2 * tramp-ia64.c: JIT trampoline code for ia64
5 * Zoltan Varga (vargaz@gmail.com)
7 * (C) 2001 Ximian, Inc.
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>
20 #include "mini-ia64.h"
22 #define GP_SCRATCH_REG 31
23 #define GP_SCRATCH_REG2 30
26 * mono_arch_get_unbox_trampoline:
27 * @gsctx: the generic sharing context
29 * @addr: pointer to native code for @m
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
36 mono_arch_get_unbox_trampoline (MonoGenericSharingContext *gsctx, MonoMethod *m, gpointer addr)
39 gpointer func_addr, func_gp;
40 Ia64CodegenState code;
43 MonoDomain *domain = mono_domain_get ();
45 /* FIXME: Optimize this */
47 if (MONO_TYPE_ISSTRUCT (mono_method_signature (m)->ret))
50 func_addr = ((gpointer*)addr) [0];
51 func_gp = ((gpointer*)addr) [1];
53 mono_domain_lock (domain);
54 buf = mono_code_manager_reserve (domain->code_mp, 256);
55 mono_domain_unlock (domain);
57 /* Since the this reg is a stacked register, its a bit hard to access it */
58 ia64_codegen_init (code, buf);
59 ia64_alloc (code, 40, 8, 1, 0, 0);
60 ia64_adds_imm (code, 32 + this_reg, sizeof (MonoObject), 32 + this_reg);
61 ia64_mov_to_ar_i (code, IA64_PFS, 40);
62 ia64_movl (code, GP_SCRATCH_REG, func_addr);
63 ia64_mov_to_br (code, IA64_B6, GP_SCRATCH_REG);
64 ia64_br_cond_reg (code, IA64_B6);
65 ia64_codegen_close (code);
67 g_assert (code.buf - buf < 256);
69 mono_arch_flush_icache (buf, code.buf - buf);
72 desc = g_malloc0 (sizeof (gpointer) * 2);
80 mono_arch_patch_callsite (guint8 *method_start, guint8 *code, guint8 *addr)
82 guint8 *callsite_begin;
83 guint64 *callsite = (guint64*)(gpointer)(code - 16);
85 guint64 ins, instructions [3];
88 gpointer func = ((gpointer*)(gpointer)addr)[0];
90 while ((ia64_bundle_template (callsite) != IA64_TEMPLATE_MLX) &&
91 (ia64_bundle_template (callsite) != IA64_TEMPLATE_MLXS))
93 callsite_begin = (guint8*)callsite;
95 next_bundle = callsite + 2;
96 ins = ia64_bundle_ins1 (next_bundle);
97 if (ia64_ins_opcode (ins) == 5) {
98 /* ld8_inc_imm -> indirect call through a function pointer */
99 g_assert (ia64_ins_r1 (ins) == GP_SCRATCH_REG2);
100 g_assert (ia64_ins_r3 (ins) == GP_SCRATCH_REG);
104 /* Patch the code generated by emit_call */
106 instructions [0] = ia64_bundle_ins1 (callsite);
107 instructions [1] = ia64_bundle_ins2 (callsite);
108 instructions [2] = ia64_bundle_ins3 (callsite);
110 ia64_codegen_init (gen, (guint8*)buf);
111 ia64_movl (gen, GP_SCRATCH_REG, func);
112 instructions [1] = gen.instructions [0];
113 instructions [2] = gen.instructions [1];
115 ia64_codegen_init (gen, (guint8*)buf);
116 ia64_emit_bundle_template (&gen, ia64_bundle_template (callsite), instructions [0], instructions [1], instructions [2]);
117 ia64_codegen_close (gen);
119 /* This might not be safe, but not all itanium processors support st16 */
120 callsite [0] = buf [0];
121 callsite [1] = buf [1];
123 mono_arch_flush_icache (callsite_begin, code - callsite_begin);
127 mono_arch_patch_plt_entry (guint8 *code, guint8 *addr)
129 g_assert_not_reached ();
133 mono_arch_nullify_class_init_trampoline (guint8 *code, gssize *regs)
135 guint8 *callsite_begin;
136 guint64 *callsite = (guint64*)(gpointer)(code - 16);
137 guint64 instructions [3];
139 Ia64CodegenState gen;
141 while ((ia64_bundle_template (callsite) != IA64_TEMPLATE_MLX) &&
142 (ia64_bundle_template (callsite) != IA64_TEMPLATE_MLXS))
144 callsite_begin = (guint8*)callsite;
146 /* Replace the code generated by emit_call with a sets of nops */
148 /* The first bundle might have other instructions in it */
149 instructions [0] = ia64_bundle_ins1 (callsite);
150 instructions [1] = IA64_NOP_X;
151 instructions [2] = IA64_NOP_X;
153 ia64_codegen_init (gen, (guint8*)buf);
154 ia64_emit_bundle_template (&gen, ia64_bundle_template (callsite), instructions [0], instructions [1], instructions [2]);
155 ia64_codegen_close (gen);
157 /* This might not be safe, but not all itanium processors support st16 */
158 callsite [0] = buf [0];
159 callsite [1] = buf [1];
163 /* The other bundles can be full replaced with nops */
165 ia64_codegen_init (gen, (guint8*)buf);
166 ia64_emit_bundle_template (&gen, IA64_TEMPLATE_MII, IA64_NOP_M, IA64_NOP_I, IA64_NOP_I);
167 ia64_codegen_close (gen);
169 while ((guint8*)callsite < code) {
170 callsite [0] = buf [0];
171 callsite [1] = buf [1];
175 mono_arch_flush_icache (callsite_begin, code - callsite_begin);
179 mono_arch_nullify_plt_entry (guint8 *code)
181 g_assert_not_reached ();
185 mono_arch_create_trampoline_code (MonoTrampolineType tramp_type)
188 int i, offset, saved_regs_offset, saved_fpregs_offset, last_offset, framesize;
189 int in0, local0, out0, l0, l1, l2, l3, l4, l5, l6, l7, l8, o0, o1, o2, o3;
191 Ia64CodegenState code;
193 unw_dyn_region_info_t *r_pro;
196 * Since jump trampolines are not patched, this trampoline is executed every
197 * time a call is made to a jump trampoline. So we try to keep things faster
200 if (tramp_type == MONO_TRAMPOLINE_JUMP)
205 buf = mono_global_codeman_reserve (2048);
207 ia64_codegen_init (code, buf);
209 /* Stacked Registers */
218 l5 = 45; /* saved ar.pfs */
221 l8 = 48; /* saved sp */
222 o0 = out0 + 0; /* regs */
223 o1 = out0 + 1; /* code */
224 o2 = out0 + 2; /* arg */
225 o3 = out0 + 3; /* tramp */
227 framesize = (128 * 8) + 1024;
228 framesize = (framesize + (MONO_ARCH_FRAME_ALIGNMENT - 1)) & ~ (MONO_ARCH_FRAME_ALIGNMENT - 1);
231 * Allocate a new register+memory stack frame.
232 * 8 input registers (the max used by the ABI)
234 * 4 output (number of parameters passed to trampoline)
236 ia64_unw_save_reg (code, UNW_IA64_AR_PFS, UNW_IA64_GR + l5);
237 ia64_alloc (code, l5, local0 - in0, out0 - local0, 4, 0);
238 ia64_unw_save_reg (code, UNW_IA64_SP, UNW_IA64_GR + l8);
239 ia64_mov (code, l8, IA64_SP);
240 ia64_adds_imm (code, IA64_SP, (-framesize), IA64_SP);
242 offset = 16; /* scratch area */
244 /* Save the argument received from the specific trampoline */
245 ia64_mov (code, l6, GP_SCRATCH_REG);
247 /* Save the calling address */
248 ia64_unw_save_reg (code, UNW_IA64_RP, UNW_IA64_GR + local0 + 7);
249 ia64_mov_from_br (code, l7, IA64_B0);
251 /* Create unwind info for the prolog */
252 ia64_begin_bundle (code);
253 r_pro = mono_ia64_create_unwind_region (&code);
256 /* Not needed for jump trampolines */
257 if (tramp_type != MONO_TRAMPOLINE_JUMP) {
258 saved_regs_offset = offset;
261 * Only the registers which are needed for computing vtable slots need
265 for (i = 0; i < 64; ++i)
266 if ((1 << i) & MONO_ARCH_CALLEE_REGS) {
267 if (last_offset != i * 8)
268 ia64_adds_imm (code, l1, saved_regs_offset + (i * 8), IA64_SP);
269 ia64_st8_spill_inc_imm_hint (code, l1, i, 8, 0);
270 last_offset = (i + 1) * 8;
274 /* Save fp registers */
275 saved_fpregs_offset = offset;
277 ia64_adds_imm (code, l1, saved_fpregs_offset, IA64_SP);
278 for (i = 0; i < 8; ++i)
279 ia64_stfd_inc_imm_hint (code, l1, i + 8, 8, 0);
281 g_assert (offset < framesize);
283 /* Arg1 is the pointer to the saved registers */
284 ia64_adds_imm (code, o0, saved_regs_offset, IA64_SP);
286 /* Arg2 is the address of the calling code */
288 ia64_mov (code, o1, l7);
290 ia64_mov (code, o1, 0);
292 /* Arg3 is the method/vtable ptr */
293 ia64_mov (code, o2, l6);
295 /* Arg4 is the trampoline address */
297 ia64_mov (code, o3, 0);
299 tramp = (guint8*)mono_get_trampoline_func (tramp_type);
301 /* Call the trampoline using an indirect call */
302 ia64_movl (code, l0, tramp);
303 ia64_ld8_inc_imm (code, l1, l0, 8);
304 ia64_mov_to_br (code, IA64_B6, l1);
305 ia64_ld8 (code, IA64_GP, l0);
306 ia64_br_call_reg (code, 0, IA64_B6);
308 /* Check for thread interruption */
309 /* This is not perf critical code so no need to check the interrupt flag */
310 ia64_mov (code, l2, IA64_R8);
312 tramp = (guint8*)mono_thread_force_interruption_checkpoint;
313 ia64_movl (code, l0, tramp);
314 ia64_ld8_inc_imm (code, l1, l0, 8);
315 ia64_mov_to_br (code, IA64_B6, l1);
316 ia64_ld8 (code, IA64_GP, l0);
317 ia64_br_call_reg (code, 0, IA64_B6);
319 ia64_mov (code, IA64_R8, l2);
321 /* Restore fp regs */
322 ia64_adds_imm (code, l1, saved_fpregs_offset, IA64_SP);
323 for (i = 0; i < 8; ++i)
324 ia64_ldfd_inc_imm (code, i + 8, l1, 8);
326 /* FIXME: Handle NATs in fp regs / scratch regs */
328 if (tramp_type != MONO_TRAMPOLINE_CLASS_INIT) {
329 /* Load method address from function descriptor */
330 ia64_ld8 (code, l0, IA64_R8);
331 ia64_mov_to_br (code, IA64_B6, l0);
334 /* Clean up register/memory stack frame */
335 ia64_adds_imm (code, IA64_SP, framesize, IA64_SP);
336 ia64_mov_to_ar_i (code, IA64_PFS, l5);
338 if (tramp_type == MONO_TRAMPOLINE_CLASS_INIT) {
339 ia64_mov_ret_to_br (code, IA64_B0, l7);
340 ia64_br_ret_reg (code, IA64_B0);
343 /* Call the compiled method */
344 ia64_mov_to_br (code, IA64_B0, l7);
345 ia64_br_cond_reg (code, IA64_B6);
348 ia64_codegen_close (code);
350 g_assert ((code.buf - buf) <= 2048);
352 /* FIXME: emit unwind info for epilog */
353 di = g_malloc0 (sizeof (unw_dyn_info_t));
354 di->start_ip = (unw_word_t) buf;
355 di->end_ip = (unw_word_t) code.buf;
357 di->format = UNW_INFO_FORMAT_DYNAMIC;
358 di->u.pi.name_ptr = (unw_word_t)"ia64_generic_trampoline";
359 di->u.pi.regions = r_pro;
361 _U_dyn_register (di);
363 mono_arch_flush_icache (buf, code.buf - buf);
368 #define TRAMPOLINE_SIZE 128
371 mono_arch_create_specific_trampoline (gpointer arg1, MonoTrampolineType tramp_type, MonoDomain *domain, guint32 *code_len)
375 Ia64CodegenState code;
377 tramp = mono_get_trampoline_code (tramp_type);
379 mono_domain_lock (domain);
380 buf = mono_code_manager_reserve (domain->code_mp, TRAMPOLINE_SIZE);
381 mono_domain_unlock (domain);
383 /* FIXME: Optimize this */
385 ia64_codegen_init (code, buf);
387 ia64_movl (code, GP_SCRATCH_REG, arg1);
389 ia64_begin_bundle (code);
390 disp = (tramp - code.buf) >> 4;
391 if (ia64_is_imm21 (disp)) {
392 ia64_br_cond (code, disp);
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);
400 ia64_codegen_close (code);
402 g_assert (code.buf - buf <= TRAMPOLINE_SIZE);
404 mono_arch_flush_icache (buf, code.buf - buf);
407 *code_len = code.buf - buf;
413 mono_arch_invalidate_method (MonoJitInfo *ji, void *func, gpointer func_arg)
419 mono_arch_create_rgctx_lazy_fetch_trampoline (guint32 encoded_offset)
421 /* FIXME: implement! */
422 g_assert_not_reached ();