diff --git a/src/dynarec/arm64/dynarec_arm64_660f.c b/src/dynarec/arm64/dynarec_arm64_660f.c
index 2d09c4bf6a..fcf0ef5384 100644
--- a/src/dynarec/arm64/dynarec_arm64_660f.c
+++ b/src/dynarec/arm64/dynarec_arm64_660f.c
@@ -1324,6 +1324,8 @@ uintptr_t dynarec64_660F(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int n
                         ed = (nextop&7)+(rex.b<<3);
                         if(ed>7)
                             sse_reflect_reg(dyn, ninst, ed);
+                        else
+                            sse_mark_used(dyn, ninst, ed);
                         ADDx_U12(x1, xEmu, offsetof(x64emu_t, xmm[ed]));
                     } else {
                         addr = geted(dyn, addr, ninst, nextop, &ed, x1, &fixedaddress, NULL, 0, 0, rex, NULL, 0, 1);
@@ -1496,11 +1498,15 @@ uintptr_t dynarec64_660F(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int n
                         SETFLAGS(X_ALL, SF_SET_DF);
                         if(gd>7)    // no need to reflect cache as xmm0-xmm7 will be saved before the function call anyway
                             sse_reflect_reg(dyn, ninst, gd);
+                        else
+                            sse_mark_used(dyn, ninst, gd);
                         ADDx_U12(x3, xEmu, offsetof(x64emu_t, xmm[gd]));
                         if(MODREG) {
                             ed = (nextop&7)+(rex.b<<3);
                             if(ed>7)
                                 sse_reflect_reg(dyn, ninst, ed);
+                            else
+                                sse_mark_used(dyn, ninst, ed);
                             ADDx_U12(x1, xEmu, offsetof(x64emu_t, xmm[ed]));
                         } else {
                             addr = geted(dyn, addr, ninst, nextop, &ed, x1, &fixedaddress, NULL, 0, 0, rex, NULL, 0, 1);
@@ -1540,6 +1546,8 @@ uintptr_t dynarec64_660F(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int n
                         ed = (nextop&7)+(rex.b<<3);
                         if(ed>7)
                             sse_reflect_reg(dyn, ninst, ed);
+                        else
+                            sse_mark_used(dyn, ninst, ed);
                         ADDx_U12(x1, xEmu, offsetof(x64emu_t, xmm[ed]));
                     } else {
                         addr = geted(dyn, addr, ninst, nextop, &ed, x1, &fixedaddress, NULL, 0, 0, rex, NULL, 0, 1);
@@ -1589,11 +1597,15 @@ uintptr_t dynarec64_660F(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int n
                     GETG;
                     if(gd>7)
                         sse_reflect_reg(dyn, ninst, gd);
+                    else
+                        sse_mark_used(dyn, ninst, gd);
                     ADDx_U12(x2, xEmu, offsetof(x64emu_t, xmm[gd]));
                     if(MODREG) {
                         ed = (nextop&7)+(rex.b<<3);
                         if(ed>7)
                             sse_reflect_reg(dyn, ninst, ed);
+                        else
+                            sse_mark_used(dyn, ninst, ed);
                         ADDx_U12(x1, xEmu, offsetof(x64emu_t, xmm[ed]));
                     } else {
                         addr = geted(dyn, addr, ninst, nextop, &ed, x1, &fixedaddress, NULL, 0, 0, rex, NULL, 0, 1);
diff --git a/src/dynarec/arm64/dynarec_arm64_functions.c b/src/dynarec/arm64/dynarec_arm64_functions.c
index c1dbe812f5..398d8aa18d 100644
--- a/src/dynarec/arm64/dynarec_arm64_functions.c
+++ b/src/dynarec/arm64/dynarec_arm64_functions.c
@@ -1514,6 +1514,11 @@ void updateYmm0s(dynarec_arm_t* dyn, int ninst, int max_ninst_reached)
 }
 
 
+void sse_mark_used(dynarec_arm_t* dyn, int ninst, int a)
+{
+    dyn->n.xmm_used |= 1<<a;
+}
+
 // AVX helpers
 void avx_mark_zero(dynarec_arm_t* dyn, int ninst, int reg)
 {
diff --git a/src/dynarec/arm64/dynarec_arm64_functions.h b/src/dynarec/arm64/dynarec_arm64_functions.h
index 3d0f01cd0f..03c66c479f 100644
--- a/src/dynarec/arm64/dynarec_arm64_functions.h
+++ b/src/dynarec/arm64/dynarec_arm64_functions.h
@@ -93,6 +93,9 @@ void updateYmm0s(dynarec_arm_t* dyn, int ninst, int max_ninst_reached);
 // add a preload of SSE/AVX register preload before i2, for a jump in ninst. Might not do it if there is jump in between that goes outside this ninst-i2 interval
 void addSSEPreload(dynarec_arm_t* dyn, int ninst, int i2, uint32_t preload);
 
+// mark a SSE reg as read by the current opcode, without touching its cache entry
+void sse_mark_used(dynarec_arm_t* dyn, int ninst, int a);
+
 // AVX helpers
 void avx_mark_zero(dynarec_arm_t* dyn, int ninst, int reg);
 int is_avx_zero(dynarec_arm_t* dyn, int ninst, int reg);