X-Git-Url: https://git.sesse.net/?p=fjl;a=blobdiff_plain;f=bitsource.h;h=4126a618d34dedbfa90940696f0489602712cceb;hp=ae29b4c4a2200abce210242a541be2c289ea8163;hb=75e1e46dafe538990da0ea3b021c82d5ed6d5e21;hpb=9069ae93f3e9e9b44f5f28968f508d09f0711737

diff --git a/bitsource.h b/bitsource.h
index ae29b4c..4126a61 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 +15,20 @@
 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(__x86_64__)
+	bitreservoir_t ret;
+	asm("bswapl %1" : "=r" (ret) : "0" (*(bitreservoir_fill_t*)(source)));
+	return ret;
+#elif defined(__GNUC__) && defined(__i386__)
+	bitreservoir_fill_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);
@@ -45,6 +57,7 @@ struct bit_source {
 	// 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,
@@ -69,7 +82,7 @@ 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->byte_read_ptr);
+		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);