0,
0};
-static unsigned int needs_clear_cache[1<<(TARGET_SIZE_2-17)];
+static u_int needs_clear_cache[1<<(TARGET_SIZE_2-17)];
/* Linker */
literalcount++;
}
-static void *kill_pointer(void *stub)
+// from a pointer to external jump stub (which was produced by emit_extjump2)
+// find where the jumping insn is
+static void *find_extjump_insn(void *stub)
{
int *ptr=(int *)(stub+4);
- assert((*ptr&0x0ff00000)==0x05900000);
+ assert((*ptr&0x0fff0000)==0x059f0000); // ldr rx, [pc, #ofs]
u_int offset=*ptr&0xfff;
- int **l_ptr=(void *)ptr+offset+8;
- int *i_ptr=*l_ptr;
- set_jump_target((int)i_ptr,(int)stub);
- return i_ptr;
+ void **l_ptr=(void *)ptr+offset+8;
+ return *l_ptr;
}
// find where external branch is liked to using addr of it's stub:
static int get_pointer(void *stub)
{
//printf("get_pointer(%x)\n",(int)stub);
- int *ptr=(int *)(stub+4);
- assert((*ptr&0x0fff0000)==0x059f0000);
- u_int offset=*ptr&0xfff;
- int **l_ptr=(void *)ptr+offset+8;
- int *i_ptr=*l_ptr;
+ int *i_ptr=find_extjump_insn(stub);
assert((*i_ptr&0x0f000000)==0x0a000000);
return (int)i_ptr+((*i_ptr<<8)>>6)+8;
}
#define wb_invalidate wb_invalidate_arm
*/
+static void mark_clear_cache(void *target)
+{
+ u_long offset = (char *)target - (char *)BASE_ADDR;
+ u_int mask = 1u << ((offset >> 12) & 31);
+ if (!(needs_clear_cache[offset >> 17] & mask)) {
+ char *start = (char *)((u_long)target & ~4095ul);
+ start_tcache_write(start, start + 4096);
+ needs_clear_cache[offset >> 17] |= mask;
+ }
+}
+
// Clearing the cache is rather slow on ARM Linux, so mark the areas
// that need to be cleared, and then only clear these areas once.
static void do_clear_cache()
end+=4096;
j++;
}else{
- __clear_cache((void *)start,(void *)end);
+ end_tcache_write((void *)start,(void *)end);
break;
}
}
#include <assert.h>
#include <errno.h>
#include <sys/mman.h>
+#ifdef __MACH__
+#include <libkern/OSCacheControl.h>
+#endif
+#include "new_dynarec_config.h"
#include "emu_if.h" //emulator interface
//#define DISASM
#include "assem_arm.h"
#endif
-#ifdef __BLACKBERRY_QNX__
-#undef __clear_cache
-#define __clear_cache(start,end) msync(start, (size_t)((void*)end - (void*)start), MS_SYNC | MS_CACHE_ONLY | MS_INVALIDATE_ICACHE);
-#elif defined(__MACH__)
-#include <libkern/OSCacheControl.h>
-#define __clear_cache mach_clear_cache
-static void __clear_cache(void *start, void *end) {
- size_t len = (char *)end - (char *)start;
- sys_dcache_flush(start, len);
- sys_icache_invalidate(start, len);
-}
-#endif
-
#define MAXBLOCK 4096
#define MAX_OUTPUT_BLOCK_SIZE 262144
static int tracedebug=0;
+static void mprotect_w_x(void *start, void *end, int is_x)
+{
+#ifdef NO_WRITE_EXEC
+ u_long mstart = (u_long)start & ~4095ul;
+ u_long mend = (u_long)end;
+ if (mprotect((void *)mstart, mend - mstart,
+ PROT_READ | (is_x ? PROT_EXEC : PROT_WRITE)) != 0)
+ SysPrintf("mprotect(%c) failed: %s\n", is_x ? 'x' : 'w', strerror(errno));
+#endif
+}
+
+static void start_tcache_write(void *start, void *end)
+{
+ mprotect_w_x(start, end, 0);
+}
+
+static void end_tcache_write(void *start, void *end)
+{
+#ifdef __arm__
+ size_t len = (char *)end - (char *)start;
+ #if defined(__BLACKBERRY_QNX__)
+ msync(start, len, MS_SYNC | MS_CACHE_ONLY | MS_INVALIDATE_ICACHE);
+ #elif defined(__MACH__)
+ sys_cache_control(kCacheFunctionPrepareForExecution, start, len);
+ #elif defined(VITA)
+ int block = sceKernelFindMemBlockByAddr(start, len);
+ sceKernelSyncVMDomain(block, start, len);
+ #else
+ __clear_cache(start, end);
+ #endif
+ (void)len;
+#endif
+
+ mprotect_w_x(start, end, 1);
+}
+
+static void *start_block(void)
+{
+ u_char *end = out + MAX_OUTPUT_BLOCK_SIZE;
+ if (end > (u_char *)BASE_ADDR + (1<<TARGET_SIZE_2))
+ end = (u_char *)BASE_ADDR + (1<<TARGET_SIZE_2);
+ start_tcache_write(out, end);
+ return out;
+}
+
+static void end_block(void *start)
+{
+ end_tcache_write(start, out);
+}
+
//#define DEBUG_CYCLE_COUNT 1
#define NO_CYCLE_PENALTY_THR 12
}
// Dereference the pointers and remove if it matches
-void ll_kill_pointers(struct ll_entry *head,int addr,int shift)
+static void ll_kill_pointers(struct ll_entry *head,int addr,int shift)
{
while(head) {
int ptr=get_pointer(head->addr);
(((ptr-MAX_OUTPUT_BLOCK_SIZE)>>shift)==(addr>>shift)))
{
inv_debug("EXP: Kill pointer at %x (%x)\n",(int)head->addr,head->vaddr);
- u_int host_addr=(u_int)kill_pointer(head->addr);
+ void *host_addr=find_extjump_insn(head->addr);
#ifdef __arm__
- needs_clear_cache[(host_addr-(u_int)BASE_ADDR)>>17]|=1<<(((host_addr-(u_int)BASE_ADDR)>>12)&31);
+ mark_clear_cache(host_addr);
#endif
+ set_jump_target((int)host_addr,(int)head->addr);
}
head=head->next;
}
jump_out[page]=0;
while(head!=NULL) {
inv_debug("INVALIDATE: kill pointer to %x (%x)\n",head->vaddr,(int)head->addr);
- u_int host_addr=(u_int)kill_pointer(head->addr);
+ void *host_addr=find_extjump_insn(head->addr);
#ifdef __arm__
- needs_clear_cache[(host_addr-(u_int)BASE_ADDR)>>17]|=1<<(((host_addr-(u_int)BASE_ADDR)>>12)&31);
+ mark_clear_cache(host_addr);
#endif
+ set_jump_target((int)host_addr,(int)head->addr);
next=head->next;
free(head);
head=next;
static int new_dynarec_test(void)
{
int (*testfunc)(void) = (void *)out;
+ void *beginning;
int ret;
+
+ beginning = start_block();
emit_movimm(DRC_TEST_VAL,0); // test
emit_jmpreg(14);
literal_pool(0);
-#ifdef __arm__
- __clear_cache((void *)testfunc, out);
-#endif
+ end_block(beginning);
SysPrintf("testing if we can run recompiled code..\n");
ret = testfunc();
if (ret == DRC_TEST_VAL)
-1, 0) <= 0) {
SysPrintf("mmap() failed: %s\n", strerror(errno));
}
-#else
+#elif !defined(NO_WRITE_EXEC)
// not all systems allow execute in data segment by default
if (mprotect(out, 1<<TARGET_SIZE_2, PROT_READ | PROT_WRITE | PROT_EXEC) != 0)
SysPrintf("mprotect() failed: %s\n", strerror(errno));
if (Config.HLE && start == 0x80001000) // hlecall
{
// XXX: is this enough? Maybe check hleSoftCall?
- u_int beginning=(u_int)out;
+ void *beginning=start_block();
u_int page=get_page(start);
+
invalid_code[start>>12]=0;
emit_movimm(start,0);
emit_writeword(0,(int)&pcaddr);
emit_jmp((int)new_dyna_leave);
literal_pool(0);
-#ifdef __arm__
- __clear_cache((void *)beginning,out);
-#endif
+ end_block(beginning);
ll_add_flags(jump_in+page,start,state_rflags,(void *)beginning);
return 0;
}
cop1_usable=0;
uint64_t is32_pre=0;
u_int dirty_pre=0;
- u_int beginning=(u_int)out;
+ void *beginning=start_block();
if((u_int)addr&1) {
ds=1;
pagespan_ds();
// Align code
if(((u_int)out)&7) emit_addnop(13);
#endif
- assert((u_int)out-beginning<MAX_OUTPUT_BLOCK_SIZE);
+ assert((u_int)out-(u_int)beginning<MAX_OUTPUT_BLOCK_SIZE);
//printf("shadow buffer: %x-%x\n",(int)copy,(int)copy+slen*4);
memcpy(copy,source,slen*4);
copy+=slen*4;
- #ifdef __arm__
- __clear_cache((void *)beginning,out);
- #endif
+ end_block(beginning);
// If we're within 256K of the end of the buffer,
// start over from the beginning. (Is 256K enough?)