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:
28 * @addr: pointer to native code for @m
30 * when value type methods are called through the vtable we need to unbox the
31 * this argument. This method returns a pointer to a trampoline which does
32 * unboxing before calling the method
35 mono_arch_get_unbox_trampoline (MonoMethod *m, gpointer addr)
38 gpointer func_addr, func_gp;
39 Ia64CodegenState code;
42 MonoDomain *domain = mono_domain_get ();
44 /* FIXME: Optimize this */
46 if (!mono_method_signature (m)->ret->byref && MONO_TYPE_ISSTRUCT (mono_method_signature (m)->ret))
49 func_addr = ((gpointer*)addr) [0];
50 func_gp = ((gpointer*)addr) [1];
52 mono_domain_lock (domain);
53 buf = mono_code_manager_reserve (domain->code_mp, 256);
54 mono_domain_unlock (domain);
56 /* Since the this reg is a stacked register, its a bit hard to access it */
57 ia64_codegen_init (code, buf);
58 ia64_alloc (code, 40, 8, 1, 0, 0);
59 ia64_adds_imm (code, 32 + this_reg, sizeof (MonoObject), 32 + this_reg);
60 ia64_mov_to_ar_i (code, IA64_PFS, 40);
61 ia64_movl (code, GP_SCRATCH_REG, func_addr);
62 ia64_mov_to_br (code, IA64_B6, GP_SCRATCH_REG);
63 ia64_br_cond_reg (code, IA64_B6);
64 ia64_codegen_close (code);
66 g_assert (code.buf - buf < 256);
68 mono_arch_flush_icache (buf, code.buf - buf);
71 desc = g_malloc0 (sizeof (gpointer) * 2);
79 mono_arch_patch_callsite (guint8 *code, guint8 *addr)
81 guint8 *callsite_begin;
82 guint64 *callsite = (guint64*)(gpointer)(code - 16);
84 guint64 ins, instructions [3];
87 gpointer func = ((gpointer*)(gpointer)addr)[0];
89 while ((ia64_bundle_template (callsite) != IA64_TEMPLATE_MLX) &&
90 (ia64_bundle_template (callsite) != IA64_TEMPLATE_MLXS))
92 callsite_begin = (guint8*)callsite;
94 next_bundle = callsite + 2;
95 ins = ia64_bundle_ins1 (next_bundle);
96 if (ia64_ins_opcode (ins) == 5) {
97 /* ld8_inc_imm -> indirect call through a function pointer */
98 g_assert (ia64_ins_r1 (ins) == GP_SCRATCH_REG2);
99 g_assert (ia64_ins_r3 (ins) == GP_SCRATCH_REG);
103 /* Patch the code generated by emit_call */
105 instructions [0] = ia64_bundle_ins1 (callsite);
106 instructions [1] = ia64_bundle_ins2 (callsite);
107 instructions [2] = ia64_bundle_ins3 (callsite);
109 ia64_codegen_init (gen, (guint8*)buf);
110 ia64_movl (gen, GP_SCRATCH_REG, func);
111 instructions [1] = gen.instructions [0];
112 instructions [2] = gen.instructions [1];
114 ia64_codegen_init (gen, (guint8*)buf);
115 ia64_emit_bundle_template (&gen, ia64_bundle_template (callsite), instructions [0], instructions [1], instructions [2]);
116 ia64_codegen_close (gen);
118 /* This might not be safe, but not all itanium processors support st16 */
119 callsite [0] = buf [0];
120 callsite [1] = buf [1];
122 mono_arch_flush_icache (callsite_begin, code - callsite_begin);
126 mono_arch_patch_plt_entry (guint8 *code, guint8 *addr)
128 g_assert_not_reached ();
132 mono_arch_nullify_class_init_trampoline (guint8 *code, gssize *regs)
134 guint8 *callsite_begin;
135 guint64 *callsite = (guint64*)(gpointer)(code - 16);
136 guint64 instructions [3];
138 Ia64CodegenState gen;
140 while ((ia64_bundle_template (callsite) != IA64_TEMPLATE_MLX) &&
141 (ia64_bundle_template (callsite) != IA64_TEMPLATE_MLXS))
143 callsite_begin = (guint8*)callsite;
145 /* Replace the code generated by emit_call with a sets of nops */
147 /* The first bundle might have other instructions in it */
148 instructions [0] = ia64_bundle_ins1 (callsite);
149 instructions [1] = IA64_NOP_X;
150 instructions [2] = IA64_NOP_X;
152 ia64_codegen_init (gen, (guint8*)buf);
153 ia64_emit_bundle_template (&gen, ia64_bundle_template (callsite), instructions [0], instructions [1], instructions [2]);
154 ia64_codegen_close (gen);
156 /* This might not be safe, but not all itanium processors support st16 */
157 callsite [0] = buf [0];
158 callsite [1] = buf [1];
162 /* The other bundles can be full replaced with nops */
164 ia64_codegen_init (gen, (guint8*)buf);
165 ia64_emit_bundle_template (&gen, IA64_TEMPLATE_MII, IA64_NOP_M, IA64_NOP_I, IA64_NOP_I);
166 ia64_codegen_close (gen);
168 while ((guint8*)callsite < code) {
169 callsite [0] = buf [0];
170 callsite [1] = buf [1];
174 mono_arch_flush_icache (callsite_begin, code - callsite_begin);
178 mono_arch_nullify_plt_entry (guint8 *code)
180 g_assert_not_reached ();
184 mono_arch_create_trampoline_code (MonoTrampolineType tramp_type)
187 int i, offset, saved_regs_offset, saved_fpregs_offset, last_offset, framesize;
188 int in0, local0, out0, l0, l1, l2, l3, l4, l5, l6, l7, l8, o0, o1, o2, o3;
190 Ia64CodegenState code;
192 unw_dyn_region_info_t *r_pro;
195 * Since jump trampolines are not patched, this trampoline is executed every
196 * time a call is made to a jump trampoline. So we try to keep things faster
199 if (tramp_type == MONO_TRAMPOLINE_JUMP)
204 buf = mono_global_codeman_reserve (2048);
206 ia64_codegen_init (code, buf);
208 /* Stacked Registers */
217 l5 = 45; /* saved ar.pfs */
220 l8 = 48; /* saved sp */
221 o0 = out0 + 0; /* regs */
222 o1 = out0 + 1; /* code */
223 o2 = out0 + 2; /* arg */
224 o3 = out0 + 3; /* tramp */
226 framesize = (128 * 8) + 1024;
227 framesize = (framesize + (MONO_ARCH_FRAME_ALIGNMENT - 1)) & ~ (MONO_ARCH_FRAME_ALIGNMENT - 1);
230 * Allocate a new register+memory stack frame.
231 * 8 input registers (the max used by the ABI)
233 * 4 output (number of parameters passed to trampoline)
235 ia64_unw_save_reg (code, UNW_IA64_AR_PFS, UNW_IA64_GR + l5);
236 ia64_alloc (code, l5, local0 - in0, out0 - local0, 4, 0);
237 ia64_unw_save_reg (code, UNW_IA64_SP, UNW_IA64_GR + l8);
238 ia64_mov (code, l8, IA64_SP);
239 ia64_adds_imm (code, IA64_SP, (-framesize), IA64_SP);
241 offset = 16; /* scratch area */
243 /* Save the argument received from the specific trampoline */
244 ia64_mov (code, l6, GP_SCRATCH_REG);
246 /* Save the calling address */
247 ia64_unw_save_reg (code, UNW_IA64_RP, UNW_IA64_GR + local0 + 7);
248 ia64_mov_from_br (code, l7, IA64_B0);
250 /* Create unwind info for the prolog */
251 ia64_begin_bundle (code);
252 r_pro = mono_ia64_create_unwind_region (&code);
255 /* Not needed for jump trampolines */
256 if (tramp_type != MONO_TRAMPOLINE_JUMP) {
257 saved_regs_offset = offset;
260 * Only the registers which are needed for computing vtable slots need
264 for (i = 0; i < 64; ++i)
265 if ((1 << i) & MONO_ARCH_CALLEE_REGS) {
266 if (last_offset != i * 8)
267 ia64_adds_imm (code, l1, saved_regs_offset + (i * 8), IA64_SP);
268 ia64_st8_spill_inc_imm_hint (code, l1, i, 8, 0);
269 last_offset = (i + 1) * 8;
273 /* Save fp registers */
274 saved_fpregs_offset = offset;
276 ia64_adds_imm (code, l1, saved_fpregs_offset, IA64_SP);
277 for (i = 0; i < 8; ++i)
278 ia64_stfd_inc_imm_hint (code, l1, i + 8, 8, 0);
280 g_assert (offset < framesize);
282 /* Arg1 is the pointer to the saved registers */
283 ia64_adds_imm (code, o0, saved_regs_offset, IA64_SP);
285 /* Arg2 is the address of the calling code */
287 ia64_mov (code, o1, l7);
289 ia64_mov (code, o1, 0);
291 /* Arg3 is the method/vtable ptr */
292 ia64_mov (code, o2, l6);
294 /* Arg4 is the trampoline address */
296 ia64_mov (code, o3, 0);
298 tramp = (guint8*)mono_get_trampoline_func (tramp_type);
300 /* Call the trampoline using an indirect call */
301 ia64_movl (code, l0, tramp);
302 ia64_ld8_inc_imm (code, l1, l0, 8);
303 ia64_mov_to_br (code, IA64_B6, l1);
304 ia64_ld8 (code, IA64_GP, l0);
305 ia64_br_call_reg (code, 0, IA64_B6);
307 /* Check for thread interruption */
308 /* This is not perf critical code so no need to check the interrupt flag */
309 ia64_mov (code, l2, IA64_R8);
311 tramp = (guint8*)mono_thread_interruption_checkpoint;
312 ia64_movl (code, l0, tramp);
313 ia64_ld8_inc_imm (code, l1, l0, 8);
314 ia64_mov_to_br (code, IA64_B6, l1);
315 ia64_ld8 (code, IA64_GP, l0);
316 ia64_br_call_reg (code, 0, IA64_B6);
318 ia64_mov (code, IA64_R8, l2);
320 /* Restore fp regs */
321 ia64_adds_imm (code, l1, saved_fpregs_offset, IA64_SP);
322 for (i = 0; i < 8; ++i)
323 ia64_ldfd_inc_imm (code, i + 8, l1, 8);
325 /* FIXME: Handle NATs in fp regs / scratch regs */
327 if (tramp_type != MONO_TRAMPOLINE_CLASS_INIT) {
328 /* Load method address from function descriptor */
329 ia64_ld8 (code, l0, IA64_R8);
330 ia64_mov_to_br (code, IA64_B6, l0);
333 /* Clean up register/memory stack frame */
334 ia64_adds_imm (code, IA64_SP, framesize, IA64_SP);
335 ia64_mov_to_ar_i (code, IA64_PFS, l5);
337 if (tramp_type == MONO_TRAMPOLINE_CLASS_INIT) {
338 ia64_mov_ret_to_br (code, IA64_B0, l7);
339 ia64_br_ret_reg (code, IA64_B0);
342 /* Call the compiled method */
343 ia64_mov_to_br (code, IA64_B0, l7);
344 ia64_br_cond_reg (code, IA64_B6);
347 ia64_codegen_close (code);
349 g_assert ((code.buf - buf) <= 2048);
351 /* FIXME: emit unwind info for epilog */
352 di = g_malloc0 (sizeof (unw_dyn_info_t));
353 di->start_ip = (unw_word_t) buf;
354 di->end_ip = (unw_word_t) code.buf;
356 di->format = UNW_INFO_FORMAT_DYNAMIC;
357 di->u.pi.name_ptr = (unw_word_t)"ia64_generic_trampoline";
358 di->u.pi.regions = r_pro;
360 _U_dyn_register (di);
362 mono_arch_flush_icache (buf, code.buf - buf);
367 #define TRAMPOLINE_SIZE 128
370 mono_arch_create_specific_trampoline (gpointer arg1, MonoTrampolineType tramp_type, MonoDomain *domain, guint32 *code_len)
374 Ia64CodegenState code;
376 tramp = mono_get_trampoline_code (tramp_type);
378 mono_domain_lock (domain);
379 buf = mono_code_manager_reserve (domain->code_mp, TRAMPOLINE_SIZE);
380 mono_domain_unlock (domain);
382 /* FIXME: Optimize this */
384 ia64_codegen_init (code, buf);
386 ia64_movl (code, GP_SCRATCH_REG, arg1);
388 ia64_begin_bundle (code);
389 disp = (tramp - code.buf) >> 4;
390 if (ia64_is_imm21 (disp)) {
391 ia64_br_cond (code, disp);
394 ia64_movl (code, GP_SCRATCH_REG2, tramp);
395 ia64_mov_to_br (code, IA64_B6, GP_SCRATCH_REG2);
396 ia64_br_cond_reg (code, IA64_B6);
399 ia64_codegen_close (code);
401 g_assert (code.buf - buf <= TRAMPOLINE_SIZE);
403 mono_arch_flush_icache (buf, code.buf - buf);
406 *code_len = code.buf - buf;
412 mono_arch_invalidate_method (MonoJitInfo *ji, void *func, gpointer func_arg)
418 mono_debugger_create_notification_function (void)