5 files changed, 10 insertions, 13 deletions
diff --git a/src/dynarec/arm64/dynarec_arm64_functions.c b/src/dynarec/arm64/dynarec_arm64_functions.c
index afb1ed6b..8fd0929e 100644
--- a/src/dynarec/arm64/dynarec_arm64_functions.c
+++ b/src/dynarec/arm64/dynarec_arm64_functions.c
@@ -688,6 +688,8 @@ void inst_name_pass3(dynarec_native_t* dyn, int ninst, const char* name, rex_t r
             dynarec_log(LOG_NONE, ", jmp=%d", dyn->insts[ninst].x64.jmp_insts);
         if(dyn->insts[ninst].x64.jmp && dyn->insts[ninst].x64.jmp_insts==-1)
             dynarec_log(LOG_NONE, ", jmp=out");
+        if(dyn->insts[ninst].x64.has_callret)
+            dynarec_log(LOG_NONE, ", callret");
         if(dyn->last_ip)
             dynarec_log(LOG_NONE, ", last_ip=%p", (void*)dyn->last_ip);
         for(int ii=0; ii<32; ++ii) {
@@ -789,14 +791,6 @@ void fpu_reset_ninst(dynarec_native_t* dyn, int ninst)
 
 }
 
-void arm64_fpu_reset(dynarec_native_t* dyn, int ninst, int step)
-{
-    if(step<2) {
-        dyn->insts[ninst].ymm0_in = 0;
-        dyn->insts[ninst].ymm0_out = 0;
-    }
-}
-
 int fpu_is_st_freed(dynarec_native_t* dyn, int ninst, int st)
 {
     return (dyn->n.tags&(0b11<<(st*2)))?1:0;
diff --git a/src/dynarec/arm64/dynarec_arm64_functions.h b/src/dynarec/arm64/dynarec_arm64_functions.h
index 0af490e4..b6c95904 100644
--- a/src/dynarec/arm64/dynarec_arm64_functions.h
+++ b/src/dynarec/arm64/dynarec_arm64_functions.h
@@ -69,7 +69,6 @@ void print_opcode(dynarec_native_t* dyn, int ninst, uint32_t opcode);
 // reset the cache
 void fpu_reset(dynarec_native_t* dyn);
 void fpu_reset_ninst(dynarec_native_t* dyn, int ninst);
-void arm64_fpu_reset(dynarec_native_t* dyn, int ninst, int step);
 
 // is st freed
 int fpu_is_st_freed(dynarec_native_t* dyn, int ninst, int st);
diff --git a/src/dynarec/arm64/dynarec_arm64_helper.c b/src/dynarec/arm64/dynarec_arm64_helper.c
index 136c0f8c..04fa97f6 100644
--- a/src/dynarec/arm64/dynarec_arm64_helper.c
+++ b/src/dynarec/arm64/dynarec_arm64_helper.c
@@ -1806,6 +1806,8 @@ int ymm_get_reg(dynarec_arm_t* dyn, int ninst, int s1, int a, int forwrite, int
             return i;
         }
     // nope, grab a new one
+    if(dyn->ymm_zero&(1<<a))
+        forwrite = 1;   // if the reg was zero, then it will need to be write back
     int ret =  fpu_get_reg_ymm(dyn, ninst, forwrite?NEON_CACHE_YMMW:NEON_CACHE_YMMR, a, k1, k2, k3);
     if(dyn->ymm_zero&(1<<a)) {
         VEORQ(ret, ret, ret);
@@ -2365,6 +2367,12 @@ void fpu_reflectcache(dynarec_arm_t* dyn, int ninst, int s1, int s2, int s3)
     x87_reflectcache(dyn, ninst, s1, s2, s3);
     mmx_reflectcache(dyn, ninst, s1);
     //sse_reflectcache(dyn, ninst, s1); // no need, it's pushed/unpushed during call
+    // but ymm0 needs to be pushed
+    if(dyn->ymm_zero) {
+        ADDx_U12(s1, xEmu, offsetof(x64emu_t, ymm[0]));
+        for(int i=0; i<16; ++i)
+            STPx_S7_offset(xZR, xZR, s1, 16*i);
+    }
 }
 
 void fpu_unreflectcache(dynarec_arm_t* dyn, int ninst, int s1, int s2, int s3)
diff --git a/src/dynarec/arm64/dynarec_arm64_pass0.h b/src/dynarec/arm64/dynarec_arm64_pass0.h
index 510dd4ab..7d4c0c2d 100644
--- a/src/dynarec/arm64/dynarec_arm64_pass0.h
+++ b/src/dynarec/arm64/dynarec_arm64_pass0.h
@@ -26,13 +26,11 @@
         dyn->n.combined1 = dyn->n.combined2 = 0;\
         dyn->n.swapped = 0; dyn->n.barrier = 0; \
         dyn->insts[ninst].f_entry = dyn->f;     \
-        dyn->insts[ninst].ymm0_in = dyn->ymm_zero;\
         if(ninst) {dyn->insts[ninst-1].x64.size = dyn->insts[ninst].x64.addr - dyn->insts[ninst-1].x64.addr;}
 
 #define INST_EPILOG                             \
         dyn->insts[ninst].f_exit = dyn->f;      \
         dyn->insts[ninst].n = dyn->n;           \
-        dyn->insts[ninst].ymm0_out = dyn->ymm_zero;\
         dyn->insts[ninst].x64.has_next = (ok>0)?1:0;
 #define INST_NAME(name) 
 #define DEFAULT                         \
diff --git a/src/dynarec/arm64/dynarec_arm64_pass1.h b/src/dynarec/arm64/dynarec_arm64_pass1.h
index ab1f5fc4..6cf92feb 100644
--- a/src/dynarec/arm64/dynarec_arm64_pass1.h
+++ b/src/dynarec/arm64/dynarec_arm64_pass1.h
@@ -5,12 +5,10 @@
 #define NEW_INST                                \
         dyn->insts[ninst].f_entry = dyn->f;     \
         dyn->n.combined1 = dyn->n.combined2 = 0;\
-        dyn->insts[ninst].ymm0_in = dyn->ymm_zero;\
         dyn->n.swapped = 0; dyn->n.barrier = 0
 
 #define INST_EPILOG                             \
         dyn->insts[ninst].n = dyn->n;           \
-        dyn->insts[ninst].ymm0_out = dyn->ymm_zero;\
         dyn->insts[ninst].f_exit = dyn->f
 
 #define INST_NAME(name)