[LA64_DYNAREC] Added more opcodes (#1481)

5 files changed, 171 insertions, 0 deletions

diff --git a/src/dynarec/la64/dynarec_la64_00.c b/src/dynarec/la64/dynarec_la64_00.c
index 47a0548d..f934c5f3 100644
--- a/src/dynarec/la64/dynarec_la64_00.c
+++ b/src/dynarec/la64/dynarec_la64_00.c
@@ -1403,6 +1403,14 @@ uintptr_t dynarec64_00(dynarec_la64_t* dyn, uintptr_t addr, uintptr_t ip, int ni
         case 0xD1:
             nextop = F8;
             switch ((nextop >> 3) & 7) {
+                case 4:
+                case 6:
+                    INST_NAME("SHL Ed, 1");
+                    SETFLAGS(X_ALL, SF_SET_PENDING); // some flags are left undefined
+                    GETED(0);
+                    emit_shl32c(dyn, ninst, rex, ed, 1, x3, x4, x5);
+                    WBACK;
+                    break;
                 case 5:
                     INST_NAME("SHR Ed, 1");
                     SETFLAGS(X_ALL, SF_SET_PENDING); // some flags are left undefined
@@ -1695,6 +1703,23 @@ uintptr_t dynarec64_00(dynarec_la64_t* dyn, uintptr_t addr, uintptr_t ip, int ni
                     MOV32w(x2, u8);
                     emit_test8(dyn, ninst, x1, x2, x3, x4, x5);
                     break;
+                case 3:
+                    INST_NAME("NEG Eb");
+                    SETFLAGS(X_ALL, SF_SET_PENDING);
+                    GETEB(x1, 0);
+                    emit_neg8(dyn, ninst, x1, x2, x4);
+                    EBBACK();
+                    break;
+                case 4:
+                    INST_NAME("MUL AL, Ed");
+                    SETFLAGS(X_ALL, SF_PENDING);
+                    UFLAG_DF(x1, d_mul8);
+                    GETEB(x1, 0);
+                    ANDI(x2, xRAX, 0xff);
+                    MUL_W(x1, x2, x1);
+                    UFLAG_RES(x1);
+                    BSTRINS_D(xRAX, x1, 15, 0);
+                    break;
                 default:
                     DEFAULT;
             }
diff --git a/src/dynarec/la64/dynarec_la64_0f.c b/src/dynarec/la64/dynarec_la64_0f.c
index 65bf7881..731e0f72 100644
--- a/src/dynarec/la64/dynarec_la64_0f.c
+++ b/src/dynarec/la64/dynarec_la64_0f.c
@@ -8,6 +8,7 @@
 #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"
@@ -118,6 +119,21 @@ uintptr_t dynarec64_0F(dynarec_la64_t* dyn, uintptr_t addr, uintptr_t ip, int ni
                 SMWRITE2();
             }
             break;
+        case 0x12:
+            nextop = F8;
+            if (MODREG) {
+                INST_NAME("MOVHLPS Gx,Ex");
+                GETGX(v0, 1);
+                v1 = sse_get_reg(dyn, ninst, x1, (nextop & 7) + (rex.b << 3), 0);
+                VEXTRINS_D(v0, v1, 1);
+            } else {
+                INST_NAME("MOVLPS Gx,Ex");
+                GETGX(v0, 1);
+                SMREAD();
+                GETEX(q1, 0, 0);
+                VEXTRINS_D(v0, q1, 0);
+            }
+            break;
         case 0x16:
             nextop = F8;
             if (MODREG) {
diff --git a/src/dynarec/la64/dynarec_la64_66.c b/src/dynarec/la64/dynarec_la64_66.c
index 626e33d7..0e0a3db0 100644
--- a/src/dynarec/la64/dynarec_la64_66.c
+++ b/src/dynarec/la64/dynarec_la64_66.c
@@ -65,6 +65,16 @@ uintptr_t dynarec64_66(dynarec_la64_t* dyn, uintptr_t addr, uintptr_t ip, int ni
                     DEFAULT;
             }
             break;
+        case 0x19:
+            INST_NAME("SBB Ew, Gw");
+            READFLAGS(X_CF);
+            SETFLAGS(X_ALL, SF_SET_PENDING);
+            nextop = F8;
+            GETGW(x2);
+            GETEW(x1, 0);
+            emit_sbb16(dyn, ninst, x1, x2, x4, x5, x6);
+            EWBACK;
+            break;
         case 0x39:
             INST_NAME("CMP Ew, Gw");
             SETFLAGS(X_ALL, SF_SET_PENDING);
diff --git a/src/dynarec/la64/dynarec_la64_emit_math.c b/src/dynarec/la64/dynarec_la64_emit_math.c
index 23a2acf9..d9958f27 100644
--- a/src/dynarec/la64/dynarec_la64_emit_math.c
+++ b/src/dynarec/la64/dynarec_la64_emit_math.c
@@ -8,6 +8,7 @@
 #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"
@@ -770,6 +771,59 @@ void emit_sbb8c(dynarec_la64_t* dyn, int ninst, int s1, int c, int s3, int s4, i
     emit_sbb8(dyn, ninst, s1, s6, s3, s4, s5);
 }
 
+// emit SBB16 instruction, from s1, s2, store result in s1 using s3, s4 and s5 as scratch
+void emit_sbb16(dynarec_la64_t* dyn, int ninst, int s1, int s2, int s3, int s4, int s5)
+{
+    IFX (X_PEND) {
+        ST_H(s1, xEmu, offsetof(x64emu_t, op1));
+        ST_H(s2, xEmu, offsetof(x64emu_t, op2));
+        SET_DF(s3, d_sbb16);
+    } else IFX (X_ALL) {
+        SET_DFNONE();
+    }
+
+    IFXA (X_ALL, la64_lbt) {
+        SBC_H(s3, s1, s2);
+
+        IFX (X_ALL) {
+            X64_SBC_H(s1, s2);
+        }
+        BSTRPICK_D(s1, s3, 15, 0);
+        IFX (X_PEND)
+            ST_H(s1, xEmu, offsetof(x64emu_t, res));
+        return;
+    }
+
+    IFX (X_AF | X_CF | X_OF) {
+        // for later flag calculation
+        NOR(s5, xZR, s1);
+    }
+
+    SUB_W(s1, s1, s2);
+    ANDI(s3, xFlags, 1 << F_CF);
+    SUB_W(s1, s1, s3);
+
+    CLEAR_FLAGS(s3);
+    SLLI_W(s1, s1, 16);
+    IFX (X_SF) {
+        BGE(s1, xZR, 8);
+        ORI(xFlags, xFlags, 1 << F_SF);
+    }
+    SRLI_W(s1, s1, 16);
+
+    IFX (X_PEND) {
+        ST_H(s1, xEmu, offsetof(x64emu_t, res));
+    }
+
+    CALC_SUB_FLAGS(s5, s2, s1, s3, s4, 16);
+    IFX (X_ZF) {
+        BNEZ(s1, 8);
+        ORI(xFlags, xFlags, 1 << F_ZF);
+    }
+    IFX (X_PF) {
+        emit_pf(dyn, ninst, s1, s3, s4);
+    }
+}
 
 // emit SBB32 instruction, from s1, s2, store result in s1 using s3, s4 and s5 as scratch
 void emit_sbb32(dynarec_la64_t* dyn, int ninst, rex_t rex, int s1, int s2, int s3, int s4, int s5)
@@ -834,6 +888,68 @@ void emit_sbb32(dynarec_la64_t* dyn, int ninst, rex_t rex, int s1, int s2, int s
 }
 
 
+// emit NEG8 instruction, from s1, store result in s1 using s2 and s3 as scratch
+void emit_neg8(dynarec_la64_t* dyn, int ninst, int s1, int s2, int s3)
+{
+    IFX (X_PEND) {
+        ST_B(s1, xEmu, offsetof(x64emu_t, op1));
+        SET_DF(s3, d_neg8);
+    } else IFX (X_ALL) {
+        SET_DFNONE();
+    }
+    IFX (X_AF | X_OF) {
+        MV(s3, s1); // s3 = op1
+    }
+
+    IFXA (X_ALL, la64_lbt) {
+        X64_SUB_B(xZR, s1);
+    }
+
+    NEG_D(s1, s1);
+    ANDI(s1, s1, 0xff);
+    IFX (X_PEND) {
+        ST_B(s1, xEmu, offsetof(x64emu_t, res));
+    }
+
+    if (la64_lbt) return;
+
+    CLEAR_FLAGS(s2);
+    IFX (X_CF) {
+        BEQZ(s1, 8);
+        ORI(xFlags, xFlags, 1 << F_CF);
+    }
+
+    IFX (X_AF | X_OF) {
+        OR(s3, s1, s3); // s3 = res | op1
+        IFX (X_AF) {
+            /* af = bc & 0x8 */
+            ANDI(s2, s3, 8);
+            BEQZ(s2, 8);
+            ORI(xFlags, xFlags, 1 << F_AF);
+        }
+        IFX (X_OF) {
+            /* of = ((bc >> (width-2)) ^ (bc >> (width-1))) & 0x1; */
+            SRLI_D(s2, s3, 6);
+            SRLI_D(s3, s2, 1);
+            XOR(s2, s2, s3);
+            ANDI(s2, s2, 1);
+            BEQZ(s2, 8);
+            ORI(xFlags, xFlags, 1 << F_OF);
+        }
+    }
+    IFX (X_SF) {
+        ANDI(s3, s1, 1 << F_SF); // 1<<F_SF is sign bit, so just mask
+        OR(xFlags, xFlags, s3);
+    }
+    IFX (X_PF) {
+        emit_pf(dyn, ninst, s1, s3, s2);
+    }
+    IFX (X_ZF) {
+        BNEZ(s1, 8);
+        ORI(xFlags, xFlags, 1 << F_ZF);
+    }
+}
+
 // emit NEG32 instruction, from s1, store result in s1 using s2 and s3 as scratch
 void emit_neg32(dynarec_la64_t* dyn, int ninst, rex_t rex, int s1, int s2, int s3)
 {
diff --git a/src/dynarec/la64/dynarec_la64_helper.h b/src/dynarec/la64/dynarec_la64_helper.h
index 50405c1c..2ac88693 100644
--- a/src/dynarec/la64/dynarec_la64_helper.h
+++ b/src/dynarec/la64/dynarec_la64_helper.h
@@ -738,7 +738,9 @@ void* la64_next(x64emu_t* emu, uintptr_t addr);
 #define emit_sub8c          STEPNAME(emit_sub8c)
 #define emit_sbb8           STEPNAME(emit_sbb8)
 #define emit_sbb8c          STEPNAME(emit_sbb8c)
+#define emit_sbb16          STEPNAME(emit_sbb16)
 #define emit_sbb32          STEPNAME(emit_sbb32)
+#define emit_neg8           STEPNAME(emit_neg8)
 #define emit_neg32          STEPNAME(emit_neg32)
 #define emit_or32           STEPNAME(emit_or32)
 #define emit_or32c          STEPNAME(emit_or32c)
@@ -816,7 +818,9 @@ void emit_sub8(dynarec_la64_t* dyn, int ninst, int s1, int s2, int s3, int s4, i
 void emit_sub8c(dynarec_la64_t* dyn, int ninst, int s1, int32_t c, int s2, int s3, int s4, int s5);
 void emit_sbb8(dynarec_la64_t* dyn, int ninst, int s1, int s2, int s3, int s4, int s5);
 void emit_sbb8c(dynarec_la64_t* dyn, int ninst, int s1, int32_t c, int s3, int s4, int s5, int s6);
+void emit_sbb16(dynarec_la64_t* dyn, int ninst, int s1, int s2, int s3, int s4, int s5);
 void emit_sbb32(dynarec_la64_t* dyn, int ninst, rex_t rex, int s1, int s2, int s3, int s4, int s5);
+void emit_neg8(dynarec_la64_t* dyn, int ninst, int s1, int s3, int s4);
 void emit_neg32(dynarec_la64_t* dyn, int ninst, rex_t rex, int s1, int s2, int s3);
 void emit_or32(dynarec_la64_t* dyn, int ninst, rex_t rex, int s1, int s2, int s3, int s4);
 void emit_or32c(dynarec_la64_t* dyn, int ninst, rex_t rex, int s1, int64_t c, int s3, int s4);