rvp: add packed instructions for rv32 - part2

Add the following instructions to rv32

 * packed add/sub register-pair instructions
 * packed cross register-pair instructions
 * packed absolute register-pair instructions
 * packed accumulation register-pair instructions
 * packed sign-extend and saturating
 * packed shift register-pair instructions
 * packed compare register-pair instructions
 * packed pack register-pair instructions
 * packed zip register-pair instructions
 * packed narrowing shift and clip register-pair
 * packed multiply register-pair instructions
 * packed accumulate register-pair instructions
 * packed multiply-add register-pair instructions

riscv/decode_macros.h

@@ -236,8 +236,8 @@ static inline bool is_aligned(const unsigned val, const unsigned pos)
 #define zext(x, pos) (((reg_t)(x) << (64 - (pos))) >> (64 - (pos)))
 #define sext_xlen(x) sext(x, xlen)
 #define zext_xlen(x) zext(x, xlen)
-#define sext_xlen_pair(x) sext(x, xlen * 2)
-#define zext_xlen_pair(x) zext(x, xlen * 2)
+#define sext_xlen_pair(x) (xlen == 32 ? sext(x, 64) : (sreg_t)(x))
+#define zext_xlen_pair(x) (xlen == 32 ? zext(x, 64) : (reg_t)(x))
 
 #define set_pc(x) \
   do { if (unlikely((x) & ~p->pc_alignment_mask())) \

riscv/insns/addd.h

@@ -0,0 +1,3 @@
+require_extension('P');
+require_rv32;
+WRITE_P_RD_PAIR(P_RS1_PAIR + P_RS2_PAIR);

riscv/insns/mqrwacc.h

@@ -0,0 +1,3 @@
+require_extension('P');
+require_rv32;
+WRITE_P_RD_PAIR(P_RD_PAIR + (((sreg_t)RS1*(sreg_t)RS2 + 0x40000000) >> 31));

riscv/insns/mqwacc.h

@@ -0,0 +1,3 @@
+require_extension('P');
+require_rv32;
+WRITE_P_RD_PAIR(P_RD_PAIR + (((sreg_t)RS1*(sreg_t)RS2) >> 31));

riscv/insns/nclip.h

@@ -0,0 +1,5 @@
+require_extension('P');
+require_rv32;
+sreg_t tmp = (sreg_t)P_RS1_PAIR >> (RS2 & 0x3f);
+int32_t result = P_SAT(32, tmp);
+WRITE_RD(result);

riscv/insns/nclipi.h

@@ -0,0 +1,5 @@
+require_extension('P');
+require_rv32;
+sreg_t tmp = (sreg_t)P_RS1_PAIR >> insn.shamtd();
+int32_t result = P_SAT(32, tmp);
+WRITE_RD(result);

riscv/insns/nclipiu.h

@@ -0,0 +1,5 @@
+require_extension('P');
+require_rv32;
+reg_t tmp = (reg_t)P_RS1_PAIR >> insn.shamtd();
+uint32_t result = P_USAT_FULL(32, (sreg_t)tmp);
+WRITE_RD(result);

riscv/insns/nclipr.h

@@ -0,0 +1,11 @@
+require_extension('P');
+require_rv32;
+sreg_t val = (sreg_t)P_RS1_PAIR;
+uint32_t shamt = RS2 & 0x3f;
+sreg_t result;
+if (shamt == 0) {
+    result = P_SAT(32, val);
+} else {
+    result = P_SAT(32, (val >> shamt) + ((val >> (shamt - 1)) & 1));
+}
+WRITE_RD(result);

riscv/insns/nclipri.h

@@ -0,0 +1,11 @@
+require_extension('P');
+require_rv32;
+sreg_t val = (sreg_t)P_RS1_PAIR;
+uint32_t shamt = insn.shamtd();
+sreg_t result;
+if (shamt == 0) {
+    result = P_SAT(32, val);
+} else {
+    result = P_SAT(32, (val >> shamt) + ((val >> (shamt - 1)) & 1));
+}
+WRITE_RD(result);

riscv/insns/nclipriu.h

@@ -0,0 +1,11 @@
+require_extension('P');
+require_rv32;
+reg_t val = (reg_t)P_RS1_PAIR;
+uint32_t shamt = insn.shamtd();
+reg_t result;
+if (shamt == 0) {
+    result = P_USAT_FULL(32, (sreg_t)val);
+} else {
+    result = P_USAT_FULL(32, (sreg_t)((val >> shamt) + ((val >> (shamt - 1)) & 1)));
+}
+WRITE_RD(result);

riscv/insns/nclipru.h

@@ -0,0 +1,11 @@
+require_extension('P');
+require_rv32;
+reg_t val = (reg_t)P_RS1_PAIR;
+uint32_t shamt = RS2 & 0x3f;
+reg_t result;
+if (shamt == 0) {
+    result = P_USAT_FULL(32, (sreg_t)val);
+} else {
+    result = P_USAT_FULL(32, (sreg_t)((val >> shamt) + ((val >> (shamt - 1)) & 1)));
+}
+WRITE_RD(result);

riscv/insns/nclipu.h

@@ -0,0 +1,5 @@
+require_extension('P');
+require_rv32;
+reg_t tmp = (reg_t)P_RS1_PAIR >> (RS2 & 0x3f);
+uint32_t result = P_USAT_FULL(32, (sreg_t)tmp);
+WRITE_RD(result);

riscv/insns/nsra.h

@@ -0,0 +1,3 @@
+require_extension('P');
+require_rv32;
+WRITE_RD((sreg_t)P_RS1_PAIR >> (RS2 & 0x3f));

riscv/insns/nsrai.h

@@ -0,0 +1,3 @@
+require_extension('P');
+require_rv32;
+WRITE_RD((sreg_t)P_RS1_PAIR >> insn.shamtd());

riscv/insns/nsrar.h

@@ -0,0 +1,11 @@
+require_extension('P');
+require_rv32;
+sreg_t val = (sreg_t)P_RS1_PAIR;
+uint32_t shamt = RS2 & 0x3f;
+sreg_t result;
+if (shamt == 0) {
+    result = val;
+} else {
+    result = (val >> shamt) + ((val >> (shamt - 1)) & 1);
+}
+WRITE_RD(result);

riscv/insns/nsrari.h

@@ -0,0 +1,11 @@
+require_extension('P');
+require_rv32;
+sreg_t val = (sreg_t)P_RS1_PAIR;
+uint32_t shamt = insn.shamtd();
+sreg_t result;
+if (shamt == 0) {
+    result = val;
+} else {
+    result = (val >> shamt) + ((val >> (shamt - 1)) & 1);
+}
+WRITE_RD(result);

riscv/insns/nsrl.h

@@ -0,0 +1,3 @@
+require_extension('P');
+require_rv32;
+WRITE_RD(P_RS1_PAIR >> (RS2 & 0x3f));

riscv/insns/nsrli.h

@@ -0,0 +1,3 @@
+require_extension('P');
+require_rv32;
+WRITE_RD(P_RS1_PAIR >> insn.shamtd());

riscv/insns/paadd_db.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_RD_RS1_RS2_DW_LOOP(8,8,8, {
+  p_rd = (p_rs1 + p_rs2) >> 1;
+})

riscv/insns/paadd_dh.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_RD_RS1_RS2_DW_LOOP(16,16,16, {
+  p_rd = (p_rs1 + p_rs2) >> 1;
+})

riscv/insns/paadd_dw.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_RD_RS1_RS2_DW_LOOP(32,32,32, {
+  p_rd = ((sreg_t)p_rs1 + (sreg_t)p_rs2) >> 1;
+})

riscv/insns/paaddu_db.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_RD_RS1_RS2_DW_ULOOP(8,8,8, {
+  p_rd = (p_rs1 + p_rs2) >> 1;
+})

riscv/insns/paaddu_dh.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_RD_RS1_RS2_DW_ULOOP(16,16,16, {
+  p_rd = (p_rs1 + p_rs2) >> 1;
+})

riscv/insns/paaddu_dw.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_RD_RS1_RS2_DW_ULOOP(32,32,32, {
+  p_rd = ((reg_t)p_rs1 + (reg_t)p_rs2) >> 1;
+})

riscv/insns/paas_dhx.h

@@ -0,0 +1,6 @@
+require_rv32;
+P_CROSS_DW_LOOP(16, {
+  p_rd = (p_rs1 + p_rs2) >> 1;
+}, {
+  p_rd = (p_rs1 - p_rs2) >> 1;
+})

riscv/insns/pabd_db.h

@@ -0,0 +1,5 @@
+require_rv32;
+P_RD_RS1_RS2_DW_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);
+})

riscv/insns/pabd_dh.h

@@ -0,0 +1,5 @@
+require_rv32;
+P_RD_RS1_RS2_DW_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);
+})

riscv/insns/pabdu_db.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_RD_RS1_RS2_DW_ULOOP(8, 8, 8, {
+  p_rd = (p_rs1 < p_rs2) ? p_rs2 - p_rs1 : p_rs1 - p_rs2;
+})

riscv/insns/pabdu_dh.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_RD_RS1_RS2_DW_ULOOP(16, 16, 16, {
+  p_rd = (p_rs1 < p_rs2) ? p_rs2 - p_rs1 : p_rs1 - p_rs2;
+})

riscv/insns/padd_db.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_RD_RS1_RS2_DW_LOOP(8, 8, 8, {
+  p_rd = p_rs1 + p_rs2;
+})

riscv/insns/padd_dbs.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_RD_RS1_DW_LOOP(8, 8, {
+  p_rd = p_rs1 + P_FIELD(RS2, 0, 8);
+})

riscv/insns/padd_dh.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_RD_RS1_RS2_DW_LOOP(16, 16, 16, {
+  p_rd = p_rs1 + p_rs2;
+})

riscv/insns/padd_dhs.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_RD_RS1_DW_LOOP(16, 16, {
+  p_rd = p_rs1 + P_FIELD(RS2, 0, 16);
+})

riscv/insns/padd_dw.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_RD_RS1_RS2_DW_LOOP(32, 32, 32, {
+  p_rd = p_rs1 + p_rs2;
+})

riscv/insns/padd_dws.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_RD_RS1_DW_LOOP(32, 32, {
+  p_rd = p_rs1 + P_FIELD(RS2, 0, 32);
+})

riscv/insns/pas_dhx.h

@@ -0,0 +1,6 @@
+require_rv32;
+P_CROSS_DW_LOOP(16, {
+  p_rd = p_rs1 + p_rs2;
+}, {
+  p_rd = p_rs1 - p_rs2;
+})

riscv/insns/pasa_dhx.h

@@ -0,0 +1,6 @@
+require_rv32;
+P_CROSS_DW_LOOP(16, {
+  p_rd = (p_rs1 - p_rs2) >> 1;
+}, {
+  p_rd = (p_rs1 + p_rs2) >> 1;
+})

riscv/insns/pasub_db.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_RD_RS1_RS2_DW_LOOP(8,8,8, {
+  p_rd = (p_rs1 - p_rs2) >> 1;
+})

riscv/insns/pasub_dh.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_RD_RS1_RS2_DW_LOOP(16,16,16, {
+  p_rd = (p_rs1 - p_rs2) >> 1;
+})

riscv/insns/pasub_dw.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_RD_RS1_RS2_DW_LOOP(32,32,32, {
+  p_rd = ((sreg_t)p_rs1 - (sreg_t)p_rs2) >> 1;
+})

riscv/insns/pasubu_db.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_RD_RS1_RS2_DW_ULOOP(8,8,8, {
+  p_rd = (p_rs1 - p_rs2) >> 1;
+})

riscv/insns/pasubu_dh.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_RD_RS1_RS2_DW_ULOOP(16,16,16, {
+  p_rd = (p_rs1 - p_rs2) >> 1;
+})

riscv/insns/pasubu_dw.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_RD_RS1_RS2_DW_ULOOP(32,32,32, {
+  p_rd = ((reg_t)p_rs1 - (reg_t)p_rs2) >> 1;
+})

riscv/insns/pm2wadd_h.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_WIDEN_REDUCTION_LOOP(32, 16, false, false, {
+  p_res += sext32(p_rs1) * sext32(p_rs2);
+})

riscv/insns/pm2wadd_hx.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_WIDEN_REDUCTION_CROSS_LOOP(32, 16, false, false, {
+  p_res += p_rs1 * p_rs2;
+})

riscv/insns/pm2wadda_h.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_WIDEN_REDUCTION_LOOP(32, 16, true, false, {
+  p_res += p_rs1 * p_rs2;
+})

riscv/insns/pm2wadda_hx.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_WIDEN_REDUCTION_CROSS_LOOP(32, 16, true, false, {
+  p_res += p_rs1 * p_rs2;
+})

riscv/insns/pm2waddasu_h.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_WIDEN_REDUCTION_SULOOP(32, 16, true, false, {
+  p_res += p_rs1 * p_rs2;
+})

riscv/insns/pm2waddau_h.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_WIDEN_REDUCTION_ULOOP(32, 16, true, false, {
+  p_res += (uint32_t)p_rs1 * (uint32_t)p_rs2;
+})

riscv/insns/pm2waddsu_h.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_WIDEN_REDUCTION_SULOOP(32, 16, false, false, {
+  p_res += p_rs1 * p_rs2;
+})

riscv/insns/pm2waddu_h.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_WIDEN_REDUCTION_ULOOP(32, 16, false, false, {
+  p_res += (uint32_t)p_rs1 * (uint32_t)p_rs2;
+})

riscv/insns/pm2wsub_h.h

@@ -0,0 +1,7 @@
+require_rv32;
+P_WIDEN_REDUCTION_LOOP(32, 16, false, false, {
+  if (j & 1)
+    p_res += p_rs1 * p_rs2;
+  else
+    p_res -= p_rs1 * p_rs2;
+})

riscv/insns/pm2wsub_hx.h

@@ -0,0 +1,7 @@
+require_rv32;
+P_WIDEN_REDUCTION_CROSS_LOOP(32, 16, false, false, {
+  if (j & 1)
+    p_res += p_rs1 * p_rs2;
+  else
+    p_res -= p_rs1 * p_rs2;
+})

riscv/insns/pm2wsuba_h.h

@@ -0,0 +1,7 @@
+require_rv32;
+P_WIDEN_REDUCTION_LOOP(32, 16, true, false, {
+  if (j & 1)
+    p_res += p_rs1 * p_rs2;
+  else
+    p_res -= p_rs1 * p_rs2;
+})

riscv/insns/pm2wsuba_hx.h

@@ -0,0 +1,7 @@
+require_rv32;
+P_WIDEN_REDUCTION_CROSS_LOOP(32, 16, true, false, {
+  if (j & 1)
+    p_res += p_rs1 * p_rs2;
+  else
+    p_res -= p_rs1 * p_rs2;
+})

riscv/insns/pmax_db.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_RD_RS1_RS2_DW_LOOP(8, 8, 8, {
+  p_rd = (p_rs1 > p_rs2) ? p_rs1 : p_rs2;
+})

riscv/insns/pmax_dh.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_RD_RS1_RS2_DW_LOOP(16, 16, 16, {
+  p_rd = (p_rs1 > p_rs2) ? p_rs1 : p_rs2;
+})

riscv/insns/pmax_dw.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_RD_RS1_RS2_DW_LOOP(32, 32, 32, {
+  p_rd = (p_rs1 > p_rs2) ? p_rs1 : p_rs2;
+})

riscv/insns/pmaxu_db.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_RD_RS1_RS2_DW_ULOOP(8, 8, 8, {
+  p_rd = (p_rs1 > p_rs2) ? p_rs1 : p_rs2;
+})

riscv/insns/pmaxu_dh.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_RD_RS1_RS2_DW_ULOOP(16, 16, 16, {
+  p_rd = (p_rs1 > p_rs2) ? p_rs1 : p_rs2;
+})

riscv/insns/pmaxu_dw.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_RD_RS1_RS2_DW_ULOOP(32, 32, 32, {
+  p_rd = (p_rs1 > p_rs2) ? p_rs1 : p_rs2;
+})

riscv/insns/pmin_db.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_RD_RS1_RS2_DW_LOOP(8, 8, 8, {
+  p_rd = (p_rs1 < p_rs2) ? p_rs1 : p_rs2;
+})

riscv/insns/pmin_dh.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_RD_RS1_RS2_DW_LOOP(16, 16, 16, {
+  p_rd = (p_rs1 < p_rs2) ? p_rs1 : p_rs2;
+})

riscv/insns/pmin_dw.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_RD_RS1_RS2_DW_LOOP(32, 32, 32, {
+  p_rd = (p_rs1 < p_rs2) ? p_rs1 : p_rs2;
+})

riscv/insns/pminu_db.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_RD_RS1_RS2_DW_ULOOP(8, 8, 8, {
+  p_rd = (p_rs1 < p_rs2) ? p_rs1 : p_rs2;
+})

riscv/insns/pminu_dh.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_RD_RS1_RS2_DW_ULOOP(16, 16, 16, {
+  p_rd = (p_rs1 < p_rs2) ? p_rs1 : p_rs2;
+})

riscv/insns/pminu_dw.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_RD_RS1_RS2_DW_ULOOP(32, 32, 32, {
+  p_rd = (p_rs1 < p_rs2) ? p_rs1 : p_rs2;
+})

riscv/insns/pmqrwacc_h.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_WIDEN_RD_RS1_RS2_LOOP(16, 16, {
+  p_rd += (p_rs1 * p_rs2 + 0x4000) >> 15;
+})

riscv/insns/pmqwacc_h.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_WIDEN_RD_RS1_RS2_LOOP(16, 16, {
+  p_rd += (p_rs1 * p_rs2) >> 15;
+})

riscv/insns/pmseq_db.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_RD_RS1_RS2_DW_LOOP(8, 8, 8, {
+  p_rd = (p_rs1 == p_rs2) ? -1 : 0;
+})

riscv/insns/pmseq_dh.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_RD_RS1_RS2_DW_LOOP(16, 16, 16, {
+  p_rd = (p_rs1 == p_rs2) ? -1 : 0;
+})

riscv/insns/pmseq_dw.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_RD_RS1_RS2_DW_LOOP(32, 32, 32, {
+  p_rd = (p_rs1 == p_rs2) ? -1 : 0;
+})

riscv/insns/pmslt_db.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_RD_RS1_RS2_DW_LOOP(8, 8, 8, {
+  p_rd = (p_rs1 < p_rs2) ? -1 : 0;
+})

riscv/insns/pmslt_dh.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_RD_RS1_RS2_DW_LOOP(16, 16, 16, {
+  p_rd = (p_rs1 < p_rs2) ? -1 : 0;
+})

riscv/insns/pmslt_dw.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_RD_RS1_RS2_DW_LOOP(32, 32, 32, {
+  p_rd = (p_rs1 < p_rs2) ? -1 : 0;
+})

riscv/insns/pmsltu_db.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_RD_RS1_RS2_DW_ULOOP(8 ,8 ,8, {
+  p_rd = (p_rs1 < p_rs2) ? -1 : 0;
+})

riscv/insns/pmsltu_dh.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_RD_RS1_RS2_DW_ULOOP(16, 16, 16, {
+  p_rd = (p_rs1 < p_rs2) ? -1 : 0;
+})

riscv/insns/pmsltu_dw.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_RD_RS1_RS2_DW_ULOOP(32, 32, 32, {
+  p_rd = (p_rs1 < p_rs2) ? -1 : 0;
+})

riscv/insns/pnclip_bs.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_NARROW_RD_RS1_LOOP(8, 16, {
+    p_rd = P_SAT(8, p_rs1 >> (P_FIELD(RS2, 0, 8) & 0xF));
+})

riscv/insns/pnclip_hs.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_NARROW_RD_RS1_LOOP(16, 32, {
+    p_rd = P_SAT(16, p_rs1 >> (P_FIELD(RS2, 0, 16) & 0X1F));
+})

riscv/insns/pnclipi_b.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_NARROW_RD_RS1_LOOP(8, 16, {
+    p_rd = P_SAT(8, p_rs1 >> insn.shamth());
+})

riscv/insns/pnclipi_h.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_NARROW_RD_RS1_LOOP(16, 32, {
+    p_rd = P_SAT(16, p_rs1 >> insn.shamtw());
+})

riscv/insns/pnclipiu_b.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_NARROW_RD_RS1_ULOOP(8, 16, {
+    p_rd = P_USAT(8, p_rs1 >> insn.shamth());
+})

riscv/insns/pnclipiu_h.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_NARROW_RD_RS1_ULOOP(16, 32, {
+    p_rd = P_USAT(16, (p_rs1 >> insn.shamtw()));
+})

riscv/insns/pnclipr_bs.h

@@ -0,0 +1,13 @@
+require_rv32;
+P_NARROW_RD_RS1_LOOP(8, 16, {
+    uint16_t shamt = P_FIELD(RS2, 0, 8) & 0xF;
+    sreg_t result;
+    if (shamt == 0) {
+        result = p_rs1;
+    } else {
+        sreg_t shifted = (sreg_t)p_rs1 >> shamt;
+        sreg_t roundbit = ((sreg_t)p_rs1 >> (shamt - 1)) & 1;
+        result = shifted + roundbit;
+    }
+    p_rd = P_SAT(8, result);
+})

riscv/insns/pnclipr_hs.h

@@ -0,0 +1,13 @@
+require_rv32;
+P_NARROW_RD_RS1_LOOP(16, 32, {
+    uint32_t shamt = P_FIELD(RS2, 0, 16) & 0x1F;
+    sreg_t result;
+    if (shamt == 0) {
+        result = p_rs1;
+    } else {
+        sreg_t shifted = (sreg_t)p_rs1 >> shamt;
+        sreg_t roundbit = ((sreg_t)p_rs1 >> (shamt - 1)) & 1;
+        result = shifted + roundbit;
+    }
+    p_rd = P_SAT(16, result);
+})

riscv/insns/pnclipri_b.h

@@ -0,0 +1,13 @@
+require_rv32;
+P_NARROW_RD_RS1_LOOP(8, 16, {
+    uint16_t shamt = insn.shamth();
+    sreg_t result;
+    if (shamt == 0) {
+        result = p_rs1;
+    } else {
+        sreg_t shifted = (sreg_t)p_rs1 >> shamt;
+        sreg_t roundbit = ((sreg_t)p_rs1 >> (shamt - 1)) & 1;
+        result = shifted + roundbit;
+    }
+    p_rd = P_SAT(8, result);
+})

riscv/insns/pnclipri_h.h

@@ -0,0 +1,13 @@
+require_rv32;
+P_NARROW_RD_RS1_LOOP(16, 32, {
+    uint32_t shamt = insn.shamtw();
+    sreg_t result;
+    if (shamt == 0) {
+        result = p_rs1;
+    } else {
+        sreg_t shifted = (sreg_t)p_rs1 >> shamt;
+        sreg_t roundbit = ((sreg_t)p_rs1 >> (shamt - 1)) & 1;
+        result = shifted + roundbit;
+    }
+    p_rd = P_SAT(16, result);
+})

riscv/insns/pnclipriu_b.h

@@ -0,0 +1,13 @@
+require_rv32;
+P_NARROW_RD_RS1_ULOOP(8, 16, {
+    uint32_t shamt = insn.shamth();
+    uint32_t result;
+    if (shamt == 0) {
+        result = p_rs1;
+    } else {
+        uint32_t shifted  = p_rs1 >> shamt;
+        uint32_t roundbit = (p_rs1 >> (shamt - 1)) & 1;
+        result = shifted + roundbit;
+    }
+    p_rd = P_USAT_FULL(8, (sreg_t)result);
+})

riscv/insns/pnclipriu_h.h

@@ -0,0 +1,13 @@
+require_rv32;
+P_NARROW_RD_RS1_ULOOP(16, 32, {
+    uint32_t shamt = insn.shamtw();
+    uint32_t result;
+    if (shamt == 0) {
+        result = p_rs1;
+    } else {
+        uint32_t shifted  = p_rs1 >> shamt;
+        uint32_t roundbit = (p_rs1 >> (shamt - 1)) & 1;
+        result = shifted + roundbit;
+    }
+    p_rd = P_USAT_FULL(16, (sreg_t)result);
+})

riscv/insns/pnclipru_bs.h

@@ -0,0 +1,13 @@
+require_rv32;
+P_NARROW_RD_RS1_ULOOP(8, 16, {
+    uint32_t shamt = P_UFIELD(RS2, 0, 8) & 0xF;
+    uint32_t result;
+    if (shamt == 0) {
+        result = p_rs1;
+    } else {
+        uint32_t shifted  = p_rs1 >> shamt;
+        uint32_t roundbit = (p_rs1 >> (shamt - 1)) & 1;
+        result = shifted + roundbit;
+    }
+    p_rd = P_USAT_FULL(8, (sreg_t)result);
+})

riscv/insns/pnclipru_hs.h

@@ -0,0 +1,13 @@
+require_rv32;
+P_NARROW_RD_RS1_ULOOP(16, 32, {
+    uint32_t shamt = P_UFIELD(RS2, 0, 16) & 0x1F;
+    uint32_t result;
+    if (shamt == 0) {
+        result = p_rs1;
+    } else {
+        uint32_t shifted  = p_rs1 >> shamt;
+        uint32_t roundbit = (p_rs1 >> (shamt - 1)) & 1;
+        result = shifted + roundbit;
+    }
+    p_rd = P_USAT_FULL(16, (sreg_t)result);
+})

riscv/insns/pnclipu_bs.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_NARROW_RD_RS1_ULOOP(8, 16, {
+    p_rd = P_USAT_FULL(8, (sreg_t)(p_rs1 >> (P_UFIELD(RS2, 0, 8) & 0xF)));
+})

riscv/insns/pnclipu_hs.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_NARROW_RD_RS1_ULOOP(16, 32, {
+    p_rd = P_USAT_FULL(16, (sreg_t)(p_rs1 >> (P_UFIELD(RS2, 0, 16) & 0X1F)));
+})

riscv/insns/pnsra_bs.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_NARROW_RD_RS1_LOOP(8, 16, {
+    p_rd = p_rs1 >> (P_FIELD(RS2, 0, 8) & 0xF);
+})

riscv/insns/pnsra_hs.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_NARROW_RD_RS1_LOOP(16, 32, {
+    p_rd = p_rs1 >> (P_FIELD(RS2, 0, 16) & 0x1F);
+})

riscv/insns/pnsrai_b.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_NARROW_RD_RS1_LOOP(8, 16, {
+    p_rd = p_rs1 >> insn.shamth();
+})

riscv/insns/pnsrai_h.h

@@ -0,0 +1,4 @@
+require_rv32;
+P_NARROW_RD_RS1_LOOP(16, 32, {
+    p_rd = p_rs1 >> insn.shamtw();
+})

riscv/insns/pnsrar_bs.h

@@ -0,0 +1,12 @@
+require_rv32;
+P_NARROW_RD_RS1_LOOP(8, 16, {
+    uint32_t shamt = P_FIELD(RS2, 0, 8) & 0xF;
+    if (shamt != 0) {
+        sreg_t val = (sreg_t)p_rs1;
+        sreg_t shifted = val >> shamt;
+        sreg_t roundbit = (val >> (shamt - 1)) & 1;
+        p_rd = (shifted + roundbit) & 0xFF;
+    } else {
+        p_rd = p_rs1 & 0xFF;
+    }
+})

riscv/insns/pnsrar_hs.h

@@ -0,0 +1,12 @@
+require_rv32;
+P_NARROW_RD_RS1_LOOP(16, 32, {
+    uint32_t shamt = P_FIELD(RS2, 0, 16) & 0x1F;
+    if (shamt != 0) {
+        sreg_t val = (sreg_t)p_rs1;
+        sreg_t shifted = val >> shamt;
+        sreg_t roundbit = (val >> (shamt - 1)) & 1;
+        p_rd = (shifted + roundbit) & 0xFFFF;
+    } else {
+        p_rd = p_rs1 & 0xFFFF;
+    }
+})