1 /*
2 * Copyright (c) 1997, 2014, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "asm/macroAssembler.hpp"
27 #include "memory/resourceArea.hpp"
28 #include "nativeInst_x86.hpp"
29 #include "oops/oop.inline.hpp"
30 #include "runtime/handles.hpp"
31 #include "runtime/sharedRuntime.hpp"
32 #include "runtime/stubRoutines.hpp"
33 #include "utilities/ostream.hpp"
34 #ifdef COMPILER1
35 #include "c1/c1_Runtime1.hpp"
36 #endif
37
38 PRAGMA_FORMAT_MUTE_WARNINGS_FOR_GCC
39
40 void NativeInstruction::wrote(int offset) {
41 ICache::invalidate_word(addr_at(offset));
42 }
43
44
45 void NativeCall::verify() {
46 // Make sure code pattern is actually a call imm32 instruction.
47 int inst = ubyte_at(0);
48 if (inst != instruction_code) {
49 tty->print_cr("Addr: " INTPTR_FORMAT " Code: 0x%x", instruction_address(),
50 inst);
51 fatal("not a call disp32");
52 }
53 }
54
55 address NativeCall::destination() const {
56 // Getting the destination of a call isn't safe because that call can
57 // be getting patched while you're calling this. There's only special
58 // places where this can be called but not automatically verifiable by
59 // checking which locks are held. The solution is true atomic patching
60 // on x86, nyi.
61 return return_address() + displacement();
62 }
63
64 void NativeCall::print() {
65 tty->print_cr(PTR_FORMAT ": call " PTR_FORMAT,
66 instruction_address(), destination());
67 }
68
69 // Inserts a native call instruction at a given pc
70 void NativeCall::insert(address code_pos, address entry) {
71 intptr_t disp = (intptr_t)entry - ((intptr_t)code_pos + 1 + 4);
72 #ifdef AMD64
73 guarantee(disp == (intptr_t)(jint)disp, "must be 32-bit offset");
74 #endif // AMD64
75 *code_pos = instruction_code;
76 *((int32_t *)(code_pos+1)) = (int32_t) disp;
77 ICache::invalidate_range(code_pos, instruction_size);
78 }
79
80 // MT-safe patching of a call instruction.
81 // First patches first word of instruction to two jmp's that jmps to them
82 // selfs (spinlock). Then patches the last byte, and then atomicly replaces
83 // the jmp's with the first 4 byte of the new instruction.
84 void NativeCall::replace_mt_safe(address instr_addr, address code_buffer) {
85 assert(Patching_lock->is_locked() ||
86 SafepointSynchronize::is_at_safepoint(), "concurrent code patching");
87 assert (instr_addr != NULL, "illegal address for code patching");
88
89 NativeCall* n_call = nativeCall_at (instr_addr); // checking that it is a call
90 if (os::is_MP()) {
91 guarantee((intptr_t)instr_addr % BytesPerWord == 0, "must be aligned");
92 }
93
94 // First patch dummy jmp in place
95 unsigned char patch[4];
96 assert(sizeof(patch)==sizeof(jint), "sanity check");
97 patch[0] = 0xEB; // jmp rel8
98 patch[1] = 0xFE; // jmp to self
99 patch[2] = 0xEB;
100 patch[3] = 0xFE;
101
102 // First patch dummy jmp in place
103 *(jint*)instr_addr = *(jint *)patch;
104
105 // Invalidate. Opteron requires a flush after every write.
106 n_call->wrote(0);
107
108 // Patch 4th byte
109 instr_addr[4] = code_buffer[4];
110
111 n_call->wrote(4);
112
113 // Patch bytes 0-3
114 *(jint*)instr_addr = *(jint *)code_buffer;
115
116 n_call->wrote(0);
117
118 #ifdef ASSERT
119 // verify patching
120 for ( int i = 0; i < instruction_size; i++) {
121 address ptr = (address)((intptr_t)code_buffer + i);
122 int a_byte = (*ptr) & 0xFF;
123 assert(*((address)((intptr_t)instr_addr + i)) == a_byte, "mt safe patching failed");
124 }
125 #endif
126
127 }
128
129
130 // Similar to replace_mt_safe, but just changes the destination. The
131 // important thing is that free-running threads are able to execute this
132 // call instruction at all times. If the displacement field is aligned
133 // we can simply rely on atomicity of 32-bit writes to make sure other threads
134 // will see no intermediate states. Otherwise, the first two bytes of the
135 // call are guaranteed to be aligned, and can be atomically patched to a
136 // self-loop to guard the instruction while we change the other bytes.
137
138 // We cannot rely on locks here, since the free-running threads must run at
139 // full speed.
140 //
141 // Used in the runtime linkage of calls; see class CompiledIC.
142 // (Cf. 4506997 and 4479829, where threads witnessed garbage displacements.)
143 void NativeCall::set_destination_mt_safe(address dest) {
144 debug_only(verify());
145 // Make sure patching code is locked. No two threads can patch at the same
146 // time but one may be executing this code.
147 assert(Patching_lock->is_locked() ||
148 SafepointSynchronize::is_at_safepoint(), "concurrent code patching");
149 // Both C1 and C2 should now be generating code which aligns the patched address
150 // to be within a single cache line except that C1 does not do the alignment on
151 // uniprocessor systems.
152 bool is_aligned = ((uintptr_t)displacement_address() + 0) / cache_line_size ==
153 ((uintptr_t)displacement_address() + 3) / cache_line_size;
154
155 guarantee(!os::is_MP() || is_aligned, "destination must be aligned");
156
157 if (is_aligned) {
158 // Simple case: The destination lies within a single cache line.
159 set_destination(dest);
160 } else if ((uintptr_t)instruction_address() / cache_line_size ==
161 ((uintptr_t)instruction_address()+1) / cache_line_size) {
162 // Tricky case: The instruction prefix lies within a single cache line.
163 intptr_t disp = dest - return_address();
164 #ifdef AMD64
165 guarantee(disp == (intptr_t)(jint)disp, "must be 32-bit offset");
166 #endif // AMD64
167
168 int call_opcode = instruction_address()[0];
169
170 // First patch dummy jump in place:
171 {
172 u_char patch_jump[2];
173 patch_jump[0] = 0xEB; // jmp rel8
174 patch_jump[1] = 0xFE; // jmp to self
175
176 assert(sizeof(patch_jump)==sizeof(short), "sanity check");
177 *(short*)instruction_address() = *(short*)patch_jump;
178 }
179 // Invalidate. Opteron requires a flush after every write.
180 wrote(0);
181
182 // (Note: We assume any reader which has already started to read
183 // the unpatched call will completely read the whole unpatched call
184 // without seeing the next writes we are about to make.)
185
186 // Next, patch the last three bytes:
187 u_char patch_disp[5];
188 patch_disp[0] = call_opcode;
189 *(int32_t*)&patch_disp[1] = (int32_t)disp;
190 assert(sizeof(patch_disp)==instruction_size, "sanity check");
191 for (int i = sizeof(short); i < instruction_size; i++)
192 instruction_address()[i] = patch_disp[i];
193
194 // Invalidate. Opteron requires a flush after every write.
195 wrote(sizeof(short));
196
197 // (Note: We assume that any reader which reads the opcode we are
198 // about to repatch will also read the writes we just made.)
199
200 // Finally, overwrite the jump:
201 *(short*)instruction_address() = *(short*)patch_disp;
202 // Invalidate. Opteron requires a flush after every write.
203 wrote(0);
204
205 debug_only(verify());
206 guarantee(destination() == dest, "patch succeeded");
207 } else {
208 // Impossible: One or the other must be atomically writable.
209 ShouldNotReachHere();
210 }
211 }
212
213
214 void NativeMovConstReg::verify() {
215 #ifdef AMD64
216 // make sure code pattern is actually a mov reg64, imm64 instruction
217 if ((ubyte_at(0) != Assembler::REX_W && ubyte_at(0) != Assembler::REX_WB) ||
218 (ubyte_at(1) & (0xff ^ register_mask)) != 0xB8) {
219 print();
220 fatal("not a REX.W[B] mov reg64, imm64");
221 }
222 #else
223 // make sure code pattern is actually a mov reg, imm32 instruction
224 u_char test_byte = *(u_char*)instruction_address();
225 u_char test_byte_2 = test_byte & ( 0xff ^ register_mask);
226 if (test_byte_2 != instruction_code) fatal("not a mov reg, imm32");
227 #endif // AMD64
228 }
229
230
231 void NativeMovConstReg::print() {
232 tty->print_cr(PTR_FORMAT ": mov reg, " INTPTR_FORMAT,
233 instruction_address(), data());
234 }
235
236 //-------------------------------------------------------------------
237
238 int NativeMovRegMem::instruction_start() const {
239 int off = 0;
240 u_char instr_0 = ubyte_at(off);
241
242 // See comment in Assembler::locate_operand() about VEX prefixes.
243 if (instr_0 == instruction_VEX_prefix_2bytes) {
244 assert((UseAVX > 0), "shouldn't have VEX prefix");
245 NOT_LP64(assert((0xC0 & ubyte_at(1)) == 0xC0, "shouldn't have LDS and LES instructions"));
246 return 2;
247 }
248 if (instr_0 == instruction_VEX_prefix_3bytes) {
249 assert((UseAVX > 0), "shouldn't have VEX prefix");
250 NOT_LP64(assert((0xC0 & ubyte_at(1)) == 0xC0, "shouldn't have LDS and LES instructions"));
251 return 3;
252 }
253
254 // First check to see if we have a (prefixed or not) xor
255 if (instr_0 >= instruction_prefix_wide_lo && // 0x40
256 instr_0 <= instruction_prefix_wide_hi) { // 0x4f
257 off++;
258 instr_0 = ubyte_at(off);
259 }
260
261 if (instr_0 == instruction_code_xor) {
262 off += 2;
263 instr_0 = ubyte_at(off);
264 }
265
266 // Now look for the real instruction and the many prefix/size specifiers.
267
268 if (instr_0 == instruction_operandsize_prefix ) { // 0x66
269 off++; // Not SSE instructions
270 instr_0 = ubyte_at(off);
271 }
272
273 if ( instr_0 == instruction_code_xmm_ss_prefix || // 0xf3
274 instr_0 == instruction_code_xmm_sd_prefix) { // 0xf2
275 off++;
276 instr_0 = ubyte_at(off);
277 }
278
279 if ( instr_0 >= instruction_prefix_wide_lo && // 0x40
280 instr_0 <= instruction_prefix_wide_hi) { // 0x4f
281 off++;
282 instr_0 = ubyte_at(off);
283 }
284
285
286 if (instr_0 == instruction_extended_prefix ) { // 0x0f
287 off++;
288 }
289
290 return off;
291 }
292
293 address NativeMovRegMem::instruction_address() const {
294 return addr_at(instruction_start());
295 }
296
297 address NativeMovRegMem::next_instruction_address() const {
298 address ret = instruction_address() + instruction_size;
299 u_char instr_0 = *(u_char*) instruction_address();
300 switch (instr_0) {
301 case instruction_operandsize_prefix:
302
303 fatal("should have skipped instruction_operandsize_prefix");
304 break;
305
306 case instruction_extended_prefix:
307 fatal("should have skipped instruction_extended_prefix");
308 break;
309
310 case instruction_code_mem2reg_movslq: // 0x63
311 case instruction_code_mem2reg_movzxb: // 0xB6
312 case instruction_code_mem2reg_movsxb: // 0xBE
313 case instruction_code_mem2reg_movzxw: // 0xB7
314 case instruction_code_mem2reg_movsxw: // 0xBF
315 case instruction_code_reg2mem: // 0x89 (q/l)
316 case instruction_code_mem2reg: // 0x8B (q/l)
317 case instruction_code_reg2memb: // 0x88
318 case instruction_code_mem2regb: // 0x8a
319
320 case instruction_code_float_s: // 0xd9 fld_s a
321 case instruction_code_float_d: // 0xdd fld_d a
322
323 case instruction_code_xmm_load: // 0x10
324 case instruction_code_xmm_store: // 0x11
325 case instruction_code_xmm_lpd: // 0x12
326 {
327 // If there is an SIB then instruction is longer than expected
328 u_char mod_rm = *(u_char*)(instruction_address() + 1);
329 if ((mod_rm & 7) == 0x4) {
330 ret++;
331 }
332 }
333 case instruction_code_xor:
334 fatal("should have skipped xor lead in");
335 break;
336
337 default:
338 fatal("not a NativeMovRegMem");
339 }
340 return ret;
341
342 }
343
344 int NativeMovRegMem::offset() const{
345 int off = data_offset + instruction_start();
346 u_char mod_rm = *(u_char*)(instruction_address() + 1);
347 // nnnn(r12|rsp) isn't coded as simple mod/rm since that is
348 // the encoding to use an SIB byte. Which will have the nnnn
349 // field off by one byte
350 if ((mod_rm & 7) == 0x4) {
351 off++;
352 }
353 return int_at(off);
354 }
355
356 void NativeMovRegMem::set_offset(int x) {
357 int off = data_offset + instruction_start();
358 u_char mod_rm = *(u_char*)(instruction_address() + 1);
359 // nnnn(r12|rsp) isn't coded as simple mod/rm since that is
360 // the encoding to use an SIB byte. Which will have the nnnn
361 // field off by one byte
362 if ((mod_rm & 7) == 0x4) {
363 off++;
364 }
365 set_int_at(off, x);
366 }
367
368 void NativeMovRegMem::verify() {
369 // make sure code pattern is actually a mov [reg+offset], reg instruction
370 u_char test_byte = *(u_char*)instruction_address();
371 switch (test_byte) {
372 case instruction_code_reg2memb: // 0x88 movb a, r
373 case instruction_code_reg2mem: // 0x89 movl a, r (can be movq in 64bit)
374 case instruction_code_mem2regb: // 0x8a movb r, a
375 case instruction_code_mem2reg: // 0x8b movl r, a (can be movq in 64bit)
376 break;
377
378 case instruction_code_mem2reg_movslq: // 0x63 movsql r, a
379 case instruction_code_mem2reg_movzxb: // 0xb6 movzbl r, a (movzxb)
380 case instruction_code_mem2reg_movzxw: // 0xb7 movzwl r, a (movzxw)
381 case instruction_code_mem2reg_movsxb: // 0xbe movsbl r, a (movsxb)
382 case instruction_code_mem2reg_movsxw: // 0xbf movswl r, a (movsxw)
383 break;
384
385 case instruction_code_float_s: // 0xd9 fld_s a
386 case instruction_code_float_d: // 0xdd fld_d a
387 case instruction_code_xmm_load: // 0x10 movsd xmm, a
388 case instruction_code_xmm_store: // 0x11 movsd a, xmm
389 case instruction_code_xmm_lpd: // 0x12 movlpd xmm, a
390 break;
391
392 default:
393 fatal ("not a mov [reg+offs], reg instruction");
394 }
395 }
396
397
398 void NativeMovRegMem::print() {
399 tty->print_cr("0x%x: mov reg, [reg + %x]", instruction_address(), offset());
400 }
401
402 //-------------------------------------------------------------------
403
404 void NativeLoadAddress::verify() {
405 // make sure code pattern is actually a mov [reg+offset], reg instruction
406 u_char test_byte = *(u_char*)instruction_address();
407 #ifdef _LP64
408 if ( (test_byte == instruction_prefix_wide ||
409 test_byte == instruction_prefix_wide_extended) ) {
410 test_byte = *(u_char*)(instruction_address() + 1);
411 }
412 #endif // _LP64
413 if ( ! ((test_byte == lea_instruction_code)
414 LP64_ONLY(|| (test_byte == mov64_instruction_code) ))) {
415 fatal ("not a lea reg, [reg+offs] instruction");
416 }
417 }
418
419
420 void NativeLoadAddress::print() {
421 tty->print_cr("0x%x: lea [reg + %x], reg", instruction_address(), offset());
422 }
423
424 //--------------------------------------------------------------------------------
425
426 void NativeJump::verify() {
427 if (*(u_char*)instruction_address() != instruction_code) {
428 fatal("not a jump instruction");
429 }
430 }
431
432
433 void NativeJump::insert(address code_pos, address entry) {
434 intptr_t disp = (intptr_t)entry - ((intptr_t)code_pos + 1 + 4);
435 #ifdef AMD64
436 guarantee(disp == (intptr_t)(int32_t)disp, "must be 32-bit offset");
437 #endif // AMD64
438
439 *code_pos = instruction_code;
440 *((int32_t*)(code_pos + 1)) = (int32_t)disp;
441
442 ICache::invalidate_range(code_pos, instruction_size);
443 }
444
445 void NativeJump::check_verified_entry_alignment(address entry, address verified_entry) {
446 // Patching to not_entrant can happen while activations of the method are
447 // in use. The patching in that instance must happen only when certain
448 // alignment restrictions are true. These guarantees check those
449 // conditions.
450 #ifdef AMD64
451 const int linesize = 64;
452 #else
453 const int linesize = 32;
454 #endif // AMD64
455
456 // Must be wordSize aligned
457 guarantee(((uintptr_t) verified_entry & (wordSize -1)) == 0,
458 "illegal address for code patching 2");
459 // First 5 bytes must be within the same cache line - 4827828
460 guarantee((uintptr_t) verified_entry / linesize ==
461 ((uintptr_t) verified_entry + 4) / linesize,
462 "illegal address for code patching 3");
463 }
464
465
466 // MT safe inserting of a jump over an unknown instruction sequence (used by nmethod::makeZombie)
467 // The problem: jmp <dest> is a 5-byte instruction. Atomical write can be only with 4 bytes.
468 // First patches the first word atomically to be a jump to itself.
469 // Then patches the last byte and then atomically patches the first word (4-bytes),
470 // thus inserting the desired jump
471 // This code is mt-safe with the following conditions: entry point is 4 byte aligned,
472 // entry point is in same cache line as unverified entry point, and the instruction being
473 // patched is >= 5 byte (size of patch).
474 //
475 // In C2 the 5+ byte sized instruction is enforced by code in MachPrologNode::emit.
476 // In C1 the restriction is enforced by CodeEmitter::method_entry
477 //
478 void NativeJump::patch_verified_entry(address entry, address verified_entry, address dest) {
479 // complete jump instruction (to be inserted) is in code_buffer;
480 unsigned char code_buffer[5];
481 code_buffer[0] = instruction_code;
482 intptr_t disp = (intptr_t)dest - ((intptr_t)verified_entry + 1 + 4);
483 #ifdef AMD64
484 guarantee(disp == (intptr_t)(int32_t)disp, "must be 32-bit offset");
485 #endif // AMD64
486 *(int32_t*)(code_buffer + 1) = (int32_t)disp;
487
488 check_verified_entry_alignment(entry, verified_entry);
489
490 // Can't call nativeJump_at() because it's asserts jump exists
491 NativeJump* n_jump = (NativeJump*) verified_entry;
492
493 //First patch dummy jmp in place
494
495 unsigned char patch[4];
496 assert(sizeof(patch)==sizeof(int32_t), "sanity check");
497 patch[0] = 0xEB; // jmp rel8
498 patch[1] = 0xFE; // jmp to self
499 patch[2] = 0xEB;
500 patch[3] = 0xFE;
501
502 // First patch dummy jmp in place
503 *(int32_t*)verified_entry = *(int32_t *)patch;
504
505 n_jump->wrote(0);
506
507 // Patch 5th byte (from jump instruction)
508 verified_entry[4] = code_buffer[4];
509
510 n_jump->wrote(4);
511
512 // Patch bytes 0-3 (from jump instruction)
513 *(int32_t*)verified_entry = *(int32_t *)code_buffer;
514 // Invalidate. Opteron requires a flush after every write.
515 n_jump->wrote(0);
516
517 }
518
519 void NativePopReg::insert(address code_pos, Register reg) {
520 assert(reg->encoding() < 8, "no space for REX");
521 assert(NativePopReg::instruction_size == sizeof(char), "right address unit for update");
522 *code_pos = (u_char)(instruction_code | reg->encoding());
523 ICache::invalidate_range(code_pos, instruction_size);
524 }
525
526
527 void NativeIllegalInstruction::insert(address code_pos) {
528 assert(NativeIllegalInstruction::instruction_size == sizeof(short), "right address unit for update");
529 *(short *)code_pos = instruction_code;
530 ICache::invalidate_range(code_pos, instruction_size);
531 }
532
533 void NativeGeneralJump::verify() {
534 assert(((NativeInstruction *)this)->is_jump() ||
535 ((NativeInstruction *)this)->is_cond_jump(), "not a general jump instruction");
536 }
537
538
539 void NativeGeneralJump::insert_unconditional(address code_pos, address entry) {
540 intptr_t disp = (intptr_t)entry - ((intptr_t)code_pos + 1 + 4);
541 #ifdef AMD64
542 guarantee(disp == (intptr_t)(int32_t)disp, "must be 32-bit offset");
543 #endif // AMD64
544
545 *code_pos = unconditional_long_jump;
546 *((int32_t *)(code_pos+1)) = (int32_t) disp;
547 ICache::invalidate_range(code_pos, instruction_size);
548 }
549
550
551 // MT-safe patching of a long jump instruction.
552 // First patches first word of instruction to two jmp's that jmps to them
553 // selfs (spinlock). Then patches the last byte, and then atomicly replaces
554 // the jmp's with the first 4 byte of the new instruction.
555 void NativeGeneralJump::replace_mt_safe(address instr_addr, address code_buffer) {
556 assert (instr_addr != NULL, "illegal address for code patching (4)");
557 NativeGeneralJump* n_jump = nativeGeneralJump_at (instr_addr); // checking that it is a jump
558
559 // Temporary code
560 unsigned char patch[4];
561 assert(sizeof(patch)==sizeof(int32_t), "sanity check");
562 patch[0] = 0xEB; // jmp rel8
563 patch[1] = 0xFE; // jmp to self
564 patch[2] = 0xEB;
565 patch[3] = 0xFE;
566
567 // First patch dummy jmp in place
568 *(int32_t*)instr_addr = *(int32_t *)patch;
569 n_jump->wrote(0);
570
571 // Patch 4th byte
572 instr_addr[4] = code_buffer[4];
573
574 n_jump->wrote(4);
575
576 // Patch bytes 0-3
577 *(jint*)instr_addr = *(jint *)code_buffer;
578
579 n_jump->wrote(0);
580
581 #ifdef ASSERT
582 // verify patching
583 for ( int i = 0; i < instruction_size; i++) {
584 address ptr = (address)((intptr_t)code_buffer + i);
585 int a_byte = (*ptr) & 0xFF;
586 assert(*((address)((intptr_t)instr_addr + i)) == a_byte, "mt safe patching failed");
587 }
588 #endif
589
590 }
591
592
593
594 address NativeGeneralJump::jump_destination() const {
595 int op_code = ubyte_at(0);
596 bool is_rel32off = (op_code == 0xE9 || op_code == 0x0F);
597 int offset = (op_code == 0x0F) ? 2 : 1;
598 int length = offset + ((is_rel32off) ? 4 : 1);
599
600 if (is_rel32off)
601 return addr_at(0) + length + int_at(offset);
602 else
603 return addr_at(0) + length + sbyte_at(offset);
604 }
605
606 bool NativeInstruction::is_dtrace_trap() {
607 return (*(int32_t*)this & 0xff) == 0xcc;
608 }