2 * mini-x86.c: x86 backend for the Mono code generator
5 * Paolo Molaro (lupus@ximian.com)
6 * Dietmar Maurer (dietmar@ximian.com)
9 * (C) 2003 Ximian, Inc.
16 #include <mono/metadata/appdomain.h>
17 #include <mono/metadata/debug-helpers.h>
18 #include <mono/metadata/threads.h>
19 #include <mono/metadata/profiler-private.h>
20 #include <mono/utils/mono-math.h>
27 /* On windows, these hold the key returned by TlsAlloc () */
28 static gint lmf_tls_offset = -1;
29 static gint appdomain_tls_offset = -1;
30 static gint thread_tls_offset = -1;
33 /* TRUE by default until we add runtime detection of Xen */
34 static gboolean optimize_for_xen = TRUE;
36 #define optimize_for_xen 0
39 #define ALIGN_TO(val,align) ((((guint64)val) + ((align) - 1)) & ~((align) - 1))
44 /* Under windows, the default pinvoke calling convention is stdcall */
45 #define CALLCONV_IS_STDCALL(sig) ((((sig)->call_convention) == MONO_CALL_STDCALL) || ((sig)->pinvoke && ((sig)->call_convention) == MONO_CALL_DEFAULT))
47 #define CALLCONV_IS_STDCALL(sig) (((sig)->call_convention) == MONO_CALL_STDCALL)
50 #define NOT_IMPLEMENTED g_assert_not_reached ()
53 mono_arch_regname (int reg) {
55 case X86_EAX: return "%eax";
56 case X86_EBX: return "%ebx";
57 case X86_ECX: return "%ecx";
58 case X86_EDX: return "%edx";
59 case X86_ESP: return "%esp"; case X86_EBP: return "%ebp";
60 case X86_EDI: return "%edi";
61 case X86_ESI: return "%esi";
67 mono_arch_fregname (int reg) {
87 /* Only if storage == ArgValuetypeInReg */
88 ArgStorage pair_storage [2];
97 gboolean need_stack_align;
98 guint32 stack_align_amount;
106 #define FLOAT_PARAM_REGS 0
108 static X86_Reg_No param_regs [] = { 0 };
110 #ifdef PLATFORM_WIN32
111 static X86_Reg_No return_regs [] = { X86_EAX, X86_EDX };
115 add_general (guint32 *gr, guint32 *stack_size, ArgInfo *ainfo)
117 ainfo->offset = *stack_size;
119 if (*gr >= PARAM_REGS) {
120 ainfo->storage = ArgOnStack;
121 (*stack_size) += sizeof (gpointer);
124 ainfo->storage = ArgInIReg;
125 ainfo->reg = param_regs [*gr];
131 add_general_pair (guint32 *gr, guint32 *stack_size, ArgInfo *ainfo)
133 ainfo->offset = *stack_size;
135 g_assert (PARAM_REGS == 0);
137 ainfo->storage = ArgOnStack;
138 (*stack_size) += sizeof (gpointer) * 2;
142 add_float (guint32 *gr, guint32 *stack_size, ArgInfo *ainfo, gboolean is_double)
144 ainfo->offset = *stack_size;
146 if (*gr >= FLOAT_PARAM_REGS) {
147 ainfo->storage = ArgOnStack;
148 (*stack_size) += is_double ? 8 : 4;
151 /* A double register */
153 ainfo->storage = ArgInDoubleSSEReg;
155 ainfo->storage = ArgInFloatSSEReg;
163 add_valuetype (MonoMethodSignature *sig, ArgInfo *ainfo, MonoType *type,
165 guint32 *gr, guint32 *fr, guint32 *stack_size)
170 klass = mono_class_from_mono_type (type);
172 size = mono_type_native_stack_size (&klass->byval_arg, NULL);
174 size = mono_type_stack_size (&klass->byval_arg, NULL);
176 #ifdef PLATFORM_WIN32
177 if (sig->pinvoke && is_return) {
178 MonoMarshalType *info;
181 * the exact rules are not very well documented, the code below seems to work with the
182 * code generated by gcc 3.3.3 -mno-cygwin.
184 info = mono_marshal_load_type_info (klass);
187 ainfo->pair_storage [0] = ainfo->pair_storage [1] = ArgNone;
189 /* Special case structs with only a float member */
190 if ((info->native_size == 8) && (info->num_fields == 1) && (info->fields [0].field->type->type == MONO_TYPE_R8)) {
191 ainfo->storage = ArgValuetypeInReg;
192 ainfo->pair_storage [0] = ArgOnDoubleFpStack;
195 if ((info->native_size == 4) && (info->num_fields == 1) && (info->fields [0].field->type->type == MONO_TYPE_R4)) {
196 ainfo->storage = ArgValuetypeInReg;
197 ainfo->pair_storage [0] = ArgOnFloatFpStack;
200 if ((info->native_size == 1) || (info->native_size == 2) || (info->native_size == 4) || (info->native_size == 8)) {
201 ainfo->storage = ArgValuetypeInReg;
202 ainfo->pair_storage [0] = ArgInIReg;
203 ainfo->pair_regs [0] = return_regs [0];
204 if (info->native_size > 4) {
205 ainfo->pair_storage [1] = ArgInIReg;
206 ainfo->pair_regs [1] = return_regs [1];
213 ainfo->offset = *stack_size;
214 ainfo->storage = ArgOnStack;
215 *stack_size += ALIGN_TO (size, sizeof (gpointer));
221 * Obtain information about a call according to the calling convention.
222 * For x86 ELF, see the "System V Application Binary Interface Intel386
223 * Architecture Processor Supplment, Fourth Edition" document for more
225 * For x86 win32, see ???.
228 get_call_info (MonoMethodSignature *sig, gboolean is_pinvoke)
232 int n = sig->hasthis + sig->param_count;
233 guint32 stack_size = 0;
236 cinfo = g_malloc0 (sizeof (CallInfo) + (sizeof (ArgInfo) * n));
243 ret_type = mono_type_get_underlying_type (sig->ret);
244 switch (ret_type->type) {
245 case MONO_TYPE_BOOLEAN:
256 case MONO_TYPE_FNPTR:
257 case MONO_TYPE_CLASS:
258 case MONO_TYPE_OBJECT:
259 case MONO_TYPE_SZARRAY:
260 case MONO_TYPE_ARRAY:
261 case MONO_TYPE_STRING:
262 cinfo->ret.storage = ArgInIReg;
263 cinfo->ret.reg = X86_EAX;
267 cinfo->ret.storage = ArgInIReg;
268 cinfo->ret.reg = X86_EAX;
271 cinfo->ret.storage = ArgOnFloatFpStack;
274 cinfo->ret.storage = ArgOnDoubleFpStack;
276 case MONO_TYPE_GENERICINST:
277 if (!mono_type_generic_inst_is_valuetype (sig->ret)) {
278 cinfo->ret.storage = ArgInIReg;
279 cinfo->ret.reg = X86_EAX;
283 case MONO_TYPE_VALUETYPE: {
284 guint32 tmp_gr = 0, tmp_fr = 0, tmp_stacksize = 0;
286 add_valuetype (sig, &cinfo->ret, sig->ret, TRUE, &tmp_gr, &tmp_fr, &tmp_stacksize);
287 if (cinfo->ret.storage == ArgOnStack)
288 /* The caller passes the address where the value is stored */
289 add_general (&gr, &stack_size, &cinfo->ret);
292 case MONO_TYPE_TYPEDBYREF:
293 /* Same as a valuetype with size 24 */
294 add_general (&gr, &stack_size, &cinfo->ret);
298 cinfo->ret.storage = ArgNone;
301 g_error ("Can't handle as return value 0x%x", sig->ret->type);
307 add_general (&gr, &stack_size, cinfo->args + 0);
309 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG) && (n == 0)) {
311 fr = FLOAT_PARAM_REGS;
313 /* Emit the signature cookie just before the implicit arguments */
314 add_general (&gr, &stack_size, &cinfo->sig_cookie);
317 for (i = 0; i < sig->param_count; ++i) {
318 ArgInfo *ainfo = &cinfo->args [sig->hasthis + i];
321 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG) && (i == sig->sentinelpos)) {
322 /* We allways pass the sig cookie on the stack for simplicity */
324 * Prevent implicit arguments + the sig cookie from being passed
328 fr = FLOAT_PARAM_REGS;
330 /* Emit the signature cookie just before the implicit arguments */
331 add_general (&gr, &stack_size, &cinfo->sig_cookie);
334 if (sig->params [i]->byref) {
335 add_general (&gr, &stack_size, ainfo);
338 ptype = mono_type_get_underlying_type (sig->params [i]);
339 switch (ptype->type) {
340 case MONO_TYPE_BOOLEAN:
343 add_general (&gr, &stack_size, ainfo);
348 add_general (&gr, &stack_size, ainfo);
352 add_general (&gr, &stack_size, ainfo);
357 case MONO_TYPE_FNPTR:
358 case MONO_TYPE_CLASS:
359 case MONO_TYPE_OBJECT:
360 case MONO_TYPE_STRING:
361 case MONO_TYPE_SZARRAY:
362 case MONO_TYPE_ARRAY:
363 add_general (&gr, &stack_size, ainfo);
365 case MONO_TYPE_GENERICINST:
366 if (!mono_type_generic_inst_is_valuetype (sig->params [i])) {
367 add_general (&gr, &stack_size, ainfo);
371 case MONO_TYPE_VALUETYPE:
372 add_valuetype (sig, ainfo, sig->params [i], FALSE, &gr, &fr, &stack_size);
374 case MONO_TYPE_TYPEDBYREF:
375 stack_size += sizeof (MonoTypedRef);
376 ainfo->storage = ArgOnStack;
380 add_general_pair (&gr, &stack_size, ainfo);
383 add_float (&fr, &stack_size, ainfo, FALSE);
386 add_float (&fr, &stack_size, ainfo, TRUE);
389 g_error ("unexpected type 0x%x", ptype->type);
390 g_assert_not_reached ();
394 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG) && (n > 0) && (sig->sentinelpos == sig->param_count)) {
396 fr = FLOAT_PARAM_REGS;
398 /* Emit the signature cookie just before the implicit arguments */
399 add_general (&gr, &stack_size, &cinfo->sig_cookie);
402 #if defined(__APPLE__)
403 if ((stack_size % 16) != 0) {
404 cinfo->need_stack_align = TRUE;
405 stack_size += cinfo->stack_align_amount = 16-(stack_size % 16);
409 cinfo->stack_usage = stack_size;
410 cinfo->reg_usage = gr;
411 cinfo->freg_usage = fr;
416 * mono_arch_get_argument_info:
417 * @csig: a method signature
418 * @param_count: the number of parameters to consider
419 * @arg_info: an array to store the result infos
421 * Gathers information on parameters such as size, alignment and
422 * padding. arg_info should be large enought to hold param_count + 1 entries.
424 * Returns the size of the activation frame.
427 mono_arch_get_argument_info (MonoMethodSignature *csig, int param_count, MonoJitArgumentInfo *arg_info)
429 int k, frame_size = 0;
435 cinfo = get_call_info (csig, FALSE);
437 if (MONO_TYPE_ISSTRUCT (csig->ret) && (cinfo->ret.storage == ArgOnStack)) {
438 frame_size += sizeof (gpointer);
442 arg_info [0].offset = offset;
445 frame_size += sizeof (gpointer);
449 arg_info [0].size = frame_size;
451 for (k = 0; k < param_count; k++) {
454 size = mono_type_native_stack_size (csig->params [k], &align);
457 size = mono_type_stack_size (csig->params [k], &ialign);
461 /* ignore alignment for now */
464 frame_size += pad = (align - (frame_size & (align - 1))) & (align - 1);
465 arg_info [k].pad = pad;
467 arg_info [k + 1].pad = 0;
468 arg_info [k + 1].size = size;
470 arg_info [k + 1].offset = offset;
474 align = MONO_ARCH_FRAME_ALIGNMENT;
475 frame_size += pad = (align - (frame_size & (align - 1))) & (align - 1);
476 arg_info [k].pad = pad;
483 static const guchar cpuid_impl [] = {
484 0x55, /* push %ebp */
485 0x89, 0xe5, /* mov %esp,%ebp */
486 0x53, /* push %ebx */
487 0x8b, 0x45, 0x08, /* mov 0x8(%ebp),%eax */
488 0x0f, 0xa2, /* cpuid */
489 0x50, /* push %eax */
490 0x8b, 0x45, 0x10, /* mov 0x10(%ebp),%eax */
491 0x89, 0x18, /* mov %ebx,(%eax) */
492 0x8b, 0x45, 0x14, /* mov 0x14(%ebp),%eax */
493 0x89, 0x08, /* mov %ecx,(%eax) */
494 0x8b, 0x45, 0x18, /* mov 0x18(%ebp),%eax */
495 0x89, 0x10, /* mov %edx,(%eax) */
497 0x8b, 0x55, 0x0c, /* mov 0xc(%ebp),%edx */
498 0x89, 0x02, /* mov %eax,(%edx) */
504 typedef void (*CpuidFunc) (int id, int* p_eax, int* p_ebx, int* p_ecx, int* p_edx);
507 cpuid (int id, int* p_eax, int* p_ebx, int* p_ecx, int* p_edx)
511 __asm__ __volatile__ (
514 "movl %%eax, %%edx\n"
515 "xorl $0x200000, %%eax\n"
520 "xorl %%edx, %%eax\n"
521 "andl $0x200000, %%eax\n"
543 /* Have to use the code manager to get around WinXP DEP */
544 MonoCodeManager *codeman = mono_code_manager_new_dynamic ();
546 void *ptr = mono_code_manager_reserve (codeman, sizeof (cpuid_impl));
547 memcpy (ptr, cpuid_impl, sizeof (cpuid_impl));
549 func = (CpuidFunc)ptr;
550 func (id, p_eax, p_ebx, p_ecx, p_edx);
552 mono_code_manager_destroy (codeman);
555 * We use this approach because of issues with gcc and pic code, see:
556 * http://gcc.gnu.org/cgi-bin/gnatsweb.pl?cmd=view%20audit-trail&database=gcc&pr=7329
557 __asm__ __volatile__ ("cpuid"
558 : "=a" (*p_eax), "=b" (*p_ebx), "=c" (*p_ecx), "=d" (*p_edx)
567 * Initialize the cpu to execute managed code.
570 mono_arch_cpu_init (void)
572 /* spec compliance requires running with double precision */
576 __asm__ __volatile__ ("fnstcw %0\n": "=m" (fpcw));
577 fpcw &= ~X86_FPCW_PRECC_MASK;
578 fpcw |= X86_FPCW_PREC_DOUBLE;
579 __asm__ __volatile__ ("fldcw %0\n": : "m" (fpcw));
580 __asm__ __volatile__ ("fnstcw %0\n": "=m" (fpcw));
582 _control87 (_PC_53, MCW_PC);
587 * This function returns the optimizations supported on this cpu.
590 mono_arch_cpu_optimizazions (guint32 *exclude_mask)
592 int eax, ebx, ecx, edx;
596 /* Feature Flags function, flags returned in EDX. */
597 if (cpuid (1, &eax, &ebx, &ecx, &edx)) {
598 if (edx & (1 << 15)) {
599 opts |= MONO_OPT_CMOV;
601 opts |= MONO_OPT_FCMOV;
603 *exclude_mask |= MONO_OPT_FCMOV;
605 *exclude_mask |= MONO_OPT_CMOV;
611 * Determine whenever the trap whose info is in SIGINFO is caused by
615 mono_arch_is_int_overflow (void *sigctx, void *info)
620 mono_arch_sigctx_to_monoctx (sigctx, &ctx);
622 ip = (guint8*)ctx.eip;
624 if ((ip [0] == 0xf7) && (x86_modrm_mod (ip [1]) == 0x3) && (x86_modrm_reg (ip [1]) == 0x7)) {
628 switch (x86_modrm_rm (ip [1])) {
648 g_assert_not_reached ();
660 is_regsize_var (MonoType *t) {
663 switch (mono_type_get_underlying_type (t)->type) {
669 case MONO_TYPE_FNPTR:
671 case MONO_TYPE_OBJECT:
672 case MONO_TYPE_STRING:
673 case MONO_TYPE_CLASS:
674 case MONO_TYPE_SZARRAY:
675 case MONO_TYPE_ARRAY:
677 case MONO_TYPE_GENERICINST:
678 if (!mono_type_generic_inst_is_valuetype (t))
681 case MONO_TYPE_VALUETYPE:
688 mono_arch_get_allocatable_int_vars (MonoCompile *cfg)
693 for (i = 0; i < cfg->num_varinfo; i++) {
694 MonoInst *ins = cfg->varinfo [i];
695 MonoMethodVar *vmv = MONO_VARINFO (cfg, i);
698 if (vmv->range.first_use.abs_pos >= vmv->range.last_use.abs_pos)
701 if ((ins->flags & (MONO_INST_IS_DEAD|MONO_INST_VOLATILE|MONO_INST_INDIRECT)) ||
702 (ins->opcode != OP_LOCAL && ins->opcode != OP_ARG))
705 /* we dont allocate I1 to registers because there is no simply way to sign extend
706 * 8bit quantities in caller saved registers on x86 */
707 if (is_regsize_var (ins->inst_vtype) || (ins->inst_vtype->type == MONO_TYPE_BOOLEAN) ||
708 (ins->inst_vtype->type == MONO_TYPE_U1) || (ins->inst_vtype->type == MONO_TYPE_U2)||
709 (ins->inst_vtype->type == MONO_TYPE_I2) || (ins->inst_vtype->type == MONO_TYPE_CHAR)) {
710 g_assert (MONO_VARINFO (cfg, i)->reg == -1);
711 g_assert (i == vmv->idx);
712 vars = g_list_prepend (vars, vmv);
716 vars = mono_varlist_sort (cfg, vars, 0);
722 mono_arch_get_global_int_regs (MonoCompile *cfg)
726 /* we can use 3 registers for global allocation */
727 regs = g_list_prepend (regs, (gpointer)X86_EBX);
728 regs = g_list_prepend (regs, (gpointer)X86_ESI);
729 regs = g_list_prepend (regs, (gpointer)X86_EDI);
735 * mono_arch_regalloc_cost:
737 * Return the cost, in number of memory references, of the action of
738 * allocating the variable VMV into a register during global register
742 mono_arch_regalloc_cost (MonoCompile *cfg, MonoMethodVar *vmv)
744 MonoInst *ins = cfg->varinfo [vmv->idx];
746 if (cfg->method->save_lmf)
747 /* The register is already saved */
748 return (ins->opcode == OP_ARG) ? 1 : 0;
750 /* push+pop+possible load if it is an argument */
751 return (ins->opcode == OP_ARG) ? 3 : 2;
755 * Set var information according to the calling convention. X86 version.
756 * The locals var stuff should most likely be split in another method.
759 mono_arch_allocate_vars (MonoCompile *cfg)
761 MonoMethodSignature *sig;
762 MonoMethodHeader *header;
764 guint32 locals_stack_size, locals_stack_align;
769 header = mono_method_get_header (cfg->method);
770 sig = mono_method_signature (cfg->method);
772 cinfo = get_call_info (sig, FALSE);
774 cfg->frame_reg = MONO_ARCH_BASEREG;
777 /* Reserve space to save LMF and caller saved registers */
779 if (cfg->method->save_lmf) {
780 offset += sizeof (MonoLMF);
782 if (cfg->used_int_regs & (1 << X86_EBX)) {
786 if (cfg->used_int_regs & (1 << X86_EDI)) {
790 if (cfg->used_int_regs & (1 << X86_ESI)) {
795 switch (cinfo->ret.storage) {
796 case ArgValuetypeInReg:
797 /* Allocate a local to hold the result, the epilog will copy it to the correct place */
799 cfg->ret->opcode = OP_REGOFFSET;
800 cfg->ret->inst_basereg = X86_EBP;
801 cfg->ret->inst_offset = - offset;
807 /* Allocate locals */
808 offsets = mono_allocate_stack_slots (cfg, &locals_stack_size, &locals_stack_align);
809 if (locals_stack_align) {
810 offset += (locals_stack_align - 1);
811 offset &= ~(locals_stack_align - 1);
813 for (i = cfg->locals_start; i < cfg->num_varinfo; i++) {
814 if (offsets [i] != -1) {
815 MonoInst *inst = cfg->varinfo [i];
816 inst->opcode = OP_REGOFFSET;
817 inst->inst_basereg = X86_EBP;
818 inst->inst_offset = - (offset + offsets [i]);
819 //printf ("allocated local %d to ", i); mono_print_tree_nl (inst);
823 offset += locals_stack_size;
827 * Allocate arguments+return value
830 switch (cinfo->ret.storage) {
832 cfg->ret->opcode = OP_REGOFFSET;
833 cfg->ret->inst_basereg = X86_EBP;
834 cfg->ret->inst_offset = cinfo->ret.offset + ARGS_OFFSET;
836 case ArgValuetypeInReg:
839 cfg->ret->opcode = OP_REGVAR;
840 cfg->ret->inst_c0 = cinfo->ret.reg;
843 case ArgOnFloatFpStack:
844 case ArgOnDoubleFpStack:
847 g_assert_not_reached ();
850 if (sig->call_convention == MONO_CALL_VARARG) {
851 g_assert (cinfo->sig_cookie.storage == ArgOnStack);
852 cfg->sig_cookie = cinfo->sig_cookie.offset + ARGS_OFFSET;
855 for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
856 ArgInfo *ainfo = &cinfo->args [i];
857 inst = cfg->varinfo [i];
858 if (inst->opcode != OP_REGVAR) {
859 inst->opcode = OP_REGOFFSET;
860 inst->inst_basereg = X86_EBP;
862 inst->inst_offset = ainfo->offset + ARGS_OFFSET;
865 offset += (MONO_ARCH_FRAME_ALIGNMENT - 1);
866 offset &= ~(MONO_ARCH_FRAME_ALIGNMENT - 1);
868 cfg->stack_offset = offset;
874 mono_arch_create_vars (MonoCompile *cfg)
876 MonoMethodSignature *sig;
879 sig = mono_method_signature (cfg->method);
881 cinfo = get_call_info (sig, FALSE);
883 if (cinfo->ret.storage == ArgValuetypeInReg)
884 cfg->ret_var_is_local = TRUE;
889 /* Fixme: we need an alignment solution for enter_method and mono_arch_call_opcode,
890 * currently alignment in mono_arch_call_opcode is computed without arch_get_argument_info
894 emit_sig_cookie (MonoCompile *cfg, MonoCallInst *call)
897 MonoMethodSignature *tmp_sig;
900 /* FIXME: Add support for signature tokens to AOT */
901 cfg->disable_aot = TRUE;
902 MONO_INST_NEW (cfg, arg, OP_OUTARG);
905 * mono_ArgIterator_Setup assumes the signature cookie is
906 * passed first and all the arguments which were before it are
907 * passed on the stack after the signature. So compensate by
908 * passing a different signature.
910 tmp_sig = mono_metadata_signature_dup (call->signature);
911 tmp_sig->param_count -= call->signature->sentinelpos;
912 tmp_sig->sentinelpos = 0;
913 memcpy (tmp_sig->params, call->signature->params + call->signature->sentinelpos, tmp_sig->param_count * sizeof (MonoType*));
915 MONO_INST_NEW (cfg, sig_arg, OP_ICONST);
916 sig_arg->inst_p0 = tmp_sig;
918 arg->inst_left = sig_arg;
919 arg->type = STACK_PTR;
920 /* prepend, so they get reversed */
921 arg->next = call->out_args;
922 call->out_args = arg;
926 * take the arguments and generate the arch-specific
927 * instructions to properly call the function in call.
928 * This includes pushing, moving arguments to the right register
932 mono_arch_call_opcode (MonoCompile *cfg, MonoBasicBlock* bb, MonoCallInst *call, int is_virtual) {
934 MonoMethodSignature *sig;
939 sig = call->signature;
940 n = sig->param_count + sig->hasthis;
942 cinfo = get_call_info (sig, FALSE);
944 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG))
945 sentinelpos = sig->sentinelpos + (is_virtual ? 1 : 0);
947 for (i = 0; i < n; ++i) {
948 ArgInfo *ainfo = cinfo->args + i;
950 /* Emit the signature cookie just before the implicit arguments */
951 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG) && (i == sentinelpos)) {
952 emit_sig_cookie (cfg, call);
955 if (is_virtual && i == 0) {
956 /* the argument will be attached to the call instrucion */
961 if (i >= sig->hasthis)
962 t = sig->params [i - sig->hasthis];
964 t = &mono_defaults.int_class->byval_arg;
965 t = mono_type_get_underlying_type (t);
967 MONO_INST_NEW (cfg, arg, OP_OUTARG);
969 arg->cil_code = in->cil_code;
971 arg->type = in->type;
972 /* prepend, so they get reversed */
973 arg->next = call->out_args;
974 call->out_args = arg;
976 if ((i >= sig->hasthis) && (MONO_TYPE_ISSTRUCT(t))) {
979 if (t->type == MONO_TYPE_TYPEDBYREF) {
980 size = sizeof (MonoTypedRef);
981 align = sizeof (gpointer);
985 size = mono_type_native_stack_size (&in->klass->byval_arg, &align);
988 size = mono_type_stack_size (&in->klass->byval_arg, &ialign);
991 arg->opcode = OP_OUTARG_VT;
992 arg->klass = in->klass;
993 arg->backend.is_pinvoke = sig->pinvoke;
994 arg->inst_imm = size;
997 switch (ainfo->storage) {
999 arg->opcode = OP_OUTARG;
1001 if (t->type == MONO_TYPE_R4)
1002 arg->opcode = OP_OUTARG_R4;
1004 if (t->type == MONO_TYPE_R8)
1005 arg->opcode = OP_OUTARG_R8;
1009 g_assert_not_reached ();
1015 /* Handle the case where there are no implicit arguments */
1016 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG) && (n == sentinelpos)) {
1017 emit_sig_cookie (cfg, call);
1020 if (sig->ret && MONO_TYPE_ISSTRUCT (sig->ret)) {
1021 if (cinfo->ret.storage == ArgValuetypeInReg) {
1022 MonoInst *zero_inst;
1024 * After the call, the struct is in registers, but needs to be saved to the memory pointed
1025 * to by vt_arg in this_vret_args. This means that vt_arg needs to be saved somewhere
1026 * before calling the function. So we add a dummy instruction to represent pushing the
1027 * struct return address to the stack. The return address will be saved to this stack slot
1028 * by the code emitted in this_vret_args.
1030 MONO_INST_NEW (cfg, arg, OP_OUTARG);
1031 MONO_INST_NEW (cfg, zero_inst, OP_ICONST);
1032 zero_inst->inst_p0 = 0;
1033 arg->inst_left = zero_inst;
1034 arg->type = STACK_PTR;
1035 /* prepend, so they get reversed */
1036 arg->next = call->out_args;
1037 call->out_args = arg;
1040 /* if the function returns a struct, the called method already does a ret $0x4 */
1041 if (sig->ret && MONO_TYPE_ISSTRUCT (sig->ret))
1042 cinfo->stack_usage -= 4;
1045 call->stack_usage = cinfo->stack_usage;
1047 #if defined(__APPLE__)
1048 if (cinfo->need_stack_align) {
1049 MONO_INST_NEW (cfg, arg, OP_X86_OUTARG_ALIGN_STACK);
1050 arg->inst_c0 = cinfo->stack_align_amount;
1051 arg->next = call->out_args;
1052 call->out_args = arg;
1062 * Allow tracing to work with this interface (with an optional argument)
1065 mono_arch_instrument_prolog (MonoCompile *cfg, void *func, void *p, gboolean enable_arguments)
1069 /* if some args are passed in registers, we need to save them here */
1070 x86_push_reg (code, X86_EBP);
1072 if (cfg->compile_aot) {
1073 x86_push_imm (code, cfg->method);
1074 x86_mov_reg_imm (code, X86_EAX, func);
1075 x86_call_reg (code, X86_EAX);
1077 mono_add_patch_info (cfg, code-cfg->native_code, MONO_PATCH_INFO_METHODCONST, cfg->method);
1078 x86_push_imm (code, cfg->method);
1079 mono_add_patch_info (cfg, code-cfg->native_code, MONO_PATCH_INFO_ABS, func);
1080 x86_call_code (code, 0);
1082 x86_alu_reg_imm (code, X86_ADD, X86_ESP, 8);
1096 mono_arch_instrument_epilog (MonoCompile *cfg, void *func, void *p, gboolean enable_arguments)
1099 int arg_size = 0, save_mode = SAVE_NONE;
1100 MonoMethod *method = cfg->method;
1102 switch (mono_type_get_underlying_type (mono_method_signature (method)->ret)->type) {
1103 case MONO_TYPE_VOID:
1104 /* special case string .ctor icall */
1105 if (strcmp (".ctor", method->name) && method->klass == mono_defaults.string_class)
1106 save_mode = SAVE_EAX;
1108 save_mode = SAVE_NONE;
1112 save_mode = SAVE_EAX_EDX;
1116 save_mode = SAVE_FP;
1118 case MONO_TYPE_GENERICINST:
1119 if (!mono_type_generic_inst_is_valuetype (mono_method_signature (method)->ret)) {
1120 save_mode = SAVE_EAX;
1124 case MONO_TYPE_VALUETYPE:
1125 save_mode = SAVE_STRUCT;
1128 save_mode = SAVE_EAX;
1132 switch (save_mode) {
1134 x86_push_reg (code, X86_EDX);
1135 x86_push_reg (code, X86_EAX);
1136 if (enable_arguments) {
1137 x86_push_reg (code, X86_EDX);
1138 x86_push_reg (code, X86_EAX);
1143 x86_push_reg (code, X86_EAX);
1144 if (enable_arguments) {
1145 x86_push_reg (code, X86_EAX);
1150 x86_alu_reg_imm (code, X86_SUB, X86_ESP, 8);
1151 x86_fst_membase (code, X86_ESP, 0, TRUE, TRUE);
1152 if (enable_arguments) {
1153 x86_alu_reg_imm (code, X86_SUB, X86_ESP, 8);
1154 x86_fst_membase (code, X86_ESP, 0, TRUE, TRUE);
1159 if (enable_arguments) {
1160 x86_push_membase (code, X86_EBP, 8);
1169 if (cfg->compile_aot) {
1170 x86_push_imm (code, method);
1171 x86_mov_reg_imm (code, X86_EAX, func);
1172 x86_call_reg (code, X86_EAX);
1174 mono_add_patch_info (cfg, code-cfg->native_code, MONO_PATCH_INFO_METHODCONST, method);
1175 x86_push_imm (code, method);
1176 mono_add_patch_info (cfg, code-cfg->native_code, MONO_PATCH_INFO_ABS, func);
1177 x86_call_code (code, 0);
1179 x86_alu_reg_imm (code, X86_ADD, X86_ESP, arg_size + 4);
1181 switch (save_mode) {
1183 x86_pop_reg (code, X86_EAX);
1184 x86_pop_reg (code, X86_EDX);
1187 x86_pop_reg (code, X86_EAX);
1190 x86_fld_membase (code, X86_ESP, 0, TRUE);
1191 x86_alu_reg_imm (code, X86_ADD, X86_ESP, 8);
1201 #define EMIT_COND_BRANCH(ins,cond,sign) \
1202 if (ins->flags & MONO_INST_BRLABEL) { \
1203 if (ins->inst_i0->inst_c0) { \
1204 x86_branch (code, cond, cfg->native_code + ins->inst_i0->inst_c0, sign); \
1206 mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_LABEL, ins->inst_i0); \
1207 if ((cfg->opt & MONO_OPT_BRANCH) && \
1208 x86_is_imm8 (ins->inst_i0->inst_c1 - cpos)) \
1209 x86_branch8 (code, cond, 0, sign); \
1211 x86_branch32 (code, cond, 0, sign); \
1214 if (ins->inst_true_bb->native_offset) { \
1215 x86_branch (code, cond, cfg->native_code + ins->inst_true_bb->native_offset, sign); \
1217 mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_BB, ins->inst_true_bb); \
1218 if ((cfg->opt & MONO_OPT_BRANCH) && \
1219 x86_is_imm8 (ins->inst_true_bb->max_offset - cpos)) \
1220 x86_branch8 (code, cond, 0, sign); \
1222 x86_branch32 (code, cond, 0, sign); \
1227 * Emit an exception if condition is fail and
1228 * if possible do a directly branch to target
1230 #define EMIT_COND_SYSTEM_EXCEPTION(cond,signed,exc_name) \
1232 MonoInst *tins = mono_branch_optimize_exception_target (cfg, bb, exc_name); \
1233 if (tins == NULL) { \
1234 mono_add_patch_info (cfg, code - cfg->native_code, \
1235 MONO_PATCH_INFO_EXC, exc_name); \
1236 x86_branch32 (code, cond, 0, signed); \
1238 EMIT_COND_BRANCH (tins, cond, signed); \
1242 #define EMIT_FPCOMPARE(code) do { \
1243 x86_fcompp (code); \
1244 x86_fnstsw (code); \
1249 emit_call (MonoCompile *cfg, guint8 *code, guint32 patch_type, gconstpointer data)
1251 mono_add_patch_info (cfg, code - cfg->native_code, patch_type, data);
1252 x86_call_code (code, 0);
1257 /* FIXME: Add more instructions */
1258 #define INST_IGNORES_CFLAGS(ins) (((ins)->opcode == CEE_BR) || ((ins)->opcode == OP_STORE_MEMBASE_IMM) || ((ins)->opcode == OP_STOREI4_MEMBASE_REG))
1261 peephole_pass (MonoCompile *cfg, MonoBasicBlock *bb)
1263 MonoInst *ins, *last_ins = NULL;
1268 switch (ins->opcode) {
1270 /* reg = 0 -> XOR (reg, reg) */
1271 /* XOR sets cflags on x86, so we cant do it always */
1272 if (ins->inst_c0 == 0 && ins->next && INST_IGNORES_CFLAGS (ins->next)) {
1273 ins->opcode = CEE_XOR;
1274 ins->sreg1 = ins->dreg;
1275 ins->sreg2 = ins->dreg;
1279 /* remove unnecessary multiplication with 1 */
1280 if (ins->inst_imm == 1) {
1281 if (ins->dreg != ins->sreg1) {
1282 ins->opcode = OP_MOVE;
1284 last_ins->next = ins->next;
1290 case OP_COMPARE_IMM:
1291 /* OP_COMPARE_IMM (reg, 0)
1293 * OP_X86_TEST_NULL (reg)
1296 ins->opcode = OP_X86_TEST_NULL;
1298 case OP_X86_COMPARE_MEMBASE_IMM:
1300 * OP_STORE_MEMBASE_REG reg, offset(basereg)
1301 * OP_X86_COMPARE_MEMBASE_IMM offset(basereg), imm
1303 * OP_STORE_MEMBASE_REG reg, offset(basereg)
1304 * OP_COMPARE_IMM reg, imm
1306 * Note: if imm = 0 then OP_COMPARE_IMM replaced with OP_X86_TEST_NULL
1308 if (last_ins && (last_ins->opcode == OP_STOREI4_MEMBASE_REG) &&
1309 ins->inst_basereg == last_ins->inst_destbasereg &&
1310 ins->inst_offset == last_ins->inst_offset) {
1311 ins->opcode = OP_COMPARE_IMM;
1312 ins->sreg1 = last_ins->sreg1;
1314 /* check if we can remove cmp reg,0 with test null */
1316 ins->opcode = OP_X86_TEST_NULL;
1320 case OP_LOAD_MEMBASE:
1321 case OP_LOADI4_MEMBASE:
1323 * Note: if reg1 = reg2 the load op is removed
1325 * OP_STORE_MEMBASE_REG reg1, offset(basereg)
1326 * OP_LOAD_MEMBASE offset(basereg), reg2
1328 * OP_STORE_MEMBASE_REG reg1, offset(basereg)
1329 * OP_MOVE reg1, reg2
1331 if (last_ins && (last_ins->opcode == OP_STOREI4_MEMBASE_REG
1332 || last_ins->opcode == OP_STORE_MEMBASE_REG) &&
1333 ins->inst_basereg == last_ins->inst_destbasereg &&
1334 ins->inst_offset == last_ins->inst_offset) {
1335 if (ins->dreg == last_ins->sreg1) {
1336 last_ins->next = ins->next;
1340 //static int c = 0; printf ("MATCHX %s %d\n", cfg->method->name,c++);
1341 ins->opcode = OP_MOVE;
1342 ins->sreg1 = last_ins->sreg1;
1346 * Note: reg1 must be different from the basereg in the second load
1347 * Note: if reg1 = reg2 is equal then second load is removed
1349 * OP_LOAD_MEMBASE offset(basereg), reg1
1350 * OP_LOAD_MEMBASE offset(basereg), reg2
1352 * OP_LOAD_MEMBASE offset(basereg), reg1
1353 * OP_MOVE reg1, reg2
1355 } if (last_ins && (last_ins->opcode == OP_LOADI4_MEMBASE
1356 || last_ins->opcode == OP_LOAD_MEMBASE) &&
1357 ins->inst_basereg != last_ins->dreg &&
1358 ins->inst_basereg == last_ins->inst_basereg &&
1359 ins->inst_offset == last_ins->inst_offset) {
1361 if (ins->dreg == last_ins->dreg) {
1362 last_ins->next = ins->next;
1366 ins->opcode = OP_MOVE;
1367 ins->sreg1 = last_ins->dreg;
1370 //g_assert_not_reached ();
1374 * OP_STORE_MEMBASE_IMM imm, offset(basereg)
1375 * OP_LOAD_MEMBASE offset(basereg), reg
1377 * OP_STORE_MEMBASE_IMM imm, offset(basereg)
1378 * OP_ICONST reg, imm
1380 } else if (last_ins && (last_ins->opcode == OP_STOREI4_MEMBASE_IMM
1381 || last_ins->opcode == OP_STORE_MEMBASE_IMM) &&
1382 ins->inst_basereg == last_ins->inst_destbasereg &&
1383 ins->inst_offset == last_ins->inst_offset) {
1384 //static int c = 0; printf ("MATCHX %s %d\n", cfg->method->name,c++);
1385 ins->opcode = OP_ICONST;
1386 ins->inst_c0 = last_ins->inst_imm;
1387 g_assert_not_reached (); // check this rule
1391 case OP_LOADU1_MEMBASE:
1392 case OP_LOADI1_MEMBASE:
1394 * OP_STORE_MEMBASE_REG reg1, offset(basereg)
1395 * OP_LOAD_MEMBASE offset(basereg), reg2
1397 * OP_STORE_MEMBASE_REG reg1, offset(basereg)
1398 * CONV_I2/U2 reg1, reg2
1400 if (last_ins && X86_IS_BYTE_REG (last_ins->sreg1) &&
1401 (last_ins->opcode == OP_STOREI1_MEMBASE_REG) &&
1402 ins->inst_basereg == last_ins->inst_destbasereg &&
1403 ins->inst_offset == last_ins->inst_offset) {
1404 ins->opcode = (ins->opcode == OP_LOADI1_MEMBASE) ? CEE_CONV_I1 : CEE_CONV_U1;
1405 ins->sreg1 = last_ins->sreg1;
1408 case OP_LOADU2_MEMBASE:
1409 case OP_LOADI2_MEMBASE:
1411 * OP_STORE_MEMBASE_REG reg1, offset(basereg)
1412 * OP_LOAD_MEMBASE offset(basereg), reg2
1414 * OP_STORE_MEMBASE_REG reg1, offset(basereg)
1415 * CONV_I2/U2 reg1, reg2
1417 if (last_ins && (last_ins->opcode == OP_STOREI2_MEMBASE_REG) &&
1418 ins->inst_basereg == last_ins->inst_destbasereg &&
1419 ins->inst_offset == last_ins->inst_offset) {
1420 ins->opcode = (ins->opcode == OP_LOADI2_MEMBASE) ? CEE_CONV_I2 : CEE_CONV_U2;
1421 ins->sreg1 = last_ins->sreg1;
1432 if (ins->dreg == ins->sreg1) {
1434 last_ins->next = ins->next;
1441 * OP_MOVE sreg, dreg
1442 * OP_MOVE dreg, sreg
1444 if (last_ins && last_ins->opcode == OP_MOVE &&
1445 ins->sreg1 == last_ins->dreg &&
1446 ins->dreg == last_ins->sreg1) {
1447 last_ins->next = ins->next;
1453 case OP_X86_PUSH_MEMBASE:
1454 if (last_ins && (last_ins->opcode == OP_STOREI4_MEMBASE_REG ||
1455 last_ins->opcode == OP_STORE_MEMBASE_REG) &&
1456 ins->inst_basereg == last_ins->inst_destbasereg &&
1457 ins->inst_offset == last_ins->inst_offset) {
1458 ins->opcode = OP_X86_PUSH;
1459 ins->sreg1 = last_ins->sreg1;
1466 bb->last_ins = last_ins;
1470 branch_cc_table [] = {
1471 X86_CC_EQ, X86_CC_GE, X86_CC_GT, X86_CC_LE, X86_CC_LT,
1472 X86_CC_NE, X86_CC_GE, X86_CC_GT, X86_CC_LE, X86_CC_LT,
1473 X86_CC_O, X86_CC_NO, X86_CC_C, X86_CC_NC
1476 static const char*const * ins_spec = x86_desc;
1478 /*#include "cprop.c"*/
1480 mono_arch_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
1482 mono_local_regalloc (cfg, bb);
1485 static unsigned char*
1486 emit_float_to_int (MonoCompile *cfg, guchar *code, int dreg, int size, gboolean is_signed)
1488 x86_alu_reg_imm (code, X86_SUB, X86_ESP, 4);
1489 x86_fnstcw_membase(code, X86_ESP, 0);
1490 x86_mov_reg_membase (code, dreg, X86_ESP, 0, 2);
1491 x86_alu_reg_imm (code, X86_OR, dreg, 0xc00);
1492 x86_mov_membase_reg (code, X86_ESP, 2, dreg, 2);
1493 x86_fldcw_membase (code, X86_ESP, 2);
1495 x86_alu_reg_imm (code, X86_SUB, X86_ESP, 8);
1496 x86_fist_pop_membase (code, X86_ESP, 0, TRUE);
1497 x86_pop_reg (code, dreg);
1498 /* FIXME: need the high register
1499 * x86_pop_reg (code, dreg_high);
1502 x86_push_reg (code, X86_EAX); // SP = SP - 4
1503 x86_fist_pop_membase (code, X86_ESP, 0, FALSE);
1504 x86_pop_reg (code, dreg);
1506 x86_fldcw_membase (code, X86_ESP, 0);
1507 x86_alu_reg_imm (code, X86_ADD, X86_ESP, 4);
1510 x86_widen_reg (code, dreg, dreg, is_signed, FALSE);
1512 x86_widen_reg (code, dreg, dreg, is_signed, TRUE);
1516 static unsigned char*
1517 mono_emit_stack_alloc (guchar *code, MonoInst* tree)
1519 int sreg = tree->sreg1;
1520 int need_touch = FALSE;
1522 #if defined(PLATFORM_WIN32) || defined(MONO_ARCH_SIGSEGV_ON_ALTSTACK)
1531 * If requested stack size is larger than one page,
1532 * perform stack-touch operation
1535 * Generate stack probe code.
1536 * Under Windows, it is necessary to allocate one page at a time,
1537 * "touching" stack after each successful sub-allocation. This is
1538 * because of the way stack growth is implemented - there is a
1539 * guard page before the lowest stack page that is currently commited.
1540 * Stack normally grows sequentially so OS traps access to the
1541 * guard page and commits more pages when needed.
1543 x86_test_reg_imm (code, sreg, ~0xFFF);
1544 br[0] = code; x86_branch8 (code, X86_CC_Z, 0, FALSE);
1546 br[2] = code; /* loop */
1547 x86_alu_reg_imm (code, X86_SUB, X86_ESP, 0x1000);
1548 x86_test_membase_reg (code, X86_ESP, 0, X86_ESP);
1551 * By the end of the loop, sreg2 is smaller than 0x1000, so the init routine
1552 * that follows only initializes the last part of the area.
1554 /* Same as the init code below with size==0x1000 */
1555 if (tree->flags & MONO_INST_INIT) {
1556 x86_push_reg (code, X86_EAX);
1557 x86_push_reg (code, X86_ECX);
1558 x86_push_reg (code, X86_EDI);
1559 x86_mov_reg_imm (code, X86_ECX, (0x1000 >> 2));
1560 x86_alu_reg_reg (code, X86_XOR, X86_EAX, X86_EAX);
1561 x86_lea_membase (code, X86_EDI, X86_ESP, 12);
1563 x86_prefix (code, X86_REP_PREFIX);
1565 x86_pop_reg (code, X86_EDI);
1566 x86_pop_reg (code, X86_ECX);
1567 x86_pop_reg (code, X86_EAX);
1570 x86_alu_reg_imm (code, X86_SUB, sreg, 0x1000);
1571 x86_alu_reg_imm (code, X86_CMP, sreg, 0x1000);
1572 br[3] = code; x86_branch8 (code, X86_CC_AE, 0, FALSE);
1573 x86_patch (br[3], br[2]);
1574 x86_test_reg_reg (code, sreg, sreg);
1575 br[4] = code; x86_branch8 (code, X86_CC_Z, 0, FALSE);
1576 x86_alu_reg_reg (code, X86_SUB, X86_ESP, sreg);
1578 br[1] = code; x86_jump8 (code, 0);
1580 x86_patch (br[0], code);
1581 x86_alu_reg_reg (code, X86_SUB, X86_ESP, sreg);
1582 x86_patch (br[1], code);
1583 x86_patch (br[4], code);
1586 x86_alu_reg_reg (code, X86_SUB, X86_ESP, tree->sreg1);
1588 if (tree->flags & MONO_INST_INIT) {
1590 if (tree->dreg != X86_EAX && sreg != X86_EAX) {
1591 x86_push_reg (code, X86_EAX);
1594 if (tree->dreg != X86_ECX && sreg != X86_ECX) {
1595 x86_push_reg (code, X86_ECX);
1598 if (tree->dreg != X86_EDI && sreg != X86_EDI) {
1599 x86_push_reg (code, X86_EDI);
1603 x86_shift_reg_imm (code, X86_SHR, sreg, 2);
1604 if (sreg != X86_ECX)
1605 x86_mov_reg_reg (code, X86_ECX, sreg, 4);
1606 x86_alu_reg_reg (code, X86_XOR, X86_EAX, X86_EAX);
1608 x86_lea_membase (code, X86_EDI, X86_ESP, offset);
1610 x86_prefix (code, X86_REP_PREFIX);
1613 if (tree->dreg != X86_EDI && sreg != X86_EDI)
1614 x86_pop_reg (code, X86_EDI);
1615 if (tree->dreg != X86_ECX && sreg != X86_ECX)
1616 x86_pop_reg (code, X86_ECX);
1617 if (tree->dreg != X86_EAX && sreg != X86_EAX)
1618 x86_pop_reg (code, X86_EAX);
1625 emit_move_return_value (MonoCompile *cfg, MonoInst *ins, guint8 *code)
1630 /* Move return value to the target register */
1631 switch (ins->opcode) {
1634 case OP_CALL_MEMBASE:
1635 if (ins->dreg != X86_EAX)
1636 x86_mov_reg_reg (code, ins->dreg, X86_EAX, 4);
1640 case OP_VCALL_MEMBASE:
1641 cinfo = get_call_info (((MonoCallInst*)ins)->signature, FALSE);
1642 if (cinfo->ret.storage == ArgValuetypeInReg) {
1643 /* Pop the destination address from the stack */
1644 x86_pop_reg (code, X86_ECX);
1646 for (quad = 0; quad < 2; quad ++) {
1647 switch (cinfo->ret.pair_storage [quad]) {
1649 g_assert (cinfo->ret.pair_regs [quad] != X86_ECX);
1650 x86_mov_membase_reg (code, X86_ECX, (quad * sizeof (gpointer)), cinfo->ret.pair_regs [quad], sizeof (gpointer));
1655 g_assert_not_reached ();
1669 * @code: buffer to store code to
1670 * @dreg: hard register where to place the result
1671 * @tls_offset: offset info
1673 * emit_tls_get emits in @code the native code that puts in the dreg register
1674 * the item in the thread local storage identified by tls_offset.
1676 * Returns: a pointer to the end of the stored code
1679 emit_tls_get (guint8* code, int dreg, int tls_offset)
1681 #ifdef PLATFORM_WIN32
1683 * See the Under the Hood article in the May 1996 issue of Microsoft Systems
1684 * Journal and/or a disassembly of the TlsGet () function.
1686 g_assert (tls_offset < 64);
1687 x86_prefix (code, X86_FS_PREFIX);
1688 x86_mov_reg_mem (code, dreg, 0x18, 4);
1689 /* Dunno what this does but TlsGetValue () contains it */
1690 x86_alu_membase_imm (code, X86_AND, dreg, 0x34, 0);
1691 x86_mov_reg_membase (code, dreg, dreg, 3600 + (tls_offset * 4), 4);
1693 if (optimize_for_xen) {
1694 x86_prefix (code, X86_GS_PREFIX);
1695 x86_mov_reg_mem (code, dreg, 0, 4);
1696 x86_mov_reg_membase (code, dreg, dreg, tls_offset, 4);
1698 x86_prefix (code, X86_GS_PREFIX);
1699 x86_mov_reg_mem (code, dreg, tls_offset, 4);
1705 #define REAL_PRINT_REG(text,reg) \
1706 mono_assert (reg >= 0); \
1707 x86_push_reg (code, X86_EAX); \
1708 x86_push_reg (code, X86_EDX); \
1709 x86_push_reg (code, X86_ECX); \
1710 x86_push_reg (code, reg); \
1711 x86_push_imm (code, reg); \
1712 x86_push_imm (code, text " %d %p\n"); \
1713 x86_mov_reg_imm (code, X86_EAX, printf); \
1714 x86_call_reg (code, X86_EAX); \
1715 x86_alu_reg_imm (code, X86_ADD, X86_ESP, 3*4); \
1716 x86_pop_reg (code, X86_ECX); \
1717 x86_pop_reg (code, X86_EDX); \
1718 x86_pop_reg (code, X86_EAX);
1720 /* benchmark and set based on cpu */
1721 #define LOOP_ALIGNMENT 8
1722 #define bb_is_loop_start(bb) ((bb)->loop_body_start && (bb)->nesting)
1725 mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
1730 guint8 *code = cfg->native_code + cfg->code_len;
1731 MonoInst *last_ins = NULL;
1732 guint last_offset = 0;
1735 if (cfg->opt & MONO_OPT_PEEPHOLE)
1736 peephole_pass (cfg, bb);
1738 if (cfg->opt & MONO_OPT_LOOP) {
1739 int pad, align = LOOP_ALIGNMENT;
1740 /* set alignment depending on cpu */
1741 if (bb_is_loop_start (bb) && (pad = (cfg->code_len & (align - 1)))) {
1743 /*g_print ("adding %d pad at %x to loop in %s\n", pad, cfg->code_len, cfg->method->name);*/
1744 x86_padding (code, pad);
1745 cfg->code_len += pad;
1746 bb->native_offset = cfg->code_len;
1750 if (cfg->verbose_level > 2)
1751 g_print ("Basic block %d starting at offset 0x%x\n", bb->block_num, bb->native_offset);
1753 cpos = bb->max_offset;
1755 if (cfg->prof_options & MONO_PROFILE_COVERAGE) {
1756 MonoProfileCoverageInfo *cov = cfg->coverage_info;
1757 g_assert (!cfg->compile_aot);
1760 cov->data [bb->dfn].cil_code = bb->cil_code;
1761 /* this is not thread save, but good enough */
1762 x86_inc_mem (code, &cov->data [bb->dfn].count);
1765 offset = code - cfg->native_code;
1767 mono_debug_open_block (cfg, bb, offset);
1771 offset = code - cfg->native_code;
1773 max_len = ((guint8 *)ins_spec [ins->opcode])[MONO_INST_LEN];
1775 if (offset > (cfg->code_size - max_len - 16)) {
1776 cfg->code_size *= 2;
1777 cfg->native_code = g_realloc (cfg->native_code, cfg->code_size);
1778 code = cfg->native_code + offset;
1779 mono_jit_stats.code_reallocs++;
1782 mono_debug_record_line_number (cfg, ins, offset);
1784 switch (ins->opcode) {
1786 x86_mul_reg (code, ins->sreg2, TRUE);
1789 x86_mul_reg (code, ins->sreg2, FALSE);
1791 case OP_X86_SETEQ_MEMBASE:
1792 case OP_X86_SETNE_MEMBASE:
1793 x86_set_membase (code, ins->opcode == OP_X86_SETEQ_MEMBASE ? X86_CC_EQ : X86_CC_NE,
1794 ins->inst_basereg, ins->inst_offset, TRUE);
1796 case OP_STOREI1_MEMBASE_IMM:
1797 x86_mov_membase_imm (code, ins->inst_destbasereg, ins->inst_offset, ins->inst_imm, 1);
1799 case OP_STOREI2_MEMBASE_IMM:
1800 x86_mov_membase_imm (code, ins->inst_destbasereg, ins->inst_offset, ins->inst_imm, 2);
1802 case OP_STORE_MEMBASE_IMM:
1803 case OP_STOREI4_MEMBASE_IMM:
1804 x86_mov_membase_imm (code, ins->inst_destbasereg, ins->inst_offset, ins->inst_imm, 4);
1806 case OP_STOREI1_MEMBASE_REG:
1807 x86_mov_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, ins->sreg1, 1);
1809 case OP_STOREI2_MEMBASE_REG:
1810 x86_mov_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, ins->sreg1, 2);
1812 case OP_STORE_MEMBASE_REG:
1813 case OP_STOREI4_MEMBASE_REG:
1814 x86_mov_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, ins->sreg1, 4);
1819 x86_mov_reg_mem (code, ins->dreg, ins->inst_p0, 4);
1822 x86_mov_reg_imm (code, ins->dreg, ins->inst_p0);
1823 x86_mov_reg_membase (code, ins->dreg, ins->dreg, 0, 4);
1825 case OP_LOAD_MEMBASE:
1826 case OP_LOADI4_MEMBASE:
1827 case OP_LOADU4_MEMBASE:
1828 x86_mov_reg_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, 4);
1830 case OP_LOADU1_MEMBASE:
1831 x86_widen_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, FALSE, FALSE);
1833 case OP_LOADI1_MEMBASE:
1834 x86_widen_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, TRUE, FALSE);
1836 case OP_LOADU2_MEMBASE:
1837 x86_widen_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, FALSE, TRUE);
1839 case OP_LOADI2_MEMBASE:
1840 x86_widen_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, TRUE, TRUE);
1843 x86_widen_reg (code, ins->dreg, ins->sreg1, TRUE, FALSE);
1846 x86_widen_reg (code, ins->dreg, ins->sreg1, TRUE, TRUE);
1849 x86_widen_reg (code, ins->dreg, ins->sreg1, FALSE, FALSE);
1852 x86_widen_reg (code, ins->dreg, ins->sreg1, FALSE, TRUE);
1855 x86_alu_reg_reg (code, X86_CMP, ins->sreg1, ins->sreg2);
1857 case OP_COMPARE_IMM:
1858 x86_alu_reg_imm (code, X86_CMP, ins->sreg1, ins->inst_imm);
1860 case OP_X86_COMPARE_MEMBASE_REG:
1861 x86_alu_membase_reg (code, X86_CMP, ins->inst_basereg, ins->inst_offset, ins->sreg2);
1863 case OP_X86_COMPARE_MEMBASE_IMM:
1864 x86_alu_membase_imm (code, X86_CMP, ins->inst_basereg, ins->inst_offset, ins->inst_imm);
1866 case OP_X86_COMPARE_MEMBASE8_IMM:
1867 x86_alu_membase8_imm (code, X86_CMP, ins->inst_basereg, ins->inst_offset, ins->inst_imm);
1869 case OP_X86_COMPARE_REG_MEMBASE:
1870 x86_alu_reg_membase (code, X86_CMP, ins->sreg1, ins->sreg2, ins->inst_offset);
1872 case OP_X86_COMPARE_MEM_IMM:
1873 x86_alu_mem_imm (code, X86_CMP, ins->inst_offset, ins->inst_imm);
1875 case OP_X86_TEST_NULL:
1876 x86_test_reg_reg (code, ins->sreg1, ins->sreg1);
1878 case OP_X86_ADD_MEMBASE_IMM:
1879 x86_alu_membase_imm (code, X86_ADD, ins->inst_basereg, ins->inst_offset, ins->inst_imm);
1881 case OP_X86_ADD_MEMBASE:
1882 x86_alu_reg_membase (code, X86_ADD, ins->sreg1, ins->sreg2, ins->inst_offset);
1884 case OP_X86_SUB_MEMBASE_IMM:
1885 x86_alu_membase_imm (code, X86_SUB, ins->inst_basereg, ins->inst_offset, ins->inst_imm);
1887 case OP_X86_SUB_MEMBASE:
1888 x86_alu_reg_membase (code, X86_SUB, ins->sreg1, ins->sreg2, ins->inst_offset);
1890 case OP_X86_AND_MEMBASE_IMM:
1891 x86_alu_membase_imm (code, X86_AND, ins->inst_basereg, ins->inst_offset, ins->inst_imm);
1893 case OP_X86_OR_MEMBASE_IMM:
1894 x86_alu_membase_imm (code, X86_OR, ins->inst_basereg, ins->inst_offset, ins->inst_imm);
1896 case OP_X86_XOR_MEMBASE_IMM:
1897 x86_alu_membase_imm (code, X86_XOR, ins->inst_basereg, ins->inst_offset, ins->inst_imm);
1899 case OP_X86_INC_MEMBASE:
1900 x86_inc_membase (code, ins->inst_basereg, ins->inst_offset);
1902 case OP_X86_INC_REG:
1903 x86_inc_reg (code, ins->dreg);
1905 case OP_X86_DEC_MEMBASE:
1906 x86_dec_membase (code, ins->inst_basereg, ins->inst_offset);
1908 case OP_X86_DEC_REG:
1909 x86_dec_reg (code, ins->dreg);
1911 case OP_X86_MUL_MEMBASE:
1912 x86_imul_reg_membase (code, ins->sreg1, ins->sreg2, ins->inst_offset);
1915 x86_breakpoint (code);
1919 x86_alu_reg_reg (code, X86_ADD, ins->sreg1, ins->sreg2);
1922 x86_alu_reg_reg (code, X86_ADC, ins->sreg1, ins->sreg2);
1926 x86_alu_reg_imm (code, X86_ADD, ins->dreg, ins->inst_imm);
1929 x86_alu_reg_imm (code, X86_ADC, ins->dreg, ins->inst_imm);
1933 x86_alu_reg_reg (code, X86_SUB, ins->sreg1, ins->sreg2);
1936 x86_alu_reg_reg (code, X86_SBB, ins->sreg1, ins->sreg2);
1940 x86_alu_reg_imm (code, X86_SUB, ins->dreg, ins->inst_imm);
1943 x86_alu_reg_imm (code, X86_SBB, ins->dreg, ins->inst_imm);
1946 x86_alu_reg_reg (code, X86_AND, ins->sreg1, ins->sreg2);
1949 x86_alu_reg_imm (code, X86_AND, ins->sreg1, ins->inst_imm);
1953 x86_div_reg (code, ins->sreg2, TRUE);
1956 x86_alu_reg_reg (code, X86_XOR, X86_EDX, X86_EDX);
1957 x86_div_reg (code, ins->sreg2, FALSE);
1960 x86_mov_reg_imm (code, ins->sreg2, ins->inst_imm);
1962 x86_div_reg (code, ins->sreg2, TRUE);
1966 x86_div_reg (code, ins->sreg2, TRUE);
1969 x86_alu_reg_reg (code, X86_XOR, X86_EDX, X86_EDX);
1970 x86_div_reg (code, ins->sreg2, FALSE);
1973 x86_mov_reg_imm (code, ins->sreg2, ins->inst_imm);
1975 x86_div_reg (code, ins->sreg2, TRUE);
1978 x86_alu_reg_reg (code, X86_OR, ins->sreg1, ins->sreg2);
1981 x86_alu_reg_imm (code, X86_OR, ins->sreg1, ins->inst_imm);
1984 x86_alu_reg_reg (code, X86_XOR, ins->sreg1, ins->sreg2);
1987 x86_alu_reg_imm (code, X86_XOR, ins->sreg1, ins->inst_imm);
1990 g_assert (ins->sreg2 == X86_ECX);
1991 x86_shift_reg (code, X86_SHL, ins->dreg);
1994 g_assert (ins->sreg2 == X86_ECX);
1995 x86_shift_reg (code, X86_SAR, ins->dreg);
1998 x86_shift_reg_imm (code, X86_SAR, ins->dreg, ins->inst_imm);
2001 x86_shift_reg_imm (code, X86_SHR, ins->dreg, ins->inst_imm);
2004 g_assert (ins->sreg2 == X86_ECX);
2005 x86_shift_reg (code, X86_SHR, ins->dreg);
2008 x86_shift_reg_imm (code, X86_SHL, ins->dreg, ins->inst_imm);
2011 guint8 *jump_to_end;
2013 /* handle shifts below 32 bits */
2014 x86_shld_reg (code, ins->backend.reg3, ins->sreg1);
2015 x86_shift_reg (code, X86_SHL, ins->sreg1);
2017 x86_test_reg_imm (code, X86_ECX, 32);
2018 jump_to_end = code; x86_branch8 (code, X86_CC_EQ, 0, TRUE);
2020 /* handle shift over 32 bit */
2021 x86_mov_reg_reg (code, ins->backend.reg3, ins->sreg1, 4);
2022 x86_clear_reg (code, ins->sreg1);
2024 x86_patch (jump_to_end, code);
2028 guint8 *jump_to_end;
2030 /* handle shifts below 32 bits */
2031 x86_shrd_reg (code, ins->sreg1, ins->backend.reg3);
2032 x86_shift_reg (code, X86_SAR, ins->backend.reg3);
2034 x86_test_reg_imm (code, X86_ECX, 32);
2035 jump_to_end = code; x86_branch8 (code, X86_CC_EQ, 0, FALSE);
2037 /* handle shifts over 31 bits */
2038 x86_mov_reg_reg (code, ins->sreg1, ins->backend.reg3, 4);
2039 x86_shift_reg_imm (code, X86_SAR, ins->backend.reg3, 31);
2041 x86_patch (jump_to_end, code);
2045 guint8 *jump_to_end;
2047 /* handle shifts below 32 bits */
2048 x86_shrd_reg (code, ins->sreg1, ins->backend.reg3);
2049 x86_shift_reg (code, X86_SHR, ins->backend.reg3);
2051 x86_test_reg_imm (code, X86_ECX, 32);
2052 jump_to_end = code; x86_branch8 (code, X86_CC_EQ, 0, FALSE);
2054 /* handle shifts over 31 bits */
2055 x86_mov_reg_reg (code, ins->sreg1, ins->backend.reg3, 4);
2056 x86_clear_reg (code, ins->backend.reg3);
2058 x86_patch (jump_to_end, code);
2062 if (ins->inst_imm >= 32) {
2063 x86_mov_reg_reg (code, ins->backend.reg3, ins->sreg1, 4);
2064 x86_clear_reg (code, ins->sreg1);
2065 x86_shift_reg_imm (code, X86_SHL, ins->backend.reg3, ins->inst_imm - 32);
2067 x86_shld_reg_imm (code, ins->backend.reg3, ins->sreg1, ins->inst_imm);
2068 x86_shift_reg_imm (code, X86_SHL, ins->sreg1, ins->inst_imm);
2072 if (ins->inst_imm >= 32) {
2073 x86_mov_reg_reg (code, ins->sreg1, ins->backend.reg3, 4);
2074 x86_shift_reg_imm (code, X86_SAR, ins->backend.reg3, 0x1f);
2075 x86_shift_reg_imm (code, X86_SAR, ins->sreg1, ins->inst_imm - 32);
2077 x86_shrd_reg_imm (code, ins->sreg1, ins->backend.reg3, ins->inst_imm);
2078 x86_shift_reg_imm (code, X86_SAR, ins->backend.reg3, ins->inst_imm);
2081 case OP_LSHR_UN_IMM:
2082 if (ins->inst_imm >= 32) {
2083 x86_mov_reg_reg (code, ins->sreg1, ins->backend.reg3, 4);
2084 x86_clear_reg (code, ins->backend.reg3);
2085 x86_shift_reg_imm (code, X86_SHR, ins->sreg1, ins->inst_imm - 32);
2087 x86_shrd_reg_imm (code, ins->sreg1, ins->backend.reg3, ins->inst_imm);
2088 x86_shift_reg_imm (code, X86_SHR, ins->backend.reg3, ins->inst_imm);
2092 x86_not_reg (code, ins->sreg1);
2095 x86_neg_reg (code, ins->sreg1);
2098 x86_widen_reg (code, ins->dreg, ins->sreg1, TRUE, FALSE);
2101 x86_widen_reg (code, ins->dreg, ins->sreg1, TRUE, TRUE);
2104 x86_imul_reg_reg (code, ins->sreg1, ins->sreg2);
2107 switch (ins->inst_imm) {
2111 if (ins->dreg != ins->sreg1)
2112 x86_mov_reg_reg (code, ins->dreg, ins->sreg1, 4);
2113 x86_alu_reg_reg (code, X86_ADD, ins->dreg, ins->dreg);
2116 /* LEA r1, [r2 + r2*2] */
2117 x86_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 1);
2120 /* LEA r1, [r2 + r2*4] */
2121 x86_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 2);
2124 /* LEA r1, [r2 + r2*2] */
2126 x86_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 1);
2127 x86_alu_reg_reg (code, X86_ADD, ins->dreg, ins->dreg);
2130 /* LEA r1, [r2 + r2*8] */
2131 x86_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 3);
2134 /* LEA r1, [r2 + r2*4] */
2136 x86_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 2);
2137 x86_alu_reg_reg (code, X86_ADD, ins->dreg, ins->dreg);
2140 /* LEA r1, [r2 + r2*2] */
2142 x86_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 1);
2143 x86_shift_reg_imm (code, X86_SHL, ins->dreg, 2);
2146 /* LEA r1, [r2 + r2*4] */
2147 /* LEA r1, [r1 + r1*4] */
2148 x86_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 2);
2149 x86_lea_memindex (code, ins->dreg, ins->dreg, 0, ins->dreg, 2);
2152 /* LEA r1, [r2 + r2*4] */
2154 /* LEA r1, [r1 + r1*4] */
2155 x86_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 2);
2156 x86_shift_reg_imm (code, X86_SHL, ins->dreg, 2);
2157 x86_lea_memindex (code, ins->dreg, ins->dreg, 0, ins->dreg, 2);
2160 x86_imul_reg_reg_imm (code, ins->dreg, ins->sreg1, ins->inst_imm);
2165 x86_imul_reg_reg (code, ins->sreg1, ins->sreg2);
2166 EMIT_COND_SYSTEM_EXCEPTION (X86_CC_O, FALSE, "OverflowException");
2168 case CEE_MUL_OVF_UN: {
2169 /* the mul operation and the exception check should most likely be split */
2170 int non_eax_reg, saved_eax = FALSE, saved_edx = FALSE;
2171 /*g_assert (ins->sreg2 == X86_EAX);
2172 g_assert (ins->dreg == X86_EAX);*/
2173 if (ins->sreg2 == X86_EAX) {
2174 non_eax_reg = ins->sreg1;
2175 } else if (ins->sreg1 == X86_EAX) {
2176 non_eax_reg = ins->sreg2;
2178 /* no need to save since we're going to store to it anyway */
2179 if (ins->dreg != X86_EAX) {
2181 x86_push_reg (code, X86_EAX);
2183 x86_mov_reg_reg (code, X86_EAX, ins->sreg1, 4);
2184 non_eax_reg = ins->sreg2;
2186 if (ins->dreg == X86_EDX) {
2189 x86_push_reg (code, X86_EAX);
2191 } else if (ins->dreg != X86_EAX) {
2193 x86_push_reg (code, X86_EDX);
2195 x86_mul_reg (code, non_eax_reg, FALSE);
2196 /* save before the check since pop and mov don't change the flags */
2197 if (ins->dreg != X86_EAX)
2198 x86_mov_reg_reg (code, ins->dreg, X86_EAX, 4);
2200 x86_pop_reg (code, X86_EDX);
2202 x86_pop_reg (code, X86_EAX);
2203 EMIT_COND_SYSTEM_EXCEPTION (X86_CC_O, FALSE, "OverflowException");
2207 x86_mov_reg_imm (code, ins->dreg, ins->inst_c0);
2210 g_assert_not_reached ();
2211 mono_add_patch_info (cfg, offset, (MonoJumpInfoType)ins->inst_i1, ins->inst_p0);
2212 x86_mov_reg_imm (code, ins->dreg, 0);
2214 case OP_LOAD_GOTADDR:
2215 x86_call_imm (code, 0);
2217 * The patch needs to point to the pop, since the GOT offset needs
2218 * to be added to that address.
2220 mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_GOT_OFFSET, NULL);
2221 x86_pop_reg (code, ins->dreg);
2222 x86_alu_reg_imm (code, X86_ADD, ins->dreg, 0xf0f0f0f0);
2225 mono_add_patch_info (cfg, offset, (MonoJumpInfoType)ins->inst_right->inst_i1, ins->inst_right->inst_p0);
2226 x86_mov_reg_membase (code, ins->dreg, ins->inst_basereg, 0xf0f0f0f0, 4);
2228 case OP_X86_PUSH_GOT_ENTRY:
2229 mono_add_patch_info (cfg, offset, (MonoJumpInfoType)ins->inst_right->inst_i1, ins->inst_right->inst_p0);
2230 x86_push_membase (code, ins->inst_basereg, 0xf0f0f0f0);
2234 x86_mov_reg_reg (code, ins->dreg, ins->sreg1, 4);
2237 g_assert_not_reached ();
2240 * Note: this 'frame destruction' logic is useful for tail calls, too.
2241 * Keep in sync with the code in emit_epilog.
2245 /* FIXME: no tracing support... */
2246 if (cfg->prof_options & MONO_PROFILE_ENTER_LEAVE)
2247 code = mono_arch_instrument_epilog (cfg, mono_profiler_method_leave, code, FALSE);
2248 /* reset offset to make max_len work */
2249 offset = code - cfg->native_code;
2251 g_assert (!cfg->method->save_lmf);
2253 if (cfg->used_int_regs & (1 << X86_EBX))
2255 if (cfg->used_int_regs & (1 << X86_EDI))
2257 if (cfg->used_int_regs & (1 << X86_ESI))
2260 x86_lea_membase (code, X86_ESP, X86_EBP, pos);
2262 if (cfg->used_int_regs & (1 << X86_ESI))
2263 x86_pop_reg (code, X86_ESI);
2264 if (cfg->used_int_regs & (1 << X86_EDI))
2265 x86_pop_reg (code, X86_EDI);
2266 if (cfg->used_int_regs & (1 << X86_EBX))
2267 x86_pop_reg (code, X86_EBX);
2269 /* restore ESP/EBP */
2271 offset = code - cfg->native_code;
2272 mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_METHOD_JUMP, ins->inst_p0);
2273 x86_jump32 (code, 0);
2277 /* ensure ins->sreg1 is not NULL
2278 * note that cmp DWORD PTR [eax], eax is one byte shorter than
2279 * cmp DWORD PTR [eax], 0
2281 x86_alu_membase_reg (code, X86_CMP, ins->sreg1, 0, ins->sreg1);
2284 int hreg = ins->sreg1 == X86_EAX? X86_ECX: X86_EAX;
2285 x86_push_reg (code, hreg);
2286 x86_lea_membase (code, hreg, X86_EBP, cfg->sig_cookie);
2287 x86_mov_membase_reg (code, ins->sreg1, 0, hreg, 4);
2288 x86_pop_reg (code, hreg);
2296 call = (MonoCallInst*)ins;
2297 if (ins->flags & MONO_INST_HAS_METHOD)
2298 code = emit_call (cfg, code, MONO_PATCH_INFO_METHOD, call->method);
2300 code = emit_call (cfg, code, MONO_PATCH_INFO_ABS, call->fptr);
2301 if (call->stack_usage && !CALLCONV_IS_STDCALL (call->signature)) {
2302 /* a pop is one byte, while an add reg, imm is 3. So if there are 4 or 8
2303 * bytes to pop, we want to use pops. GCC does this (note it won't happen
2304 * for P4 or i686 because gcc will avoid using pop push at all. But we aren't
2305 * smart enough to do that optimization yet
2307 * It turns out that on my P4, doing two pops for 8 bytes on the stack makes
2308 * mcs botstrap slow down. However, doing 1 pop for 4 bytes creates a small,
2309 * (most likely from locality benefits). People with other processors should
2310 * check on theirs to see what happens.
2312 if (call->stack_usage == 4) {
2313 /* we want to use registers that won't get used soon, so use
2314 * ecx, as eax will get allocated first. edx is used by long calls,
2315 * so we can't use that.
2318 x86_pop_reg (code, X86_ECX);
2320 x86_alu_reg_imm (code, X86_ADD, X86_ESP, call->stack_usage);
2323 code = emit_move_return_value (cfg, ins, code);
2328 case OP_VOIDCALL_REG:
2330 call = (MonoCallInst*)ins;
2331 x86_call_reg (code, ins->sreg1);
2332 if (call->stack_usage && !CALLCONV_IS_STDCALL (call->signature)) {
2333 if (call->stack_usage == 4)
2334 x86_pop_reg (code, X86_ECX);
2336 x86_alu_reg_imm (code, X86_ADD, X86_ESP, call->stack_usage);
2338 code = emit_move_return_value (cfg, ins, code);
2340 case OP_FCALL_MEMBASE:
2341 case OP_LCALL_MEMBASE:
2342 case OP_VCALL_MEMBASE:
2343 case OP_VOIDCALL_MEMBASE:
2344 case OP_CALL_MEMBASE:
2345 call = (MonoCallInst*)ins;
2346 x86_call_membase (code, ins->sreg1, ins->inst_offset);
2347 if (call->stack_usage && !CALLCONV_IS_STDCALL (call->signature)) {
2348 if (call->stack_usage == 4)
2349 x86_pop_reg (code, X86_ECX);
2351 x86_alu_reg_imm (code, X86_ADD, X86_ESP, call->stack_usage);
2353 code = emit_move_return_value (cfg, ins, code);
2357 x86_push_reg (code, ins->sreg1);
2359 case OP_X86_PUSH_IMM:
2360 x86_push_imm (code, ins->inst_imm);
2362 case OP_X86_PUSH_MEMBASE:
2363 x86_push_membase (code, ins->inst_basereg, ins->inst_offset);
2365 case OP_X86_PUSH_OBJ:
2366 x86_alu_reg_imm (code, X86_SUB, X86_ESP, ins->inst_imm);
2367 x86_push_reg (code, X86_EDI);
2368 x86_push_reg (code, X86_ESI);
2369 x86_push_reg (code, X86_ECX);
2370 if (ins->inst_offset)
2371 x86_lea_membase (code, X86_ESI, ins->inst_basereg, ins->inst_offset);
2373 x86_mov_reg_reg (code, X86_ESI, ins->inst_basereg, 4);
2374 x86_lea_membase (code, X86_EDI, X86_ESP, 12);
2375 x86_mov_reg_imm (code, X86_ECX, (ins->inst_imm >> 2));
2377 x86_prefix (code, X86_REP_PREFIX);
2379 x86_pop_reg (code, X86_ECX);
2380 x86_pop_reg (code, X86_ESI);
2381 x86_pop_reg (code, X86_EDI);
2384 x86_lea_memindex (code, ins->dreg, ins->sreg1, ins->inst_imm, ins->sreg2, ins->backend.shift_amount);
2386 case OP_X86_LEA_MEMBASE:
2387 x86_lea_membase (code, ins->dreg, ins->sreg1, ins->inst_imm);
2390 x86_xchg_reg_reg (code, ins->sreg1, ins->sreg2, 4);
2393 /* keep alignment */
2394 x86_alu_reg_imm (code, X86_ADD, ins->sreg1, MONO_ARCH_FRAME_ALIGNMENT - 1);
2395 x86_alu_reg_imm (code, X86_AND, ins->sreg1, ~(MONO_ARCH_FRAME_ALIGNMENT - 1));
2396 code = mono_emit_stack_alloc (code, ins);
2397 x86_mov_reg_reg (code, ins->dreg, X86_ESP, 4);
2403 x86_push_reg (code, ins->sreg1);
2404 code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD,
2405 (gpointer)"mono_arch_throw_exception");
2409 x86_push_reg (code, ins->sreg1);
2410 code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD,
2411 (gpointer)"mono_arch_rethrow_exception");
2414 case OP_CALL_HANDLER:
2417 x86_alu_reg_imm (code, X86_SUB, X86_ESP, 12);
2419 mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_BB, ins->inst_target_bb);
2420 x86_call_imm (code, 0);
2422 x86_alu_reg_imm (code, X86_ADD, X86_ESP, 12);
2426 ins->inst_c0 = code - cfg->native_code;
2429 //g_print ("target: %p, next: %p, curr: %p, last: %p\n", ins->inst_target_bb, bb->next_bb, ins, bb->last_ins);
2430 //if ((ins->inst_target_bb == bb->next_bb) && ins == bb->last_ins)
2432 if (ins->flags & MONO_INST_BRLABEL) {
2433 if (ins->inst_i0->inst_c0) {
2434 x86_jump_code (code, cfg->native_code + ins->inst_i0->inst_c0);
2436 mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_LABEL, ins->inst_i0);
2437 if ((cfg->opt & MONO_OPT_BRANCH) &&
2438 x86_is_imm8 (ins->inst_i0->inst_c1 - cpos))
2439 x86_jump8 (code, 0);
2441 x86_jump32 (code, 0);
2444 if (ins->inst_target_bb->native_offset) {
2445 x86_jump_code (code, cfg->native_code + ins->inst_target_bb->native_offset);
2447 mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_BB, ins->inst_target_bb);
2448 if ((cfg->opt & MONO_OPT_BRANCH) &&
2449 x86_is_imm8 (ins->inst_target_bb->max_offset - cpos))
2450 x86_jump8 (code, 0);
2452 x86_jump32 (code, 0);
2457 x86_jump_reg (code, ins->sreg1);
2460 x86_set_reg (code, X86_CC_EQ, ins->dreg, TRUE);
2461 x86_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
2464 x86_set_reg (code, X86_CC_LT, ins->dreg, TRUE);
2465 x86_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
2468 x86_set_reg (code, X86_CC_LT, ins->dreg, FALSE);
2469 x86_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
2472 x86_set_reg (code, X86_CC_GT, ins->dreg, TRUE);
2473 x86_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
2476 x86_set_reg (code, X86_CC_GT, ins->dreg, FALSE);
2477 x86_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
2480 x86_set_reg (code, X86_CC_NE, ins->dreg, TRUE);
2481 x86_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
2483 case OP_COND_EXC_EQ:
2484 case OP_COND_EXC_NE_UN:
2485 case OP_COND_EXC_LT:
2486 case OP_COND_EXC_LT_UN:
2487 case OP_COND_EXC_GT:
2488 case OP_COND_EXC_GT_UN:
2489 case OP_COND_EXC_GE:
2490 case OP_COND_EXC_GE_UN:
2491 case OP_COND_EXC_LE:
2492 case OP_COND_EXC_LE_UN:
2493 case OP_COND_EXC_OV:
2494 case OP_COND_EXC_NO:
2496 case OP_COND_EXC_NC:
2497 EMIT_COND_SYSTEM_EXCEPTION (branch_cc_table [ins->opcode - OP_COND_EXC_EQ], (ins->opcode < OP_COND_EXC_NE_UN), ins->inst_p1);
2509 EMIT_COND_BRANCH (ins, branch_cc_table [ins->opcode - CEE_BEQ], (ins->opcode < CEE_BNE_UN));
2512 /* floating point opcodes */
2514 double d = *(double *)ins->inst_p0;
2516 if ((d == 0.0) && (mono_signbit (d) == 0)) {
2518 } else if (d == 1.0) {
2521 if (cfg->compile_aot) {
2522 guint32 *val = (guint32*)&d;
2523 x86_push_imm (code, val [1]);
2524 x86_push_imm (code, val [0]);
2525 x86_fld_membase (code, X86_ESP, 0, TRUE);
2526 x86_alu_reg_imm (code, X86_ADD, X86_ESP, 8);
2529 mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_R8, ins->inst_p0);
2530 x86_fld (code, NULL, TRUE);
2536 float f = *(float *)ins->inst_p0;
2538 if ((f == 0.0) && (mono_signbit (f) == 0)) {
2540 } else if (f == 1.0) {
2543 if (cfg->compile_aot) {
2544 guint32 val = *(guint32*)&f;
2545 x86_push_imm (code, val);
2546 x86_fld_membase (code, X86_ESP, 0, FALSE);
2547 x86_alu_reg_imm (code, X86_ADD, X86_ESP, 4);
2550 mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_R4, ins->inst_p0);
2551 x86_fld (code, NULL, FALSE);
2556 case OP_STORER8_MEMBASE_REG:
2557 x86_fst_membase (code, ins->inst_destbasereg, ins->inst_offset, TRUE, TRUE);
2559 case OP_LOADR8_SPILL_MEMBASE:
2560 x86_fld_membase (code, ins->inst_basereg, ins->inst_offset, TRUE);
2563 case OP_LOADR8_MEMBASE:
2564 x86_fld_membase (code, ins->inst_basereg, ins->inst_offset, TRUE);
2566 case OP_STORER4_MEMBASE_REG:
2567 x86_fst_membase (code, ins->inst_destbasereg, ins->inst_offset, FALSE, TRUE);
2569 case OP_LOADR4_MEMBASE:
2570 x86_fld_membase (code, ins->inst_basereg, ins->inst_offset, FALSE);
2572 case CEE_CONV_R4: /* FIXME: change precision */
2574 x86_push_reg (code, ins->sreg1);
2575 x86_fild_membase (code, X86_ESP, 0, FALSE);
2576 x86_alu_reg_imm (code, X86_ADD, X86_ESP, 4);
2578 case OP_X86_FP_LOAD_I8:
2579 x86_fild_membase (code, ins->inst_basereg, ins->inst_offset, TRUE);
2581 case OP_X86_FP_LOAD_I4:
2582 x86_fild_membase (code, ins->inst_basereg, ins->inst_offset, FALSE);
2584 case OP_FCONV_TO_I1:
2585 code = emit_float_to_int (cfg, code, ins->dreg, 1, TRUE);
2587 case OP_FCONV_TO_U1:
2588 code = emit_float_to_int (cfg, code, ins->dreg, 1, FALSE);
2590 case OP_FCONV_TO_I2:
2591 code = emit_float_to_int (cfg, code, ins->dreg, 2, TRUE);
2593 case OP_FCONV_TO_U2:
2594 code = emit_float_to_int (cfg, code, ins->dreg, 2, FALSE);
2596 case OP_FCONV_TO_I4:
2598 code = emit_float_to_int (cfg, code, ins->dreg, 4, TRUE);
2600 case OP_FCONV_TO_I8:
2601 x86_alu_reg_imm (code, X86_SUB, X86_ESP, 4);
2602 x86_fnstcw_membase(code, X86_ESP, 0);
2603 x86_mov_reg_membase (code, ins->dreg, X86_ESP, 0, 2);
2604 x86_alu_reg_imm (code, X86_OR, ins->dreg, 0xc00);
2605 x86_mov_membase_reg (code, X86_ESP, 2, ins->dreg, 2);
2606 x86_fldcw_membase (code, X86_ESP, 2);
2607 x86_alu_reg_imm (code, X86_SUB, X86_ESP, 8);
2608 x86_fist_pop_membase (code, X86_ESP, 0, TRUE);
2609 x86_pop_reg (code, ins->dreg);
2610 x86_pop_reg (code, ins->backend.reg3);
2611 x86_fldcw_membase (code, X86_ESP, 0);
2612 x86_alu_reg_imm (code, X86_ADD, X86_ESP, 4);
2614 case OP_LCONV_TO_R_UN: {
2615 static guint8 mn[] = { 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x80, 0x3f, 0x40 };
2618 /* load 64bit integer to FP stack */
2619 x86_push_imm (code, 0);
2620 x86_push_reg (code, ins->sreg2);
2621 x86_push_reg (code, ins->sreg1);
2622 x86_fild_membase (code, X86_ESP, 0, TRUE);
2623 /* store as 80bit FP value */
2624 x86_fst80_membase (code, X86_ESP, 0);
2626 /* test if lreg is negative */
2627 x86_test_reg_reg (code, ins->sreg2, ins->sreg2);
2628 br = code; x86_branch8 (code, X86_CC_GEZ, 0, TRUE);
2630 /* add correction constant mn */
2631 x86_fld80_mem (code, mn);
2632 x86_fld80_membase (code, X86_ESP, 0);
2633 x86_fp_op_reg (code, X86_FADD, 1, TRUE);
2634 x86_fst80_membase (code, X86_ESP, 0);
2636 x86_patch (br, code);
2638 x86_fld80_membase (code, X86_ESP, 0);
2639 x86_alu_reg_imm (code, X86_ADD, X86_ESP, 12);
2643 case OP_LCONV_TO_OVF_I: {
2644 guint8 *br [3], *label [1];
2648 * Valid ints: 0xffffffff:8000000 to 00000000:0x7f000000
2650 x86_test_reg_reg (code, ins->sreg1, ins->sreg1);
2652 /* If the low word top bit is set, see if we are negative */
2653 br [0] = code; x86_branch8 (code, X86_CC_LT, 0, TRUE);
2654 /* We are not negative (no top bit set, check for our top word to be zero */
2655 x86_test_reg_reg (code, ins->sreg2, ins->sreg2);
2656 br [1] = code; x86_branch8 (code, X86_CC_EQ, 0, TRUE);
2659 /* throw exception */
2660 tins = mono_branch_optimize_exception_target (cfg, bb, "OverflowException");
2662 mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_BB, tins->inst_true_bb);
2663 if ((cfg->opt & MONO_OPT_BRANCH) && x86_is_imm8 (tins->inst_true_bb->max_offset - cpos))
2664 x86_jump8 (code, 0);
2666 x86_jump32 (code, 0);
2668 mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_EXC, "OverflowException");
2669 x86_jump32 (code, 0);
2673 x86_patch (br [0], code);
2674 /* our top bit is set, check that top word is 0xfffffff */
2675 x86_alu_reg_imm (code, X86_CMP, ins->sreg2, 0xffffffff);
2677 x86_patch (br [1], code);
2678 /* nope, emit exception */
2679 br [2] = code; x86_branch8 (code, X86_CC_NE, 0, TRUE);
2680 x86_patch (br [2], label [0]);
2682 if (ins->dreg != ins->sreg1)
2683 x86_mov_reg_reg (code, ins->dreg, ins->sreg1, 4);
2687 x86_fp_op_reg (code, X86_FADD, 1, TRUE);
2690 x86_fp_op_reg (code, X86_FSUB, 1, TRUE);
2693 x86_fp_op_reg (code, X86_FMUL, 1, TRUE);
2696 x86_fp_op_reg (code, X86_FDIV, 1, TRUE);
2704 x86_fp_op_reg (code, X86_FADD, 1, TRUE);
2709 x86_fp_op_reg (code, X86_FADD, 1, TRUE);
2716 * it really doesn't make sense to inline all this code,
2717 * it's here just to show that things may not be as simple
2720 guchar *check_pos, *end_tan, *pop_jump;
2721 x86_push_reg (code, X86_EAX);
2724 x86_test_reg_imm (code, X86_EAX, X86_FP_C2);
2726 x86_branch8 (code, X86_CC_NE, 0, FALSE);
2727 x86_fstp (code, 0); /* pop the 1.0 */
2729 x86_jump8 (code, 0);
2731 x86_fp_op (code, X86_FADD, 0);
2735 x86_test_reg_imm (code, X86_EAX, X86_FP_C2);
2737 x86_branch8 (code, X86_CC_NE, 0, FALSE);
2740 x86_patch (pop_jump, code);
2741 x86_fstp (code, 0); /* pop the 1.0 */
2742 x86_patch (check_pos, code);
2743 x86_patch (end_tan, code);
2745 x86_fp_op_reg (code, X86_FADD, 1, TRUE);
2746 x86_pop_reg (code, X86_EAX);
2753 x86_fp_op_reg (code, X86_FADD, 1, TRUE);
2764 x86_push_reg (code, X86_EAX);
2765 /* we need to exchange ST(0) with ST(1) */
2768 /* this requires a loop, because fprem somtimes
2769 * returns a partial remainder */
2771 /* looks like MS is using fprem instead of the IEEE compatible fprem1 */
2772 /* x86_fprem1 (code); */
2775 x86_alu_reg_imm (code, X86_AND, X86_EAX, X86_FP_C2);
2777 x86_branch8 (code, X86_CC_NE, l1 - l2, FALSE);
2782 x86_pop_reg (code, X86_EAX);
2786 if (cfg->opt & MONO_OPT_FCMOV) {
2787 x86_fcomip (code, 1);
2791 /* this overwrites EAX */
2792 EMIT_FPCOMPARE(code);
2793 x86_alu_reg_imm (code, X86_AND, X86_EAX, X86_FP_CC_MASK);
2796 if (cfg->opt & MONO_OPT_FCMOV) {
2797 /* zeroing the register at the start results in
2798 * shorter and faster code (we can also remove the widening op)
2800 guchar *unordered_check;
2801 x86_alu_reg_reg (code, X86_XOR, ins->dreg, ins->dreg);
2802 x86_fcomip (code, 1);
2804 unordered_check = code;
2805 x86_branch8 (code, X86_CC_P, 0, FALSE);
2806 x86_set_reg (code, X86_CC_EQ, ins->dreg, FALSE);
2807 x86_patch (unordered_check, code);
2810 if (ins->dreg != X86_EAX)
2811 x86_push_reg (code, X86_EAX);
2813 EMIT_FPCOMPARE(code);
2814 x86_alu_reg_imm (code, X86_AND, X86_EAX, X86_FP_CC_MASK);
2815 x86_alu_reg_imm (code, X86_CMP, X86_EAX, 0x4000);
2816 x86_set_reg (code, X86_CC_EQ, ins->dreg, TRUE);
2817 x86_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
2819 if (ins->dreg != X86_EAX)
2820 x86_pop_reg (code, X86_EAX);
2824 if (cfg->opt & MONO_OPT_FCMOV) {
2825 /* zeroing the register at the start results in
2826 * shorter and faster code (we can also remove the widening op)
2828 x86_alu_reg_reg (code, X86_XOR, ins->dreg, ins->dreg);
2829 x86_fcomip (code, 1);
2831 if (ins->opcode == OP_FCLT_UN) {
2832 guchar *unordered_check = code;
2833 guchar *jump_to_end;
2834 x86_branch8 (code, X86_CC_P, 0, FALSE);
2835 x86_set_reg (code, X86_CC_GT, ins->dreg, FALSE);
2837 x86_jump8 (code, 0);
2838 x86_patch (unordered_check, code);
2839 x86_inc_reg (code, ins->dreg);
2840 x86_patch (jump_to_end, code);
2842 x86_set_reg (code, X86_CC_GT, ins->dreg, FALSE);
2846 if (ins->dreg != X86_EAX)
2847 x86_push_reg (code, X86_EAX);
2849 EMIT_FPCOMPARE(code);
2850 x86_alu_reg_imm (code, X86_AND, X86_EAX, X86_FP_CC_MASK);
2851 if (ins->opcode == OP_FCLT_UN) {
2852 guchar *is_not_zero_check, *end_jump;
2853 is_not_zero_check = code;
2854 x86_branch8 (code, X86_CC_NZ, 0, TRUE);
2856 x86_jump8 (code, 0);
2857 x86_patch (is_not_zero_check, code);
2858 x86_alu_reg_imm (code, X86_CMP, X86_EAX, X86_FP_CC_MASK);
2860 x86_patch (end_jump, code);
2862 x86_set_reg (code, X86_CC_EQ, ins->dreg, TRUE);
2863 x86_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
2865 if (ins->dreg != X86_EAX)
2866 x86_pop_reg (code, X86_EAX);
2870 if (cfg->opt & MONO_OPT_FCMOV) {
2871 /* zeroing the register at the start results in
2872 * shorter and faster code (we can also remove the widening op)
2874 guchar *unordered_check;
2875 x86_alu_reg_reg (code, X86_XOR, ins->dreg, ins->dreg);
2876 x86_fcomip (code, 1);
2878 if (ins->opcode == OP_FCGT) {
2879 unordered_check = code;
2880 x86_branch8 (code, X86_CC_P, 0, FALSE);
2881 x86_set_reg (code, X86_CC_LT, ins->dreg, FALSE);
2882 x86_patch (unordered_check, code);
2884 x86_set_reg (code, X86_CC_LT, ins->dreg, FALSE);
2888 if (ins->dreg != X86_EAX)
2889 x86_push_reg (code, X86_EAX);
2891 EMIT_FPCOMPARE(code);
2892 x86_alu_reg_imm (code, X86_AND, X86_EAX, X86_FP_CC_MASK);
2893 x86_alu_reg_imm (code, X86_CMP, X86_EAX, X86_FP_C0);
2894 if (ins->opcode == OP_FCGT_UN) {
2895 guchar *is_not_zero_check, *end_jump;
2896 is_not_zero_check = code;
2897 x86_branch8 (code, X86_CC_NZ, 0, TRUE);
2899 x86_jump8 (code, 0);
2900 x86_patch (is_not_zero_check, code);
2901 x86_alu_reg_imm (code, X86_CMP, X86_EAX, X86_FP_CC_MASK);
2903 x86_patch (end_jump, code);
2905 x86_set_reg (code, X86_CC_EQ, ins->dreg, TRUE);
2906 x86_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
2908 if (ins->dreg != X86_EAX)
2909 x86_pop_reg (code, X86_EAX);
2912 if (cfg->opt & MONO_OPT_FCMOV) {
2913 guchar *jump = code;
2914 x86_branch8 (code, X86_CC_P, 0, TRUE);
2915 EMIT_COND_BRANCH (ins, X86_CC_EQ, FALSE);
2916 x86_patch (jump, code);
2919 x86_alu_reg_imm (code, X86_CMP, X86_EAX, 0x4000);
2920 EMIT_COND_BRANCH (ins, X86_CC_EQ, TRUE);
2923 /* Branch if C013 != 100 */
2924 if (cfg->opt & MONO_OPT_FCMOV) {
2925 /* branch if !ZF or (PF|CF) */
2926 EMIT_COND_BRANCH (ins, X86_CC_NE, FALSE);
2927 EMIT_COND_BRANCH (ins, X86_CC_P, FALSE);
2928 EMIT_COND_BRANCH (ins, X86_CC_B, FALSE);
2931 x86_alu_reg_imm (code, X86_CMP, X86_EAX, X86_FP_C3);
2932 EMIT_COND_BRANCH (ins, X86_CC_NE, FALSE);
2935 if (cfg->opt & MONO_OPT_FCMOV) {
2936 EMIT_COND_BRANCH (ins, X86_CC_GT, FALSE);
2939 EMIT_COND_BRANCH (ins, X86_CC_EQ, FALSE);
2942 if (cfg->opt & MONO_OPT_FCMOV) {
2943 EMIT_COND_BRANCH (ins, X86_CC_P, FALSE);
2944 EMIT_COND_BRANCH (ins, X86_CC_GT, FALSE);
2947 if (ins->opcode == OP_FBLT_UN) {
2948 guchar *is_not_zero_check, *end_jump;
2949 is_not_zero_check = code;
2950 x86_branch8 (code, X86_CC_NZ, 0, TRUE);
2952 x86_jump8 (code, 0);
2953 x86_patch (is_not_zero_check, code);
2954 x86_alu_reg_imm (code, X86_CMP, X86_EAX, X86_FP_CC_MASK);
2956 x86_patch (end_jump, code);
2958 EMIT_COND_BRANCH (ins, X86_CC_EQ, FALSE);
2962 if (cfg->opt & MONO_OPT_FCMOV) {
2963 EMIT_COND_BRANCH (ins, X86_CC_LT, FALSE);
2966 x86_alu_reg_imm (code, X86_CMP, X86_EAX, X86_FP_C0);
2967 if (ins->opcode == OP_FBGT_UN) {
2968 guchar *is_not_zero_check, *end_jump;
2969 is_not_zero_check = code;
2970 x86_branch8 (code, X86_CC_NZ, 0, TRUE);
2972 x86_jump8 (code, 0);
2973 x86_patch (is_not_zero_check, code);
2974 x86_alu_reg_imm (code, X86_CMP, X86_EAX, X86_FP_CC_MASK);
2976 x86_patch (end_jump, code);
2978 EMIT_COND_BRANCH (ins, X86_CC_EQ, FALSE);
2981 /* Branch if C013 == 100 or 001 */
2982 if (cfg->opt & MONO_OPT_FCMOV) {
2985 /* skip branch if C1=1 */
2987 x86_branch8 (code, X86_CC_P, 0, FALSE);
2988 /* branch if (C0 | C3) = 1 */
2989 EMIT_COND_BRANCH (ins, X86_CC_BE, FALSE);
2990 x86_patch (br1, code);
2993 x86_alu_reg_imm (code, X86_CMP, X86_EAX, X86_FP_C0);
2994 EMIT_COND_BRANCH (ins, X86_CC_EQ, FALSE);
2995 x86_alu_reg_imm (code, X86_CMP, X86_EAX, X86_FP_C3);
2996 EMIT_COND_BRANCH (ins, X86_CC_EQ, FALSE);
2999 /* Branch if C013 == 000 */
3000 if (cfg->opt & MONO_OPT_FCMOV) {
3001 EMIT_COND_BRANCH (ins, X86_CC_LE, FALSE);
3004 EMIT_COND_BRANCH (ins, X86_CC_NE, FALSE);
3007 /* Branch if C013=000 or 100 */
3008 if (cfg->opt & MONO_OPT_FCMOV) {
3011 /* skip branch if C1=1 */
3013 x86_branch8 (code, X86_CC_P, 0, FALSE);
3014 /* branch if C0=0 */
3015 EMIT_COND_BRANCH (ins, X86_CC_NB, FALSE);
3016 x86_patch (br1, code);
3019 x86_alu_reg_imm (code, X86_AND, X86_EAX, (X86_FP_C0|X86_FP_C1));
3020 x86_alu_reg_imm (code, X86_CMP, X86_EAX, 0);
3021 EMIT_COND_BRANCH (ins, X86_CC_EQ, FALSE);
3024 /* Branch if C013 != 001 */
3025 if (cfg->opt & MONO_OPT_FCMOV) {
3026 EMIT_COND_BRANCH (ins, X86_CC_P, FALSE);
3027 EMIT_COND_BRANCH (ins, X86_CC_GE, FALSE);
3030 x86_alu_reg_imm (code, X86_CMP, X86_EAX, X86_FP_C0);
3031 EMIT_COND_BRANCH (ins, X86_CC_NE, FALSE);
3033 case CEE_CKFINITE: {
3034 x86_push_reg (code, X86_EAX);
3037 x86_alu_reg_imm (code, X86_AND, X86_EAX, 0x4100);
3038 x86_alu_reg_imm (code, X86_CMP, X86_EAX, X86_FP_C0);
3039 x86_pop_reg (code, X86_EAX);
3040 EMIT_COND_SYSTEM_EXCEPTION (X86_CC_EQ, FALSE, "ArithmeticException");
3044 code = emit_tls_get (code, ins->dreg, ins->inst_offset);
3047 case OP_MEMORY_BARRIER: {
3048 /* Not needed on x86 */
3051 case OP_ATOMIC_ADD_I4: {
3052 int dreg = ins->dreg;
3054 if (dreg == ins->inst_basereg) {
3055 x86_push_reg (code, ins->sreg2);
3059 if (dreg != ins->sreg2)
3060 x86_mov_reg_reg (code, ins->dreg, ins->sreg2, 4);
3062 x86_prefix (code, X86_LOCK_PREFIX);
3063 x86_xadd_membase_reg (code, ins->inst_basereg, ins->inst_offset, dreg, 4);
3065 if (dreg != ins->dreg) {
3066 x86_mov_reg_reg (code, ins->dreg, dreg, 4);
3067 x86_pop_reg (code, dreg);
3072 case OP_ATOMIC_ADD_NEW_I4: {
3073 int dreg = ins->dreg;
3075 /* hack: limit in regalloc, dreg != sreg1 && dreg != sreg2 */
3076 if (ins->sreg2 == dreg) {
3077 if (dreg == X86_EBX) {
3079 if (ins->inst_basereg == X86_EDI)
3083 if (ins->inst_basereg == X86_EBX)
3086 } else if (ins->inst_basereg == dreg) {
3087 if (dreg == X86_EBX) {
3089 if (ins->sreg2 == X86_EDI)
3093 if (ins->sreg2 == X86_EBX)
3098 if (dreg != ins->dreg) {
3099 x86_push_reg (code, dreg);
3102 x86_mov_reg_reg (code, dreg, ins->sreg2, 4);
3103 x86_prefix (code, X86_LOCK_PREFIX);
3104 x86_xadd_membase_reg (code, ins->inst_basereg, ins->inst_offset, dreg, 4);
3105 /* dreg contains the old value, add with sreg2 value */
3106 x86_alu_reg_reg (code, X86_ADD, dreg, ins->sreg2);
3108 if (ins->dreg != dreg) {
3109 x86_mov_reg_reg (code, ins->dreg, dreg, 4);
3110 x86_pop_reg (code, dreg);
3115 case OP_ATOMIC_EXCHANGE_I4: {
3117 int sreg2 = ins->sreg2;
3118 int breg = ins->inst_basereg;
3120 /* cmpxchg uses eax as comperand, need to make sure we can use it
3121 * hack to overcome limits in x86 reg allocator
3122 * (req: dreg == eax and sreg2 != eax and breg != eax)
3124 if (ins->dreg != X86_EAX)
3125 x86_push_reg (code, X86_EAX);
3127 /* We need the EAX reg for the cmpxchg */
3128 if (ins->sreg2 == X86_EAX) {
3129 x86_push_reg (code, X86_EDX);
3130 x86_mov_reg_reg (code, X86_EDX, X86_EAX, 4);
3134 if (breg == X86_EAX) {
3135 x86_push_reg (code, X86_ESI);
3136 x86_mov_reg_reg (code, X86_ESI, X86_EAX, 4);
3140 x86_mov_reg_membase (code, X86_EAX, breg, ins->inst_offset, 4);
3142 br [0] = code; x86_prefix (code, X86_LOCK_PREFIX);
3143 x86_cmpxchg_membase_reg (code, breg, ins->inst_offset, sreg2);
3144 br [1] = code; x86_branch8 (code, X86_CC_NE, -1, FALSE);
3145 x86_patch (br [1], br [0]);
3147 if (breg != ins->inst_basereg)
3148 x86_pop_reg (code, X86_ESI);
3150 if (ins->dreg != X86_EAX) {
3151 x86_mov_reg_reg (code, ins->dreg, X86_EAX, 4);
3152 x86_pop_reg (code, X86_EAX);
3155 if (ins->sreg2 != sreg2)
3156 x86_pop_reg (code, X86_EDX);
3161 g_warning ("unknown opcode %s in %s()\n", mono_inst_name (ins->opcode), __FUNCTION__);
3162 g_assert_not_reached ();
3165 if ((code - cfg->native_code - offset) > max_len) {
3166 g_warning ("wrong maximal instruction length of instruction %s (expected %d, got %d)",
3167 mono_inst_name (ins->opcode), max_len, code - cfg->native_code - offset);
3168 g_assert_not_reached ();
3174 last_offset = offset;
3179 cfg->code_len = code - cfg->native_code;
3183 mono_arch_register_lowlevel_calls (void)
3188 mono_arch_patch_code (MonoMethod *method, MonoDomain *domain, guint8 *code, MonoJumpInfo *ji, gboolean run_cctors)
3190 MonoJumpInfo *patch_info;
3191 gboolean compile_aot = !run_cctors;
3193 for (patch_info = ji; patch_info; patch_info = patch_info->next) {
3194 unsigned char *ip = patch_info->ip.i + code;
3195 const unsigned char *target;
3197 target = mono_resolve_patch_target (method, domain, code, patch_info, run_cctors);
3200 switch (patch_info->type) {
3201 case MONO_PATCH_INFO_BB:
3202 case MONO_PATCH_INFO_LABEL:
3205 /* No need to patch these */
3210 switch (patch_info->type) {
3211 case MONO_PATCH_INFO_IP:
3212 *((gconstpointer *)(ip)) = target;
3214 case MONO_PATCH_INFO_CLASS_INIT: {
3216 /* Might already been changed to a nop */
3217 x86_call_code (code, 0);
3218 x86_patch (ip, target);
3221 case MONO_PATCH_INFO_ABS:
3222 case MONO_PATCH_INFO_METHOD:
3223 case MONO_PATCH_INFO_METHOD_JUMP:
3224 case MONO_PATCH_INFO_INTERNAL_METHOD:
3225 case MONO_PATCH_INFO_BB:
3226 case MONO_PATCH_INFO_LABEL:
3227 x86_patch (ip, target);
3229 case MONO_PATCH_INFO_NONE:
3232 guint32 offset = mono_arch_get_patch_offset (ip);
3233 *((gconstpointer *)(ip + offset)) = target;
3241 mono_arch_emit_prolog (MonoCompile *cfg)
3243 MonoMethod *method = cfg->method;
3245 MonoMethodSignature *sig;
3247 int alloc_size, pos, max_offset, i;
3250 cfg->code_size = MAX (mono_method_get_header (method)->code_size * 4, 256);
3251 code = cfg->native_code = g_malloc (cfg->code_size);
3253 x86_push_reg (code, X86_EBP);
3254 x86_mov_reg_reg (code, X86_EBP, X86_ESP, 4);
3256 alloc_size = cfg->stack_offset;
3259 if (method->wrapper_type == MONO_WRAPPER_NATIVE_TO_MANAGED) {
3260 /* Might need to attach the thread to the JIT */
3261 if (lmf_tls_offset != -1) {
3264 code = emit_tls_get ( code, X86_EAX, lmf_tls_offset);
3265 x86_test_reg_reg (code, X86_EAX, X86_EAX);
3267 x86_branch8 (code, X86_CC_NE, 0, 0);
3268 x86_push_imm (code, cfg->domain);
3269 code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD, (gpointer)"mono_jit_thread_attach");
3270 x86_alu_reg_imm (code, X86_ADD, X86_ESP, 4);
3271 x86_patch (buf, code);
3272 #ifdef PLATFORM_WIN32
3273 /* The TLS key actually contains a pointer to the MonoJitTlsData structure */
3274 /* FIXME: Add a separate key for LMF to avoid this */
3275 x86_alu_reg_imm (code, X86_ADD, X86_EAX, G_STRUCT_OFFSET (MonoJitTlsData, lmf));
3278 g_assert (!cfg->compile_aot);
3279 x86_push_imm (code, cfg->domain);
3280 code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD, (gpointer)"mono_jit_thread_attach");
3281 x86_alu_reg_imm (code, X86_ADD, X86_ESP, 4);
3285 if (method->save_lmf) {
3286 pos += sizeof (MonoLMF);
3288 /* save the current IP */
3289 mono_add_patch_info (cfg, code + 1 - cfg->native_code, MONO_PATCH_INFO_IP, NULL);
3290 x86_push_imm_template (code);
3292 /* save all caller saved regs */
3293 x86_push_reg (code, X86_EBP);
3294 x86_push_reg (code, X86_ESI);
3295 x86_push_reg (code, X86_EDI);
3296 x86_push_reg (code, X86_EBX);
3298 /* save method info */
3299 x86_push_imm (code, method);
3301 /* get the address of lmf for the current thread */
3303 * This is performance critical so we try to use some tricks to make
3306 if (lmf_tls_offset != -1) {
3307 /* Load lmf quicky using the GS register */
3308 code = emit_tls_get (code, X86_EAX, lmf_tls_offset);
3309 #ifdef PLATFORM_WIN32
3310 /* The TLS key actually contains a pointer to the MonoJitTlsData structure */
3311 /* FIXME: Add a separate key for LMF to avoid this */
3312 x86_alu_reg_imm (code, X86_ADD, X86_EAX, G_STRUCT_OFFSET (MonoJitTlsData, lmf));
3315 code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD, (gpointer)"mono_get_lmf_addr");
3319 x86_push_reg (code, X86_EAX);
3320 /* push *lfm (previous_lmf) */
3321 x86_push_membase (code, X86_EAX, 0);
3323 x86_mov_membase_reg (code, X86_EAX, 0, X86_ESP, 4);
3326 if (cfg->used_int_regs & (1 << X86_EBX)) {
3327 x86_push_reg (code, X86_EBX);
3331 if (cfg->used_int_regs & (1 << X86_EDI)) {
3332 x86_push_reg (code, X86_EDI);
3336 if (cfg->used_int_regs & (1 << X86_ESI)) {
3337 x86_push_reg (code, X86_ESI);
3345 /* the original alloc_size is already aligned: there is %ebp and retip pushed, so realign */
3347 int tot = alloc_size + pos + 4 + 4; /* ret ip + ebp */
3359 /* See mono_emit_stack_alloc */
3360 #if defined(PLATFORM_WIN32) || defined(MONO_ARCH_SIGSEGV_ON_ALTSTACK)
3361 guint32 remaining_size = alloc_size;
3362 while (remaining_size >= 0x1000) {
3363 x86_alu_reg_imm (code, X86_SUB, X86_ESP, 0x1000);
3364 x86_test_membase_reg (code, X86_ESP, 0, X86_ESP);
3365 remaining_size -= 0x1000;
3368 x86_alu_reg_imm (code, X86_SUB, X86_ESP, remaining_size);
3370 x86_alu_reg_imm (code, X86_SUB, X86_ESP, alloc_size);
3375 /* check the stack is aligned */
3376 x86_mov_reg_reg (code, X86_EDX, X86_ESP, 4);
3377 x86_alu_reg_imm (code, X86_AND, X86_EDX, 15);
3378 x86_alu_reg_imm (code, X86_CMP, X86_EDX, 0);
3379 x86_branch_disp (code, X86_CC_EQ, 3, FALSE);
3380 x86_breakpoint (code);
3383 /* compute max_offset in order to use short forward jumps */
3385 if (cfg->opt & MONO_OPT_BRANCH) {
3386 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
3387 MonoInst *ins = bb->code;
3388 bb->max_offset = max_offset;
3390 if (cfg->prof_options & MONO_PROFILE_COVERAGE)
3392 /* max alignment for loops */
3393 if ((cfg->opt & MONO_OPT_LOOP) && bb_is_loop_start (bb))
3394 max_offset += LOOP_ALIGNMENT;
3397 if (ins->opcode == OP_LABEL)
3398 ins->inst_c1 = max_offset;
3400 max_offset += ((guint8 *)ins_spec [ins->opcode])[MONO_INST_LEN];
3406 if (mono_jit_trace_calls != NULL && mono_trace_eval (method))
3407 code = mono_arch_instrument_prolog (cfg, mono_trace_enter_method, code, TRUE);
3409 /* load arguments allocated to register from the stack */
3410 sig = mono_method_signature (method);
3413 for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
3414 inst = cfg->varinfo [pos];
3415 if (inst->opcode == OP_REGVAR) {
3416 x86_mov_reg_membase (code, inst->dreg, X86_EBP, inst->inst_offset, 4);
3417 if (cfg->verbose_level > 2)
3418 g_print ("Argument %d assigned to register %s\n", pos, mono_arch_regname (inst->dreg));
3423 cfg->code_len = code - cfg->native_code;
3429 mono_arch_emit_epilog (MonoCompile *cfg)
3431 MonoMethod *method = cfg->method;
3432 MonoMethodSignature *sig = mono_method_signature (method);
3434 guint32 stack_to_pop;
3436 int max_epilog_size = 16;
3439 if (cfg->method->save_lmf)
3440 max_epilog_size += 128;
3442 if (mono_jit_trace_calls != NULL)
3443 max_epilog_size += 50;
3445 while (cfg->code_len + max_epilog_size > (cfg->code_size - 16)) {
3446 cfg->code_size *= 2;
3447 cfg->native_code = g_realloc (cfg->native_code, cfg->code_size);
3448 mono_jit_stats.code_reallocs++;
3451 code = cfg->native_code + cfg->code_len;
3453 if (mono_jit_trace_calls != NULL && mono_trace_eval (method))
3454 code = mono_arch_instrument_epilog (cfg, mono_trace_leave_method, code, TRUE);
3456 /* the code restoring the registers must be kept in sync with CEE_JMP */
3459 if (method->save_lmf) {
3460 gint32 prev_lmf_reg;
3461 gint32 lmf_offset = -sizeof (MonoLMF);
3463 /* Find a spare register */
3464 switch (sig->ret->type) {
3467 prev_lmf_reg = X86_EDI;
3468 cfg->used_int_regs |= (1 << X86_EDI);
3471 prev_lmf_reg = X86_EDX;
3475 /* reg = previous_lmf */
3476 x86_mov_reg_membase (code, prev_lmf_reg, X86_EBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, previous_lmf), 4);
3479 x86_mov_reg_membase (code, X86_ECX, X86_EBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, lmf_addr), 4);
3481 /* *(lmf) = previous_lmf */
3482 x86_mov_membase_reg (code, X86_ECX, 0, prev_lmf_reg, 4);
3484 /* restore caller saved regs */
3485 if (cfg->used_int_regs & (1 << X86_EBX)) {
3486 x86_mov_reg_membase (code, X86_EBX, X86_EBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, ebx), 4);
3489 if (cfg->used_int_regs & (1 << X86_EDI)) {
3490 x86_mov_reg_membase (code, X86_EDI, X86_EBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, edi), 4);
3492 if (cfg->used_int_regs & (1 << X86_ESI)) {
3493 x86_mov_reg_membase (code, X86_ESI, X86_EBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, esi), 4);
3496 /* EBP is restored by LEAVE */
3498 if (cfg->used_int_regs & (1 << X86_EBX)) {
3501 if (cfg->used_int_regs & (1 << X86_EDI)) {
3504 if (cfg->used_int_regs & (1 << X86_ESI)) {
3509 x86_lea_membase (code, X86_ESP, X86_EBP, pos);
3511 if (cfg->used_int_regs & (1 << X86_ESI)) {
3512 x86_pop_reg (code, X86_ESI);
3514 if (cfg->used_int_regs & (1 << X86_EDI)) {
3515 x86_pop_reg (code, X86_EDI);
3517 if (cfg->used_int_regs & (1 << X86_EBX)) {
3518 x86_pop_reg (code, X86_EBX);
3522 /* Load returned vtypes into registers if needed */
3523 cinfo = get_call_info (sig, FALSE);
3524 if (cinfo->ret.storage == ArgValuetypeInReg) {
3525 for (quad = 0; quad < 2; quad ++) {
3526 switch (cinfo->ret.pair_storage [quad]) {
3528 x86_mov_reg_membase (code, cinfo->ret.pair_regs [quad], cfg->ret->inst_basereg, cfg->ret->inst_offset + (quad * sizeof (gpointer)), 4);
3530 case ArgOnFloatFpStack:
3531 x86_fld_membase (code, cfg->ret->inst_basereg, cfg->ret->inst_offset + (quad * sizeof (gpointer)), FALSE);
3533 case ArgOnDoubleFpStack:
3534 x86_fld_membase (code, cfg->ret->inst_basereg, cfg->ret->inst_offset + (quad * sizeof (gpointer)), TRUE);
3539 g_assert_not_reached ();
3546 if (CALLCONV_IS_STDCALL (sig)) {
3547 MonoJitArgumentInfo *arg_info = alloca (sizeof (MonoJitArgumentInfo) * (sig->param_count + 1));
3549 stack_to_pop = mono_arch_get_argument_info (sig, sig->param_count, arg_info);
3550 } else if (MONO_TYPE_ISSTRUCT (mono_method_signature (cfg->method)->ret) && (cinfo->ret.storage == ArgOnStack))
3556 x86_ret_imm (code, stack_to_pop);
3562 cfg->code_len = code - cfg->native_code;
3564 g_assert (cfg->code_len < cfg->code_size);
3568 mono_arch_emit_exceptions (MonoCompile *cfg)
3570 MonoJumpInfo *patch_info;
3573 MonoClass *exc_classes [16];
3574 guint8 *exc_throw_start [16], *exc_throw_end [16];
3578 /* Compute needed space */
3579 for (patch_info = cfg->patch_info; patch_info; patch_info = patch_info->next) {
3580 if (patch_info->type == MONO_PATCH_INFO_EXC)
3585 * make sure we have enough space for exceptions
3586 * 16 is the size of two push_imm instructions and a call
3588 if (cfg->compile_aot)
3589 code_size = exc_count * 32;
3591 code_size = exc_count * 16;
3593 while (cfg->code_len + code_size > (cfg->code_size - 16)) {
3594 cfg->code_size *= 2;
3595 cfg->native_code = g_realloc (cfg->native_code, cfg->code_size);
3596 mono_jit_stats.code_reallocs++;
3599 code = cfg->native_code + cfg->code_len;
3602 for (patch_info = cfg->patch_info; patch_info; patch_info = patch_info->next) {
3603 switch (patch_info->type) {
3604 case MONO_PATCH_INFO_EXC: {
3605 MonoClass *exc_class;
3609 x86_patch (patch_info->ip.i + cfg->native_code, code);
3611 exc_class = mono_class_from_name (mono_defaults.corlib, "System", patch_info->data.name);
3612 g_assert (exc_class);
3613 throw_ip = patch_info->ip.i;
3615 /* Find a throw sequence for the same exception class */
3616 for (i = 0; i < nthrows; ++i)
3617 if (exc_classes [i] == exc_class)
3620 x86_push_imm (code, (exc_throw_end [i] - cfg->native_code) - throw_ip);
3621 x86_jump_code (code, exc_throw_start [i]);
3622 patch_info->type = MONO_PATCH_INFO_NONE;
3627 /* Compute size of code following the push <OFFSET> */
3630 if ((code - cfg->native_code) - throw_ip < 126 - size) {
3631 /* Use the shorter form */
3633 x86_push_imm (code, 0);
3637 x86_push_imm (code, 0xf0f0f0f0);
3642 exc_classes [nthrows] = exc_class;
3643 exc_throw_start [nthrows] = code;
3646 x86_push_imm (code, exc_class->type_token);
3647 patch_info->data.name = "mono_arch_throw_corlib_exception";
3648 patch_info->type = MONO_PATCH_INFO_INTERNAL_METHOD;
3649 patch_info->ip.i = code - cfg->native_code;
3650 x86_call_code (code, 0);
3651 x86_push_imm (buf, (code - cfg->native_code) - throw_ip);
3656 exc_throw_end [nthrows] = code;
3668 cfg->code_len = code - cfg->native_code;
3670 g_assert (cfg->code_len < cfg->code_size);
3674 mono_arch_flush_icache (guint8 *code, gint size)
3680 mono_arch_flush_register_windows (void)
3685 * Support for fast access to the thread-local lmf structure using the GS
3686 * segment register on NPTL + kernel 2.6.x.
3689 static gboolean tls_offset_inited = FALSE;
3692 mono_arch_setup_jit_tls_data (MonoJitTlsData *tls)
3694 if (!tls_offset_inited) {
3695 if (!getenv ("MONO_NO_TLS")) {
3696 #ifdef PLATFORM_WIN32
3698 * We need to init this multiple times, since when we are first called, the key might not
3699 * be initialized yet.
3701 appdomain_tls_offset = mono_domain_get_tls_key ();
3702 lmf_tls_offset = mono_get_jit_tls_key ();
3703 thread_tls_offset = mono_thread_get_tls_key ();
3705 /* Only 64 tls entries can be accessed using inline code */
3706 if (appdomain_tls_offset >= 64)
3707 appdomain_tls_offset = -1;
3708 if (lmf_tls_offset >= 64)
3709 lmf_tls_offset = -1;
3710 if (thread_tls_offset >= 64)
3711 thread_tls_offset = -1;
3714 optimize_for_xen = access ("/proc/xen", F_OK) == 0;
3716 tls_offset_inited = TRUE;
3717 appdomain_tls_offset = mono_domain_get_tls_offset ();
3718 lmf_tls_offset = mono_get_lmf_tls_offset ();
3719 thread_tls_offset = mono_thread_get_tls_offset ();
3726 mono_arch_free_jit_tls_data (MonoJitTlsData *tls)
3731 mono_arch_emit_this_vret_args (MonoCompile *cfg, MonoCallInst *inst, int this_reg, int this_type, int vt_reg)
3733 MonoCallInst *call = (MonoCallInst*)inst;
3734 CallInfo *cinfo = get_call_info (inst->signature, FALSE);
3736 /* add the this argument */
3737 if (this_reg != -1) {
3738 if (cinfo->args [0].storage == ArgInIReg) {
3740 MONO_INST_NEW (cfg, this, OP_MOVE);
3741 this->type = this_type;
3742 this->sreg1 = this_reg;
3743 this->dreg = mono_regstate_next_int (cfg->rs);
3744 mono_bblock_add_inst (cfg->cbb, this);
3746 mono_call_inst_add_outarg_reg (cfg, call, this->dreg, cinfo->args [0].reg, FALSE);
3750 MONO_INST_NEW (cfg, this, OP_OUTARG);
3751 this->type = this_type;
3752 this->sreg1 = this_reg;
3753 mono_bblock_add_inst (cfg->cbb, this);
3760 if (cinfo->ret.storage == ArgValuetypeInReg) {
3762 * The valuetype is in EAX:EDX after the call, needs to be copied to
3763 * the stack. Save the address here, so the call instruction can
3766 MONO_INST_NEW (cfg, vtarg, OP_STORE_MEMBASE_REG);
3767 vtarg->inst_destbasereg = X86_ESP;
3768 vtarg->inst_offset = inst->stack_usage;
3769 vtarg->sreg1 = vt_reg;
3770 mono_bblock_add_inst (cfg->cbb, vtarg);
3772 else if (cinfo->ret.storage == ArgInIReg) {
3773 /* The return address is passed in a register */
3774 MONO_INST_NEW (cfg, vtarg, OP_MOVE);
3775 vtarg->sreg1 = vt_reg;
3776 vtarg->dreg = mono_regstate_next_int (cfg->rs);
3777 mono_bblock_add_inst (cfg->cbb, vtarg);
3779 mono_call_inst_add_outarg_reg (cfg, call, vtarg->dreg, cinfo->ret.reg, FALSE);
3782 MONO_INST_NEW (cfg, vtarg, OP_OUTARG);
3783 vtarg->type = STACK_MP;
3784 vtarg->sreg1 = vt_reg;
3785 mono_bblock_add_inst (cfg->cbb, vtarg);
3793 mono_arch_get_inst_for_method (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature *fsig, MonoInst **args)
3795 MonoInst *ins = NULL;
3797 if (cmethod->klass == mono_defaults.math_class) {
3798 if (strcmp (cmethod->name, "Sin") == 0) {
3799 MONO_INST_NEW (cfg, ins, OP_SIN);
3800 ins->inst_i0 = args [0];
3801 } else if (strcmp (cmethod->name, "Cos") == 0) {
3802 MONO_INST_NEW (cfg, ins, OP_COS);
3803 ins->inst_i0 = args [0];
3804 } else if (strcmp (cmethod->name, "Tan") == 0) {
3805 MONO_INST_NEW (cfg, ins, OP_TAN);
3806 ins->inst_i0 = args [0];
3807 } else if (strcmp (cmethod->name, "Atan") == 0) {
3808 MONO_INST_NEW (cfg, ins, OP_ATAN);
3809 ins->inst_i0 = args [0];
3810 } else if (strcmp (cmethod->name, "Sqrt") == 0) {
3811 MONO_INST_NEW (cfg, ins, OP_SQRT);
3812 ins->inst_i0 = args [0];
3813 } else if (strcmp (cmethod->name, "Abs") == 0 && fsig->params [0]->type == MONO_TYPE_R8) {
3814 MONO_INST_NEW (cfg, ins, OP_ABS);
3815 ins->inst_i0 = args [0];
3818 /* OP_FREM is not IEEE compatible */
3819 else if (strcmp (cmethod->name, "IEEERemainder") == 0) {
3820 MONO_INST_NEW (cfg, ins, OP_FREM);
3821 ins->inst_i0 = args [0];
3822 ins->inst_i1 = args [1];
3825 } else if (cmethod->klass == mono_defaults.thread_class &&
3826 strcmp (cmethod->name, "MemoryBarrier") == 0) {
3827 MONO_INST_NEW (cfg, ins, OP_MEMORY_BARRIER);
3828 } else if(cmethod->klass->image == mono_defaults.corlib &&
3829 (strcmp (cmethod->klass->name_space, "System.Threading") == 0) &&
3830 (strcmp (cmethod->klass->name, "Interlocked") == 0)) {
3832 if (strcmp (cmethod->name, "Increment") == 0 && fsig->params [0]->type == MONO_TYPE_I4) {
3833 MonoInst *ins_iconst;
3835 MONO_INST_NEW (cfg, ins, OP_ATOMIC_ADD_NEW_I4);
3836 MONO_INST_NEW (cfg, ins_iconst, OP_ICONST);
3837 ins_iconst->inst_c0 = 1;
3839 ins->inst_i0 = args [0];
3840 ins->inst_i1 = ins_iconst;
3841 } else if (strcmp (cmethod->name, "Decrement") == 0 && fsig->params [0]->type == MONO_TYPE_I4) {
3842 MonoInst *ins_iconst;
3844 MONO_INST_NEW (cfg, ins, OP_ATOMIC_ADD_NEW_I4);
3845 MONO_INST_NEW (cfg, ins_iconst, OP_ICONST);
3846 ins_iconst->inst_c0 = -1;
3848 ins->inst_i0 = args [0];
3849 ins->inst_i1 = ins_iconst;
3850 } else if (strcmp (cmethod->name, "Exchange") == 0 && fsig->params [0]->type == MONO_TYPE_I4) {
3851 MONO_INST_NEW (cfg, ins, OP_ATOMIC_EXCHANGE_I4);
3853 ins->inst_i0 = args [0];
3854 ins->inst_i1 = args [1];
3855 } else if (strcmp (cmethod->name, "Add") == 0 && fsig->params [0]->type == MONO_TYPE_I4) {
3856 MONO_INST_NEW (cfg, ins, OP_ATOMIC_ADD_NEW_I4);
3858 ins->inst_i0 = args [0];
3859 ins->inst_i1 = args [1];
3868 mono_arch_print_tree (MonoInst *tree, int arity)
3873 MonoInst* mono_arch_get_domain_intrinsic (MonoCompile* cfg)
3877 if (appdomain_tls_offset == -1)
3880 MONO_INST_NEW (cfg, ins, OP_TLS_GET);
3881 ins->inst_offset = appdomain_tls_offset;
3885 MonoInst* mono_arch_get_thread_intrinsic (MonoCompile* cfg)
3889 if (thread_tls_offset == -1)
3892 MONO_INST_NEW (cfg, ins, OP_TLS_GET);
3893 ins->inst_offset = thread_tls_offset;
3898 mono_arch_get_patch_offset (guint8 *code)
3900 if ((code [0] == 0x8b) && (x86_modrm_mod (code [1]) == 0x2))
3902 else if ((code [0] == 0xba))
3904 else if ((code [0] == 0x68))
3907 else if ((code [0] == 0xff) && (x86_modrm_reg (code [1]) == 0x6))
3908 /* push <OFFSET>(<REG>) */
3910 else if ((code [0] == 0xff) && (x86_modrm_reg (code [1]) == 0x2))
3911 /* call *<OFFSET>(<REG>) */
3913 else if ((code [0] == 0xdd) || (code [0] == 0xd9))
3916 else if ((code [0] == 0x58) && (code [1] == 0x05))
3917 /* pop %eax; add <OFFSET>, %eax */
3919 else if ((code [0] >= 0x58) && (code [0] <= 0x58 + X86_NREG) && (code [1] == 0x81))
3920 /* pop <REG>; add <OFFSET>, <REG> */
3923 g_assert_not_reached ();
3929 mono_arch_get_vcall_slot_addr (guint8 *code, gpointer *regs)
3934 /* go to the start of the call instruction
3936 * address_byte = (m << 6) | (o << 3) | reg
3937 * call opcode: 0xff address_byte displacement
3939 * 0xff m=2,o=2 imm32
3944 * A given byte sequence can match more than case here, so we have to be
3945 * really careful about the ordering of the cases. Longer sequences
3948 if ((code [-2] == 0x8b) && (x86_modrm_mod (code [-1]) == 0x2) && (code [4] == 0xff) && (x86_modrm_reg (code [5]) == 0x2) && (x86_modrm_mod (code [5]) == 0x0)) {
3950 * This is an interface call
3951 * 8b 80 0c e8 ff ff mov 0xffffe80c(%eax),%eax
3952 * ff 10 call *(%eax)
3954 reg = x86_modrm_rm (code [5]);
3956 } else if ((code [1] != 0xe8) && (code [3] == 0xff) && ((code [4] & 0x18) == 0x10) && ((code [4] >> 6) == 1)) {
3957 reg = code [4] & 0x07;
3958 disp = (signed char)code [5];
3960 if ((code [0] == 0xff) && ((code [1] & 0x18) == 0x10) && ((code [1] >> 6) == 2)) {
3961 reg = code [1] & 0x07;
3962 disp = *((gint32*)(code + 2));
3963 } else if ((code [1] == 0xe8)) {
3965 } else if ((code [4] == 0xff) && (((code [5] >> 6) & 0x3) == 0) && (((code [5] >> 3) & 0x7) == 2)) {
3967 * This is a interface call
3968 * 8b 40 30 mov 0x30(%eax),%eax
3969 * ff 10 call *(%eax)
3972 reg = code [5] & 0x07;
3978 return (gpointer*)(((gint32)(regs [reg])) + disp);
3982 mono_arch_get_delegate_method_ptr_addr (guint8* code, gpointer *regs)
3988 if ((code [0] == 0x8b) && (x86_modrm_mod (code [1]) == 3) && (x86_modrm_reg (code [1]) == X86_EAX) && (code [2] == 0x8b) && (code [3] == 0x40) && (code [5] == 0xff) && (code [6] == 0xd0)) {
3989 reg = x86_modrm_rm (code [1]);
3995 return (gpointer*)(((gint32)(regs [reg])) + disp);