[Support] Add llvm::xxh3_64bits

[MaskRay]

Summary

ld.lld SHF_MERGE|SHF_STRINGS duplicate elimination is computation heavy
and utilitizes llvm::xxHash64, a simplified version of XXH64.
Externally many sources confirm that a new variant XXH3 is much faster.

I have picked a few hash implementations and computed the
proportion of time spent on hashing in the overall link time (a debug
build of clang 16 on a machine using AMD Zen 2 architecture):

• llvm::xxHash64: 3.63%
• official XXH64 (#define XXH_VECTOR XXH_SCALAR): 3.53%
• official XXH3_64bits (#define XXH_VECTOR XXH_SCALAR): 1.21%
• official XXH3_64bits (default, essentially XXH_SSE2): 1.22%
• this patch llvm::xxh3_64bits: 1.19%

The remaining part of lld remains unchanged. Consequently, a lower ratio
indicates that hashing is faster. Therefore, it is evident that XXH3 from xxhash
is significantly faster than both the official version and our llvm::xxHash64.

(
string length: count
1-3: 393434
4-8: 2084056
9-16: 2846249
17-128: 5598928
129-240: 1317989
241-: 328058
)

This patch adds heavily simplified https://github.com/Cyan4973/xxHash,
taking account of many simplification ideas from Devin Hussey's xxhash-clean.

Important x86-64 optimization ideas:

• Make XXH3_len_129to240_64b and XXH3_hashLong_64b noinline
• Unroll XXH3_len_17to128_64b
• __restrict does not affect Clang code generation

Beside SHF_MERGE|SHF_STRINGS duplicate elimination, llvm/ADT/StringMap.h
StringMapImpl::LookupBucketFor and a few places in lld can potentially be
accelerated by switching to llvm::xxh3_64bits.

Link: https://github.com/llvm/llvm-project/issues/63750

[peter.smith]

Looks like a useful optimisation, I'm not in a great position to review the code. Will try and find some time to run a benchmark on AArch64 to see if I can replicate the results on a non-X64 machine.

[MaskRay]

Ping:)

[serge-sans-paille]

LGTM with the minor nits mentioned above.

llvm/include/llvm/Support/xxhash.h

48

Why exposing this version to the user?

llvm/lib/Support/xxhash.cpp

3

you should probably update the copyright notice to 2021 to reflect upstream https://github.com/Cyan4973/xxHash/blob/dev/LICENSE#L2

38

Mention the commit number here?

151

could be constexpr size_t

198

same here.

[andrewng]

I haven't looked at the code but I have built ld.lld with this patch and the patch from D154813 on Windows with both VS 2022 cl 19.36.32537 and clang-cl 16.0.6. I have run the usual benchmarks (clang, chrome, mozilla, scylla) and also a UE5 link and I don't see any difference in performance on my Windows 10 PC, i.e. any delta is in the measurement "noise".

[MaskRay]

In D154812#4511112, @andrewng wrote:

I haven't looked at the code but I have built ld.lld with this patch and the patch from D154813 on Windows with both VS 2022 cl 19.36.32537 and clang-cl 16.0.6. I have run the usual benchmarks (clang, chrome, mozilla, scylla) and also a UE5 link and I don't see any difference in performance on my Windows 10 PC, i.e. any delta is in the measurement "noise".

Thank you for benchmarking this on Windows. Do these programs have large .debug_str sections? (Windows builds normally use CodeView/PDB, not DWARF).
I think only any effect is likely only observed with large .debug_str sections. To observe the effect of replacing the hash function, I need to compute the time ratio of hashing

The remaining part of lld remains unchanged. Consequently, a lower ratio indicates that hashing is faster.

I have observed ~2% speedup on an x86-64 machine, but the result on an aarch64 machine (Cavium ThunderX2) is so-so.
In PiotrZSL's setup the speedup seems larger?

I think xxh3 is still worth it and we can probably drop the xxh64 implementation once internal implementations have migrated.

% hyperfine --warmup 1 --min-runs 16 "numactl -C 16-23 "{/tmp/lld0,/tmp/lld1}" -flavor gnu @response.txt --threads=8"
Benchmark 1: numactl -C 16-23 /tmp/lld0 -flavor gnu @response.txt --threads=8
  Time (mean ± σ):      7.393 s ±  0.089 s    [User: 17.077 s, System: 4.274 s]
  Range (min … max):    7.208 s …  7.561 s    16 runs

Benchmark 2: numactl -C 16-23 /tmp/lld1 -flavor gnu @response.txt --threads=8
  Time (mean ± σ):      7.351 s ±  0.093 s    [User: 16.757 s, System: 4.236 s]
  Range (min … max):    7.165 s …  7.519 s    16 runs

[MaskRay]

address comment

llvm/lib/Support/xxhash.cpp

3

Yes, I'll do that. Changing this in the differential would cause clang-format to format all the text below, which is I'd want to avoid. I'll change this when landing...

151

Sounds good.

I was to retain the original code style, but since the original code has been mostly rewritten, changing the macros should be fine as well...

[MaskRay]

update description

[PiotrZSL]

"In PiotrZSL's setup the speedup seems larger?"

I was running this on Ryzen 7 3800X under Ubuntu on single thread with 1.3GB clang binary linking.
When running this with more threads, gain looks smaller because that heavy chunk is split into multiple threads, and gain on entire application seems smaller.

[andrewng]

I haven't looked at the code but I have built ld.lld with this patch and the patch from D154813 on Windows with both VS 2022 cl 19.36.32537 and clang-cl 16.0.6. I have run the usual benchmarks (clang, chrome, mozilla, scylla) and also a UE5 link and I don't see any difference in performance on my Windows 10 PC, i.e. any delta is in the measurement "noise".

Thank you for benchmarking this on Windows. Do these programs have large .debug_str sections? (Windows builds normally use CodeView/PDB, not DWARF).
I think only any effect is likely only observed with large .debug_str sections. To observe the effect of replacing the hash function, I need to compute the time ratio of hashing

The tests that I ran were all ELF links and one of the clang links included DWARF debug.

Out of interest, which compiler did you use to build your Linux lld?

[MaskRay]

In D154812#4514078, @andrewng wrote:

I haven't looked at the code but I have built ld.lld with this patch and the patch from D154813 on Windows with both VS 2022 cl 19.36.32537 and clang-cl 16.0.6. I have run the usual benchmarks (clang, chrome, mozilla, scylla) and also a UE5 link and I don't see any difference in performance on my Windows 10 PC, i.e. any delta is in the measurement "noise".

Thank you for benchmarking this on Windows. Do these programs have large .debug_str sections? (Windows builds normally use CodeView/PDB, not DWARF).
I think only any effect is likely only observed with large .debug_str sections. To observe the effect of replacing the hash function, I need to compute the time ratio of hashing

The tests that I ran were all ELF links and one of the clang links included DWARF debug.

Out of interest, which compiler did you use to build your Linux lld?

I use a close-to-trunk prebuilt Clang from Chromium: curl -s https://raw.githubusercontent.com/chromium/chromium/main/tools/clang/scripts/update.py | python3 - --output-dir=~/Stable

llvm/include/llvm/Support/xxhash.h

 #define LLVM_SUPPORT_XXHASH_H
 
 #include "llvm/ADT/ArrayRef.h"
 #include "llvm/ADT/StringRef.h"
 
 namespace llvm {
 uint64_t xxHash64(llvm::StringRef Data);
 uint64_t xxHash64(llvm::ArrayRef<uint8_t> Data);
+
+uint64_t xxh3_64bits(ArrayRef<uint8_t> data);
+inline uint64_t xxh3_64bits(StringRef data) {
+  return xxh3_64bits(ArrayRef(data.bytes_begin(), data.size()));
+}
 }
 
 #endif

llvm/lib/Support/xxhash.cpp

 /*
 *  xxHash - Fast Hash algorithm
-*  Copyright (C) 2012-2016, Yann Collet
+*  Copyright (C) 2012-2021, Yann Collet
 *
 *  BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
 *
 *  Redistribution and use in source and binary forms, with or without
 *  modification, are permitted provided that the following conditions are
 *  met:
 *
 *  * Redistributions of source code must retain the above copyright

 *  (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
 *  OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 *
 *  You can contact the author at :
 *  - xxHash homepage: http://www.xxhash.com
 *  - xxHash source repository : https://github.com/Cyan4973/xxHash
 */
 
-/* based on revision d2df04efcbef7d7f6886d345861e5dfda4edacc1 Removed
- * everything but a simple interface for computing XXh64. */
+// xxhash64 is based on commit d2df04efcbef7d7f6886d345861e5dfda4edacc1. Removed
+// everything but a simple interface for computing xxh64.
+
+// xxh3_64bits is based on commit d5891596637d21366b9b1dcf2c0007a3edb26a9e (July
+// 2023).
 
 #include "llvm/Support/xxhash.h"
+#include "llvm/Support/Compiler.h"
 #include "llvm/Support/Endian.h"
 
 #include <stdlib.h>
 
 using namespace llvm;
 using namespace support;
 
 static uint64_t rotl64(uint64_t X, size_t R) {
   return (X << R) | (X >> (64 - R));
 }
 
+constexpr uint32_t PRIME32_1 = 0x9E3779B1;
+constexpr uint32_t PRIME32_2 = 0x85EBCA77;
+constexpr uint32_t PRIME32_3 = 0xC2B2AE3D;
+
 static const uint64_t PRIME64_1 = 11400714785074694791ULL;
 static const uint64_t PRIME64_2 = 14029467366897019727ULL;
 static const uint64_t PRIME64_3 = 1609587929392839161ULL;
 static const uint64_t PRIME64_4 = 9650029242287828579ULL;
 static const uint64_t PRIME64_5 = 2870177450012600261ULL;
 
 static uint64_t round(uint64_t Acc, uint64_t Input) {
   Acc += Input * PRIME64_2;
   Acc = rotl64(Acc, 31);
   Acc *= PRIME64_1;
   return Acc;
 }
 
 static uint64_t mergeRound(uint64_t Acc, uint64_t Val) {
   Val = round(0, Val);
   Acc ^= Val;
   Acc = Acc * PRIME64_1 + PRIME64_4;
   return Acc;
 }
 
+static uint64_t XXH64_avalanche(uint64_t hash) {
+  hash ^= hash >> 33;
+  hash *= PRIME64_2;
+  hash ^= hash >> 29;
+  hash *= PRIME64_3;
+  hash ^= hash >> 32;
+  return hash;
+}
+
 uint64_t llvm::xxHash64(StringRef Data) {
   size_t Len = Data.size();
   uint64_t Seed = 0;
   const unsigned char *P = Data.bytes_begin();
   const unsigned char *const BEnd = Data.bytes_end();
   uint64_t H64;
 
   if (Len >= 32) {
uint64_t llvm::xxHash64(StringRef Data) {
   }
 
   while (P < BEnd) {
     H64 ^= (*P) * PRIME64_5;
     H64 = rotl64(H64, 11) * PRIME64_1;
     P++;
   }
 
-  H64 ^= H64 >> 33;
-  H64 *= PRIME64_2;
-  H64 ^= H64 >> 29;
-  H64 *= PRIME64_3;
-  H64 ^= H64 >> 32;
-
-  return H64;
+  return XXH64_avalanche(H64);
 }
 
 uint64_t llvm::xxHash64(ArrayRef<uint8_t> Data) {
   return xxHash64({(const char *)Data.data(), Data.size()});
 }
+
+constexpr size_t XXH3_SECRETSIZE_MIN = 136;
+constexpr size_t XXH_SECRET_DEFAULT_SIZE = 192;
+
+/* Pseudorandom data taken directly from FARSH */
+// clang-format off
+constexpr uint8_t kSecret[XXH_SECRET_DEFAULT_SIZE] = {
+    0xb8, 0xfe, 0x6c, 0x39, 0x23, 0xa4, 0x4b, 0xbe, 0x7c, 0x01, 0x81, 0x2c, 0xf7, 0x21, 0xad, 0x1c,
+    0xde, 0xd4, 0x6d, 0xe9, 0x83, 0x90, 0x97, 0xdb, 0x72, 0x40, 0xa4, 0xa4, 0xb7, 0xb3, 0x67, 0x1f,
+    0xcb, 0x79, 0xe6, 0x4e, 0xcc, 0xc0, 0xe5, 0x78, 0x82, 0x5a, 0xd0, 0x7d, 0xcc, 0xff, 0x72, 0x21,
+    0xb8, 0x08, 0x46, 0x74, 0xf7, 0x43, 0x24, 0x8e, 0xe0, 0x35, 0x90, 0xe6, 0x81, 0x3a, 0x26, 0x4c,
+    0x3c, 0x28, 0x52, 0xbb, 0x91, 0xc3, 0x00, 0xcb, 0x88, 0xd0, 0x65, 0x8b, 0x1b, 0x53, 0x2e, 0xa3,
+    0x71, 0x64, 0x48, 0x97, 0xa2, 0x0d, 0xf9, 0x4e, 0x38, 0x19, 0xef, 0x46, 0xa9, 0xde, 0xac, 0xd8,
+    0xa8, 0xfa, 0x76, 0x3f, 0xe3, 0x9c, 0x34, 0x3f, 0xf9, 0xdc, 0xbb, 0xc7, 0xc7, 0x0b, 0x4f, 0x1d,
+    0x8a, 0x51, 0xe0, 0x4b, 0xcd, 0xb4, 0x59, 0x31, 0xc8, 0x9f, 0x7e, 0xc9, 0xd9, 0x78, 0x73, 0x64,
+    0xea, 0xc5, 0xac, 0x83, 0x34, 0xd3, 0xeb, 0xc3, 0xc5, 0x81, 0xa0, 0xff, 0xfa, 0x13, 0x63, 0xeb,
+    0x17, 0x0d, 0xdd, 0x51, 0xb7, 0xf0, 0xda, 0x49, 0xd3, 0x16, 0x55, 0x26, 0x29, 0xd4, 0x68, 0x9e,
+    0x2b, 0x16, 0xbe, 0x58, 0x7d, 0x47, 0xa1, 0xfc, 0x8f, 0xf8, 0xb8, 0xd1, 0x7a, 0xd0, 0x31, 0xce,
+    0x45, 0xcb, 0x3a, 0x8f, 0x95, 0x16, 0x04, 0x28, 0xaf, 0xd7, 0xfb, 0xca, 0xbb, 0x4b, 0x40, 0x7e,
+};
+// clang-format on
+
+constexpr uint64_t PRIME_MX1 = 0x165667919E3779F9;
+constexpr uint64_t PRIME_MX2 = 0x9FB21C651E98DF25;
+
+// Calculates a 64-bit to 128-bit multiply, then XOR folds it.
+static uint64_t XXH3_mul128_fold64(uint64_t lhs, uint64_t rhs) {
+#if defined(__SIZEOF_INT128__) ||                                              \
+    (defined(_INTEGRAL_MAX_BITS) && _INTEGRAL_MAX_BITS >= 128)
+  __uint128_t product = (__uint128_t)lhs * (__uint128_t)rhs;
+  return uint64_t(product) ^ uint64_t(product >> 64);
+
+#else
+  /* First calculate all of the cross products. */
+  const uint64_t lo_lo = (lhs & 0xFFFFFFFF) * (rhs & 0xFFFFFFFF);
+  const uint64_t hi_lo = (lhs >> 32) * (rhs & 0xFFFFFFFF);
+  const uint64_t lo_hi = (lhs & 0xFFFFFFFF) * (rhs >> 32);
+  const uint64_t hi_hi = (lhs >> 32) * (rhs >> 32);
+
+  /* Now add the products together. These will never overflow. */
+  const uint64_t cross = (lo_lo >> 32) + (hi_lo & 0xFFFFFFFF) + lo_hi;
+  const uint64_t upper = (hi_lo >> 32) + (cross >> 32) + hi_hi;
+  const uint64_t lower = (cross << 32) | (lo_lo & 0xFFFFFFFF);
+
+  return upper ^ lower;
+#endif
+}
+
+constexpr size_t XXH_STRIPE_LEN = 64;
+constexpr size_t XXH_SECRET_CONSUME_RATE = 8;
+constexpr size_t XXH_ACC_NB = XXH_STRIPE_LEN / sizeof(uint64_t);
+
+static uint64_t XXH3_avalanche(uint64_t hash) {
+  hash ^= hash >> 37;
+  hash *= PRIME_MX1;
+  hash ^= hash >> 32;
+  return hash;
+}
+
+static uint64_t XXH3_len_1to3_64b(const uint8_t *input, size_t len,
+                                  const uint8_t *secret, uint64_t seed) {
+  const uint8_t c1 = input[0];
+  const uint8_t c2 = input[len >> 1];
+  const uint8_t c3 = input[len - 1];
+  uint32_t combined = ((uint32_t)c1 << 16) | ((uint32_t)c2 << 24) |
+                      ((uint32_t)c3 << 0) | ((uint32_t)len << 8);
+  uint64_t bitflip =
+      (uint64_t)(endian::read32le(secret) ^ endian::read32le(secret + 4)) +
+      seed;
+  return XXH64_avalanche(uint64_t(combined) ^ bitflip);
+}
+
+static uint64_t XXH3_len_4to8_64b(const uint8_t *input, size_t len,
+                                  const uint8_t *secret, uint64_t seed) {
+  seed ^= (uint64_t)byteswap(uint32_t(seed)) << 32;
+  const uint32_t input1 = endian::read32le(input);
+  const uint32_t input2 = endian::read32le(input + len - 4);
+  uint64_t acc =
+      (endian::read64le(secret + 8) ^ endian::read64le(secret + 16)) - seed;
+  const uint64_t input64 = (uint64_t)input2 | ((uint64_t)input1 << 32);
+  acc ^= input64;
+  // XXH3_rrmxmx(acc, len)
+  acc ^= rotl64(acc, 49) ^ rotl64(acc, 24);
+  acc *= PRIME_MX2;
+  acc ^= (acc >> 35) + (uint64_t)len;
+  acc *= PRIME_MX2;
+  return acc ^ (acc >> 28);
+}
+
+static uint64_t XXH3_len_9to16_64b(const uint8_t *input, size_t len,
+                                   const uint8_t *secret, uint64_t const seed) {
+  uint64_t input_lo =
+      (endian::read64le(secret + 24) ^ endian::read64le(secret + 32)) + seed;
+  uint64_t input_hi =
+      (endian::read64le(secret + 40) ^ endian::read64le(secret + 48)) - seed;
+  input_lo ^= endian::read64le(input);
+  input_hi ^= endian::read64le(input + len - 8);
+  uint64_t acc = uint64_t(len) + byteswap(input_lo) + input_hi +
+                 XXH3_mul128_fold64(input_lo, input_hi);
+  return XXH3_avalanche(acc);
+}
+
+LLVM_ATTRIBUTE_ALWAYS_INLINE
+static uint64_t XXH3_len_0to16_64b(const uint8_t *input, size_t len,
+                                   const uint8_t *secret, uint64_t const seed) {
+  if (LLVM_LIKELY(len > 8))
+    return XXH3_len_9to16_64b(input, len, secret, seed);
+  if (LLVM_LIKELY(len >= 4))
+    return XXH3_len_4to8_64b(input, len, secret, seed);
+  if (len != 0)
+    return XXH3_len_1to3_64b(input, len, secret, seed);
+  return XXH64_avalanche(seed ^ endian::read64le(secret + 56) ^
+                         endian::read64le(secret + 64));
+}
+
+static uint64_t XXH3_mix16B(const uint8_t *input, uint8_t const *secret,
+                            uint64_t seed) {
+  uint64_t lhs = seed;
+  uint64_t rhs = 0U - seed;
+  lhs += endian::read64le(secret);
+  rhs += endian::read64le(secret + 8);
+  lhs ^= endian::read64le(input);
+  rhs ^= endian::read64le(input + 8);
+  return XXH3_mul128_fold64(lhs, rhs);
+}
+
+/* For mid range keys, XXH3 uses a Mum-hash variant. */
+LLVM_ATTRIBUTE_ALWAYS_INLINE
+static uint64_t XXH3_len_17to128_64b(const uint8_t *input, size_t len,
+                                     const uint8_t *secret,
+                                     uint64_t const seed) {
+  uint64_t acc = len * PRIME64_1, acc_end;
+  acc += XXH3_mix16B(input + 0, secret + 0, seed);
+  acc_end = XXH3_mix16B(input + len - 16, secret + 16, seed);
+  if (len > 32) {
+    acc += XXH3_mix16B(input + 16, secret + 32, seed);
+    acc_end += XXH3_mix16B(input + len - 32, secret + 48, seed);
+    if (len > 64) {
+      acc += XXH3_mix16B(input + 32, secret + 64, seed);
+      acc_end += XXH3_mix16B(input + len - 48, secret + 80, seed);
+      if (len > 96) {
+        acc += XXH3_mix16B(input + 48, secret + 96, seed);
+        acc_end += XXH3_mix16B(input + len - 64, secret + 112, seed);
+      }
+    }
+  }
+  return XXH3_avalanche(acc + acc_end);
+}
+
+constexpr size_t XXH3_MIDSIZE_MAX = 240;
+
+LLVM_ATTRIBUTE_NOINLINE
+static uint64_t XXH3_len_129to240_64b(const uint8_t *input, size_t len,
+                                      const uint8_t *secret, uint64_t seed) {
+  constexpr size_t XXH3_MIDSIZE_STARTOFFSET = 3;
+  constexpr size_t XXH3_MIDSIZE_LASTOFFSET = 17;
+  uint64_t acc = (uint64_t)len * PRIME64_1;
+  const unsigned nbRounds = len / 16;
+  for (unsigned i = 0; i < 8; ++i)
+    acc += XXH3_mix16B(input + 16 * i, secret + 16 * i, seed);
+  acc = XXH3_avalanche(acc);
+
+  for (unsigned i = 8; i < nbRounds; ++i) {
+    acc += XXH3_mix16B(input + 16 * i,
+                       secret + 16 * (i - 8) + XXH3_MIDSIZE_STARTOFFSET, seed);
+  }
+  /* last bytes */
+  acc +=
+      XXH3_mix16B(input + len - 16,
+                  secret + XXH3_SECRETSIZE_MIN - XXH3_MIDSIZE_LASTOFFSET, seed);
+  return XXH3_avalanche(acc);
+}
+
+LLVM_ATTRIBUTE_ALWAYS_INLINE
+static void XXH3_accumulate_512_scalar(uint64_t *acc, const uint8_t *input,
+                                       const uint8_t *secret) {
+  for (size_t i = 0; i < XXH_ACC_NB; ++i) {
+    uint64_t data_val = endian::read64le(input + 8 * i);
+    uint64_t data_key = data_val ^ endian::read64le(secret + 8 * i);
+    acc[i ^ 1] += data_val;
+    acc[i] += uint32_t(data_key) * (data_key >> 32);
+  }
+}
+
+LLVM_ATTRIBUTE_ALWAYS_INLINE
+static void XXH3_accumulate_scalar(uint64_t *acc, const uint8_t *input,
+                                   const uint8_t *secret, size_t nbStripes) {
+  for (size_t n = 0; n < nbStripes; ++n)
+    XXH3_accumulate_512_scalar(acc, input + n * XXH_STRIPE_LEN,
+                               secret + n * XXH_SECRET_CONSUME_RATE);
+}
+
+static void XXH3_scrambleAcc(uint64_t *acc, const uint8_t *secret) {
+  for (size_t i = 0; i < XXH_ACC_NB; ++i) {
+    acc[i] ^= acc[i] >> 47;
+    acc[i] ^= endian::read64le(secret + 8 * i);
+    acc[i] *= PRIME32_1;
+  }
+}
+
+static uint64_t XXH3_mix2Accs(const uint64_t *acc, const uint8_t *secret) {
+  return XXH3_mul128_fold64(acc[0] ^ endian::read64le(secret),
+                            acc[1] ^ endian::read64le(secret + 8));
+}
+
+static uint64_t XXH3_mergeAccs(const uint64_t *acc, const uint8_t *key,
+                               uint64_t start) {
+  uint64_t result64 = start;
+  for (size_t i = 0; i < 4; ++i)
+    result64 += XXH3_mix2Accs(acc + 2 * i, key + 16 * i);
+  return XXH3_avalanche(result64);
+}
+
+LLVM_ATTRIBUTE_NOINLINE
+static uint64_t XXH3_hashLong_64b(const uint8_t *input, size_t len,
+                                  const uint8_t *secret, size_t secretSize) {
+  const size_t nbStripesPerBlock =
+      (secretSize - XXH_STRIPE_LEN) / XXH_SECRET_CONSUME_RATE;
+  const size_t block_len = XXH_STRIPE_LEN * nbStripesPerBlock;
+  const size_t nb_blocks = (len - 1) / block_len;
+  alignas(16) uint64_t acc[XXH_ACC_NB] = {
+      PRIME32_3, PRIME64_1, PRIME64_2, PRIME64_3,
+      PRIME64_4, PRIME32_2, PRIME64_5, PRIME32_1,
+  };
+  for (size_t n = 0; n < nb_blocks; ++n) {
+    XXH3_accumulate_scalar(acc, input + n * block_len, secret,
+                           nbStripesPerBlock);
+    XXH3_scrambleAcc(acc, secret + secretSize - XXH_STRIPE_LEN);
+  }
+
+  /* last partial block */
+  const size_t nbStripes = (len - 1 - (block_len * nb_blocks)) / XXH_STRIPE_LEN;
+  assert(nbStripes <= secretSize / XXH_SECRET_CONSUME_RATE);
+  XXH3_accumulate_scalar(acc, input + nb_blocks * block_len, secret, nbStripes);
+
+  /* last stripe */
+  constexpr size_t XXH_SECRET_LASTACC_START = 7;
+  XXH3_accumulate_512_scalar(acc, input + len - XXH_STRIPE_LEN,
+                             secret + secretSize - XXH_STRIPE_LEN -
+                                 XXH_SECRET_LASTACC_START);
+
+  /* converge into final hash */
+  constexpr size_t XXH_SECRET_MERGEACCS_START = 11;
+  return XXH3_mergeAccs(acc, secret + XXH_SECRET_MERGEACCS_START,
+                        (uint64_t)len * PRIME64_1);
+}
+
+uint64_t llvm::xxh3_64bits(ArrayRef<uint8_t> data) {
+  auto *in = data.data();
+  size_t len = data.size();
+  if (len <= 16)
+    return XXH3_len_0to16_64b(in, len, kSecret, 0);
+  if (len <= 128)
+    return XXH3_len_17to128_64b(in, len, kSecret, 0);
+  if (len <= XXH3_MIDSIZE_MAX)
+    return XXH3_len_129to240_64b(in, len, kSecret, 0);
+  return XXH3_hashLong_64b(in, len, kSecret, sizeof(kSecret));
+}

llvm/unittests/Support/xxhashTest.cpp

 
 TEST(xxhashTest, Basic) {
   EXPECT_EQ(0xef46db3751d8e999U, xxHash64(StringRef()));
   EXPECT_EQ(0x33bf00a859c4ba3fU, xxHash64("foo"));
   EXPECT_EQ(0x48a37c90ad27a659U, xxHash64("bar"));
   EXPECT_EQ(0x69196c1b3af0bff9U,
             xxHash64("0123456789abcdefghijklmnopqrstuvwxyz"));
 }
+
+TEST(xxhashTest, xxh3) {
+  constexpr size_t size = 2243;
+  uint8_t a[size];
+  uint64_t x = 1;
+  for (size_t i = 0; i < size; ++i) {
+    x ^= x << 13;
+    x ^= x >> 7;
+    x ^= x << 17;
+    a[i] = uint8_t(x);
+  }
+
+#define F(len, expected)                                                       \
+  EXPECT_EQ(uint64_t(expected), xxh3_64bits(ArrayRef(a, size_t(len))))
+  F(0, 0x2d06800538d394c2);
+  F(1, 0xd0d496e05c553485);
+  F(2, 0x84d625edb7055eac);
+  F(3, 0x6ea2d59aca5c3778);
+  F(4, 0xbf65290914e80242);
+  F(5, 0xc01fd099ad4fc8e4);
+  F(6, 0x9e3ea8187399caa5);
+  F(7, 0x9da8b60540644f5a);
+  F(8, 0xabc1413da6cd0209);
+  F(9, 0x8bc89400bfed51f6);
+  F(16, 0x7e46916754d7c9b8);
+  F(17, 0xed4be912ba5f836d);
+  F(32, 0xf59b59b58c304fd1);
+  F(33, 0x9013fb74ca603e0c);
+  F(64, 0xfa5271fcce0db1c3);
+  F(65, 0x79c42431727f1012);
+  F(96, 0x591ee0ddf9c9ccd1);
+  F(97, 0x8ffc6a3111fe19da);
+  F(128, 0x06a146ee9a2da378);
+  F(129, 0xbc7138129bf065da);
+  F(403, 0xcefeb3ffa532ad8c);
+  F(512, 0xcdfa6b6268e3650f);
+  F(513, 0x4bb5d42742f9765f);
+  F(2048, 0x330ce110cbb79eae);
+  F(2049, 0x3ba6afa0249fef9a);
+  F(2240, 0xd61d4d2a94e926a8);
+  F(2243, 0x0979f786a24edde7);
+#undef F
+}