[DYNAREC] Fixed some oversized memory load (#3051)

* [DYNAREC] Fixed some oversized memory load

* review

7 files changed, 48 insertions, 14 deletions

diff --git a/src/dynarec/arm64/dynarec_arm64_0f.c b/src/dynarec/arm64/dynarec_arm64_0f.c
index d56ea3d5..4e79d609 100644
--- a/src/dynarec/arm64/dynarec_arm64_0f.c
+++ b/src/dynarec/arm64/dynarec_arm64_0f.c
@@ -462,7 +462,7 @@ uintptr_t dynarec64_0F(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
             INST_NAME("CVTTPS2PI Gm,Ex");

             nextop = F8;

             GETGM(q0);

-            GETEX(v1, 0, 0);

+            GETEXSD(v1, 0, 0);

             if (BOX64ENV(dynarec_fastround)) {

                 VFCVTZSS(q0, v1);

             } else {

@@ -496,7 +496,7 @@ uintptr_t dynarec64_0F(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
             INST_NAME("CVTPS2PI Gm, Ex");

             nextop = F8;

             GETGM(q0);

-            GETEX(v1, 0, 0);

+            GETEXSD(v1, 0, 0);

             if (BOX64ENV(dynarec_fastround)) {

                 u8 = sse_setround(dyn, ninst, x1, x2, x3);

                 VFRINTIS(q0, v1);

@@ -1220,7 +1220,7 @@ uintptr_t dynarec64_0F(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
         case 0x5A:

             INST_NAME("CVTPS2PD Gx, Ex");

             nextop = F8;

-            GETEX(q0, 0, 0);

+            GETEXSD(q0, 0, 0);

             GETGX_empty(q1);

             FCVTL(q1, q0);

             break;

diff --git a/src/dynarec/arm64/dynarec_arm64_660f.c b/src/dynarec/arm64/dynarec_arm64_660f.c
index e486161f..6d556154 100644
--- a/src/dynarec/arm64/dynarec_arm64_660f.c
+++ b/src/dynarec/arm64/dynarec_arm64_660f.c
@@ -1154,7 +1154,15 @@ uintptr_t dynarec64_660F(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int n
                     INST_NAME("PINSRB Gx, ED, Ib");

                     nextop = F8;

                     GETGX(q0, 1);

-                    GETED(1);

+                    if (MODREG) {

+                        ed = TO_NAT((nextop & 7) + (rex.b << 3));

+                        wback = 0;

+                    } else {

+                        SMREAD();

+                        addr = geted(dyn, addr, ninst, nextop, &wback, x2, &fixedaddress, &unscaled, 0xfff, 0, rex, NULL, 0, 1);

+                        LDB(x1, wback, fixedaddress);

+                        ed = x1;

+                    }

                     u8 = F8;

                     VMOVQBfrom(q0, (u8&15), ed);

                     break;

diff --git a/src/dynarec/arm64/dynarec_arm64_avx_0f.c b/src/dynarec/arm64/dynarec_arm64_avx_0f.c
index 122b25c2..74fcf055 100644
--- a/src/dynarec/arm64/dynarec_arm64_avx_0f.c
+++ b/src/dynarec/arm64/dynarec_arm64_avx_0f.c
@@ -508,7 +508,12 @@ uintptr_t dynarec64_AVX_0F(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int
         case 0x5A:
             INST_NAME("VCVTPS2PD Gx, Ex");
             nextop = F8;
-            GETGX_empty_EX(v0, v1, 0);
+            if (vex.l) {
+                GETGX_empty_EX(v0, v1, 0);
+            } else {
+                GETEXSD(v1, 0, 0);
+                GETGX_empty(v0);
+            }
             if(vex.l) {
                 GETGY_empty(q0, -1, -1, -1);
                 FCVTL2(q0, v1);
diff --git a/src/dynarec/arm64/dynarec_arm64_avx_f2_0f.c b/src/dynarec/arm64/dynarec_arm64_avx_f2_0f.c
index b0dd73a5..6e19d8cb 100644
--- a/src/dynarec/arm64/dynarec_arm64_avx_f2_0f.c
+++ b/src/dynarec/arm64/dynarec_arm64_avx_f2_0f.c
@@ -596,11 +596,16 @@ uintptr_t dynarec64_AVX_F2_0F(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip,
                     VMOVQ(v0, v1);
                 }
             } else {
-                v0 = sse_get_reg_empty(dyn, ninst, x1, gd);
-                v1 = ymm_get_reg_empty(dyn, ninst, x1, gd, -1, -1, -1);
                 SMREAD();
-                addr = geted(dyn, addr, ninst, nextop, &ed, x2, &fixedaddress, NULL, 0x3f<<4, 15, rex, NULL, 1, 0);
-                VLDP128_I7(v0, v1, ed, fixedaddress);
+                v0 = sse_get_reg_empty(dyn, ninst, x1, gd);
+                if (vex.l) {
+                    v1 = ymm_get_reg_empty(dyn, ninst, x1, gd, -1, -1, -1);
+                    addr = geted(dyn, addr, ninst, nextop, &ed, x2, &fixedaddress, NULL, 0x3f << 4, 15, rex, NULL, 1, 0);
+                    VLDP128_I7(v0, v1, ed, fixedaddress);
+                } else {
+                    addr = geted(dyn, addr, ninst, nextop, &ed, x1, &fixedaddress, &unscaled, 0xfff << 4, 15, rex, NULL, 0, 0);
+                    VLD128(v0, ed, fixedaddress);
+                }
             }
             if(!vex.l) YMM0(gd);
             break;
diff --git a/src/dynarec/la64/dynarec_la64_0f.c b/src/dynarec/la64/dynarec_la64_0f.c
index 4391a0db..e07d650d 100644
--- a/src/dynarec/la64/dynarec_la64_0f.c
+++ b/src/dynarec/la64/dynarec_la64_0f.c
@@ -317,7 +317,7 @@ uintptr_t dynarec64_0F(dynarec_la64_t* dyn, uintptr_t addr, uintptr_t ip, int ni
             INST_NAME("CVTTPS2PI Gm, Ex");
             nextop = F8;
             GETGM(v0);
-            GETEX(v1, 0, 0);
+            GETEXSD(v1, 0, 0);
             if (BOX64ENV(dynarec_fastround)) {
                 VFTINTRZ_W_S(v0, v1);
             } else {
@@ -344,7 +344,7 @@ uintptr_t dynarec64_0F(dynarec_la64_t* dyn, uintptr_t addr, uintptr_t ip, int ni
             INST_NAME("CVTPS2PI Gm, Ex");
             nextop = F8;
             GETGM(v0);
-            GETEX(v1, 0, 0);
+            GETEXSD(v1, 0, 0);
             u8 = sse_setround(dyn, ninst, x4, x6);
             if (BOX64ENV(dynarec_fastround)) {
                 VFTINTRZ_W_S(v0, v1);
@@ -887,7 +887,7 @@ uintptr_t dynarec64_0F(dynarec_la64_t* dyn, uintptr_t addr, uintptr_t ip, int ni
         case 0x5A:
             INST_NAME("CVTPS2PD Gx, Ex");
             nextop = F8;
-            GETEX(q0, 0, 0);
+            GETEXSD(q0, 0, 0);
             GETGX(q1, 1);
             VFCVTL_D_S(q1, q0);
             break;
diff --git a/src/dynarec/la64/dynarec_la64_660f.c b/src/dynarec/la64/dynarec_la64_660f.c
index dd3d4468..8b3558d9 100644
--- a/src/dynarec/la64/dynarec_la64_660f.c
+++ b/src/dynarec/la64/dynarec_la64_660f.c
@@ -1148,7 +1148,15 @@ uintptr_t dynarec64_660F(dynarec_la64_t* dyn, uintptr_t addr, uintptr_t ip, int
                     INST_NAME("PINSRB Gx, ED, Ib");
                     nextop = F8;
                     GETGX(q0, 1);
-                    GETED(1);
+                    if (MODREG) {
+                        ed = TO_NAT((nextop & 7) + (rex.b << 3));
+                        wback = 0;
+                    } else {
+                        SMREAD();
+                        addr = geted(dyn, addr, ninst, nextop, &wback, x2, x1, &fixedaddress, rex, NULL, 1, 1);
+                        LD_B(x1, wback, fixedaddress);
+                        ed = x1;
+                    }
                     u8 = F8;
                     VINSGR2VR_B(q0, ed, (u8 & 0xf));
                     break;
diff --git a/src/dynarec/rv64/dynarec_rv64_660f38.c b/src/dynarec/rv64/dynarec_rv64_660f38.c
index 1e9a357a..cb8a234f 100644
--- a/src/dynarec/rv64/dynarec_rv64_660f38.c
+++ b/src/dynarec/rv64/dynarec_rv64_660f38.c
@@ -1096,7 +1096,15 @@ uintptr_t dynarec64_660F38(dynarec_rv64_t* dyn, uintptr_t addr, uint8_t opcode,
                     INST_NAME("PINSRB Gx, ED, Ib");
                     nextop = F8;
                     GETGX();
-                    GETED(1);
+                    if (MODREG) {
+                        ed = TO_NAT((nextop & 7) + (rex.b << 3));
+                        wback = 0;
+                    } else {
+                        SMREAD();
+                        addr = geted(dyn, addr, ninst, nextop, &wback, x2, x1, &fixedaddress, rex, NULL, 1, 1);
+                        LB(x1, wback, fixedaddress);
+                        ed = x1;
+                    }
                     u8 = F8;
                     SB(ed, gback, gdoffset + (u8 & 0xF));
                     break;