SSSE3: enc: encode src/dst size in a typedef

aklomp · Jan 16, 2024 · b955289 · b955289
1 parent 9c5c07b
commit b955289
Showing 1 changed file with 27 additions and 21 deletions.
diff --git a/lib/arch/ssse3/enc_loop.c b/lib/arch/ssse3/enc_loop.c
@@ -1,8 +1,13 @@
+// Define array types for the source and destination vectors. This enables the
+// use of post-increments to move a pointer to the next element.
+typedef uint8_t ssse3_enc_src[12];
+typedef uint8_t ssse3_enc_dst[16];
+
 static inline void
-enc_loop_ssse3_inner (const uint8_t **s, uint8_t **o)
+enc_loop_ssse3_inner (const ssse3_enc_src **s, ssse3_enc_dst **o)
 {
 	// Load input:
-	__m128i str = _mm_loadu_si128((__m128i *) *s);
+	__m128i str = _mm_loadu_si128((__m128i *) (*s)++);
 
 	// Reshuffle:
 	str = enc_reshuffle(str);
@@ -11,10 +16,7 @@ enc_loop_ssse3_inner (const uint8_t **s, uint8_t **o)
 	str = enc_translate(str);
 
 	// Store:
-	_mm_storeu_si128((__m128i *) *o, str);
-
-	*s += 12;
-	*o += 16;
+	_mm_storeu_si128((__m128i *) (*o)++, str);
 }
 
 static inline void
@@ -33,34 +35,38 @@ enc_loop_ssse3 (const uint8_t **s, size_t *slen, uint8_t **o, size_t *olen)
 	*slen -= rounds * 12;	// 12 bytes consumed per round
 	*olen += rounds * 16;	// 16 bytes produced per round
 
+	// Cast the source and destination pointers to array type pointers.
+	const ssse3_enc_src **src = (const ssse3_enc_src **) s;
+	ssse3_enc_dst       **dst = (ssse3_enc_dst **) o;
+
 	do {
 		if (rounds >= 8) {
-			enc_loop_ssse3_inner(s, o);
-			enc_loop_ssse3_inner(s, o);
-			enc_loop_ssse3_inner(s, o);
-			enc_loop_ssse3_inner(s, o);
-			enc_loop_ssse3_inner(s, o);
-			enc_loop_ssse3_inner(s, o);
-			enc_loop_ssse3_inner(s, o);
-			enc_loop_ssse3_inner(s, o);
+			enc_loop_ssse3_inner(src, dst);
+			enc_loop_ssse3_inner(src, dst);
+			enc_loop_ssse3_inner(src, dst);
+			enc_loop_ssse3_inner(src, dst);
+			enc_loop_ssse3_inner(src, dst);
+			enc_loop_ssse3_inner(src, dst);
+			enc_loop_ssse3_inner(src, dst);
+			enc_loop_ssse3_inner(src, dst);
 			rounds -= 8;
 			continue;
 		}
 		if (rounds >= 4) {
-			enc_loop_ssse3_inner(s, o);
-			enc_loop_ssse3_inner(s, o);
-			enc_loop_ssse3_inner(s, o);
-			enc_loop_ssse3_inner(s, o);
+			enc_loop_ssse3_inner(src, dst);
+			enc_loop_ssse3_inner(src, dst);
+			enc_loop_ssse3_inner(src, dst);
+			enc_loop_ssse3_inner(src, dst);
 			rounds -= 4;
 			continue;
 		}
 		if (rounds >= 2) {
-			enc_loop_ssse3_inner(s, o);
-			enc_loop_ssse3_inner(s, o);
+			enc_loop_ssse3_inner(src, dst);
+			enc_loop_ssse3_inner(src, dst);
 			rounds -= 2;
 			continue;
 		}
-		enc_loop_ssse3_inner(s, o);
+		enc_loop_ssse3_inner(src, dst);
 		break;
 
 	} while (rounds > 0);