make xof_many fall back to compress_xof instead of portable code

2 files changed, 6 insertions, 1 deletions

diff --git a/c/blake3_dispatch.c b/c/blake3_dispatch.c
index cf5bad7..5c76b14 100644
--- a/c/blake3_dispatch.c
+++ b/c/blake3_dispatch.c
@@ -241,7 +241,9 @@ void blake3_xof_many(const uint32_t cv[8],
   }
 #endif
 #endif
-  blake3_xof_many_portable(cv, block, block_len, counter, flags, out, outblocks);
+  for(size_t i = 0; i < outblocks; ++i) {
+    blake3_compress_xof(cv, block, block_len, counter + i, flags, out + 64*i);
+  }
 }
 
 void blake3_hash_many(const uint8_t *const *inputs, size_t num_inputs,
diff --git a/c/blake3_impl.h b/c/blake3_impl.h
index abd7546..3da773b 100644
--- a/c/blake3_impl.h
+++ b/c/blake3_impl.h
@@ -222,6 +222,9 @@ void blake3_compress_xof_portable(const uint32_t cv[8],
                                   uint8_t block_len, uint64_t counter,
                                   uint8_t flags, uint8_t out[64]);
 
+// This function is test-only. When blake3_xof_many doesn't have an optimized implementation,
+// it loops over blake3_compress_xof instead of falling back to this, so it still benefits
+// from compress optimizations.
 void blake3_xof_many_portable(const uint32_t cv[8],
                               const uint8_t block[BLAKE3_BLOCK_LEN],
                               uint8_t block_len, uint64_t counter, uint8_t flags,