2 * mini-amd64.c: AMD64 backend for the Mono code generator
7 * Paolo Molaro (lupus@ximian.com)
8 * Dietmar Maurer (dietmar@ximian.com)
11 * (C) 2003 Ximian, Inc.
20 #include <mono/metadata/appdomain.h>
21 #include <mono/metadata/debug-helpers.h>
22 #include <mono/metadata/threads.h>
23 #include <mono/metadata/profiler-private.h>
24 #include <mono/metadata/mono-debug.h>
25 #include <mono/utils/mono-math.h>
28 #include "mini-amd64.h"
30 #include "cpu-amd64.h"
32 static gint lmf_tls_offset = -1;
33 static gint lmf_addr_tls_offset = -1;
34 static gint appdomain_tls_offset = -1;
35 static gint thread_tls_offset = -1;
38 static gboolean optimize_for_xen = TRUE;
40 #define optimize_for_xen 0
43 static gboolean use_sse2 = !MONO_ARCH_USE_FPSTACK;
45 #define ALIGN_TO(val,align) ((((guint64)val) + ((align) - 1)) & ~((align) - 1))
47 #define IS_IMM32(val) ((((guint64)val) >> 32) == 0)
49 #define IS_REX(inst) (((inst) >= 0x40) && ((inst) <= 0x4f))
52 /* Under windows, the default pinvoke calling convention is stdcall */
53 #define CALLCONV_IS_STDCALL(call_conv) (((call_conv) == MONO_CALL_STDCALL) || ((call_conv) == MONO_CALL_DEFAULT))
55 #define CALLCONV_IS_STDCALL(call_conv) ((call_conv) == MONO_CALL_STDCALL)
58 /* This mutex protects architecture specific caches */
59 #define mono_mini_arch_lock() EnterCriticalSection (&mini_arch_mutex)
60 #define mono_mini_arch_unlock() LeaveCriticalSection (&mini_arch_mutex)
61 static CRITICAL_SECTION mini_arch_mutex;
64 mono_breakpoint_info [MONO_BREAKPOINT_ARRAY_SIZE];
66 #define ARGS_OFFSET 16
67 #define GP_SCRATCH_REG AMD64_R11
70 * AMD64 register usage:
71 * - callee saved registers are used for global register allocation
72 * - %r11 is used for materializing 64 bit constants in opcodes
73 * - the rest is used for local allocation
77 * Floating point comparison results:
86 #define NOT_IMPLEMENTED g_assert_not_reached ()
89 mono_arch_regname (int reg) {
91 case AMD64_RAX: return "%rax";
92 case AMD64_RBX: return "%rbx";
93 case AMD64_RCX: return "%rcx";
94 case AMD64_RDX: return "%rdx";
95 case AMD64_RSP: return "%rsp";
96 case AMD64_RBP: return "%rbp";
97 case AMD64_RDI: return "%rdi";
98 case AMD64_RSI: return "%rsi";
99 case AMD64_R8: return "%r8";
100 case AMD64_R9: return "%r9";
101 case AMD64_R10: return "%r10";
102 case AMD64_R11: return "%r11";
103 case AMD64_R12: return "%r12";
104 case AMD64_R13: return "%r13";
105 case AMD64_R14: return "%r14";
106 case AMD64_R15: return "%r15";
111 static const char * xmmregs [] = {
112 "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7", "xmm8",
113 "xmm9", "xmm10", "xmm11", "xmm12", "xmm13", "xmm14", "xmm15"
117 mono_arch_fregname (int reg)
119 if (reg < AMD64_XMM_NREG)
120 return xmmregs [reg];
125 G_GNUC_UNUSED static void
130 G_GNUC_UNUSED static gboolean
133 static int count = 0;
136 if (!getenv ("COUNT"))
139 if (count == atoi (getenv ("COUNT"))) {
143 if (count > atoi (getenv ("COUNT"))) {
154 return debug_count ();
160 static inline gboolean
161 amd64_is_near_call (guint8 *code)
164 if ((code [0] >= 0x40) && (code [0] <= 0x4f))
167 return code [0] == 0xe8;
171 amd64_patch (unsigned char* code, gpointer target)
174 if ((code [0] >= 0x40) && (code [0] <= 0x4f))
177 if ((code [0] & 0xf8) == 0xb8) {
178 /* amd64_set_reg_template */
179 *(guint64*)(code + 1) = (guint64)target;
181 else if (code [0] == 0x8b) {
182 /* mov 0(%rip), %dreg */
183 *(guint32*)(code + 2) = (guint32)(guint64)target - 7;
185 else if ((code [0] == 0xff) && (code [1] == 0x15)) {
186 /* call *<OFFSET>(%rip) */
187 *(guint32*)(code + 2) = ((guint32)(guint64)target) - 7;
189 else if ((code [0] == 0xe8)) {
191 gint64 disp = (guint8*)target - (guint8*)code;
192 g_assert (amd64_is_imm32 (disp));
193 x86_patch (code, (unsigned char*)target);
196 x86_patch (code, (unsigned char*)target);
200 mono_amd64_patch (unsigned char* code, gpointer target)
202 amd64_patch (code, target);
211 ArgNone /* only in pair_storage */
219 /* Only if storage == ArgValuetypeInReg */
220 ArgStorage pair_storage [2];
229 gboolean need_stack_align;
235 #define DEBUG(a) if (cfg->verbose_level > 1) a
237 #define NEW_ICONST(cfg,dest,val) do { \
238 (dest) = mono_mempool_alloc0 ((cfg)->mempool, sizeof (MonoInst)); \
239 (dest)->opcode = OP_ICONST; \
240 (dest)->inst_c0 = (val); \
241 (dest)->type = STACK_I4; \
244 #ifdef PLATFORM_WIN32
247 static AMD64_Reg_No param_regs [] = { AMD64_RCX, AMD64_RDX, AMD64_R8, AMD64_R9 };
249 static AMD64_Reg_No return_regs [] = { AMD64_RAX, AMD64_RDX };
253 static AMD64_Reg_No param_regs [] = { AMD64_RDI, AMD64_RSI, AMD64_RDX, AMD64_RCX, AMD64_R8, AMD64_R9 };
255 static AMD64_Reg_No return_regs [] = { AMD64_RAX, AMD64_RDX };
259 add_general (guint32 *gr, guint32 *stack_size, ArgInfo *ainfo)
261 ainfo->offset = *stack_size;
263 if (*gr >= PARAM_REGS) {
264 ainfo->storage = ArgOnStack;
265 (*stack_size) += sizeof (gpointer);
268 ainfo->storage = ArgInIReg;
269 ainfo->reg = param_regs [*gr];
274 #ifdef PLATFORM_WIN32
275 #define FLOAT_PARAM_REGS 4
277 #define FLOAT_PARAM_REGS 8
281 add_float (guint32 *gr, guint32 *stack_size, ArgInfo *ainfo, gboolean is_double)
283 ainfo->offset = *stack_size;
285 if (*gr >= FLOAT_PARAM_REGS) {
286 ainfo->storage = ArgOnStack;
287 (*stack_size) += sizeof (gpointer);
290 /* A double register */
292 ainfo->storage = ArgInDoubleSSEReg;
294 ainfo->storage = ArgInFloatSSEReg;
300 typedef enum ArgumentClass {
308 merge_argument_class_from_type (MonoType *type, ArgumentClass class1)
310 ArgumentClass class2 = ARG_CLASS_NO_CLASS;
313 ptype = mono_type_get_underlying_type (type);
314 switch (ptype->type) {
315 case MONO_TYPE_BOOLEAN:
325 case MONO_TYPE_STRING:
326 case MONO_TYPE_OBJECT:
327 case MONO_TYPE_CLASS:
328 case MONO_TYPE_SZARRAY:
330 case MONO_TYPE_FNPTR:
331 case MONO_TYPE_ARRAY:
334 class2 = ARG_CLASS_INTEGER;
338 class2 = ARG_CLASS_SSE;
341 case MONO_TYPE_TYPEDBYREF:
342 g_assert_not_reached ();
344 case MONO_TYPE_GENERICINST:
345 if (!mono_type_generic_inst_is_valuetype (ptype)) {
346 class2 = ARG_CLASS_INTEGER;
350 case MONO_TYPE_VALUETYPE: {
351 MonoMarshalType *info = mono_marshal_load_type_info (ptype->data.klass);
354 for (i = 0; i < info->num_fields; ++i) {
356 class2 = merge_argument_class_from_type (info->fields [i].field->type, class2);
361 g_assert_not_reached ();
365 if (class1 == class2)
367 else if (class1 == ARG_CLASS_NO_CLASS)
369 else if ((class1 == ARG_CLASS_MEMORY) || (class2 == ARG_CLASS_MEMORY))
370 class1 = ARG_CLASS_MEMORY;
371 else if ((class1 == ARG_CLASS_INTEGER) || (class2 == ARG_CLASS_INTEGER))
372 class1 = ARG_CLASS_INTEGER;
374 class1 = ARG_CLASS_SSE;
380 add_valuetype (MonoGenericSharingContext *gsctx, MonoMethodSignature *sig, ArgInfo *ainfo, MonoType *type,
382 guint32 *gr, guint32 *fr, guint32 *stack_size)
384 guint32 size, quad, nquads, i;
385 ArgumentClass args [2];
386 MonoMarshalType *info;
389 klass = mono_class_from_mono_type (type);
391 size = mono_type_native_stack_size (&klass->byval_arg, NULL);
393 size = mini_type_stack_size (gsctx, &klass->byval_arg, NULL);
395 if (!sig->pinvoke || (size == 0) || (size > 16)) {
396 /* Allways pass in memory */
397 ainfo->offset = *stack_size;
398 *stack_size += ALIGN_TO (size, 8);
399 ainfo->storage = ArgOnStack;
404 /* FIXME: Handle structs smaller than 8 bytes */
405 //if ((size % 8) != 0)
414 * Implement the algorithm from section 3.2.3 of the X86_64 ABI.
415 * The X87 and SSEUP stuff is left out since there are no such types in
418 info = mono_marshal_load_type_info (klass);
420 if (info->native_size > 16) {
421 ainfo->offset = *stack_size;
422 *stack_size += ALIGN_TO (info->native_size, 8);
423 ainfo->storage = ArgOnStack;
428 args [0] = ARG_CLASS_NO_CLASS;
429 args [1] = ARG_CLASS_NO_CLASS;
430 for (quad = 0; quad < nquads; ++quad) {
433 ArgumentClass class1;
435 class1 = ARG_CLASS_NO_CLASS;
436 for (i = 0; i < info->num_fields; ++i) {
437 size = mono_marshal_type_size (info->fields [i].field->type,
438 info->fields [i].mspec,
439 &align, TRUE, klass->unicode);
440 if ((info->fields [i].offset < 8) && (info->fields [i].offset + size) > 8) {
441 /* Unaligned field */
445 /* Skip fields in other quad */
446 if ((quad == 0) && (info->fields [i].offset >= 8))
448 if ((quad == 1) && (info->fields [i].offset < 8))
451 class1 = merge_argument_class_from_type (info->fields [i].field->type, class1);
453 g_assert (class1 != ARG_CLASS_NO_CLASS);
454 args [quad] = class1;
457 /* Post merger cleanup */
458 if ((args [0] == ARG_CLASS_MEMORY) || (args [1] == ARG_CLASS_MEMORY))
459 args [0] = args [1] = ARG_CLASS_MEMORY;
461 /* Allocate registers */
466 ainfo->storage = ArgValuetypeInReg;
467 ainfo->pair_storage [0] = ainfo->pair_storage [1] = ArgNone;
468 for (quad = 0; quad < nquads; ++quad) {
469 switch (args [quad]) {
470 case ARG_CLASS_INTEGER:
471 if (*gr >= PARAM_REGS)
472 args [quad] = ARG_CLASS_MEMORY;
474 ainfo->pair_storage [quad] = ArgInIReg;
476 ainfo->pair_regs [quad] = return_regs [*gr];
478 ainfo->pair_regs [quad] = param_regs [*gr];
483 if (*fr >= FLOAT_PARAM_REGS)
484 args [quad] = ARG_CLASS_MEMORY;
486 ainfo->pair_storage [quad] = ArgInDoubleSSEReg;
487 ainfo->pair_regs [quad] = *fr;
491 case ARG_CLASS_MEMORY:
494 g_assert_not_reached ();
498 if ((args [0] == ARG_CLASS_MEMORY) || (args [1] == ARG_CLASS_MEMORY)) {
499 /* Revert possible register assignments */
503 ainfo->offset = *stack_size;
504 *stack_size += ALIGN_TO (info->native_size, 8);
505 ainfo->storage = ArgOnStack;
513 * Obtain information about a call according to the calling convention.
514 * For AMD64, see the "System V ABI, x86-64 Architecture Processor Supplement
515 * Draft Version 0.23" document for more information.
518 get_call_info (MonoCompile *cfg, MonoMemPool *mp, MonoMethodSignature *sig, gboolean is_pinvoke)
522 int n = sig->hasthis + sig->param_count;
523 guint32 stack_size = 0;
525 MonoGenericSharingContext *gsctx = cfg ? cfg->generic_sharing_context : NULL;
528 cinfo = mono_mempool_alloc0 (mp, sizeof (CallInfo) + (sizeof (ArgInfo) * n));
530 cinfo = g_malloc0 (sizeof (CallInfo) + (sizeof (ArgInfo) * n));
537 ret_type = mono_type_get_underlying_type (sig->ret);
538 ret_type = mini_get_basic_type_from_generic (gsctx, ret_type);
539 switch (ret_type->type) {
540 case MONO_TYPE_BOOLEAN:
551 case MONO_TYPE_FNPTR:
552 case MONO_TYPE_CLASS:
553 case MONO_TYPE_OBJECT:
554 case MONO_TYPE_SZARRAY:
555 case MONO_TYPE_ARRAY:
556 case MONO_TYPE_STRING:
557 cinfo->ret.storage = ArgInIReg;
558 cinfo->ret.reg = AMD64_RAX;
562 cinfo->ret.storage = ArgInIReg;
563 cinfo->ret.reg = AMD64_RAX;
566 cinfo->ret.storage = ArgInFloatSSEReg;
567 cinfo->ret.reg = AMD64_XMM0;
570 cinfo->ret.storage = ArgInDoubleSSEReg;
571 cinfo->ret.reg = AMD64_XMM0;
573 case MONO_TYPE_GENERICINST:
574 if (!mono_type_generic_inst_is_valuetype (sig->ret)) {
575 cinfo->ret.storage = ArgInIReg;
576 cinfo->ret.reg = AMD64_RAX;
580 case MONO_TYPE_VALUETYPE: {
581 guint32 tmp_gr = 0, tmp_fr = 0, tmp_stacksize = 0;
583 add_valuetype (gsctx, sig, &cinfo->ret, sig->ret, TRUE, &tmp_gr, &tmp_fr, &tmp_stacksize);
584 if (cinfo->ret.storage == ArgOnStack)
585 /* The caller passes the address where the value is stored */
586 add_general (&gr, &stack_size, &cinfo->ret);
589 case MONO_TYPE_TYPEDBYREF:
590 /* Same as a valuetype with size 24 */
591 add_general (&gr, &stack_size, &cinfo->ret);
597 g_error ("Can't handle as return value 0x%x", sig->ret->type);
603 add_general (&gr, &stack_size, cinfo->args + 0);
605 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG) && (n == 0)) {
607 fr = FLOAT_PARAM_REGS;
609 /* Emit the signature cookie just before the implicit arguments */
610 add_general (&gr, &stack_size, &cinfo->sig_cookie);
613 for (i = 0; i < sig->param_count; ++i) {
614 ArgInfo *ainfo = &cinfo->args [sig->hasthis + i];
617 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG) && (i == sig->sentinelpos)) {
618 /* We allways pass the sig cookie on the stack for simplicity */
620 * Prevent implicit arguments + the sig cookie from being passed
624 fr = FLOAT_PARAM_REGS;
626 /* Emit the signature cookie just before the implicit arguments */
627 add_general (&gr, &stack_size, &cinfo->sig_cookie);
630 if (sig->params [i]->byref) {
631 add_general (&gr, &stack_size, ainfo);
634 ptype = mono_type_get_underlying_type (sig->params [i]);
635 ptype = mini_get_basic_type_from_generic (gsctx, ptype);
636 switch (ptype->type) {
637 case MONO_TYPE_BOOLEAN:
640 add_general (&gr, &stack_size, ainfo);
645 add_general (&gr, &stack_size, ainfo);
649 add_general (&gr, &stack_size, ainfo);
654 case MONO_TYPE_FNPTR:
655 case MONO_TYPE_CLASS:
656 case MONO_TYPE_OBJECT:
657 case MONO_TYPE_STRING:
658 case MONO_TYPE_SZARRAY:
659 case MONO_TYPE_ARRAY:
660 add_general (&gr, &stack_size, ainfo);
662 case MONO_TYPE_GENERICINST:
663 if (!mono_type_generic_inst_is_valuetype (ptype)) {
664 add_general (&gr, &stack_size, ainfo);
668 case MONO_TYPE_VALUETYPE:
669 add_valuetype (gsctx, sig, ainfo, sig->params [i], FALSE, &gr, &fr, &stack_size);
671 case MONO_TYPE_TYPEDBYREF:
672 stack_size += sizeof (MonoTypedRef);
673 ainfo->storage = ArgOnStack;
677 add_general (&gr, &stack_size, ainfo);
680 add_float (&fr, &stack_size, ainfo, FALSE);
683 add_float (&fr, &stack_size, ainfo, TRUE);
686 g_assert_not_reached ();
690 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG) && (n > 0) && (sig->sentinelpos == sig->param_count)) {
692 fr = FLOAT_PARAM_REGS;
694 /* Emit the signature cookie just before the implicit arguments */
695 add_general (&gr, &stack_size, &cinfo->sig_cookie);
698 #ifdef PLATFORM_WIN32
699 if (stack_size < 32) {
700 /* The Win64 ABI requires 32 bits */
705 if (stack_size & 0x8) {
706 /* The AMD64 ABI requires each stack frame to be 16 byte aligned */
707 cinfo->need_stack_align = TRUE;
711 cinfo->stack_usage = stack_size;
712 cinfo->reg_usage = gr;
713 cinfo->freg_usage = fr;
718 * mono_arch_get_argument_info:
719 * @csig: a method signature
720 * @param_count: the number of parameters to consider
721 * @arg_info: an array to store the result infos
723 * Gathers information on parameters such as size, alignment and
724 * padding. arg_info should be large enought to hold param_count + 1 entries.
726 * Returns the size of the argument area on the stack.
729 mono_arch_get_argument_info (MonoMethodSignature *csig, int param_count, MonoJitArgumentInfo *arg_info)
732 CallInfo *cinfo = get_call_info (NULL, NULL, csig, FALSE);
733 guint32 args_size = cinfo->stack_usage;
735 /* The arguments are saved to a stack area in mono_arch_instrument_prolog */
737 arg_info [0].offset = 0;
740 for (k = 0; k < param_count; k++) {
741 arg_info [k + 1].offset = ((k + csig->hasthis) * 8);
743 arg_info [k + 1].size = 0;
752 cpuid (int id, int* p_eax, int* p_ebx, int* p_ecx, int* p_edx)
754 __asm__ __volatile__ ("cpuid"
755 : "=a" (*p_eax), "=b" (*p_ebx), "=c" (*p_ecx), "=d" (*p_edx)
761 * Initialize the cpu to execute managed code.
764 mono_arch_cpu_init (void)
769 /* spec compliance requires running with double precision */
770 __asm__ __volatile__ ("fnstcw %0\n": "=m" (fpcw));
771 fpcw &= ~X86_FPCW_PRECC_MASK;
772 fpcw |= X86_FPCW_PREC_DOUBLE;
773 __asm__ __volatile__ ("fldcw %0\n": : "m" (fpcw));
774 __asm__ __volatile__ ("fnstcw %0\n": "=m" (fpcw));
776 _control87 (_PC_53, MCW_PC);
781 * Initialize architecture specific code.
784 mono_arch_init (void)
786 InitializeCriticalSection (&mini_arch_mutex);
790 * Cleanup architecture specific code.
793 mono_arch_cleanup (void)
795 DeleteCriticalSection (&mini_arch_mutex);
799 * This function returns the optimizations supported on this cpu.
802 mono_arch_cpu_optimizazions (guint32 *exclude_mask)
804 int eax, ebx, ecx, edx;
810 /* Feature Flags function, flags returned in EDX. */
811 if (cpuid (1, &eax, &ebx, &ecx, &edx)) {
812 if (edx & (1 << 15)) {
813 opts |= MONO_OPT_CMOV;
815 opts |= MONO_OPT_FCMOV;
817 *exclude_mask |= MONO_OPT_FCMOV;
819 *exclude_mask |= MONO_OPT_CMOV;
825 mono_amd64_is_sse2 (void)
831 mono_arch_get_allocatable_int_vars (MonoCompile *cfg)
836 for (i = 0; i < cfg->num_varinfo; i++) {
837 MonoInst *ins = cfg->varinfo [i];
838 MonoMethodVar *vmv = MONO_VARINFO (cfg, i);
841 if (vmv->range.first_use.abs_pos >= vmv->range.last_use.abs_pos)
844 if ((ins->flags & (MONO_INST_IS_DEAD|MONO_INST_VOLATILE|MONO_INST_INDIRECT)) ||
845 (ins->opcode != OP_LOCAL && ins->opcode != OP_ARG))
848 if (mono_is_regsize_var (ins->inst_vtype)) {
849 g_assert (MONO_VARINFO (cfg, i)->reg == -1);
850 g_assert (i == vmv->idx);
851 vars = g_list_prepend (vars, vmv);
855 vars = mono_varlist_sort (cfg, vars, 0);
861 * mono_arch_compute_omit_fp:
863 * Determine whenever the frame pointer can be eliminated.
866 mono_arch_compute_omit_fp (MonoCompile *cfg)
868 MonoMethodSignature *sig;
869 MonoMethodHeader *header;
873 if (cfg->arch.omit_fp_computed)
876 header = mono_method_get_header (cfg->method);
878 sig = mono_method_signature (cfg->method);
880 if (!cfg->arch.cinfo)
881 cfg->arch.cinfo = get_call_info (cfg, cfg->mempool, sig, FALSE);
882 cinfo = cfg->arch.cinfo;
885 * FIXME: Remove some of the restrictions.
887 cfg->arch.omit_fp = TRUE;
888 cfg->arch.omit_fp_computed = TRUE;
890 /* Temporarily disable this when running in the debugger until we have support
891 * for this in the debugger. */
892 if (mono_debug_using_mono_debugger ())
893 cfg->arch.omit_fp = FALSE;
895 if (!debug_omit_fp ())
896 cfg->arch.omit_fp = FALSE;
898 if (cfg->method->save_lmf)
899 cfg->arch.omit_fp = FALSE;
901 if (cfg->flags & MONO_CFG_HAS_ALLOCA)
902 cfg->arch.omit_fp = FALSE;
903 if (header->num_clauses)
904 cfg->arch.omit_fp = FALSE;
906 cfg->arch.omit_fp = FALSE;
907 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG))
908 cfg->arch.omit_fp = FALSE;
909 if ((mono_jit_trace_calls != NULL && mono_trace_eval (cfg->method)) ||
910 (cfg->prof_options & MONO_PROFILE_ENTER_LEAVE))
911 cfg->arch.omit_fp = FALSE;
912 for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
913 ArgInfo *ainfo = &cinfo->args [i];
915 if (ainfo->storage == ArgOnStack) {
917 * The stack offset can only be determined when the frame
920 cfg->arch.omit_fp = FALSE;
924 if (cinfo->ret.storage == ArgValuetypeInReg)
925 cfg->arch.omit_fp = FALSE;
928 for (i = cfg->locals_start; i < cfg->num_varinfo; i++) {
929 MonoInst *ins = cfg->varinfo [i];
932 locals_size += mono_type_size (ins->inst_vtype, &ialign);
935 if ((cfg->num_varinfo > 10000) || (locals_size >= (1 << 15))) {
936 /* Avoid hitting the stack_alloc_size < (1 << 16) assertion in emit_epilog () */
937 cfg->arch.omit_fp = FALSE;
942 mono_arch_get_global_int_regs (MonoCompile *cfg)
946 mono_arch_compute_omit_fp (cfg);
948 if (cfg->arch.omit_fp)
949 regs = g_list_prepend (regs, (gpointer)AMD64_RBP);
951 /* We use the callee saved registers for global allocation */
952 regs = g_list_prepend (regs, (gpointer)AMD64_RBX);
953 regs = g_list_prepend (regs, (gpointer)AMD64_R12);
954 regs = g_list_prepend (regs, (gpointer)AMD64_R13);
955 regs = g_list_prepend (regs, (gpointer)AMD64_R14);
956 regs = g_list_prepend (regs, (gpointer)AMD64_R15);
962 * mono_arch_regalloc_cost:
964 * Return the cost, in number of memory references, of the action of
965 * allocating the variable VMV into a register during global register
969 mono_arch_regalloc_cost (MonoCompile *cfg, MonoMethodVar *vmv)
971 MonoInst *ins = cfg->varinfo [vmv->idx];
973 if (cfg->method->save_lmf)
974 /* The register is already saved */
975 /* substract 1 for the invisible store in the prolog */
976 return (ins->opcode == OP_ARG) ? 0 : 1;
979 return (ins->opcode == OP_ARG) ? 1 : 2;
983 mono_arch_allocate_vars (MonoCompile *cfg)
985 MonoMethodSignature *sig;
986 MonoMethodHeader *header;
989 guint32 locals_stack_size, locals_stack_align;
993 header = mono_method_get_header (cfg->method);
995 sig = mono_method_signature (cfg->method);
997 cinfo = cfg->arch.cinfo;
999 mono_arch_compute_omit_fp (cfg);
1002 * We use the ABI calling conventions for managed code as well.
1003 * Exception: valuetypes are never passed or returned in registers.
1006 if (cfg->arch.omit_fp) {
1007 cfg->flags |= MONO_CFG_HAS_SPILLUP;
1008 cfg->frame_reg = AMD64_RSP;
1011 /* Locals are allocated backwards from %fp */
1012 cfg->frame_reg = AMD64_RBP;
1016 cfg->arch.reg_save_area_offset = offset;
1018 /* Reserve space for caller saved registers */
1019 for (i = 0; i < AMD64_NREG; ++i)
1020 if (AMD64_IS_CALLEE_SAVED_REG (i) && (cfg->used_int_regs & (1 << i))) {
1021 offset += sizeof (gpointer);
1024 if (cfg->method->save_lmf) {
1025 /* Reserve stack space for saving LMF + argument regs */
1026 guint32 size = sizeof (MonoLMF);
1028 if (lmf_addr_tls_offset == -1)
1029 /* Need to save argument regs too */
1030 size += (AMD64_NREG * 8) + (8 * 8);
1032 if (cfg->arch.omit_fp) {
1033 cfg->arch.lmf_offset = offset;
1038 cfg->arch.lmf_offset = -offset;
1042 if (sig->ret->type != MONO_TYPE_VOID) {
1043 switch (cinfo->ret.storage) {
1045 case ArgInFloatSSEReg:
1046 case ArgInDoubleSSEReg:
1047 if ((MONO_TYPE_ISSTRUCT (sig->ret) && !mono_class_from_mono_type (sig->ret)->enumtype) || (sig->ret->type == MONO_TYPE_TYPEDBYREF)) {
1048 /* The register is volatile */
1049 cfg->ret->opcode = OP_REGOFFSET;
1050 cfg->ret->inst_basereg = cfg->frame_reg;
1051 if (cfg->arch.omit_fp) {
1052 cfg->ret->inst_offset = offset;
1056 cfg->ret->inst_offset = -offset;
1060 cfg->ret->opcode = OP_REGVAR;
1061 cfg->ret->inst_c0 = cinfo->ret.reg;
1064 case ArgValuetypeInReg:
1065 /* Allocate a local to hold the result, the epilog will copy it to the correct place */
1066 g_assert (!cfg->arch.omit_fp);
1068 cfg->ret->opcode = OP_REGOFFSET;
1069 cfg->ret->inst_basereg = cfg->frame_reg;
1070 cfg->ret->inst_offset = - offset;
1073 g_assert_not_reached ();
1075 cfg->ret->dreg = cfg->ret->inst_c0;
1078 /* Allocate locals */
1079 offsets = mono_allocate_stack_slots_full (cfg, cfg->arch.omit_fp ? FALSE: TRUE, &locals_stack_size, &locals_stack_align);
1080 if (locals_stack_align) {
1081 offset += (locals_stack_align - 1);
1082 offset &= ~(locals_stack_align - 1);
1084 for (i = cfg->locals_start; i < cfg->num_varinfo; i++) {
1085 if (offsets [i] != -1) {
1086 MonoInst *inst = cfg->varinfo [i];
1087 inst->opcode = OP_REGOFFSET;
1088 inst->inst_basereg = cfg->frame_reg;
1089 if (cfg->arch.omit_fp)
1090 inst->inst_offset = (offset + offsets [i]);
1092 inst->inst_offset = - (offset + offsets [i]);
1093 //printf ("allocated local %d to ", i); mono_print_tree_nl (inst);
1096 offset += locals_stack_size;
1098 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG)) {
1099 g_assert (!cfg->arch.omit_fp);
1100 g_assert (cinfo->sig_cookie.storage == ArgOnStack);
1101 cfg->sig_cookie = cinfo->sig_cookie.offset + ARGS_OFFSET;
1104 for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
1105 inst = cfg->args [i];
1106 if (inst->opcode != OP_REGVAR) {
1107 ArgInfo *ainfo = &cinfo->args [i];
1108 gboolean inreg = TRUE;
1111 if (sig->hasthis && (i == 0))
1112 arg_type = &mono_defaults.object_class->byval_arg;
1114 arg_type = sig->params [i - sig->hasthis];
1116 /* FIXME: Allocate volatile arguments to registers */
1117 if (inst->flags & (MONO_INST_VOLATILE|MONO_INST_INDIRECT))
1121 * Under AMD64, all registers used to pass arguments to functions
1122 * are volatile across calls.
1123 * FIXME: Optimize this.
1125 if ((ainfo->storage == ArgInIReg) || (ainfo->storage == ArgInFloatSSEReg) || (ainfo->storage == ArgInDoubleSSEReg) || (ainfo->storage == ArgValuetypeInReg))
1128 inst->opcode = OP_REGOFFSET;
1130 switch (ainfo->storage) {
1132 case ArgInFloatSSEReg:
1133 case ArgInDoubleSSEReg:
1134 inst->opcode = OP_REGVAR;
1135 inst->dreg = ainfo->reg;
1138 g_assert (!cfg->arch.omit_fp);
1139 inst->opcode = OP_REGOFFSET;
1140 inst->inst_basereg = cfg->frame_reg;
1141 inst->inst_offset = ainfo->offset + ARGS_OFFSET;
1143 case ArgValuetypeInReg:
1149 if (!inreg && (ainfo->storage != ArgOnStack)) {
1150 inst->opcode = OP_REGOFFSET;
1151 inst->inst_basereg = cfg->frame_reg;
1152 /* These arguments are saved to the stack in the prolog */
1153 if (cfg->arch.omit_fp) {
1154 inst->inst_offset = offset;
1155 offset += (ainfo->storage == ArgValuetypeInReg) ? 2 * sizeof (gpointer) : sizeof (gpointer);
1157 offset += (ainfo->storage == ArgValuetypeInReg) ? 2 * sizeof (gpointer) : sizeof (gpointer);
1158 inst->inst_offset = - offset;
1164 cfg->stack_offset = offset;
1168 mono_arch_create_vars (MonoCompile *cfg)
1170 MonoMethodSignature *sig;
1173 sig = mono_method_signature (cfg->method);
1175 if (!cfg->arch.cinfo)
1176 cfg->arch.cinfo = get_call_info (cfg, cfg->mempool, sig, FALSE);
1177 cinfo = cfg->arch.cinfo;
1179 if (cinfo->ret.storage == ArgValuetypeInReg)
1180 cfg->ret_var_is_local = TRUE;
1184 add_outarg_reg (MonoCompile *cfg, MonoCallInst *call, MonoInst *arg, ArgStorage storage, int reg, MonoInst *tree)
1188 arg->opcode = OP_OUTARG_REG;
1189 arg->inst_left = tree;
1190 arg->inst_call = call;
1191 arg->backend.reg3 = reg;
1193 case ArgInFloatSSEReg:
1194 arg->opcode = OP_AMD64_OUTARG_XMMREG_R4;
1195 arg->inst_left = tree;
1196 arg->inst_call = call;
1197 arg->backend.reg3 = reg;
1199 case ArgInDoubleSSEReg:
1200 arg->opcode = OP_AMD64_OUTARG_XMMREG_R8;
1201 arg->inst_left = tree;
1202 arg->inst_call = call;
1203 arg->backend.reg3 = reg;
1206 g_assert_not_reached ();
1210 /* Fixme: we need an alignment solution for enter_method and mono_arch_call_opcode,
1211 * currently alignment in mono_arch_call_opcode is computed without arch_get_argument_info
1215 arg_storage_to_ldind (ArgStorage storage)
1220 case ArgInDoubleSSEReg:
1221 return CEE_LDIND_R8;
1222 case ArgInFloatSSEReg:
1223 return CEE_LDIND_R4;
1225 g_assert_not_reached ();
1232 emit_sig_cookie (MonoCompile *cfg, MonoCallInst *call, CallInfo *cinfo)
1235 MonoMethodSignature *tmp_sig;
1238 /* FIXME: Add support for signature tokens to AOT */
1239 cfg->disable_aot = TRUE;
1241 g_assert (cinfo->sig_cookie.storage == ArgOnStack);
1244 * mono_ArgIterator_Setup assumes the signature cookie is
1245 * passed first and all the arguments which were before it are
1246 * passed on the stack after the signature. So compensate by
1247 * passing a different signature.
1249 tmp_sig = mono_metadata_signature_dup (call->signature);
1250 tmp_sig->param_count -= call->signature->sentinelpos;
1251 tmp_sig->sentinelpos = 0;
1252 memcpy (tmp_sig->params, call->signature->params + call->signature->sentinelpos, tmp_sig->param_count * sizeof (MonoType*));
1254 MONO_INST_NEW (cfg, sig_arg, OP_ICONST);
1255 sig_arg->inst_p0 = tmp_sig;
1257 MONO_INST_NEW (cfg, arg, OP_OUTARG);
1258 arg->inst_left = sig_arg;
1259 arg->type = STACK_PTR;
1261 /* prepend, so they get reversed */
1262 arg->next = call->out_args;
1263 call->out_args = arg;
1267 * take the arguments and generate the arch-specific
1268 * instructions to properly call the function in call.
1269 * This includes pushing, moving arguments to the right register
1271 * Issue: who does the spilling if needed, and when?
1274 mono_arch_call_opcode (MonoCompile *cfg, MonoBasicBlock* bb, MonoCallInst *call, int is_virtual) {
1276 MonoMethodSignature *sig;
1277 int i, n, stack_size;
1283 sig = call->signature;
1284 n = sig->param_count + sig->hasthis;
1286 cinfo = get_call_info (cfg, cfg->mempool, sig, sig->pinvoke);
1288 for (i = 0; i < n; ++i) {
1289 ainfo = cinfo->args + i;
1291 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG) && (i == sig->sentinelpos)) {
1292 /* Emit the signature cookie just before the implicit arguments */
1293 emit_sig_cookie (cfg, call, cinfo);
1296 if (is_virtual && i == 0) {
1297 /* the argument will be attached to the call instruction */
1298 in = call->args [i];
1300 MONO_INST_NEW (cfg, arg, OP_OUTARG);
1301 in = call->args [i];
1302 arg->cil_code = in->cil_code;
1303 arg->inst_left = in;
1304 arg->type = in->type;
1305 /* prepend, so they get reversed */
1306 arg->next = call->out_args;
1307 call->out_args = arg;
1309 if ((i >= sig->hasthis) && (MONO_TYPE_ISSTRUCT(sig->params [i - sig->hasthis]))) {
1313 if (sig->params [i - sig->hasthis]->type == MONO_TYPE_TYPEDBYREF) {
1314 size = sizeof (MonoTypedRef);
1315 align = sizeof (gpointer);
1319 size = mono_type_native_stack_size (&in->klass->byval_arg, &align);
1322 * Other backends use mini_type_stack_size (), but that
1323 * aligns the size to 8, which is larger than the size of
1324 * the source, leading to reads of invalid memory if the
1325 * source is at the end of address space.
1327 size = mono_class_value_size (in->klass, &align);
1329 if (ainfo->storage == ArgValuetypeInReg) {
1330 if (ainfo->pair_storage [1] == ArgNone) {
1335 MONO_INST_NEW (cfg, load, arg_storage_to_ldind (ainfo->pair_storage [0]));
1336 load->inst_left = in;
1338 add_outarg_reg (cfg, call, arg, ainfo->pair_storage [0], ainfo->pair_regs [0], load);
1341 /* Trees can't be shared so make a copy */
1342 MonoInst *vtaddr = mono_compile_create_var (cfg, &mono_defaults.int_class->byval_arg, OP_LOCAL);
1343 MonoInst *load, *load2, *offset_ins;
1346 MONO_INST_NEW (cfg, load, CEE_LDIND_I);
1347 load->ssa_op = MONO_SSA_LOAD;
1348 load->inst_i0 = (cfg)->varinfo [vtaddr->inst_c0];
1350 NEW_ICONST (cfg, offset_ins, 0);
1351 MONO_INST_NEW (cfg, load2, CEE_ADD);
1352 load2->inst_left = load;
1353 load2->inst_right = offset_ins;
1355 MONO_INST_NEW (cfg, load, arg_storage_to_ldind (ainfo->pair_storage [0]));
1356 load->inst_left = load2;
1358 add_outarg_reg (cfg, call, arg, ainfo->pair_storage [0], ainfo->pair_regs [0], load);
1361 MONO_INST_NEW (cfg, load, CEE_LDIND_I);
1362 load->ssa_op = MONO_SSA_LOAD;
1363 load->inst_i0 = (cfg)->varinfo [vtaddr->inst_c0];
1365 NEW_ICONST (cfg, offset_ins, 8);
1366 MONO_INST_NEW (cfg, load2, CEE_ADD);
1367 load2->inst_left = load;
1368 load2->inst_right = offset_ins;
1370 MONO_INST_NEW (cfg, load, arg_storage_to_ldind (ainfo->pair_storage [1]));
1371 load->inst_left = load2;
1373 MONO_INST_NEW (cfg, arg, OP_OUTARG);
1374 arg->cil_code = in->cil_code;
1375 arg->type = in->type;
1376 /* prepend, so they get reversed */
1377 arg->next = call->out_args;
1378 call->out_args = arg;
1380 add_outarg_reg (cfg, call, arg, ainfo->pair_storage [1], ainfo->pair_regs [1], load);
1382 /* Prepend a copy inst */
1383 MONO_INST_NEW (cfg, arg, CEE_STIND_I);
1384 arg->cil_code = in->cil_code;
1385 arg->ssa_op = MONO_SSA_STORE;
1386 arg->inst_left = vtaddr;
1387 arg->inst_right = in;
1388 arg->type = in->type;
1390 /* prepend, so they get reversed */
1391 arg->next = call->out_args;
1392 call->out_args = arg;
1396 arg->opcode = OP_OUTARG_VT;
1397 arg->klass = in->klass;
1398 arg->backend.is_pinvoke = sig->pinvoke;
1399 arg->inst_imm = size;
1403 switch (ainfo->storage) {
1405 add_outarg_reg (cfg, call, arg, ainfo->storage, ainfo->reg, in);
1407 case ArgInFloatSSEReg:
1408 case ArgInDoubleSSEReg:
1409 add_outarg_reg (cfg, call, arg, ainfo->storage, ainfo->reg, in);
1412 arg->opcode = OP_OUTARG;
1413 if (!sig->params [i - sig->hasthis]->byref) {
1414 if (sig->params [i - sig->hasthis]->type == MONO_TYPE_R4)
1415 arg->opcode = OP_OUTARG_R4;
1417 if (sig->params [i - sig->hasthis]->type == MONO_TYPE_R8)
1418 arg->opcode = OP_OUTARG_R8;
1422 g_assert_not_reached ();
1428 /* Handle the case where there are no implicit arguments */
1429 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG) && (n == sig->sentinelpos)) {
1430 emit_sig_cookie (cfg, call, cinfo);
1433 if (cinfo->need_stack_align) {
1434 MONO_INST_NEW (cfg, arg, OP_AMD64_OUTARG_ALIGN_STACK);
1435 /* prepend, so they get reversed */
1436 arg->next = call->out_args;
1437 call->out_args = arg;
1440 call->stack_usage = cinfo->stack_usage;
1441 cfg->param_area = MAX (cfg->param_area, call->stack_usage);
1442 cfg->flags |= MONO_CFG_HAS_CALLS;
1447 #define EMIT_COND_BRANCH(ins,cond,sign) \
1448 if (ins->flags & MONO_INST_BRLABEL) { \
1449 if (ins->inst_i0->inst_c0) { \
1450 x86_branch (code, cond, cfg->native_code + ins->inst_i0->inst_c0, sign); \
1452 mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_LABEL, ins->inst_i0); \
1453 if ((cfg->opt & MONO_OPT_BRANCH) && \
1454 x86_is_imm8 (ins->inst_i0->inst_c1 - cpos)) \
1455 x86_branch8 (code, cond, 0, sign); \
1457 x86_branch32 (code, cond, 0, sign); \
1460 if (ins->inst_true_bb->native_offset) { \
1461 x86_branch (code, cond, cfg->native_code + ins->inst_true_bb->native_offset, sign); \
1463 mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_BB, ins->inst_true_bb); \
1464 if ((cfg->opt & MONO_OPT_BRANCH) && \
1465 x86_is_imm8 (ins->inst_true_bb->max_offset - cpos)) \
1466 x86_branch8 (code, cond, 0, sign); \
1468 x86_branch32 (code, cond, 0, sign); \
1472 /* emit an exception if condition is fail */
1473 #define EMIT_COND_SYSTEM_EXCEPTION(cond,signed,exc_name) \
1475 MonoInst *tins = mono_branch_optimize_exception_target (cfg, bb, exc_name); \
1476 if (tins == NULL) { \
1477 mono_add_patch_info (cfg, code - cfg->native_code, \
1478 MONO_PATCH_INFO_EXC, exc_name); \
1479 x86_branch32 (code, cond, 0, signed); \
1481 EMIT_COND_BRANCH (tins, cond, signed); \
1485 #define EMIT_FPCOMPARE(code) do { \
1486 amd64_fcompp (code); \
1487 amd64_fnstsw (code); \
1490 #define EMIT_SSE2_FPFUNC(code, op, dreg, sreg1) do { \
1491 amd64_movsd_membase_reg (code, AMD64_RSP, -8, (sreg1)); \
1492 amd64_fld_membase (code, AMD64_RSP, -8, TRUE); \
1493 amd64_ ##op (code); \
1494 amd64_fst_membase (code, AMD64_RSP, -8, TRUE, TRUE); \
1495 amd64_movsd_reg_membase (code, (dreg), AMD64_RSP, -8); \
1499 emit_call_body (MonoCompile *cfg, guint8 *code, guint32 patch_type, gconstpointer data)
1501 mono_add_patch_info (cfg, code - cfg->native_code, patch_type, data);
1504 * FIXME: Add support for thunks
1507 gboolean near_call = FALSE;
1510 * Indirect calls are expensive so try to make a near call if possible.
1511 * The caller memory is allocated by the code manager so it is
1512 * guaranteed to be at a 32 bit offset.
1515 if (patch_type != MONO_PATCH_INFO_ABS) {
1516 /* The target is in memory allocated using the code manager */
1519 if ((patch_type == MONO_PATCH_INFO_METHOD) || (patch_type == MONO_PATCH_INFO_METHOD_JUMP)) {
1520 if (((MonoMethod*)data)->klass->image->assembly->aot_module)
1521 /* The callee might be an AOT method */
1525 if (patch_type == MONO_PATCH_INFO_INTERNAL_METHOD) {
1527 * The call might go directly to a native function without
1530 MonoJitICallInfo *mi = mono_find_jit_icall_by_name (data);
1532 gconstpointer target = mono_icall_get_wrapper (mi);
1533 if ((((guint64)target) >> 32) != 0)
1539 if (mono_find_class_init_trampoline_by_addr (data))
1542 MonoJitICallInfo *info = mono_find_jit_icall_by_addr (data);
1544 if ((cfg->method->wrapper_type == MONO_WRAPPER_MANAGED_TO_NATIVE) &&
1545 strstr (cfg->method->name, info->name)) {
1546 /* A call to the wrapped function */
1547 if ((((guint64)data) >> 32) == 0)
1550 else if (info->func == info->wrapper) {
1552 if ((((guint64)info->func) >> 32) == 0)
1556 /* See the comment in mono_codegen () */
1557 if ((info->name [0] != 'v') || (strstr (info->name, "ves_array_new_va_") == NULL && strstr (info->name, "ves_array_element_address_") == NULL))
1561 else if ((((guint64)data) >> 32) == 0)
1566 if (cfg->method->dynamic)
1567 /* These methods are allocated using malloc */
1570 if (cfg->compile_aot)
1573 #ifdef MONO_ARCH_NOMAP32BIT
1578 amd64_call_code (code, 0);
1581 amd64_set_reg_template (code, GP_SCRATCH_REG);
1582 amd64_call_reg (code, GP_SCRATCH_REG);
1589 static inline guint8*
1590 emit_call (MonoCompile *cfg, guint8 *code, guint32 patch_type, gconstpointer data)
1592 mono_add_patch_info (cfg, code - cfg->native_code, patch_type, data);
1594 return emit_call_body (cfg, code, patch_type, data);
1598 store_membase_imm_to_store_membase_reg (int opcode)
1601 case OP_STORE_MEMBASE_IMM:
1602 return OP_STORE_MEMBASE_REG;
1603 case OP_STOREI4_MEMBASE_IMM:
1604 return OP_STOREI4_MEMBASE_REG;
1605 case OP_STOREI8_MEMBASE_IMM:
1606 return OP_STOREI8_MEMBASE_REG;
1612 #define INST_IGNORES_CFLAGS(opcode) (!(((opcode) == OP_ADC) || ((opcode) == OP_ADC_IMM) || ((opcode) == OP_IADC) || ((opcode) == OP_IADC_IMM) || ((opcode) == OP_SBB) || ((opcode) == OP_SBB_IMM) || ((opcode) == OP_ISBB) || ((opcode) == OP_ISBB_IMM)))
1617 * Perform peephole opts which should/can be performed before local regalloc
1620 peephole_pass_1 (MonoCompile *cfg, MonoBasicBlock *bb)
1622 MonoInst *ins, *last_ins = NULL;
1627 switch (ins->opcode) {
1631 if ((ins->sreg1 < MONO_MAX_IREGS) && (ins->dreg >= MONO_MAX_IREGS) && (ins->inst_imm > 0)) {
1633 * X86_LEA is like ADD, but doesn't have the
1634 * sreg1==dreg restriction. inst_imm > 0 is needed since LEA sign-extends
1635 * its operand to 64 bit.
1637 ins->opcode = OP_X86_LEA_MEMBASE;
1638 ins->inst_basereg = ins->sreg1;
1644 if ((ins->sreg1 == ins->sreg2) && (ins->sreg1 == ins->dreg)) {
1648 * Replace STORE_MEMBASE_IMM 0 with STORE_MEMBASE_REG since
1649 * the latter has length 2-3 instead of 6 (reverse constant
1650 * propagation). These instruction sequences are very common
1651 * in the initlocals bblock.
1653 for (ins2 = ins->next; ins2; ins2 = ins2->next) {
1654 if (((ins2->opcode == OP_STORE_MEMBASE_IMM) || (ins2->opcode == OP_STOREI4_MEMBASE_IMM) || (ins2->opcode == OP_STOREI8_MEMBASE_IMM) || (ins2->opcode == OP_STORE_MEMBASE_IMM)) && (ins2->inst_imm == 0)) {
1655 ins2->opcode = store_membase_imm_to_store_membase_reg (ins2->opcode);
1656 ins2->sreg1 = ins->dreg;
1657 } else if ((ins2->opcode == OP_STOREI1_MEMBASE_IMM) || (ins2->opcode == OP_STOREI2_MEMBASE_IMM) || (ins2->opcode == OP_STOREI8_MEMBASE_REG) || (ins2->opcode == OP_STORE_MEMBASE_REG)) {
1659 } else if (((ins2->opcode == OP_ICONST) || (ins2->opcode == OP_I8CONST)) && (ins2->dreg == ins->dreg) && (ins2->inst_c0 == 0)) {
1668 case OP_COMPARE_IMM:
1669 /* OP_COMPARE_IMM (reg, 0)
1671 * OP_AMD64_TEST_NULL (reg)
1674 ins->opcode = OP_AMD64_TEST_NULL;
1676 case OP_ICOMPARE_IMM:
1678 ins->opcode = OP_X86_TEST_NULL;
1680 case OP_AMD64_ICOMPARE_MEMBASE_IMM:
1682 * OP_STORE_MEMBASE_REG reg, offset(basereg)
1683 * OP_X86_COMPARE_MEMBASE_IMM offset(basereg), imm
1685 * OP_STORE_MEMBASE_REG reg, offset(basereg)
1686 * OP_COMPARE_IMM reg, imm
1688 * Note: if imm = 0 then OP_COMPARE_IMM replaced with OP_X86_TEST_NULL
1690 if (last_ins && (last_ins->opcode == OP_STOREI4_MEMBASE_REG) &&
1691 ins->inst_basereg == last_ins->inst_destbasereg &&
1692 ins->inst_offset == last_ins->inst_offset) {
1693 ins->opcode = OP_ICOMPARE_IMM;
1694 ins->sreg1 = last_ins->sreg1;
1696 /* check if we can remove cmp reg,0 with test null */
1698 ins->opcode = OP_X86_TEST_NULL;
1702 case OP_LOAD_MEMBASE:
1703 case OP_LOADI4_MEMBASE:
1705 * Note: if reg1 = reg2 the load op is removed
1707 * OP_STORE_MEMBASE_REG reg1, offset(basereg)
1708 * OP_LOAD_MEMBASE offset(basereg), reg2
1710 * OP_STORE_MEMBASE_REG reg1, offset(basereg)
1711 * OP_MOVE reg1, reg2
1713 if (last_ins && (last_ins->opcode == OP_STOREI4_MEMBASE_REG
1714 || last_ins->opcode == OP_STORE_MEMBASE_REG) &&
1715 ins->inst_basereg == last_ins->inst_destbasereg &&
1716 ins->inst_offset == last_ins->inst_offset) {
1717 if (ins->dreg == last_ins->sreg1) {
1718 last_ins->next = ins->next;
1722 //static int c = 0; printf ("MATCHX %s %d\n", cfg->method->name,c++);
1723 ins->opcode = OP_MOVE;
1724 ins->sreg1 = last_ins->sreg1;
1728 * Note: reg1 must be different from the basereg in the second load
1729 * Note: if reg1 = reg2 is equal then second load is removed
1731 * OP_LOAD_MEMBASE offset(basereg), reg1
1732 * OP_LOAD_MEMBASE offset(basereg), reg2
1734 * OP_LOAD_MEMBASE offset(basereg), reg1
1735 * OP_MOVE reg1, reg2
1737 } if (last_ins && (last_ins->opcode == OP_LOADI4_MEMBASE
1738 || last_ins->opcode == OP_LOAD_MEMBASE) &&
1739 ins->inst_basereg != last_ins->dreg &&
1740 ins->inst_basereg == last_ins->inst_basereg &&
1741 ins->inst_offset == last_ins->inst_offset) {
1743 if (ins->dreg == last_ins->dreg) {
1744 last_ins->next = ins->next;
1748 ins->opcode = OP_MOVE;
1749 ins->sreg1 = last_ins->dreg;
1752 //g_assert_not_reached ();
1756 * OP_STORE_MEMBASE_IMM imm, offset(basereg)
1757 * OP_LOAD_MEMBASE offset(basereg), reg
1759 * OP_STORE_MEMBASE_IMM imm, offset(basereg)
1760 * OP_ICONST reg, imm
1762 } else if (last_ins && (last_ins->opcode == OP_STOREI4_MEMBASE_IMM
1763 || last_ins->opcode == OP_STORE_MEMBASE_IMM) &&
1764 ins->inst_basereg == last_ins->inst_destbasereg &&
1765 ins->inst_offset == last_ins->inst_offset) {
1766 //static int c = 0; printf ("MATCHX %s %d\n", cfg->method->name,c++);
1767 ins->opcode = OP_ICONST;
1768 ins->inst_c0 = last_ins->inst_imm;
1769 g_assert_not_reached (); // check this rule
1773 case OP_LOADI1_MEMBASE:
1775 * Note: if reg1 = reg2 the load op is removed
1777 * OP_STORE_MEMBASE_REG reg1, offset(basereg)
1778 * OP_LOAD_MEMBASE offset(basereg), reg2
1780 * OP_STORE_MEMBASE_REG reg1, offset(basereg)
1781 * OP_MOVE reg1, reg2
1783 if (last_ins && (last_ins->opcode == OP_STOREI1_MEMBASE_REG) &&
1784 ins->inst_basereg == last_ins->inst_destbasereg &&
1785 ins->inst_offset == last_ins->inst_offset) {
1786 if (ins->dreg == last_ins->sreg1) {
1787 last_ins->next = ins->next;
1791 //static int c = 0; printf ("MATCHX %s %d\n", cfg->method->name,c++);
1792 ins->opcode = OP_MOVE;
1793 ins->sreg1 = last_ins->sreg1;
1797 case OP_LOADI2_MEMBASE:
1799 * Note: if reg1 = reg2 the load op is removed
1801 * OP_STORE_MEMBASE_REG reg1, offset(basereg)
1802 * OP_LOAD_MEMBASE offset(basereg), reg2
1804 * OP_STORE_MEMBASE_REG reg1, offset(basereg)
1805 * OP_MOVE reg1, reg2
1807 if (last_ins && (last_ins->opcode == OP_STOREI2_MEMBASE_REG) &&
1808 ins->inst_basereg == last_ins->inst_destbasereg &&
1809 ins->inst_offset == last_ins->inst_offset) {
1810 if (ins->dreg == last_ins->sreg1) {
1811 last_ins->next = ins->next;
1815 //static int c = 0; printf ("MATCHX %s %d\n", cfg->method->name,c++);
1816 ins->opcode = OP_MOVE;
1817 ins->sreg1 = last_ins->sreg1;
1830 if (ins->dreg == ins->sreg1) {
1832 last_ins->next = ins->next;
1834 bb->code = ins->next;
1841 * OP_MOVE sreg, dreg
1842 * OP_MOVE dreg, sreg
1844 if (last_ins && last_ins->opcode == OP_MOVE &&
1845 ins->sreg1 == last_ins->dreg &&
1846 ins->dreg == last_ins->sreg1) {
1847 last_ins->next = ins->next;
1856 bb->last_ins = last_ins;
1860 peephole_pass (MonoCompile *cfg, MonoBasicBlock *bb)
1862 MonoInst *ins, *last_ins = NULL;
1867 switch (ins->opcode) {
1870 /* reg = 0 -> XOR (reg, reg) */
1871 /* XOR sets cflags on x86, so we cant do it always */
1872 if (ins->inst_c0 == 0 && (!ins->next || (ins->next && INST_IGNORES_CFLAGS (ins->next->opcode)))) {
1873 ins->opcode = OP_LXOR;
1874 ins->sreg1 = ins->dreg;
1875 ins->sreg2 = ins->dreg;
1882 if ((ins->sreg1 == ins->sreg2) && (ins->sreg1 == ins->dreg)) {
1886 * Replace STORE_MEMBASE_IMM 0 with STORE_MEMBASE_REG since
1887 * the latter has length 2-3 instead of 6 (reverse constant
1888 * propagation). These instruction sequences are very common
1889 * in the initlocals bblock.
1891 for (ins2 = ins->next; ins2; ins2 = ins2->next) {
1892 if (((ins2->opcode == OP_STORE_MEMBASE_IMM) || (ins2->opcode == OP_STOREI4_MEMBASE_IMM) || (ins2->opcode == OP_STOREI8_MEMBASE_IMM) || (ins2->opcode == OP_STORE_MEMBASE_IMM)) && (ins2->inst_imm == 0)) {
1893 ins2->opcode = store_membase_imm_to_store_membase_reg (ins2->opcode);
1894 ins2->sreg1 = ins->dreg;
1895 } else if ((ins2->opcode == OP_STOREI1_MEMBASE_IMM) || (ins2->opcode == OP_STOREI2_MEMBASE_IMM) || (ins2->opcode == OP_STOREI8_MEMBASE_REG) || (ins2->opcode == OP_STORE_MEMBASE_REG)) {
1897 } else if (((ins2->opcode == OP_ICONST) || (ins2->opcode == OP_I8CONST)) && (ins2->dreg == ins->dreg) && (ins2->inst_c0 == 0)) {
1907 if ((ins->inst_imm == 1) && (ins->dreg == ins->sreg1))
1908 ins->opcode = OP_X86_INC_REG;
1911 if ((ins->inst_imm == 1) && (ins->dreg == ins->sreg1))
1912 ins->opcode = OP_X86_DEC_REG;
1915 /* remove unnecessary multiplication with 1 */
1916 if (ins->inst_imm == 1) {
1917 if (ins->dreg != ins->sreg1) {
1918 ins->opcode = OP_MOVE;
1920 last_ins->next = ins->next;
1926 case OP_COMPARE_IMM:
1927 /* OP_COMPARE_IMM (reg, 0)
1929 * OP_AMD64_TEST_NULL (reg)
1932 ins->opcode = OP_AMD64_TEST_NULL;
1934 case OP_ICOMPARE_IMM:
1936 ins->opcode = OP_X86_TEST_NULL;
1938 case OP_AMD64_ICOMPARE_MEMBASE_IMM:
1940 * OP_STORE_MEMBASE_REG reg, offset(basereg)
1941 * OP_X86_COMPARE_MEMBASE_IMM offset(basereg), imm
1943 * OP_STORE_MEMBASE_REG reg, offset(basereg)
1944 * OP_COMPARE_IMM reg, imm
1946 * Note: if imm = 0 then OP_COMPARE_IMM replaced with OP_X86_TEST_NULL
1948 if (last_ins && (last_ins->opcode == OP_STOREI4_MEMBASE_REG) &&
1949 ins->inst_basereg == last_ins->inst_destbasereg &&
1950 ins->inst_offset == last_ins->inst_offset) {
1951 ins->opcode = OP_ICOMPARE_IMM;
1952 ins->sreg1 = last_ins->sreg1;
1954 /* check if we can remove cmp reg,0 with test null */
1956 ins->opcode = OP_X86_TEST_NULL;
1960 case OP_LOAD_MEMBASE:
1961 case OP_LOADI4_MEMBASE:
1963 * Note: if reg1 = reg2 the load op is removed
1965 * OP_STORE_MEMBASE_REG reg1, offset(basereg)
1966 * OP_LOAD_MEMBASE offset(basereg), reg2
1968 * OP_STORE_MEMBASE_REG reg1, offset(basereg)
1969 * OP_MOVE reg1, reg2
1971 if (last_ins && (last_ins->opcode == OP_STOREI4_MEMBASE_REG
1972 || last_ins->opcode == OP_STORE_MEMBASE_REG) &&
1973 ins->inst_basereg == last_ins->inst_destbasereg &&
1974 ins->inst_offset == last_ins->inst_offset) {
1975 if (ins->dreg == last_ins->sreg1) {
1976 last_ins->next = ins->next;
1980 //static int c = 0; printf ("MATCHX %s %d\n", cfg->method->name,c++);
1981 ins->opcode = OP_MOVE;
1982 ins->sreg1 = last_ins->sreg1;
1986 * Note: reg1 must be different from the basereg in the second load
1987 * Note: if reg1 = reg2 is equal then second load is removed
1989 * OP_LOAD_MEMBASE offset(basereg), reg1
1990 * OP_LOAD_MEMBASE offset(basereg), reg2
1992 * OP_LOAD_MEMBASE offset(basereg), reg1
1993 * OP_MOVE reg1, reg2
1995 } if (last_ins && (last_ins->opcode == OP_LOADI4_MEMBASE
1996 || last_ins->opcode == OP_LOAD_MEMBASE) &&
1997 ins->inst_basereg != last_ins->dreg &&
1998 ins->inst_basereg == last_ins->inst_basereg &&
1999 ins->inst_offset == last_ins->inst_offset) {
2001 if (ins->dreg == last_ins->dreg) {
2002 last_ins->next = ins->next;
2006 ins->opcode = OP_MOVE;
2007 ins->sreg1 = last_ins->dreg;
2010 //g_assert_not_reached ();
2014 * OP_STORE_MEMBASE_IMM imm, offset(basereg)
2015 * OP_LOAD_MEMBASE offset(basereg), reg
2017 * OP_STORE_MEMBASE_IMM imm, offset(basereg)
2018 * OP_ICONST reg, imm
2020 } else if (last_ins && (last_ins->opcode == OP_STOREI4_MEMBASE_IMM
2021 || last_ins->opcode == OP_STORE_MEMBASE_IMM) &&
2022 ins->inst_basereg == last_ins->inst_destbasereg &&
2023 ins->inst_offset == last_ins->inst_offset) {
2024 //static int c = 0; printf ("MATCHX %s %d\n", cfg->method->name,c++);
2025 ins->opcode = OP_ICONST;
2026 ins->inst_c0 = last_ins->inst_imm;
2027 g_assert_not_reached (); // check this rule
2031 case OP_LOADI1_MEMBASE:
2033 * Note: if reg1 = reg2 the load op is removed
2035 * OP_STORE_MEMBASE_REG reg1, offset(basereg)
2036 * OP_LOAD_MEMBASE offset(basereg), reg2
2038 * OP_STORE_MEMBASE_REG reg1, offset(basereg)
2039 * OP_MOVE reg1, reg2
2041 if (last_ins && (last_ins->opcode == OP_STOREI1_MEMBASE_REG) &&
2042 ins->inst_basereg == last_ins->inst_destbasereg &&
2043 ins->inst_offset == last_ins->inst_offset) {
2044 if (ins->dreg == last_ins->sreg1) {
2045 last_ins->next = ins->next;
2049 //static int c = 0; printf ("MATCHX %s %d\n", cfg->method->name,c++);
2050 ins->opcode = OP_MOVE;
2051 ins->sreg1 = last_ins->sreg1;
2055 case OP_LOADI2_MEMBASE:
2057 * Note: if reg1 = reg2 the load op is removed
2059 * OP_STORE_MEMBASE_REG reg1, offset(basereg)
2060 * OP_LOAD_MEMBASE offset(basereg), reg2
2062 * OP_STORE_MEMBASE_REG reg1, offset(basereg)
2063 * OP_MOVE reg1, reg2
2065 if (last_ins && (last_ins->opcode == OP_STOREI2_MEMBASE_REG) &&
2066 ins->inst_basereg == last_ins->inst_destbasereg &&
2067 ins->inst_offset == last_ins->inst_offset) {
2068 if (ins->dreg == last_ins->sreg1) {
2069 last_ins->next = ins->next;
2073 //static int c = 0; printf ("MATCHX %s %d\n", cfg->method->name,c++);
2074 ins->opcode = OP_MOVE;
2075 ins->sreg1 = last_ins->sreg1;
2088 if (ins->dreg == ins->sreg1) {
2090 last_ins->next = ins->next;
2092 bb->code = ins->next;
2099 * OP_MOVE sreg, dreg
2100 * OP_MOVE dreg, sreg
2102 if (last_ins && last_ins->opcode == OP_MOVE &&
2103 ins->sreg1 == last_ins->dreg &&
2104 ins->dreg == last_ins->sreg1) {
2105 last_ins->next = ins->next;
2114 bb->last_ins = last_ins;
2118 insert_after_ins (MonoBasicBlock *bb, MonoInst *ins, MonoInst *to_insert)
2122 bb->code = to_insert;
2123 to_insert->next = ins;
2126 to_insert->next = ins->next;
2127 ins->next = to_insert;
2131 #define NEW_INS(cfg,dest,op) do { \
2132 (dest) = mono_mempool_alloc0 ((cfg)->mempool, sizeof (MonoInst)); \
2133 (dest)->opcode = (op); \
2134 insert_after_ins (bb, last_ins, (dest)); \
2138 * mono_arch_lowering_pass:
2140 * Converts complex opcodes into simpler ones so that each IR instruction
2141 * corresponds to one machine instruction.
2144 mono_arch_lowering_pass (MonoCompile *cfg, MonoBasicBlock *bb)
2146 MonoInst *ins, *temp, *last_ins = NULL;
2149 if (bb->max_vreg > cfg->rs->next_vreg)
2150 cfg->rs->next_vreg = bb->max_vreg;
2153 * FIXME: Need to add more instructions, but the current machine
2154 * description can't model some parts of the composite instructions like
2158 switch (ins->opcode) {
2163 NEW_INS (cfg, temp, OP_ICONST);
2164 temp->inst_c0 = ins->inst_imm;
2165 temp->dreg = mono_regstate_next_int (cfg->rs);
2166 switch (ins->opcode) {
2168 ins->opcode = OP_LDIV;
2171 ins->opcode = OP_LREM;
2174 ins->opcode = OP_IDIV;
2177 ins->opcode = OP_IREM;
2180 ins->sreg2 = temp->dreg;
2182 case OP_COMPARE_IMM:
2183 if (!amd64_is_imm32 (ins->inst_imm)) {
2184 NEW_INS (cfg, temp, OP_I8CONST);
2185 temp->inst_c0 = ins->inst_imm;
2186 temp->dreg = mono_regstate_next_int (cfg->rs);
2187 ins->opcode = OP_COMPARE;
2188 ins->sreg2 = temp->dreg;
2191 case OP_LOAD_MEMBASE:
2192 case OP_LOADI8_MEMBASE:
2193 if (!amd64_is_imm32 (ins->inst_offset)) {
2194 NEW_INS (cfg, temp, OP_I8CONST);
2195 temp->inst_c0 = ins->inst_offset;
2196 temp->dreg = mono_regstate_next_int (cfg->rs);
2197 ins->opcode = OP_AMD64_LOADI8_MEMINDEX;
2198 ins->inst_indexreg = temp->dreg;
2201 case OP_STORE_MEMBASE_IMM:
2202 case OP_STOREI8_MEMBASE_IMM:
2203 if (!amd64_is_imm32 (ins->inst_imm)) {
2204 NEW_INS (cfg, temp, OP_I8CONST);
2205 temp->inst_c0 = ins->inst_imm;
2206 temp->dreg = mono_regstate_next_int (cfg->rs);
2207 ins->opcode = OP_STOREI8_MEMBASE_REG;
2208 ins->sreg1 = temp->dreg;
2217 bb->last_ins = last_ins;
2219 bb->max_vreg = cfg->rs->next_vreg;
2223 branch_cc_table [] = {
2224 X86_CC_EQ, X86_CC_GE, X86_CC_GT, X86_CC_LE, X86_CC_LT,
2225 X86_CC_NE, X86_CC_GE, X86_CC_GT, X86_CC_LE, X86_CC_LT,
2226 X86_CC_O, X86_CC_NO, X86_CC_C, X86_CC_NC
2229 /* Maps CMP_... constants to X86_CC_... constants */
2232 X86_CC_EQ, X86_CC_NE, X86_CC_LE, X86_CC_GE, X86_CC_LT, X86_CC_GT,
2233 X86_CC_LE, X86_CC_GE, X86_CC_LT, X86_CC_GT
2237 cc_signed_table [] = {
2238 TRUE, TRUE, TRUE, TRUE, TRUE, TRUE,
2239 FALSE, FALSE, FALSE, FALSE
2242 /*#include "cprop.c"*/
2245 * Local register allocation.
2246 * We first scan the list of instructions and we save the liveness info of
2247 * each register (when the register is first used, when it's value is set etc.).
2248 * We also reverse the list of instructions (in the InstList list) because assigning
2249 * registers backwards allows for more tricks to be used.
2252 mono_arch_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
2257 mono_arch_lowering_pass (cfg, bb);
2259 if (cfg->opt & MONO_OPT_PEEPHOLE)
2260 peephole_pass_1 (cfg, bb);
2262 mono_local_regalloc (cfg, bb);
2265 static unsigned char*
2266 emit_float_to_int (MonoCompile *cfg, guchar *code, int dreg, int sreg, int size, gboolean is_signed)
2269 amd64_sse_cvttsd2si_reg_reg (code, dreg, sreg);
2272 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, 16);
2273 x86_fnstcw_membase(code, AMD64_RSP, 0);
2274 amd64_mov_reg_membase (code, dreg, AMD64_RSP, 0, 2);
2275 amd64_alu_reg_imm (code, X86_OR, dreg, 0xc00);
2276 amd64_mov_membase_reg (code, AMD64_RSP, 2, dreg, 2);
2277 amd64_fldcw_membase (code, AMD64_RSP, 2);
2278 amd64_push_reg (code, AMD64_RAX); // SP = SP - 8
2279 amd64_fist_pop_membase (code, AMD64_RSP, 0, size == 8);
2280 amd64_pop_reg (code, dreg);
2281 amd64_fldcw_membase (code, AMD64_RSP, 0);
2282 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, 16);
2286 amd64_widen_reg (code, dreg, dreg, is_signed, FALSE);
2288 amd64_widen_reg (code, dreg, dreg, is_signed, TRUE);
2292 static unsigned char*
2293 mono_emit_stack_alloc (guchar *code, MonoInst* tree)
2295 int sreg = tree->sreg1;
2296 int need_touch = FALSE;
2298 #if defined(PLATFORM_WIN32) || defined(MONO_ARCH_SIGSEGV_ON_ALTSTACK)
2299 if (!tree->flags & MONO_INST_INIT)
2308 * If requested stack size is larger than one page,
2309 * perform stack-touch operation
2312 * Generate stack probe code.
2313 * Under Windows, it is necessary to allocate one page at a time,
2314 * "touching" stack after each successful sub-allocation. This is
2315 * because of the way stack growth is implemented - there is a
2316 * guard page before the lowest stack page that is currently commited.
2317 * Stack normally grows sequentially so OS traps access to the
2318 * guard page and commits more pages when needed.
2320 amd64_test_reg_imm (code, sreg, ~0xFFF);
2321 br[0] = code; x86_branch8 (code, X86_CC_Z, 0, FALSE);
2323 br[2] = code; /* loop */
2324 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, 0x1000);
2325 amd64_test_membase_reg (code, AMD64_RSP, 0, AMD64_RSP);
2326 amd64_alu_reg_imm (code, X86_SUB, sreg, 0x1000);
2327 amd64_alu_reg_imm (code, X86_CMP, sreg, 0x1000);
2328 br[3] = code; x86_branch8 (code, X86_CC_AE, 0, FALSE);
2329 amd64_patch (br[3], br[2]);
2330 amd64_test_reg_reg (code, sreg, sreg);
2331 br[4] = code; x86_branch8 (code, X86_CC_Z, 0, FALSE);
2332 amd64_alu_reg_reg (code, X86_SUB, AMD64_RSP, sreg);
2334 br[1] = code; x86_jump8 (code, 0);
2336 amd64_patch (br[0], code);
2337 amd64_alu_reg_reg (code, X86_SUB, AMD64_RSP, sreg);
2338 amd64_patch (br[1], code);
2339 amd64_patch (br[4], code);
2342 amd64_alu_reg_reg (code, X86_SUB, AMD64_RSP, tree->sreg1);
2344 if (tree->flags & MONO_INST_INIT) {
2346 if (tree->dreg != AMD64_RAX && sreg != AMD64_RAX) {
2347 amd64_push_reg (code, AMD64_RAX);
2350 if (tree->dreg != AMD64_RCX && sreg != AMD64_RCX) {
2351 amd64_push_reg (code, AMD64_RCX);
2354 if (tree->dreg != AMD64_RDI && sreg != AMD64_RDI) {
2355 amd64_push_reg (code, AMD64_RDI);
2359 amd64_shift_reg_imm (code, X86_SHR, sreg, 3);
2360 if (sreg != AMD64_RCX)
2361 amd64_mov_reg_reg (code, AMD64_RCX, sreg, 8);
2362 amd64_alu_reg_reg (code, X86_XOR, AMD64_RAX, AMD64_RAX);
2364 amd64_lea_membase (code, AMD64_RDI, AMD64_RSP, offset);
2366 amd64_prefix (code, X86_REP_PREFIX);
2369 if (tree->dreg != AMD64_RDI && sreg != AMD64_RDI)
2370 amd64_pop_reg (code, AMD64_RDI);
2371 if (tree->dreg != AMD64_RCX && sreg != AMD64_RCX)
2372 amd64_pop_reg (code, AMD64_RCX);
2373 if (tree->dreg != AMD64_RAX && sreg != AMD64_RAX)
2374 amd64_pop_reg (code, AMD64_RAX);
2380 emit_move_return_value (MonoCompile *cfg, MonoInst *ins, guint8 *code)
2385 /* Move return value to the target register */
2386 /* FIXME: do this in the local reg allocator */
2387 switch (ins->opcode) {
2390 case OP_CALL_MEMBASE:
2393 case OP_LCALL_MEMBASE:
2394 g_assert (ins->dreg == AMD64_RAX);
2398 case OP_FCALL_MEMBASE:
2399 if (((MonoCallInst*)ins)->signature->ret->type == MONO_TYPE_R4) {
2401 amd64_sse_cvtss2sd_reg_reg (code, ins->dreg, AMD64_XMM0);
2403 /* FIXME: optimize this */
2404 amd64_movss_membase_reg (code, AMD64_RSP, -8, AMD64_XMM0);
2405 amd64_fld_membase (code, AMD64_RSP, -8, FALSE);
2410 if (ins->dreg != AMD64_XMM0)
2411 amd64_sse_movsd_reg_reg (code, ins->dreg, AMD64_XMM0);
2414 /* FIXME: optimize this */
2415 amd64_movsd_membase_reg (code, AMD64_RSP, -8, AMD64_XMM0);
2416 amd64_fld_membase (code, AMD64_RSP, -8, TRUE);
2422 case OP_VCALL_MEMBASE:
2423 cinfo = get_call_info (cfg, cfg->mempool, ((MonoCallInst*)ins)->signature, FALSE);
2424 if (cinfo->ret.storage == ArgValuetypeInReg) {
2425 /* Pop the destination address from the stack */
2426 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, 8);
2427 amd64_pop_reg (code, AMD64_RCX);
2429 for (quad = 0; quad < 2; quad ++) {
2430 switch (cinfo->ret.pair_storage [quad]) {
2432 amd64_mov_membase_reg (code, AMD64_RCX, (quad * 8), cinfo->ret.pair_regs [quad], 8);
2434 case ArgInFloatSSEReg:
2435 amd64_movss_membase_reg (code, AMD64_RCX, (quad * 8), cinfo->ret.pair_regs [quad]);
2437 case ArgInDoubleSSEReg:
2438 amd64_movsd_membase_reg (code, AMD64_RCX, (quad * 8), cinfo->ret.pair_regs [quad]);
2455 * @code: buffer to store code to
2456 * @dreg: hard register where to place the result
2457 * @tls_offset: offset info
2459 * emit_tls_get emits in @code the native code that puts in the dreg register
2460 * the item in the thread local storage identified by tls_offset.
2462 * Returns: a pointer to the end of the stored code
2465 emit_tls_get (guint8* code, int dreg, int tls_offset)
2467 if (optimize_for_xen) {
2468 x86_prefix (code, X86_FS_PREFIX);
2469 amd64_mov_reg_mem (code, dreg, 0, 8);
2470 amd64_mov_reg_membase (code, dreg, dreg, tls_offset, 8);
2472 x86_prefix (code, X86_FS_PREFIX);
2473 amd64_mov_reg_mem (code, dreg, tls_offset, 8);
2479 * emit_load_volatile_arguments:
2481 * Load volatile arguments from the stack to the original input registers.
2482 * Required before a tail call.
2485 emit_load_volatile_arguments (MonoCompile *cfg, guint8 *code)
2487 MonoMethod *method = cfg->method;
2488 MonoMethodSignature *sig;
2493 /* FIXME: Generate intermediate code instead */
2495 sig = mono_method_signature (method);
2497 cinfo = cfg->arch.cinfo;
2499 /* This is the opposite of the code in emit_prolog */
2501 if (sig->ret->type != MONO_TYPE_VOID) {
2502 if ((cinfo->ret.storage == ArgInIReg) && (cfg->ret->opcode != OP_REGVAR)) {
2503 amd64_mov_reg_membase (code, cinfo->ret.reg, cfg->ret->inst_basereg, cfg->ret->inst_offset, 8);
2507 for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
2508 ArgInfo *ainfo = cinfo->args + i;
2510 inst = cfg->args [i];
2512 if (sig->hasthis && (i == 0))
2513 arg_type = &mono_defaults.object_class->byval_arg;
2515 arg_type = sig->params [i - sig->hasthis];
2517 if (inst->opcode != OP_REGVAR) {
2518 switch (ainfo->storage) {
2523 amd64_mov_reg_membase (code, ainfo->reg, inst->inst_basereg, inst->inst_offset, size);
2526 case ArgInFloatSSEReg:
2527 amd64_movss_reg_membase (code, ainfo->reg, inst->inst_basereg, inst->inst_offset);
2529 case ArgInDoubleSSEReg:
2530 amd64_movsd_reg_membase (code, ainfo->reg, inst->inst_basereg, inst->inst_offset);
2537 g_assert (ainfo->storage == ArgInIReg);
2539 amd64_mov_reg_reg (code, ainfo->reg, inst->dreg, 8);
2546 #define REAL_PRINT_REG(text,reg) \
2547 mono_assert (reg >= 0); \
2548 amd64_push_reg (code, AMD64_RAX); \
2549 amd64_push_reg (code, AMD64_RDX); \
2550 amd64_push_reg (code, AMD64_RCX); \
2551 amd64_push_reg (code, reg); \
2552 amd64_push_imm (code, reg); \
2553 amd64_push_imm (code, text " %d %p\n"); \
2554 amd64_mov_reg_imm (code, AMD64_RAX, printf); \
2555 amd64_call_reg (code, AMD64_RAX); \
2556 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, 3*4); \
2557 amd64_pop_reg (code, AMD64_RCX); \
2558 amd64_pop_reg (code, AMD64_RDX); \
2559 amd64_pop_reg (code, AMD64_RAX);
2561 /* benchmark and set based on cpu */
2562 #define LOOP_ALIGNMENT 8
2563 #define bb_is_loop_start(bb) ((bb)->loop_body_start && (bb)->nesting)
2566 mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
2571 guint8 *code = cfg->native_code + cfg->code_len;
2572 MonoInst *last_ins = NULL;
2573 guint last_offset = 0;
2576 if (cfg->opt & MONO_OPT_PEEPHOLE)
2577 peephole_pass (cfg, bb);
2579 if (cfg->opt & MONO_OPT_LOOP) {
2580 int pad, align = LOOP_ALIGNMENT;
2581 /* set alignment depending on cpu */
2582 if (bb_is_loop_start (bb) && (pad = (cfg->code_len & (align - 1)))) {
2584 /*g_print ("adding %d pad at %x to loop in %s\n", pad, cfg->code_len, cfg->method->name);*/
2585 amd64_padding (code, pad);
2586 cfg->code_len += pad;
2587 bb->native_offset = cfg->code_len;
2591 if (cfg->verbose_level > 2)
2592 g_print ("Basic block %d starting at offset 0x%x\n", bb->block_num, bb->native_offset);
2594 cpos = bb->max_offset;
2596 if (cfg->prof_options & MONO_PROFILE_COVERAGE) {
2597 MonoProfileCoverageInfo *cov = cfg->coverage_info;
2598 g_assert (!cfg->compile_aot);
2601 cov->data [bb->dfn].cil_code = bb->cil_code;
2602 amd64_mov_reg_imm (code, AMD64_R11, (guint64)&cov->data [bb->dfn].count);
2603 /* this is not thread save, but good enough */
2604 amd64_inc_membase (code, AMD64_R11, 0);
2607 offset = code - cfg->native_code;
2609 mono_debug_open_block (cfg, bb, offset);
2613 offset = code - cfg->native_code;
2615 max_len = ((guint8 *)ins_get_spec (ins->opcode))[MONO_INST_LEN];
2617 if (G_UNLIKELY (offset > (cfg->code_size - max_len - 16))) {
2618 cfg->code_size *= 2;
2619 cfg->native_code = g_realloc (cfg->native_code, cfg->code_size);
2620 code = cfg->native_code + offset;
2621 mono_jit_stats.code_reallocs++;
2624 if (cfg->debug_info)
2625 mono_debug_record_line_number (cfg, ins, offset);
2627 switch (ins->opcode) {
2629 amd64_mul_reg (code, ins->sreg2, TRUE);
2632 amd64_mul_reg (code, ins->sreg2, FALSE);
2634 case OP_X86_SETEQ_MEMBASE:
2635 amd64_set_membase (code, X86_CC_EQ, ins->inst_basereg, ins->inst_offset, TRUE);
2637 case OP_STOREI1_MEMBASE_IMM:
2638 amd64_mov_membase_imm (code, ins->inst_destbasereg, ins->inst_offset, ins->inst_imm, 1);
2640 case OP_STOREI2_MEMBASE_IMM:
2641 amd64_mov_membase_imm (code, ins->inst_destbasereg, ins->inst_offset, ins->inst_imm, 2);
2643 case OP_STOREI4_MEMBASE_IMM:
2644 amd64_mov_membase_imm (code, ins->inst_destbasereg, ins->inst_offset, ins->inst_imm, 4);
2646 case OP_STOREI1_MEMBASE_REG:
2647 amd64_mov_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, ins->sreg1, 1);
2649 case OP_STOREI2_MEMBASE_REG:
2650 amd64_mov_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, ins->sreg1, 2);
2652 case OP_STORE_MEMBASE_REG:
2653 case OP_STOREI8_MEMBASE_REG:
2654 amd64_mov_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, ins->sreg1, 8);
2656 case OP_STOREI4_MEMBASE_REG:
2657 amd64_mov_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, ins->sreg1, 4);
2659 case OP_STORE_MEMBASE_IMM:
2660 case OP_STOREI8_MEMBASE_IMM:
2661 g_assert (amd64_is_imm32 (ins->inst_imm));
2662 amd64_mov_membase_imm (code, ins->inst_destbasereg, ins->inst_offset, ins->inst_imm, 8);
2665 amd64_mov_reg_mem (code, ins->dreg, (gssize)ins->inst_p0, sizeof (gpointer));
2668 amd64_mov_reg_mem (code, ins->dreg, (gssize)ins->inst_p0, 4);
2671 amd64_mov_reg_mem (code, ins->dreg, (gssize)ins->inst_p0, 4);
2674 amd64_mov_reg_imm (code, ins->dreg, ins->inst_p0);
2675 amd64_mov_reg_membase (code, ins->dreg, ins->dreg, 0, 4);
2677 case OP_LOAD_MEMBASE:
2678 case OP_LOADI8_MEMBASE:
2679 g_assert (amd64_is_imm32 (ins->inst_offset));
2680 amd64_mov_reg_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, sizeof (gpointer));
2682 case OP_LOADI4_MEMBASE:
2683 amd64_movsxd_reg_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset);
2685 case OP_LOADU4_MEMBASE:
2686 amd64_mov_reg_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, 4);
2688 case OP_LOADU1_MEMBASE:
2689 amd64_widen_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, FALSE, FALSE);
2691 case OP_LOADI1_MEMBASE:
2692 amd64_widen_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, TRUE, FALSE);
2694 case OP_LOADU2_MEMBASE:
2695 amd64_widen_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, FALSE, TRUE);
2697 case OP_LOADI2_MEMBASE:
2698 amd64_widen_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, TRUE, TRUE);
2700 case OP_AMD64_LOADI8_MEMINDEX:
2701 amd64_mov_reg_memindex_size (code, ins->dreg, ins->inst_basereg, 0, ins->inst_indexreg, 0, 8);
2705 amd64_widen_reg (code, ins->dreg, ins->sreg1, TRUE, FALSE);
2709 amd64_widen_reg (code, ins->dreg, ins->sreg1, TRUE, TRUE);
2712 amd64_widen_reg (code, ins->dreg, ins->sreg1, FALSE, FALSE);
2715 amd64_widen_reg (code, ins->dreg, ins->sreg1, FALSE, TRUE);
2719 /* Clean out the upper word */
2720 amd64_mov_reg_reg_size (code, ins->dreg, ins->sreg1, 4);
2725 amd64_movsxd_reg_reg (code, ins->dreg, ins->sreg1);
2729 amd64_alu_reg_reg (code, X86_CMP, ins->sreg1, ins->sreg2);
2731 case OP_COMPARE_IMM:
2732 g_assert (amd64_is_imm32 (ins->inst_imm));
2733 amd64_alu_reg_imm (code, X86_CMP, ins->sreg1, ins->inst_imm);
2735 case OP_X86_COMPARE_REG_MEMBASE:
2736 amd64_alu_reg_membase (code, X86_CMP, ins->sreg1, ins->sreg2, ins->inst_offset);
2738 case OP_X86_TEST_NULL:
2739 amd64_test_reg_reg_size (code, ins->sreg1, ins->sreg1, 4);
2741 case OP_AMD64_TEST_NULL:
2742 amd64_test_reg_reg (code, ins->sreg1, ins->sreg1);
2744 case OP_X86_ADD_MEMBASE_IMM:
2745 /* FIXME: Make a 64 version too */
2746 amd64_alu_membase_imm_size (code, X86_ADD, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 4);
2748 case OP_X86_ADD_MEMBASE:
2749 amd64_alu_reg_membase_size (code, X86_ADD, ins->sreg1, ins->sreg2, ins->inst_offset, 4);
2751 case OP_X86_SUB_MEMBASE_IMM:
2752 g_assert (amd64_is_imm32 (ins->inst_imm));
2753 amd64_alu_membase_imm_size (code, X86_SUB, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 4);
2755 case OP_X86_SUB_MEMBASE:
2756 amd64_alu_reg_membase_size (code, X86_SUB, ins->sreg1, ins->sreg2, ins->inst_offset, 4);
2758 case OP_X86_INC_MEMBASE:
2759 amd64_inc_membase_size (code, ins->inst_basereg, ins->inst_offset, 4);
2761 case OP_X86_INC_REG:
2762 amd64_inc_reg_size (code, ins->dreg, 4);
2764 case OP_X86_DEC_MEMBASE:
2765 amd64_dec_membase_size (code, ins->inst_basereg, ins->inst_offset, 4);
2767 case OP_X86_DEC_REG:
2768 amd64_dec_reg_size (code, ins->dreg, 4);
2770 case OP_X86_MUL_MEMBASE:
2771 amd64_imul_reg_membase_size (code, ins->sreg1, ins->sreg2, ins->inst_offset, 4);
2773 case OP_AMD64_ICOMPARE_MEMBASE_REG:
2774 amd64_alu_membase_reg_size (code, X86_CMP, ins->inst_basereg, ins->inst_offset, ins->sreg2, 4);
2776 case OP_AMD64_ICOMPARE_MEMBASE_IMM:
2777 amd64_alu_membase_imm_size (code, X86_CMP, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 4);
2779 case OP_AMD64_ICOMPARE_REG_MEMBASE:
2780 amd64_alu_reg_membase_size (code, X86_CMP, ins->sreg1, ins->sreg2, ins->inst_offset, 4);
2783 amd64_breakpoint (code);
2788 amd64_alu_reg_reg (code, X86_ADD, ins->sreg1, ins->sreg2);
2791 amd64_alu_reg_reg (code, X86_ADC, ins->sreg1, ins->sreg2);
2794 g_assert (amd64_is_imm32 (ins->inst_imm));
2795 amd64_alu_reg_imm (code, X86_ADD, ins->dreg, ins->inst_imm);
2798 g_assert (amd64_is_imm32 (ins->inst_imm));
2799 amd64_alu_reg_imm (code, X86_ADC, ins->dreg, ins->inst_imm);
2803 amd64_alu_reg_reg (code, X86_SUB, ins->sreg1, ins->sreg2);
2806 amd64_alu_reg_reg (code, X86_SBB, ins->sreg1, ins->sreg2);
2809 g_assert (amd64_is_imm32 (ins->inst_imm));
2810 amd64_alu_reg_imm (code, X86_SUB, ins->dreg, ins->inst_imm);
2813 g_assert (amd64_is_imm32 (ins->inst_imm));
2814 amd64_alu_reg_imm (code, X86_SBB, ins->dreg, ins->inst_imm);
2817 amd64_alu_reg_reg (code, X86_AND, ins->sreg1, ins->sreg2);
2820 g_assert (amd64_is_imm32 (ins->inst_imm));
2821 amd64_alu_reg_imm (code, X86_AND, ins->sreg1, ins->inst_imm);
2825 amd64_imul_reg_reg (code, ins->sreg1, ins->sreg2);
2830 guint32 size = (ins->opcode == OP_IMUL_IMM) ? 4 : 8;
2832 switch (ins->inst_imm) {
2836 if (ins->dreg != ins->sreg1)
2837 amd64_mov_reg_reg (code, ins->dreg, ins->sreg1, size);
2838 amd64_alu_reg_reg (code, X86_ADD, ins->dreg, ins->dreg);
2841 /* LEA r1, [r2 + r2*2] */
2842 amd64_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 1);
2845 /* LEA r1, [r2 + r2*4] */
2846 amd64_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 2);
2849 /* LEA r1, [r2 + r2*2] */
2851 amd64_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 1);
2852 amd64_alu_reg_reg (code, X86_ADD, ins->dreg, ins->dreg);
2855 /* LEA r1, [r2 + r2*8] */
2856 amd64_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 3);
2859 /* LEA r1, [r2 + r2*4] */
2861 amd64_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 2);
2862 amd64_alu_reg_reg (code, X86_ADD, ins->dreg, ins->dreg);
2865 /* LEA r1, [r2 + r2*2] */
2867 amd64_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 1);
2868 amd64_shift_reg_imm (code, X86_SHL, ins->dreg, 2);
2871 /* LEA r1, [r2 + r2*4] */
2872 /* LEA r1, [r1 + r1*4] */
2873 amd64_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 2);
2874 amd64_lea_memindex (code, ins->dreg, ins->dreg, 0, ins->dreg, 2);
2877 /* LEA r1, [r2 + r2*4] */
2879 /* LEA r1, [r1 + r1*4] */
2880 amd64_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 2);
2881 amd64_shift_reg_imm (code, X86_SHL, ins->dreg, 2);
2882 amd64_lea_memindex (code, ins->dreg, ins->dreg, 0, ins->dreg, 2);
2885 amd64_imul_reg_reg_imm_size (code, ins->dreg, ins->sreg1, ins->inst_imm, size);
2894 /* Regalloc magic makes the div/rem cases the same */
2895 if (ins->sreg2 == AMD64_RDX) {
2896 amd64_mov_membase_reg (code, AMD64_RSP, -8, AMD64_RDX, 8);
2898 amd64_div_membase (code, AMD64_RSP, -8, TRUE);
2901 amd64_div_reg (code, ins->sreg2, TRUE);
2908 if (ins->sreg2 == AMD64_RDX) {
2909 amd64_mov_membase_reg (code, AMD64_RSP, -8, AMD64_RDX, 8);
2910 amd64_alu_reg_reg (code, X86_XOR, AMD64_RDX, AMD64_RDX);
2911 amd64_div_membase (code, AMD64_RSP, -8, FALSE);
2913 amd64_alu_reg_reg (code, X86_XOR, AMD64_RDX, AMD64_RDX);
2914 amd64_div_reg (code, ins->sreg2, FALSE);
2919 if (ins->sreg2 == AMD64_RDX) {
2920 amd64_mov_membase_reg (code, AMD64_RSP, -8, AMD64_RDX, 8);
2921 amd64_cdq_size (code, 4);
2922 amd64_div_membase_size (code, AMD64_RSP, -8, TRUE, 4);
2924 amd64_cdq_size (code, 4);
2925 amd64_div_reg_size (code, ins->sreg2, TRUE, 4);
2930 if (ins->sreg2 == AMD64_RDX) {
2931 amd64_mov_membase_reg (code, AMD64_RSP, -8, AMD64_RDX, 8);
2932 amd64_alu_reg_reg (code, X86_XOR, AMD64_RDX, AMD64_RDX);
2933 amd64_div_membase_size (code, AMD64_RSP, -8, FALSE, 4);
2935 amd64_alu_reg_reg (code, X86_XOR, AMD64_RDX, AMD64_RDX);
2936 amd64_div_reg_size (code, ins->sreg2, FALSE, 4);
2940 amd64_imul_reg_reg (code, ins->sreg1, ins->sreg2);
2941 EMIT_COND_SYSTEM_EXCEPTION (X86_CC_O, FALSE, "OverflowException");
2944 amd64_alu_reg_reg (code, X86_OR, ins->sreg1, ins->sreg2);
2947 g_assert (amd64_is_imm32 (ins->inst_imm));
2948 amd64_alu_reg_imm (code, X86_OR, ins->sreg1, ins->inst_imm);
2952 amd64_alu_reg_reg (code, X86_XOR, ins->sreg1, ins->sreg2);
2955 g_assert (amd64_is_imm32 (ins->inst_imm));
2956 amd64_alu_reg_imm (code, X86_XOR, ins->sreg1, ins->inst_imm);
2960 g_assert (ins->sreg2 == AMD64_RCX);
2961 amd64_shift_reg (code, X86_SHL, ins->dreg);
2965 g_assert (ins->sreg2 == AMD64_RCX);
2966 amd64_shift_reg (code, X86_SAR, ins->dreg);
2969 g_assert (amd64_is_imm32 (ins->inst_imm));
2970 amd64_shift_reg_imm_size (code, X86_SAR, ins->dreg, ins->inst_imm, 4);
2973 g_assert (amd64_is_imm32 (ins->inst_imm));
2974 amd64_shift_reg_imm (code, X86_SAR, ins->dreg, ins->inst_imm);
2977 g_assert (amd64_is_imm32 (ins->inst_imm));
2978 amd64_shift_reg_imm_size (code, X86_SHR, ins->dreg, ins->inst_imm, 4);
2980 case OP_LSHR_UN_IMM:
2981 g_assert (amd64_is_imm32 (ins->inst_imm));
2982 amd64_shift_reg_imm (code, X86_SHR, ins->dreg, ins->inst_imm);
2985 g_assert (ins->sreg2 == AMD64_RCX);
2986 amd64_shift_reg_size (code, X86_SHR, ins->dreg, 4);
2989 g_assert (ins->sreg2 == AMD64_RCX);
2990 amd64_shift_reg (code, X86_SHR, ins->dreg);
2993 g_assert (amd64_is_imm32 (ins->inst_imm));
2994 amd64_shift_reg_imm_size (code, X86_SHL, ins->dreg, ins->inst_imm, 4);
2997 g_assert (amd64_is_imm32 (ins->inst_imm));
2998 amd64_shift_reg_imm (code, X86_SHL, ins->dreg, ins->inst_imm);
3003 amd64_alu_reg_reg_size (code, X86_ADD, ins->sreg1, ins->sreg2, 4);
3006 amd64_alu_reg_reg_size (code, X86_ADC, ins->sreg1, ins->sreg2, 4);
3009 amd64_alu_reg_imm_size (code, X86_ADD, ins->dreg, ins->inst_imm, 4);
3012 amd64_alu_reg_imm_size (code, X86_ADC, ins->dreg, ins->inst_imm, 4);
3016 amd64_alu_reg_reg_size (code, X86_SUB, ins->sreg1, ins->sreg2, 4);
3019 amd64_alu_reg_reg_size (code, X86_SBB, ins->sreg1, ins->sreg2, 4);
3022 amd64_alu_reg_imm_size (code, X86_SUB, ins->dreg, ins->inst_imm, 4);
3025 amd64_alu_reg_imm_size (code, X86_SBB, ins->dreg, ins->inst_imm, 4);
3028 amd64_alu_reg_reg_size (code, X86_AND, ins->sreg1, ins->sreg2, 4);
3031 amd64_alu_reg_imm_size (code, X86_AND, ins->sreg1, ins->inst_imm, 4);
3034 amd64_alu_reg_reg_size (code, X86_OR, ins->sreg1, ins->sreg2, 4);
3037 amd64_alu_reg_imm_size (code, X86_OR, ins->sreg1, ins->inst_imm, 4);
3040 amd64_alu_reg_reg_size (code, X86_XOR, ins->sreg1, ins->sreg2, 4);
3043 amd64_alu_reg_imm_size (code, X86_XOR, ins->sreg1, ins->inst_imm, 4);
3046 amd64_neg_reg_size (code, ins->sreg1, 4);
3049 amd64_not_reg_size (code, ins->sreg1, 4);
3052 g_assert (ins->sreg2 == AMD64_RCX);
3053 amd64_shift_reg_size (code, X86_SHL, ins->dreg, 4);
3056 g_assert (ins->sreg2 == AMD64_RCX);
3057 amd64_shift_reg_size (code, X86_SAR, ins->dreg, 4);
3060 amd64_shift_reg_imm_size (code, X86_SAR, ins->dreg, ins->inst_imm, 4);
3062 case OP_ISHR_UN_IMM:
3063 amd64_shift_reg_imm_size (code, X86_SHR, ins->dreg, ins->inst_imm, 4);
3066 g_assert (ins->sreg2 == AMD64_RCX);
3067 amd64_shift_reg_size (code, X86_SHR, ins->dreg, 4);
3070 amd64_shift_reg_imm_size (code, X86_SHL, ins->dreg, ins->inst_imm, 4);
3073 amd64_imul_reg_reg_size (code, ins->sreg1, ins->sreg2, 4);
3076 amd64_imul_reg_reg_size (code, ins->sreg1, ins->sreg2, 4);
3077 EMIT_COND_SYSTEM_EXCEPTION (X86_CC_O, FALSE, "OverflowException");
3079 case OP_IMUL_OVF_UN:
3080 case OP_LMUL_OVF_UN: {
3081 /* the mul operation and the exception check should most likely be split */
3082 int non_eax_reg, saved_eax = FALSE, saved_edx = FALSE;
3083 int size = (ins->opcode == OP_IMUL_OVF_UN) ? 4 : 8;
3084 /*g_assert (ins->sreg2 == X86_EAX);
3085 g_assert (ins->dreg == X86_EAX);*/
3086 if (ins->sreg2 == X86_EAX) {
3087 non_eax_reg = ins->sreg1;
3088 } else if (ins->sreg1 == X86_EAX) {
3089 non_eax_reg = ins->sreg2;
3091 /* no need to save since we're going to store to it anyway */
3092 if (ins->dreg != X86_EAX) {
3094 amd64_push_reg (code, X86_EAX);
3096 amd64_mov_reg_reg (code, X86_EAX, ins->sreg1, size);
3097 non_eax_reg = ins->sreg2;
3099 if (ins->dreg == X86_EDX) {
3102 amd64_push_reg (code, X86_EAX);
3106 amd64_push_reg (code, X86_EDX);
3108 amd64_mul_reg_size (code, non_eax_reg, FALSE, size);
3109 /* save before the check since pop and mov don't change the flags */
3110 if (ins->dreg != X86_EAX)
3111 amd64_mov_reg_reg (code, ins->dreg, X86_EAX, size);
3113 amd64_pop_reg (code, X86_EDX);
3115 amd64_pop_reg (code, X86_EAX);
3116 EMIT_COND_SYSTEM_EXCEPTION (X86_CC_O, FALSE, "OverflowException");
3120 amd64_alu_reg_reg_size (code, X86_CMP, ins->sreg1, ins->sreg2, 4);
3122 case OP_ICOMPARE_IMM:
3123 amd64_alu_reg_imm_size (code, X86_CMP, ins->sreg1, ins->inst_imm, 4);
3135 EMIT_COND_BRANCH (ins, cc_table [mono_opcode_to_cond (ins->opcode)], cc_signed_table [mono_opcode_to_cond (ins->opcode)]);
3138 amd64_not_reg (code, ins->sreg1);
3141 amd64_neg_reg (code, ins->sreg1);
3146 if ((((guint64)ins->inst_c0) >> 32) == 0)
3147 amd64_mov_reg_imm_size (code, ins->dreg, ins->inst_c0, 4);
3149 amd64_mov_reg_imm_size (code, ins->dreg, ins->inst_c0, 8);
3152 mono_add_patch_info (cfg, offset, (MonoJumpInfoType)ins->inst_i1, ins->inst_p0);
3153 amd64_mov_reg_membase (code, ins->dreg, AMD64_RIP, 0, 8);
3158 amd64_mov_reg_reg (code, ins->dreg, ins->sreg1, sizeof (gpointer));
3160 case OP_AMD64_SET_XMMREG_R4: {
3162 amd64_sse_cvtsd2ss_reg_reg (code, ins->dreg, ins->sreg1);
3165 amd64_fst_membase (code, AMD64_RSP, -8, FALSE, TRUE);
3166 /* ins->dreg is set to -1 by the reg allocator */
3167 amd64_movss_reg_membase (code, ins->backend.reg3, AMD64_RSP, -8);
3171 case OP_AMD64_SET_XMMREG_R8: {
3173 if (ins->dreg != ins->sreg1)
3174 amd64_sse_movsd_reg_reg (code, ins->dreg, ins->sreg1);
3177 amd64_fst_membase (code, AMD64_RSP, -8, TRUE, TRUE);
3178 /* ins->dreg is set to -1 by the reg allocator */
3179 amd64_movsd_reg_membase (code, ins->backend.reg3, AMD64_RSP, -8);
3185 * Note: this 'frame destruction' logic is useful for tail calls, too.
3186 * Keep in sync with the code in emit_epilog.
3190 /* FIXME: no tracing support... */
3191 if (cfg->prof_options & MONO_PROFILE_ENTER_LEAVE)
3192 code = mono_arch_instrument_epilog (cfg, mono_profiler_method_leave, code, FALSE);
3194 g_assert (!cfg->method->save_lmf);
3196 code = emit_load_volatile_arguments (cfg, code);
3198 if (cfg->arch.omit_fp) {
3199 guint32 save_offset = 0;
3200 /* Pop callee-saved registers */
3201 for (i = 0; i < AMD64_NREG; ++i)
3202 if (AMD64_IS_CALLEE_SAVED_REG (i) && (cfg->used_int_regs & (1 << i))) {
3203 amd64_mov_reg_membase (code, i, AMD64_RSP, save_offset, 8);
3206 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, cfg->arch.stack_alloc_size);
3209 for (i = 0; i < AMD64_NREG; ++i)
3210 if (AMD64_IS_CALLEE_SAVED_REG (i) && (cfg->used_int_regs & (1 << i)))
3211 pos -= sizeof (gpointer);
3214 amd64_lea_membase (code, AMD64_RSP, AMD64_RBP, pos);
3216 /* Pop registers in reverse order */
3217 for (i = AMD64_NREG - 1; i > 0; --i)
3218 if (AMD64_IS_CALLEE_SAVED_REG (i) && (cfg->used_int_regs & (1 << i))) {
3219 amd64_pop_reg (code, i);
3225 offset = code - cfg->native_code;
3226 mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_METHOD_JUMP, ins->inst_p0);
3227 if (cfg->compile_aot)
3228 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RIP, 0, 8);
3230 amd64_set_reg_template (code, AMD64_R11);
3231 amd64_jump_reg (code, AMD64_R11);
3235 /* ensure ins->sreg1 is not NULL */
3236 amd64_alu_membase_imm (code, X86_CMP, ins->sreg1, 0, 0);
3239 amd64_lea_membase (code, AMD64_R11, cfg->frame_reg, cfg->sig_cookie);
3240 amd64_mov_membase_reg (code, ins->sreg1, 0, AMD64_R11, 8);
3248 call = (MonoCallInst*)ins;
3250 * The AMD64 ABI forces callers to know about varargs.
3252 if ((call->signature->call_convention == MONO_CALL_VARARG) && (call->signature->pinvoke))
3253 amd64_alu_reg_reg (code, X86_XOR, AMD64_RAX, AMD64_RAX);
3254 else if ((cfg->method->wrapper_type == MONO_WRAPPER_MANAGED_TO_NATIVE) && (cfg->method->klass->image != mono_defaults.corlib)) {
3256 * Since the unmanaged calling convention doesn't contain a
3257 * 'vararg' entry, we have to treat every pinvoke call as a
3258 * potential vararg call.
3262 for (i = 0; i < AMD64_XMM_NREG; ++i)
3263 if (call->used_fregs & (1 << i))
3266 amd64_alu_reg_reg (code, X86_XOR, AMD64_RAX, AMD64_RAX);
3268 amd64_mov_reg_imm (code, AMD64_RAX, nregs);
3271 if (ins->flags & MONO_INST_HAS_METHOD)
3272 code = emit_call (cfg, code, MONO_PATCH_INFO_METHOD, call->method);
3274 code = emit_call (cfg, code, MONO_PATCH_INFO_ABS, call->fptr);
3275 if (call->stack_usage && !CALLCONV_IS_STDCALL (call->signature->call_convention))
3276 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, call->stack_usage);
3277 code = emit_move_return_value (cfg, ins, code);
3282 case OP_VOIDCALL_REG:
3284 call = (MonoCallInst*)ins;
3286 if (AMD64_IS_ARGUMENT_REG (ins->sreg1)) {
3287 amd64_mov_reg_reg (code, AMD64_R11, ins->sreg1, 8);
3288 ins->sreg1 = AMD64_R11;
3292 * The AMD64 ABI forces callers to know about varargs.
3294 if ((call->signature->call_convention == MONO_CALL_VARARG) && (call->signature->pinvoke)) {
3295 if (ins->sreg1 == AMD64_RAX) {
3296 amd64_mov_reg_reg (code, AMD64_R11, AMD64_RAX, 8);
3297 ins->sreg1 = AMD64_R11;
3299 amd64_alu_reg_reg (code, X86_XOR, AMD64_RAX, AMD64_RAX);
3301 amd64_call_reg (code, ins->sreg1);
3302 if (call->stack_usage && !CALLCONV_IS_STDCALL (call->signature->call_convention))
3303 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, call->stack_usage);
3304 code = emit_move_return_value (cfg, ins, code);
3306 case OP_FCALL_MEMBASE:
3307 case OP_LCALL_MEMBASE:
3308 case OP_VCALL_MEMBASE:
3309 case OP_VOIDCALL_MEMBASE:
3310 case OP_CALL_MEMBASE:
3311 call = (MonoCallInst*)ins;
3313 if (AMD64_IS_ARGUMENT_REG (ins->sreg1)) {
3315 * Can't use R11 because it is clobbered by the trampoline
3316 * code, and the reg value is needed by get_vcall_slot_addr.
3318 amd64_mov_reg_reg (code, AMD64_RAX, ins->sreg1, 8);
3319 ins->sreg1 = AMD64_RAX;
3322 amd64_call_membase (code, ins->sreg1, ins->inst_offset);
3323 if (call->stack_usage && !CALLCONV_IS_STDCALL (call->signature->call_convention))
3324 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, call->stack_usage);
3325 code = emit_move_return_value (cfg, ins, code);
3329 amd64_push_reg (code, ins->sreg1);
3331 case OP_X86_PUSH_IMM:
3332 g_assert (amd64_is_imm32 (ins->inst_imm));
3333 amd64_push_imm (code, ins->inst_imm);
3335 case OP_X86_PUSH_MEMBASE:
3336 amd64_push_membase (code, ins->inst_basereg, ins->inst_offset);
3338 case OP_X86_PUSH_OBJ:
3339 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, ins->inst_imm);
3340 amd64_push_reg (code, AMD64_RDI);
3341 amd64_push_reg (code, AMD64_RSI);
3342 amd64_push_reg (code, AMD64_RCX);
3343 if (ins->inst_offset)
3344 amd64_lea_membase (code, AMD64_RSI, ins->inst_basereg, ins->inst_offset);
3346 amd64_mov_reg_reg (code, AMD64_RSI, ins->inst_basereg, 8);
3347 amd64_lea_membase (code, AMD64_RDI, AMD64_RSP, 3 * 8);
3348 amd64_mov_reg_imm (code, AMD64_RCX, (ins->inst_imm >> 3));
3350 amd64_prefix (code, X86_REP_PREFIX);
3352 amd64_pop_reg (code, AMD64_RCX);
3353 amd64_pop_reg (code, AMD64_RSI);
3354 amd64_pop_reg (code, AMD64_RDI);
3357 amd64_lea_memindex (code, ins->dreg, ins->sreg1, ins->inst_imm, ins->sreg2, ins->backend.shift_amount);
3359 case OP_X86_LEA_MEMBASE:
3360 amd64_lea_membase (code, ins->dreg, ins->sreg1, ins->inst_imm);
3363 amd64_xchg_reg_reg (code, ins->sreg1, ins->sreg2, 4);
3366 /* keep alignment */
3367 amd64_alu_reg_imm (code, X86_ADD, ins->sreg1, MONO_ARCH_FRAME_ALIGNMENT - 1);
3368 amd64_alu_reg_imm (code, X86_AND, ins->sreg1, ~(MONO_ARCH_FRAME_ALIGNMENT - 1));
3369 code = mono_emit_stack_alloc (code, ins);
3370 amd64_mov_reg_reg (code, ins->dreg, AMD64_RSP, 8);
3376 amd64_mov_reg_reg (code, AMD64_ARG_REG1, ins->sreg1, 8);
3377 code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD,
3378 (gpointer)"mono_arch_throw_exception");
3382 amd64_mov_reg_reg (code, AMD64_ARG_REG1, ins->sreg1, 8);
3383 code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD,
3384 (gpointer)"mono_arch_rethrow_exception");
3387 case OP_CALL_HANDLER:
3389 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, 8);
3390 mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_BB, ins->inst_target_bb);
3391 amd64_call_imm (code, 0);
3392 /* Restore stack alignment */
3393 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, 8);
3395 case OP_START_HANDLER: {
3396 MonoInst *spvar = mono_find_spvar_for_region (cfg, bb->region);
3397 amd64_mov_membase_reg (code, spvar->inst_basereg, spvar->inst_offset, AMD64_RSP, 8);
3400 case OP_ENDFINALLY: {
3401 MonoInst *spvar = mono_find_spvar_for_region (cfg, bb->region);
3402 amd64_mov_reg_membase (code, AMD64_RSP, spvar->inst_basereg, spvar->inst_offset, 8);
3406 case OP_ENDFILTER: {
3407 MonoInst *spvar = mono_find_spvar_for_region (cfg, bb->region);
3408 amd64_mov_reg_membase (code, AMD64_RSP, spvar->inst_basereg, spvar->inst_offset, 8);
3409 /* The local allocator will put the result into RAX */
3415 ins->inst_c0 = code - cfg->native_code;
3420 //g_print ("target: %p, next: %p, curr: %p, last: %p\n", ins->inst_target_bb, bb->next_bb, ins, bb->last_ins);
3421 //if ((ins->inst_target_bb == bb->next_bb) && ins == bb->last_ins)
3423 if (ins->flags & MONO_INST_BRLABEL) {
3424 if (ins->inst_i0->inst_c0) {
3425 amd64_jump_code (code, cfg->native_code + ins->inst_i0->inst_c0);
3427 mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_LABEL, ins->inst_i0);
3428 if ((cfg->opt & MONO_OPT_BRANCH) &&
3429 x86_is_imm8 (ins->inst_i0->inst_c1 - cpos))
3430 x86_jump8 (code, 0);
3432 x86_jump32 (code, 0);
3435 if (ins->inst_target_bb->native_offset) {
3436 amd64_jump_code (code, cfg->native_code + ins->inst_target_bb->native_offset);
3438 mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_BB, ins->inst_target_bb);
3439 if ((cfg->opt & MONO_OPT_BRANCH) &&
3440 x86_is_imm8 (ins->inst_target_bb->max_offset - cpos))
3441 x86_jump8 (code, 0);
3443 x86_jump32 (code, 0);
3448 amd64_jump_reg (code, ins->sreg1);
3460 amd64_set_reg (code, cc_table [mono_opcode_to_cond (ins->opcode)], ins->dreg, cc_signed_table [mono_opcode_to_cond (ins->opcode)]);
3461 amd64_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
3463 case OP_COND_EXC_EQ:
3464 case OP_COND_EXC_NE_UN:
3465 case OP_COND_EXC_LT:
3466 case OP_COND_EXC_LT_UN:
3467 case OP_COND_EXC_GT:
3468 case OP_COND_EXC_GT_UN:
3469 case OP_COND_EXC_GE:
3470 case OP_COND_EXC_GE_UN:
3471 case OP_COND_EXC_LE:
3472 case OP_COND_EXC_LE_UN:
3473 EMIT_COND_SYSTEM_EXCEPTION (cc_table [mono_opcode_to_cond (ins->opcode)], cc_signed_table [mono_opcode_to_cond (ins->opcode)], ins->inst_p1);
3475 case OP_COND_EXC_OV:
3476 case OP_COND_EXC_NO:
3478 case OP_COND_EXC_NC:
3479 EMIT_COND_SYSTEM_EXCEPTION (branch_cc_table [ins->opcode - OP_COND_EXC_EQ],
3480 (ins->opcode < OP_COND_EXC_NE_UN), ins->inst_p1);
3482 case OP_COND_EXC_IOV:
3483 case OP_COND_EXC_IC:
3484 EMIT_COND_SYSTEM_EXCEPTION (branch_cc_table [ins->opcode - OP_COND_EXC_IEQ],
3485 (ins->opcode < OP_COND_EXC_INE_UN), ins->inst_p1);
3497 EMIT_COND_BRANCH (ins, cc_table [mono_opcode_to_cond (ins->opcode)], cc_signed_table [mono_opcode_to_cond (ins->opcode)]);
3500 /* floating point opcodes */
3502 double d = *(double *)ins->inst_p0;
3505 if ((d == 0.0) && (mono_signbit (d) == 0)) {
3506 amd64_sse_xorpd_reg_reg (code, ins->dreg, ins->dreg);
3509 mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_R8, ins->inst_p0);
3510 amd64_sse_movsd_reg_membase (code, ins->dreg, AMD64_RIP, 0);
3513 else if ((d == 0.0) && (mono_signbit (d) == 0)) {
3515 } else if (d == 1.0) {
3518 mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_R8, ins->inst_p0);
3519 amd64_fld_membase (code, AMD64_RIP, 0, TRUE);
3524 float f = *(float *)ins->inst_p0;
3527 if ((f == 0.0) && (mono_signbit (f) == 0)) {
3528 amd64_sse_xorpd_reg_reg (code, ins->dreg, ins->dreg);
3531 mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_R4, ins->inst_p0);
3532 amd64_sse_movss_reg_membase (code, ins->dreg, AMD64_RIP, 0);
3533 amd64_sse_cvtss2sd_reg_reg (code, ins->dreg, ins->dreg);
3536 else if ((f == 0.0) && (mono_signbit (f) == 0)) {
3538 } else if (f == 1.0) {
3541 mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_R4, ins->inst_p0);
3542 amd64_fld_membase (code, AMD64_RIP, 0, FALSE);
3546 case OP_STORER8_MEMBASE_REG:
3548 amd64_sse_movsd_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, ins->sreg1);
3550 amd64_fst_membase (code, ins->inst_destbasereg, ins->inst_offset, TRUE, TRUE);
3552 case OP_LOADR8_SPILL_MEMBASE:
3554 g_assert_not_reached ();
3555 amd64_fld_membase (code, ins->inst_basereg, ins->inst_offset, TRUE);
3556 amd64_fxch (code, 1);
3558 case OP_LOADR8_MEMBASE:
3560 amd64_sse_movsd_reg_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset);
3562 amd64_fld_membase (code, ins->inst_basereg, ins->inst_offset, TRUE);
3564 case OP_STORER4_MEMBASE_REG:
3566 /* This requires a double->single conversion */
3567 amd64_sse_cvtsd2ss_reg_reg (code, AMD64_XMM15, ins->sreg1);
3568 amd64_sse_movss_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, AMD64_XMM15);
3571 amd64_fst_membase (code, ins->inst_destbasereg, ins->inst_offset, FALSE, TRUE);
3573 case OP_LOADR4_MEMBASE:
3575 amd64_sse_movss_reg_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset);
3576 amd64_sse_cvtss2sd_reg_reg (code, ins->dreg, ins->dreg);
3579 amd64_fld_membase (code, ins->inst_basereg, ins->inst_offset, FALSE);
3581 case CEE_CONV_R4: /* FIXME: change precision */
3584 amd64_sse_cvtsi2sd_reg_reg_size (code, ins->dreg, ins->sreg1, 4);
3586 amd64_push_reg (code, ins->sreg1);
3587 amd64_fild_membase (code, AMD64_RSP, 0, FALSE);
3588 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, 8);
3593 g_assert_not_reached ();
3595 case OP_LCONV_TO_R4: /* FIXME: change precision */
3596 case OP_LCONV_TO_R8:
3598 amd64_sse_cvtsi2sd_reg_reg (code, ins->dreg, ins->sreg1);
3600 amd64_push_reg (code, ins->sreg1);
3601 amd64_fild_membase (code, AMD64_RSP, 0, TRUE);
3602 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, 8);
3605 case OP_X86_FP_LOAD_I8:
3607 g_assert_not_reached ();
3608 amd64_fild_membase (code, ins->inst_basereg, ins->inst_offset, TRUE);
3610 case OP_X86_FP_LOAD_I4:
3612 g_assert_not_reached ();
3613 amd64_fild_membase (code, ins->inst_basereg, ins->inst_offset, FALSE);
3615 case OP_FCONV_TO_I1:
3616 code = emit_float_to_int (cfg, code, ins->dreg, ins->sreg1, 1, TRUE);
3618 case OP_FCONV_TO_U1:
3619 code = emit_float_to_int (cfg, code, ins->dreg, ins->sreg1, 1, FALSE);
3621 case OP_FCONV_TO_I2:
3622 code = emit_float_to_int (cfg, code, ins->dreg, ins->sreg1, 2, TRUE);
3624 case OP_FCONV_TO_U2:
3625 code = emit_float_to_int (cfg, code, ins->dreg, ins->sreg1, 2, FALSE);
3627 case OP_FCONV_TO_I4:
3629 code = emit_float_to_int (cfg, code, ins->dreg, ins->sreg1, 4, TRUE);
3631 case OP_FCONV_TO_I8:
3632 code = emit_float_to_int (cfg, code, ins->dreg, ins->sreg1, 8, TRUE);
3634 case OP_LCONV_TO_R_UN: {
3635 static guint8 mn[] = { 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x80, 0x3f, 0x40 };
3639 g_assert_not_reached ();
3641 /* load 64bit integer to FP stack */
3642 amd64_push_imm (code, 0);
3643 amd64_push_reg (code, ins->sreg2);
3644 amd64_push_reg (code, ins->sreg1);
3645 amd64_fild_membase (code, AMD64_RSP, 0, TRUE);
3646 /* store as 80bit FP value */
3647 x86_fst80_membase (code, AMD64_RSP, 0);
3649 /* test if lreg is negative */
3650 amd64_test_reg_reg (code, ins->sreg2, ins->sreg2);
3651 br = code; x86_branch8 (code, X86_CC_GEZ, 0, TRUE);
3653 /* add correction constant mn */
3654 x86_fld80_mem (code, (gssize)mn);
3655 x86_fld80_membase (code, AMD64_RSP, 0);
3656 amd64_fp_op_reg (code, X86_FADD, 1, TRUE);
3657 x86_fst80_membase (code, AMD64_RSP, 0);
3659 amd64_patch (br, code);
3661 x86_fld80_membase (code, AMD64_RSP, 0);
3662 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, 12);
3666 case CEE_CONV_OVF_U4:
3667 amd64_alu_reg_imm (code, X86_CMP, ins->sreg1, 0);
3668 EMIT_COND_SYSTEM_EXCEPTION (X86_CC_LT, TRUE, "OverflowException");
3669 amd64_mov_reg_reg (code, ins->dreg, ins->sreg1, 8);
3671 case CEE_CONV_OVF_I4_UN:
3672 amd64_alu_reg_imm (code, X86_CMP, ins->sreg1, 0x7fffffff);
3673 EMIT_COND_SYSTEM_EXCEPTION (X86_CC_GT, FALSE, "OverflowException");
3674 amd64_mov_reg_reg (code, ins->dreg, ins->sreg1, 8);
3677 if (use_sse2 && (ins->dreg != ins->sreg1))
3678 amd64_sse_movsd_reg_reg (code, ins->dreg, ins->sreg1);
3682 amd64_sse_addsd_reg_reg (code, ins->dreg, ins->sreg2);
3684 amd64_fp_op_reg (code, X86_FADD, 1, TRUE);
3688 amd64_sse_subsd_reg_reg (code, ins->dreg, ins->sreg2);
3690 amd64_fp_op_reg (code, X86_FSUB, 1, TRUE);
3694 amd64_sse_mulsd_reg_reg (code, ins->dreg, ins->sreg2);
3696 amd64_fp_op_reg (code, X86_FMUL, 1, TRUE);
3700 amd64_sse_divsd_reg_reg (code, ins->dreg, ins->sreg2);
3702 amd64_fp_op_reg (code, X86_FDIV, 1, TRUE);
3706 static double r8_0 = -0.0;
3708 g_assert (ins->sreg1 == ins->dreg);
3710 mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_R8, &r8_0);
3711 amd64_sse_xorpd_reg_membase (code, ins->dreg, AMD64_RIP, 0);
3718 EMIT_SSE2_FPFUNC (code, fsin, ins->dreg, ins->sreg1);
3723 amd64_fp_op_reg (code, X86_FADD, 1, TRUE);
3728 EMIT_SSE2_FPFUNC (code, fcos, ins->dreg, ins->sreg1);
3733 amd64_fp_op_reg (code, X86_FADD, 1, TRUE);
3738 EMIT_SSE2_FPFUNC (code, fabs, ins->dreg, ins->sreg1);
3745 * it really doesn't make sense to inline all this code,
3746 * it's here just to show that things may not be as simple
3749 guchar *check_pos, *end_tan, *pop_jump;
3751 g_assert_not_reached ();
3752 amd64_push_reg (code, AMD64_RAX);
3754 amd64_fnstsw (code);
3755 amd64_test_reg_imm (code, AMD64_RAX, X86_FP_C2);
3757 x86_branch8 (code, X86_CC_NE, 0, FALSE);
3758 amd64_fstp (code, 0); /* pop the 1.0 */
3760 x86_jump8 (code, 0);
3762 amd64_fp_op (code, X86_FADD, 0);
3763 amd64_fxch (code, 1);
3766 amd64_test_reg_imm (code, AMD64_RAX, X86_FP_C2);
3768 x86_branch8 (code, X86_CC_NE, 0, FALSE);
3769 amd64_fstp (code, 1);
3771 amd64_patch (pop_jump, code);
3772 amd64_fstp (code, 0); /* pop the 1.0 */
3773 amd64_patch (check_pos, code);
3774 amd64_patch (end_tan, code);
3776 amd64_fp_op_reg (code, X86_FADD, 1, TRUE);
3777 amd64_pop_reg (code, AMD64_RAX);
3782 g_assert_not_reached ();
3784 amd64_fpatan (code);
3786 amd64_fp_op_reg (code, X86_FADD, 1, TRUE);
3790 EMIT_SSE2_FPFUNC (code, fsqrt, ins->dreg, ins->sreg1);
3796 g_assert (cfg->opt & MONO_OPT_CMOV);
3797 g_assert (ins->dreg == ins->sreg1);
3798 amd64_alu_reg_reg_size (code, X86_CMP, ins->sreg1, ins->sreg2, 4);
3799 amd64_cmov_reg_size (code, X86_CC_GT, TRUE, ins->dreg, ins->sreg2, 4);
3802 g_assert (cfg->opt & MONO_OPT_CMOV);
3803 g_assert (ins->dreg == ins->sreg1);
3804 amd64_alu_reg_reg_size (code, X86_CMP, ins->sreg1, ins->sreg2, 4);
3805 amd64_cmov_reg_size (code, X86_CC_LT, TRUE, ins->dreg, ins->sreg2, 4);
3808 g_assert (cfg->opt & MONO_OPT_CMOV);
3809 g_assert (ins->dreg == ins->sreg1);
3810 amd64_alu_reg_reg (code, X86_CMP, ins->sreg1, ins->sreg2);
3811 amd64_cmov_reg (code, X86_CC_GT, TRUE, ins->dreg, ins->sreg2);
3814 g_assert (cfg->opt & MONO_OPT_CMOV);
3815 g_assert (ins->dreg == ins->sreg1);
3816 amd64_alu_reg_reg (code, X86_CMP, ins->sreg1, ins->sreg2);
3817 amd64_cmov_reg (code, X86_CC_LT, TRUE, ins->dreg, ins->sreg2);
3821 amd64_fstp (code, 0);
3827 g_assert_not_reached ();
3828 amd64_push_reg (code, AMD64_RAX);
3829 /* we need to exchange ST(0) with ST(1) */
3830 amd64_fxch (code, 1);
3832 /* this requires a loop, because fprem somtimes
3833 * returns a partial remainder */
3835 /* looks like MS is using fprem instead of the IEEE compatible fprem1 */
3836 /* x86_fprem1 (code); */
3838 amd64_fnstsw (code);
3839 amd64_alu_reg_imm (code, X86_AND, AMD64_RAX, X86_FP_C2);
3841 x86_branch8 (code, X86_CC_NE, l1 - l2, FALSE);
3844 amd64_fstp (code, 1);
3846 amd64_pop_reg (code, AMD64_RAX);
3852 * The two arguments are swapped because the fbranch instructions
3853 * depend on this for the non-sse case to work.
3855 amd64_sse_comisd_reg_reg (code, ins->sreg2, ins->sreg1);
3858 if (cfg->opt & MONO_OPT_FCMOV) {
3859 amd64_fcomip (code, 1);
3860 amd64_fstp (code, 0);
3863 /* this overwrites EAX */
3864 EMIT_FPCOMPARE(code);
3865 amd64_alu_reg_imm (code, X86_AND, AMD64_RAX, X86_FP_CC_MASK);
3868 if (use_sse2 || (cfg->opt & MONO_OPT_FCMOV)) {
3869 /* zeroing the register at the start results in
3870 * shorter and faster code (we can also remove the widening op)
3872 guchar *unordered_check;
3873 amd64_alu_reg_reg (code, X86_XOR, ins->dreg, ins->dreg);
3876 amd64_sse_comisd_reg_reg (code, ins->sreg1, ins->sreg2);
3878 amd64_fcomip (code, 1);
3879 amd64_fstp (code, 0);
3881 unordered_check = code;
3882 x86_branch8 (code, X86_CC_P, 0, FALSE);
3883 amd64_set_reg (code, X86_CC_EQ, ins->dreg, FALSE);
3884 amd64_patch (unordered_check, code);
3887 if (ins->dreg != AMD64_RAX)
3888 amd64_push_reg (code, AMD64_RAX);
3890 EMIT_FPCOMPARE(code);
3891 amd64_alu_reg_imm (code, X86_AND, AMD64_RAX, X86_FP_CC_MASK);
3892 amd64_alu_reg_imm (code, X86_CMP, AMD64_RAX, 0x4000);
3893 amd64_set_reg (code, X86_CC_EQ, ins->dreg, TRUE);
3894 amd64_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
3896 if (ins->dreg != AMD64_RAX)
3897 amd64_pop_reg (code, AMD64_RAX);
3901 if (use_sse2 || (cfg->opt & MONO_OPT_FCMOV)) {
3902 /* zeroing the register at the start results in
3903 * shorter and faster code (we can also remove the widening op)
3905 amd64_alu_reg_reg (code, X86_XOR, ins->dreg, ins->dreg);
3907 amd64_sse_comisd_reg_reg (code, ins->sreg2, ins->sreg1);
3909 amd64_fcomip (code, 1);
3910 amd64_fstp (code, 0);
3912 if (ins->opcode == OP_FCLT_UN) {
3913 guchar *unordered_check = code;
3914 guchar *jump_to_end;
3915 x86_branch8 (code, X86_CC_P, 0, FALSE);
3916 amd64_set_reg (code, X86_CC_GT, ins->dreg, FALSE);
3918 x86_jump8 (code, 0);
3919 amd64_patch (unordered_check, code);
3920 amd64_inc_reg (code, ins->dreg);
3921 amd64_patch (jump_to_end, code);
3923 amd64_set_reg (code, X86_CC_GT, ins->dreg, FALSE);
3927 if (ins->dreg != AMD64_RAX)
3928 amd64_push_reg (code, AMD64_RAX);
3930 EMIT_FPCOMPARE(code);
3931 amd64_alu_reg_imm (code, X86_AND, AMD64_RAX, X86_FP_CC_MASK);
3932 if (ins->opcode == OP_FCLT_UN) {
3933 guchar *is_not_zero_check, *end_jump;
3934 is_not_zero_check = code;
3935 x86_branch8 (code, X86_CC_NZ, 0, TRUE);
3937 x86_jump8 (code, 0);
3938 amd64_patch (is_not_zero_check, code);
3939 amd64_alu_reg_imm (code, X86_CMP, AMD64_RAX, X86_FP_CC_MASK);
3941 amd64_patch (end_jump, code);
3943 amd64_set_reg (code, X86_CC_EQ, ins->dreg, TRUE);
3944 amd64_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
3946 if (ins->dreg != AMD64_RAX)
3947 amd64_pop_reg (code, AMD64_RAX);
3951 if (use_sse2 || (cfg->opt & MONO_OPT_FCMOV)) {
3952 /* zeroing the register at the start results in
3953 * shorter and faster code (we can also remove the widening op)
3955 guchar *unordered_check;
3956 amd64_alu_reg_reg (code, X86_XOR, ins->dreg, ins->dreg);
3958 amd64_sse_comisd_reg_reg (code, ins->sreg2, ins->sreg1);
3960 amd64_fcomip (code, 1);
3961 amd64_fstp (code, 0);
3963 if (ins->opcode == OP_FCGT) {
3964 unordered_check = code;
3965 x86_branch8 (code, X86_CC_P, 0, FALSE);
3966 amd64_set_reg (code, X86_CC_LT, ins->dreg, FALSE);
3967 amd64_patch (unordered_check, code);
3969 amd64_set_reg (code, X86_CC_LT, ins->dreg, FALSE);
3973 if (ins->dreg != AMD64_RAX)
3974 amd64_push_reg (code, AMD64_RAX);
3976 EMIT_FPCOMPARE(code);
3977 amd64_alu_reg_imm (code, X86_AND, AMD64_RAX, X86_FP_CC_MASK);
3978 amd64_alu_reg_imm (code, X86_CMP, AMD64_RAX, X86_FP_C0);
3979 if (ins->opcode == OP_FCGT_UN) {
3980 guchar *is_not_zero_check, *end_jump;
3981 is_not_zero_check = code;
3982 x86_branch8 (code, X86_CC_NZ, 0, TRUE);
3984 x86_jump8 (code, 0);
3985 amd64_patch (is_not_zero_check, code);
3986 amd64_alu_reg_imm (code, X86_CMP, AMD64_RAX, X86_FP_CC_MASK);
3988 amd64_patch (end_jump, code);
3990 amd64_set_reg (code, X86_CC_EQ, ins->dreg, TRUE);
3991 amd64_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
3993 if (ins->dreg != AMD64_RAX)
3994 amd64_pop_reg (code, AMD64_RAX);
3996 case OP_FCLT_MEMBASE:
3997 case OP_FCGT_MEMBASE:
3998 case OP_FCLT_UN_MEMBASE:
3999 case OP_FCGT_UN_MEMBASE:
4000 case OP_FCEQ_MEMBASE: {
4001 guchar *unordered_check, *jump_to_end;
4003 g_assert (use_sse2);
4005 amd64_alu_reg_reg (code, X86_XOR, ins->dreg, ins->dreg);
4006 amd64_sse_comisd_reg_membase (code, ins->sreg1, ins->sreg2, ins->inst_offset);
4008 switch (ins->opcode) {
4009 case OP_FCEQ_MEMBASE:
4010 x86_cond = X86_CC_EQ;
4012 case OP_FCLT_MEMBASE:
4013 case OP_FCLT_UN_MEMBASE:
4014 x86_cond = X86_CC_LT;
4016 case OP_FCGT_MEMBASE:
4017 case OP_FCGT_UN_MEMBASE:
4018 x86_cond = X86_CC_GT;
4021 g_assert_not_reached ();
4024 unordered_check = code;
4025 x86_branch8 (code, X86_CC_P, 0, FALSE);
4026 amd64_set_reg (code, x86_cond, ins->dreg, FALSE);
4028 switch (ins->opcode) {
4029 case OP_FCEQ_MEMBASE:
4030 case OP_FCLT_MEMBASE:
4031 case OP_FCGT_MEMBASE:
4032 amd64_patch (unordered_check, code);
4034 case OP_FCLT_UN_MEMBASE:
4035 case OP_FCGT_UN_MEMBASE:
4037 x86_jump8 (code, 0);
4038 amd64_patch (unordered_check, code);
4039 amd64_inc_reg (code, ins->dreg);
4040 amd64_patch (jump_to_end, code);
4048 if (use_sse2 || (cfg->opt & MONO_OPT_FCMOV)) {
4049 guchar *jump = code;
4050 x86_branch8 (code, X86_CC_P, 0, TRUE);
4051 EMIT_COND_BRANCH (ins, X86_CC_EQ, FALSE);
4052 amd64_patch (jump, code);
4055 amd64_alu_reg_imm (code, X86_CMP, AMD64_RAX, 0x4000);
4056 EMIT_COND_BRANCH (ins, X86_CC_EQ, TRUE);
4059 /* Branch if C013 != 100 */
4060 if (use_sse2 || (cfg->opt & MONO_OPT_FCMOV)) {
4061 /* branch if !ZF or (PF|CF) */
4062 EMIT_COND_BRANCH (ins, X86_CC_NE, FALSE);
4063 EMIT_COND_BRANCH (ins, X86_CC_P, FALSE);
4064 EMIT_COND_BRANCH (ins, X86_CC_B, FALSE);
4067 amd64_alu_reg_imm (code, X86_CMP, AMD64_RAX, X86_FP_C3);
4068 EMIT_COND_BRANCH (ins, X86_CC_NE, FALSE);
4071 if (use_sse2 || (cfg->opt & MONO_OPT_FCMOV)) {
4072 EMIT_COND_BRANCH (ins, X86_CC_GT, FALSE);
4075 EMIT_COND_BRANCH (ins, X86_CC_EQ, FALSE);
4078 if (use_sse2 || (cfg->opt & MONO_OPT_FCMOV)) {
4079 EMIT_COND_BRANCH (ins, X86_CC_P, FALSE);
4080 EMIT_COND_BRANCH (ins, X86_CC_GT, FALSE);
4083 if (ins->opcode == OP_FBLT_UN) {
4084 guchar *is_not_zero_check, *end_jump;
4085 is_not_zero_check = code;
4086 x86_branch8 (code, X86_CC_NZ, 0, TRUE);
4088 x86_jump8 (code, 0);
4089 amd64_patch (is_not_zero_check, code);
4090 amd64_alu_reg_imm (code, X86_CMP, AMD64_RAX, X86_FP_CC_MASK);
4092 amd64_patch (end_jump, code);
4094 EMIT_COND_BRANCH (ins, X86_CC_EQ, FALSE);
4098 if (use_sse2 || (cfg->opt & MONO_OPT_FCMOV)) {
4099 if (ins->opcode == OP_FBGT) {
4102 /* skip branch if C1=1 */
4104 x86_branch8 (code, X86_CC_P, 0, FALSE);
4105 /* branch if (C0 | C3) = 1 */
4106 EMIT_COND_BRANCH (ins, X86_CC_LT, FALSE);
4107 amd64_patch (br1, code);
4110 EMIT_COND_BRANCH (ins, X86_CC_LT, FALSE);
4114 amd64_alu_reg_imm (code, X86_CMP, AMD64_RAX, X86_FP_C0);
4115 if (ins->opcode == OP_FBGT_UN) {
4116 guchar *is_not_zero_check, *end_jump;
4117 is_not_zero_check = code;
4118 x86_branch8 (code, X86_CC_NZ, 0, TRUE);
4120 x86_jump8 (code, 0);
4121 amd64_patch (is_not_zero_check, code);
4122 amd64_alu_reg_imm (code, X86_CMP, AMD64_RAX, X86_FP_CC_MASK);
4124 amd64_patch (end_jump, code);
4126 EMIT_COND_BRANCH (ins, X86_CC_EQ, FALSE);
4129 /* Branch if C013 == 100 or 001 */
4130 if (use_sse2 || (cfg->opt & MONO_OPT_FCMOV)) {
4133 /* skip branch if C1=1 */
4135 x86_branch8 (code, X86_CC_P, 0, FALSE);
4136 /* branch if (C0 | C3) = 1 */
4137 EMIT_COND_BRANCH (ins, X86_CC_BE, FALSE);
4138 amd64_patch (br1, code);
4141 amd64_alu_reg_imm (code, X86_CMP, AMD64_RAX, X86_FP_C0);
4142 EMIT_COND_BRANCH (ins, X86_CC_EQ, FALSE);
4143 amd64_alu_reg_imm (code, X86_CMP, AMD64_RAX, X86_FP_C3);
4144 EMIT_COND_BRANCH (ins, X86_CC_EQ, FALSE);
4147 /* Branch if C013 == 000 */
4148 if (use_sse2 || (cfg->opt & MONO_OPT_FCMOV)) {
4149 EMIT_COND_BRANCH (ins, X86_CC_LE, FALSE);
4152 EMIT_COND_BRANCH (ins, X86_CC_NE, FALSE);
4155 /* Branch if C013=000 or 100 */
4156 if (use_sse2 || (cfg->opt & MONO_OPT_FCMOV)) {
4159 /* skip branch if C1=1 */
4161 x86_branch8 (code, X86_CC_P, 0, FALSE);
4162 /* branch if C0=0 */
4163 EMIT_COND_BRANCH (ins, X86_CC_NB, FALSE);
4164 amd64_patch (br1, code);
4167 amd64_alu_reg_imm (code, X86_AND, AMD64_RAX, (X86_FP_C0|X86_FP_C1));
4168 amd64_alu_reg_imm (code, X86_CMP, AMD64_RAX, 0);
4169 EMIT_COND_BRANCH (ins, X86_CC_EQ, FALSE);
4172 /* Branch if C013 != 001 */
4173 if (use_sse2 || (cfg->opt & MONO_OPT_FCMOV)) {
4174 EMIT_COND_BRANCH (ins, X86_CC_P, FALSE);
4175 EMIT_COND_BRANCH (ins, X86_CC_GE, FALSE);
4178 amd64_alu_reg_imm (code, X86_CMP, AMD64_RAX, X86_FP_C0);
4179 EMIT_COND_BRANCH (ins, X86_CC_NE, FALSE);
4183 /* Transfer value to the fp stack */
4184 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, 16);
4185 amd64_movsd_membase_reg (code, AMD64_RSP, 0, ins->sreg1);
4186 amd64_fld_membase (code, AMD64_RSP, 0, TRUE);
4188 amd64_push_reg (code, AMD64_RAX);
4190 amd64_fnstsw (code);
4191 amd64_alu_reg_imm (code, X86_AND, AMD64_RAX, 0x4100);
4192 amd64_alu_reg_imm (code, X86_CMP, AMD64_RAX, X86_FP_C0);
4193 amd64_pop_reg (code, AMD64_RAX);
4195 amd64_fstp (code, 0);
4197 EMIT_COND_SYSTEM_EXCEPTION (X86_CC_EQ, FALSE, "ArithmeticException");
4199 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, 16);
4203 code = emit_tls_get (code, ins->dreg, ins->inst_offset);
4206 case OP_MEMORY_BARRIER: {
4207 /* Not needed on amd64 */
4210 case OP_ATOMIC_ADD_I4:
4211 case OP_ATOMIC_ADD_I8: {
4212 int dreg = ins->dreg;
4213 guint32 size = (ins->opcode == OP_ATOMIC_ADD_I4) ? 4 : 8;
4215 if (dreg == ins->inst_basereg)
4218 if (dreg != ins->sreg2)
4219 amd64_mov_reg_reg (code, ins->dreg, ins->sreg2, size);
4221 x86_prefix (code, X86_LOCK_PREFIX);
4222 amd64_xadd_membase_reg (code, ins->inst_basereg, ins->inst_offset, dreg, size);
4224 if (dreg != ins->dreg)
4225 amd64_mov_reg_reg (code, ins->dreg, dreg, size);
4229 case OP_ATOMIC_ADD_NEW_I4:
4230 case OP_ATOMIC_ADD_NEW_I8: {
4231 int dreg = ins->dreg;
4232 guint32 size = (ins->opcode == OP_ATOMIC_ADD_NEW_I4) ? 4 : 8;
4234 if ((dreg == ins->sreg2) || (dreg == ins->inst_basereg))
4237 amd64_mov_reg_reg (code, dreg, ins->sreg2, size);
4238 amd64_prefix (code, X86_LOCK_PREFIX);
4239 amd64_xadd_membase_reg (code, ins->inst_basereg, ins->inst_offset, dreg, size);
4240 /* dreg contains the old value, add with sreg2 value */
4241 amd64_alu_reg_reg_size (code, X86_ADD, dreg, ins->sreg2, size);
4243 if (ins->dreg != dreg)
4244 amd64_mov_reg_reg (code, ins->dreg, dreg, size);
4248 case OP_ATOMIC_EXCHANGE_I4:
4249 case OP_ATOMIC_EXCHANGE_I8: {
4251 int sreg2 = ins->sreg2;
4252 int breg = ins->inst_basereg;
4253 guint32 size = (ins->opcode == OP_ATOMIC_EXCHANGE_I4) ? 4 : 8;
4256 * See http://msdn.microsoft.com/msdnmag/issues/0700/Win32/ for
4257 * an explanation of how this works.
4260 /* cmpxchg uses eax as comperand, need to make sure we can use it
4261 * hack to overcome limits in x86 reg allocator
4262 * (req: dreg == eax and sreg2 != eax and breg != eax)
4264 /* The pushes invalidate rsp */
4265 if ((breg == AMD64_RAX) || (breg == AMD64_RSP)) {
4266 amd64_mov_reg_reg (code, AMD64_R11, breg, 8);
4270 if (ins->dreg != AMD64_RAX)
4271 amd64_push_reg (code, AMD64_RAX);
4273 /* We need the EAX reg for the cmpxchg */
4274 if (ins->sreg2 == AMD64_RAX) {
4275 amd64_push_reg (code, AMD64_RDX);
4276 amd64_mov_reg_reg (code, AMD64_RDX, AMD64_RAX, size);
4280 amd64_mov_reg_membase (code, AMD64_RAX, breg, ins->inst_offset, size);
4282 br [0] = code; amd64_prefix (code, X86_LOCK_PREFIX);
4283 amd64_cmpxchg_membase_reg_size (code, breg, ins->inst_offset, sreg2, size);
4284 br [1] = code; amd64_branch8 (code, X86_CC_NE, -1, FALSE);
4285 amd64_patch (br [1], br [0]);
4287 if (ins->dreg != AMD64_RAX) {
4288 amd64_mov_reg_reg (code, ins->dreg, AMD64_RAX, size);
4289 amd64_pop_reg (code, AMD64_RAX);
4292 if (ins->sreg2 != sreg2)
4293 amd64_pop_reg (code, AMD64_RDX);
4298 g_warning ("unknown opcode %s in %s()\n", mono_inst_name (ins->opcode), __FUNCTION__);
4299 g_assert_not_reached ();
4302 if ((code - cfg->native_code - offset) > max_len) {
4303 g_warning ("wrong maximal instruction length of instruction %s (expected %d, got %ld)",
4304 mono_inst_name (ins->opcode), max_len, code - cfg->native_code - offset);
4305 g_assert_not_reached ();
4311 last_offset = offset;
4316 cfg->code_len = code - cfg->native_code;
4320 mono_arch_register_lowlevel_calls (void)
4325 mono_arch_patch_code (MonoMethod *method, MonoDomain *domain, guint8 *code, MonoJumpInfo *ji, gboolean run_cctors)
4327 MonoJumpInfo *patch_info;
4328 gboolean compile_aot = !run_cctors;
4330 for (patch_info = ji; patch_info; patch_info = patch_info->next) {
4331 unsigned char *ip = patch_info->ip.i + code;
4332 const unsigned char *target;
4334 target = mono_resolve_patch_target (method, domain, code, patch_info, run_cctors);
4337 switch (patch_info->type) {
4338 case MONO_PATCH_INFO_BB:
4339 case MONO_PATCH_INFO_LABEL:
4342 /* No need to patch these */
4347 switch (patch_info->type) {
4348 case MONO_PATCH_INFO_NONE:
4350 case MONO_PATCH_INFO_METHOD_REL:
4351 case MONO_PATCH_INFO_R8:
4352 case MONO_PATCH_INFO_R4:
4353 g_assert_not_reached ();
4355 case MONO_PATCH_INFO_BB:
4362 * Debug code to help track down problems where the target of a near call is
4365 if (amd64_is_near_call (ip)) {
4366 gint64 disp = (guint8*)target - (guint8*)ip;
4368 if (!amd64_is_imm32 (disp)) {
4369 printf ("TYPE: %d\n", patch_info->type);
4370 switch (patch_info->type) {
4371 case MONO_PATCH_INFO_INTERNAL_METHOD:
4372 printf ("V: %s\n", patch_info->data.name);
4374 case MONO_PATCH_INFO_METHOD_JUMP:
4375 case MONO_PATCH_INFO_METHOD:
4376 printf ("V: %s\n", patch_info->data.method->name);
4384 amd64_patch (ip, (gpointer)target);
4389 * This macro is used for testing whenever the unwinder works correctly at every point
4390 * where an async exception can happen.
4392 /* This will generate a SIGSEGV at the given point in the code */
4393 #define async_exc_point(code) do { \
4394 if (mono_inject_async_exc_method && mono_method_desc_full_match (mono_inject_async_exc_method, cfg->method)) { \
4395 if (cfg->arch.async_point_count == mono_inject_async_exc_pos) \
4396 amd64_mov_reg_mem (code, AMD64_RAX, 0, 4); \
4397 cfg->arch.async_point_count ++; \
4402 mono_arch_emit_prolog (MonoCompile *cfg)
4404 MonoMethod *method = cfg->method;
4406 MonoMethodSignature *sig;
4408 int alloc_size, pos, max_offset, i, quad;
4411 gint32 lmf_offset = cfg->arch.lmf_offset;
4413 cfg->code_size = MAX (((MonoMethodNormal *)method)->header->code_size * 4, 10240);
4415 code = cfg->native_code = g_malloc (cfg->code_size);
4417 /* Amount of stack space allocated by register saving code */
4421 * The prolog consists of the following parts:
4423 * - push rbp, mov rbp, rsp
4424 * - save callee saved regs using pushes
4426 * - save lmf if needed
4429 * - save lmf if needed
4430 * - save callee saved regs using moves
4433 async_exc_point (code);
4435 if (!cfg->arch.omit_fp) {
4436 amd64_push_reg (code, AMD64_RBP);
4437 async_exc_point (code);
4438 amd64_mov_reg_reg (code, AMD64_RBP, AMD64_RSP, sizeof (gpointer));
4439 async_exc_point (code);
4442 /* Save callee saved registers */
4443 if (!cfg->arch.omit_fp && !method->save_lmf) {
4444 for (i = 0; i < AMD64_NREG; ++i)
4445 if (AMD64_IS_CALLEE_SAVED_REG (i) && (cfg->used_int_regs & (1 << i))) {
4446 amd64_push_reg (code, i);
4447 pos += sizeof (gpointer);
4448 async_exc_point (code);
4452 alloc_size = ALIGN_TO (cfg->stack_offset, MONO_ARCH_FRAME_ALIGNMENT);
4456 if (cfg->arch.omit_fp)
4458 * On enter, the stack is misaligned by the the pushing of the return
4459 * address. It is either made aligned by the pushing of %rbp, or by
4464 cfg->arch.stack_alloc_size = alloc_size;
4466 /* Allocate stack frame */
4468 /* See mono_emit_stack_alloc */
4469 #if defined(PLATFORM_WIN32) || defined(MONO_ARCH_SIGSEGV_ON_ALTSTACK)
4470 guint32 remaining_size = alloc_size;
4471 while (remaining_size >= 0x1000) {
4472 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, 0x1000);
4473 async_exc_point (code);
4474 amd64_test_membase_reg (code, AMD64_RSP, 0, AMD64_RSP);
4475 remaining_size -= 0x1000;
4477 if (remaining_size) {
4478 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, remaining_size);
4479 async_exc_point (code);
4482 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, alloc_size);
4483 async_exc_point (code);
4487 /* Stack alignment check */
4490 amd64_mov_reg_reg (code, AMD64_RAX, AMD64_RSP, 8);
4491 amd64_alu_reg_imm (code, X86_AND, AMD64_RAX, 0xf);
4492 amd64_alu_reg_imm (code, X86_CMP, AMD64_RAX, 0);
4493 x86_branch8 (code, X86_CC_EQ, 2, FALSE);
4494 amd64_breakpoint (code);
4499 if (method->save_lmf) {
4501 amd64_lea_membase (code, AMD64_R11, AMD64_RIP, 0);
4502 amd64_mov_membase_reg (code, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rip), AMD64_R11, 8);
4504 amd64_mov_membase_reg (code, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, ebp), AMD64_RBP, 8);
4506 amd64_mov_membase_reg (code, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rsp), AMD64_RSP, 8);
4507 /* Skip method (only needed for trampoline LMF frames) */
4508 /* Save callee saved regs */
4509 amd64_mov_membase_reg (code, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rbx), AMD64_RBX, 8);
4510 amd64_mov_membase_reg (code, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r12), AMD64_R12, 8);
4511 amd64_mov_membase_reg (code, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r13), AMD64_R13, 8);
4512 amd64_mov_membase_reg (code, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r14), AMD64_R14, 8);
4513 amd64_mov_membase_reg (code, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r15), AMD64_R15, 8);
4516 /* Save callee saved registers */
4517 if (cfg->arch.omit_fp && !method->save_lmf) {
4518 gint32 save_area_offset = 0;
4520 /* Save caller saved registers after sp is adjusted */
4521 /* The registers are saved at the bottom of the frame */
4522 /* FIXME: Optimize this so the regs are saved at the end of the frame in increasing order */
4523 for (i = 0; i < AMD64_NREG; ++i)
4524 if (AMD64_IS_CALLEE_SAVED_REG (i) && (cfg->used_int_regs & (1 << i))) {
4525 amd64_mov_membase_reg (code, AMD64_RSP, save_area_offset, i, 8);
4526 save_area_offset += 8;
4527 async_exc_point (code);
4531 /* compute max_offset in order to use short forward jumps */
4533 if (cfg->opt & MONO_OPT_BRANCH) {
4534 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
4535 MonoInst *ins = bb->code;
4536 bb->max_offset = max_offset;
4538 if (cfg->prof_options & MONO_PROFILE_COVERAGE)
4540 /* max alignment for loops */
4541 if ((cfg->opt & MONO_OPT_LOOP) && bb_is_loop_start (bb))
4542 max_offset += LOOP_ALIGNMENT;
4545 if (ins->opcode == OP_LABEL)
4546 ins->inst_c1 = max_offset;
4548 max_offset += ((guint8 *)ins_get_spec (ins->opcode))[MONO_INST_LEN];
4554 sig = mono_method_signature (method);
4557 cinfo = cfg->arch.cinfo;
4559 if (sig->ret->type != MONO_TYPE_VOID) {
4560 if ((cinfo->ret.storage == ArgInIReg) && (cfg->ret->opcode != OP_REGVAR)) {
4561 /* Save volatile arguments to the stack */
4562 amd64_mov_membase_reg (code, cfg->ret->inst_basereg, cfg->ret->inst_offset, cinfo->ret.reg, 8);
4566 /* Keep this in sync with emit_load_volatile_arguments */
4567 for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
4568 ArgInfo *ainfo = cinfo->args + i;
4569 gint32 stack_offset;
4571 inst = cfg->args [i];
4573 if (sig->hasthis && (i == 0))
4574 arg_type = &mono_defaults.object_class->byval_arg;
4576 arg_type = sig->params [i - sig->hasthis];
4578 stack_offset = ainfo->offset + ARGS_OFFSET;
4580 /* Save volatile arguments to the stack */
4581 if (inst->opcode != OP_REGVAR) {
4582 switch (ainfo->storage) {
4588 if (stack_offset & 0x1)
4590 else if (stack_offset & 0x2)
4592 else if (stack_offset & 0x4)
4597 amd64_mov_membase_reg (code, inst->inst_basereg, inst->inst_offset, ainfo->reg, size);
4600 case ArgInFloatSSEReg:
4601 amd64_movss_membase_reg (code, inst->inst_basereg, inst->inst_offset, ainfo->reg);
4603 case ArgInDoubleSSEReg:
4604 amd64_movsd_membase_reg (code, inst->inst_basereg, inst->inst_offset, ainfo->reg);
4606 case ArgValuetypeInReg:
4607 for (quad = 0; quad < 2; quad ++) {
4608 switch (ainfo->pair_storage [quad]) {
4610 amd64_mov_membase_reg (code, inst->inst_basereg, inst->inst_offset + (quad * sizeof (gpointer)), ainfo->pair_regs [quad], sizeof (gpointer));
4612 case ArgInFloatSSEReg:
4613 amd64_movss_membase_reg (code, inst->inst_basereg, inst->inst_offset + (quad * sizeof (gpointer)), ainfo->pair_regs [quad]);
4615 case ArgInDoubleSSEReg:
4616 amd64_movsd_membase_reg (code, inst->inst_basereg, inst->inst_offset + (quad * sizeof (gpointer)), ainfo->pair_regs [quad]);
4621 g_assert_not_reached ();
4630 if (inst->opcode == OP_REGVAR) {
4631 /* Argument allocated to (non-volatile) register */
4632 switch (ainfo->storage) {
4634 amd64_mov_reg_reg (code, inst->dreg, ainfo->reg, 8);
4637 amd64_mov_reg_membase (code, inst->dreg, AMD64_RBP, ARGS_OFFSET + ainfo->offset, 8);
4640 g_assert_not_reached ();
4645 /* Might need to attach the thread to the JIT or change the domain for the callback */
4646 if (method->wrapper_type == MONO_WRAPPER_NATIVE_TO_MANAGED) {
4647 guint64 domain = (guint64)cfg->domain;
4650 * The call might clobber argument registers, but they are already
4651 * saved to the stack/global regs.
4653 if (appdomain_tls_offset != -1 && lmf_tls_offset != -1) {
4654 guint8 *buf, *no_domain_branch;
4656 code = emit_tls_get (code, AMD64_RAX, appdomain_tls_offset);
4657 if ((domain >> 32) == 0)
4658 amd64_mov_reg_imm_size (code, AMD64_ARG_REG1, domain, 4);
4660 amd64_mov_reg_imm_size (code, AMD64_ARG_REG1, domain, 8);
4661 amd64_alu_reg_reg (code, X86_CMP, AMD64_RAX, AMD64_ARG_REG1);
4662 no_domain_branch = code;
4663 x86_branch8 (code, X86_CC_NE, 0, 0);
4664 code = emit_tls_get ( code, AMD64_RAX, lmf_addr_tls_offset);
4665 amd64_test_reg_reg (code, AMD64_RAX, AMD64_RAX);
4667 x86_branch8 (code, X86_CC_NE, 0, 0);
4668 amd64_patch (no_domain_branch, code);
4669 code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD, (gpointer)"mono_jit_thread_attach");
4670 amd64_patch (buf, code);
4672 g_assert (!cfg->compile_aot);
4673 if ((domain >> 32) == 0)
4674 amd64_mov_reg_imm_size (code, AMD64_ARG_REG1, domain, 4);
4676 amd64_mov_reg_imm_size (code, AMD64_ARG_REG1, domain, 8);
4677 code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD, (gpointer)"mono_jit_thread_attach");
4681 if (method->save_lmf) {
4682 if ((lmf_tls_offset != -1) && !optimize_for_xen) {
4684 * Optimized version which uses the mono_lmf TLS variable instead of indirection
4685 * through the mono_lmf_addr TLS variable.
4687 /* %rax = previous_lmf */
4688 x86_prefix (code, X86_FS_PREFIX);
4689 amd64_mov_reg_mem (code, AMD64_RAX, lmf_tls_offset, 8);
4691 /* Save previous_lmf */
4692 amd64_mov_membase_reg (code, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, previous_lmf), AMD64_RAX, 8);
4694 amd64_lea_membase (code, AMD64_R11, cfg->frame_reg, lmf_offset);
4695 x86_prefix (code, X86_FS_PREFIX);
4696 amd64_mov_mem_reg (code, lmf_tls_offset, AMD64_R11, 8);
4698 if (lmf_addr_tls_offset != -1) {
4699 /* Load lmf quicky using the FS register */
4700 code = emit_tls_get (code, AMD64_RAX, lmf_addr_tls_offset);
4704 * The call might clobber argument registers, but they are already
4705 * saved to the stack/global regs.
4707 code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD,
4708 (gpointer)"mono_get_lmf_addr");
4712 amd64_mov_membase_reg (code, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, lmf_addr), AMD64_RAX, 8);
4713 /* Save previous_lmf */
4714 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RAX, 0, 8);
4715 amd64_mov_membase_reg (code, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, previous_lmf), AMD64_R11, 8);
4717 amd64_lea_membase (code, AMD64_R11, cfg->frame_reg, lmf_offset);
4718 amd64_mov_membase_reg (code, AMD64_RAX, 0, AMD64_R11, 8);
4722 if (mono_jit_trace_calls != NULL && mono_trace_eval (method))
4723 code = mono_arch_instrument_prolog (cfg, mono_trace_enter_method, code, TRUE);
4725 cfg->code_len = code - cfg->native_code;
4727 g_assert (cfg->code_len < cfg->code_size);
4733 mono_arch_emit_epilog (MonoCompile *cfg)
4735 MonoMethod *method = cfg->method;
4738 int max_epilog_size = 16;
4740 gint32 lmf_offset = cfg->arch.lmf_offset;
4742 if (cfg->method->save_lmf)
4743 max_epilog_size += 256;
4745 if (mono_jit_trace_calls != NULL)
4746 max_epilog_size += 50;
4748 if (cfg->prof_options & MONO_PROFILE_ENTER_LEAVE)
4749 max_epilog_size += 50;
4751 max_epilog_size += (AMD64_NREG * 2);
4753 while (cfg->code_len + max_epilog_size > (cfg->code_size - 16)) {
4754 cfg->code_size *= 2;
4755 cfg->native_code = g_realloc (cfg->native_code, cfg->code_size);
4756 mono_jit_stats.code_reallocs++;
4759 code = cfg->native_code + cfg->code_len;
4761 if (mono_jit_trace_calls != NULL && mono_trace_eval (method))
4762 code = mono_arch_instrument_epilog (cfg, mono_trace_leave_method, code, TRUE);
4764 /* the code restoring the registers must be kept in sync with OP_JMP */
4767 if (method->save_lmf) {
4768 if ((lmf_tls_offset != -1) && !optimize_for_xen) {
4770 * Optimized version which uses the mono_lmf TLS variable instead of indirection
4771 * through the mono_lmf_addr TLS variable.
4773 /* reg = previous_lmf */
4774 amd64_mov_reg_membase (code, AMD64_R11, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, previous_lmf), 8);
4775 x86_prefix (code, X86_FS_PREFIX);
4776 amd64_mov_mem_reg (code, lmf_tls_offset, AMD64_R11, 8);
4778 /* Restore previous lmf */
4779 amd64_mov_reg_membase (code, AMD64_RCX, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, previous_lmf), 8);
4780 amd64_mov_reg_membase (code, AMD64_R11, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, lmf_addr), 8);
4781 amd64_mov_membase_reg (code, AMD64_R11, 0, AMD64_RCX, 8);
4784 /* Restore caller saved regs */
4785 if (cfg->used_int_regs & (1 << AMD64_RBP)) {
4786 amd64_mov_reg_membase (code, AMD64_RBP, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, ebp), 8);
4788 if (cfg->used_int_regs & (1 << AMD64_RBX)) {
4789 amd64_mov_reg_membase (code, AMD64_RBX, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rbx), 8);
4791 if (cfg->used_int_regs & (1 << AMD64_R12)) {
4792 amd64_mov_reg_membase (code, AMD64_R12, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r12), 8);
4794 if (cfg->used_int_regs & (1 << AMD64_R13)) {
4795 amd64_mov_reg_membase (code, AMD64_R13, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r13), 8);
4797 if (cfg->used_int_regs & (1 << AMD64_R14)) {
4798 amd64_mov_reg_membase (code, AMD64_R14, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r14), 8);
4800 if (cfg->used_int_regs & (1 << AMD64_R15)) {
4801 amd64_mov_reg_membase (code, AMD64_R15, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r15), 8);
4805 if (cfg->arch.omit_fp) {
4806 gint32 save_area_offset = 0;
4808 for (i = 0; i < AMD64_NREG; ++i)
4809 if (AMD64_IS_CALLEE_SAVED_REG (i) && (cfg->used_int_regs & (1 << i))) {
4810 amd64_mov_reg_membase (code, i, AMD64_RSP, save_area_offset, 8);
4811 save_area_offset += 8;
4815 for (i = 0; i < AMD64_NREG; ++i)
4816 if (AMD64_IS_CALLEE_SAVED_REG (i) && (cfg->used_int_regs & (1 << i)))
4817 pos -= sizeof (gpointer);
4820 if (pos == - sizeof (gpointer)) {
4821 /* Only one register, so avoid lea */
4822 for (i = AMD64_NREG - 1; i > 0; --i)
4823 if (AMD64_IS_CALLEE_SAVED_REG (i) && (cfg->used_int_regs & (1 << i))) {
4824 amd64_mov_reg_membase (code, i, AMD64_RBP, pos, 8);
4828 amd64_lea_membase (code, AMD64_RSP, AMD64_RBP, pos);
4830 /* Pop registers in reverse order */
4831 for (i = AMD64_NREG - 1; i > 0; --i)
4832 if (AMD64_IS_CALLEE_SAVED_REG (i) && (cfg->used_int_regs & (1 << i))) {
4833 amd64_pop_reg (code, i);
4840 /* Load returned vtypes into registers if needed */
4841 cinfo = cfg->arch.cinfo;
4842 if (cinfo->ret.storage == ArgValuetypeInReg) {
4843 ArgInfo *ainfo = &cinfo->ret;
4844 MonoInst *inst = cfg->ret;
4846 for (quad = 0; quad < 2; quad ++) {
4847 switch (ainfo->pair_storage [quad]) {
4849 amd64_mov_reg_membase (code, ainfo->pair_regs [quad], inst->inst_basereg, inst->inst_offset + (quad * sizeof (gpointer)), sizeof (gpointer));
4851 case ArgInFloatSSEReg:
4852 amd64_movss_reg_membase (code, ainfo->pair_regs [quad], inst->inst_basereg, inst->inst_offset + (quad * sizeof (gpointer)));
4854 case ArgInDoubleSSEReg:
4855 amd64_movsd_reg_membase (code, ainfo->pair_regs [quad], inst->inst_basereg, inst->inst_offset + (quad * sizeof (gpointer)));
4860 g_assert_not_reached ();
4865 if (cfg->arch.omit_fp) {
4866 if (cfg->arch.stack_alloc_size)
4867 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, cfg->arch.stack_alloc_size);
4871 async_exc_point (code);
4874 cfg->code_len = code - cfg->native_code;
4876 g_assert (cfg->code_len < cfg->code_size);
4878 if (cfg->arch.omit_fp) {
4880 * Encode the stack size into used_int_regs so the exception handler
4883 g_assert (cfg->arch.stack_alloc_size < (1 << 16));
4884 cfg->used_int_regs |= (1 << 31) | (cfg->arch.stack_alloc_size << 16);
4889 mono_arch_emit_exceptions (MonoCompile *cfg)
4891 MonoJumpInfo *patch_info;
4894 MonoClass *exc_classes [16];
4895 guint8 *exc_throw_start [16], *exc_throw_end [16];
4896 guint32 code_size = 0;
4898 /* Compute needed space */
4899 for (patch_info = cfg->patch_info; patch_info; patch_info = patch_info->next) {
4900 if (patch_info->type == MONO_PATCH_INFO_EXC)
4902 if (patch_info->type == MONO_PATCH_INFO_R8)
4903 code_size += 8 + 15; /* sizeof (double) + alignment */
4904 if (patch_info->type == MONO_PATCH_INFO_R4)
4905 code_size += 4 + 15; /* sizeof (float) + alignment */
4908 while (cfg->code_len + code_size > (cfg->code_size - 16)) {
4909 cfg->code_size *= 2;
4910 cfg->native_code = g_realloc (cfg->native_code, cfg->code_size);
4911 mono_jit_stats.code_reallocs++;
4914 code = cfg->native_code + cfg->code_len;
4916 /* add code to raise exceptions */
4918 for (patch_info = cfg->patch_info; patch_info; patch_info = patch_info->next) {
4919 switch (patch_info->type) {
4920 case MONO_PATCH_INFO_EXC: {
4921 MonoClass *exc_class;
4925 amd64_patch (patch_info->ip.i + cfg->native_code, code);
4927 exc_class = mono_class_from_name (mono_defaults.corlib, "System", patch_info->data.name);
4928 g_assert (exc_class);
4929 throw_ip = patch_info->ip.i;
4931 //x86_breakpoint (code);
4932 /* Find a throw sequence for the same exception class */
4933 for (i = 0; i < nthrows; ++i)
4934 if (exc_classes [i] == exc_class)
4937 amd64_mov_reg_imm (code, AMD64_ARG_REG2, (exc_throw_end [i] - cfg->native_code) - throw_ip);
4938 x86_jump_code (code, exc_throw_start [i]);
4939 patch_info->type = MONO_PATCH_INFO_NONE;
4943 amd64_mov_reg_imm_size (code, AMD64_ARG_REG2, 0xf0f0f0f0, 4);
4947 exc_classes [nthrows] = exc_class;
4948 exc_throw_start [nthrows] = code;
4950 amd64_mov_reg_imm (code, AMD64_ARG_REG1, exc_class->type_token);
4951 patch_info->data.name = "mono_arch_throw_corlib_exception";
4952 patch_info->type = MONO_PATCH_INFO_INTERNAL_METHOD;
4953 patch_info->ip.i = code - cfg->native_code;
4955 code = emit_call_body (cfg, code, patch_info->type, patch_info->data.name);
4957 amd64_mov_reg_imm (buf, AMD64_ARG_REG2, (code - cfg->native_code) - throw_ip);
4962 exc_throw_end [nthrows] = code;
4974 /* Handle relocations with RIP relative addressing */
4975 for (patch_info = cfg->patch_info; patch_info; patch_info = patch_info->next) {
4976 gboolean remove = FALSE;
4978 switch (patch_info->type) {
4979 case MONO_PATCH_INFO_R8:
4980 case MONO_PATCH_INFO_R4: {
4984 /* The SSE opcodes require a 16 byte alignment */
4985 code = (guint8*)ALIGN_TO (code, 16);
4987 code = (guint8*)ALIGN_TO (code, 8);
4990 pos = cfg->native_code + patch_info->ip.i;
4994 if (IS_REX (pos [1]))
4995 *(guint32*)(pos + 5) = (guint8*)code - pos - 9;
4997 *(guint32*)(pos + 4) = (guint8*)code - pos - 8;
4999 *(guint32*)(pos + 3) = (guint8*)code - pos - 7;
5002 if (patch_info->type == MONO_PATCH_INFO_R8) {
5003 *(double*)code = *(double*)patch_info->data.target;
5004 code += sizeof (double);
5006 *(float*)code = *(float*)patch_info->data.target;
5007 code += sizeof (float);
5018 if (patch_info == cfg->patch_info)
5019 cfg->patch_info = patch_info->next;
5023 for (tmp = cfg->patch_info; tmp->next != patch_info; tmp = tmp->next)
5025 tmp->next = patch_info->next;
5030 cfg->code_len = code - cfg->native_code;
5032 g_assert (cfg->code_len < cfg->code_size);
5037 mono_arch_instrument_prolog (MonoCompile *cfg, void *func, void *p, gboolean enable_arguments)
5040 CallInfo *cinfo = NULL;
5041 MonoMethodSignature *sig;
5043 int i, n, stack_area = 0;
5045 /* Keep this in sync with mono_arch_get_argument_info */
5047 if (enable_arguments) {
5048 /* Allocate a new area on the stack and save arguments there */
5049 sig = mono_method_signature (cfg->method);
5051 cinfo = get_call_info (cfg, cfg->mempool, sig, FALSE);
5053 n = sig->param_count + sig->hasthis;
5055 stack_area = ALIGN_TO (n * 8, 16);
5057 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, stack_area);
5059 for (i = 0; i < n; ++i) {
5060 inst = cfg->args [i];
5062 if (inst->opcode == OP_REGVAR)
5063 amd64_mov_membase_reg (code, AMD64_RSP, (i * 8), inst->dreg, 8);
5065 amd64_mov_reg_membase (code, AMD64_R11, inst->inst_basereg, inst->inst_offset, 8);
5066 amd64_mov_membase_reg (code, AMD64_RSP, (i * 8), AMD64_R11, 8);
5071 mono_add_patch_info (cfg, code-cfg->native_code, MONO_PATCH_INFO_METHODCONST, cfg->method);
5072 amd64_set_reg_template (code, AMD64_ARG_REG1);
5073 amd64_mov_reg_reg (code, AMD64_ARG_REG2, AMD64_RSP, 8);
5074 code = emit_call (cfg, code, MONO_PATCH_INFO_ABS, (gpointer)func);
5076 if (enable_arguments)
5077 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, stack_area);
5091 mono_arch_instrument_epilog (MonoCompile *cfg, void *func, void *p, gboolean enable_arguments)
5094 int save_mode = SAVE_NONE;
5095 MonoMethod *method = cfg->method;
5096 int rtype = mono_type_get_underlying_type (mono_method_signature (method)->ret)->type;
5099 case MONO_TYPE_VOID:
5100 /* special case string .ctor icall */
5101 if (strcmp (".ctor", method->name) && method->klass == mono_defaults.string_class)
5102 save_mode = SAVE_EAX;
5104 save_mode = SAVE_NONE;
5108 save_mode = SAVE_EAX;
5112 save_mode = SAVE_XMM;
5114 case MONO_TYPE_GENERICINST:
5115 if (!mono_type_generic_inst_is_valuetype (mono_method_signature (method)->ret)) {
5116 save_mode = SAVE_EAX;
5120 case MONO_TYPE_VALUETYPE:
5121 save_mode = SAVE_STRUCT;
5124 save_mode = SAVE_EAX;
5128 /* Save the result and copy it into the proper argument register */
5129 switch (save_mode) {
5131 amd64_push_reg (code, AMD64_RAX);
5133 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, 8);
5134 if (enable_arguments)
5135 amd64_mov_reg_reg (code, AMD64_ARG_REG2, AMD64_RAX, 8);
5139 if (enable_arguments)
5140 amd64_mov_reg_imm (code, AMD64_ARG_REG2, 0);
5143 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, 8);
5144 amd64_movsd_membase_reg (code, AMD64_RSP, 0, AMD64_XMM0);
5146 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, 8);
5148 * The result is already in the proper argument register so no copying
5155 g_assert_not_reached ();
5158 /* Set %al since this is a varargs call */
5159 if (save_mode == SAVE_XMM)
5160 amd64_mov_reg_imm (code, AMD64_RAX, 1);
5162 amd64_mov_reg_imm (code, AMD64_RAX, 0);
5164 mono_add_patch_info (cfg, code-cfg->native_code, MONO_PATCH_INFO_METHODCONST, method);
5165 amd64_set_reg_template (code, AMD64_ARG_REG1);
5166 code = emit_call (cfg, code, MONO_PATCH_INFO_ABS, (gpointer)func);
5168 /* Restore result */
5169 switch (save_mode) {
5171 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, 8);
5172 amd64_pop_reg (code, AMD64_RAX);
5178 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, 8);
5179 amd64_movsd_reg_membase (code, AMD64_XMM0, AMD64_RSP, 0);
5180 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, 8);
5185 g_assert_not_reached ();
5192 mono_arch_flush_icache (guint8 *code, gint size)
5198 mono_arch_flush_register_windows (void)
5203 mono_arch_is_inst_imm (gint64 imm)
5205 return amd64_is_imm32 (imm);
5209 * Determine whenever the trap whose info is in SIGINFO is caused by
5213 mono_arch_is_int_overflow (void *sigctx, void *info)
5220 mono_arch_sigctx_to_monoctx (sigctx, &ctx);
5222 rip = (guint8*)ctx.rip;
5224 if (IS_REX (rip [0])) {
5225 reg = amd64_rex_b (rip [0]);
5231 if ((rip [0] == 0xf7) && (x86_modrm_mod (rip [1]) == 0x3) && (x86_modrm_reg (rip [1]) == 0x7)) {
5233 reg += x86_modrm_rm (rip [1]);
5273 g_assert_not_reached ();
5285 mono_arch_get_patch_offset (guint8 *code)
5291 mono_breakpoint_clean_code (guint8 *code, guint8 *buf, int size)
5294 gboolean can_write = TRUE;
5295 memcpy (buf, code, size);
5296 for (i = 0; i < MONO_BREAKPOINT_ARRAY_SIZE; ++i) {
5297 int idx = mono_breakpoint_info_index [i];
5301 ptr = mono_breakpoint_info [idx].address;
5302 if (ptr >= code && ptr < code + size) {
5303 guint8 saved_byte = mono_breakpoint_info [idx].saved_byte;
5305 /*g_print ("patching %p with 0x%02x (was: 0x%02x)\n", ptr, saved_byte, buf [ptr - code]);*/
5306 buf [ptr - code] = saved_byte;
5313 mono_arch_get_vcall_slot (guint8 *code, gpointer *regs, int *displacement)
5320 mono_breakpoint_clean_code (code - 10, buf, sizeof (buf));
5325 /* go to the start of the call instruction
5327 * address_byte = (m << 6) | (o << 3) | reg
5328 * call opcode: 0xff address_byte displacement
5330 * 0xff m=2,o=2 imm32
5335 * A given byte sequence can match more than case here, so we have to be
5336 * really careful about the ordering of the cases. Longer sequences
5339 #ifdef MONO_ARCH_HAVE_IMT
5340 if ((code [-2] == 0x41) && (code [-1] == 0xbb) && (code [4] == 0xff) && (x86_modrm_mod (code [5]) == 1) && (x86_modrm_reg (code [5]) == 2) && ((signed char)code [6] < 0)) {
5341 /* IMT-based interface calls: with MONO_ARCH_IMT_REG == r11
5342 * 41 bb 14 f8 28 08 mov $0x828f814,%r11d
5343 * ff 50 fc call *0xfffffffc(%rax)
5345 reg = amd64_modrm_rm (code [5]);
5346 disp = (signed char)code [6];
5347 /* R10 is clobbered by the IMT thunk code */
5348 g_assert (reg != AMD64_R10);
5354 else if ((code [-1] == 0x8b) && (amd64_modrm_mod (code [0]) == 0x2) && (code [5] == 0xff) && (amd64_modrm_reg (code [6]) == 0x2) && (amd64_modrm_mod (code [6]) == 0x0)) {
5356 * This is a interface call
5357 * 48 8b 80 f0 e8 ff ff mov 0xffffffffffffe8f0(%rax),%rax
5358 * ff 10 callq *(%rax)
5360 if (IS_REX (code [4]))
5362 reg = amd64_modrm_rm (code [6]);
5364 /* R10 is clobbered by the IMT thunk code */
5365 g_assert (reg != AMD64_R10);
5366 } else if ((code [0] == 0x41) && (code [1] == 0xff) && (code [2] == 0x15)) {
5367 /* call OFFSET(%rip) */
5368 disp = *(guint32*)(code + 3);
5369 return (gpointer*)(code + disp + 7);
5371 else if ((code [1] == 0xff) && (amd64_modrm_reg (code [2]) == 0x2) && (amd64_modrm_mod (code [2]) == 0x2)) {
5372 /* call *[reg+disp32] */
5373 if (IS_REX (code [0]))
5375 reg = amd64_modrm_rm (code [2]);
5376 disp = *(gint32*)(code + 3);
5377 /* R10 is clobbered by the IMT thunk code */
5378 g_assert (reg != AMD64_R10);
5380 else if (code [2] == 0xe8) {
5384 else if (IS_REX (code [4]) && (code [5] == 0xff) && (amd64_modrm_reg (code [6]) == 0x2) && (amd64_modrm_mod (code [6]) == 0x3)) {
5388 else if ((code [4] == 0xff) && (amd64_modrm_reg (code [5]) == 0x2) && (amd64_modrm_mod (code [5]) == 0x1)) {
5389 /* call *[reg+disp8] */
5390 if (IS_REX (code [3]))
5392 reg = amd64_modrm_rm (code [5]);
5393 disp = *(gint8*)(code + 6);
5394 //printf ("B: [%%r%d+0x%x]\n", reg, disp);
5396 else if ((code [5] == 0xff) && (amd64_modrm_reg (code [6]) == 0x2) && (amd64_modrm_mod (code [6]) == 0x0)) {
5398 * This is a interface call: should check the above code can't catch it earlier
5399 * 8b 40 30 mov 0x30(%eax),%eax
5400 * ff 10 call *(%eax)
5402 if (IS_REX (code [4]))
5404 reg = amd64_modrm_rm (code [6]);
5408 g_assert_not_reached ();
5410 reg += amd64_rex_b (rex);
5412 /* R11 is clobbered by the trampoline code */
5413 g_assert (reg != AMD64_R11);
5415 *displacement = disp;
5420 mono_arch_get_vcall_slot_addr (guint8* code, gpointer *regs)
5424 vt = mono_arch_get_vcall_slot (code, regs, &displacement);
5427 return (gpointer*)((char*)vt + displacement);
5431 mono_arch_get_this_arg_from_call (MonoMethodSignature *sig, gssize *regs, guint8 *code)
5433 if (MONO_TYPE_ISSTRUCT (sig->ret))
5434 return (gpointer)regs [AMD64_ARG_REG2];
5436 return (gpointer)regs [AMD64_ARG_REG1];
5439 #define MAX_ARCH_DELEGATE_PARAMS 10
5442 mono_arch_get_delegate_invoke_impl (MonoMethodSignature *sig, gboolean has_target)
5444 guint8 *code, *start;
5447 if (sig->param_count > MAX_ARCH_DELEGATE_PARAMS)
5450 /* FIXME: Support more cases */
5451 if (MONO_TYPE_ISSTRUCT (sig->ret))
5455 static guint8* cached = NULL;
5456 mono_mini_arch_lock ();
5458 mono_mini_arch_unlock ();
5462 start = code = mono_global_codeman_reserve (64);
5464 /* Replace the this argument with the target */
5465 amd64_mov_reg_reg (code, AMD64_RAX, AMD64_ARG_REG1, 8);
5466 amd64_mov_reg_membase (code, AMD64_ARG_REG1, AMD64_RAX, G_STRUCT_OFFSET (MonoDelegate, target), 8);
5467 amd64_jump_membase (code, AMD64_RAX, G_STRUCT_OFFSET (MonoDelegate, method_ptr));
5469 g_assert ((code - start) < 64);
5472 mono_debug_add_delegate_trampoline (start, code - start);
5473 mono_mini_arch_unlock ();
5475 static guint8* cache [MAX_ARCH_DELEGATE_PARAMS + 1] = {NULL};
5476 for (i = 0; i < sig->param_count; ++i)
5477 if (!mono_is_regsize_var (sig->params [i]))
5479 if (sig->param_count > 4)
5482 mono_mini_arch_lock ();
5483 code = cache [sig->param_count];
5485 mono_mini_arch_unlock ();
5489 start = code = mono_global_codeman_reserve (64);
5491 if (sig->param_count == 0) {
5492 amd64_jump_membase (code, AMD64_ARG_REG1, G_STRUCT_OFFSET (MonoDelegate, method_ptr));
5494 /* We have to shift the arguments left */
5495 amd64_mov_reg_reg (code, AMD64_RAX, AMD64_ARG_REG1, 8);
5496 for (i = 0; i < sig->param_count; ++i)
5497 amd64_mov_reg_reg (code, param_regs [i], param_regs [i + 1], 8);
5499 amd64_jump_membase (code, AMD64_RAX, G_STRUCT_OFFSET (MonoDelegate, method_ptr));
5501 g_assert ((code - start) < 64);
5503 cache [sig->param_count] = start;
5505 mono_debug_add_delegate_trampoline (start, code - start);
5506 mono_mini_arch_unlock ();
5513 * Support for fast access to the thread-local lmf structure using the GS
5514 * segment register on NPTL + kernel 2.6.x.
5517 static gboolean tls_offset_inited = FALSE;
5520 mono_arch_setup_jit_tls_data (MonoJitTlsData *tls)
5522 if (!tls_offset_inited) {
5523 tls_offset_inited = TRUE;
5525 optimize_for_xen = access ("/proc/xen", F_OK) == 0;
5527 appdomain_tls_offset = mono_domain_get_tls_offset ();
5528 lmf_tls_offset = mono_get_lmf_tls_offset ();
5529 lmf_addr_tls_offset = mono_get_lmf_addr_tls_offset ();
5530 thread_tls_offset = mono_thread_get_tls_offset ();
5535 mono_arch_free_jit_tls_data (MonoJitTlsData *tls)
5540 mono_arch_emit_this_vret_args (MonoCompile *cfg, MonoCallInst *inst, int this_reg, int this_type, int vt_reg)
5542 MonoCallInst *call = (MonoCallInst*)inst;
5543 CallInfo * cinfo = get_call_info (cfg, cfg->mempool, inst->signature, FALSE);
5548 if (cinfo->ret.storage == ArgValuetypeInReg) {
5550 * The valuetype is in RAX:RDX after the call, need to be copied to
5551 * the stack. Push the address here, so the call instruction can
5554 MONO_INST_NEW (cfg, vtarg, OP_X86_PUSH);
5555 vtarg->sreg1 = vt_reg;
5556 mono_bblock_add_inst (cfg->cbb, vtarg);
5559 MONO_EMIT_NEW_BIALU_IMM (cfg, OP_SUB_IMM, X86_ESP, X86_ESP, 8);
5562 MONO_INST_NEW (cfg, vtarg, OP_MOVE);
5563 vtarg->sreg1 = vt_reg;
5564 vtarg->dreg = mono_regstate_next_int (cfg->rs);
5565 mono_bblock_add_inst (cfg->cbb, vtarg);
5567 mono_call_inst_add_outarg_reg (cfg, call, vtarg->dreg, cinfo->ret.reg, FALSE);
5571 /* add the this argument */
5572 if (this_reg != -1) {
5574 MONO_INST_NEW (cfg, this, OP_MOVE);
5575 this->type = this_type;
5576 this->sreg1 = this_reg;
5577 this->dreg = mono_regstate_next_int (cfg->rs);
5578 mono_bblock_add_inst (cfg->cbb, this);
5580 mono_call_inst_add_outarg_reg (cfg, call, this->dreg, cinfo->args [0].reg, FALSE);
5584 #ifdef MONO_ARCH_HAVE_IMT
5586 #define CMP_SIZE (6 + 1)
5587 #define CMP_REG_REG_SIZE (4 + 1)
5588 #define BR_SMALL_SIZE 2
5589 #define BR_LARGE_SIZE 6
5590 #define MOV_REG_IMM_SIZE 10
5591 #define MOV_REG_IMM_32BIT_SIZE 6
5592 #define JUMP_REG_SIZE (2 + 1)
5595 imt_branch_distance (MonoIMTCheckItem **imt_entries, int start, int target)
5597 int i, distance = 0;
5598 for (i = start; i < target; ++i)
5599 distance += imt_entries [i]->chunk_size;
5604 * LOCKING: called with the domain lock held
5607 mono_arch_build_imt_thunk (MonoVTable *vtable, MonoDomain *domain, MonoIMTCheckItem **imt_entries, int count)
5611 guint8 *code, *start;
5612 gboolean vtable_is_32bit = ((gsize)(vtable) == (gsize)(int)(gsize)(vtable));
5614 for (i = 0; i < count; ++i) {
5615 MonoIMTCheckItem *item = imt_entries [i];
5616 if (item->is_equals) {
5617 if (item->check_target_idx) {
5618 if (!item->compare_done) {
5619 if (amd64_is_imm32 (item->method))
5620 item->chunk_size += CMP_SIZE;
5622 item->chunk_size += MOV_REG_IMM_SIZE + CMP_REG_REG_SIZE;
5624 if (vtable_is_32bit)
5625 item->chunk_size += MOV_REG_IMM_32BIT_SIZE;
5627 item->chunk_size += MOV_REG_IMM_SIZE;
5628 item->chunk_size += BR_SMALL_SIZE + JUMP_REG_SIZE;
5630 if (vtable_is_32bit)
5631 item->chunk_size += MOV_REG_IMM_32BIT_SIZE;
5633 item->chunk_size += MOV_REG_IMM_SIZE;
5634 item->chunk_size += JUMP_REG_SIZE;
5635 /* with assert below:
5636 * item->chunk_size += CMP_SIZE + BR_SMALL_SIZE + 1;
5640 if (amd64_is_imm32 (item->method))
5641 item->chunk_size += CMP_SIZE;
5643 item->chunk_size += MOV_REG_IMM_SIZE + CMP_REG_REG_SIZE;
5644 item->chunk_size += BR_LARGE_SIZE;
5645 imt_entries [item->check_target_idx]->compare_done = TRUE;
5647 size += item->chunk_size;
5649 code = mono_code_manager_reserve (domain->code_mp, size);
5651 for (i = 0; i < count; ++i) {
5652 MonoIMTCheckItem *item = imt_entries [i];
5653 item->code_target = code;
5654 if (item->is_equals) {
5655 if (item->check_target_idx) {
5656 if (!item->compare_done) {
5657 if (amd64_is_imm32 (item->method))
5658 amd64_alu_reg_imm (code, X86_CMP, MONO_ARCH_IMT_REG, (guint32)(gssize)item->method);
5660 amd64_mov_reg_imm (code, AMD64_R10, item->method);
5661 amd64_alu_reg_reg (code, X86_CMP, MONO_ARCH_IMT_REG, AMD64_R10);
5664 item->jmp_code = code;
5665 amd64_branch8 (code, X86_CC_NE, 0, FALSE);
5666 amd64_mov_reg_imm (code, AMD64_R11, & (vtable->vtable [item->vtable_slot]));
5667 amd64_jump_membase (code, AMD64_R11, 0);
5669 /* enable the commented code to assert on wrong method */
5671 if (amd64_is_imm32 (item->method))
5672 amd64_alu_reg_imm (code, X86_CMP, MONO_ARCH_IMT_REG, (guint32)(gssize)item->method);
5674 amd64_mov_reg_imm (code, AMD64_R10, item->method);
5675 amd64_alu_reg_reg (code, X86_CMP, MONO_ARCH_IMT_REG, AMD64_R10);
5677 item->jmp_code = code;
5678 amd64_branch8 (code, X86_CC_NE, 0, FALSE);
5679 amd64_mov_reg_imm (code, AMD64_R11, & (vtable->vtable [item->vtable_slot]));
5680 amd64_jump_membase (code, AMD64_R11, 0);
5681 amd64_patch (item->jmp_code, code);
5682 amd64_breakpoint (code);
5683 item->jmp_code = NULL;
5685 amd64_mov_reg_imm (code, AMD64_R11, & (vtable->vtable [item->vtable_slot]));
5686 amd64_jump_membase (code, AMD64_R11, 0);
5690 if (amd64_is_imm32 (item->method))
5691 amd64_alu_reg_imm (code, X86_CMP, MONO_ARCH_IMT_REG, (guint32)(gssize)item->method);
5693 amd64_mov_reg_imm (code, AMD64_R10, item->method);
5694 amd64_alu_reg_reg (code, X86_CMP, MONO_ARCH_IMT_REG, AMD64_R10);
5696 item->jmp_code = code;
5697 if (x86_is_imm8 (imt_branch_distance (imt_entries, i, item->check_target_idx)))
5698 x86_branch8 (code, X86_CC_GE, 0, FALSE);
5700 x86_branch32 (code, X86_CC_GE, 0, FALSE);
5702 g_assert (code - item->code_target <= item->chunk_size);
5704 /* patch the branches to get to the target items */
5705 for (i = 0; i < count; ++i) {
5706 MonoIMTCheckItem *item = imt_entries [i];
5707 if (item->jmp_code) {
5708 if (item->check_target_idx) {
5709 amd64_patch (item->jmp_code, imt_entries [item->check_target_idx]->code_target);
5714 mono_stats.imt_thunks_size += code - start;
5715 g_assert (code - start <= size);
5721 mono_arch_find_imt_method (gpointer *regs, guint8 *code)
5724 * R11 is clobbered by the trampoline code, so we have to retrieve the method
5726 * 41 bb c0 f7 89 00 mov $0x89f7c0,%r11d
5727 * ff 90 68 ff ff ff callq *0xffffffffffffff68(%rax)
5729 /* Similar to get_vcall_slot_addr () */
5731 /* Find the start of the call instruction */
5733 if ((code [-2] == 0x41) && (code [-1] == 0xbb) && (code [4] == 0xff) && (x86_modrm_mod (code [5]) == 1) && (x86_modrm_reg (code [5]) == 2) && ((signed char)code [6] < 0)) {
5734 /* IMT-based interface calls
5735 * 41 bb 14 f8 28 08 mov $0x828f814,%r11
5736 * ff 50 fc call *0xfffffffc(%rax)
5739 } else if ((code [1] == 0xff) && (amd64_modrm_reg (code [2]) == 0x2) && (amd64_modrm_mod (code [2]) == 0x2)) {
5740 /* call *[reg+disp32] */
5742 } else if ((code [4] == 0xff) && (amd64_modrm_reg (code [5]) == 0x2) && (amd64_modrm_mod (code [5]) == 0x1)) {
5743 /* call *[reg+disp8] */
5746 g_assert_not_reached ();
5748 /* Find the start of the mov instruction */
5750 if (code [0] == 0x49 && code [1] == 0xbb) {
5751 return (MonoMethod*)*(gssize*)(code + 2);
5752 } else if (code [3] == 0x4d && code [4] == 0x8b && code [5] == 0x1d) {
5753 /* mov <OFFSET>(%rip),%r11 */
5754 return (MonoMethod*)*(gssize*)(code + 10 + *(guint32*)(code + 6));
5755 } else if (code [4] == 0x41 && code [5] == 0xbb) {
5756 return (MonoMethod*)(gssize)*(guint32*)(code + 6);
5760 printf ("Unknown call sequence: ");
5761 for (i = -10; i < 20; ++i)
5762 printf ("%x ", code [i]);
5763 g_assert_not_reached ();
5769 mono_arch_find_this_argument (gpointer *regs, MonoMethod *method)
5771 return mono_arch_get_this_arg_from_call (mono_method_signature (method), (gssize*)regs, NULL);
5776 mono_arch_get_inst_for_method (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature *fsig, MonoInst **args)
5778 MonoInst *ins = NULL;
5780 if (cmethod->klass == mono_defaults.math_class) {
5781 if (strcmp (cmethod->name, "Sin") == 0) {
5782 MONO_INST_NEW (cfg, ins, OP_SIN);
5783 ins->inst_i0 = args [0];
5784 } else if (strcmp (cmethod->name, "Cos") == 0) {
5785 MONO_INST_NEW (cfg, ins, OP_COS);
5786 ins->inst_i0 = args [0];
5787 } else if (strcmp (cmethod->name, "Tan") == 0) {
5790 MONO_INST_NEW (cfg, ins, OP_TAN);
5791 ins->inst_i0 = args [0];
5792 } else if (strcmp (cmethod->name, "Atan") == 0) {
5795 MONO_INST_NEW (cfg, ins, OP_ATAN);
5796 ins->inst_i0 = args [0];
5797 } else if (strcmp (cmethod->name, "Sqrt") == 0) {
5798 MONO_INST_NEW (cfg, ins, OP_SQRT);
5799 ins->inst_i0 = args [0];
5800 } else if (strcmp (cmethod->name, "Abs") == 0 && fsig->params [0]->type == MONO_TYPE_R8) {
5801 MONO_INST_NEW (cfg, ins, OP_ABS);
5802 ins->inst_i0 = args [0];
5805 if (cfg->opt & MONO_OPT_CMOV) {
5808 if (strcmp (cmethod->name, "Min") == 0) {
5809 if (fsig->params [0]->type == MONO_TYPE_I4)
5811 else if (fsig->params [0]->type == MONO_TYPE_I8)
5813 } else if (strcmp (cmethod->name, "Max") == 0) {
5814 if (fsig->params [0]->type == MONO_TYPE_I4)
5816 else if (fsig->params [0]->type == MONO_TYPE_I8)
5821 MONO_INST_NEW (cfg, ins, opcode);
5822 ins->inst_i0 = args [0];
5823 ins->inst_i1 = args [1];
5828 /* OP_FREM is not IEEE compatible */
5829 else if (strcmp (cmethod->name, "IEEERemainder") == 0) {
5830 MONO_INST_NEW (cfg, ins, OP_FREM);
5831 ins->inst_i0 = args [0];
5832 ins->inst_i1 = args [1];
5835 } else if(cmethod->klass->image == mono_defaults.corlib &&
5836 (strcmp (cmethod->klass->name_space, "System.Threading") == 0) &&
5837 (strcmp (cmethod->klass->name, "Interlocked") == 0)) {
5839 * Can't implement CompareExchange methods this way since they have
5848 mono_arch_print_tree (MonoInst *tree, int arity)
5853 MonoInst* mono_arch_get_domain_intrinsic (MonoCompile* cfg)
5857 if (appdomain_tls_offset == -1)
5860 MONO_INST_NEW (cfg, ins, OP_TLS_GET);
5861 ins->inst_offset = appdomain_tls_offset;
5865 MonoInst* mono_arch_get_thread_intrinsic (MonoCompile* cfg)
5869 if (thread_tls_offset == -1)
5872 MONO_INST_NEW (cfg, ins, OP_TLS_GET);
5873 ins->inst_offset = thread_tls_offset;