[LA64_DYNAREC] Added some missing SM=2 store hints (#3314)

Author: ksco

src/dynarec/la64/dynarec_la64_00.c

@@ -2589,13 +2589,15 @@ uintptr_t dynarec64_00(dynarec_la64_t* dyn, uintptr_t addr, uintptr_t ip, int ni
                         tmp = 0; // TODO: removed when FP is in place
                     if ((BOX64ENV(log) < 2 && !BOX64ENV(rolling_log)) && tmp) {
                         call_n(dyn, ninst, (void*)(addr + 8), tmp);
+                        SMWRITE2();
                         addr += 8 + 8;
                     } else {
                         GETIP(ip + 1, x7); // read the 0xCC
                         STORE_XEMU_CALL();
                         ADDI_D(x3, xRIP, 8 + 8 + 2);                        // expected return address
                         ADDI_D(x1, xEmu, (uint32_t)offsetof(x64emu_t, ip)); // setup addr as &emu->ip
                         CALL_(const_int3, -1, x3, x1, 0);
+                        SMWRITE2();
                         LOAD_XEMU_CALL();
                         addr += 8 + 8;
                         BNE_MARK(xRIP, x3);
@@ -3245,13 +3247,15 @@ uintptr_t dynarec64_00(dynarec_la64_t* dyn, uintptr_t addr, uintptr_t ip, int ni
                         x87_purgecache(dyn, ninst, 0, x3, x1, x4);
                     if ((BOX64ENV(log) < 2 && !BOX64ENV(rolling_log)) && dyn->insts[ninst].natcall && tmp) {
                         call_n(dyn, ninst, (void*)(dyn->insts[ninst].natcall + 2 + 8), tmp);
+                        SMWRITE2();
                         POP1(xRIP); // pop the return address
                         dyn->last_ip = addr;
                     } else {
                         GETIP_(dyn->insts[ninst].natcall, x7); // read the 0xCC already
                         STORE_XEMU_CALL();
                         ADDI_D(x1, xEmu, (uint32_t)offsetof(x64emu_t, ip)); // setup addr as &emu->ip
                         CALL_S(const_int3, -1, x1);
+                        SMWRITE2();
                         LOAD_XEMU_CALL();
                         MOV64x(x3, dyn->insts[ninst].natcall);
                         ADDI_D(x3, x3, 2 + 8 + 8);

src/dynarec/la64/dynarec_la64_0f.c

@@ -709,6 +709,7 @@ uintptr_t dynarec64_0F(dynarec_la64_t* dyn, uintptr_t addr, uintptr_t ip, int ni
                     else
                         REVB_2W(x1, gd);
                     SDxw(x1, wback, fixedaddress);
+                    SMWRITE2();
                     break;
                 default:
                     DEFAULT;

src/dynarec/la64/dynarec_la64_660f.c

@@ -1120,6 +1120,7 @@ uintptr_t dynarec64_660F(dynarec_la64_t* dyn, uintptr_t addr, uintptr_t ip, int
                         u8 = (F8) & 15;
                         VPICKVE2GR_BU(x1, q0, u8);
                         ST_B(x1, wback, fixedaddress);
+                        SMWRITE2();
                     }
                     break;
                 case 0x15:
@@ -1136,6 +1137,7 @@ uintptr_t dynarec64_660F(dynarec_la64_t* dyn, uintptr_t addr, uintptr_t ip, int
                         u8 = (F8) & 7;
                         VPICKVE2GR_HU(x1, q0, u8);
                         ST_H(x1, wback, fixedaddress);
+                        SMWRITE2();
                     }
                     break;
                 case 0x16:

src/dynarec/la64/dynarec_la64_avx_66_0f3a.c

@@ -387,6 +387,7 @@ uintptr_t dynarec64_AVX_66_0F3A(dynarec_la64_t* dyn, uintptr_t addr, uintptr_t i
                 addr = geted(dyn, addr, ninst, nextop, &wback, x2, x4, &fixedaddress, rex, NULL, 0, 1);
                 u8 = (F8) & 15;
                 VSTELM_B(q0, wback, 0, u8);
+                SMWRITE2();
             }
             break;
         case 0x15:
@@ -402,6 +403,7 @@ uintptr_t dynarec64_AVX_66_0F3A(dynarec_la64_t* dyn, uintptr_t addr, uintptr_t i
                 addr = geted(dyn, addr, ninst, nextop, &wback, x2, x4, &fixedaddress, rex, NULL, 0, 1);
                 u8 = (F8) & 7;
                 VSTELM_H(q0, wback, 0, u8);
+                SMWRITE2();
             }
             break;
         case 0x16: