--- uae/src/genp2c.c	2018/04/24 16:39:50	1.1
+++ uae/src/genp2c.c	2018/04/24 17:14:08	1.1.1.5
@@ -12,7 +12,9 @@
 #include "sysconfig.h"
 #include "sysdeps.h"
 
-#include "custom.h"
+/* We can't include custom.h here.  */
+#define MAX_WORDS_PER_LINE 50
+
 
 static char *gen_ind (char *a, int b)
 {
@@ -38,7 +40,11 @@ static char *gen_indsx (char *a, char *s
 #define reg(a) "%" a
 #define ind(a,b) #b"("a")"
 #define imm(a) "$"#a
+#ifdef USE_UNDERSCORE
+#define sym(a) "_"#a
+#else
 #define sym(a) #a
+#endif
 #define indx(a,b,c,d) #b"("a","c","#d")"
 #define indsx(a,s,b,c,d) s"+"#b"("a","c","#d")"
 
@@ -76,13 +82,14 @@ static void orl (char *src, char *dst) {
 static void imull (unsigned long val, char *dst) { printf ("\timull $0x%08lx,%s\n", val, dst); }
 static void decl (char *dst) { printf ("\tdecl %s\n", dst); }
 static void incl (char *dst) { printf ("\tincl %s\n", dst); }
-static void bswapl (char *dst) { printf ("\tbswapl %s\n", dst); }
+static void bswapl (char *dst) { printf ("\tbswap %s\n", dst); }
 static void shrl (int count, char *dst) { printf ("\tshrl $%d,%s\n", count, dst); }
 static void shll (int count, char *dst) { printf ("\tshll $%d,%s\n", count, dst); }
 static void pushl (char *src) { printf ("\tpushl %s\n", src); }
 static void popl (char *dst) { printf ("\tpopl %s\n", dst); }
 static void ret (void) { printf ("\tret\n"); }
-static void align (int a) { printf ("\t.align %d\n", 1 << a); }
+static void align (int a) { printf ("\t.p2align %d,0x90\n", a); }
+
 static void shiftleftl (int count, char *dst)
 {
     if (count == 0)
@@ -109,7 +116,7 @@ static void shiftleftl (int count, char
 static void declare_fn (char *name)
 {
     printf ("\t.globl %s\n", name);
-    printf ("\t.type %s,@function\n", name);
+/*    printf ("\t.type %s,@function\n", name); */
     align (5);
     printf ("%s:\n", name);
 }
@@ -139,7 +146,7 @@ static void gen_x86_set_hires_h_toobad_k
     char buf[40];
     char *indb0;
 
-    sprintf (buf, "set_hires_h_%d_%d", pl, mode);
+    sprintf (buf, sym (set_hires_h_%d_%d), pl, mode);
     declare_fn (buf);
 
     pushl (ebp);
@@ -235,7 +242,7 @@ static void gen_x86_set_hires_h (int pl,
     char buf[40];
     char *indb0;
 
-    sprintf (buf, "set_hires_h_%d_%d", pl, mode);
+    sprintf (buf, sym (set_hires_h_%d_%d), pl, mode);
     declare_fn (buf);
 
     pushl (ebp);
@@ -353,7 +360,7 @@ static void gen_x86_set_hires_h (int pl,
 	    }
 	    break;
 	}
-	
+
 	if (i < pl) {
 	    orl (esi, ecx);
 	    orl (ebp, ebx);
@@ -381,7 +388,7 @@ static void gen_x86_set_hires_h (int pl,
     printf ("\n\n");
 }
 
-/* Squeeze: every second bit does not generate a pixel 
+/* Squeeze: every second bit does not generate a pixel
    Not optimized, this mode isn't useful. */
 static void gen_x86_set_hires_l (int pl, int mode)
 {
@@ -391,14 +398,14 @@ static void gen_x86_set_hires_l (int pl,
     int loop;
     char buf[40];
 
-    sprintf (buf, "set_hires_l_%d_%d", pl, mode);
+    sprintf (buf, sym (set_hires_l_%d_%d), pl, mode);
     declare_fn (buf);
 
     pushl (ebp);
     pushl (esi);
     pushl (edi);
     pushl (ebx);
-    
+
     movl (ind (esp, 20), ebp);
     movl (ind (esp, 24), esi);
     movl (imm (0), edi);
@@ -450,14 +457,14 @@ static void gen_x86_set_lores_h (int pl,
     int loop;
     char buf[40];
 
-    sprintf (buf, "set_lores_h_%d_%d", pl, mode);
+    sprintf (buf, sym (set_lores_h_%d_%d), pl, mode);
     declare_fn (buf);
 
     pushl (ebp);
     pushl (esi);
     pushl (edi);
     pushl (ebx);
-    
+
     movl (ind (esp, 20), ebp);
     movl (ind (esp, 24), esi);
     movl (imm (0), edi);
@@ -476,7 +483,7 @@ static void gen_x86_set_lores_h (int pl,
 	    char *data1 = (i == 0 && mode != 2 ? ecx : edx);
 	    char *data2 = (i == 0 && mode != 2 ? ebx : eax);
 	    char *indb0;
-	    
+
 	    indb0 = gen_indx (esi, realpl*MAX_WORDS_PER_LINE*2, edi, 1);
 	    movzbl (indb0, data2);
 	    free (indb0);
@@ -535,10 +542,10 @@ static void gen_c_set_hires_h (int pl, i
     for (i = 0; i <= pl; i++) {
 	int realpl = i * plmul + ploff;
 	char *asgn = (i == 0 && mode != 2 ? "=" : "|=");
-	
+
 	printf ("\t\t{\n");
 	printf ("\t\t\tunsigned int data = *(ptr + i  + %d);\n", MAX_WORDS_PER_LINE*2*realpl);
-	
+
 	printf ("\t\t\tv1 %s hirestab_h[data][0] << %d;\n", asgn, realpl);
 	printf ("\t\t\tv2 %s hirestab_h[data][1] << %d;\n", asgn, realpl);
 	printf ("\t\t}\n");
@@ -549,14 +556,13 @@ static void gen_c_set_hires_h (int pl, i
     printf ("}\n\n");
 }
 
-/* Squeeze: every second bit does not generate a pixel 
+/* Squeeze: every second bit does not generate a pixel
    Not optimized, this mode isn't useful. */
 static void gen_c_set_hires_l (int pl, int mode, int header)
 {
     int plmul = mode == 0 ? 1 : 2;
     int ploff = mode == 2 ? 1 : 0;
     int i;
-    int loop;
 
     if (header)
 	printf("extern ");
@@ -575,7 +581,7 @@ static void gen_c_set_hires_l (int pl, i
     for (i = 0; i <= pl; i++) {
 	int realpl = i * plmul + ploff;
 	char *asgn = (i == 0 && mode != 2 ? "=" : "|=");
-	
+
 	printf ("\t\t{\n");
 	printf ("\t\t\tunsigned int data = *(ptr + i  + %d);\n", MAX_WORDS_PER_LINE*2*realpl);
 
@@ -592,7 +598,7 @@ static void gen_c_set_lores_h (int pl, i
 {
     int plmul = mode == 0 ? 1 : 2;
     int ploff = mode == 2 ? 1 : 0;
-    int i, j;
+    int i;
 
     if (header)
 	printf("extern ");