Micro-optimize word-at-a-time loop tails

Compilers apparently aren't smart enough to infer that the odd tail
length fixup "loops" run at most once, and could be converted to ifs, so
do that manually.

2 files changed, 14 insertions, 10 deletions

diff --git a/src/bfstd.c b/src/bfstd.c
index f2938ad..219b8d0 100644
--- a/src/bfstd.c
+++ b/src/bfstd.c
@@ -786,7 +786,7 @@ size_t asciinlen(const char *str, size_t n) {
 	// Word-at-a-time isascii()
 #define CHUNK(n) CHUNK_(uint##n##_t, load8_leu##n)
 #define CHUNK_(type, load8) \
-	while (n - i >= sizeof(type)) { \
+	(n - i >= sizeof(type)) { \
 		type word = load8(ustr + i); \
 		type mask = (((type)-1) / 0xFF) << 7; /* 0x808080.. */ \
 		word &= mask; \
@@ -797,11 +797,13 @@ size_t asciinlen(const char *str, size_t n) {
 	}
 
 #if SIZE_WIDTH >= 64
-	CHUNK(64);
+	while CHUNK(64);
+	if CHUNK(32);
+#else
+	while CHUNK(32);
 #endif
-	CHUNK(32);
-	CHUNK(16);
-	CHUNK(8);
+	if CHUNK(16);
+	if CHUNK(8);
 
 #undef CHUNK_
 #undef CHUNK
diff --git a/src/trie.c b/src/trie.c
index c4bf4ba..4e0944a 100644
--- a/src/trie.c
+++ b/src/trie.c
@@ -408,7 +408,7 @@ static size_t trie_mismatch(const struct trie_leaf *rep, const void *key, size_t
 
 #define CHUNK(n) CHUNK_(uint##n##_t, load8_beu##n)
 #define CHUNK_(type, load8) \
-	while (length - i >= sizeof(type)) { \
+	(length - i >= sizeof(type)) { \
 		type rep_chunk = load8(rep_bytes + i); \
 		type key_chunk = load8(key_bytes + i); \
 		type diff = rep_chunk ^ key_chunk; \
@@ -420,11 +420,13 @@ static size_t trie_mismatch(const struct trie_leaf *rep, const void *key, size_t
 	}
 
 #if SIZE_WIDTH >= 64
-	CHUNK(64);
+	while CHUNK(64);
+	if CHUNK(32);
+#else
+	while CHUNK(32);
 #endif
-	CHUNK(32);
-	CHUNK(16);
-	CHUNK(8);
+	if CHUNK(16);
+	if CHUNK(8);
 
 #undef CHUNK_
 #undef CHUNK