[INTERPRETER] Added a few more cpu extension and associated opcodes ([ARM64_DYNAREC] added some of the new opcodes too)

5 files changed, 115 insertions, 4 deletions

diff --git a/src/dynarec/arm64/dynarec_arm64_0f.c b/src/dynarec/arm64/dynarec_arm64_0f.c
index 058478c6..6d5907f7 100644
--- a/src/dynarec/arm64/dynarec_arm64_0f.c
+++ b/src/dynarec/arm64/dynarec_arm64_0f.c
@@ -2468,6 +2468,11 @@ uintptr_t dynarec64_0F(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
                 IFX(X_CF) { BFIw(xFlags, x3, F_CF, 1); }

                 MOVxw_REG(ed, x1);

                 break;

+            case 7:

+                INST_NAME("RDPID Ed");

+                GETED(0);

+                CALL_(helper_getcpu, ed, x2);

+                break;

             default:

                 DEFAULT;

             } else switch((nextop>>3)&7) {

diff --git a/src/dynarec/arm64/dynarec_arm64_660f.c b/src/dynarec/arm64/dynarec_arm64_660f.c
index 0f336876..f2184f58 100644
--- a/src/dynarec/arm64/dynarec_arm64_660f.c
+++ b/src/dynarec/arm64/dynarec_arm64_660f.c
@@ -2455,7 +2455,37 @@ uintptr_t dynarec64_660F(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int n
             emit_shrd16(dyn, ninst, ed, gd, x4, x5, x6);

             EWBACK;

             break;

-

+        case 0xAE:

+            nextop = F8;

+            if(MODREG)

+                switch (nextop) {

+                    default:

+                        DEFAULT;

+                }

+            else

+                switch((nextop>>3)&7) {

+                    case 6:

+                        INST_NAME("CLWB Ed");

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

+                        addr = geted(dyn, addr, ninst, nextop, &ed, x1, &fixedaddress, NULL, 0, 0, rex, NULL, 0, 0);

+                        if(ed!=x1) {

+                            MOVx_REG(x1, ed);

+                        }

+                        CALL_(native_clflush, -1, 0);

+                        break;

+                    case 7:

+                        INST_NAME("CLFLUSHOPT Ed");

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

+                        addr = geted(dyn, addr, ninst, nextop, &ed, x1, &fixedaddress, NULL, 0, 0, rex, NULL, 0, 0);

+                        if(ed!=x1) {

+                            MOVx_REG(x1, ed);

+                        }

+                        CALL_(native_clflush, -1, 0);

+                        break;

+                    default:

+                        DEFAULT;

+                }

+            break;

         case 0xAF:

             INST_NAME("IMUL Gw,Ew");

             SETFLAGS(X_ALL, SF_SET);

diff --git a/src/emu/x64run0f.c b/src/emu/x64run0f.c
index 041b8320..911d80b1 100644
--- a/src/emu/x64run0f.c
+++ b/src/emu/x64run0f.c
@@ -90,6 +90,18 @@ uintptr_t Run0F(x64emu_t *emu, rex_t rex, uintptr_t addr, int *step)
             GETED(0);

             if(MODREG)

             switch(nextop) {

+                case 0xC8:  /* MONITOR */

+                    // this is a privilege opcode...

+                    #ifndef TEST_INTERPRETER

+                    emit_signal(emu, SIGSEGV, (void*)R_RIP, 0);

+                    #endif

+                    break;

+                case 0xC9:  /* MWAIT */

+                    // this is a privilege opcode...

+                    #ifndef TEST_INTERPRETER

+                    emit_signal(emu, SIGSEGV, (void*)R_RIP, 0);

+                    #endif

+                    break;

                 case 0xD0:

                     if(R_RCX) {

                         #ifndef TEST_INTERPRETER

@@ -1803,6 +1815,9 @@ uintptr_t Run0F(x64emu_t *emu, rex_t rex, uintptr_t addr, int *step)
                             ED->dword[1] = 1;

                     }

                     break;

+                case 7:     /* RDPID Ed */

+                    ED->q[0] = helper_getcpu(emu);

+                    break;

                 default:

                     return 0;

             }

diff --git a/src/emu/x64run660f.c b/src/emu/x64run660f.c
index 291a3513..6bd25132 100644
--- a/src/emu/x64run660f.c
+++ b/src/emu/x64run660f.c
@@ -21,6 +21,9 @@
 #include "box64context.h"

 #include "signals.h"

 #include "bridge.h"

+#ifdef DYNAREC

+#include "custommem.h"

+#endif

 

 #include "modrm.h"

 #include "x64compstrings.h"

@@ -699,6 +702,15 @@ uintptr_t Run660F(x64emu_t *emu, rex_t rex, uintptr_t addr)
                 GX->ud[1] = 0;

                 break;

 

+            case 0x82:  /* INVPCID */

+                nextop = F8;

+                GETED(0);

+                // this is a privilege opcode...

+                #ifndef TEST_INTERPRETER

+                emit_signal(emu, SIGSEGV, (void*)R_RIP, 0);

+                #endif

+                break;

+

             case 0xDB:  /* AESIMC Gx, Ex */

                 nextop = F8;

                 GETEX(0);

@@ -1886,7 +1898,35 @@ uintptr_t Run660F(x64emu_t *emu, rex_t rex, uintptr_t addr)
         else

             EW->word[0] = shrd16(emu, EW->word[0], GW->word[0], tmp8u);

         break;

-

+    case 0xAE:                      /* Grp Ed (SSE) */

+        nextop = F8;

+        if(MODREG)

+            switch(nextop) {

+                default:

+                    return 0;

+            }

+        else

+        switch((nextop>>3)&7) {

+            case 6:                 /* CLWB Ed */

+                // same code and CLFLUSH, is it ok?

+                _GETED(0);

+                #if defined(DYNAREC) && !defined(TEST_INTERPRETER)

+                if(box64_dynarec)

+                    cleanDBFromAddressRange((uintptr_t)ED, 8, 0);

+                #endif

+                break;

+            case 7:                 /* CLFLUSHOPT Ed */

+                // same code and CLFLUSH, is it ok?

+                _GETED(0);

+                #if defined(DYNAREC) && !defined(TEST_INTERPRETER)

+                if(box64_dynarec)

+                    cleanDBFromAddressRange((uintptr_t)ED, 8, 0);

+                #endif

+                break;

+            default:

+                return 0;

+        }

+        break;

     case 0xAF:                      /* IMUL Gw,Ew */

         nextop = F8;

         GETEW(0);

diff --git a/src/tools/my_cpuid.c b/src/tools/my_cpuid.c
index f69656fc..7316937a 100644
--- a/src/tools/my_cpuid.c
+++ b/src/tools/my_cpuid.c
@@ -268,16 +268,21 @@ void my_cpuid(x64emu_t* emu, uint32_t tmp32u)
                 R_EAX |= cpu<<24;
             }*/
             R_EDX =   1         // fpu 
+                    | 1<<1      // vme
                     | 1<<2      // debugging extension
                     | 1<<3      // pse
                     | 1<<4      // rdtsc
                     | 1<<5      // msr
                     | 1<<6      // pae
+                    | 1<<7      // mcheck extension
                     | 1<<8      // cmpxchg8
                     | 1<<11     // sep (sysenter & sysexit)
                     | 1<<12     // mtrr
+                    | 1<<13     // pgb
+                    | 1<<14     // mcheck arch
                     | 1<<15     // cmov
                     | 1<<16     // pat
+                    | 1<<17     // pse36
                     | 1<<19     // clflush (seems to be with SSE2)
                     | 1<<21     // DS, used with VMX, is that usefull?
                     | 1<<23     // mmx
@@ -288,7 +293,9 @@ void my_cpuid(x64emu_t* emu, uint32_t tmp32u)
                     ;
             R_ECX =   1<<0      // SSE3
                     | 1<<1      // PCLMULQDQ
-                    | 1<<5      // VMX  //is that usefull
+                    | (box64_cputype?0:1)<<2      // DS 64bits
+                    | 1<<3      // Monitor/MWait (priviledge instructions)
+                    | (box64_cputype?0:1)<<5      // VMX  //is that usefull
                     | 1<<9      // SSSE3
                     | box64_avx2<<12     // fma
                     | 1<<13     // cx16 (cmpxchg16)
@@ -302,6 +309,7 @@ void my_cpuid(x64emu_t* emu, uint32_t tmp32u)
                     | box64_avx<<28 // AVX
                     | box64_avx<<29 // F16C
                     | box64_avx2<<30     // RDRAND
+                    | 0<<31     // Hypervisor guest running
                     ; 
             break;
         case 0x2:
@@ -375,13 +383,22 @@ void my_cpuid(x64emu_t* emu, uint32_t tmp32u)
                 R_EBX = 
                         box64_avx<<3 |  // BMI1 
                         box64_avx2<<5 |  //AVX2
+                        (box64_cputype?0:1)<<6 | // FDP_EXCPTN_ONLY
+                        1<<7 | // SMEP
                         box64_avx2<<8 | //BMI2
+                        (box64_cputype?0:1)<<9 |    // Enhanced REP MOVSB   // is it a good idea?
+                        1<<10 | //INVPCID (priviledge instruction
+                        (box64_cputype?0:1)<<13 | // Deprecates FPU CS and FPU DS
+                        0<<18 | // RDSEED
                         box64_avx2<<19 | //ADX
+                        1<<23 | // CLFLUSHOPT
+                        1<<24 | // CLWB
                         box64_shaext<<29|  // SHA extension
                         0;
                 R_RCX = 
                         box64_avx<<9   | //VAES
                         box64_avx2<<10 | //VPCLMULQDQ.
+                        1<<22 | // RDPID
                         0;
                 R_RDX = 0;
 
@@ -523,7 +540,8 @@ void my_cpuid(x64emu_t* emu, uint32_t tmp32u)
                         //| 1<<7      // misaligned SSE
                         | 1<<8      // 3DNowPrefetch
                         //| 1<<10     // IBS
-                        //| 1<<11     // SSE5
+                        //| 1<<11     // XOP
+                        //| 1<<16     // FMA4
                         ; 
                 R_EDX =   1         // fpu 
                         | 1<<2      // debugging extension
@@ -534,9 +552,12 @@ void my_cpuid(x64emu_t* emu, uint32_t tmp32u)
                         | 1<<8      // cmpxchg8
                         | 1<<11     // sep (sysenter & sysexit)
                         | 1<<12     // mtrr
+                        | 1<<13     // pge
                         | 1<<15     // cmov
                         | 1<<16     // pat
+                        | 1<<17     // pse36
                         | 1<<19     // clflush (seems to be with SSE2)
+                        | 1<<20     // NX
                         | 1<<21     // DS, used with VMX, is that usefull?
                         | 1<<22     // MMXext
                         | 1<<23     // mmx