2 * mini-amd64.c: AMD64 backend for the Mono code generator
7 * Paolo Molaro (lupus@ximian.com)
8 * Dietmar Maurer (dietmar@ximian.com)
10 * Zoltan Varga (vargaz@gmail.com)
11 * Johan Lorensson (lateralusx.github@gmail.com)
13 * (C) 2003 Ximian, Inc.
14 * Copyright 2003-2011 Novell, Inc (http://www.novell.com)
15 * Copyright 2011 Xamarin, Inc (http://www.xamarin.com)
16 * Licensed under the MIT license. See LICENSE file in the project root for full license information.
25 #include <mono/metadata/abi-details.h>
26 #include <mono/metadata/appdomain.h>
27 #include <mono/metadata/debug-helpers.h>
28 #include <mono/metadata/threads.h>
29 #include <mono/metadata/profiler-private.h>
30 #include <mono/metadata/mono-debug.h>
31 #include <mono/metadata/gc-internals.h>
32 #include <mono/utils/mono-math.h>
33 #include <mono/utils/mono-mmap.h>
34 #include <mono/utils/mono-memory-model.h>
35 #include <mono/utils/mono-tls.h>
36 #include <mono/utils/mono-hwcap.h>
37 #include <mono/utils/mono-threads.h>
41 #include "mini-amd64.h"
42 #include "cpu-amd64.h"
43 #include "debugger-agent.h"
47 static gboolean optimize_for_xen = TRUE;
49 #define optimize_for_xen 0
52 #define ALIGN_TO(val,align) ((((guint64)val) + ((align) - 1)) & ~((align) - 1))
54 #define IS_IMM32(val) ((((guint64)val) >> 32) == 0)
56 #define IS_REX(inst) (((inst) >= 0x40) && ((inst) <= 0x4f))
59 /* Under windows, the calling convention is never stdcall */
60 #define CALLCONV_IS_STDCALL(call_conv) (FALSE)
62 #define CALLCONV_IS_STDCALL(call_conv) ((call_conv) == MONO_CALL_STDCALL)
65 /* This mutex protects architecture specific caches */
66 #define mono_mini_arch_lock() mono_os_mutex_lock (&mini_arch_mutex)
67 #define mono_mini_arch_unlock() mono_os_mutex_unlock (&mini_arch_mutex)
68 static mono_mutex_t mini_arch_mutex;
70 /* The single step trampoline */
71 static gpointer ss_trampoline;
73 /* The breakpoint trampoline */
74 static gpointer bp_trampoline;
76 /* Offset between fp and the first argument in the callee */
77 #define ARGS_OFFSET 16
78 #define GP_SCRATCH_REG AMD64_R11
81 * AMD64 register usage:
82 * - callee saved registers are used for global register allocation
83 * - %r11 is used for materializing 64 bit constants in opcodes
84 * - the rest is used for local allocation
88 * Floating point comparison results:
98 mono_arch_regname (int reg)
101 case AMD64_RAX: return "%rax";
102 case AMD64_RBX: return "%rbx";
103 case AMD64_RCX: return "%rcx";
104 case AMD64_RDX: return "%rdx";
105 case AMD64_RSP: return "%rsp";
106 case AMD64_RBP: return "%rbp";
107 case AMD64_RDI: return "%rdi";
108 case AMD64_RSI: return "%rsi";
109 case AMD64_R8: return "%r8";
110 case AMD64_R9: return "%r9";
111 case AMD64_R10: return "%r10";
112 case AMD64_R11: return "%r11";
113 case AMD64_R12: return "%r12";
114 case AMD64_R13: return "%r13";
115 case AMD64_R14: return "%r14";
116 case AMD64_R15: return "%r15";
121 static const char * packed_xmmregs [] = {
122 "p:xmm0", "p:xmm1", "p:xmm2", "p:xmm3", "p:xmm4", "p:xmm5", "p:xmm6", "p:xmm7", "p:xmm8",
123 "p:xmm9", "p:xmm10", "p:xmm11", "p:xmm12", "p:xmm13", "p:xmm14", "p:xmm15"
126 static const char * single_xmmregs [] = {
127 "s:xmm0", "s:xmm1", "s:xmm2", "s:xmm3", "s:xmm4", "s:xmm5", "s:xmm6", "s:xmm7", "s:xmm8",
128 "s:xmm9", "s:xmm10", "s:xmm11", "s:xmm12", "s:xmm13", "s:xmm14", "s:xmm15"
132 mono_arch_fregname (int reg)
134 if (reg < AMD64_XMM_NREG)
135 return single_xmmregs [reg];
141 mono_arch_xregname (int reg)
143 if (reg < AMD64_XMM_NREG)
144 return packed_xmmregs [reg];
153 return mono_debug_count ();
159 static inline gboolean
160 amd64_is_near_call (guint8 *code)
163 if ((code [0] >= 0x40) && (code [0] <= 0x4f))
166 return code [0] == 0xe8;
169 static inline gboolean
170 amd64_use_imm32 (gint64 val)
172 if (mini_get_debug_options()->single_imm_size)
175 return amd64_is_imm32 (val);
179 amd64_patch (unsigned char* code, gpointer target)
184 if ((code [0] >= 0x40) && (code [0] <= 0x4f)) {
189 if ((code [0] & 0xf8) == 0xb8) {
190 /* amd64_set_reg_template */
191 *(guint64*)(code + 1) = (guint64)target;
193 else if ((code [0] == 0x8b) && rex && x86_modrm_mod (code [1]) == 0 && x86_modrm_rm (code [1]) == 5) {
194 /* mov 0(%rip), %dreg */
195 *(guint32*)(code + 2) = (guint32)(guint64)target - 7;
197 else if ((code [0] == 0xff) && (code [1] == 0x15)) {
198 /* call *<OFFSET>(%rip) */
199 *(guint32*)(code + 2) = ((guint32)(guint64)target) - 7;
201 else if (code [0] == 0xe8) {
203 gint64 disp = (guint8*)target - (guint8*)code;
204 g_assert (amd64_is_imm32 (disp));
205 x86_patch (code, (unsigned char*)target);
208 x86_patch (code, (unsigned char*)target);
212 mono_amd64_patch (unsigned char* code, gpointer target)
214 amd64_patch (code, target);
217 #define DEBUG(a) if (cfg->verbose_level > 1) a
220 add_general (guint32 *gr, guint32 *stack_size, ArgInfo *ainfo)
222 ainfo->offset = *stack_size;
224 if (*gr >= PARAM_REGS) {
225 ainfo->storage = ArgOnStack;
226 ainfo->arg_size = sizeof (mgreg_t);
227 /* Since the same stack slot size is used for all arg */
228 /* types, it needs to be big enough to hold them all */
229 (*stack_size) += sizeof(mgreg_t);
232 ainfo->storage = ArgInIReg;
233 ainfo->reg = param_regs [*gr];
239 add_float (guint32 *gr, guint32 *stack_size, ArgInfo *ainfo, gboolean is_double)
241 ainfo->offset = *stack_size;
243 if (*gr >= FLOAT_PARAM_REGS) {
244 ainfo->storage = ArgOnStack;
245 ainfo->arg_size = sizeof (mgreg_t);
246 /* Since the same stack slot size is used for both float */
247 /* types, it needs to be big enough to hold them both */
248 (*stack_size) += sizeof(mgreg_t);
251 /* A double register */
253 ainfo->storage = ArgInDoubleSSEReg;
255 ainfo->storage = ArgInFloatSSEReg;
261 typedef enum ArgumentClass {
269 merge_argument_class_from_type (MonoType *type, ArgumentClass class1)
271 ArgumentClass class2 = ARG_CLASS_NO_CLASS;
274 ptype = mini_get_underlying_type (type);
275 switch (ptype->type) {
284 case MONO_TYPE_STRING:
285 case MONO_TYPE_OBJECT:
286 case MONO_TYPE_CLASS:
287 case MONO_TYPE_SZARRAY:
289 case MONO_TYPE_FNPTR:
290 case MONO_TYPE_ARRAY:
293 class2 = ARG_CLASS_INTEGER;
298 class2 = ARG_CLASS_INTEGER;
300 class2 = ARG_CLASS_SSE;
304 case MONO_TYPE_TYPEDBYREF:
305 g_assert_not_reached ();
307 case MONO_TYPE_GENERICINST:
308 if (!mono_type_generic_inst_is_valuetype (ptype)) {
309 class2 = ARG_CLASS_INTEGER;
313 case MONO_TYPE_VALUETYPE: {
314 MonoMarshalType *info = mono_marshal_load_type_info (ptype->data.klass);
317 for (i = 0; i < info->num_fields; ++i) {
319 class2 = merge_argument_class_from_type (info->fields [i].field->type, class2);
324 g_assert_not_reached ();
328 if (class1 == class2)
330 else if (class1 == ARG_CLASS_NO_CLASS)
332 else if ((class1 == ARG_CLASS_MEMORY) || (class2 == ARG_CLASS_MEMORY))
333 class1 = ARG_CLASS_MEMORY;
334 else if ((class1 == ARG_CLASS_INTEGER) || (class2 == ARG_CLASS_INTEGER))
335 class1 = ARG_CLASS_INTEGER;
337 class1 = ARG_CLASS_SSE;
348 * collect_field_info_nested:
350 * Collect field info from KLASS recursively into FIELDS.
353 collect_field_info_nested (MonoClass *klass, GArray *fields_array, int offset, gboolean pinvoke, gboolean unicode)
355 MonoMarshalType *info;
359 info = mono_marshal_load_type_info (klass);
361 for (i = 0; i < info->num_fields; ++i) {
362 if (MONO_TYPE_ISSTRUCT (info->fields [i].field->type)) {
363 collect_field_info_nested (mono_class_from_mono_type (info->fields [i].field->type), fields_array, info->fields [i].offset, pinvoke, unicode);
368 f.type = info->fields [i].field->type;
369 f.size = mono_marshal_type_size (info->fields [i].field->type,
370 info->fields [i].mspec,
371 &align, TRUE, unicode);
372 f.offset = offset + info->fields [i].offset;
373 if (i == info->num_fields - 1 && f.size + f.offset < info->native_size) {
374 /* This can happen with .pack directives eg. 'fixed' arrays */
375 if (MONO_TYPE_IS_PRIMITIVE (f.type)) {
376 /* Replicate the last field to fill out the remaining place, since the code in add_valuetype () needs type information */
377 g_array_append_val (fields_array, f);
378 while (f.size + f.offset < info->native_size) {
380 g_array_append_val (fields_array, f);
383 f.size = info->native_size - f.offset;
384 g_array_append_val (fields_array, f);
387 g_array_append_val (fields_array, f);
393 MonoClassField *field;
396 while ((field = mono_class_get_fields (klass, &iter))) {
397 if (field->type->attrs & FIELD_ATTRIBUTE_STATIC)
399 if (MONO_TYPE_ISSTRUCT (field->type)) {
400 collect_field_info_nested (mono_class_from_mono_type (field->type), fields_array, field->offset - sizeof (MonoObject), pinvoke, unicode);
405 f.type = field->type;
406 f.size = mono_type_size (field->type, &align);
407 f.offset = field->offset - sizeof (MonoObject) + offset;
409 g_array_append_val (fields_array, f);
417 /* Windows x64 ABI can pass/return value types in register of size 1,2,4,8 bytes. */
418 #define MONO_WIN64_VALUE_TYPE_FITS_REG(arg_size) (arg_size <= SIZEOF_REGISTER && (arg_size == 1 || arg_size == 2 || arg_size == 4 || arg_size == 8))
421 allocate_register_for_valuetype_win64 (ArgInfo *arg_info, ArgumentClass arg_class, guint32 arg_size, AMD64_Reg_No int_regs [], int int_reg_count, AMD64_XMM_Reg_No float_regs [], int float_reg_count, guint32 *current_int_reg, guint32 *current_float_reg)
423 gboolean result = FALSE;
425 assert (arg_info != NULL && int_regs != NULL && float_regs != NULL && current_int_reg != NULL && current_float_reg != NULL);
426 assert (arg_info->storage == ArgValuetypeInReg || arg_info->storage == ArgValuetypeAddrInIReg);
428 arg_info->pair_storage [0] = arg_info->pair_storage [1] = ArgNone;
429 arg_info->pair_regs [0] = arg_info->pair_regs [1] = ArgNone;
430 arg_info->pair_size [0] = 0;
431 arg_info->pair_size [1] = 0;
434 if (arg_class == ARG_CLASS_INTEGER && *current_int_reg < int_reg_count) {
435 /* Pass parameter in integer register. */
436 arg_info->pair_storage [0] = ArgInIReg;
437 arg_info->pair_regs [0] = int_regs [*current_int_reg];
438 (*current_int_reg) ++;
440 } else if (arg_class == ARG_CLASS_SSE && *current_float_reg < float_reg_count) {
441 /* Pass parameter in float register. */
442 arg_info->pair_storage [0] = (arg_size <= sizeof (gfloat)) ? ArgInFloatSSEReg : ArgInDoubleSSEReg;
443 arg_info->pair_regs [0] = float_regs [*current_float_reg];
444 (*current_float_reg) ++;
448 if (result == TRUE) {
449 arg_info->pair_size [0] = arg_size;
456 static inline gboolean
457 allocate_parameter_register_for_valuetype_win64 (ArgInfo *arg_info, ArgumentClass arg_class, guint32 arg_size, guint32 *current_int_reg, guint32 *current_float_reg)
459 return allocate_register_for_valuetype_win64 (arg_info, arg_class, arg_size, param_regs, PARAM_REGS, float_param_regs, FLOAT_PARAM_REGS, current_int_reg, current_float_reg);
462 static inline gboolean
463 allocate_return_register_for_valuetype_win64 (ArgInfo *arg_info, ArgumentClass arg_class, guint32 arg_size, guint32 *current_int_reg, guint32 *current_float_reg)
465 return allocate_register_for_valuetype_win64 (arg_info, arg_class, arg_size, return_regs, RETURN_REGS, float_return_regs, FLOAT_RETURN_REGS, current_int_reg, current_float_reg);
469 allocate_storage_for_valuetype_win64 (ArgInfo *arg_info, MonoType *type, gboolean is_return, ArgumentClass arg_class,
470 guint32 arg_size, guint32 *current_int_reg, guint32 *current_float_reg, guint32 *stack_size)
472 /* Windows x64 value type ABI.
474 * Parameters: https://msdn.microsoft.com/en-us/library/zthk2dkh.aspx
476 * Integer/Float types smaller than or equals to 8 bytes or porperly sized struct/union (1,2,4,8)
477 * Try pass in register using ArgValuetypeInReg/(ArgInIReg|ArgInFloatSSEReg|ArgInDoubleSSEReg) as storage and size of parameter(1,2,4,8), if no more registers, pass on stack using ArgOnStack as storage and size of parameter(1,2,4,8).
478 * Integer/Float types bigger than 8 bytes or struct/unions larger than 8 bytes or (3,5,6,7).
479 * Try to pass pointer in register using ArgValuetypeAddrInIReg, if no more registers, pass pointer on stack using ArgValuetypeAddrOnStack as storage and parameter size of register (8 bytes).
481 * Return values: https://msdn.microsoft.com/en-us/library/7572ztz4.aspx.
483 * Integers/Float types smaller than or equal to 8 bytes
484 * Return in corresponding register RAX/XMM0 using ArgValuetypeInReg/(ArgInIReg|ArgInFloatSSEReg|ArgInDoubleSSEReg) as storage and size of parameter(1,2,4,8).
485 * Properly sized struct/unions (1,2,4,8)
486 * Return in register RAX using ArgValuetypeInReg as storage and size of parameter(1,2,4,8).
487 * Types bigger than 8 bytes or struct/unions larger than 8 bytes or (3,5,6,7).
488 * Return pointer to allocated stack space (allocated by caller) using ArgValuetypeAddrInIReg as storage and parameter size.
491 assert (arg_info != NULL && type != NULL && current_int_reg != NULL && current_float_reg != NULL && stack_size != NULL);
495 /* Parameter cases. */
496 if (arg_class != ARG_CLASS_MEMORY && MONO_WIN64_VALUE_TYPE_FITS_REG (arg_size)) {
497 assert (arg_size == 1 || arg_size == 2 || arg_size == 4 || arg_size == 8);
499 /* First, try to use registers for parameter. If type is struct it can only be passed by value in integer register. */
500 arg_info->storage = ArgValuetypeInReg;
501 if (!allocate_parameter_register_for_valuetype_win64 (arg_info, !MONO_TYPE_ISSTRUCT (type) ? arg_class : ARG_CLASS_INTEGER, arg_size, current_int_reg, current_float_reg)) {
502 /* No more registers, fallback passing parameter on stack as value. */
503 assert (arg_info->pair_storage [0] == ArgNone && arg_info->pair_storage [1] == ArgNone && arg_info->pair_size [0] == 0 && arg_info->pair_size [1] == 0 && arg_info->nregs == 0);
505 /* Passing value directly on stack, so use size of value. */
506 arg_info->storage = ArgOnStack;
507 arg_size = ALIGN_TO (arg_size, sizeof (mgreg_t));
508 arg_info->offset = *stack_size;
509 arg_info->arg_size = arg_size;
510 *stack_size += arg_size;
513 /* Fallback to stack, try to pass address to parameter in register. Always use integer register to represent stack address. */
514 arg_info->storage = ArgValuetypeAddrInIReg;
515 if (!allocate_parameter_register_for_valuetype_win64 (arg_info, ARG_CLASS_INTEGER, arg_size, current_int_reg, current_float_reg)) {
516 /* No more registers, fallback passing address to parameter on stack. */
517 assert (arg_info->pair_storage [0] == ArgNone && arg_info->pair_storage [1] == ArgNone && arg_info->pair_size [0] == 0 && arg_info->pair_size [1] == 0 && arg_info->nregs == 0);
519 /* Passing an address to value on stack, so use size of register as argument size. */
520 arg_info->storage = ArgValuetypeAddrOnStack;
521 arg_size = sizeof (mgreg_t);
522 arg_info->offset = *stack_size;
523 arg_info->arg_size = arg_size;
524 *stack_size += arg_size;
528 /* Return value cases. */
529 if (arg_class != ARG_CLASS_MEMORY && MONO_WIN64_VALUE_TYPE_FITS_REG (arg_size)) {
530 assert (arg_size == 1 || arg_size == 2 || arg_size == 4 || arg_size == 8);
532 /* Return value fits into return registers. If type is struct it can only be returned by value in integer register. */
533 arg_info->storage = ArgValuetypeInReg;
534 allocate_return_register_for_valuetype_win64 (arg_info, !MONO_TYPE_ISSTRUCT (type) ? arg_class : ARG_CLASS_INTEGER, arg_size, current_int_reg, current_float_reg);
536 /* Only RAX/XMM0 should be used to return valuetype. */
537 assert ((arg_info->pair_regs[0] == AMD64_RAX && arg_info->pair_regs[1] == ArgNone) || (arg_info->pair_regs[0] == AMD64_XMM0 && arg_info->pair_regs[1] == ArgNone));
539 /* Return value doesn't fit into return register, return address to allocated stack space (allocated by caller and passed as input). */
540 arg_info->storage = ArgValuetypeAddrInIReg;
541 allocate_return_register_for_valuetype_win64 (arg_info, ARG_CLASS_INTEGER, arg_size, current_int_reg, current_float_reg);
543 /* Only RAX should be used to return valuetype address. */
544 assert (arg_info->pair_regs[0] == AMD64_RAX && arg_info->pair_regs[1] == ArgNone);
546 arg_size = ALIGN_TO (arg_size, sizeof (mgreg_t));
547 arg_info->offset = *stack_size;
548 *stack_size += arg_size;
554 get_valuetype_size_win64 (MonoClass *klass, gboolean pinvoke, ArgInfo *arg_info, MonoType *type, ArgumentClass *arg_class, guint32 *arg_size)
557 *arg_class = ARG_CLASS_NO_CLASS;
559 assert (klass != NULL && arg_info != NULL && type != NULL && arg_class != NULL && arg_size != NULL);
562 /* Calculate argument class type and size of marshalled type. */
563 MonoMarshalType *info = mono_marshal_load_type_info (klass);
564 *arg_size = info->native_size;
566 /* Calculate argument class type and size of managed type. */
567 *arg_size = mono_class_value_size (klass, NULL);
570 /* Windows ABI only handle value types on stack or passed in integer register (if it fits register size). */
571 *arg_class = MONO_WIN64_VALUE_TYPE_FITS_REG (*arg_size) ? ARG_CLASS_INTEGER : ARG_CLASS_MEMORY;
573 if (*arg_class == ARG_CLASS_MEMORY) {
574 /* Value type has a size that doesn't seem to fit register according to ABI. Try to used full stack size of type. */
575 *arg_size = mini_type_stack_size_full (&klass->byval_arg, NULL, pinvoke);
579 * Standard C and C++ doesn't allow empty structs, empty structs will always have a size of 1 byte.
580 * GCC have an extension to allow empty structs, https://gcc.gnu.org/onlinedocs/gcc/Empty-Structures.html.
581 * This cause a little dilemma since runtime build using none GCC compiler will not be compatible with
582 * GCC build C libraries and the other way around. On platforms where empty structs has size of 1 byte
583 * it must be represented in call and cannot be dropped.
585 if (*arg_size == 0 && MONO_TYPE_ISSTRUCT (type)) {
586 arg_info->pass_empty_struct = TRUE;
587 *arg_size = SIZEOF_REGISTER;
588 *arg_class = ARG_CLASS_INTEGER;
591 assert (*arg_class != ARG_CLASS_NO_CLASS);
595 add_valuetype_win64 (MonoMethodSignature *signature, ArgInfo *arg_info, MonoType *type,
596 gboolean is_return, guint32 *current_int_reg, guint32 *current_float_reg, guint32 *stack_size)
598 guint32 arg_size = SIZEOF_REGISTER;
599 MonoClass *klass = NULL;
600 ArgumentClass arg_class;
602 assert (signature != NULL && arg_info != NULL && type != NULL && current_int_reg != NULL && current_float_reg != NULL && stack_size != NULL);
604 klass = mono_class_from_mono_type (type);
605 get_valuetype_size_win64 (klass, signature->pinvoke, arg_info, type, &arg_class, &arg_size);
607 /* Only drop value type if its not an empty struct as input that must be represented in call */
608 if ((arg_size == 0 && !arg_info->pass_empty_struct) || (arg_size == 0 && arg_info->pass_empty_struct && is_return)) {
609 arg_info->storage = ArgValuetypeInReg;
610 arg_info->pair_storage [0] = arg_info->pair_storage [1] = ArgNone;
612 /* Alocate storage for value type. */
613 allocate_storage_for_valuetype_win64 (arg_info, type, is_return, arg_class, arg_size, current_int_reg, current_float_reg, stack_size);
617 #endif /* TARGET_WIN32 */
620 add_valuetype (MonoMethodSignature *sig, ArgInfo *ainfo, MonoType *type,
622 guint32 *gr, guint32 *fr, guint32 *stack_size)
625 add_valuetype_win64 (sig, ainfo, type, is_return, gr, fr, stack_size);
627 guint32 size, quad, nquads, i, nfields;
628 /* Keep track of the size used in each quad so we can */
629 /* use the right size when copying args/return vars. */
630 guint32 quadsize [2] = {8, 8};
631 ArgumentClass args [2];
632 StructFieldInfo *fields = NULL;
633 GArray *fields_array;
635 gboolean pass_on_stack = FALSE;
638 klass = mono_class_from_mono_type (type);
639 size = mini_type_stack_size_full (&klass->byval_arg, NULL, sig->pinvoke);
641 if (!sig->pinvoke && ((is_return && (size == 8)) || (!is_return && (size <= 16)))) {
642 /* We pass and return vtypes of size 8 in a register */
643 } else if (!sig->pinvoke || (size == 0) || (size > 16)) {
644 pass_on_stack = TRUE;
647 /* If this struct can't be split up naturally into 8-byte */
648 /* chunks (registers), pass it on the stack. */
650 MonoMarshalType *info = mono_marshal_load_type_info (klass);
652 struct_size = info->native_size;
654 struct_size = mono_class_value_size (klass, NULL);
657 * Collect field information recursively to be able to
658 * handle nested structures.
660 fields_array = g_array_new (FALSE, TRUE, sizeof (StructFieldInfo));
661 collect_field_info_nested (klass, fields_array, 0, sig->pinvoke, klass->unicode);
662 fields = (StructFieldInfo*)fields_array->data;
663 nfields = fields_array->len;
665 for (i = 0; i < nfields; ++i) {
666 if ((fields [i].offset < 8) && (fields [i].offset + fields [i].size) > 8) {
667 pass_on_stack = TRUE;
673 ainfo->storage = ArgValuetypeInReg;
674 ainfo->pair_storage [0] = ainfo->pair_storage [1] = ArgNone;
679 /* Allways pass in memory */
680 ainfo->offset = *stack_size;
681 *stack_size += ALIGN_TO (size, 8);
682 ainfo->storage = is_return ? ArgValuetypeAddrInIReg : ArgOnStack;
684 ainfo->arg_size = ALIGN_TO (size, 8);
686 g_array_free (fields_array, TRUE);
696 int n = mono_class_value_size (klass, NULL);
698 quadsize [0] = n >= 8 ? 8 : n;
699 quadsize [1] = n >= 8 ? MAX (n - 8, 8) : 0;
701 /* Always pass in 1 or 2 integer registers */
702 args [0] = ARG_CLASS_INTEGER;
703 args [1] = ARG_CLASS_INTEGER;
704 /* Only the simplest cases are supported */
705 if (is_return && nquads != 1) {
706 args [0] = ARG_CLASS_MEMORY;
707 args [1] = ARG_CLASS_MEMORY;
711 * Implement the algorithm from section 3.2.3 of the X86_64 ABI.
712 * The X87 and SSEUP stuff is left out since there are no such types in
716 ainfo->storage = ArgValuetypeInReg;
717 ainfo->pair_storage [0] = ainfo->pair_storage [1] = ArgNone;
721 if (struct_size > 16) {
722 ainfo->offset = *stack_size;
723 *stack_size += ALIGN_TO (struct_size, 8);
724 ainfo->storage = is_return ? ArgValuetypeAddrInIReg : ArgOnStack;
726 ainfo->arg_size = ALIGN_TO (struct_size, 8);
728 g_array_free (fields_array, TRUE);
732 args [0] = ARG_CLASS_NO_CLASS;
733 args [1] = ARG_CLASS_NO_CLASS;
734 for (quad = 0; quad < nquads; ++quad) {
735 ArgumentClass class1;
738 class1 = ARG_CLASS_MEMORY;
740 class1 = ARG_CLASS_NO_CLASS;
741 for (i = 0; i < nfields; ++i) {
742 if ((fields [i].offset < 8) && (fields [i].offset + fields [i].size) > 8) {
743 /* Unaligned field */
747 /* Skip fields in other quad */
748 if ((quad == 0) && (fields [i].offset >= 8))
750 if ((quad == 1) && (fields [i].offset < 8))
753 /* How far into this quad this data extends.*/
754 /* (8 is size of quad) */
755 quadsize [quad] = fields [i].offset + fields [i].size - (quad * 8);
757 class1 = merge_argument_class_from_type (fields [i].type, class1);
759 /* Empty structs have a nonzero size, causing this assert to be hit */
761 g_assert (class1 != ARG_CLASS_NO_CLASS);
762 args [quad] = class1;
766 g_array_free (fields_array, TRUE);
768 /* Post merger cleanup */
769 if ((args [0] == ARG_CLASS_MEMORY) || (args [1] == ARG_CLASS_MEMORY))
770 args [0] = args [1] = ARG_CLASS_MEMORY;
772 /* Allocate registers */
777 while (quadsize [0] != 1 && quadsize [0] != 2 && quadsize [0] != 4 && quadsize [0] != 8)
779 while (quadsize [1] != 0 && quadsize [1] != 1 && quadsize [1] != 2 && quadsize [1] != 4 && quadsize [1] != 8)
782 ainfo->storage = ArgValuetypeInReg;
783 ainfo->pair_storage [0] = ainfo->pair_storage [1] = ArgNone;
784 g_assert (quadsize [0] <= 8);
785 g_assert (quadsize [1] <= 8);
786 ainfo->pair_size [0] = quadsize [0];
787 ainfo->pair_size [1] = quadsize [1];
788 ainfo->nregs = nquads;
789 for (quad = 0; quad < nquads; ++quad) {
790 switch (args [quad]) {
791 case ARG_CLASS_INTEGER:
792 if (*gr >= PARAM_REGS)
793 args [quad] = ARG_CLASS_MEMORY;
795 ainfo->pair_storage [quad] = ArgInIReg;
797 ainfo->pair_regs [quad] = return_regs [*gr];
799 ainfo->pair_regs [quad] = param_regs [*gr];
804 if (*fr >= FLOAT_PARAM_REGS)
805 args [quad] = ARG_CLASS_MEMORY;
807 if (quadsize[quad] <= 4)
808 ainfo->pair_storage [quad] = ArgInFloatSSEReg;
809 else ainfo->pair_storage [quad] = ArgInDoubleSSEReg;
810 ainfo->pair_regs [quad] = *fr;
814 case ARG_CLASS_MEMORY:
816 case ARG_CLASS_NO_CLASS:
819 g_assert_not_reached ();
823 if ((args [0] == ARG_CLASS_MEMORY) || (args [1] == ARG_CLASS_MEMORY)) {
825 /* Revert possible register assignments */
829 ainfo->offset = *stack_size;
831 arg_size = ALIGN_TO (struct_size, 8);
833 arg_size = nquads * sizeof(mgreg_t);
834 *stack_size += arg_size;
835 ainfo->storage = is_return ? ArgValuetypeAddrInIReg : ArgOnStack;
837 ainfo->arg_size = arg_size;
840 #endif /* !TARGET_WIN32 */
846 * Obtain information about a call according to the calling convention.
847 * For AMD64 System V, see the "System V ABI, x86-64 Architecture Processor Supplement
848 * Draft Version 0.23" document for more information.
849 * For AMD64 Windows, see "Overview of x64 Calling Conventions",
850 * https://msdn.microsoft.com/en-us/library/ms235286.aspx
853 get_call_info (MonoMemPool *mp, MonoMethodSignature *sig)
855 guint32 i, gr, fr, pstart;
857 int n = sig->hasthis + sig->param_count;
858 guint32 stack_size = 0;
860 gboolean is_pinvoke = sig->pinvoke;
863 cinfo = (CallInfo *)mono_mempool_alloc0 (mp, sizeof (CallInfo) + (sizeof (ArgInfo) * n));
865 cinfo = (CallInfo *)g_malloc0 (sizeof (CallInfo) + (sizeof (ArgInfo) * n));
868 cinfo->gsharedvt = mini_is_gsharedvt_variable_signature (sig);
874 /* Reserve space where the callee can save the argument registers */
875 stack_size = 4 * sizeof (mgreg_t);
879 ret_type = mini_get_underlying_type (sig->ret);
880 switch (ret_type->type) {
890 case MONO_TYPE_FNPTR:
891 case MONO_TYPE_CLASS:
892 case MONO_TYPE_OBJECT:
893 case MONO_TYPE_SZARRAY:
894 case MONO_TYPE_ARRAY:
895 case MONO_TYPE_STRING:
896 cinfo->ret.storage = ArgInIReg;
897 cinfo->ret.reg = AMD64_RAX;
901 cinfo->ret.storage = ArgInIReg;
902 cinfo->ret.reg = AMD64_RAX;
905 cinfo->ret.storage = ArgInFloatSSEReg;
906 cinfo->ret.reg = AMD64_XMM0;
909 cinfo->ret.storage = ArgInDoubleSSEReg;
910 cinfo->ret.reg = AMD64_XMM0;
912 case MONO_TYPE_GENERICINST:
913 if (!mono_type_generic_inst_is_valuetype (ret_type)) {
914 cinfo->ret.storage = ArgInIReg;
915 cinfo->ret.reg = AMD64_RAX;
918 if (mini_is_gsharedvt_type (ret_type)) {
919 cinfo->ret.storage = ArgGsharedvtVariableInReg;
923 case MONO_TYPE_VALUETYPE:
924 case MONO_TYPE_TYPEDBYREF: {
925 guint32 tmp_gr = 0, tmp_fr = 0, tmp_stacksize = 0;
927 add_valuetype (sig, &cinfo->ret, ret_type, TRUE, &tmp_gr, &tmp_fr, &tmp_stacksize);
928 g_assert (cinfo->ret.storage != ArgInIReg);
933 g_assert (mini_is_gsharedvt_type (ret_type));
934 cinfo->ret.storage = ArgGsharedvtVariableInReg;
939 g_error ("Can't handle as return value 0x%x", ret_type->type);
944 * To simplify get_this_arg_reg () and LLVM integration, emit the vret arg after
945 * the first argument, allowing 'this' to be always passed in the first arg reg.
946 * Also do this if the first argument is a reference type, since virtual calls
947 * are sometimes made using calli without sig->hasthis set, like in the delegate
950 ArgStorage ret_storage = cinfo->ret.storage;
951 if ((ret_storage == ArgValuetypeAddrInIReg || ret_storage == ArgGsharedvtVariableInReg) && !is_pinvoke && (sig->hasthis || (sig->param_count > 0 && MONO_TYPE_IS_REFERENCE (mini_get_underlying_type (sig->params [0]))))) {
953 add_general (&gr, &stack_size, cinfo->args + 0);
955 add_general (&gr, &stack_size, &cinfo->args [sig->hasthis + 0]);
958 add_general (&gr, &stack_size, &cinfo->ret);
959 cinfo->ret.storage = ret_storage;
960 cinfo->vret_arg_index = 1;
964 add_general (&gr, &stack_size, cinfo->args + 0);
966 if (ret_storage == ArgValuetypeAddrInIReg || ret_storage == ArgGsharedvtVariableInReg) {
967 add_general (&gr, &stack_size, &cinfo->ret);
968 cinfo->ret.storage = ret_storage;
972 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG) && (n == 0)) {
974 fr = FLOAT_PARAM_REGS;
976 /* Emit the signature cookie just before the implicit arguments */
977 add_general (&gr, &stack_size, &cinfo->sig_cookie);
980 for (i = pstart; i < sig->param_count; ++i) {
981 ArgInfo *ainfo = &cinfo->args [sig->hasthis + i];
985 /* The float param registers and other param registers must be the same index on Windows x64.*/
992 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG) && (i == sig->sentinelpos)) {
993 /* We allways pass the sig cookie on the stack for simplicity */
995 * Prevent implicit arguments + the sig cookie from being passed
999 fr = FLOAT_PARAM_REGS;
1001 /* Emit the signature cookie just before the implicit arguments */
1002 add_general (&gr, &stack_size, &cinfo->sig_cookie);
1005 ptype = mini_get_underlying_type (sig->params [i]);
1006 switch (ptype->type) {
1009 add_general (&gr, &stack_size, ainfo);
1013 add_general (&gr, &stack_size, ainfo);
1017 add_general (&gr, &stack_size, ainfo);
1022 case MONO_TYPE_FNPTR:
1023 case MONO_TYPE_CLASS:
1024 case MONO_TYPE_OBJECT:
1025 case MONO_TYPE_STRING:
1026 case MONO_TYPE_SZARRAY:
1027 case MONO_TYPE_ARRAY:
1028 add_general (&gr, &stack_size, ainfo);
1030 case MONO_TYPE_GENERICINST:
1031 if (!mono_type_generic_inst_is_valuetype (ptype)) {
1032 add_general (&gr, &stack_size, ainfo);
1035 if (mini_is_gsharedvt_variable_type (ptype)) {
1036 /* gsharedvt arguments are passed by ref */
1037 add_general (&gr, &stack_size, ainfo);
1038 if (ainfo->storage == ArgInIReg)
1039 ainfo->storage = ArgGSharedVtInReg;
1041 ainfo->storage = ArgGSharedVtOnStack;
1045 case MONO_TYPE_VALUETYPE:
1046 case MONO_TYPE_TYPEDBYREF:
1047 add_valuetype (sig, ainfo, ptype, FALSE, &gr, &fr, &stack_size);
1052 add_general (&gr, &stack_size, ainfo);
1055 add_float (&fr, &stack_size, ainfo, FALSE);
1058 add_float (&fr, &stack_size, ainfo, TRUE);
1061 case MONO_TYPE_MVAR:
1062 /* gsharedvt arguments are passed by ref */
1063 g_assert (mini_is_gsharedvt_type (ptype));
1064 add_general (&gr, &stack_size, ainfo);
1065 if (ainfo->storage == ArgInIReg)
1066 ainfo->storage = ArgGSharedVtInReg;
1068 ainfo->storage = ArgGSharedVtOnStack;
1071 g_assert_not_reached ();
1075 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG) && (n > 0) && (sig->sentinelpos == sig->param_count)) {
1077 fr = FLOAT_PARAM_REGS;
1079 /* Emit the signature cookie just before the implicit arguments */
1080 add_general (&gr, &stack_size, &cinfo->sig_cookie);
1083 cinfo->stack_usage = stack_size;
1084 cinfo->reg_usage = gr;
1085 cinfo->freg_usage = fr;
1090 * mono_arch_get_argument_info:
1091 * @csig: a method signature
1092 * @param_count: the number of parameters to consider
1093 * @arg_info: an array to store the result infos
1095 * Gathers information on parameters such as size, alignment and
1096 * padding. arg_info should be large enought to hold param_count + 1 entries.
1098 * Returns the size of the argument area on the stack.
1101 mono_arch_get_argument_info (MonoMethodSignature *csig, int param_count, MonoJitArgumentInfo *arg_info)
1104 CallInfo *cinfo = get_call_info (NULL, csig);
1105 guint32 args_size = cinfo->stack_usage;
1107 /* The arguments are saved to a stack area in mono_arch_instrument_prolog */
1108 if (csig->hasthis) {
1109 arg_info [0].offset = 0;
1112 for (k = 0; k < param_count; k++) {
1113 arg_info [k + 1].offset = ((k + csig->hasthis) * 8);
1115 arg_info [k + 1].size = 0;
1124 mono_arch_tail_call_supported (MonoCompile *cfg, MonoMethodSignature *caller_sig, MonoMethodSignature *callee_sig)
1128 MonoType *callee_ret;
1130 c1 = get_call_info (NULL, caller_sig);
1131 c2 = get_call_info (NULL, callee_sig);
1132 res = c1->stack_usage >= c2->stack_usage;
1133 callee_ret = mini_get_underlying_type (callee_sig->ret);
1134 if (callee_ret && MONO_TYPE_ISSTRUCT (callee_ret) && c2->ret.storage != ArgValuetypeInReg)
1135 /* An address on the callee's stack is passed as the first argument */
1145 * Initialize the cpu to execute managed code.
1148 mono_arch_cpu_init (void)
1153 /* spec compliance requires running with double precision */
1154 __asm__ __volatile__ ("fnstcw %0\n": "=m" (fpcw));
1155 fpcw &= ~X86_FPCW_PRECC_MASK;
1156 fpcw |= X86_FPCW_PREC_DOUBLE;
1157 __asm__ __volatile__ ("fldcw %0\n": : "m" (fpcw));
1158 __asm__ __volatile__ ("fnstcw %0\n": "=m" (fpcw));
1160 /* TODO: This is crashing on Win64 right now.
1161 * _control87 (_PC_53, MCW_PC);
1167 * Initialize architecture specific code.
1170 mono_arch_init (void)
1172 mono_os_mutex_init_recursive (&mini_arch_mutex);
1174 mono_aot_register_jit_icall ("mono_amd64_throw_exception", mono_amd64_throw_exception);
1175 mono_aot_register_jit_icall ("mono_amd64_throw_corlib_exception", mono_amd64_throw_corlib_exception);
1176 mono_aot_register_jit_icall ("mono_amd64_resume_unwind", mono_amd64_resume_unwind);
1177 mono_aot_register_jit_icall ("mono_amd64_get_original_ip", mono_amd64_get_original_ip);
1178 mono_aot_register_jit_icall ("mono_amd64_handler_block_trampoline_helper", mono_amd64_handler_block_trampoline_helper);
1180 #if defined(MONO_ARCH_GSHAREDVT_SUPPORTED)
1181 mono_aot_register_jit_icall ("mono_amd64_start_gsharedvt_call", mono_amd64_start_gsharedvt_call);
1185 bp_trampoline = mini_get_breakpoint_trampoline ();
1189 * Cleanup architecture specific code.
1192 mono_arch_cleanup (void)
1194 mono_os_mutex_destroy (&mini_arch_mutex);
1198 * This function returns the optimizations supported on this cpu.
1201 mono_arch_cpu_optimizations (guint32 *exclude_mask)
1207 if (mono_hwcap_x86_has_cmov) {
1208 opts |= MONO_OPT_CMOV;
1210 if (mono_hwcap_x86_has_fcmov)
1211 opts |= MONO_OPT_FCMOV;
1213 *exclude_mask |= MONO_OPT_FCMOV;
1215 *exclude_mask |= MONO_OPT_CMOV;
1219 /* The current SIMD doesn't support the argument used by a LD_ADDR to be of type OP_VTARG_ADDR. */
1220 /* This will now be used for value types > 8 or of size 3,5,6,7 as dictated by windows x64 value type ABI. */
1221 /* Since OP_VTARG_ADDR needs to be resolved in mono_spill_global_vars and the SIMD implementation optimize */
1222 /* away the LD_ADDR in load_simd_vreg, that will cause an error in mono_spill_global_vars since incorrect opcode */
1223 /* will now have a reference to an argument that won't be fully decomposed. */
1224 *exclude_mask |= MONO_OPT_SIMD;
1231 * This function test for all SSE functions supported.
1233 * Returns a bitmask corresponding to all supported versions.
1237 mono_arch_cpu_enumerate_simd_versions (void)
1239 guint32 sse_opts = 0;
1241 if (mono_hwcap_x86_has_sse1)
1242 sse_opts |= SIMD_VERSION_SSE1;
1244 if (mono_hwcap_x86_has_sse2)
1245 sse_opts |= SIMD_VERSION_SSE2;
1247 if (mono_hwcap_x86_has_sse3)
1248 sse_opts |= SIMD_VERSION_SSE3;
1250 if (mono_hwcap_x86_has_ssse3)
1251 sse_opts |= SIMD_VERSION_SSSE3;
1253 if (mono_hwcap_x86_has_sse41)
1254 sse_opts |= SIMD_VERSION_SSE41;
1256 if (mono_hwcap_x86_has_sse42)
1257 sse_opts |= SIMD_VERSION_SSE42;
1259 if (mono_hwcap_x86_has_sse4a)
1260 sse_opts |= SIMD_VERSION_SSE4a;
1268 mono_arch_get_allocatable_int_vars (MonoCompile *cfg)
1273 for (i = 0; i < cfg->num_varinfo; i++) {
1274 MonoInst *ins = cfg->varinfo [i];
1275 MonoMethodVar *vmv = MONO_VARINFO (cfg, i);
1278 if (vmv->range.first_use.abs_pos >= vmv->range.last_use.abs_pos)
1281 if ((ins->flags & (MONO_INST_IS_DEAD|MONO_INST_VOLATILE|MONO_INST_INDIRECT)) ||
1282 (ins->opcode != OP_LOCAL && ins->opcode != OP_ARG))
1285 if (mono_is_regsize_var (ins->inst_vtype)) {
1286 g_assert (MONO_VARINFO (cfg, i)->reg == -1);
1287 g_assert (i == vmv->idx);
1288 vars = g_list_prepend (vars, vmv);
1292 vars = mono_varlist_sort (cfg, vars, 0);
1298 * mono_arch_compute_omit_fp:
1300 * Determine whenever the frame pointer can be eliminated.
1303 mono_arch_compute_omit_fp (MonoCompile *cfg)
1305 MonoMethodSignature *sig;
1306 MonoMethodHeader *header;
1310 if (cfg->arch.omit_fp_computed)
1313 header = cfg->header;
1315 sig = mono_method_signature (cfg->method);
1317 if (!cfg->arch.cinfo)
1318 cfg->arch.cinfo = get_call_info (cfg->mempool, sig);
1319 cinfo = (CallInfo *)cfg->arch.cinfo;
1322 * FIXME: Remove some of the restrictions.
1324 cfg->arch.omit_fp = TRUE;
1325 cfg->arch.omit_fp_computed = TRUE;
1327 if (cfg->disable_omit_fp)
1328 cfg->arch.omit_fp = FALSE;
1330 if (!debug_omit_fp ())
1331 cfg->arch.omit_fp = FALSE;
1333 if (cfg->method->save_lmf)
1334 cfg->arch.omit_fp = FALSE;
1336 if (cfg->flags & MONO_CFG_HAS_ALLOCA)
1337 cfg->arch.omit_fp = FALSE;
1338 if (header->num_clauses)
1339 cfg->arch.omit_fp = FALSE;
1340 if (cfg->param_area)
1341 cfg->arch.omit_fp = FALSE;
1342 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG))
1343 cfg->arch.omit_fp = FALSE;
1344 if ((mono_jit_trace_calls != NULL && mono_trace_eval (cfg->method)) ||
1345 (cfg->prof_options & MONO_PROFILE_ENTER_LEAVE))
1346 cfg->arch.omit_fp = FALSE;
1347 for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
1348 ArgInfo *ainfo = &cinfo->args [i];
1350 if (ainfo->storage == ArgOnStack || ainfo->storage == ArgValuetypeAddrInIReg || ainfo->storage == ArgValuetypeAddrOnStack) {
1352 * The stack offset can only be determined when the frame
1355 cfg->arch.omit_fp = FALSE;
1360 for (i = cfg->locals_start; i < cfg->num_varinfo; i++) {
1361 MonoInst *ins = cfg->varinfo [i];
1364 locals_size += mono_type_size (ins->inst_vtype, &ialign);
1369 mono_arch_get_global_int_regs (MonoCompile *cfg)
1373 mono_arch_compute_omit_fp (cfg);
1375 if (cfg->arch.omit_fp)
1376 regs = g_list_prepend (regs, (gpointer)AMD64_RBP);
1378 /* We use the callee saved registers for global allocation */
1379 regs = g_list_prepend (regs, (gpointer)AMD64_RBX);
1380 regs = g_list_prepend (regs, (gpointer)AMD64_R12);
1381 regs = g_list_prepend (regs, (gpointer)AMD64_R13);
1382 regs = g_list_prepend (regs, (gpointer)AMD64_R14);
1383 regs = g_list_prepend (regs, (gpointer)AMD64_R15);
1385 regs = g_list_prepend (regs, (gpointer)AMD64_RDI);
1386 regs = g_list_prepend (regs, (gpointer)AMD64_RSI);
1393 mono_arch_get_global_fp_regs (MonoCompile *cfg)
1398 /* All XMM registers */
1399 for (i = 0; i < 16; ++i)
1400 regs = g_list_prepend (regs, GINT_TO_POINTER (i));
1406 mono_arch_get_iregs_clobbered_by_call (MonoCallInst *call)
1408 static GList *r = NULL;
1413 regs = g_list_prepend (regs, (gpointer)AMD64_RBP);
1414 regs = g_list_prepend (regs, (gpointer)AMD64_RBX);
1415 regs = g_list_prepend (regs, (gpointer)AMD64_R12);
1416 regs = g_list_prepend (regs, (gpointer)AMD64_R13);
1417 regs = g_list_prepend (regs, (gpointer)AMD64_R14);
1418 regs = g_list_prepend (regs, (gpointer)AMD64_R15);
1420 regs = g_list_prepend (regs, (gpointer)AMD64_R10);
1421 regs = g_list_prepend (regs, (gpointer)AMD64_R9);
1422 regs = g_list_prepend (regs, (gpointer)AMD64_R8);
1423 regs = g_list_prepend (regs, (gpointer)AMD64_RDI);
1424 regs = g_list_prepend (regs, (gpointer)AMD64_RSI);
1425 regs = g_list_prepend (regs, (gpointer)AMD64_RDX);
1426 regs = g_list_prepend (regs, (gpointer)AMD64_RCX);
1427 regs = g_list_prepend (regs, (gpointer)AMD64_RAX);
1429 InterlockedCompareExchangePointer ((gpointer*)&r, regs, NULL);
1436 mono_arch_get_fregs_clobbered_by_call (MonoCallInst *call)
1439 static GList *r = NULL;
1444 for (i = 0; i < AMD64_XMM_NREG; ++i)
1445 regs = g_list_prepend (regs, GINT_TO_POINTER (MONO_MAX_IREGS + i));
1447 InterlockedCompareExchangePointer ((gpointer*)&r, regs, NULL);
1454 * mono_arch_regalloc_cost:
1456 * Return the cost, in number of memory references, of the action of
1457 * allocating the variable VMV into a register during global register
1461 mono_arch_regalloc_cost (MonoCompile *cfg, MonoMethodVar *vmv)
1463 MonoInst *ins = cfg->varinfo [vmv->idx];
1465 if (cfg->method->save_lmf)
1466 /* The register is already saved */
1467 /* substract 1 for the invisible store in the prolog */
1468 return (ins->opcode == OP_ARG) ? 0 : 1;
1471 return (ins->opcode == OP_ARG) ? 1 : 2;
1475 * mono_arch_fill_argument_info:
1477 * Populate cfg->args, cfg->ret and cfg->vret_addr with information about the arguments
1481 mono_arch_fill_argument_info (MonoCompile *cfg)
1484 MonoMethodSignature *sig;
1489 sig = mono_method_signature (cfg->method);
1491 cinfo = (CallInfo *)cfg->arch.cinfo;
1492 sig_ret = mini_get_underlying_type (sig->ret);
1495 * Contrary to mono_arch_allocate_vars (), the information should describe
1496 * where the arguments are at the beginning of the method, not where they can be
1497 * accessed during the execution of the method. The later makes no sense for the
1498 * global register allocator, since a variable can be in more than one location.
1500 switch (cinfo->ret.storage) {
1502 case ArgInFloatSSEReg:
1503 case ArgInDoubleSSEReg:
1504 cfg->ret->opcode = OP_REGVAR;
1505 cfg->ret->inst_c0 = cinfo->ret.reg;
1507 case ArgValuetypeInReg:
1508 cfg->ret->opcode = OP_REGOFFSET;
1509 cfg->ret->inst_basereg = -1;
1510 cfg->ret->inst_offset = -1;
1515 g_assert_not_reached ();
1518 for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
1519 ArgInfo *ainfo = &cinfo->args [i];
1521 ins = cfg->args [i];
1523 switch (ainfo->storage) {
1525 case ArgInFloatSSEReg:
1526 case ArgInDoubleSSEReg:
1527 ins->opcode = OP_REGVAR;
1528 ins->inst_c0 = ainfo->reg;
1531 ins->opcode = OP_REGOFFSET;
1532 ins->inst_basereg = -1;
1533 ins->inst_offset = -1;
1535 case ArgValuetypeInReg:
1537 ins->opcode = OP_NOP;
1540 g_assert_not_reached ();
1546 mono_arch_allocate_vars (MonoCompile *cfg)
1549 MonoMethodSignature *sig;
1552 guint32 locals_stack_size, locals_stack_align;
1556 sig = mono_method_signature (cfg->method);
1558 cinfo = (CallInfo *)cfg->arch.cinfo;
1559 sig_ret = mini_get_underlying_type (sig->ret);
1561 mono_arch_compute_omit_fp (cfg);
1564 * We use the ABI calling conventions for managed code as well.
1565 * Exception: valuetypes are only sometimes passed or returned in registers.
1569 * The stack looks like this:
1570 * <incoming arguments passed on the stack>
1572 * <lmf/caller saved registers>
1575 * <localloc area> -> grows dynamically
1579 if (cfg->arch.omit_fp) {
1580 cfg->flags |= MONO_CFG_HAS_SPILLUP;
1581 cfg->frame_reg = AMD64_RSP;
1584 /* Locals are allocated backwards from %fp */
1585 cfg->frame_reg = AMD64_RBP;
1589 cfg->arch.saved_iregs = cfg->used_int_regs;
1590 if (cfg->method->save_lmf) {
1591 /* Save all callee-saved registers normally (except RBP, if not already used), and restore them when unwinding through an LMF */
1592 guint32 iregs_to_save = AMD64_CALLEE_SAVED_REGS & ~(1<<AMD64_RBP);
1593 cfg->arch.saved_iregs |= iregs_to_save;
1596 if (cfg->arch.omit_fp)
1597 cfg->arch.reg_save_area_offset = offset;
1598 /* Reserve space for callee saved registers */
1599 for (i = 0; i < AMD64_NREG; ++i)
1600 if (AMD64_IS_CALLEE_SAVED_REG (i) && (cfg->arch.saved_iregs & (1 << i))) {
1601 offset += sizeof(mgreg_t);
1603 if (!cfg->arch.omit_fp)
1604 cfg->arch.reg_save_area_offset = -offset;
1606 if (sig_ret->type != MONO_TYPE_VOID) {
1607 switch (cinfo->ret.storage) {
1609 case ArgInFloatSSEReg:
1610 case ArgInDoubleSSEReg:
1611 cfg->ret->opcode = OP_REGVAR;
1612 cfg->ret->inst_c0 = cinfo->ret.reg;
1613 cfg->ret->dreg = cinfo->ret.reg;
1615 case ArgValuetypeAddrInIReg:
1616 case ArgGsharedvtVariableInReg:
1617 /* The register is volatile */
1618 cfg->vret_addr->opcode = OP_REGOFFSET;
1619 cfg->vret_addr->inst_basereg = cfg->frame_reg;
1620 if (cfg->arch.omit_fp) {
1621 cfg->vret_addr->inst_offset = offset;
1625 cfg->vret_addr->inst_offset = -offset;
1627 if (G_UNLIKELY (cfg->verbose_level > 1)) {
1628 printf ("vret_addr =");
1629 mono_print_ins (cfg->vret_addr);
1632 case ArgValuetypeInReg:
1633 /* Allocate a local to hold the result, the epilog will copy it to the correct place */
1634 cfg->ret->opcode = OP_REGOFFSET;
1635 cfg->ret->inst_basereg = cfg->frame_reg;
1636 if (cfg->arch.omit_fp) {
1637 cfg->ret->inst_offset = offset;
1638 offset += cinfo->ret.pair_storage [1] == ArgNone ? 8 : 16;
1640 offset += cinfo->ret.pair_storage [1] == ArgNone ? 8 : 16;
1641 cfg->ret->inst_offset = - offset;
1645 g_assert_not_reached ();
1649 /* Allocate locals */
1650 offsets = mono_allocate_stack_slots (cfg, cfg->arch.omit_fp ? FALSE: TRUE, &locals_stack_size, &locals_stack_align);
1651 if (locals_stack_size > MONO_ARCH_MAX_FRAME_SIZE) {
1652 char *mname = mono_method_full_name (cfg->method, TRUE);
1653 mono_cfg_set_exception_invalid_program (cfg, g_strdup_printf ("Method %s stack is too big.", mname));
1658 if (locals_stack_align) {
1659 offset += (locals_stack_align - 1);
1660 offset &= ~(locals_stack_align - 1);
1662 if (cfg->arch.omit_fp) {
1663 cfg->locals_min_stack_offset = offset;
1664 cfg->locals_max_stack_offset = offset + locals_stack_size;
1666 cfg->locals_min_stack_offset = - (offset + locals_stack_size);
1667 cfg->locals_max_stack_offset = - offset;
1670 for (i = cfg->locals_start; i < cfg->num_varinfo; i++) {
1671 if (offsets [i] != -1) {
1672 MonoInst *ins = cfg->varinfo [i];
1673 ins->opcode = OP_REGOFFSET;
1674 ins->inst_basereg = cfg->frame_reg;
1675 if (cfg->arch.omit_fp)
1676 ins->inst_offset = (offset + offsets [i]);
1678 ins->inst_offset = - (offset + offsets [i]);
1679 //printf ("allocated local %d to ", i); mono_print_tree_nl (ins);
1682 offset += locals_stack_size;
1684 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG)) {
1685 g_assert (!cfg->arch.omit_fp);
1686 g_assert (cinfo->sig_cookie.storage == ArgOnStack);
1687 cfg->sig_cookie = cinfo->sig_cookie.offset + ARGS_OFFSET;
1690 for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
1691 ins = cfg->args [i];
1692 if (ins->opcode != OP_REGVAR) {
1693 ArgInfo *ainfo = &cinfo->args [i];
1694 gboolean inreg = TRUE;
1696 /* FIXME: Allocate volatile arguments to registers */
1697 if (ins->flags & (MONO_INST_VOLATILE|MONO_INST_INDIRECT))
1701 * Under AMD64, all registers used to pass arguments to functions
1702 * are volatile across calls.
1703 * FIXME: Optimize this.
1705 if ((ainfo->storage == ArgInIReg) || (ainfo->storage == ArgInFloatSSEReg) || (ainfo->storage == ArgInDoubleSSEReg) || (ainfo->storage == ArgValuetypeInReg) || (ainfo->storage == ArgGSharedVtInReg))
1708 ins->opcode = OP_REGOFFSET;
1710 switch (ainfo->storage) {
1712 case ArgInFloatSSEReg:
1713 case ArgInDoubleSSEReg:
1714 case ArgGSharedVtInReg:
1716 ins->opcode = OP_REGVAR;
1717 ins->dreg = ainfo->reg;
1721 case ArgGSharedVtOnStack:
1722 g_assert (!cfg->arch.omit_fp);
1723 ins->opcode = OP_REGOFFSET;
1724 ins->inst_basereg = cfg->frame_reg;
1725 ins->inst_offset = ainfo->offset + ARGS_OFFSET;
1727 case ArgValuetypeInReg:
1729 case ArgValuetypeAddrInIReg:
1730 case ArgValuetypeAddrOnStack: {
1732 g_assert (!cfg->arch.omit_fp);
1733 g_assert (ainfo->storage == ArgValuetypeAddrInIReg || (ainfo->storage == ArgValuetypeAddrOnStack && ainfo->pair_storage [0] == ArgNone));
1734 MONO_INST_NEW (cfg, indir, 0);
1736 indir->opcode = OP_REGOFFSET;
1737 if (ainfo->pair_storage [0] == ArgInIReg) {
1738 indir->inst_basereg = cfg->frame_reg;
1739 offset = ALIGN_TO (offset, sizeof (gpointer));
1740 offset += (sizeof (gpointer));
1741 indir->inst_offset = - offset;
1744 indir->inst_basereg = cfg->frame_reg;
1745 indir->inst_offset = ainfo->offset + ARGS_OFFSET;
1748 ins->opcode = OP_VTARG_ADDR;
1749 ins->inst_left = indir;
1757 if (!inreg && (ainfo->storage != ArgOnStack) && (ainfo->storage != ArgValuetypeAddrInIReg) && (ainfo->storage != ArgValuetypeAddrOnStack) && (ainfo->storage != ArgGSharedVtOnStack)) {
1758 ins->opcode = OP_REGOFFSET;
1759 ins->inst_basereg = cfg->frame_reg;
1760 /* These arguments are saved to the stack in the prolog */
1761 offset = ALIGN_TO (offset, sizeof(mgreg_t));
1762 if (cfg->arch.omit_fp) {
1763 ins->inst_offset = offset;
1764 offset += (ainfo->storage == ArgValuetypeInReg) ? ainfo->nregs * sizeof (mgreg_t) : sizeof (mgreg_t);
1765 // Arguments are yet supported by the stack map creation code
1766 //cfg->locals_max_stack_offset = MAX (cfg->locals_max_stack_offset, offset);
1768 offset += (ainfo->storage == ArgValuetypeInReg) ? ainfo->nregs * sizeof (mgreg_t) : sizeof (mgreg_t);
1769 ins->inst_offset = - offset;
1770 //cfg->locals_min_stack_offset = MIN (cfg->locals_min_stack_offset, offset);
1776 cfg->stack_offset = offset;
1780 mono_arch_create_vars (MonoCompile *cfg)
1782 MonoMethodSignature *sig;
1786 sig = mono_method_signature (cfg->method);
1788 if (!cfg->arch.cinfo)
1789 cfg->arch.cinfo = get_call_info (cfg->mempool, sig);
1790 cinfo = (CallInfo *)cfg->arch.cinfo;
1792 if (cinfo->ret.storage == ArgValuetypeInReg)
1793 cfg->ret_var_is_local = TRUE;
1795 sig_ret = mini_get_underlying_type (sig->ret);
1796 if (cinfo->ret.storage == ArgValuetypeAddrInIReg || cinfo->ret.storage == ArgGsharedvtVariableInReg) {
1797 cfg->vret_addr = mono_compile_create_var (cfg, &mono_defaults.int_class->byval_arg, OP_ARG);
1798 if (G_UNLIKELY (cfg->verbose_level > 1)) {
1799 printf ("vret_addr = ");
1800 mono_print_ins (cfg->vret_addr);
1804 if (cfg->gen_sdb_seq_points) {
1807 if (cfg->compile_aot) {
1808 MonoInst *ins = mono_compile_create_var (cfg, &mono_defaults.int_class->byval_arg, OP_LOCAL);
1809 ins->flags |= MONO_INST_VOLATILE;
1810 cfg->arch.seq_point_info_var = ins;
1812 ins = mono_compile_create_var (cfg, &mono_defaults.int_class->byval_arg, OP_LOCAL);
1813 ins->flags |= MONO_INST_VOLATILE;
1814 cfg->arch.ss_tramp_var = ins;
1816 ins = mono_compile_create_var (cfg, &mono_defaults.int_class->byval_arg, OP_LOCAL);
1817 ins->flags |= MONO_INST_VOLATILE;
1818 cfg->arch.bp_tramp_var = ins;
1821 if (cfg->method->save_lmf)
1822 cfg->create_lmf_var = TRUE;
1824 if (cfg->method->save_lmf) {
1826 #if !defined(TARGET_WIN32)
1827 if (!optimize_for_xen)
1828 cfg->lmf_ir_mono_lmf = TRUE;
1834 add_outarg_reg (MonoCompile *cfg, MonoCallInst *call, ArgStorage storage, int reg, MonoInst *tree)
1840 MONO_INST_NEW (cfg, ins, OP_MOVE);
1841 ins->dreg = mono_alloc_ireg_copy (cfg, tree->dreg);
1842 ins->sreg1 = tree->dreg;
1843 MONO_ADD_INS (cfg->cbb, ins);
1844 mono_call_inst_add_outarg_reg (cfg, call, ins->dreg, reg, FALSE);
1846 case ArgInFloatSSEReg:
1847 MONO_INST_NEW (cfg, ins, OP_AMD64_SET_XMMREG_R4);
1848 ins->dreg = mono_alloc_freg (cfg);
1849 ins->sreg1 = tree->dreg;
1850 MONO_ADD_INS (cfg->cbb, ins);
1852 mono_call_inst_add_outarg_reg (cfg, call, ins->dreg, reg, TRUE);
1854 case ArgInDoubleSSEReg:
1855 MONO_INST_NEW (cfg, ins, OP_FMOVE);
1856 ins->dreg = mono_alloc_freg (cfg);
1857 ins->sreg1 = tree->dreg;
1858 MONO_ADD_INS (cfg->cbb, ins);
1860 mono_call_inst_add_outarg_reg (cfg, call, ins->dreg, reg, TRUE);
1864 g_assert_not_reached ();
1869 arg_storage_to_load_membase (ArgStorage storage)
1873 #if defined(__mono_ilp32__)
1874 return OP_LOADI8_MEMBASE;
1876 return OP_LOAD_MEMBASE;
1878 case ArgInDoubleSSEReg:
1879 return OP_LOADR8_MEMBASE;
1880 case ArgInFloatSSEReg:
1881 return OP_LOADR4_MEMBASE;
1883 g_assert_not_reached ();
1890 emit_sig_cookie (MonoCompile *cfg, MonoCallInst *call, CallInfo *cinfo)
1892 MonoMethodSignature *tmp_sig;
1895 if (call->tail_call)
1898 g_assert (cinfo->sig_cookie.storage == ArgOnStack);
1901 * mono_ArgIterator_Setup assumes the signature cookie is
1902 * passed first and all the arguments which were before it are
1903 * passed on the stack after the signature. So compensate by
1904 * passing a different signature.
1906 tmp_sig = mono_metadata_signature_dup_full (cfg->method->klass->image, call->signature);
1907 tmp_sig->param_count -= call->signature->sentinelpos;
1908 tmp_sig->sentinelpos = 0;
1909 memcpy (tmp_sig->params, call->signature->params + call->signature->sentinelpos, tmp_sig->param_count * sizeof (MonoType*));
1911 sig_reg = mono_alloc_ireg (cfg);
1912 MONO_EMIT_NEW_SIGNATURECONST (cfg, sig_reg, tmp_sig);
1914 MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STORE_MEMBASE_REG, AMD64_RSP, cinfo->sig_cookie.offset, sig_reg);
1918 static inline LLVMArgStorage
1919 arg_storage_to_llvm_arg_storage (MonoCompile *cfg, ArgStorage storage)
1923 return LLVMArgInIReg;
1926 case ArgGSharedVtInReg:
1927 case ArgGSharedVtOnStack:
1928 return LLVMArgGSharedVt;
1930 g_assert_not_reached ();
1936 mono_arch_get_llvm_call_info (MonoCompile *cfg, MonoMethodSignature *sig)
1942 LLVMCallInfo *linfo;
1943 MonoType *t, *sig_ret;
1945 n = sig->param_count + sig->hasthis;
1946 sig_ret = mini_get_underlying_type (sig->ret);
1948 cinfo = get_call_info (cfg->mempool, sig);
1950 linfo = mono_mempool_alloc0 (cfg->mempool, sizeof (LLVMCallInfo) + (sizeof (LLVMArgInfo) * n));
1953 * LLVM always uses the native ABI while we use our own ABI, the
1954 * only difference is the handling of vtypes:
1955 * - we only pass/receive them in registers in some cases, and only
1956 * in 1 or 2 integer registers.
1958 switch (cinfo->ret.storage) {
1960 linfo->ret.storage = LLVMArgNone;
1963 case ArgInFloatSSEReg:
1964 case ArgInDoubleSSEReg:
1965 linfo->ret.storage = LLVMArgNormal;
1967 case ArgValuetypeInReg: {
1968 ainfo = &cinfo->ret;
1971 (ainfo->pair_storage [0] == ArgInFloatSSEReg || ainfo->pair_storage [0] == ArgInDoubleSSEReg ||
1972 ainfo->pair_storage [1] == ArgInFloatSSEReg || ainfo->pair_storage [1] == ArgInDoubleSSEReg)) {
1973 cfg->exception_message = g_strdup ("pinvoke + vtype ret");
1974 cfg->disable_llvm = TRUE;
1978 linfo->ret.storage = LLVMArgVtypeInReg;
1979 for (j = 0; j < 2; ++j)
1980 linfo->ret.pair_storage [j] = arg_storage_to_llvm_arg_storage (cfg, ainfo->pair_storage [j]);
1983 case ArgValuetypeAddrInIReg:
1984 case ArgGsharedvtVariableInReg:
1985 /* Vtype returned using a hidden argument */
1986 linfo->ret.storage = LLVMArgVtypeRetAddr;
1987 linfo->vret_arg_index = cinfo->vret_arg_index;
1990 g_assert_not_reached ();
1994 for (i = 0; i < n; ++i) {
1995 ainfo = cinfo->args + i;
1997 if (i >= sig->hasthis)
1998 t = sig->params [i - sig->hasthis];
2000 t = &mono_defaults.int_class->byval_arg;
2001 t = mini_type_get_underlying_type (t);
2003 linfo->args [i].storage = LLVMArgNone;
2005 switch (ainfo->storage) {
2007 linfo->args [i].storage = LLVMArgNormal;
2009 case ArgInDoubleSSEReg:
2010 case ArgInFloatSSEReg:
2011 linfo->args [i].storage = LLVMArgNormal;
2014 if (MONO_TYPE_ISSTRUCT (t))
2015 linfo->args [i].storage = LLVMArgVtypeByVal;
2017 linfo->args [i].storage = LLVMArgNormal;
2019 case ArgValuetypeInReg:
2021 (ainfo->pair_storage [0] == ArgInFloatSSEReg || ainfo->pair_storage [0] == ArgInDoubleSSEReg ||
2022 ainfo->pair_storage [1] == ArgInFloatSSEReg || ainfo->pair_storage [1] == ArgInDoubleSSEReg)) {
2023 cfg->exception_message = g_strdup ("pinvoke + vtypes");
2024 cfg->disable_llvm = TRUE;
2028 linfo->args [i].storage = LLVMArgVtypeInReg;
2029 for (j = 0; j < 2; ++j)
2030 linfo->args [i].pair_storage [j] = arg_storage_to_llvm_arg_storage (cfg, ainfo->pair_storage [j]);
2032 case ArgGSharedVtInReg:
2033 case ArgGSharedVtOnStack:
2034 linfo->args [i].storage = LLVMArgGSharedVt;
2037 cfg->exception_message = g_strdup ("ainfo->storage");
2038 cfg->disable_llvm = TRUE;
2048 mono_arch_emit_call (MonoCompile *cfg, MonoCallInst *call)
2051 MonoMethodSignature *sig;
2057 sig = call->signature;
2058 n = sig->param_count + sig->hasthis;
2060 cinfo = get_call_info (cfg->mempool, sig);
2064 if (COMPILE_LLVM (cfg)) {
2065 /* We shouldn't be called in the llvm case */
2066 cfg->disable_llvm = TRUE;
2071 * Emit all arguments which are passed on the stack to prevent register
2072 * allocation problems.
2074 for (i = 0; i < n; ++i) {
2076 ainfo = cinfo->args + i;
2078 in = call->args [i];
2080 if (sig->hasthis && i == 0)
2081 t = &mono_defaults.object_class->byval_arg;
2083 t = sig->params [i - sig->hasthis];
2085 t = mini_get_underlying_type (t);
2086 //XXX what about ArgGSharedVtOnStack here?
2087 if (ainfo->storage == ArgOnStack && !MONO_TYPE_ISSTRUCT (t) && !call->tail_call) {
2089 if (t->type == MONO_TYPE_R4)
2090 MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STORER4_MEMBASE_REG, AMD64_RSP, ainfo->offset, in->dreg);
2091 else if (t->type == MONO_TYPE_R8)
2092 MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STORER8_MEMBASE_REG, AMD64_RSP, ainfo->offset, in->dreg);
2094 MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STORE_MEMBASE_REG, AMD64_RSP, ainfo->offset, in->dreg);
2096 MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STORE_MEMBASE_REG, AMD64_RSP, ainfo->offset, in->dreg);
2098 if (cfg->compute_gc_maps) {
2101 EMIT_NEW_GC_PARAM_SLOT_LIVENESS_DEF (cfg, def, ainfo->offset, t);
2107 * Emit all parameters passed in registers in non-reverse order for better readability
2108 * and to help the optimization in emit_prolog ().
2110 for (i = 0; i < n; ++i) {
2111 ainfo = cinfo->args + i;
2113 in = call->args [i];
2115 if (ainfo->storage == ArgInIReg)
2116 add_outarg_reg (cfg, call, ainfo->storage, ainfo->reg, in);
2119 for (i = n - 1; i >= 0; --i) {
2122 ainfo = cinfo->args + i;
2124 in = call->args [i];
2126 if (sig->hasthis && i == 0)
2127 t = &mono_defaults.object_class->byval_arg;
2129 t = sig->params [i - sig->hasthis];
2130 t = mini_get_underlying_type (t);
2132 switch (ainfo->storage) {
2136 case ArgInFloatSSEReg:
2137 case ArgInDoubleSSEReg:
2138 add_outarg_reg (cfg, call, ainfo->storage, ainfo->reg, in);
2141 case ArgValuetypeInReg:
2142 case ArgValuetypeAddrInIReg:
2143 case ArgValuetypeAddrOnStack:
2144 case ArgGSharedVtInReg:
2145 case ArgGSharedVtOnStack: {
2146 if (ainfo->storage == ArgOnStack && !MONO_TYPE_ISSTRUCT (t) && !call->tail_call)
2147 /* Already emitted above */
2149 //FIXME what about ArgGSharedVtOnStack ?
2150 if (ainfo->storage == ArgOnStack && call->tail_call) {
2151 MonoInst *call_inst = (MonoInst*)call;
2152 cfg->args [i]->flags |= MONO_INST_VOLATILE;
2153 EMIT_NEW_ARGSTORE (cfg, call_inst, i, in);
2161 size = mono_type_native_stack_size (t, &align);
2164 * Other backends use mono_type_stack_size (), but that
2165 * aligns the size to 8, which is larger than the size of
2166 * the source, leading to reads of invalid memory if the
2167 * source is at the end of address space.
2169 size = mono_class_value_size (mono_class_from_mono_type (t), &align);
2172 if (size >= 10000) {
2173 /* Avoid asserts in emit_memcpy () */
2174 mono_cfg_set_exception_invalid_program (cfg, g_strdup_printf ("Passing an argument of size '%d'.", size));
2175 /* Continue normally */
2178 if (size > 0 || ainfo->pass_empty_struct) {
2179 MONO_INST_NEW (cfg, arg, OP_OUTARG_VT);
2180 arg->sreg1 = in->dreg;
2181 arg->klass = mono_class_from_mono_type (t);
2182 arg->backend.size = size;
2183 arg->inst_p0 = call;
2184 arg->inst_p1 = mono_mempool_alloc (cfg->mempool, sizeof (ArgInfo));
2185 memcpy (arg->inst_p1, ainfo, sizeof (ArgInfo));
2187 MONO_ADD_INS (cfg->cbb, arg);
2192 g_assert_not_reached ();
2195 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG) && (i == sig->sentinelpos))
2196 /* Emit the signature cookie just before the implicit arguments */
2197 emit_sig_cookie (cfg, call, cinfo);
2200 /* Handle the case where there are no implicit arguments */
2201 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG) && (n == sig->sentinelpos))
2202 emit_sig_cookie (cfg, call, cinfo);
2204 switch (cinfo->ret.storage) {
2205 case ArgValuetypeInReg:
2206 if (cinfo->ret.pair_storage [0] == ArgInIReg && cinfo->ret.pair_storage [1] == ArgNone) {
2208 * Tell the JIT to use a more efficient calling convention: call using
2209 * OP_CALL, compute the result location after the call, and save the
2212 call->vret_in_reg = TRUE;
2214 * Nullify the instruction computing the vret addr to enable
2215 * future optimizations.
2218 NULLIFY_INS (call->vret_var);
2220 if (call->tail_call)
2223 * The valuetype is in RAX:RDX after the call, need to be copied to
2224 * the stack. Push the address here, so the call instruction can
2227 if (!cfg->arch.vret_addr_loc) {
2228 cfg->arch.vret_addr_loc = mono_compile_create_var (cfg, &mono_defaults.int_class->byval_arg, OP_LOCAL);
2229 /* Prevent it from being register allocated or optimized away */
2230 ((MonoInst*)cfg->arch.vret_addr_loc)->flags |= MONO_INST_VOLATILE;
2233 MONO_EMIT_NEW_UNALU (cfg, OP_MOVE, ((MonoInst*)cfg->arch.vret_addr_loc)->dreg, call->vret_var->dreg);
2236 case ArgValuetypeAddrInIReg:
2237 case ArgGsharedvtVariableInReg: {
2239 MONO_INST_NEW (cfg, vtarg, OP_MOVE);
2240 vtarg->sreg1 = call->vret_var->dreg;
2241 vtarg->dreg = mono_alloc_preg (cfg);
2242 MONO_ADD_INS (cfg->cbb, vtarg);
2244 mono_call_inst_add_outarg_reg (cfg, call, vtarg->dreg, cinfo->ret.reg, FALSE);
2251 if (cfg->method->save_lmf) {
2252 MONO_INST_NEW (cfg, arg, OP_AMD64_SAVE_SP_TO_LMF);
2253 MONO_ADD_INS (cfg->cbb, arg);
2256 call->stack_usage = cinfo->stack_usage;
2260 mono_arch_emit_outarg_vt (MonoCompile *cfg, MonoInst *ins, MonoInst *src)
2263 MonoCallInst *call = (MonoCallInst*)ins->inst_p0;
2264 ArgInfo *ainfo = (ArgInfo*)ins->inst_p1;
2265 int size = ins->backend.size;
2267 switch (ainfo->storage) {
2268 case ArgValuetypeInReg: {
2272 for (part = 0; part < 2; ++part) {
2273 if (ainfo->pair_storage [part] == ArgNone)
2276 if (ainfo->pass_empty_struct) {
2277 //Pass empty struct value as 0 on platforms representing empty structs as 1 byte.
2278 NEW_ICONST (cfg, load, 0);
2281 MONO_INST_NEW (cfg, load, arg_storage_to_load_membase (ainfo->pair_storage [part]));
2282 load->inst_basereg = src->dreg;
2283 load->inst_offset = part * sizeof(mgreg_t);
2285 switch (ainfo->pair_storage [part]) {
2287 load->dreg = mono_alloc_ireg (cfg);
2289 case ArgInDoubleSSEReg:
2290 case ArgInFloatSSEReg:
2291 load->dreg = mono_alloc_freg (cfg);
2294 g_assert_not_reached ();
2298 MONO_ADD_INS (cfg->cbb, load);
2300 add_outarg_reg (cfg, call, ainfo->pair_storage [part], ainfo->pair_regs [part], load);
2304 case ArgValuetypeAddrInIReg:
2305 case ArgValuetypeAddrOnStack: {
2306 MonoInst *vtaddr, *load;
2308 g_assert (ainfo->storage == ArgValuetypeAddrInIReg || (ainfo->storage == ArgValuetypeAddrOnStack && ainfo->pair_storage [0] == ArgNone));
2310 vtaddr = mono_compile_create_var (cfg, &ins->klass->byval_arg, OP_LOCAL);
2312 MONO_INST_NEW (cfg, load, OP_LDADDR);
2313 cfg->has_indirection = TRUE;
2314 load->inst_p0 = vtaddr;
2315 vtaddr->flags |= MONO_INST_INDIRECT;
2316 load->type = STACK_MP;
2317 load->klass = vtaddr->klass;
2318 load->dreg = mono_alloc_ireg (cfg);
2319 MONO_ADD_INS (cfg->cbb, load);
2320 mini_emit_memcpy (cfg, load->dreg, 0, src->dreg, 0, size, 4);
2322 if (ainfo->pair_storage [0] == ArgInIReg) {
2323 MONO_INST_NEW (cfg, arg, OP_X86_LEA_MEMBASE);
2324 arg->dreg = mono_alloc_ireg (cfg);
2325 arg->sreg1 = load->dreg;
2327 MONO_ADD_INS (cfg->cbb, arg);
2328 mono_call_inst_add_outarg_reg (cfg, call, arg->dreg, ainfo->pair_regs [0], FALSE);
2330 MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STORE_MEMBASE_REG, AMD64_RSP, ainfo->offset, load->dreg);
2334 case ArgGSharedVtInReg:
2336 mono_call_inst_add_outarg_reg (cfg, call, src->dreg, ainfo->reg, FALSE);
2338 case ArgGSharedVtOnStack:
2339 MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STORE_MEMBASE_REG, AMD64_RSP, ainfo->offset, src->dreg);
2343 int dreg = mono_alloc_ireg (cfg);
2345 MONO_EMIT_NEW_LOAD_MEMBASE (cfg, dreg, src->dreg, 0);
2346 MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STORE_MEMBASE_REG, AMD64_RSP, ainfo->offset, dreg);
2347 } else if (size <= 40) {
2348 mini_emit_memcpy (cfg, AMD64_RSP, ainfo->offset, src->dreg, 0, size, 4);
2350 // FIXME: Code growth
2351 mini_emit_memcpy (cfg, AMD64_RSP, ainfo->offset, src->dreg, 0, size, 4);
2354 if (cfg->compute_gc_maps) {
2356 EMIT_NEW_GC_PARAM_SLOT_LIVENESS_DEF (cfg, def, ainfo->offset, &ins->klass->byval_arg);
2362 mono_arch_emit_setret (MonoCompile *cfg, MonoMethod *method, MonoInst *val)
2364 MonoType *ret = mini_get_underlying_type (mono_method_signature (method)->ret);
2366 if (ret->type == MONO_TYPE_R4) {
2367 if (COMPILE_LLVM (cfg))
2368 MONO_EMIT_NEW_UNALU (cfg, OP_FMOVE, cfg->ret->dreg, val->dreg);
2370 MONO_EMIT_NEW_UNALU (cfg, OP_AMD64_SET_XMMREG_R4, cfg->ret->dreg, val->dreg);
2372 } else if (ret->type == MONO_TYPE_R8) {
2373 MONO_EMIT_NEW_UNALU (cfg, OP_FMOVE, cfg->ret->dreg, val->dreg);
2377 MONO_EMIT_NEW_UNALU (cfg, OP_MOVE, cfg->ret->dreg, val->dreg);
2380 #endif /* DISABLE_JIT */
2382 #define EMIT_COND_BRANCH(ins,cond,sign) \
2383 if (ins->inst_true_bb->native_offset) { \
2384 x86_branch (code, cond, cfg->native_code + ins->inst_true_bb->native_offset, sign); \
2386 mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_BB, ins->inst_true_bb); \
2387 if ((cfg->opt & MONO_OPT_BRANCH) && \
2388 x86_is_imm8 (ins->inst_true_bb->max_offset - offset)) \
2389 x86_branch8 (code, cond, 0, sign); \
2391 x86_branch32 (code, cond, 0, sign); \
2395 MonoMethodSignature *sig;
2400 dyn_call_supported (MonoMethodSignature *sig, CallInfo *cinfo)
2404 switch (cinfo->ret.storage) {
2407 case ArgInFloatSSEReg:
2408 case ArgInDoubleSSEReg:
2409 case ArgValuetypeAddrInIReg:
2410 case ArgValuetypeInReg:
2416 for (i = 0; i < cinfo->nargs; ++i) {
2417 ArgInfo *ainfo = &cinfo->args [i];
2418 switch (ainfo->storage) {
2420 case ArgInFloatSSEReg:
2421 case ArgInDoubleSSEReg:
2422 case ArgValuetypeInReg:
2425 if (!(ainfo->offset + (ainfo->arg_size / 8) <= DYN_CALL_STACK_ARGS))
2437 * mono_arch_dyn_call_prepare:
2439 * Return a pointer to an arch-specific structure which contains information
2440 * needed by mono_arch_get_dyn_call_args (). Return NULL if OP_DYN_CALL is not
2441 * supported for SIG.
2442 * This function is equivalent to ffi_prep_cif in libffi.
2445 mono_arch_dyn_call_prepare (MonoMethodSignature *sig)
2447 ArchDynCallInfo *info;
2450 cinfo = get_call_info (NULL, sig);
2452 if (!dyn_call_supported (sig, cinfo)) {
2457 info = g_new0 (ArchDynCallInfo, 1);
2458 // FIXME: Preprocess the info to speed up get_dyn_call_args ().
2460 info->cinfo = cinfo;
2462 return (MonoDynCallInfo*)info;
2466 * mono_arch_dyn_call_free:
2468 * Free a MonoDynCallInfo structure.
2471 mono_arch_dyn_call_free (MonoDynCallInfo *info)
2473 ArchDynCallInfo *ainfo = (ArchDynCallInfo*)info;
2475 g_free (ainfo->cinfo);
2479 #define PTR_TO_GREG(ptr) (mgreg_t)(ptr)
2480 #define GREG_TO_PTR(greg) (gpointer)(greg)
2483 * mono_arch_get_start_dyn_call:
2485 * Convert the arguments ARGS to a format which can be passed to OP_DYN_CALL, and
2486 * store the result into BUF.
2487 * ARGS should be an array of pointers pointing to the arguments.
2488 * RET should point to a memory buffer large enought to hold the result of the
2490 * This function should be as fast as possible, any work which does not depend
2491 * on the actual values of the arguments should be done in
2492 * mono_arch_dyn_call_prepare ().
2493 * start_dyn_call + OP_DYN_CALL + finish_dyn_call is equivalent to ffi_call in
2497 mono_arch_start_dyn_call (MonoDynCallInfo *info, gpointer **args, guint8 *ret, guint8 *buf, int buf_len)
2499 ArchDynCallInfo *dinfo = (ArchDynCallInfo*)info;
2500 DynCallArgs *p = (DynCallArgs*)buf;
2501 int arg_index, greg, freg, i, pindex;
2502 MonoMethodSignature *sig = dinfo->sig;
2503 int buffer_offset = 0;
2504 static int param_reg_to_index [16];
2505 static gboolean param_reg_to_index_inited;
2507 if (!param_reg_to_index_inited) {
2508 for (i = 0; i < PARAM_REGS; ++i)
2509 param_reg_to_index [param_regs [i]] = i;
2510 mono_memory_barrier ();
2511 param_reg_to_index_inited = 1;
2514 g_assert (buf_len >= sizeof (DynCallArgs));
2524 if (sig->hasthis || dinfo->cinfo->vret_arg_index == 1) {
2525 p->regs [greg ++] = PTR_TO_GREG(*(args [arg_index ++]));
2530 if (dinfo->cinfo->ret.storage == ArgValuetypeAddrInIReg || dinfo->cinfo->ret.storage == ArgGsharedvtVariableInReg)
2531 p->regs [greg ++] = PTR_TO_GREG(ret);
2533 for (; pindex < sig->param_count; pindex++) {
2534 MonoType *t = mini_get_underlying_type (sig->params [pindex]);
2535 gpointer *arg = args [arg_index ++];
2536 ArgInfo *ainfo = &dinfo->cinfo->args [pindex + sig->hasthis];
2539 if (ainfo->storage == ArgOnStack) {
2540 slot = PARAM_REGS + (ainfo->offset / sizeof (mgreg_t));
2542 slot = param_reg_to_index [ainfo->reg];
2546 p->regs [slot] = PTR_TO_GREG(*(arg));
2552 case MONO_TYPE_STRING:
2553 case MONO_TYPE_CLASS:
2554 case MONO_TYPE_ARRAY:
2555 case MONO_TYPE_SZARRAY:
2556 case MONO_TYPE_OBJECT:
2560 #if !defined(__mono_ilp32__)
2564 p->regs [slot] = PTR_TO_GREG(*(arg));
2566 #if defined(__mono_ilp32__)
2569 p->regs [slot] = *(guint64*)(arg);
2573 p->regs [slot] = *(guint8*)(arg);
2576 p->regs [slot] = *(gint8*)(arg);
2579 p->regs [slot] = *(gint16*)(arg);
2582 p->regs [slot] = *(guint16*)(arg);
2585 p->regs [slot] = *(gint32*)(arg);
2588 p->regs [slot] = *(guint32*)(arg);
2590 case MONO_TYPE_R4: {
2593 *(float*)&d = *(float*)(arg);
2595 p->fregs [freg ++] = d;
2600 p->fregs [freg ++] = *(double*)(arg);
2602 case MONO_TYPE_GENERICINST:
2603 if (MONO_TYPE_IS_REFERENCE (t)) {
2604 p->regs [slot] = PTR_TO_GREG(*(arg));
2606 } else if (t->type == MONO_TYPE_GENERICINST && mono_class_is_nullable (mono_class_from_mono_type (t))) {
2607 MonoClass *klass = mono_class_from_mono_type (t);
2608 guint8 *nullable_buf;
2611 size = mono_class_value_size (klass, NULL);
2612 nullable_buf = p->buffer + buffer_offset;
2613 buffer_offset += size;
2614 g_assert (buffer_offset <= 256);
2616 /* The argument pointed to by arg is either a boxed vtype or null */
2617 mono_nullable_init (nullable_buf, (MonoObject*)arg, klass);
2619 arg = (gpointer*)nullable_buf;
2625 case MONO_TYPE_VALUETYPE: {
2626 switch (ainfo->storage) {
2627 case ArgValuetypeInReg:
2628 for (i = 0; i < 2; ++i) {
2629 switch (ainfo->pair_storage [i]) {
2633 slot = param_reg_to_index [ainfo->pair_regs [i]];
2634 p->regs [slot] = ((mgreg_t*)(arg))[i];
2636 case ArgInDoubleSSEReg:
2638 p->fregs [ainfo->pair_regs [i]] = ((double*)(arg))[i];
2641 g_assert_not_reached ();
2647 for (i = 0; i < ainfo->arg_size / 8; ++i)
2648 p->regs [slot + i] = ((mgreg_t*)(arg))[i];
2651 g_assert_not_reached ();
2657 g_assert_not_reached ();
2663 * mono_arch_finish_dyn_call:
2665 * Store the result of a dyn call into the return value buffer passed to
2666 * start_dyn_call ().
2667 * This function should be as fast as possible, any work which does not depend
2668 * on the actual values of the arguments should be done in
2669 * mono_arch_dyn_call_prepare ().
2672 mono_arch_finish_dyn_call (MonoDynCallInfo *info, guint8 *buf)
2674 ArchDynCallInfo *dinfo = (ArchDynCallInfo*)info;
2675 MonoMethodSignature *sig = dinfo->sig;
2676 DynCallArgs *dargs = (DynCallArgs*)buf;
2677 guint8 *ret = dargs->ret;
2678 mgreg_t res = dargs->res;
2679 MonoType *sig_ret = mini_get_underlying_type (sig->ret);
2682 switch (sig_ret->type) {
2683 case MONO_TYPE_VOID:
2684 *(gpointer*)ret = NULL;
2686 case MONO_TYPE_STRING:
2687 case MONO_TYPE_CLASS:
2688 case MONO_TYPE_ARRAY:
2689 case MONO_TYPE_SZARRAY:
2690 case MONO_TYPE_OBJECT:
2694 *(gpointer*)ret = GREG_TO_PTR(res);
2700 *(guint8*)ret = res;
2703 *(gint16*)ret = res;
2706 *(guint16*)ret = res;
2709 *(gint32*)ret = res;
2712 *(guint32*)ret = res;
2715 *(gint64*)ret = res;
2718 *(guint64*)ret = res;
2721 *(float*)ret = *(float*)&(dargs->fregs [0]);
2724 *(double*)ret = dargs->fregs [0];
2726 case MONO_TYPE_GENERICINST:
2727 if (MONO_TYPE_IS_REFERENCE (sig_ret)) {
2728 *(gpointer*)ret = GREG_TO_PTR(res);
2733 case MONO_TYPE_VALUETYPE:
2734 if (dinfo->cinfo->ret.storage == ArgValuetypeAddrInIReg || dinfo->cinfo->ret.storage == ArgGsharedvtVariableInReg) {
2737 ArgInfo *ainfo = &dinfo->cinfo->ret;
2739 g_assert (ainfo->storage == ArgValuetypeInReg);
2741 for (i = 0; i < 2; ++i) {
2742 switch (ainfo->pair_storage [0]) {
2744 ((mgreg_t*)ret)[i] = res;
2746 case ArgInDoubleSSEReg:
2747 ((double*)ret)[i] = dargs->fregs [i];
2752 g_assert_not_reached ();
2759 g_assert_not_reached ();
2763 /* emit an exception if condition is fail */
2764 #define EMIT_COND_SYSTEM_EXCEPTION(cond,signed,exc_name) \
2766 MonoInst *tins = mono_branch_optimize_exception_target (cfg, bb, exc_name); \
2767 if (tins == NULL) { \
2768 mono_add_patch_info (cfg, code - cfg->native_code, \
2769 MONO_PATCH_INFO_EXC, exc_name); \
2770 x86_branch32 (code, cond, 0, signed); \
2772 EMIT_COND_BRANCH (tins, cond, signed); \
2776 #define EMIT_FPCOMPARE(code) do { \
2777 amd64_fcompp (code); \
2778 amd64_fnstsw (code); \
2781 #define EMIT_SSE2_FPFUNC(code, op, dreg, sreg1) do { \
2782 amd64_movsd_membase_reg (code, AMD64_RSP, -8, (sreg1)); \
2783 amd64_fld_membase (code, AMD64_RSP, -8, TRUE); \
2784 amd64_ ##op (code); \
2785 amd64_fst_membase (code, AMD64_RSP, -8, TRUE, TRUE); \
2786 amd64_movsd_reg_membase (code, (dreg), AMD64_RSP, -8); \
2790 emit_call_body (MonoCompile *cfg, guint8 *code, MonoJumpInfoType patch_type, gconstpointer data)
2792 gboolean no_patch = FALSE;
2795 * FIXME: Add support for thunks
2798 gboolean near_call = FALSE;
2801 * Indirect calls are expensive so try to make a near call if possible.
2802 * The caller memory is allocated by the code manager so it is
2803 * guaranteed to be at a 32 bit offset.
2806 if (patch_type != MONO_PATCH_INFO_ABS) {
2807 /* The target is in memory allocated using the code manager */
2810 if ((patch_type == MONO_PATCH_INFO_METHOD) || (patch_type == MONO_PATCH_INFO_METHOD_JUMP)) {
2811 if (((MonoMethod*)data)->klass->image->aot_module)
2812 /* The callee might be an AOT method */
2814 if (((MonoMethod*)data)->dynamic)
2815 /* The target is in malloc-ed memory */
2819 if (patch_type == MONO_PATCH_INFO_INTERNAL_METHOD) {
2821 * The call might go directly to a native function without
2824 MonoJitICallInfo *mi = mono_find_jit_icall_by_name ((const char *)data);
2826 gconstpointer target = mono_icall_get_wrapper (mi);
2827 if ((((guint64)target) >> 32) != 0)
2833 MonoJumpInfo *jinfo = NULL;
2835 if (cfg->abs_patches)
2836 jinfo = (MonoJumpInfo *)g_hash_table_lookup (cfg->abs_patches, data);
2838 if (jinfo->type == MONO_PATCH_INFO_JIT_ICALL_ADDR) {
2839 MonoJitICallInfo *mi = mono_find_jit_icall_by_name (jinfo->data.name);
2840 if (mi && (((guint64)mi->func) >> 32) == 0)
2845 * This is not really an optimization, but required because the
2846 * generic class init trampolines use R11 to pass the vtable.
2851 MonoJitICallInfo *info = mono_find_jit_icall_by_addr (data);
2853 if (info->func == info->wrapper) {
2855 if ((((guint64)info->func) >> 32) == 0)
2859 /* See the comment in mono_codegen () */
2860 if ((info->name [0] != 'v') || (strstr (info->name, "ves_array_new_va_") == NULL && strstr (info->name, "ves_array_element_address_") == NULL))
2864 else if ((((guint64)data) >> 32) == 0) {
2871 if (cfg->method->dynamic)
2872 /* These methods are allocated using malloc */
2875 #ifdef MONO_ARCH_NOMAP32BIT
2878 /* The 64bit XEN kernel does not honour the MAP_32BIT flag. (#522894) */
2879 if (optimize_for_xen)
2882 if (cfg->compile_aot) {
2889 * Align the call displacement to an address divisible by 4 so it does
2890 * not span cache lines. This is required for code patching to work on SMP
2893 if (!no_patch && ((guint32)(code + 1 - cfg->native_code) % 4) != 0) {
2894 guint32 pad_size = 4 - ((guint32)(code + 1 - cfg->native_code) % 4);
2895 amd64_padding (code, pad_size);
2897 mono_add_patch_info (cfg, code - cfg->native_code, patch_type, data);
2898 amd64_call_code (code, 0);
2901 mono_add_patch_info (cfg, code - cfg->native_code, patch_type, data);
2902 amd64_set_reg_template (code, GP_SCRATCH_REG);
2903 amd64_call_reg (code, GP_SCRATCH_REG);
2910 static inline guint8*
2911 emit_call (MonoCompile *cfg, guint8 *code, MonoJumpInfoType patch_type, gconstpointer data, gboolean win64_adjust_stack)
2914 if (win64_adjust_stack)
2915 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, 32);
2917 code = emit_call_body (cfg, code, patch_type, data);
2919 if (win64_adjust_stack)
2920 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, 32);
2927 store_membase_imm_to_store_membase_reg (int opcode)
2930 case OP_STORE_MEMBASE_IMM:
2931 return OP_STORE_MEMBASE_REG;
2932 case OP_STOREI4_MEMBASE_IMM:
2933 return OP_STOREI4_MEMBASE_REG;
2934 case OP_STOREI8_MEMBASE_IMM:
2935 return OP_STOREI8_MEMBASE_REG;
2943 #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)))
2946 * mono_arch_peephole_pass_1:
2948 * Perform peephole opts which should/can be performed before local regalloc
2951 mono_arch_peephole_pass_1 (MonoCompile *cfg, MonoBasicBlock *bb)
2955 MONO_BB_FOR_EACH_INS_SAFE (bb, n, ins) {
2956 MonoInst *last_ins = mono_inst_prev (ins, FILTER_IL_SEQ_POINT);
2958 switch (ins->opcode) {
2962 if ((ins->sreg1 < MONO_MAX_IREGS) && (ins->dreg >= MONO_MAX_IREGS) && (ins->inst_imm > 0)) {
2964 * X86_LEA is like ADD, but doesn't have the
2965 * sreg1==dreg restriction. inst_imm > 0 is needed since LEA sign-extends
2966 * its operand to 64 bit.
2968 ins->opcode = OP_X86_LEA_MEMBASE;
2969 ins->inst_basereg = ins->sreg1;
2974 if ((ins->sreg1 == ins->sreg2) && (ins->sreg1 == ins->dreg)) {
2978 * Replace STORE_MEMBASE_IMM 0 with STORE_MEMBASE_REG since
2979 * the latter has length 2-3 instead of 6 (reverse constant
2980 * propagation). These instruction sequences are very common
2981 * in the initlocals bblock.
2983 for (ins2 = ins->next; ins2; ins2 = ins2->next) {
2984 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)) {
2985 ins2->opcode = store_membase_imm_to_store_membase_reg (ins2->opcode);
2986 ins2->sreg1 = ins->dreg;
2987 } 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)) {
2989 } else if (((ins2->opcode == OP_ICONST) || (ins2->opcode == OP_I8CONST)) && (ins2->dreg == ins->dreg) && (ins2->inst_c0 == 0)) {
2992 } else if (ins2->opcode == OP_IL_SEQ_POINT) {
3000 case OP_COMPARE_IMM:
3001 case OP_LCOMPARE_IMM:
3002 /* OP_COMPARE_IMM (reg, 0)
3004 * OP_AMD64_TEST_NULL (reg)
3007 ins->opcode = OP_AMD64_TEST_NULL;
3009 case OP_ICOMPARE_IMM:
3011 ins->opcode = OP_X86_TEST_NULL;
3013 case OP_AMD64_ICOMPARE_MEMBASE_IMM:
3015 * OP_STORE_MEMBASE_REG reg, offset(basereg)
3016 * OP_X86_COMPARE_MEMBASE_IMM offset(basereg), imm
3018 * OP_STORE_MEMBASE_REG reg, offset(basereg)
3019 * OP_COMPARE_IMM reg, imm
3021 * Note: if imm = 0 then OP_COMPARE_IMM replaced with OP_X86_TEST_NULL
3023 if (last_ins && (last_ins->opcode == OP_STOREI4_MEMBASE_REG) &&
3024 ins->inst_basereg == last_ins->inst_destbasereg &&
3025 ins->inst_offset == last_ins->inst_offset) {
3026 ins->opcode = OP_ICOMPARE_IMM;
3027 ins->sreg1 = last_ins->sreg1;
3029 /* check if we can remove cmp reg,0 with test null */
3031 ins->opcode = OP_X86_TEST_NULL;
3037 mono_peephole_ins (bb, ins);
3042 mono_arch_peephole_pass_2 (MonoCompile *cfg, MonoBasicBlock *bb)
3046 MONO_BB_FOR_EACH_INS_SAFE (bb, n, ins) {
3047 switch (ins->opcode) {
3050 MonoInst *next = mono_inst_next (ins, FILTER_IL_SEQ_POINT);
3051 /* reg = 0 -> XOR (reg, reg) */
3052 /* XOR sets cflags on x86, so we cant do it always */
3053 if (ins->inst_c0 == 0 && (!next || (next && INST_IGNORES_CFLAGS (next->opcode)))) {
3054 ins->opcode = OP_LXOR;
3055 ins->sreg1 = ins->dreg;
3056 ins->sreg2 = ins->dreg;
3064 * Use IXOR to avoid a rex prefix if possible. The cpu will sign extend the
3065 * 0 result into 64 bits.
3067 if ((ins->sreg1 == ins->sreg2) && (ins->sreg1 == ins->dreg)) {
3068 ins->opcode = OP_IXOR;
3072 if ((ins->sreg1 == ins->sreg2) && (ins->sreg1 == ins->dreg)) {
3076 * Replace STORE_MEMBASE_IMM 0 with STORE_MEMBASE_REG since
3077 * the latter has length 2-3 instead of 6 (reverse constant
3078 * propagation). These instruction sequences are very common
3079 * in the initlocals bblock.
3081 for (ins2 = ins->next; ins2; ins2 = ins2->next) {
3082 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)) {
3083 ins2->opcode = store_membase_imm_to_store_membase_reg (ins2->opcode);
3084 ins2->sreg1 = ins->dreg;
3085 } else if ((ins2->opcode == OP_STOREI1_MEMBASE_IMM) || (ins2->opcode == OP_STOREI2_MEMBASE_IMM) || (ins2->opcode == OP_STOREI4_MEMBASE_REG) || (ins2->opcode == OP_STOREI8_MEMBASE_REG) || (ins2->opcode == OP_STORE_MEMBASE_REG) || (ins2->opcode == OP_LIVERANGE_START) || (ins2->opcode == OP_GC_LIVENESS_DEF) || (ins2->opcode == OP_GC_LIVENESS_USE)) {
3087 } else if (((ins2->opcode == OP_ICONST) || (ins2->opcode == OP_I8CONST)) && (ins2->dreg == ins->dreg) && (ins2->inst_c0 == 0)) {
3090 } else if (ins2->opcode == OP_IL_SEQ_POINT) {
3099 if ((ins->inst_imm == 1) && (ins->dreg == ins->sreg1))
3100 ins->opcode = OP_X86_INC_REG;
3103 if ((ins->inst_imm == 1) && (ins->dreg == ins->sreg1))
3104 ins->opcode = OP_X86_DEC_REG;
3108 mono_peephole_ins (bb, ins);
3112 #define NEW_INS(cfg,ins,dest,op) do { \
3113 MONO_INST_NEW ((cfg), (dest), (op)); \
3114 (dest)->cil_code = (ins)->cil_code; \
3115 mono_bblock_insert_before_ins (bb, ins, (dest)); \
3119 * mono_arch_lowering_pass:
3121 * Converts complex opcodes into simpler ones so that each IR instruction
3122 * corresponds to one machine instruction.
3125 mono_arch_lowering_pass (MonoCompile *cfg, MonoBasicBlock *bb)
3127 MonoInst *ins, *n, *temp;
3130 * FIXME: Need to add more instructions, but the current machine
3131 * description can't model some parts of the composite instructions like
3134 MONO_BB_FOR_EACH_INS_SAFE (bb, n, ins) {
3135 switch (ins->opcode) {
3139 case OP_IDIV_UN_IMM:
3140 case OP_IREM_UN_IMM:
3143 mono_decompose_op_imm (cfg, bb, ins);
3145 case OP_COMPARE_IMM:
3146 case OP_LCOMPARE_IMM:
3147 if (!amd64_use_imm32 (ins->inst_imm)) {
3148 NEW_INS (cfg, ins, temp, OP_I8CONST);
3149 temp->inst_c0 = ins->inst_imm;
3150 temp->dreg = mono_alloc_ireg (cfg);
3151 ins->opcode = OP_COMPARE;
3152 ins->sreg2 = temp->dreg;
3155 #ifndef __mono_ilp32__
3156 case OP_LOAD_MEMBASE:
3158 case OP_LOADI8_MEMBASE:
3159 /* Don't generate memindex opcodes (to simplify */
3160 /* read sandboxing) */
3161 if (!amd64_use_imm32 (ins->inst_offset)) {
3162 NEW_INS (cfg, ins, temp, OP_I8CONST);
3163 temp->inst_c0 = ins->inst_offset;
3164 temp->dreg = mono_alloc_ireg (cfg);
3165 ins->opcode = OP_AMD64_LOADI8_MEMINDEX;
3166 ins->inst_indexreg = temp->dreg;
3169 #ifndef __mono_ilp32__
3170 case OP_STORE_MEMBASE_IMM:
3172 case OP_STOREI8_MEMBASE_IMM:
3173 if (!amd64_use_imm32 (ins->inst_imm)) {
3174 NEW_INS (cfg, ins, temp, OP_I8CONST);
3175 temp->inst_c0 = ins->inst_imm;
3176 temp->dreg = mono_alloc_ireg (cfg);
3177 ins->opcode = OP_STOREI8_MEMBASE_REG;
3178 ins->sreg1 = temp->dreg;
3181 #ifdef MONO_ARCH_SIMD_INTRINSICS
3182 case OP_EXPAND_I1: {
3183 int temp_reg1 = mono_alloc_ireg (cfg);
3184 int temp_reg2 = mono_alloc_ireg (cfg);
3185 int original_reg = ins->sreg1;
3187 NEW_INS (cfg, ins, temp, OP_ICONV_TO_U1);
3188 temp->sreg1 = original_reg;
3189 temp->dreg = temp_reg1;
3191 NEW_INS (cfg, ins, temp, OP_SHL_IMM);
3192 temp->sreg1 = temp_reg1;
3193 temp->dreg = temp_reg2;
3196 NEW_INS (cfg, ins, temp, OP_LOR);
3197 temp->sreg1 = temp->dreg = temp_reg2;
3198 temp->sreg2 = temp_reg1;
3200 ins->opcode = OP_EXPAND_I2;
3201 ins->sreg1 = temp_reg2;
3210 bb->max_vreg = cfg->next_vreg;
3214 branch_cc_table [] = {
3215 X86_CC_EQ, X86_CC_GE, X86_CC_GT, X86_CC_LE, X86_CC_LT,
3216 X86_CC_NE, X86_CC_GE, X86_CC_GT, X86_CC_LE, X86_CC_LT,
3217 X86_CC_O, X86_CC_NO, X86_CC_C, X86_CC_NC
3220 /* Maps CMP_... constants to X86_CC_... constants */
3223 X86_CC_EQ, X86_CC_NE, X86_CC_LE, X86_CC_GE, X86_CC_LT, X86_CC_GT,
3224 X86_CC_LE, X86_CC_GE, X86_CC_LT, X86_CC_GT
3228 cc_signed_table [] = {
3229 TRUE, TRUE, TRUE, TRUE, TRUE, TRUE,
3230 FALSE, FALSE, FALSE, FALSE
3233 /*#include "cprop.c"*/
3235 static unsigned char*
3236 emit_float_to_int (MonoCompile *cfg, guchar *code, int dreg, int sreg, int size, gboolean is_signed)
3239 amd64_sse_cvttsd2si_reg_reg (code, dreg, sreg);
3241 amd64_sse_cvttsd2si_reg_reg_size (code, dreg, sreg, 4);
3244 amd64_widen_reg (code, dreg, dreg, is_signed, FALSE);
3246 amd64_widen_reg (code, dreg, dreg, is_signed, TRUE);
3250 static unsigned char*
3251 mono_emit_stack_alloc (MonoCompile *cfg, guchar *code, MonoInst* tree)
3253 int sreg = tree->sreg1;
3254 int need_touch = FALSE;
3256 #if defined(TARGET_WIN32)
3258 #elif defined(MONO_ARCH_SIGSEGV_ON_ALTSTACK)
3259 if (!tree->flags & MONO_INST_INIT)
3268 * If requested stack size is larger than one page,
3269 * perform stack-touch operation
3272 * Generate stack probe code.
3273 * Under Windows, it is necessary to allocate one page at a time,
3274 * "touching" stack after each successful sub-allocation. This is
3275 * because of the way stack growth is implemented - there is a
3276 * guard page before the lowest stack page that is currently commited.
3277 * Stack normally grows sequentially so OS traps access to the
3278 * guard page and commits more pages when needed.
3280 amd64_test_reg_imm (code, sreg, ~0xFFF);
3281 br[0] = code; x86_branch8 (code, X86_CC_Z, 0, FALSE);
3283 br[2] = code; /* loop */
3284 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, 0x1000);
3285 amd64_test_membase_reg (code, AMD64_RSP, 0, AMD64_RSP);
3286 amd64_alu_reg_imm (code, X86_SUB, sreg, 0x1000);
3287 amd64_alu_reg_imm (code, X86_CMP, sreg, 0x1000);
3288 br[3] = code; x86_branch8 (code, X86_CC_AE, 0, FALSE);
3289 amd64_patch (br[3], br[2]);
3290 amd64_test_reg_reg (code, sreg, sreg);
3291 br[4] = code; x86_branch8 (code, X86_CC_Z, 0, FALSE);
3292 amd64_alu_reg_reg (code, X86_SUB, AMD64_RSP, sreg);
3294 br[1] = code; x86_jump8 (code, 0);
3296 amd64_patch (br[0], code);
3297 amd64_alu_reg_reg (code, X86_SUB, AMD64_RSP, sreg);
3298 amd64_patch (br[1], code);
3299 amd64_patch (br[4], code);
3302 amd64_alu_reg_reg (code, X86_SUB, AMD64_RSP, tree->sreg1);
3304 if (tree->flags & MONO_INST_INIT) {
3306 if (tree->dreg != AMD64_RAX && sreg != AMD64_RAX) {
3307 amd64_push_reg (code, AMD64_RAX);
3310 if (tree->dreg != AMD64_RCX && sreg != AMD64_RCX) {
3311 amd64_push_reg (code, AMD64_RCX);
3314 if (tree->dreg != AMD64_RDI && sreg != AMD64_RDI) {
3315 amd64_push_reg (code, AMD64_RDI);
3319 amd64_shift_reg_imm (code, X86_SHR, sreg, 3);
3320 if (sreg != AMD64_RCX)
3321 amd64_mov_reg_reg (code, AMD64_RCX, sreg, 8);
3322 amd64_alu_reg_reg (code, X86_XOR, AMD64_RAX, AMD64_RAX);
3324 amd64_lea_membase (code, AMD64_RDI, AMD64_RSP, offset);
3325 if (cfg->param_area)
3326 amd64_alu_reg_imm (code, X86_ADD, AMD64_RDI, cfg->param_area);
3328 amd64_prefix (code, X86_REP_PREFIX);
3331 if (tree->dreg != AMD64_RDI && sreg != AMD64_RDI)
3332 amd64_pop_reg (code, AMD64_RDI);
3333 if (tree->dreg != AMD64_RCX && sreg != AMD64_RCX)
3334 amd64_pop_reg (code, AMD64_RCX);
3335 if (tree->dreg != AMD64_RAX && sreg != AMD64_RAX)
3336 amd64_pop_reg (code, AMD64_RAX);
3342 emit_move_return_value (MonoCompile *cfg, MonoInst *ins, guint8 *code)
3347 /* Move return value to the target register */
3348 /* FIXME: do this in the local reg allocator */
3349 switch (ins->opcode) {
3352 case OP_CALL_MEMBASE:
3355 case OP_LCALL_MEMBASE:
3356 g_assert (ins->dreg == AMD64_RAX);
3360 case OP_FCALL_MEMBASE: {
3361 MonoType *rtype = mini_get_underlying_type (((MonoCallInst*)ins)->signature->ret);
3362 if (rtype->type == MONO_TYPE_R4) {
3363 amd64_sse_cvtss2sd_reg_reg (code, ins->dreg, AMD64_XMM0);
3366 if (ins->dreg != AMD64_XMM0)
3367 amd64_sse_movsd_reg_reg (code, ins->dreg, AMD64_XMM0);
3373 case OP_RCALL_MEMBASE:
3374 if (ins->dreg != AMD64_XMM0)
3375 amd64_sse_movss_reg_reg (code, ins->dreg, AMD64_XMM0);
3379 case OP_VCALL_MEMBASE:
3382 case OP_VCALL2_MEMBASE:
3383 cinfo = get_call_info (cfg->mempool, ((MonoCallInst*)ins)->signature);
3384 if (cinfo->ret.storage == ArgValuetypeInReg) {
3385 MonoInst *loc = (MonoInst *)cfg->arch.vret_addr_loc;
3387 /* Load the destination address */
3388 g_assert (loc->opcode == OP_REGOFFSET);
3389 amd64_mov_reg_membase (code, AMD64_RCX, loc->inst_basereg, loc->inst_offset, sizeof(gpointer));
3391 for (quad = 0; quad < 2; quad ++) {
3392 switch (cinfo->ret.pair_storage [quad]) {
3394 amd64_mov_membase_reg (code, AMD64_RCX, (quad * sizeof(mgreg_t)), cinfo->ret.pair_regs [quad], sizeof(mgreg_t));
3396 case ArgInFloatSSEReg:
3397 amd64_movss_membase_reg (code, AMD64_RCX, (quad * 8), cinfo->ret.pair_regs [quad]);
3399 case ArgInDoubleSSEReg:
3400 amd64_movsd_membase_reg (code, AMD64_RCX, (quad * 8), cinfo->ret.pair_regs [quad]);
3415 #endif /* DISABLE_JIT */
3418 static int tls_gs_offset;
3422 mono_arch_have_fast_tls (void)
3425 static gboolean have_fast_tls = FALSE;
3426 static gboolean inited = FALSE;
3429 if (mini_get_debug_options ()->use_fallback_tls)
3433 return have_fast_tls;
3435 ins = (guint8*)pthread_getspecific;
3438 * We're looking for these two instructions:
3440 * mov %gs:[offset](,%rdi,8),%rax
3443 have_fast_tls = ins [0] == 0x65 &&
3453 tls_gs_offset = ins[5];
3456 * Apple now loads a different version of pthread_getspecific when launched from Xcode
3457 * For that version we're looking for these instructions:
3461 * mov %gs:[offset](,%rdi,8),%rax
3465 if (!have_fast_tls) {
3466 have_fast_tls = ins [0] == 0x55 &&
3481 tls_gs_offset = ins[9];
3485 return have_fast_tls;
3486 #elif defined(TARGET_ANDROID)
3489 if (mini_get_debug_options ()->use_fallback_tls)
3496 mono_amd64_get_tls_gs_offset (void)
3499 return tls_gs_offset;
3501 g_assert_not_reached ();
3507 * mono_amd64_emit_tls_get:
3508 * @code: buffer to store code to
3509 * @dreg: hard register where to place the result
3510 * @tls_offset: offset info
3512 * mono_amd64_emit_tls_get emits in @code the native code that puts in
3513 * the dreg register the item in the thread local storage identified
3516 * Returns: a pointer to the end of the stored code
3519 mono_amd64_emit_tls_get (guint8* code, int dreg, int tls_offset)
3522 if (tls_offset < 64) {
3523 x86_prefix (code, X86_GS_PREFIX);
3524 amd64_mov_reg_mem (code, dreg, (tls_offset * 8) + 0x1480, 8);
3528 g_assert (tls_offset < 0x440);
3529 /* Load TEB->TlsExpansionSlots */
3530 x86_prefix (code, X86_GS_PREFIX);
3531 amd64_mov_reg_mem (code, dreg, 0x1780, 8);
3532 amd64_test_reg_reg (code, dreg, dreg);
3534 amd64_branch (code, X86_CC_EQ, code, TRUE);
3535 amd64_mov_reg_membase (code, dreg, dreg, (tls_offset * 8) - 0x200, 8);
3536 amd64_patch (buf [0], code);
3538 #elif defined(TARGET_MACH)
3539 x86_prefix (code, X86_GS_PREFIX);
3540 amd64_mov_reg_mem (code, dreg, tls_gs_offset + (tls_offset * 8), 8);
3542 if (optimize_for_xen) {
3543 x86_prefix (code, X86_FS_PREFIX);
3544 amd64_mov_reg_mem (code, dreg, 0, 8);
3545 amd64_mov_reg_membase (code, dreg, dreg, tls_offset, 8);
3547 x86_prefix (code, X86_FS_PREFIX);
3548 amd64_mov_reg_mem (code, dreg, tls_offset, 8);
3555 mono_amd64_emit_tls_set (guint8 *code, int sreg, int tls_offset)
3558 g_assert_not_reached ();
3559 #elif defined(TARGET_MACH)
3560 x86_prefix (code, X86_GS_PREFIX);
3561 amd64_mov_mem_reg (code, tls_gs_offset + (tls_offset * 8), sreg, 8);
3563 g_assert (!optimize_for_xen);
3564 x86_prefix (code, X86_FS_PREFIX);
3565 amd64_mov_mem_reg (code, tls_offset, sreg, 8);
3573 * Emit code to initialize an LMF structure at LMF_OFFSET.
3576 emit_setup_lmf (MonoCompile *cfg, guint8 *code, gint32 lmf_offset, int cfa_offset)
3579 * The ip field is not set, the exception handling code will obtain it from the stack location pointed to by the sp field.
3582 * sp is saved right before calls but we need to save it here too so
3583 * async stack walks would work.
3585 amd64_mov_membase_reg (code, cfg->frame_reg, lmf_offset + MONO_STRUCT_OFFSET (MonoLMF, rsp), AMD64_RSP, 8);
3587 amd64_mov_membase_reg (code, cfg->frame_reg, lmf_offset + MONO_STRUCT_OFFSET (MonoLMF, rbp), AMD64_RBP, 8);
3588 if (cfg->arch.omit_fp && cfa_offset != -1)
3589 mono_emit_unwind_op_offset (cfg, code, AMD64_RBP, - (cfa_offset - (lmf_offset + MONO_STRUCT_OFFSET (MonoLMF, rbp))));
3591 /* These can't contain refs */
3592 mini_gc_set_slot_type_from_fp (cfg, lmf_offset + MONO_STRUCT_OFFSET (MonoLMF, previous_lmf), SLOT_NOREF);
3593 mini_gc_set_slot_type_from_fp (cfg, lmf_offset + MONO_STRUCT_OFFSET (MonoLMF, rip), SLOT_NOREF);
3594 mini_gc_set_slot_type_from_fp (cfg, lmf_offset + MONO_STRUCT_OFFSET (MonoLMF, rsp), SLOT_NOREF);
3595 /* These are handled automatically by the stack marking code */
3596 mini_gc_set_slot_type_from_fp (cfg, lmf_offset + MONO_STRUCT_OFFSET (MonoLMF, rbp), SLOT_NOREF);
3603 #define TEB_LAST_ERROR_OFFSET 0x068
3606 emit_get_last_error (guint8* code, int dreg)
3608 /* Threads last error value is located in TEB_LAST_ERROR_OFFSET. */
3609 x86_prefix (code, X86_GS_PREFIX);
3610 amd64_mov_reg_membase (code, dreg, TEB_LAST_ERROR_OFFSET, 0, sizeof (guint32));
3618 emit_get_last_error (guint8* code, int dreg)
3620 g_assert_not_reached ();
3625 /* benchmark and set based on cpu */
3626 #define LOOP_ALIGNMENT 8
3627 #define bb_is_loop_start(bb) ((bb)->loop_body_start && (bb)->nesting)
3631 mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
3636 guint8 *code = cfg->native_code + cfg->code_len;
3639 /* Fix max_offset estimate for each successor bb */
3640 if (cfg->opt & MONO_OPT_BRANCH) {
3641 int current_offset = cfg->code_len;
3642 MonoBasicBlock *current_bb;
3643 for (current_bb = bb; current_bb != NULL; current_bb = current_bb->next_bb) {
3644 current_bb->max_offset = current_offset;
3645 current_offset += current_bb->max_length;
3649 if (cfg->opt & MONO_OPT_LOOP) {
3650 int pad, align = LOOP_ALIGNMENT;
3651 /* set alignment depending on cpu */
3652 if (bb_is_loop_start (bb) && (pad = (cfg->code_len & (align - 1)))) {
3654 /*g_print ("adding %d pad at %x to loop in %s\n", pad, cfg->code_len, cfg->method->name);*/
3655 amd64_padding (code, pad);
3656 cfg->code_len += pad;
3657 bb->native_offset = cfg->code_len;
3661 if (cfg->verbose_level > 2)
3662 g_print ("Basic block %d starting at offset 0x%x\n", bb->block_num, bb->native_offset);
3664 if ((cfg->prof_options & MONO_PROFILE_COVERAGE) && cfg->coverage_info) {
3665 MonoProfileCoverageInfo *cov = cfg->coverage_info;
3666 g_assert (!cfg->compile_aot);
3668 cov->data [bb->dfn].cil_code = bb->cil_code;
3669 amd64_mov_reg_imm (code, AMD64_R11, (guint64)&cov->data [bb->dfn].count);
3670 /* this is not thread save, but good enough */
3671 amd64_inc_membase (code, AMD64_R11, 0);
3674 offset = code - cfg->native_code;
3676 mono_debug_open_block (cfg, bb, offset);
3678 if (mono_break_at_bb_method && mono_method_desc_full_match (mono_break_at_bb_method, cfg->method) && bb->block_num == mono_break_at_bb_bb_num)
3679 x86_breakpoint (code);
3681 MONO_BB_FOR_EACH_INS (bb, ins) {
3682 offset = code - cfg->native_code;
3684 max_len = ((guint8 *)ins_get_spec (ins->opcode))[MONO_INST_LEN];
3686 #define EXTRA_CODE_SPACE (16)
3688 if (G_UNLIKELY (offset > (cfg->code_size - max_len - EXTRA_CODE_SPACE))) {
3689 cfg->code_size *= 2;
3690 cfg->native_code = (unsigned char *)mono_realloc_native_code(cfg);
3691 code = cfg->native_code + offset;
3692 cfg->stat_code_reallocs++;
3695 if (cfg->debug_info)
3696 mono_debug_record_line_number (cfg, ins, offset);
3698 switch (ins->opcode) {
3700 amd64_mul_reg (code, ins->sreg2, TRUE);
3703 amd64_mul_reg (code, ins->sreg2, FALSE);
3705 case OP_X86_SETEQ_MEMBASE:
3706 amd64_set_membase (code, X86_CC_EQ, ins->inst_basereg, ins->inst_offset, TRUE);
3708 case OP_STOREI1_MEMBASE_IMM:
3709 amd64_mov_membase_imm (code, ins->inst_destbasereg, ins->inst_offset, ins->inst_imm, 1);
3711 case OP_STOREI2_MEMBASE_IMM:
3712 amd64_mov_membase_imm (code, ins->inst_destbasereg, ins->inst_offset, ins->inst_imm, 2);
3714 case OP_STOREI4_MEMBASE_IMM:
3715 amd64_mov_membase_imm (code, ins->inst_destbasereg, ins->inst_offset, ins->inst_imm, 4);
3717 case OP_STOREI1_MEMBASE_REG:
3718 amd64_mov_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, ins->sreg1, 1);
3720 case OP_STOREI2_MEMBASE_REG:
3721 amd64_mov_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, ins->sreg1, 2);
3723 /* In AMD64 NaCl, pointers are 4 bytes, */
3724 /* so STORE_* != STOREI8_*. Likewise below. */
3725 case OP_STORE_MEMBASE_REG:
3726 amd64_mov_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, ins->sreg1, sizeof(gpointer));
3728 case OP_STOREI8_MEMBASE_REG:
3729 amd64_mov_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, ins->sreg1, 8);
3731 case OP_STOREI4_MEMBASE_REG:
3732 amd64_mov_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, ins->sreg1, 4);
3734 case OP_STORE_MEMBASE_IMM:
3735 /* In NaCl, this could be a PCONST type, which could */
3736 /* mean a pointer type was copied directly into the */
3737 /* lower 32-bits of inst_imm, so for InvalidPtr==-1 */
3738 /* the value would be 0x00000000FFFFFFFF which is */
3739 /* not proper for an imm32 unless you cast it. */
3740 g_assert (amd64_is_imm32 (ins->inst_imm));
3741 amd64_mov_membase_imm (code, ins->inst_destbasereg, ins->inst_offset, (gint32)ins->inst_imm, sizeof(gpointer));
3743 case OP_STOREI8_MEMBASE_IMM:
3744 g_assert (amd64_is_imm32 (ins->inst_imm));
3745 amd64_mov_membase_imm (code, ins->inst_destbasereg, ins->inst_offset, ins->inst_imm, 8);
3748 #ifdef __mono_ilp32__
3749 /* In ILP32, pointers are 4 bytes, so separate these */
3750 /* cases, use literal 8 below where we really want 8 */
3751 amd64_mov_reg_imm (code, ins->dreg, ins->inst_imm);
3752 amd64_mov_reg_membase (code, ins->dreg, ins->dreg, 0, sizeof(gpointer));
3756 // FIXME: Decompose this earlier
3757 if (amd64_use_imm32 (ins->inst_imm))
3758 amd64_mov_reg_mem (code, ins->dreg, ins->inst_imm, 8);
3760 amd64_mov_reg_imm_size (code, ins->dreg, ins->inst_imm, sizeof(gpointer));
3761 amd64_mov_reg_membase (code, ins->dreg, ins->dreg, 0, 8);
3765 amd64_mov_reg_imm (code, ins->dreg, ins->inst_imm);
3766 amd64_movsxd_reg_membase (code, ins->dreg, ins->dreg, 0);
3769 // FIXME: Decompose this earlier
3770 if (amd64_use_imm32 (ins->inst_imm))
3771 amd64_mov_reg_mem (code, ins->dreg, ins->inst_imm, 4);
3773 amd64_mov_reg_imm_size (code, ins->dreg, ins->inst_imm, sizeof(gpointer));
3774 amd64_mov_reg_membase (code, ins->dreg, ins->dreg, 0, 4);
3778 amd64_mov_reg_imm (code, ins->dreg, ins->inst_imm);
3779 amd64_widen_membase (code, ins->dreg, ins->dreg, 0, FALSE, FALSE);
3782 /* For NaCl, pointers are 4 bytes, so separate these */
3783 /* cases, use literal 8 below where we really want 8 */
3784 amd64_mov_reg_imm (code, ins->dreg, ins->inst_imm);
3785 amd64_widen_membase (code, ins->dreg, ins->dreg, 0, FALSE, TRUE);
3787 case OP_LOAD_MEMBASE:
3788 g_assert (amd64_is_imm32 (ins->inst_offset));
3789 amd64_mov_reg_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, sizeof(gpointer));
3791 case OP_LOADI8_MEMBASE:
3792 /* Use literal 8 instead of sizeof pointer or */
3793 /* register, we really want 8 for this opcode */
3794 g_assert (amd64_is_imm32 (ins->inst_offset));
3795 amd64_mov_reg_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, 8);
3797 case OP_LOADI4_MEMBASE:
3798 amd64_movsxd_reg_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset);
3800 case OP_LOADU4_MEMBASE:
3801 amd64_mov_reg_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, 4);
3803 case OP_LOADU1_MEMBASE:
3804 /* The cpu zero extends the result into 64 bits */
3805 amd64_widen_membase_size (code, ins->dreg, ins->inst_basereg, ins->inst_offset, FALSE, FALSE, 4);
3807 case OP_LOADI1_MEMBASE:
3808 amd64_widen_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, TRUE, FALSE);
3810 case OP_LOADU2_MEMBASE:
3811 /* The cpu zero extends the result into 64 bits */
3812 amd64_widen_membase_size (code, ins->dreg, ins->inst_basereg, ins->inst_offset, FALSE, TRUE, 4);
3814 case OP_LOADI2_MEMBASE:
3815 amd64_widen_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, TRUE, TRUE);
3817 case OP_AMD64_LOADI8_MEMINDEX:
3818 amd64_mov_reg_memindex_size (code, ins->dreg, ins->inst_basereg, 0, ins->inst_indexreg, 0, 8);
3820 case OP_LCONV_TO_I1:
3821 case OP_ICONV_TO_I1:
3823 amd64_widen_reg (code, ins->dreg, ins->sreg1, TRUE, FALSE);
3825 case OP_LCONV_TO_I2:
3826 case OP_ICONV_TO_I2:
3828 amd64_widen_reg (code, ins->dreg, ins->sreg1, TRUE, TRUE);
3830 case OP_LCONV_TO_U1:
3831 case OP_ICONV_TO_U1:
3832 amd64_widen_reg (code, ins->dreg, ins->sreg1, FALSE, FALSE);
3834 case OP_LCONV_TO_U2:
3835 case OP_ICONV_TO_U2:
3836 amd64_widen_reg (code, ins->dreg, ins->sreg1, FALSE, TRUE);
3839 /* Clean out the upper word */
3840 amd64_mov_reg_reg_size (code, ins->dreg, ins->sreg1, 4);
3843 amd64_movsxd_reg_reg (code, ins->dreg, ins->sreg1);
3847 amd64_alu_reg_reg (code, X86_CMP, ins->sreg1, ins->sreg2);
3849 case OP_COMPARE_IMM:
3850 #if defined(__mono_ilp32__)
3851 /* Comparison of pointer immediates should be 4 bytes to avoid sign-extend problems */
3852 g_assert (amd64_is_imm32 (ins->inst_imm));
3853 amd64_alu_reg_imm_size (code, X86_CMP, ins->sreg1, ins->inst_imm, 4);
3856 case OP_LCOMPARE_IMM:
3857 g_assert (amd64_is_imm32 (ins->inst_imm));
3858 amd64_alu_reg_imm (code, X86_CMP, ins->sreg1, ins->inst_imm);
3860 case OP_X86_COMPARE_REG_MEMBASE:
3861 amd64_alu_reg_membase (code, X86_CMP, ins->sreg1, ins->sreg2, ins->inst_offset);
3863 case OP_X86_TEST_NULL:
3864 amd64_test_reg_reg_size (code, ins->sreg1, ins->sreg1, 4);
3866 case OP_AMD64_TEST_NULL:
3867 amd64_test_reg_reg (code, ins->sreg1, ins->sreg1);
3870 case OP_X86_ADD_REG_MEMBASE:
3871 amd64_alu_reg_membase_size (code, X86_ADD, ins->sreg1, ins->sreg2, ins->inst_offset, 4);
3873 case OP_X86_SUB_REG_MEMBASE:
3874 amd64_alu_reg_membase_size (code, X86_SUB, ins->sreg1, ins->sreg2, ins->inst_offset, 4);
3876 case OP_X86_AND_REG_MEMBASE:
3877 amd64_alu_reg_membase_size (code, X86_AND, ins->sreg1, ins->sreg2, ins->inst_offset, 4);
3879 case OP_X86_OR_REG_MEMBASE:
3880 amd64_alu_reg_membase_size (code, X86_OR, ins->sreg1, ins->sreg2, ins->inst_offset, 4);
3882 case OP_X86_XOR_REG_MEMBASE:
3883 amd64_alu_reg_membase_size (code, X86_XOR, ins->sreg1, ins->sreg2, ins->inst_offset, 4);
3886 case OP_X86_ADD_MEMBASE_IMM:
3887 /* FIXME: Make a 64 version too */
3888 amd64_alu_membase_imm_size (code, X86_ADD, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 4);
3890 case OP_X86_SUB_MEMBASE_IMM:
3891 g_assert (amd64_is_imm32 (ins->inst_imm));
3892 amd64_alu_membase_imm_size (code, X86_SUB, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 4);
3894 case OP_X86_AND_MEMBASE_IMM:
3895 g_assert (amd64_is_imm32 (ins->inst_imm));
3896 amd64_alu_membase_imm_size (code, X86_AND, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 4);
3898 case OP_X86_OR_MEMBASE_IMM:
3899 g_assert (amd64_is_imm32 (ins->inst_imm));
3900 amd64_alu_membase_imm_size (code, X86_OR, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 4);
3902 case OP_X86_XOR_MEMBASE_IMM:
3903 g_assert (amd64_is_imm32 (ins->inst_imm));
3904 amd64_alu_membase_imm_size (code, X86_XOR, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 4);
3906 case OP_X86_ADD_MEMBASE_REG:
3907 amd64_alu_membase_reg_size (code, X86_ADD, ins->inst_basereg, ins->inst_offset, ins->sreg2, 4);
3909 case OP_X86_SUB_MEMBASE_REG:
3910 amd64_alu_membase_reg_size (code, X86_SUB, ins->inst_basereg, ins->inst_offset, ins->sreg2, 4);
3912 case OP_X86_AND_MEMBASE_REG:
3913 amd64_alu_membase_reg_size (code, X86_AND, ins->inst_basereg, ins->inst_offset, ins->sreg2, 4);
3915 case OP_X86_OR_MEMBASE_REG:
3916 amd64_alu_membase_reg_size (code, X86_OR, ins->inst_basereg, ins->inst_offset, ins->sreg2, 4);
3918 case OP_X86_XOR_MEMBASE_REG:
3919 amd64_alu_membase_reg_size (code, X86_XOR, ins->inst_basereg, ins->inst_offset, ins->sreg2, 4);
3921 case OP_X86_INC_MEMBASE:
3922 amd64_inc_membase_size (code, ins->inst_basereg, ins->inst_offset, 4);
3924 case OP_X86_INC_REG:
3925 amd64_inc_reg_size (code, ins->dreg, 4);
3927 case OP_X86_DEC_MEMBASE:
3928 amd64_dec_membase_size (code, ins->inst_basereg, ins->inst_offset, 4);
3930 case OP_X86_DEC_REG:
3931 amd64_dec_reg_size (code, ins->dreg, 4);
3933 case OP_X86_MUL_REG_MEMBASE:
3934 case OP_X86_MUL_MEMBASE_REG:
3935 amd64_imul_reg_membase_size (code, ins->sreg1, ins->sreg2, ins->inst_offset, 4);
3937 case OP_AMD64_ICOMPARE_MEMBASE_REG:
3938 amd64_alu_membase_reg_size (code, X86_CMP, ins->inst_basereg, ins->inst_offset, ins->sreg2, 4);
3940 case OP_AMD64_ICOMPARE_MEMBASE_IMM:
3941 amd64_alu_membase_imm_size (code, X86_CMP, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 4);
3943 case OP_AMD64_COMPARE_MEMBASE_REG:
3944 amd64_alu_membase_reg_size (code, X86_CMP, ins->inst_basereg, ins->inst_offset, ins->sreg2, 8);
3946 case OP_AMD64_COMPARE_MEMBASE_IMM:
3947 g_assert (amd64_is_imm32 (ins->inst_imm));
3948 amd64_alu_membase_imm_size (code, X86_CMP, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 8);
3950 case OP_X86_COMPARE_MEMBASE8_IMM:
3951 amd64_alu_membase8_imm_size (code, X86_CMP, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 4);
3953 case OP_AMD64_ICOMPARE_REG_MEMBASE:
3954 amd64_alu_reg_membase_size (code, X86_CMP, ins->sreg1, ins->sreg2, ins->inst_offset, 4);
3956 case OP_AMD64_COMPARE_REG_MEMBASE:
3957 amd64_alu_reg_membase_size (code, X86_CMP, ins->sreg1, ins->sreg2, ins->inst_offset, 8);
3960 case OP_AMD64_ADD_REG_MEMBASE:
3961 amd64_alu_reg_membase_size (code, X86_ADD, ins->sreg1, ins->sreg2, ins->inst_offset, 8);
3963 case OP_AMD64_SUB_REG_MEMBASE:
3964 amd64_alu_reg_membase_size (code, X86_SUB, ins->sreg1, ins->sreg2, ins->inst_offset, 8);
3966 case OP_AMD64_AND_REG_MEMBASE:
3967 amd64_alu_reg_membase_size (code, X86_AND, ins->sreg1, ins->sreg2, ins->inst_offset, 8);
3969 case OP_AMD64_OR_REG_MEMBASE:
3970 amd64_alu_reg_membase_size (code, X86_OR, ins->sreg1, ins->sreg2, ins->inst_offset, 8);
3972 case OP_AMD64_XOR_REG_MEMBASE:
3973 amd64_alu_reg_membase_size (code, X86_XOR, ins->sreg1, ins->sreg2, ins->inst_offset, 8);
3976 case OP_AMD64_ADD_MEMBASE_REG:
3977 amd64_alu_membase_reg_size (code, X86_ADD, ins->inst_basereg, ins->inst_offset, ins->sreg2, 8);
3979 case OP_AMD64_SUB_MEMBASE_REG:
3980 amd64_alu_membase_reg_size (code, X86_SUB, ins->inst_basereg, ins->inst_offset, ins->sreg2, 8);
3982 case OP_AMD64_AND_MEMBASE_REG:
3983 amd64_alu_membase_reg_size (code, X86_AND, ins->inst_basereg, ins->inst_offset, ins->sreg2, 8);
3985 case OP_AMD64_OR_MEMBASE_REG:
3986 amd64_alu_membase_reg_size (code, X86_OR, ins->inst_basereg, ins->inst_offset, ins->sreg2, 8);
3988 case OP_AMD64_XOR_MEMBASE_REG:
3989 amd64_alu_membase_reg_size (code, X86_XOR, ins->inst_basereg, ins->inst_offset, ins->sreg2, 8);
3992 case OP_AMD64_ADD_MEMBASE_IMM:
3993 g_assert (amd64_is_imm32 (ins->inst_imm));
3994 amd64_alu_membase_imm_size (code, X86_ADD, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 8);
3996 case OP_AMD64_SUB_MEMBASE_IMM:
3997 g_assert (amd64_is_imm32 (ins->inst_imm));
3998 amd64_alu_membase_imm_size (code, X86_SUB, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 8);
4000 case OP_AMD64_AND_MEMBASE_IMM:
4001 g_assert (amd64_is_imm32 (ins->inst_imm));
4002 amd64_alu_membase_imm_size (code, X86_AND, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 8);
4004 case OP_AMD64_OR_MEMBASE_IMM:
4005 g_assert (amd64_is_imm32 (ins->inst_imm));
4006 amd64_alu_membase_imm_size (code, X86_OR, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 8);
4008 case OP_AMD64_XOR_MEMBASE_IMM:
4009 g_assert (amd64_is_imm32 (ins->inst_imm));
4010 amd64_alu_membase_imm_size (code, X86_XOR, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 8);
4014 amd64_breakpoint (code);
4016 case OP_RELAXED_NOP:
4017 x86_prefix (code, X86_REP_PREFIX);
4025 case OP_DUMMY_STORE:
4026 case OP_DUMMY_ICONST:
4027 case OP_DUMMY_R8CONST:
4028 case OP_NOT_REACHED:
4031 case OP_IL_SEQ_POINT:
4032 mono_add_seq_point (cfg, bb, ins, code - cfg->native_code);
4034 case OP_SEQ_POINT: {
4035 if (ins->flags & MONO_INST_SINGLE_STEP_LOC) {
4036 MonoInst *var = (MonoInst *)cfg->arch.ss_tramp_var;
4039 /* Load ss_tramp_var */
4040 /* This is equal to &ss_trampoline */
4041 amd64_mov_reg_membase (code, AMD64_R11, var->inst_basereg, var->inst_offset, 8);
4042 /* Load the trampoline address */
4043 amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, 0, 8);
4044 /* Call it if it is non-null */
4045 amd64_test_reg_reg (code, AMD64_R11, AMD64_R11);
4047 amd64_branch8 (code, X86_CC_Z, 0, FALSE);
4048 amd64_call_reg (code, AMD64_R11);
4049 amd64_patch (label, code);
4053 * This is the address which is saved in seq points,
4055 mono_add_seq_point (cfg, bb, ins, code - cfg->native_code);
4057 if (cfg->compile_aot) {
4058 guint32 offset = code - cfg->native_code;
4060 MonoInst *info_var = (MonoInst *)cfg->arch.seq_point_info_var;
4064 amd64_mov_reg_membase (code, AMD64_R11, info_var->inst_basereg, info_var->inst_offset, 8);
4065 val = ((offset) * sizeof (guint8*)) + MONO_STRUCT_OFFSET (SeqPointInfo, bp_addrs);
4066 /* Load the info->bp_addrs [offset], which is either NULL or the address of the breakpoint trampoline */
4067 amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, val, 8);
4068 amd64_test_reg_reg (code, AMD64_R11, AMD64_R11);
4070 amd64_branch8 (code, X86_CC_Z, 0, FALSE);
4071 /* Call the trampoline */
4072 amd64_call_reg (code, AMD64_R11);
4073 amd64_patch (label, code);
4075 MonoInst *var = (MonoInst *)cfg->arch.bp_tramp_var;
4079 * Emit a test+branch against a constant, the constant will be overwritten
4080 * by mono_arch_set_breakpoint () to cause the test to fail.
4082 amd64_mov_reg_imm (code, AMD64_R11, 0);
4083 amd64_test_reg_reg (code, AMD64_R11, AMD64_R11);
4085 amd64_branch8 (code, X86_CC_Z, 0, FALSE);
4088 g_assert (var->opcode == OP_REGOFFSET);
4089 /* Load bp_tramp_var */
4090 /* This is equal to &bp_trampoline */
4091 amd64_mov_reg_membase (code, AMD64_R11, var->inst_basereg, var->inst_offset, 8);
4092 /* Call the trampoline */
4093 amd64_call_membase (code, AMD64_R11, 0);
4094 amd64_patch (label, code);
4097 * Add an additional nop so skipping the bp doesn't cause the ip to point
4098 * to another IL offset.
4106 amd64_alu_reg_reg (code, X86_ADD, ins->sreg1, ins->sreg2);
4109 amd64_alu_reg_reg (code, X86_ADC, ins->sreg1, ins->sreg2);
4113 g_assert (amd64_is_imm32 (ins->inst_imm));
4114 amd64_alu_reg_imm (code, X86_ADD, ins->dreg, ins->inst_imm);
4117 g_assert (amd64_is_imm32 (ins->inst_imm));
4118 amd64_alu_reg_imm (code, X86_ADC, ins->dreg, ins->inst_imm);
4123 amd64_alu_reg_reg (code, X86_SUB, ins->sreg1, ins->sreg2);
4126 amd64_alu_reg_reg (code, X86_SBB, ins->sreg1, ins->sreg2);
4130 g_assert (amd64_is_imm32 (ins->inst_imm));
4131 amd64_alu_reg_imm (code, X86_SUB, ins->dreg, ins->inst_imm);
4134 g_assert (amd64_is_imm32 (ins->inst_imm));
4135 amd64_alu_reg_imm (code, X86_SBB, ins->dreg, ins->inst_imm);
4138 amd64_alu_reg_reg (code, X86_AND, ins->sreg1, ins->sreg2);
4142 g_assert (amd64_is_imm32 (ins->inst_imm));
4143 amd64_alu_reg_imm (code, X86_AND, ins->sreg1, ins->inst_imm);
4146 amd64_imul_reg_reg (code, ins->sreg1, ins->sreg2);
4151 guint32 size = (ins->opcode == OP_IMUL_IMM) ? 4 : 8;
4153 switch (ins->inst_imm) {
4157 if (ins->dreg != ins->sreg1)
4158 amd64_mov_reg_reg (code, ins->dreg, ins->sreg1, size);
4159 amd64_alu_reg_reg (code, X86_ADD, ins->dreg, ins->dreg);
4162 /* LEA r1, [r2 + r2*2] */
4163 amd64_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 1);
4166 /* LEA r1, [r2 + r2*4] */
4167 amd64_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 2);
4170 /* LEA r1, [r2 + r2*2] */
4172 amd64_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 1);
4173 amd64_alu_reg_reg (code, X86_ADD, ins->dreg, ins->dreg);
4176 /* LEA r1, [r2 + r2*8] */
4177 amd64_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 3);
4180 /* LEA r1, [r2 + r2*4] */
4182 amd64_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 2);
4183 amd64_alu_reg_reg (code, X86_ADD, ins->dreg, ins->dreg);
4186 /* LEA r1, [r2 + r2*2] */
4188 amd64_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 1);
4189 amd64_shift_reg_imm (code, X86_SHL, ins->dreg, 2);
4192 /* LEA r1, [r2 + r2*4] */
4193 /* LEA r1, [r1 + r1*4] */
4194 amd64_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 2);
4195 amd64_lea_memindex (code, ins->dreg, ins->dreg, 0, ins->dreg, 2);
4198 /* LEA r1, [r2 + r2*4] */
4200 /* LEA r1, [r1 + r1*4] */
4201 amd64_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 2);
4202 amd64_shift_reg_imm (code, X86_SHL, ins->dreg, 2);
4203 amd64_lea_memindex (code, ins->dreg, ins->dreg, 0, ins->dreg, 2);
4206 amd64_imul_reg_reg_imm_size (code, ins->dreg, ins->sreg1, ins->inst_imm, size);
4213 /* Regalloc magic makes the div/rem cases the same */
4214 if (ins->sreg2 == AMD64_RDX) {
4215 amd64_mov_membase_reg (code, AMD64_RSP, -8, AMD64_RDX, 8);
4217 amd64_div_membase (code, AMD64_RSP, -8, TRUE);
4220 amd64_div_reg (code, ins->sreg2, TRUE);
4225 if (ins->sreg2 == AMD64_RDX) {
4226 amd64_mov_membase_reg (code, AMD64_RSP, -8, AMD64_RDX, 8);
4227 amd64_alu_reg_reg (code, X86_XOR, AMD64_RDX, AMD64_RDX);
4228 amd64_div_membase (code, AMD64_RSP, -8, FALSE);
4230 amd64_alu_reg_reg (code, X86_XOR, AMD64_RDX, AMD64_RDX);
4231 amd64_div_reg (code, ins->sreg2, FALSE);
4236 if (ins->sreg2 == AMD64_RDX) {
4237 amd64_mov_membase_reg (code, AMD64_RSP, -8, AMD64_RDX, 8);
4238 amd64_cdq_size (code, 4);
4239 amd64_div_membase_size (code, AMD64_RSP, -8, TRUE, 4);
4241 amd64_cdq_size (code, 4);
4242 amd64_div_reg_size (code, ins->sreg2, TRUE, 4);
4247 if (ins->sreg2 == AMD64_RDX) {
4248 amd64_mov_membase_reg (code, AMD64_RSP, -8, AMD64_RDX, 8);
4249 amd64_alu_reg_reg (code, X86_XOR, AMD64_RDX, AMD64_RDX);
4250 amd64_div_membase_size (code, AMD64_RSP, -8, FALSE, 4);
4252 amd64_alu_reg_reg (code, X86_XOR, AMD64_RDX, AMD64_RDX);
4253 amd64_div_reg_size (code, ins->sreg2, FALSE, 4);
4257 amd64_imul_reg_reg (code, ins->sreg1, ins->sreg2);
4258 EMIT_COND_SYSTEM_EXCEPTION (X86_CC_O, FALSE, "OverflowException");
4261 amd64_alu_reg_reg (code, X86_OR, ins->sreg1, ins->sreg2);
4265 g_assert (amd64_is_imm32 (ins->inst_imm));
4266 amd64_alu_reg_imm (code, X86_OR, ins->sreg1, ins->inst_imm);
4269 amd64_alu_reg_reg (code, X86_XOR, ins->sreg1, ins->sreg2);
4273 g_assert (amd64_is_imm32 (ins->inst_imm));
4274 amd64_alu_reg_imm (code, X86_XOR, ins->sreg1, ins->inst_imm);
4277 g_assert (ins->sreg2 == AMD64_RCX);
4278 amd64_shift_reg (code, X86_SHL, ins->dreg);
4281 g_assert (ins->sreg2 == AMD64_RCX);
4282 amd64_shift_reg (code, X86_SAR, ins->dreg);
4286 g_assert (amd64_is_imm32 (ins->inst_imm));
4287 amd64_shift_reg_imm (code, X86_SAR, ins->dreg, ins->inst_imm);
4290 g_assert (amd64_is_imm32 (ins->inst_imm));
4291 amd64_shift_reg_imm_size (code, X86_SHR, ins->dreg, ins->inst_imm, 4);
4293 case OP_LSHR_UN_IMM:
4294 g_assert (amd64_is_imm32 (ins->inst_imm));
4295 amd64_shift_reg_imm (code, X86_SHR, ins->dreg, ins->inst_imm);
4298 g_assert (ins->sreg2 == AMD64_RCX);
4299 amd64_shift_reg (code, X86_SHR, ins->dreg);
4303 g_assert (amd64_is_imm32 (ins->inst_imm));
4304 amd64_shift_reg_imm (code, X86_SHL, ins->dreg, ins->inst_imm);
4309 amd64_alu_reg_reg_size (code, X86_ADD, ins->sreg1, ins->sreg2, 4);
4312 amd64_alu_reg_reg_size (code, X86_ADC, ins->sreg1, ins->sreg2, 4);
4315 amd64_alu_reg_imm_size (code, X86_ADD, ins->dreg, ins->inst_imm, 4);
4318 amd64_alu_reg_imm_size (code, X86_ADC, ins->dreg, ins->inst_imm, 4);
4322 amd64_alu_reg_reg_size (code, X86_SUB, ins->sreg1, ins->sreg2, 4);
4325 amd64_alu_reg_reg_size (code, X86_SBB, ins->sreg1, ins->sreg2, 4);
4328 amd64_alu_reg_imm_size (code, X86_SUB, ins->dreg, ins->inst_imm, 4);
4331 amd64_alu_reg_imm_size (code, X86_SBB, ins->dreg, ins->inst_imm, 4);
4334 amd64_alu_reg_reg_size (code, X86_AND, ins->sreg1, ins->sreg2, 4);
4337 amd64_alu_reg_imm_size (code, X86_AND, ins->sreg1, ins->inst_imm, 4);
4340 amd64_alu_reg_reg_size (code, X86_OR, ins->sreg1, ins->sreg2, 4);
4343 amd64_alu_reg_imm_size (code, X86_OR, ins->sreg1, ins->inst_imm, 4);
4346 amd64_alu_reg_reg_size (code, X86_XOR, ins->sreg1, ins->sreg2, 4);
4349 amd64_alu_reg_imm_size (code, X86_XOR, ins->sreg1, ins->inst_imm, 4);
4352 amd64_neg_reg_size (code, ins->sreg1, 4);
4355 amd64_not_reg_size (code, ins->sreg1, 4);
4358 g_assert (ins->sreg2 == AMD64_RCX);
4359 amd64_shift_reg_size (code, X86_SHL, ins->dreg, 4);
4362 g_assert (ins->sreg2 == AMD64_RCX);
4363 amd64_shift_reg_size (code, X86_SAR, ins->dreg, 4);
4366 amd64_shift_reg_imm_size (code, X86_SAR, ins->dreg, ins->inst_imm, 4);
4368 case OP_ISHR_UN_IMM:
4369 amd64_shift_reg_imm_size (code, X86_SHR, ins->dreg, ins->inst_imm, 4);
4372 g_assert (ins->sreg2 == AMD64_RCX);
4373 amd64_shift_reg_size (code, X86_SHR, ins->dreg, 4);
4376 amd64_shift_reg_imm_size (code, X86_SHL, ins->dreg, ins->inst_imm, 4);
4379 amd64_imul_reg_reg_size (code, ins->sreg1, ins->sreg2, 4);
4382 amd64_imul_reg_reg_size (code, ins->sreg1, ins->sreg2, 4);
4383 EMIT_COND_SYSTEM_EXCEPTION (X86_CC_O, FALSE, "OverflowException");
4385 case OP_IMUL_OVF_UN:
4386 case OP_LMUL_OVF_UN: {
4387 /* the mul operation and the exception check should most likely be split */
4388 int non_eax_reg, saved_eax = FALSE, saved_edx = FALSE;
4389 int size = (ins->opcode == OP_IMUL_OVF_UN) ? 4 : 8;
4390 /*g_assert (ins->sreg2 == X86_EAX);
4391 g_assert (ins->dreg == X86_EAX);*/
4392 if (ins->sreg2 == X86_EAX) {
4393 non_eax_reg = ins->sreg1;
4394 } else if (ins->sreg1 == X86_EAX) {
4395 non_eax_reg = ins->sreg2;
4397 /* no need to save since we're going to store to it anyway */
4398 if (ins->dreg != X86_EAX) {
4400 amd64_push_reg (code, X86_EAX);
4402 amd64_mov_reg_reg (code, X86_EAX, ins->sreg1, size);
4403 non_eax_reg = ins->sreg2;
4405 if (ins->dreg == X86_EDX) {
4408 amd64_push_reg (code, X86_EAX);
4412 amd64_push_reg (code, X86_EDX);
4414 amd64_mul_reg_size (code, non_eax_reg, FALSE, size);
4415 /* save before the check since pop and mov don't change the flags */
4416 if (ins->dreg != X86_EAX)
4417 amd64_mov_reg_reg (code, ins->dreg, X86_EAX, size);
4419 amd64_pop_reg (code, X86_EDX);
4421 amd64_pop_reg (code, X86_EAX);
4422 EMIT_COND_SYSTEM_EXCEPTION (X86_CC_O, FALSE, "OverflowException");
4426 amd64_alu_reg_reg_size (code, X86_CMP, ins->sreg1, ins->sreg2, 4);
4428 case OP_ICOMPARE_IMM:
4429 amd64_alu_reg_imm_size (code, X86_CMP, ins->sreg1, ins->inst_imm, 4);
4451 EMIT_COND_BRANCH (ins, cc_table [mono_opcode_to_cond (ins->opcode)], cc_signed_table [mono_opcode_to_cond (ins->opcode)]);
4459 case OP_CMOV_INE_UN:
4460 case OP_CMOV_IGE_UN:
4461 case OP_CMOV_IGT_UN:
4462 case OP_CMOV_ILE_UN:
4463 case OP_CMOV_ILT_UN:
4469 case OP_CMOV_LNE_UN:
4470 case OP_CMOV_LGE_UN:
4471 case OP_CMOV_LGT_UN:
4472 case OP_CMOV_LLE_UN:
4473 case OP_CMOV_LLT_UN:
4474 g_assert (ins->dreg == ins->sreg1);
4475 /* This needs to operate on 64 bit values */
4476 amd64_cmov_reg (code, cc_table [mono_opcode_to_cond (ins->opcode)], cc_signed_table [mono_opcode_to_cond (ins->opcode)], ins->dreg, ins->sreg2);
4480 amd64_not_reg (code, ins->sreg1);
4483 amd64_neg_reg (code, ins->sreg1);
4488 if ((((guint64)ins->inst_c0) >> 32) == 0 && !mini_get_debug_options()->single_imm_size)
4489 amd64_mov_reg_imm_size (code, ins->dreg, ins->inst_c0, 4);
4491 amd64_mov_reg_imm_size (code, ins->dreg, ins->inst_c0, 8);
4494 mono_add_patch_info (cfg, offset, (MonoJumpInfoType)ins->inst_i1, ins->inst_p0);
4495 amd64_mov_reg_membase (code, ins->dreg, AMD64_RIP, 0, sizeof(gpointer));
4498 mono_add_patch_info (cfg, offset, (MonoJumpInfoType)ins->inst_i1, ins->inst_p0);
4499 amd64_mov_reg_imm_size (code, ins->dreg, 0, 8);
4502 if (ins->dreg != ins->sreg1)
4503 amd64_mov_reg_reg (code, ins->dreg, ins->sreg1, sizeof(mgreg_t));
4505 case OP_AMD64_SET_XMMREG_R4: {
4507 if (ins->dreg != ins->sreg1)
4508 amd64_sse_movss_reg_reg (code, ins->dreg, ins->sreg1);
4510 amd64_sse_cvtsd2ss_reg_reg (code, ins->dreg, ins->sreg1);
4514 case OP_AMD64_SET_XMMREG_R8: {
4515 if (ins->dreg != ins->sreg1)
4516 amd64_sse_movsd_reg_reg (code, ins->dreg, ins->sreg1);
4520 MonoCallInst *call = (MonoCallInst*)ins;
4521 int i, save_area_offset;
4523 g_assert (!cfg->method->save_lmf);
4525 /* Restore callee saved registers */
4526 save_area_offset = cfg->arch.reg_save_area_offset;
4527 for (i = 0; i < AMD64_NREG; ++i)
4528 if (AMD64_IS_CALLEE_SAVED_REG (i) && (cfg->used_int_regs & (1 << i))) {
4529 amd64_mov_reg_membase (code, i, cfg->frame_reg, save_area_offset, 8);
4530 save_area_offset += 8;
4533 if (cfg->arch.omit_fp) {
4534 if (cfg->arch.stack_alloc_size)
4535 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, cfg->arch.stack_alloc_size);
4537 if (call->stack_usage)
4540 /* Copy arguments on the stack to our argument area */
4541 for (i = 0; i < call->stack_usage; i += sizeof(mgreg_t)) {
4542 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RSP, i, sizeof(mgreg_t));
4543 amd64_mov_membase_reg (code, AMD64_RBP, 16 + i, AMD64_RAX, sizeof(mgreg_t));
4549 offset = code - cfg->native_code;
4550 mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_METHOD_JUMP, call->method);
4551 if (cfg->compile_aot)
4552 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RIP, 0, 8);
4554 amd64_set_reg_template (code, AMD64_R11);
4555 amd64_jump_reg (code, AMD64_R11);
4556 ins->flags |= MONO_INST_GC_CALLSITE;
4557 ins->backend.pc_offset = code - cfg->native_code;
4561 /* ensure ins->sreg1 is not NULL */
4562 amd64_alu_membase_imm_size (code, X86_CMP, ins->sreg1, 0, 0, 4);
4565 amd64_lea_membase (code, AMD64_R11, cfg->frame_reg, cfg->sig_cookie);
4566 amd64_mov_membase_reg (code, ins->sreg1, 0, AMD64_R11, sizeof(gpointer));
4576 call = (MonoCallInst*)ins;
4578 * The AMD64 ABI forces callers to know about varargs.
4580 if ((call->signature->call_convention == MONO_CALL_VARARG) && (call->signature->pinvoke))
4581 amd64_alu_reg_reg (code, X86_XOR, AMD64_RAX, AMD64_RAX);
4582 else if ((cfg->method->wrapper_type == MONO_WRAPPER_MANAGED_TO_NATIVE) && (cfg->method->klass->image != mono_defaults.corlib)) {
4584 * Since the unmanaged calling convention doesn't contain a
4585 * 'vararg' entry, we have to treat every pinvoke call as a
4586 * potential vararg call.
4590 for (i = 0; i < AMD64_XMM_NREG; ++i)
4591 if (call->used_fregs & (1 << i))
4594 amd64_alu_reg_reg (code, X86_XOR, AMD64_RAX, AMD64_RAX);
4596 amd64_mov_reg_imm (code, AMD64_RAX, nregs);
4599 if (ins->flags & MONO_INST_HAS_METHOD)
4600 code = emit_call (cfg, code, MONO_PATCH_INFO_METHOD, call->method, FALSE);
4602 code = emit_call (cfg, code, MONO_PATCH_INFO_ABS, call->fptr, FALSE);
4603 ins->flags |= MONO_INST_GC_CALLSITE;
4604 ins->backend.pc_offset = code - cfg->native_code;
4605 code = emit_move_return_value (cfg, ins, code);
4612 case OP_VOIDCALL_REG:
4614 call = (MonoCallInst*)ins;
4616 if (AMD64_IS_ARGUMENT_REG (ins->sreg1)) {
4617 amd64_mov_reg_reg (code, AMD64_R11, ins->sreg1, 8);
4618 ins->sreg1 = AMD64_R11;
4622 * The AMD64 ABI forces callers to know about varargs.
4624 if ((call->signature->call_convention == MONO_CALL_VARARG) && (call->signature->pinvoke)) {
4625 if (ins->sreg1 == AMD64_RAX) {
4626 amd64_mov_reg_reg (code, AMD64_R11, AMD64_RAX, 8);
4627 ins->sreg1 = AMD64_R11;
4629 amd64_alu_reg_reg (code, X86_XOR, AMD64_RAX, AMD64_RAX);
4630 } else if ((cfg->method->wrapper_type == MONO_WRAPPER_MANAGED_TO_NATIVE) && (cfg->method->klass->image != mono_defaults.corlib)) {
4632 * Since the unmanaged calling convention doesn't contain a
4633 * 'vararg' entry, we have to treat every pinvoke call as a
4634 * potential vararg call.
4638 for (i = 0; i < AMD64_XMM_NREG; ++i)
4639 if (call->used_fregs & (1 << i))
4641 if (ins->sreg1 == AMD64_RAX) {
4642 amd64_mov_reg_reg (code, AMD64_R11, AMD64_RAX, 8);
4643 ins->sreg1 = AMD64_R11;
4646 amd64_alu_reg_reg (code, X86_XOR, AMD64_RAX, AMD64_RAX);
4648 amd64_mov_reg_imm (code, AMD64_RAX, nregs);
4651 amd64_call_reg (code, ins->sreg1);
4652 ins->flags |= MONO_INST_GC_CALLSITE;
4653 ins->backend.pc_offset = code - cfg->native_code;
4654 code = emit_move_return_value (cfg, ins, code);
4656 case OP_FCALL_MEMBASE:
4657 case OP_RCALL_MEMBASE:
4658 case OP_LCALL_MEMBASE:
4659 case OP_VCALL_MEMBASE:
4660 case OP_VCALL2_MEMBASE:
4661 case OP_VOIDCALL_MEMBASE:
4662 case OP_CALL_MEMBASE:
4663 call = (MonoCallInst*)ins;
4665 amd64_call_membase (code, ins->sreg1, ins->inst_offset);
4666 ins->flags |= MONO_INST_GC_CALLSITE;
4667 ins->backend.pc_offset = code - cfg->native_code;
4668 code = emit_move_return_value (cfg, ins, code);
4672 MonoInst *var = cfg->dyn_call_var;
4675 g_assert (var->opcode == OP_REGOFFSET);
4677 /* r11 = args buffer filled by mono_arch_get_dyn_call_args () */
4678 amd64_mov_reg_reg (code, AMD64_R11, ins->sreg1, 8);
4680 amd64_mov_reg_reg (code, AMD64_R10, ins->sreg2, 8);
4682 /* Save args buffer */
4683 amd64_mov_membase_reg (code, var->inst_basereg, var->inst_offset, AMD64_R11, 8);
4685 /* Set fp arg regs */
4686 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_R11, MONO_STRUCT_OFFSET (DynCallArgs, has_fp), sizeof (mgreg_t));
4687 amd64_test_reg_reg (code, AMD64_RAX, AMD64_RAX);
4689 amd64_branch8 (code, X86_CC_Z, -1, 1);
4690 for (i = 0; i < FLOAT_PARAM_REGS; ++i)
4691 amd64_sse_movsd_reg_membase (code, i, AMD64_R11, MONO_STRUCT_OFFSET (DynCallArgs, fregs) + (i * sizeof (double)));
4692 amd64_patch (label, code);
4694 /* Set stack args */
4695 for (i = 0; i < DYN_CALL_STACK_ARGS; ++i) {
4696 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_R11, MONO_STRUCT_OFFSET (DynCallArgs, regs) + ((PARAM_REGS + i) * sizeof(mgreg_t)), sizeof(mgreg_t));
4697 amd64_mov_membase_reg (code, AMD64_RSP, i * sizeof (mgreg_t), AMD64_RAX, sizeof (mgreg_t));
4700 /* Set argument registers */
4701 for (i = 0; i < PARAM_REGS; ++i)
4702 amd64_mov_reg_membase (code, param_regs [i], AMD64_R11, i * sizeof(mgreg_t), sizeof(mgreg_t));
4705 amd64_call_reg (code, AMD64_R10);
4707 ins->flags |= MONO_INST_GC_CALLSITE;
4708 ins->backend.pc_offset = code - cfg->native_code;
4711 amd64_mov_reg_membase (code, AMD64_R11, var->inst_basereg, var->inst_offset, 8);
4712 amd64_mov_membase_reg (code, AMD64_R11, MONO_STRUCT_OFFSET (DynCallArgs, res), AMD64_RAX, 8);
4713 amd64_sse_movsd_membase_reg (code, AMD64_R11, MONO_STRUCT_OFFSET (DynCallArgs, fregs), AMD64_XMM0);
4714 amd64_sse_movsd_membase_reg (code, AMD64_R11, MONO_STRUCT_OFFSET (DynCallArgs, fregs) + sizeof (double), AMD64_XMM1);
4717 case OP_AMD64_SAVE_SP_TO_LMF: {
4718 MonoInst *lmf_var = cfg->lmf_var;
4719 amd64_mov_membase_reg (code, lmf_var->inst_basereg, lmf_var->inst_offset + MONO_STRUCT_OFFSET (MonoLMF, rsp), AMD64_RSP, 8);
4723 g_assert_not_reached ();
4724 amd64_push_reg (code, ins->sreg1);
4726 case OP_X86_PUSH_IMM:
4727 g_assert_not_reached ();
4728 g_assert (amd64_is_imm32 (ins->inst_imm));
4729 amd64_push_imm (code, ins->inst_imm);
4731 case OP_X86_PUSH_MEMBASE:
4732 g_assert_not_reached ();
4733 amd64_push_membase (code, ins->inst_basereg, ins->inst_offset);
4735 case OP_X86_PUSH_OBJ: {
4736 int size = ALIGN_TO (ins->inst_imm, 8);
4738 g_assert_not_reached ();
4740 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, size);
4741 amd64_push_reg (code, AMD64_RDI);
4742 amd64_push_reg (code, AMD64_RSI);
4743 amd64_push_reg (code, AMD64_RCX);
4744 if (ins->inst_offset)
4745 amd64_lea_membase (code, AMD64_RSI, ins->inst_basereg, ins->inst_offset);
4747 amd64_mov_reg_reg (code, AMD64_RSI, ins->inst_basereg, 8);
4748 amd64_lea_membase (code, AMD64_RDI, AMD64_RSP, (3 * 8));
4749 amd64_mov_reg_imm (code, AMD64_RCX, (size >> 3));
4751 amd64_prefix (code, X86_REP_PREFIX);
4753 amd64_pop_reg (code, AMD64_RCX);
4754 amd64_pop_reg (code, AMD64_RSI);
4755 amd64_pop_reg (code, AMD64_RDI);
4758 case OP_GENERIC_CLASS_INIT: {
4761 g_assert (ins->sreg1 == MONO_AMD64_ARG_REG1);
4763 amd64_test_membase_imm_size (code, ins->sreg1, MONO_STRUCT_OFFSET (MonoVTable, initialized), 1, 1);
4765 amd64_branch8 (code, X86_CC_NZ, -1, 1);
4767 code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD, "mono_generic_class_init", FALSE);
4768 ins->flags |= MONO_INST_GC_CALLSITE;
4769 ins->backend.pc_offset = code - cfg->native_code;
4771 x86_patch (jump, code);
4776 amd64_lea_memindex (code, ins->dreg, ins->sreg1, ins->inst_imm, ins->sreg2, ins->backend.shift_amount);
4778 case OP_X86_LEA_MEMBASE:
4779 amd64_lea_membase (code, ins->dreg, ins->sreg1, ins->inst_imm);
4782 amd64_xchg_reg_reg (code, ins->sreg1, ins->sreg2, 4);
4785 /* keep alignment */
4786 amd64_alu_reg_imm (code, X86_ADD, ins->sreg1, MONO_ARCH_FRAME_ALIGNMENT - 1);
4787 amd64_alu_reg_imm (code, X86_AND, ins->sreg1, ~(MONO_ARCH_FRAME_ALIGNMENT - 1));
4788 code = mono_emit_stack_alloc (cfg, code, ins);
4789 amd64_mov_reg_reg (code, ins->dreg, AMD64_RSP, 8);
4790 if (cfg->param_area)
4791 amd64_alu_reg_imm (code, X86_ADD, ins->dreg, cfg->param_area);
4793 case OP_LOCALLOC_IMM: {
4794 guint32 size = ins->inst_imm;
4795 size = (size + (MONO_ARCH_FRAME_ALIGNMENT - 1)) & ~ (MONO_ARCH_FRAME_ALIGNMENT - 1);
4797 if (ins->flags & MONO_INST_INIT) {
4801 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, size);
4802 amd64_alu_reg_reg (code, X86_XOR, ins->dreg, ins->dreg);
4804 for (i = 0; i < size; i += 8)
4805 amd64_mov_membase_reg (code, AMD64_RSP, i, ins->dreg, 8);
4806 amd64_mov_reg_reg (code, ins->dreg, AMD64_RSP, 8);
4808 amd64_mov_reg_imm (code, ins->dreg, size);
4809 ins->sreg1 = ins->dreg;
4811 code = mono_emit_stack_alloc (cfg, code, ins);
4812 amd64_mov_reg_reg (code, ins->dreg, AMD64_RSP, 8);
4815 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, size);
4816 amd64_mov_reg_reg (code, ins->dreg, AMD64_RSP, 8);
4818 if (cfg->param_area)
4819 amd64_alu_reg_imm (code, X86_ADD, ins->dreg, cfg->param_area);
4823 amd64_mov_reg_reg (code, AMD64_ARG_REG1, ins->sreg1, 8);
4824 code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD,
4825 (gpointer)"mono_arch_throw_exception", FALSE);
4826 ins->flags |= MONO_INST_GC_CALLSITE;
4827 ins->backend.pc_offset = code - cfg->native_code;
4831 amd64_mov_reg_reg (code, AMD64_ARG_REG1, ins->sreg1, 8);
4832 code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD,
4833 (gpointer)"mono_arch_rethrow_exception", FALSE);
4834 ins->flags |= MONO_INST_GC_CALLSITE;
4835 ins->backend.pc_offset = code - cfg->native_code;
4838 case OP_CALL_HANDLER:
4840 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, 8);
4841 mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_BB, ins->inst_target_bb);
4842 amd64_call_imm (code, 0);
4843 mono_cfg_add_try_hole (cfg, ins->inst_eh_block, code, bb);
4844 /* Restore stack alignment */
4845 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, 8);
4847 case OP_START_HANDLER: {
4848 /* Even though we're saving RSP, use sizeof */
4849 /* gpointer because spvar is of type IntPtr */
4850 /* see: mono_create_spvar_for_region */
4851 MonoInst *spvar = mono_find_spvar_for_region (cfg, bb->region);
4852 amd64_mov_membase_reg (code, spvar->inst_basereg, spvar->inst_offset, AMD64_RSP, sizeof(gpointer));
4854 if ((MONO_BBLOCK_IS_IN_REGION (bb, MONO_REGION_FINALLY) ||
4855 MONO_BBLOCK_IS_IN_REGION (bb, MONO_REGION_FINALLY)) &&
4857 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, ALIGN_TO (cfg->param_area, MONO_ARCH_FRAME_ALIGNMENT));
4861 case OP_ENDFINALLY: {
4862 MonoInst *spvar = mono_find_spvar_for_region (cfg, bb->region);
4863 amd64_mov_reg_membase (code, AMD64_RSP, spvar->inst_basereg, spvar->inst_offset, sizeof(gpointer));
4867 case OP_ENDFILTER: {
4868 MonoInst *spvar = mono_find_spvar_for_region (cfg, bb->region);
4869 amd64_mov_reg_membase (code, AMD64_RSP, spvar->inst_basereg, spvar->inst_offset, sizeof(gpointer));
4870 /* The local allocator will put the result into RAX */
4875 if (ins->dreg != AMD64_RAX)
4876 amd64_mov_reg_reg (code, ins->dreg, AMD64_RAX, sizeof (gpointer));
4879 ins->inst_c0 = code - cfg->native_code;
4882 //g_print ("target: %p, next: %p, curr: %p, last: %p\n", ins->inst_target_bb, bb->next_bb, ins, bb->last_ins);
4883 //if ((ins->inst_target_bb == bb->next_bb) && ins == bb->last_ins)
4885 if (ins->inst_target_bb->native_offset) {
4886 amd64_jump_code (code, cfg->native_code + ins->inst_target_bb->native_offset);
4888 mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_BB, ins->inst_target_bb);
4889 if ((cfg->opt & MONO_OPT_BRANCH) &&
4890 x86_is_imm8 (ins->inst_target_bb->max_offset - offset))
4891 x86_jump8 (code, 0);
4893 x86_jump32 (code, 0);
4897 amd64_jump_reg (code, ins->sreg1);
4920 amd64_set_reg (code, cc_table [mono_opcode_to_cond (ins->opcode)], ins->dreg, cc_signed_table [mono_opcode_to_cond (ins->opcode)]);
4921 amd64_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
4923 case OP_COND_EXC_EQ:
4924 case OP_COND_EXC_NE_UN:
4925 case OP_COND_EXC_LT:
4926 case OP_COND_EXC_LT_UN:
4927 case OP_COND_EXC_GT:
4928 case OP_COND_EXC_GT_UN:
4929 case OP_COND_EXC_GE:
4930 case OP_COND_EXC_GE_UN:
4931 case OP_COND_EXC_LE:
4932 case OP_COND_EXC_LE_UN:
4933 case OP_COND_EXC_IEQ:
4934 case OP_COND_EXC_INE_UN:
4935 case OP_COND_EXC_ILT:
4936 case OP_COND_EXC_ILT_UN:
4937 case OP_COND_EXC_IGT:
4938 case OP_COND_EXC_IGT_UN:
4939 case OP_COND_EXC_IGE:
4940 case OP_COND_EXC_IGE_UN:
4941 case OP_COND_EXC_ILE:
4942 case OP_COND_EXC_ILE_UN:
4943 EMIT_COND_SYSTEM_EXCEPTION (cc_table [mono_opcode_to_cond (ins->opcode)], cc_signed_table [mono_opcode_to_cond (ins->opcode)], (const char *)ins->inst_p1);
4945 case OP_COND_EXC_OV:
4946 case OP_COND_EXC_NO:
4948 case OP_COND_EXC_NC:
4949 EMIT_COND_SYSTEM_EXCEPTION (branch_cc_table [ins->opcode - OP_COND_EXC_EQ],
4950 (ins->opcode < OP_COND_EXC_NE_UN), (const char *)ins->inst_p1);
4952 case OP_COND_EXC_IOV:
4953 case OP_COND_EXC_INO:
4954 case OP_COND_EXC_IC:
4955 case OP_COND_EXC_INC:
4956 EMIT_COND_SYSTEM_EXCEPTION (branch_cc_table [ins->opcode - OP_COND_EXC_IEQ],
4957 (ins->opcode < OP_COND_EXC_INE_UN), (const char *)ins->inst_p1);
4960 /* floating point opcodes */
4962 double d = *(double *)ins->inst_p0;
4964 if ((d == 0.0) && (mono_signbit (d) == 0)) {
4965 amd64_sse_xorpd_reg_reg (code, ins->dreg, ins->dreg);
4968 mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_R8, ins->inst_p0);
4969 amd64_sse_movsd_reg_membase (code, ins->dreg, AMD64_RIP, 0);
4974 float f = *(float *)ins->inst_p0;
4976 if ((f == 0.0) && (mono_signbit (f) == 0)) {
4978 amd64_sse_xorps_reg_reg (code, ins->dreg, ins->dreg);
4980 amd64_sse_xorpd_reg_reg (code, ins->dreg, ins->dreg);
4983 mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_R4, ins->inst_p0);
4984 amd64_sse_movss_reg_membase (code, ins->dreg, AMD64_RIP, 0);
4986 amd64_sse_cvtss2sd_reg_reg (code, ins->dreg, ins->dreg);
4990 case OP_STORER8_MEMBASE_REG:
4991 amd64_sse_movsd_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, ins->sreg1);
4993 case OP_LOADR8_MEMBASE:
4994 amd64_sse_movsd_reg_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset);
4996 case OP_STORER4_MEMBASE_REG:
4998 amd64_sse_movss_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, ins->sreg1);
5000 /* This requires a double->single conversion */
5001 amd64_sse_cvtsd2ss_reg_reg (code, MONO_ARCH_FP_SCRATCH_REG, ins->sreg1);
5002 amd64_sse_movss_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, MONO_ARCH_FP_SCRATCH_REG);
5005 case OP_LOADR4_MEMBASE:
5007 amd64_sse_movss_reg_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset);
5009 amd64_sse_movss_reg_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset);
5010 amd64_sse_cvtss2sd_reg_reg (code, ins->dreg, ins->dreg);
5013 case OP_ICONV_TO_R4:
5015 amd64_sse_cvtsi2ss_reg_reg_size (code, ins->dreg, ins->sreg1, 4);
5017 amd64_sse_cvtsi2ss_reg_reg_size (code, ins->dreg, ins->sreg1, 4);
5018 amd64_sse_cvtss2sd_reg_reg (code, ins->dreg, ins->dreg);
5021 case OP_ICONV_TO_R8:
5022 amd64_sse_cvtsi2sd_reg_reg_size (code, ins->dreg, ins->sreg1, 4);
5024 case OP_LCONV_TO_R4:
5026 amd64_sse_cvtsi2ss_reg_reg (code, ins->dreg, ins->sreg1);
5028 amd64_sse_cvtsi2ss_reg_reg (code, ins->dreg, ins->sreg1);
5029 amd64_sse_cvtss2sd_reg_reg (code, ins->dreg, ins->dreg);
5032 case OP_LCONV_TO_R8:
5033 amd64_sse_cvtsi2sd_reg_reg (code, ins->dreg, ins->sreg1);
5035 case OP_FCONV_TO_R4:
5037 amd64_sse_cvtsd2ss_reg_reg (code, ins->dreg, ins->sreg1);
5039 amd64_sse_cvtsd2ss_reg_reg (code, ins->dreg, ins->sreg1);
5040 amd64_sse_cvtss2sd_reg_reg (code, ins->dreg, ins->dreg);
5043 case OP_FCONV_TO_I1:
5044 code = emit_float_to_int (cfg, code, ins->dreg, ins->sreg1, 1, TRUE);
5046 case OP_FCONV_TO_U1:
5047 code = emit_float_to_int (cfg, code, ins->dreg, ins->sreg1, 1, FALSE);
5049 case OP_FCONV_TO_I2:
5050 code = emit_float_to_int (cfg, code, ins->dreg, ins->sreg1, 2, TRUE);
5052 case OP_FCONV_TO_U2:
5053 code = emit_float_to_int (cfg, code, ins->dreg, ins->sreg1, 2, FALSE);
5055 case OP_FCONV_TO_U4:
5056 code = emit_float_to_int (cfg, code, ins->dreg, ins->sreg1, 4, FALSE);
5058 case OP_FCONV_TO_I4:
5060 code = emit_float_to_int (cfg, code, ins->dreg, ins->sreg1, 4, TRUE);
5062 case OP_FCONV_TO_I8:
5063 code = emit_float_to_int (cfg, code, ins->dreg, ins->sreg1, 8, TRUE);
5066 case OP_RCONV_TO_I1:
5067 amd64_sse_cvtss2si_reg_reg_size (code, ins->dreg, ins->sreg1, 4);
5068 amd64_widen_reg (code, ins->dreg, ins->dreg, TRUE, FALSE);
5070 case OP_RCONV_TO_U1:
5071 amd64_sse_cvtss2si_reg_reg_size (code, ins->dreg, ins->sreg1, 4);
5072 amd64_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
5074 case OP_RCONV_TO_I2:
5075 amd64_sse_cvtss2si_reg_reg_size (code, ins->dreg, ins->sreg1, 4);
5076 amd64_widen_reg (code, ins->dreg, ins->dreg, TRUE, TRUE);
5078 case OP_RCONV_TO_U2:
5079 amd64_sse_cvtss2si_reg_reg_size (code, ins->dreg, ins->sreg1, 4);
5080 amd64_widen_reg (code, ins->dreg, ins->dreg, FALSE, TRUE);
5082 case OP_RCONV_TO_I4:
5083 amd64_sse_cvtss2si_reg_reg_size (code, ins->dreg, ins->sreg1, 4);
5085 case OP_RCONV_TO_U4:
5086 amd64_sse_cvtss2si_reg_reg_size (code, ins->dreg, ins->sreg1, 4);
5088 case OP_RCONV_TO_I8:
5089 amd64_sse_cvtss2si_reg_reg_size (code, ins->dreg, ins->sreg1, 8);
5091 case OP_RCONV_TO_R8:
5092 amd64_sse_cvtss2sd_reg_reg (code, ins->dreg, ins->sreg1);
5094 case OP_RCONV_TO_R4:
5095 if (ins->dreg != ins->sreg1)
5096 amd64_sse_movss_reg_reg (code, ins->dreg, ins->sreg1);
5099 case OP_LCONV_TO_R_UN: {
5102 /* Based on gcc code */
5103 amd64_test_reg_reg (code, ins->sreg1, ins->sreg1);
5104 br [0] = code; x86_branch8 (code, X86_CC_S, 0, TRUE);
5107 amd64_sse_cvtsi2sd_reg_reg (code, ins->dreg, ins->sreg1);
5108 br [1] = code; x86_jump8 (code, 0);
5109 amd64_patch (br [0], code);
5112 /* Save to the red zone */
5113 amd64_mov_membase_reg (code, AMD64_RSP, -8, AMD64_RAX, 8);
5114 amd64_mov_membase_reg (code, AMD64_RSP, -16, AMD64_RCX, 8);
5115 amd64_mov_reg_reg (code, AMD64_RCX, ins->sreg1, 8);
5116 amd64_mov_reg_reg (code, AMD64_RAX, ins->sreg1, 8);
5117 amd64_alu_reg_imm (code, X86_AND, AMD64_RCX, 1);
5118 amd64_shift_reg_imm (code, X86_SHR, AMD64_RAX, 1);
5119 amd64_alu_reg_imm (code, X86_OR, AMD64_RAX, AMD64_RCX);
5120 amd64_sse_cvtsi2sd_reg_reg (code, ins->dreg, AMD64_RAX);
5121 amd64_sse_addsd_reg_reg (code, ins->dreg, ins->dreg);
5123 amd64_mov_reg_membase (code, AMD64_RCX, AMD64_RSP, -16, 8);
5124 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RSP, -8, 8);
5125 amd64_patch (br [1], code);
5128 case OP_LCONV_TO_OVF_U4:
5129 amd64_alu_reg_imm (code, X86_CMP, ins->sreg1, 0);
5130 EMIT_COND_SYSTEM_EXCEPTION (X86_CC_LT, TRUE, "OverflowException");
5131 amd64_mov_reg_reg (code, ins->dreg, ins->sreg1, 8);
5133 case OP_LCONV_TO_OVF_I4_UN:
5134 amd64_alu_reg_imm (code, X86_CMP, ins->sreg1, 0x7fffffff);
5135 EMIT_COND_SYSTEM_EXCEPTION (X86_CC_GT, FALSE, "OverflowException");
5136 amd64_mov_reg_reg (code, ins->dreg, ins->sreg1, 8);
5139 if (ins->dreg != ins->sreg1)
5140 amd64_sse_movsd_reg_reg (code, ins->dreg, ins->sreg1);
5143 if (ins->dreg != ins->sreg1)
5144 amd64_sse_movss_reg_reg (code, ins->dreg, ins->sreg1);
5146 case OP_MOVE_F_TO_I4:
5148 amd64_movd_reg_xreg_size (code, ins->dreg, ins->sreg1, 8);
5150 amd64_sse_cvtsd2ss_reg_reg (code, MONO_ARCH_FP_SCRATCH_REG, ins->sreg1);
5151 amd64_movd_reg_xreg_size (code, ins->dreg, MONO_ARCH_FP_SCRATCH_REG, 8);
5154 case OP_MOVE_I4_TO_F:
5155 amd64_movd_xreg_reg_size (code, ins->dreg, ins->sreg1, 8);
5157 amd64_sse_cvtss2sd_reg_reg (code, ins->dreg, ins->dreg);
5159 case OP_MOVE_F_TO_I8:
5160 amd64_movd_reg_xreg_size (code, ins->dreg, ins->sreg1, 8);
5162 case OP_MOVE_I8_TO_F:
5163 amd64_movd_xreg_reg_size (code, ins->dreg, ins->sreg1, 8);
5166 amd64_sse_addsd_reg_reg (code, ins->dreg, ins->sreg2);
5169 amd64_sse_subsd_reg_reg (code, ins->dreg, ins->sreg2);
5172 amd64_sse_mulsd_reg_reg (code, ins->dreg, ins->sreg2);
5175 amd64_sse_divsd_reg_reg (code, ins->dreg, ins->sreg2);
5178 static double r8_0 = -0.0;
5180 g_assert (ins->sreg1 == ins->dreg);
5182 mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_R8, &r8_0);
5183 amd64_sse_xorpd_reg_membase (code, ins->dreg, AMD64_RIP, 0);
5187 EMIT_SSE2_FPFUNC (code, fsin, ins->dreg, ins->sreg1);
5190 EMIT_SSE2_FPFUNC (code, fcos, ins->dreg, ins->sreg1);
5193 static guint64 d = 0x7fffffffffffffffUL;
5195 g_assert (ins->sreg1 == ins->dreg);
5197 mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_R8, &d);
5198 amd64_sse_andpd_reg_membase (code, ins->dreg, AMD64_RIP, 0);
5202 EMIT_SSE2_FPFUNC (code, fsqrt, ins->dreg, ins->sreg1);
5206 amd64_sse_addss_reg_reg (code, ins->dreg, ins->sreg2);
5209 amd64_sse_subss_reg_reg (code, ins->dreg, ins->sreg2);
5212 amd64_sse_mulss_reg_reg (code, ins->dreg, ins->sreg2);
5215 amd64_sse_divss_reg_reg (code, ins->dreg, ins->sreg2);
5218 static float r4_0 = -0.0;
5220 g_assert (ins->sreg1 == ins->dreg);
5222 mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_R4, &r4_0);
5223 amd64_sse_movss_reg_membase (code, MONO_ARCH_FP_SCRATCH_REG, AMD64_RIP, 0);
5224 amd64_sse_xorps_reg_reg (code, ins->dreg, MONO_ARCH_FP_SCRATCH_REG);
5229 g_assert (cfg->opt & MONO_OPT_CMOV);
5230 g_assert (ins->dreg == ins->sreg1);
5231 amd64_alu_reg_reg_size (code, X86_CMP, ins->sreg1, ins->sreg2, 4);
5232 amd64_cmov_reg_size (code, X86_CC_GT, TRUE, ins->dreg, ins->sreg2, 4);
5235 g_assert (cfg->opt & MONO_OPT_CMOV);
5236 g_assert (ins->dreg == ins->sreg1);
5237 amd64_alu_reg_reg_size (code, X86_CMP, ins->sreg1, ins->sreg2, 4);
5238 amd64_cmov_reg_size (code, X86_CC_GT, FALSE, ins->dreg, ins->sreg2, 4);
5241 g_assert (cfg->opt & MONO_OPT_CMOV);
5242 g_assert (ins->dreg == ins->sreg1);
5243 amd64_alu_reg_reg_size (code, X86_CMP, ins->sreg1, ins->sreg2, 4);
5244 amd64_cmov_reg_size (code, X86_CC_LT, TRUE, ins->dreg, ins->sreg2, 4);
5247 g_assert (cfg->opt & MONO_OPT_CMOV);
5248 g_assert (ins->dreg == ins->sreg1);
5249 amd64_alu_reg_reg_size (code, X86_CMP, ins->sreg1, ins->sreg2, 4);
5250 amd64_cmov_reg_size (code, X86_CC_LT, FALSE, ins->dreg, ins->sreg2, 4);
5253 g_assert (cfg->opt & MONO_OPT_CMOV);
5254 g_assert (ins->dreg == ins->sreg1);
5255 amd64_alu_reg_reg (code, X86_CMP, ins->sreg1, ins->sreg2);
5256 amd64_cmov_reg (code, X86_CC_GT, TRUE, ins->dreg, ins->sreg2);
5259 g_assert (cfg->opt & MONO_OPT_CMOV);
5260 g_assert (ins->dreg == ins->sreg1);
5261 amd64_alu_reg_reg (code, X86_CMP, ins->sreg1, ins->sreg2);
5262 amd64_cmov_reg (code, X86_CC_GT, FALSE, ins->dreg, ins->sreg2);
5265 g_assert (cfg->opt & MONO_OPT_CMOV);
5266 g_assert (ins->dreg == ins->sreg1);
5267 amd64_alu_reg_reg (code, X86_CMP, ins->sreg1, ins->sreg2);
5268 amd64_cmov_reg (code, X86_CC_LT, TRUE, ins->dreg, ins->sreg2);
5271 g_assert (cfg->opt & MONO_OPT_CMOV);
5272 g_assert (ins->dreg == ins->sreg1);
5273 amd64_alu_reg_reg (code, X86_CMP, ins->sreg1, ins->sreg2);
5274 amd64_cmov_reg (code, X86_CC_LT, FALSE, ins->dreg, ins->sreg2);
5280 * The two arguments are swapped because the fbranch instructions
5281 * depend on this for the non-sse case to work.
5283 amd64_sse_comisd_reg_reg (code, ins->sreg2, ins->sreg1);
5287 * FIXME: Get rid of this.
5288 * The two arguments are swapped because the fbranch instructions
5289 * depend on this for the non-sse case to work.
5291 amd64_sse_comiss_reg_reg (code, ins->sreg2, ins->sreg1);
5295 /* zeroing the register at the start results in
5296 * shorter and faster code (we can also remove the widening op)
5298 guchar *unordered_check;
5300 amd64_alu_reg_reg (code, X86_XOR, ins->dreg, ins->dreg);
5301 amd64_sse_comisd_reg_reg (code, ins->sreg1, ins->sreg2);
5302 unordered_check = code;
5303 x86_branch8 (code, X86_CC_P, 0, FALSE);
5305 if (ins->opcode == OP_FCEQ) {
5306 amd64_set_reg (code, X86_CC_EQ, ins->dreg, FALSE);
5307 amd64_patch (unordered_check, code);
5309 guchar *jump_to_end;
5310 amd64_set_reg (code, X86_CC_NE, ins->dreg, FALSE);
5312 x86_jump8 (code, 0);
5313 amd64_patch (unordered_check, code);
5314 amd64_inc_reg (code, ins->dreg);
5315 amd64_patch (jump_to_end, code);
5321 /* zeroing the register at the start results in
5322 * shorter and faster code (we can also remove the widening op)
5324 amd64_alu_reg_reg (code, X86_XOR, ins->dreg, ins->dreg);
5325 amd64_sse_comisd_reg_reg (code, ins->sreg2, ins->sreg1);
5326 if (ins->opcode == OP_FCLT_UN) {
5327 guchar *unordered_check = code;
5328 guchar *jump_to_end;
5329 x86_branch8 (code, X86_CC_P, 0, FALSE);
5330 amd64_set_reg (code, X86_CC_GT, ins->dreg, FALSE);
5332 x86_jump8 (code, 0);
5333 amd64_patch (unordered_check, code);
5334 amd64_inc_reg (code, ins->dreg);
5335 amd64_patch (jump_to_end, code);
5337 amd64_set_reg (code, X86_CC_GT, ins->dreg, FALSE);
5342 guchar *unordered_check;
5343 amd64_alu_reg_reg (code, X86_XOR, ins->dreg, ins->dreg);
5344 amd64_sse_comisd_reg_reg (code, ins->sreg2, ins->sreg1);
5345 unordered_check = code;
5346 x86_branch8 (code, X86_CC_P, 0, FALSE);
5347 amd64_set_reg (code, X86_CC_NB, ins->dreg, FALSE);
5348 amd64_patch (unordered_check, code);
5353 /* zeroing the register at the start results in
5354 * shorter and faster code (we can also remove the widening op)
5356 guchar *unordered_check;
5358 amd64_alu_reg_reg (code, X86_XOR, ins->dreg, ins->dreg);
5359 amd64_sse_comisd_reg_reg (code, ins->sreg2, ins->sreg1);
5360 if (ins->opcode == OP_FCGT) {
5361 unordered_check = code;
5362 x86_branch8 (code, X86_CC_P, 0, FALSE);
5363 amd64_set_reg (code, X86_CC_LT, ins->dreg, FALSE);
5364 amd64_patch (unordered_check, code);
5366 amd64_set_reg (code, X86_CC_LT, ins->dreg, FALSE);
5371 guchar *unordered_check;
5372 amd64_alu_reg_reg (code, X86_XOR, ins->dreg, ins->dreg);
5373 amd64_sse_comisd_reg_reg (code, ins->sreg2, ins->sreg1);
5374 unordered_check = code;
5375 x86_branch8 (code, X86_CC_P, 0, FALSE);
5376 amd64_set_reg (code, X86_CC_NA, ins->dreg, FALSE);
5377 amd64_patch (unordered_check, code);
5387 gboolean unordered = FALSE;
5389 amd64_alu_reg_reg (code, X86_XOR, ins->dreg, ins->dreg);
5390 amd64_sse_comiss_reg_reg (code, ins->sreg2, ins->sreg1);
5392 switch (ins->opcode) {
5394 x86_cond = X86_CC_EQ;
5397 x86_cond = X86_CC_LT;
5400 x86_cond = X86_CC_GT;
5403 x86_cond = X86_CC_GT;
5407 x86_cond = X86_CC_LT;
5411 g_assert_not_reached ();
5416 guchar *unordered_check;
5417 guchar *jump_to_end;
5419 unordered_check = code;
5420 x86_branch8 (code, X86_CC_P, 0, FALSE);
5421 amd64_set_reg (code, x86_cond, ins->dreg, FALSE);
5423 x86_jump8 (code, 0);
5424 amd64_patch (unordered_check, code);
5425 amd64_inc_reg (code, ins->dreg);
5426 amd64_patch (jump_to_end, code);
5428 amd64_set_reg (code, x86_cond, ins->dreg, FALSE);
5432 case OP_FCLT_MEMBASE:
5433 case OP_FCGT_MEMBASE:
5434 case OP_FCLT_UN_MEMBASE:
5435 case OP_FCGT_UN_MEMBASE:
5436 case OP_FCEQ_MEMBASE: {
5437 guchar *unordered_check, *jump_to_end;
5440 amd64_alu_reg_reg (code, X86_XOR, ins->dreg, ins->dreg);
5441 amd64_sse_comisd_reg_membase (code, ins->sreg1, ins->sreg2, ins->inst_offset);
5443 switch (ins->opcode) {
5444 case OP_FCEQ_MEMBASE:
5445 x86_cond = X86_CC_EQ;
5447 case OP_FCLT_MEMBASE:
5448 case OP_FCLT_UN_MEMBASE:
5449 x86_cond = X86_CC_LT;
5451 case OP_FCGT_MEMBASE:
5452 case OP_FCGT_UN_MEMBASE:
5453 x86_cond = X86_CC_GT;
5456 g_assert_not_reached ();
5459 unordered_check = code;
5460 x86_branch8 (code, X86_CC_P, 0, FALSE);
5461 amd64_set_reg (code, x86_cond, ins->dreg, FALSE);
5463 switch (ins->opcode) {
5464 case OP_FCEQ_MEMBASE:
5465 case OP_FCLT_MEMBASE:
5466 case OP_FCGT_MEMBASE:
5467 amd64_patch (unordered_check, code);
5469 case OP_FCLT_UN_MEMBASE:
5470 case OP_FCGT_UN_MEMBASE:
5472 x86_jump8 (code, 0);
5473 amd64_patch (unordered_check, code);
5474 amd64_inc_reg (code, ins->dreg);
5475 amd64_patch (jump_to_end, code);
5483 guchar *jump = code;
5484 x86_branch8 (code, X86_CC_P, 0, TRUE);
5485 EMIT_COND_BRANCH (ins, X86_CC_EQ, FALSE);
5486 amd64_patch (jump, code);
5490 /* Branch if C013 != 100 */
5491 /* branch if !ZF or (PF|CF) */
5492 EMIT_COND_BRANCH (ins, X86_CC_NE, FALSE);
5493 EMIT_COND_BRANCH (ins, X86_CC_P, FALSE);
5494 EMIT_COND_BRANCH (ins, X86_CC_B, FALSE);
5497 EMIT_COND_BRANCH (ins, X86_CC_GT, FALSE);
5500 EMIT_COND_BRANCH (ins, X86_CC_P, FALSE);
5501 EMIT_COND_BRANCH (ins, X86_CC_GT, FALSE);
5505 if (ins->opcode == OP_FBGT) {
5508 /* skip branch if C1=1 */
5510 x86_branch8 (code, X86_CC_P, 0, FALSE);
5511 /* branch if (C0 | C3) = 1 */
5512 EMIT_COND_BRANCH (ins, X86_CC_LT, FALSE);
5513 amd64_patch (br1, code);
5516 EMIT_COND_BRANCH (ins, X86_CC_LT, FALSE);
5520 /* Branch if C013 == 100 or 001 */
5523 /* skip branch if C1=1 */
5525 x86_branch8 (code, X86_CC_P, 0, FALSE);
5526 /* branch if (C0 | C3) = 1 */
5527 EMIT_COND_BRANCH (ins, X86_CC_BE, FALSE);
5528 amd64_patch (br1, code);
5532 /* Branch if C013 == 000 */
5533 EMIT_COND_BRANCH (ins, X86_CC_LE, FALSE);
5536 /* Branch if C013=000 or 100 */
5539 /* skip branch if C1=1 */
5541 x86_branch8 (code, X86_CC_P, 0, FALSE);
5542 /* branch if C0=0 */
5543 EMIT_COND_BRANCH (ins, X86_CC_NB, FALSE);
5544 amd64_patch (br1, code);
5548 /* Branch if C013 != 001 */
5549 EMIT_COND_BRANCH (ins, X86_CC_P, FALSE);
5550 EMIT_COND_BRANCH (ins, X86_CC_GE, FALSE);
5553 /* Transfer value to the fp stack */
5554 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, 16);
5555 amd64_movsd_membase_reg (code, AMD64_RSP, 0, ins->sreg1);
5556 amd64_fld_membase (code, AMD64_RSP, 0, TRUE);
5558 amd64_push_reg (code, AMD64_RAX);
5560 amd64_fnstsw (code);
5561 amd64_alu_reg_imm (code, X86_AND, AMD64_RAX, 0x4100);
5562 amd64_alu_reg_imm (code, X86_CMP, AMD64_RAX, X86_FP_C0);
5563 amd64_pop_reg (code, AMD64_RAX);
5564 amd64_fstp (code, 0);
5565 EMIT_COND_SYSTEM_EXCEPTION (X86_CC_EQ, FALSE, "OverflowException");
5566 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, 16);
5569 code = mono_amd64_emit_tls_get (code, ins->dreg, ins->inst_offset);
5573 code = mono_amd64_emit_tls_set (code, ins->sreg1, ins->inst_offset);
5576 case OP_MEMORY_BARRIER: {
5577 if (ins->backend.memory_barrier_kind == MONO_MEMORY_BARRIER_SEQ)
5581 case OP_ATOMIC_ADD_I4:
5582 case OP_ATOMIC_ADD_I8: {
5583 int dreg = ins->dreg;
5584 guint32 size = (ins->opcode == OP_ATOMIC_ADD_I4) ? 4 : 8;
5586 if ((dreg == ins->sreg2) || (dreg == ins->inst_basereg))
5589 amd64_mov_reg_reg (code, dreg, ins->sreg2, size);
5590 amd64_prefix (code, X86_LOCK_PREFIX);
5591 amd64_xadd_membase_reg (code, ins->inst_basereg, ins->inst_offset, dreg, size);
5592 /* dreg contains the old value, add with sreg2 value */
5593 amd64_alu_reg_reg_size (code, X86_ADD, dreg, ins->sreg2, size);
5595 if (ins->dreg != dreg)
5596 amd64_mov_reg_reg (code, ins->dreg, dreg, size);
5600 case OP_ATOMIC_EXCHANGE_I4:
5601 case OP_ATOMIC_EXCHANGE_I8: {
5602 guint32 size = ins->opcode == OP_ATOMIC_EXCHANGE_I4 ? 4 : 8;
5604 /* LOCK prefix is implied. */
5605 amd64_mov_reg_reg (code, GP_SCRATCH_REG, ins->sreg2, size);
5606 amd64_xchg_membase_reg_size (code, ins->sreg1, ins->inst_offset, GP_SCRATCH_REG, size);
5607 amd64_mov_reg_reg (code, ins->dreg, GP_SCRATCH_REG, size);
5610 case OP_ATOMIC_CAS_I4:
5611 case OP_ATOMIC_CAS_I8: {
5614 if (ins->opcode == OP_ATOMIC_CAS_I8)
5620 * See http://msdn.microsoft.com/en-us/magazine/cc302329.aspx for
5621 * an explanation of how this works.
5623 g_assert (ins->sreg3 == AMD64_RAX);
5624 g_assert (ins->sreg1 != AMD64_RAX);
5625 g_assert (ins->sreg1 != ins->sreg2);
5627 amd64_prefix (code, X86_LOCK_PREFIX);
5628 amd64_cmpxchg_membase_reg_size (code, ins->sreg1, ins->inst_offset, ins->sreg2, size);
5630 if (ins->dreg != AMD64_RAX)
5631 amd64_mov_reg_reg (code, ins->dreg, AMD64_RAX, size);
5634 case OP_ATOMIC_LOAD_I1: {
5635 amd64_widen_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, TRUE, FALSE);
5638 case OP_ATOMIC_LOAD_U1: {
5639 amd64_widen_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, FALSE, FALSE);
5642 case OP_ATOMIC_LOAD_I2: {
5643 amd64_widen_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, TRUE, TRUE);
5646 case OP_ATOMIC_LOAD_U2: {
5647 amd64_widen_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, FALSE, TRUE);
5650 case OP_ATOMIC_LOAD_I4: {
5651 amd64_movsxd_reg_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset);
5654 case OP_ATOMIC_LOAD_U4:
5655 case OP_ATOMIC_LOAD_I8:
5656 case OP_ATOMIC_LOAD_U8: {
5657 amd64_mov_reg_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, ins->opcode == OP_ATOMIC_LOAD_U4 ? 4 : 8);
5660 case OP_ATOMIC_LOAD_R4: {
5661 amd64_sse_movss_reg_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset);
5662 amd64_sse_cvtss2sd_reg_reg (code, ins->dreg, ins->dreg);
5665 case OP_ATOMIC_LOAD_R8: {
5666 amd64_sse_movsd_reg_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset);
5669 case OP_ATOMIC_STORE_I1:
5670 case OP_ATOMIC_STORE_U1:
5671 case OP_ATOMIC_STORE_I2:
5672 case OP_ATOMIC_STORE_U2:
5673 case OP_ATOMIC_STORE_I4:
5674 case OP_ATOMIC_STORE_U4:
5675 case OP_ATOMIC_STORE_I8:
5676 case OP_ATOMIC_STORE_U8: {
5679 switch (ins->opcode) {
5680 case OP_ATOMIC_STORE_I1:
5681 case OP_ATOMIC_STORE_U1:
5684 case OP_ATOMIC_STORE_I2:
5685 case OP_ATOMIC_STORE_U2:
5688 case OP_ATOMIC_STORE_I4:
5689 case OP_ATOMIC_STORE_U4:
5692 case OP_ATOMIC_STORE_I8:
5693 case OP_ATOMIC_STORE_U8:
5698 amd64_mov_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, ins->sreg1, size);
5700 if (ins->backend.memory_barrier_kind == MONO_MEMORY_BARRIER_SEQ)
5704 case OP_ATOMIC_STORE_R4: {
5705 amd64_sse_cvtsd2ss_reg_reg (code, MONO_ARCH_FP_SCRATCH_REG, ins->sreg1);
5706 amd64_sse_movss_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, MONO_ARCH_FP_SCRATCH_REG);
5708 if (ins->backend.memory_barrier_kind == MONO_MEMORY_BARRIER_SEQ)
5712 case OP_ATOMIC_STORE_R8: {
5715 amd64_sse_movsd_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, ins->sreg1);
5719 if (ins->backend.memory_barrier_kind == MONO_MEMORY_BARRIER_SEQ)
5723 case OP_CARD_TABLE_WBARRIER: {
5724 int ptr = ins->sreg1;
5725 int value = ins->sreg2;
5727 int nursery_shift, card_table_shift;
5728 gpointer card_table_mask;
5729 size_t nursery_size;
5731 gpointer card_table = mono_gc_get_card_table (&card_table_shift, &card_table_mask);
5732 guint64 nursery_start = (guint64)mono_gc_get_nursery (&nursery_shift, &nursery_size);
5733 guint64 shifted_nursery_start = nursery_start >> nursery_shift;
5735 /*If either point to the stack we can simply avoid the WB. This happens due to
5736 * optimizations revealing a stack store that was not visible when op_cardtable was emited.
5738 if (ins->sreg1 == AMD64_RSP || ins->sreg2 == AMD64_RSP)
5742 * We need one register we can clobber, we choose EDX and make sreg1
5743 * fixed EAX to work around limitations in the local register allocator.
5744 * sreg2 might get allocated to EDX, but that is not a problem since
5745 * we use it before clobbering EDX.
5747 g_assert (ins->sreg1 == AMD64_RAX);
5750 * This is the code we produce:
5753 * edx >>= nursery_shift
5754 * cmp edx, (nursery_start >> nursery_shift)
5757 * edx >>= card_table_shift
5763 if (mono_gc_card_table_nursery_check ()) {
5764 if (value != AMD64_RDX)
5765 amd64_mov_reg_reg (code, AMD64_RDX, value, 8);
5766 amd64_shift_reg_imm (code, X86_SHR, AMD64_RDX, nursery_shift);
5767 if (shifted_nursery_start >> 31) {
5769 * The value we need to compare against is 64 bits, so we need
5770 * another spare register. We use RBX, which we save and
5773 amd64_mov_membase_reg (code, AMD64_RSP, -8, AMD64_RBX, 8);
5774 amd64_mov_reg_imm (code, AMD64_RBX, shifted_nursery_start);
5775 amd64_alu_reg_reg (code, X86_CMP, AMD64_RDX, AMD64_RBX);
5776 amd64_mov_reg_membase (code, AMD64_RBX, AMD64_RSP, -8, 8);
5778 amd64_alu_reg_imm (code, X86_CMP, AMD64_RDX, shifted_nursery_start);
5780 br = code; x86_branch8 (code, X86_CC_NE, -1, FALSE);
5782 amd64_mov_reg_reg (code, AMD64_RDX, ptr, 8);
5783 amd64_shift_reg_imm (code, X86_SHR, AMD64_RDX, card_table_shift);
5784 if (card_table_mask)
5785 amd64_alu_reg_imm (code, X86_AND, AMD64_RDX, (guint32)(guint64)card_table_mask);
5787 mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_GC_CARD_TABLE_ADDR, card_table);
5788 amd64_alu_reg_membase (code, X86_ADD, AMD64_RDX, AMD64_RIP, 0);
5790 amd64_mov_membase_imm (code, AMD64_RDX, 0, 1, 1);
5792 if (mono_gc_card_table_nursery_check ())
5793 x86_patch (br, code);
5796 #ifdef MONO_ARCH_SIMD_INTRINSICS
5797 /* TODO: Some of these IR opcodes are marked as no clobber when they indeed do. */
5799 amd64_sse_addps_reg_reg (code, ins->sreg1, ins->sreg2);
5802 amd64_sse_divps_reg_reg (code, ins->sreg1, ins->sreg2);
5805 amd64_sse_mulps_reg_reg (code, ins->sreg1, ins->sreg2);
5808 amd64_sse_subps_reg_reg (code, ins->sreg1, ins->sreg2);
5811 amd64_sse_maxps_reg_reg (code, ins->sreg1, ins->sreg2);
5814 amd64_sse_minps_reg_reg (code, ins->sreg1, ins->sreg2);
5817 g_assert (ins->inst_c0 >= 0 && ins->inst_c0 <= 7);
5818 amd64_sse_cmpps_reg_reg_imm (code, ins->sreg1, ins->sreg2, ins->inst_c0);
5821 amd64_sse_andps_reg_reg (code, ins->sreg1, ins->sreg2);
5824 amd64_sse_andnps_reg_reg (code, ins->sreg1, ins->sreg2);
5827 amd64_sse_orps_reg_reg (code, ins->sreg1, ins->sreg2);
5830 amd64_sse_xorps_reg_reg (code, ins->sreg1, ins->sreg2);
5833 amd64_sse_sqrtps_reg_reg (code, ins->dreg, ins->sreg1);
5836 amd64_sse_rsqrtps_reg_reg (code, ins->dreg, ins->sreg1);
5839 amd64_sse_rcpps_reg_reg (code, ins->dreg, ins->sreg1);
5842 amd64_sse_addsubps_reg_reg (code, ins->sreg1, ins->sreg2);
5845 amd64_sse_haddps_reg_reg (code, ins->sreg1, ins->sreg2);
5848 amd64_sse_hsubps_reg_reg (code, ins->sreg1, ins->sreg2);
5851 amd64_sse_movshdup_reg_reg (code, ins->dreg, ins->sreg1);
5854 amd64_sse_movsldup_reg_reg (code, ins->dreg, ins->sreg1);
5857 case OP_PSHUFLEW_HIGH:
5858 g_assert (ins->inst_c0 >= 0 && ins->inst_c0 <= 0xFF);
5859 amd64_sse_pshufhw_reg_reg_imm (code, ins->dreg, ins->sreg1, ins->inst_c0);
5861 case OP_PSHUFLEW_LOW:
5862 g_assert (ins->inst_c0 >= 0 && ins->inst_c0 <= 0xFF);
5863 amd64_sse_pshuflw_reg_reg_imm (code, ins->dreg, ins->sreg1, ins->inst_c0);
5866 g_assert (ins->inst_c0 >= 0 && ins->inst_c0 <= 0xFF);
5867 amd64_sse_pshufd_reg_reg_imm (code, ins->dreg, ins->sreg1, ins->inst_c0);
5870 g_assert (ins->inst_c0 >= 0 && ins->inst_c0 <= 0xFF);
5871 amd64_sse_shufps_reg_reg_imm (code, ins->sreg1, ins->sreg2, ins->inst_c0);
5874 g_assert (ins->inst_c0 >= 0 && ins->inst_c0 <= 0x3);
5875 amd64_sse_shufpd_reg_reg_imm (code, ins->sreg1, ins->sreg2, ins->inst_c0);
5879 amd64_sse_addpd_reg_reg (code, ins->sreg1, ins->sreg2);
5882 amd64_sse_divpd_reg_reg (code, ins->sreg1, ins->sreg2);
5885 amd64_sse_mulpd_reg_reg (code, ins->sreg1, ins->sreg2);
5888 amd64_sse_subpd_reg_reg (code, ins->sreg1, ins->sreg2);
5891 amd64_sse_maxpd_reg_reg (code, ins->sreg1, ins->sreg2);
5894 amd64_sse_minpd_reg_reg (code, ins->sreg1, ins->sreg2);
5897 g_assert (ins->inst_c0 >= 0 && ins->inst_c0 <= 7);
5898 amd64_sse_cmppd_reg_reg_imm (code, ins->sreg1, ins->sreg2, ins->inst_c0);
5901 amd64_sse_andpd_reg_reg (code, ins->sreg1, ins->sreg2);
5904 amd64_sse_andnpd_reg_reg (code, ins->sreg1, ins->sreg2);
5907 amd64_sse_orpd_reg_reg (code, ins->sreg1, ins->sreg2);
5910 amd64_sse_xorpd_reg_reg (code, ins->sreg1, ins->sreg2);
5913 amd64_sse_sqrtpd_reg_reg (code, ins->dreg, ins->sreg1);
5916 amd64_sse_addsubpd_reg_reg (code, ins->sreg1, ins->sreg2);
5919 amd64_sse_haddpd_reg_reg (code, ins->sreg1, ins->sreg2);
5922 amd64_sse_hsubpd_reg_reg (code, ins->sreg1, ins->sreg2);
5925 amd64_sse_movddup_reg_reg (code, ins->dreg, ins->sreg1);
5928 case OP_EXTRACT_MASK:
5929 amd64_sse_pmovmskb_reg_reg (code, ins->dreg, ins->sreg1);
5933 amd64_sse_pand_reg_reg (code, ins->sreg1, ins->sreg2);
5936 amd64_sse_por_reg_reg (code, ins->sreg1, ins->sreg2);
5939 amd64_sse_pxor_reg_reg (code, ins->sreg1, ins->sreg2);
5943 amd64_sse_paddb_reg_reg (code, ins->sreg1, ins->sreg2);
5946 amd64_sse_paddw_reg_reg (code, ins->sreg1, ins->sreg2);
5949 amd64_sse_paddd_reg_reg (code, ins->sreg1, ins->sreg2);
5952 amd64_sse_paddq_reg_reg (code, ins->sreg1, ins->sreg2);
5956 amd64_sse_psubb_reg_reg (code, ins->sreg1, ins->sreg2);
5959 amd64_sse_psubw_reg_reg (code, ins->sreg1, ins->sreg2);
5962 amd64_sse_psubd_reg_reg (code, ins->sreg1, ins->sreg2);
5965 amd64_sse_psubq_reg_reg (code, ins->sreg1, ins->sreg2);
5969 amd64_sse_pmaxub_reg_reg (code, ins->sreg1, ins->sreg2);
5972 amd64_sse_pmaxuw_reg_reg (code, ins->sreg1, ins->sreg2);
5975 amd64_sse_pmaxud_reg_reg (code, ins->sreg1, ins->sreg2);
5979 amd64_sse_pmaxsb_reg_reg (code, ins->sreg1, ins->sreg2);
5982 amd64_sse_pmaxsw_reg_reg (code, ins->sreg1, ins->sreg2);
5985 amd64_sse_pmaxsd_reg_reg (code, ins->sreg1, ins->sreg2);
5989 amd64_sse_pavgb_reg_reg (code, ins->sreg1, ins->sreg2);
5992 amd64_sse_pavgw_reg_reg (code, ins->sreg1, ins->sreg2);
5996 amd64_sse_pminub_reg_reg (code, ins->sreg1, ins->sreg2);
5999 amd64_sse_pminuw_reg_reg (code, ins->sreg1, ins->sreg2);
6002 amd64_sse_pminud_reg_reg (code, ins->sreg1, ins->sreg2);
6006 amd64_sse_pminsb_reg_reg (code, ins->sreg1, ins->sreg2);
6009 amd64_sse_pminsw_reg_reg (code, ins->sreg1, ins->sreg2);
6012 amd64_sse_pminsd_reg_reg (code, ins->sreg1, ins->sreg2);
6016 amd64_sse_pcmpeqb_reg_reg (code, ins->sreg1, ins->sreg2);
6019 amd64_sse_pcmpeqw_reg_reg (code, ins->sreg1, ins->sreg2);
6022 amd64_sse_pcmpeqd_reg_reg (code, ins->sreg1, ins->sreg2);
6025 amd64_sse_pcmpeqq_reg_reg (code, ins->sreg1, ins->sreg2);
6029 amd64_sse_pcmpgtb_reg_reg (code, ins->sreg1, ins->sreg2);
6032 amd64_sse_pcmpgtw_reg_reg (code, ins->sreg1, ins->sreg2);
6035 amd64_sse_pcmpgtd_reg_reg (code, ins->sreg1, ins->sreg2);
6038 amd64_sse_pcmpgtq_reg_reg (code, ins->sreg1, ins->sreg2);
6041 case OP_PSUM_ABS_DIFF:
6042 amd64_sse_psadbw_reg_reg (code, ins->sreg1, ins->sreg2);
6045 case OP_UNPACK_LOWB:
6046 amd64_sse_punpcklbw_reg_reg (code, ins->sreg1, ins->sreg2);
6048 case OP_UNPACK_LOWW:
6049 amd64_sse_punpcklwd_reg_reg (code, ins->sreg1, ins->sreg2);
6051 case OP_UNPACK_LOWD:
6052 amd64_sse_punpckldq_reg_reg (code, ins->sreg1, ins->sreg2);
6054 case OP_UNPACK_LOWQ:
6055 amd64_sse_punpcklqdq_reg_reg (code, ins->sreg1, ins->sreg2);
6057 case OP_UNPACK_LOWPS:
6058 amd64_sse_unpcklps_reg_reg (code, ins->sreg1, ins->sreg2);
6060 case OP_UNPACK_LOWPD:
6061 amd64_sse_unpcklpd_reg_reg (code, ins->sreg1, ins->sreg2);
6064 case OP_UNPACK_HIGHB:
6065 amd64_sse_punpckhbw_reg_reg (code, ins->sreg1, ins->sreg2);
6067 case OP_UNPACK_HIGHW:
6068 amd64_sse_punpckhwd_reg_reg (code, ins->sreg1, ins->sreg2);
6070 case OP_UNPACK_HIGHD:
6071 amd64_sse_punpckhdq_reg_reg (code, ins->sreg1, ins->sreg2);
6073 case OP_UNPACK_HIGHQ:
6074 amd64_sse_punpckhqdq_reg_reg (code, ins->sreg1, ins->sreg2);
6076 case OP_UNPACK_HIGHPS:
6077 amd64_sse_unpckhps_reg_reg (code, ins->sreg1, ins->sreg2);
6079 case OP_UNPACK_HIGHPD:
6080 amd64_sse_unpckhpd_reg_reg (code, ins->sreg1, ins->sreg2);
6084 amd64_sse_packsswb_reg_reg (code, ins->sreg1, ins->sreg2);
6087 amd64_sse_packssdw_reg_reg (code, ins->sreg1, ins->sreg2);
6090 amd64_sse_packuswb_reg_reg (code, ins->sreg1, ins->sreg2);
6093 amd64_sse_packusdw_reg_reg (code, ins->sreg1, ins->sreg2);
6096 case OP_PADDB_SAT_UN:
6097 amd64_sse_paddusb_reg_reg (code, ins->sreg1, ins->sreg2);
6099 case OP_PSUBB_SAT_UN:
6100 amd64_sse_psubusb_reg_reg (code, ins->sreg1, ins->sreg2);
6102 case OP_PADDW_SAT_UN:
6103 amd64_sse_paddusw_reg_reg (code, ins->sreg1, ins->sreg2);
6105 case OP_PSUBW_SAT_UN:
6106 amd64_sse_psubusw_reg_reg (code, ins->sreg1, ins->sreg2);
6110 amd64_sse_paddsb_reg_reg (code, ins->sreg1, ins->sreg2);
6113 amd64_sse_psubsb_reg_reg (code, ins->sreg1, ins->sreg2);
6116 amd64_sse_paddsw_reg_reg (code, ins->sreg1, ins->sreg2);
6119 amd64_sse_psubsw_reg_reg (code, ins->sreg1, ins->sreg2);
6123 amd64_sse_pmullw_reg_reg (code, ins->sreg1, ins->sreg2);
6126 amd64_sse_pmulld_reg_reg (code, ins->sreg1, ins->sreg2);
6129 amd64_sse_pmuludq_reg_reg (code, ins->sreg1, ins->sreg2);
6131 case OP_PMULW_HIGH_UN:
6132 amd64_sse_pmulhuw_reg_reg (code, ins->sreg1, ins->sreg2);
6135 amd64_sse_pmulhw_reg_reg (code, ins->sreg1, ins->sreg2);
6139 amd64_sse_psrlw_reg_imm (code, ins->dreg, ins->inst_imm);
6142 amd64_sse_psrlw_reg_reg (code, ins->dreg, ins->sreg2);
6146 amd64_sse_psraw_reg_imm (code, ins->dreg, ins->inst_imm);
6149 amd64_sse_psraw_reg_reg (code, ins->dreg, ins->sreg2);
6153 amd64_sse_psllw_reg_imm (code, ins->dreg, ins->inst_imm);
6156 amd64_sse_psllw_reg_reg (code, ins->dreg, ins->sreg2);
6160 amd64_sse_psrld_reg_imm (code, ins->dreg, ins->inst_imm);
6163 amd64_sse_psrld_reg_reg (code, ins->dreg, ins->sreg2);
6167 amd64_sse_psrad_reg_imm (code, ins->dreg, ins->inst_imm);
6170 amd64_sse_psrad_reg_reg (code, ins->dreg, ins->sreg2);
6174 amd64_sse_pslld_reg_imm (code, ins->dreg, ins->inst_imm);
6177 amd64_sse_pslld_reg_reg (code, ins->dreg, ins->sreg2);
6181 amd64_sse_psrlq_reg_imm (code, ins->dreg, ins->inst_imm);
6184 amd64_sse_psrlq_reg_reg (code, ins->dreg, ins->sreg2);
6187 /*TODO: This is appart of the sse spec but not added
6189 amd64_sse_psraq_reg_imm (code, ins->dreg, ins->inst_imm);
6192 amd64_sse_psraq_reg_reg (code, ins->dreg, ins->sreg2);
6197 amd64_sse_psllq_reg_imm (code, ins->dreg, ins->inst_imm);
6200 amd64_sse_psllq_reg_reg (code, ins->dreg, ins->sreg2);
6203 amd64_sse_cvtdq2pd_reg_reg (code, ins->dreg, ins->sreg1);
6206 amd64_sse_cvtdq2ps_reg_reg (code, ins->dreg, ins->sreg1);
6209 amd64_sse_cvtpd2dq_reg_reg (code, ins->dreg, ins->sreg1);
6212 amd64_sse_cvtpd2ps_reg_reg (code, ins->dreg, ins->sreg1);
6215 amd64_sse_cvtps2dq_reg_reg (code, ins->dreg, ins->sreg1);
6218 amd64_sse_cvtps2pd_reg_reg (code, ins->dreg, ins->sreg1);
6221 amd64_sse_cvttpd2dq_reg_reg (code, ins->dreg, ins->sreg1);
6224 amd64_sse_cvttps2dq_reg_reg (code, ins->dreg, ins->sreg1);
6228 amd64_movd_xreg_reg_size (code, ins->dreg, ins->sreg1, 4);
6231 amd64_movd_reg_xreg_size (code, ins->dreg, ins->sreg1, 4);
6235 amd64_movhlps_reg_reg (code, MONO_ARCH_FP_SCRATCH_REG, ins->sreg1);
6236 amd64_movd_reg_xreg_size (code, ins->dreg, MONO_ARCH_FP_SCRATCH_REG, 8);
6238 amd64_movd_reg_xreg_size (code, ins->dreg, ins->sreg1, 8);
6243 amd64_movd_reg_xreg_size (code, ins->dreg, ins->sreg1, 4);
6245 amd64_shift_reg_imm (code, X86_SHR, ins->dreg, ins->inst_c0 * 8);
6246 amd64_widen_reg (code, ins->dreg, ins->dreg, ins->opcode == OP_EXTRACT_I1, FALSE);
6250 /*amd64_movd_reg_xreg_size (code, ins->dreg, ins->sreg1, 4);
6252 amd64_shift_reg_imm_size (code, X86_SHR, ins->dreg, 16, 4);*/
6253 amd64_sse_pextrw_reg_reg_imm (code, ins->dreg, ins->sreg1, ins->inst_c0);
6254 amd64_widen_reg_size (code, ins->dreg, ins->dreg, ins->opcode == OP_EXTRACT_I2, TRUE, 4);
6258 amd64_movhlps_reg_reg (code, ins->dreg, ins->sreg1);
6260 amd64_sse_movsd_reg_reg (code, ins->dreg, ins->sreg1);
6263 amd64_sse_pinsrw_reg_reg_imm (code, ins->sreg1, ins->sreg2, ins->inst_c0);
6265 case OP_EXTRACTX_U2:
6266 amd64_sse_pextrw_reg_reg_imm (code, ins->dreg, ins->sreg1, ins->inst_c0);
6268 case OP_INSERTX_U1_SLOW:
6269 /*sreg1 is the extracted ireg (scratch)
6270 /sreg2 is the to be inserted ireg (scratch)
6271 /dreg is the xreg to receive the value*/
6273 /*clear the bits from the extracted word*/
6274 amd64_alu_reg_imm (code, X86_AND, ins->sreg1, ins->inst_c0 & 1 ? 0x00FF : 0xFF00);
6275 /*shift the value to insert if needed*/
6276 if (ins->inst_c0 & 1)
6277 amd64_shift_reg_imm_size (code, X86_SHL, ins->sreg2, 8, 4);
6278 /*join them together*/
6279 amd64_alu_reg_reg (code, X86_OR, ins->sreg1, ins->sreg2);
6280 amd64_sse_pinsrw_reg_reg_imm (code, ins->dreg, ins->sreg1, ins->inst_c0 / 2);
6282 case OP_INSERTX_I4_SLOW:
6283 amd64_sse_pinsrw_reg_reg_imm (code, ins->dreg, ins->sreg2, ins->inst_c0 * 2);
6284 amd64_shift_reg_imm (code, X86_SHR, ins->sreg2, 16);
6285 amd64_sse_pinsrw_reg_reg_imm (code, ins->dreg, ins->sreg2, ins->inst_c0 * 2 + 1);
6287 case OP_INSERTX_I8_SLOW:
6288 amd64_movd_xreg_reg_size(code, MONO_ARCH_FP_SCRATCH_REG, ins->sreg2, 8);
6290 amd64_movlhps_reg_reg (code, ins->dreg, MONO_ARCH_FP_SCRATCH_REG);
6292 amd64_sse_movsd_reg_reg (code, ins->dreg, MONO_ARCH_FP_SCRATCH_REG);
6295 case OP_INSERTX_R4_SLOW:
6296 switch (ins->inst_c0) {
6299 amd64_sse_movss_reg_reg (code, ins->dreg, ins->sreg2);
6301 amd64_sse_cvtsd2ss_reg_reg (code, ins->dreg, ins->sreg2);
6304 amd64_sse_pshufd_reg_reg_imm (code, ins->dreg, ins->dreg, mono_simd_shuffle_mask(1, 0, 2, 3));
6306 amd64_sse_movss_reg_reg (code, ins->dreg, ins->sreg2);
6308 amd64_sse_cvtsd2ss_reg_reg (code, ins->dreg, ins->sreg2);
6309 amd64_sse_pshufd_reg_reg_imm (code, ins->dreg, ins->dreg, mono_simd_shuffle_mask(1, 0, 2, 3));
6312 amd64_sse_pshufd_reg_reg_imm (code, ins->dreg, ins->dreg, mono_simd_shuffle_mask(2, 1, 0, 3));
6314 amd64_sse_movss_reg_reg (code, ins->dreg, ins->sreg2);
6316 amd64_sse_cvtsd2ss_reg_reg (code, ins->dreg, ins->sreg2);
6317 amd64_sse_pshufd_reg_reg_imm (code, ins->dreg, ins->dreg, mono_simd_shuffle_mask(2, 1, 0, 3));
6320 amd64_sse_pshufd_reg_reg_imm (code, ins->dreg, ins->dreg, mono_simd_shuffle_mask(3, 1, 2, 0));
6322 amd64_sse_movss_reg_reg (code, ins->dreg, ins->sreg2);
6324 amd64_sse_cvtsd2ss_reg_reg (code, ins->dreg, ins->sreg2);
6325 amd64_sse_pshufd_reg_reg_imm (code, ins->dreg, ins->dreg, mono_simd_shuffle_mask(3, 1, 2, 0));
6329 case OP_INSERTX_R8_SLOW:
6331 amd64_movlhps_reg_reg (code, ins->dreg, ins->sreg2);
6333 amd64_sse_movsd_reg_reg (code, ins->dreg, ins->sreg2);
6335 case OP_STOREX_MEMBASE_REG:
6336 case OP_STOREX_MEMBASE:
6337 amd64_sse_movups_membase_reg (code, ins->dreg, ins->inst_offset, ins->sreg1);
6339 case OP_LOADX_MEMBASE:
6340 amd64_sse_movups_reg_membase (code, ins->dreg, ins->sreg1, ins->inst_offset);
6342 case OP_LOADX_ALIGNED_MEMBASE:
6343 amd64_sse_movaps_reg_membase (code, ins->dreg, ins->sreg1, ins->inst_offset);
6345 case OP_STOREX_ALIGNED_MEMBASE_REG:
6346 amd64_sse_movaps_membase_reg (code, ins->dreg, ins->inst_offset, ins->sreg1);
6348 case OP_STOREX_NTA_MEMBASE_REG:
6349 amd64_sse_movntps_reg_membase (code, ins->dreg, ins->sreg1, ins->inst_offset);
6351 case OP_PREFETCH_MEMBASE:
6352 amd64_sse_prefetch_reg_membase (code, ins->backend.arg_info, ins->sreg1, ins->inst_offset);
6356 /*FIXME the peephole pass should have killed this*/
6357 if (ins->dreg != ins->sreg1)
6358 amd64_sse_movaps_reg_reg (code, ins->dreg, ins->sreg1);
6361 amd64_sse_pxor_reg_reg (code, ins->dreg, ins->dreg);
6364 amd64_sse_pcmpeqb_reg_reg (code, ins->dreg, ins->dreg);
6366 case OP_ICONV_TO_R4_RAW:
6367 amd64_movd_xreg_reg_size (code, ins->dreg, ins->sreg1, 4);
6369 amd64_sse_cvtss2sd_reg_reg (code, ins->dreg, ins->dreg);
6372 case OP_FCONV_TO_R8_X:
6373 amd64_sse_movsd_reg_reg (code, ins->dreg, ins->sreg1);
6376 case OP_XCONV_R8_TO_I4:
6377 amd64_sse_cvttsd2si_reg_xreg_size (code, ins->dreg, ins->sreg1, 4);
6378 switch (ins->backend.source_opcode) {
6379 case OP_FCONV_TO_I1:
6380 amd64_widen_reg (code, ins->dreg, ins->dreg, TRUE, FALSE);
6382 case OP_FCONV_TO_U1:
6383 amd64_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
6385 case OP_FCONV_TO_I2:
6386 amd64_widen_reg (code, ins->dreg, ins->dreg, TRUE, TRUE);
6388 case OP_FCONV_TO_U2:
6389 amd64_widen_reg (code, ins->dreg, ins->dreg, FALSE, TRUE);
6395 amd64_sse_pinsrw_reg_reg_imm (code, ins->dreg, ins->sreg1, 0);
6396 amd64_sse_pinsrw_reg_reg_imm (code, ins->dreg, ins->sreg1, 1);
6397 amd64_sse_pshufd_reg_reg_imm (code, ins->dreg, ins->dreg, 0);
6400 amd64_movd_xreg_reg_size (code, ins->dreg, ins->sreg1, 4);
6401 amd64_sse_pshufd_reg_reg_imm (code, ins->dreg, ins->dreg, 0);
6404 amd64_movd_xreg_reg_size (code, ins->dreg, ins->sreg1, 8);
6405 amd64_sse_pshufd_reg_reg_imm (code, ins->dreg, ins->dreg, 0x44);
6409 amd64_sse_movsd_reg_reg (code, ins->dreg, ins->sreg1);
6411 amd64_sse_movsd_reg_reg (code, ins->dreg, ins->sreg1);
6412 amd64_sse_cvtsd2ss_reg_reg (code, ins->dreg, ins->dreg);
6414 amd64_sse_pshufd_reg_reg_imm (code, ins->dreg, ins->dreg, 0);
6417 amd64_sse_movsd_reg_reg (code, ins->dreg, ins->sreg1);
6418 amd64_sse_pshufd_reg_reg_imm (code, ins->dreg, ins->dreg, 0x44);
6421 case OP_LIVERANGE_START: {
6422 if (cfg->verbose_level > 1)
6423 printf ("R%d START=0x%x\n", MONO_VARINFO (cfg, ins->inst_c0)->vreg, (int)(code - cfg->native_code));
6424 MONO_VARINFO (cfg, ins->inst_c0)->live_range_start = code - cfg->native_code;
6427 case OP_LIVERANGE_END: {
6428 if (cfg->verbose_level > 1)
6429 printf ("R%d END=0x%x\n", MONO_VARINFO (cfg, ins->inst_c0)->vreg, (int)(code - cfg->native_code));
6430 MONO_VARINFO (cfg, ins->inst_c0)->live_range_end = code - cfg->native_code;
6433 case OP_GC_SAFE_POINT: {
6436 g_assert (mono_threads_is_coop_enabled ());
6438 amd64_test_membase_imm_size (code, ins->sreg1, 0, 1, 4);
6439 br[0] = code; x86_branch8 (code, X86_CC_EQ, 0, FALSE);
6440 code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD, "mono_threads_state_poll", FALSE);
6441 amd64_patch (br[0], code);
6445 case OP_GC_LIVENESS_DEF:
6446 case OP_GC_LIVENESS_USE:
6447 case OP_GC_PARAM_SLOT_LIVENESS_DEF:
6448 ins->backend.pc_offset = code - cfg->native_code;
6450 case OP_GC_SPILL_SLOT_LIVENESS_DEF:
6451 ins->backend.pc_offset = code - cfg->native_code;
6452 bb->spill_slot_defs = g_slist_prepend_mempool (cfg->mempool, bb->spill_slot_defs, ins);
6454 case OP_GET_LAST_ERROR:
6455 emit_get_last_error(code, ins->dreg);
6458 g_warning ("unknown opcode %s in %s()\n", mono_inst_name (ins->opcode), __FUNCTION__);
6459 g_assert_not_reached ();
6462 if ((code - cfg->native_code - offset) > max_len) {
6463 g_warning ("wrong maximal instruction length of instruction %s (expected %d, got %ld)",
6464 mono_inst_name (ins->opcode), max_len, code - cfg->native_code - offset);
6465 g_assert_not_reached ();
6469 cfg->code_len = code - cfg->native_code;
6472 #endif /* DISABLE_JIT */
6475 mono_arch_register_lowlevel_calls (void)
6477 /* The signature doesn't matter */
6478 mono_register_jit_icall (mono_amd64_throw_exception, "mono_amd64_throw_exception", mono_create_icall_signature ("void"), TRUE);
6482 mono_arch_patch_code_new (MonoCompile *cfg, MonoDomain *domain, guint8 *code, MonoJumpInfo *ji, gpointer target)
6484 unsigned char *ip = ji->ip.i + code;
6487 * Debug code to help track down problems where the target of a near call is
6490 if (amd64_is_near_call (ip)) {
6491 gint64 disp = (guint8*)target - (guint8*)ip;
6493 if (!amd64_is_imm32 (disp)) {
6494 printf ("TYPE: %d\n", ji->type);
6496 case MONO_PATCH_INFO_INTERNAL_METHOD:
6497 printf ("V: %s\n", ji->data.name);
6499 case MONO_PATCH_INFO_METHOD_JUMP:
6500 case MONO_PATCH_INFO_METHOD:
6501 printf ("V: %s\n", ji->data.method->name);
6509 amd64_patch (ip, (gpointer)target);
6515 get_max_epilog_size (MonoCompile *cfg)
6517 int max_epilog_size = 16;
6519 if (cfg->method->save_lmf)
6520 max_epilog_size += 256;
6522 if (mono_jit_trace_calls != NULL)
6523 max_epilog_size += 50;
6525 if (cfg->prof_options & MONO_PROFILE_ENTER_LEAVE)
6526 max_epilog_size += 50;
6528 max_epilog_size += (AMD64_NREG * 2);
6530 return max_epilog_size;
6534 * This macro is used for testing whenever the unwinder works correctly at every point
6535 * where an async exception can happen.
6537 /* This will generate a SIGSEGV at the given point in the code */
6538 #define async_exc_point(code) do { \
6539 if (mono_inject_async_exc_method && mono_method_desc_full_match (mono_inject_async_exc_method, cfg->method)) { \
6540 if (cfg->arch.async_point_count == mono_inject_async_exc_pos) \
6541 amd64_mov_reg_mem (code, AMD64_RAX, 0, 4); \
6542 cfg->arch.async_point_count ++; \
6547 mono_arch_emit_prolog (MonoCompile *cfg)
6549 MonoMethod *method = cfg->method;
6551 MonoMethodSignature *sig;
6553 int alloc_size, pos, i, cfa_offset, quad, max_epilog_size, save_area_offset;
6556 MonoInst *lmf_var = cfg->lmf_var;
6557 gboolean args_clobbered = FALSE;
6558 gboolean trace = FALSE;
6560 cfg->code_size = MAX (cfg->header->code_size * 4, 1024);
6562 code = cfg->native_code = (unsigned char *)g_malloc (cfg->code_size);
6564 if (mono_jit_trace_calls != NULL && mono_trace_eval (method))
6567 /* Amount of stack space allocated by register saving code */
6570 /* Offset between RSP and the CFA */
6574 * The prolog consists of the following parts:
6576 * - push rbp, mov rbp, rsp
6577 * - save callee saved regs using pushes
6579 * - save rgctx if needed
6580 * - save lmf if needed
6583 * - save rgctx if needed
6584 * - save lmf if needed
6585 * - save callee saved regs using moves
6590 mono_emit_unwind_op_def_cfa (cfg, code, AMD64_RSP, 8);
6591 // IP saved at CFA - 8
6592 mono_emit_unwind_op_offset (cfg, code, AMD64_RIP, -cfa_offset);
6593 async_exc_point (code);
6594 mini_gc_set_slot_type_from_cfa (cfg, -cfa_offset, SLOT_NOREF);
6596 if (!cfg->arch.omit_fp) {
6597 amd64_push_reg (code, AMD64_RBP);
6599 mono_emit_unwind_op_def_cfa_offset (cfg, code, cfa_offset);
6600 mono_emit_unwind_op_offset (cfg, code, AMD64_RBP, - cfa_offset);
6601 async_exc_point (code);
6603 mono_arch_unwindinfo_add_push_nonvol (&cfg->arch.unwindinfo, cfg->native_code, code, AMD64_RBP);
6605 /* These are handled automatically by the stack marking code */
6606 mini_gc_set_slot_type_from_cfa (cfg, -cfa_offset, SLOT_NOREF);
6608 amd64_mov_reg_reg (code, AMD64_RBP, AMD64_RSP, sizeof(mgreg_t));
6609 mono_emit_unwind_op_def_cfa_reg (cfg, code, AMD64_RBP);
6610 async_exc_point (code);
6612 mono_arch_unwindinfo_add_set_fpreg (&cfg->arch.unwindinfo, cfg->native_code, code, AMD64_RBP);
6616 /* The param area is always at offset 0 from sp */
6617 /* This needs to be allocated here, since it has to come after the spill area */
6618 if (cfg->param_area) {
6619 if (cfg->arch.omit_fp)
6621 g_assert_not_reached ();
6622 cfg->stack_offset += ALIGN_TO (cfg->param_area, sizeof(mgreg_t));
6625 if (cfg->arch.omit_fp) {
6627 * On enter, the stack is misaligned by the pushing of the return
6628 * address. It is either made aligned by the pushing of %rbp, or by
6631 alloc_size = ALIGN_TO (cfg->stack_offset, 8);
6632 if ((alloc_size % 16) == 0) {
6634 /* Mark the padding slot as NOREF */
6635 mini_gc_set_slot_type_from_cfa (cfg, -cfa_offset - sizeof (mgreg_t), SLOT_NOREF);
6638 alloc_size = ALIGN_TO (cfg->stack_offset, MONO_ARCH_FRAME_ALIGNMENT);
6639 if (cfg->stack_offset != alloc_size) {
6640 /* Mark the padding slot as NOREF */
6641 mini_gc_set_slot_type_from_fp (cfg, -alloc_size + cfg->param_area, SLOT_NOREF);
6643 cfg->arch.sp_fp_offset = alloc_size;
6647 cfg->arch.stack_alloc_size = alloc_size;
6649 /* Allocate stack frame */
6651 /* See mono_emit_stack_alloc */
6652 #if defined(TARGET_WIN32) || defined(MONO_ARCH_SIGSEGV_ON_ALTSTACK)
6653 guint32 remaining_size = alloc_size;
6654 /*FIXME handle unbounded code expansion, we should use a loop in case of more than X interactions*/
6655 guint32 required_code_size = ((remaining_size / 0x1000) + 1) * 11; /*11 is the max size of amd64_alu_reg_imm + amd64_test_membase_reg*/
6656 guint32 offset = code - cfg->native_code;
6657 if (G_UNLIKELY (required_code_size >= (cfg->code_size - offset))) {
6658 while (required_code_size >= (cfg->code_size - offset))
6659 cfg->code_size *= 2;
6660 cfg->native_code = (unsigned char *)mono_realloc_native_code (cfg);
6661 code = cfg->native_code + offset;
6662 cfg->stat_code_reallocs++;
6665 while (remaining_size >= 0x1000) {
6666 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, 0x1000);
6667 if (cfg->arch.omit_fp) {
6668 cfa_offset += 0x1000;
6669 mono_emit_unwind_op_def_cfa_offset (cfg, code, cfa_offset);
6671 async_exc_point (code);
6673 if (cfg->arch.omit_fp)
6674 mono_arch_unwindinfo_add_alloc_stack (&cfg->arch.unwindinfo, cfg->native_code, code, 0x1000);
6677 amd64_test_membase_reg (code, AMD64_RSP, 0, AMD64_RSP);
6678 remaining_size -= 0x1000;
6680 if (remaining_size) {
6681 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, remaining_size);
6682 if (cfg->arch.omit_fp) {
6683 cfa_offset += remaining_size;
6684 mono_emit_unwind_op_def_cfa_offset (cfg, code, cfa_offset);
6685 async_exc_point (code);
6688 if (cfg->arch.omit_fp)
6689 mono_arch_unwindinfo_add_alloc_stack (&cfg->arch.unwindinfo, cfg->native_code, code, remaining_size);
6693 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, alloc_size);
6694 if (cfg->arch.omit_fp) {
6695 cfa_offset += alloc_size;
6696 mono_emit_unwind_op_def_cfa_offset (cfg, code, cfa_offset);
6697 async_exc_point (code);
6702 /* Stack alignment check */
6707 amd64_mov_reg_reg (code, AMD64_RAX, AMD64_RSP, 8);
6708 amd64_alu_reg_imm (code, X86_AND, AMD64_RAX, 0xf);
6709 amd64_alu_reg_imm (code, X86_CMP, AMD64_RAX, 0);
6711 x86_branch8 (code, X86_CC_EQ, 1, FALSE);
6712 amd64_breakpoint (code);
6713 amd64_patch (buf, code);
6717 if (mini_get_debug_options ()->init_stacks) {
6718 /* Fill the stack frame with a dummy value to force deterministic behavior */
6720 /* Save registers to the red zone */
6721 amd64_mov_membase_reg (code, AMD64_RSP, -8, AMD64_RDI, 8);
6722 amd64_mov_membase_reg (code, AMD64_RSP, -16, AMD64_RCX, 8);
6724 amd64_mov_reg_imm (code, AMD64_RAX, 0x2a2a2a2a2a2a2a2a);
6725 amd64_mov_reg_imm (code, AMD64_RCX, alloc_size / 8);
6726 amd64_mov_reg_reg (code, AMD64_RDI, AMD64_RSP, 8);
6729 amd64_prefix (code, X86_REP_PREFIX);
6732 amd64_mov_reg_membase (code, AMD64_RDI, AMD64_RSP, -8, 8);
6733 amd64_mov_reg_membase (code, AMD64_RCX, AMD64_RSP, -16, 8);
6737 if (method->save_lmf)
6738 code = emit_setup_lmf (cfg, code, lmf_var->inst_offset, cfa_offset);
6740 /* Save callee saved registers */
6741 if (cfg->arch.omit_fp) {
6742 save_area_offset = cfg->arch.reg_save_area_offset;
6743 /* Save caller saved registers after sp is adjusted */
6744 /* The registers are saved at the bottom of the frame */
6745 /* FIXME: Optimize this so the regs are saved at the end of the frame in increasing order */
6747 /* The registers are saved just below the saved rbp */
6748 save_area_offset = cfg->arch.reg_save_area_offset;
6751 for (i = 0; i < AMD64_NREG; ++i) {
6752 if (AMD64_IS_CALLEE_SAVED_REG (i) && (cfg->arch.saved_iregs & (1 << i))) {
6753 amd64_mov_membase_reg (code, cfg->frame_reg, save_area_offset, i, 8);
6755 if (cfg->arch.omit_fp) {
6756 mono_emit_unwind_op_offset (cfg, code, i, - (cfa_offset - save_area_offset));
6757 /* These are handled automatically by the stack marking code */
6758 mini_gc_set_slot_type_from_cfa (cfg, - (cfa_offset - save_area_offset), SLOT_NOREF);
6760 mono_emit_unwind_op_offset (cfg, code, i, - (-save_area_offset + (2 * 8)));
6764 save_area_offset += 8;
6765 async_exc_point (code);
6769 /* store runtime generic context */
6770 if (cfg->rgctx_var) {
6771 g_assert (cfg->rgctx_var->opcode == OP_REGOFFSET &&
6772 (cfg->rgctx_var->inst_basereg == AMD64_RBP || cfg->rgctx_var->inst_basereg == AMD64_RSP));
6774 amd64_mov_membase_reg (code, cfg->rgctx_var->inst_basereg, cfg->rgctx_var->inst_offset, MONO_ARCH_RGCTX_REG, sizeof(gpointer));
6776 mono_add_var_location (cfg, cfg->rgctx_var, TRUE, MONO_ARCH_RGCTX_REG, 0, 0, code - cfg->native_code);
6777 mono_add_var_location (cfg, cfg->rgctx_var, FALSE, cfg->rgctx_var->inst_basereg, cfg->rgctx_var->inst_offset, code - cfg->native_code, 0);
6780 /* compute max_length in order to use short forward jumps */
6781 max_epilog_size = get_max_epilog_size (cfg);
6782 if (cfg->opt & MONO_OPT_BRANCH) {
6783 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
6787 if (cfg->prof_options & MONO_PROFILE_COVERAGE)
6789 /* max alignment for loops */
6790 if ((cfg->opt & MONO_OPT_LOOP) && bb_is_loop_start (bb))
6791 max_length += LOOP_ALIGNMENT;
6793 MONO_BB_FOR_EACH_INS (bb, ins) {
6794 max_length += ((guint8 *)ins_get_spec (ins->opcode))[MONO_INST_LEN];
6797 /* Take prolog and epilog instrumentation into account */
6798 if (bb == cfg->bb_entry || bb == cfg->bb_exit)
6799 max_length += max_epilog_size;
6801 bb->max_length = max_length;
6805 sig = mono_method_signature (method);
6808 cinfo = (CallInfo *)cfg->arch.cinfo;
6810 if (sig->ret->type != MONO_TYPE_VOID) {
6811 /* Save volatile arguments to the stack */
6812 if (cfg->vret_addr && (cfg->vret_addr->opcode != OP_REGVAR))
6813 amd64_mov_membase_reg (code, cfg->vret_addr->inst_basereg, cfg->vret_addr->inst_offset, cinfo->ret.reg, 8);
6816 /* Keep this in sync with emit_load_volatile_arguments */
6817 for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
6818 ArgInfo *ainfo = cinfo->args + i;
6820 ins = cfg->args [i];
6822 if ((ins->flags & MONO_INST_IS_DEAD) && !trace)
6823 /* Unused arguments */
6826 /* Save volatile arguments to the stack */
6827 if (ins->opcode != OP_REGVAR) {
6828 switch (ainfo->storage) {
6834 if (stack_offset & 0x1)
6836 else if (stack_offset & 0x2)
6838 else if (stack_offset & 0x4)
6843 amd64_mov_membase_reg (code, ins->inst_basereg, ins->inst_offset, ainfo->reg, size);
6846 * Save the original location of 'this',
6847 * get_generic_info_from_stack_frame () needs this to properly look up
6848 * the argument value during the handling of async exceptions.
6850 if (ins == cfg->args [0]) {
6851 mono_add_var_location (cfg, ins, TRUE, ainfo->reg, 0, 0, code - cfg->native_code);
6852 mono_add_var_location (cfg, ins, FALSE, ins->inst_basereg, ins->inst_offset, code - cfg->native_code, 0);
6856 case ArgInFloatSSEReg:
6857 amd64_movss_membase_reg (code, ins->inst_basereg, ins->inst_offset, ainfo->reg);
6859 case ArgInDoubleSSEReg:
6860 amd64_movsd_membase_reg (code, ins->inst_basereg, ins->inst_offset, ainfo->reg);
6862 case ArgValuetypeInReg:
6863 for (quad = 0; quad < 2; quad ++) {
6864 switch (ainfo->pair_storage [quad]) {
6866 amd64_mov_membase_reg (code, ins->inst_basereg, ins->inst_offset + (quad * sizeof(mgreg_t)), ainfo->pair_regs [quad], sizeof(mgreg_t));
6868 case ArgInFloatSSEReg:
6869 amd64_movss_membase_reg (code, ins->inst_basereg, ins->inst_offset + (quad * sizeof(mgreg_t)), ainfo->pair_regs [quad]);
6871 case ArgInDoubleSSEReg:
6872 amd64_movsd_membase_reg (code, ins->inst_basereg, ins->inst_offset + (quad * sizeof(mgreg_t)), ainfo->pair_regs [quad]);
6877 g_assert_not_reached ();
6881 case ArgValuetypeAddrInIReg:
6882 if (ainfo->pair_storage [0] == ArgInIReg)
6883 amd64_mov_membase_reg (code, ins->inst_left->inst_basereg, ins->inst_left->inst_offset, ainfo->pair_regs [0], sizeof (gpointer));
6885 case ArgValuetypeAddrOnStack:
6887 case ArgGSharedVtInReg:
6888 amd64_mov_membase_reg (code, ins->inst_basereg, ins->inst_offset, ainfo->reg, 8);
6894 /* Argument allocated to (non-volatile) register */
6895 switch (ainfo->storage) {
6897 amd64_mov_reg_reg (code, ins->dreg, ainfo->reg, 8);
6900 amd64_mov_reg_membase (code, ins->dreg, AMD64_RBP, ARGS_OFFSET + ainfo->offset, 8);
6903 g_assert_not_reached ();
6906 if (ins == cfg->args [0]) {
6907 mono_add_var_location (cfg, ins, TRUE, ainfo->reg, 0, 0, code - cfg->native_code);
6908 mono_add_var_location (cfg, ins, TRUE, ins->dreg, 0, code - cfg->native_code, 0);
6913 if (cfg->method->save_lmf)
6914 args_clobbered = TRUE;
6917 args_clobbered = TRUE;
6918 code = (guint8 *)mono_arch_instrument_prolog (cfg, mono_trace_enter_method, code, TRUE);
6921 if (cfg->prof_options & MONO_PROFILE_ENTER_LEAVE)
6922 args_clobbered = TRUE;
6925 * Optimize the common case of the first bblock making a call with the same
6926 * arguments as the method. This works because the arguments are still in their
6927 * original argument registers.
6928 * FIXME: Generalize this
6930 if (!args_clobbered) {
6931 MonoBasicBlock *first_bb = cfg->bb_entry;
6933 int filter = FILTER_IL_SEQ_POINT;
6935 next = mono_bb_first_inst (first_bb, filter);
6936 if (!next && first_bb->next_bb) {
6937 first_bb = first_bb->next_bb;
6938 next = mono_bb_first_inst (first_bb, filter);
6941 if (first_bb->in_count > 1)
6944 for (i = 0; next && i < sig->param_count + sig->hasthis; ++i) {
6945 ArgInfo *ainfo = cinfo->args + i;
6946 gboolean match = FALSE;
6948 ins = cfg->args [i];
6949 if (ins->opcode != OP_REGVAR) {
6950 switch (ainfo->storage) {
6952 if (((next->opcode == OP_LOAD_MEMBASE) || (next->opcode == OP_LOADI4_MEMBASE)) && next->inst_basereg == ins->inst_basereg && next->inst_offset == ins->inst_offset) {
6953 if (next->dreg == ainfo->reg) {
6957 next->opcode = OP_MOVE;
6958 next->sreg1 = ainfo->reg;
6959 /* Only continue if the instruction doesn't change argument regs */
6960 if (next->dreg == ainfo->reg || next->dreg == AMD64_RAX)
6970 /* Argument allocated to (non-volatile) register */
6971 switch (ainfo->storage) {
6973 if (next->opcode == OP_MOVE && next->sreg1 == ins->dreg && next->dreg == ainfo->reg) {
6984 next = mono_inst_next (next, filter);
6985 //next = mono_inst_list_next (&next->node, &first_bb->ins_list);
6992 if (cfg->gen_sdb_seq_points) {
6993 MonoInst *info_var = (MonoInst *)cfg->arch.seq_point_info_var;
6995 /* Initialize seq_point_info_var */
6996 if (cfg->compile_aot) {
6997 /* Initialize the variable from a GOT slot */
6998 /* Same as OP_AOTCONST */
6999 mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_SEQ_POINT_INFO, cfg->method);
7000 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RIP, 0, sizeof(gpointer));
7001 g_assert (info_var->opcode == OP_REGOFFSET);
7002 amd64_mov_membase_reg (code, info_var->inst_basereg, info_var->inst_offset, AMD64_R11, 8);
7005 if (cfg->compile_aot) {
7006 /* Initialize ss_tramp_var */
7007 ins = (MonoInst *)cfg->arch.ss_tramp_var;
7008 g_assert (ins->opcode == OP_REGOFFSET);
7010 amd64_mov_reg_membase (code, AMD64_R11, info_var->inst_basereg, info_var->inst_offset, 8);
7011 amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, MONO_STRUCT_OFFSET (SeqPointInfo, ss_tramp_addr), 8);
7012 amd64_mov_membase_reg (code, ins->inst_basereg, ins->inst_offset, AMD64_R11, 8);
7014 /* Initialize ss_tramp_var */
7015 ins = (MonoInst *)cfg->arch.ss_tramp_var;
7016 g_assert (ins->opcode == OP_REGOFFSET);
7018 amd64_mov_reg_imm (code, AMD64_R11, (guint64)&ss_trampoline);
7019 amd64_mov_membase_reg (code, ins->inst_basereg, ins->inst_offset, AMD64_R11, 8);
7021 /* Initialize bp_tramp_var */
7022 ins = (MonoInst *)cfg->arch.bp_tramp_var;
7023 g_assert (ins->opcode == OP_REGOFFSET);
7025 amd64_mov_reg_imm (code, AMD64_R11, (guint64)&bp_trampoline);
7026 amd64_mov_membase_reg (code, ins->inst_basereg, ins->inst_offset, AMD64_R11, 8);
7030 cfg->code_len = code - cfg->native_code;
7032 g_assert (cfg->code_len < cfg->code_size);
7038 mono_arch_emit_epilog (MonoCompile *cfg)
7040 MonoMethod *method = cfg->method;
7043 int max_epilog_size;
7045 gint32 lmf_offset = cfg->lmf_var ? ((MonoInst*)cfg->lmf_var)->inst_offset : -1;
7046 gint32 save_area_offset = cfg->arch.reg_save_area_offset;
7048 max_epilog_size = get_max_epilog_size (cfg);
7050 while (cfg->code_len + max_epilog_size > (cfg->code_size - 16)) {
7051 cfg->code_size *= 2;
7052 cfg->native_code = (unsigned char *)mono_realloc_native_code (cfg);
7053 cfg->stat_code_reallocs++;
7055 code = cfg->native_code + cfg->code_len;
7057 cfg->has_unwind_info_for_epilog = TRUE;
7059 /* Mark the start of the epilog */
7060 mono_emit_unwind_op_mark_loc (cfg, code, 0);
7062 /* Save the uwind state which is needed by the out-of-line code */
7063 mono_emit_unwind_op_remember_state (cfg, code);
7065 if (mono_jit_trace_calls != NULL && mono_trace_eval (method))
7066 code = (guint8 *)mono_arch_instrument_epilog (cfg, mono_trace_leave_method, code, TRUE);
7068 /* the code restoring the registers must be kept in sync with OP_TAILCALL */
7070 if (method->save_lmf) {
7071 /* check if we need to restore protection of the stack after a stack overflow */
7072 if (!cfg->compile_aot && mono_arch_have_fast_tls () && mono_tls_get_tls_offset (TLS_KEY_JIT_TLS) != -1) {
7074 code = mono_amd64_emit_tls_get (code, AMD64_RCX, mono_tls_get_tls_offset (TLS_KEY_JIT_TLS));
7075 /* we load the value in a separate instruction: this mechanism may be
7076 * used later as a safer way to do thread interruption
7078 amd64_mov_reg_membase (code, AMD64_RCX, AMD64_RCX, MONO_STRUCT_OFFSET (MonoJitTlsData, restore_stack_prot), 8);
7079 x86_alu_reg_imm (code, X86_CMP, X86_ECX, 0);
7081 x86_branch8 (code, X86_CC_Z, 0, FALSE);
7082 /* note that the call trampoline will preserve eax/edx */
7083 x86_call_reg (code, X86_ECX);
7084 x86_patch (patch, code);
7086 /* FIXME: maybe save the jit tls in the prolog */
7088 if (cfg->used_int_regs & (1 << AMD64_RBP)) {
7089 amd64_mov_reg_membase (code, AMD64_RBP, cfg->frame_reg, lmf_offset + MONO_STRUCT_OFFSET (MonoLMF, rbp), 8);
7093 /* Restore callee saved regs */
7094 for (i = 0; i < AMD64_NREG; ++i) {
7095 if (AMD64_IS_CALLEE_SAVED_REG (i) && (cfg->arch.saved_iregs & (1 << i))) {
7096 /* Restore only used_int_regs, not arch.saved_iregs */
7097 #if defined(MONO_SUPPORT_TASKLETS)
7100 int restore_reg=(cfg->used_int_regs & (1 << i));
7103 amd64_mov_reg_membase (code, i, cfg->frame_reg, save_area_offset, 8);
7104 mono_emit_unwind_op_same_value (cfg, code, i);
7105 async_exc_point (code);
7107 save_area_offset += 8;
7111 /* Load returned vtypes into registers if needed */
7112 cinfo = (CallInfo *)cfg->arch.cinfo;
7113 if (cinfo->ret.storage == ArgValuetypeInReg) {
7114 ArgInfo *ainfo = &cinfo->ret;
7115 MonoInst *inst = cfg->ret;
7117 for (quad = 0; quad < 2; quad ++) {
7118 switch (ainfo->pair_storage [quad]) {
7120 amd64_mov_reg_membase (code, ainfo->pair_regs [quad], inst->inst_basereg, inst->inst_offset + (quad * sizeof(mgreg_t)), ainfo->pair_size [quad]);
7122 case ArgInFloatSSEReg:
7123 amd64_movss_reg_membase (code, ainfo->pair_regs [quad], inst->inst_basereg, inst->inst_offset + (quad * sizeof(mgreg_t)));
7125 case ArgInDoubleSSEReg:
7126 amd64_movsd_reg_membase (code, ainfo->pair_regs [quad], inst->inst_basereg, inst->inst_offset + (quad * sizeof(mgreg_t)));
7131 g_assert_not_reached ();
7136 if (cfg->arch.omit_fp) {
7137 if (cfg->arch.stack_alloc_size) {
7138 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, cfg->arch.stack_alloc_size);
7142 mono_emit_unwind_op_same_value (cfg, code, AMD64_RBP);
7144 mono_emit_unwind_op_def_cfa (cfg, code, AMD64_RSP, 8);
7145 async_exc_point (code);
7148 /* Restore the unwind state to be the same as before the epilog */
7149 mono_emit_unwind_op_restore_state (cfg, code);
7151 cfg->code_len = code - cfg->native_code;
7153 g_assert (cfg->code_len < cfg->code_size);
7157 mono_arch_emit_exceptions (MonoCompile *cfg)
7159 MonoJumpInfo *patch_info;
7162 MonoClass *exc_classes [16];
7163 guint8 *exc_throw_start [16], *exc_throw_end [16];
7164 guint32 code_size = 0;
7166 /* Compute needed space */
7167 for (patch_info = cfg->patch_info; patch_info; patch_info = patch_info->next) {
7168 if (patch_info->type == MONO_PATCH_INFO_EXC)
7170 if (patch_info->type == MONO_PATCH_INFO_R8)
7171 code_size += 8 + 15; /* sizeof (double) + alignment */
7172 if (patch_info->type == MONO_PATCH_INFO_R4)
7173 code_size += 4 + 15; /* sizeof (float) + alignment */
7174 if (patch_info->type == MONO_PATCH_INFO_GC_CARD_TABLE_ADDR)
7175 code_size += 8 + 7; /*sizeof (void*) + alignment */
7178 while (cfg->code_len + code_size > (cfg->code_size - 16)) {
7179 cfg->code_size *= 2;
7180 cfg->native_code = (unsigned char *)mono_realloc_native_code (cfg);
7181 cfg->stat_code_reallocs++;
7184 code = cfg->native_code + cfg->code_len;
7186 /* add code to raise exceptions */
7188 for (patch_info = cfg->patch_info; patch_info; patch_info = patch_info->next) {
7189 switch (patch_info->type) {
7190 case MONO_PATCH_INFO_EXC: {
7191 MonoClass *exc_class;
7195 amd64_patch (patch_info->ip.i + cfg->native_code, code);
7197 exc_class = mono_class_load_from_name (mono_defaults.corlib, "System", patch_info->data.name);
7198 throw_ip = patch_info->ip.i;
7200 //x86_breakpoint (code);
7201 /* Find a throw sequence for the same exception class */
7202 for (i = 0; i < nthrows; ++i)
7203 if (exc_classes [i] == exc_class)
7206 amd64_mov_reg_imm (code, AMD64_ARG_REG2, (exc_throw_end [i] - cfg->native_code) - throw_ip);
7207 x86_jump_code (code, exc_throw_start [i]);
7208 patch_info->type = MONO_PATCH_INFO_NONE;
7212 amd64_mov_reg_imm_size (code, AMD64_ARG_REG2, 0xf0f0f0f0, 4);
7216 exc_classes [nthrows] = exc_class;
7217 exc_throw_start [nthrows] = code;
7219 amd64_mov_reg_imm (code, AMD64_ARG_REG1, exc_class->type_token - MONO_TOKEN_TYPE_DEF);
7221 patch_info->type = MONO_PATCH_INFO_NONE;
7223 code = emit_call_body (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD, "mono_arch_throw_corlib_exception");
7225 amd64_mov_reg_imm (buf, AMD64_ARG_REG2, (code - cfg->native_code) - throw_ip);
7230 exc_throw_end [nthrows] = code;
7240 g_assert(code < cfg->native_code + cfg->code_size);
7243 /* Handle relocations with RIP relative addressing */
7244 for (patch_info = cfg->patch_info; patch_info; patch_info = patch_info->next) {
7245 gboolean remove = FALSE;
7246 guint8 *orig_code = code;
7248 switch (patch_info->type) {
7249 case MONO_PATCH_INFO_R8:
7250 case MONO_PATCH_INFO_R4: {
7251 guint8 *pos, *patch_pos;
7254 /* The SSE opcodes require a 16 byte alignment */
7255 code = (guint8*)ALIGN_TO (code, 16);
7257 pos = cfg->native_code + patch_info->ip.i;
7258 if (IS_REX (pos [1])) {
7259 patch_pos = pos + 5;
7260 target_pos = code - pos - 9;
7263 patch_pos = pos + 4;
7264 target_pos = code - pos - 8;
7267 if (patch_info->type == MONO_PATCH_INFO_R8) {
7268 *(double*)code = *(double*)patch_info->data.target;
7269 code += sizeof (double);
7271 *(float*)code = *(float*)patch_info->data.target;
7272 code += sizeof (float);
7275 *(guint32*)(patch_pos) = target_pos;
7280 case MONO_PATCH_INFO_GC_CARD_TABLE_ADDR: {
7283 if (cfg->compile_aot)
7286 /*loading is faster against aligned addresses.*/
7287 code = (guint8*)ALIGN_TO (code, 8);
7288 memset (orig_code, 0, code - orig_code);
7290 pos = cfg->native_code + patch_info->ip.i;
7292 /*alu_op [rex] modr/m imm32 - 7 or 8 bytes */
7293 if (IS_REX (pos [1]))
7294 *(guint32*)(pos + 4) = (guint8*)code - pos - 8;
7296 *(guint32*)(pos + 3) = (guint8*)code - pos - 7;
7298 *(gpointer*)code = (gpointer)patch_info->data.target;
7299 code += sizeof (gpointer);
7309 if (patch_info == cfg->patch_info)
7310 cfg->patch_info = patch_info->next;
7314 for (tmp = cfg->patch_info; tmp->next != patch_info; tmp = tmp->next)
7316 tmp->next = patch_info->next;
7319 g_assert (code < cfg->native_code + cfg->code_size);
7322 cfg->code_len = code - cfg->native_code;
7324 g_assert (cfg->code_len < cfg->code_size);
7328 #endif /* DISABLE_JIT */
7331 mono_arch_instrument_prolog (MonoCompile *cfg, void *func, void *p, gboolean enable_arguments)
7333 guchar *code = (guchar *)p;
7334 MonoMethodSignature *sig;
7336 int i, n, stack_area = 0;
7338 /* Keep this in sync with mono_arch_get_argument_info */
7340 if (enable_arguments) {
7341 /* Allocate a new area on the stack and save arguments there */
7342 sig = mono_method_signature (cfg->method);
7344 n = sig->param_count + sig->hasthis;
7346 stack_area = ALIGN_TO (n * 8, 16);
7348 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, stack_area);
7350 for (i = 0; i < n; ++i) {
7351 inst = cfg->args [i];
7353 if (inst->opcode == OP_REGVAR)
7354 amd64_mov_membase_reg (code, AMD64_RSP, (i * 8), inst->dreg, 8);
7356 amd64_mov_reg_membase (code, AMD64_R11, inst->inst_basereg, inst->inst_offset, 8);
7357 amd64_mov_membase_reg (code, AMD64_RSP, (i * 8), AMD64_R11, 8);
7362 mono_add_patch_info (cfg, code-cfg->native_code, MONO_PATCH_INFO_METHODCONST, cfg->method);
7363 amd64_set_reg_template (code, AMD64_ARG_REG1);
7364 amd64_mov_reg_reg (code, AMD64_ARG_REG2, AMD64_RSP, 8);
7365 code = emit_call (cfg, code, MONO_PATCH_INFO_ABS, (gpointer)func, TRUE);
7367 if (enable_arguments)
7368 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, stack_area);
7382 mono_arch_instrument_epilog_full (MonoCompile *cfg, void *func, void *p, gboolean enable_arguments, gboolean preserve_argument_registers)
7384 guchar *code = (guchar *)p;
7385 int save_mode = SAVE_NONE;
7386 MonoMethod *method = cfg->method;
7387 MonoType *ret_type = mini_get_underlying_type (mono_method_signature (method)->ret);
7390 switch (ret_type->type) {
7391 case MONO_TYPE_VOID:
7392 /* special case string .ctor icall */
7393 if (strcmp (".ctor", method->name) && method->klass == mono_defaults.string_class)
7394 save_mode = SAVE_EAX;
7396 save_mode = SAVE_NONE;
7400 save_mode = SAVE_EAX;
7404 save_mode = SAVE_XMM;
7406 case MONO_TYPE_GENERICINST:
7407 if (!mono_type_generic_inst_is_valuetype (ret_type)) {
7408 save_mode = SAVE_EAX;
7412 case MONO_TYPE_VALUETYPE:
7413 save_mode = SAVE_STRUCT;
7416 save_mode = SAVE_EAX;
7420 /* Save the result and copy it into the proper argument register */
7421 switch (save_mode) {
7423 amd64_push_reg (code, AMD64_RAX);
7425 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, 8);
7426 if (enable_arguments)
7427 amd64_mov_reg_reg (code, AMD64_ARG_REG2, AMD64_RAX, 8);
7431 if (enable_arguments)
7432 amd64_mov_reg_imm (code, AMD64_ARG_REG2, 0);
7435 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, 8);
7436 amd64_movsd_membase_reg (code, AMD64_RSP, 0, AMD64_XMM0);
7438 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, 8);
7440 * The result is already in the proper argument register so no copying
7447 g_assert_not_reached ();
7450 /* Set %al since this is a varargs call */
7451 if (save_mode == SAVE_XMM)
7452 amd64_mov_reg_imm (code, AMD64_RAX, 1);
7454 amd64_mov_reg_imm (code, AMD64_RAX, 0);
7456 if (preserve_argument_registers) {
7457 for (i = 0; i < PARAM_REGS; ++i)
7458 amd64_push_reg (code, param_regs [i]);
7461 mono_add_patch_info (cfg, code-cfg->native_code, MONO_PATCH_INFO_METHODCONST, method);
7462 amd64_set_reg_template (code, AMD64_ARG_REG1);
7463 code = emit_call (cfg, code, MONO_PATCH_INFO_ABS, (gpointer)func, TRUE);
7465 if (preserve_argument_registers) {
7466 for (i = PARAM_REGS - 1; i >= 0; --i)
7467 amd64_pop_reg (code, param_regs [i]);
7470 /* Restore result */
7471 switch (save_mode) {
7473 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, 8);
7474 amd64_pop_reg (code, AMD64_RAX);
7480 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, 8);
7481 amd64_movsd_reg_membase (code, AMD64_XMM0, AMD64_RSP, 0);
7482 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, 8);
7487 g_assert_not_reached ();
7494 mono_arch_flush_icache (guint8 *code, gint size)
7500 mono_arch_flush_register_windows (void)
7505 mono_arch_is_inst_imm (gint64 imm)
7507 return amd64_use_imm32 (imm);
7511 * Determine whenever the trap whose info is in SIGINFO is caused by
7515 mono_arch_is_int_overflow (void *sigctx, void *info)
7522 mono_sigctx_to_monoctx (sigctx, &ctx);
7524 rip = (guint8*)ctx.gregs [AMD64_RIP];
7526 if (IS_REX (rip [0])) {
7527 reg = amd64_rex_b (rip [0]);
7533 if ((rip [0] == 0xf7) && (x86_modrm_mod (rip [1]) == 0x3) && (x86_modrm_reg (rip [1]) == 0x7)) {
7535 reg += x86_modrm_rm (rip [1]);
7537 value = ctx.gregs [reg];
7547 mono_arch_get_patch_offset (guint8 *code)
7553 * mono_breakpoint_clean_code:
7555 * Copy @size bytes from @code - @offset to the buffer @buf. If the debugger inserted software
7556 * breakpoints in the original code, they are removed in the copy.
7558 * Returns TRUE if no sw breakpoint was present.
7561 mono_breakpoint_clean_code (guint8 *method_start, guint8 *code, int offset, guint8 *buf, int size)
7564 * If method_start is non-NULL we need to perform bound checks, since we access memory
7565 * at code - offset we could go before the start of the method and end up in a different
7566 * page of memory that is not mapped or read incorrect data anyway. We zero-fill the bytes
7569 if (!method_start || code - offset >= method_start) {
7570 memcpy (buf, code - offset, size);
7572 int diff = code - method_start;
7573 memset (buf, 0, size);
7574 memcpy (buf + offset - diff, method_start, diff + size - offset);
7580 mono_arch_get_this_arg_reg (guint8 *code)
7582 return AMD64_ARG_REG1;
7586 mono_arch_get_this_arg_from_call (mgreg_t *regs, guint8 *code)
7588 return (gpointer)regs [mono_arch_get_this_arg_reg (code)];
7591 #define MAX_ARCH_DELEGATE_PARAMS 10
7594 get_delegate_invoke_impl (MonoTrampInfo **info, gboolean has_target, guint32 param_count)
7596 guint8 *code, *start;
7597 GSList *unwind_ops = NULL;
7600 unwind_ops = mono_arch_get_cie_program ();
7603 start = code = (guint8 *)mono_global_codeman_reserve (64);
7605 /* Replace the this argument with the target */
7606 amd64_mov_reg_reg (code, AMD64_RAX, AMD64_ARG_REG1, 8);
7607 amd64_mov_reg_membase (code, AMD64_ARG_REG1, AMD64_RAX, MONO_STRUCT_OFFSET (MonoDelegate, target), 8);
7608 amd64_jump_membase (code, AMD64_RAX, MONO_STRUCT_OFFSET (MonoDelegate, method_ptr));
7610 g_assert ((code - start) < 64);
7612 start = code = (guint8 *)mono_global_codeman_reserve (64);
7614 if (param_count == 0) {
7615 amd64_jump_membase (code, AMD64_ARG_REG1, MONO_STRUCT_OFFSET (MonoDelegate, method_ptr));
7617 /* We have to shift the arguments left */
7618 amd64_mov_reg_reg (code, AMD64_RAX, AMD64_ARG_REG1, 8);
7619 for (i = 0; i < param_count; ++i) {
7622 amd64_mov_reg_reg (code, param_regs [i], param_regs [i + 1], 8);
7624 amd64_mov_reg_membase (code, param_regs [i], AMD64_RSP, 0x28, 8);
7626 amd64_mov_reg_reg (code, param_regs [i], param_regs [i + 1], 8);
7630 amd64_jump_membase (code, AMD64_RAX, MONO_STRUCT_OFFSET (MonoDelegate, method_ptr));
7632 g_assert ((code - start) < 64);
7635 mono_arch_flush_icache (start, code - start);
7638 *info = mono_tramp_info_create ("delegate_invoke_impl_has_target", start, code - start, NULL, unwind_ops);
7640 char *name = g_strdup_printf ("delegate_invoke_impl_target_%d", param_count);
7641 *info = mono_tramp_info_create (name, start, code - start, NULL, unwind_ops);
7645 if (mono_jit_map_is_enabled ()) {
7648 buff = (char*)"delegate_invoke_has_target";
7650 buff = g_strdup_printf ("delegate_invoke_no_target_%d", param_count);
7651 mono_emit_jit_tramp (start, code - start, buff);
7655 mono_profiler_code_buffer_new (start, code - start, MONO_PROFILER_CODE_BUFFER_DELEGATE_INVOKE, NULL);
7660 #define MAX_VIRTUAL_DELEGATE_OFFSET 32
7663 get_delegate_virtual_invoke_impl (MonoTrampInfo **info, gboolean load_imt_reg, int offset)
7665 guint8 *code, *start;
7670 if (offset / (int)sizeof (gpointer) > MAX_VIRTUAL_DELEGATE_OFFSET)
7673 start = code = (guint8 *)mono_global_codeman_reserve (size);
7675 unwind_ops = mono_arch_get_cie_program ();
7677 /* Replace the this argument with the target */
7678 amd64_mov_reg_reg (code, AMD64_RAX, AMD64_ARG_REG1, 8);
7679 amd64_mov_reg_membase (code, AMD64_ARG_REG1, AMD64_RAX, MONO_STRUCT_OFFSET (MonoDelegate, target), 8);
7682 /* Load the IMT reg */
7683 amd64_mov_reg_membase (code, MONO_ARCH_IMT_REG, AMD64_RAX, MONO_STRUCT_OFFSET (MonoDelegate, method), 8);
7686 /* Load the vtable */
7687 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_ARG_REG1, MONO_STRUCT_OFFSET (MonoObject, vtable), 8);
7688 amd64_jump_membase (code, AMD64_RAX, offset);
7689 mono_profiler_code_buffer_new (start, code - start, MONO_PROFILER_CODE_BUFFER_DELEGATE_INVOKE, NULL);
7691 tramp_name = mono_get_delegate_virtual_invoke_impl_name (load_imt_reg, offset);
7692 *info = mono_tramp_info_create (tramp_name, start, code - start, NULL, unwind_ops);
7693 g_free (tramp_name);
7699 * mono_arch_get_delegate_invoke_impls:
7701 * Return a list of MonoTrampInfo structures for the delegate invoke impl
7705 mono_arch_get_delegate_invoke_impls (void)
7708 MonoTrampInfo *info;
7711 get_delegate_invoke_impl (&info, TRUE, 0);
7712 res = g_slist_prepend (res, info);
7714 for (i = 0; i <= MAX_ARCH_DELEGATE_PARAMS; ++i) {
7715 get_delegate_invoke_impl (&info, FALSE, i);
7716 res = g_slist_prepend (res, info);
7719 for (i = 1; i <= MONO_IMT_SIZE; ++i) {
7720 get_delegate_virtual_invoke_impl (&info, TRUE, - i * SIZEOF_VOID_P);
7721 res = g_slist_prepend (res, info);
7724 for (i = 0; i <= MAX_VIRTUAL_DELEGATE_OFFSET; ++i) {
7725 get_delegate_virtual_invoke_impl (&info, FALSE, i * SIZEOF_VOID_P);
7726 res = g_slist_prepend (res, info);
7727 get_delegate_virtual_invoke_impl (&info, TRUE, i * SIZEOF_VOID_P);
7728 res = g_slist_prepend (res, info);
7735 mono_arch_get_delegate_invoke_impl (MonoMethodSignature *sig, gboolean has_target)
7737 guint8 *code, *start;
7740 if (sig->param_count > MAX_ARCH_DELEGATE_PARAMS)
7743 /* FIXME: Support more cases */
7744 if (MONO_TYPE_ISSTRUCT (mini_get_underlying_type (sig->ret)))
7748 static guint8* cached = NULL;
7753 if (mono_aot_only) {
7754 start = (guint8 *)mono_aot_get_trampoline ("delegate_invoke_impl_has_target");
7756 MonoTrampInfo *info;
7757 start = (guint8 *)get_delegate_invoke_impl (&info, TRUE, 0);
7758 mono_tramp_info_register (info, NULL);
7761 mono_memory_barrier ();
7765 static guint8* cache [MAX_ARCH_DELEGATE_PARAMS + 1] = {NULL};
7766 for (i = 0; i < sig->param_count; ++i)
7767 if (!mono_is_regsize_var (sig->params [i]))
7769 if (sig->param_count > 4)
7772 code = cache [sig->param_count];
7776 if (mono_aot_only) {
7777 char *name = g_strdup_printf ("delegate_invoke_impl_target_%d", sig->param_count);
7778 start = (guint8 *)mono_aot_get_trampoline (name);
7781 MonoTrampInfo *info;
7782 start = (guint8 *)get_delegate_invoke_impl (&info, FALSE, sig->param_count);
7783 mono_tramp_info_register (info, NULL);
7786 mono_memory_barrier ();
7788 cache [sig->param_count] = start;
7795 mono_arch_get_delegate_virtual_invoke_impl (MonoMethodSignature *sig, MonoMethod *method, int offset, gboolean load_imt_reg)
7797 MonoTrampInfo *info;
7800 code = get_delegate_virtual_invoke_impl (&info, load_imt_reg, offset);
7802 mono_tramp_info_register (info, NULL);
7807 mono_arch_finish_init (void)
7809 #if !defined(HOST_WIN32) && defined(MONO_XEN_OPT)
7810 optimize_for_xen = access ("/proc/xen", F_OK) == 0;
7815 mono_arch_free_jit_tls_data (MonoJitTlsData *tls)
7819 #define CMP_SIZE (6 + 1)
7820 #define CMP_REG_REG_SIZE (4 + 1)
7821 #define BR_SMALL_SIZE 2
7822 #define BR_LARGE_SIZE 6
7823 #define MOV_REG_IMM_SIZE 10
7824 #define MOV_REG_IMM_32BIT_SIZE 6
7825 #define JUMP_REG_SIZE (2 + 1)
7828 imt_branch_distance (MonoIMTCheckItem **imt_entries, int start, int target)
7830 int i, distance = 0;
7831 for (i = start; i < target; ++i)
7832 distance += imt_entries [i]->chunk_size;
7837 * LOCKING: called with the domain lock held
7840 mono_arch_build_imt_trampoline (MonoVTable *vtable, MonoDomain *domain, MonoIMTCheckItem **imt_entries, int count,
7841 gpointer fail_tramp)
7845 guint8 *code, *start;
7846 gboolean vtable_is_32bit = ((gsize)(vtable) == (gsize)(int)(gsize)(vtable));
7849 for (i = 0; i < count; ++i) {
7850 MonoIMTCheckItem *item = imt_entries [i];
7851 if (item->is_equals) {
7852 if (item->check_target_idx) {
7853 if (!item->compare_done) {
7854 if (amd64_use_imm32 ((gint64)item->key))
7855 item->chunk_size += CMP_SIZE;
7857 item->chunk_size += MOV_REG_IMM_SIZE + CMP_REG_REG_SIZE;
7859 if (item->has_target_code) {
7860 item->chunk_size += MOV_REG_IMM_SIZE;
7862 if (vtable_is_32bit)
7863 item->chunk_size += MOV_REG_IMM_32BIT_SIZE;
7865 item->chunk_size += MOV_REG_IMM_SIZE;
7867 item->chunk_size += BR_SMALL_SIZE + JUMP_REG_SIZE;
7870 item->chunk_size += MOV_REG_IMM_SIZE * 3 + CMP_REG_REG_SIZE +
7871 BR_SMALL_SIZE + JUMP_REG_SIZE * 2;
7873 if (vtable_is_32bit)
7874 item->chunk_size += MOV_REG_IMM_32BIT_SIZE;
7876 item->chunk_size += MOV_REG_IMM_SIZE;
7877 item->chunk_size += JUMP_REG_SIZE;
7878 /* with assert below:
7879 * item->chunk_size += CMP_SIZE + BR_SMALL_SIZE + 1;
7884 if (amd64_use_imm32 ((gint64)item->key))
7885 item->chunk_size += CMP_SIZE;
7887 item->chunk_size += MOV_REG_IMM_SIZE + CMP_REG_REG_SIZE;
7888 item->chunk_size += BR_LARGE_SIZE;
7889 imt_entries [item->check_target_idx]->compare_done = TRUE;
7891 size += item->chunk_size;
7894 code = (guint8 *)mono_method_alloc_generic_virtual_trampoline (domain, size);
7896 code = (guint8 *)mono_domain_code_reserve (domain, size);
7899 unwind_ops = mono_arch_get_cie_program ();
7901 for (i = 0; i < count; ++i) {
7902 MonoIMTCheckItem *item = imt_entries [i];
7903 item->code_target = code;
7904 if (item->is_equals) {
7905 gboolean fail_case = !item->check_target_idx && fail_tramp;
7907 if (item->check_target_idx || fail_case) {
7908 if (!item->compare_done || fail_case) {
7909 if (amd64_use_imm32 ((gint64)item->key))
7910 amd64_alu_reg_imm_size (code, X86_CMP, MONO_ARCH_IMT_REG, (guint32)(gssize)item->key, sizeof(gpointer));
7912 amd64_mov_reg_imm_size (code, MONO_ARCH_IMT_SCRATCH_REG, item->key, sizeof(gpointer));
7913 amd64_alu_reg_reg (code, X86_CMP, MONO_ARCH_IMT_REG, MONO_ARCH_IMT_SCRATCH_REG);
7916 item->jmp_code = code;
7917 amd64_branch8 (code, X86_CC_NE, 0, FALSE);
7918 if (item->has_target_code) {
7919 amd64_mov_reg_imm (code, MONO_ARCH_IMT_SCRATCH_REG, item->value.target_code);
7920 amd64_jump_reg (code, MONO_ARCH_IMT_SCRATCH_REG);
7922 amd64_mov_reg_imm (code, MONO_ARCH_IMT_SCRATCH_REG, & (vtable->vtable [item->value.vtable_slot]));
7923 amd64_jump_membase (code, MONO_ARCH_IMT_SCRATCH_REG, 0);
7927 amd64_patch (item->jmp_code, code);
7928 amd64_mov_reg_imm (code, MONO_ARCH_IMT_SCRATCH_REG, fail_tramp);
7929 amd64_jump_reg (code, MONO_ARCH_IMT_SCRATCH_REG);
7930 item->jmp_code = NULL;
7933 /* enable the commented code to assert on wrong method */
7935 if (amd64_is_imm32 (item->key))
7936 amd64_alu_reg_imm_size (code, X86_CMP, MONO_ARCH_IMT_REG, (guint32)(gssize)item->key, sizeof(gpointer));
7938 amd64_mov_reg_imm (code, MONO_ARCH_IMT_SCRATCH_REG, item->key);
7939 amd64_alu_reg_reg (code, X86_CMP, MONO_ARCH_IMT_REG, MONO_ARCH_IMT_SCRATCH_REG);
7941 item->jmp_code = code;
7942 amd64_branch8 (code, X86_CC_NE, 0, FALSE);
7943 /* See the comment below about R10 */
7944 amd64_mov_reg_imm (code, MONO_ARCH_IMT_SCRATCH_REG, & (vtable->vtable [item->value.vtable_slot]));
7945 amd64_jump_membase (code, MONO_ARCH_IMT_SCRATCH_REG, 0);
7946 amd64_patch (item->jmp_code, code);
7947 amd64_breakpoint (code);
7948 item->jmp_code = NULL;
7950 /* We're using R10 (MONO_ARCH_IMT_SCRATCH_REG) here because R11 (MONO_ARCH_IMT_REG)
7951 needs to be preserved. R10 needs
7952 to be preserved for calls which
7953 require a runtime generic context,
7954 but interface calls don't. */
7955 amd64_mov_reg_imm (code, MONO_ARCH_IMT_SCRATCH_REG, & (vtable->vtable [item->value.vtable_slot]));
7956 amd64_jump_membase (code, MONO_ARCH_IMT_SCRATCH_REG, 0);
7960 if (amd64_use_imm32 ((gint64)item->key))
7961 amd64_alu_reg_imm_size (code, X86_CMP, MONO_ARCH_IMT_REG, (guint32)(gssize)item->key, sizeof (gpointer));
7963 amd64_mov_reg_imm_size (code, MONO_ARCH_IMT_SCRATCH_REG, item->key, sizeof (gpointer));
7964 amd64_alu_reg_reg (code, X86_CMP, MONO_ARCH_IMT_REG, MONO_ARCH_IMT_SCRATCH_REG);
7966 item->jmp_code = code;
7967 if (x86_is_imm8 (imt_branch_distance (imt_entries, i, item->check_target_idx)))
7968 x86_branch8 (code, X86_CC_GE, 0, FALSE);
7970 x86_branch32 (code, X86_CC_GE, 0, FALSE);
7972 g_assert (code - item->code_target <= item->chunk_size);
7974 /* patch the branches to get to the target items */
7975 for (i = 0; i < count; ++i) {
7976 MonoIMTCheckItem *item = imt_entries [i];
7977 if (item->jmp_code) {
7978 if (item->check_target_idx) {
7979 amd64_patch (item->jmp_code, imt_entries [item->check_target_idx]->code_target);
7985 mono_stats.imt_trampolines_size += code - start;
7986 g_assert (code - start <= size);
7988 mono_profiler_code_buffer_new (start, code - start, MONO_PROFILER_CODE_BUFFER_IMT_TRAMPOLINE, NULL);
7990 mono_tramp_info_register (mono_tramp_info_create (NULL, start, code - start, NULL, unwind_ops), domain);
7996 mono_arch_find_imt_method (mgreg_t *regs, guint8 *code)
7998 return (MonoMethod*)regs [MONO_ARCH_IMT_REG];
8002 mono_arch_find_static_call_vtable (mgreg_t *regs, guint8 *code)
8004 return (MonoVTable*) regs [MONO_ARCH_RGCTX_REG];
8008 mono_arch_get_cie_program (void)
8012 mono_add_unwind_op_def_cfa (l, (guint8*)NULL, (guint8*)NULL, AMD64_RSP, 8);
8013 mono_add_unwind_op_offset (l, (guint8*)NULL, (guint8*)NULL, AMD64_RIP, -8);
8021 mono_arch_emit_inst_for_method (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature *fsig, MonoInst **args)
8023 MonoInst *ins = NULL;
8026 if (cmethod->klass == mono_defaults.math_class) {
8027 if (strcmp (cmethod->name, "Sin") == 0) {
8029 } else if (strcmp (cmethod->name, "Cos") == 0) {
8031 } else if (strcmp (cmethod->name, "Sqrt") == 0) {
8033 } else if (strcmp (cmethod->name, "Abs") == 0 && fsig->params [0]->type == MONO_TYPE_R8) {
8037 if (opcode && fsig->param_count == 1) {
8038 MONO_INST_NEW (cfg, ins, opcode);
8039 ins->type = STACK_R8;
8040 ins->dreg = mono_alloc_freg (cfg);
8041 ins->sreg1 = args [0]->dreg;
8042 MONO_ADD_INS (cfg->cbb, ins);
8046 if (cfg->opt & MONO_OPT_CMOV) {
8047 if (strcmp (cmethod->name, "Min") == 0) {
8048 if (fsig->params [0]->type == MONO_TYPE_I4)
8050 if (fsig->params [0]->type == MONO_TYPE_U4)
8051 opcode = OP_IMIN_UN;
8052 else if (fsig->params [0]->type == MONO_TYPE_I8)
8054 else if (fsig->params [0]->type == MONO_TYPE_U8)
8055 opcode = OP_LMIN_UN;
8056 } else if (strcmp (cmethod->name, "Max") == 0) {
8057 if (fsig->params [0]->type == MONO_TYPE_I4)
8059 if (fsig->params [0]->type == MONO_TYPE_U4)
8060 opcode = OP_IMAX_UN;
8061 else if (fsig->params [0]->type == MONO_TYPE_I8)
8063 else if (fsig->params [0]->type == MONO_TYPE_U8)
8064 opcode = OP_LMAX_UN;
8068 if (opcode && fsig->param_count == 2) {
8069 MONO_INST_NEW (cfg, ins, opcode);
8070 ins->type = fsig->params [0]->type == MONO_TYPE_I4 ? STACK_I4 : STACK_I8;
8071 ins->dreg = mono_alloc_ireg (cfg);
8072 ins->sreg1 = args [0]->dreg;
8073 ins->sreg2 = args [1]->dreg;
8074 MONO_ADD_INS (cfg->cbb, ins);
8078 /* OP_FREM is not IEEE compatible */
8079 else if (strcmp (cmethod->name, "IEEERemainder") == 0 && fsig->param_count == 2) {
8080 MONO_INST_NEW (cfg, ins, OP_FREM);
8081 ins->inst_i0 = args [0];
8082 ins->inst_i1 = args [1];
8092 mono_arch_print_tree (MonoInst *tree, int arity)
8098 mono_arch_context_get_int_reg (MonoContext *ctx, int reg)
8100 return ctx->gregs [reg];
8104 mono_arch_context_set_int_reg (MonoContext *ctx, int reg, mgreg_t val)
8106 ctx->gregs [reg] = val;
8110 mono_arch_install_handler_block_guard (MonoJitInfo *ji, MonoJitExceptionInfo *clause, MonoContext *ctx, gpointer new_value)
8112 gpointer *sp, old_value;
8116 bp = (char *)MONO_CONTEXT_GET_BP (ctx);
8117 sp = (gpointer *)*(gpointer*)(bp + clause->exvar_offset);
8120 if (old_value < ji->code_start || (char*)old_value > ((char*)ji->code_start + ji->code_size))
8129 * mono_arch_emit_load_aotconst:
8131 * Emit code to load the contents of the GOT slot identified by TRAMP_TYPE and
8132 * TARGET from the mscorlib GOT in full-aot code.
8133 * On AMD64, the result is placed into R11.
8136 mono_arch_emit_load_aotconst (guint8 *start, guint8 *code, MonoJumpInfo **ji, MonoJumpInfoType tramp_type, gconstpointer target)
8138 *ji = mono_patch_info_list_prepend (*ji, code - start, tramp_type, target);
8139 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RIP, 0, 8);
8145 * mono_arch_get_trampolines:
8147 * Return a list of MonoTrampInfo structures describing arch specific trampolines
8151 mono_arch_get_trampolines (gboolean aot)
8153 return mono_amd64_get_exception_trampolines (aot);
8156 /* Soft Debug support */
8157 #ifdef MONO_ARCH_SOFT_DEBUG_SUPPORTED
8160 * mono_arch_set_breakpoint:
8162 * Set a breakpoint at the native code corresponding to JI at NATIVE_OFFSET.
8163 * The location should contain code emitted by OP_SEQ_POINT.
8166 mono_arch_set_breakpoint (MonoJitInfo *ji, guint8 *ip)
8171 guint32 native_offset = ip - (guint8*)ji->code_start;
8172 SeqPointInfo *info = (SeqPointInfo *)mono_arch_get_seq_point_info (mono_domain_get (), (guint8 *)ji->code_start);
8174 g_assert (info->bp_addrs [native_offset] == 0);
8175 info->bp_addrs [native_offset] = mini_get_breakpoint_trampoline ();
8177 /* ip points to a mov r11, 0 */
8178 g_assert (code [0] == 0x41);
8179 g_assert (code [1] == 0xbb);
8180 amd64_mov_reg_imm (code, AMD64_R11, 1);
8185 * mono_arch_clear_breakpoint:
8187 * Clear the breakpoint at IP.
8190 mono_arch_clear_breakpoint (MonoJitInfo *ji, guint8 *ip)
8195 guint32 native_offset = ip - (guint8*)ji->code_start;
8196 SeqPointInfo *info = (SeqPointInfo *)mono_arch_get_seq_point_info (mono_domain_get (), (guint8 *)ji->code_start);
8198 info->bp_addrs [native_offset] = NULL;
8200 amd64_mov_reg_imm (code, AMD64_R11, 0);
8205 mono_arch_is_breakpoint_event (void *info, void *sigctx)
8207 /* We use soft breakpoints on amd64 */
8212 * mono_arch_skip_breakpoint:
8214 * Modify CTX so the ip is placed after the breakpoint instruction, so when
8215 * we resume, the instruction is not executed again.
8218 mono_arch_skip_breakpoint (MonoContext *ctx, MonoJitInfo *ji)
8220 g_assert_not_reached ();
8224 * mono_arch_start_single_stepping:
8226 * Start single stepping.
8229 mono_arch_start_single_stepping (void)
8231 ss_trampoline = mini_get_single_step_trampoline ();
8235 * mono_arch_stop_single_stepping:
8237 * Stop single stepping.
8240 mono_arch_stop_single_stepping (void)
8242 ss_trampoline = NULL;
8246 * mono_arch_is_single_step_event:
8248 * Return whenever the machine state in SIGCTX corresponds to a single
8252 mono_arch_is_single_step_event (void *info, void *sigctx)
8254 /* We use soft breakpoints on amd64 */
8259 * mono_arch_skip_single_step:
8261 * Modify CTX so the ip is placed after the single step trigger instruction,
8262 * we resume, the instruction is not executed again.
8265 mono_arch_skip_single_step (MonoContext *ctx)
8267 g_assert_not_reached ();
8271 * mono_arch_create_seq_point_info:
8273 * Return a pointer to a data structure which is used by the sequence
8274 * point implementation in AOTed code.
8277 mono_arch_get_seq_point_info (MonoDomain *domain, guint8 *code)
8282 // FIXME: Add a free function
8284 mono_domain_lock (domain);
8285 info = (SeqPointInfo *)g_hash_table_lookup (domain_jit_info (domain)->arch_seq_points,
8287 mono_domain_unlock (domain);
8290 ji = mono_jit_info_table_find (domain, (char*)code);
8293 // FIXME: Optimize the size
8294 info = (SeqPointInfo *)g_malloc0 (sizeof (SeqPointInfo) + (ji->code_size * sizeof (gpointer)));
8296 info->ss_tramp_addr = &ss_trampoline;
8298 mono_domain_lock (domain);
8299 g_hash_table_insert (domain_jit_info (domain)->arch_seq_points,
8301 mono_domain_unlock (domain);
8308 mono_arch_init_lmf_ext (MonoLMFExt *ext, gpointer prev_lmf)
8310 ext->lmf.previous_lmf = prev_lmf;
8311 /* Mark that this is a MonoLMFExt */
8312 ext->lmf.previous_lmf = (gpointer)(((gssize)ext->lmf.previous_lmf) | 2);
8313 ext->lmf.rsp = (gssize)ext;
8319 mono_arch_opcode_supported (int opcode)
8322 case OP_ATOMIC_ADD_I4:
8323 case OP_ATOMIC_ADD_I8:
8324 case OP_ATOMIC_EXCHANGE_I4:
8325 case OP_ATOMIC_EXCHANGE_I8:
8326 case OP_ATOMIC_CAS_I4:
8327 case OP_ATOMIC_CAS_I8:
8328 case OP_ATOMIC_LOAD_I1:
8329 case OP_ATOMIC_LOAD_I2:
8330 case OP_ATOMIC_LOAD_I4:
8331 case OP_ATOMIC_LOAD_I8:
8332 case OP_ATOMIC_LOAD_U1:
8333 case OP_ATOMIC_LOAD_U2:
8334 case OP_ATOMIC_LOAD_U4:
8335 case OP_ATOMIC_LOAD_U8:
8336 case OP_ATOMIC_LOAD_R4:
8337 case OP_ATOMIC_LOAD_R8:
8338 case OP_ATOMIC_STORE_I1:
8339 case OP_ATOMIC_STORE_I2:
8340 case OP_ATOMIC_STORE_I4:
8341 case OP_ATOMIC_STORE_I8:
8342 case OP_ATOMIC_STORE_U1:
8343 case OP_ATOMIC_STORE_U2:
8344 case OP_ATOMIC_STORE_U4:
8345 case OP_ATOMIC_STORE_U8:
8346 case OP_ATOMIC_STORE_R4:
8347 case OP_ATOMIC_STORE_R8:
8355 mono_arch_get_call_info (MonoMemPool *mp, MonoMethodSignature *sig)
8357 return get_call_info (mp, sig);