[DYNAREC] Added DF opcodes

4 files changed, 342 insertions, 12 deletions

diff --git a/src/dynarec/arm64_emitter.h b/src/dynarec/arm64_emitter.h
index 8695771b..cb00c19e 100755
--- a/src/dynarec/arm64_emitter.h
+++ b/src/dynarec/arm64_emitter.h
@@ -611,10 +611,27 @@
 #define MRS_nzvc(Rt)                    EMIT(MRS_gen(1, 1, 3, 4, 2, 0, Rt))
 // MSR : to System register
 #define MSR_nzvc(Rt)                    EMIT(MRS_gen(0, 1, 3, 4, 2, 0, Rt))
-// mrs    x0, fpcr : 1101010100 1 1 1 011 0100 0100 000 00000    o0=1(op0=3), op1=0b011(3) CRn=0b0100(4) CRm=0b0100(4) op2=2
+// mrs    x0, fpcr : 1101010100 1 1 1 011 0100 0100 000 00000    o0=1(op0=3), op1=0b011(3) CRn=0b0100(4) CRm=0b0100(4) op2=0
 #define MRS_fpcr(Rt)                    EMIT(MRS_gen(1, 1, 3, 4, 4, 0, Rt))
 #define MSR_fpcr(Rt)                    EMIT(MRS_gen(0, 1, 3, 4, 4, 0, Rt))
-
+// mrs    x0, fpsr : 1101010100 1 1 1 011 0100 0100 001 00000    o0=1(op0=3), op1=0b011(3) CRn=0b0100(4) CRm=0b0100(4) op2=1
+#define MRS_fpsr(Rt)                    EMIT(MRS_gen(1, 1, 3, 4, 4, 1, Rt))
+#define MSR_fpsr(Rt)                    EMIT(MRS_gen(0, 1, 3, 4, 4, 1, Rt))
+// NEON Saturation Bit
+#define FPSR_QC 27
+// NEON Input Denormal Cumulative
+#define FPSR_IDC    7
+// NEON IneXact Cumulative
+#define FPSR_IXC    4
+// NEON Underflow Cumulative
+#define FPSR_UFC    3
+// NEON Overflow Cumulative
+#define FPSR_OFC    2
+// NEON Divide by 0 Cumulative
+#define FPSR_DZC    1
+// NEON Invalid Operation Cumulative
+#define FPSR_IOC    0
+                     
 // FCSEL
 #define FCSEL_scalar(type, Rm, cond, Rn, Rd)    (0b11110<<24 | (type)<<22 | 1<<21 | (Rm)<<16 | (cond)<<12 | 0b11<<10 | (Rn)<<5 | (Rd))
 #define FCSELS(Sd, Sn, Sm, cond)        EMIT(FCSEL_scalar(0b00, Sm, cond, Sn, Sd))
@@ -634,6 +651,8 @@
 #define VSTR64_U12(Dt, Rn, imm15)           EMIT(VMEM_gen(0b11, 0b00, ((uint32_t)(imm15>>3))&0xfff, Rn, Dt))
 // imm16 must be 4-aligned
 #define VSTR128_U12(Qt, Rn, imm16)          EMIT(VMEM_gen(0b00, 0b10, ((uint32_t)((imm16)>>4))&0xfff, Rn, Qt))
+// (imm14) must be 1-aligned
+#define VSTR16_U12(Ht, Rn, imm14)           EMIT(VMEM_gen(0b01, 0b00, ((uint32_t)(imm14>>1))&0xfff, Rn, Ht))
 
 #define VMEMUR_vector(size, opc, imm9, Rn, Rt)  ((size)<<30 | 0b111<<27 | 1<<26 | (opc)<<22 | (imm9)<<12 | (Rn)<<5 | (Rt))
 // signed offset, no alignement!
@@ -1118,15 +1137,21 @@
 #define VFRINTISQ(Vd,Vn)            EMIT(FRINT_vector(1, 1, 1, 0, 1, Vn, Vd))
 #define VFRINTIDQ(Vd,Vn)            EMIT(FRINT_vector(1, 1, 1, 1, 1, Vn, Vd))
 
-#define FRINT_scalar(type, op, Rn, Rd)  (0b11110<<24 | (type)<<22 | 1<<21 | 0b0100<<17 | (op)<<15 | 0b10000<<10 | (Rn)<<5 | (Rd))
-#define FRINT32ZS(Sd, Sn)           EMIT(FRINT_scalar(0b00, 0b00, Sn, Sd))
-#define FRINT32ZD(Dd, Dn)           EMIT(FRINT_scalar(0b01, 0b00, Dn, Dd))
-#define FRINT32XS(Sd, Sn)           EMIT(FRINT_scalar(0b00, 0b01, Sn, Sd))
-#define FRINT32XD(Dd, Dn)           EMIT(FRINT_scalar(0b01, 0b01, Dn, Dd))
-#define FRINT64ZS(Sd, Sn)           EMIT(FRINT_scalar(0b00, 0b10, Sn, Sd))
-#define FRINT64ZD(Dd, Dn)           EMIT(FRINT_scalar(0b01, 0b10, Dn, Dd))
-#define FRINT64XS(Sd, Sn)           EMIT(FRINT_scalar(0b00, 0b11, Sn, Sd))
-#define FRINT64XD(Dd, Dn)           EMIT(FRINT_scalar(0b01, 0b11, Dn, Dd))
+#define FRINTxx_scalar(type, op, Rn, Rd)  (0b11110<<24 | (type)<<22 | 1<<21 | 0b0100<<17 | (op)<<15 | 0b10000<<10 | (Rn)<<5 | (Rd))
+#define FRINT32ZS(Sd, Sn)           EMIT(FRINTxx_scalar(0b00, 0b00, Sn, Sd))
+#define FRINT32ZD(Dd, Dn)           EMIT(FRINTxx_scalar(0b01, 0b00, Dn, Dd))
+#define FRINT32XS(Sd, Sn)           EMIT(FRINTxx_scalar(0b00, 0b01, Sn, Sd))
+#define FRINT32XD(Dd, Dn)           EMIT(FRINTxx_scalar(0b01, 0b01, Dn, Dd))
+#define FRINT64ZS(Sd, Sn)           EMIT(FRINTxx_scalar(0b00, 0b10, Sn, Sd))
+#define FRINT64ZD(Dd, Dn)           EMIT(FRINTxx_scalar(0b01, 0b10, Dn, Dd))
+#define FRINT64XS(Sd, Sn)           EMIT(FRINTxx_scalar(0b00, 0b11, Sn, Sd))
+#define FRINT64XD(Dd, Dn)           EMIT(FRINTxx_scalar(0b01, 0b11, Dn, Dd))
+
+#define FRINT_scalar(type, rmode, Rn, Rd)   (0b11110<<24 | (type)<<22 | 1<<21 | 0b001<<18 | (rmode)<<15 | 0b10000<<10 | (Rn)<<5 | (Rd))
+#define FRINTZS(Sd, Sn)             EMIT(FRINT_scalar(0b00, 0b011, Sn, Sd))
+#define FRINTZD(Sd, Sn)             EMIT(FRINT_scalar(0b01, 0b011, Sn, Sd))
+#define FRINTXS(Sd, Sn)             EMIT(FRINT_scalar(0b00, 0b110, Sn, Sd))
+#define FRINTXD(Sd, Sn)             EMIT(FRINT_scalar(0b01, 0b110, Sn, Sd))
 
 // FMAX / FMIN
 #define FMINMAX_vector(Q, U, o1, sz, Rm, Rn, Rd)    ((Q)<<30 | (U)<<29 | 0b01110<<24 | (o1)<<23 | (sz)<<22 | 0b1<<21 | (Rm)<<16 | 0b11110<<11 | 1<<10 | (Rn)<<5 | (Rd))
@@ -1217,6 +1242,14 @@
 #define VTRNQ2_8(Vd, Vn, Vm)        EMIT(TRN_gen(1, 0b00, Vm, 1, Vn, Vd))
 
 // QXTN / QXTN2
+#define QXTN_scalar(U, size, Rn, Rd)        (0b01<<30 | (U)<<29 | 0b11110<<24 | (size)<<22 | 0b10000<<17 | 0b10100<<12 | 0b10<<10 | (Rn)<<5 | (Rd))
+// Signed saturating extract Narrow, from D to S
+#define SQXTN_S_D(Sd, Dn)           EMIT(QXTN_scalar(0, 0b10, Dn, Sd))
+// Signed saturating extract Narrow, from S to H
+#define SQXTN_H_S(Hd, Sn)           EMIT(QXTN_scalar(0, 0b01, Sn, Hd))
+// Signed saturating extract Narrow, from H to B
+#define SQXTN_B_H(Bd, Hn)           EMIT(QXTN_scalar(0, 0b00, Hn, Bd))
+
 #define QXTN_vector(Q, U, size, Rn, Rd)     ((Q)<<30 | (U)<<29 | 0b01110<<24 | (size)<<22 | 0b10000<<17 | 0b10100<<12 | 0b10<<10 | (Rn)<<5 | (Rd))
 // Signed saturating extract Narrow, takes Rn element and reduce 64->32 with Signed saturation and fit lower part of Rd
 #define SQXTN_32(Rd, Rn)            EMIT(QXTN_vector(0, 0, 0b10, Rn, Rd))
diff --git a/src/dynarec/dynarec_arm64_00.c b/src/dynarec/dynarec_arm64_00.c
index 1b0c573e..783db2c3 100755
--- a/src/dynarec/dynarec_arm64_00.c
+++ b/src/dynarec/dynarec_arm64_00.c
@@ -1844,6 +1844,10 @@ uintptr_t dynarec64_00(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
             addr = dynarec64_DD(dyn, addr, ip, ninst, rex, rep, ok, need_epilog);
             break;
 
+        case 0xDF:
+            addr = dynarec64_DF(dyn, addr, ip, ninst, rex, rep, ok, need_epilog);
+            break;
+
         case 0xE8:
             INST_NAME("CALL Id");
             i32 = F32S;
diff --git a/src/dynarec/dynarec_arm64_df.c b/src/dynarec/dynarec_arm64_df.c
new file mode 100644
index 00000000..981fe830
--- /dev/null
+++ b/src/dynarec/dynarec_arm64_df.c
@@ -0,0 +1,289 @@
+#include <stdio.h>
+#include <stdlib.h>
+#include <stddef.h>
+#include <pthread.h>
+#include <errno.h>
+
+#include "debug.h"
+#include "box64context.h"
+#include "dynarec.h"
+#include "emu/x64emu_private.h"
+#include "emu/x64run_private.h"
+#include "x64run.h"
+#include "x64emu.h"
+#include "box64stack.h"
+#include "callback.h"
+#include "emu/x64run_private.h"
+#include "x64trace.h"
+#include "dynarec_arm64.h"
+#include "dynarec_arm64_private.h"
+#include "arm64_printer.h"
+#include "emu/x87emu_private.h"
+
+#include "dynarec_arm64_helper.h"
+#include "dynarec_arm64_functions.h"
+
+
+uintptr_t dynarec64_DF(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int ninst, rex_t rex, int rep, int* ok, int* need_epilog)
+{
+    uint8_t nextop = F8;
+    uint8_t ed, wback, u8;
+    int fixedaddress;
+    int v1, v2;
+    int j32;
+    int s0;
+
+    MAYUSE(s0);
+    MAYUSE(v2);
+    MAYUSE(v1);
+    MAYUSE(j32);
+
+    switch(nextop) {
+        case 0xC0:
+        case 0xC1:
+        case 0xC2:
+        case 0xC3:
+        case 0xC4:
+        case 0xC5:
+        case 0xC6:
+        case 0xC7:
+            INST_NAME("FFREEP STx");
+            // not handling Tag...
+            x87_do_pop(dyn, ninst);
+            break;
+
+        case 0xE0:
+            INST_NAME("FNSTSW AX");
+            LDRw_U12(x2, xEmu, offsetof(x64emu_t, top));
+            LDRH_U12(x1, xEmu, offsetof(x64emu_t, sw));
+            BFIw(x1, x2, 11, 3); // inject top
+            BFIw(xRAX, x1, 0, 16);
+            break;
+        case 0xE8:
+        case 0xE9:
+        case 0xEA:
+        case 0xEB:
+        case 0xEC:
+        case 0xED:
+        case 0xEE:
+        case 0xEF:
+            INST_NAME("FUCOMIP ST0, STx");
+            SETFLAGS(X_ALL, SF_SET);
+            v1 = x87_get_st(dyn, ninst, x1, x2, 0);
+            v2 = x87_get_st(dyn, ninst, x1, x2, nextop&7);
+            FCMPD(v1, v2);
+            FCOMI(x1, x2);
+            x87_do_pop(dyn, ninst);
+            break;
+        case 0xF0:
+        case 0xF1:
+        case 0xF2:
+        case 0xF3:
+        case 0xF4:
+        case 0xF5:
+        case 0xF6:
+        case 0xF7:
+            INST_NAME("FCOMIP ST0, STx");
+            SETFLAGS(X_ALL, SF_SET);
+            v1 = x87_get_st(dyn, ninst, x1, x2, 0);
+            v2 = x87_get_st(dyn, ninst, x1, x2, nextop&7);
+            FCMPD(v1, v2);
+            FCOMI(x1, x2);
+            x87_do_pop(dyn, ninst);
+            break;
+
+        case 0xC8:
+        case 0xC9:
+        case 0xCA:
+        case 0xCB:
+        case 0xCC:
+        case 0xCD:
+        case 0xCE:
+        case 0xCF:
+        case 0xD0:
+        case 0xD1:
+        case 0xD2:
+        case 0xD3:
+        case 0xD4:
+        case 0xD5:
+        case 0xD6:
+        case 0xD7:
+        case 0xD8:
+        case 0xD9:
+        case 0xDA:
+        case 0xDB:
+        case 0xDC:
+        case 0xDD:
+        case 0xDE:
+        case 0xDF:
+        case 0xE1:
+        case 0xE2:
+        case 0xE3:
+        case 0xE4:
+        case 0xE5:
+        case 0xE6:
+        case 0xE7:
+        case 0xF8:
+        case 0xF9:
+        case 0xFA:
+        case 0xFB:
+        case 0xFC:
+        case 0xFD:
+        case 0xFE:
+        case 0xFF:
+            DEFAULT;
+            break;
+
+        default:
+            switch((nextop>>3)&7) {
+                case 0:
+                    INST_NAME("FILD ST0, Ew");
+                    v1 = x87_do_push(dyn, ninst);
+                    addr = geted(dyn, addr, ninst, nextop, &wback, x3, &fixedaddress, 0xfff<<1, 1, rex, 0, 0);
+                    LDRSHw_U12(x1, wback, fixedaddress);
+                    SCVTFDw(v1, x1);
+                    break;
+                case 1:
+                    INST_NAME("FISTTP Ew, ST0");
+                    v1 = x87_get_st(dyn, ninst, x1, x2, 0);
+                    addr = geted(dyn, addr, ninst, nextop, &wback, x2, &fixedaddress, 0xfff<<1, 1, rex, 0, 0);
+                    ed = x1;
+                    s0 = fpu_get_scratch(dyn);
+                    #if 0
+                    // this version needs ARM v8.5, //TODO: add detection of this extensio to use it
+                    FRINT32ZD(s0, v1);
+                    // no saturation instruction on Arm, so using NEON
+                    VFCVTZSd(s0, s0);
+                    SQXTN_S_D(s0, s0);
+                    SQXTN_H_S(s0, s0);
+                    VSTR16_U12(s0, wback, fixedaddress);
+                    #else
+                    MSR_fpsr(x5);
+                    BFCw(x5, FPSR_IOC, 1);   // reset IOC bit
+                    MRS_fpsr(x5);
+                    VFCVTZSd(s0, v1);
+                    SQXTN_S_D(s0, s0);
+                    SQXTN_H_S(s0, s0);
+                    VSTR16_U12(s0, wback, fixedaddress);
+                    MSR_fpsr(x5);   // get back FPSR to check the IOC bit
+                    TBZ_NEXT(x5, FPSR_IOC);
+                    MOV32w(x5, 0x8000);
+                    STRH_U12(x5, wback, fixedaddress);
+                    #endif
+                    x87_do_pop(dyn, ninst);
+                    break;
+                case 2:
+                    INST_NAME("FIST Ew, ST0");
+                    v1 = x87_get_st(dyn, ninst, x1, x2, 0);
+                    u8 = x87_setround(dyn, ninst, x1, x2, x4);
+                    addr = geted(dyn, addr, ninst, nextop, &wback, x2, &fixedaddress, 0xfff<<1, 1, rex, 0, 0);
+                    ed = x1;
+                    s0 = fpu_get_scratch(dyn);
+                    #if 0
+                    FRINT32XD(s0, v1);
+                    // no saturation instruction on Arm, so using NEON
+                    VFCVTZSd(s0, s0);
+                    SQXTN_S_D(s0, s0);
+                    SQXTN_H_S(s0, s0);
+                    VSTR16_U12(s0, wback, fixedaddress);
+                    #else
+                    MSR_fpsr(x5);
+                    BFCw(x5, FPSR_IOC, 1);   // reset IOC bit
+                    MRS_fpsr(x5);
+                    FRINTXD(s0, v1);
+                    VFCVTZSd(s0, s0);
+                    SQXTN_S_D(s0, s0);
+                    SQXTN_H_S(s0, s0);
+                    VSTR16_U12(s0, wback, fixedaddress);
+                    MSR_fpsr(x5);   // get back FPSR to check the IOC bit
+                    TBZ_NEXT(x5, FPSR_IOC);
+                    MOV32w(x5, 0x8000);
+                    STRH_U12(x5, wback, fixedaddress);
+                    #endif
+                    x87_restoreround(dyn, ninst, u8);
+                    break;
+                case 3:
+                    INST_NAME("FISTP Ew, ST0");
+                    v1 = x87_get_st(dyn, ninst, x1, x2, 0);
+                    u8 = x87_setround(dyn, ninst, x1, x2, x4);
+                    addr = geted(dyn, addr, ninst, nextop, &wback, x2, &fixedaddress, 0xfff<<1, 1, rex, 0, 0);
+                    ed = x1;
+                    s0 = fpu_get_scratch(dyn);
+                    #if 0
+                    FRINT32XD(s0, v1);
+                    // no saturation instruction on Arm, so using NEON
+                    VFCVTZSd(s0, s0);
+                    SQXTN_S_D(s0, s0);
+                    SQXTN_H_S(s0, s0);
+                    VSTR16_U12(s0, wback, fixedaddress);
+                    #else
+                    MSR_fpsr(x5);
+                    BFCw(x5, FPSR_IOC, 1);   // reset IOC bit
+                    MRS_fpsr(x5);
+                    FRINTXD(s0, v1);
+                    VFCVTZSd(s0, s0);
+                    SQXTN_S_D(s0, s0);
+                    SQXTN_H_S(s0, s0);
+                    VSTR16_U12(s0, wback, fixedaddress);
+                    MSR_fpsr(x5);   // get back FPSR to check the IOC bit
+                    TBZ_NEXT(x5, FPSR_IOC);
+                    MOV32w(x5, 0x8000);
+                    STRH_U12(x5, wback, fixedaddress);
+                    #endif
+                    x87_do_pop(dyn, ninst);
+                    x87_restoreround(dyn, ninst, u8);
+                    break;
+                case 4:
+                    INST_NAME("FBLD ST0, tbytes");
+                    x87_do_push_empty(dyn, ninst, x1);
+                    addr = geted(dyn, addr, ninst, nextop, &ed, x1, &fixedaddress, 0, 0, rex, 0, 0);
+                    if(ed!=x1) {MOVx_REG(x1, ed);}
+                    CALL(fpu_fbld, -1);
+                    break;
+                case 5:
+                    INST_NAME("FILD ST0, i64");
+                    v1 = x87_do_push(dyn, ninst);
+                    addr = geted(dyn, addr, ninst, nextop, &wback, x3, &fixedaddress, 0xfff<<3, 7, rex, 0, 0);
+                    LDRx_U12(x1, wback, fixedaddress);
+                    SCVTFDx(v1, x1);
+                    break;
+                case 6:
+                    INST_NAME("FBSTP tbytes, ST0");
+                    x87_forget(dyn, ninst, x1, x2, 0);
+                    addr = geted(dyn, addr, ninst, nextop, &ed, x1, &fixedaddress, 0, 0, rex, 0, 0);
+                    if(ed!=x1) {MOVx_REG(x1, ed);}
+                    CALL(fpu_fbst, -1);
+                    x87_do_pop(dyn, ninst);
+                    break;
+                case 7:
+                    INST_NAME("FISTP i64, ST0");
+                    v1 = x87_get_st(dyn, ninst, x1, x2, 0);
+                    u8 = x87_setround(dyn, ninst, x1, x2, x4);
+                    addr = geted(dyn, addr, ninst, nextop, &wback, x2, &fixedaddress, 0xfff<<3, 7, rex, 0, 0);
+                    ed = x1;
+                    s0 = fpu_get_scratch(dyn);
+                    #if 0
+                    FRINT64XD(s0, v1);
+                    VFCVTZSd(s0, s0);
+                    VSTR64_U12(s0, wback, fixedaddress);
+                    #else
+                    MSR_fpsr(x5);
+                    BFCw(x5, FPSR_IOC, 1);   // reset IOC bit
+                    MRS_fpsr(x5);
+                    FRINTXD(s0, v1);
+                    VFCVTZSd(s0, s0);
+                    VSTR64_U12(s0, wback, fixedaddress);
+                    MSR_fpsr(x5);   // get back FPSR to check the IOC bit
+                    TBZ_NEXT(x5, FPSR_IOC);
+                    MOV64x(x5, 0x8000000000000000LL);
+                    STRx_U12(x5, wback, fixedaddress);
+                    #endif
+                    x87_restoreround(dyn, ninst, u8);
+                    break;
+                default:
+                    DEFAULT;
+            }
+    }
+    return addr;
+}
+
diff --git a/src/dynarec/dynarec_arm64_helper.h b/src/dynarec/dynarec_arm64_helper.h
index 62539654..61a115c5 100755
--- a/src/dynarec/dynarec_arm64_helper.h
+++ b/src/dynarec/dynarec_arm64_helper.h
@@ -328,6 +328,10 @@
 #define CBZx_NEXT(reg)    \
     j32 =  (dyn->insts)?(dyn->insts[ninst].epilog-(dyn->arm_size)):0; \
     CBZx(reg, j32)
+// Test bit N of A and branch to next instruction if not set
+#define TBZ_NEXT(A, N)              \
+    j32 = (dyn->insts)?(dyn->insts[ninst].epilog-(dyn->arm_size)):0; \
+    TBZ(A, N, j32)
 // Branch to MARKSEG if cond (use j32)
 #define B_MARKSEG(cond)    \
     j32 = GETMARKSEG-(dyn->arm_size);   \
@@ -845,7 +849,7 @@ uintptr_t dynarec64_DB(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
 //uintptr_t dynarec64_DC(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int ninst, rex_t rex, int rep, int* ok, int* need_epilog);
 uintptr_t dynarec64_DD(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int ninst, rex_t rex, int rep, int* ok, int* need_epilog);
 //uintptr_t dynarec64_DE(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int ninst, rex_t rex, int rep, int* ok, int* need_epilog);
-//uintptr_t dynarec64_DF(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int ninst, rex_t rex, int rep, int* ok, int* need_epilog);
+uintptr_t dynarec64_DF(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int ninst, rex_t rex, int rep, int* ok, int* need_epilog);
 uintptr_t dynarec64_F0(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int ninst, rex_t rex, int rep, int* ok, int* need_epilog);
 uintptr_t dynarec64_660F(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int ninst, rex_t rex, int rep, int* ok, int* need_epilog);
 uintptr_t dynarec64_F20F(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int ninst, rex_t rex, int* ok, int* need_epilog);