Merge branch 'master' into libretro

Author: notaz

libpcsxcore/new_dynarec/emu_if.c

@@ -17,6 +17,7 @@
 #include "../gte_arm.h"
 #include "../gte_neon.h"
 #include "compiler_features.h"
+#include "arm_features.h"
 #define FLAGLESS
 #include "../gte.h"
 #ifdef NDRC_THREAD
@@ -317,13 +318,24 @@ static void ari64_apply_config()
 #ifdef NDRC_THREAD
 static void clear_local_cache(void)
 {
-#ifdef _3DS
-	if (ndrc_g.thread.cache_dirty) {
-		ndrc_g.thread.cache_dirty = 0;
-		ctr_invalidate_icache();
-	}
+#if defined(__arm__) || defined(__aarch64__)
+	if (ndrc_g.thread.dirty_start) {
+		// see "Ensuring the visibility of updates to instructions"
+		// in v7/v8 reference manuals (DDI0406, DDI0487 etc.)
+#if defined(__aarch64__) || defined(HAVE_ARMV8)
+		// the actual clean/invalidate is broadcast to all cores,
+		// the manual only prescribes an isb
+		__asm__ volatile("isb");
+//#elif defined(_3DS)
+//		ctr_invalidate_icache();
 #else
-	// hopefully nothing is needed, as tested on r-pi4 and switch
+		// while on v6 this is always required, on v7 it depends on
+		// "Multiprocessing Extensions" being present, but that is difficult
+		// to detect so do it always for now
+		new_dyna_clear_cache(ndrc_g.thread.dirty_start, ndrc_g.thread.dirty_end);
+#endif
+		ndrc_g.thread.dirty_start = ndrc_g.thread.dirty_end = 0;
+	}
 #endif
 }
 

libpcsxcore/new_dynarec/new_dynarec.c

@@ -474,12 +474,7 @@ static void mprotect_w_x(void *start, void *end, int is_x)
 #endif
 }
 
-static void start_tcache_write(void *start, void *end)
-{
-  mprotect_w_x(start, end, 0);
-}
-
-static void end_tcache_write(void *start, void *end)
+void new_dyna_clear_cache(void *start, void *end)
 {
 #if defined(__arm__) || defined(__aarch64__)
   size_t len = (char *)end - (char *)start;
@@ -495,7 +490,6 @@ static void end_tcache_write(void *start, void *end)
     ctr_clear_cache_range(start, end);
   else
     ctr_clear_cache();
-  ndrc_g.thread.cache_dirty = 1;
   #elif defined(HAVE_LIBNX)
   if (g_jit.type == JitType_CodeMemory) {
     armDCacheClean(start, len);
@@ -512,6 +506,22 @@ static void end_tcache_write(void *start, void *end)
   #endif
   (void)len;
 #endif
+}
+
+static void start_tcache_write(void *start, void *end)
+{
+  mprotect_w_x(start, end, 0);
+}
+
+static void end_tcache_write(void *start, void *end)
+{
+#ifdef NDRC_THREAD
+  if (!ndrc_g.thread.dirty_start || (size_t)ndrc_g.thread.dirty_start > (size_t)start)
+    ndrc_g.thread.dirty_start = start;
+  if ((size_t)ndrc_g.thread.dirty_end < (size_t)end)
+    ndrc_g.thread.dirty_end = end;
+#endif
+  new_dyna_clear_cache(start, end);
 
   mprotect_w_x(start, end, 1);
 }

libpcsxcore/new_dynarec/new_dynarec.h

@@ -22,10 +22,11 @@ struct ndrc_globals
 		void *handle;
 		void *lock;
 		void *cond;
+		void *dirty_start;
+		void *dirty_end;
 		unsigned int addr;
 		int busy;
 		int exit;
-		int cache_dirty;   // 3ds only
 	} thread;
 };
 extern struct ndrc_globals ndrc_g;
@@ -40,6 +41,7 @@ void new_dynarec_print_stats(void);
 int  new_dynarec_quick_check_range(unsigned int start, unsigned int end);
 void new_dynarec_invalidate_range(unsigned int start, unsigned int end);
 void new_dynarec_invalidate_all_pages(void);
+void new_dyna_clear_cache(void *start, void *end);
 
 void new_dyna_start(void *context);
 void new_dyna_start_at(void *context, void *compiled_code);