2 * tramp-ppc.c: JIT trampoline code for PowerPC
5 * Dietmar Maurer (dietmar@ximian.com)
6 * Paolo Molaro (lupus@ximian.com)
7 * Carlos Valiente <yo@virutass.net>
8 * Andreas Faerber <andreas.faerber@web.de>
10 * (C) 2001 Ximian, Inc.
11 * (C) 2007-2008 Andreas Faerber
17 #include <mono/metadata/abi-details.h>
18 #include <mono/metadata/appdomain.h>
19 #include <mono/metadata/marshal.h>
20 #include <mono/metadata/tabledefs.h>
21 #include <mono/arch/ppc/ppc-codegen.h>
27 /* Same as mono_create_ftnptr, but doesn't require a domain */
29 mono_ppc_create_ftnptr (guint8 *code)
31 #ifdef PPC_USES_FUNCTION_DESCRIPTOR
32 MonoPPCFunctionDescriptor *ftnptr = mono_global_codeman_reserve (sizeof (MonoPPCFunctionDescriptor));
46 * Return the instruction to jump from code to target, 0 if not
47 * reachable with a single instruction
50 branch_for_target_reachable (guint8 *branch, guint8 *target)
52 gint diff = target - branch;
53 g_assert ((diff & 3) == 0);
56 return (18 << 26) | (diff);
58 /* diff between 0 and -33554432 */
59 if (diff >= -33554432)
60 return (18 << 26) | (diff & ~0xfc000000);
66 * get_unbox_trampoline:
68 * @addr: pointer to native code for @m
70 * when value type methods are called through the vtable we need to unbox the
71 * this argument. This method returns a pointer to a trampoline which does
72 * unboxing before calling the method
75 mono_arch_get_unbox_trampoline (MonoMethod *m, gpointer addr)
80 MonoDomain *domain = mono_domain_get ();
81 int size = MONO_PPC_32_64_CASE (20, 32) + PPC_FTNPTR_SIZE;
83 addr = mono_get_addr_from_ftnptr (addr);
85 mono_domain_lock (domain);
86 start = code = mono_domain_code_reserve (domain, size);
87 code = mono_ppc_create_pre_code_ftnptr (code);
88 short_branch = branch_for_target_reachable (code + 4, addr);
90 mono_domain_code_commit (domain, code, size, 8);
91 mono_domain_unlock (domain);
94 ppc_addi (code, this_pos, this_pos, sizeof (MonoObject));
95 ppc_emit32 (code, short_branch);
97 ppc_load_ptr (code, ppc_r0, addr);
98 ppc_mtctr (code, ppc_r0);
99 ppc_addi (code, this_pos, this_pos, sizeof (MonoObject));
100 ppc_bcctr (code, 20, 0);
102 mono_arch_flush_icache (start, code - start);
103 g_assert ((code - start) <= size);
104 /*g_print ("unbox trampoline at %d for %s:%s\n", this_pos, m->klass->name, m->name);
105 g_print ("unbox code is at %p for method at %p\n", start, addr);*/
107 mono_tramp_info_register (mono_tramp_info_create (NULL, start, code - start, NULL, NULL), domain);
113 * mono_arch_get_static_rgctx_trampoline:
115 * Create a trampoline which sets RGCTX_REG to MRGCTX, then jumps to ADDR.
118 mono_arch_get_static_rgctx_trampoline (MonoMethod *m, MonoMethodRuntimeGenericContext *mrgctx, gpointer addr)
120 guint8 *code, *start, *p;
121 guint8 imm_buf [128];
122 guint32 short_branch;
123 MonoDomain *domain = mono_domain_get ();
125 int size = MONO_PPC_32_64_CASE (24, (PPC_LOAD_SEQUENCE_LENGTH * 2) + 8) + PPC_FTNPTR_SIZE;
127 addr = mono_get_addr_from_ftnptr (addr);
129 /* Compute size of code needed to emit mrgctx */
131 ppc_load_ptr (p, MONO_ARCH_RGCTX_REG, mrgctx);
132 imm_size = p - imm_buf;
134 mono_domain_lock (domain);
135 start = code = mono_domain_code_reserve (domain, size);
136 code = mono_ppc_create_pre_code_ftnptr (code);
137 short_branch = branch_for_target_reachable (code + imm_size, addr);
139 mono_domain_code_commit (domain, code, size, imm_size + 4);
140 mono_domain_unlock (domain);
143 ppc_load_ptr (code, MONO_ARCH_RGCTX_REG, mrgctx);
144 ppc_emit32 (code, short_branch);
146 ppc_load_ptr (code, ppc_r0, addr);
147 ppc_mtctr (code, ppc_r0);
148 ppc_load_ptr (code, MONO_ARCH_RGCTX_REG, mrgctx);
149 ppc_bcctr (code, 20, 0);
151 mono_arch_flush_icache (start, code - start);
152 g_assert ((code - start) <= size);
154 mono_tramp_info_register (mono_tramp_info_create (NULL, start, code - start, NULL, NULL), domain);
160 mono_arch_patch_callsite (guint8 *method_start, guint8 *code_ptr, guint8 *addr)
162 guint32 *code = (guint32*)code_ptr;
164 addr = mono_get_addr_from_ftnptr (addr);
166 /* This is the 'blrl' instruction */
170 * Note that methods are called also with the bl opcode.
172 if (((*code) >> 26) == 18) {
173 /*g_print ("direct patching\n");*/
174 ppc_patch ((guint8*)code, addr);
175 mono_arch_flush_icache ((guint8*)code, 4);
180 g_assert (mono_ppc_is_direct_call_sequence (code));
182 ppc_patch ((guint8*)code, addr);
186 mono_arch_patch_plt_entry (guint8 *code, gpointer *got, mgreg_t *regs, guint8 *addr)
188 guint32 ins1, ins2, offset;
190 /* Patch the jump table entry used by the plt entry */
192 /* Should be a lis+ori */
193 ins1 = ((guint32*)code)[0];
194 g_assert (ins1 >> 26 == 15);
195 ins2 = ((guint32*)code)[1];
196 g_assert (ins2 >> 26 == 24);
197 offset = ((ins1 & 0xffff) << 16) | (ins2 & 0xffff);
199 /* Either got or regs is set */
201 got = (gpointer*)(gsize) regs [30];
202 *(guint8**)((guint8*)got + offset) = addr;
205 /* Stack size for trampoline function
206 * PPC_MINIMAL_STACK_SIZE + 16 (args + alignment to ppc_magic_trampoline)
207 * + MonoLMF + 14 fp regs + 13 gregs + alignment
209 #define STACK (((PPC_MINIMAL_STACK_SIZE + 4 * sizeof (mgreg_t) + sizeof (MonoLMF) + 14 * sizeof (double) + 31 * sizeof (mgreg_t)) + (MONO_ARCH_FRAME_ALIGNMENT - 1)) & ~(MONO_ARCH_FRAME_ALIGNMENT - 1))
211 /* Method-specific trampoline code fragment size */
212 #define METHOD_TRAMPOLINE_SIZE 64
214 /* Jump-specific trampoline code fragment size */
215 #define JUMP_TRAMPOLINE_SIZE 64
217 #ifdef PPC_USES_FUNCTION_DESCRIPTOR
218 #define PPC_TOC_REG ppc_r2
220 #define PPC_TOC_REG -1
224 * Stack frame description when the generic trampoline is called.
226 * --------------------
228 * -------------------
229 * Saved FP registers 0-13
230 * -------------------
231 * Saved general registers 0-30
232 * -------------------
233 * param area for 3 args to ppc_magic_trampoline
234 * -------------------
236 * -------------------
239 mono_arch_create_generic_trampoline (MonoTrampolineType tramp_type, MonoTrampInfo **info, gboolean aot)
242 guint8 *buf, *code = NULL;
244 gconstpointer tramp_handler;
245 int size = MONO_PPC_32_64_CASE (600, 800);
246 GSList *unwind_ops = NULL;
247 MonoJumpInfo *ji = NULL;
249 /* Now we'll create in 'buf' the PowerPC trampoline code. This
250 is the trampoline code common to all methods */
252 code = buf = mono_global_codeman_reserve (size);
254 ppc_str_update (code, ppc_r1, -STACK, ppc_r1);
256 /* start building the MonoLMF on the stack */
257 offset = STACK - sizeof (double) * MONO_SAVED_FREGS;
258 for (i = 14; i < 32; i++) {
259 ppc_stfd (code, i, offset, ppc_r1);
260 offset += sizeof (double);
263 * now the integer registers.
265 offset = STACK - sizeof (MonoLMF) + G_STRUCT_OFFSET (MonoLMF, iregs);
266 ppc_str_multiple (code, ppc_r13, offset, ppc_r1);
268 /* Now save the rest of the registers below the MonoLMF struct, first 14
269 * fp regs and then the 31 gregs.
271 offset = STACK - sizeof (MonoLMF) - (14 * sizeof (double));
272 for (i = 0; i < 14; i++) {
273 ppc_stfd (code, i, offset, ppc_r1);
274 offset += sizeof (double);
276 #define GREGS_OFFSET (STACK - sizeof (MonoLMF) - (14 * sizeof (double)) - (31 * sizeof (mgreg_t)))
277 offset = GREGS_OFFSET;
278 for (i = 0; i < 31; i++) {
279 ppc_str (code, i, offset, ppc_r1);
280 offset += sizeof (mgreg_t);
283 /* we got here through a jump to the ctr reg, we must save the lr
284 * in the parent frame (we do it here to reduce the size of the
285 * method-specific trampoline)
287 ppc_mflr (code, ppc_r0);
288 ppc_str (code, ppc_r0, STACK + PPC_RET_ADDR_OFFSET, ppc_r1);
290 /* ok, now we can continue with the MonoLMF setup, mostly untouched
291 * from emit_prolog in mini-ppc.c
294 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "mono_get_lmf_addr");
295 #ifdef PPC_USES_FUNCTION_DESCRIPTOR
296 ppc_ldptr (code, ppc_r2, sizeof (gpointer), ppc_r12);
297 ppc_ldptr (code, ppc_r12, 0, ppc_r12);
299 ppc_mtlr (code, ppc_r12);
302 ppc_load_func (code, PPC_CALL_REG, mono_get_lmf_addr);
303 ppc_mtlr (code, PPC_CALL_REG);
306 /* we build the MonoLMF structure on the stack - see mini-ppc.h
307 * The pointer to the struct is put in ppc_r12.
309 ppc_addi (code, ppc_r12, ppc_sp, STACK - sizeof (MonoLMF));
310 ppc_stptr (code, ppc_r3, G_STRUCT_OFFSET(MonoLMF, lmf_addr), ppc_r12);
311 /* new_lmf->previous_lmf = *lmf_addr */
312 ppc_ldptr (code, ppc_r0, G_STRUCT_OFFSET(MonoLMF, previous_lmf), ppc_r3);
313 ppc_stptr (code, ppc_r0, G_STRUCT_OFFSET(MonoLMF, previous_lmf), ppc_r12);
314 /* *(lmf_addr) = r12 */
315 ppc_stptr (code, ppc_r12, G_STRUCT_OFFSET(MonoLMF, previous_lmf), ppc_r3);
316 /* save method info (it's stored on the stack, so get it first). */
317 if ((tramp_type == MONO_TRAMPOLINE_JIT) || (tramp_type == MONO_TRAMPOLINE_JUMP)) {
318 ppc_ldr (code, ppc_r0, GREGS_OFFSET, ppc_r1);
319 ppc_stptr (code, ppc_r0, G_STRUCT_OFFSET(MonoLMF, method), ppc_r12);
321 ppc_load (code, ppc_r0, 0);
322 ppc_stptr (code, ppc_r0, G_STRUCT_OFFSET(MonoLMF, method), ppc_r12);
324 /* store the frame pointer of the calling method */
325 ppc_addi (code, ppc_r0, ppc_sp, STACK);
326 ppc_stptr (code, ppc_r0, G_STRUCT_OFFSET(MonoLMF, ebp), ppc_r12);
327 /* save the IP (caller ip) */
328 if (tramp_type == MONO_TRAMPOLINE_JUMP) {
329 ppc_li (code, ppc_r0, 0);
331 ppc_ldr (code, ppc_r0, STACK + PPC_RET_ADDR_OFFSET, ppc_r1);
333 ppc_stptr (code, ppc_r0, G_STRUCT_OFFSET(MonoLMF, eip), ppc_r12);
336 * Now we're ready to call trampoline (mgreg_t *regs, guint8 *code, gpointer value, guint8 *tramp)
337 * Note that the last argument is unused.
339 /* Arg 1: a pointer to the registers */
340 ppc_addi (code, ppc_r3, ppc_r1, GREGS_OFFSET);
342 /* Arg 2: code (next address to the instruction that called us) */
343 if (tramp_type == MONO_TRAMPOLINE_JUMP)
344 ppc_li (code, ppc_r4, 0);
346 ppc_ldr (code, ppc_r4, STACK + PPC_RET_ADDR_OFFSET, ppc_r1);
348 /* Arg 3: trampoline argument */
349 ppc_ldr (code, ppc_r5, GREGS_OFFSET, ppc_r1);
352 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, g_strdup_printf ("trampoline_func_%d", tramp_type));
353 #ifdef PPC_USES_FUNCTION_DESCRIPTOR
354 ppc_ldptr (code, ppc_r2, sizeof (gpointer), ppc_r12);
355 ppc_ldptr (code, ppc_r12, 0, ppc_r12);
357 ppc_mtlr (code, ppc_r12);
360 tramp_handler = mono_get_trampoline_func (tramp_type);
361 ppc_load_func (code, PPC_CALL_REG, tramp_handler);
362 ppc_mtlr (code, PPC_CALL_REG);
366 /* OK, code address is now on r3. Move it to the counter reg
367 * so it will be ready for the final jump: this is safe since we
368 * won't do any more calls.
370 if (!MONO_TRAMPOLINE_TYPE_MUST_RETURN (tramp_type)) {
371 #ifdef PPC_USES_FUNCTION_DESCRIPTOR
372 ppc_ldptr (code, ppc_r2, sizeof (gpointer), ppc_r3);
373 ppc_ldptr (code, ppc_r3, 0, ppc_r3);
375 ppc_mtctr (code, ppc_r3);
379 * Now we restore the MonoLMF (see emit_epilogue in mini-ppc.c)
380 * and the rest of the registers, so the method called will see
381 * the same state as before we executed.
382 * The pointer to MonoLMF is in ppc_r12.
384 ppc_addi (code, ppc_r12, ppc_r1, STACK - sizeof (MonoLMF));
385 /* r5 = previous_lmf */
386 ppc_ldptr (code, ppc_r5, G_STRUCT_OFFSET(MonoLMF, previous_lmf), ppc_r12);
388 ppc_ldptr (code, ppc_r6, G_STRUCT_OFFSET(MonoLMF, lmf_addr), ppc_r12);
389 /* *(lmf_addr) = previous_lmf */
390 ppc_stptr (code, ppc_r5, G_STRUCT_OFFSET(MonoLMF, previous_lmf), ppc_r6);
392 ppc_ldr_multiple (code, ppc_r13, G_STRUCT_OFFSET(MonoLMF, iregs), ppc_r12);
394 for (i = 14; i < 32; i++)
395 ppc_lfd (code, i, G_STRUCT_OFFSET(MonoLMF, fregs) + ((i-14) * sizeof (gdouble)), ppc_r12);
397 /* restore the volatile registers, we skip r1, of course */
398 offset = STACK - sizeof (MonoLMF) - (14 * sizeof (double));
399 for (i = 0; i < 14; i++) {
400 ppc_lfd (code, i, offset, ppc_r1);
401 offset += sizeof (double);
403 offset = STACK - sizeof (MonoLMF) - (14 * sizeof (double)) - (31 * sizeof (mgreg_t));
404 ppc_ldr (code, ppc_r0, offset, ppc_r1);
405 offset += 2 * sizeof (mgreg_t);
406 for (i = 2; i < 13; i++) {
407 if (i != PPC_TOC_REG && (i != 3 || tramp_type != MONO_TRAMPOLINE_RGCTX_LAZY_FETCH))
408 ppc_ldr (code, i, offset, ppc_r1);
409 offset += sizeof (mgreg_t);
412 /* Non-standard function epilogue. Instead of doing a proper
413 * return, we just jump to the compiled code.
415 /* Restore stack pointer and LR and jump to the code */
416 ppc_ldr (code, ppc_r1, 0, ppc_r1);
417 ppc_ldr (code, ppc_r12, PPC_RET_ADDR_OFFSET, ppc_r1);
418 ppc_mtlr (code, ppc_r12);
419 if (MONO_TRAMPOLINE_TYPE_MUST_RETURN (tramp_type))
422 ppc_bcctr (code, 20, 0);
424 /* Flush instruction cache, since we've generated code */
425 mono_arch_flush_icache (buf, code - buf);
428 g_assert ((code - buf) <= size);
431 tramp_name = mono_get_generic_trampoline_name (tramp_type);
432 *info = mono_tramp_info_create (tramp_name, buf, code - buf, ji, unwind_ops);
438 #define TRAMPOLINE_SIZE (MONO_PPC_32_64_CASE (24, (5+5+1+1)*4))
440 mono_arch_create_specific_trampoline (gpointer arg1, MonoTrampolineType tramp_type, MonoDomain *domain, guint32 *code_len)
442 guint8 *code, *buf, *tramp;
443 guint32 short_branch;
445 tramp = mono_get_trampoline_code (tramp_type);
447 mono_domain_lock (domain);
448 code = buf = mono_domain_code_reserve_align (domain, TRAMPOLINE_SIZE, 4);
449 short_branch = branch_for_target_reachable (code + MONO_PPC_32_64_CASE (8, 5*4), tramp);
450 #ifdef __mono_ppc64__
451 /* FIXME: make shorter if possible */
454 mono_domain_code_commit (domain, code, TRAMPOLINE_SIZE, 12);
456 mono_domain_unlock (domain);
459 ppc_load_sequence (code, ppc_r0, (mgreg_t)(gsize) arg1);
460 ppc_emit32 (code, short_branch);
462 /* Prepare the jump to the generic trampoline code.*/
463 ppc_load_ptr (code, ppc_r0, tramp);
464 ppc_mtctr (code, ppc_r0);
466 /* And finally put 'arg1' in r0 and fly! */
467 ppc_load_ptr (code, ppc_r0, arg1);
468 ppc_bcctr (code, 20, 0);
471 /* Flush instruction cache, since we've generated code */
472 mono_arch_flush_icache (buf, code - buf);
474 g_assert ((code - buf) <= TRAMPOLINE_SIZE);
477 *code_len = code - buf;
483 emit_trampoline_jump (guint8 *code, guint8 *tramp)
485 guint32 short_branch = branch_for_target_reachable (code, tramp);
487 /* FIXME: we can save a few bytes here by committing if the
488 short branch is possible */
490 ppc_emit32 (code, short_branch);
492 ppc_load_ptr (code, ppc_r0, tramp);
493 ppc_mtctr (code, ppc_r0);
494 ppc_bcctr (code, 20, 0);
501 mono_arch_create_rgctx_lazy_fetch_trampoline (guint32 slot, MonoTrampInfo **info, gboolean aot)
505 guint8 **rgctx_null_jumps;
510 MonoJumpInfo *ji = NULL;
511 GSList *unwind_ops = NULL;
513 mrgctx = MONO_RGCTX_SLOT_IS_MRGCTX (slot);
514 index = MONO_RGCTX_SLOT_INDEX (slot);
516 index += MONO_SIZEOF_METHOD_RUNTIME_GENERIC_CONTEXT / sizeof (gpointer);
517 for (depth = 0; ; ++depth) {
518 int size = mono_class_rgctx_get_array_size (depth, mrgctx);
520 if (index < size - 1)
525 tramp_size = MONO_PPC_32_64_CASE (40, 52) + 12 * depth;
533 code = buf = mono_global_codeman_reserve (tramp_size);
535 rgctx_null_jumps = g_malloc (sizeof (guint8*) * (depth + 2));
539 ppc_mr (code, ppc_r4, PPC_FIRST_ARG_REG);
541 /* load rgctx ptr from vtable */
542 ppc_ldptr (code, ppc_r4, MONO_STRUCT_OFFSET (MonoVTable, runtime_generic_context), PPC_FIRST_ARG_REG);
543 /* is the rgctx ptr null? */
544 ppc_compare_reg_imm (code, 0, ppc_r4, 0);
545 /* if yes, jump to actual trampoline */
546 rgctx_null_jumps [0] = code;
547 ppc_bc (code, PPC_BR_TRUE, PPC_BR_EQ, 0);
550 for (i = 0; i < depth; ++i) {
551 /* load ptr to next array */
552 if (mrgctx && i == 0)
553 ppc_ldptr (code, ppc_r4, MONO_SIZEOF_METHOD_RUNTIME_GENERIC_CONTEXT, ppc_r4);
555 ppc_ldptr (code, ppc_r4, 0, ppc_r4);
556 /* is the ptr null? */
557 ppc_compare_reg_imm (code, 0, ppc_r4, 0);
558 /* if yes, jump to actual trampoline */
559 rgctx_null_jumps [i + 1] = code;
560 ppc_bc (code, PPC_BR_TRUE, PPC_BR_EQ, 0);
564 ppc_ldptr (code, ppc_r4, sizeof (gpointer) * (index + 1), ppc_r4);
565 /* is the slot null? */
566 ppc_compare_reg_imm (code, 0, ppc_r4, 0);
567 /* if yes, jump to actual trampoline */
568 rgctx_null_jumps [depth + 1] = code;
569 ppc_bc (code, PPC_BR_TRUE, PPC_BR_EQ, 0);
570 /* otherwise return r4 */
571 /* FIXME: if we use r3 as the work register we can avoid this copy */
572 ppc_mr (code, ppc_r3, ppc_r4);
575 for (i = mrgctx ? 1 : 0; i <= depth + 1; ++i)
576 ppc_patch (rgctx_null_jumps [i], code);
578 g_free (rgctx_null_jumps);
580 /* move the rgctx pointer to the VTABLE register */
581 ppc_mr (code, MONO_ARCH_VTABLE_REG, ppc_r3);
584 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, g_strdup_printf ("specific_trampoline_lazy_fetch_%u", slot));
585 /* Branch to the trampoline */
586 #ifdef PPC_USES_FUNCTION_DESCRIPTOR
587 ppc_ldptr (code, ppc_r12, 0, ppc_r12);
589 ppc_mtctr (code, ppc_r12);
590 ppc_bcctr (code, PPC_BR_ALWAYS, 0);
592 tramp = mono_arch_create_specific_trampoline (GUINT_TO_POINTER (slot),
593 MONO_TRAMPOLINE_RGCTX_LAZY_FETCH, mono_get_root_domain (), NULL);
595 /* jump to the actual trampoline */
596 code = emit_trampoline_jump (code, tramp);
599 mono_arch_flush_icache (buf, code - buf);
601 g_assert (code - buf <= tramp_size);
603 char *name = mono_get_rgctx_fetch_trampoline_name (slot);
604 *info = mono_tramp_info_create (name, buf, code - buf, ji, unwind_ops);
611 mono_arch_get_call_target (guint8 *code)
614 guint32 ins = ((guint32*)(gpointer)code) [-1];
616 if ((ins >> 26 == 18) && ((ins & 1) == 1) && ((ins & 2) == 0)) {
617 gint32 disp = (((gint32)ins) >> 2) & 0xffffff;
618 guint8 *target = code - 4 + (disp * 4);
627 mono_arch_get_plt_info_offset (guint8 *plt_entry, mgreg_t *regs, guint8 *code)
629 #ifdef PPC_USES_FUNCTION_DESCRIPTOR
630 return ((guint32*)plt_entry) [8];
632 return ((guint32*)plt_entry) [6];