mono/mini/tramp-ia64.c

   1 /*
   2  * tramp-ia64.c: JIT trampoline code for ia64
   3  *
   4  * Authors:
   5  *   Zoltan Varga (vargaz@gmail.com)
   6  *
   7  * (C) 2001 Ximian, Inc.
   8  */
   9
  10 #include <config.h>
  11 #include <glib.h>
  12
  13 #include <mono/metadata/appdomain.h>
  14 #include <mono/metadata/marshal.h>
  15 #include <mono/metadata/tabledefs.h>
  16 #include <mono/metadata/mono-debug-debugger.h>
  17 #include <mono/arch/ia64/ia64-codegen.h>
  18
  19 #include "mini.h"
  20 #include "mini-ia64.h"
  21
  22 #define GP_SCRATCH_REG 31
  23 #define GP_SCRATCH_REG2 30
  24
  25 /*
  26  * mono_arch_get_unbox_trampoline:
  27  * @gsctx: the generic sharing context
  28  * @m: method pointer
  29  * @addr: pointer to native code for @m
  30  *
  31  * when value type methods are called through the vtable we need to unbox the
  32  * this argument. This method returns a pointer to a trampoline which does
  33  * unboxing before calling the method
  34  */
  35 gpointer
  36 mono_arch_get_unbox_trampoline (MonoGenericSharingContext *gsctx, MonoMethod *m, gpointer addr)
  37 {
  38         guint8 *buf;
  39         gpointer func_addr, func_gp;
  40         Ia64CodegenState code;
  41         int this_reg = 0;
  42         gpointer *desc;
  43         MonoDomain *domain = mono_domain_get ();
  44
  45         /* FIXME: Optimize this */
  46
  47         if (MONO_TYPE_ISSTRUCT (mono_method_signature (m)->ret))
  48                 this_reg = 1;
  49
  50         func_addr = ((gpointer*)addr) [0];
  51         func_gp = ((gpointer*)addr) [1];
  52
  53         buf = mono_domain_code_reserve (domain, 256);
  54
  55         /* Since the this reg is a stacked register, its a bit hard to access it */
  56         ia64_codegen_init (code, buf);
  57         ia64_alloc (code, 40, 8, 1, 0, 0);
  58         ia64_adds_imm (code, 32 + this_reg, sizeof (MonoObject), 32 + this_reg);
  59         ia64_mov_to_ar_i (code, IA64_PFS, 40);
  60         ia64_movl (code, GP_SCRATCH_REG, func_addr);
  61         ia64_mov_to_br (code, IA64_B6, GP_SCRATCH_REG);
  62         ia64_br_cond_reg (code, IA64_B6);
  63         ia64_codegen_close (code);
  64
  65         g_assert (code.buf - buf < 256);
  66
  67         mono_arch_flush_icache (buf, code.buf - buf);
  68
  69         /* FIXME: */
  70         desc = g_malloc0 (sizeof (gpointer) * 2);
  71         desc [0] = buf;
  72         desc [1] = func_gp;
  73
  74         return desc;
  75 }
  76
  77 void
  78 mono_arch_patch_callsite (guint8 *method_start, guint8 *code, guint8 *addr)
  79 {
  80         guint8 *callsite_begin;
  81         guint64 *callsite = (guint64*)(gpointer)(code - 16);
  82         guint64 *next_bundle;
  83         guint64 ins, instructions [3];
  84         guint64 buf [16];
  85         Ia64CodegenState gen;
  86         gpointer func = ((gpointer*)(gpointer)addr)[0];
  87
  88         while ((ia64_bundle_template (callsite) != IA64_TEMPLATE_MLX) &&
  89                    (ia64_bundle_template (callsite) != IA64_TEMPLATE_MLXS))
  90                 callsite -= 2;
  91         callsite_begin = (guint8*)callsite;
  92
  93         next_bundle = callsite + 2;
  94         ins = ia64_bundle_ins1 (next_bundle);
  95         if (ia64_ins_opcode (ins) == 5) {
  96                 /* ld8_inc_imm -> indirect call through a function pointer */
  97                 g_assert (ia64_ins_r1 (ins) == GP_SCRATCH_REG2);
  98                 g_assert (ia64_ins_r3 (ins) == GP_SCRATCH_REG);
  99                 return;
 100         }
 101
 102         /* Patch the code generated by emit_call */
 103
 104         instructions [0] = ia64_bundle_ins1 (callsite);
 105         instructions [1] = ia64_bundle_ins2 (callsite);
 106         instructions [2] = ia64_bundle_ins3 (callsite);
 107
 108         ia64_codegen_init (gen, (guint8*)buf);
 109         ia64_movl (gen, GP_SCRATCH_REG, func);
 110         instructions [1] = gen.instructions [0];
 111         instructions [2] = gen.instructions [1];
 112
 113         ia64_codegen_init (gen, (guint8*)buf);
 114         ia64_emit_bundle_template (&gen, ia64_bundle_template (callsite), instructions [0], instructions [1], instructions [2]);
 115         ia64_codegen_close (gen);
 116
 117         /* This might not be safe, but not all itanium processors support st16 */
 118         callsite [0] = buf [0];
 119         callsite [1] = buf [1];
 120
 121         mono_arch_flush_icache (callsite_begin, code - callsite_begin);
 122 }
 123
 124 void
 125 mono_arch_patch_plt_entry (guint8 *code, gpointer *got, mgreg_t *regs, guint8 *addr)
 126 {
 127         g_assert_not_reached ();
 128 }
 129
 130 void
 131 mono_arch_nullify_class_init_trampoline (guint8 *code, mgreg_t *regs)
 132 {
 133         guint8 *callsite_begin;
 134         guint64 *callsite = (guint64*)(gpointer)(code - 16);
 135         guint64 instructions [3];
 136         guint64 buf [16];
 137         Ia64CodegenState gen;
 138
 139         while ((ia64_bundle_template (callsite) != IA64_TEMPLATE_MLX) &&
 140                    (ia64_bundle_template (callsite) != IA64_TEMPLATE_MLXS))
 141                 callsite -= 2;
 142         callsite_begin = (guint8*)callsite;
 143
 144         /* Replace the code generated by emit_call with a sets of nops */
 145
 146         /* The first bundle might have other instructions in it */
 147         instructions [0] = ia64_bundle_ins1 (callsite);
 148         instructions [1] = IA64_NOP_X;
 149         instructions [2] = IA64_NOP_X;
 150
 151         ia64_codegen_init (gen, (guint8*)buf);
 152         ia64_emit_bundle_template (&gen, ia64_bundle_template (callsite), instructions [0], instructions [1], instructions [2]);
 153         ia64_codegen_close (gen);
 154
 155         /* This might not be safe, but not all itanium processors support st16 */
 156         callsite [0] = buf [0];
 157         callsite [1] = buf [1];
 158
 159         callsite += 2;
 160
 161         /* The other bundles can be full replaced with nops */
 162
 163         ia64_codegen_init (gen, (guint8*)buf);
 164         ia64_emit_bundle_template (&gen, IA64_TEMPLATE_MII, IA64_NOP_M, IA64_NOP_I, IA64_NOP_I);
 165         ia64_codegen_close (gen);
 166
 167         while ((guint8*)callsite < code) {
 168                 callsite [0] = buf [0];
 169                 callsite [1] = buf [1];
 170                 callsite += 2;
 171         }
 172
 173         mono_arch_flush_icache (callsite_begin, code - callsite_begin);
 174 }
 175
 176 void
 177 mono_arch_nullify_plt_entry (guint8 *code, mgreg_t *regs)
 178 {
 179         g_assert_not_reached ();
 180 }
 181
 182 guchar*
 183 mono_arch_create_generic_trampoline (MonoTrampolineType tramp_type, MonoTrampInfo **info, gboolean aot)
 184 {
 185         guint8 *buf, *tramp;
 186         int i, offset, saved_regs_offset, saved_fpregs_offset, last_offset, framesize;
 187         int in0, local0, out0, l0, l1, l2, l3, l4, l5, l6, l7, l8, o0, o1, o2, o3;
 188         gboolean has_caller;
 189         Ia64CodegenState code;
 190         unw_dyn_info_t *di;
 191         unw_dyn_region_info_t *r_pro;
 192
 193         g_assert (!aot);
 194         if (info)
 195                 *info = NULL;
 196
 197         /*
 198          * Since jump trampolines are not patched, this trampoline is executed every
 199          * time a call is made to a jump trampoline. So we try to keep things faster
 200          * in that case.
 201          */
 202         if (tramp_type == MONO_TRAMPOLINE_JUMP)
 203                 has_caller = FALSE;
 204         else
 205                 has_caller = TRUE;
 206
 207         buf = mono_global_codeman_reserve (2048);
 208
 209         ia64_codegen_init (code, buf);
 210
 211         /* Stacked Registers */
 212         in0 = 32;
 213         local0 = in0 + 8;
 214         out0 = local0 + 16;
 215         l0 = 40;
 216         l1 = 41;
 217         l2 = 42;
 218         l3 = 43;
 219         l4 = 44;
 220         l5 = 45; /* saved ar.pfs */
 221         l6 = 46; /* arg */
 222         l7 = 47; /* code */
 223         l8 = 48; /* saved sp */
 224         o0 = out0 + 0; /* regs */
 225         o1 = out0 + 1; /* code */
 226         o2 = out0 + 2; /* arg */
 227         o3 = out0 + 3; /* tramp */
 228
 229         framesize = (128 * 8) + 1024;
 230         framesize = (framesize + (MONO_ARCH_FRAME_ALIGNMENT - 1)) & ~ (MONO_ARCH_FRAME_ALIGNMENT - 1);
 231
 232         /*
 233          * Allocate a new register+memory stack frame.
 234          * 8 input registers (the max used by the ABI)
 235          * 16 locals
 236          * 4 output (number of parameters passed to trampoline)
 237          */
 238         ia64_unw_save_reg (code, UNW_IA64_AR_PFS, UNW_IA64_GR + l5);
 239         ia64_alloc (code, l5, local0 - in0, out0 - local0, 4, 0);
 240         ia64_unw_save_reg (code, UNW_IA64_SP, UNW_IA64_GR + l8);
 241         ia64_mov (code, l8, IA64_SP);
 242         ia64_adds_imm (code, IA64_SP, (-framesize), IA64_SP);
 243
 244         offset = 16; /* scratch area */
 245
 246         /* Save the argument received from the specific trampoline */
 247         ia64_mov (code, l6, GP_SCRATCH_REG);
 248
 249         /* Save the calling address */
 250         ia64_unw_save_reg (code, UNW_IA64_RP, UNW_IA64_GR + local0 + 7);
 251         ia64_mov_from_br (code, l7, IA64_B0);
 252
 253         /* Create unwind info for the prolog */
 254         ia64_begin_bundle (code);
 255         r_pro = mono_ia64_create_unwind_region (&code);
 256
 257         /* Save registers */
 258         /* Not needed for jump trampolines */
 259         if (tramp_type != MONO_TRAMPOLINE_JUMP) {
 260                 saved_regs_offset = offset;
 261                 offset += 128 * 8;
 262                 /*
 263                  * Only the registers which are needed for computing vtable slots need
 264                  * to be saved.
 265                  */
 266                 last_offset = -1;
 267                 for (i = 0; i < 64; ++i)
 268                         if ((1 << i) & MONO_ARCH_CALLEE_REGS) {
 269                                 if (last_offset != i * 8)
 270                                         ia64_adds_imm (code, l1, saved_regs_offset + (i * 8), IA64_SP);
 271                                 ia64_st8_spill_inc_imm_hint (code, l1, i, 8, 0);
 272                                 last_offset = (i + 1) * 8;
 273                         }
 274         }
 275
 276         /* Save fp registers */
 277         saved_fpregs_offset = offset;
 278         offset += 8 * 8;
 279         ia64_adds_imm (code, l1, saved_fpregs_offset, IA64_SP);
 280         for (i = 0; i < 8; ++i)
 281                 ia64_stfd_inc_imm_hint (code, l1, i + 8, 8, 0);
 282
 283         g_assert (offset < framesize);
 284
 285         /* Arg1 is the pointer to the saved registers */
 286         ia64_adds_imm (code, o0, saved_regs_offset, IA64_SP);
 287
 288         /* Arg2 is the address of the calling code */
 289         if (has_caller)
 290                 ia64_mov (code, o1, l7);
 291         else
 292                 ia64_mov (code, o1, 0);
 293
 294         /* Arg3 is the method/vtable ptr */
 295         ia64_mov (code, o2, l6);
 296
 297         /* Arg4 is the trampoline address */
 298         /* FIXME: */
 299         ia64_mov (code, o3, 0);
 300
 301         tramp = (guint8*)mono_get_trampoline_func (tramp_type);
 302
 303         /* Call the trampoline using an indirect call */
 304         ia64_movl (code, l0, tramp);
 305         ia64_ld8_inc_imm (code, l1, l0, 8);
 306         ia64_mov_to_br (code, IA64_B6, l1);
 307         ia64_ld8 (code, IA64_GP, l0);
 308         ia64_br_call_reg (code, 0, IA64_B6);
 309
 310         /* Check for thread interruption */
 311         /* This is not perf critical code so no need to check the interrupt flag */
 312         ia64_mov (code, l2, IA64_R8);
 313
 314         tramp = (guint8*)mono_thread_force_interruption_checkpoint;
 315         ia64_movl (code, l0, tramp);
 316         ia64_ld8_inc_imm (code, l1, l0, 8);
 317         ia64_mov_to_br (code, IA64_B6, l1);
 318         ia64_ld8 (code, IA64_GP, l0);
 319         ia64_br_call_reg (code, 0, IA64_B6);
 320
 321         ia64_mov (code, IA64_R8, l2);
 322
 323         /* Restore fp regs */
 324         ia64_adds_imm (code, l1, saved_fpregs_offset, IA64_SP);
 325         for (i = 0; i < 8; ++i)
 326                 ia64_ldfd_inc_imm (code, i + 8, l1, 8);
 327
 328         /* FIXME: Handle NATs in fp regs / scratch regs */
 329
 330         if (tramp_type != MONO_TRAMPOLINE_CLASS_INIT) {
 331                 /* Load method address from function descriptor */
 332                 ia64_ld8 (code, l0, IA64_R8);
 333                 ia64_mov_to_br (code, IA64_B6, l0);
 334         }
 335
 336         /* Clean up register/memory stack frame */
 337         ia64_adds_imm (code, IA64_SP, framesize, IA64_SP);
 338         ia64_mov_to_ar_i (code, IA64_PFS, l5);
 339
 340         if (tramp_type == MONO_TRAMPOLINE_CLASS_INIT) {
 341                 ia64_mov_ret_to_br (code, IA64_B0, l7);
 342                 ia64_br_ret_reg (code, IA64_B0);
 343         }
 344         else {
 345                 /* Call the compiled method */
 346                 ia64_mov_to_br (code, IA64_B0, l7);
 347                 ia64_br_cond_reg (code, IA64_B6);
 348         }
 349
 350         ia64_codegen_close (code);
 351
 352         g_assert ((code.buf - buf) <= 2048);
 353
 354         /* FIXME: emit unwind info for epilog */
 355         di = g_malloc0 (sizeof (unw_dyn_info_t));
 356         di->start_ip = (unw_word_t) buf;
 357         di->end_ip = (unw_word_t) code.buf;
 358         di->gp = 0;
 359         di->format = UNW_INFO_FORMAT_DYNAMIC;
 360         di->u.pi.name_ptr = (unw_word_t)"ia64_generic_trampoline";
 361         di->u.pi.regions = r_pro;
 362
 363         _U_dyn_register (di);
 364
 365         mono_arch_flush_icache (buf, code.buf - buf);
 366
 367         return buf;
 368 }
 369
 370 #define TRAMPOLINE_SIZE 128
 371
 372 gpointer
 373 mono_arch_create_specific_trampoline (gpointer arg1, MonoTrampolineType tramp_type, MonoDomain *domain, guint32 *code_len)
 374 {
 375         guint8 *buf, *tramp;
 376         gint64 disp;
 377         Ia64CodegenState code;
 378
 379         tramp = mono_get_trampoline_code (tramp_type);
 380
 381         buf = mono_domain_code_reserve (domain, TRAMPOLINE_SIZE);
 382
 383         /* FIXME: Optimize this */
 384
 385         ia64_codegen_init (code, buf);
 386
 387         ia64_movl (code, GP_SCRATCH_REG, arg1);
 388
 389         ia64_begin_bundle (code);
 390         disp = (tramp - code.buf) >> 4;
 391         if (ia64_is_imm21 (disp)) {
 392                 ia64_br_cond (code, disp);
 393         }
 394         else {
 395                 ia64_movl (code, GP_SCRATCH_REG2, tramp);
 396                 ia64_mov_to_br (code, IA64_B6, GP_SCRATCH_REG2);
 397                 ia64_br_cond_reg (code, IA64_B6);
 398         }
 399
 400         ia64_codegen_close (code);
 401
 402         g_assert (code.buf - buf <= TRAMPOLINE_SIZE);
 403
 404         mono_arch_flush_icache (buf, code.buf - buf);
 405
 406         if (code_len)
 407                 *code_len = code.buf - buf;
 408
 409         return buf;
 410 }
 411
 412 void
 413 mono_arch_invalidate_method (MonoJitInfo *ji, void *func, gpointer func_arg)
 414 {
 415         NOT_IMPLEMENTED;
 416 }
 417
 418 gpointer
 419 mono_arch_create_rgctx_lazy_fetch_trampoline (guint32 slot, MonoTrampInfo **info, gboolean aot)
 420 {
 421         /* FIXME: implement! */
 422         g_assert_not_reached ();
 423         return NULL;
 424 }