From e12f7e3faf4f370ea0dd09bd7bce17d1dabc1df8 Mon Sep 17 00:00:00 2001
From: Hans Henrik Bergan <hans@easy-ad.no>
Date: Fri, 23 Oct 2020 06:56:11 +0200
Subject: [PATCH 1/3] initial blake3 code import (v0.3.7)

this time i'm only importing the portable version..
this will result in decreased performance, obiously..

if i'm importing the assembly optimized versions, it will be some ~30,000 lines ( see https://github.com/php/php-src/pull/6358 )
---
 ext/hash/blake3/blake3.c                      | 603 ++++++++++++++++++
 ext/hash/blake3/blake3.h                      |  58 ++
 ext/hash/blake3/blake3_dispatch.c             | 277 ++++++++
 ext/hash/blake3/blake3_impl.h                 | 269 ++++++++
 ext/hash/blake3/blake3_portable.c             | 160 +++++
 .../helper_script_blake3_version_fetcher.php  |  61 ++
 6 files changed, 1428 insertions(+)
 create mode 100644 ext/hash/blake3/blake3.c
 create mode 100644 ext/hash/blake3/blake3.h
 create mode 100644 ext/hash/blake3/blake3_dispatch.c
 create mode 100644 ext/hash/blake3/blake3_impl.h
 create mode 100644 ext/hash/blake3/blake3_portable.c
 create mode 100644 ext/hash/blake3/helper_script_blake3_version_fetcher.php

diff --git a/ext/hash/blake3/blake3.c b/ext/hash/blake3/blake3.c
new file mode 100644
index 0000000000000..741a76d08bc93
--- /dev/null
+++ b/ext/hash/blake3/blake3.c
@@ -0,0 +1,603 @@
+#include <assert.h>
+#include <stdbool.h>
+#include <string.h>
+
+#include "blake3.h"
+#include "blake3_impl.h"
+
+INLINE void chunk_state_init(blake3_chunk_state *self, const uint32_t key[8],
+                             uint8_t flags) {
+  memcpy(self->cv, key, BLAKE3_KEY_LEN);
+  self->chunk_counter = 0;
+  memset(self->buf, 0, BLAKE3_BLOCK_LEN);
+  self->buf_len = 0;
+  self->blocks_compressed = 0;
+  self->flags = flags;
+}
+
+INLINE void chunk_state_reset(blake3_chunk_state *self, const uint32_t key[8],
+                              uint64_t chunk_counter) {
+  memcpy(self->cv, key, BLAKE3_KEY_LEN);
+  self->chunk_counter = chunk_counter;
+  self->blocks_compressed = 0;
+  memset(self->buf, 0, BLAKE3_BLOCK_LEN);
+  self->buf_len = 0;
+}
+
+INLINE size_t chunk_state_len(const blake3_chunk_state *self) {
+  return (BLAKE3_BLOCK_LEN * (size_t)self->blocks_compressed) +
+         ((size_t)self->buf_len);
+}
+
+INLINE size_t chunk_state_fill_buf(blake3_chunk_state *self,
+                                   const uint8_t *input, size_t input_len) {
+  size_t take = BLAKE3_BLOCK_LEN - ((size_t)self->buf_len);
+  if (take > input_len) {
+    take = input_len;
+  }
+  uint8_t *dest = self->buf + ((size_t)self->buf_len);
+  memcpy(dest, input, take);
+  self->buf_len += (uint8_t)take;
+  return take;
+}
+
+INLINE uint8_t chunk_state_maybe_start_flag(const blake3_chunk_state *self) {
+  if (self->blocks_compressed == 0) {
+    return CHUNK_START;
+  } else {
+    return 0;
+  }
+}
+
+typedef struct {
+  uint32_t input_cv[8];
+  uint64_t counter;
+  uint8_t block[BLAKE3_BLOCK_LEN];
+  uint8_t block_len;
+  uint8_t flags;
+} output_t;
+
+INLINE output_t make_output(const uint32_t input_cv[8],
+                            const uint8_t block[BLAKE3_BLOCK_LEN],
+                            uint8_t block_len, uint64_t counter,
+                            uint8_t flags) {
+  output_t ret;
+  memcpy(ret.input_cv, input_cv, 32);
+  memcpy(ret.block, block, BLAKE3_BLOCK_LEN);
+  ret.block_len = block_len;
+  ret.counter = counter;
+  ret.flags = flags;
+  return ret;
+}
+
+// Chaining values within a given chunk (specifically the compress_in_place
+// interface) are represented as words. This avoids unnecessary bytes<->words
+// conversion overhead in the portable implementation. However, the hash_many
+// interface handles both user input and parent node blocks, so it accepts
+// bytes. For that reason, chaining values in the CV stack are represented as
+// bytes.
+INLINE void output_chaining_value(const output_t *self, uint8_t cv[32]) {
+  uint32_t cv_words[8];
+  memcpy(cv_words, self->input_cv, 32);
+  blake3_compress_in_place(cv_words, self->block, self->block_len,
+                           self->counter, self->flags);
+  store_cv_words(cv, cv_words);
+}
+
+INLINE void output_root_bytes(const output_t *self, uint64_t seek, uint8_t *out,
+                              size_t out_len) {
+  uint64_t output_block_counter = seek / 64;
+  size_t offset_within_block = seek % 64;
+  uint8_t wide_buf[64];
+  while (out_len > 0) {
+    blake3_compress_xof(self->input_cv, self->block, self->block_len,
+                        output_block_counter, self->flags | ROOT, wide_buf);
+    size_t available_bytes = 64 - offset_within_block;
+    size_t memcpy_len;
+    if (out_len > available_bytes) {
+      memcpy_len = available_bytes;
+    } else {
+      memcpy_len = out_len;
+    }
+    memcpy(out, wide_buf + offset_within_block, memcpy_len);
+    out += memcpy_len;
+    out_len -= memcpy_len;
+    output_block_counter += 1;
+    offset_within_block = 0;
+  }
+}
+
+INLINE void chunk_state_update(blake3_chunk_state *self, const uint8_t *input,
+                               size_t input_len) {
+  if (self->buf_len > 0) {
+    size_t take = chunk_state_fill_buf(self, input, input_len);
+    input += take;
+    input_len -= take;
+    if (input_len > 0) {
+      blake3_compress_in_place(
+          self->cv, self->buf, BLAKE3_BLOCK_LEN, self->chunk_counter,
+          self->flags | chunk_state_maybe_start_flag(self));
+      self->blocks_compressed += 1;
+      self->buf_len = 0;
+      memset(self->buf, 0, BLAKE3_BLOCK_LEN);
+    }
+  }
+
+  while (input_len > BLAKE3_BLOCK_LEN) {
+    blake3_compress_in_place(self->cv, input, BLAKE3_BLOCK_LEN,
+                             self->chunk_counter,
+                             self->flags | chunk_state_maybe_start_flag(self));
+    self->blocks_compressed += 1;
+    input += BLAKE3_BLOCK_LEN;
+    input_len -= BLAKE3_BLOCK_LEN;
+  }
+
+  size_t take = chunk_state_fill_buf(self, input, input_len);
+  input += take;
+  input_len -= take;
+}
+
+INLINE output_t chunk_state_output(const blake3_chunk_state *self) {
+  uint8_t block_flags =
+      self->flags | chunk_state_maybe_start_flag(self) | CHUNK_END;
+  return make_output(self->cv, self->buf, self->buf_len, self->chunk_counter,
+                     block_flags);
+}
+
+INLINE output_t parent_output(const uint8_t block[BLAKE3_BLOCK_LEN],
+                              const uint32_t key[8], uint8_t flags) {
+  return make_output(key, block, BLAKE3_BLOCK_LEN, 0, flags | PARENT);
+}
+
+// Given some input larger than one chunk, return the number of bytes that
+// should go in the left subtree. This is the largest power-of-2 number of
+// chunks that leaves at least 1 byte for the right subtree.
+INLINE size_t left_len(size_t content_len) {
+  // Subtract 1 to reserve at least one byte for the right side. content_len
+  // should always be greater than BLAKE3_CHUNK_LEN.
+  size_t full_chunks = (content_len - 1) / BLAKE3_CHUNK_LEN;
+  return round_down_to_power_of_2(full_chunks) * BLAKE3_CHUNK_LEN;
+}
+
+// Use SIMD parallelism to hash up to MAX_SIMD_DEGREE chunks at the same time
+// on a single thread. Write out the chunk chaining values and return the
+// number of chunks hashed. These chunks are never the root and never empty;
+// those cases use a different codepath.
+INLINE size_t compress_chunks_parallel(const uint8_t *input, size_t input_len,
+                                       const uint32_t key[8],
+                                       uint64_t chunk_counter, uint8_t flags,
+                                       uint8_t *out) {
+#if defined(BLAKE3_TESTING)
+  assert(0 < input_len);
+  assert(input_len <= MAX_SIMD_DEGREE * BLAKE3_CHUNK_LEN);
+#endif
+
+  const uint8_t *chunks_array[MAX_SIMD_DEGREE];
+  size_t input_position = 0;
+  size_t chunks_array_len = 0;
+  while (input_len - input_position >= BLAKE3_CHUNK_LEN) {
+    chunks_array[chunks_array_len] = &input[input_position];
+    input_position += BLAKE3_CHUNK_LEN;
+    chunks_array_len += 1;
+  }
+
+  blake3_hash_many(chunks_array, chunks_array_len,
+                   BLAKE3_CHUNK_LEN / BLAKE3_BLOCK_LEN, key, chunk_counter,
+                   true, flags, CHUNK_START, CHUNK_END, out);
+
+  // Hash the remaining partial chunk, if there is one. Note that the empty
+  // chunk (meaning the empty message) is a different codepath.
+  if (input_len > input_position) {
+    uint64_t counter = chunk_counter + (uint64_t)chunks_array_len;
+    blake3_chunk_state chunk_state;
+    chunk_state_init(&chunk_state, key, flags);
+    chunk_state.chunk_counter = counter;
+    chunk_state_update(&chunk_state, &input[input_position],
+                       input_len - input_position);
+    output_t output = chunk_state_output(&chunk_state);
+    output_chaining_value(&output, &out[chunks_array_len * BLAKE3_OUT_LEN]);
+    return chunks_array_len + 1;
+  } else {
+    return chunks_array_len;
+  }
+}
+
+// Use SIMD parallelism to hash up to MAX_SIMD_DEGREE parents at the same time
+// on a single thread. Write out the parent chaining values and return the
+// number of parents hashed. (If there's an odd input chaining value left over,
+// return it as an additional output.) These parents are never the root and
+// never empty; those cases use a different codepath.
+INLINE size_t compress_parents_parallel(const uint8_t *child_chaining_values,
+                                        size_t num_chaining_values,
+                                        const uint32_t key[8], uint8_t flags,
+                                        uint8_t *out) {
+#if defined(BLAKE3_TESTING)
+  assert(2 <= num_chaining_values);
+  assert(num_chaining_values <= 2 * MAX_SIMD_DEGREE_OR_2);
+#endif
+
+  const uint8_t *parents_array[MAX_SIMD_DEGREE_OR_2];
+  size_t parents_array_len = 0;
+  while (num_chaining_values - (2 * parents_array_len) >= 2) {
+    parents_array[parents_array_len] =
+        &child_chaining_values[2 * parents_array_len * BLAKE3_OUT_LEN];
+    parents_array_len += 1;
+  }
+
+  blake3_hash_many(parents_array, parents_array_len, 1, key,
+                   0, // Parents always use counter 0.
+                   false, flags | PARENT,
+                   0, // Parents have no start flags.
+                   0, // Parents have no end flags.
+                   out);
+
+  // If there's an odd child left over, it becomes an output.
+  if (num_chaining_values > 2 * parents_array_len) {
+    memcpy(&out[parents_array_len * BLAKE3_OUT_LEN],
+           &child_chaining_values[2 * parents_array_len * BLAKE3_OUT_LEN],
+           BLAKE3_OUT_LEN);
+    return parents_array_len + 1;
+  } else {
+    return parents_array_len;
+  }
+}
+
+// The wide helper function returns (writes out) an array of chaining values
+// and returns the length of that array. The number of chaining values returned
+// is the dyanmically detected SIMD degree, at most MAX_SIMD_DEGREE. Or fewer,
+// if the input is shorter than that many chunks. The reason for maintaining a
+// wide array of chaining values going back up the tree, is to allow the
+// implementation to hash as many parents in parallel as possible.
+//
+// As a special case when the SIMD degree is 1, this function will still return
+// at least 2 outputs. This guarantees that this function doesn't perform the
+// root compression. (If it did, it would use the wrong flags, and also we
+// wouldn't be able to implement exendable ouput.) Note that this function is
+// not used when the whole input is only 1 chunk long; that's a different
+// codepath.
+//
+// Why not just have the caller split the input on the first update(), instead
+// of implementing this special rule? Because we don't want to limit SIMD or
+// multi-threading parallelism for that update().
+static size_t blake3_compress_subtree_wide(const uint8_t *input,
+                                           size_t input_len,
+                                           const uint32_t key[8],
+                                           uint64_t chunk_counter,
+                                           uint8_t flags, uint8_t *out) {
+  // Note that the single chunk case does *not* bump the SIMD degree up to 2
+  // when it is 1. If this implementation adds multi-threading in the future,
+  // this gives us the option of multi-threading even the 2-chunk case, which
+  // can help performance on smaller platforms.
+  if (input_len <= blake3_simd_degree() * BLAKE3_CHUNK_LEN) {
+    return compress_chunks_parallel(input, input_len, key, chunk_counter, flags,
+                                    out);
+  }
+
+  // With more than simd_degree chunks, we need to recurse. Start by dividing
+  // the input into left and right subtrees. (Note that this is only optimal
+  // as long as the SIMD degree is a power of 2. If we ever get a SIMD degree
+  // of 3 or something, we'll need a more complicated strategy.)
+  size_t left_input_len = left_len(input_len);
+  size_t right_input_len = input_len - left_input_len;
+  const uint8_t *right_input = &input[left_input_len];
+  uint64_t right_chunk_counter =
+      chunk_counter + (uint64_t)(left_input_len / BLAKE3_CHUNK_LEN);
+
+  // Make space for the child outputs. Here we use MAX_SIMD_DEGREE_OR_2 to
+  // account for the special case of returning 2 outputs when the SIMD degree
+  // is 1.
+  uint8_t cv_array[2 * MAX_SIMD_DEGREE_OR_2 * BLAKE3_OUT_LEN];
+  size_t degree = blake3_simd_degree();
+  if (left_input_len > BLAKE3_CHUNK_LEN && degree == 1) {
+    // The special case: We always use a degree of at least two, to make
+    // sure there are two outputs. Except, as noted above, at the chunk
+    // level, where we allow degree=1. (Note that the 1-chunk-input case is
+    // a different codepath.)
+    degree = 2;
+  }
+  uint8_t *right_cvs = &cv_array[degree * BLAKE3_OUT_LEN];
+
+  // Recurse! If this implementation adds multi-threading support in the
+  // future, this is where it will go.
+  size_t left_n = blake3_compress_subtree_wide(input, left_input_len, key,
+                                               chunk_counter, flags, cv_array);
+  size_t right_n = blake3_compress_subtree_wide(
+      right_input, right_input_len, key, right_chunk_counter, flags, right_cvs);
+
+  // The special case again. If simd_degree=1, then we'll have left_n=1 and
+  // right_n=1. Rather than compressing them into a single output, return
+  // them directly, to make sure we always have at least two outputs.
+  if (left_n == 1) {
+    memcpy(out, cv_array, 2 * BLAKE3_OUT_LEN);
+    return 2;
+  }
+
+  // Otherwise, do one layer of parent node compression.
+  size_t num_chaining_values = left_n + right_n;
+  return compress_parents_parallel(cv_array, num_chaining_values, key, flags,
+                                   out);
+}
+
+// Hash a subtree with compress_subtree_wide(), and then condense the resulting
+// list of chaining values down to a single parent node. Don't compress that
+// last parent node, however. Instead, return its message bytes (the
+// concatenated chaining values of its children). This is necessary when the
+// first call to update() supplies a complete subtree, because the topmost
+// parent node of that subtree could end up being the root. It's also necessary
+// for extended output in the general case.
+//
+// As with compress_subtree_wide(), this function is not used on inputs of 1
+// chunk or less. That's a different codepath.
+INLINE void compress_subtree_to_parent_node(
+    const uint8_t *input, size_t input_len, const uint32_t key[8],
+    uint64_t chunk_counter, uint8_t flags, uint8_t out[2 * BLAKE3_OUT_LEN]) {
+#if defined(BLAKE3_TESTING)
+  assert(input_len > BLAKE3_CHUNK_LEN);
+#endif
+
+  uint8_t cv_array[MAX_SIMD_DEGREE_OR_2 * BLAKE3_OUT_LEN];
+  size_t num_cvs = blake3_compress_subtree_wide(input, input_len, key,
+                                                chunk_counter, flags, cv_array);
+
+  // If MAX_SIMD_DEGREE is greater than 2 and there's enough input,
+  // compress_subtree_wide() returns more than 2 chaining values. Condense
+  // them into 2 by forming parent nodes repeatedly.
+  uint8_t out_array[MAX_SIMD_DEGREE_OR_2 * BLAKE3_OUT_LEN / 2];
+  while (num_cvs > 2) {
+    num_cvs =
+        compress_parents_parallel(cv_array, num_cvs, key, flags, out_array);
+    memcpy(cv_array, out_array, num_cvs * BLAKE3_OUT_LEN);
+  }
+  memcpy(out, cv_array, 2 * BLAKE3_OUT_LEN);
+}
+
+INLINE void hasher_init_base(blake3_hasher *self, const uint32_t key[8],
+                             uint8_t flags) {
+  memcpy(self->key, key, BLAKE3_KEY_LEN);
+  chunk_state_init(&self->chunk, key, flags);
+  self->cv_stack_len = 0;
+}
+
+void blake3_hasher_init(blake3_hasher *self) { hasher_init_base(self, IV, 0); }
+
+void blake3_hasher_init_keyed(blake3_hasher *self,
+                              const uint8_t key[BLAKE3_KEY_LEN]) {
+  uint32_t key_words[8];
+  load_key_words(key, key_words);
+  hasher_init_base(self, key_words, KEYED_HASH);
+}
+
+void blake3_hasher_init_derive_key_raw(blake3_hasher *self, const void *context,
+                                       size_t context_len) {
+  blake3_hasher context_hasher;
+  hasher_init_base(&context_hasher, IV, DERIVE_KEY_CONTEXT);
+  blake3_hasher_update(&context_hasher, context, context_len);
+  uint8_t context_key[BLAKE3_KEY_LEN];
+  blake3_hasher_finalize(&context_hasher, context_key, BLAKE3_KEY_LEN);
+  uint32_t context_key_words[8];
+  load_key_words(context_key, context_key_words);
+  hasher_init_base(self, context_key_words, DERIVE_KEY_MATERIAL);
+}
+
+void blake3_hasher_init_derive_key(blake3_hasher *self, const char *context) {
+  blake3_hasher_init_derive_key_raw(self, context, strlen(context));
+}
+
+// As described in hasher_push_cv() below, we do "lazy merging", delaying
+// merges until right before the next CV is about to be added. This is
+// different from the reference implementation. Another difference is that we
+// aren't always merging 1 chunk at a time. Instead, each CV might represent
+// any power-of-two number of chunks, as long as the smaller-above-larger stack
+// order is maintained. Instead of the "count the trailing 0-bits" algorithm
+// described in the spec, we use a "count the total number of 1-bits" variant
+// that doesn't require us to retain the subtree size of the CV on top of the
+// stack. The principle is the same: each CV that should remain in the stack is
+// represented by a 1-bit in the total number of chunks (or bytes) so far.
+INLINE void hasher_merge_cv_stack(blake3_hasher *self, uint64_t total_len) {
+  size_t post_merge_stack_len = (size_t)popcnt(total_len);
+  while (self->cv_stack_len > post_merge_stack_len) {
+    uint8_t *parent_node =
+        &self->cv_stack[(self->cv_stack_len - 2) * BLAKE3_OUT_LEN];
+    output_t output = parent_output(parent_node, self->key, self->chunk.flags);
+    output_chaining_value(&output, parent_node);
+    self->cv_stack_len -= 1;
+  }
+}
+
+// In reference_impl.rs, we merge the new CV with existing CVs from the stack
+// before pushing it. We can do that because we know more input is coming, so
+// we know none of the merges are root.
+//
+// This setting is different. We want to feed as much input as possible to
+// compress_subtree_wide(), without setting aside anything for the chunk_state.
+// If the user gives us 64 KiB, we want to parallelize over all 64 KiB at once
+// as a single subtree, if at all possible.
+//
+// This leads to two problems:
+// 1) This 64 KiB input might be the only call that ever gets made to update.
+//    In this case, the root node of the 64 KiB subtree would be the root node
+//    of the whole tree, and it would need to be ROOT finalized. We can't
+//    compress it until we know.
+// 2) This 64 KiB input might complete a larger tree, whose root node is
+//    similarly going to be the the root of the whole tree. For example, maybe
+//    we have 196 KiB (that is, 128 + 64) hashed so far. We can't compress the
+//    node at the root of the 256 KiB subtree until we know how to finalize it.
+//
+// The second problem is solved with "lazy merging". That is, when we're about
+// to add a CV to the stack, we don't merge it with anything first, as the
+// reference impl does. Instead we do merges using the *previous* CV that was
+// added, which is sitting on top of the stack, and we put the new CV
+// (unmerged) on top of the stack afterwards. This guarantees that we never
+// merge the root node until finalize().
+//
+// Solving the first problem requires an additional tool,
+// compress_subtree_to_parent_node(). That function always returns the top
+// *two* chaining values of the subtree it's compressing. We then do lazy
+// merging with each of them separately, so that the second CV will always
+// remain unmerged. (That also helps us support extendable output when we're
+// hashing an input all-at-once.)
+INLINE void hasher_push_cv(blake3_hasher *self, uint8_t new_cv[BLAKE3_OUT_LEN],
+                           uint64_t chunk_counter) {
+  hasher_merge_cv_stack(self, chunk_counter);
+  memcpy(&self->cv_stack[self->cv_stack_len * BLAKE3_OUT_LEN], new_cv,
+         BLAKE3_OUT_LEN);
+  self->cv_stack_len += 1;
+}
+
+void blake3_hasher_update(blake3_hasher *self, const void *input,
+                          size_t input_len) {
+  // Explicitly checking for zero avoids causing UB by passing a null pointer
+  // to memcpy. This comes up in practice with things like:
+  //   std::vector<uint8_t> v;
+  //   blake3_hasher_update(&hasher, v.data(), v.size());
+  if (input_len == 0) {
+    return;
+  }
+
+  const uint8_t *input_bytes = (const uint8_t *)input;
+
+  // If we have some partial chunk bytes in the internal chunk_state, we need
+  // to finish that chunk first.
+  if (chunk_state_len(&self->chunk) > 0) {
+    size_t take = BLAKE3_CHUNK_LEN - chunk_state_len(&self->chunk);
+    if (take > input_len) {
+      take = input_len;
+    }
+    chunk_state_update(&self->chunk, input_bytes, take);
+    input_bytes += take;
+    input_len -= take;
+    // If we've filled the current chunk and there's more coming, finalize this
+    // chunk and proceed. In this case we know it's not the root.
+    if (input_len > 0) {
+      output_t output = chunk_state_output(&self->chunk);
+      uint8_t chunk_cv[32];
+      output_chaining_value(&output, chunk_cv);
+      hasher_push_cv(self, chunk_cv, self->chunk.chunk_counter);
+      chunk_state_reset(&self->chunk, self->key, self->chunk.chunk_counter + 1);
+    } else {
+      return;
+    }
+  }
+
+  // Now the chunk_state is clear, and we have more input. If there's more than
+  // a single chunk (so, definitely not the root chunk), hash the largest whole
+  // subtree we can, with the full benefits of SIMD (and maybe in the future,
+  // multi-threading) parallelism. Two restrictions:
+  // - The subtree has to be a power-of-2 number of chunks. Only subtrees along
+  //   the right edge can be incomplete, and we don't know where the right edge
+  //   is going to be until we get to finalize().
+  // - The subtree must evenly divide the total number of chunks up until this
+  //   point (if total is not 0). If the current incomplete subtree is only
+  //   waiting for 1 more chunk, we can't hash a subtree of 4 chunks. We have
+  //   to complete the current subtree first.
+  // Because we might need to break up the input to form powers of 2, or to
+  // evenly divide what we already have, this part runs in a loop.
+  while (input_len > BLAKE3_CHUNK_LEN) {
+    size_t subtree_len = round_down_to_power_of_2(input_len);
+    uint64_t count_so_far = self->chunk.chunk_counter * BLAKE3_CHUNK_LEN;
+    // Shrink the subtree_len until it evenly divides the count so far. We know
+    // that subtree_len itself is a power of 2, so we can use a bitmasking
+    // trick instead of an actual remainder operation. (Note that if the caller
+    // consistently passes power-of-2 inputs of the same size, as is hopefully
+    // typical, this loop condition will always fail, and subtree_len will
+    // always be the full length of the input.)
+    //
+    // An aside: We don't have to shrink subtree_len quite this much. For
+    // example, if count_so_far is 1, we could pass 2 chunks to
+    // compress_subtree_to_parent_node. Since we'll get 2 CVs back, we'll still
+    // get the right answer in the end, and we might get to use 2-way SIMD
+    // parallelism. The problem with this optimization, is that it gets us
+    // stuck always hashing 2 chunks. The total number of chunks will remain
+    // odd, and we'll never graduate to higher degrees of parallelism. See
+    // https://github.com/BLAKE3-team/BLAKE3/issues/69.
+    while ((((uint64_t)(subtree_len - 1)) & count_so_far) != 0) {
+      subtree_len /= 2;
+    }
+    // The shrunken subtree_len might now be 1 chunk long. If so, hash that one
+    // chunk by itself. Otherwise, compress the subtree into a pair of CVs.
+    uint64_t subtree_chunks = subtree_len / BLAKE3_CHUNK_LEN;
+    if (subtree_len <= BLAKE3_CHUNK_LEN) {
+      blake3_chunk_state chunk_state;
+      chunk_state_init(&chunk_state, self->key, self->chunk.flags);
+      chunk_state.chunk_counter = self->chunk.chunk_counter;
+      chunk_state_update(&chunk_state, input_bytes, subtree_len);
+      output_t output = chunk_state_output(&chunk_state);
+      uint8_t cv[BLAKE3_OUT_LEN];
+      output_chaining_value(&output, cv);
+      hasher_push_cv(self, cv, chunk_state.chunk_counter);
+    } else {
+      // This is the high-performance happy path, though getting here depends
+      // on the caller giving us a long enough input.
+      uint8_t cv_pair[2 * BLAKE3_OUT_LEN];
+      compress_subtree_to_parent_node(input_bytes, subtree_len, self->key,
+                                      self->chunk.chunk_counter,
+                                      self->chunk.flags, cv_pair);
+      hasher_push_cv(self, cv_pair, self->chunk.chunk_counter);
+      hasher_push_cv(self, &cv_pair[BLAKE3_OUT_LEN],
+                     self->chunk.chunk_counter + (subtree_chunks / 2));
+    }
+    self->chunk.chunk_counter += subtree_chunks;
+    input_bytes += subtree_len;
+    input_len -= subtree_len;
+  }
+
+  // If there's any remaining input less than a full chunk, add it to the chunk
+  // state. In that case, also do a final merge loop to make sure the subtree
+  // stack doesn't contain any unmerged pairs. The remaining input means we
+  // know these merges are non-root. This merge loop isn't strictly necessary
+  // here, because hasher_push_chunk_cv already does its own merge loop, but it
+  // simplifies blake3_hasher_finalize below.
+  if (input_len > 0) {
+    chunk_state_update(&self->chunk, input_bytes, input_len);
+    hasher_merge_cv_stack(self, self->chunk.chunk_counter);
+  }
+}
+
+void blake3_hasher_finalize(const blake3_hasher *self, uint8_t *out,
+                            size_t out_len) {
+  blake3_hasher_finalize_seek(self, 0, out, out_len);
+}
+
+void blake3_hasher_finalize_seek(const blake3_hasher *self, uint64_t seek,
+                                 uint8_t *out, size_t out_len) {
+  // Explicitly checking for zero avoids causing UB by passing a null pointer
+  // to memcpy. This comes up in practice with things like:
+  //   std::vector<uint8_t> v;
+  //   blake3_hasher_finalize(&hasher, v.data(), v.size());
+  if (out_len == 0) {
+    return;
+  }
+
+  // If the subtree stack is empty, then the current chunk is the root.
+  if (self->cv_stack_len == 0) {
+    output_t output = chunk_state_output(&self->chunk);
+    output_root_bytes(&output, seek, out, out_len);
+    return;
+  }
+  // If there are any bytes in the chunk state, finalize that chunk and do a
+  // roll-up merge between that chunk hash and every subtree in the stack. In
+  // this case, the extra merge loop at the end of blake3_hasher_update
+  // guarantees that none of the subtrees in the stack need to be merged with
+  // each other first. Otherwise, if there are no bytes in the chunk state,
+  // then the top of the stack is a chunk hash, and we start the merge from
+  // that.
+  output_t output;
+  size_t cvs_remaining;
+  if (chunk_state_len(&self->chunk) > 0) {
+    cvs_remaining = self->cv_stack_len;
+    output = chunk_state_output(&self->chunk);
+  } else {
+    // There are always at least 2 CVs in the stack in this case.
+    cvs_remaining = self->cv_stack_len - 2;
+    output = parent_output(&self->cv_stack[cvs_remaining * 32], self->key,
+                           self->chunk.flags);
+  }
+  while (cvs_remaining > 0) {
+    cvs_remaining -= 1;
+    uint8_t parent_block[BLAKE3_BLOCK_LEN];
+    memcpy(parent_block, &self->cv_stack[cvs_remaining * 32], 32);
+    output_chaining_value(&output, &parent_block[32]);
+    output = parent_output(parent_block, self->key, self->chunk.flags);
+  }
+  output_root_bytes(&output, seek, out, out_len);
+}
diff --git a/ext/hash/blake3/blake3.h b/ext/hash/blake3/blake3.h
new file mode 100644
index 0000000000000..51f1d2a8672bb
--- /dev/null
+++ b/ext/hash/blake3/blake3.h
@@ -0,0 +1,58 @@
+#ifndef BLAKE3_H
+#define BLAKE3_H
+
+#include <stddef.h>
+#include <stdint.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define BLAKE3_KEY_LEN 32
+#define BLAKE3_OUT_LEN 32
+#define BLAKE3_BLOCK_LEN 64
+#define BLAKE3_CHUNK_LEN 1024
+#define BLAKE3_MAX_DEPTH 54
+#define BLAKE3_MAX_SIMD_DEGREE 16
+
+// This struct is a private implementation detail. It has to be here because
+// it's part of blake3_hasher below.
+typedef struct {
+  uint32_t cv[8];
+  uint64_t chunk_counter;
+  uint8_t buf[BLAKE3_BLOCK_LEN];
+  uint8_t buf_len;
+  uint8_t blocks_compressed;
+  uint8_t flags;
+} blake3_chunk_state;
+
+typedef struct {
+  uint32_t key[8];
+  blake3_chunk_state chunk;
+  uint8_t cv_stack_len;
+  // The stack size is MAX_DEPTH + 1 because we do lazy merging. For example,
+  // with 7 chunks, we have 3 entries in the stack. Adding an 8th chunk
+  // requires a 4th entry, rather than merging everything down to 1, because we
+  // don't know whether more input is coming. This is different from how the
+  // reference implementation does things.
+  uint8_t cv_stack[(BLAKE3_MAX_DEPTH + 1) * BLAKE3_OUT_LEN];
+} blake3_hasher;
+
+void blake3_hasher_init(blake3_hasher *self);
+void blake3_hasher_init_keyed(blake3_hasher *self,
+                              const uint8_t key[BLAKE3_KEY_LEN]);
+void blake3_hasher_init_derive_key(blake3_hasher *self, const char *context);
+void blake3_hasher_init_derive_key_raw(blake3_hasher *self, const void *context, 
+                                       size_t context_len);
+void blake3_hasher_update(blake3_hasher *self, const void *input,
+                          size_t input_len);
+void blake3_hasher_finalize(const blake3_hasher *self, uint8_t *out,
+                            size_t out_len);
+void blake3_hasher_finalize_seek(const blake3_hasher *self, uint64_t seek,
+                                 uint8_t *out, size_t out_len);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /* BLAKE3_H */
diff --git a/ext/hash/blake3/blake3_dispatch.c b/ext/hash/blake3/blake3_dispatch.c
new file mode 100644
index 0000000000000..f7e6a787612e8
--- /dev/null
+++ b/ext/hash/blake3/blake3_dispatch.c
@@ -0,0 +1,277 @@
+#include <stdbool.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#include "blake3_impl.h"
+
+#if defined(IS_X86)
+#if defined(_MSC_VER)
+#include <intrin.h>
+#elif defined(__GNUC__)
+#include <immintrin.h>
+#else
+#error "Unimplemented!"
+#endif
+#endif
+
+#define MAYBE_UNUSED(x) (void)((x))
+
+#if defined(IS_X86)
+static uint64_t xgetbv() {
+#if defined(_MSC_VER)
+  return _xgetbv(0);
+#else
+  uint32_t eax = 0, edx = 0;
+  __asm__ __volatile__("xgetbv\n" : "=a"(eax), "=d"(edx) : "c"(0));
+  return ((uint64_t)edx << 32) | eax;
+#endif
+}
+
+static void cpuid(uint32_t out[4], uint32_t id) {
+#if defined(_MSC_VER)
+  __cpuid((int *)out, id);
+#elif defined(__i386__) || defined(_M_IX86)
+  __asm__ __volatile__("movl %%ebx, %1\n"
+                       "cpuid\n"
+                       "xchgl %1, %%ebx\n"
+                       : "=a"(out[0]), "=r"(out[1]), "=c"(out[2]), "=d"(out[3])
+                       : "a"(id));
+#else
+  __asm__ __volatile__("cpuid\n"
+                       : "=a"(out[0]), "=b"(out[1]), "=c"(out[2]), "=d"(out[3])
+                       : "a"(id));
+#endif
+}
+
+static void cpuidex(uint32_t out[4], uint32_t id, uint32_t sid) {
+#if defined(_MSC_VER)
+  __cpuidex((int *)out, id, sid);
+#elif defined(__i386__) || defined(_M_IX86)
+  __asm__ __volatile__("movl %%ebx, %1\n"
+                       "cpuid\n"
+                       "xchgl %1, %%ebx\n"
+                       : "=a"(out[0]), "=r"(out[1]), "=c"(out[2]), "=d"(out[3])
+                       : "a"(id), "c"(sid));
+#else
+  __asm__ __volatile__("cpuid\n"
+                       : "=a"(out[0]), "=b"(out[1]), "=c"(out[2]), "=d"(out[3])
+                       : "a"(id), "c"(sid));
+#endif
+}
+
+#endif
+
+enum cpu_feature {
+  SSE2 = 1 << 0,
+  SSSE3 = 1 << 1,
+  SSE41 = 1 << 2,
+  AVX = 1 << 3,
+  AVX2 = 1 << 4,
+  AVX512F = 1 << 5,
+  AVX512VL = 1 << 6,
+  /* ... */
+  UNDEFINED = 1 << 30
+};
+
+#if !defined(BLAKE3_TESTING)
+static /* Allow the variable to be controlled manually for testing */
+#endif
+    enum cpu_feature g_cpu_features = UNDEFINED;
+
+#if !defined(BLAKE3_TESTING)
+static
+#endif
+    enum cpu_feature
+    get_cpu_features() {
+
+  if (g_cpu_features != UNDEFINED) {
+    return g_cpu_features;
+  } else {
+#if defined(IS_X86)
+    uint32_t regs[4] = {0};
+    uint32_t *eax = &regs[0], *ebx = &regs[1], *ecx = &regs[2], *edx = &regs[3];
+    (void)edx;
+    enum cpu_feature features = 0;
+    cpuid(regs, 0);
+    const int max_id = *eax;
+    cpuid(regs, 1);
+#if defined(__amd64__) || defined(_M_X64)
+    features |= SSE2;
+#else
+    if (*edx & (1UL << 26))
+      features |= SSE2;
+#endif
+    if (*ecx & (1UL << 0))
+      features |= SSSE3;
+    if (*ecx & (1UL << 19))
+      features |= SSE41;
+
+    if (*ecx & (1UL << 27)) { // OSXSAVE
+      const uint64_t mask = xgetbv();
+      if ((mask & 6) == 6) { // SSE and AVX states
+        if (*ecx & (1UL << 28))
+          features |= AVX;
+        if (max_id >= 7) {
+          cpuidex(regs, 7, 0);
+          if (*ebx & (1UL << 5))
+            features |= AVX2;
+          if ((mask & 224) == 224) { // Opmask, ZMM_Hi256, Hi16_Zmm
+            if (*ebx & (1UL << 31))
+              features |= AVX512VL;
+            if (*ebx & (1UL << 16))
+              features |= AVX512F;
+          }
+        }
+      }
+    }
+    g_cpu_features = features;
+    return features;
+#else
+    /* How to detect NEON? */
+    return 0;
+#endif
+  }
+}
+
+void blake3_compress_in_place(uint32_t cv[8],
+                              const uint8_t block[BLAKE3_BLOCK_LEN],
+                              uint8_t block_len, uint64_t counter,
+                              uint8_t flags) {
+#if defined(IS_X86)
+  const enum cpu_feature features = get_cpu_features();
+  MAYBE_UNUSED(features);
+#if !defined(BLAKE3_NO_AVX512)
+  if (features & AVX512VL) {
+    blake3_compress_in_place_avx512(cv, block, block_len, counter, flags);
+    return;
+  }
+#endif
+#if !defined(BLAKE3_NO_SSE41)
+  if (features & SSE41) {
+    blake3_compress_in_place_sse41(cv, block, block_len, counter, flags);
+    return;
+  }
+#endif
+#if !defined(BLAKE3_NO_SSE2)
+  if (features & SSE2) {
+    blake3_compress_in_place_sse2(cv, block, block_len, counter, flags);
+    return;
+  }
+#endif
+#endif
+  blake3_compress_in_place_portable(cv, block, block_len, counter, flags);
+}
+
+void blake3_compress_xof(const uint32_t cv[8],
+                         const uint8_t block[BLAKE3_BLOCK_LEN],
+                         uint8_t block_len, uint64_t counter, uint8_t flags,
+                         uint8_t out[64]) {
+#if defined(IS_X86)
+  const enum cpu_feature features = get_cpu_features();
+  MAYBE_UNUSED(features);
+#if !defined(BLAKE3_NO_AVX512)
+  if (features & AVX512VL) {
+    blake3_compress_xof_avx512(cv, block, block_len, counter, flags, out);
+    return;
+  }
+#endif
+#if !defined(BLAKE3_NO_SSE41)
+  if (features & SSE41) {
+    blake3_compress_xof_sse41(cv, block, block_len, counter, flags, out);
+    return;
+  }
+#endif
+#if !defined(BLAKE3_NO_SSE2)
+  if (features & SSE2) {
+    blake3_compress_xof_sse2(cv, block, block_len, counter, flags, out);
+    return;
+  }
+#endif
+#endif
+  blake3_compress_xof_portable(cv, block, block_len, counter, flags, out);
+}
+
+void blake3_hash_many(const uint8_t *const *inputs, size_t num_inputs,
+                      size_t blocks, const uint32_t key[8], uint64_t counter,
+                      bool increment_counter, uint8_t flags,
+                      uint8_t flags_start, uint8_t flags_end, uint8_t *out) {
+#if defined(IS_X86)
+  const enum cpu_feature features = get_cpu_features();
+  MAYBE_UNUSED(features);
+#if !defined(BLAKE3_NO_AVX512)
+  if ((features & (AVX512F|AVX512VL)) == (AVX512F|AVX512VL)) {
+    blake3_hash_many_avx512(inputs, num_inputs, blocks, key, counter,
+                            increment_counter, flags, flags_start, flags_end,
+                            out);
+    return;
+  }
+#endif
+#if !defined(BLAKE3_NO_AVX2)
+  if (features & AVX2) {
+    blake3_hash_many_avx2(inputs, num_inputs, blocks, key, counter,
+                          increment_counter, flags, flags_start, flags_end,
+                          out);
+    return;
+  }
+#endif
+#if !defined(BLAKE3_NO_SSE41)
+  if (features & SSE41) {
+    blake3_hash_many_sse41(inputs, num_inputs, blocks, key, counter,
+                           increment_counter, flags, flags_start, flags_end,
+                           out);
+    return;
+  }
+#endif
+#if !defined(BLAKE3_NO_SSE2)
+  if (features & SSE2) {
+    blake3_hash_many_sse2(inputs, num_inputs, blocks, key, counter,
+                          increment_counter, flags, flags_start, flags_end,
+                          out);
+    return;
+  }
+#endif
+#endif
+
+#if defined(BLAKE3_USE_NEON)
+  blake3_hash_many_neon(inputs, num_inputs, blocks, key, counter,
+                        increment_counter, flags, flags_start, flags_end, out);
+  return;
+#endif
+
+  blake3_hash_many_portable(inputs, num_inputs, blocks, key, counter,
+                            increment_counter, flags, flags_start, flags_end,
+                            out);
+}
+
+// The dynamically detected SIMD degree of the current platform.
+size_t blake3_simd_degree(void) {
+#if defined(IS_X86)
+  const enum cpu_feature features = get_cpu_features();
+  MAYBE_UNUSED(features);
+#if !defined(BLAKE3_NO_AVX512)
+  if ((features & (AVX512F|AVX512VL)) == (AVX512F|AVX512VL)) {
+    return 16;
+  }
+#endif
+#if !defined(BLAKE3_NO_AVX2)
+  if (features & AVX2) {
+    return 8;
+  }
+#endif
+#if !defined(BLAKE3_NO_SSE41)
+  if (features & SSE41) {
+    return 4;
+  }
+#endif
+#if !defined(BLAKE3_NO_SSE2)
+  if (features & SSE2) {
+    return 4;
+  }
+#endif
+#endif
+#if defined(BLAKE3_USE_NEON)
+  return 4;
+#endif
+  return 1;
+}
+
diff --git a/ext/hash/blake3/blake3_impl.h b/ext/hash/blake3/blake3_impl.h
new file mode 100644
index 0000000000000..86ab6aa253a38
--- /dev/null
+++ b/ext/hash/blake3/blake3_impl.h
@@ -0,0 +1,269 @@
+#ifndef BLAKE3_IMPL_H
+#define BLAKE3_IMPL_H
+
+#include <assert.h>
+#include <stdbool.h>
+#include <stddef.h>
+#include <stdint.h>
+#include <string.h>
+
+#include "blake3.h"
+
+// internal flags
+enum blake3_flags {
+  CHUNK_START         = 1 << 0,
+  CHUNK_END           = 1 << 1,
+  PARENT              = 1 << 2,
+  ROOT                = 1 << 3,
+  KEYED_HASH          = 1 << 4,
+  DERIVE_KEY_CONTEXT  = 1 << 5,
+  DERIVE_KEY_MATERIAL = 1 << 6,
+};
+
+// This C implementation tries to support recent versions of GCC, Clang, and
+// MSVC.
+#if defined(_MSC_VER)
+#define INLINE static __forceinline
+#else
+#define INLINE static inline __attribute__((always_inline))
+#endif
+
+#if defined(__x86_64__) || defined(_M_X64) 
+#define IS_X86
+#define IS_X86_64
+#endif
+
+#if defined(__i386__) || defined(_M_IX86)
+#define IS_X86
+#define IS_X86_32
+#endif
+
+#if defined(IS_X86)
+#if defined(_MSC_VER)
+#include <intrin.h>
+#endif
+#include <immintrin.h>
+#endif
+
+#if defined(IS_X86)
+#define MAX_SIMD_DEGREE 16
+#elif defined(BLAKE3_USE_NEON)
+#define MAX_SIMD_DEGREE 4
+#else
+#define MAX_SIMD_DEGREE 1
+#endif
+
+// There are some places where we want a static size that's equal to the
+// MAX_SIMD_DEGREE, but also at least 2.
+#define MAX_SIMD_DEGREE_OR_2 (MAX_SIMD_DEGREE > 2 ? MAX_SIMD_DEGREE : 2)
+
+static const uint32_t IV[8] = {0x6A09E667UL, 0xBB67AE85UL, 0x3C6EF372UL,
+                               0xA54FF53AUL, 0x510E527FUL, 0x9B05688CUL,
+                               0x1F83D9ABUL, 0x5BE0CD19UL};
+
+static const uint8_t MSG_SCHEDULE[7][16] = {
+    {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15},
+    {2, 6, 3, 10, 7, 0, 4, 13, 1, 11, 12, 5, 9, 14, 15, 8},
+    {3, 4, 10, 12, 13, 2, 7, 14, 6, 5, 9, 0, 11, 15, 8, 1},
+    {10, 7, 12, 9, 14, 3, 13, 15, 4, 0, 11, 2, 5, 8, 1, 6},
+    {12, 13, 9, 11, 15, 10, 14, 8, 7, 2, 5, 3, 0, 1, 6, 4},
+    {9, 14, 11, 5, 8, 12, 15, 1, 13, 3, 0, 10, 2, 6, 4, 7},
+    {11, 15, 5, 0, 1, 9, 8, 6, 14, 10, 2, 12, 3, 4, 7, 13},
+};
+
+/* Find index of the highest set bit */
+/* x is assumed to be nonzero.       */
+static unsigned int highest_one(uint64_t x) {
+#if defined(__GNUC__) || defined(__clang__)
+  return 63 ^ __builtin_clzll(x);
+#elif defined(_MSC_VER) && defined(IS_X86_64)
+  unsigned long index;
+  _BitScanReverse64(&index, x);
+  return index;
+#elif defined(_MSC_VER) && defined(IS_X86_32)
+  if(x >> 32) {
+    unsigned long index;
+    _BitScanReverse(&index, x >> 32);
+    return 32 + index;
+  } else {
+    unsigned long index;
+    _BitScanReverse(&index, x);
+    return index;
+  }
+#else
+  unsigned int c = 0;
+  if(x & 0xffffffff00000000ULL) { x >>= 32; c += 32; }
+  if(x & 0x00000000ffff0000ULL) { x >>= 16; c += 16; }
+  if(x & 0x000000000000ff00ULL) { x >>=  8; c +=  8; }
+  if(x & 0x00000000000000f0ULL) { x >>=  4; c +=  4; }
+  if(x & 0x000000000000000cULL) { x >>=  2; c +=  2; }
+  if(x & 0x0000000000000002ULL) {           c +=  1; }
+  return c;
+#endif
+}
+
+// Count the number of 1 bits.
+INLINE unsigned int popcnt(uint64_t x) {
+#if defined(__GNUC__) || defined(__clang__)
+  return __builtin_popcountll(x);
+#else
+  unsigned int count = 0;
+  while (x != 0) {
+    count += 1;
+    x &= x - 1;
+  }
+  return count;
+#endif
+}
+
+// Largest power of two less than or equal to x. As a special case, returns 1
+// when x is 0. 
+INLINE uint64_t round_down_to_power_of_2(uint64_t x) {
+  return 1ULL << highest_one(x | 1);
+}
+
+INLINE uint32_t counter_low(uint64_t counter) { return (uint32_t)counter; }
+
+INLINE uint32_t counter_high(uint64_t counter) {
+  return (uint32_t)(counter >> 32);
+}
+
+INLINE uint32_t load32(const void *src) {
+  const uint8_t *p = (const uint8_t *)src;
+  return ((uint32_t)(p[0]) << 0) | ((uint32_t)(p[1]) << 8) |
+         ((uint32_t)(p[2]) << 16) | ((uint32_t)(p[3]) << 24);
+}
+
+INLINE void load_key_words(const uint8_t key[BLAKE3_KEY_LEN],
+                           uint32_t key_words[8]) {
+  key_words[0] = load32(&key[0 * 4]);
+  key_words[1] = load32(&key[1 * 4]);
+  key_words[2] = load32(&key[2 * 4]);
+  key_words[3] = load32(&key[3 * 4]);
+  key_words[4] = load32(&key[4 * 4]);
+  key_words[5] = load32(&key[5 * 4]);
+  key_words[6] = load32(&key[6 * 4]);
+  key_words[7] = load32(&key[7 * 4]);
+}
+
+INLINE void store32(void *dst, uint32_t w) {
+  uint8_t *p = (uint8_t *)dst;
+  p[0] = (uint8_t)(w >> 0);
+  p[1] = (uint8_t)(w >> 8);
+  p[2] = (uint8_t)(w >> 16);
+  p[3] = (uint8_t)(w >> 24);
+}
+
+INLINE void store_cv_words(uint8_t bytes_out[32], uint32_t cv_words[8]) {
+  store32(&bytes_out[0 * 4], cv_words[0]);
+  store32(&bytes_out[1 * 4], cv_words[1]);
+  store32(&bytes_out[2 * 4], cv_words[2]);
+  store32(&bytes_out[3 * 4], cv_words[3]);
+  store32(&bytes_out[4 * 4], cv_words[4]);
+  store32(&bytes_out[5 * 4], cv_words[5]);
+  store32(&bytes_out[6 * 4], cv_words[6]);
+  store32(&bytes_out[7 * 4], cv_words[7]);
+}
+
+void blake3_compress_in_place(uint32_t cv[8],
+                              const uint8_t block[BLAKE3_BLOCK_LEN],
+                              uint8_t block_len, uint64_t counter,
+                              uint8_t flags);
+
+void blake3_compress_xof(const uint32_t cv[8],
+                         const uint8_t block[BLAKE3_BLOCK_LEN],
+                         uint8_t block_len, uint64_t counter, uint8_t flags,
+                         uint8_t out[64]);
+
+void blake3_hash_many(const uint8_t *const *inputs, size_t num_inputs,
+                      size_t blocks, const uint32_t key[8], uint64_t counter,
+                      bool increment_counter, uint8_t flags,
+                      uint8_t flags_start, uint8_t flags_end, uint8_t *out);
+
+size_t blake3_simd_degree(void);
+
+
+// Declarations for implementation-specific functions.
+void blake3_compress_in_place_portable(uint32_t cv[8],
+                                       const uint8_t block[BLAKE3_BLOCK_LEN],
+                                       uint8_t block_len, uint64_t counter,
+                                       uint8_t flags);
+
+void blake3_compress_xof_portable(const uint32_t cv[8],
+                                  const uint8_t block[BLAKE3_BLOCK_LEN],
+                                  uint8_t block_len, uint64_t counter,
+                                  uint8_t flags, uint8_t out[64]);
+
+void blake3_hash_many_portable(const uint8_t *const *inputs, size_t num_inputs,
+                               size_t blocks, const uint32_t key[8],
+                               uint64_t counter, bool increment_counter,
+                               uint8_t flags, uint8_t flags_start,
+                               uint8_t flags_end, uint8_t *out);
+
+#if defined(IS_X86)
+#if !defined(BLAKE3_NO_SSE2)
+void blake3_compress_in_place_sse2(uint32_t cv[8],
+                                   const uint8_t block[BLAKE3_BLOCK_LEN],
+                                   uint8_t block_len, uint64_t counter,
+                                   uint8_t flags);
+void blake3_compress_xof_sse2(const uint32_t cv[8],
+                              const uint8_t block[BLAKE3_BLOCK_LEN],
+                              uint8_t block_len, uint64_t counter,
+                              uint8_t flags, uint8_t out[64]);
+void blake3_hash_many_sse2(const uint8_t *const *inputs, size_t num_inputs,
+                           size_t blocks, const uint32_t key[8],
+                           uint64_t counter, bool increment_counter,
+                           uint8_t flags, uint8_t flags_start,
+                           uint8_t flags_end, uint8_t *out);
+#endif
+#if !defined(BLAKE3_NO_SSE41)
+void blake3_compress_in_place_sse41(uint32_t cv[8],
+                                    const uint8_t block[BLAKE3_BLOCK_LEN],
+                                    uint8_t block_len, uint64_t counter,
+                                    uint8_t flags);
+void blake3_compress_xof_sse41(const uint32_t cv[8],
+                               const uint8_t block[BLAKE3_BLOCK_LEN],
+                               uint8_t block_len, uint64_t counter,
+                               uint8_t flags, uint8_t out[64]);
+void blake3_hash_many_sse41(const uint8_t *const *inputs, size_t num_inputs,
+                            size_t blocks, const uint32_t key[8],
+                            uint64_t counter, bool increment_counter,
+                            uint8_t flags, uint8_t flags_start,
+                            uint8_t flags_end, uint8_t *out);
+#endif
+#if !defined(BLAKE3_NO_AVX2)
+void blake3_hash_many_avx2(const uint8_t *const *inputs, size_t num_inputs,
+                           size_t blocks, const uint32_t key[8],
+                           uint64_t counter, bool increment_counter,
+                           uint8_t flags, uint8_t flags_start,
+                           uint8_t flags_end, uint8_t *out);
+#endif
+#if !defined(BLAKE3_NO_AVX512)
+void blake3_compress_in_place_avx512(uint32_t cv[8],
+                                     const uint8_t block[BLAKE3_BLOCK_LEN],
+                                     uint8_t block_len, uint64_t counter,
+                                     uint8_t flags);
+
+void blake3_compress_xof_avx512(const uint32_t cv[8],
+                                const uint8_t block[BLAKE3_BLOCK_LEN],
+                                uint8_t block_len, uint64_t counter,
+                                uint8_t flags, uint8_t out[64]);
+
+void blake3_hash_many_avx512(const uint8_t *const *inputs, size_t num_inputs,
+                             size_t blocks, const uint32_t key[8],
+                             uint64_t counter, bool increment_counter,
+                             uint8_t flags, uint8_t flags_start,
+                             uint8_t flags_end, uint8_t *out);
+#endif
+#endif
+
+#if defined(BLAKE3_USE_NEON)
+void blake3_hash_many_neon(const uint8_t *const *inputs, size_t num_inputs,
+                           size_t blocks, const uint32_t key[8],
+                           uint64_t counter, bool increment_counter,
+                           uint8_t flags, uint8_t flags_start,
+                           uint8_t flags_end, uint8_t *out);
+#endif
+
+
+#endif /* BLAKE3_IMPL_H */
diff --git a/ext/hash/blake3/blake3_portable.c b/ext/hash/blake3/blake3_portable.c
new file mode 100644
index 0000000000000..062dd1b47fb64
--- /dev/null
+++ b/ext/hash/blake3/blake3_portable.c
@@ -0,0 +1,160 @@
+#include "blake3_impl.h"
+#include <string.h>
+
+INLINE uint32_t rotr32(uint32_t w, uint32_t c) {
+  return (w >> c) | (w << (32 - c));
+}
+
+INLINE void g(uint32_t *state, size_t a, size_t b, size_t c, size_t d,
+              uint32_t x, uint32_t y) {
+  state[a] = state[a] + state[b] + x;
+  state[d] = rotr32(state[d] ^ state[a], 16);
+  state[c] = state[c] + state[d];
+  state[b] = rotr32(state[b] ^ state[c], 12);
+  state[a] = state[a] + state[b] + y;
+  state[d] = rotr32(state[d] ^ state[a], 8);
+  state[c] = state[c] + state[d];
+  state[b] = rotr32(state[b] ^ state[c], 7);
+}
+
+INLINE void round_fn(uint32_t state[16], const uint32_t *msg, size_t round) {
+  // Select the message schedule based on the round.
+  const uint8_t *schedule = MSG_SCHEDULE[round];
+
+  // Mix the columns.
+  g(state, 0, 4, 8, 12, msg[schedule[0]], msg[schedule[1]]);
+  g(state, 1, 5, 9, 13, msg[schedule[2]], msg[schedule[3]]);
+  g(state, 2, 6, 10, 14, msg[schedule[4]], msg[schedule[5]]);
+  g(state, 3, 7, 11, 15, msg[schedule[6]], msg[schedule[7]]);
+
+  // Mix the rows.
+  g(state, 0, 5, 10, 15, msg[schedule[8]], msg[schedule[9]]);
+  g(state, 1, 6, 11, 12, msg[schedule[10]], msg[schedule[11]]);
+  g(state, 2, 7, 8, 13, msg[schedule[12]], msg[schedule[13]]);
+  g(state, 3, 4, 9, 14, msg[schedule[14]], msg[schedule[15]]);
+}
+
+INLINE void compress_pre(uint32_t state[16], const uint32_t cv[8],
+                         const uint8_t block[BLAKE3_BLOCK_LEN],
+                         uint8_t block_len, uint64_t counter, uint8_t flags) {
+  uint32_t block_words[16];
+  block_words[0] = load32(block + 4 * 0);
+  block_words[1] = load32(block + 4 * 1);
+  block_words[2] = load32(block + 4 * 2);
+  block_words[3] = load32(block + 4 * 3);
+  block_words[4] = load32(block + 4 * 4);
+  block_words[5] = load32(block + 4 * 5);
+  block_words[6] = load32(block + 4 * 6);
+  block_words[7] = load32(block + 4 * 7);
+  block_words[8] = load32(block + 4 * 8);
+  block_words[9] = load32(block + 4 * 9);
+  block_words[10] = load32(block + 4 * 10);
+  block_words[11] = load32(block + 4 * 11);
+  block_words[12] = load32(block + 4 * 12);
+  block_words[13] = load32(block + 4 * 13);
+  block_words[14] = load32(block + 4 * 14);
+  block_words[15] = load32(block + 4 * 15);
+
+  state[0] = cv[0];
+  state[1] = cv[1];
+  state[2] = cv[2];
+  state[3] = cv[3];
+  state[4] = cv[4];
+  state[5] = cv[5];
+  state[6] = cv[6];
+  state[7] = cv[7];
+  state[8] = IV[0];
+  state[9] = IV[1];
+  state[10] = IV[2];
+  state[11] = IV[3];
+  state[12] = counter_low(counter);
+  state[13] = counter_high(counter);
+  state[14] = (uint32_t)block_len;
+  state[15] = (uint32_t)flags;
+
+  round_fn(state, &block_words[0], 0);
+  round_fn(state, &block_words[0], 1);
+  round_fn(state, &block_words[0], 2);
+  round_fn(state, &block_words[0], 3);
+  round_fn(state, &block_words[0], 4);
+  round_fn(state, &block_words[0], 5);
+  round_fn(state, &block_words[0], 6);
+}
+
+void blake3_compress_in_place_portable(uint32_t cv[8],
+                                       const uint8_t block[BLAKE3_BLOCK_LEN],
+                                       uint8_t block_len, uint64_t counter,
+                                       uint8_t flags) {
+  uint32_t state[16];
+  compress_pre(state, cv, block, block_len, counter, flags);
+  cv[0] = state[0] ^ state[8];
+  cv[1] = state[1] ^ state[9];
+  cv[2] = state[2] ^ state[10];
+  cv[3] = state[3] ^ state[11];
+  cv[4] = state[4] ^ state[12];
+  cv[5] = state[5] ^ state[13];
+  cv[6] = state[6] ^ state[14];
+  cv[7] = state[7] ^ state[15];
+}
+
+void blake3_compress_xof_portable(const uint32_t cv[8],
+                                  const uint8_t block[BLAKE3_BLOCK_LEN],
+                                  uint8_t block_len, uint64_t counter,
+                                  uint8_t flags, uint8_t out[64]) {
+  uint32_t state[16];
+  compress_pre(state, cv, block, block_len, counter, flags);
+
+  store32(&out[0 * 4], state[0] ^ state[8]);
+  store32(&out[1 * 4], state[1] ^ state[9]);
+  store32(&out[2 * 4], state[2] ^ state[10]);
+  store32(&out[3 * 4], state[3] ^ state[11]);
+  store32(&out[4 * 4], state[4] ^ state[12]);
+  store32(&out[5 * 4], state[5] ^ state[13]);
+  store32(&out[6 * 4], state[6] ^ state[14]);
+  store32(&out[7 * 4], state[7] ^ state[15]);
+  store32(&out[8 * 4], state[8] ^ cv[0]);
+  store32(&out[9 * 4], state[9] ^ cv[1]);
+  store32(&out[10 * 4], state[10] ^ cv[2]);
+  store32(&out[11 * 4], state[11] ^ cv[3]);
+  store32(&out[12 * 4], state[12] ^ cv[4]);
+  store32(&out[13 * 4], state[13] ^ cv[5]);
+  store32(&out[14 * 4], state[14] ^ cv[6]);
+  store32(&out[15 * 4], state[15] ^ cv[7]);
+}
+
+INLINE void hash_one_portable(const uint8_t *input, size_t blocks,
+                              const uint32_t key[8], uint64_t counter,
+                              uint8_t flags, uint8_t flags_start,
+                              uint8_t flags_end, uint8_t out[BLAKE3_OUT_LEN]) {
+  uint32_t cv[8];
+  memcpy(cv, key, BLAKE3_KEY_LEN);
+  uint8_t block_flags = flags | flags_start;
+  while (blocks > 0) {
+    if (blocks == 1) {
+      block_flags |= flags_end;
+    }
+    blake3_compress_in_place_portable(cv, input, BLAKE3_BLOCK_LEN, counter,
+                                      block_flags);
+    input = &input[BLAKE3_BLOCK_LEN];
+    blocks -= 1;
+    block_flags = flags;
+  }
+  store_cv_words(out, cv);
+}
+
+void blake3_hash_many_portable(const uint8_t *const *inputs, size_t num_inputs,
+                               size_t blocks, const uint32_t key[8],
+                               uint64_t counter, bool increment_counter,
+                               uint8_t flags, uint8_t flags_start,
+                               uint8_t flags_end, uint8_t *out) {
+  while (num_inputs > 0) {
+    hash_one_portable(inputs[0], blocks, key, counter, flags, flags_start,
+                      flags_end, out);
+    if (increment_counter) {
+      counter += 1;
+    }
+    inputs += 1;
+    num_inputs -= 1;
+    out = &out[BLAKE3_OUT_LEN];
+  }
+}
diff --git a/ext/hash/blake3/helper_script_blake3_version_fetcher.php b/ext/hash/blake3/helper_script_blake3_version_fetcher.php
new file mode 100644
index 0000000000000..9b149bcc5062d
--- /dev/null
+++ b/ext/hash/blake3/helper_script_blake3_version_fetcher.php
@@ -0,0 +1,61 @@
+<?php
+declare(strict_types = 1);
+
+// this script is intended to help developers update to newer versions of
+// the upstream "official blake3 optimized c implementation"..
+// not something to be used often
+const BLAKE3_TARGET_VERSION = "0.3.7";
+
+/**
+ * fetch url with user-agent
+ * (file_get_contents doesn't use user-agent, and api.github.com block people without UA)
+ *
+ * @param string $url
+ * @throws \LogicException
+ * @return string
+ */
+function fetch(string $url): string
+{
+    $ch = curl_init();
+    curl_setopt_array($ch, array(
+        CURLOPT_URL => $url,
+        CURLOPT_USERAGENT => 'php-src/ext/hash/blake3/helper_script bot',
+        CURLOPT_ENCODING => '',
+        CURLOPT_RETURNTRANSFER => 1
+    ));
+    $ret = curl_exec($ch);
+    if (curl_errno($ch) !== CURLE_OK || curl_getinfo($ch, CURLINFO_HTTP_CODE) !== 200) {
+        throw new \LogicException();
+    }
+    curl_close($ch);
+    return $ret;
+}
+
+$files = json_decode(fetch("https://api.github.com/repos/BLAKE3-team/BLAKE3/contents/c?ref=" . urlencode(BLAKE3_TARGET_VERSION)), true);
+foreach ($files as $file) {
+    // $file looks like:
+    array(
+        'name' => 'blake3_avx512.c',
+        'path' => 'c/blake3_avx512.c',
+        'sha' => '77a5c385cb897f712d55dd884d8ae54f14ed2570',
+        'size' => 47960,
+        'url' => 'https://api.github.com/repos/BLAKE3-team/BLAKE3/contents/c/blake3_avx512.c?ref=0.3.7',
+        'html_url' => 'https://github.com/BLAKE3-team/BLAKE3/blob/0.3.7/c/blake3_avx512.c',
+        'git_url' => 'https://api.github.com/repos/BLAKE3-team/BLAKE3/git/blobs/77a5c385cb897f712d55dd884d8ae54f14ed2570',
+        'download_url' => 'https://raw.githubusercontent.com/BLAKE3-team/BLAKE3/0.3.7/c/blake3_avx512.c',
+        'type' => 'file',
+        '_links' => array(
+            'self' => 'https://api.github.com/repos/BLAKE3-team/BLAKE3/contents/c/blake3_avx512.c?ref=0.3.7',
+            'git' => 'https://api.github.com/repos/BLAKE3-team/BLAKE3/git/blobs/77a5c385cb897f712d55dd884d8ae54f14ed2570',
+            'html' => 'https://github.com/BLAKE3-team/BLAKE3/blob/0.3.7/c/blake3_avx512.c'
+        )
+    );
+    $name = $file["name"];
+    if ($file["type"] !== "file" || ! preg_match('/^blake3.*\\.(?:c|h|S|asm)$/', $name)) {
+        // the dir contains several things we're not interested in.
+        continue;
+    }
+    var_dump($file["name"]);
+    $content = fetch($file['download_url']);
+    file_put_contents($name, $content, LOCK_EX);
+}

From 7b51423f0162823afa6f99f6a85241d7083e8111 Mon Sep 17 00:00:00 2001
From: Hans Henrik Bergan <hans@easy-ad.no>
Date: Wed, 28 Oct 2020 22:24:47 +0100
Subject: [PATCH 2/3] blake3 portable + upstream-support

this is a significantly smaller alternative to the PR at https://github.com/php/php-src/pull/6358 ,
now blake3 "portable c" version is bundled (specifically version 0.3.7 plus a few patches that will be part of the 0.3.8 release..), and ./configure supports a new optional --with-blake3-upstream-c-source-dir=DIR argument for specifying the location of the upstream BLAKE3 C implementation, if invoked, the SSE2/SSE4.1/AVX2/AVX512 optimized versions of BLAKE3 will be compiled in when applicable (this has not been added to MSVC, i don't know how to do it on MSVC, and i don't have a MSVC system to figure it out out on, if someone think getting those optimizations available on MSVC is important, take it up with the windows php mailing list.. just getting the portable version to compile on MSVC was good enough for me.)

also userland scripts can detect at runtime if the portable version or the upstream version, of BLAKE was compiled by consulting phpinfo(), it will either say "blake3 implementation: portable 0.3.7" or "blake3 implementation: upstream X.X.X"
---
 build/php.m4                           | 80 ++++++++++++++++++++++++++
 configure.ac                           |  9 +++
 ext/hash/blake3/blake3.c               |  4 ++
 ext/hash/blake3/blake3.h               |  2 +
 ext/hash/config.m4                     | 52 ++++++++++++++++-
 ext/hash/config.w32                    | 10 +++-
 ext/hash/hash.c                        |  8 +++
 ext/hash/hash_blake3.c                 | 27 +++++++++
 ext/hash/php_hash.h                    |  1 +
 ext/hash/php_hash_blake3.h             | 35 +++++++++++
 ext/hash/tests/hash-clone.phpt         |  6 ++
 ext/hash/tests/hash_algos.phpt         |  4 +-
 ext/hash/tests/hash_copy_001.phpt      |  6 ++
 ext/hash/tests/hash_hmac_algos.phpt    |  1 +
 ext/hash/tests/hash_serialize_001.phpt |  6 ++
 ext/hash/tests/hash_serialize_002.phpt |  2 +
 16 files changed, 247 insertions(+), 6 deletions(-)
 create mode 100644 ext/hash/hash_blake3.c
 create mode 100644 ext/hash/php_hash_blake3.h

diff --git a/build/php.m4 b/build/php.m4
index 16c5e25fbaab0..42616f6f213ec 100644
--- a/build/php.m4
+++ b/build/php.m4
@@ -2703,3 +2703,83 @@ AC_DEFUN([PHP_PATCH_CONFIG_HEADERS], [
   $SED -e 's/^#undef PACKAGE_[^ ]*/\/\* & \*\//g' < $srcdir/$1 \
     > $srcdir/$1.tmp && mv $srcdir/$1.tmp $srcdir/$1
 ])
+
+
+
+
+
+dnl
+dnl PHP_CHECK_X86_TARGET
+dnl
+dnl check if we're compiling for x86/x86_64
+dnl
+AC_DEFUN([PHP_CHECK_X86_TARGET], [
+  AC_CACHE_CHECK([for x86 target],ac_cv_target_x86,[
+  AC_RUN_IFELSE([AC_LANG_SOURCE([[
+int main(void) {
+#if defined(__x86_64__) || defined(__i386__)
+    return 0;
+#else
+    return 1;
+#endif
+}
+]])],[
+  ac_cv_target_x86=yes
+],[
+  ac_cv_target_x86=no
+],[
+  ac_cv_target_x86=no
+])])
+
+])
+
+
+dnl
+dnl PHP_CHECK_WINDOWS_TARGET
+dnl
+dnl check if we're compiling for windows
+dnl
+AC_DEFUN([PHP_CHECK_WINDOWS_TARGET], [
+  AC_CACHE_CHECK([for windows target],ac_cv_target_windows,[
+  AC_RUN_IFELSE([AC_LANG_SOURCE([[
+int main(void) {
+#if defined(_WIN32)
+    return 0;
+#else
+    return 1;
+#endif
+}
+]])],[
+  ac_cv_target_windows=yes
+],[
+  ac_cv_target_windows=no
+],[
+  ac_cv_target_windows=no
+])])
+
+])
+
+dnl
+dnl PHP_CHECK_UNIX_TARGET
+dnl
+dnl check if we're compiling for a unix-ish target
+dnl
+AC_DEFUN([PHP_CHECK_UNIX_TARGET], [
+  AC_CACHE_CHECK([for unix-ish target],ac_cv_target_unix,[
+  AC_RUN_IFELSE([AC_LANG_SOURCE([[
+int main(void) {
+#if defined(unix) || defined(__unix) || defined(__unix__)
+    return 0;
+#else
+    return 1;
+#endif
+}
+]])],[
+  ac_cv_target_unix=yes
+],[
+  ac_cv_target_unix=no
+],[
+  ac_cv_target_unix=no
+])])
+
+])
diff --git a/configure.ac b/configure.ac
index 6b33047538fe2..9d51d88e64cbd 100644
--- a/configure.ac
+++ b/configure.ac
@@ -324,6 +324,15 @@ PHP_EBCDIC
 dnl Check whether the system byte ordering is bigendian.
 PHP_C_BIGENDIAN
 
+dnl Check if we're targeting x86 / x86_64
+PHP_CHECK_X86_TARGET
+
+dnl Check if we're targeting Windows
+PHP_CHECK_WINDOWS_TARGET
+
+dnl Check whether we're targeting a unix-ish system
+PHP_CHECK_UNIX_TARGET
+
 dnl Check whether writing to stdout works.
 PHP_TEST_WRITE_STDOUT
 
diff --git a/ext/hash/blake3/blake3.c b/ext/hash/blake3/blake3.c
index 741a76d08bc93..7abf5324ecd05 100644
--- a/ext/hash/blake3/blake3.c
+++ b/ext/hash/blake3/blake3.c
@@ -5,6 +5,10 @@
 #include "blake3.h"
 #include "blake3_impl.h"
 
+const char * blake3_version(void) {
+  return BLAKE3_VERSION_STRING;
+}
+
 INLINE void chunk_state_init(blake3_chunk_state *self, const uint32_t key[8],
                              uint8_t flags) {
   memcpy(self->cv, key, BLAKE3_KEY_LEN);
diff --git a/ext/hash/blake3/blake3.h b/ext/hash/blake3/blake3.h
index 51f1d2a8672bb..57ebd5adc2c35 100644
--- a/ext/hash/blake3/blake3.h
+++ b/ext/hash/blake3/blake3.h
@@ -8,6 +8,7 @@
 extern "C" {
 #endif
 
+#define BLAKE3_VERSION_STRING "0.3.7"
 #define BLAKE3_KEY_LEN 32
 #define BLAKE3_OUT_LEN 32
 #define BLAKE3_BLOCK_LEN 64
@@ -38,6 +39,7 @@ typedef struct {
   uint8_t cv_stack[(BLAKE3_MAX_DEPTH + 1) * BLAKE3_OUT_LEN];
 } blake3_hasher;
 
+const char * blake3_version(void);
 void blake3_hasher_init(blake3_hasher *self);
 void blake3_hasher_init_keyed(blake3_hasher *self,
                               const uint8_t key[BLAKE3_KEY_LEN]);
diff --git a/ext/hash/config.m4 b/ext/hash/config.m4
index d84e89cc6b333..43cab675d56ba 100644
--- a/ext/hash/config.m4
+++ b/ext/hash/config.m4
@@ -11,7 +11,28 @@ if test "$PHP_MHASH" != "no"; then
   AC_DEFINE(PHP_MHASH_BC, 1, [ ])
 fi
 
+
+
+PHP_ARG_WITH([blake3-upstream-c-source-dir],
+  [for upstream blake3 implementation, github.com/blake3-team/BLAKE3/tree/0.3.7/c ],
+  [AS_HELP_STRING([[--with-blake3-upstream-c-source-dir=DIR]],
+    [Include upstream blake3 source code, github.com/blake3-team/BLAKE3/tree/0.3.7/c])],
+    "", "")
+
+if test "$PHP_BLAKE3_UPSTREAM_C_SOURCE_DIR" = ""; then
+  EXT_HASH_BLAKE3_SOURCES="hash_blake3.c blake3/blake3.c blake3/blake3_dispatch.c blake3/blake3_portable.c"
+  EXT_HASH_BLAKE3_CFLAGS="-I@ext_srcdir@/blake3"
+  PHP_ADD_BUILD_DIR(ext/hash/blake3, 1)
+else
+  AC_DEFINE(PHP_BLAKE3_UPSTREAM_C_SOURCE_DIR, $PHP_BLAKE3_UPSTREAM_C_SOURCE_DIR, [ ])
+  EXT_HASH_BLAKE3_SOURCES="hash_blake3.c $PHP_BLAKE3_UPSTREAM_C_SOURCE_DIR/blake3.c $PHP_BLAKE3_UPSTREAM_C_SOURCE_DIR/blake3_dispatch.c $PHP_BLAKE3_UPSTREAM_C_SOURCE_DIR/blake3_portable.c"
+  PHP_ADD_BUILD_DIR($PHP_BLAKE3_UPSTREAM_C_SOURCE_DIR, 1)
+  EXT_HASH_BLAKE3_CFLAGS="-I$PHP_BLAKE3_UPSTREAM_C_SOURCE_DIR"
+fi
+
+
 if test $ac_cv_c_bigendian_php = yes; then
+  dnl todo: check if $PHP_BLAKE3_UPSTREAM_C_SOURCE_DIR/blake3_neon.c is applicable ?
   EXT_HASH_SHA3_SOURCES="hash_sha3.c"
   AC_DEFINE(HAVE_SLOW_HASH3, 1, [Define is hash3 algo is available])
   AC_MSG_WARN("Use SHA3 slow implementation on bigendian")
@@ -20,26 +41,53 @@ else
   AC_MSG_CHECKING([if we're at 64-bit platform])
   AS_IF([test "$ac_cv_sizeof_long" -eq 4],[
     AC_MSG_RESULT([no])
+    if test $ac_cv_target_x86 = yes; then 
+      dnl i think there are some 32bit x86 cpus with SSE2 but.. cba
+      EXT_HASH_BLAKE3_CFLAGS="$EXT_HASH_BLAKE3_CFLAGS -DBLAKE3_NO_SSE2 -DBLAKE3_NO_SSE41 -DBLAKE3_NO_AVX2 -DBLAKE3_NO_AVX512"
+    fi
     SHA3_DIR="sha3/generic32lc"
     SHA3_OPT_SRC="$SHA3_DIR/KeccakP-1600-inplace32BI.c"
   ],[
     AC_MSG_RESULT([yes])
     SHA3_DIR="sha3/generic64lc"
     SHA3_OPT_SRC="$SHA3_DIR/KeccakP-1600-opt64.c"
+    if test $ac_cv_target_x86 = yes; then 
+      if test "$PHP_BLAKE3_UPSTREAM_C_SOURCE_DIR" = ""; then
+        dnl AC_MSG_WARN("Use BLAKE3 slow/portable implementation")
+        EXT_HASH_BLAKE3_CFLAGS="$EXT_HASH_BLAKE3_CFLAGS -DBLAKE3_NO_SSE2 -DBLAKE3_NO_SSE41 -DBLAKE3_NO_AVX2 -DBLAKE3_NO_AVX512"
+      else
+        if test $ac_cv_target_windows = yes; then
+          dnl x86_64 windows gnuc
+          EXT_HASH_BLAKE3_SOURCES="$EXT_HASH_BLAKE3_SOURCES $PHP_BLAKE3_UPSTREAM_C_SOURCE_DIR/blake3_avx512_x86-64_windows_gnu.S $PHP_BLAKE3_UPSTREAM_C_SOURCE_DIR/blake3_avx2_x86-64_windows_gnu.S $PHP_BLAKE3_UPSTREAM_C_SOURCE_DIR/blake3_sse41_x86-64_windows_gnu.S $PHP_BLAKE3_UPSTREAM_C_SOURCE_DIR/blake3_sse2_x86-64_windows_gnu.S"
+        else
+          if test $ac_cv_target_unix = yes; then 
+            dnl x86_64 unix gnuc
+            EXT_HASH_BLAKE3_SOURCES="$EXT_HASH_BLAKE3_SOURCES $PHP_BLAKE3_UPSTREAM_C_SOURCE_DIR/blake3_avx512_x86-64_unix.S $PHP_BLAKE3_UPSTREAM_C_SOURCE_DIR/blake3_avx2_x86-64_unix.S $PHP_BLAKE3_UPSTREAM_C_SOURCE_DIR/blake3_sse41_x86-64_unix.S $PHP_BLAKE3_UPSTREAM_C_SOURCE_DIR/blake3_sse2_x86-64_unix.S"
+          else
+            dnl blake still has C instrictics versions (eg $PHP_BLAKE3_UPSTREAM_C_SOURCE_DIR/blake3_avx512.c  ) which COULD be used in this situation, 
+            dnl we're targeting x86_64 but neither windows nor unix-ish.. 
+            dnl i have nothing to test this on, so i'm just disabling it for now, feel free to fix it.
+            EXT_HASH_BLAKE3_CFLAGS="$EXT_HASH_BLAKE3_CFLAGS -DBLAKE3_NO_SSE2 -DBLAKE3_NO_SSE41 -DBLAKE3_NO_AVX2 -DBLAKE3_NO_AVX512"
+          fi
+        fi
+      fi
+    fi
   ])
   EXT_HASH_SHA3_SOURCES="$SHA3_OPT_SRC $SHA3_DIR/KeccakHash.c $SHA3_DIR/KeccakSponge.c hash_sha3.c"
   PHP_HASH_CFLAGS="-I@ext_srcdir@/$SHA3_DIR -DKeccakP200_excluded -DKeccakP400_excluded -DKeccakP800_excluded -DZEND_ENABLE_STATIC_TSRMLS_CACHE=1"
 
   PHP_ADD_BUILD_DIR(ext/hash/$SHA3_DIR, 1)
 fi
+PHP_HASH_CFLAGS="$PHP_HASH_CFLAGS $EXT_HASH_BLAKE3_CFLAGS"
 
 EXT_HASH_SOURCES="hash.c hash_md.c hash_sha.c hash_ripemd.c hash_haval.c \
   hash_tiger.c hash_gost.c hash_snefru.c hash_whirlpool.c hash_adler32.c \
-  hash_crc32.c hash_fnv.c hash_joaat.c $EXT_HASH_SHA3_SOURCES"
+  hash_crc32.c hash_fnv.c hash_joaat.c $EXT_HASH_SHA3_SOURCES \
+  $EXT_HASH_BLAKE3_SOURCES"
 EXT_HASH_HEADERS="php_hash.h php_hash_md.h php_hash_sha.h php_hash_ripemd.h \
   php_hash_haval.h php_hash_tiger.h php_hash_gost.h php_hash_snefru.h \
   php_hash_whirlpool.h php_hash_adler32.h php_hash_crc32.h \
-  php_hash_fnv.h php_hash_joaat.h php_hash_sha3.h"
+  php_hash_fnv.h php_hash_joaat.h php_hash_sha3.h php_hash_blake3.h"
 
 PHP_NEW_EXTENSION(hash, $EXT_HASH_SOURCES, 0,,$PHP_HASH_CFLAGS)
 PHP_INSTALL_HEADERS(ext/hash, $EXT_HASH_HEADERS)
diff --git a/ext/hash/config.w32 b/ext/hash/config.w32
index 5d6be6b92249a..4503f86654fa0 100644
--- a/ext/hash/config.w32
+++ b/ext/hash/config.w32
@@ -11,7 +11,9 @@ PHP_HASH = 'yes';
 EXTENSION('hash',	'hash.c hash_md.c hash_sha.c hash_ripemd.c hash_haval.c ' +
 					'hash_tiger.c hash_gost.c hash_snefru.c hash_whirlpool.c ' +
 					'hash_adler32.c hash_crc32.c hash_joaat.c hash_fnv.c ' +
-					'hash_sha3.c', false);
+					'hash_sha3.c hash_blake3.c', false);
+
+ADD_SOURCES('ext/hash/blake3', 'blake3.c blake3_dispatch.c blake3_portable.c');
 
 var hash_sha3_dir = 'ext/hash/sha3/generic' + (X64 ? '64' : '32') + 'lc';
 
@@ -26,9 +28,11 @@ if (!CHECK_HEADER_ADD_INCLUDE('KeccakHash.h', 'CFLAGS_HASH', hash_sha3_dir)) {
 	ERROR('Unable to locate SHA3 headers');
 }
 
-ADD_FLAG('CFLAGS_HASH', '/DKeccakP200_excluded /DKeccakP400_excluded /DKeccakP800_excluded /DZEND_ENABLE_STATIC_TSRMLS_CACHE=1');
+ADD_FLAG('CFLAGS_HASH', '/DKeccakP200_excluded /DKeccakP400_excluded /DKeccakP800_excluded /DZEND_ENABLE_STATIC_TSRMLS_CACHE=1 ' +
+                        '/DBLAKE3_NO_SSE2 /DBLAKE3_NO_SSE41 /DBLAKE3_NO_AVX2 /DBLAKE3_NO_AVX512');
 
 PHP_INSTALL_HEADERS('ext/hash/',	'php_hash.h php_hash_md.h php_hash_sha.h ' +
 									'php_hash_ripemd.h php_hash_haval.h php_hash_tiger.h ' +
 									'php_hash_gost.h php_hash_snefru.h php_hash_whirlpool.h ' +
-									'php_hash_adler32.h php_hash_crc32.h php_hash_sha3.h');
+									'php_hash_adler32.h php_hash_crc32.h php_hash_sha3.h ' +
+									'php_hash_blake3.h');
diff --git a/ext/hash/hash.c b/ext/hash/hash.c
index 4e3820f35e6b4..f920f8a3784b8 100644
--- a/ext/hash/hash.c
+++ b/ext/hash/hash.c
@@ -19,6 +19,8 @@
 #include "config.h"
 #endif
 
+#include "blake3.h"
+
 #include <math.h>
 #include "php_hash.h"
 #include "ext/standard/info.h"
@@ -1586,6 +1588,7 @@ PHP_MINIT_FUNCTION(hash)
 	php_hash_register_algo("fnv164",		&php_hash_fnv164_ops);
 	php_hash_register_algo("fnv1a64",		&php_hash_fnv1a64_ops);
 	php_hash_register_algo("joaat",			&php_hash_joaat_ops);
+	php_hash_register_algo("blake3",		&php_hash_blake3_ops);
 
 	PHP_HASH_HAVAL_REGISTER(3,128);
 	PHP_HASH_HAVAL_REGISTER(3,160);
@@ -1650,6 +1653,11 @@ PHP_MINFO_FUNCTION(hash)
 	php_info_print_table_start();
 	php_info_print_table_row(2, "hash support", "enabled");
 	php_info_print_table_row(2, "Hashing Engines", buffer);
+#ifdef PHP_BLAKE3_UPSTREAM_C_SOURCE_DIR
+	php_info_print_table_row(2, "blake3 implementation", "upstream " BLAKE3_VERSION_STRING);
+#else
+	php_info_print_table_row(2, "blake3 implementation", "portable " BLAKE3_VERSION_STRING);
+#endif
 	php_info_print_table_end();
 
 #ifdef PHP_MHASH_BC
diff --git a/ext/hash/hash_blake3.c b/ext/hash/hash_blake3.c
new file mode 100644
index 0000000000000..360ddaac47834
--- /dev/null
+++ b/ext/hash/hash_blake3.c
@@ -0,0 +1,27 @@
+#include "php_hash.h"
+#include "php_hash_blake3.h"
+#include "blake3.h"
+
+
+#include <string.h> // memcpy
+
+PHP_HASH_API void PHP_BLAKE3Init(PHP_BLAKE3_CTX *context)
+{
+	blake3_hasher_init(context);
+}
+
+PHP_HASH_API void PHP_BLAKE3Update(PHP_BLAKE3_CTX *context, const unsigned char *input, size_t len)
+{
+	blake3_hasher_update(context, input, len);
+}
+
+PHP_HASH_API void PHP_BLAKE3Final(unsigned char digest[BLAKE3_OUT_LEN/*32*/], PHP_BLAKE3_CTX *context)
+{
+	blake3_hasher_finalize(context, digest, BLAKE3_OUT_LEN);
+}
+
+PHP_HASH_API int PHP_BLAKE3Copy(const php_hash_ops *ops, PHP_BLAKE3_CTX *orig_context, PHP_BLAKE3_CTX *copy_context)
+{
+	memcpy(copy_context, orig_context, sizeof(*orig_context));
+	return SUCCESS;
+}
diff --git a/ext/hash/php_hash.h b/ext/hash/php_hash.h
index 33cc6a473a7a1..70807b7f1ba6c 100644
--- a/ext/hash/php_hash.h
+++ b/ext/hash/php_hash.h
@@ -105,6 +105,7 @@ extern const php_hash_ops php_hash_fnv1a32_ops;
 extern const php_hash_ops php_hash_fnv164_ops;
 extern const php_hash_ops php_hash_fnv1a64_ops;
 extern const php_hash_ops php_hash_joaat_ops;
+extern const php_hash_ops php_hash_blake3_ops;
 
 #define PHP_HASH_HAVAL_OPS(p,b)	extern const php_hash_ops php_hash_##p##haval##b##_ops;
 
diff --git a/ext/hash/php_hash_blake3.h b/ext/hash/php_hash_blake3.h
new file mode 100644
index 0000000000000..f977b6818eb4c
--- /dev/null
+++ b/ext/hash/php_hash_blake3.h
@@ -0,0 +1,35 @@
+#ifndef PHP_HASH_BLAKE3_H
+#define PHP_HASH_BLAKE3_H
+
+#include "ext/standard/basic_functions.h"
+#include "php_hash.h"
+#include "blake3.h"
+
+
+// typedef struct blake3_hasher PHP_BLAKE3_CTX;
+#define PHP_BLAKE3_CTX blake3_hasher
+// help: is V correct?
+#define PHP_BLAKE3_SPEC "b8b8qb64bbbbb1760"
+
+PHP_HASH_API void PHP_BLAKE3Init(PHP_BLAKE3_CTX *context);
+PHP_HASH_API void PHP_BLAKE3Update(PHP_BLAKE3_CTX *context, const unsigned char *input, size_t len);
+PHP_HASH_API void PHP_BLAKE3Final(unsigned char digest[BLAKE3_OUT_LEN/*32*/], PHP_BLAKE3_CTX *context);
+PHP_HASH_API int PHP_BLAKE3Copy(const php_hash_ops *ops, PHP_BLAKE3_CTX *orig_context, PHP_BLAKE3_CTX *copy_context);
+
+const php_hash_ops php_hash_blake3_ops = {
+	"blake3",
+	(php_hash_init_func_t) PHP_BLAKE3Init,
+	(php_hash_update_func_t) PHP_BLAKE3Update,
+	(php_hash_final_func_t) PHP_BLAKE3Final,
+	(php_hash_copy_func_t) PHP_BLAKE3Copy,
+	php_hash_serialize,
+	php_hash_unserialize,
+	PHP_BLAKE3_SPEC, // << don't know what this should be, hopefully a dev that knows can remove this comment
+	BLAKE3_OUT_LEN /*32*/,
+	BLAKE3_CHUNK_LEN /*1024*/,
+	sizeof(PHP_BLAKE3_CTX),
+	1
+};
+
+#endif
+
diff --git a/ext/hash/tests/hash-clone.phpt b/ext/hash/tests/hash-clone.phpt
index 57567c0bc5014..271e310418346 100644
--- a/ext/hash/tests/hash-clone.phpt
+++ b/ext/hash/tests/hash-clone.phpt
@@ -143,6 +143,9 @@ string(16) "bebc746a33b6ab62"
 string(5) "joaat"
 string(8) "aaebf370"
 string(8) "aaebf370"
+string(6) "blake3"
+string(64) "e232493e3d416e24ffc588b24a1772ccc6f80290f1920cf15f21313bc3543a51"
+string(64) "e232493e3d416e24ffc588b24a1772ccc6f80290f1920cf15f21313bc3543a51"
 string(10) "haval128,3"
 string(32) "86362472c8895e68e223ef8b3711d8d9"
 string(32) "86362472c8895e68e223ef8b3711d8d9"
@@ -302,6 +305,9 @@ string(16) "893899e4415a920f"
 string(5) "joaat"
 string(8) "aaebf370"
 string(8) "836fb0e5"
+string(6) "blake3"
+string(64) "e232493e3d416e24ffc588b24a1772ccc6f80290f1920cf15f21313bc3543a51"
+string(64) "dbdea45e5a6c3bad18a4f96d9d4b9105e4cceaa4fc06568f69829435c47587fb"
 string(10) "haval128,3"
 string(32) "86362472c8895e68e223ef8b3711d8d9"
 string(32) "ebeeeb05c18af1e53d2d127b561d5e0d"
diff --git a/ext/hash/tests/hash_algos.phpt b/ext/hash/tests/hash_algos.phpt
index cff71a7771235..daf9a4fc9b4a7 100644
--- a/ext/hash/tests/hash_algos.phpt
+++ b/ext/hash/tests/hash_algos.phpt
@@ -9,7 +9,7 @@ var_dump(hash_algos());
 ?>
 --EXPECTF--
 *** Testing hash_algos() : basic functionality ***
-array(53) {
+array(54) {
   [%d]=>
   string(3) "md2"
   [%d]=>
@@ -87,6 +87,8 @@ array(53) {
   [%d]=>
   string(5) "joaat"
   [%d]=>
+  string(6) "blake3"
+  [%d]=>
   string(10) "haval128,3"
   [%d]=>
   string(10) "haval160,3"
diff --git a/ext/hash/tests/hash_copy_001.phpt b/ext/hash/tests/hash_copy_001.phpt
index 271326178d523..edc66f04f92a0 100644
--- a/ext/hash/tests/hash_copy_001.phpt
+++ b/ext/hash/tests/hash_copy_001.phpt
@@ -143,6 +143,9 @@ string(16) "bebc746a33b6ab62"
 string(5) "joaat"
 string(8) "aaebf370"
 string(8) "aaebf370"
+string(6) "blake3"
+string(64) "e232493e3d416e24ffc588b24a1772ccc6f80290f1920cf15f21313bc3543a51"
+string(64) "e232493e3d416e24ffc588b24a1772ccc6f80290f1920cf15f21313bc3543a51"
 string(10) "haval128,3"
 string(32) "86362472c8895e68e223ef8b3711d8d9"
 string(32) "86362472c8895e68e223ef8b3711d8d9"
@@ -302,6 +305,9 @@ string(16) "893899e4415a920f"
 string(5) "joaat"
 string(8) "aaebf370"
 string(8) "836fb0e5"
+string(6) "blake3"
+string(64) "e232493e3d416e24ffc588b24a1772ccc6f80290f1920cf15f21313bc3543a51"
+string(64) "dbdea45e5a6c3bad18a4f96d9d4b9105e4cceaa4fc06568f69829435c47587fb"
 string(10) "haval128,3"
 string(32) "86362472c8895e68e223ef8b3711d8d9"
 string(32) "ebeeeb05c18af1e53d2d127b561d5e0d"
diff --git a/ext/hash/tests/hash_hmac_algos.phpt b/ext/hash/tests/hash_hmac_algos.phpt
index 16aec26df6d9d..9435bae60e655 100644
--- a/ext/hash/tests/hash_hmac_algos.phpt
+++ b/ext/hash/tests/hash_hmac_algos.phpt
@@ -37,6 +37,7 @@ Array
     [%d] => snefru256
     [%d] => gost
     [%d] => gost-crypto
+    [%d] => blake3
     [%d] => haval128,3
     [%d] => haval160,3
     [%d] => haval192,3
diff --git a/ext/hash/tests/hash_serialize_001.phpt b/ext/hash/tests/hash_serialize_001.phpt
index d515d2c5c55cd..4ef4c518b702a 100644
--- a/ext/hash/tests/hash_serialize_001.phpt
+++ b/ext/hash/tests/hash_serialize_001.phpt
@@ -154,6 +154,9 @@ string(16) "bebc746a33b6ab62"
 string(5) "joaat"
 string(8) "aaebf370"
 string(8) "aaebf370"
+string(6) "blake3"
+string(64) "e232493e3d416e24ffc588b24a1772ccc6f80290f1920cf15f21313bc3543a51"
+string(64) "e232493e3d416e24ffc588b24a1772ccc6f80290f1920cf15f21313bc3543a51"
 string(10) "haval128,3"
 string(32) "86362472c8895e68e223ef8b3711d8d9"
 string(32) "86362472c8895e68e223ef8b3711d8d9"
@@ -313,6 +316,9 @@ string(16) "893899e4415a920f"
 string(5) "joaat"
 string(8) "836fb0e5"
 string(8) "836fb0e5"
+string(6) "blake3"
+string(64) "dbdea45e5a6c3bad18a4f96d9d4b9105e4cceaa4fc06568f69829435c47587fb"
+string(64) "dbdea45e5a6c3bad18a4f96d9d4b9105e4cceaa4fc06568f69829435c47587fb"
 string(10) "haval128,3"
 string(32) "ebeeeb05c18af1e53d2d127b561d5e0d"
 string(32) "ebeeeb05c18af1e53d2d127b561d5e0d"
diff --git a/ext/hash/tests/hash_serialize_002.phpt b/ext/hash/tests/hash_serialize_002.phpt
index b6b4ccfb5eab7..20e3533ae1f23 100644
--- a/ext/hash/tests/hash_serialize_002.phpt
+++ b/ext/hash/tests/hash_serialize_002.phpt
@@ -83,6 +83,8 @@ string(4) "gost"
 HashContext with HASH_HMAC option cannot be serialized
 string(11) "gost-crypto"
 HashContext with HASH_HMAC option cannot be serialized
+string(6) "blake3"
+HashContext with HASH_HMAC option cannot be serialized
 string(10) "haval128,3"
 HashContext with HASH_HMAC option cannot be serialized
 string(10) "haval160,3"

From 3998738bf1d1f085e43eb2bea13c9e40de099906 Mon Sep 17 00:00:00 2001
From: Hans Henrik Bergan <hans@easy-ad.no>
Date: Fri, 30 Oct 2020 15:20:42 +0100
Subject: [PATCH 3/3] attempt msvc build fix

---
 ext/hash/config.w32 | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/ext/hash/config.w32 b/ext/hash/config.w32
index 4503f86654fa0..bf7fb71832acf 100644
--- a/ext/hash/config.w32
+++ b/ext/hash/config.w32
@@ -27,6 +27,12 @@ if (!CHECK_HEADER_ADD_INCLUDE('KeccakHash.h', 'CFLAGS_HASH', hash_sha3_dir)) {
 	// Should NEVER happen
 	ERROR('Unable to locate SHA3 headers');
 }
+var hash_blake3_dir = 'ext/hash/blake3';
+if (!CHECK_HEADER_ADD_INCLUDE('blake3.h', 'CFLAGS_HASH', hash_blake3_dir)) {
+	// Should NEVER happen
+	ERROR('Unable to locate BLAKE3 headers');
+}
+
 
 ADD_FLAG('CFLAGS_HASH', '/DKeccakP200_excluded /DKeccakP400_excluded /DKeccakP800_excluded /DZEND_ENABLE_STATIC_TSRMLS_CACHE=1 ' +
                         '/DBLAKE3_NO_SSE2 /DBLAKE3_NO_SSE41 /DBLAKE3_NO_AVX2 /DBLAKE3_NO_AVX512');