int arch_ppc_probe(void) {
+  arch_ppc_crc32 = 0;
+
+#if defined(__powerpc64__)
+  if (getauxval(AT_HWCAP2) & PPC_FEATURE2_VEC_CRYPTO) arch_ppc_crc32 = 1;
+#endif /* __powerpc64__ */
+
+  return arch_ppc_crc32;
+}
+#endif  // __linux__
+
+static bool isAltiVec() {
+  if (arch_ppc_probe()) {
+    return true;
+  } else {
+    return false;
+  }
+}
+#endif
+
 static inline Function Choose_Extend() {
+#ifndef HAVE_POWER8
   return isSSE42() ? ExtendImpl<Fast_CRC32> : ExtendImpl<Slow_CRC32>;
+#else
+  return isAltiVec() ? ExtendPPCImpl : ExtendImpl<Slow_CRC32>;
+#endif
 }
 
-bool IsFastCrc32Supported() {
-  return isSSE42();
+std::string IsFastCrc32Supported() {
+  bool has_fast_crc = false;
+  std::string fast_zero_msg;
+  std::string arch;
+#ifdef HAVE_POWER8
+#ifdef HAS_ALTIVEC
+  if (arch_ppc_probe()) {
+    has_fast_crc = true;
+    arch = "PPC";
+  }
+#else
+  has_fast_crc = false;
+  arch = "PPC";
+#endif
+#else
+  has_fast_crc = isSSE42();
+  arch = "x86";
+#endif
+  if (has_fast_crc){
+    fast_zero_msg.append("Supported on " + arch);
+  }
+  fast_zero_msg.append("Not supported on " + arch);
+  return fast_zero_msg;  
 }
 
 Function ChosenExtend = Choose_Extend();

util/crc32c.h

@@ -10,11 +10,12 @@
 #pragma once
 #include <stddef.h>
 #include <stdint.h>
+#include <string>
 
 namespace rocksdb {
 namespace crc32c {
 
-extern bool IsFastCrc32Supported();
+extern std::string IsFastCrc32Supported();
 
 // Return the crc32c of concat(A, data[0,n-1]) where init_crc is the
 // crc32c of some string A.  Extend() is often used to maintain the

util/crc32c_ppc.c

+//  Copyright (c) 2017 International Business Machines Corp.
+//  All rights reserved.
+//  This source code is licensed under the BSD-style license found in the
+//  LICENSE file in the root directory of this source tree. An additional grant
+//  of patent rights can be found in the PATENTS file in the same directory.
+//  This source code is also licensed under the GPLv2 license found in the
+//  COPYING file in the root directory of this source tree.
+
+#define CRC_TABLE
+#include <inttypes.h>
+#include <stdlib.h>
+#include <strings.h>
+#include "util/crc32c_ppc_constants.h"
+
+#define VMX_ALIGN 16
+#define VMX_ALIGN_MASK (VMX_ALIGN - 1)
+
+#ifdef REFLECT
+static unsigned int crc32_align(unsigned int crc, unsigned char const *p,
+                                unsigned long len) {
+  while (len--) crc = crc_table[(crc ^ *p++) & 0xff] ^ (crc >> 8);
+  return crc;
+}
+#endif
+
+#ifdef HAVE_POWER8
+unsigned int __crc32_vpmsum(unsigned int crc, unsigned char const *p,
+                            unsigned long len);
+
+static uint32_t crc32_vpmsum(uint32_t crc, unsigned char const *data,
+                             unsigned len) {
+  unsigned int prealign;
+  unsigned int tail;
+
+#ifdef CRC_XOR
+  crc ^= 0xffffffff;
+#endif
+
+  if (len < VMX_ALIGN + VMX_ALIGN_MASK) {
+    crc = crc32_align(crc, data, (unsigned long)len);
+    goto out;
+  }
+
+  if ((unsigned long)data & VMX_ALIGN_MASK) {
+    prealign = VMX_ALIGN - ((unsigned long)data & VMX_ALIGN_MASK);
+    crc = crc32_align(crc, data, prealign);
+    len -= prealign;
+    data += prealign;
+  }
+
+  crc = __crc32_vpmsum(crc, data, (unsigned long)len & ~VMX_ALIGN_MASK);
+
+  tail = len & VMX_ALIGN_MASK;
+  if (tail) {
+    data += len & ~VMX_ALIGN_MASK;
+    crc = crc32_align(crc, data, tail);
+  }
+
+out:
+#ifdef CRC_XOR
+  crc ^= 0xffffffff;
+#endif
+
+  return crc;
+}
+
+/* This wrapper function works around the fact that crc32_vpmsum
+ * does not gracefully handle the case where the data pointer is NULL.  There
+ * may be room for performance improvement here.
+ */
+uint32_t crc32c_ppc(uint32_t crc, unsigned char const *data, unsigned len) {
+  unsigned char *buf2;
+
+  if (!data) {
+    buf2 = (unsigned char *)malloc(len);
+    bzero(buf2, len);
+    crc = crc32_vpmsum(crc, buf2, len);
+    free(buf2);
+  } else {
+    crc = crc32_vpmsum(crc, data, (unsigned long)len);
+  }
+  return crc;
+}
+
+#else /* HAVE_POWER8 */
+
+/* This symbol has to exist on non-ppc architectures (and on legacy
+ * ppc systems using power7 or below) in order to compile properly
+ * there, even though it won't be called.
+ */
+uint32_t crc32c_ppc(uint32_t crc, unsigned char const *data, unsigned len) {
+  return 0;
+}
+
+#endif /* HAVE_POWER8 */

util/crc32c_ppc.h

+//  Copyright (c) 2017 International Business Machines Corp.
+//  All rights reserved.
+//  This source code is licensed under the BSD-style license found in the
+//  LICENSE file in the root directory of this source tree. An additional grant
+//  of patent rights can be found in the PATENTS file in the same directory.
+//  This source code is also licensed under the GPLv2 license found in the
+//  COPYING file in the root directory of this source tree.
+
+#ifndef CRC32C_PPC_H
+#define CRC32C_PPC_H
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+extern uint32_t crc32c_ppc(uint32_t crc, unsigned char const *buffer,
+                           unsigned len);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif

util/crc32c_ppc_asm.S

+//  Copyright (c) 2015 Anton Blanchard <anton@au.ibm.com>, IBM
+//  Copyright (c) 2017 International Business Machines Corp.
+//  All rights reserved.
+//  This source code is licensed under the BSD-style license found in the
+//  LICENSE file in the root directory of this source tree. An additional grant
+//  of patent rights can be found in the PATENTS file in the same directory.
+//  This source code is also licensed under the GPLv2 license found in the
+//  COPYING file in the root directory of this source tree.
+
+#include <ppc-asm.h>
+#include "ppc-opcode.h"
+
+#undef toc
+
+#ifndef r1
+#define r1 1
+#endif
+
+#ifndef r2
+#define r2 2
+#endif
+
+	.section	.rodata
+.balign 16
+
+.byteswap_constant:
+	/* byte reverse permute constant */
+	.octa 0x0F0E0D0C0B0A09080706050403020100
+
+#define __ASSEMBLY__
+#include "crc32c_ppc_constants.h"
+
+	.text
+
+#if defined(__BIG_ENDIAN__) && defined(REFLECT)
+#define BYTESWAP_DATA
+#elif defined(__LITTLE_ENDIAN__) && !defined(REFLECT)
+#define BYTESWAP_DATA
+#else
+#undef BYTESWAP_DATA
+#endif
+
+#define off16		r25
+#define off32		r26
+#define off48		r27
+#define off64		r28
+#define off80		r29
+#define off96		r30
+#define off112		r31
+
+#define const1		v24
+#define const2		v25
+
+#define byteswap	v26
+#define	mask_32bit	v27
+#define	mask_64bit	v28
+#define zeroes		v29
+
+#ifdef BYTESWAP_DATA
+#define VPERM(A, B, C, D) vperm	A, B, C, D
+#else
+#define VPERM(A, B, C, D)
+#endif
+
+/* unsigned int __crc32_vpmsum(unsigned int crc, void *p, unsigned long len) */
+FUNC_START(__crc32_vpmsum)
+	std	r31,-8(r1)
+	std	r30,-16(r1)
+	std	r29,-24(r1)
+	std	r28,-32(r1)
+	std	r27,-40(r1)
+	std	r26,-48(r1)
+	std	r25,-56(r1)
+
+	li	off16,16
+	li	off32,32
+	li	off48,48
+	li	off64,64
+	li	off80,80
+	li	off96,96
+	li	off112,112
+	li	r0,0
+
+	/* Enough room for saving 10 non volatile VMX registers */
+	subi	r6,r1,56+10*16
+	subi	r7,r1,56+2*16
+
+	stvx	v20,0,r6
+	stvx	v21,off16,r6
+	stvx	v22,off32,r6
+	stvx	v23,off48,r6
+	stvx	v24,off64,r6
+	stvx	v25,off80,r6
+	stvx	v26,off96,r6
+	stvx	v27,off112,r6
+	stvx	v28,0,r7
+	stvx	v29,off16,r7
+
+	mr	r10,r3
+
+	vxor	zeroes,zeroes,zeroes
+	vspltisw v0,-1
+
+	vsldoi	mask_32bit,zeroes,v0,4
+	vsldoi	mask_64bit,zeroes,v0,8
+
+	/* Get the initial value into v8 */
+	vxor	v8,v8,v8
+	MTVRD(v8, r3)
+#ifdef REFLECT
+	vsldoi	v8,zeroes,v8,8	/* shift into bottom 32 bits */
+#else
+	vsldoi	v8,v8,zeroes,4	/* shift into top 32 bits */
+#endif
+
+#ifdef BYTESWAP_DATA
+	addis	r3,r2,.byteswap_constant@toc@ha
+	addi	r3,r3,.byteswap_constant@toc@l
+
+	lvx	byteswap,0,r3
+	addi	r3,r3,16
+#endif
+
+	cmpdi	r5,256
+	blt	.Lshort
+
+	rldicr	r6,r5,0,56
+
+	/* Checksum in blocks of MAX_SIZE */
+1:	lis	r7,MAX_SIZE@h
+	ori	r7,r7,MAX_SIZE@l
+	mr	r9,r7
+	cmpd	r6,r7
+	bgt	2f
+	mr	r7,r6
+2:	subf	r6,r7,r6
+
+	/* our main loop does 128 bytes at a time */
+	srdi	r7,r7,7
+
+	/*
+	 * Work out the offset into the constants table to start at. Each
+	 * constant is 16 bytes, and it is used against 128 bytes of input
+	 * data - 128 / 16 = 8
+	 */
+	sldi	r8,r7,4
+	srdi	r9,r9,3
+	subf	r8,r8,r9
+
+	/* We reduce our final 128 bytes in a separate step */
+	addi	r7,r7,-1
+	mtctr	r7
+
+	addis	r3,r2,.constants@toc@ha
+	addi	r3,r3,.constants@toc@l
+
+	/* Find the start of our constants */
+	add	r3,r3,r8
+
+	/* zero v0-v7 which will contain our checksums */
+	vxor	v0,v0,v0
+	vxor	v1,v1,v1
+	vxor	v2,v2,v2
+	vxor	v3,v3,v3
+	vxor	v4,v4,v4
+	vxor	v5,v5,v5
+	vxor	v6,v6,v6
+	vxor	v7,v7,v7
+
+	lvx	const1,0,r3
+
+	/*
+	 * If we are looping back to consume more data we use the values
+	 * already in v16-v23.
+	 */
+	cmpdi	r0,1
+	beq	2f
+
+	/* First warm up pass */
+	lvx	v16,0,r4
+	lvx	v17,off16,r4
+	VPERM(v16,v16,v16,byteswap)
+	VPERM(v17,v17,v17,byteswap)
+	lvx	v18,off32,r4
+	lvx	v19,off48,r4
+	VPERM(v18,v18,v18,byteswap)
+	VPERM(v19,v19,v19,byteswap)
+	lvx	v20,off64,r4
+	lvx	v21,off80,r4
+	VPERM(v20,v20,v20,byteswap)
+	VPERM(v21,v21,v21,byteswap)
+	lvx	v22,off96,r4
+	lvx	v23,off112,r4
+	VPERM(v22,v22,v22,byteswap)
+	VPERM(v23,v23,v23,byteswap)
+	addi	r4,r4,8*16
+
+	/* xor in initial value */
+	vxor	v16,v16,v8
+
+2:	bdz	.Lfirst_warm_up_done
+
+	addi	r3,r3,16
+	lvx	const2,0,r3
+
+	/* Second warm up pass */
+	VPMSUMD(v8,v16,const1)
+	lvx	v16,0,r4
+	VPERM(v16,v16,v16,byteswap)
+	ori	r2,r2,0
+
+	VPMSUMD(v9,v17,const1)
+	lvx	v17,off16,r4
+	VPERM(v17,v17,v17,byteswap)
+	ori	r2,r2,0
+
+	VPMSUMD(v10,v18,const1)
+	lvx	v18,off32,r4
+	VPERM(v18,v18,v18,byteswap)
+	ori	r2,r2,0
+
+	VPMSUMD(v11,v19,const1)
+	lvx	v19,off48,r4
+	VPERM(v19,v19,v19,byteswap)
+	ori	r2,r2,0
+
+	VPMSUMD(v12,v20,const1)
+	lvx	v20,off64,r4
+	VPERM(v20,v20,v20,byteswap)
+	ori	r2,r2,0
+
+	VPMSUMD(v13,v21,const1)
+	lvx	v21,off80,r4
+	VPERM(v21,v21,v21,byteswap)
+	ori	r2,r2,0
+
+	VPMSUMD(v14,v22,const1)
+	lvx	v22,off96,r4
+	VPERM(v22,v22,v22,byteswap)
+	ori	r2,r2,0
+
+	VPMSUMD(v15,v23,const1)
+	lvx	v23,off112,r4
+	VPERM(v23,v23,v23,byteswap)
+
+	addi	r4,r4,8*16
+
+	bdz	.Lfirst_cool_down
+
+	/*
+	 * main loop. We modulo schedule it such that it takes three iterations
+	 * to complete - first iteration load, second iteration vpmsum, third
+	 * iteration xor.
+	 */
+	.balign	16
+4:	lvx	const1,0,r3
+	addi	r3,r3,16
+	ori	r2,r2,0
+
+	vxor	v0,v0,v8
+	VPMSUMD(v8,v16,const2)
+	lvx	v16,0,r4
+	VPERM(v16,v16,v16,byteswap)
+	ori	r2,r2,0
+
+	vxor	v1,v1,v9
+	VPMSUMD(v9,v17,const2)
+	lvx	v17,off16,r4
+	VPERM(v17,v17,v17,byteswap)
+	ori	r2,r2,0
+
+	vxor	v2,v2,v10
+	VPMSUMD(v10,v18,const2)
+	lvx	v18,off32,r4
+	VPERM(v18,v18,v18,byteswap)
+	ori	r2,r2,0
+
+	vxor	v3,v3,v11
+	VPMSUMD(v11,v19,const2)
+	lvx	v19,off48,r4
+	VPERM(v19,v19,v19,byteswap)
+	lvx	const2,0,r3
+	ori	r2,r2,0
+
+	vxor	v4,v4,v12
+	VPMSUMD(v12,v20,const1)
+	lvx	v20,off64,r4
+	VPERM(v20,v20,v20,byteswap)
+	ori	r2,r2,0
+
+	vxor	v5,v5,v13
+	VPMSUMD(v13,v21,const1)
+	lvx	v21,off80,r4
+	VPERM(v21,v21,v21,byteswap)
+	ori	r2,r2,0
+
+	vxor	v6,v6,v14
+	VPMSUMD(v14,v22,const1)
+	lvx	v22,off96,r4
+	VPERM(v22,v22,v22,byteswap)
+	ori	r2,r2,0
+
+	vxor	v7,v7,v15
+	VPMSUMD(v15,v23,const1)
+	lvx	v23,off112,r4
+	VPERM(v23,v23,v23,byteswap)
+
+	addi	r4,r4,8*16
+
+	bdnz	4b
+
+.Lfirst_cool_down:
+	/* First cool down pass */
+	lvx	const1,0,r3
+	addi	r3,r3,16
+
+	vxor	v0,v0,v8
+	VPMSUMD(v8,v16,const1)
+	ori	r2,r2,0
+
+	vxor	v1,v1,v9
+	VPMSUMD(v9,v17,const1)
+	ori	r2,r2,0
+
+	vxor	v2,v2,v10
+	VPMSUMD(v10,v18,const1)
+	ori	r2,r2,0
+
+	vxor	v3,v3,v11
+	VPMSUMD(v11,v19,const1)
+	ori	r2,r2,0
+
+	vxor	v4,v4,v12
+	VPMSUMD(v12,v20,const1)
+	ori	r2,r2,0
+
+	vxor	v5,v5,v13
+	VPMSUMD(v13,v21,const1)
+	ori	r2,r2,0
+
+	vxor	v6,v6,v14
+	VPMSUMD(v14,v22,const1)
+	ori	r2,r2,0
+
+	vxor	v7,v7,v15
+	VPMSUMD(v15,v23,const1)
+	ori	r2,r2,0
+
+.Lsecond_cool_down:
+	/* Second cool down pass */
+	vxor	v0,v0,v8
+	vxor	v1,v1,v9
+	vxor	v2,v2,v10
+	vxor	v3,v3,v11
+	vxor	v4,v4,v12
+	vxor	v5,v5,v13
+	vxor	v6,v6,v14
+	vxor	v7,v7,v15
+
+#ifdef REFLECT
+	/*
+	 * vpmsumd produces a 96 bit result in the least significant bits
+	 * of the register. Since we are bit reflected we have to shift it
+	 * left 32 bits so it occupies the least significant bits in the
+	 * bit reflected domain.
+	 */
+	vsldoi	v0,v0,zeroes,4
+	vsldoi	v1,v1,zeroes,4
+	vsldoi	v2,v2,zeroes,4
+	vsldoi	v3,v3,zeroes,4
+	vsldoi	v4,v4,zeroes,4
+	vsldoi	v5,v5,zeroes,4
+	vsldoi	v6,v6,zeroes,4
+	vsldoi	v7,v7,zeroes,4
+#endif
+
+	/* xor with last 1024 bits */
+	lvx	v8,0,r4
+	lvx	v9,off16,r4
+	VPERM(v8,v8,v8,byteswap)
+	VPERM(v9,v9,v9,byteswap)
+	lvx	v10,off32,r4
+	lvx	v11,off48,r4
+	VPERM(v10,v10,v10,byteswap)
+	VPERM(v11,v11,v11,byteswap)
+	lvx	v12,off64,r4
+	lvx	v13,off80,r4
+	VPERM(v12,v12,v12,byteswap)
+	VPERM(v13,v13,v13,byteswap)
+	lvx	v14,off96,r4
+	lvx	v15,off112,r4
+	VPERM(v14,v14,v14,byteswap)
+	VPERM(v15,v15,v15,byteswap)
+
+	addi	r4,r4,8*16
+
+	vxor	v16,v0,v8
+	vxor	v17,v1,v9
+	vxor	v18,v2,v10
+	vxor	v19,v3,v11
+	vxor	v20,v4,v12
+	vxor	v21,v5,v13
+	vxor	v22,v6,v14
+	vxor	v23,v7,v15
+
+	li	r0,1
+	cmpdi	r6,0
+	addi	r6,r6,128
+	bne	1b
+
+	/* Work out how many bytes we have left */
+	andi.	r5,r5,127
+
+	/* Calculate where in the constant table we need to start */
+	subfic	r6,r5,128
+	add	r3,r3,r6
+
+	/* How many 16 byte chunks are in the tail */
+	srdi	r7,r5,4
+	mtctr	r7
+
+	/*
+	 * Reduce the previously calculated 1024 bits to 64 bits, shifting
+	 * 32 bits to include the trailing 32 bits of zeros
+	 */
+	lvx	v0,0,r3
+	lvx	v1,off16,r3
+	lvx	v2,off32,r3
+	lvx	v3,off48,r3
+	lvx	v4,off64,r3
+	lvx	v5,off80,r3
+	lvx	v6,off96,r3
+	lvx	v7,off112,r3
+	addi	r3,r3,8*16
+
+	VPMSUMW(v0,v16,v0)
+	VPMSUMW(v1,v17,v1)
+	VPMSUMW(v2,v18,v2)
+	VPMSUMW(v3,v19,v3)
+	VPMSUMW(v4,v20,v4)
+	VPMSUMW(v5,v21,v5)
+	VPMSUMW(v6,v22,v6)
+	VPMSUMW(v7,v23,v7)
+
+	/* Now reduce the tail (0 - 112 bytes) */
+	cmpdi	r7,0
+	beq	1f
+
+	lvx	v16,0,r4
+	lvx	v17,0,r3
+	VPERM(v16,v16,v16,byteswap)
+	VPMSUMW(v16,v16,v17)
+	vxor	v0,v0,v16
+	bdz	1f
+
+	lvx	v16,off16,r4
+	lvx	v17,off16,r3
+	VPERM(v16,v16,v16,byteswap)
+	VPMSUMW(v16,v16,v17)
+	vxor	v0,v0,v16
+	bdz	1f
+
+	lvx	v16,off32,r4
+	lvx	v17,off32,r3
+	VPERM(v16,v16,v16,byteswap)
+	VPMSUMW(v16,v16,v17)
+	vxor	v0,v0,v16
+	bdz	1f
+
+	lvx	v16,off48,r4
+	lvx	v17,off48,r3
+	VPERM(v16,v16,v16,byteswap)
+	VPMSUMW(v16,v16,v17)
+	vxor	v0,v0,v16
+	bdz	1f
+
+	lvx	v16,off64,r4
+	lvx	v17,off64,r3
+	VPERM(v16,v16,v16,byteswap)
+	VPMSUMW(v16,v16,v17)
+	vxor	v0,v0,v16
+	bdz	1f
+
+	lvx	v16,off80,r4
+	lvx	v17,off80,r3
+	VPERM(v16,v16,v16,byteswap)
+	VPMSUMW(v16,v16,v17)
+	vxor	v0,v0,v16
+	bdz	1f
+
+	lvx	v16,off96,r4
+	lvx	v17,off96,r3
+	VPERM(v16,v16,v16,byteswap)
+	VPMSUMW(v16,v16,v17)
+	vxor	v0,v0,v16
+
+	/* Now xor all the parallel chunks together */
+1:	vxor	v0,v0,v1
+	vxor	v2,v2,v3
+	vxor	v4,v4,v5
+	vxor	v6,v6,v7
+
+	vxor	v0,v0,v2
+	vxor	v4,v4,v6
+
+	vxor	v0,v0,v4
+
+.Lbarrett_reduction:
+	/* Barrett constants */
+	addis	r3,r2,.barrett_constants@toc@ha
+	addi	r3,r3,.barrett_constants@toc@l
+
+	lvx	const1,0,r3
+	lvx	const2,off16,r3
+
+	vsldoi	v1,v0,v0,8
+	vxor	v0,v0,v1		/* xor two 64 bit results together */
+
+#ifdef REFLECT
+	/* shift left one bit */
+	vspltisb v1,1
+	vsl	v0,v0,v1
+#endif
+
+	vand	v0,v0,mask_64bit
+
+#ifndef REFLECT
+	/*
+	 * Now for the Barrett reduction algorithm. The idea is to calculate q,
+	 * the multiple of our polynomial that we need to subtract. By
+	 * doing the computation 2x bits higher (ie 64 bits) and shifting the
+	 * result back down 2x bits, we round down to the nearest multiple.
+	 */
+	VPMSUMD(v1,v0,const1)	/* ma */
+	vsldoi	v1,zeroes,v1,8	/* q = floor(ma/(2^64)) */
+	VPMSUMD(v1,v1,const2)	/* qn */
+	vxor	v0,v0,v1	/* a - qn, subtraction is xor in GF(2) */
+
+	/*
+	 * Get the result into r3. We need to shift it left 8 bytes:
+	 * V0 [ 0 1 2 X ]
+	 * V0 [ 0 X 2 3 ]
+	 */
+	vsldoi	v0,v0,zeroes,8	/* shift result into top 64 bits */
+#else
+	/*
+	 * The reflected version of Barrett reduction. Instead of bit
+	 * reflecting our data (which is expensive to do), we bit reflect our
+	 * constants and our algorithm, which means the intermediate data in
+	 * our vector registers goes from 0-63 instead of 63-0. We can reflect
+	 * the algorithm because we don't carry in mod 2 arithmetic.
+	 */
+	vand	v1,v0,mask_32bit	/* bottom 32 bits of a */
+	VPMSUMD(v1,v1,const1)		/* ma */
+	vand	v1,v1,mask_32bit	/* bottom 32bits of ma */
+	VPMSUMD(v1,v1,const2)		/* qn */
+	vxor	v0,v0,v1		/* a - qn, subtraction is xor in GF(2) */
+
+	/*
+	 * Since we are bit reflected, the result (ie the low 32 bits) is in
+	 * the high 32 bits. We just need to shift it left 4 bytes
+	 * V0 [ 0 1 X 3 ]
+	 * V0 [ 0 X 2 3 ]
+	 */
+	vsldoi	v0,v0,zeroes,4		/* shift result into top 64 bits of */
+#endif
+
+	/* Get it into r3 */
+	MFVRD(r3, v0)
+
+.Lout:
+	subi	r6,r1,56+10*16
+	subi	r7,r1,56+2*16
+
+	lvx	v20,0,r6
+	lvx	v21,off16,r6
+	lvx	v22,off32,r6
+	lvx	v23,off48,r6
+	lvx	v24,off64,r6
+	lvx	v25,off80,r6
+	lvx	v26,off96,r6
+	lvx	v27,off112,r6
+	lvx	v28,0,r7
+	lvx	v29,off16,r7
+
+	ld	r31,-8(r1)
+	ld	r30,-16(r1)
+	ld	r29,-24(r1)
+	ld	r28,-32(r1)
+	ld	r27,-40(r1)
+	ld	r26,-48(r1)
+	ld	r25,-56(r1)
+
+	blr
+
+.Lfirst_warm_up_done:
+	lvx	const1,0,r3
+	addi	r3,r3,16
+
+	VPMSUMD(v8,v16,const1)
+	VPMSUMD(v9,v17,const1)
+	VPMSUMD(v10,v18,const1)
+	VPMSUMD(v11,v19,const1)
+	VPMSUMD(v12,v20,const1)
+	VPMSUMD(v13,v21,const1)
+	VPMSUMD(v14,v22,const1)
+	VPMSUMD(v15,v23,const1)
+
+	b	.Lsecond_cool_down
+
+.Lshort:
+	cmpdi	r5,0
+	beq	.Lzero
+
+	addis	r3,r2,.short_constants@toc@ha
+	addi	r3,r3,.short_constants@toc@l
+
+	/* Calculate where in the constant table we need to start */
+	subfic	r6,r5,256
+	add	r3,r3,r6
+
+	/* How many 16 byte chunks? */
+	srdi	r7,r5,4
+	mtctr	r7
+
+	vxor	v19,v19,v19
+	vxor	v20,v20,v20
+
+	lvx	v0,0,r4
+	lvx	v16,0,r3
+	VPERM(v0,v0,v16,byteswap)
+	vxor	v0,v0,v8	/* xor in initial value */
+	VPMSUMW(v0,v0,v16)
+	bdz	.Lv0
+
+	lvx	v1,off16,r4
+	lvx	v17,off16,r3
+	VPERM(v1,v1,v17,byteswap)
+	VPMSUMW(v1,v1,v17)
+	bdz	.Lv1
+
+	lvx	v2,off32,r4
+	lvx	v16,off32,r3
+	VPERM(v2,v2,v16,byteswap)
+	VPMSUMW(v2,v2,v16)
+	bdz	.Lv2
+
+	lvx	v3,off48,r4
+	lvx	v17,off48,r3
+	VPERM(v3,v3,v17,byteswap)
+	VPMSUMW(v3,v3,v17)
+	bdz	.Lv3
+
+	lvx	v4,off64,r4
+	lvx	v16,off64,r3
+	VPERM(v4,v4,v16,byteswap)
+	VPMSUMW(v4,v4,v16)
+	bdz	.Lv4
+
+	lvx	v5,off80,r4
+	lvx	v17,off80,r3
+	VPERM(v5,v5,v17,byteswap)
+	VPMSUMW(v5,v5,v17)
+	bdz	.Lv5
+
+	lvx	v6,off96,r4
+	lvx	v16,off96,r3
+	VPERM(v6,v6,v16,byteswap)
+	VPMSUMW(v6,v6,v16)
+	bdz	.Lv6
+
+	lvx	v7,off112,r4
+	lvx	v17,off112,r3
+	VPERM(v7,v7,v17,byteswap)
+	VPMSUMW(v7,v7,v17)
+	bdz	.Lv7
+
+	addi	r3,r3,128
+	addi	r4,r4,128
+
+	lvx	v8,0,r4
+	lvx	v16,0,r3
+	VPERM(v8,v8,v16,byteswap)
+	VPMSUMW(v8,v8,v16)
+	bdz	.Lv8
+
+	lvx	v9,off16,r4
+	lvx	v17,off16,r3
+	VPERM(v9,v9,v17,byteswap)
+	VPMSUMW(v9,v9,v17)
+	bdz	.Lv9
+
+	lvx	v10,off32,r4
+	lvx	v16,off32,r3
+	VPERM(v10,v10,v16,byteswap)
+	VPMSUMW(v10,v10,v16)
+	bdz	.Lv10
+
+	lvx	v11,off48,r4
+	lvx	v17,off48,r3
+	VPERM(v11,v11,v17,byteswap)
+	VPMSUMW(v11,v11,v17)
+	bdz	.Lv11
+
+	lvx	v12,off64,r4
+	lvx	v16,off64,r3
+	VPERM(v12,v12,v16,byteswap)
+	VPMSUMW(v12,v12,v16)
+	bdz	.Lv12
+
+	lvx	v13,off80,r4
+	lvx	v17,off80,r3
+	VPERM(v13,v13,v17,byteswap)
+	VPMSUMW(v13,v13,v17)
+	bdz	.Lv13
+
+	lvx	v14,off96,r4
+	lvx	v16,off96,r3
+	VPERM(v14,v14,v16,byteswap)
+	VPMSUMW(v14,v14,v16)
+	bdz	.Lv14
+
+	lvx	v15,off112,r4
+	lvx	v17,off112,r3
+	VPERM(v15,v15,v17,byteswap)
+	VPMSUMW(v15,v15,v17)
+
+.Lv15:	vxor	v19,v19,v15
+.Lv14:	vxor	v20,v20,v14
+.Lv13:	vxor	v19,v19,v13
+.Lv12:	vxor	v20,v20,v12
+.Lv11:	vxor	v19,v19,v11
+.Lv10:	vxor	v20,v20,v10
+.Lv9:	vxor	v19,v19,v9
+.Lv8:	vxor	v20,v20,v8
+.Lv7:	vxor	v19,v19,v7
+.Lv6:	vxor	v20,v20,v6
+.Lv5:	vxor	v19,v19,v5
+.Lv4:	vxor	v20,v20,v4
+.Lv3:	vxor	v19,v19,v3
+.Lv2:	vxor	v20,v20,v2
+.Lv1:	vxor	v19,v19,v1
+.Lv0:	vxor	v20,v20,v0
+
+	vxor	v0,v19,v20
+
+	b	.Lbarrett_reduction
+
+.Lzero:
+	mr	r3,r10
+	b	.Lout
+
+FUNC_END(__crc32_vpmsum)

util/ppc-opcode.h

+//  Copyright (c) 2017 International Business Machines Corp.
+//  All rights reserved.
+//  This source code is licensed under the BSD-style license found in the
+//  LICENSE file in the root directory of this source tree. An additional grant
+//  of patent rights can be found in the PATENTS file in the same directory.
+//  This source code is also licensed under the GPLv2 license found in the
+//  COPYING file in the root directory of this source tree.
+
+#ifndef __OPCODES_H
+#define __OPCODES_H
+
+#define __PPC_RA(a) (((a)&0x1f) << 16)
+#define __PPC_RB(b) (((b)&0x1f) << 11)
+#define __PPC_XA(a) ((((a)&0x1f) << 16) | (((a)&0x20) >> 3))
+#define __PPC_XB(b) ((((b)&0x1f) << 11) | (((b)&0x20) >> 4))
+#define __PPC_XS(s) ((((s)&0x1f) << 21) | (((s)&0x20) >> 5))
+#define __PPC_XT(s) __PPC_XS(s)
+#define VSX_XX3(t, a, b) (__PPC_XT(t) | __PPC_XA(a) | __PPC_XB(b))
+#define VSX_XX1(s, a, b) (__PPC_XS(s) | __PPC_RA(a) | __PPC_RB(b))
+
+#define PPC_INST_VPMSUMW 0x10000488
+#define PPC_INST_VPMSUMD 0x100004c8
+#define PPC_INST_MFVSRD 0x7c000066
+#define PPC_INST_MTVSRD 0x7c000166
+
+#define VPMSUMW(t, a, b) .long PPC_INST_VPMSUMW | VSX_XX3((t), a, b)
+#define VPMSUMD(t, a, b) .long PPC_INST_VPMSUMD | VSX_XX3((t), a, b)
+#define MFVRD(a, t) .long PPC_INST_MFVSRD | VSX_XX1((t) + 32, a, 0)
+#define MTVRD(t, a) .long PPC_INST_MTVSRD | VSX_XX1((t) + 32, a, 0)
+
+#endif