2 * mini-x86.c: x86 backend for the Mono code generator
5 * Paolo Molaro (lupus@ximian.com)
6 * Dietmar Maurer (dietmar@ximian.com)
9 * (C) 2003 Ximian, Inc.
16 #include <mono/metadata/appdomain.h>
17 #include <mono/metadata/debug-helpers.h>
18 #include <mono/metadata/threads.h>
19 #include <mono/metadata/profiler-private.h>
20 #include <mono/utils/mono-math.h>
25 #include "cpu-pentium.h"
27 /* On windows, these hold the key returned by TlsAlloc () */
28 static gint lmf_tls_offset = -1;
29 static gint appdomain_tls_offset = -1;
30 static gint thread_tls_offset = -1;
33 /* TRUE by default until we add runtime detection of Xen */
34 static gboolean optimize_for_xen = TRUE;
36 #define optimize_for_xen 0
39 #define ALIGN_TO(val,align) ((((guint64)val) + ((align) - 1)) & ~((align) - 1))
44 /* Under windows, the default pinvoke calling convention is stdcall */
45 #define CALLCONV_IS_STDCALL(sig) ((((sig)->call_convention) == MONO_CALL_STDCALL) || ((sig)->pinvoke && ((sig)->call_convention) == MONO_CALL_DEFAULT))
47 #define CALLCONV_IS_STDCALL(sig) (((sig)->call_convention) == MONO_CALL_STDCALL)
50 #define NOT_IMPLEMENTED g_assert_not_reached ()
53 mono_arch_regname (int reg) {
55 case X86_EAX: return "%eax";
56 case X86_EBX: return "%ebx";
57 case X86_ECX: return "%ecx";
58 case X86_EDX: return "%edx";
59 case X86_ESP: return "%esp"; case X86_EBP: return "%ebp";
60 case X86_EDI: return "%edi";
61 case X86_ESI: return "%esi";
67 mono_arch_fregname (int reg) {
87 /* Only if storage == ArgValuetypeInReg */
88 ArgStorage pair_storage [2];
97 gboolean need_stack_align;
98 guint32 stack_align_amount;
106 #define FLOAT_PARAM_REGS 0
108 static X86_Reg_No param_regs [] = { 0 };
110 #ifdef PLATFORM_WIN32
111 static X86_Reg_No return_regs [] = { X86_EAX, X86_EDX };
115 add_general (guint32 *gr, guint32 *stack_size, ArgInfo *ainfo)
117 ainfo->offset = *stack_size;
119 if (*gr >= PARAM_REGS) {
120 ainfo->storage = ArgOnStack;
121 (*stack_size) += sizeof (gpointer);
124 ainfo->storage = ArgInIReg;
125 ainfo->reg = param_regs [*gr];
131 add_general_pair (guint32 *gr, guint32 *stack_size, ArgInfo *ainfo)
133 ainfo->offset = *stack_size;
135 g_assert (PARAM_REGS == 0);
137 ainfo->storage = ArgOnStack;
138 (*stack_size) += sizeof (gpointer) * 2;
142 add_float (guint32 *gr, guint32 *stack_size, ArgInfo *ainfo, gboolean is_double)
144 ainfo->offset = *stack_size;
146 if (*gr >= FLOAT_PARAM_REGS) {
147 ainfo->storage = ArgOnStack;
148 (*stack_size) += is_double ? 8 : 4;
151 /* A double register */
153 ainfo->storage = ArgInDoubleSSEReg;
155 ainfo->storage = ArgInFloatSSEReg;
163 add_valuetype (MonoMethodSignature *sig, ArgInfo *ainfo, MonoType *type,
165 guint32 *gr, guint32 *fr, guint32 *stack_size)
170 klass = mono_class_from_mono_type (type);
172 size = mono_type_native_stack_size (&klass->byval_arg, NULL);
174 size = mono_type_stack_size (&klass->byval_arg, NULL);
176 #ifdef PLATFORM_WIN32
177 if (sig->pinvoke && is_return) {
178 MonoMarshalType *info;
181 * the exact rules are not very well documented, the code below seems to work with the
182 * code generated by gcc 3.3.3 -mno-cygwin.
184 info = mono_marshal_load_type_info (klass);
187 ainfo->pair_storage [0] = ainfo->pair_storage [1] = ArgNone;
189 /* Special case structs with only a float member */
190 if ((info->native_size == 8) && (info->num_fields == 1) && (info->fields [0].field->type->type == MONO_TYPE_R8)) {
191 ainfo->storage = ArgValuetypeInReg;
192 ainfo->pair_storage [0] = ArgOnDoubleFpStack;
195 if ((info->native_size == 4) && (info->num_fields == 1) && (info->fields [0].field->type->type == MONO_TYPE_R4)) {
196 ainfo->storage = ArgValuetypeInReg;
197 ainfo->pair_storage [0] = ArgOnFloatFpStack;
200 if ((info->native_size == 1) || (info->native_size == 2) || (info->native_size == 4) || (info->native_size == 8)) {
201 ainfo->storage = ArgValuetypeInReg;
202 ainfo->pair_storage [0] = ArgInIReg;
203 ainfo->pair_regs [0] = return_regs [0];
204 if (info->native_size > 4) {
205 ainfo->pair_storage [1] = ArgInIReg;
206 ainfo->pair_regs [1] = return_regs [1];
213 ainfo->offset = *stack_size;
214 ainfo->storage = ArgOnStack;
215 *stack_size += ALIGN_TO (size, sizeof (gpointer));
221 * Obtain information about a call according to the calling convention.
222 * For x86 ELF, see the "System V Application Binary Interface Intel386
223 * Architecture Processor Supplment, Fourth Edition" document for more
225 * For x86 win32, see ???.
228 get_call_info (MonoMethodSignature *sig, gboolean is_pinvoke)
232 int n = sig->hasthis + sig->param_count;
233 guint32 stack_size = 0;
236 cinfo = g_malloc0 (sizeof (CallInfo) + (sizeof (ArgInfo) * n));
243 ret_type = mono_type_get_underlying_type (sig->ret);
244 switch (ret_type->type) {
245 case MONO_TYPE_BOOLEAN:
256 case MONO_TYPE_FNPTR:
257 case MONO_TYPE_CLASS:
258 case MONO_TYPE_OBJECT:
259 case MONO_TYPE_SZARRAY:
260 case MONO_TYPE_ARRAY:
261 case MONO_TYPE_STRING:
262 cinfo->ret.storage = ArgInIReg;
263 cinfo->ret.reg = X86_EAX;
267 cinfo->ret.storage = ArgInIReg;
268 cinfo->ret.reg = X86_EAX;
271 cinfo->ret.storage = ArgOnFloatFpStack;
274 cinfo->ret.storage = ArgOnDoubleFpStack;
276 case MONO_TYPE_GENERICINST:
277 if (!mono_type_generic_inst_is_valuetype (sig->ret)) {
278 cinfo->ret.storage = ArgInIReg;
279 cinfo->ret.reg = X86_EAX;
283 case MONO_TYPE_VALUETYPE: {
284 guint32 tmp_gr = 0, tmp_fr = 0, tmp_stacksize = 0;
286 add_valuetype (sig, &cinfo->ret, sig->ret, TRUE, &tmp_gr, &tmp_fr, &tmp_stacksize);
287 if (cinfo->ret.storage == ArgOnStack)
288 /* The caller passes the address where the value is stored */
289 add_general (&gr, &stack_size, &cinfo->ret);
292 case MONO_TYPE_TYPEDBYREF:
293 /* Same as a valuetype with size 24 */
294 add_general (&gr, &stack_size, &cinfo->ret);
298 cinfo->ret.storage = ArgNone;
301 g_error ("Can't handle as return value 0x%x", sig->ret->type);
307 add_general (&gr, &stack_size, cinfo->args + 0);
309 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG) && (n == 0)) {
311 fr = FLOAT_PARAM_REGS;
313 /* Emit the signature cookie just before the implicit arguments */
314 add_general (&gr, &stack_size, &cinfo->sig_cookie);
317 for (i = 0; i < sig->param_count; ++i) {
318 ArgInfo *ainfo = &cinfo->args [sig->hasthis + i];
321 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG) && (i == sig->sentinelpos)) {
322 /* We allways pass the sig cookie on the stack for simplicity */
324 * Prevent implicit arguments + the sig cookie from being passed
328 fr = FLOAT_PARAM_REGS;
330 /* Emit the signature cookie just before the implicit arguments */
331 add_general (&gr, &stack_size, &cinfo->sig_cookie);
334 if (sig->params [i]->byref) {
335 add_general (&gr, &stack_size, ainfo);
338 ptype = mono_type_get_underlying_type (sig->params [i]);
339 switch (ptype->type) {
340 case MONO_TYPE_BOOLEAN:
343 add_general (&gr, &stack_size, ainfo);
348 add_general (&gr, &stack_size, ainfo);
352 add_general (&gr, &stack_size, ainfo);
357 case MONO_TYPE_FNPTR:
358 case MONO_TYPE_CLASS:
359 case MONO_TYPE_OBJECT:
360 case MONO_TYPE_STRING:
361 case MONO_TYPE_SZARRAY:
362 case MONO_TYPE_ARRAY:
363 add_general (&gr, &stack_size, ainfo);
365 case MONO_TYPE_GENERICINST:
366 if (!mono_type_generic_inst_is_valuetype (sig->params [i])) {
367 add_general (&gr, &stack_size, ainfo);
371 case MONO_TYPE_VALUETYPE:
372 add_valuetype (sig, ainfo, sig->params [i], FALSE, &gr, &fr, &stack_size);
374 case MONO_TYPE_TYPEDBYREF:
375 stack_size += sizeof (MonoTypedRef);
376 ainfo->storage = ArgOnStack;
380 add_general_pair (&gr, &stack_size, ainfo);
383 add_float (&fr, &stack_size, ainfo, FALSE);
386 add_float (&fr, &stack_size, ainfo, TRUE);
389 g_error ("unexpected type 0x%x", ptype->type);
390 g_assert_not_reached ();
394 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG) && (n > 0) && (sig->sentinelpos == sig->param_count)) {
396 fr = FLOAT_PARAM_REGS;
398 /* Emit the signature cookie just before the implicit arguments */
399 add_general (&gr, &stack_size, &cinfo->sig_cookie);
402 #if defined(__APPLE__)
403 if ((stack_size % 16) != 0) {
404 cinfo->need_stack_align = TRUE;
405 stack_size += cinfo->stack_align_amount = 16-(stack_size % 16);
409 cinfo->stack_usage = stack_size;
410 cinfo->reg_usage = gr;
411 cinfo->freg_usage = fr;
416 * mono_arch_get_argument_info:
417 * @csig: a method signature
418 * @param_count: the number of parameters to consider
419 * @arg_info: an array to store the result infos
421 * Gathers information on parameters such as size, alignment and
422 * padding. arg_info should be large enought to hold param_count + 1 entries.
424 * Returns the size of the activation frame.
427 mono_arch_get_argument_info (MonoMethodSignature *csig, int param_count, MonoJitArgumentInfo *arg_info)
429 int k, frame_size = 0;
435 cinfo = get_call_info (csig, FALSE);
437 if (MONO_TYPE_ISSTRUCT (csig->ret) && (cinfo->ret.storage == ArgOnStack)) {
438 frame_size += sizeof (gpointer);
442 arg_info [0].offset = offset;
445 frame_size += sizeof (gpointer);
449 arg_info [0].size = frame_size;
451 for (k = 0; k < param_count; k++) {
454 size = mono_type_native_stack_size (csig->params [k], &align);
457 size = mono_type_stack_size (csig->params [k], &ialign);
461 /* ignore alignment for now */
464 frame_size += pad = (align - (frame_size & (align - 1))) & (align - 1);
465 arg_info [k].pad = pad;
467 arg_info [k + 1].pad = 0;
468 arg_info [k + 1].size = size;
470 arg_info [k + 1].offset = offset;
474 align = MONO_ARCH_FRAME_ALIGNMENT;
475 frame_size += pad = (align - (frame_size & (align - 1))) & (align - 1);
476 arg_info [k].pad = pad;
483 static const guchar cpuid_impl [] = {
484 0x55, /* push %ebp */
485 0x89, 0xe5, /* mov %esp,%ebp */
486 0x53, /* push %ebx */
487 0x8b, 0x45, 0x08, /* mov 0x8(%ebp),%eax */
488 0x0f, 0xa2, /* cpuid */
489 0x50, /* push %eax */
490 0x8b, 0x45, 0x10, /* mov 0x10(%ebp),%eax */
491 0x89, 0x18, /* mov %ebx,(%eax) */
492 0x8b, 0x45, 0x14, /* mov 0x14(%ebp),%eax */
493 0x89, 0x08, /* mov %ecx,(%eax) */
494 0x8b, 0x45, 0x18, /* mov 0x18(%ebp),%eax */
495 0x89, 0x10, /* mov %edx,(%eax) */
497 0x8b, 0x55, 0x0c, /* mov 0xc(%ebp),%edx */
498 0x89, 0x02, /* mov %eax,(%edx) */
504 typedef void (*CpuidFunc) (int id, int* p_eax, int* p_ebx, int* p_ecx, int* p_edx);
507 cpuid (int id, int* p_eax, int* p_ebx, int* p_ecx, int* p_edx)
511 __asm__ __volatile__ (
514 "movl %%eax, %%edx\n"
515 "xorl $0x200000, %%eax\n"
520 "xorl %%edx, %%eax\n"
521 "andl $0x200000, %%eax\n"
543 /* Have to use the code manager to get around WinXP DEP */
544 MonoCodeManager *codeman = mono_code_manager_new_dynamic ();
546 void *ptr = mono_code_manager_reserve (codeman, sizeof (cpuid_impl));
547 memcpy (ptr, cpuid_impl, sizeof (cpuid_impl));
549 func = (CpuidFunc)ptr;
550 func (id, p_eax, p_ebx, p_ecx, p_edx);
552 mono_code_manager_destroy (codeman);
555 * We use this approach because of issues with gcc and pic code, see:
556 * http://gcc.gnu.org/cgi-bin/gnatsweb.pl?cmd=view%20audit-trail&database=gcc&pr=7329
557 __asm__ __volatile__ ("cpuid"
558 : "=a" (*p_eax), "=b" (*p_ebx), "=c" (*p_ecx), "=d" (*p_edx)
567 * Initialize the cpu to execute managed code.
570 mono_arch_cpu_init (void)
572 /* spec compliance requires running with double precision */
576 __asm__ __volatile__ ("fnstcw %0\n": "=m" (fpcw));
577 fpcw &= ~X86_FPCW_PRECC_MASK;
578 fpcw |= X86_FPCW_PREC_DOUBLE;
579 __asm__ __volatile__ ("fldcw %0\n": : "m" (fpcw));
580 __asm__ __volatile__ ("fnstcw %0\n": "=m" (fpcw));
582 _control87 (_PC_53, MCW_PC);
587 * This function returns the optimizations supported on this cpu.
590 mono_arch_cpu_optimizazions (guint32 *exclude_mask)
592 int eax, ebx, ecx, edx;
596 /* Feature Flags function, flags returned in EDX. */
597 if (cpuid (1, &eax, &ebx, &ecx, &edx)) {
598 if (edx & (1 << 15)) {
599 opts |= MONO_OPT_CMOV;
601 opts |= MONO_OPT_FCMOV;
603 *exclude_mask |= MONO_OPT_FCMOV;
605 *exclude_mask |= MONO_OPT_CMOV;
611 * Determine whenever the trap whose info is in SIGINFO is caused by
615 mono_arch_is_int_overflow (void *sigctx, void *info)
620 mono_arch_sigctx_to_monoctx (sigctx, &ctx);
622 ip = (guint8*)ctx.eip;
624 if ((ip [0] == 0xf7) && (x86_modrm_mod (ip [1]) == 0x3) && (x86_modrm_reg (ip [1]) == 0x7)) {
628 switch (x86_modrm_rm (ip [1])) {
648 g_assert_not_reached ();
660 is_regsize_var (MonoType *t) {
663 switch (mono_type_get_underlying_type (t)->type) {
669 case MONO_TYPE_FNPTR:
671 case MONO_TYPE_OBJECT:
672 case MONO_TYPE_STRING:
673 case MONO_TYPE_CLASS:
674 case MONO_TYPE_SZARRAY:
675 case MONO_TYPE_ARRAY:
677 case MONO_TYPE_GENERICINST:
678 if (!mono_type_generic_inst_is_valuetype (t))
681 case MONO_TYPE_VALUETYPE:
688 mono_arch_get_allocatable_int_vars (MonoCompile *cfg)
693 for (i = 0; i < cfg->num_varinfo; i++) {
694 MonoInst *ins = cfg->varinfo [i];
695 MonoMethodVar *vmv = MONO_VARINFO (cfg, i);
698 if (vmv->range.first_use.abs_pos >= vmv->range.last_use.abs_pos)
701 if ((ins->flags & (MONO_INST_IS_DEAD|MONO_INST_VOLATILE|MONO_INST_INDIRECT)) ||
702 (ins->opcode != OP_LOCAL && ins->opcode != OP_ARG))
705 /* we dont allocate I1 to registers because there is no simply way to sign extend
706 * 8bit quantities in caller saved registers on x86 */
707 if (is_regsize_var (ins->inst_vtype) || (ins->inst_vtype->type == MONO_TYPE_BOOLEAN) ||
708 (ins->inst_vtype->type == MONO_TYPE_U1) || (ins->inst_vtype->type == MONO_TYPE_U2)||
709 (ins->inst_vtype->type == MONO_TYPE_I2) || (ins->inst_vtype->type == MONO_TYPE_CHAR)) {
710 g_assert (MONO_VARINFO (cfg, i)->reg == -1);
711 g_assert (i == vmv->idx);
712 vars = g_list_prepend (vars, vmv);
716 vars = mono_varlist_sort (cfg, vars, 0);
722 mono_arch_get_global_int_regs (MonoCompile *cfg)
726 /* we can use 3 registers for global allocation */
727 regs = g_list_prepend (regs, (gpointer)X86_EBX);
728 regs = g_list_prepend (regs, (gpointer)X86_ESI);
729 regs = g_list_prepend (regs, (gpointer)X86_EDI);
735 * mono_arch_regalloc_cost:
737 * Return the cost, in number of memory references, of the action of
738 * allocating the variable VMV into a register during global register
742 mono_arch_regalloc_cost (MonoCompile *cfg, MonoMethodVar *vmv)
744 MonoInst *ins = cfg->varinfo [vmv->idx];
746 if (cfg->method->save_lmf)
747 /* The register is already saved */
748 return (ins->opcode == OP_ARG) ? 1 : 0;
750 /* push+pop+possible load if it is an argument */
751 return (ins->opcode == OP_ARG) ? 3 : 2;
755 * Set var information according to the calling convention. X86 version.
756 * The locals var stuff should most likely be split in another method.
759 mono_arch_allocate_vars (MonoCompile *cfg)
761 MonoMethodSignature *sig;
762 MonoMethodHeader *header;
764 guint32 locals_stack_size, locals_stack_align;
769 header = mono_method_get_header (cfg->method);
770 sig = mono_method_signature (cfg->method);
772 cinfo = get_call_info (sig, FALSE);
774 cfg->frame_reg = MONO_ARCH_BASEREG;
777 /* Reserve space to save LMF and caller saved registers */
779 if (cfg->method->save_lmf) {
780 offset += sizeof (MonoLMF);
782 if (cfg->used_int_regs & (1 << X86_EBX)) {
786 if (cfg->used_int_regs & (1 << X86_EDI)) {
790 if (cfg->used_int_regs & (1 << X86_ESI)) {
795 switch (cinfo->ret.storage) {
796 case ArgValuetypeInReg:
797 /* Allocate a local to hold the result, the epilog will copy it to the correct place */
799 cfg->ret->opcode = OP_REGOFFSET;
800 cfg->ret->inst_basereg = X86_EBP;
801 cfg->ret->inst_offset = - offset;
807 /* Allocate locals */
808 offsets = mono_allocate_stack_slots (cfg, &locals_stack_size, &locals_stack_align);
809 if (locals_stack_align) {
810 offset += (locals_stack_align - 1);
811 offset &= ~(locals_stack_align - 1);
813 for (i = cfg->locals_start; i < cfg->num_varinfo; i++) {
814 if (offsets [i] != -1) {
815 MonoInst *inst = cfg->varinfo [i];
816 inst->opcode = OP_REGOFFSET;
817 inst->inst_basereg = X86_EBP;
818 inst->inst_offset = - (offset + offsets [i]);
819 //printf ("allocated local %d to ", i); mono_print_tree_nl (inst);
823 offset += locals_stack_size;
827 * Allocate arguments+return value
830 switch (cinfo->ret.storage) {
832 cfg->ret->opcode = OP_REGOFFSET;
833 cfg->ret->inst_basereg = X86_EBP;
834 cfg->ret->inst_offset = cinfo->ret.offset + ARGS_OFFSET;
836 case ArgValuetypeInReg:
839 cfg->ret->opcode = OP_REGVAR;
840 cfg->ret->inst_c0 = cinfo->ret.reg;
843 case ArgOnFloatFpStack:
844 case ArgOnDoubleFpStack:
847 g_assert_not_reached ();
850 if (sig->call_convention == MONO_CALL_VARARG) {
851 g_assert (cinfo->sig_cookie.storage == ArgOnStack);
852 cfg->sig_cookie = cinfo->sig_cookie.offset + ARGS_OFFSET;
855 for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
856 ArgInfo *ainfo = &cinfo->args [i];
857 inst = cfg->varinfo [i];
858 if (inst->opcode != OP_REGVAR) {
859 inst->opcode = OP_REGOFFSET;
860 inst->inst_basereg = X86_EBP;
862 inst->inst_offset = ainfo->offset + ARGS_OFFSET;
865 offset += (MONO_ARCH_FRAME_ALIGNMENT - 1);
866 offset &= ~(MONO_ARCH_FRAME_ALIGNMENT - 1);
868 cfg->stack_offset = offset;
874 mono_arch_create_vars (MonoCompile *cfg)
876 MonoMethodSignature *sig;
879 sig = mono_method_signature (cfg->method);
881 cinfo = get_call_info (sig, FALSE);
883 if (cinfo->ret.storage == ArgValuetypeInReg)
884 cfg->ret_var_is_local = TRUE;
889 /* Fixme: we need an alignment solution for enter_method and mono_arch_call_opcode,
890 * currently alignment in mono_arch_call_opcode is computed without arch_get_argument_info
894 * take the arguments and generate the arch-specific
895 * instructions to properly call the function in call.
896 * This includes pushing, moving arguments to the right register
900 mono_arch_call_opcode (MonoCompile *cfg, MonoBasicBlock* bb, MonoCallInst *call, int is_virtual) {
902 MonoMethodSignature *sig;
907 sig = call->signature;
908 n = sig->param_count + sig->hasthis;
910 cinfo = get_call_info (sig, FALSE);
912 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG))
913 sentinelpos = sig->sentinelpos + (is_virtual ? 1 : 0);
915 for (i = 0; i < n; ++i) {
916 ArgInfo *ainfo = cinfo->args + i;
918 /* Emit the signature cookie just before the implicit arguments */
919 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG) && (i == sentinelpos)) {
920 MonoMethodSignature *tmp_sig;
923 /* FIXME: Add support for signature tokens to AOT */
924 cfg->disable_aot = TRUE;
925 MONO_INST_NEW (cfg, arg, OP_OUTARG);
928 * mono_ArgIterator_Setup assumes the signature cookie is
929 * passed first and all the arguments which were before it are
930 * passed on the stack after the signature. So compensate by
931 * passing a different signature.
933 tmp_sig = mono_metadata_signature_dup (call->signature);
934 tmp_sig->param_count -= call->signature->sentinelpos;
935 tmp_sig->sentinelpos = 0;
936 memcpy (tmp_sig->params, call->signature->params + call->signature->sentinelpos, tmp_sig->param_count * sizeof (MonoType*));
938 MONO_INST_NEW (cfg, sig_arg, OP_ICONST);
939 sig_arg->inst_p0 = tmp_sig;
941 arg->inst_left = sig_arg;
942 arg->type = STACK_PTR;
943 /* prepend, so they get reversed */
944 arg->next = call->out_args;
945 call->out_args = arg;
948 if (is_virtual && i == 0) {
949 /* the argument will be attached to the call instrucion */
954 if (i >= sig->hasthis)
955 t = sig->params [i - sig->hasthis];
957 t = &mono_defaults.int_class->byval_arg;
958 t = mono_type_get_underlying_type (t);
960 MONO_INST_NEW (cfg, arg, OP_OUTARG);
962 arg->cil_code = in->cil_code;
964 arg->type = in->type;
965 /* prepend, so they get reversed */
966 arg->next = call->out_args;
967 call->out_args = arg;
969 if ((i >= sig->hasthis) && (MONO_TYPE_ISSTRUCT(t))) {
972 if (t->type == MONO_TYPE_TYPEDBYREF) {
973 size = sizeof (MonoTypedRef);
974 align = sizeof (gpointer);
978 size = mono_type_native_stack_size (&in->klass->byval_arg, &align);
981 size = mono_type_stack_size (&in->klass->byval_arg, &ialign);
984 arg->opcode = OP_OUTARG_VT;
985 arg->klass = in->klass;
986 arg->unused = sig->pinvoke;
987 arg->inst_imm = size;
990 switch (ainfo->storage) {
992 arg->opcode = OP_OUTARG;
994 if (t->type == MONO_TYPE_R4)
995 arg->opcode = OP_OUTARG_R4;
997 if (t->type == MONO_TYPE_R8)
998 arg->opcode = OP_OUTARG_R8;
1002 g_assert_not_reached ();
1008 if (sig->ret && MONO_TYPE_ISSTRUCT (sig->ret)) {
1009 if (cinfo->ret.storage == ArgValuetypeInReg) {
1010 MonoInst *zero_inst;
1012 * After the call, the struct is in registers, but needs to be saved to the memory pointed
1013 * to by vt_arg in this_vret_args. This means that vt_arg needs to be saved somewhere
1014 * before calling the function. So we add a dummy instruction to represent pushing the
1015 * struct return address to the stack. The return address will be saved to this stack slot
1016 * by the code emitted in this_vret_args.
1018 MONO_INST_NEW (cfg, arg, OP_OUTARG);
1019 MONO_INST_NEW (cfg, zero_inst, OP_ICONST);
1020 zero_inst->inst_p0 = 0;
1021 arg->inst_left = zero_inst;
1022 arg->type = STACK_PTR;
1023 /* prepend, so they get reversed */
1024 arg->next = call->out_args;
1025 call->out_args = arg;
1028 /* if the function returns a struct, the called method already does a ret $0x4 */
1029 if (sig->ret && MONO_TYPE_ISSTRUCT (sig->ret))
1030 cinfo->stack_usage -= 4;
1033 call->stack_usage = cinfo->stack_usage;
1035 #if defined(__APPLE__)
1036 if (cinfo->need_stack_align) {
1037 MONO_INST_NEW (cfg, arg, OP_X86_OUTARG_ALIGN_STACK);
1038 arg->inst_c0 = cinfo->stack_align_amount;
1039 arg->next = call->out_args;
1040 call->out_args = arg;
1050 * Allow tracing to work with this interface (with an optional argument)
1053 mono_arch_instrument_prolog (MonoCompile *cfg, void *func, void *p, gboolean enable_arguments)
1057 /* if some args are passed in registers, we need to save them here */
1058 x86_push_reg (code, X86_EBP);
1060 if (cfg->compile_aot) {
1061 x86_push_imm (code, cfg->method);
1062 x86_mov_reg_imm (code, X86_EAX, func);
1063 x86_call_reg (code, X86_EAX);
1065 mono_add_patch_info (cfg, code-cfg->native_code, MONO_PATCH_INFO_METHODCONST, cfg->method);
1066 x86_push_imm (code, cfg->method);
1067 mono_add_patch_info (cfg, code-cfg->native_code, MONO_PATCH_INFO_ABS, func);
1068 x86_call_code (code, 0);
1070 x86_alu_reg_imm (code, X86_ADD, X86_ESP, 8);
1084 mono_arch_instrument_epilog (MonoCompile *cfg, void *func, void *p, gboolean enable_arguments)
1087 int arg_size = 0, save_mode = SAVE_NONE;
1088 MonoMethod *method = cfg->method;
1090 switch (mono_type_get_underlying_type (mono_method_signature (method)->ret)->type) {
1091 case MONO_TYPE_VOID:
1092 /* special case string .ctor icall */
1093 if (strcmp (".ctor", method->name) && method->klass == mono_defaults.string_class)
1094 save_mode = SAVE_EAX;
1096 save_mode = SAVE_NONE;
1100 save_mode = SAVE_EAX_EDX;
1104 save_mode = SAVE_FP;
1106 case MONO_TYPE_GENERICINST:
1107 if (!mono_type_generic_inst_is_valuetype (mono_method_signature (method)->ret)) {
1108 save_mode = SAVE_EAX;
1112 case MONO_TYPE_VALUETYPE:
1113 save_mode = SAVE_STRUCT;
1116 save_mode = SAVE_EAX;
1120 switch (save_mode) {
1122 x86_push_reg (code, X86_EDX);
1123 x86_push_reg (code, X86_EAX);
1124 if (enable_arguments) {
1125 x86_push_reg (code, X86_EDX);
1126 x86_push_reg (code, X86_EAX);
1131 x86_push_reg (code, X86_EAX);
1132 if (enable_arguments) {
1133 x86_push_reg (code, X86_EAX);
1138 x86_alu_reg_imm (code, X86_SUB, X86_ESP, 8);
1139 x86_fst_membase (code, X86_ESP, 0, TRUE, TRUE);
1140 if (enable_arguments) {
1141 x86_alu_reg_imm (code, X86_SUB, X86_ESP, 8);
1142 x86_fst_membase (code, X86_ESP, 0, TRUE, TRUE);
1147 if (enable_arguments) {
1148 x86_push_membase (code, X86_EBP, 8);
1157 if (cfg->compile_aot) {
1158 x86_push_imm (code, method);
1159 x86_mov_reg_imm (code, X86_EAX, func);
1160 x86_call_reg (code, X86_EAX);
1162 mono_add_patch_info (cfg, code-cfg->native_code, MONO_PATCH_INFO_METHODCONST, method);
1163 x86_push_imm (code, method);
1164 mono_add_patch_info (cfg, code-cfg->native_code, MONO_PATCH_INFO_ABS, func);
1165 x86_call_code (code, 0);
1167 x86_alu_reg_imm (code, X86_ADD, X86_ESP, arg_size + 4);
1169 switch (save_mode) {
1171 x86_pop_reg (code, X86_EAX);
1172 x86_pop_reg (code, X86_EDX);
1175 x86_pop_reg (code, X86_EAX);
1178 x86_fld_membase (code, X86_ESP, 0, TRUE);
1179 x86_alu_reg_imm (code, X86_ADD, X86_ESP, 8);
1189 #define EMIT_COND_BRANCH(ins,cond,sign) \
1190 if (ins->flags & MONO_INST_BRLABEL) { \
1191 if (ins->inst_i0->inst_c0) { \
1192 x86_branch (code, cond, cfg->native_code + ins->inst_i0->inst_c0, sign); \
1194 mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_LABEL, ins->inst_i0); \
1195 if ((cfg->opt & MONO_OPT_BRANCH) && \
1196 x86_is_imm8 (ins->inst_i0->inst_c1 - cpos)) \
1197 x86_branch8 (code, cond, 0, sign); \
1199 x86_branch32 (code, cond, 0, sign); \
1202 if (ins->inst_true_bb->native_offset) { \
1203 x86_branch (code, cond, cfg->native_code + ins->inst_true_bb->native_offset, sign); \
1205 mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_BB, ins->inst_true_bb); \
1206 if ((cfg->opt & MONO_OPT_BRANCH) && \
1207 x86_is_imm8 (ins->inst_true_bb->max_offset - cpos)) \
1208 x86_branch8 (code, cond, 0, sign); \
1210 x86_branch32 (code, cond, 0, sign); \
1215 * Emit an exception if condition is fail and
1216 * if possible do a directly branch to target
1218 #define EMIT_COND_SYSTEM_EXCEPTION(cond,signed,exc_name) \
1220 MonoInst *tins = mono_branch_optimize_exception_target (cfg, bb, exc_name); \
1221 if (tins == NULL) { \
1222 mono_add_patch_info (cfg, code - cfg->native_code, \
1223 MONO_PATCH_INFO_EXC, exc_name); \
1224 x86_branch32 (code, cond, 0, signed); \
1226 EMIT_COND_BRANCH (tins, cond, signed); \
1230 #define EMIT_FPCOMPARE(code) do { \
1231 x86_fcompp (code); \
1232 x86_fnstsw (code); \
1237 emit_call (MonoCompile *cfg, guint8 *code, guint32 patch_type, gconstpointer data)
1239 mono_add_patch_info (cfg, code - cfg->native_code, patch_type, data);
1240 x86_call_code (code, 0);
1245 /* FIXME: Add more instructions */
1246 #define INST_IGNORES_CFLAGS(ins) (((ins)->opcode == CEE_BR) || ((ins)->opcode == OP_STORE_MEMBASE_IMM) || ((ins)->opcode == OP_STOREI4_MEMBASE_REG))
1249 peephole_pass (MonoCompile *cfg, MonoBasicBlock *bb)
1251 MonoInst *ins, *last_ins = NULL;
1256 switch (ins->opcode) {
1258 /* reg = 0 -> XOR (reg, reg) */
1259 /* XOR sets cflags on x86, so we cant do it always */
1260 if (ins->inst_c0 == 0 && ins->next && INST_IGNORES_CFLAGS (ins->next)) {
1261 ins->opcode = CEE_XOR;
1262 ins->sreg1 = ins->dreg;
1263 ins->sreg2 = ins->dreg;
1267 /* remove unnecessary multiplication with 1 */
1268 if (ins->inst_imm == 1) {
1269 if (ins->dreg != ins->sreg1) {
1270 ins->opcode = OP_MOVE;
1272 last_ins->next = ins->next;
1278 case OP_COMPARE_IMM:
1279 /* OP_COMPARE_IMM (reg, 0)
1281 * OP_X86_TEST_NULL (reg)
1284 ins->opcode = OP_X86_TEST_NULL;
1286 case OP_X86_COMPARE_MEMBASE_IMM:
1288 * OP_STORE_MEMBASE_REG reg, offset(basereg)
1289 * OP_X86_COMPARE_MEMBASE_IMM offset(basereg), imm
1291 * OP_STORE_MEMBASE_REG reg, offset(basereg)
1292 * OP_COMPARE_IMM reg, imm
1294 * Note: if imm = 0 then OP_COMPARE_IMM replaced with OP_X86_TEST_NULL
1296 if (last_ins && (last_ins->opcode == OP_STOREI4_MEMBASE_REG) &&
1297 ins->inst_basereg == last_ins->inst_destbasereg &&
1298 ins->inst_offset == last_ins->inst_offset) {
1299 ins->opcode = OP_COMPARE_IMM;
1300 ins->sreg1 = last_ins->sreg1;
1302 /* check if we can remove cmp reg,0 with test null */
1304 ins->opcode = OP_X86_TEST_NULL;
1308 case OP_LOAD_MEMBASE:
1309 case OP_LOADI4_MEMBASE:
1311 * Note: if reg1 = reg2 the load op is removed
1313 * OP_STORE_MEMBASE_REG reg1, offset(basereg)
1314 * OP_LOAD_MEMBASE offset(basereg), reg2
1316 * OP_STORE_MEMBASE_REG reg1, offset(basereg)
1317 * OP_MOVE reg1, reg2
1319 if (last_ins && (last_ins->opcode == OP_STOREI4_MEMBASE_REG
1320 || last_ins->opcode == OP_STORE_MEMBASE_REG) &&
1321 ins->inst_basereg == last_ins->inst_destbasereg &&
1322 ins->inst_offset == last_ins->inst_offset) {
1323 if (ins->dreg == last_ins->sreg1) {
1324 last_ins->next = ins->next;
1328 //static int c = 0; printf ("MATCHX %s %d\n", cfg->method->name,c++);
1329 ins->opcode = OP_MOVE;
1330 ins->sreg1 = last_ins->sreg1;
1334 * Note: reg1 must be different from the basereg in the second load
1335 * Note: if reg1 = reg2 is equal then second load is removed
1337 * OP_LOAD_MEMBASE offset(basereg), reg1
1338 * OP_LOAD_MEMBASE offset(basereg), reg2
1340 * OP_LOAD_MEMBASE offset(basereg), reg1
1341 * OP_MOVE reg1, reg2
1343 } if (last_ins && (last_ins->opcode == OP_LOADI4_MEMBASE
1344 || last_ins->opcode == OP_LOAD_MEMBASE) &&
1345 ins->inst_basereg != last_ins->dreg &&
1346 ins->inst_basereg == last_ins->inst_basereg &&
1347 ins->inst_offset == last_ins->inst_offset) {
1349 if (ins->dreg == last_ins->dreg) {
1350 last_ins->next = ins->next;
1354 ins->opcode = OP_MOVE;
1355 ins->sreg1 = last_ins->dreg;
1358 //g_assert_not_reached ();
1362 * OP_STORE_MEMBASE_IMM imm, offset(basereg)
1363 * OP_LOAD_MEMBASE offset(basereg), reg
1365 * OP_STORE_MEMBASE_IMM imm, offset(basereg)
1366 * OP_ICONST reg, imm
1368 } else if (last_ins && (last_ins->opcode == OP_STOREI4_MEMBASE_IMM
1369 || last_ins->opcode == OP_STORE_MEMBASE_IMM) &&
1370 ins->inst_basereg == last_ins->inst_destbasereg &&
1371 ins->inst_offset == last_ins->inst_offset) {
1372 //static int c = 0; printf ("MATCHX %s %d\n", cfg->method->name,c++);
1373 ins->opcode = OP_ICONST;
1374 ins->inst_c0 = last_ins->inst_imm;
1375 g_assert_not_reached (); // check this rule
1379 case OP_LOADU1_MEMBASE:
1380 case OP_LOADI1_MEMBASE:
1382 * OP_STORE_MEMBASE_REG reg1, offset(basereg)
1383 * OP_LOAD_MEMBASE offset(basereg), reg2
1385 * OP_STORE_MEMBASE_REG reg1, offset(basereg)
1386 * CONV_I2/U2 reg1, reg2
1388 if (last_ins && X86_IS_BYTE_REG (last_ins->sreg1) &&
1389 (last_ins->opcode == OP_STOREI1_MEMBASE_REG) &&
1390 ins->inst_basereg == last_ins->inst_destbasereg &&
1391 ins->inst_offset == last_ins->inst_offset) {
1392 ins->opcode = (ins->opcode == OP_LOADI1_MEMBASE) ? CEE_CONV_I1 : CEE_CONV_U1;
1393 ins->sreg1 = last_ins->sreg1;
1396 case OP_LOADU2_MEMBASE:
1397 case OP_LOADI2_MEMBASE:
1399 * OP_STORE_MEMBASE_REG reg1, offset(basereg)
1400 * OP_LOAD_MEMBASE offset(basereg), reg2
1402 * OP_STORE_MEMBASE_REG reg1, offset(basereg)
1403 * CONV_I2/U2 reg1, reg2
1405 if (last_ins && (last_ins->opcode == OP_STOREI2_MEMBASE_REG) &&
1406 ins->inst_basereg == last_ins->inst_destbasereg &&
1407 ins->inst_offset == last_ins->inst_offset) {
1408 ins->opcode = (ins->opcode == OP_LOADI2_MEMBASE) ? CEE_CONV_I2 : CEE_CONV_U2;
1409 ins->sreg1 = last_ins->sreg1;
1420 if (ins->dreg == ins->sreg1) {
1422 last_ins->next = ins->next;
1429 * OP_MOVE sreg, dreg
1430 * OP_MOVE dreg, sreg
1432 if (last_ins && last_ins->opcode == OP_MOVE &&
1433 ins->sreg1 == last_ins->dreg &&
1434 ins->dreg == last_ins->sreg1) {
1435 last_ins->next = ins->next;
1441 case OP_X86_PUSH_MEMBASE:
1442 if (last_ins && (last_ins->opcode == OP_STOREI4_MEMBASE_REG ||
1443 last_ins->opcode == OP_STORE_MEMBASE_REG) &&
1444 ins->inst_basereg == last_ins->inst_destbasereg &&
1445 ins->inst_offset == last_ins->inst_offset) {
1446 ins->opcode = OP_X86_PUSH;
1447 ins->sreg1 = last_ins->sreg1;
1454 bb->last_ins = last_ins;
1458 branch_cc_table [] = {
1459 X86_CC_EQ, X86_CC_GE, X86_CC_GT, X86_CC_LE, X86_CC_LT,
1460 X86_CC_NE, X86_CC_GE, X86_CC_GT, X86_CC_LE, X86_CC_LT,
1461 X86_CC_O, X86_CC_NO, X86_CC_C, X86_CC_NC
1464 static const char*const * ins_spec = pentium_desc;
1466 /*#include "cprop.c"*/
1468 mono_arch_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
1470 mono_local_regalloc (cfg, bb);
1473 static unsigned char*
1474 emit_float_to_int (MonoCompile *cfg, guchar *code, int dreg, int size, gboolean is_signed)
1476 x86_alu_reg_imm (code, X86_SUB, X86_ESP, 4);
1477 x86_fnstcw_membase(code, X86_ESP, 0);
1478 x86_mov_reg_membase (code, dreg, X86_ESP, 0, 2);
1479 x86_alu_reg_imm (code, X86_OR, dreg, 0xc00);
1480 x86_mov_membase_reg (code, X86_ESP, 2, dreg, 2);
1481 x86_fldcw_membase (code, X86_ESP, 2);
1483 x86_alu_reg_imm (code, X86_SUB, X86_ESP, 8);
1484 x86_fist_pop_membase (code, X86_ESP, 0, TRUE);
1485 x86_pop_reg (code, dreg);
1486 /* FIXME: need the high register
1487 * x86_pop_reg (code, dreg_high);
1490 x86_push_reg (code, X86_EAX); // SP = SP - 4
1491 x86_fist_pop_membase (code, X86_ESP, 0, FALSE);
1492 x86_pop_reg (code, dreg);
1494 x86_fldcw_membase (code, X86_ESP, 0);
1495 x86_alu_reg_imm (code, X86_ADD, X86_ESP, 4);
1498 x86_widen_reg (code, dreg, dreg, is_signed, FALSE);
1500 x86_widen_reg (code, dreg, dreg, is_signed, TRUE);
1504 static unsigned char*
1505 mono_emit_stack_alloc (guchar *code, MonoInst* tree)
1507 int sreg = tree->sreg1;
1508 int need_touch = FALSE;
1510 #if defined(PLATFORM_WIN32) || defined(MONO_ARCH_SIGSEGV_ON_ALTSTACK)
1519 * If requested stack size is larger than one page,
1520 * perform stack-touch operation
1523 * Generate stack probe code.
1524 * Under Windows, it is necessary to allocate one page at a time,
1525 * "touching" stack after each successful sub-allocation. This is
1526 * because of the way stack growth is implemented - there is a
1527 * guard page before the lowest stack page that is currently commited.
1528 * Stack normally grows sequentially so OS traps access to the
1529 * guard page and commits more pages when needed.
1531 x86_test_reg_imm (code, sreg, ~0xFFF);
1532 br[0] = code; x86_branch8 (code, X86_CC_Z, 0, FALSE);
1534 br[2] = code; /* loop */
1535 x86_alu_reg_imm (code, X86_SUB, X86_ESP, 0x1000);
1536 x86_test_membase_reg (code, X86_ESP, 0, X86_ESP);
1539 * By the end of the loop, sreg2 is smaller than 0x1000, so the init routine
1540 * that follows only initializes the last part of the area.
1542 /* Same as the init code below with size==0x1000 */
1543 if (tree->flags & MONO_INST_INIT) {
1544 x86_push_reg (code, X86_EAX);
1545 x86_push_reg (code, X86_ECX);
1546 x86_push_reg (code, X86_EDI);
1547 x86_mov_reg_imm (code, X86_ECX, (0x1000 >> 2));
1548 x86_alu_reg_reg (code, X86_XOR, X86_EAX, X86_EAX);
1549 x86_lea_membase (code, X86_EDI, X86_ESP, 12);
1551 x86_prefix (code, X86_REP_PREFIX);
1553 x86_pop_reg (code, X86_EDI);
1554 x86_pop_reg (code, X86_ECX);
1555 x86_pop_reg (code, X86_EAX);
1558 x86_alu_reg_imm (code, X86_SUB, sreg, 0x1000);
1559 x86_alu_reg_imm (code, X86_CMP, sreg, 0x1000);
1560 br[3] = code; x86_branch8 (code, X86_CC_AE, 0, FALSE);
1561 x86_patch (br[3], br[2]);
1562 x86_test_reg_reg (code, sreg, sreg);
1563 br[4] = code; x86_branch8 (code, X86_CC_Z, 0, FALSE);
1564 x86_alu_reg_reg (code, X86_SUB, X86_ESP, sreg);
1566 br[1] = code; x86_jump8 (code, 0);
1568 x86_patch (br[0], code);
1569 x86_alu_reg_reg (code, X86_SUB, X86_ESP, sreg);
1570 x86_patch (br[1], code);
1571 x86_patch (br[4], code);
1574 x86_alu_reg_reg (code, X86_SUB, X86_ESP, tree->sreg1);
1576 if (tree->flags & MONO_INST_INIT) {
1578 if (tree->dreg != X86_EAX && sreg != X86_EAX) {
1579 x86_push_reg (code, X86_EAX);
1582 if (tree->dreg != X86_ECX && sreg != X86_ECX) {
1583 x86_push_reg (code, X86_ECX);
1586 if (tree->dreg != X86_EDI && sreg != X86_EDI) {
1587 x86_push_reg (code, X86_EDI);
1591 x86_shift_reg_imm (code, X86_SHR, sreg, 2);
1592 if (sreg != X86_ECX)
1593 x86_mov_reg_reg (code, X86_ECX, sreg, 4);
1594 x86_alu_reg_reg (code, X86_XOR, X86_EAX, X86_EAX);
1596 x86_lea_membase (code, X86_EDI, X86_ESP, offset);
1598 x86_prefix (code, X86_REP_PREFIX);
1601 if (tree->dreg != X86_EDI && sreg != X86_EDI)
1602 x86_pop_reg (code, X86_EDI);
1603 if (tree->dreg != X86_ECX && sreg != X86_ECX)
1604 x86_pop_reg (code, X86_ECX);
1605 if (tree->dreg != X86_EAX && sreg != X86_EAX)
1606 x86_pop_reg (code, X86_EAX);
1613 emit_move_return_value (MonoCompile *cfg, MonoInst *ins, guint8 *code)
1618 /* Move return value to the target register */
1619 switch (ins->opcode) {
1622 case OP_CALL_MEMBASE:
1623 if (ins->dreg != X86_EAX)
1624 x86_mov_reg_reg (code, ins->dreg, X86_EAX, 4);
1628 case OP_VCALL_MEMBASE:
1629 cinfo = get_call_info (((MonoCallInst*)ins)->signature, FALSE);
1630 if (cinfo->ret.storage == ArgValuetypeInReg) {
1631 /* Pop the destination address from the stack */
1632 x86_pop_reg (code, X86_ECX);
1634 for (quad = 0; quad < 2; quad ++) {
1635 switch (cinfo->ret.pair_storage [quad]) {
1637 g_assert (cinfo->ret.pair_regs [quad] != X86_ECX);
1638 x86_mov_membase_reg (code, X86_ECX, (quad * sizeof (gpointer)), cinfo->ret.pair_regs [quad], sizeof (gpointer));
1643 g_assert_not_reached ();
1657 * @code: buffer to store code to
1658 * @dreg: hard register where to place the result
1659 * @tls_offset: offset info
1661 * emit_tls_get emits in @code the native code that puts in the dreg register
1662 * the item in the thread local storage identified by tls_offset.
1664 * Returns: a pointer to the end of the stored code
1667 emit_tls_get (guint8* code, int dreg, int tls_offset)
1669 #ifdef PLATFORM_WIN32
1671 * See the Under the Hood article in the May 1996 issue of Microsoft Systems
1672 * Journal and/or a disassembly of the TlsGet () function.
1674 g_assert (tls_offset < 64);
1675 x86_prefix (code, X86_FS_PREFIX);
1676 x86_mov_reg_mem (code, dreg, 0x18, 4);
1677 /* Dunno what this does but TlsGetValue () contains it */
1678 x86_alu_membase_imm (code, X86_AND, dreg, 0x34, 0);
1679 x86_mov_reg_membase (code, dreg, dreg, 3600 + (tls_offset * 4), 4);
1681 if (optimize_for_xen) {
1682 x86_prefix (code, X86_GS_PREFIX);
1683 x86_mov_reg_mem (code, dreg, 0, 4);
1684 x86_mov_reg_membase (code, dreg, dreg, tls_offset, 4);
1686 x86_prefix (code, X86_GS_PREFIX);
1687 x86_mov_reg_mem (code, dreg, tls_offset, 4);
1693 #define REAL_PRINT_REG(text,reg) \
1694 mono_assert (reg >= 0); \
1695 x86_push_reg (code, X86_EAX); \
1696 x86_push_reg (code, X86_EDX); \
1697 x86_push_reg (code, X86_ECX); \
1698 x86_push_reg (code, reg); \
1699 x86_push_imm (code, reg); \
1700 x86_push_imm (code, text " %d %p\n"); \
1701 x86_mov_reg_imm (code, X86_EAX, printf); \
1702 x86_call_reg (code, X86_EAX); \
1703 x86_alu_reg_imm (code, X86_ADD, X86_ESP, 3*4); \
1704 x86_pop_reg (code, X86_ECX); \
1705 x86_pop_reg (code, X86_EDX); \
1706 x86_pop_reg (code, X86_EAX);
1708 /* benchmark and set based on cpu */
1709 #define LOOP_ALIGNMENT 8
1710 #define bb_is_loop_start(bb) ((bb)->loop_body_start && (bb)->nesting)
1713 mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
1718 guint8 *code = cfg->native_code + cfg->code_len;
1719 MonoInst *last_ins = NULL;
1720 guint last_offset = 0;
1723 if (cfg->opt & MONO_OPT_PEEPHOLE)
1724 peephole_pass (cfg, bb);
1726 if (cfg->opt & MONO_OPT_LOOP) {
1727 int pad, align = LOOP_ALIGNMENT;
1728 /* set alignment depending on cpu */
1729 if (bb_is_loop_start (bb) && (pad = (cfg->code_len & (align - 1)))) {
1731 /*g_print ("adding %d pad at %x to loop in %s\n", pad, cfg->code_len, cfg->method->name);*/
1732 x86_padding (code, pad);
1733 cfg->code_len += pad;
1734 bb->native_offset = cfg->code_len;
1738 if (cfg->verbose_level > 2)
1739 g_print ("Basic block %d starting at offset 0x%x\n", bb->block_num, bb->native_offset);
1741 cpos = bb->max_offset;
1743 if (cfg->prof_options & MONO_PROFILE_COVERAGE) {
1744 MonoProfileCoverageInfo *cov = cfg->coverage_info;
1745 g_assert (!cfg->compile_aot);
1748 cov->data [bb->dfn].cil_code = bb->cil_code;
1749 /* this is not thread save, but good enough */
1750 x86_inc_mem (code, &cov->data [bb->dfn].count);
1753 offset = code - cfg->native_code;
1755 mono_debug_open_block (cfg, bb, offset);
1759 offset = code - cfg->native_code;
1761 max_len = ((guint8 *)ins_spec [ins->opcode])[MONO_INST_LEN];
1763 if (offset > (cfg->code_size - max_len - 16)) {
1764 cfg->code_size *= 2;
1765 cfg->native_code = g_realloc (cfg->native_code, cfg->code_size);
1766 code = cfg->native_code + offset;
1767 mono_jit_stats.code_reallocs++;
1770 mono_debug_record_line_number (cfg, ins, offset);
1772 switch (ins->opcode) {
1774 x86_mul_reg (code, ins->sreg2, TRUE);
1777 x86_mul_reg (code, ins->sreg2, FALSE);
1779 case OP_X86_SETEQ_MEMBASE:
1780 case OP_X86_SETNE_MEMBASE:
1781 x86_set_membase (code, ins->opcode == OP_X86_SETEQ_MEMBASE ? X86_CC_EQ : X86_CC_NE,
1782 ins->inst_basereg, ins->inst_offset, TRUE);
1784 case OP_STOREI1_MEMBASE_IMM:
1785 x86_mov_membase_imm (code, ins->inst_destbasereg, ins->inst_offset, ins->inst_imm, 1);
1787 case OP_STOREI2_MEMBASE_IMM:
1788 x86_mov_membase_imm (code, ins->inst_destbasereg, ins->inst_offset, ins->inst_imm, 2);
1790 case OP_STORE_MEMBASE_IMM:
1791 case OP_STOREI4_MEMBASE_IMM:
1792 x86_mov_membase_imm (code, ins->inst_destbasereg, ins->inst_offset, ins->inst_imm, 4);
1794 case OP_STOREI1_MEMBASE_REG:
1795 x86_mov_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, ins->sreg1, 1);
1797 case OP_STOREI2_MEMBASE_REG:
1798 x86_mov_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, ins->sreg1, 2);
1800 case OP_STORE_MEMBASE_REG:
1801 case OP_STOREI4_MEMBASE_REG:
1802 x86_mov_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, ins->sreg1, 4);
1807 x86_mov_reg_mem (code, ins->dreg, ins->inst_p0, 4);
1810 x86_mov_reg_imm (code, ins->dreg, ins->inst_p0);
1811 x86_mov_reg_membase (code, ins->dreg, ins->dreg, 0, 4);
1813 case OP_LOAD_MEMBASE:
1814 case OP_LOADI4_MEMBASE:
1815 case OP_LOADU4_MEMBASE:
1816 x86_mov_reg_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, 4);
1818 case OP_LOADU1_MEMBASE:
1819 x86_widen_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, FALSE, FALSE);
1821 case OP_LOADI1_MEMBASE:
1822 x86_widen_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, TRUE, FALSE);
1824 case OP_LOADU2_MEMBASE:
1825 x86_widen_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, FALSE, TRUE);
1827 case OP_LOADI2_MEMBASE:
1828 x86_widen_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, TRUE, TRUE);
1831 x86_widen_reg (code, ins->dreg, ins->sreg1, TRUE, FALSE);
1834 x86_widen_reg (code, ins->dreg, ins->sreg1, TRUE, TRUE);
1837 x86_widen_reg (code, ins->dreg, ins->sreg1, FALSE, FALSE);
1840 x86_widen_reg (code, ins->dreg, ins->sreg1, FALSE, TRUE);
1843 x86_alu_reg_reg (code, X86_CMP, ins->sreg1, ins->sreg2);
1845 case OP_COMPARE_IMM:
1846 x86_alu_reg_imm (code, X86_CMP, ins->sreg1, ins->inst_imm);
1848 case OP_X86_COMPARE_MEMBASE_REG:
1849 x86_alu_membase_reg (code, X86_CMP, ins->inst_basereg, ins->inst_offset, ins->sreg2);
1851 case OP_X86_COMPARE_MEMBASE_IMM:
1852 x86_alu_membase_imm (code, X86_CMP, ins->inst_basereg, ins->inst_offset, ins->inst_imm);
1854 case OP_X86_COMPARE_MEMBASE8_IMM:
1855 x86_alu_membase8_imm (code, X86_CMP, ins->inst_basereg, ins->inst_offset, ins->inst_imm);
1857 case OP_X86_COMPARE_REG_MEMBASE:
1858 x86_alu_reg_membase (code, X86_CMP, ins->sreg1, ins->sreg2, ins->inst_offset);
1860 case OP_X86_COMPARE_MEM_IMM:
1861 x86_alu_mem_imm (code, X86_CMP, ins->inst_offset, ins->inst_imm);
1863 case OP_X86_TEST_NULL:
1864 x86_test_reg_reg (code, ins->sreg1, ins->sreg1);
1866 case OP_X86_ADD_MEMBASE_IMM:
1867 x86_alu_membase_imm (code, X86_ADD, ins->inst_basereg, ins->inst_offset, ins->inst_imm);
1869 case OP_X86_ADD_MEMBASE:
1870 x86_alu_reg_membase (code, X86_ADD, ins->sreg1, ins->sreg2, ins->inst_offset);
1872 case OP_X86_SUB_MEMBASE_IMM:
1873 x86_alu_membase_imm (code, X86_SUB, ins->inst_basereg, ins->inst_offset, ins->inst_imm);
1875 case OP_X86_SUB_MEMBASE:
1876 x86_alu_reg_membase (code, X86_SUB, ins->sreg1, ins->sreg2, ins->inst_offset);
1878 case OP_X86_AND_MEMBASE_IMM:
1879 x86_alu_membase_imm (code, X86_AND, ins->inst_basereg, ins->inst_offset, ins->inst_imm);
1881 case OP_X86_OR_MEMBASE_IMM:
1882 x86_alu_membase_imm (code, X86_OR, ins->inst_basereg, ins->inst_offset, ins->inst_imm);
1884 case OP_X86_XOR_MEMBASE_IMM:
1885 x86_alu_membase_imm (code, X86_XOR, ins->inst_basereg, ins->inst_offset, ins->inst_imm);
1887 case OP_X86_INC_MEMBASE:
1888 x86_inc_membase (code, ins->inst_basereg, ins->inst_offset);
1890 case OP_X86_INC_REG:
1891 x86_inc_reg (code, ins->dreg);
1893 case OP_X86_DEC_MEMBASE:
1894 x86_dec_membase (code, ins->inst_basereg, ins->inst_offset);
1896 case OP_X86_DEC_REG:
1897 x86_dec_reg (code, ins->dreg);
1899 case OP_X86_MUL_MEMBASE:
1900 x86_imul_reg_membase (code, ins->sreg1, ins->sreg2, ins->inst_offset);
1903 x86_breakpoint (code);
1907 x86_alu_reg_reg (code, X86_ADD, ins->sreg1, ins->sreg2);
1910 x86_alu_reg_reg (code, X86_ADC, ins->sreg1, ins->sreg2);
1914 x86_alu_reg_imm (code, X86_ADD, ins->dreg, ins->inst_imm);
1917 x86_alu_reg_imm (code, X86_ADC, ins->dreg, ins->inst_imm);
1921 x86_alu_reg_reg (code, X86_SUB, ins->sreg1, ins->sreg2);
1924 x86_alu_reg_reg (code, X86_SBB, ins->sreg1, ins->sreg2);
1928 x86_alu_reg_imm (code, X86_SUB, ins->dreg, ins->inst_imm);
1931 x86_alu_reg_imm (code, X86_SBB, ins->dreg, ins->inst_imm);
1934 x86_alu_reg_reg (code, X86_AND, ins->sreg1, ins->sreg2);
1937 x86_alu_reg_imm (code, X86_AND, ins->sreg1, ins->inst_imm);
1941 x86_div_reg (code, ins->sreg2, TRUE);
1944 x86_alu_reg_reg (code, X86_XOR, X86_EDX, X86_EDX);
1945 x86_div_reg (code, ins->sreg2, FALSE);
1948 x86_mov_reg_imm (code, ins->sreg2, ins->inst_imm);
1950 x86_div_reg (code, ins->sreg2, TRUE);
1954 x86_div_reg (code, ins->sreg2, TRUE);
1957 x86_alu_reg_reg (code, X86_XOR, X86_EDX, X86_EDX);
1958 x86_div_reg (code, ins->sreg2, FALSE);
1961 x86_mov_reg_imm (code, ins->sreg2, ins->inst_imm);
1963 x86_div_reg (code, ins->sreg2, TRUE);
1966 x86_alu_reg_reg (code, X86_OR, ins->sreg1, ins->sreg2);
1969 x86_alu_reg_imm (code, X86_OR, ins->sreg1, ins->inst_imm);
1972 x86_alu_reg_reg (code, X86_XOR, ins->sreg1, ins->sreg2);
1975 x86_alu_reg_imm (code, X86_XOR, ins->sreg1, ins->inst_imm);
1978 g_assert (ins->sreg2 == X86_ECX);
1979 x86_shift_reg (code, X86_SHL, ins->dreg);
1982 g_assert (ins->sreg2 == X86_ECX);
1983 x86_shift_reg (code, X86_SAR, ins->dreg);
1986 x86_shift_reg_imm (code, X86_SAR, ins->dreg, ins->inst_imm);
1989 x86_shift_reg_imm (code, X86_SHR, ins->dreg, ins->inst_imm);
1992 g_assert (ins->sreg2 == X86_ECX);
1993 x86_shift_reg (code, X86_SHR, ins->dreg);
1996 x86_shift_reg_imm (code, X86_SHL, ins->dreg, ins->inst_imm);
1999 guint8 *jump_to_end;
2001 /* handle shifts below 32 bits */
2002 x86_shld_reg (code, ins->unused, ins->sreg1);
2003 x86_shift_reg (code, X86_SHL, ins->sreg1);
2005 x86_test_reg_imm (code, X86_ECX, 32);
2006 jump_to_end = code; x86_branch8 (code, X86_CC_EQ, 0, TRUE);
2008 /* handle shift over 32 bit */
2009 x86_mov_reg_reg (code, ins->unused, ins->sreg1, 4);
2010 x86_clear_reg (code, ins->sreg1);
2012 x86_patch (jump_to_end, code);
2016 guint8 *jump_to_end;
2018 /* handle shifts below 32 bits */
2019 x86_shrd_reg (code, ins->sreg1, ins->unused);
2020 x86_shift_reg (code, X86_SAR, ins->unused);
2022 x86_test_reg_imm (code, X86_ECX, 32);
2023 jump_to_end = code; x86_branch8 (code, X86_CC_EQ, 0, FALSE);
2025 /* handle shifts over 31 bits */
2026 x86_mov_reg_reg (code, ins->sreg1, ins->unused, 4);
2027 x86_shift_reg_imm (code, X86_SAR, ins->unused, 31);
2029 x86_patch (jump_to_end, code);
2033 guint8 *jump_to_end;
2035 /* handle shifts below 32 bits */
2036 x86_shrd_reg (code, ins->sreg1, ins->unused);
2037 x86_shift_reg (code, X86_SHR, ins->unused);
2039 x86_test_reg_imm (code, X86_ECX, 32);
2040 jump_to_end = code; x86_branch8 (code, X86_CC_EQ, 0, FALSE);
2042 /* handle shifts over 31 bits */
2043 x86_mov_reg_reg (code, ins->sreg1, ins->unused, 4);
2044 x86_shift_reg_imm (code, X86_SHR, ins->unused, 31);
2046 x86_patch (jump_to_end, code);
2050 if (ins->inst_imm >= 32) {
2051 x86_mov_reg_reg (code, ins->unused, ins->sreg1, 4);
2052 x86_clear_reg (code, ins->sreg1);
2053 x86_shift_reg_imm (code, X86_SHL, ins->unused, ins->inst_imm - 32);
2055 x86_shld_reg_imm (code, ins->unused, ins->sreg1, ins->inst_imm);
2056 x86_shift_reg_imm (code, X86_SHL, ins->sreg1, ins->inst_imm);
2060 if (ins->inst_imm >= 32) {
2061 x86_mov_reg_reg (code, ins->sreg1, ins->unused, 4);
2062 x86_shift_reg_imm (code, X86_SAR, ins->unused, 0x1f);
2063 x86_shift_reg_imm (code, X86_SAR, ins->sreg1, ins->inst_imm - 32);
2065 x86_shrd_reg_imm (code, ins->sreg1, ins->unused, ins->inst_imm);
2066 x86_shift_reg_imm (code, X86_SAR, ins->unused, ins->inst_imm);
2069 case OP_LSHR_UN_IMM:
2070 if (ins->inst_imm >= 32) {
2071 x86_mov_reg_reg (code, ins->sreg1, ins->unused, 4);
2072 x86_clear_reg (code, ins->unused);
2073 x86_shift_reg_imm (code, X86_SHR, ins->sreg1, ins->inst_imm - 32);
2075 x86_shrd_reg_imm (code, ins->sreg1, ins->unused, ins->inst_imm);
2076 x86_shift_reg_imm (code, X86_SHR, ins->unused, ins->inst_imm);
2080 x86_not_reg (code, ins->sreg1);
2083 x86_neg_reg (code, ins->sreg1);
2086 x86_widen_reg (code, ins->dreg, ins->sreg1, TRUE, FALSE);
2089 x86_widen_reg (code, ins->dreg, ins->sreg1, TRUE, TRUE);
2092 x86_imul_reg_reg (code, ins->sreg1, ins->sreg2);
2095 switch (ins->inst_imm) {
2099 if (ins->dreg != ins->sreg1)
2100 x86_mov_reg_reg (code, ins->dreg, ins->sreg1, 4);
2101 x86_alu_reg_reg (code, X86_ADD, ins->dreg, ins->dreg);
2104 /* LEA r1, [r2 + r2*2] */
2105 x86_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 1);
2108 /* LEA r1, [r2 + r2*4] */
2109 x86_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 2);
2112 /* LEA r1, [r2 + r2*2] */
2114 x86_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 1);
2115 x86_alu_reg_reg (code, X86_ADD, ins->dreg, ins->dreg);
2118 /* LEA r1, [r2 + r2*8] */
2119 x86_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 3);
2122 /* LEA r1, [r2 + r2*4] */
2124 x86_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 2);
2125 x86_alu_reg_reg (code, X86_ADD, ins->dreg, ins->dreg);
2128 /* LEA r1, [r2 + r2*2] */
2130 x86_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 1);
2131 x86_shift_reg_imm (code, X86_SHL, ins->dreg, 2);
2134 /* LEA r1, [r2 + r2*4] */
2135 /* LEA r1, [r1 + r1*4] */
2136 x86_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 2);
2137 x86_lea_memindex (code, ins->dreg, ins->dreg, 0, ins->dreg, 2);
2140 /* LEA r1, [r2 + r2*4] */
2142 /* LEA r1, [r1 + r1*4] */
2143 x86_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 2);
2144 x86_shift_reg_imm (code, X86_SHL, ins->dreg, 2);
2145 x86_lea_memindex (code, ins->dreg, ins->dreg, 0, ins->dreg, 2);
2148 x86_imul_reg_reg_imm (code, ins->dreg, ins->sreg1, ins->inst_imm);
2153 x86_imul_reg_reg (code, ins->sreg1, ins->sreg2);
2154 EMIT_COND_SYSTEM_EXCEPTION (X86_CC_O, FALSE, "OverflowException");
2156 case CEE_MUL_OVF_UN: {
2157 /* the mul operation and the exception check should most likely be split */
2158 int non_eax_reg, saved_eax = FALSE, saved_edx = FALSE;
2159 /*g_assert (ins->sreg2 == X86_EAX);
2160 g_assert (ins->dreg == X86_EAX);*/
2161 if (ins->sreg2 == X86_EAX) {
2162 non_eax_reg = ins->sreg1;
2163 } else if (ins->sreg1 == X86_EAX) {
2164 non_eax_reg = ins->sreg2;
2166 /* no need to save since we're going to store to it anyway */
2167 if (ins->dreg != X86_EAX) {
2169 x86_push_reg (code, X86_EAX);
2171 x86_mov_reg_reg (code, X86_EAX, ins->sreg1, 4);
2172 non_eax_reg = ins->sreg2;
2174 if (ins->dreg == X86_EDX) {
2177 x86_push_reg (code, X86_EAX);
2179 } else if (ins->dreg != X86_EAX) {
2181 x86_push_reg (code, X86_EDX);
2183 x86_mul_reg (code, non_eax_reg, FALSE);
2184 /* save before the check since pop and mov don't change the flags */
2185 if (ins->dreg != X86_EAX)
2186 x86_mov_reg_reg (code, ins->dreg, X86_EAX, 4);
2188 x86_pop_reg (code, X86_EDX);
2190 x86_pop_reg (code, X86_EAX);
2191 EMIT_COND_SYSTEM_EXCEPTION (X86_CC_O, FALSE, "OverflowException");
2195 x86_mov_reg_imm (code, ins->dreg, ins->inst_c0);
2198 g_assert_not_reached ();
2199 mono_add_patch_info (cfg, offset, (MonoJumpInfoType)ins->inst_i1, ins->inst_p0);
2200 x86_mov_reg_imm (code, ins->dreg, 0);
2202 case OP_LOAD_GOTADDR:
2203 x86_call_imm (code, 0);
2205 * The patch needs to point to the pop, since the GOT offset needs
2206 * to be added to that address.
2208 mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_GOT_OFFSET, NULL);
2209 x86_pop_reg (code, ins->dreg);
2210 x86_alu_reg_imm (code, X86_ADD, ins->dreg, 0xf0f0f0f0);
2213 mono_add_patch_info (cfg, offset, (MonoJumpInfoType)ins->inst_right->inst_i1, ins->inst_right->inst_p0);
2214 x86_mov_reg_membase (code, ins->dreg, ins->inst_basereg, 0xf0f0f0f0, 4);
2216 case OP_X86_PUSH_GOT_ENTRY:
2217 mono_add_patch_info (cfg, offset, (MonoJumpInfoType)ins->inst_right->inst_i1, ins->inst_right->inst_p0);
2218 x86_push_membase (code, ins->inst_basereg, 0xf0f0f0f0);
2222 x86_mov_reg_reg (code, ins->dreg, ins->sreg1, 4);
2225 g_assert_not_reached ();
2228 * Note: this 'frame destruction' logic is useful for tail calls, too.
2229 * Keep in sync with the code in emit_epilog.
2233 /* FIXME: no tracing support... */
2234 if (cfg->prof_options & MONO_PROFILE_ENTER_LEAVE)
2235 code = mono_arch_instrument_epilog (cfg, mono_profiler_method_leave, code, FALSE);
2236 /* reset offset to make max_len work */
2237 offset = code - cfg->native_code;
2239 g_assert (!cfg->method->save_lmf);
2241 if (cfg->used_int_regs & (1 << X86_EBX))
2243 if (cfg->used_int_regs & (1 << X86_EDI))
2245 if (cfg->used_int_regs & (1 << X86_ESI))
2248 x86_lea_membase (code, X86_ESP, X86_EBP, pos);
2250 if (cfg->used_int_regs & (1 << X86_ESI))
2251 x86_pop_reg (code, X86_ESI);
2252 if (cfg->used_int_regs & (1 << X86_EDI))
2253 x86_pop_reg (code, X86_EDI);
2254 if (cfg->used_int_regs & (1 << X86_EBX))
2255 x86_pop_reg (code, X86_EBX);
2257 /* restore ESP/EBP */
2259 offset = code - cfg->native_code;
2260 mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_METHOD_JUMP, ins->inst_p0);
2261 x86_jump32 (code, 0);
2265 /* ensure ins->sreg1 is not NULL
2266 * note that cmp DWORD PTR [eax], eax is one byte shorter than
2267 * cmp DWORD PTR [eax], 0
2269 x86_alu_membase_reg (code, X86_CMP, ins->sreg1, 0, ins->sreg1);
2272 int hreg = ins->sreg1 == X86_EAX? X86_ECX: X86_EAX;
2273 x86_push_reg (code, hreg);
2274 x86_lea_membase (code, hreg, X86_EBP, cfg->sig_cookie);
2275 x86_mov_membase_reg (code, ins->sreg1, 0, hreg, 4);
2276 x86_pop_reg (code, hreg);
2284 call = (MonoCallInst*)ins;
2285 if (ins->flags & MONO_INST_HAS_METHOD)
2286 code = emit_call (cfg, code, MONO_PATCH_INFO_METHOD, call->method);
2288 code = emit_call (cfg, code, MONO_PATCH_INFO_ABS, call->fptr);
2289 if (call->stack_usage && !CALLCONV_IS_STDCALL (call->signature)) {
2290 /* a pop is one byte, while an add reg, imm is 3. So if there are 4 or 8
2291 * bytes to pop, we want to use pops. GCC does this (note it won't happen
2292 * for P4 or i686 because gcc will avoid using pop push at all. But we aren't
2293 * smart enough to do that optimization yet
2295 * It turns out that on my P4, doing two pops for 8 bytes on the stack makes
2296 * mcs botstrap slow down. However, doing 1 pop for 4 bytes creates a small,
2297 * (most likely from locality benefits). People with other processors should
2298 * check on theirs to see what happens.
2300 if (call->stack_usage == 4) {
2301 /* we want to use registers that won't get used soon, so use
2302 * ecx, as eax will get allocated first. edx is used by long calls,
2303 * so we can't use that.
2306 x86_pop_reg (code, X86_ECX);
2308 x86_alu_reg_imm (code, X86_ADD, X86_ESP, call->stack_usage);
2311 code = emit_move_return_value (cfg, ins, code);
2316 case OP_VOIDCALL_REG:
2318 call = (MonoCallInst*)ins;
2319 x86_call_reg (code, ins->sreg1);
2320 if (call->stack_usage && !CALLCONV_IS_STDCALL (call->signature)) {
2321 if (call->stack_usage == 4)
2322 x86_pop_reg (code, X86_ECX);
2324 x86_alu_reg_imm (code, X86_ADD, X86_ESP, call->stack_usage);
2326 code = emit_move_return_value (cfg, ins, code);
2328 case OP_FCALL_MEMBASE:
2329 case OP_LCALL_MEMBASE:
2330 case OP_VCALL_MEMBASE:
2331 case OP_VOIDCALL_MEMBASE:
2332 case OP_CALL_MEMBASE:
2333 call = (MonoCallInst*)ins;
2334 x86_call_membase (code, ins->sreg1, ins->inst_offset);
2335 if (call->stack_usage && !CALLCONV_IS_STDCALL (call->signature)) {
2336 if (call->stack_usage == 4)
2337 x86_pop_reg (code, X86_ECX);
2339 x86_alu_reg_imm (code, X86_ADD, X86_ESP, call->stack_usage);
2341 code = emit_move_return_value (cfg, ins, code);
2345 x86_push_reg (code, ins->sreg1);
2347 case OP_X86_PUSH_IMM:
2348 x86_push_imm (code, ins->inst_imm);
2350 case OP_X86_PUSH_MEMBASE:
2351 x86_push_membase (code, ins->inst_basereg, ins->inst_offset);
2353 case OP_X86_PUSH_OBJ:
2354 x86_alu_reg_imm (code, X86_SUB, X86_ESP, ins->inst_imm);
2355 x86_push_reg (code, X86_EDI);
2356 x86_push_reg (code, X86_ESI);
2357 x86_push_reg (code, X86_ECX);
2358 if (ins->inst_offset)
2359 x86_lea_membase (code, X86_ESI, ins->inst_basereg, ins->inst_offset);
2361 x86_mov_reg_reg (code, X86_ESI, ins->inst_basereg, 4);
2362 x86_lea_membase (code, X86_EDI, X86_ESP, 12);
2363 x86_mov_reg_imm (code, X86_ECX, (ins->inst_imm >> 2));
2365 x86_prefix (code, X86_REP_PREFIX);
2367 x86_pop_reg (code, X86_ECX);
2368 x86_pop_reg (code, X86_ESI);
2369 x86_pop_reg (code, X86_EDI);
2372 x86_lea_memindex (code, ins->dreg, ins->sreg1, ins->inst_imm, ins->sreg2, ins->unused);
2374 case OP_X86_LEA_MEMBASE:
2375 x86_lea_membase (code, ins->dreg, ins->sreg1, ins->inst_imm);
2378 x86_xchg_reg_reg (code, ins->sreg1, ins->sreg2, 4);
2381 /* keep alignment */
2382 x86_alu_reg_imm (code, X86_ADD, ins->sreg1, MONO_ARCH_FRAME_ALIGNMENT - 1);
2383 x86_alu_reg_imm (code, X86_AND, ins->sreg1, ~(MONO_ARCH_FRAME_ALIGNMENT - 1));
2384 code = mono_emit_stack_alloc (code, ins);
2385 x86_mov_reg_reg (code, ins->dreg, X86_ESP, 4);
2391 x86_push_reg (code, ins->sreg1);
2392 code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD,
2393 (gpointer)"mono_arch_throw_exception");
2397 x86_push_reg (code, ins->sreg1);
2398 code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD,
2399 (gpointer)"mono_arch_rethrow_exception");
2402 case OP_CALL_HANDLER:
2405 x86_alu_reg_imm (code, X86_SUB, X86_ESP, 12);
2407 mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_BB, ins->inst_target_bb);
2408 x86_call_imm (code, 0);
2410 x86_alu_reg_imm (code, X86_ADD, X86_ESP, 12);
2414 ins->inst_c0 = code - cfg->native_code;
2417 //g_print ("target: %p, next: %p, curr: %p, last: %p\n", ins->inst_target_bb, bb->next_bb, ins, bb->last_ins);
2418 //if ((ins->inst_target_bb == bb->next_bb) && ins == bb->last_ins)
2420 if (ins->flags & MONO_INST_BRLABEL) {
2421 if (ins->inst_i0->inst_c0) {
2422 x86_jump_code (code, cfg->native_code + ins->inst_i0->inst_c0);
2424 mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_LABEL, ins->inst_i0);
2425 if ((cfg->opt & MONO_OPT_BRANCH) &&
2426 x86_is_imm8 (ins->inst_i0->inst_c1 - cpos))
2427 x86_jump8 (code, 0);
2429 x86_jump32 (code, 0);
2432 if (ins->inst_target_bb->native_offset) {
2433 x86_jump_code (code, cfg->native_code + ins->inst_target_bb->native_offset);
2435 mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_BB, ins->inst_target_bb);
2436 if ((cfg->opt & MONO_OPT_BRANCH) &&
2437 x86_is_imm8 (ins->inst_target_bb->max_offset - cpos))
2438 x86_jump8 (code, 0);
2440 x86_jump32 (code, 0);
2445 x86_jump_reg (code, ins->sreg1);
2448 x86_set_reg (code, X86_CC_EQ, ins->dreg, TRUE);
2449 x86_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
2452 x86_set_reg (code, X86_CC_LT, ins->dreg, TRUE);
2453 x86_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
2456 x86_set_reg (code, X86_CC_LT, ins->dreg, FALSE);
2457 x86_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
2460 x86_set_reg (code, X86_CC_GT, ins->dreg, TRUE);
2461 x86_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
2464 x86_set_reg (code, X86_CC_GT, ins->dreg, FALSE);
2465 x86_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
2468 x86_set_reg (code, X86_CC_NE, ins->dreg, TRUE);
2469 x86_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
2471 case OP_COND_EXC_EQ:
2472 case OP_COND_EXC_NE_UN:
2473 case OP_COND_EXC_LT:
2474 case OP_COND_EXC_LT_UN:
2475 case OP_COND_EXC_GT:
2476 case OP_COND_EXC_GT_UN:
2477 case OP_COND_EXC_GE:
2478 case OP_COND_EXC_GE_UN:
2479 case OP_COND_EXC_LE:
2480 case OP_COND_EXC_LE_UN:
2481 case OP_COND_EXC_OV:
2482 case OP_COND_EXC_NO:
2484 case OP_COND_EXC_NC:
2485 EMIT_COND_SYSTEM_EXCEPTION (branch_cc_table [ins->opcode - OP_COND_EXC_EQ], (ins->opcode < OP_COND_EXC_NE_UN), ins->inst_p1);
2497 EMIT_COND_BRANCH (ins, branch_cc_table [ins->opcode - CEE_BEQ], (ins->opcode < CEE_BNE_UN));
2500 /* floating point opcodes */
2502 double d = *(double *)ins->inst_p0;
2504 if ((d == 0.0) && (mono_signbit (d) == 0)) {
2506 } else if (d == 1.0) {
2509 if (cfg->compile_aot) {
2510 guint32 *val = (guint32*)&d;
2511 x86_push_imm (code, val [1]);
2512 x86_push_imm (code, val [0]);
2513 x86_fld_membase (code, X86_ESP, 0, TRUE);
2514 x86_alu_reg_imm (code, X86_ADD, X86_ESP, 8);
2517 mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_R8, ins->inst_p0);
2518 x86_fld (code, NULL, TRUE);
2524 float f = *(float *)ins->inst_p0;
2526 if ((f == 0.0) && (mono_signbit (f) == 0)) {
2528 } else if (f == 1.0) {
2531 if (cfg->compile_aot) {
2532 guint32 val = *(guint32*)&f;
2533 x86_push_imm (code, val);
2534 x86_fld_membase (code, X86_ESP, 0, FALSE);
2535 x86_alu_reg_imm (code, X86_ADD, X86_ESP, 4);
2538 mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_R4, ins->inst_p0);
2539 x86_fld (code, NULL, FALSE);
2544 case OP_STORER8_MEMBASE_REG:
2545 x86_fst_membase (code, ins->inst_destbasereg, ins->inst_offset, TRUE, TRUE);
2547 case OP_LOADR8_SPILL_MEMBASE:
2548 x86_fld_membase (code, ins->inst_basereg, ins->inst_offset, TRUE);
2551 case OP_LOADR8_MEMBASE:
2552 x86_fld_membase (code, ins->inst_basereg, ins->inst_offset, TRUE);
2554 case OP_STORER4_MEMBASE_REG:
2555 x86_fst_membase (code, ins->inst_destbasereg, ins->inst_offset, FALSE, TRUE);
2557 case OP_LOADR4_MEMBASE:
2558 x86_fld_membase (code, ins->inst_basereg, ins->inst_offset, FALSE);
2560 case CEE_CONV_R4: /* FIXME: change precision */
2562 x86_push_reg (code, ins->sreg1);
2563 x86_fild_membase (code, X86_ESP, 0, FALSE);
2564 x86_alu_reg_imm (code, X86_ADD, X86_ESP, 4);
2566 case OP_X86_FP_LOAD_I8:
2567 x86_fild_membase (code, ins->inst_basereg, ins->inst_offset, TRUE);
2569 case OP_X86_FP_LOAD_I4:
2570 x86_fild_membase (code, ins->inst_basereg, ins->inst_offset, FALSE);
2572 case OP_FCONV_TO_I1:
2573 code = emit_float_to_int (cfg, code, ins->dreg, 1, TRUE);
2575 case OP_FCONV_TO_U1:
2576 code = emit_float_to_int (cfg, code, ins->dreg, 1, FALSE);
2578 case OP_FCONV_TO_I2:
2579 code = emit_float_to_int (cfg, code, ins->dreg, 2, TRUE);
2581 case OP_FCONV_TO_U2:
2582 code = emit_float_to_int (cfg, code, ins->dreg, 2, FALSE);
2584 case OP_FCONV_TO_I4:
2586 code = emit_float_to_int (cfg, code, ins->dreg, 4, TRUE);
2588 case OP_FCONV_TO_I8:
2589 x86_alu_reg_imm (code, X86_SUB, X86_ESP, 4);
2590 x86_fnstcw_membase(code, X86_ESP, 0);
2591 x86_mov_reg_membase (code, ins->dreg, X86_ESP, 0, 2);
2592 x86_alu_reg_imm (code, X86_OR, ins->dreg, 0xc00);
2593 x86_mov_membase_reg (code, X86_ESP, 2, ins->dreg, 2);
2594 x86_fldcw_membase (code, X86_ESP, 2);
2595 x86_alu_reg_imm (code, X86_SUB, X86_ESP, 8);
2596 x86_fist_pop_membase (code, X86_ESP, 0, TRUE);
2597 x86_pop_reg (code, ins->dreg);
2598 x86_pop_reg (code, ins->unused);
2599 x86_fldcw_membase (code, X86_ESP, 0);
2600 x86_alu_reg_imm (code, X86_ADD, X86_ESP, 4);
2602 case OP_LCONV_TO_R_UN: {
2603 static guint8 mn[] = { 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x80, 0x3f, 0x40 };
2606 /* load 64bit integer to FP stack */
2607 x86_push_imm (code, 0);
2608 x86_push_reg (code, ins->sreg2);
2609 x86_push_reg (code, ins->sreg1);
2610 x86_fild_membase (code, X86_ESP, 0, TRUE);
2611 /* store as 80bit FP value */
2612 x86_fst80_membase (code, X86_ESP, 0);
2614 /* test if lreg is negative */
2615 x86_test_reg_reg (code, ins->sreg2, ins->sreg2);
2616 br = code; x86_branch8 (code, X86_CC_GEZ, 0, TRUE);
2618 /* add correction constant mn */
2619 x86_fld80_mem (code, mn);
2620 x86_fld80_membase (code, X86_ESP, 0);
2621 x86_fp_op_reg (code, X86_FADD, 1, TRUE);
2622 x86_fst80_membase (code, X86_ESP, 0);
2624 x86_patch (br, code);
2626 x86_fld80_membase (code, X86_ESP, 0);
2627 x86_alu_reg_imm (code, X86_ADD, X86_ESP, 12);
2631 case OP_LCONV_TO_OVF_I: {
2632 guint8 *br [3], *label [1];
2636 * Valid ints: 0xffffffff:8000000 to 00000000:0x7f000000
2638 x86_test_reg_reg (code, ins->sreg1, ins->sreg1);
2640 /* If the low word top bit is set, see if we are negative */
2641 br [0] = code; x86_branch8 (code, X86_CC_LT, 0, TRUE);
2642 /* We are not negative (no top bit set, check for our top word to be zero */
2643 x86_test_reg_reg (code, ins->sreg2, ins->sreg2);
2644 br [1] = code; x86_branch8 (code, X86_CC_EQ, 0, TRUE);
2647 /* throw exception */
2648 tins = mono_branch_optimize_exception_target (cfg, bb, "OverflowException");
2650 mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_BB, tins->inst_true_bb);
2651 if ((cfg->opt & MONO_OPT_BRANCH) && x86_is_imm8 (tins->inst_true_bb->max_offset - cpos))
2652 x86_jump8 (code, 0);
2654 x86_jump32 (code, 0);
2656 mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_EXC, "OverflowException");
2657 x86_jump32 (code, 0);
2661 x86_patch (br [0], code);
2662 /* our top bit is set, check that top word is 0xfffffff */
2663 x86_alu_reg_imm (code, X86_CMP, ins->sreg2, 0xffffffff);
2665 x86_patch (br [1], code);
2666 /* nope, emit exception */
2667 br [2] = code; x86_branch8 (code, X86_CC_NE, 0, TRUE);
2668 x86_patch (br [2], label [0]);
2670 if (ins->dreg != ins->sreg1)
2671 x86_mov_reg_reg (code, ins->dreg, ins->sreg1, 4);
2675 x86_fp_op_reg (code, X86_FADD, 1, TRUE);
2678 x86_fp_op_reg (code, X86_FSUB, 1, TRUE);
2681 x86_fp_op_reg (code, X86_FMUL, 1, TRUE);
2684 x86_fp_op_reg (code, X86_FDIV, 1, TRUE);
2692 x86_fp_op_reg (code, X86_FADD, 1, TRUE);
2697 x86_fp_op_reg (code, X86_FADD, 1, TRUE);
2704 * it really doesn't make sense to inline all this code,
2705 * it's here just to show that things may not be as simple
2708 guchar *check_pos, *end_tan, *pop_jump;
2709 x86_push_reg (code, X86_EAX);
2712 x86_test_reg_imm (code, X86_EAX, X86_FP_C2);
2714 x86_branch8 (code, X86_CC_NE, 0, FALSE);
2715 x86_fstp (code, 0); /* pop the 1.0 */
2717 x86_jump8 (code, 0);
2719 x86_fp_op (code, X86_FADD, 0);
2723 x86_test_reg_imm (code, X86_EAX, X86_FP_C2);
2725 x86_branch8 (code, X86_CC_NE, 0, FALSE);
2728 x86_patch (pop_jump, code);
2729 x86_fstp (code, 0); /* pop the 1.0 */
2730 x86_patch (check_pos, code);
2731 x86_patch (end_tan, code);
2733 x86_fp_op_reg (code, X86_FADD, 1, TRUE);
2734 x86_pop_reg (code, X86_EAX);
2741 x86_fp_op_reg (code, X86_FADD, 1, TRUE);
2752 x86_push_reg (code, X86_EAX);
2753 /* we need to exchange ST(0) with ST(1) */
2756 /* this requires a loop, because fprem somtimes
2757 * returns a partial remainder */
2759 /* looks like MS is using fprem instead of the IEEE compatible fprem1 */
2760 /* x86_fprem1 (code); */
2763 x86_alu_reg_imm (code, X86_AND, X86_EAX, X86_FP_C2);
2765 x86_branch8 (code, X86_CC_NE, l1 - l2, FALSE);
2770 x86_pop_reg (code, X86_EAX);
2774 if (cfg->opt & MONO_OPT_FCMOV) {
2775 x86_fcomip (code, 1);
2779 /* this overwrites EAX */
2780 EMIT_FPCOMPARE(code);
2781 x86_alu_reg_imm (code, X86_AND, X86_EAX, X86_FP_CC_MASK);
2784 if (cfg->opt & MONO_OPT_FCMOV) {
2785 /* zeroing the register at the start results in
2786 * shorter and faster code (we can also remove the widening op)
2788 guchar *unordered_check;
2789 x86_alu_reg_reg (code, X86_XOR, ins->dreg, ins->dreg);
2790 x86_fcomip (code, 1);
2792 unordered_check = code;
2793 x86_branch8 (code, X86_CC_P, 0, FALSE);
2794 x86_set_reg (code, X86_CC_EQ, ins->dreg, FALSE);
2795 x86_patch (unordered_check, code);
2798 if (ins->dreg != X86_EAX)
2799 x86_push_reg (code, X86_EAX);
2801 EMIT_FPCOMPARE(code);
2802 x86_alu_reg_imm (code, X86_AND, X86_EAX, X86_FP_CC_MASK);
2803 x86_alu_reg_imm (code, X86_CMP, X86_EAX, 0x4000);
2804 x86_set_reg (code, X86_CC_EQ, ins->dreg, TRUE);
2805 x86_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
2807 if (ins->dreg != X86_EAX)
2808 x86_pop_reg (code, X86_EAX);
2812 if (cfg->opt & MONO_OPT_FCMOV) {
2813 /* zeroing the register at the start results in
2814 * shorter and faster code (we can also remove the widening op)
2816 x86_alu_reg_reg (code, X86_XOR, ins->dreg, ins->dreg);
2817 x86_fcomip (code, 1);
2819 if (ins->opcode == OP_FCLT_UN) {
2820 guchar *unordered_check = code;
2821 guchar *jump_to_end;
2822 x86_branch8 (code, X86_CC_P, 0, FALSE);
2823 x86_set_reg (code, X86_CC_GT, ins->dreg, FALSE);
2825 x86_jump8 (code, 0);
2826 x86_patch (unordered_check, code);
2827 x86_inc_reg (code, ins->dreg);
2828 x86_patch (jump_to_end, code);
2830 x86_set_reg (code, X86_CC_GT, ins->dreg, FALSE);
2834 if (ins->dreg != X86_EAX)
2835 x86_push_reg (code, X86_EAX);
2837 EMIT_FPCOMPARE(code);
2838 x86_alu_reg_imm (code, X86_AND, X86_EAX, X86_FP_CC_MASK);
2839 if (ins->opcode == OP_FCLT_UN) {
2840 guchar *is_not_zero_check, *end_jump;
2841 is_not_zero_check = code;
2842 x86_branch8 (code, X86_CC_NZ, 0, TRUE);
2844 x86_jump8 (code, 0);
2845 x86_patch (is_not_zero_check, code);
2846 x86_alu_reg_imm (code, X86_CMP, X86_EAX, X86_FP_CC_MASK);
2848 x86_patch (end_jump, code);
2850 x86_set_reg (code, X86_CC_EQ, ins->dreg, TRUE);
2851 x86_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
2853 if (ins->dreg != X86_EAX)
2854 x86_pop_reg (code, X86_EAX);
2858 if (cfg->opt & MONO_OPT_FCMOV) {
2859 /* zeroing the register at the start results in
2860 * shorter and faster code (we can also remove the widening op)
2862 guchar *unordered_check;
2863 x86_alu_reg_reg (code, X86_XOR, ins->dreg, ins->dreg);
2864 x86_fcomip (code, 1);
2866 if (ins->opcode == OP_FCGT) {
2867 unordered_check = code;
2868 x86_branch8 (code, X86_CC_P, 0, FALSE);
2869 x86_set_reg (code, X86_CC_LT, ins->dreg, FALSE);
2870 x86_patch (unordered_check, code);
2872 x86_set_reg (code, X86_CC_LT, ins->dreg, FALSE);
2876 if (ins->dreg != X86_EAX)
2877 x86_push_reg (code, X86_EAX);
2879 EMIT_FPCOMPARE(code);
2880 x86_alu_reg_imm (code, X86_AND, X86_EAX, X86_FP_CC_MASK);
2881 x86_alu_reg_imm (code, X86_CMP, X86_EAX, X86_FP_C0);
2882 if (ins->opcode == OP_FCGT_UN) {
2883 guchar *is_not_zero_check, *end_jump;
2884 is_not_zero_check = code;
2885 x86_branch8 (code, X86_CC_NZ, 0, TRUE);
2887 x86_jump8 (code, 0);
2888 x86_patch (is_not_zero_check, code);
2889 x86_alu_reg_imm (code, X86_CMP, X86_EAX, X86_FP_CC_MASK);
2891 x86_patch (end_jump, code);
2893 x86_set_reg (code, X86_CC_EQ, ins->dreg, TRUE);
2894 x86_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
2896 if (ins->dreg != X86_EAX)
2897 x86_pop_reg (code, X86_EAX);
2900 if (cfg->opt & MONO_OPT_FCMOV) {
2901 guchar *jump = code;
2902 x86_branch8 (code, X86_CC_P, 0, TRUE);
2903 EMIT_COND_BRANCH (ins, X86_CC_EQ, FALSE);
2904 x86_patch (jump, code);
2907 x86_alu_reg_imm (code, X86_CMP, X86_EAX, 0x4000);
2908 EMIT_COND_BRANCH (ins, X86_CC_EQ, TRUE);
2911 /* Branch if C013 != 100 */
2912 if (cfg->opt & MONO_OPT_FCMOV) {
2913 /* branch if !ZF or (PF|CF) */
2914 EMIT_COND_BRANCH (ins, X86_CC_NE, FALSE);
2915 EMIT_COND_BRANCH (ins, X86_CC_P, FALSE);
2916 EMIT_COND_BRANCH (ins, X86_CC_B, FALSE);
2919 x86_alu_reg_imm (code, X86_CMP, X86_EAX, X86_FP_C3);
2920 EMIT_COND_BRANCH (ins, X86_CC_NE, FALSE);
2923 if (cfg->opt & MONO_OPT_FCMOV) {
2924 EMIT_COND_BRANCH (ins, X86_CC_GT, FALSE);
2927 EMIT_COND_BRANCH (ins, X86_CC_EQ, FALSE);
2930 if (cfg->opt & MONO_OPT_FCMOV) {
2931 EMIT_COND_BRANCH (ins, X86_CC_P, FALSE);
2932 EMIT_COND_BRANCH (ins, X86_CC_GT, FALSE);
2935 if (ins->opcode == OP_FBLT_UN) {
2936 guchar *is_not_zero_check, *end_jump;
2937 is_not_zero_check = code;
2938 x86_branch8 (code, X86_CC_NZ, 0, TRUE);
2940 x86_jump8 (code, 0);
2941 x86_patch (is_not_zero_check, code);
2942 x86_alu_reg_imm (code, X86_CMP, X86_EAX, X86_FP_CC_MASK);
2944 x86_patch (end_jump, code);
2946 EMIT_COND_BRANCH (ins, X86_CC_EQ, FALSE);
2950 if (cfg->opt & MONO_OPT_FCMOV) {
2951 EMIT_COND_BRANCH (ins, X86_CC_LT, FALSE);
2954 x86_alu_reg_imm (code, X86_CMP, X86_EAX, X86_FP_C0);
2955 if (ins->opcode == OP_FBGT_UN) {
2956 guchar *is_not_zero_check, *end_jump;
2957 is_not_zero_check = code;
2958 x86_branch8 (code, X86_CC_NZ, 0, TRUE);
2960 x86_jump8 (code, 0);
2961 x86_patch (is_not_zero_check, code);
2962 x86_alu_reg_imm (code, X86_CMP, X86_EAX, X86_FP_CC_MASK);
2964 x86_patch (end_jump, code);
2966 EMIT_COND_BRANCH (ins, X86_CC_EQ, FALSE);
2969 /* Branch if C013 == 100 or 001 */
2970 if (cfg->opt & MONO_OPT_FCMOV) {
2973 /* skip branch if C1=1 */
2975 x86_branch8 (code, X86_CC_P, 0, FALSE);
2976 /* branch if (C0 | C3) = 1 */
2977 EMIT_COND_BRANCH (ins, X86_CC_BE, FALSE);
2978 x86_patch (br1, code);
2981 x86_alu_reg_imm (code, X86_CMP, X86_EAX, X86_FP_C0);
2982 EMIT_COND_BRANCH (ins, X86_CC_EQ, FALSE);
2983 x86_alu_reg_imm (code, X86_CMP, X86_EAX, X86_FP_C3);
2984 EMIT_COND_BRANCH (ins, X86_CC_EQ, FALSE);
2987 /* Branch if C013 == 000 */
2988 if (cfg->opt & MONO_OPT_FCMOV) {
2989 EMIT_COND_BRANCH (ins, X86_CC_LE, FALSE);
2992 EMIT_COND_BRANCH (ins, X86_CC_NE, FALSE);
2995 /* Branch if C013=000 or 100 */
2996 if (cfg->opt & MONO_OPT_FCMOV) {
2999 /* skip branch if C1=1 */
3001 x86_branch8 (code, X86_CC_P, 0, FALSE);
3002 /* branch if C0=0 */
3003 EMIT_COND_BRANCH (ins, X86_CC_NB, FALSE);
3004 x86_patch (br1, code);
3007 x86_alu_reg_imm (code, X86_AND, X86_EAX, (X86_FP_C0|X86_FP_C1));
3008 x86_alu_reg_imm (code, X86_CMP, X86_EAX, 0);
3009 EMIT_COND_BRANCH (ins, X86_CC_EQ, FALSE);
3012 /* Branch if C013 != 001 */
3013 if (cfg->opt & MONO_OPT_FCMOV) {
3014 EMIT_COND_BRANCH (ins, X86_CC_P, FALSE);
3015 EMIT_COND_BRANCH (ins, X86_CC_GE, FALSE);
3018 x86_alu_reg_imm (code, X86_CMP, X86_EAX, X86_FP_C0);
3019 EMIT_COND_BRANCH (ins, X86_CC_NE, FALSE);
3021 case CEE_CKFINITE: {
3022 x86_push_reg (code, X86_EAX);
3025 x86_alu_reg_imm (code, X86_AND, X86_EAX, 0x4100);
3026 x86_alu_reg_imm (code, X86_CMP, X86_EAX, X86_FP_C0);
3027 x86_pop_reg (code, X86_EAX);
3028 EMIT_COND_SYSTEM_EXCEPTION (X86_CC_EQ, FALSE, "ArithmeticException");
3032 code = emit_tls_get (code, ins->dreg, ins->inst_offset);
3035 case OP_MEMORY_BARRIER: {
3036 /* Not needed on x86 */
3039 case OP_ATOMIC_ADD_I4: {
3040 int dreg = ins->dreg;
3042 if (dreg == ins->inst_basereg) {
3043 x86_push_reg (code, ins->sreg2);
3047 if (dreg != ins->sreg2)
3048 x86_mov_reg_reg (code, ins->dreg, ins->sreg2, 4);
3050 x86_prefix (code, X86_LOCK_PREFIX);
3051 x86_xadd_membase_reg (code, ins->inst_basereg, ins->inst_offset, dreg, 4);
3053 if (dreg != ins->dreg) {
3054 x86_mov_reg_reg (code, ins->dreg, dreg, 4);
3055 x86_pop_reg (code, dreg);
3060 case OP_ATOMIC_ADD_NEW_I4: {
3061 int dreg = ins->dreg;
3063 /* hack: limit in regalloc, dreg != sreg1 && dreg != sreg2 */
3064 if (ins->sreg2 == dreg) {
3065 if (dreg == X86_EBX) {
3067 if (ins->inst_basereg == X86_EDI)
3071 if (ins->inst_basereg == X86_EBX)
3074 } else if (ins->inst_basereg == dreg) {
3075 if (dreg == X86_EBX) {
3077 if (ins->sreg2 == X86_EDI)
3081 if (ins->sreg2 == X86_EBX)
3086 if (dreg != ins->dreg) {
3087 x86_push_reg (code, dreg);
3090 x86_mov_reg_reg (code, dreg, ins->sreg2, 4);
3091 x86_prefix (code, X86_LOCK_PREFIX);
3092 x86_xadd_membase_reg (code, ins->inst_basereg, ins->inst_offset, dreg, 4);
3093 /* dreg contains the old value, add with sreg2 value */
3094 x86_alu_reg_reg (code, X86_ADD, dreg, ins->sreg2);
3096 if (ins->dreg != dreg) {
3097 x86_mov_reg_reg (code, ins->dreg, dreg, 4);
3098 x86_pop_reg (code, dreg);
3103 case OP_ATOMIC_EXCHANGE_I4: {
3105 int sreg2 = ins->sreg2;
3106 int breg = ins->inst_basereg;
3108 /* cmpxchg uses eax as comperand, need to make sure we can use it
3109 * hack to overcome limits in x86 reg allocator
3110 * (req: dreg == eax and sreg2 != eax and breg != eax)
3112 if (ins->dreg != X86_EAX)
3113 x86_push_reg (code, X86_EAX);
3115 /* We need the EAX reg for the cmpxchg */
3116 if (ins->sreg2 == X86_EAX) {
3117 x86_push_reg (code, X86_EDX);
3118 x86_mov_reg_reg (code, X86_EDX, X86_EAX, 4);
3122 if (breg == X86_EAX) {
3123 x86_push_reg (code, X86_ESI);
3124 x86_mov_reg_reg (code, X86_ESI, X86_EAX, 4);
3128 x86_mov_reg_membase (code, X86_EAX, breg, ins->inst_offset, 4);
3130 br [0] = code; x86_prefix (code, X86_LOCK_PREFIX);
3131 x86_cmpxchg_membase_reg (code, breg, ins->inst_offset, sreg2);
3132 br [1] = code; x86_branch8 (code, X86_CC_NE, -1, FALSE);
3133 x86_patch (br [1], br [0]);
3135 if (breg != ins->inst_basereg)
3136 x86_pop_reg (code, X86_ESI);
3138 if (ins->dreg != X86_EAX) {
3139 x86_mov_reg_reg (code, ins->dreg, X86_EAX, 4);
3140 x86_pop_reg (code, X86_EAX);
3143 if (ins->sreg2 != sreg2)
3144 x86_pop_reg (code, X86_EDX);
3149 g_warning ("unknown opcode %s in %s()\n", mono_inst_name (ins->opcode), __FUNCTION__);
3150 g_assert_not_reached ();
3153 if ((code - cfg->native_code - offset) > max_len) {
3154 g_warning ("wrong maximal instruction length of instruction %s (expected %d, got %d)",
3155 mono_inst_name (ins->opcode), max_len, code - cfg->native_code - offset);
3156 g_assert_not_reached ();
3162 last_offset = offset;
3167 cfg->code_len = code - cfg->native_code;
3171 mono_arch_register_lowlevel_calls (void)
3176 mono_arch_patch_code (MonoMethod *method, MonoDomain *domain, guint8 *code, MonoJumpInfo *ji, gboolean run_cctors)
3178 MonoJumpInfo *patch_info;
3179 gboolean compile_aot = !run_cctors;
3181 for (patch_info = ji; patch_info; patch_info = patch_info->next) {
3182 unsigned char *ip = patch_info->ip.i + code;
3183 const unsigned char *target;
3185 target = mono_resolve_patch_target (method, domain, code, patch_info, run_cctors);
3188 switch (patch_info->type) {
3189 case MONO_PATCH_INFO_BB:
3190 case MONO_PATCH_INFO_LABEL:
3193 /* No need to patch these */
3198 switch (patch_info->type) {
3199 case MONO_PATCH_INFO_IP:
3200 *((gconstpointer *)(ip)) = target;
3202 case MONO_PATCH_INFO_CLASS_INIT: {
3204 /* Might already been changed to a nop */
3205 x86_call_code (code, 0);
3206 x86_patch (ip, target);
3209 case MONO_PATCH_INFO_ABS:
3210 case MONO_PATCH_INFO_METHOD:
3211 case MONO_PATCH_INFO_METHOD_JUMP:
3212 case MONO_PATCH_INFO_INTERNAL_METHOD:
3213 case MONO_PATCH_INFO_BB:
3214 case MONO_PATCH_INFO_LABEL:
3215 x86_patch (ip, target);
3217 case MONO_PATCH_INFO_NONE:
3220 guint32 offset = mono_arch_get_patch_offset (ip);
3221 *((gconstpointer *)(ip + offset)) = target;
3229 mono_arch_emit_prolog (MonoCompile *cfg)
3231 MonoMethod *method = cfg->method;
3233 MonoMethodSignature *sig;
3235 int alloc_size, pos, max_offset, i;
3238 cfg->code_size = MAX (mono_method_get_header (method)->code_size * 4, 256);
3239 code = cfg->native_code = g_malloc (cfg->code_size);
3241 x86_push_reg (code, X86_EBP);
3242 x86_mov_reg_reg (code, X86_EBP, X86_ESP, 4);
3244 alloc_size = cfg->stack_offset;
3247 if (method->wrapper_type == MONO_WRAPPER_NATIVE_TO_MANAGED) {
3248 /* Might need to attach the thread to the JIT */
3249 if (lmf_tls_offset != -1) {
3252 code = emit_tls_get ( code, X86_EAX, lmf_tls_offset);
3253 x86_test_reg_reg (code, X86_EAX, X86_EAX);
3255 x86_branch8 (code, X86_CC_NE, 0, 0);
3256 x86_push_imm (code, cfg->domain);
3257 code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD, (gpointer)"mono_jit_thread_attach");
3258 x86_alu_reg_imm (code, X86_ADD, X86_ESP, 4);
3259 x86_patch (buf, code);
3260 #ifdef PLATFORM_WIN32
3261 /* The TLS key actually contains a pointer to the MonoJitTlsData structure */
3262 /* FIXME: Add a separate key for LMF to avoid this */
3263 x86_alu_reg_imm (code, X86_ADD, X86_EAX, G_STRUCT_OFFSET (MonoJitTlsData, lmf));
3266 g_assert (!cfg->compile_aot);
3267 x86_push_imm (code, cfg->domain);
3268 code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD, (gpointer)"mono_jit_thread_attach");
3269 x86_alu_reg_imm (code, X86_ADD, X86_ESP, 4);
3273 if (method->save_lmf) {
3274 pos += sizeof (MonoLMF);
3276 /* save the current IP */
3277 mono_add_patch_info (cfg, code + 1 - cfg->native_code, MONO_PATCH_INFO_IP, NULL);
3278 x86_push_imm_template (code);
3280 /* save all caller saved regs */
3281 x86_push_reg (code, X86_EBP);
3282 x86_push_reg (code, X86_ESI);
3283 x86_push_reg (code, X86_EDI);
3284 x86_push_reg (code, X86_EBX);
3286 /* save method info */
3287 x86_push_imm (code, method);
3289 /* get the address of lmf for the current thread */
3291 * This is performance critical so we try to use some tricks to make
3294 if (lmf_tls_offset != -1) {
3295 /* Load lmf quicky using the GS register */
3296 code = emit_tls_get (code, X86_EAX, lmf_tls_offset);
3297 #ifdef PLATFORM_WIN32
3298 /* The TLS key actually contains a pointer to the MonoJitTlsData structure */
3299 /* FIXME: Add a separate key for LMF to avoid this */
3300 x86_alu_reg_imm (code, X86_ADD, X86_EAX, G_STRUCT_OFFSET (MonoJitTlsData, lmf));
3303 code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD, (gpointer)"mono_get_lmf_addr");
3307 x86_push_reg (code, X86_EAX);
3308 /* push *lfm (previous_lmf) */
3309 x86_push_membase (code, X86_EAX, 0);
3311 x86_mov_membase_reg (code, X86_EAX, 0, X86_ESP, 4);
3314 if (cfg->used_int_regs & (1 << X86_EBX)) {
3315 x86_push_reg (code, X86_EBX);
3319 if (cfg->used_int_regs & (1 << X86_EDI)) {
3320 x86_push_reg (code, X86_EDI);
3324 if (cfg->used_int_regs & (1 << X86_ESI)) {
3325 x86_push_reg (code, X86_ESI);
3333 /* the original alloc_size is already aligned: there is %ebp and retip pushed, so realign */
3335 int tot = alloc_size + pos + 4 + 4; /* ret ip + ebp */
3347 /* See mono_emit_stack_alloc */
3348 #if defined(PLATFORM_WIN32) || defined(MONO_ARCH_SIGSEGV_ON_ALTSTACK)
3349 guint32 remaining_size = alloc_size;
3350 while (remaining_size >= 0x1000) {
3351 x86_alu_reg_imm (code, X86_SUB, X86_ESP, 0x1000);
3352 x86_test_membase_reg (code, X86_ESP, 0, X86_ESP);
3353 remaining_size -= 0x1000;
3356 x86_alu_reg_imm (code, X86_SUB, X86_ESP, remaining_size);
3358 x86_alu_reg_imm (code, X86_SUB, X86_ESP, alloc_size);
3363 /* check the stack is aligned */
3364 x86_mov_reg_reg (code, X86_EDX, X86_ESP, 4);
3365 x86_alu_reg_imm (code, X86_AND, X86_EDX, 15);
3366 x86_alu_reg_imm (code, X86_CMP, X86_EDX, 0);
3367 x86_branch_disp (code, X86_CC_EQ, 3, FALSE);
3368 x86_breakpoint (code);
3371 /* compute max_offset in order to use short forward jumps */
3373 if (cfg->opt & MONO_OPT_BRANCH) {
3374 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
3375 MonoInst *ins = bb->code;
3376 bb->max_offset = max_offset;
3378 if (cfg->prof_options & MONO_PROFILE_COVERAGE)
3380 /* max alignment for loops */
3381 if ((cfg->opt & MONO_OPT_LOOP) && bb_is_loop_start (bb))
3382 max_offset += LOOP_ALIGNMENT;
3385 if (ins->opcode == OP_LABEL)
3386 ins->inst_c1 = max_offset;
3388 max_offset += ((guint8 *)ins_spec [ins->opcode])[MONO_INST_LEN];
3394 if (mono_jit_trace_calls != NULL && mono_trace_eval (method))
3395 code = mono_arch_instrument_prolog (cfg, mono_trace_enter_method, code, TRUE);
3397 /* load arguments allocated to register from the stack */
3398 sig = mono_method_signature (method);
3401 for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
3402 inst = cfg->varinfo [pos];
3403 if (inst->opcode == OP_REGVAR) {
3404 x86_mov_reg_membase (code, inst->dreg, X86_EBP, inst->inst_offset, 4);
3405 if (cfg->verbose_level > 2)
3406 g_print ("Argument %d assigned to register %s\n", pos, mono_arch_regname (inst->dreg));
3411 cfg->code_len = code - cfg->native_code;
3417 mono_arch_emit_epilog (MonoCompile *cfg)
3419 MonoMethod *method = cfg->method;
3420 MonoMethodSignature *sig = mono_method_signature (method);
3422 guint32 stack_to_pop;
3424 int max_epilog_size = 16;
3427 if (cfg->method->save_lmf)
3428 max_epilog_size += 128;
3430 if (mono_jit_trace_calls != NULL)
3431 max_epilog_size += 50;
3433 while (cfg->code_len + max_epilog_size > (cfg->code_size - 16)) {
3434 cfg->code_size *= 2;
3435 cfg->native_code = g_realloc (cfg->native_code, cfg->code_size);
3436 mono_jit_stats.code_reallocs++;
3439 code = cfg->native_code + cfg->code_len;
3441 if (mono_jit_trace_calls != NULL && mono_trace_eval (method))
3442 code = mono_arch_instrument_epilog (cfg, mono_trace_leave_method, code, TRUE);
3444 /* the code restoring the registers must be kept in sync with CEE_JMP */
3447 if (method->save_lmf) {
3448 gint32 prev_lmf_reg;
3449 gint32 lmf_offset = -sizeof (MonoLMF);
3451 /* Find a spare register */
3452 switch (sig->ret->type) {
3455 prev_lmf_reg = X86_EDI;
3456 cfg->used_int_regs |= (1 << X86_EDI);
3459 prev_lmf_reg = X86_EDX;
3463 /* reg = previous_lmf */
3464 x86_mov_reg_membase (code, prev_lmf_reg, X86_EBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, previous_lmf), 4);
3467 x86_mov_reg_membase (code, X86_ECX, X86_EBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, lmf_addr), 4);
3469 /* *(lmf) = previous_lmf */
3470 x86_mov_membase_reg (code, X86_ECX, 0, prev_lmf_reg, 4);
3472 /* restore caller saved regs */
3473 if (cfg->used_int_regs & (1 << X86_EBX)) {
3474 x86_mov_reg_membase (code, X86_EBX, X86_EBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, ebx), 4);
3477 if (cfg->used_int_regs & (1 << X86_EDI)) {
3478 x86_mov_reg_membase (code, X86_EDI, X86_EBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, edi), 4);
3480 if (cfg->used_int_regs & (1 << X86_ESI)) {
3481 x86_mov_reg_membase (code, X86_ESI, X86_EBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, esi), 4);
3484 /* EBP is restored by LEAVE */
3486 if (cfg->used_int_regs & (1 << X86_EBX)) {
3489 if (cfg->used_int_regs & (1 << X86_EDI)) {
3492 if (cfg->used_int_regs & (1 << X86_ESI)) {
3497 x86_lea_membase (code, X86_ESP, X86_EBP, pos);
3499 if (cfg->used_int_regs & (1 << X86_ESI)) {
3500 x86_pop_reg (code, X86_ESI);
3502 if (cfg->used_int_regs & (1 << X86_EDI)) {
3503 x86_pop_reg (code, X86_EDI);
3505 if (cfg->used_int_regs & (1 << X86_EBX)) {
3506 x86_pop_reg (code, X86_EBX);
3510 /* Load returned vtypes into registers if needed */
3511 cinfo = get_call_info (sig, FALSE);
3512 if (cinfo->ret.storage == ArgValuetypeInReg) {
3513 for (quad = 0; quad < 2; quad ++) {
3514 switch (cinfo->ret.pair_storage [quad]) {
3516 x86_mov_reg_membase (code, cinfo->ret.pair_regs [quad], cfg->ret->inst_basereg, cfg->ret->inst_offset + (quad * sizeof (gpointer)), 4);
3518 case ArgOnFloatFpStack:
3519 x86_fld_membase (code, cfg->ret->inst_basereg, cfg->ret->inst_offset + (quad * sizeof (gpointer)), FALSE);
3521 case ArgOnDoubleFpStack:
3522 x86_fld_membase (code, cfg->ret->inst_basereg, cfg->ret->inst_offset + (quad * sizeof (gpointer)), TRUE);
3527 g_assert_not_reached ();
3534 if (CALLCONV_IS_STDCALL (sig)) {
3535 MonoJitArgumentInfo *arg_info = alloca (sizeof (MonoJitArgumentInfo) * (sig->param_count + 1));
3537 stack_to_pop = mono_arch_get_argument_info (sig, sig->param_count, arg_info);
3538 } else if (MONO_TYPE_ISSTRUCT (mono_method_signature (cfg->method)->ret) && (cinfo->ret.storage == ArgOnStack))
3544 x86_ret_imm (code, stack_to_pop);
3550 cfg->code_len = code - cfg->native_code;
3552 g_assert (cfg->code_len < cfg->code_size);
3556 mono_arch_emit_exceptions (MonoCompile *cfg)
3558 MonoJumpInfo *patch_info;
3561 MonoClass *exc_classes [16];
3562 guint8 *exc_throw_start [16], *exc_throw_end [16];
3566 /* Compute needed space */
3567 for (patch_info = cfg->patch_info; patch_info; patch_info = patch_info->next) {
3568 if (patch_info->type == MONO_PATCH_INFO_EXC)
3573 * make sure we have enough space for exceptions
3574 * 16 is the size of two push_imm instructions and a call
3576 if (cfg->compile_aot)
3577 code_size = exc_count * 32;
3579 code_size = exc_count * 16;
3581 while (cfg->code_len + code_size > (cfg->code_size - 16)) {
3582 cfg->code_size *= 2;
3583 cfg->native_code = g_realloc (cfg->native_code, cfg->code_size);
3584 mono_jit_stats.code_reallocs++;
3587 code = cfg->native_code + cfg->code_len;
3590 for (patch_info = cfg->patch_info; patch_info; patch_info = patch_info->next) {
3591 switch (patch_info->type) {
3592 case MONO_PATCH_INFO_EXC: {
3593 MonoClass *exc_class;
3597 x86_patch (patch_info->ip.i + cfg->native_code, code);
3599 exc_class = mono_class_from_name (mono_defaults.corlib, "System", patch_info->data.name);
3600 g_assert (exc_class);
3601 throw_ip = patch_info->ip.i;
3603 /* Find a throw sequence for the same exception class */
3604 for (i = 0; i < nthrows; ++i)
3605 if (exc_classes [i] == exc_class)
3608 x86_push_imm (code, (exc_throw_end [i] - cfg->native_code) - throw_ip);
3609 x86_jump_code (code, exc_throw_start [i]);
3610 patch_info->type = MONO_PATCH_INFO_NONE;
3615 /* Compute size of code following the push <OFFSET> */
3618 if ((code - cfg->native_code) - throw_ip < 126 - size) {
3619 /* Use the shorter form */
3621 x86_push_imm (code, 0);
3625 x86_push_imm (code, 0xf0f0f0f0);
3630 exc_classes [nthrows] = exc_class;
3631 exc_throw_start [nthrows] = code;
3634 x86_push_imm (code, exc_class->type_token);
3635 patch_info->data.name = "mono_arch_throw_corlib_exception";
3636 patch_info->type = MONO_PATCH_INFO_INTERNAL_METHOD;
3637 patch_info->ip.i = code - cfg->native_code;
3638 x86_call_code (code, 0);
3639 x86_push_imm (buf, (code - cfg->native_code) - throw_ip);
3644 exc_throw_end [nthrows] = code;
3656 cfg->code_len = code - cfg->native_code;
3658 g_assert (cfg->code_len < cfg->code_size);
3662 mono_arch_flush_icache (guint8 *code, gint size)
3668 mono_arch_flush_register_windows (void)
3673 * Support for fast access to the thread-local lmf structure using the GS
3674 * segment register on NPTL + kernel 2.6.x.
3677 static gboolean tls_offset_inited = FALSE;
3680 mono_arch_setup_jit_tls_data (MonoJitTlsData *tls)
3682 if (!tls_offset_inited) {
3683 if (!getenv ("MONO_NO_TLS")) {
3684 #ifdef PLATFORM_WIN32
3686 * We need to init this multiple times, since when we are first called, the key might not
3687 * be initialized yet.
3689 appdomain_tls_offset = mono_domain_get_tls_key ();
3690 lmf_tls_offset = mono_get_jit_tls_key ();
3691 thread_tls_offset = mono_thread_get_tls_key ();
3693 /* Only 64 tls entries can be accessed using inline code */
3694 if (appdomain_tls_offset >= 64)
3695 appdomain_tls_offset = -1;
3696 if (lmf_tls_offset >= 64)
3697 lmf_tls_offset = -1;
3698 if (thread_tls_offset >= 64)
3699 thread_tls_offset = -1;
3702 optimize_for_xen = access ("/proc/xen", F_OK) == 0;
3704 tls_offset_inited = TRUE;
3705 appdomain_tls_offset = mono_domain_get_tls_offset ();
3706 lmf_tls_offset = mono_get_lmf_tls_offset ();
3707 thread_tls_offset = mono_thread_get_tls_offset ();
3714 mono_arch_free_jit_tls_data (MonoJitTlsData *tls)
3719 mono_arch_emit_this_vret_args (MonoCompile *cfg, MonoCallInst *inst, int this_reg, int this_type, int vt_reg)
3721 MonoCallInst *call = (MonoCallInst*)inst;
3722 CallInfo *cinfo = get_call_info (inst->signature, FALSE);
3724 /* add the this argument */
3725 if (this_reg != -1) {
3726 if (cinfo->args [0].storage == ArgInIReg) {
3728 MONO_INST_NEW (cfg, this, OP_MOVE);
3729 this->type = this_type;
3730 this->sreg1 = this_reg;
3731 this->dreg = mono_regstate_next_int (cfg->rs);
3732 mono_bblock_add_inst (cfg->cbb, this);
3734 mono_call_inst_add_outarg_reg (cfg, call, this->dreg, cinfo->args [0].reg, FALSE);
3738 MONO_INST_NEW (cfg, this, OP_OUTARG);
3739 this->type = this_type;
3740 this->sreg1 = this_reg;
3741 mono_bblock_add_inst (cfg->cbb, this);
3748 if (cinfo->ret.storage == ArgValuetypeInReg) {
3750 * The valuetype is in EAX:EDX after the call, needs to be copied to
3751 * the stack. Save the address here, so the call instruction can
3754 MONO_INST_NEW (cfg, vtarg, OP_STORE_MEMBASE_REG);
3755 vtarg->inst_destbasereg = X86_ESP;
3756 vtarg->inst_offset = inst->stack_usage;
3757 vtarg->sreg1 = vt_reg;
3758 mono_bblock_add_inst (cfg->cbb, vtarg);
3760 else if (cinfo->ret.storage == ArgInIReg) {
3761 /* The return address is passed in a register */
3762 MONO_INST_NEW (cfg, vtarg, OP_MOVE);
3763 vtarg->sreg1 = vt_reg;
3764 vtarg->dreg = mono_regstate_next_int (cfg->rs);
3765 mono_bblock_add_inst (cfg->cbb, vtarg);
3767 mono_call_inst_add_outarg_reg (cfg, call, vtarg->dreg, cinfo->ret.reg, FALSE);
3770 MONO_INST_NEW (cfg, vtarg, OP_OUTARG);
3771 vtarg->type = STACK_MP;
3772 vtarg->sreg1 = vt_reg;
3773 mono_bblock_add_inst (cfg->cbb, vtarg);
3781 mono_arch_get_inst_for_method (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature *fsig, MonoInst **args)
3783 MonoInst *ins = NULL;
3785 if (cmethod->klass == mono_defaults.math_class) {
3786 if (strcmp (cmethod->name, "Sin") == 0) {
3787 MONO_INST_NEW (cfg, ins, OP_SIN);
3788 ins->inst_i0 = args [0];
3789 } else if (strcmp (cmethod->name, "Cos") == 0) {
3790 MONO_INST_NEW (cfg, ins, OP_COS);
3791 ins->inst_i0 = args [0];
3792 } else if (strcmp (cmethod->name, "Tan") == 0) {
3793 MONO_INST_NEW (cfg, ins, OP_TAN);
3794 ins->inst_i0 = args [0];
3795 } else if (strcmp (cmethod->name, "Atan") == 0) {
3796 MONO_INST_NEW (cfg, ins, OP_ATAN);
3797 ins->inst_i0 = args [0];
3798 } else if (strcmp (cmethod->name, "Sqrt") == 0) {
3799 MONO_INST_NEW (cfg, ins, OP_SQRT);
3800 ins->inst_i0 = args [0];
3801 } else if (strcmp (cmethod->name, "Abs") == 0 && fsig->params [0]->type == MONO_TYPE_R8) {
3802 MONO_INST_NEW (cfg, ins, OP_ABS);
3803 ins->inst_i0 = args [0];
3806 /* OP_FREM is not IEEE compatible */
3807 else if (strcmp (cmethod->name, "IEEERemainder") == 0) {
3808 MONO_INST_NEW (cfg, ins, OP_FREM);
3809 ins->inst_i0 = args [0];
3810 ins->inst_i1 = args [1];
3813 } else if (cmethod->klass == mono_defaults.thread_class &&
3814 strcmp (cmethod->name, "MemoryBarrier") == 0) {
3815 MONO_INST_NEW (cfg, ins, OP_MEMORY_BARRIER);
3816 } else if(cmethod->klass->image == mono_defaults.corlib &&
3817 (strcmp (cmethod->klass->name_space, "System.Threading") == 0) &&
3818 (strcmp (cmethod->klass->name, "Interlocked") == 0)) {
3820 if (strcmp (cmethod->name, "Increment") == 0 && fsig->params [0]->type == MONO_TYPE_I4) {
3821 MonoInst *ins_iconst;
3823 MONO_INST_NEW (cfg, ins, OP_ATOMIC_ADD_NEW_I4);
3824 MONO_INST_NEW (cfg, ins_iconst, OP_ICONST);
3825 ins_iconst->inst_c0 = 1;
3827 ins->inst_i0 = args [0];
3828 ins->inst_i1 = ins_iconst;
3829 } else if (strcmp (cmethod->name, "Decrement") == 0 && fsig->params [0]->type == MONO_TYPE_I4) {
3830 MonoInst *ins_iconst;
3832 MONO_INST_NEW (cfg, ins, OP_ATOMIC_ADD_NEW_I4);
3833 MONO_INST_NEW (cfg, ins_iconst, OP_ICONST);
3834 ins_iconst->inst_c0 = -1;
3836 ins->inst_i0 = args [0];
3837 ins->inst_i1 = ins_iconst;
3838 } else if (strcmp (cmethod->name, "Exchange") == 0 && fsig->params [0]->type == MONO_TYPE_I4) {
3839 MONO_INST_NEW (cfg, ins, OP_ATOMIC_EXCHANGE_I4);
3841 ins->inst_i0 = args [0];
3842 ins->inst_i1 = args [1];
3843 } else if (strcmp (cmethod->name, "Add") == 0 && fsig->params [0]->type == MONO_TYPE_I4) {
3844 MONO_INST_NEW (cfg, ins, OP_ATOMIC_ADD_NEW_I4);
3846 ins->inst_i0 = args [0];
3847 ins->inst_i1 = args [1];
3856 mono_arch_print_tree (MonoInst *tree, int arity)
3861 MonoInst* mono_arch_get_domain_intrinsic (MonoCompile* cfg)
3865 if (appdomain_tls_offset == -1)
3868 MONO_INST_NEW (cfg, ins, OP_TLS_GET);
3869 ins->inst_offset = appdomain_tls_offset;
3873 MonoInst* mono_arch_get_thread_intrinsic (MonoCompile* cfg)
3877 if (thread_tls_offset == -1)
3880 MONO_INST_NEW (cfg, ins, OP_TLS_GET);
3881 ins->inst_offset = thread_tls_offset;
3886 mono_arch_get_patch_offset (guint8 *code)
3888 if ((code [0] == 0x8b) && (x86_modrm_mod (code [1]) == 0x2))
3890 else if ((code [0] == 0xba))
3892 else if ((code [0] == 0x68))
3895 else if ((code [0] == 0xff) && (x86_modrm_reg (code [1]) == 0x6))
3896 /* push <OFFSET>(<REG>) */
3898 else if ((code [0] == 0xff) && (x86_modrm_reg (code [1]) == 0x2))
3899 /* call *<OFFSET>(<REG>) */
3901 else if ((code [0] == 0xdd) || (code [0] == 0xd9))
3904 else if ((code [0] == 0x58) && (code [1] == 0x05))
3905 /* pop %eax; add <OFFSET>, %eax */
3907 else if ((code [0] >= 0x58) && (code [0] <= 0x58 + X86_NREG) && (code [1] == 0x81))
3908 /* pop <REG>; add <OFFSET>, <REG> */
3911 g_assert_not_reached ();
3917 mono_arch_get_vcall_slot_addr (guint8 *code, gpointer *regs)
3922 /* go to the start of the call instruction
3924 * address_byte = (m << 6) | (o << 3) | reg
3925 * call opcode: 0xff address_byte displacement
3927 * 0xff m=2,o=2 imm32
3932 * A given byte sequence can match more than case here, so we have to be
3933 * really careful about the ordering of the cases. Longer sequences
3936 if ((code [-2] == 0x8b) && (x86_modrm_mod (code [-1]) == 0x2) && (code [4] == 0xff) && (x86_modrm_reg (code [5]) == 0x2) && (x86_modrm_mod (code [5]) == 0x0)) {
3938 * This is an interface call
3939 * 8b 80 0c e8 ff ff mov 0xffffe80c(%eax),%eax
3940 * ff 10 call *(%eax)
3942 reg = x86_modrm_rm (code [5]);
3944 } else if ((code [1] != 0xe8) && (code [3] == 0xff) && ((code [4] & 0x18) == 0x10) && ((code [4] >> 6) == 1)) {
3945 reg = code [4] & 0x07;
3946 disp = (signed char)code [5];
3948 if ((code [0] == 0xff) && ((code [1] & 0x18) == 0x10) && ((code [1] >> 6) == 2)) {
3949 reg = code [1] & 0x07;
3950 disp = *((gint32*)(code + 2));
3951 } else if ((code [1] == 0xe8)) {
3953 } else if ((code [4] == 0xff) && (((code [5] >> 6) & 0x3) == 0) && (((code [5] >> 3) & 0x7) == 2)) {
3955 * This is a interface call
3956 * 8b 40 30 mov 0x30(%eax),%eax
3957 * ff 10 call *(%eax)
3960 reg = code [5] & 0x07;
3966 return (gpointer*)(((gint32)(regs [reg])) + disp);
3970 mono_arch_get_delegate_method_ptr_addr (guint8* code, gpointer *regs)
3976 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)) {
3977 reg = x86_modrm_rm (code [1]);
3983 return (gpointer*)(((gint32)(regs [reg])) + disp);