borrow cache prefetching tricks from the Marvell port, seem to improve performance by 5-28%

src/fdr/fdr.c

@@ -147,6 +147,7 @@ void get_conf_stride_1(const u8 *itPtr, UNUSED const u8 *start_ptr,
                        const u64a *ft, u64a *conf0, u64a *conf8, m128 *s) {
     /* +1: the zones ensure that we can read the byte at z->end */
     assert(itPtr >= start_ptr && itPtr + ITER_BYTES <= end_ptr);
+
     u64a reach0 = andn(domain_mask_flipped, itPtr);
     u64a reach1 = andn(domain_mask_flipped, itPtr + 1);
     u64a reach2 = andn(domain_mask_flipped, itPtr + 2);
@@ -184,9 +185,8 @@ void get_conf_stride_1(const u8 *itPtr, UNUSED const u8 *start_ptr,
     st0 = or128(st0, st4);
     *s = or128(*s, st0);
 
-    *conf0 = movq(*s);
+    *conf0 = movq(*s) ^ ~0ULL;
     *s = rshiftbyte_m128(*s, 8);
-    *conf0 ^= ~0ULL;
 
     u64a reach8 = andn(domain_mask_flipped, itPtr + 8);
     u64a reach9 = andn(domain_mask_flipped, itPtr + 9);
@@ -225,9 +225,8 @@ void get_conf_stride_1(const u8 *itPtr, UNUSED const u8 *start_ptr,
     st8 = or128(st8, st12);
     *s = or128(*s, st8);
 
-    *conf8 = movq(*s);
+    *conf8 = movq(*s) ^ ~0ULL;
     *s = rshiftbyte_m128(*s, 8);
-    *conf8 ^= ~0ULL;
 }
 
 static really_inline
@@ -235,6 +234,7 @@ void get_conf_stride_2(const u8 *itPtr, UNUSED const u8 *start_ptr,
                        UNUSED const u8 *end_ptr, u32 domain_mask_flipped,
                        const u64a *ft, u64a *conf0, u64a *conf8, m128 *s) {
     assert(itPtr >= start_ptr && itPtr + ITER_BYTES <= end_ptr);
+
     u64a reach0 = andn(domain_mask_flipped, itPtr);
     u64a reach2 = andn(domain_mask_flipped, itPtr + 2);
     u64a reach4 = andn(domain_mask_flipped, itPtr + 4);
@@ -287,6 +287,7 @@ void get_conf_stride_4(const u8 *itPtr, UNUSED const u8 *start_ptr,
                        UNUSED const u8 *end_ptr, u32 domain_mask_flipped,
                        const u64a *ft, u64a *conf0, u64a *conf8, m128 *s) {
     assert(itPtr >= start_ptr && itPtr + ITER_BYTES <= end_ptr);
+
     u64a reach0 = andn(domain_mask_flipped, itPtr);
     u64a reach4 = andn(domain_mask_flipped, itPtr + 4);
     u64a reach8 = andn(domain_mask_flipped, itPtr + 8);
@@ -683,6 +684,10 @@ size_t prepareZones(const u8 *buf, size_t len, const u8 *hend,
         const u8 *tryFloodDetect = zz->floodPtr;                            \
         const u8 *start_ptr = zz->start;                                    \
         const u8 *end_ptr = zz->end;                                        \
+        for (const u8 *itPtr = start_ptr; itPtr + 4*ITER_BYTES <= end_ptr;  \
+            itPtr += 4*ITER_BYTES) {                                        \
+            __builtin_prefetch(itPtr);                                      \
+        }                                                                   \
                                                                             \
         for (const u8 *itPtr = start_ptr; itPtr + ITER_BYTES <= end_ptr;    \
             itPtr += ITER_BYTES) {                                          \

src/nfa/mcclellan.c

@@ -634,6 +634,11 @@ char nfaExecMcClellan16_Q2i(const struct NFA *n, u64a offset, const u8 *buffer,
     assert(ISALIGNED_N(q->state, 2));
     u32 s = *(u16 *)q->state;
 
+    __builtin_prefetch(&m->remap[0]);
+    __builtin_prefetch(&m->remap[64]);
+    __builtin_prefetch(&m->remap[128]);
+    __builtin_prefetch(&m->remap[192]);
+
     if (q->report_current) {
         assert(s);
         assert(get_aux(m, s)->accept);
@@ -790,6 +795,11 @@ char nfaExecMcClellan8_Q2i(const struct NFA *n, u64a offset, const u8 *buffer,
 
     u32 s = *(u8 *)q->state;
 
+    __builtin_prefetch(&m->remap[0]);
+    __builtin_prefetch(&m->remap[64]);
+    __builtin_prefetch(&m->remap[128]);
+    __builtin_prefetch(&m->remap[192]);
+
     if (q->report_current) {
         assert(s);
         assert(s >= m->accept_limit_8);

src/nfa/mcsheng.c

@@ -889,6 +889,11 @@ char nfaExecMcSheng16_Q2i(const struct NFA *n, u64a offset, const u8 *buffer,
         return MO_ALIVE;
     }
 
+    __builtin_prefetch(&m->remap[0]);
+    __builtin_prefetch(&m->remap[64]);
+    __builtin_prefetch(&m->remap[128]);
+    __builtin_prefetch(&m->remap[192]);
+
     while (1) {
         assert(q->cur < q->end);
         s64a ep = q->items[q->cur].location;
@@ -1017,6 +1022,11 @@ char nfaExecMcSheng8_Q2i(const struct NFA *n, u64a offset, const u8 *buffer,
         return MO_ALIVE;
     }
 
+    __builtin_prefetch(&m->remap[0]);
+    __builtin_prefetch(&m->remap[64]);
+    __builtin_prefetch(&m->remap[128]);
+    __builtin_prefetch(&m->remap[192]);
+
     while (1) {
         DEBUG_PRINTF("%s @ %llu\n", q->items[q->cur].type == MQE_TOP ? "TOP" :
                      q->items[q->cur].type == MQE_END ? "END" : "???",

src/nfa/shufti.c

@@ -109,7 +109,8 @@ DUMP_MSK(128)
 #endif
 
 #define GET_LO_4(chars) and128(chars, low4bits)
-#define GET_HI_4(chars) rshift64_m128(andnot128(low4bits, chars), 4)
+#define GET_HI_4(chars) and128(rshift64_m128(chars, 4), low4bits)
+//#define GET_HI_4(chars) rshift64_m128(andnot128(low4bits, chars), 4)
 
 static really_inline
 u32 block(m128 mask_lo, m128 mask_hi, m128 chars, const m128 low4bits,
@@ -177,6 +178,10 @@ const u8 *shuftiExec(m128 mask_lo, m128 mask_hi, const u8 *buf,
     // Reroll FTW.
 
     const u8 *last_block = buf_end - 16;
+
+    for (const u8 *itPtr = buf; itPtr + 4*16 <= last_block; itPtr += 4*16) {
+        __builtin_prefetch(itPtr);
+    }
     while (buf < last_block) {
         m128 lchars = load128(buf);
         rv = fwdBlock(mask_lo, mask_hi, lchars, buf, low4bits, zeroes);