Merge pull request #2310 from chihminchao/fix-rvp

Fix rvp

riscv/check-opcode-overlap.t.cc

@@ -33,8 +33,11 @@ int main()
   std::unordered_map<std::string, bool> overlap_list;
   #define DECLARE_OVERLAP_INSN(name, ext) \
     overlap_list[std::string(#name)] = false;
+  #define DECLARE_OVERLAP_INSN_RV64(name, ext) \
+    overlap_list[std::string(#name)] = false;
     #include "overlap_list.h"
   #undef DECLARE_OVERLAP_INSN
+  #undef DECLARE_OVERLAP_INSN_RV64
 
   std::vector<const opcode*> list;
   for (size_t i = 0; i < std::size(static_list); i++) {

riscv/insns/aadd.h

@@ -1,3 +1,3 @@
 require_extension('P');
 require_rv32;
-WRITE_RD((RS1 + RS2)>>1);
+WRITE_RD(sext32((RS1 + RS2)>>1));

riscv/insns/aaddu.h

@@ -1,3 +1,3 @@
 require_extension('P');
 require_rv32;
-WRITE_RD((0ULL + (uint32_t)RS1 + (uint32_t)RS2)>>1);
+WRITE_RD(sext32((0ULL + (uint32_t)RS1 + (uint32_t)RS2)>>1));

riscv/insns/asub.h

@@ -1,3 +1,3 @@
 require_extension('P');
 require_rv32;
-WRITE_RD((RS1 - RS2)>>1);
+WRITE_RD(sext32((RS1 - RS2)>>1));

riscv/insns/asubu.h

@@ -1,3 +1,3 @@
 require_extension('P');
 require_rv32;
-WRITE_RD((0ULL + (uint32_t)RS1 - (uint32_t)RS2)>>1);
+WRITE_RD((sext32(0ULL + (uint32_t)RS1 - (uint32_t)RS2)>>1));

riscv/insns/mhacc.h

@@ -1,4 +1,4 @@
 require_extension('P');
 require_rv32;
 sreg_t mres = sext32(RS1) * sext32(RS2);
-WRITE_RD(RD + (mres >> 32));
+WRITE_RD(sext32(RD + (mres >> 32)));

riscv/insns/mhacc_h0.h

@@ -1,4 +1,4 @@
 require_extension('P');
 require_rv32;
 sreg_t mres = sext32(RS1) * sext32(P_FIELD(RS2, 0, 16));
-WRITE_RD(RD + (mres >> 32));
+WRITE_RD(sext32(RD + (mres >> 16)));

riscv/insns/mhacc_h1.h

@@ -1,4 +1,4 @@
 require_extension('P');
 require_rv32;
 sreg_t mres = sext32(RS1) * sext32(P_FIELD(RS2, 1, 16));
-WRITE_RD(RD + (mres >> 32));
+WRITE_RD(sext32(RD + (mres >> 16)));

riscv/insns/mhaccsu.h

@@ -1,4 +1,4 @@
 require_extension('P');
 require_rv32;
 sreg_t mres = sext32(RS1) * reg_t((uint32_t)RS2);
-WRITE_RD(RD + (mres >> 32));
+WRITE_RD(sext32(RD + (mres >> 32)));

riscv/insns/mhaccsu_h0.h

@@ -1,4 +1,4 @@
 require_extension('P');
 require_rv32;
 sreg_t mres = sext32(RS1) * (uint32_t)P_FIELD(RS2, 0, 16);
-WRITE_RD(RD + (mres >> 32));
+WRITE_RD(sext32(RD + (mres >> 16)));

riscv/insns/mhaccsu_h1.h

@@ -1,4 +1,4 @@
 require_extension('P');
 require_rv32;
 sreg_t mres = sext32(RS1) * (uint32_t)P_FIELD(RS2, 1, 16);
-WRITE_RD(RD + (mres >> 32));
+WRITE_RD(sext32(RD + (mres >> 16)));

riscv/insns/mhaccu.h

@@ -1,4 +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));
+WRITE_RD(sext32(RD + (mres >> 32)));

riscv/insns/mhracc.h

@@ -2,4 +2,4 @@ require_extension('P');
 require_rv32;
 sreg_t mres = sext32(RS1) * sext32(RS2);
 int32_t round = ((mres >> 31) + 1) >> 1;
-WRITE_RD(RD + round);
+WRITE_RD(sext32(RD + round));

riscv/insns/mhraccsu.h

@@ -2,4 +2,4 @@ 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);
+WRITE_RD(sext32(RD + round));

riscv/insns/mhraccu.h

@@ -2,4 +2,4 @@ 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);
+WRITE_RD(sext32(RD + round));

riscv/insns/mqacc_h00.h

@@ -1,3 +1,3 @@
 require_extension('P');
 require_rv32;
-WRITE_RD(RD + (((int32_t)P_FIELD(RS1, 0, 16) * P_FIELD(RS2, 0, 16)) >> 15));
+WRITE_RD(sext32(RD + (((int32_t)P_FIELD(RS1, 0, 16) * P_FIELD(RS2, 0, 16)) >> 15)));

riscv/insns/mqacc_h01.h

@@ -1,3 +1,3 @@
 require_extension('P');
 require_rv32;
-WRITE_RD(RD + (((int32_t)P_FIELD(RS1, 0, 16) * P_FIELD(RS2, 1, 16)) >> 15));
+WRITE_RD(sext32(RD + (((int32_t)P_FIELD(RS1, 0, 16) * P_FIELD(RS2, 1, 16)) >> 15)));

riscv/insns/mqacc_h11.h

@@ -1,3 +1,3 @@
 require_extension('P');
 require_rv32;
-WRITE_RD(RD + (((int32_t)P_FIELD(RS1, 1, 16) * P_FIELD(RS2, 1, 16)) >> 15));
+WRITE_RD(sext32(RD + (((int32_t)P_FIELD(RS1, 1, 16) * P_FIELD(RS2, 1, 16)) >> 15)));

riscv/insns/mqracc_h00.h

@@ -1,3 +1,3 @@
 require_extension('P');
 require_rv32;
-WRITE_RD(RD + (((int32_t)P_FIELD(RS1, 0, 16) * P_FIELD(RS2, 0, 16) + 0x4000) >> 15));
+WRITE_RD(sext32(RD + (((int32_t)P_FIELD(RS1, 0, 16) * P_FIELD(RS2, 0, 16) + 0x4000) >> 15)));

riscv/insns/mqracc_h01.h

@@ -1,3 +1,3 @@
 require_extension('P');
 require_rv32;
-WRITE_RD(RD + (((int32_t)P_FIELD(RS1, 0, 16) * P_FIELD(RS2, 1, 16) + 0x4000) >> 15));
+WRITE_RD(sext32(RD + (((int32_t)P_FIELD(RS1, 0, 16) * P_FIELD(RS2, 1, 16) + 0x4000) >> 15)));

riscv/insns/mqracc_h11.h

@@ -1,3 +1,3 @@
 require_extension('P');
 require_rv32;
-WRITE_RD(RD + (((int32_t)P_FIELD(RS1, 1, 16) * P_FIELD(RS2, 1, 16) + 0x4000) >> 15));
+WRITE_RD(sext32(RD + (((int32_t)P_FIELD(RS1, 1, 16) * P_FIELD(RS2, 1, 16) + 0x4000) >> 15)));

riscv/insns/mqrwacc.h

@@ -1,3 +1,4 @@
 require_extension('P');
 require_rv32;
-WRITE_P_RD_PAIR(P_RD_PAIR + (((sreg_t)RS1*(sreg_t)RS2 + 0x40000000) >> 31));
+int64_t tmp = (((int128_t)(int32_t)RS1 * (int32_t)RS2) + (1 << 30)) >> 31;
+WRITE_P_RD_PAIR(P_RD_PAIR + tmp);

riscv/insns/mqwacc.h

@@ -1,3 +1,4 @@
 require_extension('P');
 require_rv32;
-WRITE_P_RD_PAIR(P_RD_PAIR + (((sreg_t)RS1*(sreg_t)RS2) >> 31));
+int64_t tmp = (((int128_t)(int32_t)RS1 * (int32_t)RS2)) >> 31;
+WRITE_P_RD_PAIR((P_RD_PAIR + tmp));

riscv/insns/mulh_h0.h

@@ -1,4 +1,4 @@
 require_extension('P');
 require_rv32;
 int64_t mres = sext(RS1,64) * sext(P_FIELD(RS2, 0, 16),64);
-WRITE_RD(mres>>32);
+WRITE_RD(sext32(mres>>16));

riscv/insns/mulh_h1.h

@@ -1,4 +1,4 @@
 require_extension('P');
 require_rv32;
 int64_t mres = sext(RS1,64) * sext(P_FIELD(RS2, 1, 16),64);
-WRITE_RD(mres>>32);
+WRITE_RD(sext32(mres>>16));

riscv/insns/mulhr.h

@@ -1,4 +1,4 @@
 require_extension('P');
 require_rv32;
 int64_t mres = sext(RS1,64) * sext(RS2,64);
-WRITE_RD(((mres >> 31) + 1) >> 1);
+WRITE_RD(sext32(((mres >> 31) + 1) >> 1));

riscv/insns/mulhrsu.h

@@ -1,4 +1,4 @@
 require_extension('P');
 require_rv32;
 sreg_t mres = sext(RS1,64) * reg_t((uint32_t)RS2);
-WRITE_RD(((mres >> 31) + 1) >> 1);
+WRITE_RD(sext32(((mres >> 31) + 1) >> 1));

riscv/insns/mulhru.h

@@ -1,4 +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);
+WRITE_RD(sext32(((mres >> 31) + 1) >> 1));

riscv/insns/mulhsu_h0.h

@@ -1,4 +1,4 @@
 require_extension('P');
 require_rv32;
 sreg_t mres = sext32(RS1) * (uint32_t)P_FIELD(RS2, 0, 16);
-WRITE_RD(mres >> 32);
+WRITE_RD(sext32(mres >> 16));

riscv/insns/mulhsu_h1.h

@@ -1,4 +1,4 @@
 require_extension('P');
 require_rv32;
 sreg_t mres = sext32(RS1) * (uint32_t)P_FIELD(RS2, 1, 16);
-WRITE_RD(mres >> 32);
+WRITE_RD(sext32(mres >> 16));

riscv/insns/mulq.h

@@ -1,7 +1,7 @@
 require_extension('P');
 require_rv32;
 if ((RS1 != (reg_t)INT32_MIN) || (RS2 != (reg_t)INT32_MIN)) {
-    WRITE_RD((RS1 * RS2) >> 31);
+    WRITE_RD(sext32((RS1 * RS2) >> 31));
   } else {
     WRITE_RD(INT32_MAX);
     P.set_vxsat();

riscv/insns/mulqr.h

@@ -1,7 +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);
+    WRITE_RD(sext32((((RS1 * RS2) >> 30) + 1) >> 1));
   } else {
     WRITE_RD(INT32_MAX);
     P.set_vxsat();

riscv/insns/nclip.h

@@ -2,4 +2,4 @@ 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);
+WRITE_RD(sext32(result));

riscv/insns/nclipi.h

@@ -2,4 +2,4 @@ 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);
+WRITE_RD(sext32(result));

riscv/insns/nclipiu.h

@@ -1,5 +1,8 @@
 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);
+if (tmp > UINT32_MAX) {
+  tmp = UINT32_MAX;
+  P.set_vxsat();
+}
+WRITE_RD(sext32(tmp));

riscv/insns/nclipr.h

@@ -8,4 +8,4 @@ if (shamt == 0) {
 } else {
     result = P_SAT(32, (val >> shamt) + ((val >> (shamt - 1)) & 1));
 }
-WRITE_RD(result);
+WRITE_RD(sext32(result));

riscv/insns/nclipri.h

@@ -4,8 +4,8 @@ sreg_t val = (sreg_t)P_RS1_PAIR;
 uint32_t shamt = insn.shamtd();
 sreg_t result;
 if (shamt == 0) {
-    result = P_SAT(32, val);
+  result = P_SAT(32, val);
 } else {
-    result = P_SAT(32, (val >> shamt) + ((val >> (shamt - 1)) & 1));
+  result = P_SAT(32, (val >> shamt) + ((val >> (shamt - 1)) & 1));
 }
-WRITE_RD(result);
+WRITE_RD(sext32(result));

riscv/insns/nclipriu.h

@@ -1,11 +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)));
+reg_t tmp = ((((uint128_t)P_RS1_PAIR << 1) >> shamt) + 1) >> 1;
+
+if (tmp > UINT32_MAX) {
+  tmp = UINT32_MAX;
+  P.set_vxsat();
 }
-WRITE_RD(result);
+WRITE_RD(sext32(tmp));
+

riscv/insns/nclipru.h

@@ -1,11 +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);
+reg_t tmp = ((((uint128_t)P_RS1_PAIR << 1) >> shamt) + 1) >> 1;
+
+if (tmp > UINT32_MAX) {
+  tmp = UINT32_MAX;
+  P.set_vxsat();
+} 
+WRITE_RD(sext32(tmp));
+

riscv/insns/nclipu.h

@@ -1,5 +1,9 @@
 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);
+if (tmp > UINT32_MAX) {
+  tmp = UINT32_MAX;
+  P.set_vxsat();
+}
+
+WRITE_RD(sext32(tmp));

riscv/insns/nsra.h

@@ -1,3 +1,3 @@
 require_extension('P');
 require_rv32;
-WRITE_RD((sreg_t)P_RS1_PAIR >> (RS2 & 0x3f));
+WRITE_RD(sext32((sreg_t)P_RS1_PAIR >> (RS2 & 0x3f)));

riscv/insns/nsrai.h

@@ -1,3 +1,3 @@
 require_extension('P');
 require_rv32;
-WRITE_RD((sreg_t)P_RS1_PAIR >> insn.shamtd());
+WRITE_RD(sext32((sreg_t)P_RS1_PAIR >> insn.shamtd()));

riscv/insns/nsrar.h

@@ -4,8 +4,8 @@ sreg_t val = (sreg_t)P_RS1_PAIR;
 uint32_t shamt = RS2 & 0x3f;
 sreg_t result;
 if (shamt == 0) {
-    result = val;
+  result = val;
 } else {
-    result = (val >> shamt) + ((val >> (shamt - 1)) & 1);
+  result = (val >> shamt) + ((val >> (shamt - 1)) & 1);
 }
-WRITE_RD(result);
+WRITE_RD(sext32(result));

riscv/insns/nsrari.h

@@ -4,8 +4,8 @@ sreg_t val = (sreg_t)P_RS1_PAIR;
 uint32_t shamt = insn.shamtd();
 sreg_t result;
 if (shamt == 0) {
-    result = val;
+  result = val;
 } else {
-    result = (val >> shamt) + ((val >> (shamt - 1)) & 1);
+  result = (val >> shamt) + ((val >> (shamt - 1)) & 1);
 }
-WRITE_RD(result);
+WRITE_RD(sext32(result));

riscv/insns/nsrl.h

@@ -1,3 +1,3 @@
 require_extension('P');
 require_rv32;
-WRITE_RD(P_RS1_PAIR >> (RS2 & 0x3f));
+WRITE_RD(sext32(P_RS1_PAIR >> (RS2 & 0x3f)));

riscv/insns/nsrli.h

@@ -1,3 +1,3 @@
 require_extension('P');
 require_rv32;
-WRITE_RD(P_RS1_PAIR >> insn.shamtd());
+WRITE_RD(sext32(P_RS1_PAIR >> insn.shamtd()));

riscv/insns/pm2wsub_h.h

@@ -1,7 +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;
+  else
+    p_res += p_rs1 * p_rs2;
 })

riscv/insns/pm2wsub_hx.h

@@ -1,7 +1,7 @@
 require_rv32;
 P_WIDEN_REDUCTION_CROSS_LOOP(32, 16, false, false, {
   if (j & 1)
-    p_res += p_rs1 * p_rs2;
+    p_res -= (int64_t)p_rs1 * p_rs2;
   else
-    p_res -= p_rs1 * p_rs2;
-})
+    p_res += (int64_t)p_rs1 * p_rs2;
+})

riscv/insns/pm2wsuba_h.h

@@ -1,7 +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;
+  else
+    p_res += p_rs1 * p_rs2;
 })

riscv/insns/pm2wsuba_hx.h

@@ -1,7 +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;
-})
+  else
+    p_res += p_rs1 * p_rs2;
+})

riscv/insns/pmhacc_h.h

@@ -1,4 +1,4 @@
 P_RD_RS1_RS2_LOOP(16,16,16, {
   int32_t mres = sext32(p_rs1) * sext32(p_rs2);
-  p_rd += mres>>16; 
-})
+  p_rd += mres>>16;
+})

riscv/insns/pmhacc_h_b0.h

@@ -1,4 +1,4 @@
 P_RD_RS1_RS2_E_LOOP(16,16,8, {
   int32_t mres = sext32(p_rs1) * sext32(p_rs2);
-  p_rd += mres>>16; 
-})
+  p_rd += mres>>8;
+})

riscv/insns/pmhacc_h_b1.h

@@ -1,4 +1,4 @@
 P_RD_RS1_RS2_O_LOOP(16,16,8, {
   int32_t mres = sext32(p_rs1) * sext32(p_rs2);
-  p_rd += mres>>16; 
-})
+  p_rd += mres>>8;
+})

riscv/insns/pmhacc_w_h0.h

@@ -1,6 +1,6 @@
 require_rv64;
 P_RD_RS1_RS2_E_LOOP(32,32,16, {
   sreg_t mres = sext(p_rs1,64) * sext(p_rs2,64);
-  p_rd += mres>>32; 
+  p_rd += mres>>16;
 }
 )

riscv/insns/pmhacc_w_h1.h

@@ -1,6 +1,6 @@
 require_rv64;
 P_RD_RS1_RS2_O_LOOP(32,32,16, {
   sreg_t mres = sext(p_rs1,64) * sext(p_rs2,64);
-  p_rd += mres>>32; 
+  p_rd += mres>>16;
 }
 )

riscv/insns/pmhaccsu_h_b0.h

@@ -1,4 +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;
-})
+  p_rd += mres>>8;
+})

riscv/insns/pmhaccsu_h_b1.h

@@ -1,4 +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;
-})
+  p_rd += mres>>8;
+})

riscv/insns/pmhaccsu_w_h0.h

@@ -1,6 +1,6 @@
 require_rv64;
 P_RD_RS1_RS2_E_SULOOP(32,32,16, {
   sreg_t mres = sext(p_rs1,64) * zext(p_rs2,64);
-  p_rd += mres>>32;
+  p_rd += mres>>16;
 }
 )

riscv/insns/pmhaccsu_w_h1.h

@@ -1,6 +1,6 @@
 require_rv64;
 P_RD_RS1_RS2_O_SULOOP(32,32,16, {
   sreg_t mres = sext(p_rs1,64) * zext(p_rs2,64);
-  p_rd += mres>>32;
+  p_rd += mres>>16;
 }
 )

riscv/insns/pmqrwacc_h.h

@@ -1,4 +1,5 @@
 require_rv32;
 P_WIDEN_RD_RS1_RS2_LOOP(16, 16, {
-  p_rd += (p_rs1 * p_rs2 + 0x4000) >> 15;
-})
+  int64_t val = ((int64_t)(int16_t)p_rs1 * (int64_t)(int16_t)p_rs2 + 0x4000) >> 15;
+  p_rd += val;
+})

riscv/insns/pmqwacc_h.h

@@ -1,4 +1,5 @@
 require_rv32;
 P_WIDEN_RD_RS1_RS2_LOOP(16, 16, {
-  p_rd += (p_rs1 * p_rs2) >> 15;
-})
+  int64_t val = ((int64_t)(int16_t)p_rs1 * (int64_t)(int16_t)p_rs2) >> 15;
+  p_rd += val;
+})

riscv/insns/pmulh_h_b0.h

@@ -1,4 +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;
-})
+  p_rd = mres >> 8;
+})

riscv/insns/pmulh_h_b1.h

@@ -1,4 +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;
-})
+  p_rd = mres >> 8;
+})

riscv/insns/pmulh_w_h0.h

@@ -1,6 +1,6 @@
 require_rv64;
 P_RD_RS1_RS2_E_LOOP(32,32,16, {
   sreg_t mres = sext(p_rs1,64) * sext(p_rs2,64);
-  p_rd = mres >> 32;
+  p_rd = mres >> 16;
 }
 )

riscv/insns/pmulh_w_h1.h

@@ -1,6 +1,6 @@
 require_rv64;
 P_RD_RS1_RS2_O_LOOP(32,32,16, {
   sreg_t mres = sext(p_rs1,64) * sext(p_rs2,64);
-  p_rd = mres >> 32;
+  p_rd = mres >> 16;
 }
 )

riscv/insns/pmulhsu_h_b0.h

@@ -1,4 +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;
-})
+  p_rd = mres >> 8;
+})

riscv/insns/pmulhsu_h_b1.h

@@ -1,4 +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;
-})
+  p_rd = mres >> 8;
+})

riscv/insns/pmulhsu_w_h1.h

@@ -1,6 +1,6 @@
 require_rv64;
 P_RD_RS1_RS2_O_SULOOP(32,32,16, {
   sreg_t mres = sext(p_rs1,64) * zext(p_rs2,64);
-  p_rd = mres >> 32;
+  p_rd = mres >> 16;
 }
 )

riscv/insns/psadd_b.h

@@ -1,4 +1,6 @@
 P_RD_RS1_RS2_LOOP(8,8,8, {
   bool sat = false;
   p_rd = (sat_add<int8_t, uint8_t>(p_rs1, p_rs2, sat));
-})
+  if (sat)
+    P.set_vxsat();
+})

riscv/insns/psadd_db.h

@@ -2,4 +2,6 @@ require_rv32;
 P_RD_RS1_RS2_DW_LOOP(8,8,8, {
   bool sat = false;
   p_rd = (sat_add<int8_t, uint8_t>(p_rs1, p_rs2, sat));
-})
+  if (sat)
+    P.set_vxsat();
+})

riscv/insns/psadd_dh.h

@@ -2,4 +2,6 @@ require_rv32;
 P_RD_RS1_RS2_DW_LOOP(16,16,16, {
   bool sat = false;
   p_rd = (sat_add<int16_t, uint16_t>(p_rs1, p_rs2, sat));
-})
+  if (sat)
+    P.set_vxsat();
+})

riscv/insns/psadd_dw.h

@@ -2,4 +2,6 @@ require_rv32;
 P_RD_RS1_RS2_DW_LOOP(32,32,32, {
   bool sat = false;
   p_rd = (sat_add<int32_t, uint32_t>(p_rs1, p_rs2, sat));
-})
+  if (sat)
+    P.set_vxsat();
+})

riscv/insns/psadd_h.h

@@ -1,4 +1,6 @@
 P_RD_RS1_RS2_LOOP(16,16,16, {
   bool sat = false;
   p_rd = (sat_add<int16_t, uint16_t>(p_rs1, p_rs2, sat));
-})
+  if (sat)
+    P.set_vxsat();
+})

riscv/insns/psadd_w.h

@@ -2,5 +2,7 @@ require_rv64;
 P_RD_RS1_RS2_LOOP(32,32,32, {
   bool sat = false;
   p_rd = (sat_add<int32_t, uint32_t>(p_rs1, p_rs2, sat));
+  if (sat)
+    P.set_vxsat();
 }
 )

riscv/insns/psaddu_b.h

@@ -1,4 +1,6 @@
 P_RD_RS1_RS2_ULOOP(8,8,8, {
   bool sat = false;
   p_rd = (sat_addu<uint8_t>(p_rs1, p_rs2, sat));
-})
+  if (sat)
+    P.set_vxsat();
+})

riscv/insns/psaddu_db.h

@@ -2,4 +2,6 @@ require_rv32;
 P_RD_RS1_RS2_DW_ULOOP(8,8,8, {
   bool sat = false;
   p_rd = (sat_addu<uint8_t>(p_rs1, p_rs2, sat));
-})
+  if (sat)
+    P.set_vxsat();
+})

riscv/insns/psaddu_dh.h

@@ -2,4 +2,6 @@ require_rv32;
 P_RD_RS1_RS2_DW_ULOOP(16,16,16, {
   bool sat = false;
   p_rd = (sat_addu<uint16_t>(p_rs1, p_rs2, sat));
-})
+  if (sat)
+    P.set_vxsat();
+})

riscv/insns/psaddu_dw.h

@@ -2,4 +2,6 @@ require_rv32;
 P_RD_RS1_RS2_DW_ULOOP(32,32,32, {
   bool sat = false;
   p_rd = (sat_addu<uint32_t>(p_rs1, p_rs2, sat));
-})
+  if (sat)
+    P.set_vxsat();
+})

riscv/insns/psaddu_h.h

@@ -1,4 +1,6 @@
 P_RD_RS1_RS2_ULOOP(16,16,16, {
   bool sat = false;
   p_rd = (sat_addu<uint16_t>(p_rs1, p_rs2, sat));
-})
+  if (sat)
+    P.set_vxsat();
+})

riscv/insns/psaddu_w.h

@@ -2,5 +2,7 @@ require_rv64;
 P_RD_RS1_RS2_ULOOP(32,32,32, {
   bool sat = false;
   p_rd = (sat_addu<uint32_t>(p_rs1, p_rs2, sat));
+  if (sat)
+    P.set_vxsat();
 }
 )

riscv/insns/psll_dws.h

@@ -1,7 +1,6 @@
 require_rv32;
 P_RD_RS1_DW_LOOP(32, 32, {
-  uint8_t m = P_FIELD(RS2, 0, 8);
+  uint8_t m = RS2 & 0x1F;
   const uint64_t maskN = 0xFFFFFFFFull;
-  if (m >= 32) p_rd = 0;
-  else p_rd = (uint32_t)((p_rs1 << m) & maskN);
+  p_rd = (uint32_t)((p_rs1 << m) & maskN);
 })

riscv/insns/pssh1sadd_w.h

@@ -1,5 +1,5 @@
 require_rv64;
 P_RD_RS1_RS2_LOOP(32, 32, 32, {
-  p_rd = P_SAT(32, P_SAT(32, p_rs1 << 1) + p_rs2);
+  p_rd = P_SAT(32, P_SAT(32, (int64_t)p_rs1 << 1) + p_rs2);
 }
 )

riscv/insns/pssha_dhs.h

@@ -1,45 +1,26 @@
 require_rv32;
 P_RD_RS1_DW_LOOP(16, 16, {
-  uint64_t bits_SMIN = (uint64_t{1} << (16 - 1));
-  uint64_t bits_SMAX = ((uint64_t{1} << (16 - 1)) - 1);
   bool ov = false;
-  int8_t m = P_FIELD(RS2, 0, 8);
-  int8_t  rev = static_cast<int8_t>(m);
-  rev = (m < 0) ? static_cast<uint8_t>(~m + 1u) : m;
-  uint64_t mask = ((uint64_t{1} << 16) - 1);
-  p_rs1 &= mask;
-  if(m < 0){
-    unsigned sh = ((unsigned)(uint8_t)rev > 255u) ? 255u : (unsigned)(uint8_t)rev;
-    uint64_t sign = (p_rs1 >> (16 - 1)) & 1u;
-
-    if(sh >= 16) p_rd = (uint16_t)(sign ? mask : 0u);
-    else{
-      uint64_t shifted = (p_rs1 >> sh);
-      uint64_t fill = (~uint64_t{0}) << (16 - sh);
-      shifted |= fill;
-      p_rd = (uint16_t)(shifted & mask);
-    }
-  }
-  else{
-    if(rev==0) p_rd = (uint16_t)p_rs1;
-    else if(rev >= 16){
-      if(p_rs1==0) p_rd = 0;
-      else{
-        ov = true;
-        uint64_t sign = (p_rs1 >> (16 - 1)) & 1u;
-        p_rd = (uint16_t)(sign ? bits_SMIN : bits_SMAX);
-      }
-    }
-    else{
-      uint64_t sign = (p_rs1 >> (16 - 1)) & 1u;
-      uint64_t top  = (p_rs1 >> (16 - rev));
-      uint64_t need = sign ? ((uint64_t{1} << rev) - 1) : 0u;
-      ov = (top != need);
-      if(ov)
-        p_rd = (uint16_t)(sign ? bits_SMIN : bits_SMAX);
-      else
-        p_rd =  (uint16_t)((p_rs1 << rev) & mask);
+  int8_t sshamt = P_FIELD(RS2, 0, 8);
+  int32_t val = p_rs1;
+  if (sshamt < 0) {
+    val = (int16_t)(val >> std::min(-sshamt, 16));
+  } else if (sshamt >= 16 && val != 0){
+    val = val > 0 ? INT16_MAX : INT16_MIN;
+    ov = true;
+  } else if (val != 0) {
+    int32_t tmp = (int32_t)val << sshamt;
+    if (tmp > INT16_MAX) {
+      val = INT16_MAX;
+      ov = true;
+    } else if (tmp < INT16_MIN) {
+      val = INT16_MIN;
+      ov = true;
+    } else {
+      val = tmp;
     }
   }
+  p_rd = val;
+
   if (ov) P.set_vxsat();
-})
+})

riscv/insns/pssha_dws.h

@@ -1,45 +1,26 @@
 require_rv32;
 P_RD_RS1_DW_LOOP(32, 32, {
-  uint64_t bits_SMIN = (uint64_t{1} << (32 - 1));
-  uint64_t bits_SMAX = ((uint64_t{1} << (32 - 1)) - 1);
   bool ov = false;
-  int8_t m = P_FIELD(RS2, 0, 8);
-  int8_t  rev = static_cast<int8_t>(m);
-  rev = (m < 0) ? static_cast<uint8_t>(~m + 1u) : m;
-  uint64_t mask = ((uint64_t{1} << 32) - 1);
-  p_rs1 &= mask;
-  if(m < 0){
-    unsigned sh = ((unsigned)(uint8_t)rev > 255u) ? 255u : (unsigned)(uint8_t)rev;
-    uint64_t sign = (p_rs1 >> (32 - 1)) & 1u;
-
-    if(sh >= 32) p_rd = (uint32_t)(sign ? mask : 0u);
-    else{
-      uint64_t shifted = (p_rs1 >> sh);
-      uint64_t fill = (~uint64_t{0}) << (32 - sh);
-      shifted |= fill;
-      p_rd = (uint32_t)(shifted & mask);
-    }
-  }
-  else{
-    if(rev==0) p_rd = (uint32_t)p_rs1;
-    else if(rev >= 32){
-      if(p_rs1==0) p_rd = 0;
-      else{
-        ov = true;
-        uint64_t sign = (p_rs1 >> (32 - 1)) & 1u;
-        p_rd = (uint32_t)(sign ? bits_SMIN : bits_SMAX);
-      }
-    }
-    else{
-      uint64_t sign = (p_rs1 >> (32 - 1)) & 1u;
-      uint64_t top  = (p_rs1 >> (32 - rev));
-      uint64_t need = sign ? ((uint64_t{1} << rev) - 1) : 0u;
-      ov = (top != need);
-      if(ov)
-        p_rd = (uint32_t)(sign ? bits_SMIN : bits_SMAX);
-      else
-        p_rd =  (uint32_t)((p_rs1 << rev) & mask);
+  int8_t sshamt = P_FIELD(RS2, 0, 8);
+  int64_t val = p_rs1;
+  if (sshamt < 0) {
+    val = (int32_t)(val >> std::min(-sshamt, 32));
+  } else if (sshamt >= 32 && val != 0){
+    val = val > 0 ? INT32_MAX : INT32_MIN;
+    ov = true;
+  } else if (val != 0) {
+    int64_t tmp = (int64_t)val << sshamt;
+    if (tmp > INT32_MAX) {
+      val = INT32_MAX;
+      ov = true;
+    } else if (tmp < INT32_MIN) {
+      val = INT32_MIN;
+      ov = true;
+    } else {
+      val = tmp;
     }
   }
+  p_rd = val;
+
   if (ov) P.set_vxsat();
 })

riscv/insns/pssha_hs.h

@@ -3,11 +3,11 @@ P_RD_RS1_LOOP(16, 16, {
   if (p_rs1 == 0)
     p_rd = 0;
   else if (sshamt >= 16) {
-    p_rd = (p_rs1 & 0x8000) ? 0x8000 : 0x7fff;
+    p_rd = (p_rs1 & 0x8000) ? INT16_MIN : INT16_MAX;
     P.set_vxsat();
   }
   else if (sshamt <= -16)
-    p_rd = (p_rs1 & 0x8000) ? 0xffff : 0;
+    p_rd = (p_rs1 & 0x8000) ? (int16_t)UINT16_MAX : 0;
   else
     p_rd = sshamt >= 0 ? P_SAT(16, sext32(p_rs1) << sshamt) : (p_rs1 >> -sshamt);
-})
+})

riscv/insns/pssha_ws.h

@@ -4,11 +4,11 @@ P_RD_RS1_LOOP(32, 32, {
   if (p_rs1 == 0)
     p_rd = 0;
   else if (sshamt >= 32) {
-    p_rd = (p_rs1 & 0x80000000) ? 0x80000000 : 0x7fffffff;
+    p_rd = (p_rs1 & 0x80000000) ? INT32_MIN : INT32_MAX;
     P.set_vxsat();
   }
   else if (sshamt <= -32)
-    p_rd = (p_rs1 & 0x80000000) ? 0xffffffff : 0;
+    p_rd = (p_rs1 & 0x80000000) ? (int32_t)UINT32_MAX : 0;
   else
     p_rd = sshamt >= 0 ? P_SAT(32, sext32(p_rs1) << sshamt) : (p_rs1 >> -sshamt);
 }

riscv/insns/psshar_dhs.h

@@ -1,64 +1,27 @@
 require_rv32;
 P_RD_RS1_DW_LOOP(16, 16, {
-    uint64_t bits_SMIN = (uint64_t{1} << (16 - 1));
-    uint64_t bits_SMAX = ((uint64_t{1} << (16 - 1)) - 1);
-    bool ov = false;
-    int8_t m = P_FIELD(RS2, 0, 8);
-    int8_t  rev = static_cast<int8_t>(m);
-    rev = (m < 0) ? static_cast<uint8_t>(~m + 1u) : m;
-    uint64_t mask = ((uint64_t{1} << 16) - 1);
-    p_rs1 &= mask;
-    if(m < 0){
-        if ((rev & 0xFFu) == 0u) 
-            p_rd = (uint16_t)p_rs1;
-        else{  
-            int128_t v_sext;
-            bool neg = ((p_rs1 >> (16 - 1)) & 1u);
-            if(!neg) v_sext = static_cast<int128_t>(p_rs1);
-            else v_sext = static_cast<int128_t>((~static_cast<uint128_t>(0) << 16) | static_cast<uint128_t>(p_rs1));
-            int128_t v_cat0 = v_sext << 1;
-
-            unsigned sh = ((unsigned)(uint8_t)rev > 255u) ? 255u : (unsigned)(uint8_t)rev;
-
-            int128_t sra_val;
-            if(sh == 0)
-                sra_val = v_cat0;
-            else if(sh >=127)
-                sra_val = (v_cat0 < 0) ? static_cast<int128_t>(-1) : static_cast<int128_t>(0);
-            else{
-                int128_t ux = static_cast<uint128_t>(v_cat0);
-                int128_t shifted = ux >> sh;
-                if(v_cat0 < 0)
-                    shifted |= (~static_cast<uint128_t>(0)) << (128 - sh);
-                sra_val = static_cast<int128_t>(shifted);
-            }
-
-            int128_t plus1 = sra_val + static_cast<int128_t>(1);
-            uint128_t ures = static_cast<uint128_t>(plus1);
-            p_rd = (uint16_t)(static_cast<uint64_t>((ures >> 1) & static_cast<uint128_t>(mask)));
-        }
-    }
-    else{
-        if(rev==0) p_rd = (uint16_t)p_rs1;
-        else if(rev >= 16){
-            if(p_rs1==0) 
-                p_rd = 0;
-            else{
-                ov = true;
-                uint64_t sign = (p_rs1 >> (16 - 1)) & 1u;
-                p_rd = (uint16_t)(sign ? bits_SMIN : bits_SMAX);
-            }
-        }
-        else{
-            uint64_t sign = (p_rs1 >> (16 - 1)) & 1u;
-            uint64_t top  = (p_rs1 >> (16 - rev));
-            uint64_t need = sign ? ((uint64_t{1} << rev) - 1) : 0u;
-            ov = (top != need);
-            if(ov)
-                p_rd = (uint16_t)(sign ? bits_SMIN : bits_SMAX);
-            else
-                p_rd =  (uint16_t)((p_rs1 << rev) & mask);
-        }
+  bool ov = false;
+  int8_t sshamt = P_FIELD(RS2, 0, 8);
+  int32_t val = p_rs1;
+  if (sshamt < 0) {
+    val = (val << 1) >> std::min(-sshamt, 16);
+    val = (int16_t)((val + 1) >> 1);
+  } else if (sshamt >= 16 && val != 0){
+    val = val > 0 ? INT16_MAX : INT16_MIN;
+    ov = true;
+  } else if (val != 0) {
+    int32_t tmp = (int32_t)val << sshamt;
+    if (tmp > INT16_MAX) {
+      val = INT16_MAX;
+      ov = true;
+    } else if (tmp < INT16_MIN) {
+      val = INT16_MIN;
+      ov = true;
+    } else {
+      val = tmp;
     }
+  }
+  p_rd = val;
+
   if (ov) P.set_vxsat();
 })

riscv/insns/psshar_dws.h

@@ -1,64 +1,27 @@
 require_rv32;
 P_RD_RS1_DW_LOOP(32, 32, {
-    uint64_t bits_SMIN = (uint64_t{1} << (32 - 1));
-    uint64_t bits_SMAX = ((uint64_t{1} << (32 - 1)) - 1);
-    bool ov = false;
-    int8_t m = P_FIELD(RS2, 0, 8);
-    int8_t  rev = static_cast<int8_t>(m);
-    rev = (m < 0) ? static_cast<uint8_t>(~m + 1u) : m;
-    uint64_t mask = ((uint64_t{1} << 32) - 1);
-    p_rs1 &= mask;
-    if(m < 0){
-        if ((rev & 0xFFu) == 0u) 
-            p_rd = (uint32_t)p_rs1;
-        else{  
-            int128_t v_sext;
-            bool neg = ((p_rs1 >> (32 - 1)) & 1u);
-            if(!neg) v_sext = static_cast<int128_t>(p_rs1);
-            else v_sext = static_cast<int128_t>((~static_cast<uint128_t>(0) << 32) | static_cast<uint128_t>(p_rs1));
-            int128_t v_cat0 = v_sext << 1;
-
-            unsigned sh = ((unsigned)(uint8_t)rev > 255u) ? 255u : (unsigned)(uint8_t)rev;
-
-            int128_t sra_val;
-            if(sh == 0)
-                sra_val = v_cat0;
-            else if(sh >=127)
-                sra_val = (v_cat0 < 0) ? static_cast<int128_t>(-1) : static_cast<int128_t>(0);
-            else{
-                int128_t ux = static_cast<uint128_t>(v_cat0);
-                int128_t shifted = ux >> sh;
-                if(v_cat0 < 0)
-                    shifted |= (~static_cast<uint128_t>(0)) << (128 - sh);
-                sra_val = static_cast<int128_t>(shifted);
-            }
-
-            int128_t plus1 = sra_val + static_cast<int128_t>(1);
-            uint128_t ures = static_cast<uint128_t>(plus1);
-            p_rd = (uint32_t)(static_cast<uint64_t>((ures >> 1) & static_cast<uint128_t>(mask)));
-        }
-    }
-    else{
-        if(rev==0) p_rd = (uint32_t)p_rs1;
-        else if(rev >= 32){
-            if(p_rs1==0)
-                p_rd = 0;
-            else{
-                ov = true;
-                uint64_t sign = (p_rs1 >> (32 - 1)) & 1u;
-                p_rd = (uint32_t)(sign ? bits_SMIN : bits_SMAX);
-            }
-        }
-        else{
-            uint64_t sign = (p_rs1 >> (32 - 1)) & 1u;
-            uint64_t top  = (p_rs1 >> (32 - rev));
-            uint64_t need = sign ? ((uint64_t{1} << rev) - 1) : 0u;
-            ov = (top != need);
-            if(ov)
-                p_rd = (uint32_t)(sign ? bits_SMIN : bits_SMAX);
-            else
-                p_rd =  (uint32_t)((p_rs1 << rev) & mask);
-        }
+  bool ov = false;
+  int8_t sshamt = P_FIELD(RS2, 0, 8);
+  int64_t val = p_rs1;
+  if (sshamt < 0) {
+    val = (val << 1) >> std::min(-sshamt, 32);
+    val = (int32_t)((val + 1) >> 1);
+  } else if (sshamt >= 32 && val != 0){
+    val = val > 0 ? INT32_MAX : INT32_MIN;
+    ov = true;
+  } else if (val != 0) {
+    int64_t tmp = (int64_t)val << sshamt;
+    if (tmp > INT32_MAX) {
+      val = INT32_MAX;
+      ov = true;
+    } else if (tmp < INT32_MIN) {
+      val = INT32_MIN;
+      ov = true;
+    } else {
+      val = tmp;
     }
+  }
+  p_rd = val;
+
   if (ov) P.set_vxsat();
 })

riscv/insns/psshar_hs.h

@@ -3,11 +3,11 @@ P_RD_RS1_LOOP(16, 16, {
   if (p_rs1 == 0)
     p_rd = 0;
   else if (sshamt >= 16) {
-    p_rd = (p_rs1 & 0x8000) ? 0x8000 : 0x7fff;
+    p_rd = (p_rs1 & 0x8000) ? INT16_MIN : INT16_MAX;
     P.set_vxsat();
   }
   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));
-})
+})

riscv/insns/psshar_ws.h

@@ -4,7 +4,7 @@ P_RD_RS1_LOOP(32, 32, {
   if (p_rs1 == 0)
     p_rd = 0;
   else if (sshamt >= 32) {
-    p_rd = (p_rs1 & 0x80000000) ? 0x80000000 : 0x7fffffff;
+    p_rd = (p_rs1 & 0x80000000) ? INT32_MIN : INT32_MAX;
     P.set_vxsat();
   }
   else if (sshamt <= -32)

riscv/insns/psshl_dhs.h

@@ -1,16 +1,13 @@
 require_rv32;
-sreg_t sshamt = P_FIELD(RS2, 0, 8);
+int sshamt = P_FIELD(RS2, 0, 8);
 P_RD_RS1_DW_LOOP(16, 16, {
   if (sshamt < 0) {
-    if (sshamt <= -16)
-      p_rd = 0;
-    else
-      p_rd = (uint16_t)p_rs1 >> (-sshamt);
+    p_rd = (uint32_t)(uint16_t)p_rs1 >> std::min(-sshamt, 16);
   } else {
-    uint32_t shx = (sshamt >= 16) ? ((uint32_t)(uint16_t)p_rs1 << 16) : ((uint32_t)(uint16_t)p_rs1 << sshamt);
-    if (shx > 0xFFFF) {
+    uint32_t shx = (uint32_t)(uint16_t)p_rs1 << std::min(sshamt, 16);
+    if (shx > UINT16_MAX) {
       P.set_vxsat();
-      p_rd = 0xFFFF;
+      p_rd = (uint16_t)UINT16_MAX;
     } else {
       p_rd = (uint16_t)shx;
     }

riscv/insns/psshl_dws.h

@@ -1,16 +1,13 @@
 require_rv32;
-sreg_t sshamt = P_FIELD(RS2, 0, 8);
+int sshamt = P_FIELD(RS2, 0, 8);
 P_RD_RS1_DW_LOOP(32, 32, {
   if (sshamt < 0) {
-    if (sshamt <= -32)
-      p_rd = 0;
-    else
-      p_rd = (uint32_t)p_rs1 >> (-sshamt);
+    p_rd = (uint64_t)(uint32_t)p_rs1 >> std::min(-sshamt, 32);
   } else {
-    uint64_t shx = (sshamt >= 32) ? ((uint64_t)(uint32_t)p_rs1 << 32) : ((uint64_t)(uint32_t)p_rs1 << sshamt);
-    if (shx > 0xFFFFFFFFULL) {
+    uint64_t shx = (uint64_t)(uint32_t)p_rs1 << std::min(sshamt, 32);
+    if (shx > UINT32_MAX) {
       P.set_vxsat();
-      p_rd = 0xFFFFFFFF;
+      p_rd = UINT32_MAX;
     } else {
       p_rd = (uint32_t)shx;
     }

riscv/insns/psshl_hs.h

@@ -1,15 +1,12 @@
-sreg_t sshamt = P_FIELD(RS2, 0, 8);
+int sshamt = P_FIELD(RS2, 0, 8);
 P_RD_RS1_LOOP(16, 16, {
   if (sshamt < 0) {
-    if (sshamt <= -16)
-      p_rd = 0;
-    else
-      p_rd = (uint16_t)p_rs1 >> (-sshamt);
+    p_rd = (uint32_t)(uint16_t)p_rs1 >> std::min(-sshamt, 16);
   } else {
-    uint32_t shx = (sshamt >= 16) ? ((uint32_t)(uint16_t)p_rs1 << 16) : ((uint32_t)(uint16_t)p_rs1 << sshamt);
-    if (shx > 0xFFFF) {
+    uint32_t shx = (uint32_t)(uint16_t)p_rs1 << std::min(sshamt, 16);
+    if (shx > UINT16_MAX) {
       P.set_vxsat();
-      p_rd = 0xFFFF;
+      p_rd = (uint16_t)UINT16_MAX;
     } else {
       p_rd = (uint16_t)shx;
     }

riscv/insns/psshl_ws.h

@@ -1,16 +1,13 @@
 require_rv64;
-sreg_t sshamt = P_FIELD(RS2, 0, 8);
+int sshamt = P_FIELD(RS2, 0, 8);
 P_RD_RS1_LOOP(32, 32, {
   if (sshamt < 0) {
-    if (sshamt <= -32)
-      p_rd = 0;
-    else
-      p_rd = (uint32_t)p_rs1 >> (-sshamt);
+    p_rd = (uint64_t)(uint32_t)p_rs1 >> std::min(-sshamt, 32);
   } else {
-    uint64_t shx = (sshamt >= 32) ? ((uint64_t)(uint32_t)p_rs1 << 32) : ((uint64_t)(uint32_t)p_rs1 << sshamt);
-    if (shx > 0xFFFFFFFFULL) {
+    uint64_t shx = (uint64_t)(uint32_t)p_rs1 << std::min(sshamt, 32);
+    if (shx > UINT32_MAX) {
       P.set_vxsat();
-      p_rd = 0xFFFFFFFF;
+      p_rd = UINT32_MAX;
     } else {
       p_rd = (uint32_t)shx;
     }

riscv/insns/psshlr_dhs.h

@@ -1,20 +1,14 @@
 require_rv32;
-sreg_t sshamt = P_FIELD(RS2, 0, 8);
+int sshamt = P_FIELD(RS2, 0, 8);
 P_RD_RS1_DW_LOOP(16, 16, {
   if (sshamt < 0) {
-    uint32_t shx;
-    if (sshamt < -16)
-      shx = 0;
-    else if (sshamt == -16)
-      shx = ((uint16_t)p_rs1 >> 15) & 1;
-    else
-      shx = ((uint32_t)(uint16_t)p_rs1 << 1) >> (-sshamt);
+    uint32_t shx = ((uint32_t)(uint16_t)p_rs1 << 1) >> std::min(-sshamt, 16);
     p_rd = (uint16_t)((shx + 1) >> 1);
   } else {
-    uint32_t shx = (sshamt >= 16) ? ((uint32_t)(uint16_t)p_rs1 << 16) : ((uint32_t)(uint16_t)p_rs1 << sshamt);
-    if (shx > 0xFFFF) {
+    uint32_t shx = (uint32_t)(uint16_t)p_rs1 << std::min(sshamt, 16);
+    if (shx > UINT16_MAX) {
       P.set_vxsat();
-      p_rd = 0xFFFF;
+      p_rd = (uint16_t)UINT16_MAX;
     } else {
       p_rd = (uint16_t)shx;
     }

riscv/insns/psshlr_dws.h

@@ -1,20 +1,14 @@
 require_rv32;
-sreg_t sshamt = P_FIELD(RS2, 0, 8);
+int sshamt = P_FIELD(RS2, 0, 8);
 P_RD_RS1_DW_LOOP(32, 32, {
   if (sshamt < 0) {
-    uint64_t shx;
-    if (sshamt < -32)
-      shx = 0;
-    else if (sshamt == -32)
-      shx = ((uint32_t)p_rs1 >> 31) & 1;
-    else
-      shx = ((uint64_t)(uint32_t)p_rs1 << 1) >> (-sshamt);
+    uint64_t shx = ((uint64_t)(uint32_t)p_rs1 << 1) >> std::min(-sshamt, 32);
     p_rd = (uint32_t)((shx + 1) >> 1);
   } else {
-    uint64_t shx = (sshamt >= 32) ? ((uint64_t)(uint32_t)p_rs1 << 32) : ((uint64_t)(uint32_t)p_rs1 << sshamt);
-    if (shx > 0xFFFFFFFFULL) {
+    uint64_t shx = (uint64_t)(uint32_t)p_rs1 << std::min(sshamt, 32);
+    if (shx > UINT32_MAX) {
       P.set_vxsat();
-      p_rd = 0xFFFFFFFF;
+      p_rd = UINT32_MAX;
     } else {
       p_rd = (uint32_t)shx;
     }

riscv/insns/psshlr_hs.h

@@ -1,19 +1,13 @@
-sreg_t sshamt = P_FIELD(RS2, 0, 8);
+int sshamt = P_FIELD(RS2, 0, 8);
 P_RD_RS1_LOOP(16, 16, {
   if (sshamt < 0) {
-    uint32_t shx;
-    if (sshamt < -16)
-      shx = 0;
-    else if (sshamt == -16)
-      shx = ((uint16_t)p_rs1 >> 15) & 1;
-    else
-      shx = ((uint32_t)(uint16_t)p_rs1 << 1) >> (-sshamt);
+    uint32_t shx = ((uint32_t)(uint16_t)p_rs1 << 1) >> std::min(-sshamt, 16);
     p_rd = (uint16_t)((shx + 1) >> 1);
   } else {
-    uint32_t shx = (sshamt >= 16) ? ((uint32_t)(uint16_t)p_rs1 << 16) : ((uint32_t)(uint16_t)p_rs1 << sshamt);
-    if (shx > 0xFFFF) {
+    uint32_t shx = (uint32_t)(uint16_t)p_rs1 << std::min(sshamt, 16);
+    if (shx > UINT16_MAX) {
       P.set_vxsat();
-      p_rd = 0xFFFF;
+      p_rd = (uint16_t)UINT16_MAX;
     } else {
       p_rd = (uint16_t)shx;
     }

riscv/insns/psshlr_ws.h

@@ -1,20 +1,14 @@
 require_rv64;
-sreg_t sshamt = P_FIELD(RS2, 0, 8);
+int sshamt = P_FIELD(RS2, 0, 8);
 P_RD_RS1_LOOP(32, 32, {
   if (sshamt < 0) {
-    uint64_t shx;
-    if (sshamt < -32)
-      shx = 0;
-    else if (sshamt == -32)
-      shx = ((uint32_t)p_rs1 >> 31) & 1;
-    else
-      shx = ((uint64_t)(uint32_t)p_rs1 << 1) >> (-sshamt);
+    uint64_t shx = ((uint64_t)(uint32_t)p_rs1 << 1) >> std::min(-sshamt, 32);
     p_rd = (uint32_t)((shx + 1) >> 1);
   } else {
-    uint64_t shx = (sshamt >= 32) ? ((uint64_t)(uint32_t)p_rs1 << 32) : ((uint64_t)(uint32_t)p_rs1 << sshamt);
-    if (shx > 0xFFFFFFFFULL) {
+    uint64_t shx = (uint64_t)(uint32_t)p_rs1 << std::min(sshamt, 32);
+    if (shx > UINT32_MAX) {
       P.set_vxsat();
-      p_rd = 0xFFFFFFFF;
+      p_rd = UINT32_MAX;
     } else {
       p_rd = (uint32_t)shx;
     }

riscv/insns/pssub_b.h

@@ -1,4 +1,6 @@
 P_RD_RS1_RS2_LOOP(8,8,8, {
   bool sat = false;
   p_rd = (sat_sub<int8_t, uint8_t>(p_rs1, p_rs2, sat));
-})
+  if (sat)
+    P.set_vxsat();
+})

riscv/insns/pssub_db.h

@@ -2,4 +2,6 @@ require_rv32;
 P_RD_RS1_RS2_DW_LOOP(8,8,8, {
   bool sat = false;
   p_rd = (sat_sub<int8_t, uint8_t>(p_rs1, p_rs2, sat));
-})
+  if (sat)
+    P.set_vxsat();
+})