From b66f5972f9ca126bdd2a204b8df253720fc750a9 Mon Sep 17 00:00:00 2001
From: gbeauche <>
Date: Fri, 3 Oct 2003 18:18:15 +0000
Subject: [PATCH] Make sure a 32-bit B2/JIT works reasonnably well on AMD64
 too. This implies to force RAMBaseHost < 0x80000000. This is empirically
 determined to work on Linux/x86 and Linux/amd64.

---
 BasiliskII/src/Unix/vm_alloc.cpp              | 26 ++++++--
 .../src/uae_cpu/compiler/compemu_support.cpp  | 59 ++++++++++++++++++-
 2 files changed, 80 insertions(+), 5 deletions(-)

diff --git a/BasiliskII/src/Unix/vm_alloc.cpp b/BasiliskII/src/Unix/vm_alloc.cpp
index ba97f238..5a68d0f3 100644
--- a/BasiliskII/src/Unix/vm_alloc.cpp
+++ b/BasiliskII/src/Unix/vm_alloc.cpp
@@ -38,20 +38,38 @@
 #endif
 #endif
 
+/* We want MAP_32BIT, if available, for SheepShaver and BasiliskII
+   because the emulated target is 32-bit and this helps to allocate
+   memory so that branches could be resolved more easily (32-bit
+   displacement to code in .text), on AMD64 for example.  */
+#ifndef MAP_32BIT
+#define MAP_32BIT 0
+#endif
+
+#define MAP_EXTRA_FLAGS (MAP_32BIT)
+
 #ifdef HAVE_MMAP_VM
-static char * next_address = 0;
+#if defined(__linux__) && defined(__i386__)
+/* Force a reasonnable address below 0x80000000 on x86 so that we
+   don't get addresses above when the program is run on AMD64.
+   NOTE: this is empirically determined on Linux/x86.  */
+#define MAP_BASE	0x10000000
+#else
+#define MAP_BASE	0x00000000
+#endif
+static char * next_address = (char *)MAP_BASE;
 #ifdef HAVE_MMAP_ANON
-#define map_flags	(MAP_PRIVATE | MAP_ANON)
+#define map_flags	(MAP_PRIVATE | MAP_ANON | MAP_EXTRA_FLAGS)
 #define zero_fd		-1
 #else
 #ifdef HAVE_MMAP_ANONYMOUS
-#define map_flags	(MAP_PRIVATE | MAP_ANONYMOUS)
+#define map_flags	(MAP_PRIVATE | MAP_ANONYMOUS | MAP_EXTRA_FLAGS)
 #define zero_fd		-1
 #else
 #ifdef HAVE_FCNTL_H
 #include <fcntl.h>
 #endif
-#define map_flags	(MAP_PRIVATE)
+#define map_flags	(MAP_PRIVATE | MAP_EXTRA_FLAGS)
 static int zero_fd	= -1;
 #endif
 #endif
diff --git a/BasiliskII/src/uae_cpu/compiler/compemu_support.cpp b/BasiliskII/src/uae_cpu/compiler/compemu_support.cpp
index 17b8c458..385f96eb 100644
--- a/BasiliskII/src/uae_cpu/compiler/compemu_support.cpp
+++ b/BasiliskII/src/uae_cpu/compiler/compemu_support.cpp
@@ -37,6 +37,7 @@
 #define USE_NORMAL_CALLING_CONVENTION 0
 
 #ifndef WIN32
+#include <unistd.h>
 #include <sys/types.h>
 #include <sys/mman.h>
 #endif
@@ -5530,6 +5531,62 @@ uae_u32 get_jitted_size(void)
     return 0;
 }
 
+const int CODE_ALLOC_MAX_ATTEMPTS = 10;
+const int CODE_ALLOC_BOUNDARIES   = 128 * 1024; // 128 KB
+
+static uint8 *do_alloc_code(uint32 size, int depth)
+{
+#if defined(__linux__) && 0
+	/*
+	  This is a really awful hack that is known to work on Linux at
+	  least.
+	  
+	  The trick here is to make sure the allocated cache is nearby
+	  code segment, and more precisely in the positive half of a
+	  32-bit address space. i.e. addr < 0x80000000. Actually, it
+	  turned out that a 32-bit binary run on AMD64 yields a cache
+	  allocated around 0xa0000000, thus causing some troubles when
+	  translating addresses from m68k to x86.
+	*/
+	static uint8 * code_base = NULL;
+	if (code_base == NULL) {
+		uintptr page_size = getpagesize();
+		uintptr boundaries = CODE_ALLOC_BOUNDARIES;
+		if (boundaries < page_size)
+			boundaries = page_size;
+		code_base = (uint8 *)sbrk(0);
+		for (int attempts = 0; attempts < CODE_ALLOC_MAX_ATTEMPTS; attempts++) {
+			if (vm_acquire_fixed(code_base, size) == 0) {
+				uint8 *code = code_base;
+				code_base += size;
+				return code;
+			}
+			code_base += boundaries;
+		}
+		return NULL;
+	}
+
+	if (vm_acquire_fixed(code_base, size) == 0) {
+		uint8 *code = code_base;
+		code_base += size;
+		return code;
+	}
+
+	if (depth >= CODE_ALLOC_MAX_ATTEMPTS)
+		return NULL;
+
+	return do_alloc_code(size, depth + 1);
+#else
+	uint8 *code = (uint8 *)vm_acquire(size);
+	return code == VM_MAP_FAILED ? NULL : code;
+#endif
+}
+
+static inline uint8 *alloc_code(uint32 size)
+{
+	return do_alloc_code(size, 0);
+}
+
 void alloc_cache(void)
 {
 	if (compiled_code) {
@@ -5542,7 +5599,7 @@ void alloc_cache(void)
 		return;
 	
 	while (!compiled_code && cache_size) {
-		if ((compiled_code = (uae_u8 *)vm_acquire(cache_size * 1024)) == VM_MAP_FAILED) {
+		if ((compiled_code = alloc_code(cache_size * 1024)) == NULL) {
 			compiled_code = 0;
 			cache_size /= 2;
 		}