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/arch/ia64/ia64-codegen.h>
19 #include "mini-ia64.h"
20 #include "jit-icalls.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 func_addr = ((gpointer*)addr) [0];
47 func_gp = ((gpointer*)addr) [1];
49 buf = mono_domain_code_reserve (domain, 256);
51 /* Since the this reg is a stacked register, its a bit hard to access it */
52 ia64_codegen_init (code, buf);
53 ia64_alloc (code, 40, 8, 1, 0, 0);
54 ia64_adds_imm (code, 32 + this_reg, sizeof (MonoObject), 32 + this_reg);
55 ia64_mov_to_ar_i (code, IA64_PFS, 40);
56 ia64_movl (code, GP_SCRATCH_REG, func_addr);
57 ia64_mov_to_br (code, IA64_B6, GP_SCRATCH_REG);
58 ia64_br_cond_reg (code, IA64_B6);
59 ia64_codegen_close (code);
61 g_assert (code.buf - buf < 256);
63 mono_arch_flush_icache (buf, code.buf - buf);
66 desc = g_malloc0 (sizeof (gpointer) * 2);
70 mono_tramp_info_register (mono_tramp_info_create (NULL, buf, code.buf - buf, NULL, NULL), domain);
76 mono_arch_patch_callsite (guint8 *method_start, guint8 *code, guint8 *addr)
78 guint8 *callsite_begin;
79 guint64 *callsite = (guint64*)(gpointer)(code - 16);
81 guint64 ins, instructions [3];
84 gpointer func = ((gpointer*)(gpointer)addr)[0];
86 while ((ia64_bundle_template (callsite) != IA64_TEMPLATE_MLX) &&
87 (ia64_bundle_template (callsite) != IA64_TEMPLATE_MLXS))
89 callsite_begin = (guint8*)callsite;
91 next_bundle = callsite + 2;
92 ins = ia64_bundle_ins1 (next_bundle);
93 if (ia64_ins_opcode (ins) == 5) {
94 /* ld8_inc_imm -> indirect call through a function pointer */
95 g_assert (ia64_ins_r1 (ins) == GP_SCRATCH_REG2);
96 g_assert (ia64_ins_r3 (ins) == GP_SCRATCH_REG);
100 /* Patch the code generated by emit_call */
102 instructions [0] = ia64_bundle_ins1 (callsite);
103 instructions [1] = ia64_bundle_ins2 (callsite);
104 instructions [2] = ia64_bundle_ins3 (callsite);
106 ia64_codegen_init (gen, (guint8*)buf);
107 ia64_movl (gen, GP_SCRATCH_REG, func);
108 instructions [1] = gen.instructions [0];
109 instructions [2] = gen.instructions [1];
111 ia64_codegen_init (gen, (guint8*)buf);
112 ia64_emit_bundle_template (&gen, ia64_bundle_template (callsite), instructions [0], instructions [1], instructions [2]);
113 ia64_codegen_close (gen);
115 /* This might not be safe, but not all itanium processors support st16 */
116 callsite [0] = buf [0];
117 callsite [1] = buf [1];
119 mono_arch_flush_icache (callsite_begin, code - callsite_begin);
123 mono_arch_patch_plt_entry (guint8 *code, gpointer *got, mgreg_t *regs, guint8 *addr)
125 g_assert_not_reached ();
129 mono_arch_create_generic_trampoline (MonoTrampolineType tramp_type, MonoTrampInfo **info, gboolean aot)
132 int i, offset, saved_regs_offset, saved_fpregs_offset, last_offset, framesize;
133 int in0, local0, out0, l0, l1, l2, l3, l4, l5, l6, l7, l8, o0, o1, o2, o3;
135 Ia64CodegenState code;
137 unw_dyn_region_info_t *r_pro;
143 * Since jump trampolines are not patched, this trampoline is executed every
144 * time a call is made to a jump trampoline. So we try to keep things faster
147 if (tramp_type == MONO_TRAMPOLINE_JUMP)
152 buf = mono_global_codeman_reserve (2048);
154 ia64_codegen_init (code, buf);
156 /* Stacked Registers */
165 l5 = 45; /* saved ar.pfs */
168 l8 = 48; /* saved sp */
169 o0 = out0 + 0; /* regs */
170 o1 = out0 + 1; /* code */
171 o2 = out0 + 2; /* arg */
172 o3 = out0 + 3; /* tramp */
174 framesize = (128 * 8) + 1024;
175 framesize = (framesize + (MONO_ARCH_FRAME_ALIGNMENT - 1)) & ~ (MONO_ARCH_FRAME_ALIGNMENT - 1);
178 * Allocate a new register+memory stack frame.
179 * 8 input registers (the max used by the ABI)
181 * 4 output (number of parameters passed to trampoline)
183 ia64_unw_save_reg (code, UNW_IA64_AR_PFS, UNW_IA64_GR + l5);
184 ia64_alloc (code, l5, local0 - in0, out0 - local0, 4, 0);
185 ia64_unw_save_reg (code, UNW_IA64_SP, UNW_IA64_GR + l8);
186 ia64_mov (code, l8, IA64_SP);
187 ia64_adds_imm (code, IA64_SP, (-framesize), IA64_SP);
189 offset = 16; /* scratch area */
191 /* Save the argument received from the specific trampoline */
192 ia64_mov (code, l6, GP_SCRATCH_REG);
194 /* Save the calling address */
195 ia64_unw_save_reg (code, UNW_IA64_RP, UNW_IA64_GR + local0 + 7);
196 ia64_mov_from_br (code, l7, IA64_B0);
198 /* Create unwind info for the prolog */
199 ia64_begin_bundle (code);
200 r_pro = mono_ia64_create_unwind_region (&code);
203 /* Not needed for jump trampolines */
204 if (tramp_type != MONO_TRAMPOLINE_JUMP) {
205 saved_regs_offset = offset;
208 * Only the registers which are needed for computing vtable slots need
212 for (i = 0; i < 64; ++i)
213 if ((1 << i) & MONO_ARCH_CALLEE_REGS) {
214 if (last_offset != i * 8)
215 ia64_adds_imm (code, l1, saved_regs_offset + (i * 8), IA64_SP);
216 ia64_st8_spill_inc_imm_hint (code, l1, i, 8, 0);
217 last_offset = (i + 1) * 8;
221 /* Save fp registers */
222 saved_fpregs_offset = offset;
224 ia64_adds_imm (code, l1, saved_fpregs_offset, IA64_SP);
225 for (i = 0; i < 8; ++i)
226 ia64_stfd_inc_imm_hint (code, l1, i + 8, 8, 0);
228 g_assert (offset < framesize);
230 /* Arg1 is the pointer to the saved registers */
231 ia64_adds_imm (code, o0, saved_regs_offset, IA64_SP);
233 /* Arg2 is the address of the calling code */
235 ia64_mov (code, o1, l7);
237 ia64_mov (code, o1, 0);
239 /* Arg3 is the method/vtable ptr */
240 ia64_mov (code, o2, l6);
242 /* Arg4 is the trampoline address */
244 ia64_mov (code, o3, 0);
246 tramp = (guint8*)mono_get_trampoline_func (tramp_type);
248 /* Call the trampoline using an indirect call */
249 ia64_movl (code, l0, tramp);
250 ia64_ld8_inc_imm (code, l1, l0, 8);
251 ia64_mov_to_br (code, IA64_B6, l1);
252 ia64_ld8 (code, IA64_GP, l0);
253 ia64_br_call_reg (code, 0, IA64_B6);
255 /* Check for thread interruption */
256 /* This is not perf critical code so no need to check the interrupt flag */
257 ia64_mov (code, l2, IA64_R8);
259 tramp = (guint8*)mono_interruption_checkpoint_from_trampoline;
260 ia64_movl (code, l0, tramp);
261 ia64_ld8_inc_imm (code, l1, l0, 8);
262 ia64_mov_to_br (code, IA64_B6, l1);
263 ia64_ld8 (code, IA64_GP, l0);
264 ia64_br_call_reg (code, 0, IA64_B6);
266 ia64_mov (code, IA64_R8, l2);
268 /* Restore fp regs */
269 ia64_adds_imm (code, l1, saved_fpregs_offset, IA64_SP);
270 for (i = 0; i < 8; ++i)
271 ia64_ldfd_inc_imm (code, i + 8, l1, 8);
273 /* FIXME: Handle NATs in fp regs / scratch regs */
275 /* Load method address from function descriptor */
276 ia64_ld8 (code, l0, IA64_R8);
277 ia64_mov_to_br (code, IA64_B6, l0);
279 /* Clean up register/memory stack frame */
280 ia64_adds_imm (code, IA64_SP, framesize, IA64_SP);
281 ia64_mov_to_ar_i (code, IA64_PFS, l5);
283 /* Call the compiled method */
284 ia64_mov_to_br (code, IA64_B0, l7);
285 ia64_br_cond_reg (code, IA64_B6);
287 ia64_codegen_close (code);
289 g_assert ((code.buf - buf) <= 2048);
291 /* FIXME: emit unwind info for epilog */
292 di = g_malloc0 (sizeof (unw_dyn_info_t));
293 di->start_ip = (unw_word_t) buf;
294 di->end_ip = (unw_word_t) code.buf;
296 di->format = UNW_INFO_FORMAT_DYNAMIC;
297 di->u.pi.name_ptr = (unw_word_t)"ia64_generic_trampoline";
298 di->u.pi.regions = r_pro;
300 _U_dyn_register (di);
302 mono_arch_flush_icache (buf, code.buf - buf);
307 #define TRAMPOLINE_SIZE 128
310 mono_arch_create_specific_trampoline (gpointer arg1, MonoTrampolineType tramp_type, MonoDomain *domain, guint32 *code_len)
314 Ia64CodegenState code;
316 tramp = mono_get_trampoline_code (tramp_type);
318 buf = mono_domain_code_reserve (domain, TRAMPOLINE_SIZE);
320 /* FIXME: Optimize this */
322 ia64_codegen_init (code, buf);
324 ia64_movl (code, GP_SCRATCH_REG, arg1);
326 ia64_begin_bundle (code);
327 disp = (tramp - code.buf) >> 4;
328 if (ia64_is_imm21 (disp)) {
329 ia64_br_cond (code, disp);
332 ia64_movl (code, GP_SCRATCH_REG2, tramp);
333 ia64_mov_to_br (code, IA64_B6, GP_SCRATCH_REG2);
334 ia64_br_cond_reg (code, IA64_B6);
337 ia64_codegen_close (code);
339 g_assert (code.buf - buf <= TRAMPOLINE_SIZE);
341 mono_arch_flush_icache (buf, code.buf - buf);
344 *code_len = code.buf - buf;
350 mono_arch_invalidate_method (MonoJitInfo *ji, void *func, gpointer func_arg)
356 mono_arch_create_rgctx_lazy_fetch_trampoline (guint32 slot, MonoTrampInfo **info, gboolean aot)
358 /* FIXME: implement! */
359 g_assert_not_reached ();