brintos

brintos / llvm-project-archived public Read only

0
0
Text · 29.0 KiB · ec76dd7 Raw
662 lines · c
1/*===-- blake3.c - BLAKE3 C Implementation ------------------------*- C -*-===*\2|*                                                                            *|3|* Released into the public domain with CC0 1.0                               *|4|* See 'llvm/lib/Support/BLAKE3/LICENSE' for info.                            *|5|* SPDX-License-Identifier: CC0-1.0                                           *|6|*                                                                            *|7\*===----------------------------------------------------------------------===*/8 9#include <assert.h>10#include <stdbool.h>11#include <string.h>12 13#include "blake3_impl.h"14 15const char *llvm_blake3_version(void) { return BLAKE3_VERSION_STRING; }16 17INLINE void chunk_state_init(blake3_chunk_state *self, const uint32_t key[8],18                             uint8_t flags) {19  memcpy(self->cv, key, BLAKE3_KEY_LEN);20  self->chunk_counter = 0;21  memset(self->buf, 0, BLAKE3_BLOCK_LEN);22  self->buf_len = 0;23  self->blocks_compressed = 0;24  self->flags = flags;25}26 27INLINE void chunk_state_reset(blake3_chunk_state *self, const uint32_t key[8],28                              uint64_t chunk_counter) {29  memcpy(self->cv, key, BLAKE3_KEY_LEN);30  self->chunk_counter = chunk_counter;31  self->blocks_compressed = 0;32  memset(self->buf, 0, BLAKE3_BLOCK_LEN);33  self->buf_len = 0;34}35 36INLINE size_t chunk_state_len(const blake3_chunk_state *self) {37  return (BLAKE3_BLOCK_LEN * (size_t)self->blocks_compressed) +38         ((size_t)self->buf_len);39}40 41INLINE size_t chunk_state_fill_buf(blake3_chunk_state *self,42                                   const uint8_t *input, size_t input_len) {43  size_t take = BLAKE3_BLOCK_LEN - ((size_t)self->buf_len);44  if (take > input_len) {45    take = input_len;46  }47  uint8_t *dest = self->buf + ((size_t)self->buf_len);48  memcpy(dest, input, take);49  self->buf_len += (uint8_t)take;50  return take;51}52 53INLINE uint8_t chunk_state_maybe_start_flag(const blake3_chunk_state *self) {54  if (self->blocks_compressed == 0) {55    return CHUNK_START;56  } else {57    return 0;58  }59}60 61typedef struct {62  uint32_t input_cv[8];63  uint64_t counter;64  uint8_t block[BLAKE3_BLOCK_LEN];65  uint8_t block_len;66  uint8_t flags;67} output_t;68 69INLINE output_t make_output(const uint32_t input_cv[8],70                            const uint8_t block[BLAKE3_BLOCK_LEN],71                            uint8_t block_len, uint64_t counter,72                            uint8_t flags) {73  output_t ret;74  memcpy(ret.input_cv, input_cv, 32);75  memcpy(ret.block, block, BLAKE3_BLOCK_LEN);76  ret.block_len = block_len;77  ret.counter = counter;78  ret.flags = flags;79  return ret;80}81 82// Chaining values within a given chunk (specifically the compress_in_place83// interface) are represented as words. This avoids unnecessary bytes<->words84// conversion overhead in the portable implementation. However, the hash_many85// interface handles both user input and parent node blocks, so it accepts86// bytes. For that reason, chaining values in the CV stack are represented as87// bytes.88INLINE void output_chaining_value(const output_t *self, uint8_t cv[32]) {89  uint32_t cv_words[8];90  memcpy(cv_words, self->input_cv, 32);91  blake3_compress_in_place(cv_words, self->block, self->block_len,92                           self->counter, self->flags);93  store_cv_words(cv, cv_words);94}95 96INLINE void output_root_bytes(const output_t *self, uint64_t seek, uint8_t *out,97                              size_t out_len) {98  if (out_len == 0) {99      return;100  }101  uint64_t output_block_counter = seek / 64;102  size_t offset_within_block = seek % 64;103  uint8_t wide_buf[64];104  if(offset_within_block) {105    blake3_compress_xof(self->input_cv, self->block, self->block_len, output_block_counter, self->flags | ROOT, wide_buf);106    const size_t available_bytes = 64 - offset_within_block;107    const size_t bytes = out_len > available_bytes ? available_bytes : out_len;108    memcpy(out, wide_buf + offset_within_block, bytes);109    out += bytes;110    out_len -= bytes;111    output_block_counter += 1;112  }113  if(out_len / 64) {114    blake3_xof_many(self->input_cv, self->block, self->block_len, output_block_counter, self->flags | ROOT, out, out_len / 64);115  }116  output_block_counter += out_len / 64;117  out += out_len & -64;118  out_len -= out_len & -64;119  if(out_len) {120    blake3_compress_xof(self->input_cv, self->block, self->block_len, output_block_counter, self->flags | ROOT, wide_buf);121    memcpy(out, wide_buf, out_len);122  }123}124 125INLINE void chunk_state_update(blake3_chunk_state *self, const uint8_t *input,126                               size_t input_len) {127  if (self->buf_len > 0) {128    size_t take = chunk_state_fill_buf(self, input, input_len);129    input += take;130    input_len -= take;131    if (input_len > 0) {132      blake3_compress_in_place(133          self->cv, self->buf, BLAKE3_BLOCK_LEN, self->chunk_counter,134          self->flags | chunk_state_maybe_start_flag(self));135      self->blocks_compressed += 1;136      self->buf_len = 0;137      memset(self->buf, 0, BLAKE3_BLOCK_LEN);138    }139  }140 141  while (input_len > BLAKE3_BLOCK_LEN) {142    blake3_compress_in_place(self->cv, input, BLAKE3_BLOCK_LEN,143                             self->chunk_counter,144                             self->flags | chunk_state_maybe_start_flag(self));145    self->blocks_compressed += 1;146    input += BLAKE3_BLOCK_LEN;147    input_len -= BLAKE3_BLOCK_LEN;148  }149 150  chunk_state_fill_buf(self, input, input_len);151}152 153INLINE output_t chunk_state_output(const blake3_chunk_state *self) {154  uint8_t block_flags =155      self->flags | chunk_state_maybe_start_flag(self) | CHUNK_END;156  return make_output(self->cv, self->buf, self->buf_len, self->chunk_counter,157                     block_flags);158}159 160INLINE output_t parent_output(const uint8_t block[BLAKE3_BLOCK_LEN],161                              const uint32_t key[8], uint8_t flags) {162  return make_output(key, block, BLAKE3_BLOCK_LEN, 0, flags | PARENT);163}164 165// Given some input larger than one chunk, return the number of bytes that166// should go in the left subtree. This is the largest power-of-2 number of167// chunks that leaves at least 1 byte for the right subtree.168INLINE size_t left_subtree_len(size_t input_len) {169  // Subtract 1 to reserve at least one byte for the right side. input_len170  // should always be greater than BLAKE3_CHUNK_LEN.171  size_t full_chunks = (input_len - 1) / BLAKE3_CHUNK_LEN;172  return round_down_to_power_of_2(full_chunks) * BLAKE3_CHUNK_LEN;173}174 175// Use SIMD parallelism to hash up to MAX_SIMD_DEGREE chunks at the same time176// on a single thread. Write out the chunk chaining values and return the177// number of chunks hashed. These chunks are never the root and never empty;178// those cases use a different codepath.179INLINE size_t compress_chunks_parallel(const uint8_t *input, size_t input_len,180                                       const uint32_t key[8],181                                       uint64_t chunk_counter, uint8_t flags,182                                       uint8_t *out) {183#if defined(BLAKE3_TESTING)184  assert(0 < input_len);185  assert(input_len <= MAX_SIMD_DEGREE * BLAKE3_CHUNK_LEN);186#endif187 188  const uint8_t *chunks_array[MAX_SIMD_DEGREE];189  size_t input_position = 0;190  size_t chunks_array_len = 0;191  while (input_len - input_position >= BLAKE3_CHUNK_LEN) {192    chunks_array[chunks_array_len] = &input[input_position];193    input_position += BLAKE3_CHUNK_LEN;194    chunks_array_len += 1;195  }196 197  blake3_hash_many(chunks_array, chunks_array_len,198                   BLAKE3_CHUNK_LEN / BLAKE3_BLOCK_LEN, key, chunk_counter,199                   true, flags, CHUNK_START, CHUNK_END, out);200 201  // Hash the remaining partial chunk, if there is one. Note that the empty202  // chunk (meaning the empty message) is a different codepath.203  if (input_len > input_position) {204    uint64_t counter = chunk_counter + (uint64_t)chunks_array_len;205    blake3_chunk_state chunk_state;206    chunk_state_init(&chunk_state, key, flags);207    chunk_state.chunk_counter = counter;208    chunk_state_update(&chunk_state, &input[input_position],209                       input_len - input_position);210    output_t output = chunk_state_output(&chunk_state);211    output_chaining_value(&output, &out[chunks_array_len * BLAKE3_OUT_LEN]);212    return chunks_array_len + 1;213  } else {214    return chunks_array_len;215  }216}217 218// Use SIMD parallelism to hash up to MAX_SIMD_DEGREE parents at the same time219// on a single thread. Write out the parent chaining values and return the220// number of parents hashed. (If there's an odd input chaining value left over,221// return it as an additional output.) These parents are never the root and222// never empty; those cases use a different codepath.223INLINE size_t compress_parents_parallel(const uint8_t *child_chaining_values,224                                        size_t num_chaining_values,225                                        const uint32_t key[8], uint8_t flags,226                                        uint8_t *out) {227#if defined(BLAKE3_TESTING)228  assert(2 <= num_chaining_values);229  assert(num_chaining_values <= 2 * MAX_SIMD_DEGREE_OR_2);230#endif231 232  const uint8_t *parents_array[MAX_SIMD_DEGREE_OR_2];233  size_t parents_array_len = 0;234  while (num_chaining_values - (2 * parents_array_len) >= 2) {235    parents_array[parents_array_len] =236        &child_chaining_values[2 * parents_array_len * BLAKE3_OUT_LEN];237    parents_array_len += 1;238  }239 240  blake3_hash_many(parents_array, parents_array_len, 1, key,241                   0, // Parents always use counter 0.242                   false, flags | PARENT,243                   0, // Parents have no start flags.244                   0, // Parents have no end flags.245                   out);246 247  // If there's an odd child left over, it becomes an output.248  if (num_chaining_values > 2 * parents_array_len) {249    memcpy(&out[parents_array_len * BLAKE3_OUT_LEN],250           &child_chaining_values[2 * parents_array_len * BLAKE3_OUT_LEN],251           BLAKE3_OUT_LEN);252    return parents_array_len + 1;253  } else {254    return parents_array_len;255  }256}257 258// The wide helper function returns (writes out) an array of chaining values259// and returns the length of that array. The number of chaining values returned260// is the dynamically detected SIMD degree, at most MAX_SIMD_DEGREE. Or fewer,261// if the input is shorter than that many chunks. The reason for maintaining a262// wide array of chaining values going back up the tree, is to allow the263// implementation to hash as many parents in parallel as possible.264//265// As a special case when the SIMD degree is 1, this function will still return266// at least 2 outputs. This guarantees that this function doesn't perform the267// root compression. (If it did, it would use the wrong flags, and also we268// wouldn't be able to implement extendable output.) Note that this function is269// not used when the whole input is only 1 chunk long; that's a different270// codepath.271//272// Why not just have the caller split the input on the first update(), instead273// of implementing this special rule? Because we don't want to limit SIMD or274// multi-threading parallelism for that update().275size_t blake3_compress_subtree_wide(const uint8_t *input, size_t input_len,276                                    const uint32_t key[8],277                                    uint64_t chunk_counter, uint8_t flags,278                                    uint8_t *out, bool use_tbb) {279  // Note that the single chunk case does *not* bump the SIMD degree up to 2280  // when it is 1. If this implementation adds multi-threading in the future,281  // this gives us the option of multi-threading even the 2-chunk case, which282  // can help performance on smaller platforms.283  if (input_len <= blake3_simd_degree() * BLAKE3_CHUNK_LEN) {284    return compress_chunks_parallel(input, input_len, key, chunk_counter, flags,285                                    out);286  }287 288  // With more than simd_degree chunks, we need to recurse. Start by dividing289  // the input into left and right subtrees. (Note that this is only optimal290  // as long as the SIMD degree is a power of 2. If we ever get a SIMD degree291  // of 3 or something, we'll need a more complicated strategy.)292  size_t left_input_len = left_subtree_len(input_len);293  size_t right_input_len = input_len - left_input_len;294  const uint8_t *right_input = &input[left_input_len];295  uint64_t right_chunk_counter =296      chunk_counter + (uint64_t)(left_input_len / BLAKE3_CHUNK_LEN);297 298  // Make space for the child outputs. Here we use MAX_SIMD_DEGREE_OR_2 to299  // account for the special case of returning 2 outputs when the SIMD degree300  // is 1.301  uint8_t cv_array[2 * MAX_SIMD_DEGREE_OR_2 * BLAKE3_OUT_LEN];302  size_t degree = blake3_simd_degree();303  if (left_input_len > BLAKE3_CHUNK_LEN && degree == 1) {304    // The special case: We always use a degree of at least two, to make305    // sure there are two outputs. Except, as noted above, at the chunk306    // level, where we allow degree=1. (Note that the 1-chunk-input case is307    // a different codepath.)308    degree = 2;309  }310  uint8_t *right_cvs = &cv_array[degree * BLAKE3_OUT_LEN];311 312  // Recurse!313  size_t left_n = -1;314  size_t right_n = -1;315 316#if defined(BLAKE3_USE_TBB)317  blake3_compress_subtree_wide_join_tbb(318      key, flags, use_tbb,319      // left-hand side320      input, left_input_len, chunk_counter, cv_array, &left_n,321      // right-hand side322      right_input, right_input_len, right_chunk_counter, right_cvs, &right_n);323#else324  left_n = blake3_compress_subtree_wide(325      input, left_input_len, key, chunk_counter, flags, cv_array, use_tbb);326  right_n = blake3_compress_subtree_wide(right_input, right_input_len, key,327                                         right_chunk_counter, flags, right_cvs,328                                         use_tbb);329#endif // BLAKE3_USE_TBB330 331  // The special case again. If simd_degree=1, then we'll have left_n=1 and332  // right_n=1. Rather than compressing them into a single output, return333  // them directly, to make sure we always have at least two outputs.334  if (left_n == 1) {335    memcpy(out, cv_array, 2 * BLAKE3_OUT_LEN);336    return 2;337  }338 339  // Otherwise, do one layer of parent node compression.340  size_t num_chaining_values = left_n + right_n;341  return compress_parents_parallel(cv_array, num_chaining_values, key, flags,342                                   out);343}344 345// Hash a subtree with compress_subtree_wide(), and then condense the resulting346// list of chaining values down to a single parent node. Don't compress that347// last parent node, however. Instead, return its message bytes (the348// concatenated chaining values of its children). This is necessary when the349// first call to update() supplies a complete subtree, because the topmost350// parent node of that subtree could end up being the root. It's also necessary351// for extended output in the general case.352//353// As with compress_subtree_wide(), this function is not used on inputs of 1354// chunk or less. That's a different codepath.355INLINE void356compress_subtree_to_parent_node(const uint8_t *input, size_t input_len,357                                const uint32_t key[8], uint64_t chunk_counter,358                                uint8_t flags, uint8_t out[2 * BLAKE3_OUT_LEN],359                                bool use_tbb) {360#if defined(BLAKE3_TESTING)361  assert(input_len > BLAKE3_CHUNK_LEN);362#endif363 364  uint8_t cv_array[MAX_SIMD_DEGREE_OR_2 * BLAKE3_OUT_LEN];365  size_t num_cvs = blake3_compress_subtree_wide(input, input_len, key,366                                                chunk_counter, flags, cv_array, use_tbb);367  assert(num_cvs <= MAX_SIMD_DEGREE_OR_2);368  // The following loop never executes when MAX_SIMD_DEGREE_OR_2 is 2, because369  // as we just asserted, num_cvs will always be <=2 in that case. But GCC370  // (particularly GCC 8.5) can't tell that it never executes, and if NDEBUG is371  // set then it emits incorrect warnings here. We tried a few different372  // hacks to silence these, but in the end our hacks just produced different373  // warnings (see https://github.com/BLAKE3-team/BLAKE3/pull/380). Out of374  // desperation, we ifdef out this entire loop when we know it's not needed.375#if MAX_SIMD_DEGREE_OR_2 > 2376  // If MAX_SIMD_DEGREE_OR_2 is greater than 2 and there's enough input,377  // compress_subtree_wide() returns more than 2 chaining values. Condense378  // them into 2 by forming parent nodes repeatedly.379  uint8_t out_array[MAX_SIMD_DEGREE_OR_2 * BLAKE3_OUT_LEN / 2];380  while (num_cvs > 2) {381    num_cvs =382        compress_parents_parallel(cv_array, num_cvs, key, flags, out_array);383    memcpy(cv_array, out_array, num_cvs * BLAKE3_OUT_LEN);384  }385#endif386  memcpy(out, cv_array, 2 * BLAKE3_OUT_LEN);387}388 389INLINE void hasher_init_base(blake3_hasher *self, const uint32_t key[8],390                             uint8_t flags) {391  memcpy(self->key, key, BLAKE3_KEY_LEN);392  chunk_state_init(&self->chunk, key, flags);393  self->cv_stack_len = 0;394}395 396void llvm_blake3_hasher_init(blake3_hasher *self) { hasher_init_base(self, IV, 0); }397 398void llvm_blake3_hasher_init_keyed(blake3_hasher *self,399                              const uint8_t key[BLAKE3_KEY_LEN]) {400  uint32_t key_words[8];401  load_key_words(key, key_words);402  hasher_init_base(self, key_words, KEYED_HASH);403}404 405void llvm_blake3_hasher_init_derive_key_raw(blake3_hasher *self, const void *context,406                                       size_t context_len) {407  blake3_hasher context_hasher;408  hasher_init_base(&context_hasher, IV, DERIVE_KEY_CONTEXT);409  llvm_blake3_hasher_update(&context_hasher, context, context_len);410  uint8_t context_key[BLAKE3_KEY_LEN];411  llvm_blake3_hasher_finalize(&context_hasher, context_key, BLAKE3_KEY_LEN);412  uint32_t context_key_words[8];413  load_key_words(context_key, context_key_words);414  hasher_init_base(self, context_key_words, DERIVE_KEY_MATERIAL);415}416 417void llvm_blake3_hasher_init_derive_key(blake3_hasher *self, const char *context) {418  llvm_blake3_hasher_init_derive_key_raw(self, context, strlen(context));419}420 421// As described in hasher_push_cv() below, we do "lazy merging", delaying422// merges until right before the next CV is about to be added. This is423// different from the reference implementation. Another difference is that we424// aren't always merging 1 chunk at a time. Instead, each CV might represent425// any power-of-two number of chunks, as long as the smaller-above-larger stack426// order is maintained. Instead of the "count the trailing 0-bits" algorithm427// described in the spec, we use a "count the total number of 1-bits" variant428// that doesn't require us to retain the subtree size of the CV on top of the429// stack. The principle is the same: each CV that should remain in the stack is430// represented by a 1-bit in the total number of chunks (or bytes) so far.431INLINE void hasher_merge_cv_stack(blake3_hasher *self, uint64_t total_len) {432  size_t post_merge_stack_len = (size_t)popcnt(total_len);433  while (self->cv_stack_len > post_merge_stack_len) {434    uint8_t *parent_node =435        &self->cv_stack[(self->cv_stack_len - 2) * BLAKE3_OUT_LEN];436    output_t output = parent_output(parent_node, self->key, self->chunk.flags);437    output_chaining_value(&output, parent_node);438    self->cv_stack_len -= 1;439  }440}441 442// In reference_impl.rs, we merge the new CV with existing CVs from the stack443// before pushing it. We can do that because we know more input is coming, so444// we know none of the merges are root.445//446// This setting is different. We want to feed as much input as possible to447// compress_subtree_wide(), without setting aside anything for the chunk_state.448// If the user gives us 64 KiB, we want to parallelize over all 64 KiB at once449// as a single subtree, if at all possible.450//451// This leads to two problems:452// 1) This 64 KiB input might be the only call that ever gets made to update.453//    In this case, the root node of the 64 KiB subtree would be the root node454//    of the whole tree, and it would need to be ROOT finalized. We can't455//    compress it until we know.456// 2) This 64 KiB input might complete a larger tree, whose root node is457//    similarly going to be the root of the whole tree. For example, maybe458//    we have 196 KiB (that is, 128 + 64) hashed so far. We can't compress the459//    node at the root of the 256 KiB subtree until we know how to finalize it.460//461// The second problem is solved with "lazy merging". That is, when we're about462// to add a CV to the stack, we don't merge it with anything first, as the463// reference impl does. Instead we do merges using the *previous* CV that was464// added, which is sitting on top of the stack, and we put the new CV465// (unmerged) on top of the stack afterwards. This guarantees that we never466// merge the root node until finalize().467//468// Solving the first problem requires an additional tool,469// compress_subtree_to_parent_node(). That function always returns the top470// *two* chaining values of the subtree it's compressing. We then do lazy471// merging with each of them separately, so that the second CV will always472// remain unmerged. (That also helps us support extendable output when we're473// hashing an input all-at-once.)474INLINE void hasher_push_cv(blake3_hasher *self, uint8_t new_cv[BLAKE3_OUT_LEN],475                           uint64_t chunk_counter) {476  hasher_merge_cv_stack(self, chunk_counter);477  memcpy(&self->cv_stack[self->cv_stack_len * BLAKE3_OUT_LEN], new_cv,478         BLAKE3_OUT_LEN);479  self->cv_stack_len += 1;480}481 482INLINE void blake3_hasher_update_base(blake3_hasher *self, const void *input,483                                      size_t input_len, bool use_tbb) {484  // Explicitly checking for zero avoids causing UB by passing a null pointer485  // to memcpy. This comes up in practice with things like:486  //   std::vector<uint8_t> v;487  //   blake3_hasher_update(&hasher, v.data(), v.size());488  if (input_len == 0) {489    return;490  }491 492  const uint8_t *input_bytes = (const uint8_t *)input;493 494  // If we have some partial chunk bytes in the internal chunk_state, we need495  // to finish that chunk first.496  if (chunk_state_len(&self->chunk) > 0) {497    size_t take = BLAKE3_CHUNK_LEN - chunk_state_len(&self->chunk);498    if (take > input_len) {499      take = input_len;500    }501    chunk_state_update(&self->chunk, input_bytes, take);502    input_bytes += take;503    input_len -= take;504    // If we've filled the current chunk and there's more coming, finalize this505    // chunk and proceed. In this case we know it's not the root.506    if (input_len > 0) {507      output_t output = chunk_state_output(&self->chunk);508      uint8_t chunk_cv[32];509      output_chaining_value(&output, chunk_cv);510      hasher_push_cv(self, chunk_cv, self->chunk.chunk_counter);511      chunk_state_reset(&self->chunk, self->key, self->chunk.chunk_counter + 1);512    } else {513      return;514    }515  }516 517  // Now the chunk_state is clear, and we have more input. If there's more than518  // a single chunk (so, definitely not the root chunk), hash the largest whole519  // subtree we can, with the full benefits of SIMD (and maybe in the future,520  // multi-threading) parallelism. Two restrictions:521  // - The subtree has to be a power-of-2 number of chunks. Only subtrees along522  //   the right edge can be incomplete, and we don't know where the right edge523  //   is going to be until we get to finalize().524  // - The subtree must evenly divide the total number of chunks up until this525  //   point (if total is not 0). If the current incomplete subtree is only526  //   waiting for 1 more chunk, we can't hash a subtree of 4 chunks. We have527  //   to complete the current subtree first.528  // Because we might need to break up the input to form powers of 2, or to529  // evenly divide what we already have, this part runs in a loop.530  while (input_len > BLAKE3_CHUNK_LEN) {531    size_t subtree_len = round_down_to_power_of_2(input_len);532    uint64_t count_so_far = self->chunk.chunk_counter * BLAKE3_CHUNK_LEN;533    // Shrink the subtree_len until it evenly divides the count so far. We know534    // that subtree_len itself is a power of 2, so we can use a bitmasking535    // trick instead of an actual remainder operation. (Note that if the caller536    // consistently passes power-of-2 inputs of the same size, as is hopefully537    // typical, this loop condition will always fail, and subtree_len will538    // always be the full length of the input.)539    //540    // An aside: We don't have to shrink subtree_len quite this much. For541    // example, if count_so_far is 1, we could pass 2 chunks to542    // compress_subtree_to_parent_node. Since we'll get 2 CVs back, we'll still543    // get the right answer in the end, and we might get to use 2-way SIMD544    // parallelism. The problem with this optimization, is that it gets us545    // stuck always hashing 2 chunks. The total number of chunks will remain546    // odd, and we'll never graduate to higher degrees of parallelism. See547    // https://github.com/BLAKE3-team/BLAKE3/issues/69.548    while ((((uint64_t)(subtree_len - 1)) & count_so_far) != 0) {549      subtree_len /= 2;550    }551    // The shrunken subtree_len might now be 1 chunk long. If so, hash that one552    // chunk by itself. Otherwise, compress the subtree into a pair of CVs.553    uint64_t subtree_chunks = subtree_len / BLAKE3_CHUNK_LEN;554    if (subtree_len <= BLAKE3_CHUNK_LEN) {555      blake3_chunk_state chunk_state;556      chunk_state_init(&chunk_state, self->key, self->chunk.flags);557      chunk_state.chunk_counter = self->chunk.chunk_counter;558      chunk_state_update(&chunk_state, input_bytes, subtree_len);559      output_t output = chunk_state_output(&chunk_state);560      uint8_t cv[BLAKE3_OUT_LEN];561      output_chaining_value(&output, cv);562      hasher_push_cv(self, cv, chunk_state.chunk_counter);563    } else {564      // This is the high-performance happy path, though getting here depends565      // on the caller giving us a long enough input.566      uint8_t cv_pair[2 * BLAKE3_OUT_LEN];567      compress_subtree_to_parent_node(input_bytes, subtree_len, self->key,568                                      self->chunk.chunk_counter,569                                      self->chunk.flags, cv_pair, use_tbb);570      hasher_push_cv(self, cv_pair, self->chunk.chunk_counter);571      hasher_push_cv(self, &cv_pair[BLAKE3_OUT_LEN],572                     self->chunk.chunk_counter + (subtree_chunks / 2));573    }574    self->chunk.chunk_counter += subtree_chunks;575    input_bytes += subtree_len;576    input_len -= subtree_len;577  }578 579  // If there's any remaining input less than a full chunk, add it to the chunk580  // state. In that case, also do a final merge loop to make sure the subtree581  // stack doesn't contain any unmerged pairs. The remaining input means we582  // know these merges are non-root. This merge loop isn't strictly necessary583  // here, because hasher_push_chunk_cv already does its own merge loop, but it584  // simplifies blake3_hasher_finalize below.585  if (input_len > 0) {586    chunk_state_update(&self->chunk, input_bytes, input_len);587    hasher_merge_cv_stack(self, self->chunk.chunk_counter);588  }589}590 591void llvm_blake3_hasher_update(blake3_hasher *self, const void *input,592                          size_t input_len) {593  bool use_tbb = false;594  blake3_hasher_update_base(self, input, input_len, use_tbb);595}596 597#if defined(BLAKE3_USE_TBB)598void blake3_hasher_update_tbb(blake3_hasher *self, const void *input,599                              size_t input_len) {600  bool use_tbb = true;601  blake3_hasher_update_base(self, input, input_len, use_tbb);602}603#endif // BLAKE3_USE_TBB604 605void llvm_blake3_hasher_finalize(const blake3_hasher *self, uint8_t *out,606                            size_t out_len) {607  llvm_blake3_hasher_finalize_seek(self, 0, out, out_len);608#if LLVM_MEMORY_SANITIZER_BUILD609  // Avoid false positives due to uninstrumented assembly code.610  __msan_unpoison(out, out_len);611#endif612}613 614void llvm_blake3_hasher_finalize_seek(const blake3_hasher *self, uint64_t seek,615                                 uint8_t *out, size_t out_len) {616  // Explicitly checking for zero avoids causing UB by passing a null pointer617  // to memcpy. This comes up in practice with things like:618  //   std::vector<uint8_t> v;619  //   blake3_hasher_finalize(&hasher, v.data(), v.size());620  if (out_len == 0) {621    return;622  }623 624  // If the subtree stack is empty, then the current chunk is the root.625  if (self->cv_stack_len == 0) {626    output_t output = chunk_state_output(&self->chunk);627    output_root_bytes(&output, seek, out, out_len);628    return;629  }630  // If there are any bytes in the chunk state, finalize that chunk and do a631  // roll-up merge between that chunk hash and every subtree in the stack. In632  // this case, the extra merge loop at the end of blake3_hasher_update633  // guarantees that none of the subtrees in the stack need to be merged with634  // each other first. Otherwise, if there are no bytes in the chunk state,635  // then the top of the stack is a chunk hash, and we start the merge from636  // that.637  output_t output;638  size_t cvs_remaining;639  if (chunk_state_len(&self->chunk) > 0) {640    cvs_remaining = self->cv_stack_len;641    output = chunk_state_output(&self->chunk);642  } else {643    // There are always at least 2 CVs in the stack in this case.644    cvs_remaining = self->cv_stack_len - 2;645    output = parent_output(&self->cv_stack[cvs_remaining * 32], self->key,646                           self->chunk.flags);647  }648  while (cvs_remaining > 0) {649    cvs_remaining -= 1;650    uint8_t parent_block[BLAKE3_BLOCK_LEN];651    memcpy(parent_block, &self->cv_stack[cvs_remaining * 32], 32);652    output_chaining_value(&output, &parent_block[32]);653    output = parent_output(parent_block, self->key, self->chunk.flags);654  }655  output_root_bytes(&output, seek, out, out_len);656}657 658void llvm_blake3_hasher_reset(blake3_hasher *self) {659  chunk_state_reset(&self->chunk, self->key, 0);660  self->cv_stack_len = 0;661}662