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.
15 #ifndef PLATFORM_WIN32
20 #include <mono/metadata/appdomain.h>
21 #include <mono/metadata/debug-helpers.h>
22 #include <mono/metadata/threads.h>
23 #include <mono/metadata/profiler-private.h>
24 #include <mono/utils/mono-math.h>
29 #include "cpu-pentium.h"
31 /* On windows, these hold the key returned by TlsAlloc () */
32 static gint lmf_tls_offset = -1;
33 static gint appdomain_tls_offset = -1;
34 static gint thread_tls_offset = -1;
36 #define ALIGN_TO(val,align) ((((guint64)val) + ((align) - 1)) & ~((align) - 1))
39 /* Under windows, the default pinvoke calling convention is stdcall */
40 #define CALLCONV_IS_STDCALL(sig) ((((sig)->call_convention) == MONO_CALL_STDCALL) || ((sig)->pinvoke && ((sig)->call_convention) == MONO_CALL_DEFAULT))
42 #define CALLCONV_IS_STDCALL(sig) (((sig)->call_convention) == MONO_CALL_STDCALL)
45 #define SIGNAL_STACK_SIZE (64 * 1024)
47 #define NOT_IMPLEMENTED g_assert_not_reached ()
50 mono_arch_regname (int reg) {
52 case X86_EAX: return "%eax";
53 case X86_EBX: return "%ebx";
54 case X86_ECX: return "%ecx";
55 case X86_EDX: return "%edx";
56 case X86_ESP: return "%esp"; case X86_EBP: return "%ebp";
57 case X86_EDI: return "%edi";
58 case X86_ESI: return "%esi";
79 /* Only if storage == ArgValuetypeInReg */
80 ArgStorage pair_storage [2];
89 gboolean need_stack_align;
97 #define FLOAT_PARAM_REGS 0
99 static X86_Reg_No param_regs [] = { 0 };
101 #ifdef PLATFORM_WIN32
102 static X86_Reg_No return_regs [] = { X86_EAX, X86_EDX };
106 add_general (guint32 *gr, guint32 *stack_size, ArgInfo *ainfo)
108 ainfo->offset = *stack_size;
110 if (*gr >= PARAM_REGS) {
111 ainfo->storage = ArgOnStack;
112 (*stack_size) += sizeof (gpointer);
115 ainfo->storage = ArgInIReg;
116 ainfo->reg = param_regs [*gr];
122 add_general_pair (guint32 *gr, guint32 *stack_size, ArgInfo *ainfo)
124 ainfo->offset = *stack_size;
126 g_assert (PARAM_REGS == 0);
128 ainfo->storage = ArgOnStack;
129 (*stack_size) += sizeof (gpointer) * 2;
133 add_float (guint32 *gr, guint32 *stack_size, ArgInfo *ainfo, gboolean is_double)
135 ainfo->offset = *stack_size;
137 if (*gr >= FLOAT_PARAM_REGS) {
138 ainfo->storage = ArgOnStack;
139 (*stack_size) += sizeof (gpointer);
142 /* A double register */
144 ainfo->storage = ArgInDoubleSSEReg;
146 ainfo->storage = ArgInFloatSSEReg;
154 add_valuetype (MonoMethodSignature *sig, ArgInfo *ainfo, MonoType *type,
156 guint32 *gr, guint32 *fr, guint32 *stack_size)
161 klass = mono_class_from_mono_type (type);
163 size = mono_type_native_stack_size (&klass->byval_arg, NULL);
165 size = mono_type_stack_size (&klass->byval_arg, NULL);
167 #ifdef PLATFORM_WIN32
168 if (sig->pinvoke && is_return) {
169 MonoMarshalType *info;
172 * the exact rules are not very well documented, the code below seems to work with the
173 * code generated by gcc 3.3.3 -mno-cygwin.
175 info = mono_marshal_load_type_info (klass);
178 ainfo->pair_storage [0] = ainfo->pair_storage [1] = ArgNone;
180 /* Special case structs with only a float member */
181 if ((info->native_size == 8) && (info->num_fields == 1) && (info->fields [0].field->type->type == MONO_TYPE_R8)) {
182 ainfo->storage = ArgValuetypeInReg;
183 ainfo->pair_storage [0] = ArgOnDoubleFpStack;
186 if ((info->native_size == 4) && (info->num_fields == 1) && (info->fields [0].field->type->type == MONO_TYPE_R4)) {
187 ainfo->storage = ArgValuetypeInReg;
188 ainfo->pair_storage [0] = ArgOnFloatFpStack;
191 if ((info->native_size == 1) || (info->native_size == 2) || (info->native_size == 4) || (info->native_size == 8)) {
192 ainfo->storage = ArgValuetypeInReg;
193 ainfo->pair_storage [0] = ArgInIReg;
194 ainfo->pair_regs [0] = return_regs [0];
195 if (info->native_size > 4) {
196 ainfo->pair_storage [1] = ArgInIReg;
197 ainfo->pair_regs [1] = return_regs [1];
204 ainfo->offset = *stack_size;
205 ainfo->storage = ArgOnStack;
206 *stack_size += ALIGN_TO (size, sizeof (gpointer));
212 * Obtain information about a call according to the calling convention.
213 * For x86 ELF, see the "System V Application Binary Interface Intel386
214 * Architecture Processor Supplment, Fourth Edition" document for more
216 * For x86 win32, see ???.
219 get_call_info (MonoMethodSignature *sig, gboolean is_pinvoke)
223 int n = sig->hasthis + sig->param_count;
224 guint32 stack_size = 0;
227 cinfo = g_malloc0 (sizeof (CallInfo) + (sizeof (ArgInfo) * n));
234 ret_type = mono_type_get_underlying_type (sig->ret);
235 switch (ret_type->type) {
236 case MONO_TYPE_BOOLEAN:
247 case MONO_TYPE_FNPTR:
248 case MONO_TYPE_CLASS:
249 case MONO_TYPE_OBJECT:
250 case MONO_TYPE_SZARRAY:
251 case MONO_TYPE_ARRAY:
252 case MONO_TYPE_STRING:
253 cinfo->ret.storage = ArgInIReg;
254 cinfo->ret.reg = X86_EAX;
258 cinfo->ret.storage = ArgInIReg;
259 cinfo->ret.reg = X86_EAX;
262 cinfo->ret.storage = ArgOnFloatFpStack;
265 cinfo->ret.storage = ArgOnDoubleFpStack;
267 case MONO_TYPE_VALUETYPE: {
268 guint32 tmp_gr = 0, tmp_fr = 0, tmp_stacksize = 0;
270 add_valuetype (sig, &cinfo->ret, sig->ret, TRUE, &tmp_gr, &tmp_fr, &tmp_stacksize);
271 if (cinfo->ret.storage == ArgOnStack)
272 /* The caller passes the address where the value is stored */
273 add_general (&gr, &stack_size, &cinfo->ret);
276 case MONO_TYPE_TYPEDBYREF:
277 /* Same as a valuetype with size 24 */
278 add_general (&gr, &stack_size, &cinfo->ret);
282 cinfo->ret.storage = ArgNone;
285 g_error ("Can't handle as return value 0x%x", sig->ret->type);
291 add_general (&gr, &stack_size, cinfo->args + 0);
293 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG) && (n == 0)) {
295 fr = FLOAT_PARAM_REGS;
297 /* Emit the signature cookie just before the implicit arguments */
298 add_general (&gr, &stack_size, &cinfo->sig_cookie);
301 for (i = 0; i < sig->param_count; ++i) {
302 ArgInfo *ainfo = &cinfo->args [sig->hasthis + i];
305 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG) && (i == sig->sentinelpos)) {
306 /* We allways pass the sig cookie on the stack for simplicity */
308 * Prevent implicit arguments + the sig cookie from being passed
312 fr = FLOAT_PARAM_REGS;
314 /* Emit the signature cookie just before the implicit arguments */
315 add_general (&gr, &stack_size, &cinfo->sig_cookie);
318 if (sig->params [i]->byref) {
319 add_general (&gr, &stack_size, ainfo);
322 ptype = mono_type_get_underlying_type (sig->params [i]);
323 switch (ptype->type) {
324 case MONO_TYPE_BOOLEAN:
327 add_general (&gr, &stack_size, ainfo);
332 add_general (&gr, &stack_size, ainfo);
336 add_general (&gr, &stack_size, ainfo);
341 case MONO_TYPE_FNPTR:
342 case MONO_TYPE_CLASS:
343 case MONO_TYPE_OBJECT:
344 case MONO_TYPE_STRING:
345 case MONO_TYPE_SZARRAY:
346 case MONO_TYPE_ARRAY:
347 add_general (&gr, &stack_size, ainfo);
349 case MONO_TYPE_VALUETYPE:
350 add_valuetype (sig, ainfo, sig->params [i], FALSE, &gr, &fr, &stack_size);
352 case MONO_TYPE_TYPEDBYREF:
353 stack_size += sizeof (MonoTypedRef);
354 ainfo->storage = ArgOnStack;
358 add_general_pair (&gr, &stack_size, ainfo);
361 add_float (&fr, &stack_size, ainfo, FALSE);
364 add_float (&fr, &stack_size, ainfo, TRUE);
367 g_error ("unexpected type 0x%x", ptype->type);
368 g_assert_not_reached ();
372 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG) && (n > 0) && (sig->sentinelpos == sig->param_count)) {
374 fr = FLOAT_PARAM_REGS;
376 /* Emit the signature cookie just before the implicit arguments */
377 add_general (&gr, &stack_size, &cinfo->sig_cookie);
380 cinfo->stack_usage = stack_size;
381 cinfo->reg_usage = gr;
382 cinfo->freg_usage = fr;
387 * mono_arch_get_argument_info:
388 * @csig: a method signature
389 * @param_count: the number of parameters to consider
390 * @arg_info: an array to store the result infos
392 * Gathers information on parameters such as size, alignment and
393 * padding. arg_info should be large enought to hold param_count + 1 entries.
395 * Returns the size of the activation frame.
398 mono_arch_get_argument_info (MonoMethodSignature *csig, int param_count, MonoJitArgumentInfo *arg_info)
400 int k, frame_size = 0;
401 int size, align, pad;
405 cinfo = get_call_info (csig, FALSE);
407 if (MONO_TYPE_ISSTRUCT (csig->ret) && (cinfo->ret.storage == ArgOnStack)) {
408 frame_size += sizeof (gpointer);
412 arg_info [0].offset = offset;
415 frame_size += sizeof (gpointer);
419 arg_info [0].size = frame_size;
421 for (k = 0; k < param_count; k++) {
424 size = mono_type_native_stack_size (csig->params [k], &align);
426 size = mono_type_stack_size (csig->params [k], &align);
428 /* ignore alignment for now */
431 frame_size += pad = (align - (frame_size & (align - 1))) & (align - 1);
432 arg_info [k].pad = pad;
434 arg_info [k + 1].pad = 0;
435 arg_info [k + 1].size = size;
437 arg_info [k + 1].offset = offset;
441 align = MONO_ARCH_FRAME_ALIGNMENT;
442 frame_size += pad = (align - (frame_size & (align - 1))) & (align - 1);
443 arg_info [k].pad = pad;
450 static const guchar cpuid_impl [] = {
451 0x55, /* push %ebp */
452 0x89, 0xe5, /* mov %esp,%ebp */
453 0x53, /* push %ebx */
454 0x8b, 0x45, 0x08, /* mov 0x8(%ebp),%eax */
455 0x0f, 0xa2, /* cpuid */
456 0x50, /* push %eax */
457 0x8b, 0x45, 0x10, /* mov 0x10(%ebp),%eax */
458 0x89, 0x18, /* mov %ebx,(%eax) */
459 0x8b, 0x45, 0x14, /* mov 0x14(%ebp),%eax */
460 0x89, 0x08, /* mov %ecx,(%eax) */
461 0x8b, 0x45, 0x18, /* mov 0x18(%ebp),%eax */
462 0x89, 0x10, /* mov %edx,(%eax) */
464 0x8b, 0x55, 0x0c, /* mov 0xc(%ebp),%edx */
465 0x89, 0x02, /* mov %eax,(%edx) */
471 typedef void (*CpuidFunc) (int id, int* p_eax, int* p_ebx, int* p_ecx, int* p_edx);
474 cpuid (int id, int* p_eax, int* p_ebx, int* p_ecx, int* p_edx)
478 __asm__ __volatile__ (
481 "movl %%eax, %%edx\n"
482 "xorl $0x200000, %%eax\n"
487 "xorl %%edx, %%eax\n"
488 "andl $0x200000, %%eax\n"
510 /* Have to use the code manager to get around WinXP DEP */
511 MonoCodeManager *codeman = mono_code_manager_new_dynamic ();
513 void *ptr = mono_code_manager_reserve (codeman, sizeof (cpuid_impl));
514 memcpy (ptr, cpuid_impl, sizeof (cpuid_impl));
516 func = (CpuidFunc)ptr;
517 func (id, p_eax, p_ebx, p_ecx, p_edx);
519 mono_code_manager_destroy (codeman);
522 * We use this approach because of issues with gcc and pic code, see:
523 * http://gcc.gnu.org/cgi-bin/gnatsweb.pl?cmd=view%20audit-trail&database=gcc&pr=7329
524 __asm__ __volatile__ ("cpuid"
525 : "=a" (*p_eax), "=b" (*p_ebx), "=c" (*p_ecx), "=d" (*p_edx)
534 * Initialize the cpu to execute managed code.
537 mono_arch_cpu_init (void)
539 /* spec compliance requires running with double precision */
543 __asm__ __volatile__ ("fnstcw %0\n": "=m" (fpcw));
544 fpcw &= ~X86_FPCW_PRECC_MASK;
545 fpcw |= X86_FPCW_PREC_DOUBLE;
546 __asm__ __volatile__ ("fldcw %0\n": : "m" (fpcw));
547 __asm__ __volatile__ ("fnstcw %0\n": "=m" (fpcw));
549 _control87 (_PC_53, MCW_PC);
554 * This function returns the optimizations supported on this cpu.
557 mono_arch_cpu_optimizazions (guint32 *exclude_mask)
559 int eax, ebx, ecx, edx;
563 /* Feature Flags function, flags returned in EDX. */
564 if (cpuid (1, &eax, &ebx, &ecx, &edx)) {
565 if (edx & (1 << 15)) {
566 opts |= MONO_OPT_CMOV;
568 opts |= MONO_OPT_FCMOV;
570 *exclude_mask |= MONO_OPT_FCMOV;
572 *exclude_mask |= MONO_OPT_CMOV;
578 * Determine whenever the trap whose info is in SIGINFO is caused by
582 mono_arch_is_int_overflow (void *sigctx, void *info)
587 mono_arch_sigctx_to_monoctx (sigctx, &ctx);
589 ip = (guint8*)ctx.eip;
591 if ((ip [0] == 0xf7) && (x86_modrm_mod (ip [1]) == 0x3) && (x86_modrm_reg (ip [1]) == 0x7)) {
595 switch (x86_modrm_rm (ip [1])) {
603 g_assert_not_reached ();
615 is_regsize_var (MonoType *t) {
618 switch (mono_type_get_underlying_type (t)->type) {
624 case MONO_TYPE_FNPTR:
626 case MONO_TYPE_OBJECT:
627 case MONO_TYPE_STRING:
628 case MONO_TYPE_CLASS:
629 case MONO_TYPE_SZARRAY:
630 case MONO_TYPE_ARRAY:
632 case MONO_TYPE_VALUETYPE:
639 mono_arch_get_allocatable_int_vars (MonoCompile *cfg)
644 for (i = 0; i < cfg->num_varinfo; i++) {
645 MonoInst *ins = cfg->varinfo [i];
646 MonoMethodVar *vmv = MONO_VARINFO (cfg, i);
649 if (vmv->range.first_use.abs_pos >= vmv->range.last_use.abs_pos)
652 if ((ins->flags & (MONO_INST_IS_DEAD|MONO_INST_VOLATILE|MONO_INST_INDIRECT)) ||
653 (ins->opcode != OP_LOCAL && ins->opcode != OP_ARG))
656 /* we dont allocate I1 to registers because there is no simply way to sign extend
657 * 8bit quantities in caller saved registers on x86 */
658 if (is_regsize_var (ins->inst_vtype) || (ins->inst_vtype->type == MONO_TYPE_BOOLEAN) ||
659 (ins->inst_vtype->type == MONO_TYPE_U1) || (ins->inst_vtype->type == MONO_TYPE_U2)||
660 (ins->inst_vtype->type == MONO_TYPE_I2) || (ins->inst_vtype->type == MONO_TYPE_CHAR)) {
661 g_assert (MONO_VARINFO (cfg, i)->reg == -1);
662 g_assert (i == vmv->idx);
663 vars = g_list_prepend (vars, vmv);
667 vars = mono_varlist_sort (cfg, vars, 0);
673 mono_arch_get_global_int_regs (MonoCompile *cfg)
677 /* we can use 3 registers for global allocation */
678 regs = g_list_prepend (regs, (gpointer)X86_EBX);
679 regs = g_list_prepend (regs, (gpointer)X86_ESI);
680 regs = g_list_prepend (regs, (gpointer)X86_EDI);
686 * mono_arch_regalloc_cost:
688 * Return the cost, in number of memory references, of the action of
689 * allocating the variable VMV into a register during global register
693 mono_arch_regalloc_cost (MonoCompile *cfg, MonoMethodVar *vmv)
695 MonoInst *ins = cfg->varinfo [vmv->idx];
697 if (cfg->method->save_lmf)
698 /* The register is already saved */
699 return (ins->opcode == OP_ARG) ? 1 : 0;
701 /* push+pop+possible load if it is an argument */
702 return (ins->opcode == OP_ARG) ? 3 : 2;
706 * Set var information according to the calling convention. X86 version.
707 * The locals var stuff should most likely be split in another method.
710 mono_arch_allocate_vars (MonoCompile *m)
712 MonoMethodSignature *sig;
713 MonoMethodHeader *header;
715 guint32 locals_stack_size, locals_stack_align;
716 int i, offset, curinst, size, align;
720 header = mono_method_get_header (m->method);
721 sig = mono_method_signature (m->method);
726 cinfo = get_call_info (sig, FALSE);
728 switch (cinfo->ret.storage) {
730 m->ret->opcode = OP_REGOFFSET;
731 m->ret->inst_basereg = X86_EBP;
732 m->ret->inst_offset = offset;
733 offset += sizeof (gpointer);
735 case ArgValuetypeInReg:
738 m->ret->opcode = OP_REGVAR;
739 m->ret->inst_c0 = cinfo->ret.reg;
742 case ArgOnFloatFpStack:
743 case ArgOnDoubleFpStack:
746 g_assert_not_reached ();
750 inst = m->varinfo [curinst];
751 if (inst->opcode != OP_REGVAR) {
752 inst->opcode = OP_REGOFFSET;
753 inst->inst_basereg = X86_EBP;
755 inst->inst_offset = offset;
756 offset += sizeof (gpointer);
760 if (sig->call_convention == MONO_CALL_VARARG) {
761 m->sig_cookie = offset;
762 offset += sizeof (gpointer);
765 for (i = 0; i < sig->param_count; ++i) {
766 inst = m->varinfo [curinst];
767 if (inst->opcode != OP_REGVAR) {
768 inst->opcode = OP_REGOFFSET;
769 inst->inst_basereg = X86_EBP;
771 inst->inst_offset = offset;
772 size = mono_type_size (sig->params [i], &align);
781 /* reserve space to save LMF and caller saved registers */
783 if (m->method->save_lmf) {
784 offset += sizeof (MonoLMF);
786 if (m->used_int_regs & (1 << X86_EBX)) {
790 if (m->used_int_regs & (1 << X86_EDI)) {
794 if (m->used_int_regs & (1 << X86_ESI)) {
799 switch (cinfo->ret.storage) {
800 case ArgValuetypeInReg:
801 /* Allocate a local to hold the result, the epilog will copy it to the correct place */
803 m->ret->opcode = OP_REGOFFSET;
804 m->ret->inst_basereg = X86_EBP;
805 m->ret->inst_offset = - offset;
811 /* Allocate locals */
812 offsets = mono_allocate_stack_slots (m, &locals_stack_size, &locals_stack_align);
813 if (locals_stack_align) {
814 offset += (locals_stack_align - 1);
815 offset &= ~(locals_stack_align - 1);
817 for (i = m->locals_start; i < m->num_varinfo; i++) {
818 if (offsets [i] != -1) {
819 MonoInst *inst = m->varinfo [i];
820 inst->opcode = OP_REGOFFSET;
821 inst->inst_basereg = X86_EBP;
822 inst->inst_offset = - (offset + offsets [i]);
823 //printf ("allocated local %d to ", i); mono_print_tree_nl (inst);
827 offset += locals_stack_size;
829 offset += (MONO_ARCH_FRAME_ALIGNMENT - 1);
830 offset &= ~(MONO_ARCH_FRAME_ALIGNMENT - 1);
835 m->stack_offset = -offset;
839 mono_arch_create_vars (MonoCompile *cfg)
841 MonoMethodSignature *sig;
844 sig = mono_method_signature (cfg->method);
846 cinfo = get_call_info (sig, FALSE);
848 if (cinfo->ret.storage == ArgValuetypeInReg)
849 cfg->ret_var_is_local = TRUE;
854 /* Fixme: we need an alignment solution for enter_method and mono_arch_call_opcode,
855 * currently alignment in mono_arch_call_opcode is computed without arch_get_argument_info
859 * take the arguments and generate the arch-specific
860 * instructions to properly call the function in call.
861 * This includes pushing, moving arguments to the right register
863 * Issue: who does the spilling if needed, and when?
866 mono_arch_call_opcode (MonoCompile *cfg, MonoBasicBlock* bb, MonoCallInst *call, int is_virtual) {
868 MonoMethodSignature *sig;
869 int i, n, stack_size, type;
874 /* add the vararg cookie before the non-implicit args */
875 if (call->signature->call_convention == MONO_CALL_VARARG) {
877 /* FIXME: Add support for signature tokens to AOT */
878 cfg->disable_aot = TRUE;
879 MONO_INST_NEW (cfg, arg, OP_OUTARG);
880 MONO_INST_NEW (cfg, sig_arg, OP_ICONST);
881 sig_arg->inst_p0 = call->signature;
882 arg->inst_left = sig_arg;
883 arg->type = STACK_PTR;
884 /* prepend, so they get reversed */
885 arg->next = call->out_args;
886 call->out_args = arg;
887 stack_size += sizeof (gpointer);
889 sig = call->signature;
890 n = sig->param_count + sig->hasthis;
892 cinfo = get_call_info (sig, FALSE);
894 if (sig->ret && MONO_TYPE_ISSTRUCT (sig->ret)) {
895 if (cinfo->ret.storage == ArgOnStack)
896 stack_size += sizeof (gpointer);
899 for (i = 0; i < n; ++i) {
900 if (is_virtual && i == 0) {
901 /* the argument will be attached to the call instrucion */
905 MONO_INST_NEW (cfg, arg, OP_OUTARG);
907 arg->cil_code = in->cil_code;
909 arg->type = in->type;
910 /* prepend, so they get reversed */
911 arg->next = call->out_args;
912 call->out_args = arg;
913 if (i >= sig->hasthis) {
914 MonoType *t = sig->params [i - sig->hasthis];
915 ptype = mono_type_get_underlying_type (t);
920 /* FIXME: validate arguments... */
924 case MONO_TYPE_BOOLEAN:
932 case MONO_TYPE_STRING:
933 case MONO_TYPE_CLASS:
934 case MONO_TYPE_OBJECT:
936 case MONO_TYPE_FNPTR:
937 case MONO_TYPE_ARRAY:
938 case MONO_TYPE_SZARRAY:
947 arg->opcode = OP_OUTARG_R4;
951 arg->opcode = OP_OUTARG_R8;
953 case MONO_TYPE_VALUETYPE: {
956 size = mono_type_native_stack_size (&in->klass->byval_arg, NULL);
958 size = mono_type_stack_size (&in->klass->byval_arg, NULL);
961 arg->opcode = OP_OUTARG_VT;
962 arg->klass = in->klass;
963 arg->unused = sig->pinvoke;
964 arg->inst_imm = size;
967 case MONO_TYPE_TYPEDBYREF:
968 stack_size += sizeof (MonoTypedRef);
969 arg->opcode = OP_OUTARG_VT;
970 arg->klass = in->klass;
971 arg->unused = sig->pinvoke;
972 arg->inst_imm = sizeof (MonoTypedRef);
975 g_error ("unknown type 0x%02x in mono_arch_call_opcode\n", type);
978 /* the this argument */
984 if (sig->ret && MONO_TYPE_ISSTRUCT (sig->ret)) {
985 if (cinfo->ret.storage == ArgValuetypeInReg) {
988 * After the call, the struct is in registers, but needs to be saved to the memory pointed
989 * to by vt_arg in this_vret_args. This means that vt_arg needs to be saved somewhere
990 * before calling the function. So we add a dummy instruction to represent pushing the
991 * struct return address to the stack. The return address will be saved to this stack slot
992 * by the code emitted in this_vret_args.
994 MONO_INST_NEW (cfg, arg, OP_OUTARG);
995 MONO_INST_NEW (cfg, zero_inst, OP_ICONST);
996 zero_inst->inst_p0 = 0;
997 arg->inst_left = zero_inst;
998 arg->type = STACK_PTR;
999 /* prepend, so they get reversed */
1000 arg->next = call->out_args;
1001 call->out_args = arg;
1004 /* if the function returns a struct, the called method already does a ret $0x4 */
1005 if (sig->ret && MONO_TYPE_ISSTRUCT (sig->ret))
1009 call->stack_usage = stack_size;
1013 * should set more info in call, such as the stack space
1014 * used by the args that needs to be added back to esp
1021 * Allow tracing to work with this interface (with an optional argument)
1024 mono_arch_instrument_prolog (MonoCompile *cfg, void *func, void *p, gboolean enable_arguments)
1028 /* if some args are passed in registers, we need to save them here */
1029 x86_push_reg (code, X86_EBP);
1031 if (cfg->compile_aot) {
1032 x86_push_imm (code, cfg->method);
1033 x86_mov_reg_imm (code, X86_EAX, func);
1034 x86_call_reg (code, X86_EAX);
1036 mono_add_patch_info (cfg, code-cfg->native_code, MONO_PATCH_INFO_METHODCONST, cfg->method);
1037 x86_push_imm (code, cfg->method);
1038 mono_add_patch_info (cfg, code-cfg->native_code, MONO_PATCH_INFO_ABS, func);
1039 x86_call_code (code, 0);
1041 x86_alu_reg_imm (code, X86_ADD, X86_ESP, 8);
1055 mono_arch_instrument_epilog (MonoCompile *cfg, void *func, void *p, gboolean enable_arguments)
1058 int arg_size = 0, save_mode = SAVE_NONE;
1059 MonoMethod *method = cfg->method;
1061 switch (mono_type_get_underlying_type (mono_method_signature (method)->ret)->type) {
1062 case MONO_TYPE_VOID:
1063 /* special case string .ctor icall */
1064 if (strcmp (".ctor", method->name) && method->klass == mono_defaults.string_class)
1065 save_mode = SAVE_EAX;
1067 save_mode = SAVE_NONE;
1071 save_mode = SAVE_EAX_EDX;
1075 save_mode = SAVE_FP;
1077 case MONO_TYPE_VALUETYPE:
1078 save_mode = SAVE_STRUCT;
1081 save_mode = SAVE_EAX;
1085 switch (save_mode) {
1087 x86_push_reg (code, X86_EDX);
1088 x86_push_reg (code, X86_EAX);
1089 if (enable_arguments) {
1090 x86_push_reg (code, X86_EDX);
1091 x86_push_reg (code, X86_EAX);
1096 x86_push_reg (code, X86_EAX);
1097 if (enable_arguments) {
1098 x86_push_reg (code, X86_EAX);
1103 x86_alu_reg_imm (code, X86_SUB, X86_ESP, 8);
1104 x86_fst_membase (code, X86_ESP, 0, TRUE, TRUE);
1105 if (enable_arguments) {
1106 x86_alu_reg_imm (code, X86_SUB, X86_ESP, 8);
1107 x86_fst_membase (code, X86_ESP, 0, TRUE, TRUE);
1112 if (enable_arguments) {
1113 x86_push_membase (code, X86_EBP, 8);
1122 if (cfg->compile_aot) {
1123 x86_push_imm (code, method);
1124 x86_mov_reg_imm (code, X86_EAX, func);
1125 x86_call_reg (code, X86_EAX);
1127 mono_add_patch_info (cfg, code-cfg->native_code, MONO_PATCH_INFO_METHODCONST, method);
1128 x86_push_imm (code, method);
1129 mono_add_patch_info (cfg, code-cfg->native_code, MONO_PATCH_INFO_ABS, func);
1130 x86_call_code (code, 0);
1132 x86_alu_reg_imm (code, X86_ADD, X86_ESP, arg_size + 4);
1134 switch (save_mode) {
1136 x86_pop_reg (code, X86_EAX);
1137 x86_pop_reg (code, X86_EDX);
1140 x86_pop_reg (code, X86_EAX);
1143 x86_fld_membase (code, X86_ESP, 0, TRUE);
1144 x86_alu_reg_imm (code, X86_ADD, X86_ESP, 8);
1154 #define EMIT_COND_BRANCH(ins,cond,sign) \
1155 if (ins->flags & MONO_INST_BRLABEL) { \
1156 if (ins->inst_i0->inst_c0) { \
1157 x86_branch (code, cond, cfg->native_code + ins->inst_i0->inst_c0, sign); \
1159 mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_LABEL, ins->inst_i0); \
1160 if ((cfg->opt & MONO_OPT_BRANCH) && \
1161 x86_is_imm8 (ins->inst_i0->inst_c1 - cpos)) \
1162 x86_branch8 (code, cond, 0, sign); \
1164 x86_branch32 (code, cond, 0, sign); \
1167 if (ins->inst_true_bb->native_offset) { \
1168 x86_branch (code, cond, cfg->native_code + ins->inst_true_bb->native_offset, sign); \
1170 mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_BB, ins->inst_true_bb); \
1171 if ((cfg->opt & MONO_OPT_BRANCH) && \
1172 x86_is_imm8 (ins->inst_true_bb->max_offset - cpos)) \
1173 x86_branch8 (code, cond, 0, sign); \
1175 x86_branch32 (code, cond, 0, sign); \
1179 /* emit an exception if condition is fail */
1180 #define EMIT_COND_SYSTEM_EXCEPTION(cond,signed,exc_name) \
1182 mono_add_patch_info (cfg, code - cfg->native_code, \
1183 MONO_PATCH_INFO_EXC, exc_name); \
1184 x86_branch32 (code, cond, 0, signed); \
1187 #define EMIT_FPCOMPARE(code) do { \
1188 x86_fcompp (code); \
1189 x86_fnstsw (code); \
1194 emit_call (MonoCompile *cfg, guint8 *code, guint32 patch_type, gconstpointer data)
1196 if (cfg->compile_aot) {
1197 guint32 got_reg = X86_EAX;
1199 if (cfg->compile_aot) {
1201 * Since the patches are generated by the back end, there is
1202 * no way to generate a got_var at this point.
1204 g_assert (cfg->got_var);
1206 if (cfg->got_var->opcode == OP_REGOFFSET)
1207 x86_mov_reg_membase (code, X86_EAX, cfg->got_var->inst_basereg, cfg->got_var->inst_offset, 4);
1209 got_reg = cfg->got_var->dreg;
1212 mono_add_patch_info (cfg, code - cfg->native_code, patch_type, data);
1213 x86_call_membase (code, got_reg, 0xf0f0f0f0);
1216 mono_add_patch_info (cfg, code - cfg->native_code, patch_type, data);
1217 x86_call_code (code, 0);
1223 /* FIXME: Add more instructions */
1224 #define INST_IGNORES_CFLAGS(ins) (((ins)->opcode == CEE_BR) || ((ins)->opcode == OP_STORE_MEMBASE_IMM) || ((ins)->opcode == OP_STOREI4_MEMBASE_REG))
1227 peephole_pass (MonoCompile *cfg, MonoBasicBlock *bb)
1229 MonoInst *ins, *last_ins = NULL;
1234 switch (ins->opcode) {
1236 /* reg = 0 -> XOR (reg, reg) */
1237 /* XOR sets cflags on x86, so we cant do it always */
1238 if (ins->inst_c0 == 0 && ins->next && INST_IGNORES_CFLAGS (ins->next)) {
1239 ins->opcode = CEE_XOR;
1240 ins->sreg1 = ins->dreg;
1241 ins->sreg2 = ins->dreg;
1245 /* remove unnecessary multiplication with 1 */
1246 if (ins->inst_imm == 1) {
1247 if (ins->dreg != ins->sreg1) {
1248 ins->opcode = OP_MOVE;
1250 last_ins->next = ins->next;
1256 case OP_COMPARE_IMM:
1257 /* OP_COMPARE_IMM (reg, 0)
1259 * OP_X86_TEST_NULL (reg)
1262 ins->opcode = OP_X86_TEST_NULL;
1264 case OP_X86_COMPARE_MEMBASE_IMM:
1266 * OP_STORE_MEMBASE_REG reg, offset(basereg)
1267 * OP_X86_COMPARE_MEMBASE_IMM offset(basereg), imm
1269 * OP_STORE_MEMBASE_REG reg, offset(basereg)
1270 * OP_COMPARE_IMM reg, imm
1272 * Note: if imm = 0 then OP_COMPARE_IMM replaced with OP_X86_TEST_NULL
1274 if (last_ins && (last_ins->opcode == OP_STOREI4_MEMBASE_REG) &&
1275 ins->inst_basereg == last_ins->inst_destbasereg &&
1276 ins->inst_offset == last_ins->inst_offset) {
1277 ins->opcode = OP_COMPARE_IMM;
1278 ins->sreg1 = last_ins->sreg1;
1280 /* check if we can remove cmp reg,0 with test null */
1282 ins->opcode = OP_X86_TEST_NULL;
1286 case OP_LOAD_MEMBASE:
1287 case OP_LOADI4_MEMBASE:
1289 * Note: if reg1 = reg2 the load op is removed
1291 * OP_STORE_MEMBASE_REG reg1, offset(basereg)
1292 * OP_LOAD_MEMBASE offset(basereg), reg2
1294 * OP_STORE_MEMBASE_REG reg1, offset(basereg)
1295 * OP_MOVE reg1, reg2
1297 if (last_ins && (last_ins->opcode == OP_STOREI4_MEMBASE_REG
1298 || last_ins->opcode == OP_STORE_MEMBASE_REG) &&
1299 ins->inst_basereg == last_ins->inst_destbasereg &&
1300 ins->inst_offset == last_ins->inst_offset) {
1301 if (ins->dreg == last_ins->sreg1) {
1302 last_ins->next = ins->next;
1306 //static int c = 0; printf ("MATCHX %s %d\n", cfg->method->name,c++);
1307 ins->opcode = OP_MOVE;
1308 ins->sreg1 = last_ins->sreg1;
1312 * Note: reg1 must be different from the basereg in the second load
1313 * Note: if reg1 = reg2 is equal then second load is removed
1315 * OP_LOAD_MEMBASE offset(basereg), reg1
1316 * OP_LOAD_MEMBASE offset(basereg), reg2
1318 * OP_LOAD_MEMBASE offset(basereg), reg1
1319 * OP_MOVE reg1, reg2
1321 } if (last_ins && (last_ins->opcode == OP_LOADI4_MEMBASE
1322 || last_ins->opcode == OP_LOAD_MEMBASE) &&
1323 ins->inst_basereg != last_ins->dreg &&
1324 ins->inst_basereg == last_ins->inst_basereg &&
1325 ins->inst_offset == last_ins->inst_offset) {
1327 if (ins->dreg == last_ins->dreg) {
1328 last_ins->next = ins->next;
1332 ins->opcode = OP_MOVE;
1333 ins->sreg1 = last_ins->dreg;
1336 //g_assert_not_reached ();
1340 * OP_STORE_MEMBASE_IMM imm, offset(basereg)
1341 * OP_LOAD_MEMBASE offset(basereg), reg
1343 * OP_STORE_MEMBASE_IMM imm, offset(basereg)
1344 * OP_ICONST reg, imm
1346 } else if (last_ins && (last_ins->opcode == OP_STOREI4_MEMBASE_IMM
1347 || last_ins->opcode == OP_STORE_MEMBASE_IMM) &&
1348 ins->inst_basereg == last_ins->inst_destbasereg &&
1349 ins->inst_offset == last_ins->inst_offset) {
1350 //static int c = 0; printf ("MATCHX %s %d\n", cfg->method->name,c++);
1351 ins->opcode = OP_ICONST;
1352 ins->inst_c0 = last_ins->inst_imm;
1353 g_assert_not_reached (); // check this rule
1357 case OP_LOADU1_MEMBASE:
1358 case OP_LOADI1_MEMBASE:
1360 * OP_STORE_MEMBASE_REG reg1, offset(basereg)
1361 * OP_LOAD_MEMBASE offset(basereg), reg2
1363 * OP_STORE_MEMBASE_REG reg1, offset(basereg)
1364 * CONV_I2/U2 reg1, reg2
1366 if (last_ins && (last_ins->opcode == OP_STOREI1_MEMBASE_REG) &&
1367 ins->inst_basereg == last_ins->inst_destbasereg &&
1368 ins->inst_offset == last_ins->inst_offset) {
1369 ins->opcode = (ins->opcode == OP_LOADI1_MEMBASE) ? CEE_CONV_I1 : CEE_CONV_U1;
1370 ins->sreg1 = last_ins->sreg1;
1373 case OP_LOADU2_MEMBASE:
1374 case OP_LOADI2_MEMBASE:
1376 * OP_STORE_MEMBASE_REG reg1, offset(basereg)
1377 * OP_LOAD_MEMBASE offset(basereg), reg2
1379 * OP_STORE_MEMBASE_REG reg1, offset(basereg)
1380 * CONV_I2/U2 reg1, reg2
1382 if (last_ins && (last_ins->opcode == OP_STOREI2_MEMBASE_REG) &&
1383 ins->inst_basereg == last_ins->inst_destbasereg &&
1384 ins->inst_offset == last_ins->inst_offset) {
1385 ins->opcode = (ins->opcode == OP_LOADI2_MEMBASE) ? CEE_CONV_I2 : CEE_CONV_U2;
1386 ins->sreg1 = last_ins->sreg1;
1397 if (ins->dreg == ins->sreg1) {
1399 last_ins->next = ins->next;
1406 * OP_MOVE sreg, dreg
1407 * OP_MOVE dreg, sreg
1409 if (last_ins && last_ins->opcode == OP_MOVE &&
1410 ins->sreg1 == last_ins->dreg &&
1411 ins->dreg == last_ins->sreg1) {
1412 last_ins->next = ins->next;
1418 case OP_X86_PUSH_MEMBASE:
1419 if (last_ins && (last_ins->opcode == OP_STOREI4_MEMBASE_REG ||
1420 last_ins->opcode == OP_STORE_MEMBASE_REG) &&
1421 ins->inst_basereg == last_ins->inst_destbasereg &&
1422 ins->inst_offset == last_ins->inst_offset) {
1423 ins->opcode = OP_X86_PUSH;
1424 ins->sreg1 = last_ins->sreg1;
1431 bb->last_ins = last_ins;
1435 branch_cc_table [] = {
1436 X86_CC_EQ, X86_CC_GE, X86_CC_GT, X86_CC_LE, X86_CC_LT,
1437 X86_CC_NE, X86_CC_GE, X86_CC_GT, X86_CC_LE, X86_CC_LT,
1438 X86_CC_O, X86_CC_NO, X86_CC_C, X86_CC_NC
1441 #define DEBUG(a) if (cfg->verbose_level > 1) a
1445 * returns the offset used by spillvar. It allocates a new
1446 * spill variable if necessary.
1449 mono_spillvar_offset (MonoCompile *cfg, int spillvar)
1451 MonoSpillInfo **si, *info;
1454 si = &cfg->spill_info;
1456 while (i <= spillvar) {
1459 *si = info = mono_mempool_alloc (cfg->mempool, sizeof (MonoSpillInfo));
1461 cfg->stack_offset -= sizeof (gpointer);
1462 info->offset = cfg->stack_offset;
1466 return (*si)->offset;
1472 g_assert_not_reached ();
1477 * returns the offset used by spillvar. It allocates a new
1478 * spill float variable if necessary.
1479 * (same as mono_spillvar_offset but for float)
1482 mono_spillvar_offset_float (MonoCompile *cfg, int spillvar)
1484 MonoSpillInfo **si, *info;
1487 si = &cfg->spill_info_float;
1489 while (i <= spillvar) {
1492 *si = info = mono_mempool_alloc (cfg->mempool, sizeof (MonoSpillInfo));
1494 cfg->stack_offset -= sizeof (double);
1495 info->offset = cfg->stack_offset;
1499 return (*si)->offset;
1505 g_assert_not_reached ();
1510 * Creates a store for spilled floating point items
1513 create_spilled_store_float (MonoCompile *cfg, int spill, int reg, MonoInst *ins)
1516 MONO_INST_NEW (cfg, store, OP_STORER8_MEMBASE_REG);
1518 store->inst_destbasereg = X86_EBP;
1519 store->inst_offset = mono_spillvar_offset_float (cfg, spill);
1521 DEBUG (g_print ("SPILLED FLOAT STORE (%d at 0x%08x(%%sp)) (from %d)\n", spill, store->inst_offset, reg));
1526 * Creates a load for spilled floating point items
1529 create_spilled_load_float (MonoCompile *cfg, int spill, int reg, MonoInst *ins)
1532 MONO_INST_NEW (cfg, load, OP_LOADR8_SPILL_MEMBASE);
1534 load->inst_basereg = X86_EBP;
1535 load->inst_offset = mono_spillvar_offset_float (cfg, spill);
1537 DEBUG (g_print ("SPILLED FLOAT LOAD (%d at 0x%08x(%%sp)) (from %d)\n", spill, load->inst_offset, reg));
1541 #define is_global_ireg(r) ((r) >= 0 && (r) < MONO_MAX_IREGS && !X86_IS_CALLEE ((r)))
1542 #define reg_is_freeable(r) ((r) >= 0 && (r) < MONO_MAX_IREGS && X86_IS_CALLEE ((r)))
1549 int flags; /* used to track fp spill/load */
1552 static const char*const * ins_spec = pentium_desc;
1555 print_ins (int i, MonoInst *ins)
1557 const char *spec = ins_spec [ins->opcode];
1558 g_print ("\t%-2d %s", i, mono_inst_name (ins->opcode));
1559 if (spec [MONO_INST_DEST]) {
1560 if (ins->dreg >= MONO_MAX_IREGS)
1561 g_print (" R%d <-", ins->dreg);
1563 g_print (" %s <-", mono_arch_regname (ins->dreg));
1565 if (spec [MONO_INST_SRC1]) {
1566 if (ins->sreg1 >= MONO_MAX_IREGS)
1567 g_print (" R%d", ins->sreg1);
1569 g_print (" %s", mono_arch_regname (ins->sreg1));
1571 if (spec [MONO_INST_SRC2]) {
1572 if (ins->sreg2 >= MONO_MAX_IREGS)
1573 g_print (" R%d", ins->sreg2);
1575 g_print (" %s", mono_arch_regname (ins->sreg2));
1577 if (spec [MONO_INST_CLOB])
1578 g_print (" clobbers: %c", spec [MONO_INST_CLOB]);
1583 print_regtrack (RegTrack *t, int num)
1589 for (i = 0; i < num; ++i) {
1592 if (i >= MONO_MAX_IREGS) {
1593 g_snprintf (buf, sizeof(buf), "R%d", i);
1596 r = mono_arch_regname (i);
1597 g_print ("liveness: %s [%d - %d]\n", r, t [i].born_in, t[i].last_use);
1601 typedef struct InstList InstList;
1609 static inline InstList*
1610 inst_list_prepend (MonoMemPool *pool, InstList *list, MonoInst *data)
1612 InstList *item = mono_mempool_alloc (pool, sizeof (InstList));
1622 * Force the spilling of the variable in the symbolic register 'reg'.
1625 get_register_force_spilling (MonoCompile *cfg, InstList *item, MonoInst *ins, int reg)
1630 sel = cfg->rs->iassign [reg];
1631 /*i = cfg->rs->isymbolic [sel];
1632 g_assert (i == reg);*/
1634 spill = ++cfg->spill_count;
1635 cfg->rs->iassign [i] = -spill - 1;
1636 mono_regstate_free_int (cfg->rs, sel);
1637 /* we need to create a spill var and insert a load to sel after the current instruction */
1638 MONO_INST_NEW (cfg, load, OP_LOAD_MEMBASE);
1640 load->inst_basereg = X86_EBP;
1641 load->inst_offset = mono_spillvar_offset (cfg, spill);
1643 while (ins->next != item->prev->data)
1646 load->next = ins->next;
1648 DEBUG (g_print ("SPILLED LOAD (%d at 0x%08x(%%ebp)) R%d (freed %s)\n", spill, load->inst_offset, i, mono_arch_regname (sel)));
1649 i = mono_regstate_alloc_int (cfg->rs, 1 << sel);
1650 g_assert (i == sel);
1656 get_register_spilling (MonoCompile *cfg, InstList *item, MonoInst *ins, guint32 regmask, int reg)
1661 DEBUG (g_print ("\tstart regmask to assign R%d: 0x%08x (R%d <- R%d R%d)\n", reg, regmask, ins->dreg, ins->sreg1, ins->sreg2));
1662 /* exclude the registers in the current instruction */
1663 if (reg != ins->sreg1 && (reg_is_freeable (ins->sreg1) || (ins->sreg1 >= MONO_MAX_IREGS && cfg->rs->iassign [ins->sreg1] >= 0))) {
1664 if (ins->sreg1 >= MONO_MAX_IREGS)
1665 regmask &= ~ (1 << cfg->rs->iassign [ins->sreg1]);
1667 regmask &= ~ (1 << ins->sreg1);
1668 DEBUG (g_print ("\t\texcluding sreg1 %s\n", mono_arch_regname (ins->sreg1)));
1670 if (reg != ins->sreg2 && (reg_is_freeable (ins->sreg2) || (ins->sreg2 >= MONO_MAX_IREGS && cfg->rs->iassign [ins->sreg2] >= 0))) {
1671 if (ins->sreg2 >= MONO_MAX_IREGS)
1672 regmask &= ~ (1 << cfg->rs->iassign [ins->sreg2]);
1674 regmask &= ~ (1 << ins->sreg2);
1675 DEBUG (g_print ("\t\texcluding sreg2 %s %d\n", mono_arch_regname (ins->sreg2), ins->sreg2));
1677 if (reg != ins->dreg && reg_is_freeable (ins->dreg)) {
1678 regmask &= ~ (1 << ins->dreg);
1679 DEBUG (g_print ("\t\texcluding dreg %s\n", mono_arch_regname (ins->dreg)));
1682 DEBUG (g_print ("\t\tavailable regmask: 0x%08x\n", regmask));
1683 g_assert (regmask); /* need at least a register we can free */
1685 /* we should track prev_use and spill the register that's farther */
1686 for (i = 0; i < MONO_MAX_IREGS; ++i) {
1687 if (regmask & (1 << i)) {
1689 DEBUG (g_print ("\t\tselected register %s has assignment %d\n", mono_arch_regname (sel), cfg->rs->iassign [sel]));
1693 i = cfg->rs->isymbolic [sel];
1694 spill = ++cfg->spill_count;
1695 cfg->rs->iassign [i] = -spill - 1;
1696 mono_regstate_free_int (cfg->rs, sel);
1697 /* we need to create a spill var and insert a load to sel after the current instruction */
1698 MONO_INST_NEW (cfg, load, OP_LOAD_MEMBASE);
1700 load->inst_basereg = X86_EBP;
1701 load->inst_offset = mono_spillvar_offset (cfg, spill);
1703 while (ins->next != item->prev->data)
1706 load->next = ins->next;
1708 DEBUG (g_print ("\tSPILLED LOAD (%d at 0x%08x(%%ebp)) R%d (freed %s)\n", spill, load->inst_offset, i, mono_arch_regname (sel)));
1709 i = mono_regstate_alloc_int (cfg->rs, 1 << sel);
1710 g_assert (i == sel);
1716 create_copy_ins (MonoCompile *cfg, int dest, int src, MonoInst *ins)
1719 MONO_INST_NEW (cfg, copy, OP_MOVE);
1723 copy->next = ins->next;
1726 DEBUG (g_print ("\tforced copy from %s to %s\n", mono_arch_regname (src), mono_arch_regname (dest)));
1731 create_spilled_store (MonoCompile *cfg, int spill, int reg, int prev_reg, MonoInst *ins)
1734 MONO_INST_NEW (cfg, store, OP_STORE_MEMBASE_REG);
1736 store->inst_destbasereg = X86_EBP;
1737 store->inst_offset = mono_spillvar_offset (cfg, spill);
1739 store->next = ins->next;
1742 DEBUG (g_print ("\tSPILLED STORE (%d at 0x%08x(%%ebp)) R%d (from %s)\n", spill, store->inst_offset, prev_reg, mono_arch_regname (reg)));
1747 insert_before_ins (MonoInst *ins, InstList *item, MonoInst* to_insert)
1751 prev = item->next->data;
1753 while (prev->next != ins)
1755 to_insert->next = ins;
1756 prev->next = to_insert;
1758 to_insert->next = ins;
1761 * needed otherwise in the next instruction we can add an ins to the
1762 * end and that would get past this instruction.
1764 item->data = to_insert;
1770 alloc_int_reg (MonoCompile *cfg, InstList *curinst, MonoInst *ins, int sym_reg, guint32 allow_mask)
1772 int val = cfg->rs->iassign [sym_reg];
1776 /* the register gets spilled after this inst */
1779 val = mono_regstate_alloc_int (cfg->rs, allow_mask);
1781 val = get_register_spilling (cfg, curinst, ins, allow_mask, sym_reg);
1782 cfg->rs->iassign [sym_reg] = val;
1783 /* add option to store before the instruction for src registers */
1785 create_spilled_store (cfg, spill, val, sym_reg, ins);
1787 cfg->rs->isymbolic [val] = sym_reg;
1792 /* flags used in reginfo->flags */
1794 MONO_X86_FP_NEEDS_LOAD_SPILL = 1 << 0,
1795 MONO_X86_FP_NEEDS_SPILL = 1 << 1,
1796 MONO_X86_FP_NEEDS_LOAD = 1 << 2,
1797 MONO_X86_REG_NOT_ECX = 1 << 3,
1798 MONO_X86_REG_EAX = 1 << 4,
1799 MONO_X86_REG_EDX = 1 << 5,
1800 MONO_X86_REG_ECX = 1 << 6
1804 mono_x86_alloc_int_reg (MonoCompile *cfg, InstList *tmp, MonoInst *ins, guint32 dest_mask, int sym_reg, int flags)
1807 int test_mask = dest_mask;
1809 if (flags & MONO_X86_REG_EAX)
1810 test_mask &= (1 << X86_EAX);
1811 else if (flags & MONO_X86_REG_EDX)
1812 test_mask &= (1 << X86_EDX);
1813 else if (flags & MONO_X86_REG_ECX)
1814 test_mask &= (1 << X86_ECX);
1815 else if (flags & MONO_X86_REG_NOT_ECX)
1816 test_mask &= ~ (1 << X86_ECX);
1818 val = mono_regstate_alloc_int (cfg->rs, test_mask);
1819 if (val >= 0 && test_mask != dest_mask)
1820 DEBUG(g_print ("\tUsed flag to allocate reg %s for R%u\n", mono_arch_regname (val), sym_reg));
1822 if (val < 0 && (flags & MONO_X86_REG_NOT_ECX)) {
1823 DEBUG(g_print ("\tFailed to allocate flag suggested mask (%u) but exluding ECX\n", test_mask));
1824 val = mono_regstate_alloc_int (cfg->rs, (dest_mask & (~1 << X86_ECX)));
1828 val = mono_regstate_alloc_int (cfg->rs, dest_mask);
1830 val = get_register_spilling (cfg, tmp, ins, dest_mask, sym_reg);
1837 assign_ireg (MonoRegState *rs, int reg, int hreg)
1839 g_assert (reg >= MONO_MAX_IREGS);
1840 g_assert (hreg < MONO_MAX_IREGS);
1841 g_assert (! is_global_ireg (hreg));
1843 rs->iassign [reg] = hreg;
1844 rs->isymbolic [hreg] = reg;
1845 rs->ifree_mask &= ~ (1 << hreg);
1848 /*#include "cprop.c"*/
1851 * Local register allocation.
1852 * We first scan the list of instructions and we save the liveness info of
1853 * each register (when the register is first used, when it's value is set etc.).
1854 * We also reverse the list of instructions (in the InstList list) because assigning
1855 * registers backwards allows for more tricks to be used.
1858 mono_arch_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
1861 MonoRegState *rs = cfg->rs;
1862 int i, val, fpcount;
1863 RegTrack *reginfo, *reginfof;
1864 RegTrack *reginfo1, *reginfo2, *reginfod;
1865 InstList *tmp, *reversed = NULL;
1867 guint32 src1_mask, src2_mask, dest_mask;
1868 GList *fspill_list = NULL;
1873 rs->next_vireg = bb->max_ireg;
1874 rs->next_vfreg = bb->max_freg;
1875 mono_regstate_assign (rs);
1876 reginfo = g_malloc0 (sizeof (RegTrack) * rs->next_vireg);
1877 reginfof = g_malloc0 (sizeof (RegTrack) * rs->next_vfreg);
1878 rs->ifree_mask = X86_CALLEE_REGS;
1882 /*if (cfg->opt & MONO_OPT_COPYPROP)
1883 local_copy_prop (cfg, ins);*/
1887 DEBUG (g_print ("LOCAL regalloc: basic block: %d\n", bb->block_num));
1888 /* forward pass on the instructions to collect register liveness info */
1890 spec = ins_spec [ins->opcode];
1892 DEBUG (print_ins (i, ins));
1894 if (spec [MONO_INST_SRC1]) {
1895 if (spec [MONO_INST_SRC1] == 'f') {
1897 reginfo1 = reginfof;
1899 spill = g_list_first (fspill_list);
1900 if (spill && fpcount < MONO_MAX_FREGS) {
1901 reginfo1 [ins->sreg1].flags |= MONO_X86_FP_NEEDS_LOAD;
1902 fspill_list = g_list_remove (fspill_list, spill->data);
1908 reginfo1 [ins->sreg1].prev_use = reginfo1 [ins->sreg1].last_use;
1909 reginfo1 [ins->sreg1].last_use = i;
1910 if (spec [MONO_INST_SRC1] == 'L') {
1911 /* The virtual register is allocated sequentially */
1912 reginfo1 [ins->sreg1 + 1].prev_use = reginfo1 [ins->sreg1 + 1].last_use;
1913 reginfo1 [ins->sreg1 + 1].last_use = i;
1914 if (reginfo1 [ins->sreg1 + 1].born_in == 0 || reginfo1 [ins->sreg1 + 1].born_in > i)
1915 reginfo1 [ins->sreg1 + 1].born_in = i;
1917 reginfo1 [ins->sreg1].flags |= MONO_X86_REG_EAX;
1918 reginfo1 [ins->sreg1 + 1].flags |= MONO_X86_REG_EDX;
1923 if (spec [MONO_INST_SRC2]) {
1924 if (spec [MONO_INST_SRC2] == 'f') {
1926 reginfo2 = reginfof;
1927 spill = g_list_first (fspill_list);
1929 reginfo2 [ins->sreg2].flags |= MONO_X86_FP_NEEDS_LOAD;
1930 fspill_list = g_list_remove (fspill_list, spill->data);
1931 if (fpcount >= MONO_MAX_FREGS) {
1933 fspill_list = g_list_prepend (fspill_list, GINT_TO_POINTER(fspill));
1934 reginfo2 [ins->sreg2].flags |= MONO_X86_FP_NEEDS_LOAD_SPILL;
1941 reginfo2 [ins->sreg2].prev_use = reginfo2 [ins->sreg2].last_use;
1942 reginfo2 [ins->sreg2].last_use = i;
1943 if (spec [MONO_INST_SRC2] == 'L') {
1944 /* The virtual register is allocated sequentially */
1945 reginfo2 [ins->sreg2 + 1].prev_use = reginfo2 [ins->sreg2 + 1].last_use;
1946 reginfo2 [ins->sreg2 + 1].last_use = i;
1947 if (reginfo2 [ins->sreg2 + 1].born_in == 0 || reginfo2 [ins->sreg2 + 1].born_in > i)
1948 reginfo2 [ins->sreg2 + 1].born_in = i;
1950 if (spec [MONO_INST_CLOB] == 's') {
1951 reginfo2 [ins->sreg1].flags |= MONO_X86_REG_NOT_ECX;
1952 reginfo2 [ins->sreg2].flags |= MONO_X86_REG_ECX;
1957 if (spec [MONO_INST_DEST]) {
1958 if (spec [MONO_INST_DEST] == 'f') {
1959 reginfod = reginfof;
1960 if (fpcount >= MONO_MAX_FREGS) {
1961 reginfod [ins->dreg].flags |= MONO_X86_FP_NEEDS_SPILL;
1963 fspill_list = g_list_prepend (fspill_list, GINT_TO_POINTER(fspill));
1970 if (spec [MONO_INST_DEST] != 'b') /* it's not just a base register */
1971 reginfod [ins->dreg].killed_in = i;
1972 reginfod [ins->dreg].prev_use = reginfod [ins->dreg].last_use;
1973 reginfod [ins->dreg].last_use = i;
1974 if (reginfod [ins->dreg].born_in == 0 || reginfod [ins->dreg].born_in > i)
1975 reginfod [ins->dreg].born_in = i;
1976 if (spec [MONO_INST_DEST] == 'l' || spec [MONO_INST_DEST] == 'L') {
1977 /* The virtual register is allocated sequentially */
1978 reginfod [ins->dreg + 1].prev_use = reginfod [ins->dreg + 1].last_use;
1979 reginfod [ins->dreg + 1].last_use = i;
1980 if (reginfod [ins->dreg + 1].born_in == 0 || reginfod [ins->dreg + 1].born_in > i)
1981 reginfod [ins->dreg + 1].born_in = i;
1983 reginfod [ins->dreg].flags |= MONO_X86_REG_EAX;
1984 reginfod [ins->dreg + 1].flags |= MONO_X86_REG_EDX;
1990 reversed = inst_list_prepend (cfg->mempool, reversed, ins);
1995 // todo: check if we have anything left on fp stack, in verify mode?
1998 DEBUG (print_regtrack (reginfo, rs->next_vireg));
1999 DEBUG (print_regtrack (reginfof, rs->next_vfreg));
2002 int prev_dreg, prev_sreg1, prev_sreg2, clob_dreg;
2003 dest_mask = src1_mask = src2_mask = X86_CALLEE_REGS;
2006 spec = ins_spec [ins->opcode];
2009 DEBUG (g_print ("processing:"));
2010 DEBUG (print_ins (i, ins));
2011 if (spec [MONO_INST_CLOB] == 's') {
2013 * Shift opcodes, SREG2 must be RCX
2015 if (rs->ifree_mask & (1 << X86_ECX)) {
2016 if (ins->sreg2 < MONO_MAX_IREGS) {
2017 /* Argument already in hard reg, need to copy */
2018 MonoInst *copy = create_copy_ins (cfg, X86_ECX, ins->sreg2, NULL);
2019 insert_before_ins (ins, tmp, copy);
2022 DEBUG (g_print ("\tshortcut assignment of R%d to ECX\n", ins->sreg2));
2023 assign_ireg (rs, ins->sreg2, X86_ECX);
2026 int need_ecx_spill = TRUE;
2028 * we first check if src1/dreg is already assigned a register
2029 * and then we force a spill of the var assigned to ECX.
2031 /* the destination register can't be ECX */
2032 dest_mask &= ~ (1 << X86_ECX);
2033 src1_mask &= ~ (1 << X86_ECX);
2034 val = rs->iassign [ins->dreg];
2036 * the destination register is already assigned to ECX:
2037 * we need to allocate another register for it and then
2038 * copy from this to ECX.
2040 if (val == X86_ECX && ins->dreg != ins->sreg2) {
2042 new_dest = mono_x86_alloc_int_reg (cfg, tmp, ins, dest_mask, ins->dreg, reginfo [ins->dreg].flags);
2043 g_assert (new_dest >= 0);
2044 DEBUG (g_print ("\tclob:s changing dreg R%d to %s from ECX\n", ins->dreg, mono_arch_regname (new_dest)));
2046 rs->isymbolic [new_dest] = ins->dreg;
2047 rs->iassign [ins->dreg] = new_dest;
2048 clob_dreg = ins->dreg;
2049 ins->dreg = new_dest;
2050 create_copy_ins (cfg, X86_ECX, new_dest, ins);
2051 need_ecx_spill = FALSE;
2052 /*DEBUG (g_print ("\tforced spill of R%d\n", ins->dreg));
2053 val = get_register_force_spilling (cfg, tmp, ins, ins->dreg);
2054 rs->iassign [ins->dreg] = val;
2055 rs->isymbolic [val] = prev_dreg;
2058 if (is_global_ireg (ins->sreg2)) {
2059 MonoInst *copy = create_copy_ins (cfg, X86_ECX, ins->sreg2, NULL);
2060 insert_before_ins (ins, tmp, copy);
2063 val = rs->iassign [ins->sreg2];
2064 if (val >= 0 && val != X86_ECX) {
2065 MonoInst *move = create_copy_ins (cfg, X86_ECX, val, NULL);
2066 DEBUG (g_print ("\tmoved arg from R%d (%d) to ECX\n", val, ins->sreg2));
2068 g_assert_not_reached ();
2069 /* FIXME: where is move connected to the instruction list? */
2070 //tmp->prev->data->next = move;
2074 need_ecx_spill = FALSE;
2077 if (need_ecx_spill && !(rs->ifree_mask & (1 << X86_ECX))) {
2078 DEBUG (g_print ("\tforced spill of R%d\n", rs->isymbolic [X86_ECX]));
2079 get_register_force_spilling (cfg, tmp, ins, rs->isymbolic [X86_ECX]);
2080 mono_regstate_free_int (rs, X86_ECX);
2082 if (!is_global_ireg (ins->sreg2))
2083 /* force-set sreg2 */
2084 assign_ireg (rs, ins->sreg2, X86_ECX);
2086 ins->sreg2 = X86_ECX;
2087 } else if (spec [MONO_INST_CLOB] == 'd') {
2091 int dest_reg = X86_EAX;
2092 int clob_reg = X86_EDX;
2093 if (spec [MONO_INST_DEST] == 'd') {
2094 dest_reg = X86_EDX; /* reminder */
2097 if (is_global_ireg (ins->dreg))
2100 val = rs->iassign [ins->dreg];
2101 if (0 && val >= 0 && val != dest_reg && !(rs->ifree_mask & (1 << dest_reg))) {
2102 DEBUG (g_print ("\tforced spill of R%d\n", rs->isymbolic [dest_reg]));
2103 get_register_force_spilling (cfg, tmp, ins, rs->isymbolic [dest_reg]);
2104 mono_regstate_free_int (rs, dest_reg);
2108 /* the register gets spilled after this inst */
2109 int spill = -val -1;
2110 dest_mask = 1 << dest_reg;
2111 prev_dreg = ins->dreg;
2112 val = mono_regstate_alloc_int (rs, dest_mask);
2114 val = get_register_spilling (cfg, tmp, ins, dest_mask, ins->dreg);
2115 rs->iassign [ins->dreg] = val;
2117 create_spilled_store (cfg, spill, val, prev_dreg, ins);
2118 DEBUG (g_print ("\tassigned dreg %s to dest R%d\n", mono_arch_regname (val), ins->dreg));
2119 rs->isymbolic [val] = prev_dreg;
2122 DEBUG (g_print ("\tshortcut assignment of R%d to %s\n", ins->dreg, mono_arch_regname (dest_reg)));
2123 prev_dreg = ins->dreg;
2124 assign_ireg (rs, ins->dreg, dest_reg);
2125 ins->dreg = dest_reg;
2130 //DEBUG (g_print ("dest reg in div assigned: %s\n", mono_arch_regname (val)));
2131 if (val != dest_reg) { /* force a copy */
2132 create_copy_ins (cfg, val, dest_reg, ins);
2133 if (!(rs->ifree_mask & (1 << dest_reg)) && rs->isymbolic [dest_reg] >= MONO_MAX_IREGS) {
2134 DEBUG (g_print ("\tforced spill of R%d\n", rs->isymbolic [dest_reg]));
2135 get_register_force_spilling (cfg, tmp, ins, rs->isymbolic [dest_reg]);
2136 mono_regstate_free_int (rs, dest_reg);
2139 if (!(rs->ifree_mask & (1 << clob_reg)) && (clob_reg != val) && (rs->isymbolic [clob_reg] >= 8)) {
2140 DEBUG (g_print ("\tforced spill of clobbered reg R%d\n", rs->isymbolic [clob_reg]));
2141 get_register_force_spilling (cfg, tmp, ins, rs->isymbolic [clob_reg]);
2142 mono_regstate_free_int (rs, clob_reg);
2144 src1_mask = 1 << X86_EAX;
2145 src2_mask = 1 << X86_ECX;
2146 } else if (spec [MONO_INST_DEST] == 'l') {
2148 val = rs->iassign [ins->dreg];
2149 /* check special case when dreg have been moved from ecx (clob shift) */
2150 if (spec [MONO_INST_CLOB] == 's' && clob_dreg != -1)
2151 hreg = clob_dreg + 1;
2153 hreg = ins->dreg + 1;
2155 /* base prev_dreg on fixed hreg, handle clob case */
2158 if (val != rs->isymbolic [X86_EAX] && !(rs->ifree_mask & (1 << X86_EAX))) {
2159 DEBUG (g_print ("\t(long-low) forced spill of R%d\n", rs->isymbolic [X86_EAX]));
2160 get_register_force_spilling (cfg, tmp, ins, rs->isymbolic [X86_EAX]);
2161 mono_regstate_free_int (rs, X86_EAX);
2163 if (hreg != rs->isymbolic [X86_EDX] && !(rs->ifree_mask & (1 << X86_EDX))) {
2164 DEBUG (g_print ("\t(long-high) forced spill of R%d\n", rs->isymbolic [X86_EDX]));
2165 get_register_force_spilling (cfg, tmp, ins, rs->isymbolic [X86_EDX]);
2166 mono_regstate_free_int (rs, X86_EDX);
2168 } else if (spec [MONO_INST_CLOB] == 'b') {
2170 * x86_set_reg instructions, dreg needs to be EAX..EDX
2172 dest_mask = (1 << X86_EAX) | (1 << X86_EBX) | (1 << X86_ECX) | (1 << X86_EDX);
2173 if ((ins->dreg < MONO_MAX_IREGS) && (! (dest_mask & (1 << ins->dreg)))) {
2175 * ins->dreg is already a hard reg, need to allocate another
2176 * suitable hard reg and make a copy.
2178 int new_dest = mono_x86_alloc_int_reg (cfg, tmp, ins, dest_mask, ins->dreg, reginfo [ins->dreg].flags);
2179 g_assert (new_dest >= 0);
2181 create_copy_ins (cfg, ins->dreg, new_dest, ins);
2182 DEBUG (g_print ("\tclob:b changing dreg R%d to %s\n", ins->dreg, mono_arch_regname (new_dest)));
2183 ins->dreg = new_dest;
2185 /* The hard reg is no longer needed */
2186 mono_regstate_free_int (rs, new_dest);
2193 if (spec [MONO_INST_DEST] == 'f') {
2194 if (reginfof [ins->dreg].flags & MONO_X86_FP_NEEDS_SPILL) {
2197 spill_node = g_list_first (fspill_list);
2198 g_assert (spill_node);
2200 store = create_spilled_store_float (cfg, GPOINTER_TO_INT (spill_node->data), ins->dreg, ins);
2201 insert_before_ins (ins, tmp, store);
2202 fspill_list = g_list_remove (fspill_list, spill_node->data);
2205 } else if (spec [MONO_INST_DEST] == 'L') {
2207 val = rs->iassign [ins->dreg];
2208 /* check special case when dreg have been moved from ecx (clob shift) */
2209 if (spec [MONO_INST_CLOB] == 's' && clob_dreg != -1)
2210 hreg = clob_dreg + 1;
2212 hreg = ins->dreg + 1;
2214 /* base prev_dreg on fixed hreg, handle clob case */
2215 prev_dreg = hreg - 1;
2220 /* the register gets spilled after this inst */
2223 val = mono_x86_alloc_int_reg (cfg, tmp, ins, dest_mask, ins->dreg, reginfo [ins->dreg].flags);
2224 rs->iassign [ins->dreg] = val;
2226 create_spilled_store (cfg, spill, val, prev_dreg, ins);
2229 DEBUG (g_print ("\tassigned dreg (long) %s to dest R%d\n", mono_arch_regname (val), hreg - 1));
2231 rs->isymbolic [val] = hreg - 1;
2234 val = rs->iassign [hreg];
2238 /* the register gets spilled after this inst */
2241 val = mono_x86_alloc_int_reg (cfg, tmp, ins, dest_mask, hreg, reginfo [hreg].flags);
2242 rs->iassign [hreg] = val;
2244 create_spilled_store (cfg, spill, val, hreg, ins);
2247 DEBUG (g_print ("\tassigned hreg (long-high) %s to dest R%d\n", mono_arch_regname (val), hreg));
2248 rs->isymbolic [val] = hreg;
2249 /* save reg allocating into unused */
2252 /* check if we can free our long reg */
2253 if (reg_is_freeable (val) && hreg >= 0 && reginfo [hreg].born_in >= i) {
2254 DEBUG (g_print ("\tfreeable %s (R%d) (born in %d)\n", mono_arch_regname (val), hreg, reginfo [hreg].born_in));
2255 mono_regstate_free_int (rs, val);
2258 else if (ins->dreg >= MONO_MAX_IREGS) {
2260 val = rs->iassign [ins->dreg];
2261 if (spec [MONO_INST_DEST] == 'l') {
2262 /* check special case when dreg have been moved from ecx (clob shift) */
2263 if (spec [MONO_INST_CLOB] == 's' && clob_dreg != -1)
2264 hreg = clob_dreg + 1;
2266 hreg = ins->dreg + 1;
2268 /* base prev_dreg on fixed hreg, handle clob case */
2269 prev_dreg = hreg - 1;
2271 prev_dreg = ins->dreg;
2276 /* the register gets spilled after this inst */
2279 val = mono_x86_alloc_int_reg (cfg, tmp, ins, dest_mask, ins->dreg, reginfo [ins->dreg].flags);
2280 rs->iassign [ins->dreg] = val;
2282 create_spilled_store (cfg, spill, val, prev_dreg, ins);
2284 DEBUG (g_print ("\tassigned dreg %s to dest R%d\n", mono_arch_regname (val), ins->dreg));
2285 rs->isymbolic [val] = prev_dreg;
2287 /* handle cases where lreg needs to be eax:edx */
2288 if (spec [MONO_INST_DEST] == 'l') {
2289 /* check special case when dreg have been moved from ecx (clob shift) */
2290 int hreg = prev_dreg + 1;
2291 val = rs->iassign [hreg];
2295 /* the register gets spilled after this inst */
2298 val = mono_x86_alloc_int_reg (cfg, tmp, ins, dest_mask, hreg, reginfo [hreg].flags);
2299 rs->iassign [hreg] = val;
2301 create_spilled_store (cfg, spill, val, hreg, ins);
2303 DEBUG (g_print ("\tassigned hreg %s to dest R%d\n", mono_arch_regname (val), hreg));
2304 rs->isymbolic [val] = hreg;
2305 if (ins->dreg == X86_EAX) {
2307 create_copy_ins (cfg, val, X86_EDX, ins);
2308 } else if (ins->dreg == X86_EDX) {
2309 if (val == X86_EAX) {
2311 g_assert_not_reached ();
2313 /* two forced copies */
2314 create_copy_ins (cfg, val, X86_EDX, ins);
2315 create_copy_ins (cfg, ins->dreg, X86_EAX, ins);
2318 if (val == X86_EDX) {
2319 create_copy_ins (cfg, ins->dreg, X86_EAX, ins);
2321 /* two forced copies */
2322 create_copy_ins (cfg, val, X86_EDX, ins);
2323 create_copy_ins (cfg, ins->dreg, X86_EAX, ins);
2326 if (reg_is_freeable (val) && hreg >= 0 && reginfo [hreg].born_in >= i) {
2327 DEBUG (g_print ("\tfreeable %s (R%d)\n", mono_arch_regname (val), hreg));
2328 mono_regstate_free_int (rs, val);
2330 } else if (spec [MONO_INST_DEST] == 'a' && ins->dreg != X86_EAX && spec [MONO_INST_CLOB] != 'd') {
2331 /* this instruction only outputs to EAX, need to copy */
2332 create_copy_ins (cfg, ins->dreg, X86_EAX, ins);
2333 } else if (spec [MONO_INST_DEST] == 'd' && ins->dreg != X86_EDX && spec [MONO_INST_CLOB] != 'd') {
2334 create_copy_ins (cfg, ins->dreg, X86_EDX, ins);
2337 if (spec [MONO_INST_DEST] != 'f' && reg_is_freeable (ins->dreg) && prev_dreg >= 0 && reginfo [prev_dreg].born_in >= i) {
2338 DEBUG (g_print ("\tfreeable %s (R%d) (born in %d)\n", mono_arch_regname (ins->dreg), prev_dreg, reginfo [prev_dreg].born_in));
2339 mono_regstate_free_int (rs, ins->dreg);
2341 /* put src1 in EAX if it needs to be */
2342 if (spec [MONO_INST_SRC1] == 'a') {
2343 if (!(rs->ifree_mask & (1 << X86_EAX))) {
2344 DEBUG (g_print ("\tforced spill of R%d\n", rs->isymbolic [X86_EAX]));
2345 get_register_force_spilling (cfg, tmp, ins, rs->isymbolic [X86_EAX]);
2346 mono_regstate_free_int (rs, X86_EAX);
2348 if (ins->sreg1 < MONO_MAX_IREGS) {
2349 /* The argument is already in a hard reg, need to copy */
2350 MonoInst *copy = create_copy_ins (cfg, X86_EAX, ins->sreg1, NULL);
2351 insert_before_ins (ins, tmp, copy);
2354 /* force-set sreg1 */
2355 assign_ireg (rs, ins->sreg1, X86_EAX);
2356 ins->sreg1 = X86_EAX;
2362 if (spec [MONO_INST_SRC1] == 'f') {
2363 if (reginfof [ins->sreg1].flags & MONO_X86_FP_NEEDS_LOAD) {
2365 MonoInst *store = NULL;
2367 if (reginfof [ins->sreg1].flags & MONO_X86_FP_NEEDS_LOAD_SPILL) {
2369 spill_node = g_list_first (fspill_list);
2370 g_assert (spill_node);
2372 store = create_spilled_store_float (cfg, GPOINTER_TO_INT (spill_node->data), ins->sreg1, ins);
2373 fspill_list = g_list_remove (fspill_list, spill_node->data);
2377 fspill_list = g_list_prepend (fspill_list, GINT_TO_POINTER(fspill));
2378 load = create_spilled_load_float (cfg, fspill, ins->sreg1, ins);
2379 insert_before_ins (ins, tmp, load);
2381 insert_before_ins (load, tmp, store);
2383 } else if ((spec [MONO_INST_DEST] == 'L') && (spec [MONO_INST_SRC1] == 'L')) {
2384 /* force source to be same as dest */
2385 assign_ireg (rs, ins->sreg1, ins->dreg);
2386 assign_ireg (rs, ins->sreg1 + 1, ins->unused);
2388 DEBUG (g_print ("\tassigned sreg1 (long) %s to sreg1 R%d\n", mono_arch_regname (ins->dreg), ins->sreg1));
2389 DEBUG (g_print ("\tassigned sreg1 (long-high) %s to sreg1 R%d\n", mono_arch_regname (ins->unused), ins->sreg1 + 1));
2391 ins->sreg1 = ins->dreg;
2393 * No need for saving the reg, we know that src1=dest in this cases
2394 * ins->inst_c0 = ins->unused;
2397 else if (ins->sreg1 >= MONO_MAX_IREGS) {
2398 val = rs->iassign [ins->sreg1];
2399 prev_sreg1 = ins->sreg1;
2403 /* the register gets spilled after this inst */
2406 if (0 && ins->opcode == OP_MOVE) {
2408 * small optimization: the dest register is already allocated
2409 * but the src one is not: we can simply assign the same register
2410 * here and peephole will get rid of the instruction later.
2411 * This optimization may interfere with the clobbering handling:
2412 * it removes a mov operation that will be added again to handle clobbering.
2413 * There are also some other issues that should with make testjit.
2415 mono_regstate_alloc_int (rs, 1 << ins->dreg);
2416 val = rs->iassign [ins->sreg1] = ins->dreg;
2417 //g_assert (val >= 0);
2418 DEBUG (g_print ("\tfast assigned sreg1 %s to R%d\n", mono_arch_regname (val), ins->sreg1));
2420 //g_assert (val == -1); /* source cannot be spilled */
2421 val = mono_x86_alloc_int_reg (cfg, tmp, ins, src1_mask, ins->sreg1, reginfo [ins->sreg1].flags);
2422 rs->iassign [ins->sreg1] = val;
2423 DEBUG (g_print ("\tassigned sreg1 %s to R%d\n", mono_arch_regname (val), ins->sreg1));
2426 MonoInst *store = create_spilled_store (cfg, spill, val, prev_sreg1, NULL);
2427 insert_before_ins (ins, tmp, store);
2430 rs->isymbolic [val] = prev_sreg1;
2435 /* handle clobbering of sreg1 */
2436 if ((spec [MONO_INST_CLOB] == '1' || spec [MONO_INST_CLOB] == 's') && ins->dreg != ins->sreg1) {
2437 MonoInst *sreg2_copy = NULL;
2438 MonoInst *copy = NULL;
2440 if (ins->dreg == ins->sreg2) {
2442 * copying sreg1 to dreg could clobber sreg2, so allocate a new
2447 reg2 = mono_x86_alloc_int_reg (cfg, tmp, ins, dest_mask, ins->sreg2, 0);
2449 DEBUG (g_print ("\tneed to copy sreg2 %s to reg %s\n", mono_arch_regname (ins->sreg2), mono_arch_regname (reg2)));
2450 sreg2_copy = create_copy_ins (cfg, reg2, ins->sreg2, NULL);
2451 prev_sreg2 = ins->sreg2 = reg2;
2453 mono_regstate_free_int (rs, reg2);
2456 copy = create_copy_ins (cfg, ins->dreg, ins->sreg1, NULL);
2457 DEBUG (g_print ("\tneed to copy sreg1 %s to dreg %s\n", mono_arch_regname (ins->sreg1), mono_arch_regname (ins->dreg)));
2458 insert_before_ins (ins, tmp, copy);
2461 insert_before_ins (copy, tmp, sreg2_copy);
2464 * Need to prevent sreg2 to be allocated to sreg1, since that
2465 * would screw up the previous copy.
2467 src2_mask &= ~ (1 << ins->sreg1);
2468 /* we set sreg1 to dest as well */
2469 prev_sreg1 = ins->sreg1 = ins->dreg;
2470 src2_mask &= ~ (1 << ins->dreg);
2476 if (spec [MONO_INST_SRC2] == 'f') {
2477 if (reginfof [ins->sreg2].flags & MONO_X86_FP_NEEDS_LOAD) {
2479 MonoInst *store = NULL;
2481 if (reginfof [ins->sreg2].flags & MONO_X86_FP_NEEDS_LOAD_SPILL) {
2484 spill_node = g_list_first (fspill_list);
2485 g_assert (spill_node);
2486 if (spec [MONO_INST_SRC1] == 'f' && (reginfof [ins->sreg1].flags & MONO_X86_FP_NEEDS_LOAD_SPILL))
2487 spill_node = g_list_next (spill_node);
2489 store = create_spilled_store_float (cfg, GPOINTER_TO_INT (spill_node->data), ins->sreg2, ins);
2490 fspill_list = g_list_remove (fspill_list, spill_node->data);
2494 fspill_list = g_list_prepend (fspill_list, GINT_TO_POINTER(fspill));
2495 load = create_spilled_load_float (cfg, fspill, ins->sreg2, ins);
2496 insert_before_ins (ins, tmp, load);
2498 insert_before_ins (load, tmp, store);
2501 else if (ins->sreg2 >= MONO_MAX_IREGS) {
2502 val = rs->iassign [ins->sreg2];
2503 prev_sreg2 = ins->sreg2;
2507 /* the register gets spilled after this inst */
2510 val = mono_x86_alloc_int_reg (cfg, tmp, ins, src2_mask, ins->sreg2, reginfo [ins->sreg2].flags);
2511 rs->iassign [ins->sreg2] = val;
2512 DEBUG (g_print ("\tassigned sreg2 %s to R%d\n", mono_arch_regname (val), ins->sreg2));
2514 create_spilled_store (cfg, spill, val, prev_sreg2, ins);
2516 rs->isymbolic [val] = prev_sreg2;
2518 if (spec [MONO_INST_CLOB] == 's' && ins->sreg2 != X86_ECX) {
2519 DEBUG (g_print ("\tassigned sreg2 %s to R%d, but ECX is needed (R%d)\n", mono_arch_regname (val), ins->sreg2, rs->iassign [X86_ECX]));
2525 if (spec [MONO_INST_CLOB] == 'c') {
2527 guint32 clob_mask = X86_CALLEE_REGS;
2528 for (j = 0; j < MONO_MAX_IREGS; ++j) {
2530 if ((clob_mask & s) && !(rs->ifree_mask & s) && j != ins->sreg1) {
2531 //g_warning ("register %s busy at call site\n", mono_arch_regname (j));
2535 if (spec [MONO_INST_CLOB] == 'a') {
2536 guint32 clob_reg = X86_EAX;
2537 if (!(rs->ifree_mask & (1 << clob_reg)) && (rs->isymbolic [clob_reg] >= 8)) {
2538 DEBUG (g_print ("\tforced spill of clobbered reg R%d\n", rs->isymbolic [clob_reg]));
2539 get_register_force_spilling (cfg, tmp, ins, rs->isymbolic [clob_reg]);
2540 mono_regstate_free_int (rs, clob_reg);
2543 /*if (reg_is_freeable (ins->sreg1) && prev_sreg1 >= 0 && reginfo [prev_sreg1].born_in >= i) {
2544 DEBUG (g_print ("freeable %s\n", mono_arch_regname (ins->sreg1)));
2545 mono_regstate_free_int (rs, ins->sreg1);
2547 if (reg_is_freeable (ins->sreg2) && prev_sreg2 >= 0 && reginfo [prev_sreg2].born_in >= i) {
2548 DEBUG (g_print ("freeable %s\n", mono_arch_regname (ins->sreg2)));
2549 mono_regstate_free_int (rs, ins->sreg2);
2552 //DEBUG (print_ins (i, ins));
2553 /* this may result from a insert_before call */
2555 bb->code = tmp->data;
2561 g_list_free (fspill_list);
2564 static unsigned char*
2565 emit_float_to_int (MonoCompile *cfg, guchar *code, int dreg, int size, gboolean is_signed)
2567 x86_alu_reg_imm (code, X86_SUB, X86_ESP, 4);
2568 x86_fnstcw_membase(code, X86_ESP, 0);
2569 x86_mov_reg_membase (code, dreg, X86_ESP, 0, 2);
2570 x86_alu_reg_imm (code, X86_OR, dreg, 0xc00);
2571 x86_mov_membase_reg (code, X86_ESP, 2, dreg, 2);
2572 x86_fldcw_membase (code, X86_ESP, 2);
2574 x86_alu_reg_imm (code, X86_SUB, X86_ESP, 8);
2575 x86_fist_pop_membase (code, X86_ESP, 0, TRUE);
2576 x86_pop_reg (code, dreg);
2577 /* FIXME: need the high register
2578 * x86_pop_reg (code, dreg_high);
2581 x86_push_reg (code, X86_EAX); // SP = SP - 4
2582 x86_fist_pop_membase (code, X86_ESP, 0, FALSE);
2583 x86_pop_reg (code, dreg);
2585 x86_fldcw_membase (code, X86_ESP, 0);
2586 x86_alu_reg_imm (code, X86_ADD, X86_ESP, 4);
2589 x86_widen_reg (code, dreg, dreg, is_signed, FALSE);
2591 x86_widen_reg (code, dreg, dreg, is_signed, TRUE);
2595 static unsigned char*
2596 mono_emit_stack_alloc (guchar *code, MonoInst* tree)
2598 int sreg = tree->sreg1;
2599 int need_touch = FALSE;
2601 #if defined(PLATFORM_WIN32) || defined(MONO_ARCH_SIGSEGV_ON_ALTSTACK)
2610 * If requested stack size is larger than one page,
2611 * perform stack-touch operation
2614 * Generate stack probe code.
2615 * Under Windows, it is necessary to allocate one page at a time,
2616 * "touching" stack after each successful sub-allocation. This is
2617 * because of the way stack growth is implemented - there is a
2618 * guard page before the lowest stack page that is currently commited.
2619 * Stack normally grows sequentially so OS traps access to the
2620 * guard page and commits more pages when needed.
2622 x86_test_reg_imm (code, sreg, ~0xFFF);
2623 br[0] = code; x86_branch8 (code, X86_CC_Z, 0, FALSE);
2625 br[2] = code; /* loop */
2626 x86_alu_reg_imm (code, X86_SUB, X86_ESP, 0x1000);
2627 x86_test_membase_reg (code, X86_ESP, 0, X86_ESP);
2630 * By the end of the loop, sreg2 is smaller than 0x1000, so the init routine
2631 * that follows only initializes the last part of the area.
2633 /* Same as the init code below with size==0x1000 */
2634 if (tree->flags & MONO_INST_INIT) {
2635 x86_push_reg (code, X86_EAX);
2636 x86_push_reg (code, X86_ECX);
2637 x86_push_reg (code, X86_EDI);
2638 x86_mov_reg_imm (code, X86_ECX, (0x1000 >> 2));
2639 x86_alu_reg_reg (code, X86_XOR, X86_EAX, X86_EAX);
2640 x86_lea_membase (code, X86_EDI, X86_ESP, 12);
2642 x86_prefix (code, X86_REP_PREFIX);
2644 x86_pop_reg (code, X86_EDI);
2645 x86_pop_reg (code, X86_ECX);
2646 x86_pop_reg (code, X86_EAX);
2649 x86_alu_reg_imm (code, X86_SUB, sreg, 0x1000);
2650 x86_alu_reg_imm (code, X86_CMP, sreg, 0x1000);
2651 br[3] = code; x86_branch8 (code, X86_CC_AE, 0, FALSE);
2652 x86_patch (br[3], br[2]);
2653 x86_test_reg_reg (code, sreg, sreg);
2654 br[4] = code; x86_branch8 (code, X86_CC_Z, 0, FALSE);
2655 x86_alu_reg_reg (code, X86_SUB, X86_ESP, sreg);
2657 br[1] = code; x86_jump8 (code, 0);
2659 x86_patch (br[0], code);
2660 x86_alu_reg_reg (code, X86_SUB, X86_ESP, sreg);
2661 x86_patch (br[1], code);
2662 x86_patch (br[4], code);
2665 x86_alu_reg_reg (code, X86_SUB, X86_ESP, tree->sreg1);
2667 if (tree->flags & MONO_INST_INIT) {
2669 if (tree->dreg != X86_EAX && sreg != X86_EAX) {
2670 x86_push_reg (code, X86_EAX);
2673 if (tree->dreg != X86_ECX && sreg != X86_ECX) {
2674 x86_push_reg (code, X86_ECX);
2677 if (tree->dreg != X86_EDI && sreg != X86_EDI) {
2678 x86_push_reg (code, X86_EDI);
2682 x86_shift_reg_imm (code, X86_SHR, sreg, 2);
2683 if (sreg != X86_ECX)
2684 x86_mov_reg_reg (code, X86_ECX, sreg, 4);
2685 x86_alu_reg_reg (code, X86_XOR, X86_EAX, X86_EAX);
2687 x86_lea_membase (code, X86_EDI, X86_ESP, offset);
2689 x86_prefix (code, X86_REP_PREFIX);
2692 if (tree->dreg != X86_EDI && sreg != X86_EDI)
2693 x86_pop_reg (code, X86_EDI);
2694 if (tree->dreg != X86_ECX && sreg != X86_ECX)
2695 x86_pop_reg (code, X86_ECX);
2696 if (tree->dreg != X86_EAX && sreg != X86_EAX)
2697 x86_pop_reg (code, X86_EAX);
2704 emit_move_return_value (MonoCompile *cfg, MonoInst *ins, guint8 *code)
2709 /* Move return value to the target register */
2710 switch (ins->opcode) {
2713 case OP_CALL_MEMBASE:
2714 if (ins->dreg != X86_EAX)
2715 x86_mov_reg_reg (code, ins->dreg, X86_EAX, 4);
2719 case OP_VCALL_MEMBASE:
2720 cinfo = get_call_info (((MonoCallInst*)ins)->signature, FALSE);
2721 if (cinfo->ret.storage == ArgValuetypeInReg) {
2722 /* Pop the destination address from the stack */
2723 x86_pop_reg (code, X86_ECX);
2725 for (quad = 0; quad < 2; quad ++) {
2726 switch (cinfo->ret.pair_storage [quad]) {
2728 g_assert (cinfo->ret.pair_regs [quad] != X86_ECX);
2729 x86_mov_membase_reg (code, X86_ECX, (quad * sizeof (gpointer)), cinfo->ret.pair_regs [quad], sizeof (gpointer));
2734 g_assert_not_reached ();
2747 emit_tls_get (guint8* code, int dreg, int tls_offset)
2749 #ifdef PLATFORM_WIN32
2751 * See the Under the Hood article in the May 1996 issue of Microsoft Systems
2752 * Journal and/or a disassembly of the TlsGet () function.
2754 g_assert (tls_offset < 64);
2755 x86_prefix (code, X86_FS_PREFIX);
2756 x86_mov_reg_mem (code, dreg, 0x18, 4);
2757 /* Dunno what this does but TlsGetValue () contains it */
2758 x86_alu_membase_imm (code, X86_AND, dreg, 0x34, 0);
2759 x86_mov_reg_membase (code, dreg, dreg, 3600 + (tls_offset * 4), 4);
2761 x86_prefix (code, X86_GS_PREFIX);
2762 x86_mov_reg_mem (code, dreg, tls_offset, 4);
2767 #define REAL_PRINT_REG(text,reg) \
2768 mono_assert (reg >= 0); \
2769 x86_push_reg (code, X86_EAX); \
2770 x86_push_reg (code, X86_EDX); \
2771 x86_push_reg (code, X86_ECX); \
2772 x86_push_reg (code, reg); \
2773 x86_push_imm (code, reg); \
2774 x86_push_imm (code, text " %d %p\n"); \
2775 x86_mov_reg_imm (code, X86_EAX, printf); \
2776 x86_call_reg (code, X86_EAX); \
2777 x86_alu_reg_imm (code, X86_ADD, X86_ESP, 3*4); \
2778 x86_pop_reg (code, X86_ECX); \
2779 x86_pop_reg (code, X86_EDX); \
2780 x86_pop_reg (code, X86_EAX);
2782 /* benchmark and set based on cpu */
2783 #define LOOP_ALIGNMENT 8
2784 #define bb_is_loop_start(bb) ((bb)->loop_body_start && (bb)->nesting)
2787 mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
2792 guint8 *code = cfg->native_code + cfg->code_len;
2793 MonoInst *last_ins = NULL;
2794 guint last_offset = 0;
2797 if (cfg->opt & MONO_OPT_PEEPHOLE)
2798 peephole_pass (cfg, bb);
2800 if (cfg->opt & MONO_OPT_LOOP) {
2801 int pad, align = LOOP_ALIGNMENT;
2802 /* set alignment depending on cpu */
2803 if (bb_is_loop_start (bb) && (pad = (cfg->code_len & (align - 1)))) {
2805 /*g_print ("adding %d pad at %x to loop in %s\n", pad, cfg->code_len, cfg->method->name);*/
2806 x86_padding (code, pad);
2807 cfg->code_len += pad;
2808 bb->native_offset = cfg->code_len;
2812 if (cfg->verbose_level > 2)
2813 g_print ("Basic block %d starting at offset 0x%x\n", bb->block_num, bb->native_offset);
2815 cpos = bb->max_offset;
2817 if (cfg->prof_options & MONO_PROFILE_COVERAGE) {
2818 MonoProfileCoverageInfo *cov = cfg->coverage_info;
2819 g_assert (!cfg->compile_aot);
2822 cov->data [bb->dfn].cil_code = bb->cil_code;
2823 /* this is not thread save, but good enough */
2824 x86_inc_mem (code, &cov->data [bb->dfn].count);
2827 offset = code - cfg->native_code;
2831 offset = code - cfg->native_code;
2833 max_len = ((guint8 *)ins_spec [ins->opcode])[MONO_INST_LEN];
2835 if (offset > (cfg->code_size - max_len - 16)) {
2836 cfg->code_size *= 2;
2837 cfg->native_code = g_realloc (cfg->native_code, cfg->code_size);
2838 code = cfg->native_code + offset;
2839 mono_jit_stats.code_reallocs++;
2842 mono_debug_record_line_number (cfg, ins, offset);
2844 switch (ins->opcode) {
2846 x86_mul_reg (code, ins->sreg2, TRUE);
2849 x86_mul_reg (code, ins->sreg2, FALSE);
2851 case OP_X86_SETEQ_MEMBASE:
2852 case OP_X86_SETNE_MEMBASE:
2853 x86_set_membase (code, ins->opcode == OP_X86_SETEQ_MEMBASE ? X86_CC_EQ : X86_CC_NE,
2854 ins->inst_basereg, ins->inst_offset, TRUE);
2856 case OP_STOREI1_MEMBASE_IMM:
2857 x86_mov_membase_imm (code, ins->inst_destbasereg, ins->inst_offset, ins->inst_imm, 1);
2859 case OP_STOREI2_MEMBASE_IMM:
2860 x86_mov_membase_imm (code, ins->inst_destbasereg, ins->inst_offset, ins->inst_imm, 2);
2862 case OP_STORE_MEMBASE_IMM:
2863 case OP_STOREI4_MEMBASE_IMM:
2864 x86_mov_membase_imm (code, ins->inst_destbasereg, ins->inst_offset, ins->inst_imm, 4);
2866 case OP_STOREI1_MEMBASE_REG:
2867 x86_mov_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, ins->sreg1, 1);
2869 case OP_STOREI2_MEMBASE_REG:
2870 x86_mov_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, ins->sreg1, 2);
2872 case OP_STORE_MEMBASE_REG:
2873 case OP_STOREI4_MEMBASE_REG:
2874 x86_mov_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, ins->sreg1, 4);
2879 x86_mov_reg_mem (code, ins->dreg, ins->inst_p0, 4);
2882 x86_mov_reg_imm (code, ins->dreg, ins->inst_p0);
2883 x86_mov_reg_membase (code, ins->dreg, ins->dreg, 0, 4);
2885 case OP_LOAD_MEMBASE:
2886 case OP_LOADI4_MEMBASE:
2887 case OP_LOADU4_MEMBASE:
2888 x86_mov_reg_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, 4);
2890 case OP_LOADU1_MEMBASE:
2891 x86_widen_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, FALSE, FALSE);
2893 case OP_LOADI1_MEMBASE:
2894 x86_widen_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, TRUE, FALSE);
2896 case OP_LOADU2_MEMBASE:
2897 x86_widen_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, FALSE, TRUE);
2899 case OP_LOADI2_MEMBASE:
2900 x86_widen_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, TRUE, TRUE);
2903 x86_widen_reg (code, ins->dreg, ins->sreg1, TRUE, FALSE);
2906 x86_widen_reg (code, ins->dreg, ins->sreg1, TRUE, TRUE);
2909 x86_widen_reg (code, ins->dreg, ins->sreg1, FALSE, FALSE);
2912 x86_widen_reg (code, ins->dreg, ins->sreg1, FALSE, TRUE);
2915 x86_alu_reg_reg (code, X86_CMP, ins->sreg1, ins->sreg2);
2917 case OP_COMPARE_IMM:
2918 x86_alu_reg_imm (code, X86_CMP, ins->sreg1, ins->inst_imm);
2920 case OP_X86_COMPARE_MEMBASE_REG:
2921 x86_alu_membase_reg (code, X86_CMP, ins->inst_basereg, ins->inst_offset, ins->sreg2);
2923 case OP_X86_COMPARE_MEMBASE_IMM:
2924 x86_alu_membase_imm (code, X86_CMP, ins->inst_basereg, ins->inst_offset, ins->inst_imm);
2926 case OP_X86_COMPARE_MEMBASE8_IMM:
2927 x86_alu_membase8_imm (code, X86_CMP, ins->inst_basereg, ins->inst_offset, ins->inst_imm);
2929 case OP_X86_COMPARE_REG_MEMBASE:
2930 x86_alu_reg_membase (code, X86_CMP, ins->sreg1, ins->sreg2, ins->inst_offset);
2932 case OP_X86_COMPARE_MEM_IMM:
2933 x86_alu_mem_imm (code, X86_CMP, ins->inst_offset, ins->inst_imm);
2935 case OP_X86_TEST_NULL:
2936 x86_test_reg_reg (code, ins->sreg1, ins->sreg1);
2938 case OP_X86_ADD_MEMBASE_IMM:
2939 x86_alu_membase_imm (code, X86_ADD, ins->inst_basereg, ins->inst_offset, ins->inst_imm);
2941 case OP_X86_ADD_MEMBASE:
2942 x86_alu_reg_membase (code, X86_ADD, ins->sreg1, ins->sreg2, ins->inst_offset);
2944 case OP_X86_SUB_MEMBASE_IMM:
2945 x86_alu_membase_imm (code, X86_SUB, ins->inst_basereg, ins->inst_offset, ins->inst_imm);
2947 case OP_X86_SUB_MEMBASE:
2948 x86_alu_reg_membase (code, X86_SUB, ins->sreg1, ins->sreg2, ins->inst_offset);
2950 case OP_X86_INC_MEMBASE:
2951 x86_inc_membase (code, ins->inst_basereg, ins->inst_offset);
2953 case OP_X86_INC_REG:
2954 x86_inc_reg (code, ins->dreg);
2956 case OP_X86_DEC_MEMBASE:
2957 x86_dec_membase (code, ins->inst_basereg, ins->inst_offset);
2959 case OP_X86_DEC_REG:
2960 x86_dec_reg (code, ins->dreg);
2962 case OP_X86_MUL_MEMBASE:
2963 x86_imul_reg_membase (code, ins->sreg1, ins->sreg2, ins->inst_offset);
2966 x86_breakpoint (code);
2970 x86_alu_reg_reg (code, X86_ADD, ins->sreg1, ins->sreg2);
2973 x86_alu_reg_reg (code, X86_ADC, ins->sreg1, ins->sreg2);
2977 x86_alu_reg_imm (code, X86_ADD, ins->dreg, ins->inst_imm);
2980 x86_alu_reg_imm (code, X86_ADC, ins->dreg, ins->inst_imm);
2984 x86_alu_reg_reg (code, X86_SUB, ins->sreg1, ins->sreg2);
2987 x86_alu_reg_reg (code, X86_SBB, ins->sreg1, ins->sreg2);
2991 x86_alu_reg_imm (code, X86_SUB, ins->dreg, ins->inst_imm);
2994 x86_alu_reg_imm (code, X86_SBB, ins->dreg, ins->inst_imm);
2997 x86_alu_reg_reg (code, X86_AND, ins->sreg1, ins->sreg2);
3000 x86_alu_reg_imm (code, X86_AND, ins->sreg1, ins->inst_imm);
3004 x86_div_reg (code, ins->sreg2, TRUE);
3007 x86_alu_reg_reg (code, X86_XOR, X86_EDX, X86_EDX);
3008 x86_div_reg (code, ins->sreg2, FALSE);
3011 x86_mov_reg_imm (code, ins->sreg2, ins->inst_imm);
3013 x86_div_reg (code, ins->sreg2, TRUE);
3017 x86_div_reg (code, ins->sreg2, TRUE);
3020 x86_alu_reg_reg (code, X86_XOR, X86_EDX, X86_EDX);
3021 x86_div_reg (code, ins->sreg2, FALSE);
3024 x86_mov_reg_imm (code, ins->sreg2, ins->inst_imm);
3026 x86_div_reg (code, ins->sreg2, TRUE);
3029 x86_alu_reg_reg (code, X86_OR, ins->sreg1, ins->sreg2);
3032 x86_alu_reg_imm (code, X86_OR, ins->sreg1, ins->inst_imm);
3035 x86_alu_reg_reg (code, X86_XOR, ins->sreg1, ins->sreg2);
3038 x86_alu_reg_imm (code, X86_XOR, ins->sreg1, ins->inst_imm);
3041 g_assert (ins->sreg2 == X86_ECX);
3042 x86_shift_reg (code, X86_SHL, ins->dreg);
3045 g_assert (ins->sreg2 == X86_ECX);
3046 x86_shift_reg (code, X86_SAR, ins->dreg);
3049 x86_shift_reg_imm (code, X86_SAR, ins->dreg, ins->inst_imm);
3052 x86_shift_reg_imm (code, X86_SHR, ins->dreg, ins->inst_imm);
3055 g_assert (ins->sreg2 == X86_ECX);
3056 x86_shift_reg (code, X86_SHR, ins->dreg);
3059 x86_shift_reg_imm (code, X86_SHL, ins->dreg, ins->inst_imm);
3062 guint8 *jump_to_end;
3064 /* handle shifts below 32 bits */
3065 x86_shld_reg (code, ins->unused, ins->sreg1);
3066 x86_shift_reg (code, X86_SHL, ins->sreg1);
3068 x86_test_reg_imm (code, X86_ECX, 32);
3069 jump_to_end = code; x86_branch8 (code, X86_CC_EQ, 0, TRUE);
3071 /* handle shift over 32 bit */
3072 x86_mov_reg_reg (code, ins->unused, ins->sreg1, 4);
3073 x86_clear_reg (code, ins->sreg1);
3075 x86_patch (jump_to_end, code);
3079 guint8 *jump_to_end;
3081 /* handle shifts below 32 bits */
3082 x86_shrd_reg (code, ins->sreg1, ins->unused);
3083 x86_shift_reg (code, X86_SAR, ins->unused);
3085 x86_test_reg_imm (code, X86_ECX, 32);
3086 jump_to_end = code; x86_branch8 (code, X86_CC_EQ, 0, FALSE);
3088 /* handle shifts over 31 bits */
3089 x86_mov_reg_reg (code, ins->sreg1, ins->unused, 4);
3090 x86_shift_reg_imm (code, X86_SAR, ins->unused, 31);
3092 x86_patch (jump_to_end, code);
3096 guint8 *jump_to_end;
3098 /* handle shifts below 32 bits */
3099 x86_shrd_reg (code, ins->sreg1, ins->unused);
3100 x86_shift_reg (code, X86_SHR, ins->unused);
3102 x86_test_reg_imm (code, X86_ECX, 32);
3103 jump_to_end = code; x86_branch8 (code, X86_CC_EQ, 0, FALSE);
3105 /* handle shifts over 31 bits */
3106 x86_mov_reg_reg (code, ins->sreg1, ins->unused, 4);
3107 x86_shift_reg_imm (code, X86_SHR, ins->unused, 31);
3109 x86_patch (jump_to_end, code);
3113 if (ins->inst_imm >= 32) {
3114 x86_mov_reg_reg (code, ins->unused, ins->sreg1, 4);
3115 x86_clear_reg (code, ins->sreg1);
3116 x86_shift_reg_imm (code, X86_SHL, ins->unused, ins->inst_imm - 32);
3118 x86_shld_reg_imm (code, ins->unused, ins->sreg1, ins->inst_imm);
3119 x86_shift_reg_imm (code, X86_SHL, ins->sreg1, ins->inst_imm);
3123 if (ins->inst_imm >= 32) {
3124 x86_mov_reg_reg (code, ins->sreg1, ins->unused, 4);
3125 x86_shift_reg_imm (code, X86_SAR, ins->unused, 0x1f);
3126 x86_shift_reg_imm (code, X86_SAR, ins->sreg1, ins->inst_imm - 32);
3128 x86_shrd_reg_imm (code, ins->sreg1, ins->unused, ins->inst_imm);
3129 x86_shift_reg_imm (code, X86_SAR, ins->unused, ins->inst_imm);
3132 case OP_LSHR_UN_IMM:
3133 if (ins->inst_imm >= 32) {
3134 x86_mov_reg_reg (code, ins->sreg1, ins->unused, 4);
3135 x86_clear_reg (code, ins->unused);
3136 x86_shift_reg_imm (code, X86_SHR, ins->sreg1, ins->inst_imm - 32);
3138 x86_shrd_reg_imm (code, ins->sreg1, ins->unused, ins->inst_imm);
3139 x86_shift_reg_imm (code, X86_SHR, ins->unused, ins->inst_imm);
3143 x86_not_reg (code, ins->sreg1);
3146 x86_neg_reg (code, ins->sreg1);
3149 x86_widen_reg (code, ins->dreg, ins->sreg1, TRUE, FALSE);
3152 x86_widen_reg (code, ins->dreg, ins->sreg1, TRUE, TRUE);
3155 x86_imul_reg_reg (code, ins->sreg1, ins->sreg2);
3158 x86_imul_reg_reg_imm (code, ins->dreg, ins->sreg1, ins->inst_imm);
3161 x86_imul_reg_reg (code, ins->sreg1, ins->sreg2);
3162 EMIT_COND_SYSTEM_EXCEPTION (X86_CC_O, FALSE, "OverflowException");
3164 case CEE_MUL_OVF_UN: {
3165 /* the mul operation and the exception check should most likely be split */
3166 int non_eax_reg, saved_eax = FALSE, saved_edx = FALSE;
3167 /*g_assert (ins->sreg2 == X86_EAX);
3168 g_assert (ins->dreg == X86_EAX);*/
3169 if (ins->sreg2 == X86_EAX) {
3170 non_eax_reg = ins->sreg1;
3171 } else if (ins->sreg1 == X86_EAX) {
3172 non_eax_reg = ins->sreg2;
3174 /* no need to save since we're going to store to it anyway */
3175 if (ins->dreg != X86_EAX) {
3177 x86_push_reg (code, X86_EAX);
3179 x86_mov_reg_reg (code, X86_EAX, ins->sreg1, 4);
3180 non_eax_reg = ins->sreg2;
3182 if (ins->dreg == X86_EDX) {
3185 x86_push_reg (code, X86_EAX);
3187 } else if (ins->dreg != X86_EAX) {
3189 x86_push_reg (code, X86_EDX);
3191 x86_mul_reg (code, non_eax_reg, FALSE);
3192 /* save before the check since pop and mov don't change the flags */
3193 if (ins->dreg != X86_EAX)
3194 x86_mov_reg_reg (code, ins->dreg, X86_EAX, 4);
3196 x86_pop_reg (code, X86_EDX);
3198 x86_pop_reg (code, X86_EAX);
3199 EMIT_COND_SYSTEM_EXCEPTION (X86_CC_O, FALSE, "OverflowException");
3203 x86_mov_reg_imm (code, ins->dreg, ins->inst_c0);
3206 g_assert_not_reached ();
3207 mono_add_patch_info (cfg, offset, (MonoJumpInfoType)ins->inst_i1, ins->inst_p0);
3208 x86_mov_reg_imm (code, ins->dreg, 0);
3210 case OP_LOAD_GOTADDR:
3211 x86_call_imm (code, 0);
3213 * The patch needs to point to the pop, since the GOT offset needs
3214 * to be added to that address.
3216 mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_GOT_OFFSET, NULL);
3217 x86_pop_reg (code, ins->dreg);
3218 x86_alu_reg_imm (code, X86_ADD, ins->dreg, 0xf0f0f0f0);
3221 mono_add_patch_info (cfg, offset, (MonoJumpInfoType)ins->inst_right->inst_i1, ins->inst_right->inst_p0);
3222 x86_mov_reg_membase (code, ins->dreg, ins->inst_basereg, 0xf0f0f0f0, 4);
3224 case OP_X86_PUSH_GOT_ENTRY:
3225 mono_add_patch_info (cfg, offset, (MonoJumpInfoType)ins->inst_right->inst_i1, ins->inst_right->inst_p0);
3226 x86_push_membase (code, ins->inst_basereg, 0xf0f0f0f0);
3230 x86_mov_reg_reg (code, ins->dreg, ins->sreg1, 4);
3233 g_assert_not_reached ();
3236 * Note: this 'frame destruction' logic is useful for tail calls, too.
3237 * Keep in sync with the code in emit_epilog.
3241 /* FIXME: no tracing support... */
3242 if (cfg->prof_options & MONO_PROFILE_ENTER_LEAVE)
3243 code = mono_arch_instrument_epilog (cfg, mono_profiler_method_leave, code, FALSE);
3244 /* reset offset to make max_len work */
3245 offset = code - cfg->native_code;
3247 g_assert (!cfg->method->save_lmf);
3249 if (cfg->used_int_regs & (1 << X86_EBX))
3251 if (cfg->used_int_regs & (1 << X86_EDI))
3253 if (cfg->used_int_regs & (1 << X86_ESI))
3256 x86_lea_membase (code, X86_ESP, X86_EBP, pos);
3258 if (cfg->used_int_regs & (1 << X86_ESI))
3259 x86_pop_reg (code, X86_ESI);
3260 if (cfg->used_int_regs & (1 << X86_EDI))
3261 x86_pop_reg (code, X86_EDI);
3262 if (cfg->used_int_regs & (1 << X86_EBX))
3263 x86_pop_reg (code, X86_EBX);
3265 /* restore ESP/EBP */
3267 offset = code - cfg->native_code;
3268 mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_METHOD_JUMP, ins->inst_p0);
3269 x86_jump32 (code, 0);
3273 /* ensure ins->sreg1 is not NULL
3274 * note that cmp DWORD PTR [eax], eax is one byte shorter than
3275 * cmp DWORD PTR [eax], 0
3277 x86_alu_membase_reg (code, X86_CMP, ins->sreg1, 0, ins->sreg1);
3280 int hreg = ins->sreg1 == X86_EAX? X86_ECX: X86_EAX;
3281 x86_push_reg (code, hreg);
3282 x86_lea_membase (code, hreg, X86_EBP, cfg->sig_cookie);
3283 x86_mov_membase_reg (code, ins->sreg1, 0, hreg, 4);
3284 x86_pop_reg (code, hreg);
3292 call = (MonoCallInst*)ins;
3293 if (ins->flags & MONO_INST_HAS_METHOD)
3294 code = emit_call (cfg, code, MONO_PATCH_INFO_METHOD, call->method);
3296 code = emit_call (cfg, code, MONO_PATCH_INFO_ABS, call->fptr);
3297 if (call->stack_usage && !CALLCONV_IS_STDCALL (call->signature)) {
3298 /* a pop is one byte, while an add reg, imm is 3. So if there are 4 or 8
3299 * bytes to pop, we want to use pops. GCC does this (note it won't happen
3300 * for P4 or i686 because gcc will avoid using pop push at all. But we aren't
3301 * smart enough to do that optimization yet
3303 * It turns out that on my P4, doing two pops for 8 bytes on the stack makes
3304 * mcs botstrap slow down. However, doing 1 pop for 4 bytes creates a small,
3305 * (most likely from locality benefits). People with other processors should
3306 * check on theirs to see what happens.
3308 if (call->stack_usage == 4) {
3309 /* we want to use registers that won't get used soon, so use
3310 * ecx, as eax will get allocated first. edx is used by long calls,
3311 * so we can't use that.
3314 x86_pop_reg (code, X86_ECX);
3316 x86_alu_reg_imm (code, X86_ADD, X86_ESP, call->stack_usage);
3319 code = emit_move_return_value (cfg, ins, code);
3324 case OP_VOIDCALL_REG:
3326 call = (MonoCallInst*)ins;
3327 x86_call_reg (code, ins->sreg1);
3328 if (call->stack_usage && !CALLCONV_IS_STDCALL (call->signature)) {
3329 if (call->stack_usage == 4)
3330 x86_pop_reg (code, X86_ECX);
3332 x86_alu_reg_imm (code, X86_ADD, X86_ESP, call->stack_usage);
3334 code = emit_move_return_value (cfg, ins, code);
3336 case OP_FCALL_MEMBASE:
3337 case OP_LCALL_MEMBASE:
3338 case OP_VCALL_MEMBASE:
3339 case OP_VOIDCALL_MEMBASE:
3340 case OP_CALL_MEMBASE:
3341 call = (MonoCallInst*)ins;
3342 x86_call_membase (code, ins->sreg1, ins->inst_offset);
3343 if (call->stack_usage && !CALLCONV_IS_STDCALL (call->signature)) {
3344 if (call->stack_usage == 4)
3345 x86_pop_reg (code, X86_ECX);
3347 x86_alu_reg_imm (code, X86_ADD, X86_ESP, call->stack_usage);
3349 code = emit_move_return_value (cfg, ins, code);
3353 x86_push_reg (code, ins->sreg1);
3355 case OP_X86_PUSH_IMM:
3356 x86_push_imm (code, ins->inst_imm);
3358 case OP_X86_PUSH_MEMBASE:
3359 x86_push_membase (code, ins->inst_basereg, ins->inst_offset);
3361 case OP_X86_PUSH_OBJ:
3362 x86_alu_reg_imm (code, X86_SUB, X86_ESP, ins->inst_imm);
3363 x86_push_reg (code, X86_EDI);
3364 x86_push_reg (code, X86_ESI);
3365 x86_push_reg (code, X86_ECX);
3366 if (ins->inst_offset)
3367 x86_lea_membase (code, X86_ESI, ins->inst_basereg, ins->inst_offset);
3369 x86_mov_reg_reg (code, X86_ESI, ins->inst_basereg, 4);
3370 x86_lea_membase (code, X86_EDI, X86_ESP, 12);
3371 x86_mov_reg_imm (code, X86_ECX, (ins->inst_imm >> 2));
3373 x86_prefix (code, X86_REP_PREFIX);
3375 x86_pop_reg (code, X86_ECX);
3376 x86_pop_reg (code, X86_ESI);
3377 x86_pop_reg (code, X86_EDI);
3380 x86_lea_memindex (code, ins->dreg, ins->sreg1, ins->inst_imm, ins->sreg2, ins->unused);
3382 case OP_X86_LEA_MEMBASE:
3383 x86_lea_membase (code, ins->dreg, ins->sreg1, ins->inst_imm);
3386 x86_xchg_reg_reg (code, ins->sreg1, ins->sreg2, 4);
3389 /* keep alignment */
3390 x86_alu_reg_imm (code, X86_ADD, ins->sreg1, MONO_ARCH_FRAME_ALIGNMENT - 1);
3391 x86_alu_reg_imm (code, X86_AND, ins->sreg1, ~(MONO_ARCH_FRAME_ALIGNMENT - 1));
3392 code = mono_emit_stack_alloc (code, ins);
3393 x86_mov_reg_reg (code, ins->dreg, X86_ESP, 4);
3399 x86_push_reg (code, ins->sreg1);
3400 code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD,
3401 (gpointer)"mono_arch_throw_exception");
3405 x86_push_reg (code, ins->sreg1);
3406 code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD,
3407 (gpointer)"mono_arch_rethrow_exception");
3410 case OP_CALL_HANDLER:
3411 mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_BB, ins->inst_target_bb);
3412 x86_call_imm (code, 0);
3415 ins->inst_c0 = code - cfg->native_code;
3418 //g_print ("target: %p, next: %p, curr: %p, last: %p\n", ins->inst_target_bb, bb->next_bb, ins, bb->last_ins);
3419 //if ((ins->inst_target_bb == bb->next_bb) && ins == bb->last_ins)
3421 if (ins->flags & MONO_INST_BRLABEL) {
3422 if (ins->inst_i0->inst_c0) {
3423 x86_jump_code (code, cfg->native_code + ins->inst_i0->inst_c0);
3425 mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_LABEL, ins->inst_i0);
3426 if ((cfg->opt & MONO_OPT_BRANCH) &&
3427 x86_is_imm8 (ins->inst_i0->inst_c1 - cpos))
3428 x86_jump8 (code, 0);
3430 x86_jump32 (code, 0);
3433 if (ins->inst_target_bb->native_offset) {
3434 x86_jump_code (code, cfg->native_code + ins->inst_target_bb->native_offset);
3436 mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_BB, ins->inst_target_bb);
3437 if ((cfg->opt & MONO_OPT_BRANCH) &&
3438 x86_is_imm8 (ins->inst_target_bb->max_offset - cpos))
3439 x86_jump8 (code, 0);
3441 x86_jump32 (code, 0);
3446 x86_jump_reg (code, ins->sreg1);
3449 x86_set_reg (code, X86_CC_EQ, ins->dreg, TRUE);
3450 x86_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
3453 x86_set_reg (code, X86_CC_LT, ins->dreg, TRUE);
3454 x86_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
3457 x86_set_reg (code, X86_CC_LT, ins->dreg, FALSE);
3458 x86_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
3461 x86_set_reg (code, X86_CC_GT, ins->dreg, TRUE);
3462 x86_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
3465 x86_set_reg (code, X86_CC_GT, ins->dreg, FALSE);
3466 x86_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
3469 x86_set_reg (code, X86_CC_NE, ins->dreg, TRUE);
3470 x86_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
3472 case OP_COND_EXC_EQ:
3473 case OP_COND_EXC_NE_UN:
3474 case OP_COND_EXC_LT:
3475 case OP_COND_EXC_LT_UN:
3476 case OP_COND_EXC_GT:
3477 case OP_COND_EXC_GT_UN:
3478 case OP_COND_EXC_GE:
3479 case OP_COND_EXC_GE_UN:
3480 case OP_COND_EXC_LE:
3481 case OP_COND_EXC_LE_UN:
3482 case OP_COND_EXC_OV:
3483 case OP_COND_EXC_NO:
3485 case OP_COND_EXC_NC:
3486 EMIT_COND_SYSTEM_EXCEPTION (branch_cc_table [ins->opcode - OP_COND_EXC_EQ],
3487 (ins->opcode < OP_COND_EXC_NE_UN), ins->inst_p1);
3499 EMIT_COND_BRANCH (ins, branch_cc_table [ins->opcode - CEE_BEQ], (ins->opcode < CEE_BNE_UN));
3502 /* floating point opcodes */
3504 double d = *(double *)ins->inst_p0;
3506 if ((d == 0.0) && (mono_signbit (d) == 0)) {
3508 } else if (d == 1.0) {
3511 if (cfg->compile_aot) {
3512 guint32 *val = (guint32*)&d;
3513 x86_push_imm (code, val [1]);
3514 x86_push_imm (code, val [0]);
3515 x86_fld_membase (code, X86_ESP, 0, TRUE);
3516 x86_alu_reg_imm (code, X86_ADD, X86_ESP, 8);
3519 mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_R8, ins->inst_p0);
3520 x86_fld (code, NULL, TRUE);
3526 float f = *(float *)ins->inst_p0;
3528 if ((f == 0.0) && (mono_signbit (f) == 0)) {
3530 } else if (f == 1.0) {
3533 if (cfg->compile_aot) {
3534 guint32 val = *(guint32*)&f;
3535 x86_push_imm (code, val);
3536 x86_fld_membase (code, X86_ESP, 0, FALSE);
3537 x86_alu_reg_imm (code, X86_ADD, X86_ESP, 4);
3540 mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_R4, ins->inst_p0);
3541 x86_fld (code, NULL, FALSE);
3546 case OP_STORER8_MEMBASE_REG:
3547 x86_fst_membase (code, ins->inst_destbasereg, ins->inst_offset, TRUE, TRUE);
3549 case OP_LOADR8_SPILL_MEMBASE:
3550 x86_fld_membase (code, ins->inst_basereg, ins->inst_offset, TRUE);
3553 case OP_LOADR8_MEMBASE:
3554 x86_fld_membase (code, ins->inst_basereg, ins->inst_offset, TRUE);
3556 case OP_STORER4_MEMBASE_REG:
3557 x86_fst_membase (code, ins->inst_destbasereg, ins->inst_offset, FALSE, TRUE);
3559 case OP_LOADR4_MEMBASE:
3560 x86_fld_membase (code, ins->inst_basereg, ins->inst_offset, FALSE);
3562 case CEE_CONV_R4: /* FIXME: change precision */
3564 x86_push_reg (code, ins->sreg1);
3565 x86_fild_membase (code, X86_ESP, 0, FALSE);
3566 x86_alu_reg_imm (code, X86_ADD, X86_ESP, 4);
3568 case OP_X86_FP_LOAD_I8:
3569 x86_fild_membase (code, ins->inst_basereg, ins->inst_offset, TRUE);
3571 case OP_X86_FP_LOAD_I4:
3572 x86_fild_membase (code, ins->inst_basereg, ins->inst_offset, FALSE);
3574 case OP_FCONV_TO_I1:
3575 code = emit_float_to_int (cfg, code, ins->dreg, 1, TRUE);
3577 case OP_FCONV_TO_U1:
3578 code = emit_float_to_int (cfg, code, ins->dreg, 1, FALSE);
3580 case OP_FCONV_TO_I2:
3581 code = emit_float_to_int (cfg, code, ins->dreg, 2, TRUE);
3583 case OP_FCONV_TO_U2:
3584 code = emit_float_to_int (cfg, code, ins->dreg, 2, FALSE);
3586 case OP_FCONV_TO_I4:
3588 code = emit_float_to_int (cfg, code, ins->dreg, 4, TRUE);
3590 case OP_FCONV_TO_I8:
3591 x86_alu_reg_imm (code, X86_SUB, X86_ESP, 4);
3592 x86_fnstcw_membase(code, X86_ESP, 0);
3593 x86_mov_reg_membase (code, ins->dreg, X86_ESP, 0, 2);
3594 x86_alu_reg_imm (code, X86_OR, ins->dreg, 0xc00);
3595 x86_mov_membase_reg (code, X86_ESP, 2, ins->dreg, 2);
3596 x86_fldcw_membase (code, X86_ESP, 2);
3597 x86_alu_reg_imm (code, X86_SUB, X86_ESP, 8);
3598 x86_fist_pop_membase (code, X86_ESP, 0, TRUE);
3599 x86_pop_reg (code, ins->dreg);
3600 x86_pop_reg (code, ins->unused);
3601 x86_fldcw_membase (code, X86_ESP, 0);
3602 x86_alu_reg_imm (code, X86_ADD, X86_ESP, 4);
3604 case OP_LCONV_TO_R_UN: {
3605 static guint8 mn[] = { 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x80, 0x3f, 0x40 };
3608 /* load 64bit integer to FP stack */
3609 x86_push_imm (code, 0);
3610 x86_push_reg (code, ins->sreg2);
3611 x86_push_reg (code, ins->sreg1);
3612 x86_fild_membase (code, X86_ESP, 0, TRUE);
3613 /* store as 80bit FP value */
3614 x86_fst80_membase (code, X86_ESP, 0);
3616 /* test if lreg is negative */
3617 x86_test_reg_reg (code, ins->sreg2, ins->sreg2);
3618 br = code; x86_branch8 (code, X86_CC_GEZ, 0, TRUE);
3620 /* add correction constant mn */
3621 x86_fld80_mem (code, mn);
3622 x86_fld80_membase (code, X86_ESP, 0);
3623 x86_fp_op_reg (code, X86_FADD, 1, TRUE);
3624 x86_fst80_membase (code, X86_ESP, 0);
3626 x86_patch (br, code);
3628 x86_fld80_membase (code, X86_ESP, 0);
3629 x86_alu_reg_imm (code, X86_ADD, X86_ESP, 12);
3633 case OP_LCONV_TO_OVF_I: {
3634 guint8 *br [3], *label [1];
3637 * Valid ints: 0xffffffff:8000000 to 00000000:0x7f000000
3639 x86_test_reg_reg (code, ins->sreg1, ins->sreg1);
3641 /* If the low word top bit is set, see if we are negative */
3642 br [0] = code; x86_branch8 (code, X86_CC_LT, 0, TRUE);
3643 /* We are not negative (no top bit set, check for our top word to be zero */
3644 x86_test_reg_reg (code, ins->sreg2, ins->sreg2);
3645 br [1] = code; x86_branch8 (code, X86_CC_EQ, 0, TRUE);
3648 /* throw exception */
3649 mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_EXC, "OverflowException");
3650 x86_jump32 (code, 0);
3652 x86_patch (br [0], code);
3653 /* our top bit is set, check that top word is 0xfffffff */
3654 x86_alu_reg_imm (code, X86_CMP, ins->sreg2, 0xffffffff);
3656 x86_patch (br [1], code);
3657 /* nope, emit exception */
3658 br [2] = code; x86_branch8 (code, X86_CC_NE, 0, TRUE);
3659 x86_patch (br [2], label [0]);
3661 if (ins->dreg != ins->sreg1)
3662 x86_mov_reg_reg (code, ins->dreg, ins->sreg1, 4);
3666 x86_fp_op_reg (code, X86_FADD, 1, TRUE);
3669 x86_fp_op_reg (code, X86_FSUB, 1, TRUE);
3672 x86_fp_op_reg (code, X86_FMUL, 1, TRUE);
3675 x86_fp_op_reg (code, X86_FDIV, 1, TRUE);
3683 x86_fp_op_reg (code, X86_FADD, 1, TRUE);
3688 x86_fp_op_reg (code, X86_FADD, 1, TRUE);
3695 * it really doesn't make sense to inline all this code,
3696 * it's here just to show that things may not be as simple
3699 guchar *check_pos, *end_tan, *pop_jump;
3700 x86_push_reg (code, X86_EAX);
3703 x86_test_reg_imm (code, X86_EAX, X86_FP_C2);
3705 x86_branch8 (code, X86_CC_NE, 0, FALSE);
3706 x86_fstp (code, 0); /* pop the 1.0 */
3708 x86_jump8 (code, 0);
3710 x86_fp_op (code, X86_FADD, 0);
3714 x86_test_reg_imm (code, X86_EAX, X86_FP_C2);
3716 x86_branch8 (code, X86_CC_NE, 0, FALSE);
3719 x86_patch (pop_jump, code);
3720 x86_fstp (code, 0); /* pop the 1.0 */
3721 x86_patch (check_pos, code);
3722 x86_patch (end_tan, code);
3724 x86_fp_op_reg (code, X86_FADD, 1, TRUE);
3725 x86_pop_reg (code, X86_EAX);
3732 x86_fp_op_reg (code, X86_FADD, 1, TRUE);
3743 x86_push_reg (code, X86_EAX);
3744 /* we need to exchange ST(0) with ST(1) */
3747 /* this requires a loop, because fprem somtimes
3748 * returns a partial remainder */
3750 /* looks like MS is using fprem instead of the IEEE compatible fprem1 */
3751 /* x86_fprem1 (code); */
3754 x86_alu_reg_imm (code, X86_AND, X86_EAX, X86_FP_C2);
3756 x86_branch8 (code, X86_CC_NE, l1 - l2, FALSE);
3761 x86_pop_reg (code, X86_EAX);
3765 if (cfg->opt & MONO_OPT_FCMOV) {
3766 x86_fcomip (code, 1);
3770 /* this overwrites EAX */
3771 EMIT_FPCOMPARE(code);
3772 x86_alu_reg_imm (code, X86_AND, X86_EAX, X86_FP_CC_MASK);
3775 if (cfg->opt & MONO_OPT_FCMOV) {
3776 /* zeroing the register at the start results in
3777 * shorter and faster code (we can also remove the widening op)
3779 guchar *unordered_check;
3780 x86_alu_reg_reg (code, X86_XOR, ins->dreg, ins->dreg);
3781 x86_fcomip (code, 1);
3783 unordered_check = code;
3784 x86_branch8 (code, X86_CC_P, 0, FALSE);
3785 x86_set_reg (code, X86_CC_EQ, ins->dreg, FALSE);
3786 x86_patch (unordered_check, code);
3789 if (ins->dreg != X86_EAX)
3790 x86_push_reg (code, X86_EAX);
3792 EMIT_FPCOMPARE(code);
3793 x86_alu_reg_imm (code, X86_AND, X86_EAX, X86_FP_CC_MASK);
3794 x86_alu_reg_imm (code, X86_CMP, X86_EAX, 0x4000);
3795 x86_set_reg (code, X86_CC_EQ, ins->dreg, TRUE);
3796 x86_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
3798 if (ins->dreg != X86_EAX)
3799 x86_pop_reg (code, X86_EAX);
3803 if (cfg->opt & MONO_OPT_FCMOV) {
3804 /* zeroing the register at the start results in
3805 * shorter and faster code (we can also remove the widening op)
3807 x86_alu_reg_reg (code, X86_XOR, ins->dreg, ins->dreg);
3808 x86_fcomip (code, 1);
3810 if (ins->opcode == OP_FCLT_UN) {
3811 guchar *unordered_check = code;
3812 guchar *jump_to_end;
3813 x86_branch8 (code, X86_CC_P, 0, FALSE);
3814 x86_set_reg (code, X86_CC_GT, ins->dreg, FALSE);
3816 x86_jump8 (code, 0);
3817 x86_patch (unordered_check, code);
3818 x86_inc_reg (code, ins->dreg);
3819 x86_patch (jump_to_end, code);
3821 x86_set_reg (code, X86_CC_GT, ins->dreg, FALSE);
3825 if (ins->dreg != X86_EAX)
3826 x86_push_reg (code, X86_EAX);
3828 EMIT_FPCOMPARE(code);
3829 x86_alu_reg_imm (code, X86_AND, X86_EAX, X86_FP_CC_MASK);
3830 if (ins->opcode == OP_FCLT_UN) {
3831 guchar *is_not_zero_check, *end_jump;
3832 is_not_zero_check = code;
3833 x86_branch8 (code, X86_CC_NZ, 0, TRUE);
3835 x86_jump8 (code, 0);
3836 x86_patch (is_not_zero_check, code);
3837 x86_alu_reg_imm (code, X86_CMP, X86_EAX, X86_FP_CC_MASK);
3839 x86_patch (end_jump, code);
3841 x86_set_reg (code, X86_CC_EQ, ins->dreg, TRUE);
3842 x86_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
3844 if (ins->dreg != X86_EAX)
3845 x86_pop_reg (code, X86_EAX);
3849 if (cfg->opt & MONO_OPT_FCMOV) {
3850 /* zeroing the register at the start results in
3851 * shorter and faster code (we can also remove the widening op)
3853 guchar *unordered_check;
3854 x86_alu_reg_reg (code, X86_XOR, ins->dreg, ins->dreg);
3855 x86_fcomip (code, 1);
3857 if (ins->opcode == OP_FCGT) {
3858 unordered_check = code;
3859 x86_branch8 (code, X86_CC_P, 0, FALSE);
3860 x86_set_reg (code, X86_CC_LT, ins->dreg, FALSE);
3861 x86_patch (unordered_check, code);
3863 x86_set_reg (code, X86_CC_LT, ins->dreg, FALSE);
3867 if (ins->dreg != X86_EAX)
3868 x86_push_reg (code, X86_EAX);
3870 EMIT_FPCOMPARE(code);
3871 x86_alu_reg_imm (code, X86_AND, X86_EAX, X86_FP_CC_MASK);
3872 x86_alu_reg_imm (code, X86_CMP, X86_EAX, X86_FP_C0);
3873 if (ins->opcode == OP_FCGT_UN) {
3874 guchar *is_not_zero_check, *end_jump;
3875 is_not_zero_check = code;
3876 x86_branch8 (code, X86_CC_NZ, 0, TRUE);
3878 x86_jump8 (code, 0);
3879 x86_patch (is_not_zero_check, code);
3880 x86_alu_reg_imm (code, X86_CMP, X86_EAX, X86_FP_CC_MASK);
3882 x86_patch (end_jump, code);
3884 x86_set_reg (code, X86_CC_EQ, ins->dreg, TRUE);
3885 x86_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
3887 if (ins->dreg != X86_EAX)
3888 x86_pop_reg (code, X86_EAX);
3891 if (cfg->opt & MONO_OPT_FCMOV) {
3892 guchar *jump = code;
3893 x86_branch8 (code, X86_CC_P, 0, TRUE);
3894 EMIT_COND_BRANCH (ins, X86_CC_EQ, FALSE);
3895 x86_patch (jump, code);
3898 x86_alu_reg_imm (code, X86_CMP, X86_EAX, 0x4000);
3899 EMIT_COND_BRANCH (ins, X86_CC_EQ, TRUE);
3902 /* Branch if C013 != 100 */
3903 if (cfg->opt & MONO_OPT_FCMOV) {
3904 /* branch if !ZF or (PF|CF) */
3905 EMIT_COND_BRANCH (ins, X86_CC_NE, FALSE);
3906 EMIT_COND_BRANCH (ins, X86_CC_P, FALSE);
3907 EMIT_COND_BRANCH (ins, X86_CC_B, FALSE);
3910 x86_alu_reg_imm (code, X86_CMP, X86_EAX, X86_FP_C3);
3911 EMIT_COND_BRANCH (ins, X86_CC_NE, FALSE);
3914 if (cfg->opt & MONO_OPT_FCMOV) {
3915 EMIT_COND_BRANCH (ins, X86_CC_GT, FALSE);
3918 EMIT_COND_BRANCH (ins, X86_CC_EQ, FALSE);
3921 if (cfg->opt & MONO_OPT_FCMOV) {
3922 EMIT_COND_BRANCH (ins, X86_CC_P, FALSE);
3923 EMIT_COND_BRANCH (ins, X86_CC_GT, FALSE);
3926 if (ins->opcode == OP_FBLT_UN) {
3927 guchar *is_not_zero_check, *end_jump;
3928 is_not_zero_check = code;
3929 x86_branch8 (code, X86_CC_NZ, 0, TRUE);
3931 x86_jump8 (code, 0);
3932 x86_patch (is_not_zero_check, code);
3933 x86_alu_reg_imm (code, X86_CMP, X86_EAX, X86_FP_CC_MASK);
3935 x86_patch (end_jump, code);
3937 EMIT_COND_BRANCH (ins, X86_CC_EQ, FALSE);
3941 if (cfg->opt & MONO_OPT_FCMOV) {
3942 EMIT_COND_BRANCH (ins, X86_CC_LT, FALSE);
3945 x86_alu_reg_imm (code, X86_CMP, X86_EAX, X86_FP_C0);
3946 if (ins->opcode == OP_FBGT_UN) {
3947 guchar *is_not_zero_check, *end_jump;
3948 is_not_zero_check = code;
3949 x86_branch8 (code, X86_CC_NZ, 0, TRUE);
3951 x86_jump8 (code, 0);
3952 x86_patch (is_not_zero_check, code);
3953 x86_alu_reg_imm (code, X86_CMP, X86_EAX, X86_FP_CC_MASK);
3955 x86_patch (end_jump, code);
3957 EMIT_COND_BRANCH (ins, X86_CC_EQ, FALSE);
3960 /* Branch if C013 == 100 or 001 */
3961 if (cfg->opt & MONO_OPT_FCMOV) {
3964 /* skip branch if C1=1 */
3966 x86_branch8 (code, X86_CC_P, 0, FALSE);
3967 /* branch if (C0 | C3) = 1 */
3968 EMIT_COND_BRANCH (ins, X86_CC_BE, FALSE);
3969 x86_patch (br1, code);
3972 x86_alu_reg_imm (code, X86_CMP, X86_EAX, X86_FP_C0);
3973 EMIT_COND_BRANCH (ins, X86_CC_EQ, FALSE);
3974 x86_alu_reg_imm (code, X86_CMP, X86_EAX, X86_FP_C3);
3975 EMIT_COND_BRANCH (ins, X86_CC_EQ, FALSE);
3978 /* Branch if C013 == 000 */
3979 if (cfg->opt & MONO_OPT_FCMOV) {
3980 EMIT_COND_BRANCH (ins, X86_CC_LE, FALSE);
3983 EMIT_COND_BRANCH (ins, X86_CC_NE, FALSE);
3986 /* Branch if C013=000 or 100 */
3987 if (cfg->opt & MONO_OPT_FCMOV) {
3990 /* skip branch if C1=1 */
3992 x86_branch8 (code, X86_CC_P, 0, FALSE);
3993 /* branch if C0=0 */
3994 EMIT_COND_BRANCH (ins, X86_CC_NB, FALSE);
3995 x86_patch (br1, code);
3998 x86_alu_reg_imm (code, X86_AND, X86_EAX, (X86_FP_C0|X86_FP_C1));
3999 x86_alu_reg_imm (code, X86_CMP, X86_EAX, 0);
4000 EMIT_COND_BRANCH (ins, X86_CC_EQ, FALSE);
4003 /* Branch if C013 != 001 */
4004 if (cfg->opt & MONO_OPT_FCMOV) {
4005 EMIT_COND_BRANCH (ins, X86_CC_P, FALSE);
4006 EMIT_COND_BRANCH (ins, X86_CC_GE, FALSE);
4009 x86_alu_reg_imm (code, X86_CMP, X86_EAX, X86_FP_C0);
4010 EMIT_COND_BRANCH (ins, X86_CC_NE, FALSE);
4012 case CEE_CKFINITE: {
4013 x86_push_reg (code, X86_EAX);
4016 x86_alu_reg_imm (code, X86_AND, X86_EAX, 0x4100);
4017 x86_alu_reg_imm (code, X86_CMP, X86_EAX, X86_FP_C0);
4018 x86_pop_reg (code, X86_EAX);
4019 EMIT_COND_SYSTEM_EXCEPTION (X86_CC_EQ, FALSE, "ArithmeticException");
4023 code = emit_tls_get (code, ins->dreg, ins->inst_offset);
4026 case OP_ATOMIC_ADD_I4: {
4027 int dreg = ins->dreg;
4029 if (dreg == ins->inst_basereg) {
4030 x86_push_reg (code, ins->sreg2);
4034 if (dreg != ins->sreg2)
4035 x86_mov_reg_reg (code, ins->dreg, ins->sreg2, 4);
4037 x86_prefix (code, X86_LOCK_PREFIX);
4038 x86_xadd_membase_reg (code, ins->inst_basereg, ins->inst_offset, dreg, 4);
4040 if (dreg != ins->dreg) {
4041 x86_mov_reg_reg (code, ins->dreg, dreg, 4);
4042 x86_pop_reg (code, dreg);
4047 case OP_ATOMIC_ADD_NEW_I4: {
4048 int dreg = ins->dreg;
4050 /* hack: limit in regalloc, dreg != sreg1 && dreg != sreg2 */
4051 if (ins->sreg2 == dreg) {
4052 if (dreg == X86_EBX) {
4054 if (ins->inst_basereg == X86_EDI)
4058 if (ins->inst_basereg == X86_EBX)
4061 } else if (ins->inst_basereg == dreg) {
4062 if (dreg == X86_EBX) {
4064 if (ins->sreg2 == X86_EDI)
4068 if (ins->sreg2 == X86_EBX)
4073 if (dreg != ins->dreg) {
4074 x86_push_reg (code, dreg);
4077 x86_mov_reg_reg (code, dreg, ins->sreg2, 4);
4078 x86_prefix (code, X86_LOCK_PREFIX);
4079 x86_xadd_membase_reg (code, ins->inst_basereg, ins->inst_offset, dreg, 4);
4080 /* dreg contains the old value, add with sreg2 value */
4081 x86_alu_reg_reg (code, X86_ADD, dreg, ins->sreg2);
4083 if (ins->dreg != dreg) {
4084 x86_mov_reg_reg (code, ins->dreg, dreg, 4);
4085 x86_pop_reg (code, dreg);
4090 case OP_ATOMIC_EXCHANGE_I4: {
4092 int sreg2 = ins->sreg2;
4093 int breg = ins->inst_basereg;
4095 /* cmpxchg uses eax as comperand, need to make sure we can use it
4096 * hack to overcome limits in x86 reg allocator
4097 * (req: dreg == eax and sreg2 != eax and breg != eax)
4099 if (ins->dreg != X86_EAX)
4100 x86_push_reg (code, X86_EAX);
4102 /* We need the EAX reg for the cmpxchg */
4103 if (ins->sreg2 == X86_EAX) {
4104 x86_push_reg (code, X86_EDX);
4105 x86_mov_reg_reg (code, X86_EDX, X86_EAX, 4);
4109 if (breg == X86_EAX) {
4110 x86_push_reg (code, X86_ESI);
4111 x86_mov_reg_reg (code, X86_ESI, X86_EAX, 4);
4115 x86_mov_reg_membase (code, X86_EAX, breg, ins->inst_offset, 4);
4117 br [0] = code; x86_prefix (code, X86_LOCK_PREFIX);
4118 x86_cmpxchg_membase_reg (code, breg, ins->inst_offset, sreg2);
4119 br [1] = code; x86_branch8 (code, X86_CC_NE, -1, FALSE);
4120 x86_patch (br [1], br [0]);
4122 if (breg != ins->inst_basereg)
4123 x86_pop_reg (code, X86_ESI);
4125 if (ins->dreg != X86_EAX) {
4126 x86_mov_reg_reg (code, ins->dreg, X86_EAX, 4);
4127 x86_pop_reg (code, X86_EAX);
4130 if (ins->sreg2 != sreg2)
4131 x86_pop_reg (code, X86_EDX);
4136 g_warning ("unknown opcode %s in %s()\n", mono_inst_name (ins->opcode), __FUNCTION__);
4137 g_assert_not_reached ();
4140 if ((code - cfg->native_code - offset) > max_len) {
4141 g_warning ("wrong maximal instruction length of instruction %s (expected %d, got %d)",
4142 mono_inst_name (ins->opcode), max_len, code - cfg->native_code - offset);
4143 g_assert_not_reached ();
4149 last_offset = offset;
4154 cfg->code_len = code - cfg->native_code;
4158 mono_arch_register_lowlevel_calls (void)
4163 mono_arch_patch_code (MonoMethod *method, MonoDomain *domain, guint8 *code, MonoJumpInfo *ji, gboolean run_cctors)
4165 MonoJumpInfo *patch_info;
4166 gboolean compile_aot = !run_cctors;
4168 for (patch_info = ji; patch_info; patch_info = patch_info->next) {
4169 unsigned char *ip = patch_info->ip.i + code;
4170 const unsigned char *target;
4172 target = mono_resolve_patch_target (method, domain, code, patch_info, run_cctors);
4175 switch (patch_info->type) {
4176 case MONO_PATCH_INFO_BB:
4177 case MONO_PATCH_INFO_LABEL:
4180 /* No need to patch these */
4185 switch (patch_info->type) {
4186 case MONO_PATCH_INFO_IP:
4187 *((gconstpointer *)(ip)) = target;
4189 case MONO_PATCH_INFO_CLASS_INIT: {
4191 /* Might already been changed to a nop */
4192 x86_call_code (code, 0);
4193 x86_patch (ip, target);
4196 case MONO_PATCH_INFO_ABS:
4197 case MONO_PATCH_INFO_METHOD:
4198 case MONO_PATCH_INFO_METHOD_JUMP:
4199 case MONO_PATCH_INFO_INTERNAL_METHOD:
4200 case MONO_PATCH_INFO_BB:
4201 case MONO_PATCH_INFO_LABEL:
4202 x86_patch (ip, target);
4204 case MONO_PATCH_INFO_NONE:
4207 guint32 offset = mono_arch_get_patch_offset (ip);
4208 *((gconstpointer *)(ip + offset)) = target;
4216 mono_arch_emit_prolog (MonoCompile *cfg)
4218 MonoMethod *method = cfg->method;
4220 MonoMethodSignature *sig;
4222 int alloc_size, pos, max_offset, i;
4225 cfg->code_size = MAX (mono_method_get_header (method)->code_size * 4, 256);
4226 code = cfg->native_code = g_malloc (cfg->code_size);
4228 x86_push_reg (code, X86_EBP);
4229 x86_mov_reg_reg (code, X86_EBP, X86_ESP, 4);
4231 alloc_size = - cfg->stack_offset;
4234 if (method->wrapper_type == MONO_WRAPPER_NATIVE_TO_MANAGED) {
4235 /* Might need to attach the thread to the JIT */
4236 if (lmf_tls_offset != -1) {
4239 code = emit_tls_get ( code, X86_EAX, lmf_tls_offset);
4240 #ifdef PLATFORM_WIN32
4241 /* The TLS key actually contains a pointer to the MonoJitTlsData structure */
4242 /* FIXME: Add a separate key for LMF to avoid this */
4243 x86_alu_reg_imm (code, X86_ADD, X86_EAX, G_STRUCT_OFFSET (MonoJitTlsData, lmf));
4245 x86_test_reg_reg (code, X86_EAX, X86_EAX);
4247 x86_branch8 (code, X86_CC_NE, 0, 0);
4248 x86_push_imm (code, cfg->domain);
4249 code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD, (gpointer)"mono_jit_thread_attach");
4250 x86_alu_reg_imm (code, X86_ADD, X86_ESP, 4);
4251 x86_patch (buf, code);
4254 g_assert (!cfg->compile_aot);
4255 x86_push_imm (code, cfg->domain);
4256 code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD, (gpointer)"mono_jit_thread_attach");
4257 x86_alu_reg_imm (code, X86_ADD, X86_ESP, 4);
4261 if (method->save_lmf) {
4262 pos += sizeof (MonoLMF);
4264 /* save the current IP */
4265 mono_add_patch_info (cfg, code + 1 - cfg->native_code, MONO_PATCH_INFO_IP, NULL);
4266 x86_push_imm_template (code);
4268 /* save all caller saved regs */
4269 x86_push_reg (code, X86_EBP);
4270 x86_push_reg (code, X86_ESI);
4271 x86_push_reg (code, X86_EDI);
4272 x86_push_reg (code, X86_EBX);
4274 /* save method info */
4275 x86_push_imm (code, method);
4277 /* get the address of lmf for the current thread */
4279 * This is performance critical so we try to use some tricks to make
4282 if (lmf_tls_offset != -1) {
4283 /* Load lmf quicky using the GS register */
4284 code = emit_tls_get (code, X86_EAX, lmf_tls_offset);
4285 #ifdef PLATFORM_WIN32
4286 /* The TLS key actually contains a pointer to the MonoJitTlsData structure */
4287 /* FIXME: Add a separate key for LMF to avoid this */
4288 x86_alu_reg_imm (code, X86_ADD, X86_EAX, G_STRUCT_OFFSET (MonoJitTlsData, lmf));
4292 if (cfg->compile_aot) {
4293 /* The GOT var does not exist yet */
4294 x86_call_imm (code, 0);
4295 mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_GOT_OFFSET, NULL);
4296 x86_pop_reg (code, X86_EAX);
4297 x86_alu_reg_imm (code, X86_ADD, X86_EAX, 0);
4298 mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_INTERNAL_METHOD, (gpointer)"mono_get_lmf_addr");
4299 x86_call_membase (code, X86_EAX, 0xf0f0f0f0);
4302 code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD, (gpointer)"mono_get_lmf_addr");
4306 x86_push_reg (code, X86_EAX);
4307 /* push *lfm (previous_lmf) */
4308 x86_push_membase (code, X86_EAX, 0);
4310 x86_mov_membase_reg (code, X86_EAX, 0, X86_ESP, 4);
4313 if (cfg->used_int_regs & (1 << X86_EBX)) {
4314 x86_push_reg (code, X86_EBX);
4318 if (cfg->used_int_regs & (1 << X86_EDI)) {
4319 x86_push_reg (code, X86_EDI);
4323 if (cfg->used_int_regs & (1 << X86_ESI)) {
4324 x86_push_reg (code, X86_ESI);
4332 /* See mono_emit_stack_alloc */
4333 #if defined(PLATFORM_WIN32) || defined(MONO_ARCH_SIGSEGV_ON_ALTSTACK)
4334 guint32 remaining_size = alloc_size;
4335 while (remaining_size >= 0x1000) {
4336 x86_alu_reg_imm (code, X86_SUB, X86_ESP, 0x1000);
4337 x86_test_membase_reg (code, X86_ESP, 0, X86_ESP);
4338 remaining_size -= 0x1000;
4341 x86_alu_reg_imm (code, X86_SUB, X86_ESP, remaining_size);
4343 x86_alu_reg_imm (code, X86_SUB, X86_ESP, alloc_size);
4347 /* compute max_offset in order to use short forward jumps */
4349 if (cfg->opt & MONO_OPT_BRANCH) {
4350 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
4351 MonoInst *ins = bb->code;
4352 bb->max_offset = max_offset;
4354 if (cfg->prof_options & MONO_PROFILE_COVERAGE)
4356 /* max alignment for loops */
4357 if ((cfg->opt & MONO_OPT_LOOP) && bb_is_loop_start (bb))
4358 max_offset += LOOP_ALIGNMENT;
4361 if (ins->opcode == OP_LABEL)
4362 ins->inst_c1 = max_offset;
4364 max_offset += ((guint8 *)ins_spec [ins->opcode])[MONO_INST_LEN];
4370 if (mono_jit_trace_calls != NULL && mono_trace_eval (method))
4371 code = mono_arch_instrument_prolog (cfg, mono_trace_enter_method, code, TRUE);
4373 /* load arguments allocated to register from the stack */
4374 sig = mono_method_signature (method);
4377 for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
4378 inst = cfg->varinfo [pos];
4379 if (inst->opcode == OP_REGVAR) {
4380 x86_mov_reg_membase (code, inst->dreg, X86_EBP, inst->inst_offset, 4);
4381 if (cfg->verbose_level > 2)
4382 g_print ("Argument %d assigned to register %s\n", pos, mono_arch_regname (inst->dreg));
4387 cfg->code_len = code - cfg->native_code;
4393 mono_arch_emit_epilog (MonoCompile *cfg)
4395 MonoMethod *method = cfg->method;
4396 MonoMethodSignature *sig = mono_method_signature (method);
4398 guint32 stack_to_pop;
4400 int max_epilog_size = 16;
4403 if (cfg->method->save_lmf)
4404 max_epilog_size += 128;
4406 if (mono_jit_trace_calls != NULL)
4407 max_epilog_size += 50;
4409 while (cfg->code_len + max_epilog_size > (cfg->code_size - 16)) {
4410 cfg->code_size *= 2;
4411 cfg->native_code = g_realloc (cfg->native_code, cfg->code_size);
4412 mono_jit_stats.code_reallocs++;
4415 code = cfg->native_code + cfg->code_len;
4417 if (mono_jit_trace_calls != NULL && mono_trace_eval (method))
4418 code = mono_arch_instrument_epilog (cfg, mono_trace_leave_method, code, TRUE);
4420 /* the code restoring the registers must be kept in sync with CEE_JMP */
4423 if (method->save_lmf) {
4424 gint32 prev_lmf_reg;
4426 /* Find a spare register */
4427 switch (sig->ret->type) {
4430 prev_lmf_reg = X86_EDI;
4431 cfg->used_int_regs |= (1 << X86_EDI);
4434 prev_lmf_reg = X86_EDX;
4438 /* reg = previous_lmf */
4439 x86_mov_reg_membase (code, prev_lmf_reg, X86_EBP, -32, 4);
4442 x86_mov_reg_membase (code, X86_ECX, X86_EBP, -28, 4);
4444 /* *(lmf) = previous_lmf */
4445 x86_mov_membase_reg (code, X86_ECX, 0, prev_lmf_reg, 4);
4447 /* restore caller saved regs */
4448 if (cfg->used_int_regs & (1 << X86_EBX)) {
4449 x86_mov_reg_membase (code, X86_EBX, X86_EBP, -20, 4);
4452 if (cfg->used_int_regs & (1 << X86_EDI)) {
4453 x86_mov_reg_membase (code, X86_EDI, X86_EBP, -16, 4);
4455 if (cfg->used_int_regs & (1 << X86_ESI)) {
4456 x86_mov_reg_membase (code, X86_ESI, X86_EBP, -12, 4);
4459 /* EBP is restored by LEAVE */
4461 if (cfg->used_int_regs & (1 << X86_EBX)) {
4464 if (cfg->used_int_regs & (1 << X86_EDI)) {
4467 if (cfg->used_int_regs & (1 << X86_ESI)) {
4472 x86_lea_membase (code, X86_ESP, X86_EBP, pos);
4474 if (cfg->used_int_regs & (1 << X86_ESI)) {
4475 x86_pop_reg (code, X86_ESI);
4477 if (cfg->used_int_regs & (1 << X86_EDI)) {
4478 x86_pop_reg (code, X86_EDI);
4480 if (cfg->used_int_regs & (1 << X86_EBX)) {
4481 x86_pop_reg (code, X86_EBX);
4485 /* Load returned vtypes into registers if needed */
4486 cinfo = get_call_info (sig, FALSE);
4487 if (cinfo->ret.storage == ArgValuetypeInReg) {
4488 for (quad = 0; quad < 2; quad ++) {
4489 switch (cinfo->ret.pair_storage [quad]) {
4491 x86_mov_reg_membase (code, cinfo->ret.pair_regs [quad], cfg->ret->inst_basereg, cfg->ret->inst_offset + (quad * sizeof (gpointer)), 4);
4493 case ArgOnFloatFpStack:
4494 x86_fld_membase (code, cfg->ret->inst_basereg, cfg->ret->inst_offset + (quad * sizeof (gpointer)), FALSE);
4496 case ArgOnDoubleFpStack:
4497 x86_fld_membase (code, cfg->ret->inst_basereg, cfg->ret->inst_offset + (quad * sizeof (gpointer)), TRUE);
4502 g_assert_not_reached ();
4509 if (CALLCONV_IS_STDCALL (sig)) {
4510 MonoJitArgumentInfo *arg_info = alloca (sizeof (MonoJitArgumentInfo) * (sig->param_count + 1));
4512 stack_to_pop = mono_arch_get_argument_info (sig, sig->param_count, arg_info);
4513 } else if (MONO_TYPE_ISSTRUCT (mono_method_signature (cfg->method)->ret) && (cinfo->ret.storage == ArgOnStack))
4519 x86_ret_imm (code, stack_to_pop);
4525 cfg->code_len = code - cfg->native_code;
4527 g_assert (cfg->code_len < cfg->code_size);
4531 mono_arch_emit_exceptions (MonoCompile *cfg)
4533 MonoJumpInfo *patch_info;
4536 MonoClass *exc_classes [16];
4537 guint8 *exc_throw_start [16], *exc_throw_end [16];
4541 /* Compute needed space */
4542 for (patch_info = cfg->patch_info; patch_info; patch_info = patch_info->next) {
4543 if (patch_info->type == MONO_PATCH_INFO_EXC)
4548 * make sure we have enough space for exceptions
4549 * 16 is the size of two push_imm instructions and a call
4551 if (cfg->compile_aot)
4552 code_size = exc_count * 32;
4554 code_size = exc_count * 16;
4556 while (cfg->code_len + code_size > (cfg->code_size - 16)) {
4557 cfg->code_size *= 2;
4558 cfg->native_code = g_realloc (cfg->native_code, cfg->code_size);
4559 mono_jit_stats.code_reallocs++;
4562 code = cfg->native_code + cfg->code_len;
4565 for (patch_info = cfg->patch_info; patch_info; patch_info = patch_info->next) {
4566 switch (patch_info->type) {
4567 case MONO_PATCH_INFO_EXC: {
4568 MonoClass *exc_class;
4572 x86_patch (patch_info->ip.i + cfg->native_code, code);
4574 exc_class = mono_class_from_name (mono_defaults.corlib, "System", patch_info->data.name);
4575 g_assert (exc_class);
4576 throw_ip = patch_info->ip.i;
4578 /* Find a throw sequence for the same exception class */
4579 for (i = 0; i < nthrows; ++i)
4580 if (exc_classes [i] == exc_class)
4583 x86_push_imm (code, (exc_throw_end [i] - cfg->native_code) - throw_ip);
4584 x86_jump_code (code, exc_throw_start [i]);
4585 patch_info->type = MONO_PATCH_INFO_NONE;
4588 guint32 got_reg = X86_EAX;
4591 /* Compute size of code following the push <OFFSET> */
4592 if (cfg->compile_aot) {
4596 else if (cfg->got_var->opcode == OP_REGOFFSET)
4602 if ((code - cfg->native_code) - throw_ip < 126 - size) {
4603 /* Use the shorter form */
4605 x86_push_imm (code, 0);
4609 x86_push_imm (code, 0xf0f0f0f0);
4614 exc_classes [nthrows] = exc_class;
4615 exc_throw_start [nthrows] = code;
4618 if (cfg->compile_aot) {
4620 * Since the patches are generated by the back end, there is * no way to generate a got_var at this point.
4622 if (!cfg->got_var) {
4623 x86_call_imm (code, 0);
4624 mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_GOT_OFFSET, NULL);
4625 x86_pop_reg (code, X86_EAX);
4626 x86_alu_reg_imm (code, X86_ADD, X86_EAX, 0);
4629 if (cfg->got_var->opcode == OP_REGOFFSET)
4630 x86_mov_reg_membase (code, X86_EAX, cfg->got_var->inst_basereg, cfg->got_var->inst_offset, 4);
4632 got_reg = cfg->got_var->dreg;
4636 x86_push_imm (code, exc_class->type_token);
4637 patch_info->data.name = "mono_arch_throw_corlib_exception";
4638 patch_info->type = MONO_PATCH_INFO_INTERNAL_METHOD;
4639 patch_info->ip.i = code - cfg->native_code;
4640 if (cfg->compile_aot)
4641 x86_call_membase (code, got_reg, 0xf0f0f0f0);
4643 x86_call_code (code, 0);
4644 x86_push_imm (buf, (code - cfg->native_code) - throw_ip);
4649 exc_throw_end [nthrows] = code;
4661 cfg->code_len = code - cfg->native_code;
4663 g_assert (cfg->code_len < cfg->code_size);
4667 mono_arch_flush_icache (guint8 *code, gint size)
4673 mono_arch_flush_register_windows (void)
4677 #ifdef MONO_ARCH_SIGSEGV_ON_ALTSTACK
4680 setup_stack (MonoJitTlsData *tls)
4682 pthread_t self = pthread_self();
4683 pthread_attr_t attr;
4685 struct sigaltstack sa;
4686 guint8 *staddr = NULL;
4687 guint8 *current = (guint8*)&staddr;
4689 if (mono_running_on_valgrind ())
4692 /* Determine stack boundaries */
4693 pthread_attr_init( &attr );
4694 #ifdef HAVE_PTHREAD_GETATTR_NP
4695 pthread_getattr_np( self, &attr );
4697 #ifdef HAVE_PTHREAD_ATTR_GET_NP
4698 pthread_attr_get_np( self, &attr );
4700 pthread_attr_getstacksize( &attr, &stsize );
4702 #error "Not implemented"
4706 pthread_attr_getstack( &attr, (void**)&staddr, &stsize );
4711 g_assert ((current > staddr) && (current < staddr + stsize));
4713 tls->end_of_stack = staddr + stsize;
4716 * threads created by nptl does not seem to have a guard page, and
4717 * since the main thread is not created by us, we can't even set one.
4718 * Increasing stsize fools the SIGSEGV signal handler into thinking this
4719 * is a stack overflow exception.
4721 tls->stack_size = stsize + getpagesize ();
4723 /* Setup an alternate signal stack */
4724 tls->signal_stack = mmap (0, SIGNAL_STACK_SIZE, PROT_READ|PROT_WRITE|PROT_EXEC, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0);
4725 tls->signal_stack_size = SIGNAL_STACK_SIZE;
4727 g_assert (tls->signal_stack);
4729 sa.ss_sp = tls->signal_stack;
4730 sa.ss_size = SIGNAL_STACK_SIZE;
4731 sa.ss_flags = SS_ONSTACK;
4732 sigaltstack (&sa, NULL);
4738 * Support for fast access to the thread-local lmf structure using the GS
4739 * segment register on NPTL + kernel 2.6.x.
4742 static gboolean tls_offset_inited = FALSE;
4745 mono_arch_setup_jit_tls_data (MonoJitTlsData *tls)
4747 if (!tls_offset_inited) {
4748 if (!getenv ("MONO_NO_TLS")) {
4749 #ifdef PLATFORM_WIN32
4751 * We need to init this multiple times, since when we are first called, the key might not
4752 * be initialized yet.
4754 appdomain_tls_offset = mono_domain_get_tls_key ();
4755 lmf_tls_offset = mono_get_jit_tls_key ();
4756 thread_tls_offset = mono_thread_get_tls_key ();
4758 /* Only 64 tls entries can be accessed using inline code */
4759 if (appdomain_tls_offset >= 64)
4760 appdomain_tls_offset = -1;
4761 if (lmf_tls_offset >= 64)
4762 lmf_tls_offset = -1;
4763 if (thread_tls_offset >= 64)
4764 thread_tls_offset = -1;
4766 tls_offset_inited = TRUE;
4767 appdomain_tls_offset = mono_domain_get_tls_offset ();
4768 lmf_tls_offset = mono_get_lmf_tls_offset ();
4769 thread_tls_offset = mono_thread_get_tls_offset ();
4774 #ifdef MONO_ARCH_SIGSEGV_ON_ALTSTACK
4780 mono_arch_free_jit_tls_data (MonoJitTlsData *tls)
4782 #ifdef MONO_ARCH_SIGSEGV_ON_ALTSTACK
4783 struct sigaltstack sa;
4785 sa.ss_sp = tls->signal_stack;
4786 sa.ss_size = SIGNAL_STACK_SIZE;
4787 sa.ss_flags = SS_DISABLE;
4788 sigaltstack (&sa, NULL);
4790 if (tls->signal_stack)
4791 munmap (tls->signal_stack, SIGNAL_STACK_SIZE);
4796 mono_arch_emit_this_vret_args (MonoCompile *cfg, MonoCallInst *inst, int this_reg, int this_type, int vt_reg)
4799 /* add the this argument */
4800 if (this_reg != -1) {
4802 MONO_INST_NEW (cfg, this, OP_OUTARG);
4803 this->type = this_type;
4804 this->sreg1 = this_reg;
4805 mono_bblock_add_inst (cfg->cbb, this);
4809 CallInfo * cinfo = get_call_info (inst->signature, FALSE);
4812 if (cinfo->ret.storage == ArgValuetypeInReg) {
4814 * The valuetype is in EAX:EDX after the call, needs to be copied to
4815 * the stack. Save the address here, so the call instruction can
4818 MONO_INST_NEW (cfg, vtarg, OP_STORE_MEMBASE_REG);
4819 vtarg->inst_destbasereg = X86_ESP;
4820 vtarg->inst_offset = inst->stack_usage;
4821 vtarg->sreg1 = vt_reg;
4822 mono_bblock_add_inst (cfg->cbb, vtarg);
4826 MONO_INST_NEW (cfg, vtarg, OP_OUTARG);
4827 vtarg->type = STACK_MP;
4828 vtarg->sreg1 = vt_reg;
4829 mono_bblock_add_inst (cfg->cbb, vtarg);
4838 mono_arch_get_inst_for_method (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature *fsig, MonoInst **args)
4840 MonoInst *ins = NULL;
4842 if (cmethod->klass == mono_defaults.math_class) {
4843 if (strcmp (cmethod->name, "Sin") == 0) {
4844 MONO_INST_NEW (cfg, ins, OP_SIN);
4845 ins->inst_i0 = args [0];
4846 } else if (strcmp (cmethod->name, "Cos") == 0) {
4847 MONO_INST_NEW (cfg, ins, OP_COS);
4848 ins->inst_i0 = args [0];
4849 } else if (strcmp (cmethod->name, "Tan") == 0) {
4850 MONO_INST_NEW (cfg, ins, OP_TAN);
4851 ins->inst_i0 = args [0];
4852 } else if (strcmp (cmethod->name, "Atan") == 0) {
4853 MONO_INST_NEW (cfg, ins, OP_ATAN);
4854 ins->inst_i0 = args [0];
4855 } else if (strcmp (cmethod->name, "Sqrt") == 0) {
4856 MONO_INST_NEW (cfg, ins, OP_SQRT);
4857 ins->inst_i0 = args [0];
4858 } else if (strcmp (cmethod->name, "Abs") == 0 && fsig->params [0]->type == MONO_TYPE_R8) {
4859 MONO_INST_NEW (cfg, ins, OP_ABS);
4860 ins->inst_i0 = args [0];
4863 /* OP_FREM is not IEEE compatible */
4864 else if (strcmp (cmethod->name, "IEEERemainder") == 0) {
4865 MONO_INST_NEW (cfg, ins, OP_FREM);
4866 ins->inst_i0 = args [0];
4867 ins->inst_i1 = args [1];
4870 } else if(cmethod->klass->image == mono_defaults.corlib &&
4871 (strcmp (cmethod->klass->name_space, "System.Threading") == 0) &&
4872 (strcmp (cmethod->klass->name, "Interlocked") == 0)) {
4874 if (strcmp (cmethod->name, "Increment") == 0 && fsig->params [0]->type == MONO_TYPE_I4) {
4875 MonoInst *ins_iconst;
4877 MONO_INST_NEW (cfg, ins, OP_ATOMIC_ADD_NEW_I4);
4878 MONO_INST_NEW (cfg, ins_iconst, OP_ICONST);
4879 ins_iconst->inst_c0 = 1;
4881 ins->inst_i0 = args [0];
4882 ins->inst_i1 = ins_iconst;
4883 } else if (strcmp (cmethod->name, "Decrement") == 0 && fsig->params [0]->type == MONO_TYPE_I4) {
4884 MonoInst *ins_iconst;
4886 MONO_INST_NEW (cfg, ins, OP_ATOMIC_ADD_NEW_I4);
4887 MONO_INST_NEW (cfg, ins_iconst, OP_ICONST);
4888 ins_iconst->inst_c0 = -1;
4890 ins->inst_i0 = args [0];
4891 ins->inst_i1 = ins_iconst;
4892 } else if (strcmp (cmethod->name, "Exchange") == 0 && fsig->params [0]->type == MONO_TYPE_I4) {
4893 MONO_INST_NEW (cfg, ins, OP_ATOMIC_EXCHANGE_I4);
4895 ins->inst_i0 = args [0];
4896 ins->inst_i1 = args [1];
4897 } else if (strcmp (cmethod->name, "Add") == 0 && fsig->params [0]->type == MONO_TYPE_I4) {
4898 MONO_INST_NEW (cfg, ins, OP_ATOMIC_ADD_I4);
4900 ins->inst_i0 = args [0];
4901 ins->inst_i1 = args [1];
4910 mono_arch_print_tree (MonoInst *tree, int arity)
4915 MonoInst* mono_arch_get_domain_intrinsic (MonoCompile* cfg)
4919 if (appdomain_tls_offset == -1)
4922 MONO_INST_NEW (cfg, ins, OP_TLS_GET);
4923 ins->inst_offset = appdomain_tls_offset;
4927 MonoInst* mono_arch_get_thread_intrinsic (MonoCompile* cfg)
4931 if (thread_tls_offset == -1)
4934 MONO_INST_NEW (cfg, ins, OP_TLS_GET);
4935 ins->inst_offset = thread_tls_offset;
4940 mono_arch_get_patch_offset (guint8 *code)
4942 if ((code [0] == 0x8b) && (x86_modrm_mod (code [1]) == 0x2))
4944 else if ((code [0] == 0xba))
4946 else if ((code [0] == 0x68))
4949 else if ((code [0] == 0xff) && (x86_modrm_reg (code [1]) == 0x6))
4950 /* push <OFFSET>(<REG>) */
4952 else if ((code [0] == 0xff) && (x86_modrm_reg (code [1]) == 0x2))
4953 /* call *<OFFSET>(<REG>) */
4955 else if ((code [0] == 0xdd) || (code [0] == 0xd9))
4958 else if ((code [0] == 0x58) && (code [1] == 0x05))
4959 /* pop %eax; add <OFFSET>, %eax */
4961 else if ((code [0] >= 0x58) && (code [0] <= 0x58 + X86_NREG) && (code [1] == 0x81))
4962 /* pop <REG>; add <OFFSET>, <REG> */
4965 g_assert_not_reached ();
4971 mono_arch_get_vcall_slot_addr (guint8 *code, gpointer *regs)
4976 /* go to the start of the call instruction
4978 * address_byte = (m << 6) | (o << 3) | reg
4979 * call opcode: 0xff address_byte displacement
4981 * 0xff m=2,o=2 imm32
4984 if ((code [1] != 0xe8) && (code [3] == 0xff) && ((code [4] & 0x18) == 0x10) && ((code [4] >> 6) == 1)) {
4985 reg = code [4] & 0x07;
4986 disp = (signed char)code [5];
4988 if ((code [0] == 0xff) && ((code [1] & 0x18) == 0x10) && ((code [1] >> 6) == 2)) {
4989 reg = code [1] & 0x07;
4990 disp = *((gint32*)(code + 2));
4991 } else if ((code [1] == 0xe8)) {
4993 } else if ((code [4] == 0xff) && (((code [5] >> 6) & 0x3) == 0) && (((code [5] >> 3) & 0x7) == 2)) {
4995 * This is a interface call: should check the above code can't catch it earlier
4996 * 8b 40 30 mov 0x30(%eax),%eax
4997 * ff 10 call *(%eax)
5000 reg = code [5] & 0x07;
5006 return (gpointer*)(((gint32)(regs [reg])) + disp);
5010 mono_arch_get_delegate_method_ptr_addr (guint8* code, gpointer *regs)
5016 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)) {
5017 reg = x86_modrm_rm (code [1]);
5023 return (gpointer*)(((gint32)(regs [reg])) + disp);