X-Git-Url: https://wimlib.net/git/?a=blobdiff_plain;f=src%2Flzx_decompress.c;h=5747d327270717ae6892c257374088e4dfec789d;hb=38bd45bb7e08f2072e256afd5bcc21ceb0d97b8e;hp=9f93fcf781441cc815e155dda41b6a339910c879;hpb=908381d2809a48acd9490ec080e51087ae1529fd;p=wimlib

diff --git a/src/lzx_decompress.c b/src/lzx_decompress.c
index 9f93fcf7..5747d327 100644
--- a/src/lzx_decompress.c
+++ b/src/lzx_decompress.c
@@ -93,6 +93,21 @@ struct lzx_decompressor {
 		u8 extra_offset_bits[LZX_MAX_OFFSET_SLOTS];
 	};
 
+	union {
+		DECODE_TABLE_WORKING_SPACE(maincode_working_space,
+					   LZX_MAINCODE_MAX_NUM_SYMBOLS,
+					   LZX_MAX_MAIN_CODEWORD_LEN);
+		DECODE_TABLE_WORKING_SPACE(lencode_working_space,
+					   LZX_LENCODE_NUM_SYMBOLS,
+					   LZX_MAX_LEN_CODEWORD_LEN);
+		DECODE_TABLE_WORKING_SPACE(alignedcode_working_space,
+					   LZX_ALIGNEDCODE_NUM_SYMBOLS,
+					   LZX_MAX_ALIGNED_CODEWORD_LEN);
+		DECODE_TABLE_WORKING_SPACE(precode_working_space,
+					   LZX_PRECODE_NUM_SYMBOLS,
+					   LZX_MAX_PRE_CODEWORD_LEN);
+	};
+
 	unsigned window_order;
 	unsigned num_main_syms;
 
@@ -103,7 +118,7 @@ struct lzx_decompressor {
 } _aligned_attribute(DECODE_TABLE_ALIGNMENT);
 
 /* Read a Huffman-encoded symbol using the precode. */
-static inline unsigned
+static forceinline unsigned
 read_presym(const struct lzx_decompressor *d, struct input_bitstream *is)
 {
 	return read_huffsym(is, d->precode_decode_table,
@@ -111,7 +126,7 @@ read_presym(const struct lzx_decompressor *d, struct input_bitstream *is)
 }
 
 /* Read a Huffman-encoded symbol using the main code. */
-static inline unsigned
+static forceinline unsigned
 read_mainsym(const struct lzx_decompressor *d, struct input_bitstream *is)
 {
 	return read_huffsym(is, d->maincode_decode_table,
@@ -119,7 +134,7 @@ read_mainsym(const struct lzx_decompressor *d, struct input_bitstream *is)
 }
 
 /* Read a Huffman-encoded symbol using the length code. */
-static inline unsigned
+static forceinline unsigned
 read_lensym(const struct lzx_decompressor *d, struct input_bitstream *is)
 {
 	return read_huffsym(is, d->lencode_decode_table,
@@ -127,7 +142,7 @@ read_lensym(const struct lzx_decompressor *d, struct input_bitstream *is)
 }
 
 /* Read a Huffman-encoded symbol using the aligned offset code. */
-static inline unsigned
+static forceinline unsigned
 read_alignedsym(const struct lzx_decompressor *d, struct input_bitstream *is)
 {
 	return read_huffsym(is, d->alignedcode_decode_table,
@@ -157,7 +172,8 @@ lzx_read_codeword_lens(struct lzx_decompressor *d, struct input_bitstream *is,
 				      LZX_PRECODE_NUM_SYMBOLS,
 				      LZX_PRECODE_TABLEBITS,
 				      d->precode_lens,
-				      LZX_MAX_PRE_CODEWORD_LEN))
+				      LZX_MAX_PRE_CODEWORD_LEN,
+				      d->precode_working_space))
 		return -1;
 
 	/* Decode the codeword lengths.  */
@@ -316,10 +332,19 @@ lzx_read_block_header(struct lzx_decompressor *d, struct input_bitstream *is,
 
 /* Decompress a block of LZX-compressed data. */
 static int
-lzx_decompress_block(struct lzx_decompressor *d, struct input_bitstream *is,
+lzx_decompress_block(struct lzx_decompressor *d, struct input_bitstream *_is,
 		     int block_type, u32 block_size,
 		     u8 * const out_begin, u8 *out_next, u32 recent_offsets[])
 {
+	/*
+	 * Redeclare the input bitstream on the stack.  This shouldn't be
+	 * needed, but it can improve the main loop's performance significantly
+	 * with both gcc and clang, apparently because the compiler otherwise
+	 * gets confused and doesn't properly allocate registers for
+	 * 'is->bitbuf' et al. and/or thinks 'is->next' may point into 'is'.
+	 */
+	struct input_bitstream is_onstack = *_is;
+	struct input_bitstream *is = &is_onstack;
 	u8 * const block_end = out_next + block_size;
 	unsigned min_aligned_offset_slot;
 
@@ -333,14 +358,16 @@ lzx_decompress_block(struct lzx_decompressor *d, struct input_bitstream *is,
 				      d->num_main_syms,
 				      LZX_MAINCODE_TABLEBITS,
 				      d->maincode_lens,
-				      LZX_MAX_MAIN_CODEWORD_LEN))
+				      LZX_MAX_MAIN_CODEWORD_LEN,
+				      d->maincode_working_space))
 		return -1;
 
 	if (make_huffman_decode_table(d->lencode_decode_table,
 				      LZX_LENCODE_NUM_SYMBOLS,
 				      LZX_LENCODE_TABLEBITS,
 				      d->lencode_lens,
-				      LZX_MAX_LEN_CODEWORD_LEN))
+				      LZX_MAX_LEN_CODEWORD_LEN,
+				      d->lencode_working_space))
 		return -1;
 
 	if (block_type == LZX_BLOCKTYPE_ALIGNED) {
@@ -348,7 +375,8 @@ lzx_decompress_block(struct lzx_decompressor *d, struct input_bitstream *is,
 					      LZX_ALIGNEDCODE_NUM_SYMBOLS,
 					      LZX_ALIGNEDCODE_TABLEBITS,
 					      d->alignedcode_lens,
-					      LZX_MAX_ALIGNED_CODEWORD_LEN))
+					      LZX_MAX_ALIGNED_CODEWORD_LEN,
+					      d->alignedcode_working_space))
 			return -1;
 		min_aligned_offset_slot = LZX_MIN_ALIGNED_OFFSET_SLOT;
 		memcpy(d->extra_offset_bits, d->extra_offset_bits_minus_aligned,
@@ -409,20 +437,14 @@ lzx_decompress_block(struct lzx_decompressor *d, struct input_bitstream *is,
 		}
 		recent_offsets[0] = offset;
 
-		/* Validate the match, then copy it to the current position.  */
-
-		if (unlikely(length > block_end - out_next))
+		/* Validate the match and copy it to the current position.  */
+		if (unlikely(lz_copy(length, offset, out_begin,
+				     out_next, block_end, LZX_MIN_MATCH_LEN)))
 			return -1;
-
-		if (unlikely(offset > out_next - out_begin))
-			return -1;
-
-		lz_copy(out_next, length, offset, block_end, LZX_MIN_MATCH_LEN);
-
 		out_next += length;
-
 	} while (out_next != block_end);
 
+	*_is = is_onstack;
 	return 0;
 }