[ARM64_DYNAREC] Improved 0F A4 opcode

3 files changed, 44 insertions, 8 deletions

diff --git a/src/dynarec/arm64/dynarec_arm64_emit_shift.c b/src/dynarec/arm64/dynarec_arm64_emit_shift.c
index 19343416..7c626eb1 100644
--- a/src/dynarec/arm64/dynarec_arm64_emit_shift.c
+++ b/src/dynarec/arm64/dynarec_arm64_emit_shift.c
@@ -1056,8 +1056,7 @@ void emit_shld32c(dynarec_arm_t* dyn, int ninst, rex_t rex, int s1, int s2, uint
         MOV32w(s3, c);
         STRxw_U12(s1, xEmu, offsetof(x64emu_t, op1));
         STRxw_U12(s3, xEmu, offsetof(x64emu_t, op2));
-        // same flags computation as with shl64/shl32
-        SET_DF(s4, rex.w?d_shl64:d_shl32);
+        SET_DF(s4, rex.w?d_shld64:d_shld32);
     } else IFX(X_ALL) {
         SET_DFNONE(s4);
     }
@@ -1070,12 +1069,15 @@ void emit_shld32c(dynarec_arm_t* dyn, int ninst, rex_t rex, int s1, int s2, uint
         }
         return;
     }
-    IFX(X_CF|X_OF) {
+    IFX(X_CF) {
         LSRxw(s3, s1, (rex.w?64:32)-c);
         BFIxw(xFlags, s3, F_CF, 1);
     }
-    LSLxw(s3, s1, c);
-    ORRxw_REG_LSR(s1, s3, s2, (rex.w?64:32)-c);
+    IFX(X_OF) {
+        LSRxw(s3, s1, rex.w?63:31);
+        BFIw(xFlags, s3, F_OF, 1);  // store current sign for later use
+    }
+    EXTRxw(s1, s1, s2, (rex.w?64:32)-c);
 
     IFX(X_PEND) {
         STRxw_U12(s1, xEmu, offsetof(x64emu_t, res));
@@ -1091,8 +1093,8 @@ void emit_shld32c(dynarec_arm_t* dyn, int ninst, rex_t rex, int s1, int s2, uint
     }
     IFX(X_OF) {
         if(c==1) {
-            UBFXxw(s3, s1, rex.w?63:31, 1);
-            EORxw_REG(s3, s3, xFlags);  // CF is set if OF is asked
+            LSRxw(s3, s1, rex.w?63:31);
+            EORxw_REG_LSR(s3, s3, xFlags, F_OF);  // OF is set if sign changed
             BFIw(xFlags, s3, F_OF, 1);
         } else {
             BFCw(xFlags, F_OF, 1);
diff --git a/src/emu/x64run_private.c b/src/emu/x64run_private.c
index a5a588c2..1eaa6301 100644
--- a/src/emu/x64run_private.c
+++ b/src/emu/x64run_private.c
@@ -731,6 +731,36 @@ void UpdateFlags(x64emu_t *emu)
                 }
             }
             break;
+        case d_shld32:
+            cnt = emu->op2.u32;
+            if (cnt > 0) {
+                cc = emu->op1.u32 & (1 << (32 - cnt));
+                CONDITIONAL_SET_FLAG(cc, F_CF);
+                CONDITIONAL_SET_FLAG(!emu->res.u32, F_ZF);
+                CONDITIONAL_SET_FLAG(emu->res.u32 & 0x80000000, F_SF);
+                CONDITIONAL_SET_FLAG(PARITY(emu->res.u32 & 0xff), F_PF);
+                if (cnt == 1) {
+                    CONDITIONAL_SET_FLAG((emu->op1.u32 ^ emu->res.u32) & 0x80000000, F_OF);
+                } else {
+                    CLEAR_FLAG(F_OF);
+                }
+            }
+            break;
+        case d_shld64:
+            cnt = emu->op2.u64;
+            if (cnt > 0) {
+                cc = emu->op1.u64 & (1LL << (64 - cnt));
+                CONDITIONAL_SET_FLAG(cc, F_CF);
+                CONDITIONAL_SET_FLAG(!emu->res.u64, F_ZF);
+                CONDITIONAL_SET_FLAG(emu->res.u64 & 0x8000000000000000LL, F_SF);
+                CONDITIONAL_SET_FLAG(PARITY(emu->res.u64 & 0xff), F_PF);
+                if (cnt == 1) {
+                    CONDITIONAL_SET_FLAG((emu->op1.u64 ^ emu->res.u64) & 0x8000000000000000LL, F_OF);
+                } else {
+                    CLEAR_FLAG(F_OF);
+                }
+            }
+            break;
         case d_sub8:
             CONDITIONAL_SET_FLAG(emu->res.u8 & 0x80, F_SF);
             CONDITIONAL_SET_FLAG((emu->res.u8 & 0xff) == 0, F_ZF);
@@ -1014,6 +1044,7 @@ void UpdateFlags(x64emu_t *emu)
         case d_rcr16:
         case d_rcr32:
         case d_rcr64:
+        case d_shld16:
         case d_unknown:
             printf_log(LOG_NONE, "Box64: %p trying to evaluate Unknown deferred Flags\n", (void*)R_RIP);
             break;
diff --git a/src/include/regs.h b/src/include/regs.h
index 433b352c..dc72a648 100644
--- a/src/include/regs.h
+++ b/src/include/regs.h
@@ -142,7 +142,10 @@ typedef enum {
 	d_shrd16,
 	d_shrd32,
 	d_shrd64,
-	d_unknown	//98
+	d_shld16,
+	d_shld32,
+	d_shld64,
+	d_unknown	//101
 } deferred_flags_t;
 
 #pragma pack(push, 1)