// Copyright 2017 The Wuffs Authors. // // Licensed under the Apache License, Version 2.0 or the MIT license // , at your // option. This file may not be copied, modified, or distributed // except according to those terms. // // SPDX-License-Identifier: Apache-2.0 OR MIT // ---------------- // This file contains a hand-written C implementation of // release/c/wuffs-unsupported-snapshot.c's generated // wuffs_deflate__decoder__decode_huffman_fast function. // // It is not intended to be used in production settings, on untrusted data. Its // purpose is to give a rough upper bound on how fast Wuffs' generated C code // can be, with a sufficiently smart Wuffs compiler. // // To repeat, substituting in this C implementation is NOT SAFE, and may result // in buffer overflows. This code exists only to aid optimization of the (safe) // std/deflate/*.wuffs code and the Wuffs compiler's code generation. // ---------------- // Having said that, to generate the benchmark numbers with this hand-written C // implementation, edit release/c/wuffs-unsupported-snapshot.c and find the // lines that say // // // ‼ WUFFS C HEADER ENDS HERE. // #ifdef WUFFS_IMPLEMENTATION // // After that, add this line: // // #include "../../script/wuffs-deflate-decoder-decode-huffman.c" // // Then find the calls to wuffs_deflate__decoder__decode_huffman_fastxx. They // should be inside the wuffs_deflate__decoder__decode_blocks function body, // and the lines of code should look something like // v_status = wuffs_deflate__decoder__decode_huffman_fast32(etc); // and // v_status = wuffs_deflate__decoder__decode_huffman_fast64(etc); // // Change the "wuffs_etc_fastxx" to "c_wuffs_etc_fast", i.e. add a "c_" prefix // and drop the "32" or "64" suffix. The net result should look something like: // // clang-format off /* $ git diff release/c/wuffs-unsupported-snapshot.c diff --git a/release/c/wuffs-unsupported-snapshot.c b/release/c/wuffs-unsupported-snapshot.c index d255a788..2a112728 100644 --- a/release/c/wuffs-unsupported-snapshot.c +++ b/release/c/wuffs-unsupported-snapshot.c @@ -9758,6 +9758,8 @@ DecodeJson(DecodeJsonCallbacks& callbacks, // ‼ WUFFS C HEADER ENDS HERE. #ifdef WUFFS_IMPLEMENTATION +#include "../../script/wuffs-deflate-decoder-decode-huffman.c" + #ifdef __cplusplus extern "C" { #endif @@ -25326,7 +25328,7 @@ wuffs_deflate__decoder__decode_blocks( if (a_src) { a_src->meta.ri = ((size_t)(iop_a_src - a_src->data.ptr)); } - v_status = wuffs_deflate__decoder__decode_huffman_fast32(self, a_dst, a_src); + v_status = c_wuffs_deflate__decoder__decode_huffman_fast(self, a_dst, a_src); if (a_src) { iop_a_src = a_src->data.ptr + a_src->meta.ri; } @@ -25334,7 +25336,7 @@ wuffs_deflate__decoder__decode_blocks( if (a_src) { a_src->meta.ri = ((size_t)(iop_a_src - a_src->data.ptr)); } - v_status = wuffs_deflate__decoder__decode_huffman_fast64(self, a_dst, a_src); + v_status = c_wuffs_deflate__decoder__decode_huffman_fast(self, a_dst, a_src); if (a_src) { iop_a_src = a_src->data.ptr + a_src->meta.ri; } */ // clang-format on // // That concludes the edits to release/c/wuffs-unsupported-snapshot.c. Run the // tests and benchmarks with the "-skipgen" flag, otherwise the "wuffs" tool // will re-generate the C code and override your edits: // // wuffs test -skipgen std/deflate // wuffs bench -skipgen std/deflate // // You may also want to focus on one specific test, e.g.: // // wuffs test -skipgen -focus=wuffs_deflate_decode_midsummer std/deflate // wuffs bench -skipgen -focus=wuffs_deflate_decode_100k_just std/deflate #include #include #include #include // For manual printf debugging. #include extern const char wuffs_deflate__error__internal_error_inconsistent_distance[]; extern const char wuffs_deflate__error__internal_error_inconsistent_huffman_decoder_state[]; // ---------------- // Define this macro to make this file compilable standalone, instead of // hooking it into release/c/wuffs-unsupported-snapshot.c per the instructions // above. Standalone means that you can run a "does it compile" check by // // gcc -c script/wuffs-deflate-decoder-decode-huffman.c -o /dev/null // // and also drop this file into disassembly tools like https://godbolt.org/ //#define MAKE_THIS_FILE_COMPILABLE_STANDALONE #ifdef MAKE_THIS_FILE_COMPILABLE_STANDALONE typedef struct { const char* repr; } wuffs_base__status; static inline wuffs_base__status // wuffs_base__make_status(const char* repr) { wuffs_base__status z; z.repr = repr; return z; } typedef struct { uint8_t* ptr; size_t len; } wuffs_base__slice_u8; typedef struct { size_t wi; size_t ri; uint64_t pos; bool closed; } wuffs_base__io_buffer_meta; typedef struct { wuffs_base__slice_u8 data; wuffs_base__io_buffer_meta meta; } wuffs_base__io_buffer; typedef struct wuffs_deflate__decoder__struct { struct { uint32_t magic; uint32_t f_bits; uint32_t f_n_bits; uint32_t f_history_index; uint32_t f_n_huffs_bits[2]; bool f_end_of_block; } private_impl; struct { // 1024 is huffs_table_size in std/deflate/decode_deflate.wuffs. uint32_t f_huffs[2][1024]; uint8_t f_history[32768]; uint8_t f_code_lengths[320]; } private_data; } wuffs_deflate__decoder; const char wuffs_base__error__bad_argument[] = "?base: bad argument"; const char wuffs_deflate__error__bad_distance[] = "?deflate: bad distance"; const char wuffs_deflate__error__internal_error_inconsistent_distance[] = "?deflate: internal error: inconsistent distance"; const char wuffs_deflate__error__internal_error_inconsistent_huffman_decoder_state[] = "?deflate: internal error: inconsistent Huffman decoder state"; static wuffs_base__status // wuffs_deflate__decoder__decode_huffman_fast(wuffs_deflate__decoder* self, wuffs_base__io_buffer* a_dst, wuffs_base__io_buffer* a_src) { return wuffs_base__make_status(NULL); } #endif // MAKE_THIS_FILE_COMPILABLE_STANDALONE // ---------------- // Define this macro to further speed up this C implementation, using two // techniques that are not used by the zlib-the-library C implementation (as of // version 1.2.11, current as of November 2017). Doing so gives data for "how // fast can Wuffs' zlib-the-format implementation be" instead of "is Wuffs' // generated C code as fast as zlib-the-library's hand-written C code". // // Whether the same techniques could apply to zlib-the-library is discussed at // https://github.com/madler/zlib/pull/292 #ifdef __x86_64__ //#define WUFFS_DEFLATE__HAVE_64_BIT_UNALIGNED_LITTLE_ENDIAN_LOADS #endif // wuffs_base__width_to_mask_table is the width_to_mask_table from // https://fgiesen.wordpress.com/2018/02/19/reading-bits-in-far-too-many-ways-part-1/ // // Look for "It may feel ridiculous" on that page for the rationale. static const uint32_t wuffs_base__width_to_mask_table[33] = { 0x00000000u, 0x00000001u, 0x00000003u, 0x00000007u, 0x0000000Fu, 0x0000001Fu, 0x0000003Fu, 0x0000007Fu, 0x000000FFu, 0x000001FFu, 0x000003FFu, 0x000007FFu, 0x00000FFFu, 0x00001FFFu, 0x00003FFFu, 0x00007FFFu, 0x0000FFFFu, 0x0001FFFFu, 0x0003FFFFu, 0x0007FFFFu, 0x000FFFFFu, 0x001FFFFFu, 0x003FFFFFu, 0x007FFFFFu, 0x00FFFFFFu, 0x01FFFFFFu, 0x03FFFFFFu, 0x07FFFFFFu, 0x0FFFFFFFu, 0x1FFFFFFFu, 0x3FFFFFFFu, 0x7FFFFFFFu, 0xFFFFFFFFu, }; // Uncomment exactly one of these. Choosing between them lets us measure the // effect of the table-based mask. //#define MASK(n) ((((uint32_t)1) << (n)) - 1) #define MASK(n) wuffs_base__width_to_mask_table[n] // ---------------- // These are the generated functions that we are explicitly overriding. Note // that the function name is "wuffs_etc", not "c_wuffs_etc". static wuffs_base__status // wuffs_deflate__decoder__decode_huffman_fast32(wuffs_deflate__decoder* self, wuffs_base__io_buffer* a_dst, wuffs_base__io_buffer* a_src); static wuffs_base__status // wuffs_deflate__decoder__decode_huffman_fast64(wuffs_deflate__decoder* self, wuffs_base__io_buffer* a_dst, wuffs_base__io_buffer* a_src); // This is the overriding implementation. wuffs_base__status // c_wuffs_deflate__decoder__decode_huffman_fast(wuffs_deflate__decoder* self, wuffs_base__io_buffer* a_dst, wuffs_base__io_buffer* a_src) { // Avoid the -Werror=unused-function warning for the now-unused // overridden wuffs_deflate__decoder__decode_huffman_fastxx functions. (void)(wuffs_deflate__decoder__decode_huffman_fast32); (void)(wuffs_deflate__decoder__decode_huffman_fast64); if (!a_dst || !a_src) { return wuffs_base__make_status(wuffs_base__error__bad_argument); } wuffs_base__status status = wuffs_base__make_status(NULL); // Load contextual state. Prepare to check that pdst and psrc remain within // a_dst's and a_src's bounds. uint8_t* pdst = a_dst->data.ptr + a_dst->meta.wi; uint8_t* qdst = a_dst->data.ptr + a_dst->data.len; if ((qdst - pdst) < 258) { return wuffs_base__make_status(NULL); } else { qdst -= 258; } uint8_t* psrc = a_src->data.ptr + a_src->meta.ri; uint8_t* qsrc = a_src->data.ptr + a_src->meta.wi; if ((qsrc - psrc) < 12) { return wuffs_base__make_status(NULL); } else { qsrc -= 12; } #if defined(WUFFS_DEFLATE__HAVE_64_BIT_UNALIGNED_LITTLE_ENDIAN_LOADS) uint64_t bits = self->private_impl.f_bits; #else uint32_t bits = self->private_impl.f_bits; #endif uint32_t n_bits = self->private_impl.f_n_bits; // Initialize other local variables. uint8_t* pdst_mark = a_dst->data.ptr; uint32_t lmask = MASK(self->private_impl.f_n_huffs_bits[0]); uint32_t dmask = MASK(self->private_impl.f_n_huffs_bits[1]); outer_loop: while ((pdst <= qdst) && (psrc <= qsrc)) { #if defined(WUFFS_DEFLATE__HAVE_64_BIT_UNALIGNED_LITTLE_ENDIAN_LOADS) // Ensure that we have at least 56 bits of input. // // This is "Variant 4" of // https://fgiesen.wordpress.com/2018/02/20/reading-bits-in-far-too-many-ways-part-2/ // // 56, the number of bits in 7 bytes, is a property of that "Variant 4" // bit-reading technique, and not related to the DEFLATE format per se. // // Specifically for DEFLATE, we need only up to 48 bits per outer_loop // iteration. The maximum input bits used by a length/distance pair is 15 // bits for the length code, 5 bits for the length extra, 15 bits for the // distance code, and 13 bits for the distance extra. This totals 48 bits. // // The fact that the 48 we need is less than the 56 we get is a happy // coincidence. It lets us eliminate any other loads in the loop body. uint64_t src_u64; memcpy(&src_u64, psrc, 8); bits |= src_u64 << n_bits; psrc += (63 - n_bits) >> 3; n_bits |= 56; #else // Ensure that we have at least 15 bits of input. if (n_bits < 15) { bits |= ((uint32_t)(*psrc++)) << n_bits; n_bits += 8; bits |= ((uint32_t)(*psrc++)) << n_bits; n_bits += 8; } #endif // Decode an lcode symbol from H-L. uint32_t table_entry = self->private_data.f_huffs[0][bits & lmask]; while (true) { uint32_t n = table_entry & 0x0F; bits >>= n; n_bits -= n; if ((table_entry >> 31) != 0) { // Literal. *pdst++ = (uint8_t)(table_entry >> 8); goto outer_loop; } if ((table_entry >> 30) != 0) { // Back-reference; length = base_number + extra_bits. break; } if ((table_entry >> 29) != 0) { // End of block. self->private_impl.f_end_of_block = true; goto end; } if ((table_entry >> 24) != 0x10) { status = wuffs_base__make_status( wuffs_deflate__error__internal_error_inconsistent_huffman_decoder_state); goto end; } uint32_t top = (table_entry >> 8) & 0xFFFF; uint32_t mask = MASK((table_entry >> 4) & 0x0F); table_entry = self->private_data.f_huffs[0][top + (bits & mask)]; } // length = base_number_minus_3 + 3 + extra_bits. uint32_t length = ((table_entry >> 8) & 0xFF) + 3; { uint32_t n = (table_entry >> 4) & 0x0F; if (n) { #if !defined(WUFFS_DEFLATE__HAVE_64_BIT_UNALIGNED_LITTLE_ENDIAN_LOADS) if (n_bits < n) { bits |= ((uint32_t)(*psrc++)) << n_bits; n_bits += 8; } #endif length += bits & MASK(n); bits >>= n; n_bits -= n; } } #if !defined(WUFFS_DEFLATE__HAVE_64_BIT_UNALIGNED_LITTLE_ENDIAN_LOADS) // Ensure that we have at least 15 bits of input. if (n_bits < 15) { bits |= ((uint32_t)(*psrc++)) << n_bits; n_bits += 8; bits |= ((uint32_t)(*psrc++)) << n_bits; n_bits += 8; } #endif // Decode a dcode symbol from H-D. table_entry = self->private_data.f_huffs[1][bits & dmask]; while (true) { uint32_t n = table_entry & 0x0F; bits >>= n; n_bits -= n; if ((table_entry >> 30) != 0) { // Back-reference; distance = base_number + extra_bits. break; } if ((table_entry >> 24) != 0x10) { status = wuffs_base__make_status( wuffs_deflate__error__internal_error_inconsistent_huffman_decoder_state); goto end; } uint32_t top = (table_entry >> 8) & 0xFFFF; uint32_t mask = MASK((table_entry >> 4) & 0x0F); table_entry = self->private_data.f_huffs[1][top + (bits & mask)]; } // dist_minus_1 = base_number_minus_1 + extra_bits. // distance = dist_minus_1 + 1. uint32_t dist_minus_1 = (table_entry >> 8) & 0xFFFF; { uint32_t n = (table_entry >> 4) & 0x0F; #if !defined(WUFFS_DEFLATE__HAVE_64_BIT_UNALIGNED_LITTLE_ENDIAN_LOADS) // Ensure that we have at least n bits of input. if (n_bits < n) { bits |= ((uint32_t)(*psrc++)) << n_bits; n_bits += 8; bits |= ((uint32_t)(*psrc++)) << n_bits; n_bits += 8; } #endif dist_minus_1 += bits & MASK(n); bits >>= n; n_bits -= n; } // Copy from the history buffer, if necessary. if ((ptrdiff_t)(dist_minus_1 + 1) > (pdst - pdst_mark)) { // Set (hlen, hdist) to be the length-distance pair to copy from // this.history, and (length, distance) to be the remaining // length-distance pair to copy from args.dst. uint32_t hlen = 0; uint32_t hdist = ((dist_minus_1 + 1) - (pdst - pdst_mark)); if (length > hdist) { length -= hdist; hlen = hdist; } else { hlen = length; length = 0; } if (self->private_impl.f_history_index < hdist) { status = wuffs_base__make_status(wuffs_deflate__error__bad_distance); goto end; } // Copy up to the right edge of the f_history array. uint32_t history_index = (self->private_impl.f_history_index - hdist) & 0x7FFF; uint32_t available = 0x8000 - history_index; uint32_t n_copied = (hlen < available) ? hlen : available; memmove(pdst, self->private_data.f_history + history_index, n_copied); pdst += n_copied; // Copy from the left edge of the f_history array. if (hlen > n_copied) { hlen -= n_copied; memmove(pdst, self->private_data.f_history, hlen); pdst += hlen; } if (length == 0) { goto outer_loop; } if ((ptrdiff_t)(dist_minus_1 + 1) > (pdst - pdst_mark)) { status = wuffs_base__make_status( wuffs_deflate__error__internal_error_inconsistent_distance); goto end; } } uint8_t* pback = pdst - (dist_minus_1 + 1); #if defined(WUFFS_DEFLATE__HAVE_64_BIT_UNALIGNED_LITTLE_ENDIAN_LOADS) // Back-copy fast path, copying 8 instead of 1 bytes at a time. // // This always copies 8*N bytes (where N is the smallest integer such that // 8*N >= length, i.e. we round length up to a multiple of 8), instead of // only length bytes, but that's OK, as subsequent iterations will fix up // the overrun. if (dist_minus_1 + 1 >= 8) { while (1) { memcpy(pdst, pback, 8); if (length <= 8) { pdst += length; break; } pdst += 8; pback += 8; length -= 8; } continue; } #endif // Back-copy slow path. for (; length >= 3; length -= 3) { *pdst++ = *pback++; *pdst++ = *pback++; *pdst++ = *pback++; } for (; length; length--) { *pdst++ = *pback++; } } end: // Return unused input bytes. for (; n_bits >= 8; n_bits -= 8) { psrc--; } bits &= MASK(n_bits); // Save contextual state. a_dst->meta.wi = pdst - a_dst->data.ptr; a_src->meta.ri = psrc - a_src->data.ptr; self->private_impl.f_bits = bits; self->private_impl.f_n_bits = n_bits; return status; }