simx86: stk16 and stk32 in cpatch don't need to lock e_invalidate()

Instead call prejit_sync() like in m_munprotect, and also do that
in the other cpatch write functions.

SPEC_PREJIT is about ~40% slower in my tests than non-prejit.

Author: bartoldeman

src/base/emu-i386/simx86/cpatch.c

@@ -53,7 +53,7 @@ int CpatchInvalidates;
 /*
  * Return address of the stub function is passed into eip
  */
-void m_munprotect(unsigned int addr, unsigned int len, unsigned char *eip)
+static void m_munprotect(unsigned int addr, unsigned int len, unsigned char *eip)
 {
 	/* Shut down prejitter before invalidating, or it may crash.
 	 * Also since mprot API currently lacks mutex, shut down prejitter
@@ -241,7 +241,8 @@ void rep_movs_stos(struct rep_stack *stack)
 
 void stk_16(dosaddr_t addr, Bit16u value)
 {
-	e_invalidate(addr, 2);
+	prejit_sync();
+	e_invalidate_unlocked(addr, 2);
 	WRITE_WORD(addr, value);
 #if PROFILE
 	CpatchStkWrites++;
@@ -250,7 +251,8 @@ void stk_16(dosaddr_t addr, Bit16u value)
 
 void stk_32(dosaddr_t addr, Bit32u value)
 {
-	e_invalidate(addr, 4);
+	prejit_sync();
+	e_invalidate_unlocked(addr, 4);
 	WRITE_DWORD(addr, value);
 #if PROFILE
 	CpatchStkWrites++;
@@ -325,6 +327,7 @@ void wri_8(dosaddr_t addr, Bit8u value, unsigned char *rip)
 #if PROFILE
 	CpatchWrites++;
 #endif
+	prejit_sync();
 	if (e_querymprot(addr)) {
 		wri8_slow(addr, value, eip);
 		return;
@@ -354,6 +357,7 @@ void wri_16(dosaddr_t addr, Bit16u value, unsigned char *rip)
 #if PROFILE
 	CpatchWrites++;
 #endif
+	prejit_sync();
 	if (e_querymprotrange(addr, 2)) {
 		wri16_slow(addr, value, eip);
 		return;
@@ -383,6 +387,7 @@ void wri_32(dosaddr_t addr, Bit32u value, unsigned char *rip)
 #if PROFILE
 	CpatchWrites++;
 #endif
+	prejit_sync();
 	if (e_querymprotrange(addr, 4)) {
 		wri32_slow(addr, value, eip);
 		return;

src/base/emu-i386/simx86/emu86.h

@@ -748,7 +748,6 @@ int e_querymprotrange(unsigned int addr, size_t len);
 int e_querymprotrange_full(unsigned int addr, size_t len);
 int e_markpage(unsigned int addr, size_t len);
 int e_unmarkpage(unsigned int addr, size_t len);
-void m_munprotect(unsigned int addr, unsigned int len, unsigned char *eip);
 
 void InitGenCodeBuf(void);
 void EndGenCodeBuf(void);
@@ -773,7 +772,11 @@ int64_t m_findnode(unsigned int addr, size_t len);
 int64_t m_findnodestart(unsigned int addr);
 void mprot_init(void);
 void mprot_end(void);
+#if SPEC_PREJIT
 void prejit_sync(void);
+#else
+#define prejit_sync()
+#endif
 void init_emu_npu(void);
 
 unsigned int Sim_helper(unsigned int mem_ref, unsigned int data, int mode,

src/base/emu-i386/simx86/interp.c

@@ -2809,15 +2809,15 @@ static void prejit_run(TNode *G)
 }
 #endif
 
+#if SPEC_PREJIT
 void prejit_sync(void)
 {
-#if SPEC_PREJIT
   pthread_mutex_lock(&run_mtx);
   while (prejit_running)
     cond_wait(&run_cnd, &run_mtx);
   pthread_mutex_unlock(&run_mtx);
-#endif
 }
+#endif
 
 void prejit_init(void)
 {

src/base/emu-i386/simx86/trees.c

@@ -1398,7 +1398,7 @@ static void do_invalidate(unsigned data, int cnt)
 	InvalidateNodeRange(data, cnt, NULL);
 }
 
-static void _e_invalidate(unsigned data, int cnt)
+void e_invalidate_unlocked(unsigned data, int cnt)
 {
 	/* nothing to invalidate if there are no page protections */
 	if (!e_querymprotrange(data, cnt))
@@ -1413,7 +1413,7 @@ static void _e_invalidate(unsigned data, int cnt)
 void e_invalidate(unsigned data, int cnt)
 {
 	prejit_lock();
-	_e_invalidate(data, cnt);
+	e_invalidate_unlocked(data, cnt);
 	prejit_unlock();
 }
 

src/include/cpu-emu.h

@@ -64,6 +64,7 @@ char *e_scp_disasm(cpuctx_t *scp, int pmode);
 
 /* called from mfs.c, fatfs.c and some places that memcpy */
 void e_invalidate(unsigned data, int cnt);
+void e_invalidate_unlocked(unsigned data, int cnt);
 void e_invalidate_full(unsigned data, int cnt);
 void e_invalidate_full_pa(unsigned data, int cnt);
 int e_invalidate_page_full(unsigned data);