X-Git-Url: https://git.sesse.net/?p=fjl;a=blobdiff_plain;f=bitsource.h;h=f2f164db43821652acaf281c76f2b050f4041c7a;hp=6cabbf004245139d6f3f5a8f5f5fac5cfdb646e6;hb=a4009687c73083dd0290285a065740a83e27e855;hpb=352c52d38c212a2aa93a9970731c9d7a2f8d4dac

diff --git a/bitsource.h b/bitsource.h
index 6cabbf0..f2f164d 100644
--- a/bitsource.h
+++ b/bitsource.h
@@ -2,6 +2,7 @@
 #define _BITSOURCE_H 1
 
 #include <assert.h>
+#include <stdbool.h>
 #include <stddef.h>
 #include <stdint.h>
 #include <sys/types.h>
@@ -14,9 +15,16 @@
 typedef uint64_t bitreservoir_t;
 typedef uint32_t bitreservoir_fill_t;
 
-static inline bitreservoir_fill_t read_bitreservoir_fill(uint8_t* source)
+// Note: We return bitreservoir_t here, so we can get implicit zero extension on amd64.
+static inline bitreservoir_t read_bitreservoir_fill(uint8_t* source)
 {
+#if defined(__GNUC__) && (defined(__i386__) || defined(__x86_64__))
+	bitreservoir_t ret;
+	asm("bswapl %1" : "=r" (ret) : "0" (*(bitreservoir_fill_t*)(source)));
+	return ret;
+#else
 	return ntohl(*(bitreservoir_fill_t*)(source));
+#endif
 }
 
 static const unsigned BITRESERVOIR_SIZE = 8 * sizeof(bitreservoir_t);
@@ -34,14 +42,22 @@ struct bit_source {
 	// When this is empty, it needs to be refilled from the input
 	// stream.
 	uint8_t* bytes;
+	uint8_t* byte_read_ptr;
 	unsigned bytes_available;
 
+	// Some clients will purposedly read a bit ahead of the stream, causing
+	// problems at EOF. Thus, the client is allowed to request that we pad
+	// the end stream with a few bytes after the source reports EOF.
+	int padding_bytes_available;
+
 	// Data source.
 	input_func_t* input_func;
 	void* userdata;
+	bool source_eof;
 };
 
-void init_bit_source(struct bit_source* source, input_func_t* input_func, void* userdata);
+void init_bit_source(struct bit_source* source, input_func_t* input_func,
+                     unsigned padding_bytes, void* userdata);
 
 // Internal function. Do not use.
 void possibly_refill_slow_path(struct bit_source* source, unsigned num_bits);
@@ -62,8 +78,8 @@ static inline void possibly_refill(struct bit_source* source, unsigned num_bits)
 	// Slower path (~99% of remaining invocations?)
 	assert(source->bits_available + BITRESERVOIR_FILL_SIZE < BITRESERVOIR_SIZE);
 	if (source->bytes_available >= sizeof(bitreservoir_fill_t)) {
-		bitreservoir_fill_t fill = read_bitreservoir_fill(source->bytes);
-		source->bytes += sizeof(bitreservoir_fill_t);
+		bitreservoir_t fill = read_bitreservoir_fill(source->byte_read_ptr);
+		source->byte_read_ptr += sizeof(bitreservoir_fill_t);
 		source->bytes_available -= sizeof(bitreservoir_fill_t);
 		source->bits |= (bitreservoir_t)fill << (BITRESERVOIR_SIZE - BITRESERVOIR_FILL_SIZE - source->bits_available);
 		source->bits_available += BITRESERVOIR_FILL_SIZE;