[LA64_DYNAREC] Added more opcodes (#2339)

6 files changed, 204 insertions, 2 deletions

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 711dd035..23a9eaa4 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -1044,11 +1044,12 @@ if(LARCH64_DYNAREC)
     "${BOX64_ROOT}/src/dynarec/la64/dynarec_la64_64.c"
     "${BOX64_ROOT}/src/dynarec/la64/dynarec_la64_66.c"
     "${BOX64_ROOT}/src/dynarec/la64/dynarec_la64_6664.c"
-    "${BOX64_ROOT}/src/dynarec/la64/dynarec_la64_67.c"
-    "${BOX64_ROOT}/src/dynarec/la64/dynarec_la64_f30f.c"
     "${BOX64_ROOT}/src/dynarec/la64/dynarec_la64_660f.c"
+    "${BOX64_ROOT}/src/dynarec/la64/dynarec_la64_66f0.c"
+    "${BOX64_ROOT}/src/dynarec/la64/dynarec_la64_67.c"
     "${BOX64_ROOT}/src/dynarec/la64/dynarec_la64_f0.c"
     "${BOX64_ROOT}/src/dynarec/la64/dynarec_la64_f20f.c"
+    "${BOX64_ROOT}/src/dynarec/la64/dynarec_la64_f30f.c"
     )
 endif()
 
diff --git a/src/dynarec/la64/dynarec_la64_0f.c b/src/dynarec/la64/dynarec_la64_0f.c
index e0366227..192c9445 100644
--- a/src/dynarec/la64/dynarec_la64_0f.c
+++ b/src/dynarec/la64/dynarec_la64_0f.c
@@ -416,6 +416,32 @@ uintptr_t dynarec64_0F(dynarec_la64_t* dyn, uintptr_t addr, uintptr_t ip, int ni
                             break;
                     }
                     break;
+                case 0xF0:
+                    INST_NAME("MOVBE Gd, Ed");
+                    nextop = F8;
+                    GETGD;
+                    SMREAD();
+                    addr = geted(dyn, addr, ninst, nextop, &ed, x2, x1, &fixedaddress, rex, NULL, 1, 0);
+                    LDxw(gd, ed, fixedaddress);
+                    if (rex.w) {
+                        REVB_D(gd, gd);
+                    } else {
+                        REVB_2W(gd, gd);
+                        ZEROUP(gd);
+                    }
+                    break;
+                case 0xF1:
+                    INST_NAME("MOVBE Ed, Gd");
+                    nextop = F8;
+                    GETGD;
+                    SMREAD();
+                    addr = geted(dyn, addr, ninst, nextop, &wback, x2, x1, &fixedaddress, rex, NULL, 1, 0);
+                    if (rex.w)
+                        REVB_D(x1, gd);
+                    else
+                        REVB_2W(x1, gd);
+                    SDxw(x1, wback, fixedaddress);
+                    break;
                 default:
                     DEFAULT;
             }
diff --git a/src/dynarec/la64/dynarec_la64_66.c b/src/dynarec/la64/dynarec_la64_66.c
index 548f01bf..8d3ae2e5 100644
--- a/src/dynarec/la64/dynarec_la64_66.c
+++ b/src/dynarec/la64/dynarec_la64_66.c
@@ -859,6 +859,10 @@ uintptr_t dynarec64_66(dynarec_la64_t* dyn, uintptr_t addr, uintptr_t ip, int ni
                     DEFAULT;
             }
             break;
+
+        case 0xF0:
+            return dynarec64_66F0(dyn, addr, ip, ninst, rex, rep, ok, need_epilog);
+
         case 0xF7:
             nextop = F8;
             switch ((nextop >> 3) & 7) {
diff --git a/src/dynarec/la64/dynarec_la64_66f0.c b/src/dynarec/la64/dynarec_la64_66f0.c
new file mode 100644
index 00000000..e2c74406
--- /dev/null
+++ b/src/dynarec/la64/dynarec_la64_66f0.c
@@ -0,0 +1,132 @@
+#include <stdio.h>
+#include <stdlib.h>
+#include <stddef.h>
+#include <errno.h>
+
+#include "debug.h"
+#include "box64context.h"
+#include "dynarec.h"
+#include "emu/x64emu_private.h"
+#include "emu/x64run_private.h"
+#include "la64_emitter.h"
+#include "x64run.h"
+#include "x64emu.h"
+#include "box64stack.h"
+#include "callback.h"
+#include "emu/x64run_private.h"
+#include "x64trace.h"
+#include "dynarec_native.h"
+
+#include "la64_printer.h"
+#include "dynarec_la64_private.h"
+#include "../dynarec_helper.h"
+#include "dynarec_la64_functions.h"
+
+
+uintptr_t dynarec64_66F0(dynarec_la64_t* dyn, uintptr_t addr, uintptr_t ip, int ninst, rex_t rex, int rep, int* ok, int* need_epilog)
+{
+    (void)ip;
+    (void)rep;
+    (void)need_epilog;
+
+    uint8_t opcode = F8;
+    uint8_t nextop;
+    uint8_t gd, ed, u8;
+    uint8_t wback, wb1, wb2, gb1, gb2;
+    int32_t i32;
+    int64_t i64, j64;
+    uint64_t u64;
+    int64_t fixedaddress;
+    int unscaled;
+    MAYUSE(gb1);
+    MAYUSE(gb2);
+    MAYUSE(wb1);
+    MAYUSE(wb2);
+    MAYUSE(j64);
+
+    while ((opcode == 0xF2) || (opcode == 0xF3)) {
+        rep = opcode - 0xF1;
+        opcode = F8;
+    }
+
+    GETREX();
+
+    switch (opcode) {
+        case 0x81:
+        case 0x83:
+            nextop = F8;
+            SMDMB();
+            switch ((nextop >> 3) & 7) {
+                case 0: // ADD
+                    if (opcode == 0x81) {
+                        INST_NAME("LOCK ADD Ew, Iw");
+                    } else {
+                        INST_NAME("LOCK ADD Ew, Ib");
+                    }
+                    SETFLAGS(X_ALL, SF_SET_PENDING, NAT_FLAGS_FUSION);
+                    if (MODREG) {
+                        if (opcode == 0x81)
+                            u64 = F16;
+                        else
+                            u64 = (uint16_t)(int16_t)F8S;
+                        ed = TO_NAT((nextop & 7) + (rex.b << 3));
+                        MOV64x(x5, u64);
+                        BSTRPICK_D(x6, ed, 15, 0);
+                        emit_add16(dyn, ninst, x6, x5, x3, x4, x2);
+                        BSTRINS_D(ed, x6, 15, 0);
+                    } else {
+                        addr = geted(dyn, addr, ninst, nextop, &wback, x2, x1, &fixedaddress, rex, LOCK_LOCK, 0, (opcode == 0x81) ? 2 : 1);
+                        if (opcode == 0x81)
+                            u64 = F16;
+                        else
+                            u64 = (uint16_t)(int16_t)F8S;
+                        MOV64x(x5, u64);
+
+                        if (la64_lam_bh) {
+                            AMADD_DB_H(x1, x5, wback);
+                        } else {
+                            ANDI(x3, wback, 0b10);
+                            BNEZ_MARK(x3);
+                            // lower 16 bits
+                            MARKLOCK;
+                            LL_W(x1, wback, 0);
+                            ADD_D(x4, x1, x5);
+                            MV(x3, x1);
+                            BSTRINS_D(x3, x4, 15, 0);
+                            SC_W(x3, wback, 0);
+                            BEQZ_MARKLOCK(x3);
+                            IFXORNAT (X_ALL | X_PEND) {
+                                BSTRPICK_D(x1, x1, 15, 0);
+                            }
+                            B_MARK3_nocond;
+                            MARK;
+                            // upper 16 bits
+                            XORI(wback, wback, 0b10);
+                            MARK2;
+                            LL_W(x1, wback, 0);
+                            BSTRPICK_D(x3, x1, 15, 0);
+                            SRLI_W(x1, x1, 16);
+                            ADD_D(x4, x1, x5);
+                            BSTRINS_D(x3, x4, 31, 15);
+                            SC_W(x3, wback, 0);
+                            BEQZ_MARK2(x3);
+                            MARK3;
+                        }
+                        // final
+                        IFXORNAT (X_ALL | X_PEND) {
+                            emit_add16(dyn, ninst, x1, x5, x3, x4, x6);
+                        }
+                    }
+                    break;
+                default:
+                    DEFAULT;
+            }
+            SMDMB();
+            break;
+
+        default:
+            DEFAULT;
+    }
+
+    return addr;
+}
diff --git a/src/dynarec/la64/dynarec_la64_f0.c b/src/dynarec/la64/dynarec_la64_f0.c
index 4ff48729..d4786552 100644
--- a/src/dynarec/la64/dynarec_la64_f0.c
+++ b/src/dynarec/la64/dynarec_la64_f0.c
@@ -462,6 +462,43 @@ uintptr_t dynarec64_F0(dynarec_la64_t* dyn, uintptr_t addr, uintptr_t ip, int ni
             }
             SMDMB();
             break;
+
+        case 0x66:
+            return dynarec64_66F0(dyn, addr, ip, ninst, rex, rep, ok, need_epilog);
+
+        case 0x80:
+            nextop = F8;
+            SMDMB();
+            switch ((nextop >> 3) & 7) {
+                case 1: // OR
+                    INST_NAME("LOCK OR Eb, Ib");
+                    SETFLAGS(X_ALL, SF_SET_PENDING, NAT_FLAGS_FUSION);
+                    if (MODREG) {
+                        GETEB(x1, 1);
+                        u8 = F8;
+                        emit_or8c(dyn, ninst, x1, u8, x2, x4, x5);
+                        EBBACK();
+                    } else {
+                        addr = geted(dyn, addr, ninst, nextop, &wback, x5, x1, &fixedaddress, rex, LOCK_LOCK, 0, 1);
+                        u8 = F8;
+                        ANDI(x2, wback, 3);
+                        SLLI_D(x2, x2, 3);   // offset in bits
+                        ANDI(x3, wback, ~3); // aligned addr
+                        ADDI_D(x1, xZR, u8);
+                        SLL_D(x1, x1, x2); // Ib << offset
+                        AMOR_DB_W(x4, x1, x3);
+                        IFXORNAT (X_ALL | X_PEND) {
+                            SRL_D(x1, x4, x2);
+                            ANDI(x1, x1, 0xFF);
+                            emit_or8c(dyn, ninst, x1, u8, x2, x4, x5);
+                        }
+                    }
+                    break;
+                default:
+                    DEFAULT;
+            }
+            SMDMB();
+            break;
         case 0x81:
         case 0x83:
             nextop = F8;
diff --git a/src/dynarec/la64/dynarec_la64_helper.h b/src/dynarec/la64/dynarec_la64_helper.h
index 09d6cad5..eaa385cb 100644
--- a/src/dynarec/la64/dynarec_la64_helper.h
+++ b/src/dynarec/la64/dynarec_la64_helper.h
@@ -843,6 +843,7 @@ void* la64_next(x64emu_t* emu, uintptr_t addr);
 #define dynarec64_67   STEPNAME(dynarec64_67)
 #define dynarec64_F30F STEPNAME(dynarec64_F30F)
 #define dynarec64_660F STEPNAME(dynarec64_660F)
+#define dynarec64_66F0 STEPNAME(dynarec64_66F0)
 #define dynarec64_F0   STEPNAME(dynarec64_F0)
 #define dynarec64_F20F STEPNAME(dynarec64_F20F)
 
@@ -1091,6 +1092,7 @@ uintptr_t dynarec64_66(dynarec_la64_t* dyn, uintptr_t addr, uintptr_t ip, int ni
 uintptr_t dynarec64_6664(dynarec_la64_t* dyn, uintptr_t addr, uintptr_t ip, int ninst, rex_t rex, int seg, int* ok, int* need_epilog);
 uintptr_t dynarec64_67(dynarec_la64_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_la64_t* dyn, uintptr_t addr, uintptr_t ip, int ninst, rex_t rex, int* ok, int* need_epilog);
+uintptr_t dynarec64_66F0(dynarec_la64_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_la64_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_la64_t* dyn, uintptr_t addr, uintptr_t ip, int ninst, rex_t rex, int* ok, int* need_epilog);