--- uae/src/md-i386-gcc/support.c	2018/04/24 16:39:33	1.1.1.1
+++ uae/src/md-i386-gcc/support.c	2018/04/24 16:40:31	1.1.1.2
@@ -14,6 +14,12 @@
 #include "options.h"
 #include "machdep/m68k.h"
 
+#ifndef USE_UNDERSCORE
+#define LARGE_ALIGNMENT ".align 16\n"
+#else
+#define LARGE_ALIGNMENT ".align 4,0x90\n"
+#endif
+
 struct flag_struct regflags;
 
 /*
@@ -59,37 +65,32 @@ int fast_memcmp(const void *foo, const v
 
 int memcmpy(void *foo, const void *bar, int len)
 {
-    int differs, baz = 0, uupzuq = 0;
+    int differs = 0, baz = 0, uupzuq = 0;
 
     __asm__ __volatile__ ("subl %1, %2\n"
-			  "movl $0, %0\n"
 			  "subl $16, %3\n"
 			  "jc LLB7\n"
-			  
+			  LARGE_ALIGNMENT
 			  "LLB8:\n"
-			  "movl (%1),%%ecx\n"
 			  "movl (%2,%1),%%ebx\n"
-			  "xorl %%ebx, %%ecx\n"
-			  "movl %%ebx, (%1)\n"
-			  "orl %%ecx, %0\n"
+			  "movl (%1),%%ecx\n"
+			  "cmpl %%ebx, %%ecx\n"
+			  "jne LLC1\n"
 			  
 			  "movl 4(%2,%1),%%ebx\n"
 			  "movl 4(%1),%%ecx\n"
-			  "xorl %%ebx, %%ecx\n"
-			  "movl %%ebx, 4(%1)\n"
-			  "orl %%ecx, %0\n"
+			  "cmpl %%ebx, %%ecx\n"
+			  "jne LLC2\n"
 			  
 			  "movl 8(%2,%1),%%ebx\n"
 			  "movl 8(%1),%%ecx\n"
-			  "xorl %%ebx, %%ecx\n"
-			  "movl %%ebx, 8(%1)\n"
-			  "orl %%ecx, %0\n"
+			  "cmpl %%ebx, %%ecx\n"
+			  "jne LLC3\n"
 			  
 			  "movl 12(%2,%1),%%ebx\n"
 			  "movl 12(%1),%%ecx\n"
-			  "xorl %%ebx, %%ecx\n"
-			  "movl %%ebx, 12(%1)\n"
-			  "orl %%ecx, %0\n"
+			  "cmpl %%ebx, %%ecx\n"
+			  "jne LLC4\n"
 			  
 			  "addl $16, %1\n"
 			  "subl $16, %3\n"
@@ -99,35 +100,109 @@ int memcmpy(void *foo, const void *bar,
 			  "addl $16, %3\n"
 			  "subl $4, %3\n"
 			  "jc  LLB2\n"
-			  
+
 			  "LLB1:\n"
 			  "movl (%2,%1),%%ebx\n"
 			  "movl (%1),%%ecx\n"
-			  "xorl %%ebx, %%ecx\n"
-			  "movl %%ebx, (%1)\n"
-			  "orl %%ecx, %0\n"
+			  "cmpl %%ebx, %%ecx\n"
+			  "jne LLC5\n"
 			  "addl $4, %1\n"
 			  "subl $4, %3\n"
 			  "jnc  LLB1\n"
 			  
 			  "LLB2:\n"
 			  "addl $4, %3\n"
-			  "jz LLB6\n"
-			  "xorl %%ecx, %%ecx\n"
+			  "jz LLB9\n"
 			  
 			  "LLB3:\n"
 			  "movb (%2,%1),%%bl\n"
 			  "movb (%1),%%cl\n"
-			  "xorb %%bl,%%cl\n"
-			  "movb %%bl,(%1)\n"
-			  "orl %%ecx, %0\n"
+			  "cmpb %%bl,%%cl\n"
+			  "jne LLC6\n"
 			  "incl %1\n"
 			  "decl %3\n"
 			  "jnz LLB3\n"
 			  
-			  "LLB6:\n"
+			  "jmp LLB9\n"
+
+			  LARGE_ALIGNMENT
+			  /* Once we find a difference, we switch to plain memcpy() */
+			  "LLC01:\n"
+			  "movl (%2,%1),%%ebx\n"
+			  "LLC1:\n"
+			  "movl %%ebx, (%1)\n"
+			  
+			  "movl 4(%2,%1),%%ebx\n"
+			  "LLC2:\n"
+			  "movl %%ebx, 4(%1)\n"
+			  
+			  "movl 8(%2,%1),%%ebx\n"
+			  "LLC3:\n"
+			  "movl %%ebx, 8(%1)\n"
+			  
+			  "movl 12(%2,%1),%%ebx\n"
+			  "LLC4:\n"
+			  "movl %%ebx, 12(%1)\n"
+			  
+			  "addl $16, %1\n"
+#if 0
+			  "movl $1,%0\n"
+			  
+			  "addl %1,%2\n"
+			  "movl %3,%%ecx\n"
+			  "shrl $2,%%ecx\n"
+			  "je LLC02a\n"
+			  "rep\n"
+			  "movsl\n"
+			  "andl $3,%3\n"
+			  "je LLB9\n"
+			  "LLC02a:\n"
+			  "movb (%2),%%bl\n"
+			  
+			  "movb %%bl,(%1)\n"
+			  "incl %1\n"
+			  "decl %3\n"
+			  "jnz LLC02a\n"
+			  "jmp LLB9\n"
+#else
+#if 0
+			  "movl $1,%0\n"
+			  "jnc  LLB8\n"
+#else
+			  "subl $16, %3\n"
+			  "jnc LLC01\n"
+#endif
+
+			  "addl $16, %3\n"
+			  "subl $4, %3\n"
+			  "jc  LLC03\n"
+#endif
+			  
+			  "LLC02:\n"
+			  "movl (%2,%1),%%ebx\n"
+			  "LLC5:\n"
+			  "movl %%ebx, (%1)\n"
+			  "addl $4, %1\n"
+			  "subl $4, %3\n"
+			  "jnc  LLC02\n"
+			  
+			  "LLC03:\n"
+			  "addl $4, %3\n"
+			  "jz LLC05\n"
+			  
+			  "LLC04:\n"
+			  "movb (%2,%1),%%bl\n"
+			  "LLC6:\n"
+			  "movb %%bl,(%1)\n"
+			  "incl %1\n"
+			  "decl %3\n"
+			  "jnz LLC04\n"
+			  
+			  "LLC05:\n"
+			  "movl $1,%0\n"
+			  "LLB9:"
 			  : "=m" (differs)
-			  : "r" (foo), "r" (bar), "r" (len), "b" (baz), "c" (uupzuq) : "cc", "memory");
+			  : "D" (foo), "S" (bar), "r" (len), "b" (baz), "c" (uupzuq), "0" (differs) : "cc", "memory");
     /* Now tell the compiler that foo, bar and len have been modified 
      * If someone finds a way to express all this cleaner in constraints that
      * GCC 2.7.2 understands, please FIXME */
@@ -135,3 +210,53 @@ int memcmpy(void *foo, const void *bar,
 
     return differs;
 }
+
+#ifdef HAVE_RDTSC
+
+#include <signal.h>
+
+static volatile frame_time_t last_time, best_time;
+static volatile int loops_to_go;
+
+#ifdef __cplusplus
+static RETSIGTYPE alarmhandler(...)
+#else
+static RETSIGTYPE alarmhandler(int foo)
+#endif
+{
+    frame_time_t bar;
+    bar = read_processor_time ();
+    if (bar - last_time < best_time)
+	best_time = bar - last_time;
+    if (--loops_to_go > 0) {
+	signal (SIGALRM, alarmhandler);
+	last_time = read_processor_time();
+	alarm (1);
+    }
+}
+
+void machdep_init (void)
+{
+    fprintf(stderr, "Calibrating delay loop.. ");
+    fflush(stderr);
+    best_time = (frame_time_t)-1;
+    loops_to_go = 5;
+    signal (SIGALRM, alarmhandler);
+    /* We want exact values... */
+    sync (); sync (); sync ();
+    last_time = read_processor_time();
+    alarm (1);
+    while (loops_to_go != 0)
+	usleep(1000000);
+    fprintf (stderr, "ok - %.2f BogoMIPS\n",
+	     ((double)best_time / 1000000), best_time);
+    vsynctime = best_time / 50;
+}
+
+#else
+
+void machdep_init (void)
+{
+}
+
+#endif