[DYNAREC] Small optim on shl/shr 16bits, and marked all opcode that need optimization in dynarec dumps

7 files changed, 67 insertions, 2 deletions

diff --git a/src/dynarec/arm64_emitter.h b/src/dynarec/arm64_emitter.h
index 3114d214..d2e7e3fa 100755
--- a/src/dynarec/arm64_emitter.h
+++ b/src/dynarec/arm64_emitter.h
@@ -434,12 +434,15 @@
 #define ORRx_REG_LSL(Rd, Rn, Rm, lsl)   EMIT(LOGIC_REG_gen(1, 0b01, 0b00, 0, Rm, lsl, Rn, Rd))
 #define ORRw_REG_LSL(Rd, Rn, Rm, lsl)   EMIT(LOGIC_REG_gen(0, 0b01, 0b00, 0, Rm, lsl, Rn, Rd))
 #define ORRxw_REG_LSL(Rd, Rn, Rm, lsl)  EMIT(LOGIC_REG_gen(rex.w, 0b01, 0b00, 0, Rm, lsl, Rn, Rd))
+#define ORRx_REG_LSR(Rd, Rn, Rm, lsr)   EMIT(LOGIC_REG_gen(1, 0b01, 0b01, 0, Rm, lsr, Rn, Rd))
+#define ORRw_REG_LSR(Rd, Rn, Rm, lsr)   EMIT(LOGIC_REG_gen(0, 0b01, 0b01, 0, Rm, lsr, Rn, Rd))
 #define ORRxw_REG_LSR(Rd, Rn, Rm, lsr)  EMIT(LOGIC_REG_gen(rex.w, 0b01, 0b01, 0, Rm, lsr, Rn, Rd))
 #define ORRxw_REG(Rd, Rn, Rm)           EMIT(LOGIC_REG_gen(rex.w, 0b01, 0b00, 0, Rm, 0, Rn, Rd))
 #define ORRw_REG(Rd, Rn, Rm)            EMIT(LOGIC_REG_gen(0, 0b01, 0b00, 0, Rm, 0, Rn, Rd))
 #define ORNx_REG(Rd, Rn, Rm)            EMIT(LOGIC_REG_gen(1, 0b01, 0b00, 1, Rm, 0, Rn, Rd))
 #define ORNw_REG(Rd, Rn, Rm)            EMIT(LOGIC_REG_gen(0, 0b01, 0b00, 1, Rm, 0, Rn, Rd))
 #define ORNxw_REG(Rd, Rn, Rm)           EMIT(LOGIC_REG_gen(rex.w, 0b01, 0b00, 1, Rm, 0, Rn, Rd))
+#define ORNx_REG_LSL(Rd, Rn, Rm, lsl)   EMIT(LOGIC_REG_gen(1, 0b01, 0b00, 1, Rm, lsl, Rn, Rd))
 #define EORx_REG(Rd, Rn, Rm)            EMIT(LOGIC_REG_gen(1, 0b10, 0b00, 0, Rm, 0, Rn, Rd))
 #define EORw_REG(Rd, Rn, Rm)            EMIT(LOGIC_REG_gen(0, 0b10, 0b00, 0, Rm, 0, Rn, Rd))
 #define EORxw_REG(Rd, Rn, Rm)           EMIT(LOGIC_REG_gen(rex.w, 0b10, 0b00, 0, Rm, 0, Rn, Rd))
@@ -452,6 +455,12 @@
 #define MOVx_REG(Rd, Rm)                ORRx_REG(Rd, xZR, Rm)
 #define MOVw_REG(Rd, Rm)                ORRw_REG(Rd, xZR, Rm)
 #define MOVxw_REG(Rd, Rm)               ORRxw_REG(Rd, xZR, Rm)
+#define LSLw_IMM(Rd, Rm, lsl)           ORRw_REG_LSL(Rd, xZR, Rm, lsl)
+#define LSLx_IMM(Rd, Rm, lsl)           ORRx_REG_LSL(Rd, xZR, Rm, lsl)
+#define LSLxw_IMM(Rd, Rm, lsl)          ORRxw_REG_LSL(Rd, xZR, Rm, lsl)
+#define LSRw_IMM(Rd, Rm, lsr)           ORRw_REG_LSR(Rd, xZR, Rm, lsr)
+#define LSRx_IMM(Rd, Rm, lsr)           ORRx_REG_LSR(Rd, xZR, Rm, lsr)
+#define LSRxw_IMM(Rd, Rm, lsr)          ORRxw_REG_LSR(Rd, xZR, Rm, lsr)
 #define MVNx_REG(Rd, Rm)                ORNx_REG(Rd, xZR, Rm)
 #define MVNx_REG_LSL(Rd, Rm, lsl)       ORNx_REG_LSL(Rd, xZR, Rm, lsl)
 #define MVNw_REG(Rd, Rm)                ORNw_REG(Rd, xZR, Rm)
diff --git a/src/dynarec/dynarec_arm64_00.c b/src/dynarec/dynarec_arm64_00.c
index 3a136b12..df621729 100755
--- a/src/dynarec/dynarec_arm64_00.c
+++ b/src/dynarec/dynarec_arm64_00.c
@@ -1312,6 +1312,7 @@ uintptr_t dynarec64_00(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
             switch((nextop>>3)&7) {
                 case 0:
                     INST_NAME("ROL Eb, Ib");
+                    MESSAGE(LOG_DUMP, "Need Optimization\n");
                     SETFLAGS(X_OF|X_CF, SF_SET);
                     GETEB(x1, 1);
                     u8 = F8;
@@ -1321,6 +1322,7 @@ uintptr_t dynarec64_00(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
                     break;
                 case 1:
                     INST_NAME("ROR Eb, Ib");
+                    MESSAGE(LOG_DUMP, "Need Optimization\n");
                     SETFLAGS(X_OF|X_CF, SF_SET);
                     GETEB(x1, 1);
                     u8 = F8;
@@ -1330,6 +1332,7 @@ uintptr_t dynarec64_00(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
                     break;
                 case 2:
                     INST_NAME("RCL Eb, Ib");
+                    MESSAGE(LOG_DUMP, "Need Optimization\n");
                     READFLAGS(X_CF);
                     SETFLAGS(X_OF|X_CF, SF_SET);
                     GETEB(x1, 1);
@@ -1340,6 +1343,7 @@ uintptr_t dynarec64_00(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
                     break;
                 case 3:
                     INST_NAME("RCR Eb, Ib");
+                    MESSAGE(LOG_DUMP, "Need Optimization\n");
                     READFLAGS(X_CF);
                     SETFLAGS(X_OF|X_CF, SF_SET);
                     GETEB(x1, 1);
@@ -1418,6 +1422,7 @@ uintptr_t dynarec64_00(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
                     break;
                 case 2:
                     INST_NAME("RCL Ed, Ib");
+                    MESSAGE(LOG_DUMP, "Need Optimization\n");
                     READFLAGS(X_CF);
                     SETFLAGS(X_OF|X_CF, SF_SET);
                     GETEDW(x4, x1, 1);
@@ -1428,6 +1433,7 @@ uintptr_t dynarec64_00(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
                     break;
                 case 3:
                     INST_NAME("RCR Ed, Ib");
+                    MESSAGE(LOG_DUMP, "Need Optimization\n");
                     READFLAGS(X_CF);
                     SETFLAGS(X_OF|X_CF, SF_SET);
                     GETEDW(x4, x1, 1);
@@ -1593,6 +1599,7 @@ uintptr_t dynarec64_00(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
                         INST_NAME("ROL Eb, CL");
                         ANDSw_mask(x2, xRCX, 0, 0b00100);
                     }
+                    MESSAGE(LOG_DUMP, "Need Optimization\n");
                     SETFLAGS(X_OF|X_CF, SF_SET);
                     GETEB(x1, 0);
                     CALL_(rol8, x1, x3);
@@ -1606,6 +1613,7 @@ uintptr_t dynarec64_00(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
                         INST_NAME("ROR Eb, CL");
                         ANDSw_mask(x2, xRCX, 0, 0b00100);
                     }
+                    MESSAGE(LOG_DUMP, "Need Optimization\n");
                     SETFLAGS(X_OF|X_CF, SF_SET);
                     GETEB(x1, 0);
                     CALL_(ror8, x1, x3);
@@ -1613,6 +1621,7 @@ uintptr_t dynarec64_00(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
                     break;
                 case 2:
                     if(opcode==0xD0) {INST_NAME("RCL Eb, 1");} else {INST_NAME("RCL Eb, CL");}
+                    MESSAGE(LOG_DUMP, "Need Optimization\n");
                     READFLAGS(X_CF);
                     SETFLAGS(X_OF|X_CF, SF_SET);
                     if(opcode==0xD0) {MOV32w(x2, 1);} else {ANDSw_mask(x2, xRCX, 0, 0b00100);}
@@ -1622,6 +1631,7 @@ uintptr_t dynarec64_00(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
                     break;
                 case 3:
                     if(opcode==0xD0) {INST_NAME("RCR Eb, 1");} else {INST_NAME("RCR Eb, CL");}
+                    MESSAGE(LOG_DUMP, "Need Optimization\n");
                     READFLAGS(X_CF);
                     SETFLAGS(X_OF|X_CF, SF_SET);
                     if(opcode==0xD0) {MOV32w(x2, 1);} else {ANDSw_mask(x2, xRCX, 0, 0b00100);}
@@ -1699,6 +1709,7 @@ uintptr_t dynarec64_00(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
                     break;
                 case 2:
                     INST_NAME("RCL Ed, 1");
+                    MESSAGE(LOG_DUMP, "Need Optimization\n");
                     READFLAGS(X_CF);
                     SETFLAGS(X_OF|X_CF, SF_SET);
                     MOV32w(x2, 1);
@@ -1708,6 +1719,7 @@ uintptr_t dynarec64_00(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
                     break;
                 case 3:
                     INST_NAME("RCR Ed, 1");
+                    MESSAGE(LOG_DUMP, "Need Optimization\n");
                     READFLAGS(X_CF);
                     SETFLAGS(X_OF|X_CF, SF_SET);
                     MOV32w(x2, 1);
@@ -1795,6 +1807,7 @@ uintptr_t dynarec64_00(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
                     break;
                 case 2:
                     INST_NAME("RCL Ed, CL");
+                    MESSAGE(LOG_DUMP, "Need Optimization\n");
                     READFLAGS(X_CF);
                     SETFLAGS(X_OF|X_CF, SF_SET);
                     if(rex.w) {
@@ -1810,6 +1823,7 @@ uintptr_t dynarec64_00(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
                     break;
                 case 3:
                     INST_NAME("RCR Ed, CL");
+                    MESSAGE(LOG_DUMP, "Need Optimization\n");
                     READFLAGS(X_CF);
                     SETFLAGS(X_OF|X_CF, SF_SET);
                     if(rex.w) {
@@ -2096,12 +2110,14 @@ uintptr_t dynarec64_00(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
                     break;
                 case 6:
                     INST_NAME("DIV Eb");
+                    MESSAGE(LOG_DUMP, "Need Optimization\n");
                     SETFLAGS(X_ALL, SF_SET);
                     GETEB(x1, 0);
                     CALL(div8, -1);
                     break;
                 case 7:
                     INST_NAME("IDIV Eb");
+                    MESSAGE(LOG_DUMP, "Need Optimization\n");
                     SETFLAGS(X_ALL, SF_SET);
                     GETEB(x1, 0);
                     CALL(idiv8, -1);
diff --git a/src/dynarec/dynarec_arm64_0f.c b/src/dynarec/dynarec_arm64_0f.c
index 26ac8b62..8c648d6e 100755
--- a/src/dynarec/dynarec_arm64_0f.c
+++ b/src/dynarec/dynarec_arm64_0f.c
@@ -313,6 +313,7 @@ uintptr_t dynarec64_0F(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
 

         case 0x31:

             INST_NAME("RDTSC");

+            MESSAGE(LOG_DUMP, "Need Optimization\n");

             CALL(ReadTSC, xRAX);   // will return the u64 in xEAX

             LSRx(xRDX, xRAX, 32);

             MOVw_REG(xRAX, xRAX);   // wipe upper part

@@ -1125,6 +1126,7 @@ uintptr_t dynarec64_0F(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
                 switch((nextop>>3)&7) {

                     case 0:

                         INST_NAME("FXSAVE Ed");

+                        MESSAGE(LOG_DUMP, "Need Optimization\n");

                         fpu_purgecache(dyn, ninst, x1, x2, x3);

                         if(MODREG) {

                             DEFAULT;

@@ -1136,6 +1138,7 @@ uintptr_t dynarec64_0F(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
                         break;

                     case 1:

                         INST_NAME("FXRSTOR Ed");

+                        MESSAGE(LOG_DUMP, "Need Optimization\n");

                         fpu_purgecache(dyn, ninst, x1, x2, x3);

                         if(MODREG) {

                             DEFAULT;

diff --git a/src/dynarec/dynarec_arm64_66.c b/src/dynarec/dynarec_arm64_66.c
index e5429363..a8debcb8 100755
--- a/src/dynarec/dynarec_arm64_66.c
+++ b/src/dynarec/dynarec_arm64_66.c
@@ -552,6 +552,7 @@ uintptr_t dynarec64_66(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
             switch((nextop>>3)&7) {

                 case 0:

                     INST_NAME("ROL Ew, Ib");

+                    MESSAGE(LOG_DUMP, "Need Optimization\n");

                     SETFLAGS(X_OF|X_CF, SF_SET);

                     GETEW(x1, 1);

                     u8 = F8;

@@ -561,6 +562,7 @@ uintptr_t dynarec64_66(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
                     break;

                 case 1:

                     INST_NAME("ROR Ew, Ib");

+                    MESSAGE(LOG_DUMP, "Need Optimization\n");

                     SETFLAGS(X_OF|X_CF, SF_SET);

                     GETEW(x1, 1);

                     u8 = F8;

@@ -570,6 +572,7 @@ uintptr_t dynarec64_66(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
                     break;

                 case 2:

                     INST_NAME("RCL Ew, Ib");

+                    MESSAGE(LOG_DUMP, "Need Optimization\n");

                     READFLAGS(X_CF);

                     SETFLAGS(X_OF|X_CF, SF_SET);

                     GETEW(x1, 1);

@@ -580,6 +583,7 @@ uintptr_t dynarec64_66(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
                     break;

                 case 3:

                     INST_NAME("RCR Ew, Ib");

+                    MESSAGE(LOG_DUMP, "Need Optimization\n");

                     READFLAGS(X_CF);

                     SETFLAGS(X_OF|X_CF, SF_SET);

                     GETEW(x1, 1);

@@ -591,24 +595,26 @@ uintptr_t dynarec64_66(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
                 case 4:

                 case 6:

                     INST_NAME("SHL Ew, Ib");

+                    UFLAG_IF {MESSAGE(LOG_DUMP, "Need Optimization for flags\n");}

                     SETFLAGS(X_ALL, SF_PENDING);

                     GETEW(x1, 1);

                     u8 = F8;

                     MOV32w(x2, (u8&0x1f));

                     UFLAG_OP12(ed, x2)

-                    LSLw_REG(ed, ed, x2);

+                    LSLw_IMM(ed, ed, u8&0x1f);

                     EWBACK;

                     UFLAG_RES(ed);

                     UFLAG_DF(x3, d_shl16);

                     break;

                 case 5:

                     INST_NAME("SHR Ed, Ib");

+                    UFLAG_IF {MESSAGE(LOG_DUMP, "Need Optimization for flags\n");}

                     SETFLAGS(X_ALL, SF_PENDING);

                     GETEW(x1, 1);

                     u8 = F8;

                     MOV32w(x2, (u8&0x1f));

                     UFLAG_OP12(ed, x2)

-                    LSRw_REG(ed, ed, x2);

+                    LSRw_IMM(ed, ed, u8&0x1f);

                     EWBACK;

                     UFLAG_RES(ed);

                     UFLAG_DF(x3, d_shr16);

@@ -616,6 +622,7 @@ uintptr_t dynarec64_66(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
                 case 7:

                     INST_NAME("SAR Ed, Ib");

                     SETFLAGS(X_ALL, SF_PENDING);

+                    UFLAG_IF {MESSAGE(LOG_DUMP, "Need Optimization for flags\n");}

                     GETSEW(x1, 0);

                     u8 = F8;

                     MOV32w(x2, (u8&0x1f));

@@ -656,6 +663,7 @@ uintptr_t dynarec64_66(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
                         INST_NAME("ROL Ew, CL");

                         ANDSw_mask(x2, xRCX, 0, 0b00100);

                     }

+                    MESSAGE(LOG_DUMP, "Need Optimization\n");

                     SETFLAGS(X_OF|X_CF, SF_SET);

                     GETEW(x1, 0);

                     CALL_(rol16, x1, x3);

@@ -669,6 +677,7 @@ uintptr_t dynarec64_66(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
                         INST_NAME("ROR Ew, CL");

                         ANDSw_mask(x2, xRCX, 0, 0b00100);

                     }

+                    MESSAGE(LOG_DUMP, "Need Optimization\n");

                     SETFLAGS(X_OF|X_CF, SF_SET);

                     GETEW(x1, 0);

                     CALL_(ror16, x1, x3);

@@ -676,6 +685,7 @@ uintptr_t dynarec64_66(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
                     break;

                 case 2:

                     if(opcode==0xD1) {INST_NAME("RCL Ew, 1"); } else { INST_NAME("RCL Ew, CL");}

+                    MESSAGE(LOG_DUMP, "Need Optimization\n");

                     READFLAGS(X_CF);

                     SETFLAGS(X_OF|X_CF, SF_SET);

                     if(opcode==0xD1) {MOV32w(x2, 1);} else {ANDSw_mask(x2, xRCX, 0, 0b00100);}

@@ -685,6 +695,7 @@ uintptr_t dynarec64_66(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
                     break;

                 case 3:

                     if(opcode==0xD1) {INST_NAME("RCR Ew, 1");} else {INST_NAME("RCR Ew, CL");}

+                    MESSAGE(LOG_DUMP, "Need Optimization\n");

                     READFLAGS(X_CF);

                     SETFLAGS(X_OF|X_CF, SF_SET);

                     if(opcode==0xD1) {MOV32w(x2, 1);} else {ANDSw_mask(x2, xRCX, 0, 0b00100);}

@@ -701,6 +712,7 @@ uintptr_t dynarec64_66(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
                         INST_NAME("SHL Ew, CL");

                         ANDSw_mask(x4, xRCX, 0, 0b00100);

                     }

+                    UFLAG_IF {MESSAGE(LOG_DUMP, "Need Optimization for flags\n");}

                     SETFLAGS(X_ALL, SF_PENDING);

                     GETEW(x1, 0);

                     UFLAG_OP12(ed, x4)

@@ -717,6 +729,7 @@ uintptr_t dynarec64_66(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
                         INST_NAME("SHR Ew, CL");

                         ANDSw_mask(x4, xRCX, 0, 0b00100);

                     }

+                    UFLAG_IF {MESSAGE(LOG_DUMP, "Need Optimization for flags\n");}

                     SETFLAGS(X_ALL, SF_PENDING);

                     GETEW(x1, 0);

                     UFLAG_OP12(ed, x4)

@@ -733,6 +746,7 @@ uintptr_t dynarec64_66(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
                         INST_NAME("SAR Ew, CL");

                         ANDSw_mask(x4, xRCX, 0, 0b00100);

                     }

+                    UFLAG_IF {MESSAGE(LOG_DUMP, "Need Optimization for flags\n");}

                     SETFLAGS(X_ALL, SF_PENDING);

                     GETSEW(x1, 0);

                     UFLAG_OP12(ed, x4)

@@ -793,12 +807,14 @@ uintptr_t dynarec64_66(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
                     break;

                 case 6:

                     INST_NAME("DIV Ew");

+                    MESSAGE(LOG_DUMP, "Need Optimization\n");

                     SETFLAGS(X_ALL, SF_SET);

                     GETEW(x1, 0);

                     CALL(div16, -1);

                     break;

                 case 7:

                     INST_NAME("IDIV Ew");

+                    MESSAGE(LOG_DUMP, "Need Optimization\n");

                     SETFLAGS(X_ALL, SF_SET);

                     GETEW(x1, 0);

                     CALL(idiv16, -1);

diff --git a/src/dynarec/dynarec_arm64_660f.c b/src/dynarec/dynarec_arm64_660f.c
index 39bf8229..577c3d2f 100755
--- a/src/dynarec/dynarec_arm64_660f.c
+++ b/src/dynarec/dynarec_arm64_660f.c
@@ -967,6 +967,7 @@ uintptr_t dynarec64_660F(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int n
                 INST_NAME("SHLD Ew, Gw, CL");

                 UXTBw(x3, xRCX);

             }

+            MESSAGE(LOG_DUMP, "Need Optimization\n");

             SETFLAGS(X_ALL, SF_SET);

             GETEWW(x4, x1, (opcode==0xA4)?1:0);

             GETGW(x2);

@@ -1015,6 +1016,7 @@ uintptr_t dynarec64_660F(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int n
                 INST_NAME("SHRD Ew, Gw, CL");

                 UXTBw(x3, xRCX);

             }

+            MESSAGE(LOG_DUMP, "Need Optimization\n");

             SETFLAGS(X_ALL, SF_SET);

             GETEWW(x4, x1, (opcode==0xAC)?1:0);

             GETGW(x2);

diff --git a/src/dynarec/dynarec_arm64_d9.c b/src/dynarec/dynarec_arm64_d9.c
index f152860a..8089b68d 100644
--- a/src/dynarec/dynarec_arm64_d9.c
+++ b/src/dynarec/dynarec_arm64_d9.c
@@ -95,6 +95,7 @@ uintptr_t dynarec64_D9(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
             break;
         case 0xE5:
             INST_NAME("FXAM");
+            MESSAGE(LOG_DUMP, "Need Optimization\n");
             x87_refresh(dyn, ninst, x1, x2, 0);
             CALL(fpu_fxam, -1);  // should be possible inline, but is it worth it?
             break;
@@ -143,6 +144,7 @@ uintptr_t dynarec64_D9(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
 
         case 0xFC:
             INST_NAME("FRNDINT");
+            MESSAGE(LOG_DUMP, "Need Optimization\n");
             // use C helper for now, nothing staightforward is available
             x87_forget(dyn, ninst, x1, x2, 0);
             CALL(arm_frndint, -1);
@@ -160,11 +162,13 @@ uintptr_t dynarec64_D9(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
             break;
         case 0xF0:
             INST_NAME("F2XM1");
+            MESSAGE(LOG_DUMP, "Need Optimization\n");
             x87_forget(dyn, ninst, x1, x2, 0);
             CALL(arm_f2xm1, -1);
             break;
         case 0xF1:
             INST_NAME("FYL2X");
+            MESSAGE(LOG_DUMP, "Need Optimization\n");
             x87_forget(dyn, ninst, x1, x2, 0);
             x87_forget(dyn, ninst, x1, x2, 1);
             CALL(arm_fyl2x, -1);
@@ -172,6 +176,7 @@ uintptr_t dynarec64_D9(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
             break;
         case 0xF2:
             INST_NAME("FTAN");
+            MESSAGE(LOG_DUMP, "Need Optimization\n");
             x87_forget(dyn, ninst, x1, x2, 0);
             CALL(arm_ftan, -1);
             v1 = x87_do_push(dyn, ninst);
@@ -179,6 +184,7 @@ uintptr_t dynarec64_D9(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
             break;
         case 0xF3:
             INST_NAME("FPATAN");
+            MESSAGE(LOG_DUMP, "Need Optimization\n");
             x87_forget(dyn, ninst, x1, x2, 0);
             x87_forget(dyn, ninst, x1, x2, 1);
             CALL(arm_fpatan, -1);
@@ -186,12 +192,14 @@ uintptr_t dynarec64_D9(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
             break;
         case 0xF4:
             INST_NAME("FXTRACT");
+            MESSAGE(LOG_DUMP, "Need Optimization\n");
             x87_do_push_empty(dyn, ninst, 0);
             x87_forget(dyn, ninst, x1, x2, 1);
             CALL(arm_fxtract, -1);
             break;
         case 0xF5:
             INST_NAME("FPREM1");
+            MESSAGE(LOG_DUMP, "Need Optimization\n");
             x87_forget(dyn, ninst, x1, x2, 0);
             x87_forget(dyn, ninst, x1, x2, 1);
             CALL(arm_fprem1, -1);
@@ -214,12 +222,14 @@ uintptr_t dynarec64_D9(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
             break;
         case 0xF8:
             INST_NAME("FPREM");
+            MESSAGE(LOG_DUMP, "Need Optimization\n");
             x87_forget(dyn, ninst, x1, x2, 0);
             x87_forget(dyn, ninst, x1, x2, 1);
             CALL(arm_fprem, -1);
             break;
         case 0xF9:
             INST_NAME("FYL2XP1");
+            MESSAGE(LOG_DUMP, "Need Optimization\n");
             x87_forget(dyn, ninst, x1, x2, 0);
             x87_forget(dyn, ninst, x1, x2, 1);
             CALL(arm_fyl2xp1, -1);
@@ -227,23 +237,27 @@ uintptr_t dynarec64_D9(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
             break;
         case 0xFB:
             INST_NAME("FSINCOS");
+            MESSAGE(LOG_DUMP, "Need Optimization\n");
             x87_do_push_empty(dyn, ninst, 0);
             x87_forget(dyn, ninst, x1, x2, 1);
             CALL(arm_fsincos, -1);
             break;
         case 0xFD:
             INST_NAME("FSCALE");
+            MESSAGE(LOG_DUMP, "Need Optimization\n");
             x87_forget(dyn, ninst, x1, x2, 0);
             x87_forget(dyn, ninst, x1, x2, 1);
             CALL(arm_fscale, -1);
             break;
         case 0xFE:
             INST_NAME("FSIN");
+            MESSAGE(LOG_DUMP, "Need Optimization\n");
             x87_forget(dyn, ninst, x1, x2, 0);
             CALL(arm_fsin, -1);
             break;
         case 0xFF:
             INST_NAME("FCOS");
+            MESSAGE(LOG_DUMP, "Need Optimization\n");
             x87_forget(dyn, ninst, x1, x2, 0);
             CALL(arm_fcos, -1);
             break;
@@ -299,6 +313,7 @@ uintptr_t dynarec64_D9(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
                     break;
                 case 4:
                     INST_NAME("FLDENV Ed");
+                    MESSAGE(LOG_DUMP, "Need Optimization\n");
                     fpu_purgecache(dyn, ninst, x1, x2, x3); // maybe only x87, not SSE?
                     addr = geted(dyn, addr, ninst, nextop, &ed, x1, &fixedaddress, 0, 0, rex, 0, 0);
                     if(ed!=x1) {
@@ -316,6 +331,7 @@ uintptr_t dynarec64_D9(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
                     break;
                 case 6:
                     INST_NAME("FNSTENV Ed");
+                    MESSAGE(LOG_DUMP, "Need Optimization\n");
                     fpu_purgecache(dyn, ninst, x1, x2, x3); // maybe only x87, not SSE?
                     addr = geted(dyn, addr, ninst, nextop, &ed, x1, &fixedaddress, 0, 0, rex, 0, 0);
                     if(ed!=x1) {
diff --git a/src/dynarec/dynarec_arm64_dd.c b/src/dynarec/dynarec_arm64_dd.c
index 887cd3fc..4068af44 100644
--- a/src/dynarec/dynarec_arm64_dd.c
+++ b/src/dynarec/dynarec_arm64_dd.c
@@ -48,6 +48,7 @@ uintptr_t dynarec64_DD(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
         case 0xC6:
         case 0xC7:
             INST_NAME("FFREE STx");
+            MESSAGE(LOG_DUMP, "Need Optimization\n");
             x87_purgecache(dyn, ninst, x1, x2, x3);
             MOV32w(x1, nextop-0xC0);
             CALL(fpu_do_free, -1);
@@ -173,6 +174,7 @@ uintptr_t dynarec64_DD(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
                     break;
                 case 4: 
                     INST_NAME("FRSTOR m108byte");
+                    MESSAGE(LOG_DUMP, "Need Optimization\n");
                     fpu_purgecache(dyn, ninst, x1, x2, x3);
                     addr = geted(dyn, addr, ninst, nextop, &ed, x1, &fixedaddress, 0, 0, rex, 0, 0);
                     if(ed!=x1) {MOVx_REG(x1, ed);}
@@ -180,6 +182,7 @@ uintptr_t dynarec64_DD(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
                     break;
                 case 6: 
                     INST_NAME("FSAVE m108byte");
+                    MESSAGE(LOG_DUMP, "Need Optimization\n");
                     fpu_purgecache(dyn, ninst, x1, x2, x3);
                     addr = geted(dyn, addr, ninst, nextop, &ed, x1, &fixedaddress, 0, 0, rex, 0, 0);
                     if(ed!=x1) {MOVx_REG(x1, ed);}