FFI: Fix code generation for replay of sunk float fields.
[luajit-2.0.git] / src / lj_mcode.c
blob34405b5aff59e57d2ffdce3cd55cdd18ccfec4be
1 /*
2 ** Machine code management.
3 ** Copyright (C) 2005-2012 Mike Pall. See Copyright Notice in luajit.h
4 */
6 #define lj_mcode_c
7 #define LUA_CORE
9 #include "lj_obj.h"
10 #if LJ_HASJIT
11 #include "lj_gc.h"
12 #include "lj_jit.h"
13 #include "lj_mcode.h"
14 #include "lj_trace.h"
15 #include "lj_dispatch.h"
16 #endif
17 #if LJ_HASJIT || LJ_HASFFI
18 #include "lj_vm.h"
19 #endif
21 /* -- OS-specific functions ----------------------------------------------- */
23 #if LJ_HASJIT || LJ_HASFFI
25 /* Define this if you want to run LuaJIT with Valgrind. */
26 #ifdef LUAJIT_USE_VALGRIND
27 #include <valgrind/valgrind.h>
28 #endif
30 #if LJ_TARGET_IOS
31 void sys_icache_invalidate(void *start, size_t len);
32 #endif
34 /* Synchronize data/instruction cache. */
35 void lj_mcode_sync(void *start, void *end)
37 #ifdef LUAJIT_USE_VALGRIND
38 VALGRIND_DISCARD_TRANSLATIONS(start, (char *)end-(char *)start);
39 #endif
40 #if LJ_TARGET_X86ORX64
41 UNUSED(start); UNUSED(end);
42 #elif LJ_TARGET_IOS
43 sys_icache_invalidate(start, (char *)end-(char *)start);
44 #elif LJ_TARGET_PPC
45 lj_vm_cachesync(start, end);
46 #elif defined(__GNUC__)
47 __clear_cache(start, end);
48 #else
49 #error "Missing builtin to flush instruction cache"
50 #endif
53 #endif
55 #if LJ_HASJIT
57 #if LJ_TARGET_WINDOWS
59 #define WIN32_LEAN_AND_MEAN
60 #include <windows.h>
62 #define MCPROT_RW PAGE_READWRITE
63 #define MCPROT_RX PAGE_EXECUTE_READ
64 #define MCPROT_RWX PAGE_EXECUTE_READWRITE
66 static void *mcode_alloc_at(jit_State *J, uintptr_t hint, size_t sz, DWORD prot)
68 void *p = VirtualAlloc((void *)hint, sz,
69 MEM_RESERVE|MEM_COMMIT|MEM_TOP_DOWN, prot);
70 if (!p && !hint)
71 lj_trace_err(J, LJ_TRERR_MCODEAL);
72 return p;
75 static void mcode_free(jit_State *J, void *p, size_t sz)
77 UNUSED(J); UNUSED(sz);
78 VirtualFree(p, 0, MEM_RELEASE);
81 static void mcode_setprot(void *p, size_t sz, DWORD prot)
83 DWORD oprot;
84 VirtualProtect(p, sz, prot, &oprot);
87 #elif LJ_TARGET_POSIX
89 #include <sys/mman.h>
91 #ifndef MAP_ANONYMOUS
92 #define MAP_ANONYMOUS MAP_ANON
93 #endif
95 #define MCPROT_RW (PROT_READ|PROT_WRITE)
96 #define MCPROT_RX (PROT_READ|PROT_EXEC)
97 #define MCPROT_RWX (PROT_READ|PROT_WRITE|PROT_EXEC)
99 static void *mcode_alloc_at(jit_State *J, uintptr_t hint, size_t sz, int prot)
101 void *p = mmap((void *)hint, sz, prot, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0);
102 if (p == MAP_FAILED && !hint)
103 lj_trace_err(J, LJ_TRERR_MCODEAL);
104 return p;
107 static void mcode_free(jit_State *J, void *p, size_t sz)
109 UNUSED(J);
110 munmap(p, sz);
113 static void mcode_setprot(void *p, size_t sz, int prot)
115 mprotect(p, sz, prot);
118 #elif LJ_64
120 #error "Missing OS support for explicit placement of executable memory"
122 #else
124 /* Fallback allocator. This will fail if memory is not executable by default. */
125 #define LUAJIT_UNPROTECT_MCODE
126 #define MCPROT_RW 0
127 #define MCPROT_RX 0
128 #define MCPROT_RWX 0
130 static void *mcode_alloc_at(jit_State *J, uintptr_t hint, size_t sz, int prot)
132 UNUSED(hint); UNUSED(prot);
133 return lj_mem_new(J->L, sz);
136 static void mcode_free(jit_State *J, void *p, size_t sz)
138 lj_mem_free(J2G(J), p, sz);
141 #define mcode_setprot(p, sz, prot) UNUSED(p)
143 #endif
145 /* -- MCode area protection ----------------------------------------------- */
147 /* Define this ONLY if the page protection twiddling becomes a bottleneck. */
148 #ifdef LUAJIT_UNPROTECT_MCODE
150 /* It's generally considered to be a potential security risk to have
151 ** pages with simultaneous write *and* execute access in a process.
153 ** Do not even think about using this mode for server processes or
154 ** apps handling untrusted external data (such as a browser).
156 ** The security risk is not in LuaJIT itself -- but if an adversary finds
157 ** any *other* flaw in your C application logic, then any RWX memory page
158 ** simplifies writing an exploit considerably.
160 #define MCPROT_GEN MCPROT_RWX
161 #define MCPROT_RUN MCPROT_RWX
163 static void mcode_protect(jit_State *J, int prot)
165 UNUSED(J); UNUSED(prot);
168 #else
170 /* This is the default behaviour and much safer:
172 ** Most of the time the memory pages holding machine code are executable,
173 ** but NONE of them is writable.
175 ** The current memory area is marked read-write (but NOT executable) only
176 ** during the short time window while the assembler generates machine code.
178 #define MCPROT_GEN MCPROT_RW
179 #define MCPROT_RUN MCPROT_RX
181 /* Change protection of MCode area. */
182 static void mcode_protect(jit_State *J, int prot)
184 if (J->mcprot != prot) {
185 mcode_setprot(J->mcarea, J->szmcarea, prot);
186 J->mcprot = prot;
190 #endif
192 /* -- MCode area allocation ----------------------------------------------- */
194 #if LJ_TARGET_X64
195 #define mcode_validptr(p) ((p) && (uintptr_t)(p) < (uintptr_t)1<<47)
196 #else
197 #define mcode_validptr(p) ((p) && (uintptr_t)(p) < 0xffff0000)
198 #endif
200 #ifdef LJ_TARGET_JUMPRANGE
202 /* Get memory within relative jump distance of our code in 64 bit mode. */
203 static void *mcode_alloc(jit_State *J, size_t sz)
205 /* Target an address in the static assembler code (64K aligned).
206 ** Try addresses within a distance of target-range/2+1MB..target+range/2-1MB.
208 #if LJ_TARGET_MIPS
209 /* Use the middle of the 256MB-aligned region. */
210 uintptr_t target = ((uintptr_t)(void *)lj_vm_exit_handler & 0xf0000000u) +
211 0x08000000u;
212 #else
213 uintptr_t target = (uintptr_t)(void *)lj_vm_exit_handler & ~(uintptr_t)0xffff;
214 #endif
215 const uintptr_t range = (1u << LJ_TARGET_JUMPRANGE) - (1u << 21);
216 /* First try a contiguous area below the last one. */
217 uintptr_t hint = J->mcarea ? (uintptr_t)J->mcarea - sz : 0;
218 int i;
219 for (i = 0; i < 32; i++) { /* 32 attempts ought to be enough ... */
220 if (mcode_validptr(hint)) {
221 void *p = mcode_alloc_at(J, hint, sz, MCPROT_GEN);
223 if (mcode_validptr(p)) {
224 if ((uintptr_t)p + sz - target < range || target - (uintptr_t)p < range)
225 return p;
226 mcode_free(J, p, sz); /* Free badly placed area. */
229 /* Next try probing pseudo-random addresses. */
230 do {
231 hint = (0x78fb ^ LJ_PRNG_BITS(J, 15)) << 16; /* 64K aligned. */
232 } while (!(hint + sz < range));
233 hint = target + hint - (range>>1);
235 lj_trace_err(J, LJ_TRERR_MCODEAL); /* Give up. OS probably ignores hints? */
236 return NULL;
239 #else
241 /* All memory addresses are reachable by relative jumps. */
242 #define mcode_alloc(J, sz) mcode_alloc_at((J), 0, (sz), MCPROT_GEN)
244 #endif
246 /* -- MCode area management ----------------------------------------------- */
248 /* Linked list of MCode areas. */
249 typedef struct MCLink {
250 MCode *next; /* Next area. */
251 size_t size; /* Size of current area. */
252 } MCLink;
254 /* Allocate a new MCode area. */
255 static void mcode_allocarea(jit_State *J)
257 MCode *oldarea = J->mcarea;
258 size_t sz = (size_t)J->param[JIT_P_sizemcode] << 10;
259 sz = (sz + LJ_PAGESIZE-1) & ~(size_t)(LJ_PAGESIZE - 1);
260 J->mcarea = (MCode *)mcode_alloc(J, sz);
261 J->szmcarea = sz;
262 J->mcprot = MCPROT_GEN;
263 J->mctop = (MCode *)((char *)J->mcarea + J->szmcarea);
264 J->mcbot = (MCode *)((char *)J->mcarea + sizeof(MCLink));
265 ((MCLink *)J->mcarea)->next = oldarea;
266 ((MCLink *)J->mcarea)->size = sz;
267 J->szallmcarea += sz;
270 /* Free all MCode areas. */
271 void lj_mcode_free(jit_State *J)
273 MCode *mc = J->mcarea;
274 J->mcarea = NULL;
275 J->szallmcarea = 0;
276 while (mc) {
277 MCode *next = ((MCLink *)mc)->next;
278 mcode_free(J, mc, ((MCLink *)mc)->size);
279 mc = next;
283 /* -- MCode transactions -------------------------------------------------- */
285 /* Reserve the remainder of the current MCode area. */
286 MCode *lj_mcode_reserve(jit_State *J, MCode **lim)
288 if (!J->mcarea)
289 mcode_allocarea(J);
290 else
291 mcode_protect(J, MCPROT_GEN);
292 *lim = J->mcbot;
293 return J->mctop;
296 /* Commit the top part of the current MCode area. */
297 void lj_mcode_commit(jit_State *J, MCode *top)
299 J->mctop = top;
300 mcode_protect(J, MCPROT_RUN);
303 /* Abort the reservation. */
304 void lj_mcode_abort(jit_State *J)
306 mcode_protect(J, MCPROT_RUN);
309 /* Set/reset protection to allow patching of MCode areas. */
310 MCode *lj_mcode_patch(jit_State *J, MCode *ptr, int finish)
312 #ifdef LUAJIT_UNPROTECT_MCODE
313 UNUSED(J); UNUSED(ptr); UNUSED(finish);
314 return NULL;
315 #else
316 if (finish) {
317 if (J->mcarea == ptr)
318 mcode_protect(J, MCPROT_RUN);
319 else
320 mcode_setprot(ptr, ((MCLink *)ptr)->size, MCPROT_RUN);
321 return NULL;
322 } else {
323 MCode *mc = J->mcarea;
324 /* Try current area first to use the protection cache. */
325 if (ptr >= mc && ptr < (MCode *)((char *)mc + J->szmcarea)) {
326 mcode_protect(J, MCPROT_GEN);
327 return mc;
329 /* Otherwise search through the list of MCode areas. */
330 for (;;) {
331 mc = ((MCLink *)mc)->next;
332 lua_assert(mc != NULL);
333 if (ptr >= mc && ptr < (MCode *)((char *)mc + ((MCLink *)mc)->size)) {
334 mcode_setprot(mc, ((MCLink *)mc)->size, MCPROT_GEN);
335 return mc;
339 #endif
342 /* Limit of MCode reservation reached. */
343 void lj_mcode_limiterr(jit_State *J, size_t need)
345 size_t sizemcode, maxmcode;
346 lj_mcode_abort(J);
347 sizemcode = (size_t)J->param[JIT_P_sizemcode] << 10;
348 sizemcode = (sizemcode + LJ_PAGESIZE-1) & ~(size_t)(LJ_PAGESIZE - 1);
349 maxmcode = (size_t)J->param[JIT_P_maxmcode] << 10;
350 if ((size_t)need > sizemcode)
351 lj_trace_err(J, LJ_TRERR_MCODEOV); /* Too long for any area. */
352 if (J->szallmcarea + sizemcode > maxmcode)
353 lj_trace_err(J, LJ_TRERR_MCODEAL);
354 mcode_allocarea(J);
355 lj_trace_err(J, LJ_TRERR_MCODELM); /* Retry with new area. */
358 #endif