[ARM64_DYNAREC] Re-enable weakbarrier for dmb.ishst (#2048)

docs/USAGE.md

@@ -169,9 +169,14 @@ Define Box64's Dynarec max allowed forward value when building Block.
 #### BOX64_DYNAREC_STRONGMEM *
 Enable/Disable simulation of Strong Memory model
 * 0 : Don't try anything special (Default.)
-* 1 : Enable some Memory Barrier when writting to memory (on some MOV opcode) to simulate Strong Memory Model while trying to limit performance impact (Default when libmonobdwgc-2.0.so is loaded)
+* 1 : Enable some memory barriers when writting to memory to simulate the Strong Memory Model in a limited way (Default when libmonobdwgc-2.0.so is loaded)
-* 2 : All 1. plus a memory barrier on every write to memory using MOV
+* 2 : All 1. plus memory barriers on SIMD instructions
-* 3 : All 2. plus Memory Barrier when reading from memory and on some SSE/SSE2 opcodes too
+* 3 : All 2. plus more memory barriers on a regular basis
+
+#### BOX64_DYNAREC_WEAKBARRIER *
+Use weak memory barriers to reduce the performance impact by STRONGMEM
+* 0 : Use regular safe barrier (Default.)
+* 1 : Use weak barriers to have more performance boost
 
 #### BOX64_DYNAREC_X87DOUBLE *
 Force the use of Double for x87 emulation

docs/box64.pod

@@ -296,8 +296,16 @@ Define Box64's Dynarec max allowed forward value when building Block.
 Enable/Disable simulation of Strong Memory model
 
     * 0 : Don't try anything special (Default.)
-    * 1 : Enable some Memory Barrier when reading from memory (on some MOV opcode) to simulate Strong Memory Model while trying to limit performance impact (Default when libmonobdwgc-2.0.so is loaded)
+    * 1 : Enable some memory barriers when writting to memory to simulate the Strong Memory Model in a limited way (Default when libmonobdwgc-2.0.so is loaded)
-    * 2 : Enable some Memory Barrier when reading from memory (on some MOV opcode) to simulate Strong Memory Model
+    * 2 : All 1. plus memory barriers on SIMD instructions
+    * 3 : All 2. plus more memory barriers on a regular basis
+
+=item B<BOX64_DYNAREC_WEAKBARRIER>=I<0|1>
+
+Use weak memory barriers to reduce the performance impact by STRONGMEM
+
+    * 0 : Use regular safe barrier (Default.)
+    * 1 : Use weak barriers to have more performance boost
 
 =item B<BOX64_DYNAREC_X87DOUBLE>=I<0|1>
 

src/dynarec/arm64/dynarec_arm64_helper.h

@@ -60,7 +60,7 @@
  */
 
 #define STRONGMEM_SIMD_WRITE 2 // The level of SIMD memory writes will be tracked
-#define STRONGMEM_LAST_WRITE 2 // The level of a barrier before the last guest memory store will be put
+#define STRONGMEM_LAST_WRITE 1 // The level of a barrier before the last guest memory store will be put
 #define STRONGMEM_SEQ_WRITE  3 // The level of a barrier at every third memory store will be  put
 
 #if STEP == 1
@@ -165,9 +165,15 @@
     do {                                                                                              \
         if (box64_dynarec_strongmem >= dyn->insts[ninst].will_write && dyn->smwrite == 0) {           \
             /* Will write but never written, this is the start of a SEQ, put a barrier. */            \
+            if (box64_dynarec_weakbarrier)                                                            \
+                DMB_ISHST();                                                                          \
+            else                                                                                      \
                 DMB_ISH();                                                                            \
         } else if (box64_dynarec_strongmem >= STRONGMEM_LAST_WRITE && dyn->insts[ninst].last_write) { \
             /* Last write, put a barrier */                                                           \
+            if (box64_dynarec_weakbarrier)                                                            \
+                DMB_ISHST();                                                                          \
+            else                                                                                      \
                 DMB_ISH();                                                                            \
         }                                                                                             \
     } while (0)
@@ -198,6 +204,9 @@
                 --i;                                        \
             if (i >= 0) {                                   \
                 /* It's a SEQ, put a barrier here. */       \
+                if (box64_dynarec_weakbarrier)              \
+                    DMB_ISHST();                            \
+                else                                        \
                     DMB_ISH();                              \
             }                                               \
         }                                                   \