[RV64_DYNAREC] Added (66) 0F 2A/2C/2D opcodes

4 files changed, 147 insertions, 13 deletions

diff --git a/src/dynarec/rv64/dynarec_rv64_0f.c b/src/dynarec/rv64/dynarec_rv64_0f.c
index 199068cf..9e225bd2 100644
--- a/src/dynarec/rv64/dynarec_rv64_0f.c
+++ b/src/dynarec/rv64/dynarec_rv64_0f.c
@@ -262,7 +262,20 @@ uintptr_t dynarec64_0F(dynarec_rv64_t* dyn, uintptr_t addr, uintptr_t ip, int ni
             if(!MODREG)
                 SMWRITE2();
             break;
-
+        case 0x2A:
+            INST_NAME("CVTPI2PS Gx,Em");
+            nextop = F8;
+            GETGX();
+            GETEM(x2, 0);
+            d0 = fpu_get_scratch(dyn);
+            u8 = sse_setround(dyn, ninst, x4, x5);
+            for (int i=0; i<2; ++i) {
+                LW(x3, wback, fixedaddress+i*4);
+                FCVTSW(d0, x3, RD_DYN);
+                FSW(d0, gback, gdoffset+i*4);
+            }
+            x87_restoreround(dyn, ninst, u8);
+            break;
         case 0x2B:
             INST_NAME("MOVNTPS Ex,Gx");
             nextop = F8;
@@ -273,6 +286,52 @@ uintptr_t dynarec64_0F(dynarec_rv64_t* dyn, uintptr_t addr, uintptr_t ip, int ni
             SD(x3, wback, fixedaddress+0);
             SD(x4, wback, fixedaddress+8);
             break;
+        case 0x2C:
+            INST_NAME("CVTTPS2PI Gm,Ex");
+            nextop = F8;
+            GETGM();
+            GETEX(x2, 0);
+            d0 = fpu_get_scratch(dyn);
+            for (int i=0; i<2; ++i) {
+                if(!box64_dynarec_fastround) {
+                    FSFLAGSI(0);  // // reset all bits
+                }
+                FLW(d0, wback, fixedaddress+i*4);
+                FCVTWS(x1, d0, RD_RTZ);
+                if(!box64_dynarec_fastround) {
+                    FRFLAGS(x5);   // get back FPSR to check the IOC bit
+                    ANDI(x5, x5, (1<<FR_NV)|(1<<FR_OF));
+                    BEQ_MARKi(x5, xZR, i);
+                    MOV32w(x1, 0x80000000);
+                    MARKi(i);
+                }
+                SW(x1, gback, gdoffset+i*4);
+            }
+            break;
+        case 0x2D:
+            INST_NAME("CVTPS2PI Gm, Ex");
+            nextop = F8;
+            GETGM();
+            GETEX(x2, 0);
+            d0 = fpu_get_scratch(dyn);
+            u8 = sse_setround(dyn, ninst, x6, x4);
+            for (int i=0; i<2; ++i) {
+                if(!box64_dynarec_fastround) {
+                    FSFLAGSI(0);  // // reset all bits
+                }
+                FLW(d0, wback, fixedaddress+i*4);
+                FCVTWS(x1, d0, RD_DYN);
+                if(!box64_dynarec_fastround) {
+                    FRFLAGS(x5);   // get back FPSR to check the IOC bit
+                    ANDI(x5, x5, (1<<FR_NV)|(1<<FR_OF));
+                    BEQ_MARKi(x5, xZR, i);
+                    MOV32w(x1, 0x80000000);
+                    MARKi(i);
+                }
+                SW(x1, gback, gdoffset+i*4);
+            }
+            x87_restoreround(dyn, ninst, u8);
+            break;
         case 0x2E:
             // no special check...
         case 0x2F:
diff --git a/src/dynarec/rv64/dynarec_rv64_660f.c b/src/dynarec/rv64/dynarec_rv64_660f.c
index c60d214d..7baddf4f 100644
--- a/src/dynarec/rv64/dynarec_rv64_660f.c
+++ b/src/dynarec/rv64/dynarec_rv64_660f.c
@@ -148,6 +148,18 @@ uintptr_t dynarec64_660F(dynarec_rv64_t* dyn, uintptr_t addr, uintptr_t ip, int
             SSE_LOOP_MV_Q2(x3);
             if(!MODREG) SMWRITE2();
             break;
+        case 0x2A:
+            INST_NAME("CVTPI2PD Gx,Em");
+            nextop = F8;
+            GETGX();
+            GETEM(x2, 0);
+            d0 = fpu_get_scratch(dyn);
+            for (int i=0; i<2; ++i) {
+                LW(x1, wback, fixedaddress+i*4);
+                FCVTDW(d0, x1, RD_RTZ);
+                FSD(d0, gback, gdoffset+i*8);
+            }
+            break;
         case 0x2B:
             INST_NAME("MOVNTPD Ex, Gx");
             nextop = F8;
@@ -155,6 +167,52 @@ uintptr_t dynarec64_660F(dynarec_rv64_t* dyn, uintptr_t addr, uintptr_t ip, int
             GETEX(x2, 0);
             SSE_LOOP_MV_Q2(x3);
             break;
+        case 0x2C:
+            INST_NAME("CVTTPD2PI Gm,Ex");
+            nextop = F8;
+            GETGM();
+            GETEX(x2, 0);
+            d0 = fpu_get_scratch(dyn);
+            for (int i=0; i<2; ++i) {
+                if(!box64_dynarec_fastround) {
+                    FSFLAGSI(0);  // // reset all bits
+                }
+                FLD(d0, wback, fixedaddress+i*8);
+                FCVTWD(x1, d0, RD_RTZ);
+                if(!box64_dynarec_fastround) {
+                    FRFLAGS(x5);   // get back FPSR to check the IOC bit
+                    ANDI(x5, x5, (1<<FR_NV)|(1<<FR_OF));
+                    BEQ_MARKi(x5, xZR, i);
+                    MOV32w(x1, 0x80000000);
+                    MARKi(i);
+                }
+                SW(x1, gback, gdoffset+i*4);
+            }
+            break;
+        case 0x2D:
+            INST_NAME("CVTPD2PI Gm,Ex");
+            nextop = F8;
+            GETGM();
+            GETEX(x2, 0);
+            d0 = fpu_get_scratch(dyn);
+            u8 = sse_setround(dyn, ninst, x4, x5);
+            for (int i=0; i<2; ++i) {
+                if(!box64_dynarec_fastround) {
+                    FSFLAGSI(0);  // // reset all bits
+                }
+                FLD(d0, wback, fixedaddress+i*8);
+                FCVTWD(x1, d0, RD_DYN);
+                if(!box64_dynarec_fastround) {
+                    FRFLAGS(x5);   // get back FPSR to check the IOC bit
+                    ANDI(x5, x5, (1<<FR_NV)|(1<<FR_OF));
+                    BEQ_MARKi(x5, xZR, i);
+                    MOV32w(x1, 0x80000000);
+                    MARKi(i);
+                }
+                SW(x1, gback, gdoffset+i*4);
+            }
+            x87_restoreround(dyn, ninst, u8);
+            break;
         case 0x2E:
             // no special check...
         case 0x2F:
diff --git a/src/dynarec/rv64/dynarec_rv64_helper.h b/src/dynarec/rv64/dynarec_rv64_helper.h
index 0b1023b3..51d7817f 100644
--- a/src/dynarec/rv64/dynarec_rv64_helper.h
+++ b/src/dynarec/rv64/dynarec_rv64_helper.h
@@ -531,16 +531,22 @@
 // R0 will not be pushed/popd if ret is -2. Flags are not save/restored
 #define CALL_S(F, ret) call_c(dyn, ninst, F, x6, ret, 0, 0)
 
-#define MARK        dyn->insts[ninst].mark = dyn->native_size
-#define GETMARK     dyn->insts[ninst].mark
-#define MARK2       dyn->insts[ninst].mark2 = dyn->native_size
-#define GETMARK2    dyn->insts[ninst].mark2
-#define MARK3       dyn->insts[ninst].mark3 = dyn->native_size
-#define GETMARK3    dyn->insts[ninst].mark3
-#define MARKF       dyn->insts[ninst].markf = dyn->native_size
-#define GETMARKF    dyn->insts[ninst].markf
-#define MARKF2      dyn->insts[ninst].markf2 = dyn->native_size
-#define GETMARKF2   dyn->insts[ninst].markf2
+#define MARKi(i)   dyn->insts[ninst].mark[i] = dyn->native_size
+#define GETMARKi(i) dyn->insts[ninst].mark[i]
+#define MARK        MARKi(0)
+#define GETMARK     GETMARKi(0)
+#define MARK2       MARKi(1)
+#define GETMARK2    GETMARKi(1)
+#define MARK3       MARKi(2)
+#define GETMARK3    GETMARKi(2)
+
+#define MARKFi(i)   dyn->insts[ninst].markf[i] = dyn->native_size
+#define GETMARKFi(i) dyn->insts[ninst].markf[i]
+#define MARKF       MARKFi(0)
+#define GETMARKF    GETMARKFi(0)
+#define MARKF2      MARKFi(1)
+#define GETMARKF2   GETMARKFi(1)
+
 #define MARKSEG     dyn->insts[ninst].markseg = dyn->native_size
 #define GETMARKSEG  dyn->insts[ninst].markseg
 #define MARKLOCK    dyn->insts[ninst].marklock = dyn->native_size
@@ -550,20 +556,31 @@
     j64 = GET##M - dyn->native_size;    \
     B##OP (reg1, reg2, j64)
 
+#define Bxx_geni(OP, M, reg1, reg2, i)      \
+    j64 = GET##M##i(i) - dyn->native_size;    \
+    B##OP (reg1, reg2, j64)
+
 // Branch to MARK if reg1==reg2 (use j64)
 #define BEQ_MARK(reg1, reg2) Bxx_gen(EQ, MARK, reg1, reg2)
+#define BEQ_MARKi(reg1, reg2, i) Bxx_geni(EQ, MARK, reg1, reg2, i)
 // Branch to MARK if reg1!=reg2 (use j64)
 #define BNE_MARK(reg1, reg2) Bxx_gen(NE, MARK, reg1, reg2)
+#define BNE_MARKi(reg1, reg2, i) Bxx_geni(NE, MARK, reg1, reg2, i)
 // Branch to MARK if reg1!=0 (use j64)
 #define BNEZ_MARK(reg) BNE_MARK(reg, xZR)
+#define BNEZ_MARKi(reg, i) BNE_MARKi(reg, xZR, i)
 // Branch to MARK instruction unconditionnal (use j64)
 #define B_MARK_nocond   Bxx_gen(__, MARK, 0, 0)
+#define B_MARKi_nocond   Bxx_geni(__, MARK, 0, 0, i)
 // Branch to MARK if reg1<reg2 (use j64)
 #define BLT_MARK(reg1, reg2) Bxx_gen(LT, MARK, reg1, reg2)
+#define BLT_MARKi(reg1, reg2) Bxx_geni(LT, MARK, reg1, reg2, i)
 // Branch to MARK if reg1<reg2 (use j64)
 #define BLTU_MARK(reg1, reg2) Bxx_gen(LTU, MARK, reg1, reg2)
+#define BLTU_MARKi(reg1, reg2) Bxx_geni(LTU, MARK, reg1, reg2, i)
 // Branch to MARK if reg1>=reg2 (use j64)
 #define BGE_MARK(reg1, reg2) Bxx_gen(GE, MARK, reg1, reg2)
+#define BGE_MARKi(reg1, reg2) Bxx_geni(GE, MARK, reg1, reg2, i)
 // Branch to MARK2 if reg1==reg2 (use j64)
 #define BEQ_MARK2(reg1, reg2) Bxx_gen(EQ, MARK2, reg1,reg2)
 // Branch to MARK2 if reg1!=reg2 (use j64)
diff --git a/src/dynarec/rv64/dynarec_rv64_private.h b/src/dynarec/rv64/dynarec_rv64_private.h
index 99135297..d377309d 100644
--- a/src/dynarec/rv64/dynarec_rv64_private.h
+++ b/src/dynarec/rv64/dynarec_rv64_private.h
@@ -79,8 +79,8 @@ typedef struct instruction_rv64_s {
     int                 size2;      // size of the arm emitted instrucion after pass2
     int                 pred_sz;    // size of predecessor list
     int                 *pred;      // predecessor array
-    uintptr_t           mark, mark2, mark3;
-    uintptr_t           markf, markf2;
+    uintptr_t           mark[3];
+    uintptr_t           markf[2];
     uintptr_t           markseg;
     uintptr_t           marklock;
     int                 pass2choice;// value for choices that are fixed on pass2 for pass3