diff --git a/src/asmjit/arm/a64emitter.h b/src/asmjit/arm/a64emitter.h
index c1964e8..34fe092 100644
--- a/src/asmjit/arm/a64emitter.h
+++ b/src/asmjit/arm/a64emitter.h
@@ -133,22 +133,22 @@ struct EmitterExplicitT {
   ASMJIT_INST_4x(bfm, Bfm, Gp, Gp, Imm, Imm)
   ASMJIT_INST_4x(bfxil, Bfxil, Gp, Gp, Imm, Imm)
 
-  ASMJIT_INST_3x(bic, Bic, Gp, Gp, Imm);
-  ASMJIT_INST_3x(bic, Bic, Gp, Gp, Gp);
-  ASMJIT_INST_4x(bic, Bic, Gp, Gp, Gp, Imm);
-  ASMJIT_INST_3x(bics, Bics, Gp, Gp, Imm);
-  ASMJIT_INST_3x(bics, Bics, Gp, Gp, Gp);
-  ASMJIT_INST_4x(bics, Bics, Gp, Gp, Gp, Imm);
+  ASMJIT_INST_3x(bic, Bic, Gp, Gp, Imm)
+  ASMJIT_INST_3x(bic, Bic, Gp, Gp, Gp)
+  ASMJIT_INST_4x(bic, Bic, Gp, Gp, Gp, Imm)
+  ASMJIT_INST_3x(bics, Bics, Gp, Gp, Imm)
+  ASMJIT_INST_3x(bics, Bics, Gp, Gp, Gp)
+  ASMJIT_INST_4x(bics, Bics, Gp, Gp, Gp, Imm)
 
   ASMJIT_INST_1x(brk, Brk, Imm)
 
-  ASMJIT_INST_4x(ccmn, Ccmn, Gp, Gp, Imm, Imm);
-  ASMJIT_INST_4x(ccmn, Ccmn, Gp, Imm, Imm, Imm);
-  ASMJIT_INST_4x(ccmp, Ccmp, Gp, Gp, Imm, Imm);
-  ASMJIT_INST_4x(ccmp, Ccmp, Gp, Imm, Imm, Imm);
+  ASMJIT_INST_4x(ccmn, Ccmn, Gp, Gp, Imm, Imm)
+  ASMJIT_INST_4x(ccmn, Ccmn, Gp, Imm, Imm, Imm)
+  ASMJIT_INST_4x(ccmp, Ccmp, Gp, Gp, Imm, Imm)
+  ASMJIT_INST_4x(ccmp, Ccmp, Gp, Imm, Imm, Imm)
 
-  ASMJIT_INST_3x(cinc, Cinc, Gp, Gp, Imm);
-  ASMJIT_INST_3x(cinv, Cinv, Gp, Gp, Imm);
+  ASMJIT_INST_3x(cinc, Cinc, Gp, Gp, Imm)
+  ASMJIT_INST_3x(cinv, Cinv, Gp, Gp, Imm)
 
   ASMJIT_INST_1x(clrex, Clrex, Imm)
 
@@ -164,15 +164,15 @@ struct EmitterExplicitT {
   ASMJIT_INST_2x(cmp, Cmp, Gp, Imm)
   ASMJIT_INST_3x(cmp, Cmp, Gp, Imm, Imm)
 
-  ASMJIT_INST_3x(cneg, Cneg, Gp, Gp, Imm);
+  ASMJIT_INST_3x(cneg, Cneg, Gp, Gp, Imm)
 
-  ASMJIT_INST_4x(csel, Csel, Gp, Gp, Gp, Imm);
-  ASMJIT_INST_2x(cset, Cset, Gp, Imm);
-  ASMJIT_INST_2x(csetm, Csetm, Gp, Imm);
+  ASMJIT_INST_4x(csel, Csel, Gp, Gp, Gp, Imm)
+  ASMJIT_INST_2x(cset, Cset, Gp, Imm)
+  ASMJIT_INST_2x(csetm, Csetm, Gp, Imm)
 
-  ASMJIT_INST_4x(csinc, Csinc, Gp, Gp, Gp, Imm);
-  ASMJIT_INST_4x(csinv, Csinv, Gp, Gp, Gp, Imm);
-  ASMJIT_INST_4x(csneg, Csneg, Gp, Gp, Gp, Imm);
+  ASMJIT_INST_4x(csinc, Csinc, Gp, Gp, Gp, Imm)
+  ASMJIT_INST_4x(csinv, Csinv, Gp, Gp, Gp, Imm)
+  ASMJIT_INST_4x(csneg, Csneg, Gp, Gp, Gp, Imm)
 
   ASMJIT_INST_2x(dc, Dc, Imm, Gp)
   ASMJIT_INST_1x(dmb, Dmb, Imm)
@@ -619,53 +619,53 @@ struct EmitterExplicitT {
   //! \name BTI Instructions
   //! \{
 
-  ASMJIT_INST_1x(bti, Bti, Imm);
+  ASMJIT_INST_1x(bti, Bti, Imm)
 
   //! \}
 
   //! \name CHK Instructions
   //! \{
 
-  ASMJIT_INST_1x(chkfeat, Chkfeat, Gp);
+  ASMJIT_INST_1x(chkfeat, Chkfeat, Gp)
 
   //! \}
 
   //! \name CLRBHB Instructions
   //! \{
 
-  ASMJIT_INST_0x(clrbhb, Clrbhb);
+  ASMJIT_INST_0x(clrbhb, Clrbhb)
 
   //! \}
 
   //! \name CRC Instructions (ARMv8.1-A, optional in ARMv8.0-A)
   //! \{
 
-  ASMJIT_INST_3x(crc32b, Crc32b, Gp, Gp, Gp);
-  ASMJIT_INST_3x(crc32h, Crc32h, Gp, Gp, Gp);
-  ASMJIT_INST_3x(crc32w, Crc32w, Gp, Gp, Gp);
-  ASMJIT_INST_3x(crc32x, Crc32x, Gp, Gp, Gp);
+  ASMJIT_INST_3x(crc32b, Crc32b, Gp, Gp, Gp)
+  ASMJIT_INST_3x(crc32h, Crc32h, Gp, Gp, Gp)
+  ASMJIT_INST_3x(crc32w, Crc32w, Gp, Gp, Gp)
+  ASMJIT_INST_3x(crc32x, Crc32x, Gp, Gp, Gp)
 
-  ASMJIT_INST_3x(crc32cb, Crc32cb, Gp, Gp, Gp);
-  ASMJIT_INST_3x(crc32ch, Crc32ch, Gp, Gp, Gp);
-  ASMJIT_INST_3x(crc32cw, Crc32cw, Gp, Gp, Gp);
-  ASMJIT_INST_3x(crc32cx, Crc32cx, Gp, Gp, Gp);
+  ASMJIT_INST_3x(crc32cb, Crc32cb, Gp, Gp, Gp)
+  ASMJIT_INST_3x(crc32ch, Crc32ch, Gp, Gp, Gp)
+  ASMJIT_INST_3x(crc32cw, Crc32cw, Gp, Gp, Gp)
+  ASMJIT_INST_3x(crc32cx, Crc32cx, Gp, Gp, Gp)
 
   //! \}
 
   //! \name CSSC Instructions
   //! \{
 
-  ASMJIT_INST_2x(abs, Abs, Gp, Gp);
-  ASMJIT_INST_2x(cnt, Cnt, Gp, Gp);
-  ASMJIT_INST_2x(ctz, Ctz, Gp, Gp);
-  ASMJIT_INST_3x(smax, Smax, Gp, Gp, Gp);
-  ASMJIT_INST_3x(smax, Smax, Gp, Gp, Imm);
-  ASMJIT_INST_3x(smin, Smin, Gp, Gp, Gp);
-  ASMJIT_INST_3x(smin, Smin, Gp, Gp, Imm);
-  ASMJIT_INST_3x(umax, Umax, Gp, Gp, Gp);
-  ASMJIT_INST_3x(umax, Umax, Gp, Gp, Imm);
-  ASMJIT_INST_3x(umin, Umin, Gp, Gp, Gp);
-  ASMJIT_INST_3x(umin, Umin, Gp, Gp, Imm);
+  ASMJIT_INST_2x(abs, Abs, Gp, Gp)
+  ASMJIT_INST_2x(cnt, Cnt, Gp, Gp)
+  ASMJIT_INST_2x(ctz, Ctz, Gp, Gp)
+  ASMJIT_INST_3x(smax, Smax, Gp, Gp, Gp)
+  ASMJIT_INST_3x(smax, Smax, Gp, Gp, Imm)
+  ASMJIT_INST_3x(smin, Smin, Gp, Gp, Gp)
+  ASMJIT_INST_3x(smin, Smin, Gp, Gp, Imm)
+  ASMJIT_INST_3x(umax, Umax, Gp, Gp, Gp)
+  ASMJIT_INST_3x(umax, Umax, Gp, Gp, Imm)
+  ASMJIT_INST_3x(umin, Umin, Gp, Gp, Gp)
+  ASMJIT_INST_3x(umin, Umin, Gp, Gp, Imm)
 
   //! \}
 
@@ -704,40 +704,40 @@ struct EmitterExplicitT {
   //! \name MTE Instructions
   //! \{
 
-  ASMJIT_INST_2x(autda, Autda, Gp, Gp);
-  ASMJIT_INST_2x(autdb, Autdb, Gp, Gp);
-  ASMJIT_INST_1x(autdza, Autdza, Gp);
-  ASMJIT_INST_1x(autdzb, Autdzb, Gp);
-  ASMJIT_INST_2x(autia, Autia, Gp, Gp);
-  ASMJIT_INST_0x(autia1716, Autia1716);
-  ASMJIT_INST_0x(autiasp, Autiasp);
-  ASMJIT_INST_0x(autiaz, Autiaz);
-  ASMJIT_INST_2x(autib, Autib, Gp, Gp);
-  ASMJIT_INST_0x(autib1716, Autib1716);
-  ASMJIT_INST_0x(autibsp, Autibsp);
-  ASMJIT_INST_0x(autibz, Autibz);
-  ASMJIT_INST_1x(autiza, Autiza, Gp);
-  ASMJIT_INST_1x(autizb, Autizb, Gp);
+  ASMJIT_INST_2x(autda, Autda, Gp, Gp)
+  ASMJIT_INST_2x(autdb, Autdb, Gp, Gp)
+  ASMJIT_INST_1x(autdza, Autdza, Gp)
+  ASMJIT_INST_1x(autdzb, Autdzb, Gp)
+  ASMJIT_INST_2x(autia, Autia, Gp, Gp)
+  ASMJIT_INST_0x(autia1716, Autia1716)
+  ASMJIT_INST_0x(autiasp, Autiasp)
+  ASMJIT_INST_0x(autiaz, Autiaz)
+  ASMJIT_INST_2x(autib, Autib, Gp, Gp)
+  ASMJIT_INST_0x(autib1716, Autib1716)
+  ASMJIT_INST_0x(autibsp, Autibsp)
+  ASMJIT_INST_0x(autibz, Autibz)
+  ASMJIT_INST_1x(autiza, Autiza, Gp)
+  ASMJIT_INST_1x(autizb, Autizb, Gp)
 
-  ASMJIT_INST_3x(gmi, Gmi, Gp, Gp, Gp);
+  ASMJIT_INST_3x(gmi, Gmi, Gp, Gp, Gp)
 
-  ASMJIT_INST_2x(cmpp, Cmpp, Gp, Gp);
-  ASMJIT_INST_4x(addg, Addg, Gp, Gp, Imm, Imm);
+  ASMJIT_INST_2x(cmpp, Cmpp, Gp, Gp)
+  ASMJIT_INST_4x(addg, Addg, Gp, Gp, Imm, Imm)
 
   ASMJIT_INST_2x(ldg, Ldg, Gp, Mem)
   ASMJIT_INST_2x(ldgm, Ldgm, Gp, Mem)
   ASMJIT_INST_2x(ldraa, Ldraa, Gp, Mem)
   ASMJIT_INST_2x(ldrab, Ldrab, Gp, Mem)
 
-  ASMJIT_INST_2x(pacda, Pacda, Gp, Gp);
-  ASMJIT_INST_2x(pacdb, Pacdb, Gp, Gp);
-  ASMJIT_INST_1x(pacdza, Pacdza, Gp);
-  ASMJIT_INST_1x(pacdzb, Pacdzb, Gp);
-  ASMJIT_INST_3x(pacga, Pacga, Gp, Gp, Gp);
+  ASMJIT_INST_2x(pacda, Pacda, Gp, Gp)
+  ASMJIT_INST_2x(pacdb, Pacdb, Gp, Gp)
+  ASMJIT_INST_1x(pacdza, Pacdza, Gp)
+  ASMJIT_INST_1x(pacdzb, Pacdzb, Gp)
+  ASMJIT_INST_3x(pacga, Pacga, Gp, Gp, Gp)
 
-  ASMJIT_INST_3x(subp, Subp, Gp, Gp, Gp);
-  ASMJIT_INST_3x(subps, Subps, Gp, Gp, Gp);
-  ASMJIT_INST_4x(subg, Subg, Gp, Gp, Imm, Imm);
+  ASMJIT_INST_3x(subp, Subp, Gp, Gp, Gp)
+  ASMJIT_INST_3x(subps, Subps, Gp, Gp, Gp)
+  ASMJIT_INST_4x(subg, Subg, Gp, Gp, Imm, Imm)
 
   ASMJIT_INST_2x(st2g, St2g, Gp, Mem)
   ASMJIT_INST_2x(stg, Stg, Gp, Mem)
@@ -747,9 +747,9 @@ struct EmitterExplicitT {
   ASMJIT_INST_2x(stz2g, Stz2g, Gp, Mem)
   ASMJIT_INST_2x(stzgm, Stzgm, Gp, Mem)
 
-  ASMJIT_INST_1x(xpacd, Xpacd, Gp);
-  ASMJIT_INST_1x(xpaci, Xpaci, Gp);
-  ASMJIT_INST_0x(xpaclri, Xpaclri);
+  ASMJIT_INST_1x(xpacd, Xpacd, Gp)
+  ASMJIT_INST_1x(xpaci, Xpaci, Gp)
+  ASMJIT_INST_0x(xpaclri, Xpaclri)
 
   //! \}
 
@@ -769,428 +769,428 @@ struct EmitterExplicitT {
   //! \name SIMD & FP Instructions
   //! \{
 
-  ASMJIT_INST_2x(abs, Abs_v, Vec, Vec);
-  ASMJIT_INST_3x(add, Add_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(addhn, Addhn_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(addhn2, Addhn2_v, Vec, Vec, Vec);
-  ASMJIT_INST_2x(addp, Addp_v, Vec, Vec);
-  ASMJIT_INST_3x(addp, Addp_v, Vec, Vec, Vec);
-  ASMJIT_INST_2x(addv, Addv_v, Vec, Vec);
-  ASMJIT_INST_3x(and_, And_v, Vec, Vec, Vec);
-  ASMJIT_INST_2x(bic, Bic_v, Vec, Imm);
-  ASMJIT_INST_3x(bic, Bic_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(bic, Bic_v, Vec, Imm, Imm);
-  ASMJIT_INST_3x(bif, Bif_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(bit, Bit_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(bsl, Bsl_v, Vec, Vec, Vec);
-  ASMJIT_INST_2x(cls, Cls_v, Vec, Vec);
-  ASMJIT_INST_2x(clz, Clz_v, Vec, Vec);
-  ASMJIT_INST_3x(cmeq, Cmeq_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(cmeq, Cmeq_v, Vec, Vec, Imm);
-  ASMJIT_INST_3x(cmge, Cmge_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(cmge, Cmge_v, Vec, Vec, Imm);
-  ASMJIT_INST_3x(cmgt, Cmgt_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(cmgt, Cmgt_v, Vec, Vec, Imm);
-  ASMJIT_INST_3x(cmhi, Cmhi_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(cmhs, Cmhs_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(cmle, Cmle_v, Vec, Vec, Imm);
-  ASMJIT_INST_3x(cmlt, Cmlt_v, Vec, Vec, Imm);
-  ASMJIT_INST_3x(cmtst, Cmtst_v, Vec, Vec, Vec);
-  ASMJIT_INST_2x(cnt, Cnt_v, Vec, Vec);
-  ASMJIT_INST_2x(dup, Dup_v, Vec, Gp);
-  ASMJIT_INST_2x(dup, Dup_v, Vec, Vec);
-  ASMJIT_INST_3x(eor, Eor_v, Vec, Vec, Vec);
-  ASMJIT_INST_4x(ext, Ext_v, Vec, Vec, Vec, Imm);
-  ASMJIT_INST_3x(fabd, Fabd_v, Vec, Vec, Vec);
-  ASMJIT_INST_2x(fabs, Fabs_v, Vec, Vec);
-  ASMJIT_INST_3x(facge, Facge_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(facgt, Facgt_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(fadd, Fadd_v, Vec, Vec, Vec);
-  ASMJIT_INST_2x(faddp, Faddp_v, Vec, Vec);
-  ASMJIT_INST_3x(faddp, Faddp_v, Vec, Vec, Vec);
-  ASMJIT_INST_4x(fccmp, Fccmp_v, Vec, Vec, Imm, Imm);
-  ASMJIT_INST_4x(fccmpe, Fccmpe_v, Vec, Vec, Imm, Imm);
-  ASMJIT_INST_3x(fcmeq, Fcmeq_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(fcmeq, Fcmeq_v, Vec, Vec, Imm);
-  ASMJIT_INST_3x(fcmge, Fcmge_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(fcmge, Fcmge_v, Vec, Vec, Imm);
-  ASMJIT_INST_3x(fcmgt, Fcmgt_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(fcmgt, Fcmgt_v, Vec, Vec, Imm);
-  ASMJIT_INST_3x(fcmle, Fcmle_v, Vec, Vec, Imm);
-  ASMJIT_INST_3x(fcmlt, Fcmlt_v, Vec, Vec, Imm);
-  ASMJIT_INST_2x(fcmp, Fcmp_v, Vec, Vec);
-  ASMJIT_INST_2x(fcmp, Fcmp_v, Vec, Imm);
-  ASMJIT_INST_2x(fcmpe, Fcmpe_v, Vec, Vec);
-  ASMJIT_INST_2x(fcmpe, Fcmpe_v, Vec, Imm);
-  ASMJIT_INST_4x(fcsel, Fcsel_v, Vec, Vec, Vec, Imm);
-  ASMJIT_INST_2x(fcvt, Fcvt_v, Vec, Vec);
-  ASMJIT_INST_2x(fcvtas, Fcvtas_v, Gp, Vec);
-  ASMJIT_INST_2x(fcvtas, Fcvtas_v, Vec, Vec);
-  ASMJIT_INST_2x(fcvtau, Fcvtau_v, Gp, Vec);
-  ASMJIT_INST_2x(fcvtau, Fcvtau_v, Vec, Vec);
-  ASMJIT_INST_2x(fcvtl, Fcvtl_v, Vec, Vec);
-  ASMJIT_INST_2x(fcvtl2, Fcvtl2_v, Vec, Vec);
-  ASMJIT_INST_2x(fcvtms, Fcvtms_v, Gp, Vec);
-  ASMJIT_INST_2x(fcvtms, Fcvtms_v, Vec, Vec);
-  ASMJIT_INST_2x(fcvtmu, Fcvtmu_v, Gp, Vec);
-  ASMJIT_INST_2x(fcvtmu, Fcvtmu_v, Vec, Vec);
-  ASMJIT_INST_2x(fcvtn, Fcvtn_v, Vec, Vec);
-  ASMJIT_INST_2x(fcvtn2, Fcvtn2_v, Vec, Vec);
-  ASMJIT_INST_2x(fcvtns, Fcvtns_v, Gp, Vec);
-  ASMJIT_INST_2x(fcvtns, Fcvtns_v, Vec, Vec);
-  ASMJIT_INST_2x(fcvtnu, Fcvtnu_v, Gp, Vec);
-  ASMJIT_INST_2x(fcvtnu, Fcvtnu_v, Vec, Vec);
-  ASMJIT_INST_2x(fcvtps, Fcvtps_v, Gp, Vec);
-  ASMJIT_INST_2x(fcvtps, Fcvtps_v, Vec, Vec);
-  ASMJIT_INST_2x(fcvtpu, Fcvtpu_v, Gp, Vec);
-  ASMJIT_INST_2x(fcvtpu, Fcvtpu_v, Vec, Vec);
-  ASMJIT_INST_2x(fcvtxn, Fcvtxn_v, Vec, Vec);
-  ASMJIT_INST_2x(fcvtxn2, Fcvtxn2_v, Vec, Vec);
-  ASMJIT_INST_2x(fcvtzs, Fcvtzs_v, Gp, Vec);
-  ASMJIT_INST_3x(fcvtzs, Fcvtzs_v, Gp, Vec, Imm);
-  ASMJIT_INST_2x(fcvtzs, Fcvtzs_v, Vec, Vec);
-  ASMJIT_INST_3x(fcvtzs, Fcvtzs_v, Vec, Vec, Imm);
-  ASMJIT_INST_2x(fcvtzu, Fcvtzu_v, Gp, Vec);
-  ASMJIT_INST_3x(fcvtzu, Fcvtzu_v, Gp, Vec, Imm);
-  ASMJIT_INST_2x(fcvtzu, Fcvtzu_v, Vec, Vec);
-  ASMJIT_INST_3x(fcvtzu, Fcvtzu_v, Vec, Vec, Imm);
-  ASMJIT_INST_3x(fdiv, Fdiv_v, Vec, Vec, Vec);
-  ASMJIT_INST_4x(fmadd, Fmadd_v, Vec, Vec, Vec, Vec);
-  ASMJIT_INST_3x(fmax, Fmax_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(fmaxnm, Fmaxnm_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(fmaxnmp, Fmaxnmp_v, Vec, Vec, Vec);
-  ASMJIT_INST_2x(fmaxnmp, Fmaxnmp_v, Vec, Vec);
-  ASMJIT_INST_2x(fmaxnmv, Fmaxnmv_v, Vec, Vec);
-  ASMJIT_INST_3x(fmaxp, Fmaxp_v, Vec, Vec, Vec);
-  ASMJIT_INST_2x(fmaxp, Fmaxp_v, Vec, Vec);
-  ASMJIT_INST_2x(fmaxv, Fmaxv_v, Vec, Vec);
-  ASMJIT_INST_3x(fmin, Fmin_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(fminnm, Fminnm_v, Vec, Vec, Vec);
-  ASMJIT_INST_2x(fminnmv, Fminnmv_v, Vec, Vec);
-  ASMJIT_INST_3x(fminnmp, Fminnmp_v, Vec, Vec, Vec);
-  ASMJIT_INST_2x(fminnmp, Fminnmp_v, Vec, Vec);
-  ASMJIT_INST_2x(fminp, Fminp_v, Vec, Vec);
-  ASMJIT_INST_3x(fminp, Fminp_v, Vec, Vec, Vec);
-  ASMJIT_INST_2x(fminv, Fminv_v, Vec, Vec);
-  ASMJIT_INST_3x(fmla, Fmla_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(fmls, Fmls_v, Vec, Vec, Vec);
-  ASMJIT_INST_2x(fmov, Fmov_v, Gp, Vec);
-  ASMJIT_INST_2x(fmov, Fmov_v, Vec, Gp);
-  ASMJIT_INST_2x(fmov, Fmov_v, Vec, Vec);
-  ASMJIT_INST_2x(fmov, Fmov_v, Vec, Imm);
-  ASMJIT_INST_4x(fmsub, Fmsub_v, Vec, Vec, Vec, Vec);
-  ASMJIT_INST_3x(fmul, Fmul_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(fmulx, Fmulx_v, Vec, Vec, Vec);
-  ASMJIT_INST_2x(fneg, Fneg_v, Vec, Vec);
-  ASMJIT_INST_4x(fnmadd, Fnmadd_v, Vec, Vec, Vec, Vec);
-  ASMJIT_INST_4x(fnmsub, Fnmsub_v, Vec, Vec, Vec, Vec);
-  ASMJIT_INST_3x(fnmul, Fnmul_v, Vec, Vec, Vec);
-  ASMJIT_INST_2x(frecpe, Frecpe_v, Vec, Vec);
-  ASMJIT_INST_3x(frecps, Frecps_v, Vec, Vec, Vec);
-  ASMJIT_INST_2x(frecpx, Frecpx_v, Vec, Vec);
-  ASMJIT_INST_2x(frint32x, Frint32x_v, Vec, Vec);
-  ASMJIT_INST_2x(frint32z, Frint32z_v, Vec, Vec);
-  ASMJIT_INST_2x(frint64x, Frint64x_v, Vec, Vec);
-  ASMJIT_INST_2x(frint64z, Frint64z_v, Vec, Vec);
-  ASMJIT_INST_2x(frinta, Frinta_v, Vec, Vec);
-  ASMJIT_INST_2x(frinti, Frinti_v, Vec, Vec);
-  ASMJIT_INST_2x(frintm, Frintm_v, Vec, Vec);
-  ASMJIT_INST_2x(frintn, Frintn_v, Vec, Vec);
-  ASMJIT_INST_2x(frintp, Frintp_v, Vec, Vec);
-  ASMJIT_INST_2x(frintx, Frintx_v, Vec, Vec);
-  ASMJIT_INST_2x(frintz, Frintz_v, Vec, Vec);
-  ASMJIT_INST_2x(frsqrte, Frsqrte_v, Vec, Vec);
-  ASMJIT_INST_3x(frsqrts, Frsqrts_v, Vec, Vec, Vec);
-  ASMJIT_INST_2x(fsqrt, Fsqrt_v, Vec, Vec);
-  ASMJIT_INST_3x(fsub, Fsub_v, Vec, Vec, Vec);
-  ASMJIT_INST_2x(ins, Ins_v, Vec, Gp);
-  ASMJIT_INST_2x(ins, Ins_v, Vec, Vec);
-  ASMJIT_INST_2x(ld1, Ld1_v, Vec, Mem);
-  ASMJIT_INST_3x(ld1, Ld1_v, Vec, Vec, Mem);
-  ASMJIT_INST_4x(ld1, Ld1_v, Vec, Vec, Vec, Mem);
-  ASMJIT_INST_5x(ld1, Ld1_v, Vec, Vec, Vec, Vec, Mem);
-  ASMJIT_INST_2x(ld1r, Ld1r_v, Vec, Mem);
-  ASMJIT_INST_3x(ld2, Ld2_v, Vec, Vec, Mem);
-  ASMJIT_INST_3x(ld2r, Ld2r_v, Vec, Vec, Mem);
-  ASMJIT_INST_4x(ld3, Ld3_v, Vec, Vec, Vec, Mem);
-  ASMJIT_INST_4x(ld3r, Ld3r_v, Vec, Vec, Vec, Mem);
-  ASMJIT_INST_5x(ld4, Ld4_v, Vec, Vec, Vec, Vec, Mem);
-  ASMJIT_INST_5x(ld4r, Ld4r_v, Vec, Vec, Vec, Vec, Mem);
-  ASMJIT_INST_3x(ldnp, Ldnp_v, Vec, Vec, Mem);
-  ASMJIT_INST_3x(ldp, Ldp_v, Vec, Vec, Mem);
-  ASMJIT_INST_2x(ldr, Ldr_v, Vec, Mem);
-  ASMJIT_INST_2x(ldur, Ldur_v, Vec, Mem);
-  ASMJIT_INST_3x(mla, Mla_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(mls, Mls_v, Vec, Vec, Vec);
-  ASMJIT_INST_2x(mov, Mov_v, Vec, Vec);
-  ASMJIT_INST_2x(mov, Mov_v, Gp, Vec);
-  ASMJIT_INST_2x(mov, Mov_v, Vec, Gp);
-  ASMJIT_INST_2x(movi, Movi_v, Vec, Imm);
-  ASMJIT_INST_3x(movi, Movi_v, Vec, Imm, Imm);
-  ASMJIT_INST_3x(mul, Mul_v, Vec, Vec, Vec);
-  ASMJIT_INST_2x(mvn, Mvn_v, Vec, Vec);
-  ASMJIT_INST_2x(mvn_, Mvn_v, Vec, Vec);
-  ASMJIT_INST_2x(mvni, Mvni_v, Vec, Imm);
-  ASMJIT_INST_3x(mvni, Mvni_v, Vec, Imm, Imm);
-  ASMJIT_INST_2x(neg, Neg_v, Vec, Vec);
-  ASMJIT_INST_2x(not_, Not_v, Vec, Vec);
-  ASMJIT_INST_3x(orn, Orn_v, Vec, Vec, Vec);
-  ASMJIT_INST_2x(orr, Orr_v, Vec, Imm);
-  ASMJIT_INST_3x(orr, Orr_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(orr, Orr_v, Vec, Imm, Imm);
-  ASMJIT_INST_3x(pmul, Pmul_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(pmull, Pmull_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(pmull2, Pmull2_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(raddhn, Raddhn_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(raddhn2, Raddhn2_v, Vec, Vec, Vec);
-  ASMJIT_INST_2x(rbit, Rbit_v, Vec, Vec);
-  ASMJIT_INST_2x(rev16, Rev16_v, Vec, Vec);
-  ASMJIT_INST_2x(rev32, Rev32_v, Vec, Vec);
-  ASMJIT_INST_2x(rev64, Rev64_v, Vec, Vec);
-  ASMJIT_INST_3x(rshrn, Rshrn_v, Vec, Vec, Imm);
-  ASMJIT_INST_3x(rshrn2, Rshrn2_v, Vec, Vec, Imm);
-  ASMJIT_INST_3x(rsubhn, Rsubhn_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(rsubhn2, Rsubhn2_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(saba, Saba_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(sabal, Sabal_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(sabal2, Sabal2_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(sabd, Sabd_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(sabdl, Sabdl_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(sabdl2, Sabdl2_v, Vec, Vec, Vec);
-  ASMJIT_INST_2x(sadalp, Sadalp_v, Vec, Vec);
-  ASMJIT_INST_3x(saddl, Saddl_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(saddl2, Saddl2_v, Vec, Vec, Vec);
-  ASMJIT_INST_2x(saddlp, Saddlp_v, Vec, Vec);
-  ASMJIT_INST_2x(saddlv, Saddlv_v, Vec, Vec);
-  ASMJIT_INST_3x(saddw, Saddw_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(saddw2, Saddw2_v, Vec, Vec, Vec);
-  ASMJIT_INST_2x(scvtf, Scvtf_v, Vec, Gp);
-  ASMJIT_INST_3x(scvtf, Scvtf_v, Vec, Gp, Imm);
-  ASMJIT_INST_2x(scvtf, Scvtf_v, Vec, Vec);
-  ASMJIT_INST_3x(scvtf, Scvtf_v, Vec, Vec, Imm);
-  ASMJIT_INST_3x(shadd, Shadd_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(shl, Shl_v, Vec, Vec, Imm);
-  ASMJIT_INST_3x(shll, Shll_v, Vec, Vec, Imm);
-  ASMJIT_INST_3x(shll2, Shll2_v, Vec, Vec, Imm);
-  ASMJIT_INST_3x(shrn, Shrn_v, Vec, Vec, Imm);
-  ASMJIT_INST_3x(shrn2, Shrn2_v, Vec, Vec, Imm);
-  ASMJIT_INST_3x(shsub, Shsub_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(sli, Sli_v, Vec, Vec, Imm);
-  ASMJIT_INST_3x(smax, Smax_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(smaxp, Smaxp_v, Vec, Vec, Vec);
-  ASMJIT_INST_2x(smaxv, Smaxv_v, Vec, Vec);
-  ASMJIT_INST_3x(smin, Smin_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(sminp, Sminp_v, Vec, Vec, Vec);
-  ASMJIT_INST_2x(sminv, Sminv_v, Vec, Vec);
-  ASMJIT_INST_3x(smlal, Smlal_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(smlal2, Smlal2_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(smlsl, Smlsl_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(smlsl2, Smlsl2_v, Vec, Vec, Vec);
-  ASMJIT_INST_2x(smov, Smov_v, Gp, Vec);
-  ASMJIT_INST_3x(smull, Smull_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(smull2, Smull2_v, Vec, Vec, Vec);
-  ASMJIT_INST_2x(sqabs, Sqabs_v, Vec, Vec);
-  ASMJIT_INST_3x(sqadd, Sqadd_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(sqdmlal, Sqdmlal_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(sqdmlal2, Sqdmlal2_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(sqdmlsl, Sqdmlsl_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(sqdmlsl2, Sqdmlsl2_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(sqdmulh, Sqdmulh_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(sqdmull, Sqdmull_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(sqdmull2, Sqdmull2_v, Vec, Vec, Vec);
-  ASMJIT_INST_2x(sqneg, Sqneg_v, Vec, Vec);
-  ASMJIT_INST_3x(sqrdmulh, Sqrdmulh_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(sqrshl, Sqrshl_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(sqrshrn, Sqrshrn_v, Vec, Vec, Imm);
-  ASMJIT_INST_3x(sqrshrn2, Sqrshrn2_v, Vec, Vec, Imm);
-  ASMJIT_INST_3x(sqrshrun, Sqrshrun_v, Vec, Vec, Imm);
-  ASMJIT_INST_3x(sqrshrun2, Sqrshrun2_v, Vec, Vec, Imm);
-  ASMJIT_INST_3x(sqshl, Sqshl_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(sqshl, Sqshl_v, Vec, Vec, Imm);
-  ASMJIT_INST_3x(sqshlu, Sqshlu_v, Vec, Vec, Imm);
-  ASMJIT_INST_3x(sqshrn, Sqshrn_v, Vec, Vec, Imm);
-  ASMJIT_INST_3x(sqshrn2, Sqshrn2_v, Vec, Vec, Imm);
-  ASMJIT_INST_3x(sqshrun, Sqshrun_v, Vec, Vec, Imm);
-  ASMJIT_INST_3x(sqshrun2, Sqshrun2_v, Vec, Vec, Imm);
-  ASMJIT_INST_3x(sqsub, Sqsub_v, Vec, Vec, Vec);
-  ASMJIT_INST_2x(sqxtn, Sqxtn_v, Vec, Vec);
-  ASMJIT_INST_2x(sqxtn2, Sqxtn2_v, Vec, Vec);
-  ASMJIT_INST_2x(sqxtun, Sqxtun_v, Vec, Vec);
-  ASMJIT_INST_2x(sqxtun2, Sqxtun2_v, Vec, Vec);
-  ASMJIT_INST_3x(srhadd, Srhadd_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(sri, Sri_v, Vec, Vec, Imm);
-  ASMJIT_INST_3x(srshl, Srshl_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(srshr, Srshr_v, Vec, Vec, Imm);
-  ASMJIT_INST_3x(srsra, Srsra_v, Vec, Vec, Imm);
-  ASMJIT_INST_3x(sshl, Sshl_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(sshll, Sshll_v, Vec, Vec, Imm);
-  ASMJIT_INST_3x(sshll2, Sshll2_v, Vec, Vec, Imm);
-  ASMJIT_INST_3x(sshr, Sshr_v, Vec, Vec, Imm);
-  ASMJIT_INST_3x(ssra, Ssra_v, Vec, Vec, Imm);
-  ASMJIT_INST_3x(ssubl, Ssubl_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(ssubl2, Ssubl2_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(ssubw, Ssubw_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(ssubw2, Ssubw2_v, Vec, Vec, Vec);
-  ASMJIT_INST_2x(st1, St1_v, Vec, Mem);
-  ASMJIT_INST_3x(st1, St1_v, Vec, Vec, Mem);
-  ASMJIT_INST_4x(st1, St1_v, Vec, Vec, Vec, Mem);
-  ASMJIT_INST_5x(st1, St1_v, Vec, Vec, Vec, Vec, Mem);
-  ASMJIT_INST_3x(st2, St2_v, Vec, Vec, Mem);
-  ASMJIT_INST_4x(st3, St3_v, Vec, Vec, Vec, Mem);
-  ASMJIT_INST_5x(st4, St4_v, Vec, Vec, Vec, Vec, Mem);
-  ASMJIT_INST_3x(stnp, Stnp_v, Vec, Vec, Mem);
-  ASMJIT_INST_3x(stp, Stp_v, Vec, Vec, Mem);
-  ASMJIT_INST_2x(str, Str_v, Vec, Mem);
-  ASMJIT_INST_2x(stur, Stur_v, Vec, Mem);
-  ASMJIT_INST_3x(sub, Sub_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(subhn, Subhn_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(subhn2, Subhn2_v, Vec, Vec, Vec);
-  ASMJIT_INST_2x(suqadd, Suqadd_v, Vec, Vec);
-  ASMJIT_INST_2x(sxtl, Sxtl_v, Vec, Vec);
-  ASMJIT_INST_2x(sxtl2, Sxtl2_v, Vec, Vec);
-  ASMJIT_INST_3x(tbl, Tbl_v, Vec, Vec, Vec);
-  ASMJIT_INST_4x(tbl, Tbl_v, Vec, Vec, Vec, Vec);
-  ASMJIT_INST_5x(tbl, Tbl_v, Vec, Vec, Vec, Vec, Vec);
-  ASMJIT_INST_6x(tbl, Tbl_v, Vec, Vec, Vec, Vec, Vec, Vec);
-  ASMJIT_INST_3x(tbx, Tbx_v, Vec, Vec, Vec);
-  ASMJIT_INST_4x(tbx, Tbx_v, Vec, Vec, Vec, Vec);
-  ASMJIT_INST_5x(tbx, Tbx_v, Vec, Vec, Vec, Vec, Vec);
-  ASMJIT_INST_6x(tbx, Tbx_v, Vec, Vec, Vec, Vec, Vec, Vec);
-  ASMJIT_INST_3x(trn1, Trn1_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(trn2, Trn2_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(uaba, Uaba_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(uabal, Uabal_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(uabal2, Uabal2_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(uabd, Uabd_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(uabdl, Uabdl_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(uabdl2, Uabdl2_v, Vec, Vec, Vec);
-  ASMJIT_INST_2x(uadalp, Uadalp_v, Vec, Vec);
-  ASMJIT_INST_3x(uaddl, Uaddl_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(uaddl2, Uaddl2_v, Vec, Vec, Vec);
-  ASMJIT_INST_2x(uaddlp, Uaddlp_v, Vec, Vec);
-  ASMJIT_INST_2x(uaddlv, Uaddlv_v, Vec, Vec);
-  ASMJIT_INST_3x(uaddw, Uaddw_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(uaddw2, Uaddw2_v, Vec, Vec, Vec);
-  ASMJIT_INST_2x(ucvtf, Ucvtf_v, Vec, Gp);
-  ASMJIT_INST_3x(ucvtf, Ucvtf_v, Vec, Gp, Imm);
-  ASMJIT_INST_2x(ucvtf, Ucvtf_v, Vec, Vec);
-  ASMJIT_INST_3x(ucvtf, Ucvtf_v, Vec, Vec, Imm);
-  ASMJIT_INST_3x(uhadd, Uhadd_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(uhsub, Uhsub_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(umax, Umax_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(umaxp, Umaxp_v, Vec, Vec, Vec);
-  ASMJIT_INST_2x(umaxv, Umaxv_v, Vec, Vec);
-  ASMJIT_INST_3x(umin, Umin_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(uminp, Uminp_v, Vec, Vec, Vec);
-  ASMJIT_INST_2x(uminv, Uminv_v, Vec, Vec);
-  ASMJIT_INST_3x(umlal, Umlal_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(umlal2, Umlal2_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(umlsl, Umlsl_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(umlsl2, Umlsl2_v, Vec, Vec, Vec);
-  ASMJIT_INST_2x(umov, Umov_v, Gp, Vec);
-  ASMJIT_INST_3x(umull, Umull_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(umull2, Umull2_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(uqadd, Uqadd_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(uqrshl, Uqrshl_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(uqrshl, Uqrshl_v, Vec, Vec, Imm);
-  ASMJIT_INST_3x(uqrshrn, Uqrshrn_v, Vec, Vec, Imm);
-  ASMJIT_INST_3x(uqrshrn2, Uqrshrn2_v, Vec, Vec, Imm);
-  ASMJIT_INST_3x(uqshl, Uqshl_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(uqshl, Uqshl_v, Vec, Vec, Imm);
-  ASMJIT_INST_3x(uqshrn, Uqshrn_v, Vec, Vec, Imm);
-  ASMJIT_INST_3x(uqshrn2, Uqshrn2_v, Vec, Vec, Imm);
-  ASMJIT_INST_3x(uqsub, Uqsub_v, Vec, Vec, Vec);
-  ASMJIT_INST_2x(uqxtn, Uqxtn_v, Vec, Vec);
-  ASMJIT_INST_2x(uqxtn2, Uqxtn2_v, Vec, Vec);
-  ASMJIT_INST_2x(urecpe, Urecpe_v, Vec, Vec);
-  ASMJIT_INST_3x(urhadd, Urhadd_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(urshl, Urshl_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(urshr, Urshr_v, Vec, Vec, Imm);
-  ASMJIT_INST_2x(ursqrte, Ursqrte_v, Vec, Vec);
-  ASMJIT_INST_3x(ursra, Ursra_v, Vec, Vec, Imm);
-  ASMJIT_INST_3x(ushl, Ushl_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(ushll, Ushll_v, Vec, Vec, Imm);
-  ASMJIT_INST_3x(ushll2, Ushll2_v, Vec, Vec, Imm);
-  ASMJIT_INST_3x(ushr, Ushr_v, Vec, Vec, Imm);
-  ASMJIT_INST_2x(usqadd, Usqadd_v, Vec, Vec);
-  ASMJIT_INST_3x(usra, Usra_v, Vec, Vec, Imm);
-  ASMJIT_INST_3x(usubl, Usubl_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(usubl2, Usubl2_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(usubw, Usubw_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(usubw2, Usubw2_v, Vec, Vec, Vec);
-  ASMJIT_INST_2x(uxtl, Uxtl_v, Vec, Vec);
-  ASMJIT_INST_2x(uxtl2, Uxtl2_v, Vec, Vec);
-  ASMJIT_INST_3x(uzp1, Uzp1_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(uzp2, Uzp2_v, Vec, Vec, Vec);
-  ASMJIT_INST_2x(xtn, Xtn_v, Vec, Vec);
-  ASMJIT_INST_2x(xtn2, Xtn2_v, Vec, Vec);
-  ASMJIT_INST_3x(zip1, Zip1_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(zip2, Zip2_v, Vec, Vec, Vec);
+  ASMJIT_INST_2x(abs, Abs_v, Vec, Vec)
+  ASMJIT_INST_3x(add, Add_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(addhn, Addhn_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(addhn2, Addhn2_v, Vec, Vec, Vec)
+  ASMJIT_INST_2x(addp, Addp_v, Vec, Vec)
+  ASMJIT_INST_3x(addp, Addp_v, Vec, Vec, Vec)
+  ASMJIT_INST_2x(addv, Addv_v, Vec, Vec)
+  ASMJIT_INST_3x(and_, And_v, Vec, Vec, Vec)
+  ASMJIT_INST_2x(bic, Bic_v, Vec, Imm)
+  ASMJIT_INST_3x(bic, Bic_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(bic, Bic_v, Vec, Imm, Imm)
+  ASMJIT_INST_3x(bif, Bif_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(bit, Bit_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(bsl, Bsl_v, Vec, Vec, Vec)
+  ASMJIT_INST_2x(cls, Cls_v, Vec, Vec)
+  ASMJIT_INST_2x(clz, Clz_v, Vec, Vec)
+  ASMJIT_INST_3x(cmeq, Cmeq_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(cmeq, Cmeq_v, Vec, Vec, Imm)
+  ASMJIT_INST_3x(cmge, Cmge_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(cmge, Cmge_v, Vec, Vec, Imm)
+  ASMJIT_INST_3x(cmgt, Cmgt_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(cmgt, Cmgt_v, Vec, Vec, Imm)
+  ASMJIT_INST_3x(cmhi, Cmhi_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(cmhs, Cmhs_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(cmle, Cmle_v, Vec, Vec, Imm)
+  ASMJIT_INST_3x(cmlt, Cmlt_v, Vec, Vec, Imm)
+  ASMJIT_INST_3x(cmtst, Cmtst_v, Vec, Vec, Vec)
+  ASMJIT_INST_2x(cnt, Cnt_v, Vec, Vec)
+  ASMJIT_INST_2x(dup, Dup_v, Vec, Gp)
+  ASMJIT_INST_2x(dup, Dup_v, Vec, Vec)
+  ASMJIT_INST_3x(eor, Eor_v, Vec, Vec, Vec)
+  ASMJIT_INST_4x(ext, Ext_v, Vec, Vec, Vec, Imm)
+  ASMJIT_INST_3x(fabd, Fabd_v, Vec, Vec, Vec)
+  ASMJIT_INST_2x(fabs, Fabs_v, Vec, Vec)
+  ASMJIT_INST_3x(facge, Facge_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(facgt, Facgt_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(fadd, Fadd_v, Vec, Vec, Vec)
+  ASMJIT_INST_2x(faddp, Faddp_v, Vec, Vec)
+  ASMJIT_INST_3x(faddp, Faddp_v, Vec, Vec, Vec)
+  ASMJIT_INST_4x(fccmp, Fccmp_v, Vec, Vec, Imm, Imm)
+  ASMJIT_INST_4x(fccmpe, Fccmpe_v, Vec, Vec, Imm, Imm)
+  ASMJIT_INST_3x(fcmeq, Fcmeq_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(fcmeq, Fcmeq_v, Vec, Vec, Imm)
+  ASMJIT_INST_3x(fcmge, Fcmge_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(fcmge, Fcmge_v, Vec, Vec, Imm)
+  ASMJIT_INST_3x(fcmgt, Fcmgt_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(fcmgt, Fcmgt_v, Vec, Vec, Imm)
+  ASMJIT_INST_3x(fcmle, Fcmle_v, Vec, Vec, Imm)
+  ASMJIT_INST_3x(fcmlt, Fcmlt_v, Vec, Vec, Imm)
+  ASMJIT_INST_2x(fcmp, Fcmp_v, Vec, Vec)
+  ASMJIT_INST_2x(fcmp, Fcmp_v, Vec, Imm)
+  ASMJIT_INST_2x(fcmpe, Fcmpe_v, Vec, Vec)
+  ASMJIT_INST_2x(fcmpe, Fcmpe_v, Vec, Imm)
+  ASMJIT_INST_4x(fcsel, Fcsel_v, Vec, Vec, Vec, Imm)
+  ASMJIT_INST_2x(fcvt, Fcvt_v, Vec, Vec)
+  ASMJIT_INST_2x(fcvtas, Fcvtas_v, Gp, Vec)
+  ASMJIT_INST_2x(fcvtas, Fcvtas_v, Vec, Vec)
+  ASMJIT_INST_2x(fcvtau, Fcvtau_v, Gp, Vec)
+  ASMJIT_INST_2x(fcvtau, Fcvtau_v, Vec, Vec)
+  ASMJIT_INST_2x(fcvtl, Fcvtl_v, Vec, Vec)
+  ASMJIT_INST_2x(fcvtl2, Fcvtl2_v, Vec, Vec)
+  ASMJIT_INST_2x(fcvtms, Fcvtms_v, Gp, Vec)
+  ASMJIT_INST_2x(fcvtms, Fcvtms_v, Vec, Vec)
+  ASMJIT_INST_2x(fcvtmu, Fcvtmu_v, Gp, Vec)
+  ASMJIT_INST_2x(fcvtmu, Fcvtmu_v, Vec, Vec)
+  ASMJIT_INST_2x(fcvtn, Fcvtn_v, Vec, Vec)
+  ASMJIT_INST_2x(fcvtn2, Fcvtn2_v, Vec, Vec)
+  ASMJIT_INST_2x(fcvtns, Fcvtns_v, Gp, Vec)
+  ASMJIT_INST_2x(fcvtns, Fcvtns_v, Vec, Vec)
+  ASMJIT_INST_2x(fcvtnu, Fcvtnu_v, Gp, Vec)
+  ASMJIT_INST_2x(fcvtnu, Fcvtnu_v, Vec, Vec)
+  ASMJIT_INST_2x(fcvtps, Fcvtps_v, Gp, Vec)
+  ASMJIT_INST_2x(fcvtps, Fcvtps_v, Vec, Vec)
+  ASMJIT_INST_2x(fcvtpu, Fcvtpu_v, Gp, Vec)
+  ASMJIT_INST_2x(fcvtpu, Fcvtpu_v, Vec, Vec)
+  ASMJIT_INST_2x(fcvtxn, Fcvtxn_v, Vec, Vec)
+  ASMJIT_INST_2x(fcvtxn2, Fcvtxn2_v, Vec, Vec)
+  ASMJIT_INST_2x(fcvtzs, Fcvtzs_v, Gp, Vec)
+  ASMJIT_INST_3x(fcvtzs, Fcvtzs_v, Gp, Vec, Imm)
+  ASMJIT_INST_2x(fcvtzs, Fcvtzs_v, Vec, Vec)
+  ASMJIT_INST_3x(fcvtzs, Fcvtzs_v, Vec, Vec, Imm)
+  ASMJIT_INST_2x(fcvtzu, Fcvtzu_v, Gp, Vec)
+  ASMJIT_INST_3x(fcvtzu, Fcvtzu_v, Gp, Vec, Imm)
+  ASMJIT_INST_2x(fcvtzu, Fcvtzu_v, Vec, Vec)
+  ASMJIT_INST_3x(fcvtzu, Fcvtzu_v, Vec, Vec, Imm)
+  ASMJIT_INST_3x(fdiv, Fdiv_v, Vec, Vec, Vec)
+  ASMJIT_INST_4x(fmadd, Fmadd_v, Vec, Vec, Vec, Vec)
+  ASMJIT_INST_3x(fmax, Fmax_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(fmaxnm, Fmaxnm_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(fmaxnmp, Fmaxnmp_v, Vec, Vec, Vec)
+  ASMJIT_INST_2x(fmaxnmp, Fmaxnmp_v, Vec, Vec)
+  ASMJIT_INST_2x(fmaxnmv, Fmaxnmv_v, Vec, Vec)
+  ASMJIT_INST_3x(fmaxp, Fmaxp_v, Vec, Vec, Vec)
+  ASMJIT_INST_2x(fmaxp, Fmaxp_v, Vec, Vec)
+  ASMJIT_INST_2x(fmaxv, Fmaxv_v, Vec, Vec)
+  ASMJIT_INST_3x(fmin, Fmin_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(fminnm, Fminnm_v, Vec, Vec, Vec)
+  ASMJIT_INST_2x(fminnmv, Fminnmv_v, Vec, Vec)
+  ASMJIT_INST_3x(fminnmp, Fminnmp_v, Vec, Vec, Vec)
+  ASMJIT_INST_2x(fminnmp, Fminnmp_v, Vec, Vec)
+  ASMJIT_INST_2x(fminp, Fminp_v, Vec, Vec)
+  ASMJIT_INST_3x(fminp, Fminp_v, Vec, Vec, Vec)
+  ASMJIT_INST_2x(fminv, Fminv_v, Vec, Vec)
+  ASMJIT_INST_3x(fmla, Fmla_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(fmls, Fmls_v, Vec, Vec, Vec)
+  ASMJIT_INST_2x(fmov, Fmov_v, Gp, Vec)
+  ASMJIT_INST_2x(fmov, Fmov_v, Vec, Gp)
+  ASMJIT_INST_2x(fmov, Fmov_v, Vec, Vec)
+  ASMJIT_INST_2x(fmov, Fmov_v, Vec, Imm)
+  ASMJIT_INST_4x(fmsub, Fmsub_v, Vec, Vec, Vec, Vec)
+  ASMJIT_INST_3x(fmul, Fmul_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(fmulx, Fmulx_v, Vec, Vec, Vec)
+  ASMJIT_INST_2x(fneg, Fneg_v, Vec, Vec)
+  ASMJIT_INST_4x(fnmadd, Fnmadd_v, Vec, Vec, Vec, Vec)
+  ASMJIT_INST_4x(fnmsub, Fnmsub_v, Vec, Vec, Vec, Vec)
+  ASMJIT_INST_3x(fnmul, Fnmul_v, Vec, Vec, Vec)
+  ASMJIT_INST_2x(frecpe, Frecpe_v, Vec, Vec)
+  ASMJIT_INST_3x(frecps, Frecps_v, Vec, Vec, Vec)
+  ASMJIT_INST_2x(frecpx, Frecpx_v, Vec, Vec)
+  ASMJIT_INST_2x(frint32x, Frint32x_v, Vec, Vec)
+  ASMJIT_INST_2x(frint32z, Frint32z_v, Vec, Vec)
+  ASMJIT_INST_2x(frint64x, Frint64x_v, Vec, Vec)
+  ASMJIT_INST_2x(frint64z, Frint64z_v, Vec, Vec)
+  ASMJIT_INST_2x(frinta, Frinta_v, Vec, Vec)
+  ASMJIT_INST_2x(frinti, Frinti_v, Vec, Vec)
+  ASMJIT_INST_2x(frintm, Frintm_v, Vec, Vec)
+  ASMJIT_INST_2x(frintn, Frintn_v, Vec, Vec)
+  ASMJIT_INST_2x(frintp, Frintp_v, Vec, Vec)
+  ASMJIT_INST_2x(frintx, Frintx_v, Vec, Vec)
+  ASMJIT_INST_2x(frintz, Frintz_v, Vec, Vec)
+  ASMJIT_INST_2x(frsqrte, Frsqrte_v, Vec, Vec)
+  ASMJIT_INST_3x(frsqrts, Frsqrts_v, Vec, Vec, Vec)
+  ASMJIT_INST_2x(fsqrt, Fsqrt_v, Vec, Vec)
+  ASMJIT_INST_3x(fsub, Fsub_v, Vec, Vec, Vec)
+  ASMJIT_INST_2x(ins, Ins_v, Vec, Gp)
+  ASMJIT_INST_2x(ins, Ins_v, Vec, Vec)
+  ASMJIT_INST_2x(ld1, Ld1_v, Vec, Mem)
+  ASMJIT_INST_3x(ld1, Ld1_v, Vec, Vec, Mem)
+  ASMJIT_INST_4x(ld1, Ld1_v, Vec, Vec, Vec, Mem)
+  ASMJIT_INST_5x(ld1, Ld1_v, Vec, Vec, Vec, Vec, Mem)
+  ASMJIT_INST_2x(ld1r, Ld1r_v, Vec, Mem)
+  ASMJIT_INST_3x(ld2, Ld2_v, Vec, Vec, Mem)
+  ASMJIT_INST_3x(ld2r, Ld2r_v, Vec, Vec, Mem)
+  ASMJIT_INST_4x(ld3, Ld3_v, Vec, Vec, Vec, Mem)
+  ASMJIT_INST_4x(ld3r, Ld3r_v, Vec, Vec, Vec, Mem)
+  ASMJIT_INST_5x(ld4, Ld4_v, Vec, Vec, Vec, Vec, Mem)
+  ASMJIT_INST_5x(ld4r, Ld4r_v, Vec, Vec, Vec, Vec, Mem)
+  ASMJIT_INST_3x(ldnp, Ldnp_v, Vec, Vec, Mem)
+  ASMJIT_INST_3x(ldp, Ldp_v, Vec, Vec, Mem)
+  ASMJIT_INST_2x(ldr, Ldr_v, Vec, Mem)
+  ASMJIT_INST_2x(ldur, Ldur_v, Vec, Mem)
+  ASMJIT_INST_3x(mla, Mla_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(mls, Mls_v, Vec, Vec, Vec)
+  ASMJIT_INST_2x(mov, Mov_v, Vec, Vec)
+  ASMJIT_INST_2x(mov, Mov_v, Gp, Vec)
+  ASMJIT_INST_2x(mov, Mov_v, Vec, Gp)
+  ASMJIT_INST_2x(movi, Movi_v, Vec, Imm)
+  ASMJIT_INST_3x(movi, Movi_v, Vec, Imm, Imm)
+  ASMJIT_INST_3x(mul, Mul_v, Vec, Vec, Vec)
+  ASMJIT_INST_2x(mvn, Mvn_v, Vec, Vec)
+  ASMJIT_INST_2x(mvn_, Mvn_v, Vec, Vec)
+  ASMJIT_INST_2x(mvni, Mvni_v, Vec, Imm)
+  ASMJIT_INST_3x(mvni, Mvni_v, Vec, Imm, Imm)
+  ASMJIT_INST_2x(neg, Neg_v, Vec, Vec)
+  ASMJIT_INST_2x(not_, Not_v, Vec, Vec)
+  ASMJIT_INST_3x(orn, Orn_v, Vec, Vec, Vec)
+  ASMJIT_INST_2x(orr, Orr_v, Vec, Imm)
+  ASMJIT_INST_3x(orr, Orr_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(orr, Orr_v, Vec, Imm, Imm)
+  ASMJIT_INST_3x(pmul, Pmul_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(pmull, Pmull_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(pmull2, Pmull2_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(raddhn, Raddhn_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(raddhn2, Raddhn2_v, Vec, Vec, Vec)
+  ASMJIT_INST_2x(rbit, Rbit_v, Vec, Vec)
+  ASMJIT_INST_2x(rev16, Rev16_v, Vec, Vec)
+  ASMJIT_INST_2x(rev32, Rev32_v, Vec, Vec)
+  ASMJIT_INST_2x(rev64, Rev64_v, Vec, Vec)
+  ASMJIT_INST_3x(rshrn, Rshrn_v, Vec, Vec, Imm)
+  ASMJIT_INST_3x(rshrn2, Rshrn2_v, Vec, Vec, Imm)
+  ASMJIT_INST_3x(rsubhn, Rsubhn_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(rsubhn2, Rsubhn2_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(saba, Saba_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(sabal, Sabal_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(sabal2, Sabal2_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(sabd, Sabd_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(sabdl, Sabdl_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(sabdl2, Sabdl2_v, Vec, Vec, Vec)
+  ASMJIT_INST_2x(sadalp, Sadalp_v, Vec, Vec)
+  ASMJIT_INST_3x(saddl, Saddl_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(saddl2, Saddl2_v, Vec, Vec, Vec)
+  ASMJIT_INST_2x(saddlp, Saddlp_v, Vec, Vec)
+  ASMJIT_INST_2x(saddlv, Saddlv_v, Vec, Vec)
+  ASMJIT_INST_3x(saddw, Saddw_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(saddw2, Saddw2_v, Vec, Vec, Vec)
+  ASMJIT_INST_2x(scvtf, Scvtf_v, Vec, Gp)
+  ASMJIT_INST_3x(scvtf, Scvtf_v, Vec, Gp, Imm)
+  ASMJIT_INST_2x(scvtf, Scvtf_v, Vec, Vec)
+  ASMJIT_INST_3x(scvtf, Scvtf_v, Vec, Vec, Imm)
+  ASMJIT_INST_3x(shadd, Shadd_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(shl, Shl_v, Vec, Vec, Imm)
+  ASMJIT_INST_3x(shll, Shll_v, Vec, Vec, Imm)
+  ASMJIT_INST_3x(shll2, Shll2_v, Vec, Vec, Imm)
+  ASMJIT_INST_3x(shrn, Shrn_v, Vec, Vec, Imm)
+  ASMJIT_INST_3x(shrn2, Shrn2_v, Vec, Vec, Imm)
+  ASMJIT_INST_3x(shsub, Shsub_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(sli, Sli_v, Vec, Vec, Imm)
+  ASMJIT_INST_3x(smax, Smax_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(smaxp, Smaxp_v, Vec, Vec, Vec)
+  ASMJIT_INST_2x(smaxv, Smaxv_v, Vec, Vec)
+  ASMJIT_INST_3x(smin, Smin_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(sminp, Sminp_v, Vec, Vec, Vec)
+  ASMJIT_INST_2x(sminv, Sminv_v, Vec, Vec)
+  ASMJIT_INST_3x(smlal, Smlal_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(smlal2, Smlal2_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(smlsl, Smlsl_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(smlsl2, Smlsl2_v, Vec, Vec, Vec)
+  ASMJIT_INST_2x(smov, Smov_v, Gp, Vec)
+  ASMJIT_INST_3x(smull, Smull_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(smull2, Smull2_v, Vec, Vec, Vec)
+  ASMJIT_INST_2x(sqabs, Sqabs_v, Vec, Vec)
+  ASMJIT_INST_3x(sqadd, Sqadd_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(sqdmlal, Sqdmlal_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(sqdmlal2, Sqdmlal2_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(sqdmlsl, Sqdmlsl_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(sqdmlsl2, Sqdmlsl2_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(sqdmulh, Sqdmulh_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(sqdmull, Sqdmull_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(sqdmull2, Sqdmull2_v, Vec, Vec, Vec)
+  ASMJIT_INST_2x(sqneg, Sqneg_v, Vec, Vec)
+  ASMJIT_INST_3x(sqrdmulh, Sqrdmulh_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(sqrshl, Sqrshl_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(sqrshrn, Sqrshrn_v, Vec, Vec, Imm)
+  ASMJIT_INST_3x(sqrshrn2, Sqrshrn2_v, Vec, Vec, Imm)
+  ASMJIT_INST_3x(sqrshrun, Sqrshrun_v, Vec, Vec, Imm)
+  ASMJIT_INST_3x(sqrshrun2, Sqrshrun2_v, Vec, Vec, Imm)
+  ASMJIT_INST_3x(sqshl, Sqshl_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(sqshl, Sqshl_v, Vec, Vec, Imm)
+  ASMJIT_INST_3x(sqshlu, Sqshlu_v, Vec, Vec, Imm)
+  ASMJIT_INST_3x(sqshrn, Sqshrn_v, Vec, Vec, Imm)
+  ASMJIT_INST_3x(sqshrn2, Sqshrn2_v, Vec, Vec, Imm)
+  ASMJIT_INST_3x(sqshrun, Sqshrun_v, Vec, Vec, Imm)
+  ASMJIT_INST_3x(sqshrun2, Sqshrun2_v, Vec, Vec, Imm)
+  ASMJIT_INST_3x(sqsub, Sqsub_v, Vec, Vec, Vec)
+  ASMJIT_INST_2x(sqxtn, Sqxtn_v, Vec, Vec)
+  ASMJIT_INST_2x(sqxtn2, Sqxtn2_v, Vec, Vec)
+  ASMJIT_INST_2x(sqxtun, Sqxtun_v, Vec, Vec)
+  ASMJIT_INST_2x(sqxtun2, Sqxtun2_v, Vec, Vec)
+  ASMJIT_INST_3x(srhadd, Srhadd_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(sri, Sri_v, Vec, Vec, Imm)
+  ASMJIT_INST_3x(srshl, Srshl_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(srshr, Srshr_v, Vec, Vec, Imm)
+  ASMJIT_INST_3x(srsra, Srsra_v, Vec, Vec, Imm)
+  ASMJIT_INST_3x(sshl, Sshl_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(sshll, Sshll_v, Vec, Vec, Imm)
+  ASMJIT_INST_3x(sshll2, Sshll2_v, Vec, Vec, Imm)
+  ASMJIT_INST_3x(sshr, Sshr_v, Vec, Vec, Imm)
+  ASMJIT_INST_3x(ssra, Ssra_v, Vec, Vec, Imm)
+  ASMJIT_INST_3x(ssubl, Ssubl_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(ssubl2, Ssubl2_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(ssubw, Ssubw_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(ssubw2, Ssubw2_v, Vec, Vec, Vec)
+  ASMJIT_INST_2x(st1, St1_v, Vec, Mem)
+  ASMJIT_INST_3x(st1, St1_v, Vec, Vec, Mem)
+  ASMJIT_INST_4x(st1, St1_v, Vec, Vec, Vec, Mem)
+  ASMJIT_INST_5x(st1, St1_v, Vec, Vec, Vec, Vec, Mem)
+  ASMJIT_INST_3x(st2, St2_v, Vec, Vec, Mem)
+  ASMJIT_INST_4x(st3, St3_v, Vec, Vec, Vec, Mem)
+  ASMJIT_INST_5x(st4, St4_v, Vec, Vec, Vec, Vec, Mem)
+  ASMJIT_INST_3x(stnp, Stnp_v, Vec, Vec, Mem)
+  ASMJIT_INST_3x(stp, Stp_v, Vec, Vec, Mem)
+  ASMJIT_INST_2x(str, Str_v, Vec, Mem)
+  ASMJIT_INST_2x(stur, Stur_v, Vec, Mem)
+  ASMJIT_INST_3x(sub, Sub_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(subhn, Subhn_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(subhn2, Subhn2_v, Vec, Vec, Vec)
+  ASMJIT_INST_2x(suqadd, Suqadd_v, Vec, Vec)
+  ASMJIT_INST_2x(sxtl, Sxtl_v, Vec, Vec)
+  ASMJIT_INST_2x(sxtl2, Sxtl2_v, Vec, Vec)
+  ASMJIT_INST_3x(tbl, Tbl_v, Vec, Vec, Vec)
+  ASMJIT_INST_4x(tbl, Tbl_v, Vec, Vec, Vec, Vec)
+  ASMJIT_INST_5x(tbl, Tbl_v, Vec, Vec, Vec, Vec, Vec)
+  ASMJIT_INST_6x(tbl, Tbl_v, Vec, Vec, Vec, Vec, Vec, Vec)
+  ASMJIT_INST_3x(tbx, Tbx_v, Vec, Vec, Vec)
+  ASMJIT_INST_4x(tbx, Tbx_v, Vec, Vec, Vec, Vec)
+  ASMJIT_INST_5x(tbx, Tbx_v, Vec, Vec, Vec, Vec, Vec)
+  ASMJIT_INST_6x(tbx, Tbx_v, Vec, Vec, Vec, Vec, Vec, Vec)
+  ASMJIT_INST_3x(trn1, Trn1_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(trn2, Trn2_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(uaba, Uaba_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(uabal, Uabal_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(uabal2, Uabal2_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(uabd, Uabd_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(uabdl, Uabdl_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(uabdl2, Uabdl2_v, Vec, Vec, Vec)
+  ASMJIT_INST_2x(uadalp, Uadalp_v, Vec, Vec)
+  ASMJIT_INST_3x(uaddl, Uaddl_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(uaddl2, Uaddl2_v, Vec, Vec, Vec)
+  ASMJIT_INST_2x(uaddlp, Uaddlp_v, Vec, Vec)
+  ASMJIT_INST_2x(uaddlv, Uaddlv_v, Vec, Vec)
+  ASMJIT_INST_3x(uaddw, Uaddw_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(uaddw2, Uaddw2_v, Vec, Vec, Vec)
+  ASMJIT_INST_2x(ucvtf, Ucvtf_v, Vec, Gp)
+  ASMJIT_INST_3x(ucvtf, Ucvtf_v, Vec, Gp, Imm)
+  ASMJIT_INST_2x(ucvtf, Ucvtf_v, Vec, Vec)
+  ASMJIT_INST_3x(ucvtf, Ucvtf_v, Vec, Vec, Imm)
+  ASMJIT_INST_3x(uhadd, Uhadd_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(uhsub, Uhsub_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(umax, Umax_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(umaxp, Umaxp_v, Vec, Vec, Vec)
+  ASMJIT_INST_2x(umaxv, Umaxv_v, Vec, Vec)
+  ASMJIT_INST_3x(umin, Umin_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(uminp, Uminp_v, Vec, Vec, Vec)
+  ASMJIT_INST_2x(uminv, Uminv_v, Vec, Vec)
+  ASMJIT_INST_3x(umlal, Umlal_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(umlal2, Umlal2_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(umlsl, Umlsl_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(umlsl2, Umlsl2_v, Vec, Vec, Vec)
+  ASMJIT_INST_2x(umov, Umov_v, Gp, Vec)
+  ASMJIT_INST_3x(umull, Umull_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(umull2, Umull2_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(uqadd, Uqadd_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(uqrshl, Uqrshl_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(uqrshl, Uqrshl_v, Vec, Vec, Imm)
+  ASMJIT_INST_3x(uqrshrn, Uqrshrn_v, Vec, Vec, Imm)
+  ASMJIT_INST_3x(uqrshrn2, Uqrshrn2_v, Vec, Vec, Imm)
+  ASMJIT_INST_3x(uqshl, Uqshl_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(uqshl, Uqshl_v, Vec, Vec, Imm)
+  ASMJIT_INST_3x(uqshrn, Uqshrn_v, Vec, Vec, Imm)
+  ASMJIT_INST_3x(uqshrn2, Uqshrn2_v, Vec, Vec, Imm)
+  ASMJIT_INST_3x(uqsub, Uqsub_v, Vec, Vec, Vec)
+  ASMJIT_INST_2x(uqxtn, Uqxtn_v, Vec, Vec)
+  ASMJIT_INST_2x(uqxtn2, Uqxtn2_v, Vec, Vec)
+  ASMJIT_INST_2x(urecpe, Urecpe_v, Vec, Vec)
+  ASMJIT_INST_3x(urhadd, Urhadd_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(urshl, Urshl_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(urshr, Urshr_v, Vec, Vec, Imm)
+  ASMJIT_INST_2x(ursqrte, Ursqrte_v, Vec, Vec)
+  ASMJIT_INST_3x(ursra, Ursra_v, Vec, Vec, Imm)
+  ASMJIT_INST_3x(ushl, Ushl_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(ushll, Ushll_v, Vec, Vec, Imm)
+  ASMJIT_INST_3x(ushll2, Ushll2_v, Vec, Vec, Imm)
+  ASMJIT_INST_3x(ushr, Ushr_v, Vec, Vec, Imm)
+  ASMJIT_INST_2x(usqadd, Usqadd_v, Vec, Vec)
+  ASMJIT_INST_3x(usra, Usra_v, Vec, Vec, Imm)
+  ASMJIT_INST_3x(usubl, Usubl_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(usubl2, Usubl2_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(usubw, Usubw_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(usubw2, Usubw2_v, Vec, Vec, Vec)
+  ASMJIT_INST_2x(uxtl, Uxtl_v, Vec, Vec)
+  ASMJIT_INST_2x(uxtl2, Uxtl2_v, Vec, Vec)
+  ASMJIT_INST_3x(uzp1, Uzp1_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(uzp2, Uzp2_v, Vec, Vec, Vec)
+  ASMJIT_INST_2x(xtn, Xtn_v, Vec, Vec)
+  ASMJIT_INST_2x(xtn2, Xtn2_v, Vec, Vec)
+  ASMJIT_INST_3x(zip1, Zip1_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(zip2, Zip2_v, Vec, Vec, Vec)
 
   //! \}
 
   //! \name AES Instructions
   //! \{
 
-  ASMJIT_INST_2x(aesd, Aesd_v, Vec, Vec);
-  ASMJIT_INST_2x(aese, Aese_v, Vec, Vec);
-  ASMJIT_INST_2x(aesimc, Aesimc_v, Vec, Vec);
-  ASMJIT_INST_2x(aesmc, Aesmc_v, Vec, Vec);
+  ASMJIT_INST_2x(aesd, Aesd_v, Vec, Vec)
+  ASMJIT_INST_2x(aese, Aese_v, Vec, Vec)
+  ASMJIT_INST_2x(aesimc, Aesimc_v, Vec, Vec)
+  ASMJIT_INST_2x(aesmc, Aesmc_v, Vec, Vec)
 
   //! \}
 
   //! \name SHA1 Instructions
   //! \{
 
-  ASMJIT_INST_3x(sha1c, Sha1c_v, Vec, Vec, Vec);
-  ASMJIT_INST_2x(sha1h, Sha1h_v, Vec, Vec);
-  ASMJIT_INST_3x(sha1m, Sha1m_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(sha1p, Sha1p_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(sha1su0, Sha1su0_v, Vec, Vec, Vec);
-  ASMJIT_INST_2x(sha1su1, Sha1su1_v, Vec, Vec);
+  ASMJIT_INST_3x(sha1c, Sha1c_v, Vec, Vec, Vec)
+  ASMJIT_INST_2x(sha1h, Sha1h_v, Vec, Vec)
+  ASMJIT_INST_3x(sha1m, Sha1m_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(sha1p, Sha1p_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(sha1su0, Sha1su0_v, Vec, Vec, Vec)
+  ASMJIT_INST_2x(sha1su1, Sha1su1_v, Vec, Vec)
 
   //! \}
 
   //! \name SHA2 Instructions
   //! \{
 
-  ASMJIT_INST_3x(sha256h, Sha256h_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(sha256h2, Sha256h2_v, Vec, Vec, Vec);
-  ASMJIT_INST_2x(sha256su0, Sha256su0_v, Vec, Vec);
-  ASMJIT_INST_3x(sha256su1, Sha256su1_v, Vec, Vec, Vec);
+  ASMJIT_INST_3x(sha256h, Sha256h_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(sha256h2, Sha256h2_v, Vec, Vec, Vec)
+  ASMJIT_INST_2x(sha256su0, Sha256su0_v, Vec, Vec)
+  ASMJIT_INST_3x(sha256su1, Sha256su1_v, Vec, Vec, Vec)
 
   //! \}
 
   //! \name RDMA Instructions (ARMv8.1-A)
   //! \{
 
-  ASMJIT_INST_3x(sqrdmlah, Sqrdmlah_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(sqrdmlsh, Sqrdmlsh_v, Vec, Vec, Vec);
+  ASMJIT_INST_3x(sqrdmlah, Sqrdmlah_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(sqrdmlsh, Sqrdmlsh_v, Vec, Vec, Vec)
 
   //! \}
 
   //! \name FCMA Instruction (ARMv8.3-A)
   //! \{
 
-  ASMJIT_INST_4x(fcadd, Fcadd_v, Vec, Vec, Vec, Imm);
-  ASMJIT_INST_4x(fcmla, Fcmla_v, Vec, Vec, Vec, Imm);
+  ASMJIT_INST_4x(fcadd, Fcadd_v, Vec, Vec, Vec, Imm)
+  ASMJIT_INST_4x(fcmla, Fcmla_v, Vec, Vec, Vec, Imm)
 
   //! \}
 
   //! \name JSCVT Instruction (ARMv8.3-A)
   //! \{
 
-  ASMJIT_INST_2x(fjcvtzs, Fjcvtzs_v, Gp, Vec);
+  ASMJIT_INST_2x(fjcvtzs, Fjcvtzs_v, Gp, Vec)
 
   //! \}
 
   //! \name FHM Instructions
   //! \{
 
-  ASMJIT_INST_3x(fmlal, Fmlal_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(fmlal2, Fmlal2_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(fmlsl, Fmlsl_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(fmlsl2, Fmlsl2_v, Vec, Vec, Vec);
+  ASMJIT_INST_3x(fmlal, Fmlal_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(fmlal2, Fmlal2_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(fmlsl, Fmlsl_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(fmlsl2, Fmlsl2_v, Vec, Vec, Vec)
 
 
   //! \}
@@ -1198,73 +1198,73 @@ struct EmitterExplicitT {
   //! \name SHA3 Instructions (ARMv8.4-A, optional in ARMv8.2-A)
   //! \{
 
-  ASMJIT_INST_4x(bcax, Bcax_v, Vec, Vec, Vec, Vec);
-  ASMJIT_INST_4x(eor3, Eor3_v, Vec, Vec, Vec, Vec);
-  ASMJIT_INST_3x(rax1, Rax1_v, Vec, Vec, Vec);
-  ASMJIT_INST_4x(xar, Xar_v, Vec, Vec, Vec, Imm);
+  ASMJIT_INST_4x(bcax, Bcax_v, Vec, Vec, Vec, Vec)
+  ASMJIT_INST_4x(eor3, Eor3_v, Vec, Vec, Vec, Vec)
+  ASMJIT_INST_3x(rax1, Rax1_v, Vec, Vec, Vec)
+  ASMJIT_INST_4x(xar, Xar_v, Vec, Vec, Vec, Imm)
 
   //! \}
 
   //! \name SHA512 Instructions (ARMv8.4-A)
   //! \{
 
-  ASMJIT_INST_3x(sha512h, Sha512h_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(sha512h2, Sha512h2_v, Vec, Vec, Vec);
-  ASMJIT_INST_2x(sha512su0, Sha512su0_v, Vec, Vec);
-  ASMJIT_INST_3x(sha512su1, Sha512su1_v, Vec, Vec, Vec);
+  ASMJIT_INST_3x(sha512h, Sha512h_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(sha512h2, Sha512h2_v, Vec, Vec, Vec)
+  ASMJIT_INST_2x(sha512su0, Sha512su0_v, Vec, Vec)
+  ASMJIT_INST_3x(sha512su1, Sha512su1_v, Vec, Vec, Vec)
 
   //! \}
 
   //! \name SM3 Instructions (ARMv8.4-A)
   //! \{
 
-  ASMJIT_INST_3x(sm3partw1, Sm3partw1_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(sm3partw2, Sm3partw2_v, Vec, Vec, Vec);
-  ASMJIT_INST_4x(sm3ss1, Sm3ss1_v, Vec, Vec, Vec, Vec);
-  ASMJIT_INST_3x(sm3tt1a, Sm3tt1a_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(sm3tt1b, Sm3tt1b_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(sm3tt2a, Sm3tt2a_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(sm3tt2b, Sm3tt2b_v, Vec, Vec, Vec);
+  ASMJIT_INST_3x(sm3partw1, Sm3partw1_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(sm3partw2, Sm3partw2_v, Vec, Vec, Vec)
+  ASMJIT_INST_4x(sm3ss1, Sm3ss1_v, Vec, Vec, Vec, Vec)
+  ASMJIT_INST_3x(sm3tt1a, Sm3tt1a_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(sm3tt1b, Sm3tt1b_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(sm3tt2a, Sm3tt2a_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(sm3tt2b, Sm3tt2b_v, Vec, Vec, Vec)
 
   //! \}
 
   //! \name SM4 Instructions (ARMv8.4-A)
   //! \{
 
-  ASMJIT_INST_2x(sm4e, Sm4e_v, Vec, Vec);
-  ASMJIT_INST_3x(sm4ekey, Sm4ekey_v, Vec, Vec, Vec);
+  ASMJIT_INST_2x(sm4e, Sm4e_v, Vec, Vec)
+  ASMJIT_INST_3x(sm4ekey, Sm4ekey_v, Vec, Vec, Vec)
 
   //! \}
 
   //! \name DOTPROD Instructions (ARMv8.4-A, optional in ARMv8.2-A)
   //! \{
 
-  ASMJIT_INST_3x(sdot, Sdot_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(udot, Udot_v, Vec, Vec, Vec);
+  ASMJIT_INST_3x(sdot, Sdot_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(udot, Udot_v, Vec, Vec, Vec)
 
   //! \}
 
   //! \name BF16 Instructions (ARMv8.6-A)
   //! \{
 
-  ASMJIT_INST_2x(bfcvt, Bfcvt_v, Vec, Vec);
-  ASMJIT_INST_2x(bfcvtn, Bfcvtn_v, Vec, Vec);
-  ASMJIT_INST_2x(bfcvtn2, Bfcvtn2_v, Vec, Vec);
-  ASMJIT_INST_3x(bfmlalb, Bfmlalb_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(bfmlalt, Bfmlalt_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(bfmmla, Bfmmla_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(bfdot, Bfdot_v, Vec, Vec, Vec);
+  ASMJIT_INST_2x(bfcvt, Bfcvt_v, Vec, Vec)
+  ASMJIT_INST_2x(bfcvtn, Bfcvtn_v, Vec, Vec)
+  ASMJIT_INST_2x(bfcvtn2, Bfcvtn2_v, Vec, Vec)
+  ASMJIT_INST_3x(bfmlalb, Bfmlalb_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(bfmlalt, Bfmlalt_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(bfmmla, Bfmmla_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(bfdot, Bfdot_v, Vec, Vec, Vec)
 
   //! \}
 
   //! \name I8MM Instructions (ARMv8.6-A)
   //! \{
 
-  ASMJIT_INST_3x(smmla, Smmla_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(sudot, Sudot_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(ummla, Ummla_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(usdot, Usdot_v, Vec, Vec, Vec);
-  ASMJIT_INST_3x(usmmla, Usmmla_v, Vec, Vec, Vec);
+  ASMJIT_INST_3x(smmla, Smmla_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(sudot, Sudot_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(ummla, Ummla_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(usdot, Usdot_v, Vec, Vec, Vec)
+  ASMJIT_INST_3x(usmmla, Usmmla_v, Vec, Vec, Vec)
 
   //! \}
 };
diff --git a/src/asmjit/x86/x86emitter.h b/src/asmjit/x86/x86emitter.h
index 06bae27..1c018af 100644
--- a/src/asmjit/x86/x86emitter.h
+++ b/src/asmjit/x86/x86emitter.h
@@ -450,8 +450,8 @@ public:
   ASMJIT_INST_2x(cmps, Cmps, DS_ZSI, ES_ZDI)                           // ANY [EXPLICIT]
   ASMJIT_INST_3x(cmpxchg, Cmpxchg, Gp, Gp, Gp_ZAX)                     // I486 [EXPLICIT]
   ASMJIT_INST_3x(cmpxchg, Cmpxchg, Mem, Gp, Gp_ZAX)                    // I486 [EXPLICIT]
-  ASMJIT_INST_5x(cmpxchg16b, Cmpxchg16b, Mem, Gp_RDX, Gp_RAX, Gp_RCX, Gp_RBX); // CMPXCHG16B [EXPLICIT] m == EDX:EAX ? m <- ECX:EBX
-  ASMJIT_INST_5x(cmpxchg8b, Cmpxchg8b, Mem, Gp_EDX, Gp_EAX, Gp_ECX, Gp_EBX);   // CMPXCHG8B  [EXPLICIT] m == RDX:RAX ? m <- RCX:RBX
+  ASMJIT_INST_5x(cmpxchg16b, Cmpxchg16b, Mem, Gp_RDX, Gp_RAX, Gp_RCX, Gp_RBX) // CMPXCHG16B [EXPLICIT] m == EDX:EAX ? m <- ECX:EBX
+  ASMJIT_INST_5x(cmpxchg8b, Cmpxchg8b, Mem, Gp_EDX, Gp_EAX, Gp_ECX, Gp_EBX)   // CMPXCHG8B  [EXPLICIT] m == RDX:RAX ? m <- RCX:RBX
   ASMJIT_INST_1x(dec, Dec, Gp)                                         // ANY
   ASMJIT_INST_1x(dec, Dec, Mem)                                        // ANY
   ASMJIT_INST_2x(div, Div, Gp, Gp)                                     // ANY [EXPLICIT]  AH[Rem]: AL[Quot] <- AX / r8
@@ -534,7 +534,7 @@ public:
   ASMJIT_INST_2x(or_, Or, Mem, Imm)                                    // ANY
   ASMJIT_INST_1x(pop, Pop, Gp)                                         // ANY
   ASMJIT_INST_1x(pop, Pop, Mem)                                        // ANY
-  ASMJIT_INST_1x(pop, Pop, SReg);                                      // ANY
+  ASMJIT_INST_1x(pop, Pop, SReg)                                       // ANY
   ASMJIT_INST_0x(popa, Popa)                                           // X86
   ASMJIT_INST_0x(popad, Popad)                                         // X86
   ASMJIT_INST_0x(popf, Popf)                                           // ANY