Adapt x86 GCC assembly copybufreverse() to x86_64 and enable unconditionally.

For a 1680x1050 scene standing directly in front of a mirror, this increases
FPS from ~90 to ~95 for me.

git-svn-id: https://svn.eduke32.com/eduke32@4363 1a8010ca-5511-0410-912e-c29ae57300e0

polymer/eduke32/build/src/pragmas.c

@@ -390,6 +390,50 @@ void copybufbyte(const void *S, void *D, int32_t c)
     while ((c--) > 0) *(q++) = *(p++);
 }
 
+
+// copybufreverse() is a special case: use the assembly version for GCC on x86
+// *and* x86_64, and the C version otherwise.
+// XXX: we don't honor NOASM in the x86_64 case.
+
+#if defined(__GNUC__) && defined(__x86_64__)
+// NOTE: Almost CODEDUP from x86 GCC assembly version, except that
+// - %%esi -> %%rsi
+// - %%edi -> %%rdi
+// - (dec,inc,sub,add)l suffix removed where necessary
+void copybufreverse(const void *S, void *D, int32_t c)
+{
+    __asm__ __volatile__(
+        "shrl $1, %%ecx\n\t"
+        "jnc 0f\n\t"		// jnc skipit1
+        "movb (%%rsi), %%al\n\t"
+        "dec %%rsi\n\t"
+        "movb %%al, (%%rdi)\n\t"
+        "inc %%rdi\n\t"
+        "0:\n\t"		// skipit1:
+        "shrl $1, %%ecx\n\t"
+        "jnc 1f\n\t"		// jnc skipit2
+        "movw -1(%%rsi), %%ax\n\t"
+        "sub $2, %%rsi\n\t"
+        "rorw $8, %%ax\n\t"
+        "movw %%ax, (%%rdi)\n\t"
+        "add $2, %%rdi\n\t"
+        "1:\n\t"		// skipit2
+        "testl %%ecx, %%ecx\n\t"
+        "jz 3f\n\t"		// jz endloop
+        "2:\n\t"		// begloop
+        "movl -3(%%rsi), %%eax\n\t"
+        "sub $4, %%rsi\n\t"
+        "bswapl %%eax\n\t"
+        "movl %%eax, (%%rdi)\n\t"
+        "add $4, %%rdi\n\t"
+        "decl %%ecx\n\t"
+        "jnz 2b\n\t"		// jnz begloop
+        "3:"
+    : "+S"(S), "+D"(D), "+c"(c) :
+            : "eax", "memory", "cc"
+        );
+}
+#else
 void copybufreverse(const void *S, void *D, int32_t c)
 {
     const char *p = (const char *)S;
@@ -397,7 +441,7 @@ void copybufreverse(const void *S, void *D, int32_t c)
 
     while ((c--) > 0) *(q++) = *(p--);
 }
-
 #endif
 
 
+#endif