[RV64_DYNAREC] Optimized 16bit constant ROL opcodes (#2267)

3 files changed, 124 insertions, 61 deletions

diff --git a/src/dynarec/rv64/dynarec_rv64_66.c b/src/dynarec/rv64/dynarec_rv64_66.c
index 45a3a3a9..5a35e1f2 100644
--- a/src/dynarec/rv64/dynarec_rv64_66.c
+++ b/src/dynarec/rv64/dynarec_rv64_66.c
@@ -1062,13 +1062,18 @@ uintptr_t dynarec64_66(dynarec_rv64_t* dyn, uintptr_t addr, uintptr_t ip, int ni
             switch ((nextop >> 3) & 7) {
                 case 0:
                     INST_NAME("ROL Ew, Ib");
-                    MESSAGE(LOG_DUMP, "Need Optimization\n");
-                    SETFLAGS(X_OF | X_CF, SF_SET_DF, NAT_FLAGS_NOFUSION);
-                    GETEW(x1, 1);
-                    u8 = F8;
-                    MOV32w(x2, u8);
-                    CALL_(rol16, x1, x3, x1, x2);
-                    EWBACK;
+                    u8 = geted_ib(dyn, addr, ninst, nextop) & 15;
+                    if (u8) {
+                        // removed PENDING on purpose
+                        SETFLAGS(X_OF | X_CF, SF_SUBSET, NAT_FLAGS_FUSION);
+                        GETEW(x1, 1);
+                        u8 = (F8) & 0x1f;
+                        emit_rol16c(dyn, ninst, x1, u8, x4, x5);
+                        EWBACK;
+                    } else {
+                        FAKEED;
+                        F8;
+                    }
                     break;
                 case 1:
                     INST_NAME("ROR Ew, Ib");
@@ -1178,17 +1183,76 @@ uintptr_t dynarec64_66(dynarec_rv64_t* dyn, uintptr_t addr, uintptr_t ip, int ni
             }
             break;
         case 0xD1:
+            nextop = F8;
+            switch ((nextop >> 3) & 7) {
+                case 0:
+                    INST_NAME("ROL Ew, 1");
+                    // removed PENDING on purpose
+                    SETFLAGS(X_OF | X_CF, SF_SUBSET, NAT_FLAGS_FUSION);
+                    GETEW(x1, 0);
+                    emit_rol16c(dyn, ninst, x1, 1, x5, x4);
+                    EWBACK;
+                    break;
+                case 1:
+                    INST_NAME("ROR Ew, 1");
+                    MOV32w(x2, 1);
+                    MESSAGE(LOG_DUMP, "Need Optimization\n");
+                    SETFLAGS(X_OF | X_CF, SF_SET_DF, NAT_FLAGS_NOFUSION);
+                    GETEW(x1, 1);
+                    CALL_(ror16, x1, x3, x1, x2);
+                    EWBACK;
+                    break;
+                case 2:
+                    INST_NAME("RCL Ew, 1");
+                    MOV32w(x2, 1);
+                    MESSAGE("LOG_DUMP", "Need optimization\n");
+                    READFLAGS(X_CF);
+                    SETFLAGS(X_OF | X_CF, SF_SET_DF, NAT_FLAGS_NOFUSION);
+                    GETEW(x1, 1);
+                    CALL_(rcl16, x1, x3, x1, x2);
+                    EWBACK;
+                    break;
+                case 3:
+                    INST_NAME("RCR Ew, 1");
+                    MOV32w(x2, 1);
+                    MESSAGE("LOG_DUMP", "Need optimization\n");
+                    READFLAGS(X_CF);
+                    SETFLAGS(X_OF | X_CF, SF_SET_DF, NAT_FLAGS_NOFUSION);
+                    GETEW(x1, 1);
+                    CALL_(rcr16, x1, x3, x1, x2);
+                    EWBACK;
+                    break;
+                case 5:
+                    INST_NAME("SHR Ew, 1");
+                    SETFLAGS(X_ALL, SF_SET_PENDING, NAT_FLAGS_FUSION); // some flags are left undefined
+                    GETEW(x1, 0);
+                    emit_shr16c(dyn, ninst, x1, 1, x5, x4, x6);
+                    EWBACK;
+                    break;
+                case 4:
+                case 6:
+                    INST_NAME("SHL Ew, 1");
+                    SETFLAGS(X_ALL, SF_SET_PENDING, NAT_FLAGS_FUSION); // some flags are left undefined
+                    GETEW(x1, 0);
+                    emit_shl16c(dyn, ninst, x1, 1, x5, x4, x6);
+                    EWBACK;
+                    break;
+                case 7:
+                    INST_NAME("SAR Ew, 1");
+                    SETFLAGS(X_ALL, SF_SET_PENDING, NAT_FLAGS_FUSION); // some flags are left undefined
+                    GETSEW(x1, 0);
+                    emit_sar16c(dyn, ninst, x1, 1, x5, x4, x6);
+                    EWBACK;
+                    break;
+            }
+            break;
+
         case 0xD3:
             nextop = F8;
             switch ((nextop >> 3) & 7) {
                 case 0:
-                    if (opcode == 0xD1) {
-                        INST_NAME("ROL Ew, 1");
-                        MOV32w(x2, 1);
-                    } else {
-                        INST_NAME("ROL Ew, CL");
-                        ANDI(x2, xRCX, 0x1f);
-                    }
+                    INST_NAME("ROL Ew, CL");
+                    ANDI(x2, xRCX, 0x1f);
                     MESSAGE(LOG_DUMP, "Need Optimization\n");
                     SETFLAGS(X_OF | X_CF, SF_SET_DF, NAT_FLAGS_NOFUSION);
                     GETEW(x1, 1);
@@ -1196,13 +1260,8 @@ uintptr_t dynarec64_66(dynarec_rv64_t* dyn, uintptr_t addr, uintptr_t ip, int ni
                     EWBACK;
                     break;
                 case 1:
-                    if (opcode == 0xD1) {
-                        INST_NAME("ROR Ew, 1");
-                        MOV32w(x2, 1);
-                    } else {
-                        INST_NAME("ROR Ew, CL");
-                        ANDI(x2, xRCX, 0x1f);
-                    }
+                    INST_NAME("ROR Ew, CL");
+                    ANDI(x2, xRCX, 0x1f);
                     MESSAGE(LOG_DUMP, "Need Optimization\n");
                     SETFLAGS(X_OF | X_CF, SF_SET_DF, NAT_FLAGS_NOFUSION);
                     GETEW(x1, 1);
@@ -1210,13 +1269,8 @@ uintptr_t dynarec64_66(dynarec_rv64_t* dyn, uintptr_t addr, uintptr_t ip, int ni
                     EWBACK;
                     break;
                 case 2:
-                    if (opcode == 0xD1) {
-                        INST_NAME("RCL Ew, 1");
-                        MOV32w(x2, 1);
-                    } else {
-                        INST_NAME("RCL Ew, CL");
-                        ANDI(x2, xRCX, 0x1f);
-                    }
+                    INST_NAME("RCL Ew, CL");
+                    ANDI(x2, xRCX, 0x1f);
                     MESSAGE("LOG_DUMP", "Need optimization\n");
                     READFLAGS(X_CF);
                     SETFLAGS(X_OF | X_CF, SF_SET_DF, NAT_FLAGS_NOFUSION);
@@ -1225,13 +1279,8 @@ uintptr_t dynarec64_66(dynarec_rv64_t* dyn, uintptr_t addr, uintptr_t ip, int ni
                     EWBACK;
                     break;
                 case 3:
-                    if (opcode == 0xD1) {
-                        INST_NAME("RCR Ew, 1");
-                        MOV32w(x2, 1);
-                    } else {
-                        INST_NAME("RCR Ew, CL");
-                        ANDI(x2, xRCX, 0x1f);
-                    }
+                    INST_NAME("RCR Ew, CL");
+                    ANDI(x2, xRCX, 0x1f);
                     MESSAGE("LOG_DUMP", "Need optimization\n");
                     READFLAGS(X_CF);
                     SETFLAGS(X_OF | X_CF, SF_SET_DF, NAT_FLAGS_NOFUSION);
@@ -1240,14 +1289,9 @@ uintptr_t dynarec64_66(dynarec_rv64_t* dyn, uintptr_t addr, uintptr_t ip, int ni
                     EWBACK;
                     break;
                 case 5:
-                    if (opcode == 0xD1) {
-                        INST_NAME("SHR Ew, 1");
-                        MOV32w(x2, 1);
-                    } else {
-                        INST_NAME("SHR Ew, CL");
-                        ANDI(x2, xRCX, 0x1f);
-                        BEQ_NEXT(x2, xZR);
-                    }
+                    INST_NAME("SHR Ew, CL");
+                    ANDI(x2, xRCX, 0x1f);
+                    BEQ_NEXT(x2, xZR);
                     SETFLAGS(X_ALL, SF_SET_PENDING, NAT_FLAGS_FUSION); // some flags are left undefined
                     if (box64_dynarec_safeflags > 1)
                         MAYSETFLAGS();
@@ -1257,14 +1301,9 @@ uintptr_t dynarec64_66(dynarec_rv64_t* dyn, uintptr_t addr, uintptr_t ip, int ni
                     break;
                 case 4:
                 case 6:
-                    if (opcode == 0xD1) {
-                        INST_NAME("SHL Ew, 1");
-                        MOV32w(x2, 1);
-                    } else {
-                        INST_NAME("SHL Ew, CL");
-                        ANDI(x2, xRCX, 0x1f);
-                        BEQ_NEXT(x2, xZR);
-                    }
+                    INST_NAME("SHL Ew, CL");
+                    ANDI(x2, xRCX, 0x1f);
+                    BEQ_NEXT(x2, xZR);
                     SETFLAGS(X_ALL, SF_SET_PENDING, NAT_FLAGS_FUSION); // some flags are left undefined
                     if (box64_dynarec_safeflags > 1)
                         MAYSETFLAGS();
@@ -1273,14 +1312,9 @@ uintptr_t dynarec64_66(dynarec_rv64_t* dyn, uintptr_t addr, uintptr_t ip, int ni
                     EWBACK;
                     break;
                 case 7:
-                    if (opcode == 0xD1) {
-                        INST_NAME("SAR Ew, 1");
-                        MOV32w(x2, 1);
-                    } else {
-                        INST_NAME("SAR Ew, CL");
-                        ANDI(x2, xRCX, 0x1f);
-                        BEQ_NEXT(x2, xZR);
-                    }
+                    INST_NAME("SAR Ew, CL");
+                    ANDI(x2, xRCX, 0x1f);
+                    BEQ_NEXT(x2, xZR);
                     SETFLAGS(X_ALL, SF_SET_PENDING, NAT_FLAGS_FUSION); // some flags are left undefined
                     if (box64_dynarec_safeflags > 1)
                         MAYSETFLAGS();
@@ -1288,8 +1322,6 @@ uintptr_t dynarec64_66(dynarec_rv64_t* dyn, uintptr_t addr, uintptr_t ip, int ni
                     emit_sar16(dyn, ninst, x1, x2, x5, x4, x6);
                     EWBACK;
                     break;
-                default:
-                    DEFAULT;
             }
             break;
 
diff --git a/src/dynarec/rv64/dynarec_rv64_emit_shift.c b/src/dynarec/rv64/dynarec_rv64_emit_shift.c
index 267bf1cf..8daa62dc 100644
--- a/src/dynarec/rv64/dynarec_rv64_emit_shift.c
+++ b/src/dynarec/rv64/dynarec_rv64_emit_shift.c
@@ -1061,6 +1061,35 @@ void emit_ror32(dynarec_rv64_t* dyn, int ninst, rex_t rex, int s1, int s2, int s
     }
 }
 
+// emit ROL16 instruction, from s1, constant c, store result in s1 using s3 and s4 as scratch
+void emit_rol16c(dynarec_rv64_t* dyn, int ninst, int s1, uint32_t c, int s3, int s4)
+{
+    if (!c) return;
+
+    SET_DFNONE();
+
+    SLLI(s3, s1, 48 + c);
+    SRLI(s3, s3, 48);
+    SRLI(s1, s1, 16 - c);
+    OR(s1, s1, s3);
+
+    if (dyn->insts[ninst].nat_flags_fusion) NAT_FLAGS_OPS(s1, xZR);
+
+    IFX (X_CF | X_OF) {
+        ANDI(xFlags, xFlags, ~(1UL << F_CF | 1UL << F_OF2));
+        ANDI(s4, s1, 1 << F_CF);
+        IFX (X_CF) OR(xFlags, xFlags, s4);
+    }
+    IFX (X_OF) {
+        if (c == 1) {
+            SRLI(s3, s1, 15);
+            XOR(s3, s3, s4);
+            SLLI(s3, s3, F_OF2);
+            OR(xFlags, xFlags, s3);
+        }
+    }
+}
+
 // emit ROL32 instruction, from s1 , constant c, store result in s1 using s3 and s4 as scratch
 void emit_rol32c(dynarec_rv64_t* dyn, int ninst, rex_t rex, int s1, uint32_t c, int s3, int s4)
 {
diff --git a/src/dynarec/rv64/dynarec_rv64_helper.h b/src/dynarec/rv64/dynarec_rv64_helper.h
index 40146b77..a389f7b1 100644
--- a/src/dynarec/rv64/dynarec_rv64_helper.h
+++ b/src/dynarec/rv64/dynarec_rv64_helper.h
@@ -1301,6 +1301,7 @@ void* rv64_next(void);
 #define emit_shr32          STEPNAME(emit_shr32)
 #define emit_shr32c         STEPNAME(emit_shr32c)
 #define emit_sar32c         STEPNAME(emit_sar32c)
+#define emit_rol16c         STEPNAME(emit_rol16c)
 #define emit_rol32          STEPNAME(emit_rol32)
 #define emit_ror32          STEPNAME(emit_ror32)
 #define emit_rol32c         STEPNAME(emit_rol32c)
@@ -1470,6 +1471,7 @@ void emit_shr32c(dynarec_rv64_t* dyn, int ninst, rex_t rex, int s1, uint32_t c,
 void emit_sar32c(dynarec_rv64_t* dyn, int ninst, rex_t rex, int s1, uint32_t c, int s3, int s4);
 void emit_rol32(dynarec_rv64_t* dyn, int ninst, rex_t rex, int s1, int s2, int s3, int s4);
 void emit_ror32(dynarec_rv64_t* dyn, int ninst, rex_t rex, int s1, int s2, int s3, int s4);
+void emit_rol16c(dynarec_rv64_t* dyn, int ninst, int s1, uint32_t c, int s3, int s4);
 void emit_rol32c(dynarec_rv64_t* dyn, int ninst, rex_t rex, int s1, uint32_t c, int s3, int s4);
 void emit_ror32c(dynarec_rv64_t* dyn, int ninst, rex_t rex, int s1, uint32_t c, int s3, int s4);
 void emit_shrd32c(dynarec_rv64_t* dyn, int ninst, rex_t rex, int s1, int s2, uint32_t c, int s3, int s4, int s5);