Mercurial > hg > graal-compiler
annotate src/cpu/x86/vm/nativeInst_x86.cpp @ 144:e3729351c946
6697534: Premature GC and invalid lgrp selection with NUMA-aware allocator.
Summary: Don't move tops of the chunks in ensure_parsibility(). Handle the situation with Solaris when a machine has a locality group with no memory.
Reviewed-by: apetrusenko, jcoomes, ysr
author | iveresov |
---|---|
date | Fri, 09 May 2008 16:34:08 +0400 |
parents | 018d5b58dd4f |
children | d1605aabd0a1 |
rev | line source |
---|---|
0 | 1 /* |
2 * Copyright 1997-2007 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, | |
20 * CA 95054 USA or visit www.sun.com if you need additional information or | |
21 * have any questions. | |
22 * | |
23 */ | |
24 | |
25 # include "incls/_precompiled.incl" | |
26 # include "incls/_nativeInst_x86.cpp.incl" | |
27 | |
28 void NativeInstruction::wrote(int offset) { | |
29 ICache::invalidate_word(addr_at(offset)); | |
30 } | |
31 | |
32 | |
33 void NativeCall::verify() { | |
34 // Make sure code pattern is actually a call imm32 instruction. | |
35 int inst = ubyte_at(0); | |
36 if (inst != instruction_code) { | |
37 tty->print_cr("Addr: " INTPTR_FORMAT " Code: 0x%x", instruction_address(), | |
38 inst); | |
39 fatal("not a call disp32"); | |
40 } | |
41 } | |
42 | |
43 address NativeCall::destination() const { | |
44 // Getting the destination of a call isn't safe because that call can | |
45 // be getting patched while you're calling this. There's only special | |
46 // places where this can be called but not automatically verifiable by | |
47 // checking which locks are held. The solution is true atomic patching | |
48 // on x86, nyi. | |
49 return return_address() + displacement(); | |
50 } | |
51 | |
52 void NativeCall::print() { | |
53 tty->print_cr(PTR_FORMAT ": call " PTR_FORMAT, | |
54 instruction_address(), destination()); | |
55 } | |
56 | |
57 // Inserts a native call instruction at a given pc | |
58 void NativeCall::insert(address code_pos, address entry) { | |
59 intptr_t disp = (intptr_t)entry - ((intptr_t)code_pos + 1 + 4); | |
60 #ifdef AMD64 | |
61 guarantee(disp == (intptr_t)(jint)disp, "must be 32-bit offset"); | |
62 #endif // AMD64 | |
63 *code_pos = instruction_code; | |
64 *((int32_t *)(code_pos+1)) = (int32_t) disp; | |
65 ICache::invalidate_range(code_pos, instruction_size); | |
66 } | |
67 | |
68 // MT-safe patching of a call instruction. | |
69 // First patches first word of instruction to two jmp's that jmps to them | |
70 // selfs (spinlock). Then patches the last byte, and then atomicly replaces | |
71 // the jmp's with the first 4 byte of the new instruction. | |
72 void NativeCall::replace_mt_safe(address instr_addr, address code_buffer) { | |
73 assert(Patching_lock->is_locked() || | |
74 SafepointSynchronize::is_at_safepoint(), "concurrent code patching"); | |
75 assert (instr_addr != NULL, "illegal address for code patching"); | |
76 | |
77 NativeCall* n_call = nativeCall_at (instr_addr); // checking that it is a call | |
78 if (os::is_MP()) { | |
79 guarantee((intptr_t)instr_addr % BytesPerWord == 0, "must be aligned"); | |
80 } | |
81 | |
82 // First patch dummy jmp in place | |
83 unsigned char patch[4]; | |
84 assert(sizeof(patch)==sizeof(jint), "sanity check"); | |
85 patch[0] = 0xEB; // jmp rel8 | |
86 patch[1] = 0xFE; // jmp to self | |
87 patch[2] = 0xEB; | |
88 patch[3] = 0xFE; | |
89 | |
90 // First patch dummy jmp in place | |
91 *(jint*)instr_addr = *(jint *)patch; | |
92 | |
93 // Invalidate. Opteron requires a flush after every write. | |
94 n_call->wrote(0); | |
95 | |
96 // Patch 4th byte | |
97 instr_addr[4] = code_buffer[4]; | |
98 | |
99 n_call->wrote(4); | |
100 | |
101 // Patch bytes 0-3 | |
102 *(jint*)instr_addr = *(jint *)code_buffer; | |
103 | |
104 n_call->wrote(0); | |
105 | |
106 #ifdef ASSERT | |
107 // verify patching | |
108 for ( int i = 0; i < instruction_size; i++) { | |
109 address ptr = (address)((intptr_t)code_buffer + i); | |
110 int a_byte = (*ptr) & 0xFF; | |
111 assert(*((address)((intptr_t)instr_addr + i)) == a_byte, "mt safe patching failed"); | |
112 } | |
113 #endif | |
114 | |
115 } | |
116 | |
117 | |
118 // Similar to replace_mt_safe, but just changes the destination. The | |
119 // important thing is that free-running threads are able to execute this | |
120 // call instruction at all times. If the displacement field is aligned | |
121 // we can simply rely on atomicity of 32-bit writes to make sure other threads | |
122 // will see no intermediate states. Otherwise, the first two bytes of the | |
123 // call are guaranteed to be aligned, and can be atomically patched to a | |
124 // self-loop to guard the instruction while we change the other bytes. | |
125 | |
126 // We cannot rely on locks here, since the free-running threads must run at | |
127 // full speed. | |
128 // | |
129 // Used in the runtime linkage of calls; see class CompiledIC. | |
130 // (Cf. 4506997 and 4479829, where threads witnessed garbage displacements.) | |
131 void NativeCall::set_destination_mt_safe(address dest) { | |
132 debug_only(verify()); | |
133 // Make sure patching code is locked. No two threads can patch at the same | |
134 // time but one may be executing this code. | |
135 assert(Patching_lock->is_locked() || | |
136 SafepointSynchronize::is_at_safepoint(), "concurrent code patching"); | |
137 // Both C1 and C2 should now be generating code which aligns the patched address | |
138 // to be within a single cache line except that C1 does not do the alignment on | |
139 // uniprocessor systems. | |
140 bool is_aligned = ((uintptr_t)displacement_address() + 0) / cache_line_size == | |
141 ((uintptr_t)displacement_address() + 3) / cache_line_size; | |
142 | |
143 guarantee(!os::is_MP() || is_aligned, "destination must be aligned"); | |
144 | |
145 if (is_aligned) { | |
146 // Simple case: The destination lies within a single cache line. | |
147 set_destination(dest); | |
148 } else if ((uintptr_t)instruction_address() / cache_line_size == | |
149 ((uintptr_t)instruction_address()+1) / cache_line_size) { | |
150 // Tricky case: The instruction prefix lies within a single cache line. | |
151 intptr_t disp = dest - return_address(); | |
152 #ifdef AMD64 | |
153 guarantee(disp == (intptr_t)(jint)disp, "must be 32-bit offset"); | |
154 #endif // AMD64 | |
155 | |
156 int call_opcode = instruction_address()[0]; | |
157 | |
158 // First patch dummy jump in place: | |
159 { | |
160 u_char patch_jump[2]; | |
161 patch_jump[0] = 0xEB; // jmp rel8 | |
162 patch_jump[1] = 0xFE; // jmp to self | |
163 | |
164 assert(sizeof(patch_jump)==sizeof(short), "sanity check"); | |
165 *(short*)instruction_address() = *(short*)patch_jump; | |
166 } | |
167 // Invalidate. Opteron requires a flush after every write. | |
168 wrote(0); | |
169 | |
170 // (Note: We assume any reader which has already started to read | |
171 // the unpatched call will completely read the whole unpatched call | |
172 // without seeing the next writes we are about to make.) | |
173 | |
174 // Next, patch the last three bytes: | |
175 u_char patch_disp[5]; | |
176 patch_disp[0] = call_opcode; | |
177 *(int32_t*)&patch_disp[1] = (int32_t)disp; | |
178 assert(sizeof(patch_disp)==instruction_size, "sanity check"); | |
179 for (int i = sizeof(short); i < instruction_size; i++) | |
180 instruction_address()[i] = patch_disp[i]; | |
181 | |
182 // Invalidate. Opteron requires a flush after every write. | |
183 wrote(sizeof(short)); | |
184 | |
185 // (Note: We assume that any reader which reads the opcode we are | |
186 // about to repatch will also read the writes we just made.) | |
187 | |
188 // Finally, overwrite the jump: | |
189 *(short*)instruction_address() = *(short*)patch_disp; | |
190 // Invalidate. Opteron requires a flush after every write. | |
191 wrote(0); | |
192 | |
193 debug_only(verify()); | |
194 guarantee(destination() == dest, "patch succeeded"); | |
195 } else { | |
196 // Impossible: One or the other must be atomically writable. | |
197 ShouldNotReachHere(); | |
198 } | |
199 } | |
200 | |
201 | |
202 void NativeMovConstReg::verify() { | |
203 #ifdef AMD64 | |
204 // make sure code pattern is actually a mov reg64, imm64 instruction | |
205 if ((ubyte_at(0) != Assembler::REX_W && ubyte_at(0) != Assembler::REX_WB) || | |
206 (ubyte_at(1) & (0xff ^ register_mask)) != 0xB8) { | |
207 print(); | |
208 fatal("not a REX.W[B] mov reg64, imm64"); | |
209 } | |
210 #else | |
211 // make sure code pattern is actually a mov reg, imm32 instruction | |
212 u_char test_byte = *(u_char*)instruction_address(); | |
213 u_char test_byte_2 = test_byte & ( 0xff ^ register_mask); | |
214 if (test_byte_2 != instruction_code) fatal("not a mov reg, imm32"); | |
215 #endif // AMD64 | |
216 } | |
217 | |
218 | |
219 void NativeMovConstReg::print() { | |
220 tty->print_cr(PTR_FORMAT ": mov reg, " INTPTR_FORMAT, | |
221 instruction_address(), data()); | |
222 } | |
223 | |
224 //------------------------------------------------------------------- | |
225 | |
226 #ifndef AMD64 | |
227 | |
228 void NativeMovRegMem::copy_instruction_to(address new_instruction_address) { | |
229 int inst_size = instruction_size; | |
230 | |
231 // See if there's an instruction size prefix override. | |
232 if ( *(address(this)) == instruction_operandsize_prefix && | |
233 *(address(this)+1) != instruction_code_xmm_code ) { // Not SSE instr | |
234 inst_size += 1; | |
235 } | |
236 if ( *(address(this)) == instruction_extended_prefix ) inst_size += 1; | |
237 | |
238 for (int i = 0; i < instruction_size; i++) { | |
239 *(new_instruction_address + i) = *(address(this) + i); | |
240 } | |
241 } | |
242 | |
243 void NativeMovRegMem::verify() { | |
244 // make sure code pattern is actually a mov [reg+offset], reg instruction | |
245 u_char test_byte = *(u_char*)instruction_address(); | |
246 if ( ! ( (test_byte == instruction_code_reg2memb) | |
247 || (test_byte == instruction_code_mem2regb) | |
248 || (test_byte == instruction_code_mem2regl) | |
249 || (test_byte == instruction_code_reg2meml) | |
250 || (test_byte == instruction_code_mem2reg_movzxb ) | |
251 || (test_byte == instruction_code_mem2reg_movzxw ) | |
252 || (test_byte == instruction_code_mem2reg_movsxb ) | |
253 || (test_byte == instruction_code_mem2reg_movsxw ) | |
254 || (test_byte == instruction_code_float_s) | |
255 || (test_byte == instruction_code_float_d) | |
256 || (test_byte == instruction_code_long_volatile) ) ) | |
257 { | |
258 u_char byte1 = ((u_char*)instruction_address())[1]; | |
259 u_char byte2 = ((u_char*)instruction_address())[2]; | |
260 if ((test_byte != instruction_code_xmm_ss_prefix && | |
261 test_byte != instruction_code_xmm_sd_prefix && | |
262 test_byte != instruction_operandsize_prefix) || | |
263 byte1 != instruction_code_xmm_code || | |
264 (byte2 != instruction_code_xmm_load && | |
265 byte2 != instruction_code_xmm_lpd && | |
266 byte2 != instruction_code_xmm_store)) { | |
267 fatal ("not a mov [reg+offs], reg instruction"); | |
268 } | |
269 } | |
270 } | |
271 | |
272 | |
273 void NativeMovRegMem::print() { | |
274 tty->print_cr("0x%x: mov reg, [reg + %x]", instruction_address(), offset()); | |
275 } | |
276 | |
277 //------------------------------------------------------------------- | |
278 | |
279 void NativeLoadAddress::verify() { | |
280 // make sure code pattern is actually a mov [reg+offset], reg instruction | |
281 u_char test_byte = *(u_char*)instruction_address(); | |
282 if ( ! (test_byte == instruction_code) ) { | |
283 fatal ("not a lea reg, [reg+offs] instruction"); | |
284 } | |
285 } | |
286 | |
287 | |
288 void NativeLoadAddress::print() { | |
289 tty->print_cr("0x%x: lea [reg + %x], reg", instruction_address(), offset()); | |
290 } | |
291 | |
292 #endif // !AMD64 | |
293 | |
294 //-------------------------------------------------------------------------------- | |
295 | |
296 void NativeJump::verify() { | |
297 if (*(u_char*)instruction_address() != instruction_code) { | |
298 fatal("not a jump instruction"); | |
299 } | |
300 } | |
301 | |
302 | |
303 void NativeJump::insert(address code_pos, address entry) { | |
304 intptr_t disp = (intptr_t)entry - ((intptr_t)code_pos + 1 + 4); | |
305 #ifdef AMD64 | |
306 guarantee(disp == (intptr_t)(int32_t)disp, "must be 32-bit offset"); | |
307 #endif // AMD64 | |
308 | |
309 *code_pos = instruction_code; | |
310 *((int32_t*)(code_pos + 1)) = (int32_t)disp; | |
311 | |
312 ICache::invalidate_range(code_pos, instruction_size); | |
313 } | |
314 | |
315 void NativeJump::check_verified_entry_alignment(address entry, address verified_entry) { | |
316 // Patching to not_entrant can happen while activations of the method are | |
317 // in use. The patching in that instance must happen only when certain | |
318 // alignment restrictions are true. These guarantees check those | |
319 // conditions. | |
320 #ifdef AMD64 | |
321 const int linesize = 64; | |
322 #else | |
323 const int linesize = 32; | |
324 #endif // AMD64 | |
325 | |
326 // Must be wordSize aligned | |
327 guarantee(((uintptr_t) verified_entry & (wordSize -1)) == 0, | |
328 "illegal address for code patching 2"); | |
329 // First 5 bytes must be within the same cache line - 4827828 | |
330 guarantee((uintptr_t) verified_entry / linesize == | |
331 ((uintptr_t) verified_entry + 4) / linesize, | |
332 "illegal address for code patching 3"); | |
333 } | |
334 | |
335 | |
336 // MT safe inserting of a jump over an unknown instruction sequence (used by nmethod::makeZombie) | |
337 // The problem: jmp <dest> is a 5-byte instruction. Atomical write can be only with 4 bytes. | |
338 // First patches the first word atomically to be a jump to itself. | |
339 // Then patches the last byte and then atomically patches the first word (4-bytes), | |
340 // thus inserting the desired jump | |
341 // This code is mt-safe with the following conditions: entry point is 4 byte aligned, | |
342 // entry point is in same cache line as unverified entry point, and the instruction being | |
343 // patched is >= 5 byte (size of patch). | |
344 // | |
345 // In C2 the 5+ byte sized instruction is enforced by code in MachPrologNode::emit. | |
346 // In C1 the restriction is enforced by CodeEmitter::method_entry | |
347 // | |
348 void NativeJump::patch_verified_entry(address entry, address verified_entry, address dest) { | |
349 // complete jump instruction (to be inserted) is in code_buffer; | |
350 unsigned char code_buffer[5]; | |
351 code_buffer[0] = instruction_code; | |
352 intptr_t disp = (intptr_t)dest - ((intptr_t)verified_entry + 1 + 4); | |
353 #ifdef AMD64 | |
354 guarantee(disp == (intptr_t)(int32_t)disp, "must be 32-bit offset"); | |
355 #endif // AMD64 | |
356 *(int32_t*)(code_buffer + 1) = (int32_t)disp; | |
357 | |
358 check_verified_entry_alignment(entry, verified_entry); | |
359 | |
360 // Can't call nativeJump_at() because it's asserts jump exists | |
361 NativeJump* n_jump = (NativeJump*) verified_entry; | |
362 | |
363 //First patch dummy jmp in place | |
364 | |
365 unsigned char patch[4]; | |
366 assert(sizeof(patch)==sizeof(int32_t), "sanity check"); | |
367 patch[0] = 0xEB; // jmp rel8 | |
368 patch[1] = 0xFE; // jmp to self | |
369 patch[2] = 0xEB; | |
370 patch[3] = 0xFE; | |
371 | |
372 // First patch dummy jmp in place | |
373 *(int32_t*)verified_entry = *(int32_t *)patch; | |
374 | |
375 n_jump->wrote(0); | |
376 | |
377 // Patch 5th byte (from jump instruction) | |
378 verified_entry[4] = code_buffer[4]; | |
379 | |
380 n_jump->wrote(4); | |
381 | |
382 // Patch bytes 0-3 (from jump instruction) | |
383 *(int32_t*)verified_entry = *(int32_t *)code_buffer; | |
384 // Invalidate. Opteron requires a flush after every write. | |
385 n_jump->wrote(0); | |
386 | |
387 } | |
388 | |
389 void NativePopReg::insert(address code_pos, Register reg) { | |
390 assert(reg->encoding() < 8, "no space for REX"); | |
391 assert(NativePopReg::instruction_size == sizeof(char), "right address unit for update"); | |
392 *code_pos = (u_char)(instruction_code | reg->encoding()); | |
393 ICache::invalidate_range(code_pos, instruction_size); | |
394 } | |
395 | |
396 | |
397 void NativeIllegalInstruction::insert(address code_pos) { | |
398 assert(NativeIllegalInstruction::instruction_size == sizeof(short), "right address unit for update"); | |
399 *(short *)code_pos = instruction_code; | |
400 ICache::invalidate_range(code_pos, instruction_size); | |
401 } | |
402 | |
403 void NativeGeneralJump::verify() { | |
404 assert(((NativeInstruction *)this)->is_jump() || | |
405 ((NativeInstruction *)this)->is_cond_jump(), "not a general jump instruction"); | |
406 } | |
407 | |
408 | |
409 void NativeGeneralJump::insert_unconditional(address code_pos, address entry) { | |
410 intptr_t disp = (intptr_t)entry - ((intptr_t)code_pos + 1 + 4); | |
411 #ifdef AMD64 | |
412 guarantee(disp == (intptr_t)(int32_t)disp, "must be 32-bit offset"); | |
413 #endif // AMD64 | |
414 | |
415 *code_pos = unconditional_long_jump; | |
416 *((int32_t *)(code_pos+1)) = (int32_t) disp; | |
417 ICache::invalidate_range(code_pos, instruction_size); | |
418 } | |
419 | |
420 | |
421 // MT-safe patching of a long jump instruction. | |
422 // First patches first word of instruction to two jmp's that jmps to them | |
423 // selfs (spinlock). Then patches the last byte, and then atomicly replaces | |
424 // the jmp's with the first 4 byte of the new instruction. | |
425 void NativeGeneralJump::replace_mt_safe(address instr_addr, address code_buffer) { | |
426 assert (instr_addr != NULL, "illegal address for code patching (4)"); | |
427 NativeGeneralJump* n_jump = nativeGeneralJump_at (instr_addr); // checking that it is a jump | |
428 | |
429 // Temporary code | |
430 unsigned char patch[4]; | |
431 assert(sizeof(patch)==sizeof(int32_t), "sanity check"); | |
432 patch[0] = 0xEB; // jmp rel8 | |
433 patch[1] = 0xFE; // jmp to self | |
434 patch[2] = 0xEB; | |
435 patch[3] = 0xFE; | |
436 | |
437 // First patch dummy jmp in place | |
438 *(int32_t*)instr_addr = *(int32_t *)patch; | |
439 n_jump->wrote(0); | |
440 | |
441 // Patch 4th byte | |
442 instr_addr[4] = code_buffer[4]; | |
443 | |
444 n_jump->wrote(4); | |
445 | |
446 // Patch bytes 0-3 | |
447 *(jint*)instr_addr = *(jint *)code_buffer; | |
448 | |
449 n_jump->wrote(0); | |
450 | |
451 #ifdef ASSERT | |
452 // verify patching | |
453 for ( int i = 0; i < instruction_size; i++) { | |
454 address ptr = (address)((intptr_t)code_buffer + i); | |
455 int a_byte = (*ptr) & 0xFF; | |
456 assert(*((address)((intptr_t)instr_addr + i)) == a_byte, "mt safe patching failed"); | |
457 } | |
458 #endif | |
459 | |
460 } | |
461 | |
462 | |
463 | |
464 address NativeGeneralJump::jump_destination() const { | |
465 int op_code = ubyte_at(0); | |
466 bool is_rel32off = (op_code == 0xE9 || op_code == 0x0F); | |
467 int offset = (op_code == 0x0F) ? 2 : 1; | |
468 int length = offset + ((is_rel32off) ? 4 : 1); | |
469 | |
470 if (is_rel32off) | |
471 return addr_at(0) + length + int_at(offset); | |
472 else | |
473 return addr_at(0) + length + sbyte_at(offset); | |
474 } | |
116
018d5b58dd4f
6537506: Provide a mechanism for specifying Java-level USDT-like dtrace probes
kamg
parents:
0
diff
changeset
|
475 |
018d5b58dd4f
6537506: Provide a mechanism for specifying Java-level USDT-like dtrace probes
kamg
parents:
0
diff
changeset
|
476 bool NativeInstruction::is_dtrace_trap() { |
018d5b58dd4f
6537506: Provide a mechanism for specifying Java-level USDT-like dtrace probes
kamg
parents:
0
diff
changeset
|
477 return (*(int32_t*)this & 0xff) == 0xcc; |
018d5b58dd4f
6537506: Provide a mechanism for specifying Java-level USDT-like dtrace probes
kamg
parents:
0
diff
changeset
|
478 } |