[ARM64_DYNAREC] Small change and optims to various (V)MOVNT* opcodes

4 files changed, 24 insertions, 33 deletions

diff --git a/src/dynarec/arm64/dynarec_arm64_0f.c b/src/dynarec/arm64/dynarec_arm64_0f.c
index 33a9abb5..9e62417c 100644
--- a/src/dynarec/arm64/dynarec_arm64_0f.c
+++ b/src/dynarec/arm64/dynarec_arm64_0f.c
@@ -433,9 +433,7 @@ uintptr_t dynarec64_0F(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
             GETG;

             v0 = sse_get_reg(dyn, ninst, x1, gd, 0);

             if(MODREG) {

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

-                v1 = sse_get_reg_empty(dyn, ninst, x1, ed);

-                VMOVQ(v1, v0);

+                DEFAULT;

             } else {

                 addr = geted(dyn, addr, ninst, nextop, &ed, x1, &fixedaddress, &unscaled, 0xfff<<4, 15, rex, NULL, 0, 0);

                 VST128(v0, ed, fixedaddress);

diff --git a/src/dynarec/arm64/dynarec_arm64_660f.c b/src/dynarec/arm64/dynarec_arm64_660f.c
index dec21e10..03157f77 100644
--- a/src/dynarec/arm64/dynarec_arm64_660f.c
+++ b/src/dynarec/arm64/dynarec_arm64_660f.c
@@ -216,12 +216,10 @@ uintptr_t dynarec64_660F(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int n
         case 0x2B:

             INST_NAME("MOVNTPD Ex,Gx");

             nextop = F8;

-            GETGX(v0, 0);

             if(MODREG) {

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

-                v1 = sse_get_reg_empty(dyn, ninst, x1, ed);

-                VMOVQ(v1, v0);

+                DEFAULT;

             } else {

+                GETGX(v0, 0);

                 addr = geted(dyn, addr, ninst, nextop, &ed, x1, &fixedaddress, &unscaled, 0xfff<<4, 15, rex, NULL, 0, 0);

                 VST128(v0, ed, fixedaddress);

             }

@@ -626,9 +624,7 @@ uintptr_t dynarec64_660F(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int n
                     INST_NAME("MOVNTDQA Gx, Ex");

                     nextop = F8;

                     if (MODREG) {

-                        v1 = sse_get_reg(dyn, ninst, x1, (nextop & 7) + (rex.b << 3), 0);

-                        GETGX_empty(v0);

-                        VMOVQ(v0, v1);

+                        DEFAULT;

                     } else {

                         GETGX_empty(v0);

                         SMREAD();

@@ -3186,11 +3182,10 @@ uintptr_t dynarec64_660F(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int n
         case 0xE7:

             INST_NAME("MOVNTDQ Ex, Gx");

             nextop = F8;

-            GETGX(v0, 0);

             if(MODREG) {

-                v1 = sse_get_reg_empty(dyn, ninst, x1, (nextop&7)+(rex.b<<3));

-                VMOVQ(v1, v0);

+                DEFAULT;

             } else {

+                GETGX(v0, 0);

                 addr = geted(dyn, addr, ninst, nextop, &ed, x1, &fixedaddress, &unscaled, 0xfff<<4, 15, rex, NULL, 0, 0);

                 VST128(v0, ed, fixedaddress);

             }

diff --git a/src/dynarec/arm64/dynarec_arm64_avx_66_0f.c b/src/dynarec/arm64/dynarec_arm64_avx_66_0f.c
index 2d80f5b6..cb2cc235 100644
--- a/src/dynarec/arm64/dynarec_arm64_avx_66_0f.c
+++ b/src/dynarec/arm64/dynarec_arm64_avx_66_0f.c
@@ -245,16 +245,10 @@ uintptr_t dynarec64_AVX_66_0F(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip,
             INST_NAME("VMOVNTPD Ex,Gx");
             nextop = F8;
             GETG;
-            v0 = sse_get_reg(dyn, ninst, x1, gd, 0);
             if(MODREG) {
-                ed = (nextop&7) + (rex.b<<3);
-                v1 = sse_get_reg_empty(dyn, ninst, x1, ed);
-                VMOVQ(v1, v0);
-                if(vex.l) {
-                    GETGYEY_empty(v0, v1);
-                    VMOVQ(v1, v0);
-                } else YMM0(ed);
+                DEFAULT;
             } else {
+                v0 = sse_get_reg(dyn, ninst, x1, gd, 0);
                 if(vex.l) {
                     GETGY(v1, 0, -1, -1, -1);
                     addr = geted(dyn, addr, ninst, nextop, &ed, x1, &fixedaddress, NULL, 0x3f<<4, 15, rex, NULL, 1, 0);
@@ -1702,16 +1696,10 @@ uintptr_t dynarec64_AVX_66_0F(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip,
             INST_NAME("VMOVNTDQ Ex,Gx");
             nextop = F8;
             GETG;
-            v0 = sse_get_reg(dyn, ninst, x1, gd, 0);
             if(MODREG) {
-                ed = (nextop&7)+(rex.b<<3);
-                v1 = sse_get_reg_empty(dyn, ninst, x1, ed);
-                VMOVQ(v1, v0);
-                if(vex.l) {
-                    GETGYEY_empty(v0, v1);
-                    VMOVQ(v1, v0);
-                } else YMM0(ed);
+                DEFAULT;
             } else {
+                v0 = sse_get_reg(dyn, ninst, x1, gd, 0);
                 if(vex.l) {
                     GETGY(v1, 0, -1, -1, -1);
                     addr = geted(dyn, addr, ninst, nextop, &ed, x1, &fixedaddress, NULL, 0x3f<<4, 15, rex, NULL, 1, 0);
diff --git a/src/dynarec/arm64/dynarec_arm64_avx_66_0f38.c b/src/dynarec/arm64/dynarec_arm64_avx_66_0f38.c
index fe254a73..35c2fe29 100644
--- a/src/dynarec/arm64/dynarec_arm64_avx_66_0f38.c
+++ b/src/dynarec/arm64/dynarec_arm64_avx_66_0f38.c
@@ -645,11 +645,21 @@ uintptr_t dynarec64_AVX_66_0F38(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip
         case 0x2A:
             INST_NAME("VMOVNTDQA Gx, Ex");
             nextop = F8;
-            for(int l=0; l<1+vex.l; ++l) {
-                if(!l) { GETGX_empty_EX(v0, v1, 0); } else { GETGY_empty_EY(v0, v1); }
-                VMOVQ(v0, v1);
+            GETG;
+            if(MODREG) {
+                DEFAULT;
+            } else {
+                v0 = sse_get_reg_empty(dyn, ninst, x1, gd);
+                if(vex.l) {
+                    GETGY_empty(v1, -1, -1, -1);
+                    addr = geted(dyn, addr, ninst, nextop, &ed, x1, &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);
+                    YMM0(gd);
+                }
             }
-            if(!vex.l) YMM0(gd);
             break;
         case 0x2B:
             INST_NAME("VPACKUSDW Gx, Ex, Vx");