Flatten ZSTD_row_getMatchMask (#2681)

* Flatten ZSTD_row_getMatchMask

* Remove the SIMD abstraction layer.
* Add big endian support.
* Align `hashTags` within `tagRow` to a 16-byte boundary. 
* Switch SSE2 to use aligned reads.
* Optimize scalar path using SWAR.
* Optimize neon path for `n == 32`
* Work around minor clang issue for NEON (https://bugs.llvm.org/show_bug.cgi?id=49577)

* replace memcpy with MEM_readST

* silence alignment warnings

* fix neon casts

* Update zstd_lazy.c

* unify simd preprocessor detection (#3)

* remove duplicate asserts

* tweak rotates

* improve endian detection

* add cast

there is a fun little catch-22 with gcc: result from pmovmskb has to be cast to uint32_t to avoid a zero-extension
but must be uint16_t to get gcc to generate a rotate instruction..

* more casts

* fix casts

better work-around for the (bogus) warning: unary minus on unsigned

lib/common/compiler.h

@@ -197,6 +197,22 @@
     #define STATIC_BMI2 0
 #endif
 
+/* compile time determination of SIMD support */
+#if !defined(ZSTD_NO_INTRINSICS)
+#  if defined(__SSE2__) || defined(_M_AMD64) || (defined (_M_IX86) && defined(_M_IX86_FP) && (_M_IX86_FP >= 2))
+#    define ZSTD_ARCH_X86_SSE2
+#  endif
+#  if defined(__ARM_NEON) || defined(_M_ARM64)
+#    define ZSTD_ARCH_ARM_NEON
+#  endif
+#
+#  if defined(ZSTD_ARCH_X86_SSE2)
+#    include <emmintrin.h>
+#  elif defined(ZSTD_ARCH_ARM_NEON)
+#    include <arm_neon.h>
+#  endif
+#endif  
+
 /* compat. with non-clang compilers */
 #ifndef __has_builtin
 #  define __has_builtin(x) 0

lib/common/mem.h

@@ -153,8 +153,22 @@ MEM_STATIC unsigned MEM_64bits(void) { return sizeof(size_t)==8; }
 
 MEM_STATIC unsigned MEM_isLittleEndian(void)
 {
+#if defined(__BYTE_ORDER__) && defined(__ORDER_LITTLE_ENDIAN__) && (__BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__)
+    return 1;
+#elif defined(__BYTE_ORDER__) && defined(__ORDER_BIG_ENDIAN__) && (__BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
+    return 0;
+#elif defined(__clang__) && __LITTLE_ENDIAN__
+    return 1;
+#elif defined(__clang__) && __BIG_ENDIAN__
+    return 0;
+#elif defined(_MSC_VER) && (_M_AMD64 || _M_IX86)
+    return 1;
+#elif defined(__DMC__) && defined(_M_IX86)
+    return 1;
+#else
     const union { U32 u; BYTE c[4]; } one = { 1 };   /* don't use static : performance detrimental  */
     return one.c[0];
+#endif
 }
 
 #if defined(MEM_FORCE_MEMORY_ACCESS) && (MEM_FORCE_MEMORY_ACCESS==2)

lib/common/zstd_internal.h

@@ -19,9 +19,6 @@
 /*-*************************************
 *  Dependencies
 ***************************************/
-#if !defined(ZSTD_NO_INTRINSICS) && defined(__ARM_NEON)
-#include <arm_neon.h>
-#endif
 #include "compiler.h"
 #include "mem.h"
 #include "debug.h"                 /* assert, DEBUGLOG, RAWLOG, g_debuglevel */
@@ -247,7 +244,7 @@ static UNUSED_ATTR const U32 OF_defaultNormLog = OF_DEFAULTNORMLOG;
 *  Shared functions to include for inlining
 *********************************************/
 static void ZSTD_copy8(void* dst, const void* src) {
-#if !defined(ZSTD_NO_INTRINSICS) && defined(__ARM_NEON)
+#if defined(ZSTD_ARCH_ARM_NEON)
     vst1_u8((uint8_t*)dst, vld1_u8((const uint8_t*)src));
 #else
     ZSTD_memcpy(dst, src, 8);
@@ -256,7 +253,7 @@ static void ZSTD_copy8(void* dst, const void* src) {
 
 #define COPY8(d,s) { ZSTD_copy8(d,s); d+=8; s+=8; }
 static void ZSTD_copy16(void* dst, const void* src) {
-#if !defined(ZSTD_NO_INTRINSICS) && defined(__ARM_NEON)
+#if defined(ZSTD_ARCH_ARM_NEON)
     vst1q_u8((uint8_t*)dst, vld1q_u8((const uint8_t*)src));
 #else
     ZSTD_memcpy(dst, src, 16);

lib/compress/zstd_compress.c

@@ -222,7 +222,7 @@ static int ZSTD_rowMatchFinderUsed(const ZSTD_strategy strategy, const ZSTD_useR
 /* Returns row matchfinder usage enum given an initial mode and cParams */
 static ZSTD_useRowMatchFinderMode_e ZSTD_resolveRowMatchFinderMode(ZSTD_useRowMatchFinderMode_e mode,
                                                                    const ZSTD_compressionParameters* const cParams) {
-#if !defined(ZSTD_NO_INTRINSICS) && (defined(__SSE2__) || defined(_M_AMD64)  || defined(__ARM_NEON))
+#if defined(ZSTD_ARCH_X86_SSE2) || defined(ZSTD_ARCH_ARM_NEON)
     int const kHasSIMD128 = 1;
 #else
     int const kHasSIMD128 = 0;

lib/compress/zstd_lazy.c

@@ -865,7 +865,7 @@ FORCE_INLINE_TEMPLATE size_t ZSTD_HcFindBestMatch_extDict_selectMLS (
 * (SIMD) Row-based matchfinder
 ***********************************/
 /* Constants for row-based hash */
-#define ZSTD_ROW_HASH_TAG_OFFSET 1                               /* byte offset of hashes in the match state's tagTable from the beginning of a row */
+#define ZSTD_ROW_HASH_TAG_OFFSET 16                              /* byte offset of hashes in the match state's tagTable from the beginning of a row */
 #define ZSTD_ROW_HASH_TAG_BITS 8                                 /* nb bits to use for the tag */
 #define ZSTD_ROW_HASH_TAG_MASK ((1u << ZSTD_ROW_HASH_TAG_BITS) - 1)
 
@@ -873,197 +873,6 @@ FORCE_INLINE_TEMPLATE size_t ZSTD_HcFindBestMatch_extDict_selectMLS (
 
 typedef U32 ZSTD_VecMask;   /* Clarifies when we are interacting with a U32 representing a mask of matches */
 
-#if !defined(ZSTD_NO_INTRINSICS) && (defined(__SSE2__) || defined(_M_AMD64)) /* SIMD SSE version*/
-
-#include <emmintrin.h>
-typedef __m128i ZSTD_Vec128;
-
-/* Returns a 128-bit container with 128-bits from src */
-static ZSTD_Vec128 ZSTD_Vec128_read(const void* const src) {
-  return _mm_loadu_si128((ZSTD_Vec128 const*)src);
-}
-
-/* Returns a ZSTD_Vec128 with the byte "val" packed 16 times */
-static ZSTD_Vec128 ZSTD_Vec128_set8(BYTE val) {
-  return _mm_set1_epi8((char)val);
-}
-
-/* Do byte-by-byte comparison result of x and y. Then collapse 128-bit resultant mask
- * into a 32-bit mask that is the MSB of each byte.
- * */
-static ZSTD_VecMask ZSTD_Vec128_cmpMask8(ZSTD_Vec128 x, ZSTD_Vec128 y) {
-  return (ZSTD_VecMask)_mm_movemask_epi8(_mm_cmpeq_epi8(x, y));
-}
-
-typedef struct {
-  __m128i fst;
-  __m128i snd;
-} ZSTD_Vec256;
-
-static ZSTD_Vec256 ZSTD_Vec256_read(const void* const ptr) {
-  ZSTD_Vec256 v;
-  v.fst = ZSTD_Vec128_read(ptr);
-  v.snd = ZSTD_Vec128_read((ZSTD_Vec128 const*)ptr + 1);
-  return v;
-}
-
-static ZSTD_Vec256 ZSTD_Vec256_set8(BYTE val) {
-  ZSTD_Vec256 v;
-  v.fst = ZSTD_Vec128_set8(val);
-  v.snd = ZSTD_Vec128_set8(val);
-  return v;
-}
-
-static ZSTD_VecMask ZSTD_Vec256_cmpMask8(ZSTD_Vec256 x, ZSTD_Vec256 y) {
-  ZSTD_VecMask fstMask;
-  ZSTD_VecMask sndMask;
-  fstMask = ZSTD_Vec128_cmpMask8(x.fst, y.fst);
-  sndMask = ZSTD_Vec128_cmpMask8(x.snd, y.snd);
-  return fstMask | (sndMask << 16);
-}
-
-#elif !defined(ZSTD_NO_INTRINSICS) && defined(__ARM_NEON) /* SIMD ARM NEON Version */
-
-#include <arm_neon.h>
-typedef uint8x16_t ZSTD_Vec128;
-
-static ZSTD_Vec128 ZSTD_Vec128_read(const void* const src) {
-  return vld1q_u8((const BYTE* const)src);
-}
-
-static ZSTD_Vec128 ZSTD_Vec128_set8(BYTE val) {
-  return vdupq_n_u8(val);
-}
-
-/* Mimics '_mm_movemask_epi8()' from SSE */
-static U32 ZSTD_vmovmaskq_u8(ZSTD_Vec128 val) {
-    /* Shift out everything but the MSB bits in each byte */
-    uint16x8_t highBits = vreinterpretq_u16_u8(vshrq_n_u8(val, 7));
-    /* Merge the even lanes together with vsra (right shift and add) */
-    uint32x4_t paired16 = vreinterpretq_u32_u16(vsraq_n_u16(highBits, highBits, 7));
-    uint64x2_t paired32 = vreinterpretq_u64_u32(vsraq_n_u32(paired16, paired16, 14));
-    uint8x16_t paired64 = vreinterpretq_u8_u64(vsraq_n_u64(paired32, paired32, 28));
-    /* Extract the low 8 bits from each lane, merge */
-    return vgetq_lane_u8(paired64, 0) | ((U32)vgetq_lane_u8(paired64, 8) << 8);
-}
-
-static ZSTD_VecMask ZSTD_Vec128_cmpMask8(ZSTD_Vec128 x, ZSTD_Vec128 y) {
-  return (ZSTD_VecMask)ZSTD_vmovmaskq_u8(vceqq_u8(x, y));
-}
-
-typedef struct {
-    uint8x16_t fst;
-    uint8x16_t snd;
-} ZSTD_Vec256;
-
-static ZSTD_Vec256 ZSTD_Vec256_read(const void* const ptr) {
-  ZSTD_Vec256 v;
-  v.fst = ZSTD_Vec128_read(ptr);
-  v.snd = ZSTD_Vec128_read((ZSTD_Vec128 const*)ptr + 1);
-  return v;
-}
-
-static ZSTD_Vec256 ZSTD_Vec256_set8(BYTE val) {
-  ZSTD_Vec256 v;
-  v.fst = ZSTD_Vec128_set8(val);
-  v.snd = ZSTD_Vec128_set8(val);
-  return v;
-}
-
-static ZSTD_VecMask ZSTD_Vec256_cmpMask8(ZSTD_Vec256 x, ZSTD_Vec256 y) {
-  ZSTD_VecMask fstMask;
-  ZSTD_VecMask sndMask;
-  fstMask = ZSTD_Vec128_cmpMask8(x.fst, y.fst);
-  sndMask = ZSTD_Vec128_cmpMask8(x.snd, y.snd);
-  return fstMask | (sndMask << 16);
-}
-
-#else /* Scalar fallback version */
-
-#define VEC128_NB_SIZE_T (16 / sizeof(size_t))
-typedef struct {
-    size_t vec[VEC128_NB_SIZE_T];
-} ZSTD_Vec128;
-
-static ZSTD_Vec128 ZSTD_Vec128_read(const void* const src) {
-    ZSTD_Vec128 ret;
-    ZSTD_memcpy(ret.vec, src, VEC128_NB_SIZE_T*sizeof(size_t));
-    return ret;
-}
-
-static ZSTD_Vec128 ZSTD_Vec128_set8(BYTE val) {
-    ZSTD_Vec128 ret = { {0} };
-    int startBit = sizeof(size_t) * 8 - 8;
-    for (;startBit >= 0; startBit -= 8) {
-        unsigned j = 0;
-        for (;j < VEC128_NB_SIZE_T; ++j) {
-            ret.vec[j] |= ((size_t)val << startBit);
-        }
-    }
-    return ret;
-}
-
-/* Compare x to y, byte by byte, generating a "matches" bitfield */
-static ZSTD_VecMask ZSTD_Vec128_cmpMask8(ZSTD_Vec128 x, ZSTD_Vec128 y) {
-    ZSTD_VecMask res = 0;
-    unsigned i = 0;
-    unsigned l = 0;
-    for (; i < VEC128_NB_SIZE_T; ++i) {
-        const size_t cmp1 = x.vec[i];
-        const size_t cmp2 = y.vec[i];
-        unsigned j = 0;
-        for (; j < sizeof(size_t); ++j, ++l) {
-            if (((cmp1 >> j*8) & 0xFF) == ((cmp2 >> j*8) & 0xFF)) {
-                res |= ((U32)1 << (j+i*sizeof(size_t)));
-            }
-        }
-    }
-    return res;
-}
-
-#define VEC256_NB_SIZE_T 2*VEC128_NB_SIZE_T
-typedef struct {
-    size_t vec[VEC256_NB_SIZE_T];
-} ZSTD_Vec256;
-
-static ZSTD_Vec256 ZSTD_Vec256_read(const void* const src) {
-    ZSTD_Vec256 ret;
-    ZSTD_memcpy(ret.vec, src, VEC256_NB_SIZE_T*sizeof(size_t));
-    return ret;
-}
-
-static ZSTD_Vec256 ZSTD_Vec256_set8(BYTE val) {
-    ZSTD_Vec256 ret = { {0} };
-    int startBit = sizeof(size_t) * 8 - 8;
-    for (;startBit >= 0; startBit -= 8) {
-        unsigned j = 0;
-        for (;j < VEC256_NB_SIZE_T; ++j) {
-            ret.vec[j] |= ((size_t)val << startBit);
-        }
-    }
-    return ret;
-}
-
-/* Compare x to y, byte by byte, generating a "matches" bitfield */
-static ZSTD_VecMask ZSTD_Vec256_cmpMask8(ZSTD_Vec256 x, ZSTD_Vec256 y) {
-    ZSTD_VecMask res = 0;
-    unsigned i = 0;
-    unsigned l = 0;
-    for (; i < VEC256_NB_SIZE_T; ++i) {
-        const size_t cmp1 = x.vec[i];
-        const size_t cmp2 = y.vec[i];
-        unsigned j = 0;
-        for (; j < sizeof(size_t); ++j, ++l) {
-            if (((cmp1 >> j*8) & 0xFF) == ((cmp2 >> j*8) & 0xFF)) {
-                res |= ((U32)1 << (j+i*sizeof(size_t)));
-            }
-        }
-    }
-    return res;
-}
-
-#endif /* !defined(ZSTD_NO_INTRINSICS) && defined(__SSE2__) */
-
 /* ZSTD_VecMask_next():
  * Starting from the LSB, returns the idx of the next non-zero bit.
  * Basically counting the nb of trailing zeroes.
@@ -1085,22 +894,22 @@ static U32 ZSTD_VecMask_next(ZSTD_VecMask val) {
 #   endif
 }
 
-/* ZSTD_VecMask_rotateRight():
+/* ZSTD_rotateRight_U32():
- * Rotates a bitfield to the right by "rotation" bits.
+ * Rotates a bitfield to the right by "count" bits.
- * If the rotation is greater than totalBits, the returned mask is 0.
+ * https://en.wikipedia.org/w/index.php?title=Circular_shift&oldid=991635599#Implementing_circular_shifts
  */
-FORCE_INLINE_TEMPLATE ZSTD_VecMask
+FORCE_INLINE_TEMPLATE
-ZSTD_VecMask_rotateRight(ZSTD_VecMask mask, U32 const rotation, U32 const totalBits) {
+U32 ZSTD_rotateRight_U32(U32 const value, U32 count) {
-  if (rotation == 0)
+    assert(count < 32);
-    return mask;
+    count &= 0x1F; /* for fickle pattern recognition */
-  switch (totalBits) {
+    return (value >> count) | (U32)(value << ((0U - count) & 0x1F));
-    default:
+}
-      assert(0);
+
-    case 16:
+FORCE_INLINE_TEMPLATE
-      return (mask >> rotation) | (U16)(mask << (16 - rotation));
+U16 ZSTD_rotateRight_U16(U16 const value, U32 count) {
-    case 32:
+    assert(count < 16);
-      return (mask >> rotation) | (U32)(mask << (32 - rotation));
+    count &= 0x0F; /* for fickle pattern recognition */
-  }
+    return (value >> count) | (U16)(value << ((0U - count) & 0x0F));
 }
 
 /* ZSTD_row_nextIndex():
@@ -1226,24 +1035,98 @@ void ZSTD_row_update(ZSTD_matchState_t* const ms, const BYTE* ip) {
 
 /* Returns a ZSTD_VecMask (U32) that has the nth bit set to 1 if the newly-computed "tag" matches
  * the hash at the nth position in a row of the tagTable.
- */
+ * Each row is a circular buffer beginning at the value of "head". So we must rotate the "matches" bitfield
+ * to match up with the actual layout of the entries within the hashTable */
 FORCE_INLINE_TEMPLATE
 ZSTD_VecMask ZSTD_row_getMatchMask(const BYTE* const tagRow, const BYTE tag, const U32 head, const U32 rowEntries) {
-    ZSTD_VecMask matches = 0;
+    const BYTE* const src = tagRow + ZSTD_ROW_HASH_TAG_OFFSET;
+    assert((rowEntries == 16) || (rowEntries == 32));
+#if defined(ZSTD_ARCH_X86_SSE2)
     if (rowEntries == 16) {
-        ZSTD_Vec128 hashes        = ZSTD_Vec128_read(tagRow + ZSTD_ROW_HASH_TAG_OFFSET);
+        const __m128i chunk = _mm_loadu_si128((const __m128i*)(const void*)src);
-        ZSTD_Vec128 expandedTags  = ZSTD_Vec128_set8(tag);
+        const __m128i equalMask = _mm_cmpeq_epi8(chunk, _mm_set1_epi8(tag));
-        matches                   = ZSTD_Vec128_cmpMask8(hashes, expandedTags);
+        const U16 matches = (U16)_mm_movemask_epi8(equalMask);
-    } else if (rowEntries == 32) {
+        return ZSTD_rotateRight_U16(matches, head);
-        ZSTD_Vec256 hashes        = ZSTD_Vec256_read(tagRow + ZSTD_ROW_HASH_TAG_OFFSET);
+    } else { /* rowEntries == 32 */
-        ZSTD_Vec256 expandedTags  = ZSTD_Vec256_set8(tag);
+        const __m128i chunk0 = _mm_loadu_si128((const __m128i*)(const void*)&src[0]);
-        matches                   = ZSTD_Vec256_cmpMask8(hashes, expandedTags);
+        const __m128i chunk1 = _mm_loadu_si128((const __m128i*)(const void*)&src[16]);
-    } else {
+        const __m128i equalMask0 = _mm_cmpeq_epi8(chunk0, _mm_set1_epi8(tag));
-        assert(0);
+        const __m128i equalMask1 = _mm_cmpeq_epi8(chunk1, _mm_set1_epi8(tag));
+        const U32 lo = (U32)_mm_movemask_epi8(equalMask0);
+        const U32 hi = (U32)_mm_movemask_epi8(equalMask1);
+        return ZSTD_rotateRight_U32((hi << 16) | lo, head);
     }
-    /* Each row is a circular buffer beginning at the value of "head". So we must rotate the "matches" bitfield
+#else
-        to match up with the actual layout of the entries within the hashTable */
+#  if defined(ZSTD_ARCH_ARM_NEON)
-    return ZSTD_VecMask_rotateRight(matches, head, rowEntries);
+    if (MEM_isLittleEndian()) {
+        if (rowEntries == 16) {
+            const uint8x16_t chunk = vld1q_u8(src);
+            const uint16x8_t equalMask = vreinterpretq_u16_u8(vceqq_u8(chunk, vdupq_n_u8(tag)));
+            const uint16x8_t t0 = vshlq_n_u16(equalMask, 7);
+            const uint32x4_t t1 = vreinterpretq_u32_u16(vsriq_n_u16(t0, t0, 14));
+            const uint64x2_t t2 = vreinterpretq_u64_u32(vshrq_n_u32(t1, 14));
+            const uint8x16_t t3 = vreinterpretq_u8_u64(vsraq_n_u64(t2, t2, 28));
+            const U16 hi = (U16)vgetq_lane_u8(t3, 8);
+            const U16 lo = (U16)vgetq_lane_u8(t3, 0);
+            return ZSTD_rotateRight_U16((hi << 8) | lo, head);
+        } else { /* rowEntries == 32 */
+            const uint16x8x2_t chunk = vld2q_u16((const U16*)(const void*)src);
+            const uint8x16_t chunk0 = vreinterpretq_u8_u16(chunk.val[0]);
+            const uint8x16_t chunk1 = vreinterpretq_u8_u16(chunk.val[1]);
+            const uint8x16_t equalMask0 = vceqq_u8(chunk0, vdupq_n_u8(tag));
+            const uint8x16_t equalMask1 = vceqq_u8(chunk1, vdupq_n_u8(tag));
+            const int8x8_t pack0 = vqmovn_s16(vreinterpretq_s16_u8(equalMask0));
+            const int8x8_t pack1 = vqmovn_s16(vreinterpretq_s16_u8(equalMask1));
+            const uint8x8_t t0 = vreinterpret_u8_s8(pack0);
+            const uint8x8_t t1 = vreinterpret_u8_s8(pack1);
+            const uint8x8_t t2 = vsri_n_u8(t1, t0, 2);
+            const uint8x8x2_t t3 = vuzp_u8(t2, t0);
+            const uint8x8_t t4 = vsri_n_u8(t3.val[1], t3.val[0], 4);
+            const U32 matches = vget_lane_u32(vreinterpret_u32_u8(t4), 0);
+            return ZSTD_rotateRight_U32(matches, head);
+        }
+    }
+#  endif
+    { /* SWAR */
+        const size_t chunkSize = sizeof(size_t);
+        const size_t shiftAmount = ((chunkSize * 8) - chunkSize);
+        const size_t xFF = ~((size_t)0);
+        const size_t x01 = xFF / 0xFF;
+        const size_t x80 = x01 << 7;
+        const size_t splatChar = tag * x01;
+        size_t matches = 0;
+        int i = rowEntries - chunkSize;
+        assert((sizeof(size_t) == 4) || (sizeof(size_t) == 8));
+        if (MEM_isLittleEndian()) { /* runtime check so have two loops */
+            const size_t extractMagic = (xFF / 0x7F) >> chunkSize;
+            do {
+                size_t chunk = MEM_readST(&src[i]);
+                chunk ^= splatChar;
+                chunk = (((chunk | x80) - x01) | chunk) & x80;
+                matches <<= chunkSize;
+                matches |= (chunk * extractMagic) >> shiftAmount;
+                i -= chunkSize;
+            } while (i >= 0);
+        } else { /* big endian: reverse bits during extraction */
+            const size_t msb = xFF ^ (xFF >> 1);
+            const size_t extractMagic = (msb / 0x1FF) | msb;
+            do {
+                size_t chunk = MEM_readST(&src[i]);
+                chunk ^= splatChar;
+                chunk = (((chunk | x80) - x01) | chunk) & x80;
+                matches <<= chunkSize;
+                matches |= ((chunk >> 7) * extractMagic) >> shiftAmount;
+                i -= chunkSize;
+            } while (i >= 0);
+        }
+        matches = ~matches;
+        if (rowEntries == 16) {
+            return ZSTD_rotateRight_U16((U16)matches, head);
+        } else { /* rowEntries == 32 */
+            return ZSTD_rotateRight_U32((U32)matches, head);
+        }
+    }
+#endif
 }
 
 /* The high-level approach of the SIMD row based match finder is as follows: