diff --git a/riscv/insn_template.h b/riscv/insn_template.h
index bf9d9d74..7a9aadd8 100644
--- a/riscv/insn_template.h
+++ b/riscv/insn_template.h
@@ -7,6 +7,7 @@
 #include "internals.h"
 #include "specialize.h"
 #include "tracer.h"
+#include "p_ext_macros.h"
 #include "v_ext_macros.h"
 #include "debug_defines.h"
 #include <assert.h>
diff --git a/riscv/insns/aadd.h b/riscv/insns/aadd.h
new file mode 100644
index 00000000..5456cef6
--- /dev/null
+++ b/riscv/insns/aadd.h
@@ -0,0 +1,3 @@
+require_extension('P');
+require_rv32;
+WRITE_RD((RS1 + RS2)>>1);
\ No newline at end of file
diff --git a/riscv/insns/aaddu.h b/riscv/insns/aaddu.h
new file mode 100644
index 00000000..bf011496
--- /dev/null
+++ b/riscv/insns/aaddu.h
@@ -0,0 +1,3 @@
+require_extension('P');
+require_rv32;
+WRITE_RD((0ULL + (uint32_t)RS1 + (uint32_t)RS2)>>1);
\ No newline at end of file
diff --git a/riscv/insns/abs.h b/riscv/insns/abs.h
new file mode 100644
index 00000000..1b747ffe
--- /dev/null
+++ b/riscv/insns/abs.h
@@ -0,0 +1,4 @@
+require_extension('P');
+reg_t s1 = RS1;
+reg_t result = (int64_t(s1) < 0) ? -s1 : s1;
+WRITE_RD(sext_xlen(result));
\ No newline at end of file
diff --git a/riscv/insns/asub.h b/riscv/insns/asub.h
new file mode 100644
index 00000000..b7a7ab53
--- /dev/null
+++ b/riscv/insns/asub.h
@@ -0,0 +1,3 @@
+require_extension('P');
+require_rv32;
+WRITE_RD((RS1 - RS2)>>1);
\ No newline at end of file
diff --git a/riscv/insns/asubu.h b/riscv/insns/asubu.h
new file mode 100644
index 00000000..6139feb2
--- /dev/null
+++ b/riscv/insns/asubu.h
@@ -0,0 +1,3 @@
+require_extension('P');
+require_rv32;
+WRITE_RD((0ULL + (uint32_t)RS1 - (uint32_t)RS2)>>1);
\ No newline at end of file
diff --git a/riscv/insns/cls.h b/riscv/insns/cls.h
new file mode 100644
index 00000000..3d80b618
--- /dev/null
+++ b/riscv/insns/cls.h
@@ -0,0 +1,10 @@
+require_extension('P');
+reg_t x = xlen - 1;
+reg_t msb = (RS1 >> (xlen - 1)) & 1;
+for (int i = 0; i < xlen - 1; i++) {
+  if (msb != ((RS1 >> (xlen - i - 2)) & 1)) {
+    x = i;
+    break;
+  }
+}
+WRITE_RD(sext_xlen(x));
\ No newline at end of file
diff --git a/riscv/insns/macc_h00.h b/riscv/insns/macc_h00.h
new file mode 100644
index 00000000..bddbc4cd
--- /dev/null
+++ b/riscv/insns/macc_h00.h
@@ -0,0 +1,3 @@
+require_extension('P');
+require_rv32;
+WRITE_RD(RD + sext(RS1, 16) * sext(RS2, 16));
\ No newline at end of file
diff --git a/riscv/insns/macc_h01.h b/riscv/insns/macc_h01.h
new file mode 100644
index 00000000..551e4568
--- /dev/null
+++ b/riscv/insns/macc_h01.h
@@ -0,0 +1,3 @@
+require_extension('P');
+require_rv32;
+WRITE_RD(RD + sext(RS1, 16) * sext(RS2 >> 16, 16));
\ No newline at end of file
diff --git a/riscv/insns/macc_h11.h b/riscv/insns/macc_h11.h
new file mode 100644
index 00000000..73104545
--- /dev/null
+++ b/riscv/insns/macc_h11.h
@@ -0,0 +1,3 @@
+require_extension('P');
+require_rv32;
+WRITE_RD(RD + sext(RS1 >> 16, 16) * sext(RS2 >> 16, 16));
\ No newline at end of file
diff --git a/riscv/insns/maccsu_h00.h b/riscv/insns/maccsu_h00.h
new file mode 100644
index 00000000..4039c255
--- /dev/null
+++ b/riscv/insns/maccsu_h00.h
@@ -0,0 +1,3 @@
+require_extension('P');
+require_rv32;
+WRITE_RD(RD + sext(RS1, 16) * zext(RS2, 16));
\ No newline at end of file
diff --git a/riscv/insns/maccsu_h11.h b/riscv/insns/maccsu_h11.h
new file mode 100644
index 00000000..1bbde234
--- /dev/null
+++ b/riscv/insns/maccsu_h11.h
@@ -0,0 +1,3 @@
+require_extension('P');
+require_rv32;
+WRITE_RD(RD + sext(RS1 >> 16, 16) * zext(RS2 >> 16, 16));
\ No newline at end of file
diff --git a/riscv/insns/maccu_h00.h b/riscv/insns/maccu_h00.h
new file mode 100644
index 00000000..22002959
--- /dev/null
+++ b/riscv/insns/maccu_h00.h
@@ -0,0 +1,3 @@
+require_extension('P');
+require_rv32;
+WRITE_RD(RD + zext(RS1, 16) * zext(RS2, 16));
\ No newline at end of file
diff --git a/riscv/insns/maccu_h01.h b/riscv/insns/maccu_h01.h
new file mode 100644
index 00000000..a35865bf
--- /dev/null
+++ b/riscv/insns/maccu_h01.h
@@ -0,0 +1,3 @@
+require_extension('P');
+require_rv32;
+WRITE_RD(RD + zext(RS1, 16) * zext(RS2 >> 16, 16));
\ No newline at end of file
diff --git a/riscv/insns/maccu_h11.h b/riscv/insns/maccu_h11.h
new file mode 100644
index 00000000..a30a8760
--- /dev/null
+++ b/riscv/insns/maccu_h11.h
@@ -0,0 +1,3 @@
+require_extension('P');
+require_rv32;
+WRITE_RD(RD + zext(RS1 >> 16, 16) * zext(RS2 >> 16, 16));
\ No newline at end of file
diff --git a/riscv/insns/merge.h b/riscv/insns/merge.h
new file mode 100644
index 00000000..d184c5c8
--- /dev/null
+++ b/riscv/insns/merge.h
@@ -0,0 +1,2 @@
+require_extension('P');
+WRITE_RD((RS2 & RD) | (RS1 & ~RD));
\ No newline at end of file
diff --git a/riscv/insns/mhacc.h b/riscv/insns/mhacc.h
new file mode 100644
index 00000000..77de911d
--- /dev/null
+++ b/riscv/insns/mhacc.h
@@ -0,0 +1,4 @@
+require_extension('P');
+require_rv32;
+sreg_t mres = sext32(RS1) * sext32(RS2);
+WRITE_RD(RD + (mres >> 32));
\ No newline at end of file
diff --git a/riscv/insns/mhacc_h0.h b/riscv/insns/mhacc_h0.h
new file mode 100644
index 00000000..699e9165
--- /dev/null
+++ b/riscv/insns/mhacc_h0.h
@@ -0,0 +1,4 @@
+require_extension('P');
+require_rv32;
+sreg_t mres = sext32(RS1) * sext32(P_FIELD(RS2, 0, 16));
+WRITE_RD(RD + (mres >> 32));
\ No newline at end of file
diff --git a/riscv/insns/mhacc_h1.h b/riscv/insns/mhacc_h1.h
new file mode 100644
index 00000000..ea228fe6
--- /dev/null
+++ b/riscv/insns/mhacc_h1.h
@@ -0,0 +1,4 @@
+require_extension('P');
+require_rv32;
+sreg_t mres = sext32(RS1) * sext32(P_FIELD(RS2, 1, 16));
+WRITE_RD(RD + (mres >> 32));
\ No newline at end of file
diff --git a/riscv/insns/mhaccsu.h b/riscv/insns/mhaccsu.h
new file mode 100644
index 00000000..328355c1
--- /dev/null
+++ b/riscv/insns/mhaccsu.h
@@ -0,0 +1,4 @@
+require_extension('P');
+require_rv32;
+sreg_t mres = sext32(RS1) * reg_t((uint32_t)RS2);
+WRITE_RD(RD + (mres >> 32));
\ No newline at end of file
diff --git a/riscv/insns/mhaccsu_h0.h b/riscv/insns/mhaccsu_h0.h
new file mode 100644
index 00000000..e89f309b
--- /dev/null
+++ b/riscv/insns/mhaccsu_h0.h
@@ -0,0 +1,4 @@
+require_extension('P');
+require_rv32;
+sreg_t mres = sext32(RS1) * (uint32_t)P_FIELD(RS2, 0, 16);
+WRITE_RD(RD + (mres >> 32));
\ No newline at end of file
diff --git a/riscv/insns/mhaccsu_h1.h b/riscv/insns/mhaccsu_h1.h
new file mode 100644
index 00000000..e49a15f6
--- /dev/null
+++ b/riscv/insns/mhaccsu_h1.h
@@ -0,0 +1,4 @@
+require_extension('P');
+require_rv32;
+sreg_t mres = sext32(RS1) * (uint32_t)P_FIELD(RS2, 1, 16);
+WRITE_RD(RD + (mres >> 32));
\ No newline at end of file
diff --git a/riscv/insns/mhaccu.h b/riscv/insns/mhaccu.h
new file mode 100644
index 00000000..e9fb8020
--- /dev/null
+++ b/riscv/insns/mhaccu.h
@@ -0,0 +1,4 @@
+require_extension('P');
+require_rv32;
+uint64_t mres = reg_t((uint32_t)RS1) * reg_t((uint32_t)RS2);
+WRITE_RD(RD + (mres >> 32));
\ No newline at end of file
diff --git a/riscv/insns/mhracc.h b/riscv/insns/mhracc.h
new file mode 100644
index 00000000..902b0e56
--- /dev/null
+++ b/riscv/insns/mhracc.h
@@ -0,0 +1,5 @@
+require_extension('P');
+require_rv32;
+sreg_t mres = sext32(RS1) * sext32(RS2);
+int32_t round = ((mres >> 31) + 1) >> 1;
+WRITE_RD(RD + round);
\ No newline at end of file
diff --git a/riscv/insns/mhraccsu.h b/riscv/insns/mhraccsu.h
new file mode 100644
index 00000000..5d676ba6
--- /dev/null
+++ b/riscv/insns/mhraccsu.h
@@ -0,0 +1,5 @@
+require_extension('P');
+require_rv32;
+sreg_t mres = sext(RS1,64) * reg_t((uint32_t)RS2);
+int32_t round = ((mres >> 31) + 1) >> 1;
+WRITE_RD(RD + round);
\ No newline at end of file
diff --git a/riscv/insns/mhraccu.h b/riscv/insns/mhraccu.h
new file mode 100644
index 00000000..9cf5e5a4
--- /dev/null
+++ b/riscv/insns/mhraccu.h
@@ -0,0 +1,5 @@
+require_extension('P');
+require_rv32;
+reg_t mres = reg_t((uint32_t)RS1) * reg_t((uint32_t)RS2);
+uint32_t round = ((mres >> 31) + 1) >> 1;
+WRITE_RD(RD + round);
\ No newline at end of file
diff --git a/riscv/insns/mqacc_h00.h b/riscv/insns/mqacc_h00.h
new file mode 100644
index 00000000..7b4b8e7c
--- /dev/null
+++ b/riscv/insns/mqacc_h00.h
@@ -0,0 +1,3 @@
+require_extension('P');
+require_rv32;
+WRITE_RD(RD + (((int32_t)P_FIELD(RS1, 0, 16) * P_FIELD(RS2, 0, 16)) >> 15));
\ No newline at end of file
diff --git a/riscv/insns/mqacc_h01.h b/riscv/insns/mqacc_h01.h
new file mode 100644
index 00000000..ff618720
--- /dev/null
+++ b/riscv/insns/mqacc_h01.h
@@ -0,0 +1,3 @@
+require_extension('P');
+require_rv32;
+WRITE_RD(RD + (((int32_t)P_FIELD(RS1, 0, 16) * P_FIELD(RS2, 1, 16)) >> 15));
\ No newline at end of file
diff --git a/riscv/insns/mqacc_h11.h b/riscv/insns/mqacc_h11.h
new file mode 100644
index 00000000..ac9e18bc
--- /dev/null
+++ b/riscv/insns/mqacc_h11.h
@@ -0,0 +1,3 @@
+require_extension('P');
+require_rv32;
+WRITE_RD(RD + (((int32_t)P_FIELD(RS1, 1, 16) * P_FIELD(RS2, 1, 16)) >> 15));
\ No newline at end of file
diff --git a/riscv/insns/mqracc_h00.h b/riscv/insns/mqracc_h00.h
new file mode 100644
index 00000000..37d49bde
--- /dev/null
+++ b/riscv/insns/mqracc_h00.h
@@ -0,0 +1,3 @@
+require_extension('P');
+require_rv32;
+WRITE_RD(RD + (((int32_t)P_FIELD(RS1, 0, 16) * P_FIELD(RS2, 0, 16) + 0x4000) >> 15));
\ No newline at end of file
diff --git a/riscv/insns/mqracc_h01.h b/riscv/insns/mqracc_h01.h
new file mode 100644
index 00000000..3e7358e5
--- /dev/null
+++ b/riscv/insns/mqracc_h01.h
@@ -0,0 +1,3 @@
+require_extension('P');
+require_rv32;
+WRITE_RD(RD + (((int32_t)P_FIELD(RS1, 0, 16) * P_FIELD(RS2, 1, 16) + 0x4000) >> 15));
\ No newline at end of file
diff --git a/riscv/insns/mqracc_h11.h b/riscv/insns/mqracc_h11.h
new file mode 100644
index 00000000..c4eb4860
--- /dev/null
+++ b/riscv/insns/mqracc_h11.h
@@ -0,0 +1,3 @@
+require_extension('P');
+require_rv32;
+WRITE_RD(RD + (((int32_t)P_FIELD(RS1, 1, 16) * P_FIELD(RS2, 1, 16) + 0x4000) >> 15));
\ No newline at end of file
diff --git a/riscv/insns/mseq.h b/riscv/insns/mseq.h
new file mode 100644
index 00000000..cf540b20
--- /dev/null
+++ b/riscv/insns/mseq.h
@@ -0,0 +1,3 @@
+require_extension('P');
+require_rv32;
+WRITE_RD(RS1 == RS2 ? -1 : 0);
\ No newline at end of file
diff --git a/riscv/insns/mslt.h b/riscv/insns/mslt.h
new file mode 100644
index 00000000..18aa387a
--- /dev/null
+++ b/riscv/insns/mslt.h
@@ -0,0 +1,3 @@
+require_extension('P');
+require_rv32;
+WRITE_RD((int32_t)RS1 < (int32_t)RS2 ? -1 : 0);
\ No newline at end of file
diff --git a/riscv/insns/msltu.h b/riscv/insns/msltu.h
new file mode 100644
index 00000000..062f12df
--- /dev/null
+++ b/riscv/insns/msltu.h
@@ -0,0 +1,3 @@
+require_extension('P');
+require_rv32;
+WRITE_RD((uint32_t)RS1 < (uint32_t)RS2 ? -1 : 0);
\ No newline at end of file
diff --git a/riscv/insns/mul_h00.h b/riscv/insns/mul_h00.h
new file mode 100644
index 00000000..69bfaae6
--- /dev/null
+++ b/riscv/insns/mul_h00.h
@@ -0,0 +1,3 @@
+require_extension('P');
+require_rv32;
+WRITE_RD(sext(RS1, 16) * sext(RS2, 16));
\ No newline at end of file
diff --git a/riscv/insns/mul_h01.h b/riscv/insns/mul_h01.h
new file mode 100644
index 00000000..ec95fe17
--- /dev/null
+++ b/riscv/insns/mul_h01.h
@@ -0,0 +1,3 @@
+require_extension('P');
+require_rv32;
+WRITE_RD(sext(RS1, 16) * sext(RS2 >> 16, 16));
\ No newline at end of file
diff --git a/riscv/insns/mul_h11.h b/riscv/insns/mul_h11.h
new file mode 100644
index 00000000..4ea0d680
--- /dev/null
+++ b/riscv/insns/mul_h11.h
@@ -0,0 +1,3 @@
+require_extension('P');
+require_rv32;
+WRITE_RD(sext(RS1 >> 16, 16) * sext(RS2 >> 16, 16));
\ No newline at end of file
diff --git a/riscv/insns/mulh_h0.h b/riscv/insns/mulh_h0.h
new file mode 100644
index 00000000..2143d3d2
--- /dev/null
+++ b/riscv/insns/mulh_h0.h
@@ -0,0 +1,4 @@
+require_extension('P');
+require_rv32;
+int64_t mres = sext(RS1,64) * sext(P_FIELD(RS2, 0, 16),64);
+WRITE_RD(mres>>32);
\ No newline at end of file
diff --git a/riscv/insns/mulh_h1.h b/riscv/insns/mulh_h1.h
new file mode 100644
index 00000000..5e5bf994
--- /dev/null
+++ b/riscv/insns/mulh_h1.h
@@ -0,0 +1,4 @@
+require_extension('P');
+require_rv32;
+int64_t mres = sext(RS1,64) * sext(P_FIELD(RS2, 1, 16),64);
+WRITE_RD(mres>>32);
\ No newline at end of file
diff --git a/riscv/insns/mulhr.h b/riscv/insns/mulhr.h
new file mode 100644
index 00000000..527d2a2f
--- /dev/null
+++ b/riscv/insns/mulhr.h
@@ -0,0 +1,4 @@
+require_extension('P');
+require_rv32;
+int64_t mres = sext(RS1,64) * sext(RS2,64);
+WRITE_RD(((mres >> 31) + 1) >> 1);
\ No newline at end of file
diff --git a/riscv/insns/mulhrsu.h b/riscv/insns/mulhrsu.h
new file mode 100644
index 00000000..bbd7619b
--- /dev/null
+++ b/riscv/insns/mulhrsu.h
@@ -0,0 +1,4 @@
+require_extension('P');
+require_rv32;
+sreg_t mres = sext(RS1,64) * reg_t((uint32_t)RS2);
+WRITE_RD(((mres >> 31) + 1) >> 1);
\ No newline at end of file
diff --git a/riscv/insns/mulhru.h b/riscv/insns/mulhru.h
new file mode 100644
index 00000000..51cf38cb
--- /dev/null
+++ b/riscv/insns/mulhru.h
@@ -0,0 +1,4 @@
+require_extension('P');
+require_rv32;
+sreg_t mres = reg_t((uint32_t)RS1) * reg_t((uint32_t)RS2);
+WRITE_RD(((mres >> 31) + 1) >> 1);
\ No newline at end of file
diff --git a/riscv/insns/mulhsu_h0.h b/riscv/insns/mulhsu_h0.h
new file mode 100644
index 00000000..2cdb292c
--- /dev/null
+++ b/riscv/insns/mulhsu_h0.h
@@ -0,0 +1,4 @@
+require_extension('P');
+require_rv32;
+sreg_t mres = sext32(RS1) * (uint32_t)P_FIELD(RS2, 0, 16);
+WRITE_RD(mres >> 32);
\ No newline at end of file
diff --git a/riscv/insns/mulhsu_h1.h b/riscv/insns/mulhsu_h1.h
new file mode 100644
index 00000000..4cc9fc32
--- /dev/null
+++ b/riscv/insns/mulhsu_h1.h
@@ -0,0 +1,4 @@
+require_extension('P');
+require_rv32;
+sreg_t mres = sext32(RS1) * (uint32_t)P_FIELD(RS2, 1, 16);
+WRITE_RD(mres >> 32);
\ No newline at end of file
diff --git a/riscv/insns/mulq.h b/riscv/insns/mulq.h
new file mode 100644
index 00000000..cb362db9
--- /dev/null
+++ b/riscv/insns/mulq.h
@@ -0,0 +1,7 @@
+require_extension('P');
+require_rv32;
+if ((RS1 != (reg_t)INT32_MIN) || (RS2 != (reg_t)INT32_MIN)) {
+    WRITE_RD((RS1 * RS2) >> 31);
+  } else {
+    WRITE_RD(INT32_MAX);
+}
diff --git a/riscv/insns/mulqr.h b/riscv/insns/mulqr.h
new file mode 100644
index 00000000..070b8fc4
--- /dev/null
+++ b/riscv/insns/mulqr.h
@@ -0,0 +1,7 @@
+require_extension('P');
+require_rv32;
+if ((RS1 != (reg_t)INT32_MIN) || (RS2 != (reg_t)INT32_MIN)) {
+    WRITE_RD((((RS1 * RS2) >> 30) + 1) >> 1);
+  } else {
+    WRITE_RD(INT32_MAX);
+}
diff --git a/riscv/insns/mulsu_h00.h b/riscv/insns/mulsu_h00.h
new file mode 100644
index 00000000..f02de759
--- /dev/null
+++ b/riscv/insns/mulsu_h00.h
@@ -0,0 +1,3 @@
+require_extension('P');
+require_rv32;
+WRITE_RD(sext(RS1, 16) * zext(RS2, 16));
\ No newline at end of file
diff --git a/riscv/insns/mulsu_h11.h b/riscv/insns/mulsu_h11.h
new file mode 100644
index 00000000..d4954037
--- /dev/null
+++ b/riscv/insns/mulsu_h11.h
@@ -0,0 +1,3 @@
+require_extension('P');
+require_rv32;
+WRITE_RD(sext(RS1 >> 16, 16) * zext(RS2 >> 16, 16));
\ No newline at end of file
diff --git a/riscv/insns/mulu_h00.h b/riscv/insns/mulu_h00.h
new file mode 100644
index 00000000..d92c605b
--- /dev/null
+++ b/riscv/insns/mulu_h00.h
@@ -0,0 +1,3 @@
+require_extension('P');
+require_rv32;
+WRITE_RD(zext(RS1, 16) * zext(RS2, 16));
\ No newline at end of file
diff --git a/riscv/insns/mulu_h01.h b/riscv/insns/mulu_h01.h
new file mode 100644
index 00000000..75a8302b
--- /dev/null
+++ b/riscv/insns/mulu_h01.h
@@ -0,0 +1,3 @@
+require_extension('P');
+require_rv32;
+WRITE_RD(zext(RS1, 16) * zext(RS2 >> 16, 16));
\ No newline at end of file
diff --git a/riscv/insns/mulu_h11.h b/riscv/insns/mulu_h11.h
new file mode 100644
index 00000000..28db7c86
--- /dev/null
+++ b/riscv/insns/mulu_h11.h
@@ -0,0 +1,3 @@
+require_extension('P');
+require_rv32;
+WRITE_RD(zext(RS1 >> 16, 16) * zext(RS2 >> 16, 16));
\ No newline at end of file
diff --git a/riscv/insns/mvm.h b/riscv/insns/mvm.h
new file mode 100644
index 00000000..bf2ddad1
--- /dev/null
+++ b/riscv/insns/mvm.h
@@ -0,0 +1,2 @@
+require_extension('P');
+WRITE_RD((RS1 & RS2) | (RD & ~RS2));
\ No newline at end of file
diff --git a/riscv/insns/mvmn.h b/riscv/insns/mvmn.h
new file mode 100644
index 00000000..8c863d91
--- /dev/null
+++ b/riscv/insns/mvmn.h
@@ -0,0 +1,2 @@
+require_extension('P');
+WRITE_RD((RD & RS2) | (RS1 & ~RS2));
\ No newline at end of file
diff --git a/riscv/insns/paadd_b.h b/riscv/insns/paadd_b.h
new file mode 100644
index 00000000..9f5f043c
--- /dev/null
+++ b/riscv/insns/paadd_b.h
@@ -0,0 +1,3 @@
+P_RD_RS1_RS2_LOOP(8,8,8, {
+  p_rd = (p_rs1 + p_rs2) >> 1;
+})
\ No newline at end of file
diff --git a/riscv/insns/paadd_h.h b/riscv/insns/paadd_h.h
new file mode 100644
index 00000000..bc01ea16
--- /dev/null
+++ b/riscv/insns/paadd_h.h
@@ -0,0 +1,3 @@
+P_RD_RS1_RS2_LOOP(16,16,16, {
+  p_rd = (p_rs1 + p_rs2) >> 1;
+})
\ No newline at end of file
diff --git a/riscv/insns/paaddu_b.h b/riscv/insns/paaddu_b.h
new file mode 100644
index 00000000..0a413d55
--- /dev/null
+++ b/riscv/insns/paaddu_b.h
@@ -0,0 +1,3 @@
+P_RD_RS1_RS2_ULOOP(8,8,8, {
+  p_rd = (p_rs1 + p_rs2) >> 1;
+})
\ No newline at end of file
diff --git a/riscv/insns/paaddu_h.h b/riscv/insns/paaddu_h.h
new file mode 100644
index 00000000..dbef824b
--- /dev/null
+++ b/riscv/insns/paaddu_h.h
@@ -0,0 +1,3 @@
+P_RD_RS1_RS2_ULOOP(16,16,16, {
+  p_rd = (p_rs1 + p_rs2) >> 1;
+})
\ No newline at end of file
diff --git a/riscv/insns/paas_hx.h b/riscv/insns/paas_hx.h
new file mode 100644
index 00000000..5c393a85
--- /dev/null
+++ b/riscv/insns/paas_hx.h
@@ -0,0 +1,5 @@
+P_CROSS_LOOP(16, {
+  p_rd = (p_rs1 + p_rs2) >> 1;
+}, {
+  p_rd = (p_rs1 - p_rs2) >> 1;
+})
\ No newline at end of file
diff --git a/riscv/insns/pabd_b.h b/riscv/insns/pabd_b.h
new file mode 100644
index 00000000..978adbac
--- /dev/null
+++ b/riscv/insns/pabd_b.h
@@ -0,0 +1,4 @@
+P_RD_RS1_RS2_LOOP(8, 8, 8, {
+  p_rd = (p_rs1 < p_rs2) ? (int8_t)((uint8_t)p_rs2 - (uint8_t)p_rs1)
+                          : (int8_t)((uint8_t)p_rs1 - (uint8_t)p_rs2);
+})
\ No newline at end of file
diff --git a/riscv/insns/pabd_h.h b/riscv/insns/pabd_h.h
new file mode 100644
index 00000000..cd18fac7
--- /dev/null
+++ b/riscv/insns/pabd_h.h
@@ -0,0 +1,4 @@
+P_RD_RS1_RS2_LOOP(16, 16, 16, {
+  p_rd = (p_rs1 < p_rs2) ? (int16_t)((uint16_t)p_rs2 - (uint16_t)p_rs1)
+                          : (int16_t)((uint16_t)p_rs1 - (uint16_t)p_rs2);
+})
\ No newline at end of file
diff --git a/riscv/insns/pabdsumau_b.h b/riscv/insns/pabdsumau_b.h
new file mode 100644
index 00000000..6ffc6888
--- /dev/null
+++ b/riscv/insns/pabdsumau_b.h
@@ -0,0 +1,3 @@
+P_REDUCTION_ULOOP(64, 8, true, false, {
+  p_res += (p_rs1 > p_rs2 ? p_rs1 - p_rs2 : p_rs2 - p_rs1);
+})
\ No newline at end of file
diff --git a/riscv/insns/pabdsumu_b.h b/riscv/insns/pabdsumu_b.h
new file mode 100644
index 00000000..b7effa47
--- /dev/null
+++ b/riscv/insns/pabdsumu_b.h
@@ -0,0 +1,3 @@
+P_REDUCTION_ULOOP(64, 8, false, false, {
+  p_res += (p_rs1 > p_rs2 ? p_rs1 - p_rs2 : p_rs2 - p_rs1);
+})
\ No newline at end of file
diff --git a/riscv/insns/pabdu_b.h b/riscv/insns/pabdu_b.h
new file mode 100644
index 00000000..d0092550
--- /dev/null
+++ b/riscv/insns/pabdu_b.h
@@ -0,0 +1,3 @@
+P_RD_RS1_RS2_ULOOP(8, 8, 8, {
+  p_rd = (p_rs1 < p_rs2) ? p_rs2 - p_rs1 : p_rs1 - p_rs2;
+})
\ No newline at end of file
diff --git a/riscv/insns/pabdu_h.h b/riscv/insns/pabdu_h.h
new file mode 100644
index 00000000..83b1445d
--- /dev/null
+++ b/riscv/insns/pabdu_h.h
@@ -0,0 +1,3 @@
+P_RD_RS1_RS2_ULOOP(16, 16, 16, {
+  p_rd = (p_rs1 < p_rs2) ? p_rs2 - p_rs1 : p_rs1 - p_rs2;
+})
\ No newline at end of file
diff --git a/riscv/insns/pack.h b/riscv/insns/pack.h
index acc49877..25f5145b 100644
--- a/riscv/insns/pack.h
+++ b/riscv/insns/pack.h
@@ -1,6 +1,7 @@
 // RV32Zbb contains zext.h but not general pack
 require(((xlen == 32) && (insn.rs2() == 0) && p->extension_enabled(EXT_ZBB))
-  || p->extension_enabled(EXT_ZBKB));
+  || p->extension_enabled(EXT_ZBKB)
+  || p->extension_enabled('P'));
 reg_t lo = zext_xlen(RS1 << (xlen/2)) >> (xlen/2);
 reg_t hi = zext_xlen(RS2 << (xlen/2));
-WRITE_RD(sext_xlen(lo | hi));
+WRITE_RD(sext_xlen(lo | hi));
\ No newline at end of file
diff --git a/riscv/insns/padd_b.h b/riscv/insns/padd_b.h
new file mode 100644
index 00000000..05f890ca
--- /dev/null
+++ b/riscv/insns/padd_b.h
@@ -0,0 +1,3 @@
+P_RD_RS1_RS2_LOOP(8,8,8, {
+  p_rd = p_rs1 + p_rs2;
+})
\ No newline at end of file
diff --git a/riscv/insns/padd_bs.h b/riscv/insns/padd_bs.h
new file mode 100644
index 00000000..9070fc9b
--- /dev/null
+++ b/riscv/insns/padd_bs.h
@@ -0,0 +1,3 @@
+P_RD_RS1_LOOP(8, 8, {
+  p_rd = p_rs1 + P_FIELD(RS2, 0, 8);
+})
\ No newline at end of file
diff --git a/riscv/insns/padd_h.h b/riscv/insns/padd_h.h
new file mode 100644
index 00000000..5ff1d324
--- /dev/null
+++ b/riscv/insns/padd_h.h
@@ -0,0 +1,3 @@
+P_RD_RS1_RS2_LOOP(16, 16, 16, {
+  p_rd = p_rs1 + p_rs2;
+})
\ No newline at end of file
diff --git a/riscv/insns/padd_hs.h b/riscv/insns/padd_hs.h
new file mode 100644
index 00000000..e68d0044
--- /dev/null
+++ b/riscv/insns/padd_hs.h
@@ -0,0 +1,3 @@
+P_RD_RS1_LOOP(16, 16, {
+  p_rd = p_rs1 + P_FIELD(RS2, 0, 16);
+})
\ No newline at end of file
diff --git a/riscv/insns/pas_hx.h b/riscv/insns/pas_hx.h
new file mode 100644
index 00000000..fbe4a244
--- /dev/null
+++ b/riscv/insns/pas_hx.h
@@ -0,0 +1,5 @@
+P_CROSS_LOOP(16, {
+  p_rd = p_rs1 + p_rs2;
+}, {
+  p_rd = p_rs1 - p_rs2;
+})
\ No newline at end of file
diff --git a/riscv/insns/pasa_hx.h b/riscv/insns/pasa_hx.h
new file mode 100644
index 00000000..eea09357
--- /dev/null
+++ b/riscv/insns/pasa_hx.h
@@ -0,0 +1,5 @@
+P_CROSS_LOOP(16, {
+  p_rd = (p_rs1 - p_rs2) >> 1;
+}, {
+  p_rd = (p_rs1 + p_rs2) >> 1;
+})
\ No newline at end of file
diff --git a/riscv/insns/pasub_b.h b/riscv/insns/pasub_b.h
new file mode 100644
index 00000000..48173164
--- /dev/null
+++ b/riscv/insns/pasub_b.h
@@ -0,0 +1,3 @@
+P_RD_RS1_RS2_LOOP(8,8,8, {
+  p_rd = (p_rs1 - p_rs2) >> 1;
+})
\ No newline at end of file
diff --git a/riscv/insns/pasub_h.h b/riscv/insns/pasub_h.h
new file mode 100644
index 00000000..57c271ae
--- /dev/null
+++ b/riscv/insns/pasub_h.h
@@ -0,0 +1,3 @@
+P_RD_RS1_RS2_LOOP(16,16,16, {
+  p_rd = (p_rs1 - p_rs2) >> 1;
+})
\ No newline at end of file
diff --git a/riscv/insns/pasubu_b.h b/riscv/insns/pasubu_b.h
new file mode 100644
index 00000000..7543305a
--- /dev/null
+++ b/riscv/insns/pasubu_b.h
@@ -0,0 +1,3 @@
+P_RD_RS1_RS2_ULOOP(8,8,8, {
+  p_rd = (p_rs1 - p_rs2) >> 1;
+})
\ No newline at end of file
diff --git a/riscv/insns/pasubu_h.h b/riscv/insns/pasubu_h.h
new file mode 100644
index 00000000..453fe57b
--- /dev/null
+++ b/riscv/insns/pasubu_h.h
@@ -0,0 +1,3 @@
+P_RD_RS1_RS2_ULOOP(16,16,16, {
+  p_rd = (p_rs1 - p_rs2) >> 1;
+})
\ No newline at end of file
diff --git a/riscv/insns/pli_b.h b/riscv/insns/pli_b.h
new file mode 100644
index 00000000..b04452e3
--- /dev/null
+++ b/riscv/insns/pli_b.h
@@ -0,0 +1,3 @@
+P_RD_LOOP(8, {
+  p_rd = insn.p_imm8();
+})
\ No newline at end of file
diff --git a/riscv/insns/pli_db.h b/riscv/insns/pli_db.h
new file mode 100644
index 00000000..e411fe0f
--- /dev/null
+++ b/riscv/insns/pli_db.h
@@ -0,0 +1,4 @@
+require_rv32;
+P_RD_DW_LOOP(8, {
+  p_rd = insn.p_imm8();
+})
\ No newline at end of file
diff --git a/riscv/insns/pli_dh.h b/riscv/insns/pli_dh.h
new file mode 100644
index 00000000..215b03df
--- /dev/null
+++ b/riscv/insns/pli_dh.h
@@ -0,0 +1,4 @@
+require_rv32;
+P_RD_DW_LOOP(16, {
+  p_rd = (insn.p_imm10csl() & 0x200) ? (0xfc00 | insn.p_imm10csl()) : insn.p_imm10csl();
+})
\ No newline at end of file
diff --git a/riscv/insns/pli_h.h b/riscv/insns/pli_h.h
new file mode 100644
index 00000000..4a891624
--- /dev/null
+++ b/riscv/insns/pli_h.h
@@ -0,0 +1,3 @@
+P_RD_LOOP(16, {
+  p_rd = (insn.p_imm10csl() & 0x200) ? (0xfc00 | insn.p_imm10csl()) : insn.p_imm10csl();
+})
\ No newline at end of file
diff --git a/riscv/insns/plui_dh.h b/riscv/insns/plui_dh.h
new file mode 100644
index 00000000..61f3d6b6
--- /dev/null
+++ b/riscv/insns/plui_dh.h
@@ -0,0 +1,4 @@
+require_rv32;
+P_RD_DW_LOOP(16, {
+  p_rd = insn.p_imm10csr();
+})
\ No newline at end of file
diff --git a/riscv/insns/plui_h.h b/riscv/insns/plui_h.h
new file mode 100644
index 00000000..c11af3e3
--- /dev/null
+++ b/riscv/insns/plui_h.h
@@ -0,0 +1,3 @@
+P_RD_LOOP(16, {
+  p_rd = insn.p_imm10csr();
+})
\ No newline at end of file
diff --git a/riscv/insns/pm2add_h.h b/riscv/insns/pm2add_h.h
new file mode 100644
index 00000000..c0218159
--- /dev/null
+++ b/riscv/insns/pm2add_h.h
@@ -0,0 +1,3 @@
+P_REDUCTION_LOOP(32, 16, false, false, {
+  p_res += p_rs1 * p_rs2;
+})
\ No newline at end of file
diff --git a/riscv/insns/pm2add_hx.h b/riscv/insns/pm2add_hx.h
new file mode 100644
index 00000000..a84dbcc7
--- /dev/null
+++ b/riscv/insns/pm2add_hx.h
@@ -0,0 +1,3 @@
+P_REDUCTION_CROSS_LOOP(32, 16, false, false, {
+  p_res += p_rs1 * p_rs2;
+})
\ No newline at end of file
diff --git a/riscv/insns/pm2adda_h.h b/riscv/insns/pm2adda_h.h
new file mode 100644
index 00000000..ed2e0eb5
--- /dev/null
+++ b/riscv/insns/pm2adda_h.h
@@ -0,0 +1,3 @@
+P_REDUCTION_LOOP(32, 16, true, false, {
+  p_res += p_rs1 * p_rs2;
+})
\ No newline at end of file
diff --git a/riscv/insns/pm2adda_hx.h b/riscv/insns/pm2adda_hx.h
new file mode 100644
index 00000000..2a5d96cb
--- /dev/null
+++ b/riscv/insns/pm2adda_hx.h
@@ -0,0 +1,3 @@
+P_REDUCTION_CROSS_LOOP(32, 16, true, false, {
+  p_res += p_rs1 * p_rs2;
+})
\ No newline at end of file
diff --git a/riscv/insns/pm2addasu_h.h b/riscv/insns/pm2addasu_h.h
new file mode 100644
index 00000000..8ddbcf21
--- /dev/null
+++ b/riscv/insns/pm2addasu_h.h
@@ -0,0 +1,3 @@
+P_REDUCTION_SULOOP(32, 16, true, false, {
+  p_res += p_rs1 * p_rs2;
+})
\ No newline at end of file
diff --git a/riscv/insns/pm2addau_h.h b/riscv/insns/pm2addau_h.h
new file mode 100644
index 00000000..bc8dba63
--- /dev/null
+++ b/riscv/insns/pm2addau_h.h
@@ -0,0 +1,3 @@
+P_REDUCTION_ULOOP(32, 16, true, false, {
+  p_res += p_rs1 * p_rs2;
+})
\ No newline at end of file
diff --git a/riscv/insns/pm2addsu_h.h b/riscv/insns/pm2addsu_h.h
new file mode 100644
index 00000000..dfbc5274
--- /dev/null
+++ b/riscv/insns/pm2addsu_h.h
@@ -0,0 +1,3 @@
+P_REDUCTION_SULOOP(32, 16, false, false, {
+  p_res += p_rs1 * p_rs2;
+})
\ No newline at end of file
diff --git a/riscv/insns/pm2addu_h.h b/riscv/insns/pm2addu_h.h
new file mode 100644
index 00000000..97944410
--- /dev/null
+++ b/riscv/insns/pm2addu_h.h
@@ -0,0 +1,3 @@
+P_REDUCTION_ULOOP(32, 16, false, false, {
+  p_res += p_rs1 * p_rs2;
+})
\ No newline at end of file
diff --git a/riscv/insns/pm2sadd_h.h b/riscv/insns/pm2sadd_h.h
new file mode 100644
index 00000000..01ba8c66
--- /dev/null
+++ b/riscv/insns/pm2sadd_h.h
@@ -0,0 +1,3 @@
+P_REDUCTION_LOOP(32, 16, false, true, {
+  p_res += p_rs1 * p_rs2;
+})
\ No newline at end of file
diff --git a/riscv/insns/pm2sadd_hx.h b/riscv/insns/pm2sadd_hx.h
new file mode 100644
index 00000000..d7dca373
--- /dev/null
+++ b/riscv/insns/pm2sadd_hx.h
@@ -0,0 +1,3 @@
+P_REDUCTION_CROSS_LOOP(32, 16, false, true, {
+  p_res += p_rs1 * p_rs2;
+})
\ No newline at end of file
diff --git a/riscv/insns/pm2sub_h.h b/riscv/insns/pm2sub_h.h
new file mode 100644
index 00000000..48f6ce06
--- /dev/null
+++ b/riscv/insns/pm2sub_h.h
@@ -0,0 +1,6 @@
+P_REDUCTION_LOOP(32, 16, false, false, {
+  if (j & 1)
+    p_res -= p_rs1 * p_rs2;
+  else
+    p_res += p_rs1 * p_rs2;
+})
\ No newline at end of file
diff --git a/riscv/insns/pm2sub_hx.h b/riscv/insns/pm2sub_hx.h
new file mode 100644
index 00000000..1ff687c0
--- /dev/null
+++ b/riscv/insns/pm2sub_hx.h
@@ -0,0 +1,6 @@
+P_REDUCTION_CROSS_LOOP(32, 16, false, false, {
+  if (j & 1)
+    p_res -= p_rs1 * p_rs2;
+  else
+    p_res += p_rs1 * p_rs2;
+})
\ No newline at end of file
diff --git a/riscv/insns/pm2suba_h.h b/riscv/insns/pm2suba_h.h
new file mode 100644
index 00000000..d8c923e2
--- /dev/null
+++ b/riscv/insns/pm2suba_h.h
@@ -0,0 +1,6 @@
+P_REDUCTION_LOOP(32, 16, true, false, {
+  if (j & 1)
+    p_res -= p_rs1 * p_rs2;
+  else
+    p_res += p_rs1 * p_rs2;
+})
\ No newline at end of file
diff --git a/riscv/insns/pm2suba_hx.h b/riscv/insns/pm2suba_hx.h
new file mode 100644
index 00000000..13971e26
--- /dev/null
+++ b/riscv/insns/pm2suba_hx.h
@@ -0,0 +1,6 @@
+P_REDUCTION_CROSS_LOOP(32, 16, true, false, {
+  if (j & 1)
+    p_res -= p_rs1 * p_rs2;
+  else
+    p_res += p_rs1 * p_rs2;
+})
\ No newline at end of file
diff --git a/riscv/insns/pm4add_b.h b/riscv/insns/pm4add_b.h
new file mode 100644
index 00000000..ac1c49ce
--- /dev/null
+++ b/riscv/insns/pm4add_b.h
@@ -0,0 +1,3 @@
+P_REDUCTION_LOOP(32, 8, false, false, {
+  p_res += p_rs1 * p_rs2;
+})
\ No newline at end of file
diff --git a/riscv/insns/pm4adda_b.h b/riscv/insns/pm4adda_b.h
new file mode 100644
index 00000000..d3e00eee
--- /dev/null
+++ b/riscv/insns/pm4adda_b.h
@@ -0,0 +1,3 @@
+P_REDUCTION_LOOP(32, 8, true, false, {
+  p_res += p_rs1 * p_rs2;
+})
\ No newline at end of file
diff --git a/riscv/insns/pm4addasu_b.h b/riscv/insns/pm4addasu_b.h
new file mode 100644
index 00000000..cac07f07
--- /dev/null
+++ b/riscv/insns/pm4addasu_b.h
@@ -0,0 +1,3 @@
+P_REDUCTION_SULOOP(32, 8, true, false, {
+  p_res += p_rs1 * p_rs2;
+})
\ No newline at end of file
diff --git a/riscv/insns/pm4addau_b.h b/riscv/insns/pm4addau_b.h
new file mode 100644
index 00000000..37b29c33
--- /dev/null
+++ b/riscv/insns/pm4addau_b.h
@@ -0,0 +1,3 @@
+P_REDUCTION_ULOOP(32, 8, true, false, {
+  p_res += p_rs1 * p_rs2;
+})
\ No newline at end of file
diff --git a/riscv/insns/pm4addsu_b.h b/riscv/insns/pm4addsu_b.h
new file mode 100644
index 00000000..02fb9b74
--- /dev/null
+++ b/riscv/insns/pm4addsu_b.h
@@ -0,0 +1,3 @@
+P_REDUCTION_SULOOP(32, 8, false, false, {
+  p_res += p_rs1 * p_rs2;
+})
\ No newline at end of file
diff --git a/riscv/insns/pm4addu_b.h b/riscv/insns/pm4addu_b.h
new file mode 100644
index 00000000..a6880e57
--- /dev/null
+++ b/riscv/insns/pm4addu_b.h
@@ -0,0 +1,3 @@
+P_REDUCTION_ULOOP(32, 8, false, false, {
+  p_res += p_rs1 * p_rs2;
+})
\ No newline at end of file
diff --git a/riscv/insns/pmax_b.h b/riscv/insns/pmax_b.h
new file mode 100644
index 00000000..9ab92eb3
--- /dev/null
+++ b/riscv/insns/pmax_b.h
@@ -0,0 +1,3 @@
+P_RD_RS1_RS2_LOOP(8,8,8, {
+  p_rd = (p_rs1 > p_rs2) ? p_rs1 : p_rs2;
+})
\ No newline at end of file
diff --git a/riscv/insns/pmax_h.h b/riscv/insns/pmax_h.h
new file mode 100644
index 00000000..b1e9c0ad
--- /dev/null
+++ b/riscv/insns/pmax_h.h
@@ -0,0 +1,3 @@
+P_RD_RS1_RS2_LOOP(16,16,16, {
+  p_rd = (p_rs1 > p_rs2) ? p_rs1 : p_rs2;
+})
\ No newline at end of file
diff --git a/riscv/insns/pmaxu_b.h b/riscv/insns/pmaxu_b.h
new file mode 100644
index 00000000..82b60f67
--- /dev/null
+++ b/riscv/insns/pmaxu_b.h
@@ -0,0 +1,3 @@
+P_RD_RS1_RS2_ULOOP(8,8,8, {
+  p_rd = (p_rs1 > p_rs2) ? p_rs1 : p_rs2;
+})
\ No newline at end of file
diff --git a/riscv/insns/pmaxu_h.h b/riscv/insns/pmaxu_h.h
new file mode 100644
index 00000000..ac6f6028
--- /dev/null
+++ b/riscv/insns/pmaxu_h.h
@@ -0,0 +1,3 @@
+P_RD_RS1_RS2_ULOOP(16,16,16, {
+  p_rd = (p_rs1 > p_rs2) ? p_rs1 : p_rs2;
+})
\ No newline at end of file
diff --git a/riscv/insns/pmhacc_h.h b/riscv/insns/pmhacc_h.h
new file mode 100644
index 00000000..053c2858
--- /dev/null
+++ b/riscv/insns/pmhacc_h.h
@@ -0,0 +1,4 @@
+P_RD_RS1_RS2_LOOP(16,16,16, {
+  int32_t mres = sext32(p_rs1) * sext32(p_rs2);
+  p_rd += mres>>16; 
+})
\ No newline at end of file
diff --git a/riscv/insns/pmhacc_h_b0.h b/riscv/insns/pmhacc_h_b0.h
new file mode 100644
index 00000000..60a1d5c2
--- /dev/null
+++ b/riscv/insns/pmhacc_h_b0.h
@@ -0,0 +1,4 @@
+P_RD_RS1_RS2_E_LOOP(16,16,8, {
+  int32_t mres = sext32(p_rs1) * sext32(p_rs2);
+  p_rd += mres>>16; 
+})
\ No newline at end of file
diff --git a/riscv/insns/pmhacc_h_b1.h b/riscv/insns/pmhacc_h_b1.h
new file mode 100644
index 00000000..c7aa4d7a
--- /dev/null
+++ b/riscv/insns/pmhacc_h_b1.h
@@ -0,0 +1,4 @@
+P_RD_RS1_RS2_O_LOOP(16,16,8, {
+  int32_t mres = sext32(p_rs1) * sext32(p_rs2);
+  p_rd += mres>>16; 
+})
\ No newline at end of file
diff --git a/riscv/insns/pmhaccsu_h.h b/riscv/insns/pmhaccsu_h.h
new file mode 100644
index 00000000..10b0211c
--- /dev/null
+++ b/riscv/insns/pmhaccsu_h.h
@@ -0,0 +1,4 @@
+P_RD_RS1_RS2_SULOOP(16,16,16, {
+  int32_t mres = sext(p_rs1,32) * zext(p_rs2,32);
+  p_rd += mres>>16;
+})
\ No newline at end of file
diff --git a/riscv/insns/pmhaccsu_h_b0.h b/riscv/insns/pmhaccsu_h_b0.h
new file mode 100644
index 00000000..3d14229b
--- /dev/null
+++ b/riscv/insns/pmhaccsu_h_b0.h
@@ -0,0 +1,4 @@
+P_RD_RS1_RS2_E_SULOOP(16,16,8, {
+  int32_t mres = sext(p_rs1,32) * zext(p_rs2,32);
+  p_rd += mres>>16;
+})
\ No newline at end of file
diff --git a/riscv/insns/pmhaccsu_h_b1.h b/riscv/insns/pmhaccsu_h_b1.h
new file mode 100644
index 00000000..6a5a478f
--- /dev/null
+++ b/riscv/insns/pmhaccsu_h_b1.h
@@ -0,0 +1,4 @@
+P_RD_RS1_RS2_O_SULOOP(16,16,8, {
+  int32_t mres = sext(p_rs1,32) * zext(p_rs2,32);
+  p_rd += mres>>16;
+})
\ No newline at end of file
diff --git a/riscv/insns/pmhaccu_h.h b/riscv/insns/pmhaccu_h.h
new file mode 100644
index 00000000..de1c6eb1
--- /dev/null
+++ b/riscv/insns/pmhaccu_h.h
@@ -0,0 +1,4 @@
+P_RD_RS1_RS2_ULOOP(16,16,16, {
+  uint32_t mres = zext(p_rs1,32) * zext(p_rs2,32);
+  p_rd += mres>>16;
+})
\ No newline at end of file
diff --git a/riscv/insns/pmhracc_h.h b/riscv/insns/pmhracc_h.h
new file mode 100644
index 00000000..1c79a71a
--- /dev/null
+++ b/riscv/insns/pmhracc_h.h
@@ -0,0 +1,5 @@
+P_RD_RS1_RS2_LOOP(16,16,16, {
+  int32_t mres = sext(p_rs1,32) * sext(p_rs2,32);
+  int16_t round = ((mres >> 15) + 1) >> 1;
+  p_rd += round;
+})
\ No newline at end of file
diff --git a/riscv/insns/pmhraccsu_h.h b/riscv/insns/pmhraccsu_h.h
new file mode 100644
index 00000000..c4302837
--- /dev/null
+++ b/riscv/insns/pmhraccsu_h.h
@@ -0,0 +1,5 @@
+P_RD_RS1_RS2_SULOOP(16,16,16, {
+  int32_t mres = sext(p_rs1,32) * zext(p_rs2,32);
+  int16_t round = ((mres >> 15) + 1) >> 1;
+  p_rd += round;
+})
\ No newline at end of file
diff --git a/riscv/insns/pmhraccu_h.h b/riscv/insns/pmhraccu_h.h
new file mode 100644
index 00000000..41efaad8
--- /dev/null
+++ b/riscv/insns/pmhraccu_h.h
@@ -0,0 +1,5 @@
+P_RD_RS1_RS2_ULOOP(16,16,16, {
+  uint32_t mres = zext(p_rs1,32) * zext(p_rs2,32);
+  uint16_t round = ((mres >> 15) + 1) >> 1;
+  p_rd += round;
+})
\ No newline at end of file
diff --git a/riscv/insns/pmin_b.h b/riscv/insns/pmin_b.h
new file mode 100644
index 00000000..402301c0
--- /dev/null
+++ b/riscv/insns/pmin_b.h
@@ -0,0 +1,3 @@
+P_RD_RS1_RS2_LOOP(8,8,8, {
+  p_rd = (p_rs1 < p_rs2) ? p_rs1 : p_rs2;
+})
\ No newline at end of file
diff --git a/riscv/insns/pmin_h.h b/riscv/insns/pmin_h.h
new file mode 100644
index 00000000..2634f3cd
--- /dev/null
+++ b/riscv/insns/pmin_h.h
@@ -0,0 +1,3 @@
+P_RD_RS1_RS2_LOOP(16,16,16, {
+  p_rd = (p_rs1 < p_rs2) ? p_rs1 : p_rs2;
+})
\ No newline at end of file
diff --git a/riscv/insns/pminu_b.h b/riscv/insns/pminu_b.h
new file mode 100644
index 00000000..1156ff22
--- /dev/null
+++ b/riscv/insns/pminu_b.h
@@ -0,0 +1,3 @@
+P_RD_RS1_RS2_ULOOP(8,8,8, {
+  p_rd = (p_rs1 < p_rs2) ? p_rs1 : p_rs2;
+})
\ No newline at end of file
diff --git a/riscv/insns/pminu_h.h b/riscv/insns/pminu_h.h
new file mode 100644
index 00000000..734339c8
--- /dev/null
+++ b/riscv/insns/pminu_h.h
@@ -0,0 +1,3 @@
+P_RD_RS1_RS2_ULOOP(16,16,16, {
+  p_rd = (p_rs1 < p_rs2) ? p_rs1 : p_rs2;
+})
\ No newline at end of file
diff --git a/riscv/insns/pmq2add_h.h b/riscv/insns/pmq2add_h.h
new file mode 100644
index 00000000..efa6e8ff
--- /dev/null
+++ b/riscv/insns/pmq2add_h.h
@@ -0,0 +1,3 @@
+P_REDUCTION_LOOP(32, 16, false, false, {
+  p_res += (p_rs1 * p_rs2) >> 15;
+})
\ No newline at end of file
diff --git a/riscv/insns/pmq2adda_h.h b/riscv/insns/pmq2adda_h.h
new file mode 100644
index 00000000..70e4a5c2
--- /dev/null
+++ b/riscv/insns/pmq2adda_h.h
@@ -0,0 +1,3 @@
+P_REDUCTION_LOOP(32, 16, true, false, {
+  p_res += (p_rs1 * p_rs2) >> 15;
+})
\ No newline at end of file
diff --git a/riscv/insns/pmqr2add_h.h b/riscv/insns/pmqr2add_h.h
new file mode 100644
index 00000000..ee8ac64a
--- /dev/null
+++ b/riscv/insns/pmqr2add_h.h
@@ -0,0 +1,3 @@
+P_REDUCTION_LOOP(32, 16, false, false, {
+  p_res += ((p_rs1 * p_rs2) + 0x4000) >> 15;
+})
\ No newline at end of file
diff --git a/riscv/insns/pmqr2adda_h.h b/riscv/insns/pmqr2adda_h.h
new file mode 100644
index 00000000..7352e174
--- /dev/null
+++ b/riscv/insns/pmqr2adda_h.h
@@ -0,0 +1,3 @@
+P_REDUCTION_LOOP(32, 16, true, false, {
+  p_res += ((p_rs1 * p_rs2) + 0x4000) >> 15;
+})
\ No newline at end of file
diff --git a/riscv/insns/pmseq_b.h b/riscv/insns/pmseq_b.h
new file mode 100644
index 00000000..bd66522d
--- /dev/null
+++ b/riscv/insns/pmseq_b.h
@@ -0,0 +1,3 @@
+P_RD_RS1_RS2_LOOP(8,8,8, {
+  p_rd = (p_rs1 == p_rs2) ? -1 : 0;
+})
\ No newline at end of file
diff --git a/riscv/insns/pmseq_h.h b/riscv/insns/pmseq_h.h
new file mode 100644
index 00000000..3a9b42f8
--- /dev/null
+++ b/riscv/insns/pmseq_h.h
@@ -0,0 +1,3 @@
+P_RD_RS1_RS2_LOOP(16,16,16, {
+  p_rd = (p_rs1 == p_rs2) ? -1 : 0;
+})
\ No newline at end of file
diff --git a/riscv/insns/pmslt_b.h b/riscv/insns/pmslt_b.h
new file mode 100644
index 00000000..d0c1d1a6
--- /dev/null
+++ b/riscv/insns/pmslt_b.h
@@ -0,0 +1,3 @@
+P_RD_RS1_RS2_LOOP(8,8,8, {
+  p_rd = (p_rs1 < p_rs2) ? -1 : 0;
+})
\ No newline at end of file
diff --git a/riscv/insns/pmslt_h.h b/riscv/insns/pmslt_h.h
new file mode 100644
index 00000000..73f71531
--- /dev/null
+++ b/riscv/insns/pmslt_h.h
@@ -0,0 +1,3 @@
+P_RD_RS1_RS2_LOOP(16,16,16, {
+  p_rd = (p_rs1 < p_rs2) ? -1 : 0;
+})
\ No newline at end of file
diff --git a/riscv/insns/pmsltu_b.h b/riscv/insns/pmsltu_b.h
new file mode 100644
index 00000000..8e39925b
--- /dev/null
+++ b/riscv/insns/pmsltu_b.h
@@ -0,0 +1,3 @@
+P_RD_RS1_RS2_ULOOP(8,8,8, {
+  p_rd = (p_rs1 < p_rs2) ? -1 : 0;
+})
\ No newline at end of file
diff --git a/riscv/insns/pmsltu_h.h b/riscv/insns/pmsltu_h.h
new file mode 100644
index 00000000..90d21792
--- /dev/null
+++ b/riscv/insns/pmsltu_h.h
@@ -0,0 +1,3 @@
+P_RD_RS1_RS2_ULOOP(16,16,16, {
+  p_rd = (p_rs1 < p_rs2) ? -1 : 0;
+})
\ No newline at end of file
diff --git a/riscv/insns/pmul_h_b00.h b/riscv/insns/pmul_h_b00.h
new file mode 100644
index 00000000..41af6b34
--- /dev/null
+++ b/riscv/insns/pmul_h_b00.h
@@ -0,0 +1,3 @@
+P_RD_RS1_RS2_EE_LOOP(16, 8, 8, {
+  p_rd = sext32(p_rs1) * sext32(p_rs2);
+})
\ No newline at end of file
diff --git a/riscv/insns/pmul_h_b01.h b/riscv/insns/pmul_h_b01.h
new file mode 100644
index 00000000..1cb85a1b
--- /dev/null
+++ b/riscv/insns/pmul_h_b01.h
@@ -0,0 +1,3 @@
+P_RD_RS1_RS2_EO_LOOP(16, 8, 8, {
+    p_rd = sext32(p_rs1) * sext32(p_rs2);
+})
\ No newline at end of file
diff --git a/riscv/insns/pmul_h_b11.h b/riscv/insns/pmul_h_b11.h
new file mode 100644
index 00000000..8a1765d2
--- /dev/null
+++ b/riscv/insns/pmul_h_b11.h
@@ -0,0 +1,3 @@
+P_RD_RS1_RS2_OO_LOOP(16, 8, 8, {
+  p_rd = sext32(p_rs1) * sext32(p_rs2);
+})
\ No newline at end of file
diff --git a/riscv/insns/pmulh_h.h b/riscv/insns/pmulh_h.h
new file mode 100644
index 00000000..0e199f7b
--- /dev/null
+++ b/riscv/insns/pmulh_h.h
@@ -0,0 +1,4 @@
+P_RD_RS1_RS2_LOOP(16,16,16, {
+  int32_t mres = sext(p_rs1,32) * sext(p_rs2,32);
+  p_rd = mres >> 16;
+})
\ No newline at end of file
diff --git a/riscv/insns/pmulh_h_b0.h b/riscv/insns/pmulh_h_b0.h
new file mode 100644
index 00000000..5c69a9c6
--- /dev/null
+++ b/riscv/insns/pmulh_h_b0.h
@@ -0,0 +1,4 @@
+P_RD_RS1_RS2_E_LOOP(16,16,8, {
+  int32_t mres = sext(p_rs1,32) * sext(p_rs2,32);
+  p_rd = mres >> 16;
+})
\ No newline at end of file
diff --git a/riscv/insns/pmulh_h_b1.h b/riscv/insns/pmulh_h_b1.h
new file mode 100644
index 00000000..1cd5d680
--- /dev/null
+++ b/riscv/insns/pmulh_h_b1.h
@@ -0,0 +1,4 @@
+P_RD_RS1_RS2_O_LOOP(16,16,8, {
+  int32_t mres = sext(p_rs1,32) * sext(p_rs2,32);
+  p_rd = mres >> 16;
+})
\ No newline at end of file
diff --git a/riscv/insns/pmulhr_h.h b/riscv/insns/pmulhr_h.h
new file mode 100644
index 00000000..fbccf0fe
--- /dev/null
+++ b/riscv/insns/pmulhr_h.h
@@ -0,0 +1,4 @@
+P_RD_RS1_RS2_LOOP(16,16,16, {
+  int32_t mres = sext(p_rs1,32) * sext(p_rs2,32);
+  p_rd = ((mres >> 15) + 1) >> 1;
+})
\ No newline at end of file
diff --git a/riscv/insns/pmulhrsu_h.h b/riscv/insns/pmulhrsu_h.h
new file mode 100644
index 00000000..ab681ef0
--- /dev/null
+++ b/riscv/insns/pmulhrsu_h.h
@@ -0,0 +1,4 @@
+P_RD_RS1_RS2_SULOOP(16,16,16, {
+  int32_t mres = sext(p_rs1,32) * zext(p_rs2,32);
+  p_rd = ((mres >> 15) + 1) >> 1;
+})
\ No newline at end of file
diff --git a/riscv/insns/pmulhru_h.h b/riscv/insns/pmulhru_h.h
new file mode 100644
index 00000000..82f1bfe5
--- /dev/null
+++ b/riscv/insns/pmulhru_h.h
@@ -0,0 +1,4 @@
+P_RD_RS1_RS2_ULOOP(16,16,16, {
+  uint32_t mres = zext(p_rs1,32) * zext(p_rs2,32);
+  p_rd = ((mres >> 15) + 1) >> 1;
+})
\ No newline at end of file
diff --git a/riscv/insns/pmulhsu_h.h b/riscv/insns/pmulhsu_h.h
new file mode 100644
index 00000000..d4f07462
--- /dev/null
+++ b/riscv/insns/pmulhsu_h.h
@@ -0,0 +1,4 @@
+P_RD_RS1_RS2_SULOOP(16,16,16, {
+  int32_t mres = sext(p_rs1,32) * zext(p_rs2,32);
+  p_rd = mres >> 16;
+})
\ No newline at end of file
diff --git a/riscv/insns/pmulhsu_h_b0.h b/riscv/insns/pmulhsu_h_b0.h
new file mode 100644
index 00000000..3cfdd4b4
--- /dev/null
+++ b/riscv/insns/pmulhsu_h_b0.h
@@ -0,0 +1,4 @@
+P_RD_RS1_RS2_E_SULOOP(16,16,8, {
+  int32_t mres = sext(p_rs1,32) * zext(p_rs2,32);
+  p_rd = mres >> 16;
+})
\ No newline at end of file
diff --git a/riscv/insns/pmulhsu_h_b1.h b/riscv/insns/pmulhsu_h_b1.h
new file mode 100644
index 00000000..4e6b5bf3
--- /dev/null
+++ b/riscv/insns/pmulhsu_h_b1.h
@@ -0,0 +1,4 @@
+P_RD_RS1_RS2_O_SULOOP(16,16,8, {
+  int32_t mres = sext(p_rs1,32) * zext(p_rs2,32);
+  p_rd = mres >> 16;
+})
\ No newline at end of file
diff --git a/riscv/insns/pmulhu_h.h b/riscv/insns/pmulhu_h.h
new file mode 100644
index 00000000..32a1b280
--- /dev/null
+++ b/riscv/insns/pmulhu_h.h
@@ -0,0 +1,4 @@
+P_RD_RS1_RS2_ULOOP(16,16,16, {
+  uint32_t mres = zext(p_rs1,32) * zext(p_rs2,32);
+  p_rd = mres >> 16;
+})
\ No newline at end of file
diff --git a/riscv/insns/pmulq_h.h b/riscv/insns/pmulq_h.h
new file mode 100644
index 00000000..ad5f5f9e
--- /dev/null
+++ b/riscv/insns/pmulq_h.h
@@ -0,0 +1,7 @@
+P_RD_RS1_RS2_LOOP(16,16,16, {
+  if ((p_rs1 != INT16_MIN) | (p_rs2 != INT16_MIN)) {
+    p_rd = (p_rs1 * p_rs2) >> 15;
+  } else {
+    p_rd = INT16_MAX;
+  }
+})
\ No newline at end of file
diff --git a/riscv/insns/pmulqr_h.h b/riscv/insns/pmulqr_h.h
new file mode 100644
index 00000000..8dc8b472
--- /dev/null
+++ b/riscv/insns/pmulqr_h.h
@@ -0,0 +1,7 @@
+P_RD_RS1_RS2_LOOP(16,16,16, {
+  if ((p_rs1 != INT16_MIN) | (p_rs2 != INT16_MIN)) {
+    p_rd = (((p_rs1 * p_rs2) >> 14) + 1) >> 1;
+  } else {
+    p_rd = INT16_MAX;
+  }
+})
\ No newline at end of file
diff --git a/riscv/insns/pmulsu_h_b00.h b/riscv/insns/pmulsu_h_b00.h
new file mode 100644
index 00000000..4682257e
--- /dev/null
+++ b/riscv/insns/pmulsu_h_b00.h
@@ -0,0 +1,3 @@
+P_RD_RS1_RS2_EE_SULOOP(16, 8, 8, {
+  p_rd = sext32(p_rs1) * zext32(p_rs2);
+})
\ No newline at end of file
diff --git a/riscv/insns/pmulsu_h_b11.h b/riscv/insns/pmulsu_h_b11.h
new file mode 100644
index 00000000..57436181
--- /dev/null
+++ b/riscv/insns/pmulsu_h_b11.h
@@ -0,0 +1,3 @@
+P_RD_RS1_RS2_OO_SULOOP(16, 8, 8, {
+  p_rd = sext32(p_rs1) * zext32(p_rs2);
+})
\ No newline at end of file
diff --git a/riscv/insns/pmulu_h_b00.h b/riscv/insns/pmulu_h_b00.h
new file mode 100644
index 00000000..8ff863f0
--- /dev/null
+++ b/riscv/insns/pmulu_h_b00.h
@@ -0,0 +1,3 @@
+P_RD_RS1_RS2_EE_ULOOP(16, 8, 8, {
+  p_rd = zext32(p_rs1) * zext32(p_rs2);
+})
\ No newline at end of file
diff --git a/riscv/insns/pmulu_h_b01.h b/riscv/insns/pmulu_h_b01.h
new file mode 100644
index 00000000..24d782cf
--- /dev/null
+++ b/riscv/insns/pmulu_h_b01.h
@@ -0,0 +1,3 @@
+P_RD_RS1_RS2_EO_ULOOP(16, 8, 8, {
+  p_rd = zext32(p_rs1) * zext32(p_rs2);
+})
\ No newline at end of file
diff --git a/riscv/insns/pmulu_h_b11.h b/riscv/insns/pmulu_h_b11.h
new file mode 100644
index 00000000..31316a1d
--- /dev/null
+++ b/riscv/insns/pmulu_h_b11.h
@@ -0,0 +1,3 @@
+P_RD_RS1_RS2_OO_ULOOP(16, 8, 8, {
+  p_rd = zext32(p_rs1) * zext32(p_rs2);
+})
\ No newline at end of file
diff --git a/riscv/insns/ppaire_b.h b/riscv/insns/ppaire_b.h
new file mode 100644
index 00000000..a1b9665a
--- /dev/null
+++ b/riscv/insns/ppaire_b.h
@@ -0,0 +1 @@
+P_PACK(8, 0, 0);
\ No newline at end of file
diff --git a/riscv/insns/ppaireo_b.h b/riscv/insns/ppaireo_b.h
new file mode 100644
index 00000000..1e02dc4c
--- /dev/null
+++ b/riscv/insns/ppaireo_b.h
@@ -0,0 +1 @@
+P_PACK(8, 0, 1);
\ No newline at end of file
diff --git a/riscv/insns/ppaireo_h.h b/riscv/insns/ppaireo_h.h
new file mode 100644
index 00000000..c5f043f5
--- /dev/null
+++ b/riscv/insns/ppaireo_h.h
@@ -0,0 +1 @@
+P_PACK(16, 0, 1);
\ No newline at end of file
diff --git a/riscv/insns/ppairo_b.h b/riscv/insns/ppairo_b.h
new file mode 100644
index 00000000..c691850b
--- /dev/null
+++ b/riscv/insns/ppairo_b.h
@@ -0,0 +1 @@
+P_PACK(8, 1, 1);
\ No newline at end of file
diff --git a/riscv/insns/ppairo_h.h b/riscv/insns/ppairo_h.h
new file mode 100644
index 00000000..600ee873
--- /dev/null
+++ b/riscv/insns/ppairo_h.h
@@ -0,0 +1 @@
+P_PACK(16, 1, 1);
\ No newline at end of file
diff --git a/riscv/insns/ppairoe_b.h b/riscv/insns/ppairoe_b.h
new file mode 100644
index 00000000..37df7381
--- /dev/null
+++ b/riscv/insns/ppairoe_b.h
@@ -0,0 +1 @@
+P_PACK(8, 1, 0);
\ No newline at end of file
diff --git a/riscv/insns/ppairoe_h.h b/riscv/insns/ppairoe_h.h
new file mode 100644
index 00000000..ecca5386
--- /dev/null
+++ b/riscv/insns/ppairoe_h.h
@@ -0,0 +1 @@
+P_PACK(16, 1, 0);
\ No newline at end of file
diff --git a/riscv/insns/predsum_bs.h b/riscv/insns/predsum_bs.h
new file mode 100644
index 00000000..9f110b6d
--- /dev/null
+++ b/riscv/insns/predsum_bs.h
@@ -0,0 +1,5 @@
+reg_t rd_tmp = RS2; \
+P_RS1_LOOP_BASE(8)
+  P_RS1_PARAMS(8)
+  rd_tmp += sext_xlen(p_rs1);
+P_RD_LOOP_END()
\ No newline at end of file
diff --git a/riscv/insns/predsum_hs.h b/riscv/insns/predsum_hs.h
new file mode 100644
index 00000000..dd2c8242
--- /dev/null
+++ b/riscv/insns/predsum_hs.h
@@ -0,0 +1,5 @@
+reg_t rd_tmp = RS2; \
+P_RS1_LOOP_BASE(16)
+  P_RS1_PARAMS(16)
+  rd_tmp += sext_xlen(p_rs1);
+P_RD_LOOP_END()
\ No newline at end of file
diff --git a/riscv/insns/predsumu_bs.h b/riscv/insns/predsumu_bs.h
new file mode 100644
index 00000000..e5c2f53a
--- /dev/null
+++ b/riscv/insns/predsumu_bs.h
@@ -0,0 +1,5 @@
+reg_t rd_tmp = RS2; \
+P_RS1_LOOP_BASE(8)
+  P_RS1_UPARAMS(8)
+  rd_tmp += zext_xlen(p_rs1);
+P_RD_LOOP_END()
\ No newline at end of file
diff --git a/riscv/insns/predsumu_hs.h b/riscv/insns/predsumu_hs.h
new file mode 100644
index 00000000..bf2dc457
--- /dev/null
+++ b/riscv/insns/predsumu_hs.h
@@ -0,0 +1,5 @@
+reg_t rd_tmp = RS2; \
+P_RS1_LOOP_BASE(16)
+  P_RS1_UPARAMS(16)
+  rd_tmp += zext_xlen(p_rs1);
+P_RD_LOOP_END()
\ No newline at end of file
diff --git a/riscv/insns/psa_hx.h b/riscv/insns/psa_hx.h
new file mode 100644
index 00000000..864671e4
--- /dev/null
+++ b/riscv/insns/psa_hx.h
@@ -0,0 +1,5 @@
+P_CROSS_LOOP(16, {
+  p_rd = p_rs1 - p_rs2;
+}, {
+  p_rd = p_rs1 + p_rs2;
+})
\ No newline at end of file
diff --git a/riscv/insns/psabs_b.h b/riscv/insns/psabs_b.h
new file mode 100644
index 00000000..8ca5085b
--- /dev/null
+++ b/riscv/insns/psabs_b.h
@@ -0,0 +1,5 @@
+P_RD_RS1_LOOP(8, 8, {
+  sreg_t abs_val = p_rs1 > 0 ? p_rs1 : -sext32(p_rs1);
+  p_rd = P_SAT(8, abs_val);
+  if (p_rd != abs_val) P.VU.vxsat->write(1);
+})
\ No newline at end of file
diff --git a/riscv/insns/psabs_h.h b/riscv/insns/psabs_h.h
new file mode 100644
index 00000000..54564afe
--- /dev/null
+++ b/riscv/insns/psabs_h.h
@@ -0,0 +1,5 @@
+P_RD_RS1_LOOP(16, 16, {
+  sreg_t abs_val = p_rs1 > 0 ? p_rs1 : -sext32(p_rs1);
+  p_rd = P_SAT(16, abs_val);
+  if (p_rd != abs_val) P.VU.vxsat->write(1);
+})
\ No newline at end of file
diff --git a/riscv/insns/psadd_b.h b/riscv/insns/psadd_b.h
new file mode 100644
index 00000000..a024d4d0
--- /dev/null
+++ b/riscv/insns/psadd_b.h
@@ -0,0 +1,4 @@
+P_RD_RS1_RS2_LOOP(8,8,8, {
+  bool sat = false;
+  p_rd = (sat_add<int8_t, uint8_t>(p_rs1, p_rs2, sat));
+})
\ No newline at end of file
diff --git a/riscv/insns/psadd_h.h b/riscv/insns/psadd_h.h
new file mode 100644
index 00000000..47af97ad
--- /dev/null
+++ b/riscv/insns/psadd_h.h
@@ -0,0 +1,4 @@
+P_RD_RS1_RS2_LOOP(16,16,16, {
+  bool sat = false;
+  p_rd = (sat_add<int16_t, uint16_t>(p_rs1, p_rs2, sat));
+})
\ No newline at end of file
diff --git a/riscv/insns/psaddu_b.h b/riscv/insns/psaddu_b.h
new file mode 100644
index 00000000..491fc2d3
--- /dev/null
+++ b/riscv/insns/psaddu_b.h
@@ -0,0 +1,4 @@
+P_RD_RS1_RS2_ULOOP(8,8,8, {
+  bool sat = false;
+  p_rd = (sat_addu<uint8_t>(p_rs1, p_rs2, sat));
+})
\ No newline at end of file
diff --git a/riscv/insns/psaddu_h.h b/riscv/insns/psaddu_h.h
new file mode 100644
index 00000000..f28d0f32
--- /dev/null
+++ b/riscv/insns/psaddu_h.h
@@ -0,0 +1,4 @@
+P_RD_RS1_RS2_ULOOP(16,16,16, {
+  bool sat = false;
+  p_rd = (sat_addu<uint16_t>(p_rs1, p_rs2, sat));
+})
\ No newline at end of file
diff --git a/riscv/insns/psas_hx.h b/riscv/insns/psas_hx.h
new file mode 100644
index 00000000..3be87430
--- /dev/null
+++ b/riscv/insns/psas_hx.h
@@ -0,0 +1,7 @@
+P_CROSS_ULOOP(16, {
+  bool sat = false;
+  p_rd = (sat_add<int16_t, uint16_t>(p_rs1, p_rs2, sat));
+}, {
+  bool sat = false;
+  p_rd = (sat_sub<int16_t, uint16_t>(p_rs1, p_rs2, sat));
+})
\ No newline at end of file
diff --git a/riscv/insns/psati_h.h b/riscv/insns/psati_h.h
new file mode 100644
index 00000000..b63f2233
--- /dev/null
+++ b/riscv/insns/psati_h.h
@@ -0,0 +1,3 @@
+P_RD_RS1_LOOP(16, 16, {
+  p_rd = P_SAT(insn.shamth() + 1, p_rs1);
+})
\ No newline at end of file
diff --git a/riscv/insns/psext_h_b.h b/riscv/insns/psext_h_b.h
new file mode 100644
index 00000000..1bc3dd0b
--- /dev/null
+++ b/riscv/insns/psext_h_b.h
@@ -0,0 +1,3 @@
+P_RD_RS1_LOOP(16, 16, {
+  p_rd = (int16_t)(int8_t)p_rs1;
+})
\ No newline at end of file
diff --git a/riscv/insns/psh1add_h.h b/riscv/insns/psh1add_h.h
new file mode 100644
index 00000000..db41f6e4
--- /dev/null
+++ b/riscv/insns/psh1add_h.h
@@ -0,0 +1,3 @@
+P_RD_RS1_RS2_LOOP(16, 16, 16, {
+  p_rd = (p_rs1 << 1) + p_rs2;
+})
\ No newline at end of file
diff --git a/riscv/insns/psll_bs.h b/riscv/insns/psll_bs.h
new file mode 100644
index 00000000..de0fa0df
--- /dev/null
+++ b/riscv/insns/psll_bs.h
@@ -0,0 +1,3 @@
+P_RD_RS1_LOOP(8, 8, {
+  p_rd = p_rs1 << (RS2 & (8 - 1));
+})
\ No newline at end of file
diff --git a/riscv/insns/psll_hs.h b/riscv/insns/psll_hs.h
new file mode 100644
index 00000000..5470506f
--- /dev/null
+++ b/riscv/insns/psll_hs.h
@@ -0,0 +1,3 @@
+P_RD_RS1_LOOP(16, 16, {
+  p_rd = p_rs1 << (RS2 & (16 - 1));
+})
\ No newline at end of file
diff --git a/riscv/insns/pslli_b.h b/riscv/insns/pslli_b.h
new file mode 100644
index 00000000..79f8ebc5
--- /dev/null
+++ b/riscv/insns/pslli_b.h
@@ -0,0 +1,3 @@
+P_RD_RS1_LOOP(8, 8, {
+  p_rd = p_rs1 << insn.shamtb();
+})
\ No newline at end of file
diff --git a/riscv/insns/pslli_h.h b/riscv/insns/pslli_h.h
new file mode 100644
index 00000000..65127f7e
--- /dev/null
+++ b/riscv/insns/pslli_h.h
@@ -0,0 +1,3 @@
+P_RD_RS1_LOOP(16, 16, {
+  p_rd = p_rs1 << insn.shamth();
+})
\ No newline at end of file
diff --git a/riscv/insns/psra_bs.h b/riscv/insns/psra_bs.h
new file mode 100644
index 00000000..bdc00144
--- /dev/null
+++ b/riscv/insns/psra_bs.h
@@ -0,0 +1,3 @@
+P_RD_RS1_LOOP(8, 8, {
+  p_rd = p_rs1 >> (RS2 & (8 - 1));
+})
\ No newline at end of file
diff --git a/riscv/insns/psra_hs.h b/riscv/insns/psra_hs.h
new file mode 100644
index 00000000..3a719ac2
--- /dev/null
+++ b/riscv/insns/psra_hs.h
@@ -0,0 +1,3 @@
+P_RD_RS1_LOOP(16, 16, {
+  p_rd = p_rs1 >> (RS2 & (16 - 1));
+})
\ No newline at end of file
diff --git a/riscv/insns/psrai_b.h b/riscv/insns/psrai_b.h
new file mode 100644
index 00000000..dd109927
--- /dev/null
+++ b/riscv/insns/psrai_b.h
@@ -0,0 +1,3 @@
+P_RD_RS1_LOOP(8, 8, {
+  p_rd = p_rs1 >> insn.shamtb();
+})
\ No newline at end of file
diff --git a/riscv/insns/psrai_h.h b/riscv/insns/psrai_h.h
new file mode 100644
index 00000000..89a1a686
--- /dev/null
+++ b/riscv/insns/psrai_h.h
@@ -0,0 +1,3 @@
+P_RD_RS1_LOOP(16, 16, {
+  p_rd = p_rs1 >> insn.shamth();
+})
\ No newline at end of file
diff --git a/riscv/insns/psrari_h.h b/riscv/insns/psrari_h.h
new file mode 100644
index 00000000..ace18a9a
--- /dev/null
+++ b/riscv/insns/psrari_h.h
@@ -0,0 +1,3 @@
+P_RD_RS1_LOOP(16, 16, {
+  p_rd = insn.shamth() ? ((p_rs1 >> insn.shamth()) + ((p_rs1 >> (insn.shamth() - 1)) & 1)) : p_rs1;
+})
\ No newline at end of file
diff --git a/riscv/insns/psrl_bs.h b/riscv/insns/psrl_bs.h
new file mode 100644
index 00000000..5e3c5953
--- /dev/null
+++ b/riscv/insns/psrl_bs.h
@@ -0,0 +1,3 @@
+P_RD_RS1_ULOOP(8, 8, {
+  p_rd = p_rs1 >> (RS2 & (8 - 1));
+})
\ No newline at end of file
diff --git a/riscv/insns/psrl_hs.h b/riscv/insns/psrl_hs.h
new file mode 100644
index 00000000..d3389c05
--- /dev/null
+++ b/riscv/insns/psrl_hs.h
@@ -0,0 +1,3 @@
+P_RD_RS1_ULOOP(16, 16, {
+  p_rd = p_rs1 >> (RS2 & (16 - 1));
+})
\ No newline at end of file
diff --git a/riscv/insns/psrli_b.h b/riscv/insns/psrli_b.h
new file mode 100644
index 00000000..68b79e06
--- /dev/null
+++ b/riscv/insns/psrli_b.h
@@ -0,0 +1,3 @@
+P_RD_RS1_ULOOP(8, 8, {
+  p_rd = p_rs1 >> insn.shamtb();
+})
\ No newline at end of file
diff --git a/riscv/insns/psrli_h.h b/riscv/insns/psrli_h.h
new file mode 100644
index 00000000..b7b2a763
--- /dev/null
+++ b/riscv/insns/psrli_h.h
@@ -0,0 +1,3 @@
+P_RD_RS1_ULOOP(16, 16, {
+  p_rd = p_rs1 >> insn.shamth();
+})
\ No newline at end of file
diff --git a/riscv/insns/pssa_hx.h b/riscv/insns/pssa_hx.h
new file mode 100644
index 00000000..0a8b7c7b
--- /dev/null
+++ b/riscv/insns/pssa_hx.h
@@ -0,0 +1,7 @@
+P_CROSS_ULOOP(16, {
+  bool sat = false;
+  p_rd = (sat_sub<int16_t, uint16_t>(p_rs1, p_rs2, sat));
+}, {
+  bool sat = false;
+  p_rd = (sat_add<int16_t, uint16_t>(p_rs1, p_rs2, sat));
+})
\ No newline at end of file
diff --git a/riscv/insns/pssh1sadd_h.h b/riscv/insns/pssh1sadd_h.h
new file mode 100644
index 00000000..dead4c71
--- /dev/null
+++ b/riscv/insns/pssh1sadd_h.h
@@ -0,0 +1,3 @@
+P_RD_RS1_RS2_LOOP(16, 16, 16, {
+  p_rd = P_SAT(16, P_SAT(16, p_rs1 << 1) + p_rs2);
+})
\ No newline at end of file
diff --git a/riscv/insns/pssha_hs.h b/riscv/insns/pssha_hs.h
new file mode 100644
index 00000000..34d9852e
--- /dev/null
+++ b/riscv/insns/pssha_hs.h
@@ -0,0 +1,11 @@
+sreg_t sshamt = P_FIELD(RS2, 0, 8);
+P_RD_RS1_LOOP(16, 16, {
+  if (p_rs1 == 0)
+    p_rd = 0;
+  else if (sshamt >= 16)
+    p_rd = (p_rs1 & 0x8000) ? 0x8000 : 0x7fff;
+  else if (sshamt <= -16)
+    p_rd = (p_rs1 & 0x8000) ? 0xffff : 0;
+  else
+    p_rd = sshamt >= 0 ? P_SAT(16, sext32(p_rs1) << sshamt) : (p_rs1 >> -sshamt);
+})
\ No newline at end of file
diff --git a/riscv/insns/psshar_hs.h b/riscv/insns/psshar_hs.h
new file mode 100644
index 00000000..73d220aa
--- /dev/null
+++ b/riscv/insns/psshar_hs.h
@@ -0,0 +1,11 @@
+sreg_t sshamt = P_FIELD(RS2, 0, 8);
+P_RD_RS1_LOOP(16, 16, {
+  if (p_rs1 == 0)
+    p_rd = 0;
+  else if (sshamt >= 16)
+    p_rd = (p_rs1 & 0x8000) ? 0x8000 : 0x7fff;
+  else if (sshamt <= -16)
+    p_rd = 0;
+  else
+    p_rd = sshamt >= 0 ? P_SAT(16, sext32(p_rs1) << sshamt) : ((p_rs1 >> -sshamt) + ((p_rs1 >> (-sshamt - 1)) & 1));
+})
\ No newline at end of file
diff --git a/riscv/insns/psslai_h.h b/riscv/insns/psslai_h.h
new file mode 100644
index 00000000..b5ae37e1
--- /dev/null
+++ b/riscv/insns/psslai_h.h
@@ -0,0 +1,3 @@
+P_RD_RS1_LOOP(16, 16, {
+  p_rd = P_SAT(16, sext32(p_rs1) << insn.shamth());
+})
\ No newline at end of file
diff --git a/riscv/insns/pssub_b.h b/riscv/insns/pssub_b.h
new file mode 100644
index 00000000..91a89aa1
--- /dev/null
+++ b/riscv/insns/pssub_b.h
@@ -0,0 +1,4 @@
+P_RD_RS1_RS2_LOOP(8,8,8, {
+  bool sat = false;
+  p_rd = (sat_sub<int8_t, uint8_t>(p_rs1, p_rs2, sat));
+})
\ No newline at end of file
diff --git a/riscv/insns/pssub_h.h b/riscv/insns/pssub_h.h
new file mode 100644
index 00000000..db88d680
--- /dev/null
+++ b/riscv/insns/pssub_h.h
@@ -0,0 +1,4 @@
+P_RD_RS1_RS2_LOOP(16,16,16, {
+  bool sat = false;
+  p_rd = (sat_sub<int16_t, uint16_t>(p_rs1, p_rs2, sat));
+})
\ No newline at end of file
diff --git a/riscv/insns/pssubu_b.h b/riscv/insns/pssubu_b.h
new file mode 100644
index 00000000..3fd44bee
--- /dev/null
+++ b/riscv/insns/pssubu_b.h
@@ -0,0 +1,4 @@
+P_RD_RS1_RS2_ULOOP(8,8,8, {
+  bool sat = false;
+  p_rd = (sat_subu<uint8_t>(p_rs1, p_rs2, sat));
+})
\ No newline at end of file
diff --git a/riscv/insns/pssubu_h.h b/riscv/insns/pssubu_h.h
new file mode 100644
index 00000000..de3ff72a
--- /dev/null
+++ b/riscv/insns/pssubu_h.h
@@ -0,0 +1,4 @@
+P_RD_RS1_RS2_ULOOP(16,16,16, {
+  bool sat = false;
+  p_rd = (sat_subu<uint16_t>(p_rs1, p_rs2, sat));
+})
\ No newline at end of file
diff --git a/riscv/insns/psub_b.h b/riscv/insns/psub_b.h
new file mode 100644
index 00000000..1dce95cf
--- /dev/null
+++ b/riscv/insns/psub_b.h
@@ -0,0 +1,3 @@
+P_RD_RS1_RS2_LOOP(8, 8, 8, {
+  p_rd = p_rs1 - p_rs2;
+})
\ No newline at end of file
diff --git a/riscv/insns/psub_h.h b/riscv/insns/psub_h.h
new file mode 100644
index 00000000..a2d7c3b1
--- /dev/null
+++ b/riscv/insns/psub_h.h
@@ -0,0 +1,3 @@
+P_RD_RS1_RS2_LOOP(16, 16, 16, {
+  p_rd = p_rs1 - p_rs2;
+})
\ No newline at end of file
diff --git a/riscv/insns/pusati_h.h b/riscv/insns/pusati_h.h
new file mode 100644
index 00000000..d5857448
--- /dev/null
+++ b/riscv/insns/pusati_h.h
@@ -0,0 +1,3 @@
+P_RD_RS1_LOOP(16, 16, {
+  p_rd = P_USAT(insn.shamth() + 1, p_rs1);
+})
\ No newline at end of file
diff --git a/riscv/insns/sadd.h b/riscv/insns/sadd.h
new file mode 100644
index 00000000..94cc264c
--- /dev/null
+++ b/riscv/insns/sadd.h
@@ -0,0 +1,3 @@
+require_extension('P');
+require_rv32;
+WRITE_RD(P_SAT(xlen, sext_xlen((RS1 << 1) + RS2)));
\ No newline at end of file
diff --git a/riscv/insns/saddu.h b/riscv/insns/saddu.h
new file mode 100644
index 00000000..dec16b84
--- /dev/null
+++ b/riscv/insns/saddu.h
@@ -0,0 +1,4 @@
+require_extension('P');
+require_rv32;
+bool sat = false;
+WRITE_RD(sat_addu<uint32_t>(RS1, RS2, sat));
\ No newline at end of file
diff --git a/riscv/insns/sati_rv32.h b/riscv/insns/sati_rv32.h
new file mode 100644
index 00000000..a950d95c
--- /dev/null
+++ b/riscv/insns/sati_rv32.h
@@ -0,0 +1,3 @@
+require_extension('P');
+require_rv32;
+WRITE_RD(P_SAT(insn.shamtw() + 1, (sreg_t)RS1));
\ No newline at end of file
diff --git a/riscv/insns/slx.h b/riscv/insns/slx.h
new file mode 100644
index 00000000..a6eb3e9c
--- /dev/null
+++ b/riscv/insns/slx.h
@@ -0,0 +1,16 @@
+require_extension('P');
+if(xlen == 64){
+    int shamt = RS2 & 63;
+    if(shamt == 0){
+        WRITE_RD((uint64_t)RD);
+    }else{
+        WRITE_RD(((uint64_t)RS1 >> (64 - shamt)) | ((uint64_t)RD << shamt));
+    }
+}else{
+    int shamt = (uint32_t)RS2 & 31;
+    if (shamt == 0) {
+        WRITE_RD(sext_xlen((uint32_t)RD));
+    } else {
+        WRITE_RD(sext_xlen(((uint32_t)RS1 >> (32 - shamt)) | ((uint32_t)RD << shamt)));
+    }
+}
diff --git a/riscv/insns/srari_rv32.h b/riscv/insns/srari_rv32.h
new file mode 100644
index 00000000..25d4bee4
--- /dev/null
+++ b/riscv/insns/srari_rv32.h
@@ -0,0 +1,3 @@
+require_extension('P');
+require_rv32;
+WRITE_RD(insn.shamtw() ? ((sext_xlen(RS1) >> insn.shamtw()) + ((sext_xlen(RS1) >> (insn.shamtw() - 1)) & 1)) : RS1);
\ No newline at end of file
diff --git a/riscv/insns/srx.h b/riscv/insns/srx.h
new file mode 100644
index 00000000..19460c74
--- /dev/null
+++ b/riscv/insns/srx.h
@@ -0,0 +1,16 @@
+require_extension('P');
+if(xlen == 64){
+    int shamt = RS2 & 63;
+    if(shamt == 0){
+        WRITE_RD((uint64_t)RD);
+    }else{
+        WRITE_RD(((uint64_t)RS1 << (64 - shamt)) | ((uint64_t)RD >> shamt));
+    }
+}else{
+    int shamt = (uint32_t)RS2 & 31;
+    if (shamt == 0) {
+        WRITE_RD(sext_xlen((uint32_t)RD));
+    } else {
+        WRITE_RD(sext_xlen(((uint32_t)RS1 << (32 - shamt)) | ((uint32_t)RD >> shamt)));
+    }
+}
diff --git a/riscv/insns/ssh1sadd.h b/riscv/insns/ssh1sadd.h
new file mode 100644
index 00000000..94cc264c
--- /dev/null
+++ b/riscv/insns/ssh1sadd.h
@@ -0,0 +1,3 @@
+require_extension('P');
+require_rv32;
+WRITE_RD(P_SAT(xlen, sext_xlen((RS1 << 1) + RS2)));
\ No newline at end of file
diff --git a/riscv/insns/ssha.h b/riscv/insns/ssha.h
new file mode 100644
index 00000000..ed1247f6
--- /dev/null
+++ b/riscv/insns/ssha.h
@@ -0,0 +1,11 @@
+require_extension('P');
+require_rv32;
+sreg_t sshamt = P_FIELD(RS2, 0, 8);
+if (RS1 == 0)
+  WRITE_RD(0);
+else if (sshamt >= 32)
+  WRITE_RD((RS1 & 0x80000000) ? 0x80000000 : 0x7fffffff);
+else if (sshamt <= -32)
+  WRITE_RD((RS1 & 0x80000000) ? 0xffffffff : 0);
+else
+  WRITE_RD(sshamt >= 0 ? P_SAT(32, static_cast<sreg_t> (RS1) << sshamt) : (RS1 >> -sshamt));
\ No newline at end of file
diff --git a/riscv/insns/sshar.h b/riscv/insns/sshar.h
new file mode 100644
index 00000000..d602edb7
--- /dev/null
+++ b/riscv/insns/sshar.h
@@ -0,0 +1,11 @@
+require_extension('P');
+require_rv32;
+sreg_t sshamt = P_FIELD(RS2, 0, 8);
+if (RS1 == 0)
+  WRITE_RD(0);
+else if (sshamt >= 32)
+  WRITE_RD((RS1 & 0x80000000) ? 0x80000000 : 0x7fffffff);
+else if (sshamt <= -32)
+  WRITE_RD(0);
+else
+  WRITE_RD(sshamt >= 0 ? P_SAT(32, static_cast<sreg_t> (RS1) << sshamt) : ((RS1 >> -sshamt) + ((RS1 >> (-sshamt - 1)) & 1)));
\ No newline at end of file
diff --git a/riscv/insns/sslai.h b/riscv/insns/sslai.h
new file mode 100644
index 00000000..5db6ab1a
--- /dev/null
+++ b/riscv/insns/sslai.h
@@ -0,0 +1,3 @@
+require_extension('P');
+require_rv32;
+WRITE_RD(P_SAT(32, static_cast<sreg_t> (RS1) << insn.shamtw()));
\ No newline at end of file
diff --git a/riscv/insns/ssub.h b/riscv/insns/ssub.h
new file mode 100644
index 00000000..f3db9b4b
--- /dev/null
+++ b/riscv/insns/ssub.h
@@ -0,0 +1,5 @@
+require_extension('P');
+require_rv32;
+bool sat = false;
+int32_t p_rd = sat_sub<int32_t, uint32_t>(RS1, RS2, sat);
+WRITE_RD(p_rd);
\ No newline at end of file
diff --git a/riscv/insns/ssubu.h b/riscv/insns/ssubu.h
new file mode 100644
index 00000000..c4aac5d3
--- /dev/null
+++ b/riscv/insns/ssubu.h
@@ -0,0 +1,4 @@
+require_extension('P');
+require_rv32;
+bool sat = false;
+WRITE_RD(sat_subu<uint32_t>(RS1, RS2, sat));
\ No newline at end of file
diff --git a/riscv/insns/usati_rv32.h b/riscv/insns/usati_rv32.h
new file mode 100644
index 00000000..84ba83a1
--- /dev/null
+++ b/riscv/insns/usati_rv32.h
@@ -0,0 +1,3 @@
+require_extension('P');
+require_rv32;
+WRITE_RD(P_USAT(insn.shamtw() + 1, (sreg_t)RS1));
\ No newline at end of file
diff --git a/riscv/p_ext_macros.h b/riscv/p_ext_macros.h
new file mode 100644
index 00000000..40bf1f6f
--- /dev/null
+++ b/riscv/p_ext_macros.h
@@ -0,0 +1,512 @@
+#ifndef _RISCV_P_EXT_MACROS_H_
+#define _RISCV_P_EXT_MACROS_H_
+
+// rd temp
+#define WRITE_P_RD() \
+  rd_tmp = set_field(rd_tmp, make_mask64((i * sizeof(p_rd) * 8), sizeof(p_rd) * 8), p_rd);
+
+// Field
+#define P_FIELD(R, INDEX, SIZE) \
+  (type_sew_t<SIZE>::type)get_field(R, make_mask64(((INDEX) * SIZE), SIZE))
+
+#define P_UFIELD(R, INDEX, SIZE) \
+  (type_usew_t<SIZE>::type)get_field(R, make_mask64(((INDEX) * SIZE), SIZE))
+
+// Params
+#define P_RD_PARAMS(BIT) \
+  auto p_rd = P_FIELD(rd_tmp, i, BIT);
+
+#define P_RD_UPARAMS(BIT) \
+  auto p_rd = P_UFIELD(rd_tmp, i, BIT);  
+
+#define P_RS1_PARAMS(BIT) \
+  auto p_rs1 = P_FIELD(rs1, i, BIT);
+
+#define P_RS1_UPARAMS(BIT) \
+  auto p_rs1 = P_UFIELD(rs1, i, BIT);
+
+#define P_RS1_INNER_PARAMS(BIT_INNER) \
+  auto p_rs1 = P_FIELD(rs1, j, BIT_INNER);
+
+#define P_RS1_INNER_UPARAMS(BIT_INNER) \
+  auto p_rs1 = P_UFIELD(rs1, j, BIT_INNER);
+
+#define P_RS1_EVEN_PARAMS(BIT) \
+  auto p_rs1 = P_FIELD(rs1, i * 2, BIT);
+
+#define P_RS1_ODD_PARAMS(BIT) \
+  auto p_rs1 = P_FIELD(rs1, i * 2 + 1, BIT);
+
+#define P_RS1_EVEN_UPARAMS(BIT) \
+  auto p_rs1 = P_UFIELD(rs1, i * 2, BIT);
+
+#define P_RS1_ODD_UPARAMS(BIT) \
+  auto p_rs1 = P_UFIELD(rs1, i * 2 + 1, BIT);
+
+#define P_RS2_PARAMS(BIT) \
+  auto p_rs2 = P_FIELD(rs2, i, BIT);  
+
+#define P_RS2_UPARAMS(BIT) \
+  auto p_rs2 = P_UFIELD(rs2, i, BIT);  
+
+#define P_RS2_CROSS_PARAMS(BIT) \
+  auto p_rs2 = P_FIELD(rs2, (i ^ 1), BIT); 
+ 
+#define P_RS2_CROSS_UPARAMS(BIT) \
+  auto p_rs2 = P_UFIELD(rs2, (i ^ 1), BIT);
+
+#define P_RS2_INNER_PARAMS(BIT_INNER) \
+  auto p_rs2 = P_FIELD(rs2, j, BIT_INNER);
+
+#define P_RS2_INNER_UPARAMS(BIT_INNER) \
+  auto p_rs2 = P_UFIELD(rs2, j, BIT_INNER);
+
+#define P_RS2_INNER_CROSS_PARAMS(BIT_INNER) \
+  auto p_rs2 = P_FIELD(rs2, (j ^ 1), BIT_INNER);
+
+#define P_RS2_EVEN_PARAMS(BIT) \
+  auto p_rs2 = P_FIELD(rs2, i * 2, BIT);
+
+#define P_RS2_ODD_PARAMS(BIT) \
+  auto p_rs2 = P_FIELD(rs2, i * 2 + 1, BIT);  
+
+#define P_RS2_EVEN_UPARAMS(BIT) \
+  auto p_rs2 = P_UFIELD(rs2, i * 2, BIT);
+
+#define P_RS2_ODD_UPARAMS(BIT) \
+  auto p_rs2 = P_UFIELD(rs2, i * 2 + 1, BIT);
+
+// Loop base
+#define P_RD_LOOP_BASE(BIT) \
+  require_extension('P'); \
+  require((BIT) == e8 || (BIT) == e16 || (BIT) == e32); \
+  reg_t rd_tmp = RD; \
+  sreg_t len = xlen / (BIT); \
+  for (sreg_t i = len - 1; i >= 0; --i) {
+
+#define P_RD_RS1_LOOP_BASE(BIT) \
+  require_extension('P'); \
+  require((BIT) == e8 || (BIT) == e16 || (BIT) == e32); \
+  reg_t rd_tmp = RD; \
+  reg_t rs1 = RS1; \
+  sreg_t len = xlen / (BIT); \
+  for (sreg_t i = len - 1; i >= 0; --i) {
+
+#define P_RD_RS1_RS2_LOOP_BASE(BIT) \
+  require_extension('P'); \
+  require((BIT) == e8 || (BIT) == e16 || (BIT) == e32); \
+  reg_t rd_tmp = RD; \
+  reg_t rs1 = RS1; \
+  reg_t rs2 = RS2; \
+  sreg_t len = xlen / (BIT); \
+  for (sreg_t i = len - 1; i >= 0; --i) {
+
+#define P_RS1_LOOP_BASE(BIT) \
+  require_extension('P'); \
+  require((BIT) == e8 || (BIT) == e16 || (BIT) == e32); \
+  reg_t rs1 = RS1; \
+  sreg_t len = xlen / (BIT); \
+  for (sreg_t i = len - 1; i >= 0; --i) {
+
+#define P_REDUCTION_LOOP_BASE(BIT, BIT_INNER, USE_RD) \
+  require_extension('P'); \
+  require(BIT == e16 || BIT == e32 || BIT == e64); \
+  reg_t rd_tmp = USE_RD ? zext_xlen(RD) : 0; \
+  reg_t rs1 = zext_xlen(RS1); \
+  reg_t rs2 = zext_xlen(RS2); \
+  sreg_t len = 64 / BIT; \
+  sreg_t len_inner = BIT / BIT_INNER; \
+  for (sreg_t i = len - 1; i >= 0; --i) { \
+    sreg_t p_res = P_FIELD(rd_tmp, i, BIT); \
+    for (sreg_t j = i * len_inner; j < (i + 1) * len_inner; ++j) {
+
+#define P_REDUCTION_ULOOP_BASE(BIT, BIT_INNER, USE_RD) \
+  require_extension('P'); \
+  require(BIT == e16 || BIT == e32 || BIT == e64); \
+  reg_t rd_tmp = USE_RD ? zext_xlen(RD) : 0; \
+  reg_t rs1 = zext_xlen(RS1); \
+  reg_t rs2 = zext_xlen(RS2); \
+  sreg_t len = 64 / BIT; \
+  sreg_t len_inner = BIT / BIT_INNER; \
+  for (sreg_t i = len - 1; i >= 0; --i) { \
+    sreg_t p_res = P_UFIELD(rd_tmp, i, BIT); \
+    for (sreg_t j = i * len_inner; j < (i + 1) * len_inner; ++j) {
+
+#define P_REDUCTION_ULOOP_BASE(BIT, BIT_INNER, USE_RD) \
+  require_extension('P'); \
+  require(BIT == e16 || BIT == e32 || BIT == e64); \
+  reg_t rd_tmp = USE_RD ? zext_xlen(RD) : 0; \
+  reg_t rs1 = zext_xlen(RS1); \
+  reg_t rs2 = zext_xlen(RS2); \
+  sreg_t len = 64 / BIT; \
+  sreg_t len_inner = BIT / BIT_INNER; \
+  for (sreg_t i = len - 1; i >= 0; --i) { \
+    sreg_t p_res = P_UFIELD(rd_tmp, i, BIT); \
+    for (sreg_t j = i * len_inner; j < (i + 1) * len_inner; ++j) {
+
+#define P_RD_DW_LOOP_BASE(BIT) \
+  require_extension('P'); \
+  require((BIT) == e8 || (BIT) == e16 || (BIT) == e32); \
+  reg_t rd_tmp = P_RD_PAIR; \
+  sreg_t len = xlen / (BIT) * 2; \
+  for (sreg_t i = len - 1; i >= 0; --i) {
+
+// Loop body
+#define P_RD_LOOP_BODY(BIT, BODY) { \
+  P_RD_PARAMS(BIT) \
+  BODY \
+  WRITE_P_RD(); \
+}
+
+#define P_RD_RS1_LOOP_BODY(BIT_RD, BIT_RS1, BODY) { \
+  P_RD_PARAMS(BIT_RD) \
+  P_RS1_PARAMS(BIT_RS1) \
+  BODY \
+  WRITE_P_RD(); \
+}
+
+#define P_RD_RS1_ULOOP_BODY(BIT_RD, BIT_RS1, BODY) { \
+  P_RD_UPARAMS(BIT_RD) \
+  P_RS1_UPARAMS(BIT_RS1) \
+  BODY \
+  WRITE_P_RD(); \
+}
+
+#define P_RD_RS1_RS2_LOOP_BODY(BIT_RD, BIT_RS1, BIT_RS2, BODY) { \
+  P_RD_PARAMS(BIT_RD) \
+  P_RS1_PARAMS(BIT_RS1) \
+  P_RS2_PARAMS(BIT_RS2) \
+  BODY \
+  WRITE_P_RD(); \
+}
+
+#define P_RD_RS1_RS2_ULOOP_BODY(BIT_RD, BIT_RS1, BIT_RS2, BODY) { \
+  P_RD_UPARAMS(BIT_RD) \
+  P_RS1_UPARAMS(BIT_RS1) \
+  P_RS2_UPARAMS(BIT_RS2) \
+  BODY \
+  WRITE_P_RD(); \
+}
+
+#define P_RD_RS1_RS2_SULOOP_BODY(BIT_RD, BIT_RS1, BIT_RS2, BODY) { \
+  P_RD_PARAMS(BIT_RD) \
+  P_RS1_PARAMS(BIT_RS1) \
+  P_RS2_UPARAMS(BIT_RS2) \
+  BODY \
+  WRITE_P_RD(); \
+}
+
+#define P_CROSS_LOOP_BODY(BIT, BODY) { \
+  P_RD_PARAMS(BIT) \
+  P_RS1_PARAMS(BIT) \
+  P_RS2_CROSS_PARAMS(BIT) \
+  BODY \
+  WRITE_P_RD(); \
+}
+
+#define P_CROSS_ULOOP_BODY(BIT, BODY) { \
+  P_RD_UPARAMS(BIT) \
+  P_RS1_UPARAMS(BIT) \
+  P_RS2_CROSS_UPARAMS(BIT) \
+  BODY \
+  WRITE_P_RD(); \
+}
+
+#define P_RD_RS1_RS2_EE_LOOP_BODY(BIT_RD, BIT_RS1, BIT_RS2, BODY) { \
+  P_RD_PARAMS(BIT_RD) \
+  P_RS1_EVEN_PARAMS(BIT_RS1) \
+  P_RS2_EVEN_PARAMS(BIT_RS2) \
+  BODY \
+  WRITE_P_RD(); \
+}
+
+#define P_RD_RS1_RS2_EO_LOOP_BODY(BIT_RD, BIT_RS1, BIT_RS2, BODY) { \
+  P_RD_PARAMS(BIT_RD) \
+  P_RS1_EVEN_PARAMS(BIT_RS1) \
+  P_RS2_ODD_PARAMS(BIT_RS2) \
+  BODY \
+  WRITE_P_RD(); \
+}
+
+#define P_RD_RS1_RS2_OO_LOOP_BODY(BIT_RD, BIT_RS1, BIT_RS2, BODY) { \
+  P_RD_PARAMS(BIT_RD) \
+  P_RS1_ODD_PARAMS(BIT_RS1) \
+  P_RS2_ODD_PARAMS(BIT_RS2) \
+  BODY \
+  WRITE_P_RD(); \
+}
+
+#define P_RD_RS1_RS2_EE_ULOOP_BODY(BIT_RD, BIT_RS1, BIT_RS2, BODY) { \
+  P_RD_UPARAMS(BIT_RD) \
+  P_RS1_EVEN_UPARAMS(BIT_RS1) \
+  P_RS2_EVEN_UPARAMS(BIT_RS2) \
+  BODY \
+  WRITE_P_RD(); \
+}
+
+#define P_RD_RS1_RS2_EO_ULOOP_BODY(BIT_RD, BIT_RS1, BIT_RS2, BODY) { \
+  P_RD_UPARAMS(BIT_RD) \
+  P_RS1_EVEN_UPARAMS(BIT_RS1) \
+  P_RS2_ODD_UPARAMS(BIT_RS2) \
+  BODY \
+  WRITE_P_RD(); \
+}
+
+#define P_RD_RS1_RS2_OO_ULOOP_BODY(BIT_RD, BIT_RS1, BIT_RS2, BODY) { \
+  P_RD_UPARAMS(BIT_RD) \
+  P_RS1_ODD_UPARAMS(BIT_RS1) \
+  P_RS2_ODD_UPARAMS(BIT_RS2) \
+  BODY \
+  WRITE_P_RD(); \
+}
+
+#define P_RD_RS1_RS2_EE_SULOOP_BODY(BIT_RD, BIT_RS1, BIT_RS2, BODY) { \
+  P_RD_PARAMS(BIT_RD) \
+  P_RS1_EVEN_PARAMS(BIT_RS1) \
+  P_RS2_EVEN_UPARAMS(BIT_RS2) \
+  BODY \
+  WRITE_P_RD(); \
+}
+
+#define P_RD_RS1_RS2_OO_SULOOP_BODY(BIT_RD, BIT_RS1, BIT_RS2, BODY) { \
+  P_RD_PARAMS(BIT_RD) \
+  P_RS1_ODD_PARAMS(BIT_RS1) \
+  P_RS2_ODD_UPARAMS(BIT_RS2) \
+  BODY \
+  WRITE_P_RD(); \
+}
+
+#define P_RD_RS1_RS2_E_LOOP_BODY(BIT_RD, BIT_RS1, BIT_RS2, BODY) { \
+  P_RD_PARAMS(BIT_RD) \
+  P_RS1_PARAMS(BIT_RS1) \
+  P_RS2_EVEN_PARAMS(BIT_RS2) \
+  BODY \
+  WRITE_P_RD(); \
+}
+
+#define P_RD_RS1_RS2_O_LOOP_BODY(BIT_RD, BIT_RS1, BIT_RS2, BODY) { \
+  P_RD_PARAMS(BIT_RD) \
+  P_RS1_PARAMS(BIT_RS1) \
+  P_RS2_ODD_PARAMS(BIT_RS2) \
+  BODY \
+  WRITE_P_RD(); \
+}
+
+#define P_RD_RS1_RS2_E_SULOOP_BODY(BIT_RD, BIT_RS1, BIT_RS2, BODY) { \
+  P_RD_PARAMS(BIT_RD) \
+  P_RS1_PARAMS(BIT_RS1) \
+  P_RS2_EVEN_UPARAMS(BIT_RS2) \
+  BODY \
+  WRITE_P_RD(); \
+}
+
+#define P_RD_RS1_RS2_O_SULOOP_BODY(BIT_RD, BIT_RS1, BIT_RS2, BODY) { \
+  P_RD_PARAMS(BIT_RD) \
+  P_RS1_PARAMS(BIT_RS1) \
+  P_RS2_ODD_UPARAMS(BIT_RS2) \
+  BODY \
+  WRITE_P_RD(); \
+}
+
+// Loop end
+#define P_RD_LOOP_END() \
+  } \
+  WRITE_RD(sext_xlen(rd_tmp));
+
+#define P_REDUCTION_LOOP_END(BIT, IS_SAT) \
+    } \
+    if (IS_SAT) { \
+      p_res = P_SAT(BIT, p_res); \
+    } \
+    type_usew_t<BIT>::type p_rd = p_res; \
+    WRITE_P_RD(); \
+  } \
+  WRITE_RD(sext_xlen(rd_tmp));
+
+#define P_REDUCTION_ULOOP_END(BIT, IS_SAT) \
+    } \
+    type_usew_t<BIT>::type p_rd = p_res; \
+    WRITE_P_RD(); \
+  } \
+  WRITE_RD(sext_xlen(rd_tmp));
+
+#define P_RD_DW_LOOP_END() \
+  } \
+  WRITE_P_RD_PAIR(rd_tmp);
+
+// Loop
+#define P_RD_LOOP(BIT_RD, BODY) \
+  P_RD_LOOP_BASE(BIT_RD) \
+  P_RD_LOOP_BODY(BIT_RD, BODY) \
+  P_RD_LOOP_END()
+
+#define P_RD_RS1_LOOP(BIT_RD, BIT_RS1, BODY) \
+  P_RD_RS1_LOOP_BASE(BIT_RD) \
+  P_RD_RS1_LOOP_BODY(BIT_RD, BIT_RS1, BODY) \
+  P_RD_LOOP_END()
+
+#define P_RD_RS1_ULOOP(BIT_RD, BIT_RS1, BODY) \
+  P_RD_RS1_LOOP_BASE(BIT_RD) \
+  P_RD_RS1_ULOOP_BODY(BIT_RD, BIT_RS1, BODY) \
+  P_RD_LOOP_END()
+
+#define P_RD_RS1_RS2_LOOP(BIT_RD, BIT_RS1, BIT_RS2, BODY) \
+  P_RD_RS1_RS2_LOOP_BASE(BIT_RD) \
+  P_RD_RS1_RS2_LOOP_BODY(BIT_RD, BIT_RS1, BIT_RS2, BODY) \
+  P_RD_LOOP_END()
+
+#define P_RD_RS1_RS2_ULOOP(BIT_RD, BIT_RS1, BIT_RS2, BODY) \
+  P_RD_RS1_RS2_LOOP_BASE(BIT_RD) \
+  P_RD_RS1_RS2_ULOOP_BODY(BIT_RD, BIT_RS1, BIT_RS2, BODY) \
+  P_RD_LOOP_END()
+
+#define P_RD_RS1_RS2_SULOOP(BIT_RD, BIT_RS1, BIT_RS2, BODY) \
+  P_RD_RS1_RS2_LOOP_BASE(BIT_RD) \
+  P_RD_RS1_RS2_SULOOP_BODY(BIT_RD, BIT_RS1, BIT_RS2, BODY) \
+  P_RD_LOOP_END()
+
+#define P_CROSS_LOOP(BIT, BODY1, BODY2) \
+  P_RD_RS1_RS2_LOOP_BASE(BIT) \
+  P_CROSS_LOOP_BODY(BIT, BODY1) \
+  --i; \
+  if (sizeof(#BODY2) == 1) { \
+    P_CROSS_LOOP_BODY(BIT, BODY1) \
+  } \
+  else { \
+    P_CROSS_LOOP_BODY(BIT, BODY2) \
+  } \
+  P_RD_LOOP_END()
+
+#define P_CROSS_ULOOP(BIT, BODY1, BODY2) \
+  P_RD_RS1_RS2_LOOP_BASE(BIT) \
+  P_CROSS_ULOOP_BODY(BIT, BODY1) \
+  --i; \
+  if (sizeof(#BODY2) == 1) { \
+    P_CROSS_ULOOP_BODY(BIT, BODY1) \
+  } \
+  else { \
+    P_CROSS_ULOOP_BODY(BIT, BODY2) \
+  } \
+  P_RD_LOOP_END()
+
+#define P_RD_RS1_RS2_EE_LOOP(BIT_RD, BIT_RS1, BIT_RS2, BODY) \
+  P_RD_RS1_RS2_LOOP_BASE(BIT_RD) \
+  P_RD_RS1_RS2_EE_LOOP_BODY(BIT_RD, BIT_RS1, BIT_RS2, BODY) \
+  P_RD_LOOP_END()
+
+#define P_RD_RS1_RS2_EO_LOOP(BIT_RD, BIT_RS1, BIT_RS2, BODY) \
+  P_RD_RS1_RS2_LOOP_BASE(BIT_RD) \
+  P_RD_RS1_RS2_EO_LOOP_BODY(BIT_RD, BIT_RS1, BIT_RS2, BODY) \
+  P_RD_LOOP_END()
+
+#define P_RD_RS1_RS2_OO_LOOP(BIT_RD, BIT_RS1, BIT_RS2, BODY) \
+  P_RD_RS1_RS2_LOOP_BASE(BIT_RD) \
+  P_RD_RS1_RS2_OO_LOOP_BODY(BIT_RD, BIT_RS1, BIT_RS2, BODY) \
+  P_RD_LOOP_END()
+
+#define P_RD_RS1_RS2_EE_ULOOP(BIT_RD, BIT_RS1, BIT_RS2, BODY) \
+  P_RD_RS1_RS2_LOOP_BASE(BIT_RD) \
+  P_RD_RS1_RS2_EE_ULOOP_BODY(BIT_RD, BIT_RS1, BIT_RS2, BODY) \
+  P_RD_LOOP_END()
+
+#define P_RD_RS1_RS2_EO_ULOOP(BIT_RD, BIT_RS1, BIT_RS2, BODY) \
+  P_RD_RS1_RS2_LOOP_BASE(BIT_RD) \
+  P_RD_RS1_RS2_EO_ULOOP_BODY(BIT_RD, BIT_RS1, BIT_RS2, BODY) \
+  P_RD_LOOP_END()
+
+#define P_RD_RS1_RS2_OO_ULOOP(BIT_RD, BIT_RS1, BIT_RS2, BODY) \
+  P_RD_RS1_RS2_LOOP_BASE(BIT_RD) \
+  P_RD_RS1_RS2_OO_ULOOP_BODY(BIT_RD, BIT_RS1, BIT_RS2, BODY) \
+  P_RD_LOOP_END()
+
+#define P_RD_RS1_RS2_EE_SULOOP(BIT_RD, BIT_RS1, BIT_RS2, BODY) \
+  P_RD_RS1_RS2_LOOP_BASE(BIT_RD) \
+  P_RD_RS1_RS2_EE_SULOOP_BODY(BIT_RD, BIT_RS1, BIT_RS2, BODY) \
+  P_RD_LOOP_END()
+
+#define P_RD_RS1_RS2_OO_SULOOP(BIT_RD, BIT_RS1, BIT_RS2, BODY) \
+  P_RD_RS1_RS2_LOOP_BASE(BIT_RD) \
+  P_RD_RS1_RS2_OO_SULOOP_BODY(BIT_RD, BIT_RS1, BIT_RS2, BODY) \
+  P_RD_LOOP_END()
+
+#define P_RD_RS1_RS2_E_LOOP(BIT_RD, BIT_RS1, BIT_RS2, BODY) \
+  P_RD_RS1_RS2_LOOP_BASE(BIT_RD) \
+  P_RD_RS1_RS2_E_LOOP_BODY(BIT_RD, BIT_RS1, BIT_RS2, BODY) \
+  P_RD_LOOP_END()
+
+#define P_RD_RS1_RS2_O_LOOP(BIT_RD, BIT_RS1, BIT_RS2, BODY) \
+  P_RD_RS1_RS2_LOOP_BASE(BIT_RD) \
+  P_RD_RS1_RS2_O_LOOP_BODY(BIT_RD, BIT_RS1, BIT_RS2, BODY) \
+  P_RD_LOOP_END()
+
+#define P_RD_RS1_RS2_E_SULOOP(BIT_RD, BIT_RS1, BIT_RS2, BODY) \
+  P_RD_RS1_RS2_LOOP_BASE(BIT_RD) \
+  P_RD_RS1_RS2_E_SULOOP_BODY(BIT_RD, BIT_RS1, BIT_RS2, BODY) \
+  P_RD_LOOP_END()
+
+#define P_RD_RS1_RS2_O_SULOOP(BIT_RD, BIT_RS1, BIT_RS2, BODY) \
+  P_RD_RS1_RS2_LOOP_BASE(BIT_RD) \
+  P_RD_RS1_RS2_O_SULOOP_BODY(BIT_RD, BIT_RS1, BIT_RS2, BODY) \
+  P_RD_LOOP_END()
+
+#define P_REDUCTION_LOOP(BIT, BIT_INNER, USE_RD, IS_SAT, BODY) \
+  P_REDUCTION_LOOP_BASE(BIT, BIT_INNER, USE_RD) \
+  P_RS1_INNER_PARAMS(BIT_INNER) \
+  P_RS2_INNER_PARAMS(BIT_INNER) \
+  BODY \
+  P_REDUCTION_LOOP_END(BIT, IS_SAT)
+
+#define P_REDUCTION_SULOOP(BIT, BIT_INNER, USE_RD, IS_SAT, BODY) \
+  P_REDUCTION_LOOP_BASE(BIT, BIT_INNER, USE_RD) \
+  P_RS1_INNER_PARAMS(BIT_INNER) \
+  P_RS2_INNER_UPARAMS(BIT_INNER) \
+  BODY \
+  P_REDUCTION_LOOP_END(BIT, IS_SAT)
+
+#define P_REDUCTION_ULOOP(BIT, BIT_INNER, USE_RD, IS_SAT, BODY) \
+  P_REDUCTION_ULOOP_BASE(BIT, BIT_INNER, USE_RD) \
+  P_RS1_INNER_UPARAMS(BIT_INNER) \
+  P_RS2_INNER_UPARAMS(BIT_INNER) \
+  BODY \
+  P_REDUCTION_ULOOP_END(BIT, IS_SAT)
+
+#define P_REDUCTION_CROSS_LOOP(BIT, BIT_INNER, USE_RD, IS_SAT, BODY) \
+  P_REDUCTION_LOOP_BASE(BIT, BIT_INNER, USE_RD) \
+  P_RS1_INNER_PARAMS(BIT_INNER) \
+  P_RS2_INNER_CROSS_PARAMS(BIT_INNER) \
+  BODY \
+  P_REDUCTION_LOOP_END(BIT, IS_SAT)
+  
+#define P_RD_DW_LOOP(BIT_RD, BODY) \
+  P_RD_DW_LOOP_BASE(BIT_RD) \
+  P_RD_LOOP_BODY(BIT_RD, BODY) \
+  P_RD_DW_LOOP_END()
+
+// Misc
+#define P_SAT(BIT, R) ( \
+  ((BIT) == 64) ? (R) : \
+  ((R) > ((1LL << ((BIT) - 1)) - 1)) ? ((1LL << ((BIT) - 1)) - 1) : \
+  ((R) < -(1LL << ((BIT) - 1))) ? -(1LL << ((BIT) - 1)) : \
+  (R) \
+)
+
+#define P_USAT(BIT, R) ( \
+  ((R) < 0) ? 0 : \
+  ((BIT) == 64) ? (R) : \
+  ((R) > ((1LL << ((BIT) - 1)) - 1)) ? ((1LL << ((BIT) - 1)) - 1) : \
+  (R) \
+)
+
+#define P_PACK(BIT, X, Y) \
+  require_extension('P'); \
+  require(BIT == e8 || BIT == e16 || BIT == e32); \
+  reg_t rd_tmp = 0; \
+  for (sreg_t i = 0; i < xlen / BIT / 2; i++) { \
+    rd_tmp = set_field(rd_tmp, make_mask64((i * 2 + 1) * BIT, BIT), \
+      P_UFIELD(RS2, i * 2 + Y, BIT)); \
+    rd_tmp = set_field(rd_tmp, make_mask64(i * 2 * BIT, BIT), \
+      P_UFIELD(RS1, i * 2 + X, BIT)); \
+  } \
+  WRITE_RD(sext_xlen(rd_tmp));
+
+#endif
\ No newline at end of file
diff --git a/riscv/riscv.mk.in b/riscv/riscv.mk.in
index a996b889..e1f1add8 100644
--- a/riscv/riscv.mk.in
+++ b/riscv/riscv.mk.in
@@ -1150,6 +1150,216 @@ riscv_insn_ext_zvzip = \
 	vpaire_vv \
 	vpairo_vv \
 
+riscv_insn_ext_p = \
+	pli_b \
+	pli_h \
+	plui_h \
+	padd_bs \
+	padd_hs \
+	padd_b \
+	padd_h \
+	psub_b \
+	psub_h \
+	psadd_b \
+	psadd_h \
+	sadd \
+	psaddu_b \
+	psaddu_h \
+	saddu \
+	pssub_b \
+	pssub_h \
+	ssub \
+	pssubu_b \
+	pssubu_h \
+	ssubu \
+	paadd_b \
+	paadd_h \
+	aadd \
+	paaddu_b \
+	paaddu_h \
+	aaddu \
+	pasub_b \
+	pasub_h \
+	asub \
+	pasubu_b \
+	pasubu_h \
+	asubu \
+	psh1add_h \
+	pssh1sadd_h \
+	ssh1sadd \
+	pas_hx \
+	psa_hx \
+	psas_hx \
+	pssa_hx \
+	paas_hx \
+	pasa_hx \
+	pabd_b \
+	pabd_h \
+	pabdu_b \
+	pabdu_h \
+	psabs_b \
+	psabs_h \
+	predsum_bs \
+	predsum_hs \
+	predsumu_bs \
+	predsumu_hs \
+	pabdsumu_b \
+	pabdsumau_b \
+	psext_h_b \
+	psati_h \
+	sati_rv32 \
+	pusati_h \
+	usati_rv32 \
+	pslli_b \
+	pslli_h \
+	psll_bs \
+	psll_hs \
+	psrli_b \
+	psrli_h \
+	psrl_bs \
+	psrl_hs \
+	psrai_b \
+	psrai_h \
+	psra_bs \
+	psra_hs \
+	psslai_h \
+	psrari_h \
+	pssha_hs \
+	psshar_hs \
+	sslai \
+	srari_rv32 \
+	ssha \
+	sshar \
+	pmin_b \
+	pmin_h \
+	pminu_b \
+	pminu_h \
+	pmax_b \
+	pmax_h \
+	pmaxu_b \
+	pmaxu_h \
+	pmseq_b \
+	pmseq_h \
+	mseq \
+	pmslt_b \
+	pmslt_h \
+	mslt \
+	pmsltu_b \
+	pmsltu_h \
+	msltu \
+	ppaire_b \
+	ppaireo_b \
+	ppaireo_h \
+	ppairoe_b \
+	ppairoe_h \
+	ppairo_b \
+	ppairo_h \
+	abs \
+	cls \
+	slx \
+	srx \
+	mvm \
+	mvmn \
+	merge \
+	pmulh_h \
+	pmulhr_h \
+	pmulhsu_h \
+	pmulhrsu_h \
+	pmulhu_h \
+	pmulhru_h \
+	pmulq_h \
+	pmulqr_h \
+	mulhr \
+	mulhrsu \
+	mulhru \
+	mulq \
+	mulqr \
+	pmul_h_b00 \
+	pmul_h_b01 \
+	pmul_h_b11 \
+	pmulsu_h_b00 \
+	pmulsu_h_b11 \
+	pmulu_h_b00 \
+	pmulu_h_b01 \
+	pmulu_h_b11 \
+	mul_h00 \
+	mul_h01 \
+	mul_h11 \
+	mulsu_h00 \
+	mulsu_h11 \
+	mulu_h00 \
+	mulu_h01 \
+	mulu_h11 \
+	pmulh_h_b0 \
+	pmulh_h_b1 \
+	pmulhsu_h_b0 \
+	pmulhsu_h_b1 \
+	mulh_h0 \
+	mulh_h1 \
+	mulhsu_h0 \
+	mulhsu_h1 \
+	pmhacc_h \
+	pmhracc_h \
+	pmhaccsu_h \
+	pmhraccsu_h \
+	pmhaccu_h \
+	pmhraccu_h \
+	mhacc \
+	mhracc \
+	mhaccsu \
+	mhraccsu \
+	mhaccu \
+	mhraccu \
+	mqacc_h00 \
+	mqacc_h01 \
+	mqacc_h11 \
+	mqracc_h00 \
+	mqracc_h01 \
+	mqracc_h11 \
+	macc_h00 \
+	macc_h01 \
+	macc_h11 \
+	maccsu_h00 \
+	maccsu_h11 \
+	maccu_h00 \
+	maccu_h01 \
+	maccu_h11 \
+	pmhacc_h_b0 \
+	pmhacc_h_b1 \
+	pmhaccsu_h_b0 \
+	pmhaccsu_h_b1 \
+	mhacc_h0 \
+	mhacc_h1 \
+	mhaccsu_h0 \
+	mhaccsu_h1 \
+	pmq2add_h \
+	pmq2adda_h \
+	pmqr2add_h \
+	pmqr2adda_h \
+	pm2add_h \
+	pm2adda_h \
+	pm2addsu_h \
+	pm2addasu_h \
+	pm2addu_h \
+	pm2addau_h \
+	pm2add_hx \
+	pm2adda_hx \
+	pm2sadd_h \
+	pm2sadd_hx \
+	pm2sub_h \
+	pm2suba_h \
+	pm2sub_hx \
+	pm2suba_hx \
+	pm4add_b \
+	pm4adda_b \
+	pm4addsu_b \
+	pm4addasu_b \
+	pm4addu_b \
+	pm4addau_b \
+	pli_db \
+	pli_dh \
+	plui_dh \
+
 riscv_insn_list = \
 	$(riscv_insn_ext_i) \
 	$(riscv_insn_ext_c) \
@@ -1165,6 +1375,7 @@ riscv_insn_list = \
 	$(riscv_insn_ext_cmo) \
 	$(riscv_insn_ext_d_zfa) \
 	$(riscv_insn_ext_f_zfa) \
+	$(riscv_insn_ext_p) \
 	$(riscv_insn_ext_h) \
 	$(riscv_insn_ext_k) \
 	$(if $(HAVE_INT128),$(riscv_insn_ext_q),) \