Text file
src/runtime/race_ppc64le.s
1 // Copyright 2018 The Go Authors. All rights reserved.
2 // Use of this source code is governed by a BSD-style
3 // license that can be found in the LICENSE file.
4
5 //go:build race
6
7 #include "go_asm.h"
8 #include "go_tls.h"
9 #include "funcdata.h"
10 #include "textflag.h"
11 #include "asm_ppc64x.h"
12 #include "cgo/abi_ppc64x.h"
13
14 // The following functions allow calling the clang-compiled race runtime directly
15 // from Go code without going all the way through cgo.
16 // First, it's much faster (up to 50% speedup for real Go programs).
17 // Second, it eliminates race-related special cases from cgocall and scheduler.
18 // Third, in long-term it will allow to remove cyclic runtime/race dependency on cmd/go.
19
20 // A brief recap of the ppc64le calling convention.
21 // Arguments are passed in R3, R4, R5 ...
22 // SP must be 16-byte aligned.
23
24 // Note that for ppc64x, LLVM follows the standard ABI and
25 // expects arguments in registers, so these functions move
26 // the arguments from storage to the registers expected
27 // by the ABI.
28
29 // When calling from Go to Clang tsan code:
30 // R3 is the 1st argument and is usually the ThreadState*
31 // R4-? are the 2nd, 3rd, 4th, etc. arguments
32
33 // When calling racecalladdr:
34 // R8 is the call target address
35
36 // The race ctx is passed in R3 and loaded in
37 // racecalladdr.
38 //
39 // The sequence used to get the race ctx:
40 // MOVD runtime·tls_g(SB), R10 // Address of TLS variable
41 // MOVD 0(R10), g // g = R30
42 // MOVD g_racectx(g), R3 // racectx == ThreadState
43
44 // func runtime·RaceRead(addr uintptr)
45 // Called from instrumented Go code
46 TEXT runtime·raceread<ABIInternal>(SB), NOSPLIT, $0-8
47 MOVD R3, R4 // addr
48 MOVD LR, R5 // caller of this?
49 // void __tsan_read(ThreadState *thr, void *addr, void *pc);
50 MOVD $__tsan_read(SB), R8
51 BR racecalladdr<>(SB)
52
53 TEXT runtime·RaceRead(SB), NOSPLIT, $0-8
54 BR runtime·raceread(SB)
55
56 // void runtime·racereadpc(void *addr, void *callpc, void *pc)
57 TEXT runtime·racereadpc(SB), NOSPLIT, $0-24
58 MOVD addr+0(FP), R4
59 MOVD callpc+8(FP), R5
60 MOVD pc+16(FP), R6
61 // void __tsan_read_pc(ThreadState *thr, void *addr, void *callpc, void *pc);
62 MOVD $__tsan_read_pc(SB), R8
63 BR racecalladdr<>(SB)
64
65 // func runtime·RaceWrite(addr uintptr)
66 // Called from instrumented Go code
67 TEXT runtime·racewrite<ABIInternal>(SB), NOSPLIT, $0-8
68 MOVD R3, R4 // addr
69 MOVD LR, R5 // caller has set LR via BL inst
70 // void __tsan_write(ThreadState *thr, void *addr, void *pc);
71 MOVD $__tsan_write(SB), R8
72 BR racecalladdr<>(SB)
73
74 TEXT runtime·RaceWrite(SB), NOSPLIT, $0-8
75 JMP runtime·racewrite(SB)
76
77 // void runtime·racewritepc(void *addr, void *callpc, void *pc)
78 TEXT runtime·racewritepc(SB), NOSPLIT, $0-24
79 MOVD addr+0(FP), R4
80 MOVD callpc+8(FP), R5
81 MOVD pc+16(FP), R6
82 // void __tsan_write_pc(ThreadState *thr, void *addr, void *callpc, void *pc);
83 MOVD $__tsan_write_pc(SB), R8
84 BR racecalladdr<>(SB)
85
86 // func runtime·RaceReadRange(addr, size uintptr)
87 // Called from instrumented Go code.
88 TEXT runtime·racereadrange<ABIInternal>(SB), NOSPLIT, $0-16
89 MOVD R4, R5 // size
90 MOVD R3, R4 // addr
91 MOVD LR, R6
92 // void __tsan_read_range(ThreadState *thr, void *addr, uintptr size, void *pc);
93 MOVD $__tsan_read_range(SB), R8
94 BR racecalladdr<>(SB)
95
96 // void runtime·racereadrangepc1(void *addr, uintptr sz, void *pc)
97 TEXT runtime·racereadrangepc1(SB), NOSPLIT, $0-24
98 MOVD addr+0(FP), R4
99 MOVD size+8(FP), R5
100 MOVD pc+16(FP), R6
101 ADD $4, R6 // tsan wants return addr
102 // void __tsan_read_range(ThreadState *thr, void *addr, uintptr size, void *pc);
103 MOVD $__tsan_read_range(SB), R8
104 BR racecalladdr<>(SB)
105
106 TEXT runtime·RaceReadRange(SB), NOSPLIT, $0-16
107 BR runtime·racereadrange(SB)
108
109 // func runtime·RaceWriteRange(addr, size uintptr)
110 // Called from instrumented Go code.
111 TEXT runtime·racewriterange<ABIInternal>(SB), NOSPLIT, $0-16
112 MOVD R4, R5 // size
113 MOVD R3, R4 // addr
114 MOVD LR, R6
115 // void __tsan_write_range(ThreadState *thr, void *addr, uintptr size, void *pc);
116 MOVD $__tsan_write_range(SB), R8
117 BR racecalladdr<>(SB)
118
119 TEXT runtime·RaceWriteRange(SB), NOSPLIT, $0-16
120 BR runtime·racewriterange(SB)
121
122 // void runtime·racewriterangepc1(void *addr, uintptr sz, void *pc)
123 // Called from instrumented Go code
124 TEXT runtime·racewriterangepc1(SB), NOSPLIT, $0-24
125 MOVD addr+0(FP), R4
126 MOVD size+8(FP), R5
127 MOVD pc+16(FP), R6
128 ADD $4, R6 // add 4 to inst offset?
129 // void __tsan_write_range(ThreadState *thr, void *addr, uintptr size, void *pc);
130 MOVD $__tsan_write_range(SB), R8
131 BR racecalladdr<>(SB)
132
133 // Call a __tsan function from Go code.
134 // R8 = tsan function address
135 // R3 = *ThreadState a.k.a. g_racectx from g
136 // R4 = addr passed to __tsan function
137 //
138 // Otherwise, setup goroutine context and invoke racecall. Other arguments already set.
139 TEXT racecalladdr<>(SB), NOSPLIT, $0-0
140 MOVD runtime·tls_g(SB), R10
141 MOVD 0(R10), g
142 MOVD g_racectx(g), R3 // goroutine context
143 // Check that addr is within [arenastart, arenaend) or within [racedatastart, racedataend).
144 MOVD runtime·racearenastart(SB), R9
145 CMP R4, R9
146 BLT data
147 MOVD runtime·racearenaend(SB), R9
148 CMP R4, R9
149 BLT call
150 data:
151 MOVD runtime·racedatastart(SB), R9
152 CMP R4, R9
153 BLT ret
154 MOVD runtime·racedataend(SB), R9
155 CMP R4, R9
156 BGT ret
157 call:
158 // Careful!! racecall will save LR on its
159 // stack, which is OK as long as racecalladdr
160 // doesn't change in a way that generates a stack.
161 // racecall should return to the caller of
162 // recalladdr.
163 BR racecall<>(SB)
164 ret:
165 RET
166
167 // func runtime·racefuncenter(pc uintptr)
168 // Called from instrumented Go code.
169 TEXT runtime·racefuncenter(SB), NOSPLIT, $0-8
170 MOVD callpc+0(FP), R8
171 BR racefuncenter<>(SB)
172
173 // Common code for racefuncenter
174 // R11 = caller's return address
175 TEXT racefuncenter<>(SB), NOSPLIT, $0-0
176 MOVD runtime·tls_g(SB), R10
177 MOVD 0(R10), g
178 MOVD g_racectx(g), R3 // goroutine racectx aka *ThreadState
179 MOVD R8, R4 // caller pc set by caller in R8
180 // void __tsan_func_enter(ThreadState *thr, void *pc);
181 MOVD $__tsan_func_enter(SB), R8
182 BR racecall<>(SB)
183 RET
184
185 // func runtime·racefuncexit()
186 // Called from Go instrumented code.
187 TEXT runtime·racefuncexit(SB), NOSPLIT, $0-0
188 MOVD runtime·tls_g(SB), R10
189 MOVD 0(R10), g
190 MOVD g_racectx(g), R3 // goroutine racectx aka *ThreadState
191 // void __tsan_func_exit(ThreadState *thr);
192 MOVD $__tsan_func_exit(SB), R8
193 BR racecall<>(SB)
194
195 // Atomic operations for sync/atomic package.
196 // Some use the __tsan versions instead
197 // R6 = addr of arguments passed to this function
198 // R3, R4, R5 set in racecallatomic
199
200 // Load atomic in tsan
201 TEXT sync∕atomic·LoadInt32(SB), NOSPLIT, $0-12
202 GO_ARGS
203 // void __tsan_go_atomic32_load(ThreadState *thr, uptr cpc, uptr pc, u8 *a);
204 MOVD $__tsan_go_atomic32_load(SB), R8
205 ADD $32, R1, R6 // addr of caller's 1st arg
206 BR racecallatomic<>(SB)
207 RET
208
209 TEXT sync∕atomic·LoadInt64(SB), NOSPLIT, $0-16
210 GO_ARGS
211 // void __tsan_go_atomic64_load(ThreadState *thr, uptr cpc, uptr pc, u8 *a);
212 MOVD $__tsan_go_atomic64_load(SB), R8
213 ADD $32, R1, R6 // addr of caller's 1st arg
214 BR racecallatomic<>(SB)
215 RET
216
217 TEXT sync∕atomic·LoadUint32(SB), NOSPLIT, $0-12
218 GO_ARGS
219 BR sync∕atomic·LoadInt32(SB)
220
221 TEXT sync∕atomic·LoadUint64(SB), NOSPLIT, $0-16
222 GO_ARGS
223 BR sync∕atomic·LoadInt64(SB)
224
225 TEXT sync∕atomic·LoadUintptr(SB), NOSPLIT, $0-16
226 GO_ARGS
227 BR sync∕atomic·LoadInt64(SB)
228
229 TEXT sync∕atomic·LoadPointer(SB), NOSPLIT, $0-16
230 GO_ARGS
231 BR sync∕atomic·LoadInt64(SB)
232
233 // Store atomic in tsan
234 TEXT sync∕atomic·StoreInt32(SB), NOSPLIT, $0-12
235 GO_ARGS
236 // void __tsan_go_atomic32_store(ThreadState *thr, uptr cpc, uptr pc, u8 *a);
237 MOVD $__tsan_go_atomic32_store(SB), R8
238 ADD $32, R1, R6 // addr of caller's 1st arg
239 BR racecallatomic<>(SB)
240
241 TEXT sync∕atomic·StoreInt64(SB), NOSPLIT, $0-16
242 GO_ARGS
243 // void __tsan_go_atomic64_store(ThreadState *thr, uptr cpc, uptr pc, u8 *a);
244 MOVD $__tsan_go_atomic64_store(SB), R8
245 ADD $32, R1, R6 // addr of caller's 1st arg
246 BR racecallatomic<>(SB)
247
248 TEXT sync∕atomic·StoreUint32(SB), NOSPLIT, $0-12
249 GO_ARGS
250 BR sync∕atomic·StoreInt32(SB)
251
252 TEXT sync∕atomic·StoreUint64(SB), NOSPLIT, $0-16
253 GO_ARGS
254 BR sync∕atomic·StoreInt64(SB)
255
256 TEXT sync∕atomic·StoreUintptr(SB), NOSPLIT, $0-16
257 GO_ARGS
258 BR sync∕atomic·StoreInt64(SB)
259
260 // Swap in tsan
261 TEXT sync∕atomic·SwapInt32(SB), NOSPLIT, $0-20
262 GO_ARGS
263 // void __tsan_go_atomic32_exchange(ThreadState *thr, uptr cpc, uptr pc, u8 *a);
264 MOVD $__tsan_go_atomic32_exchange(SB), R8
265 ADD $32, R1, R6 // addr of caller's 1st arg
266 BR racecallatomic<>(SB)
267
268 TEXT sync∕atomic·SwapInt64(SB), NOSPLIT, $0-24
269 GO_ARGS
270 // void __tsan_go_atomic64_exchange(ThreadState *thr, uptr cpc, uptr pc, u8 *a)
271 MOVD $__tsan_go_atomic64_exchange(SB), R8
272 ADD $32, R1, R6 // addr of caller's 1st arg
273 BR racecallatomic<>(SB)
274
275 TEXT sync∕atomic·SwapUint32(SB), NOSPLIT, $0-20
276 GO_ARGS
277 BR sync∕atomic·SwapInt32(SB)
278
279 TEXT sync∕atomic·SwapUint64(SB), NOSPLIT, $0-24
280 GO_ARGS
281 BR sync∕atomic·SwapInt64(SB)
282
283 TEXT sync∕atomic·SwapUintptr(SB), NOSPLIT, $0-24
284 GO_ARGS
285 BR sync∕atomic·SwapInt64(SB)
286
287 // Add atomic in tsan
288 TEXT sync∕atomic·AddInt32(SB), NOSPLIT, $0-20
289 GO_ARGS
290 // void __tsan_go_atomic32_fetch_add(ThreadState *thr, uptr cpc, uptr pc, u8 *a);
291 MOVD $__tsan_go_atomic32_fetch_add(SB), R8
292 ADD $64, R1, R6 // addr of caller's 1st arg
293 BL racecallatomic<>(SB)
294 // The tsan fetch_add result is not as expected by Go,
295 // so the 'add' must be added to the result.
296 MOVW add+8(FP), R3 // The tsa fetch_add does not return the
297 MOVW ret+16(FP), R4 // result as expected by go, so fix it.
298 ADD R3, R4, R3
299 MOVW R3, ret+16(FP)
300 RET
301
302 TEXT sync∕atomic·AddInt64(SB), NOSPLIT, $0-24
303 GO_ARGS
304 // void __tsan_go_atomic64_fetch_add(ThreadState *thr, uptr cpc, uptr pc, u8 *a);
305 MOVD $__tsan_go_atomic64_fetch_add(SB), R8
306 ADD $64, R1, R6 // addr of caller's 1st arg
307 BL racecallatomic<>(SB)
308 // The tsan fetch_add result is not as expected by Go,
309 // so the 'add' must be added to the result.
310 MOVD add+8(FP), R3
311 MOVD ret+16(FP), R4
312 ADD R3, R4, R3
313 MOVD R3, ret+16(FP)
314 RET
315
316 TEXT sync∕atomic·AddUint32(SB), NOSPLIT, $0-20
317 GO_ARGS
318 BR sync∕atomic·AddInt32(SB)
319
320 TEXT sync∕atomic·AddUint64(SB), NOSPLIT, $0-24
321 GO_ARGS
322 BR sync∕atomic·AddInt64(SB)
323
324 TEXT sync∕atomic·AddUintptr(SB), NOSPLIT, $0-24
325 GO_ARGS
326 BR sync∕atomic·AddInt64(SB)
327
328 // CompareAndSwap in tsan
329 TEXT sync∕atomic·CompareAndSwapInt32(SB), NOSPLIT, $0-17
330 GO_ARGS
331 // void __tsan_go_atomic32_compare_exchange(
332 // ThreadState *thr, uptr cpc, uptr pc, u8 *a)
333 MOVD $__tsan_go_atomic32_compare_exchange(SB), R8
334 ADD $32, R1, R6 // addr of caller's 1st arg
335 BR racecallatomic<>(SB)
336
337 TEXT sync∕atomic·CompareAndSwapInt64(SB), NOSPLIT, $0-25
338 GO_ARGS
339 // void __tsan_go_atomic32_compare_exchange(
340 // ThreadState *thr, uptr cpc, uptr pc, u8 *a)
341 MOVD $__tsan_go_atomic64_compare_exchange(SB), R8
342 ADD $32, R1, R6 // addr of caller's 1st arg
343 BR racecallatomic<>(SB)
344
345 TEXT sync∕atomic·CompareAndSwapUint32(SB), NOSPLIT, $0-17
346 GO_ARGS
347 BR sync∕atomic·CompareAndSwapInt32(SB)
348
349 TEXT sync∕atomic·CompareAndSwapUint64(SB), NOSPLIT, $0-25
350 GO_ARGS
351 BR sync∕atomic·CompareAndSwapInt64(SB)
352
353 TEXT sync∕atomic·CompareAndSwapUintptr(SB), NOSPLIT, $0-25
354 GO_ARGS
355 BR sync∕atomic·CompareAndSwapInt64(SB)
356
357 // Common function used to call tsan's atomic functions
358 // R3 = *ThreadState
359 // R4 = TODO: What's this supposed to be?
360 // R5 = caller pc
361 // R6 = addr of incoming arg list
362 // R8 contains addr of target function.
363 TEXT racecallatomic<>(SB), NOSPLIT, $0-0
364 // Trigger SIGSEGV early if address passed to atomic function is bad.
365 MOVD (R6), R7 // 1st arg is addr
366 MOVB (R7), R9 // segv here if addr is bad
367 // Check that addr is within [arenastart, arenaend) or within [racedatastart, racedataend).
368 MOVD runtime·racearenastart(SB), R9
369 CMP R7, R9
370 BLT racecallatomic_data
371 MOVD runtime·racearenaend(SB), R9
372 CMP R7, R9
373 BLT racecallatomic_ok
374 racecallatomic_data:
375 MOVD runtime·racedatastart(SB), R9
376 CMP R7, R9
377 BLT racecallatomic_ignore
378 MOVD runtime·racedataend(SB), R9
379 CMP R7, R9
380 BGE racecallatomic_ignore
381 racecallatomic_ok:
382 // Addr is within the good range, call the atomic function.
383 MOVD runtime·tls_g(SB), R10
384 MOVD 0(R10), g
385 MOVD g_racectx(g), R3 // goroutine racectx aka *ThreadState
386 MOVD R8, R5 // pc is the function called
387 MOVD (R1), R4 // caller pc from stack
388 BL racecall<>(SB) // BL needed to maintain stack consistency
389 RET //
390 racecallatomic_ignore:
391 // Addr is outside the good range.
392 // Call __tsan_go_ignore_sync_begin to ignore synchronization during the atomic op.
393 // An attempt to synchronize on the address would cause crash.
394 MOVD R8, R15 // save the original function
395 MOVD R6, R17 // save the original arg list addr
396 MOVD $__tsan_go_ignore_sync_begin(SB), R8 // func addr to call
397 MOVD runtime·tls_g(SB), R10
398 MOVD 0(R10), g
399 MOVD g_racectx(g), R3 // goroutine context
400 BL racecall<>(SB)
401 MOVD R15, R8 // restore the original function
402 MOVD R17, R6 // restore arg list addr
403 // Call the atomic function.
404 // racecall will call LLVM race code which might clobber r30 (g)
405 MOVD runtime·tls_g(SB), R10
406 MOVD 0(R10), g
407
408 MOVD g_racectx(g), R3
409 MOVD R8, R4 // pc being called same TODO as above
410 MOVD (R1), R5 // caller pc from latest LR
411 BL racecall<>(SB)
412 // Call __tsan_go_ignore_sync_end.
413 MOVD $__tsan_go_ignore_sync_end(SB), R8
414 MOVD g_racectx(g), R3 // goroutine context g should still be good?
415 BL racecall<>(SB)
416 RET
417
418 // void runtime·racecall(void(*f)(...), ...)
419 // Calls C function f from race runtime and passes up to 4 arguments to it.
420 // The arguments are never heap-object-preserving pointers, so we pretend there are no arguments.
421 TEXT runtime·racecall(SB), NOSPLIT, $0-0
422 MOVD fn+0(FP), R8
423 MOVD arg0+8(FP), R3
424 MOVD arg1+16(FP), R4
425 MOVD arg2+24(FP), R5
426 MOVD arg3+32(FP), R6
427 JMP racecall<>(SB)
428
429 // Finds g0 and sets its stack
430 // Arguments were loaded for call from Go to C
431 TEXT racecall<>(SB), NOSPLIT, $0-0
432 // Set the LR slot for the ppc64 ABI
433 MOVD LR, R10
434 MOVD R10, 0(R1) // Go expectation
435 MOVD R10, 16(R1) // C ABI
436 // Get info from the current goroutine
437 MOVD runtime·tls_g(SB), R10 // g offset in TLS
438 MOVD 0(R10), g
439 MOVD g_m(g), R7 // m for g
440 MOVD R1, R16 // callee-saved, preserved across C call
441 MOVD m_g0(R7), R10 // g0 for m
442 CMP R10, g // same g0?
443 BEQ call // already on g0
444 MOVD (g_sched+gobuf_sp)(R10), R1 // switch R1
445 call:
446 // prepare frame for C ABI
447 SUB $32, R1 // create frame for callee saving LR, CR, R2 etc.
448 RLDCR $0, R1, $~15, R1 // align SP to 16 bytes
449 MOVD R8, CTR // R8 = caller addr
450 MOVD R8, R12 // expected by PPC64 ABI
451 BL (CTR)
452 XOR R0, R0 // clear R0 on return from Clang
453 MOVD R16, R1 // restore R1; R16 nonvol in Clang
454 MOVD runtime·tls_g(SB), R10 // find correct g
455 MOVD 0(R10), g
456 MOVD 16(R1), R10 // LR was saved away, restore for return
457 MOVD R10, LR
458 RET
459
460 // C->Go callback thunk that allows to call runtime·racesymbolize from C code.
461 // Direct Go->C race call has only switched SP, finish g->g0 switch by setting correct g.
462 // The overall effect of Go->C->Go call chain is similar to that of mcall.
463 // RARG0 contains command code. RARG1 contains command-specific context.
464 // See racecallback for command codes.
465 TEXT runtime·racecallbackthunk(SB), NOSPLIT|NOFRAME, $0
466 // Handle command raceGetProcCmd (0) here.
467 // First, code below assumes that we are on curg, while raceGetProcCmd
468 // can be executed on g0. Second, it is called frequently, so will
469 // benefit from this fast path.
470 MOVD $0, R0 // clear R0 since we came from C code
471 CMP R3, $0
472 BNE rest
473 // Inline raceGetProdCmd without clobbering callee-save registers.
474 MOVD runtime·tls_g(SB), R10
475 MOVD 0(R10), R11
476 MOVD g_m(R11), R3
477 MOVD m_p(R3), R3
478 MOVD p_raceprocctx(R3), R3
479 MOVD R3, (R4)
480 RET
481
482 rest:
483 // Save registers according to the host PPC64 ABI
484 // and reserve 16B for argument storage.
485 STACK_AND_SAVE_HOST_TO_GO_ABI(16)
486
487 // Load g, and switch to g0 if not already on it.
488 MOVD runtime·tls_g(SB), R10
489 MOVD 0(R10), g
490
491 MOVD g_m(g), R7
492 MOVD m_g0(R7), R8
493 CMP g, R8
494 BEQ noswitch
495
496 MOVD R8, g // set g = m->g0
497
498 noswitch:
499 BL runtime·racecallback<ABIInternal>(SB)
500
501 UNSTACK_AND_RESTORE_GO_TO_HOST_ABI(16)
502 RET
503
504 // tls_g, g value for each thread in TLS
505 GLOBL runtime·tls_g+0(SB), TLSBSS+DUPOK, $8
506
View as plain text