[RV64_DYNAREC] Optimized 16bit constant RCL/RCR opcodes (#2270)

3 files changed, 120 insertions, 26 deletions

diff --git a/src/dynarec/rv64/dynarec_rv64_66.c b/src/dynarec/rv64/dynarec_rv64_66.c
index dfb4a4b5..3aac15f0 100644
--- a/src/dynarec/rv64/dynarec_rv64_66.c
+++ b/src/dynarec/rv64/dynarec_rv64_66.c
@@ -1091,25 +1091,33 @@ uintptr_t dynarec64_66(dynarec_rv64_t* dyn, uintptr_t addr, uintptr_t ip, int ni
                     break;
                 case 2:
                     INST_NAME("RCL Ew, Ib");
-                    MESSAGE(LOG_DUMP, "Need Optimization\n");
-                    READFLAGS(X_CF);
-                    SETFLAGS(X_OF | X_CF, SF_SET_DF, NAT_FLAGS_NOFUSION);
-                    GETEW(x1, 1);
-                    u8 = F8;
-                    MOV32w(x2, u8);
-                    CALL_(rcl16, x1, x3, x1, x2);
-                    EWBACK;
+                    if (geted_ib(dyn, addr, ninst, nextop) & 31) {
+                        READFLAGS(X_CF);
+                        // removed PENDING on purpose
+                        SETFLAGS(X_OF | X_CF, SF_SUBSET, NAT_FLAGS_FUSION);
+                        GETEW(x1, 1);
+                        u8 = (F8) & 0x1f;
+                        emit_rcl16c(dyn, ninst, ed, u8, x4, x5);
+                        EWBACK;
+                    } else {
+                        FAKEED;
+                        F8;
+                    }
                     break;
                 case 3:
                     INST_NAME("RCR Ew, Ib");
-                    MESSAGE(LOG_DUMP, "Need Optimization\n");
-                    READFLAGS(X_CF);
-                    SETFLAGS(X_OF | X_CF, SF_SET_DF, NAT_FLAGS_NOFUSION);
-                    GETEW(x1, 1);
-                    u8 = F8;
-                    MOV32w(x2, u8);
-                    CALL_(rcr16, x1, x3, x1, x2);
-                    EWBACK;
+                    if (geted_ib(dyn, addr, ninst, nextop) & 31) {
+                        READFLAGS(X_CF);
+                        // removed PENDING on purpose
+                        SETFLAGS(X_OF | X_CF, SF_SUBSET, NAT_FLAGS_FUSION);
+                        GETEW(x1, 1);
+                        u8 = (F8) & 0x1f;
+                        emit_rcr16c(dyn, ninst, ed, u8, x4, x5);
+                        EWBACK;
+                    } else {
+                        FAKEED;
+                        F8;
+                    }
                     break;
                 case 4:
                 case 6:
@@ -1207,22 +1215,20 @@ uintptr_t dynarec64_66(dynarec_rv64_t* dyn, uintptr_t addr, uintptr_t ip, int ni
                     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);
+                    // removed PENDING on purpose
+                    SETFLAGS(X_OF | X_CF, SF_SUBSET, NAT_FLAGS_FUSION);
+                    GETEW(x1, 0);
+                    emit_rcl16c(dyn, ninst, x1, 1, x5, x4);
                     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);
+                    // removed PENDING on purpose
+                    SETFLAGS(X_OF | X_CF, SF_SUBSET, NAT_FLAGS_FUSION);
+                    GETEW(x1, 0);
+                    emit_rcr16c(dyn, ninst, x1, 1, x5, x4);
                     EWBACK;
                     break;
                 case 5:
diff --git a/src/dynarec/rv64/dynarec_rv64_emit_shift.c b/src/dynarec/rv64/dynarec_rv64_emit_shift.c
index 54d4782e..91c9bb6c 100644
--- a/src/dynarec/rv64/dynarec_rv64_emit_shift.c
+++ b/src/dynarec/rv64/dynarec_rv64_emit_shift.c
@@ -1635,3 +1635,87 @@ void emit_shld16c(dynarec_rv64_t* dyn, int ninst, rex_t rex, int s1, int s2, uin
         emit_pf(dyn, ninst, s1, s3, s4);
     }
 }
+
+// emit RCL16 instruction, from s1 , constant c, store result in s1 using s3 and s4 as scratch
+void emit_rcl16c(dynarec_rv64_t* dyn, int ninst, int s1, uint32_t c, int s3, int s4)
+{
+    if (!(c % 17)) return;
+
+    SET_DFNONE();
+
+    c %= 17;
+
+    ANDI(s3, xFlags, 1 << F_CF);
+    SLLI(s3, s3, 16);
+    OR(s1, s1, s3); // insert CF to bit 16
+
+    SLLI(s4, s1, 47 + c);
+    SLLI(s3, s4, 1);
+    SRLI(s3, s3, 48);
+    SRLI(s1, s1, 17 - 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));
+        SRLI(s4, s4, 63);
+        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 RCR16 instruction, from s1 , constant c, store result in s1 using s3 and s4 as scratch
+void emit_rcr16c(dynarec_rv64_t* dyn, int ninst, int s1, uint32_t c, int s3, int s4)
+{
+    MAYUSE(s1);
+    MAYUSE(s3);
+    MAYUSE(s4);
+
+    if (!(c % 17)) return;
+
+    SET_DFNONE();
+
+    c %= 17;
+
+
+    ANDI(s3, xFlags, 1 << F_CF);
+    SLLI(s3, s3, 16);
+    OR(s1, s1, s3); // insert CF to bit 16
+
+    SRLI(s3, s1, c);
+    SLLI(s4, s1, 63 - c);
+    SLLI(s1, s4, s1);
+    SRLI(s1, s1, 48);
+    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));
+    }
+
+    IFX (X_CF) {
+        SRLI(s4, s4, 63);
+        OR(xFlags, xFlags, s4);
+    }
+
+    IFX (X_OF) {
+        if (c == 1) {
+            SRLI(s3, s1, 14);
+            SRLI(s4, s3, 1);
+            XOR(s3, s3, s4);
+            ANDI(s3, s3, 1);
+            SLLI(s3, s3, F_OF2);
+            OR(xFlags, xFlags, s3);
+        }
+    }
+}
diff --git a/src/dynarec/rv64/dynarec_rv64_helper.h b/src/dynarec/rv64/dynarec_rv64_helper.h
index b58caf0e..40ef0717 100644
--- a/src/dynarec/rv64/dynarec_rv64_helper.h
+++ b/src/dynarec/rv64/dynarec_rv64_helper.h
@@ -1313,6 +1313,8 @@ void* rv64_next(void);
 #define emit_shld32         STEPNAME(emit_shld32)
 #define emit_shld16c        STEPNAME(emit_shld16c)
 #define emit_shrd16c        STEPNAME(emit_shrd16c)
+#define emit_rcl16c         STEPNAME(emit_rcl16c)
+#define emit_rcr16c         STEPNAME(emit_rcr16c)
 
 #define emit_pf STEPNAME(emit_pf)
 
@@ -1482,6 +1484,8 @@ void emit_shrd32(dynarec_rv64_t* dyn, int ninst, rex_t rex, int s1, int s2, int
 void emit_shld32(dynarec_rv64_t* dyn, int ninst, rex_t rex, int s1, int s2, int s5, int s3, int s4, int s6);
 void emit_shrd16c(dynarec_rv64_t* dyn, int ninst, rex_t rex, int s1, int s2, uint32_t c, int s3, int s4, int s5);
 void emit_shld16c(dynarec_rv64_t* dyn, int ninst, rex_t rex, int s1, int s2, uint32_t c, int s3, int s4, int s5);
+void emit_rcl16c(dynarec_rv64_t* dyn, int ninst, int s1, uint32_t c, int s3, int s4);
+void emit_rcr16c(dynarec_rv64_t* dyn, int ninst, int s1, uint32_t c, int s3, int s4);
 
 void emit_pf(dynarec_rv64_t* dyn, int ninst, int s1, int s3, int s4);