[LA64_DYNAREC] Add la64 avx BMI2 shift ops. (#2807)

*  VEX.66.0f.38  SHLX
  *  VEX.f2.0f.38  SHRX
  *  VEX.f3.0f.38  SARX

5 files changed, 164 insertions, 5 deletions

diff --git a/src/dynarec/la64/dynarec_la64_avx.c b/src/dynarec/la64/dynarec_la64_avx.c
index 12e00db2..f395c80e 100644
--- a/src/dynarec/la64/dynarec_la64_avx.c
+++ b/src/dynarec/la64/dynarec_la64_avx.c
@@ -63,12 +63,12 @@ uintptr_t dynarec64_AVX(dynarec_la64_t* dyn, uintptr_t addr, uintptr_t ip, int n
         addr = dynarec64_AVX_66_0F38(dyn, addr, ip, ninst, vex, ok, need_epilog);
     else if ((vex.m == VEX_M_0F3A) && (vex.p == VEX_P_66))
         addr = dynarec64_AVX_66_0F3A(dyn, addr, ip, ninst, vex, ok, need_epilog);
-    // else if( (vex.m==VEX_M_0F38) && (vex.p==VEX_P_F2))
-    //     addr = dynarec64_AVX_F2_0F38(dyn, addr, ip, ninst, vex, ok, need_epilog);
+    else if( (vex.m == VEX_M_0F38) && (vex.p == VEX_P_F2))
+        addr = dynarec64_AVX_F2_0F38(dyn, addr, ip, ninst, vex, ok, need_epilog);
     else if ((vex.m == VEX_M_0F3A) && (vex.p == VEX_P_F2))
         addr = dynarec64_AVX_F2_0F3A(dyn, addr, ip, ninst, vex, ok, need_epilog);
-    // else if( (vex.m==VEX_M_0F38) && (vex.p==VEX_P_F3))
-    //     addr = dynarec64_AVX_F3_0F38(dyn, addr, ip, ninst, vex, ok, need_epilog);
+    else if( (vex.m==VEX_M_0F38) && (vex.p==VEX_P_F3))
+        addr = dynarec64_AVX_F3_0F38(dyn, addr, ip, ninst, vex, ok, need_epilog);
     else {
         DEFAULT;
     }
diff --git a/src/dynarec/la64/dynarec_la64_avx_66_0f38.c b/src/dynarec/la64/dynarec_la64_avx_66_0f38.c
index 48d4e4e1..e8f82715 100644
--- a/src/dynarec/la64/dynarec_la64_avx_66_0f38.c
+++ b/src/dynarec/la64/dynarec_la64_avx_66_0f38.c
@@ -27,7 +27,7 @@ uintptr_t dynarec64_AVX_66_0F38(dynarec_la64_t* dyn, uintptr_t addr, uintptr_t i
 
     uint8_t opcode = F8;
     uint8_t nextop, u8;
-    uint8_t gd, ed;
+    uint8_t gd, ed, vd;
     uint8_t wback, wb1, wb2;
     uint8_t eb1, eb2, gb1, gb2;
     int32_t i32, i32_;
@@ -354,6 +354,15 @@ uintptr_t dynarec64_AVX_66_0F38(dynarec_la64_t* dyn, uintptr_t addr, uintptr_t i
                 PUTEYx(v0);
             }
             break;
+        case 0xF7:
+            INST_NAME("SHLX Gd, Ed, Vd");
+            nextop = F8;
+            GETGD;
+            GETED(0);
+            GETVD;
+            ANDI(x5, vd, rex.w ? 0x3f : 0x1f);
+            SLLxw(gd, ed, x5);
+            break;
         default:
             DEFAULT;
     }
diff --git a/src/dynarec/la64/dynarec_la64_avx_f2_0f38.c b/src/dynarec/la64/dynarec_la64_avx_f2_0f38.c
new file mode 100644
index 00000000..f8a29542
--- /dev/null
+++ b/src/dynarec/la64/dynarec_la64_avx_f2_0f38.c
@@ -0,0 +1,73 @@
+#include <stdio.h>
+#include <stdlib.h>
+#include <stddef.h>
+#include <errno.h>
+
+#include "debug.h"
+#include "env.h"
+#include "box64context.h"
+#include "box64cpu.h"
+#include "emu/x64emu_private.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_la64_functions.h"
+#include "../dynarec_helper.h"
+
+uintptr_t dynarec64_AVX_F2_0F38(dynarec_la64_t* dyn, uintptr_t addr, uintptr_t ip, int ninst, vex_t vex, int* ok, int* need_epilog)
+{
+    (void)ip;
+    (void)need_epilog;
+
+    uint8_t opcode = F8;
+    uint8_t nextop, u8;
+    uint8_t gd, ed, vd;
+    uint8_t wback, wb1, wb2;
+    uint8_t eb1, eb2, gb1, gb2;
+    int32_t i32, i32_;
+    int cacheupd = 0;
+    int v0, v1, v2;
+    int q0, q1, q2;
+    int d0, d1, d2;
+    int s0;
+    uint64_t tmp64u;
+    int64_t j64;
+    int64_t fixedaddress;
+    int unscaled;
+    MAYUSE(wb1);
+    MAYUSE(wb2);
+    MAYUSE(eb1);
+    MAYUSE(eb2);
+    MAYUSE(gb1);
+    MAYUSE(gb2);
+    MAYUSE(q0);
+    MAYUSE(q1);
+    MAYUSE(d0);
+    MAYUSE(d1);
+    MAYUSE(s0);
+    MAYUSE(j64);
+    MAYUSE(cacheupd);
+
+    rex_t rex = vex.rex;
+
+    switch (opcode) {
+        case 0xF7:
+            INST_NAME("SHRX Gd, Ed, Vd");
+            nextop = F8;
+            GETGD;
+            GETED(0);
+            GETVD;
+            ANDI(x5, vd, rex.w ? 0x3f : 0x1f);
+            SRLxw(gd, ed, x5);
+            break;
+        default:
+            DEFAULT;
+    }
+    return addr;
+}
diff --git a/src/dynarec/la64/dynarec_la64_avx_f3_0f38.c b/src/dynarec/la64/dynarec_la64_avx_f3_0f38.c
new file mode 100644
index 00000000..8e8e6781
--- /dev/null
+++ b/src/dynarec/la64/dynarec_la64_avx_f3_0f38.c
@@ -0,0 +1,73 @@
+#include <stdio.h>
+#include <stdlib.h>
+#include <stddef.h>
+#include <errno.h>
+
+#include "debug.h"
+#include "env.h"
+#include "box64context.h"
+#include "box64cpu.h"
+#include "emu/x64emu_private.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_la64_functions.h"
+#include "../dynarec_helper.h"
+
+uintptr_t dynarec64_AVX_F3_0F38(dynarec_la64_t* dyn, uintptr_t addr, uintptr_t ip, int ninst, vex_t vex, int* ok, int* need_epilog)
+{
+    (void)ip;
+    (void)need_epilog;
+
+    uint8_t opcode = F8;
+    uint8_t nextop, u8;
+    uint8_t gd, ed, vd;
+    uint8_t wback, wb1, wb2;
+    uint8_t eb1, eb2, gb1, gb2;
+    int32_t i32, i32_;
+    int cacheupd = 0;
+    int v0, v1, v2;
+    int q0, q1, q2;
+    int d0, d1, d2;
+    int s0;
+    uint64_t tmp64u;
+    int64_t j64;
+    int64_t fixedaddress;
+    int unscaled;
+    MAYUSE(wb1);
+    MAYUSE(wb2);
+    MAYUSE(eb1);
+    MAYUSE(eb2);
+    MAYUSE(gb1);
+    MAYUSE(gb2);
+    MAYUSE(q0);
+    MAYUSE(q1);
+    MAYUSE(d0);
+    MAYUSE(d1);
+    MAYUSE(s0);
+    MAYUSE(j64);
+    MAYUSE(cacheupd);
+
+    rex_t rex = vex.rex;
+
+    switch (opcode) {
+        case 0xF7:
+            INST_NAME("SARX Gd, Ed, Vd");
+            nextop = F8;
+            GETGD;
+            GETED(0);
+            GETVD;
+            ANDI(x5, vd, rex.w ? 0x3f : 0x1f);
+            SRAxw(gd, ed, x5);
+            break;
+        default:
+            DEFAULT;
+    }
+    return addr;
+}
diff --git a/src/dynarec/la64/dynarec_la64_helper.h b/src/dynarec/la64/dynarec_la64_helper.h
index b6b646a8..acd22031 100644
--- a/src/dynarec/la64/dynarec_la64_helper.h
+++ b/src/dynarec/la64/dynarec_la64_helper.h
@@ -1100,8 +1100,10 @@
 #define dynarec64_AVX_66_0F38 STEPNAME(dynarec64_AVX_66_0F38)
 #define dynarec64_AVX_66_0F3A STEPNAME(dynarec64_AVX_66_0F3A)
 #define dynarec64_AVX_F2_0F   STEPNAME(dynarec64_AVX_F2_0F)
+#define dynarec64_AVX_F2_0F38 STEPNAME(dynarec64_AVX_F2_0F38)
 #define dynarec64_AVX_F2_0F3A STEPNAME(dynarec64_AVX_F2_0F3A)
 #define dynarec64_AVX_F3_0F   STEPNAME(dynarec64_AVX_F3_0F)
+#define dynarec64_AVX_F3_0F38 STEPNAME(dynarec64_AVX_F3_0F38)
 
 #define geted               STEPNAME(geted)
 #define geted32             STEPNAME(geted32)
@@ -1392,8 +1394,10 @@ uintptr_t dynarec64_AVX_66_0F(dynarec_la64_t* dyn, uintptr_t addr, uintptr_t ip,
 uintptr_t dynarec64_AVX_66_0F38(dynarec_la64_t* dyn, uintptr_t addr, uintptr_t ip, int ninst, vex_t vex, int* ok, int* need_epilog);
 uintptr_t dynarec64_AVX_66_0F3A(dynarec_la64_t* dyn, uintptr_t addr, uintptr_t ip, int ninst, vex_t vex, int* ok, int* need_epilog);
 uintptr_t dynarec64_AVX_F2_0F(dynarec_la64_t* dyn, uintptr_t addr, uintptr_t ip, int ninst, vex_t vex, int* ok, int* need_epilog);
+uintptr_t dynarec64_AVX_F2_0F38(dynarec_la64_t* dyn, uintptr_t addr, uintptr_t ip, int ninst, vex_t vex, int* ok, int* need_epilog);
 uintptr_t dynarec64_AVX_F2_0F3A(dynarec_la64_t* dyn, uintptr_t addr, uintptr_t ip, int ninst, vex_t vex, int* ok, int* need_epilog);
 uintptr_t dynarec64_AVX_F3_0F(dynarec_la64_t* dyn, uintptr_t addr, uintptr_t ip, int ninst, vex_t vex, int* ok, int* need_epilog);
+uintptr_t dynarec64_AVX_F3_0F38(dynarec_la64_t* dyn, uintptr_t addr, uintptr_t ip, int ninst, vex_t vex, int* ok, int* need_epilog);
 
 
 #if STEP < 3