Enable sheng32 and sheng64 on Arm

Signed-off-by: Yoan Picchi <yoan.picchi@arm.com>

src/nfa/sheng.c

@@ -154,7 +154,7 @@ char fireReports(const struct sheng *sh, NfaCallback cb, void *ctxt,
     return MO_CONTINUE_MATCHING; /* continue execution */
 }
 
-#if defined(HAVE_AVX512VBMI)
+#if defined(HAVE_AVX512VBMI) || defined(HAVE_SVE)
 // Sheng32
 static really_inline
 const struct sheng32 *get_sheng32(const struct NFA *n) {
@@ -351,7 +351,7 @@ char fireReports64(const struct sheng64 *sh, NfaCallback cb, void *ctxt,
     }
     return MO_CONTINUE_MATCHING; /* continue execution */
 }
-#endif // end of HAVE_AVX512VBMI
+#endif // end of HAVE_AVX512VBMI || HAVE_SVE
 
 /* include Sheng function definitions */
 #include "sheng_defs.h"
@@ -871,7 +871,7 @@ char nfaExecSheng_expandState(UNUSED const struct NFA *nfa, void *dest,
     return 0;
 }
 
-#if defined(HAVE_AVX512VBMI)
+#if defined(HAVE_AVX512VBMI) || defined(HAVE_SVE)
 // Sheng32
 static really_inline
 char runSheng32Cb(const struct sheng32 *sh, NfaCallback cb, void *ctxt,
@@ -1874,4 +1874,4 @@ char nfaExecSheng64_expandState(UNUSED const struct NFA *nfa, void *dest,
     *(u8 *)dest = *(const u8 *)src;
     return 0;
 }
-#endif // end of HAVE_AVX512VBMI
+#endif // end of HAVE_AVX512VBMI || HAVE_SVE

src/nfa/sheng.h

@@ -58,7 +58,7 @@ char nfaExecSheng_reportCurrent(const struct NFA *n, struct mq *q);
 char nfaExecSheng_B(const struct NFA *n, u64a offset, const u8 *buffer,
                     size_t length, NfaCallback cb, void *context);
 
-#if defined(HAVE_AVX512VBMI)
+#if defined(HAVE_AVX512VBMI) || defined(HAVE_SVE)
 #define nfaExecSheng32_B_Reverse NFA_API_NO_IMPL
 #define nfaExecSheng32_zombie_status NFA_API_ZOMBIE_NO_IMPL
 
@@ -106,8 +106,7 @@ char nfaExecSheng64_reportCurrent(const struct NFA *n, struct mq *q);
 
 char nfaExecSheng64_B(const struct NFA *n, u64a offset, const u8 *buffer,
                       size_t length, NfaCallback cb, void *context);
-
+#else // !HAVE_AVX512VBMI && !HAVE_SVE
-#else // !HAVE_AVX512VBMI
 
 #define nfaExecSheng32_B_Reverse NFA_API_NO_IMPL
 #define nfaExecSheng32_zombie_status NFA_API_ZOMBIE_NO_IMPL
@@ -138,6 +137,7 @@ char nfaExecSheng64_B(const struct NFA *n, u64a offset, const u8 *buffer,
 #define nfaExecSheng64_testEOD NFA_API_NO_IMPL
 #define nfaExecSheng64_reportCurrent NFA_API_NO_IMPL
 #define nfaExecSheng64_B NFA_API_NO_IMPL
-#endif // end of HAVE_AVX512VBMI
+#endif // end of HAVE_AVX512VBMI || defined(HAVE_SVE)
+
 
 #endif /* SHENG_H_ */

src/nfa/sheng_defs.h

@@ -52,7 +52,7 @@ u8 hasInterestingStates(const u8 a, const u8 b, const u8 c, const u8 d) {
     return (a | b | c | d) & (SHENG_STATE_FLAG_MASK);
 }
 
-#if defined(HAVE_AVX512VBMI)
+#if defined(HAVE_AVX512VBMI) || defined(HAVE_SVE)
 static really_inline
 u8 isDeadState32(const u8 a) {
     return a & SHENG32_STATE_DEAD;
@@ -108,7 +108,7 @@ u8 dummyFunc(UNUSED const u8 a) {
 #define SHENG_IMPL sheng_cod
 #define DEAD_FUNC isDeadState
 #define ACCEPT_FUNC isAcceptState
-#if defined(HAVE_AVX512VBMI)
+#if defined(HAVE_AVX512VBMI) || defined(HAVE_SVE)
 #define SHENG32_IMPL sheng32_cod
 #define DEAD_FUNC32 isDeadState32
 #define ACCEPT_FUNC32 isAcceptState32
@@ -121,7 +121,7 @@ u8 dummyFunc(UNUSED const u8 a) {
 #undef SHENG_IMPL
 #undef DEAD_FUNC
 #undef ACCEPT_FUNC
-#if defined(HAVE_AVX512VBMI)
+#if defined(HAVE_AVX512VBMI) || defined(HAVE_SVE)
 #undef SHENG32_IMPL
 #undef DEAD_FUNC32
 #undef ACCEPT_FUNC32
@@ -135,7 +135,7 @@ u8 dummyFunc(UNUSED const u8 a) {
 #define SHENG_IMPL sheng_co
 #define DEAD_FUNC dummyFunc
 #define ACCEPT_FUNC isAcceptState
-#if defined(HAVE_AVX512VBMI)
+#if defined(HAVE_AVX512VBMI) || defined(HAVE_SVE)
 #define SHENG32_IMPL sheng32_co
 #define DEAD_FUNC32 dummyFunc
 #define ACCEPT_FUNC32 isAcceptState32
@@ -148,7 +148,7 @@ u8 dummyFunc(UNUSED const u8 a) {
 #undef SHENG_IMPL
 #undef DEAD_FUNC
 #undef ACCEPT_FUNC
-#if defined(HAVE_AVX512VBMI)
+#if defined(HAVE_AVX512VBMI) || defined(HAVE_SVE)
 #undef SHENG32_IMPL
 #undef DEAD_FUNC32
 #undef ACCEPT_FUNC32
@@ -162,7 +162,7 @@ u8 dummyFunc(UNUSED const u8 a) {
 #define SHENG_IMPL sheng_samd
 #define DEAD_FUNC isDeadState
 #define ACCEPT_FUNC isAcceptState
-#if defined(HAVE_AVX512VBMI)
+#if defined(HAVE_AVX512VBMI) || defined(HAVE_SVE)
 #define SHENG32_IMPL sheng32_samd
 #define DEAD_FUNC32 isDeadState32
 #define ACCEPT_FUNC32 isAcceptState32
@@ -175,7 +175,7 @@ u8 dummyFunc(UNUSED const u8 a) {
 #undef SHENG_IMPL
 #undef DEAD_FUNC
 #undef ACCEPT_FUNC
-#if defined(HAVE_AVX512VBMI)
+#if defined(HAVE_AVX512VBMI) || defined(HAVE_SVE)
 #undef SHENG32_IMPL
 #undef DEAD_FUNC32
 #undef ACCEPT_FUNC32
@@ -189,7 +189,7 @@ u8 dummyFunc(UNUSED const u8 a) {
 #define SHENG_IMPL sheng_sam
 #define DEAD_FUNC dummyFunc
 #define ACCEPT_FUNC isAcceptState
-#if defined(HAVE_AVX512VBMI)
+#if defined(HAVE_AVX512VBMI) || defined(HAVE_SVE)
 #define SHENG32_IMPL sheng32_sam
 #define DEAD_FUNC32 dummyFunc
 #define ACCEPT_FUNC32 isAcceptState32
@@ -202,7 +202,7 @@ u8 dummyFunc(UNUSED const u8 a) {
 #undef SHENG_IMPL
 #undef DEAD_FUNC
 #undef ACCEPT_FUNC
-#if defined(HAVE_AVX512VBMI)
+#if defined(HAVE_AVX512VBMI) || defined(HAVE_SVE)
 #undef SHENG32_IMPL
 #undef DEAD_FUNC32
 #undef ACCEPT_FUNC32
@@ -216,7 +216,7 @@ u8 dummyFunc(UNUSED const u8 a) {
 #define SHENG_IMPL sheng_nmd
 #define DEAD_FUNC isDeadState
 #define ACCEPT_FUNC dummyFunc
-#if defined(HAVE_AVX512VBMI)
+#if defined(HAVE_AVX512VBMI) || defined(HAVE_SVE)
 #define SHENG32_IMPL sheng32_nmd
 #define DEAD_FUNC32 isDeadState32
 #define ACCEPT_FUNC32 dummyFunc
@@ -229,7 +229,7 @@ u8 dummyFunc(UNUSED const u8 a) {
 #undef SHENG_IMPL
 #undef DEAD_FUNC
 #undef ACCEPT_FUNC
-#if defined(HAVE_AVX512VBMI)
+#if defined(HAVE_AVX512VBMI) || defined(HAVE_SVE)
 #undef SHENG32_IMPL
 #undef DEAD_FUNC32
 #undef ACCEPT_FUNC32
@@ -243,7 +243,7 @@ u8 dummyFunc(UNUSED const u8 a) {
 #define SHENG_IMPL sheng_nm
 #define DEAD_FUNC dummyFunc
 #define ACCEPT_FUNC dummyFunc
-#if defined(HAVE_AVX512VBMI)
+#if defined(HAVE_AVX512VBMI) || defined(HAVE_SVE)
 #define SHENG32_IMPL sheng32_nm
 #define DEAD_FUNC32 dummyFunc
 #define ACCEPT_FUNC32 dummyFunc
@@ -256,7 +256,7 @@ u8 dummyFunc(UNUSED const u8 a) {
 #undef SHENG_IMPL
 #undef DEAD_FUNC
 #undef ACCEPT_FUNC
-#if defined(HAVE_AVX512VBMI)
+#if defined(HAVE_AVX512VBMI) || defined(HAVE_SVE)
 #undef SHENG32_IMPL
 #undef DEAD_FUNC32
 #undef ACCEPT_FUNC32
@@ -277,7 +277,7 @@ u8 dummyFunc(UNUSED const u8 a) {
 #define INNER_ACCEL_FUNC isAccelState
 #define OUTER_ACCEL_FUNC dummyFunc
 #define ACCEPT_FUNC isAcceptState
-#if defined(HAVE_AVX512VBMI)
+#if defined(HAVE_AVX512VBMI) || defined(HAVE_SVE)
 #define SHENG32_IMPL sheng32_4_coda
 #define INTERESTING_FUNC32 hasInterestingStates32
 #define INNER_DEAD_FUNC32 isDeadState32
@@ -296,7 +296,7 @@ u8 dummyFunc(UNUSED const u8 a) {
 #undef INNER_ACCEL_FUNC
 #undef OUTER_ACCEL_FUNC
 #undef ACCEPT_FUNC
-#if defined(HAVE_AVX512VBMI)
+#if defined(HAVE_AVX512VBMI) || defined(HAVE_SVE)
 #undef SHENG32_IMPL
 #undef INTERESTING_FUNC32
 #undef INNER_DEAD_FUNC32
@@ -316,7 +316,7 @@ u8 dummyFunc(UNUSED const u8 a) {
 #define INNER_ACCEL_FUNC dummyFunc
 #define OUTER_ACCEL_FUNC dummyFunc
 #define ACCEPT_FUNC isAcceptState
-#if defined(HAVE_AVX512VBMI)
+#if defined(HAVE_AVX512VBMI) || defined(HAVE_SVE)
 #define SHENG32_IMPL sheng32_4_cod
 #define INTERESTING_FUNC32 hasInterestingStates32
 #define INNER_DEAD_FUNC32 isDeadState32
@@ -339,7 +339,7 @@ u8 dummyFunc(UNUSED const u8 a) {
 #undef INNER_ACCEL_FUNC
 #undef OUTER_ACCEL_FUNC
 #undef ACCEPT_FUNC
-#if defined(HAVE_AVX512VBMI)
+#if defined(HAVE_AVX512VBMI) || defined(HAVE_SVE)
 #undef SHENG32_IMPL
 #undef INTERESTING_FUNC32
 #undef INNER_DEAD_FUNC32
@@ -363,7 +363,7 @@ u8 dummyFunc(UNUSED const u8 a) {
 #define INNER_ACCEL_FUNC isAccelState
 #define OUTER_ACCEL_FUNC dummyFunc
 #define ACCEPT_FUNC isAcceptState
-#if defined(HAVE_AVX512VBMI)
+#if defined(HAVE_AVX512VBMI) || defined(HAVE_SVE)
 #define SHENG32_IMPL sheng32_4_coa
 #define INTERESTING_FUNC32 hasInterestingStates32
 #define INNER_DEAD_FUNC32 dummyFunc
@@ -382,7 +382,7 @@ u8 dummyFunc(UNUSED const u8 a) {
 #undef INNER_ACCEL_FUNC
 #undef OUTER_ACCEL_FUNC
 #undef ACCEPT_FUNC
-#if defined(HAVE_AVX512VBMI)
+#if defined(HAVE_AVX512VBMI) || defined(HAVE_SVE)
 #undef SHENG32_IMPL
 #undef INTERESTING_FUNC32
 #undef INNER_DEAD_FUNC32
@@ -402,7 +402,7 @@ u8 dummyFunc(UNUSED const u8 a) {
 #define INNER_ACCEL_FUNC dummyFunc
 #define OUTER_ACCEL_FUNC dummyFunc
 #define ACCEPT_FUNC isAcceptState
-#if defined(HAVE_AVX512VBMI)
+#if defined(HAVE_AVX512VBMI) || defined(HAVE_SVE)
 #define SHENG32_IMPL sheng32_4_co
 #define INTERESTING_FUNC32 hasInterestingStates32
 #define INNER_DEAD_FUNC32 dummyFunc
@@ -425,7 +425,7 @@ u8 dummyFunc(UNUSED const u8 a) {
 #undef INNER_ACCEL_FUNC
 #undef OUTER_ACCEL_FUNC
 #undef ACCEPT_FUNC
-#if defined(HAVE_AVX512VBMI)
+#if defined(HAVE_AVX512VBMI) || defined(HAVE_SVE)
 #undef SHENG32_IMPL
 #undef INTERESTING_FUNC32
 #undef INNER_DEAD_FUNC32
@@ -449,7 +449,7 @@ u8 dummyFunc(UNUSED const u8 a) {
 #define INNER_ACCEL_FUNC isAccelState
 #define OUTER_ACCEL_FUNC dummyFunc
 #define ACCEPT_FUNC isAcceptState
-#if defined(HAVE_AVX512VBMI)
+#if defined(HAVE_AVX512VBMI) || defined(HAVE_SVE)
 #define SHENG32_IMPL sheng32_4_samda
 #define INTERESTING_FUNC32 hasInterestingStates32
 #define INNER_DEAD_FUNC32 isDeadState32
@@ -468,7 +468,7 @@ u8 dummyFunc(UNUSED const u8 a) {
 #undef INNER_ACCEL_FUNC
 #undef OUTER_ACCEL_FUNC
 #undef ACCEPT_FUNC
-#if defined(HAVE_AVX512VBMI)
+#if defined(HAVE_AVX512VBMI) || defined(HAVE_SVE)
 #undef SHENG32_IMPL
 #undef INTERESTING_FUNC32
 #undef INNER_DEAD_FUNC32
@@ -488,7 +488,7 @@ u8 dummyFunc(UNUSED const u8 a) {
 #define INNER_ACCEL_FUNC dummyFunc
 #define OUTER_ACCEL_FUNC dummyFunc
 #define ACCEPT_FUNC isAcceptState
-#if defined(HAVE_AVX512VBMI)
+#if defined(HAVE_AVX512VBMI) || defined(HAVE_SVE)
 #define SHENG32_IMPL sheng32_4_samd
 #define INTERESTING_FUNC32 hasInterestingStates32
 #define INNER_DEAD_FUNC32 isDeadState32
@@ -511,7 +511,7 @@ u8 dummyFunc(UNUSED const u8 a) {
 #undef INNER_ACCEL_FUNC
 #undef OUTER_ACCEL_FUNC
 #undef ACCEPT_FUNC
-#if defined(HAVE_AVX512VBMI)
+#if defined(HAVE_AVX512VBMI) || defined(HAVE_SVE)
 #undef SHENG32_IMPL
 #undef INTERESTING_FUNC32
 #undef INNER_DEAD_FUNC32
@@ -535,7 +535,7 @@ u8 dummyFunc(UNUSED const u8 a) {
 #define INNER_ACCEL_FUNC isAccelState
 #define OUTER_ACCEL_FUNC dummyFunc
 #define ACCEPT_FUNC isAcceptState
-#if defined(HAVE_AVX512VBMI)
+#if defined(HAVE_AVX512VBMI) || defined(HAVE_SVE)
 #define SHENG32_IMPL sheng32_4_sama
 #define INTERESTING_FUNC32 hasInterestingStates32
 #define INNER_DEAD_FUNC32 dummyFunc
@@ -554,7 +554,7 @@ u8 dummyFunc(UNUSED const u8 a) {
 #undef INNER_ACCEL_FUNC
 #undef OUTER_ACCEL_FUNC
 #undef ACCEPT_FUNC
-#if defined(HAVE_AVX512VBMI)
+#if defined(HAVE_AVX512VBMI) || defined(HAVE_SVE)
 #undef SHENG32_IMPL
 #undef INTERESTING_FUNC32
 #undef INNER_DEAD_FUNC32
@@ -574,7 +574,7 @@ u8 dummyFunc(UNUSED const u8 a) {
 #define INNER_ACCEL_FUNC dummyFunc
 #define OUTER_ACCEL_FUNC dummyFunc
 #define ACCEPT_FUNC isAcceptState
-#if defined(HAVE_AVX512VBMI)
+#if defined(HAVE_AVX512VBMI) || defined(HAVE_SVE)
 #define SHENG32_IMPL sheng32_4_sam
 #define INTERESTING_FUNC32 hasInterestingStates32
 #define INNER_DEAD_FUNC32 dummyFunc
@@ -597,7 +597,7 @@ u8 dummyFunc(UNUSED const u8 a) {
 #undef INNER_ACCEL_FUNC
 #undef OUTER_ACCEL_FUNC
 #undef ACCEPT_FUNC
-#if defined(HAVE_AVX512VBMI)
+#if defined(HAVE_AVX512VBMI) || defined(HAVE_SVE)
 #undef SHENG32_IMPL
 #undef INTERESTING_FUNC32
 #undef INNER_DEAD_FUNC32
@@ -623,7 +623,7 @@ u8 dummyFunc(UNUSED const u8 a) {
 #define INNER_ACCEL_FUNC dummyFunc
 #define OUTER_ACCEL_FUNC isAccelState
 #define ACCEPT_FUNC dummyFunc
-#if defined(HAVE_AVX512VBMI)
+#if defined(HAVE_AVX512VBMI) || defined(HAVE_SVE)
 #define SHENG32_IMPL sheng32_4_nmda
 #define INTERESTING_FUNC32 dummyFunc4
 #define INNER_DEAD_FUNC32 dummyFunc
@@ -642,7 +642,7 @@ u8 dummyFunc(UNUSED const u8 a) {
 #undef INNER_ACCEL_FUNC
 #undef OUTER_ACCEL_FUNC
 #undef ACCEPT_FUNC
-#if defined(HAVE_AVX512VBMI)
+#if defined(HAVE_AVX512VBMI) || defined(HAVE_SVE)
 #undef SHENG32_IMPL
 #undef INTERESTING_FUNC32
 #undef INNER_DEAD_FUNC32
@@ -662,7 +662,7 @@ u8 dummyFunc(UNUSED const u8 a) {
 #define INNER_ACCEL_FUNC dummyFunc
 #define OUTER_ACCEL_FUNC dummyFunc
 #define ACCEPT_FUNC dummyFunc
-#if defined(HAVE_AVX512VBMI)
+#if defined(HAVE_AVX512VBMI) || defined(HAVE_SVE)
 #define SHENG32_IMPL sheng32_4_nmd
 #define INTERESTING_FUNC32 dummyFunc4
 #define INNER_DEAD_FUNC32 dummyFunc
@@ -685,7 +685,7 @@ u8 dummyFunc(UNUSED const u8 a) {
 #undef INNER_ACCEL_FUNC
 #undef OUTER_ACCEL_FUNC
 #undef ACCEPT_FUNC
-#if defined(HAVE_AVX512VBMI)
+#if defined(HAVE_AVX512VBMI) || defined(HAVE_SVE)
 #undef SHENG32_IMPL
 #undef INTERESTING_FUNC32
 #undef INNER_DEAD_FUNC32
@@ -712,7 +712,7 @@ u8 dummyFunc(UNUSED const u8 a) {
 #define INNER_ACCEL_FUNC dummyFunc
 #define OUTER_ACCEL_FUNC dummyFunc
 #define ACCEPT_FUNC dummyFunc
-#if defined(HAVE_AVX512VBMI)
+#if defined(HAVE_AVX512VBMI) || defined(HAVE_SVE)
 #define SHENG32_IMPL sheng32_4_nm
 #define INTERESTING_FUNC32 dummyFunc4
 #define INNER_DEAD_FUNC32 dummyFunc
@@ -735,7 +735,7 @@ u8 dummyFunc(UNUSED const u8 a) {
 #undef INNER_ACCEL_FUNC
 #undef OUTER_ACCEL_FUNC
 #undef ACCEPT_FUNC
-#if defined(HAVE_AVX512VBMI)
+#if defined(HAVE_AVX512VBMI) || defined(HAVE_SVE)
 #undef SHENG32_IMPL
 #undef INTERESTING_FUNC32
 #undef INNER_DEAD_FUNC32

src/nfa/sheng_impl.h

@@ -96,7 +96,7 @@ char SHENG_IMPL(u8 *state, NfaCallback cb, void *ctxt, const struct sheng *s,
     return MO_CONTINUE_MATCHING;
 }
 
-#if defined(HAVE_AVX512VBMI)
+#if defined(HAVE_AVX512VBMI) || defined(HAVE_SVE)
 static really_inline
 char SHENG32_IMPL(u8 *state, NfaCallback cb, void *ctxt,
                   const struct sheng32 *s,
@@ -114,14 +114,28 @@ char SHENG32_IMPL(u8 *state, NfaCallback cb, void *ctxt,
     }
     DEBUG_PRINTF("Scanning %lli bytes\n", (s64a)(end - start));
 
+#if defined(HAVE_SVE)
+    const svbool_t lane_pred_32 = svwhilelt_b8(0, 32);
+    svuint8_t cur_state = svdup_u8(*state);
+    svuint8_t tbl_mask = svdup_u8((unsigned char)0x1F);
+    const m512 *masks = s->succ_masks;
+#else
     m512 cur_state = set1_64x8(*state);
     const m512 *masks = s->succ_masks;
+#endif
 
     while (likely(cur_buf != end)) {
         const u8 c = *cur_buf;
+
+#if defined(HAVE_SVE)
+        svuint8_t succ_mask = svld1(lane_pred_32, (const u8*)(masks + c));
+        cur_state = svtbl(succ_mask, svand_x(svptrue_b8(), tbl_mask, cur_state));
+        const u8 tmp = svlastb(lane_pred_32, cur_state);
+#else
         const m512 succ_mask = masks[c];
         cur_state = vpermb512(cur_state, succ_mask);
         const u8 tmp = movd512(cur_state);
+#endif
 
         DEBUG_PRINTF("c: %02hhx '%c'\n", c, ourisprint(c) ? c : '?');
         DEBUG_PRINTF("s: %u (flag: %u)\n", tmp & SHENG32_STATE_MASK,
@@ -153,7 +167,11 @@ char SHENG32_IMPL(u8 *state, NfaCallback cb, void *ctxt,
         }
         cur_buf++;
     }
+#if defined(HAVE_SVE)
+    *state = svlastb(lane_pred_32, cur_state);
+#else
     *state = movd512(cur_state);
+#endif
     *scan_end = cur_buf;
     return MO_CONTINUE_MATCHING;
 }
@@ -175,14 +193,28 @@ char SHENG64_IMPL(u8 *state, NfaCallback cb, void *ctxt,
     }
     DEBUG_PRINTF("Scanning %lli bytes\n", (s64a)(end - start));
 
+#if defined(HAVE_SVE)
+    const svbool_t lane_pred_64 = svwhilelt_b8(0, 64);
+    svuint8_t cur_state = svdup_u8(*state);
+    svuint8_t tbl_mask = svdup_u8((unsigned char)0x3F);
+    const m512 *masks = s->succ_masks;
+#else
     m512 cur_state = set1_64x8(*state);
     const m512 *masks = s->succ_masks;
+#endif
 
     while (likely(cur_buf != end)) {
         const u8 c = *cur_buf;
+
+#if defined(HAVE_SVE)
+        svuint8_t succ_mask = svld1(lane_pred_64, (const u8*)(masks + c));
+        cur_state = svtbl(succ_mask, svand_x(svptrue_b8(), tbl_mask, cur_state));
+        const u8 tmp = svlastb(lane_pred_64, cur_state);
+#else
         const m512 succ_mask = masks[c];
         cur_state = vpermb512(cur_state, succ_mask);
         const u8 tmp = movd512(cur_state);
+#endif
 
         DEBUG_PRINTF("c: %02hhx '%c'\n", c, ourisprint(c) ? c : '?');
         DEBUG_PRINTF("s: %u (flag: %u)\n", tmp & SHENG64_STATE_MASK,
@@ -214,7 +246,11 @@ char SHENG64_IMPL(u8 *state, NfaCallback cb, void *ctxt,
         }
         cur_buf++;
     }
+#if defined(HAVE_SVE)
+    *state = svlastb(lane_pred_64, cur_state);
+#else
     *state = movd512(cur_state);
+#endif
     *scan_end = cur_buf;
     return MO_CONTINUE_MATCHING;
 }

src/nfa/sheng_impl4.h

@@ -283,7 +283,7 @@ char SHENG_IMPL(u8 *state, NfaCallback cb, void *ctxt, const struct sheng *s,
     return MO_CONTINUE_MATCHING;
 }
 
-#if defined(HAVE_AVX512VBMI)
+#if defined(HAVE_AVX512VBMI) || defined(HAVE_SVE)
 static really_inline
 char SHENG32_IMPL(u8 *state, NfaCallback cb, void *ctxt,
                   const struct sheng32 *s,
@@ -320,8 +320,15 @@ char SHENG32_IMPL(u8 *state, NfaCallback cb, void *ctxt,
         return MO_CONTINUE_MATCHING;
     }
 
+#if defined(HAVE_SVE)
+    const svbool_t lane_pred_32 = svwhilelt_b8(0, 32);
+    svuint8_t cur_state = svdup_u8(*state);
+    svuint8_t tbl_mask = svdup_u8((unsigned char)0x1F);
+    const m512 *masks = s->succ_masks;
+#else
     m512 cur_state = set1_64x8(*state);
     const m512 *masks = s->succ_masks;
+#endif
 
     while (likely(end - cur_buf >= 4)) {
         const u8 *b1 = cur_buf;
@@ -333,6 +340,23 @@ char SHENG32_IMPL(u8 *state, NfaCallback cb, void *ctxt,
         const u8 c3 = *b3;
         const u8 c4 = *b4;
 
+#if defined(HAVE_SVE)
+        svuint8_t succ_mask1 = svld1(lane_pred_32, (const u8*)(masks+c1));
+        cur_state = svtbl(succ_mask1, svand_x(svptrue_b8(), tbl_mask, cur_state));
+        const u8 a1 = svlastb(lane_pred_32, cur_state);
+
+        svuint8_t succ_mask2 = svld1(lane_pred_32, (const u8*)(masks+c2));
+        cur_state = svtbl(succ_mask2, svand_x(svptrue_b8(), tbl_mask, cur_state));
+        const u8 a2 = svlastb(lane_pred_32, cur_state);
+
+        svuint8_t succ_mask3 = svld1(lane_pred_32, (const u8*)(masks+c3));
+        cur_state = svtbl(succ_mask3, svand_x(svptrue_b8(), tbl_mask, cur_state));
+        const u8 a3 = svlastb(lane_pred_32, cur_state);
+
+        svuint8_t succ_mask4 = svld1(lane_pred_32, (const u8*)(masks+c4));
+        cur_state = svtbl(succ_mask4, svand_x(svptrue_b8(), tbl_mask, cur_state));
+        const u8 a4 = svlastb(lane_pred_32, cur_state);
+#else
         const m512 succ_mask1 = masks[c1];
         cur_state = vpermb512(cur_state, succ_mask1);
         const u8 a1 = movd512(cur_state);
@@ -348,6 +372,7 @@ char SHENG32_IMPL(u8 *state, NfaCallback cb, void *ctxt,
         const m512 succ_mask4 = masks[c4];
         cur_state = vpermb512(cur_state, succ_mask4);
         const u8 a4 = movd512(cur_state);
+#endif
 
         DEBUG_PRINTF("c: %02hhx '%c'\n", c1, ourisprint(c1) ? c1 : '?');
         DEBUG_PRINTF("s: %u (flag: %u)\n", a1 & SHENG32_STATE_MASK,
@@ -517,7 +542,11 @@ char SHENG32_IMPL(u8 *state, NfaCallback cb, void *ctxt,
         };
         cur_buf += 4;
     }
+#if defined(HAVE_SVE)
+    *state = svlastb(lane_pred_32, cur_state);
+#else
     *state = movd512(cur_state);
+#endif
     *scan_end = cur_buf;
     return MO_CONTINUE_MATCHING;
 }
@@ -541,9 +570,15 @@ char SHENG64_IMPL(u8 *state, NfaCallback cb, void *ctxt,
         *scan_end = end;
         return MO_CONTINUE_MATCHING;
     }
-
+#if defined(HAVE_SVE)
+    const svbool_t lane_pred_64 = svwhilelt_b8(0, 64);
+    svuint8_t cur_state = svdup_u8(*state);
+    svuint8_t tbl_mask = svdup_u8((unsigned char)0x3F);
+    const m512 *masks = s->succ_masks;
+#else
     m512 cur_state = set1_64x8(*state);
     const m512 *masks = s->succ_masks;
+#endif
 
     while (likely(end - cur_buf >= 4)) {
         const u8 *b1 = cur_buf;
@@ -555,6 +590,23 @@ char SHENG64_IMPL(u8 *state, NfaCallback cb, void *ctxt,
         const u8 c3 = *b3;
         const u8 c4 = *b4;
 
+#if defined(HAVE_SVE)
+        svuint8_t succ_mask1 = svld1(lane_pred_64, (const u8*)(masks+c1));
+        cur_state = svtbl(succ_mask1, svand_x(svptrue_b8(), tbl_mask, cur_state));
+        const u8 a1 = svlastb(lane_pred_64, cur_state);
+
+        svuint8_t succ_mask2 = svld1(lane_pred_64, (const u8*)(masks+c2));
+        cur_state = svtbl(succ_mask2, svand_x(svptrue_b8(), tbl_mask, cur_state));
+        const u8 a2 = svlastb(lane_pred_64, cur_state);
+
+        svuint8_t succ_mask3 = svld1(lane_pred_64, (const u8*)(masks+c3));
+        cur_state = svtbl(succ_mask3, svand_x(svptrue_b8(), tbl_mask, cur_state));
+        const u8 a3 = svlastb(lane_pred_64, cur_state);
+
+        svuint8_t succ_mask4 = svld1(lane_pred_64, (const u8*)(masks+c4));
+        cur_state = svtbl(succ_mask4, svand_x(svptrue_b8(), tbl_mask, cur_state));
+        const u8 a4 = svlastb(lane_pred_64, cur_state);
+#else
         const m512 succ_mask1 = masks[c1];
         cur_state = vpermb512(cur_state, succ_mask1);
         const u8 a1 = movd512(cur_state);
@@ -570,6 +622,7 @@ char SHENG64_IMPL(u8 *state, NfaCallback cb, void *ctxt,
         const m512 succ_mask4 = masks[c4];
         cur_state = vpermb512(cur_state, succ_mask4);
         const u8 a4 = movd512(cur_state);
+#endif
 
         DEBUG_PRINTF("c: %02hhx '%c'\n", c1, ourisprint(c1) ? c1 : '?');
         DEBUG_PRINTF("s: %u (flag: %u)\n", a1 & SHENG64_STATE_MASK,
@@ -703,7 +756,11 @@ char SHENG64_IMPL(u8 *state, NfaCallback cb, void *ctxt,
         }
         cur_buf += 4;
     }
+#if defined(HAVE_SVE)
+    *state = svlastb(lane_pred_64, cur_state);
+#else
     *state = movd512(cur_state);
+#endif
     *scan_end = cur_buf;
     return MO_CONTINUE_MATCHING;
 }

src/nfa/shengcompile.cpp

@@ -730,10 +730,17 @@ bytecode_ptr<NFA> sheng32Compile(raw_dfa &raw, const CompileContext &cc,
         return nullptr;
     }
 
+#ifdef HAVE_SVE
+    if (svcntb()<32) {
+        DEBUG_PRINTF("Sheng32 failed, SVE width is too small!\n");
+        return nullptr;
+    }
+#else
     if (!cc.target_info.has_avx512vbmi()) {
         DEBUG_PRINTF("Sheng32 failed, no HS_CPU_FEATURES_AVX512VBMI!\n");
         return nullptr;
     }
+#endif
 
     sheng_build_strat strat(raw, rm, only_accel_init);
     dfa_info info(strat);
@@ -762,10 +769,17 @@ bytecode_ptr<NFA> sheng64Compile(raw_dfa &raw, const CompileContext &cc,
         return nullptr;
     }
 
+#ifdef HAVE_SVE
+    if (svcntb()<64) {
+        DEBUG_PRINTF("Sheng64 failed, SVE width is too small!\n");
+        return nullptr;
+    }
+#else
     if (!cc.target_info.has_avx512vbmi()) {
         DEBUG_PRINTF("Sheng64 failed, no HS_CPU_FEATURES_AVX512VBMI!\n");
         return nullptr;
     }
+#endif
 
     sheng_build_strat strat(raw, rm, only_accel_init);
     dfa_info info(strat);