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)
12 * (C) 2003 Ximian, Inc.
13 * Copyright 2003-2011 Novell, Inc (http://www.novell.com)
14 * Copyright 2011 Xamarin, Inc (http://www.xamarin.com)
23 #include <mono/metadata/abi-details.h>
24 #include <mono/metadata/appdomain.h>
25 #include <mono/metadata/debug-helpers.h>
26 #include <mono/metadata/threads.h>
27 #include <mono/metadata/profiler-private.h>
28 #include <mono/metadata/mono-debug.h>
29 #include <mono/metadata/gc-internals.h>
30 #include <mono/utils/mono-math.h>
31 #include <mono/utils/mono-mmap.h>
32 #include <mono/utils/mono-memory-model.h>
33 #include <mono/utils/mono-tls.h>
34 #include <mono/utils/mono-hwcap-x86.h>
35 #include <mono/utils/mono-threads.h>
39 #include "mini-amd64.h"
40 #include "cpu-amd64.h"
41 #include "debugger-agent.h"
45 static gboolean optimize_for_xen = TRUE;
47 #define optimize_for_xen 0
50 #define ALIGN_TO(val,align) ((((guint64)val) + ((align) - 1)) & ~((align) - 1))
52 #define IS_IMM32(val) ((((guint64)val) >> 32) == 0)
54 #define IS_REX(inst) (((inst) >= 0x40) && ((inst) <= 0x4f))
57 /* Under windows, the calling convention is never stdcall */
58 #define CALLCONV_IS_STDCALL(call_conv) (FALSE)
60 #define CALLCONV_IS_STDCALL(call_conv) ((call_conv) == MONO_CALL_STDCALL)
63 /* This mutex protects architecture specific caches */
64 #define mono_mini_arch_lock() mono_mutex_lock (&mini_arch_mutex)
65 #define mono_mini_arch_unlock() mono_mutex_unlock (&mini_arch_mutex)
66 static mono_mutex_t mini_arch_mutex;
68 /* The single step trampoline */
69 static gpointer ss_trampoline;
71 /* The breakpoint trampoline */
72 static gpointer bp_trampoline;
74 /* Offset between fp and the first argument in the callee */
75 #define ARGS_OFFSET 16
76 #define GP_SCRATCH_REG AMD64_R11
79 * AMD64 register usage:
80 * - callee saved registers are used for global register allocation
81 * - %r11 is used for materializing 64 bit constants in opcodes
82 * - the rest is used for local allocation
86 * Floating point comparison results:
96 mono_arch_regname (int reg)
99 case AMD64_RAX: return "%rax";
100 case AMD64_RBX: return "%rbx";
101 case AMD64_RCX: return "%rcx";
102 case AMD64_RDX: return "%rdx";
103 case AMD64_RSP: return "%rsp";
104 case AMD64_RBP: return "%rbp";
105 case AMD64_RDI: return "%rdi";
106 case AMD64_RSI: return "%rsi";
107 case AMD64_R8: return "%r8";
108 case AMD64_R9: return "%r9";
109 case AMD64_R10: return "%r10";
110 case AMD64_R11: return "%r11";
111 case AMD64_R12: return "%r12";
112 case AMD64_R13: return "%r13";
113 case AMD64_R14: return "%r14";
114 case AMD64_R15: return "%r15";
119 static const char * packed_xmmregs [] = {
120 "p:xmm0", "p:xmm1", "p:xmm2", "p:xmm3", "p:xmm4", "p:xmm5", "p:xmm6", "p:xmm7", "p:xmm8",
121 "p:xmm9", "p:xmm10", "p:xmm11", "p:xmm12", "p:xmm13", "p:xmm14", "p:xmm15"
124 static const char * single_xmmregs [] = {
125 "s:xmm0", "s:xmm1", "s:xmm2", "s:xmm3", "s:xmm4", "s:xmm5", "s:xmm6", "s:xmm7", "s:xmm8",
126 "s:xmm9", "s:xmm10", "s:xmm11", "s:xmm12", "s:xmm13", "s:xmm14", "s:xmm15"
130 mono_arch_fregname (int reg)
132 if (reg < AMD64_XMM_NREG)
133 return single_xmmregs [reg];
139 mono_arch_xregname (int reg)
141 if (reg < AMD64_XMM_NREG)
142 return packed_xmmregs [reg];
151 return mono_debug_count ();
157 static inline gboolean
158 amd64_is_near_call (guint8 *code)
161 if ((code [0] >= 0x40) && (code [0] <= 0x4f))
164 return code [0] == 0xe8;
167 static inline gboolean
168 amd64_use_imm32 (gint64 val)
170 if (mini_get_debug_options()->single_imm_size)
173 return amd64_is_imm32 (val);
176 #ifdef __native_client_codegen__
178 /* Keep track of instruction "depth", that is, the level of sub-instruction */
179 /* for any given instruction. For instance, amd64_call_reg resolves to */
180 /* amd64_call_reg_internal, which uses amd64_alu_* macros, etc. */
181 /* We only want to force bundle alignment for the top level instruction, */
182 /* so NaCl pseudo-instructions can be implemented with sub instructions. */
183 static MonoNativeTlsKey nacl_instruction_depth;
185 static MonoNativeTlsKey nacl_rex_tag;
186 static MonoNativeTlsKey nacl_legacy_prefix_tag;
189 amd64_nacl_clear_legacy_prefix_tag ()
191 mono_native_tls_set_value (nacl_legacy_prefix_tag, NULL);
195 amd64_nacl_tag_legacy_prefix (guint8* code)
197 if (mono_native_tls_get_value (nacl_legacy_prefix_tag) == NULL)
198 mono_native_tls_set_value (nacl_legacy_prefix_tag, code);
202 amd64_nacl_tag_rex (guint8* code)
204 mono_native_tls_set_value (nacl_rex_tag, code);
208 amd64_nacl_get_legacy_prefix_tag ()
210 return (guint8*)mono_native_tls_get_value (nacl_legacy_prefix_tag);
214 amd64_nacl_get_rex_tag ()
216 return (guint8*)mono_native_tls_get_value (nacl_rex_tag);
219 /* Increment the instruction "depth" described above */
221 amd64_nacl_instruction_pre ()
223 intptr_t depth = (intptr_t) mono_native_tls_get_value (nacl_instruction_depth);
225 mono_native_tls_set_value (nacl_instruction_depth, (gpointer)depth);
228 /* amd64_nacl_instruction_post: Decrement instruction "depth", force bundle */
229 /* alignment if depth == 0 (top level instruction) */
230 /* IN: start, end pointers to instruction beginning and end */
231 /* OUT: start, end pointers to beginning and end after possible alignment */
232 /* GLOBALS: nacl_instruction_depth defined above */
234 amd64_nacl_instruction_post (guint8 **start, guint8 **end)
236 intptr_t depth = (intptr_t) mono_native_tls_get_value (nacl_instruction_depth);
238 mono_native_tls_set_value (nacl_instruction_depth, (void*)depth);
240 g_assert ( depth >= 0 );
242 uintptr_t space_in_block;
244 guint8 *prefix = amd64_nacl_get_legacy_prefix_tag ();
245 /* if legacy prefix is present, and if it was emitted before */
246 /* the start of the instruction sequence, adjust the start */
247 if (prefix != NULL && prefix < *start) {
248 g_assert (*start - prefix <= 3);/* only 3 are allowed */
251 space_in_block = kNaClAlignment - ((uintptr_t)(*start) & kNaClAlignmentMask);
252 instlen = (uintptr_t)(*end - *start);
253 /* Only check for instructions which are less than */
254 /* kNaClAlignment. The only instructions that should ever */
255 /* be that long are call sequences, which are already */
256 /* padded out to align the return to the next bundle. */
257 if (instlen > space_in_block && instlen < kNaClAlignment) {
258 const size_t MAX_NACL_INST_LENGTH = kNaClAlignment;
259 guint8 copy_of_instruction[MAX_NACL_INST_LENGTH];
260 const size_t length = (size_t)((*end)-(*start));
261 g_assert (length < MAX_NACL_INST_LENGTH);
263 memcpy (copy_of_instruction, *start, length);
264 *start = mono_arch_nacl_pad (*start, space_in_block);
265 memcpy (*start, copy_of_instruction, length);
266 *end = *start + length;
268 amd64_nacl_clear_legacy_prefix_tag ();
269 amd64_nacl_tag_rex (NULL);
273 /* amd64_nacl_membase_handler: ensure all access to memory of the form */
274 /* OFFSET(%rXX) is sandboxed. For allowable base registers %rip, %rbp, */
275 /* %rsp, and %r15, emit the membase as usual. For all other registers, */
276 /* make sure the upper 32-bits are cleared, and use that register in the */
277 /* index field of a new address of this form: OFFSET(%r15,%eXX,1) */
279 /* pointer to current instruction stream (in the */
280 /* middle of an instruction, after opcode is emitted) */
281 /* basereg/offset/dreg */
282 /* operands of normal membase address */
284 /* pointer to the end of the membase/memindex emit */
285 /* GLOBALS: nacl_rex_tag */
286 /* position in instruction stream that rex prefix was emitted */
287 /* nacl_legacy_prefix_tag */
288 /* (possibly NULL) position in instruction of legacy x86 prefix */
290 amd64_nacl_membase_handler (guint8** code, gint8 basereg, gint32 offset, gint8 dreg)
292 gint8 true_basereg = basereg;
294 /* Cache these values, they might change */
295 /* as new instructions are emitted below. */
296 guint8* rex_tag = amd64_nacl_get_rex_tag ();
297 guint8* legacy_prefix_tag = amd64_nacl_get_legacy_prefix_tag ();
299 /* 'basereg' is given masked to 0x7 at this point, so check */
300 /* the rex prefix to see if this is an extended register. */
301 if ((rex_tag != NULL) && IS_REX(*rex_tag) && (*rex_tag & AMD64_REX_B)) {
305 #define X86_LEA_OPCODE (0x8D)
307 if (!amd64_is_valid_nacl_base (true_basereg) && (*(*code-1) != X86_LEA_OPCODE)) {
308 guint8* old_instruction_start;
310 /* This will hold the 'mov %eXX, %eXX' that clears the upper */
311 /* 32-bits of the old base register (new index register) */
313 guint8* buf_ptr = buf;
316 g_assert (rex_tag != NULL);
318 if (IS_REX(*rex_tag)) {
319 /* The old rex.B should be the new rex.X */
320 if (*rex_tag & AMD64_REX_B) {
321 *rex_tag |= AMD64_REX_X;
323 /* Since our new base is %r15 set rex.B */
324 *rex_tag |= AMD64_REX_B;
326 /* Shift the instruction by one byte */
327 /* so we can insert a rex prefix */
328 memmove (rex_tag + 1, rex_tag, (size_t)(*code - rex_tag));
330 /* New rex prefix only needs rex.B for %r15 base */
331 *rex_tag = AMD64_REX(AMD64_REX_B);
334 if (legacy_prefix_tag) {
335 old_instruction_start = legacy_prefix_tag;
337 old_instruction_start = rex_tag;
340 /* Clears the upper 32-bits of the previous base register */
341 amd64_mov_reg_reg_size (buf_ptr, true_basereg, true_basereg, 4);
342 insert_len = buf_ptr - buf;
344 /* Move the old instruction forward to make */
345 /* room for 'mov' stored in 'buf_ptr' */
346 memmove (old_instruction_start + insert_len, old_instruction_start, (size_t)(*code - old_instruction_start));
348 memcpy (old_instruction_start, buf, insert_len);
350 /* Sandboxed replacement for the normal membase_emit */
351 x86_memindex_emit (*code, dreg, AMD64_R15, offset, basereg, 0);
354 /* Normal default behavior, emit membase memory location */
355 x86_membase_emit_body (*code, dreg, basereg, offset);
360 static inline unsigned char*
361 amd64_skip_nops (unsigned char* code)
366 if ( code[0] == 0x90) {
370 if ( code[0] == 0x66 && code[1] == 0x90) {
374 if (code[0] == 0x0f && code[1] == 0x1f
375 && code[2] == 0x00) {
379 if (code[0] == 0x0f && code[1] == 0x1f
380 && code[2] == 0x40 && code[3] == 0x00) {
384 if (code[0] == 0x0f && code[1] == 0x1f
385 && code[2] == 0x44 && code[3] == 0x00
386 && code[4] == 0x00) {
390 if (code[0] == 0x66 && code[1] == 0x0f
391 && code[2] == 0x1f && code[3] == 0x44
392 && code[4] == 0x00 && code[5] == 0x00) {
396 if (code[0] == 0x0f && code[1] == 0x1f
397 && code[2] == 0x80 && code[3] == 0x00
398 && code[4] == 0x00 && code[5] == 0x00
399 && code[6] == 0x00) {
403 if (code[0] == 0x0f && code[1] == 0x1f
404 && code[2] == 0x84 && code[3] == 0x00
405 && code[4] == 0x00 && code[5] == 0x00
406 && code[6] == 0x00 && code[7] == 0x00) {
415 mono_arch_nacl_skip_nops (guint8* code)
417 return amd64_skip_nops(code);
420 #endif /*__native_client_codegen__*/
423 amd64_patch (unsigned char* code, gpointer target)
427 #ifdef __native_client_codegen__
428 code = amd64_skip_nops (code);
430 #if defined(__native_client_codegen__) && defined(__native_client__)
431 if (nacl_is_code_address (code)) {
432 /* For tail calls, code is patched after being installed */
433 /* but not through the normal "patch callsite" method. */
434 unsigned char buf[kNaClAlignment];
435 unsigned char *aligned_code = (uintptr_t)code & ~kNaClAlignmentMask;
437 memcpy (buf, aligned_code, kNaClAlignment);
438 /* Patch a temp buffer of bundle size, */
439 /* then install to actual location. */
440 amd64_patch (buf + ((uintptr_t)code - (uintptr_t)aligned_code), target);
441 ret = nacl_dyncode_modify (aligned_code, buf, kNaClAlignment);
445 target = nacl_modify_patch_target (target);
449 if ((code [0] >= 0x40) && (code [0] <= 0x4f)) {
454 if ((code [0] & 0xf8) == 0xb8) {
455 /* amd64_set_reg_template */
456 *(guint64*)(code + 1) = (guint64)target;
458 else if ((code [0] == 0x8b) && rex && x86_modrm_mod (code [1]) == 0 && x86_modrm_rm (code [1]) == 5) {
459 /* mov 0(%rip), %dreg */
460 *(guint32*)(code + 2) = (guint32)(guint64)target - 7;
462 else if ((code [0] == 0xff) && (code [1] == 0x15)) {
463 /* call *<OFFSET>(%rip) */
464 *(guint32*)(code + 2) = ((guint32)(guint64)target) - 7;
466 else if (code [0] == 0xe8) {
468 gint64 disp = (guint8*)target - (guint8*)code;
469 g_assert (amd64_is_imm32 (disp));
470 x86_patch (code, (unsigned char*)target);
473 x86_patch (code, (unsigned char*)target);
477 mono_amd64_patch (unsigned char* code, gpointer target)
479 amd64_patch (code, target);
488 ArgValuetypeAddrInIReg,
489 /* gsharedvt argument passed by addr */
492 ArgNone /* only in pair_storage */
500 /* Only if storage == ArgValuetypeInReg */
501 ArgStorage pair_storage [2];
503 /* The size of each pair */
513 gboolean need_stack_align;
514 /* The index of the vret arg in the argument list */
521 #define DEBUG(a) if (cfg->verbose_level > 1) a
524 static AMD64_Reg_No param_regs [] = { AMD64_RCX, AMD64_RDX, AMD64_R8, AMD64_R9 };
526 static AMD64_Reg_No return_regs [] = { AMD64_RAX, AMD64_RDX };
528 static AMD64_Reg_No param_regs [] = { AMD64_RDI, AMD64_RSI, AMD64_RDX, AMD64_RCX, AMD64_R8, AMD64_R9 };
530 static AMD64_Reg_No return_regs [] = { AMD64_RAX, AMD64_RDX };
534 add_general (guint32 *gr, guint32 *stack_size, ArgInfo *ainfo)
536 ainfo->offset = *stack_size;
538 if (*gr >= PARAM_REGS) {
539 ainfo->storage = ArgOnStack;
540 /* Since the same stack slot size is used for all arg */
541 /* types, it needs to be big enough to hold them all */
542 (*stack_size) += sizeof(mgreg_t);
545 ainfo->storage = ArgInIReg;
546 ainfo->reg = param_regs [*gr];
552 #define FLOAT_PARAM_REGS 4
554 #define FLOAT_PARAM_REGS 8
558 add_float (guint32 *gr, guint32 *stack_size, ArgInfo *ainfo, gboolean is_double)
560 ainfo->offset = *stack_size;
562 if (*gr >= FLOAT_PARAM_REGS) {
563 ainfo->storage = ArgOnStack;
564 /* Since the same stack slot size is used for both float */
565 /* types, it needs to be big enough to hold them both */
566 (*stack_size) += sizeof(mgreg_t);
569 /* A double register */
571 ainfo->storage = ArgInDoubleSSEReg;
573 ainfo->storage = ArgInFloatSSEReg;
579 typedef enum ArgumentClass {
587 merge_argument_class_from_type (MonoType *type, ArgumentClass class1)
589 ArgumentClass class2 = ARG_CLASS_NO_CLASS;
592 ptype = mini_get_underlying_type (type);
593 switch (ptype->type) {
602 case MONO_TYPE_STRING:
603 case MONO_TYPE_OBJECT:
604 case MONO_TYPE_CLASS:
605 case MONO_TYPE_SZARRAY:
607 case MONO_TYPE_FNPTR:
608 case MONO_TYPE_ARRAY:
611 class2 = ARG_CLASS_INTEGER;
616 class2 = ARG_CLASS_INTEGER;
618 class2 = ARG_CLASS_SSE;
622 case MONO_TYPE_TYPEDBYREF:
623 g_assert_not_reached ();
625 case MONO_TYPE_GENERICINST:
626 if (!mono_type_generic_inst_is_valuetype (ptype)) {
627 class2 = ARG_CLASS_INTEGER;
631 case MONO_TYPE_VALUETYPE: {
632 MonoMarshalType *info = mono_marshal_load_type_info (ptype->data.klass);
635 for (i = 0; i < info->num_fields; ++i) {
637 class2 = merge_argument_class_from_type (info->fields [i].field->type, class2);
642 g_assert_not_reached ();
646 if (class1 == class2)
648 else if (class1 == ARG_CLASS_NO_CLASS)
650 else if ((class1 == ARG_CLASS_MEMORY) || (class2 == ARG_CLASS_MEMORY))
651 class1 = ARG_CLASS_MEMORY;
652 else if ((class1 == ARG_CLASS_INTEGER) || (class2 == ARG_CLASS_INTEGER))
653 class1 = ARG_CLASS_INTEGER;
655 class1 = ARG_CLASS_SSE;
659 #ifdef __native_client_codegen__
661 /* Default alignment for Native Client is 32-byte. */
662 gint8 nacl_align_byte = -32; /* signed version of 0xe0 */
664 /* mono_arch_nacl_pad: Add pad bytes of alignment instructions at code, */
665 /* Check that alignment doesn't cross an alignment boundary. */
667 mono_arch_nacl_pad(guint8 *code, int pad)
669 const int kMaxPadding = 8; /* see amd64-codegen.h:amd64_padding_size() */
671 if (pad == 0) return code;
672 /* assertion: alignment cannot cross a block boundary */
673 g_assert (((uintptr_t)code & (~kNaClAlignmentMask)) ==
674 (((uintptr_t)code + pad - 1) & (~kNaClAlignmentMask)));
675 while (pad >= kMaxPadding) {
676 amd64_padding (code, kMaxPadding);
679 if (pad != 0) amd64_padding (code, pad);
685 count_fields_nested (MonoClass *klass)
687 MonoMarshalType *info;
690 info = mono_marshal_load_type_info (klass);
693 for (i = 0; i < info->num_fields; ++i) {
694 if (MONO_TYPE_ISSTRUCT (info->fields [i].field->type))
695 count += count_fields_nested (mono_class_from_mono_type (info->fields [i].field->type));
703 collect_field_info_nested (MonoClass *klass, MonoMarshalField *fields, int index, int offset)
705 MonoMarshalType *info;
708 info = mono_marshal_load_type_info (klass);
710 for (i = 0; i < info->num_fields; ++i) {
711 if (MONO_TYPE_ISSTRUCT (info->fields [i].field->type)) {
712 index = collect_field_info_nested (mono_class_from_mono_type (info->fields [i].field->type), fields, index, info->fields [i].offset);
714 memcpy (&fields [index], &info->fields [i], sizeof (MonoMarshalField));
715 fields [index].offset += offset;
724 add_valuetype_win64 (MonoMethodSignature *sig, ArgInfo *ainfo, MonoType *type,
726 guint32 *gr, guint32 *fr, guint32 *stack_size)
728 guint32 size, i, nfields;
730 ArgumentClass arg_class;
731 MonoMarshalType *info = NULL;
732 MonoMarshalField *fields = NULL;
734 gboolean pass_on_stack = FALSE;
736 klass = mono_class_from_mono_type (type);
737 size = mini_type_stack_size_full (&klass->byval_arg, NULL, sig->pinvoke);
739 pass_on_stack = TRUE;
741 /* If this struct can't be split up naturally into 8-byte */
742 /* chunks (registers), pass it on the stack. */
743 if (sig->pinvoke && !pass_on_stack) {
747 info = mono_marshal_load_type_info (klass);
751 * Collect field information recursively to be able to
752 * handle nested structures.
754 nfields = count_fields_nested (klass);
755 fields = g_new0 (MonoMarshalField, nfields);
756 collect_field_info_nested (klass, fields, 0, 0);
758 for (i = 0; i < nfields; ++i) {
759 field_size = mono_marshal_type_size (fields [i].field->type,
761 &align, TRUE, klass->unicode);
762 if ((fields [i].offset < 8) && (fields [i].offset + field_size) > 8) {
763 pass_on_stack = TRUE;
770 /* Allways pass in memory */
771 ainfo->offset = *stack_size;
772 *stack_size += ALIGN_TO (size, 8);
773 ainfo->storage = is_return ? ArgValuetypeAddrInIReg : ArgOnStack;
780 int n = mono_class_value_size (klass, NULL);
785 arg_class = ARG_CLASS_MEMORY;
787 /* Always pass in 1 integer register */
788 arg_class = ARG_CLASS_INTEGER;
793 ainfo->storage = ArgValuetypeInReg;
794 ainfo->pair_storage [0] = ainfo->pair_storage [1] = ArgNone;
798 switch (info->native_size) {
799 case 1: case 2: case 4: case 8:
803 ainfo->storage = ArgValuetypeAddrInIReg;
804 ainfo->offset = *stack_size;
805 *stack_size += ALIGN_TO (info->native_size, 8);
808 ainfo->storage = ArgValuetypeAddrInIReg;
810 if (*gr < PARAM_REGS) {
811 ainfo->pair_storage [0] = ArgInIReg;
812 ainfo->pair_regs [0] = param_regs [*gr];
816 ainfo->pair_storage [0] = ArgOnStack;
817 ainfo->offset = *stack_size;
828 ArgumentClass class1;
831 class1 = ARG_CLASS_MEMORY;
833 class1 = ARG_CLASS_NO_CLASS;
834 for (i = 0; i < nfields; ++i) {
835 size = mono_marshal_type_size (fields [i].field->type,
837 &align, TRUE, klass->unicode);
838 /* How far into this quad this data extends.*/
839 /* (8 is size of quad) */
840 argsize = fields [i].offset + size;
842 class1 = merge_argument_class_from_type (fields [i].field->type, class1);
844 g_assert (class1 != ARG_CLASS_NO_CLASS);
850 /* Allocate registers */
855 while (argsize != 1 && argsize != 2 && argsize != 4 && argsize != 8)
858 ainfo->storage = ArgValuetypeInReg;
859 ainfo->pair_storage [0] = ainfo->pair_storage [1] = ArgNone;
860 ainfo->pair_size [0] = argsize;
861 ainfo->pair_size [1] = 0;
864 case ARG_CLASS_INTEGER:
865 if (*gr >= PARAM_REGS)
866 arg_class = ARG_CLASS_MEMORY;
868 ainfo->pair_storage [0] = ArgInIReg;
870 ainfo->pair_regs [0] = return_regs [*gr];
872 ainfo->pair_regs [0] = param_regs [*gr];
877 if (*fr >= FLOAT_PARAM_REGS)
878 arg_class = ARG_CLASS_MEMORY;
881 ainfo->pair_storage [0] = ArgInFloatSSEReg;
883 ainfo->pair_storage [0] = ArgInDoubleSSEReg;
884 ainfo->pair_regs [0] = *fr;
888 case ARG_CLASS_MEMORY:
891 g_assert_not_reached ();
894 if (arg_class == ARG_CLASS_MEMORY) {
895 /* Revert possible register assignments */
899 ainfo->offset = *stack_size;
900 *stack_size += sizeof (mgreg_t);
901 ainfo->storage = is_return ? ArgValuetypeAddrInIReg : ArgOnStack;
905 #endif /* TARGET_WIN32 */
908 add_valuetype (MonoMethodSignature *sig, ArgInfo *ainfo, MonoType *type,
910 guint32 *gr, guint32 *fr, guint32 *stack_size)
913 add_valuetype_win64 (sig, ainfo, type, is_return, gr, fr, stack_size);
915 guint32 size, quad, nquads, i, nfields;
916 /* Keep track of the size used in each quad so we can */
917 /* use the right size when copying args/return vars. */
918 guint32 quadsize [2] = {8, 8};
919 ArgumentClass args [2];
920 MonoMarshalType *info = NULL;
921 MonoMarshalField *fields = NULL;
923 gboolean pass_on_stack = FALSE;
925 klass = mono_class_from_mono_type (type);
926 size = mini_type_stack_size_full (&klass->byval_arg, NULL, sig->pinvoke);
927 if (!sig->pinvoke && ((is_return && (size == 8)) || (!is_return && (size <= 16)))) {
928 /* We pass and return vtypes of size 8 in a register */
929 } else if (!sig->pinvoke || (size == 0) || (size > 16)) {
930 pass_on_stack = TRUE;
933 /* If this struct can't be split up naturally into 8-byte */
934 /* chunks (registers), pass it on the stack. */
935 if (sig->pinvoke && !pass_on_stack) {
939 info = mono_marshal_load_type_info (klass);
943 * Collect field information recursively to be able to
944 * handle nested structures.
946 nfields = count_fields_nested (klass);
947 fields = g_new0 (MonoMarshalField, nfields);
948 collect_field_info_nested (klass, fields, 0, 0);
950 for (i = 0; i < nfields; ++i) {
951 field_size = mono_marshal_type_size (fields [i].field->type,
953 &align, TRUE, klass->unicode);
954 if ((fields [i].offset < 8) && (fields [i].offset + field_size) > 8) {
955 pass_on_stack = TRUE;
962 ainfo->storage = ArgValuetypeInReg;
963 ainfo->pair_storage [0] = ainfo->pair_storage [1] = ArgNone;
968 /* Allways pass in memory */
969 ainfo->offset = *stack_size;
970 *stack_size += ALIGN_TO (size, 8);
971 ainfo->storage = is_return ? ArgValuetypeAddrInIReg : ArgOnStack;
983 int n = mono_class_value_size (klass, NULL);
985 quadsize [0] = n >= 8 ? 8 : n;
986 quadsize [1] = n >= 8 ? MAX (n - 8, 8) : 0;
988 /* Always pass in 1 or 2 integer registers */
989 args [0] = ARG_CLASS_INTEGER;
990 args [1] = ARG_CLASS_INTEGER;
991 /* Only the simplest cases are supported */
992 if (is_return && nquads != 1) {
993 args [0] = ARG_CLASS_MEMORY;
994 args [1] = ARG_CLASS_MEMORY;
998 * Implement the algorithm from section 3.2.3 of the X86_64 ABI.
999 * The X87 and SSEUP stuff is left out since there are no such types in
1005 ainfo->storage = ArgValuetypeInReg;
1006 ainfo->pair_storage [0] = ainfo->pair_storage [1] = ArgNone;
1010 if (info->native_size > 16) {
1011 ainfo->offset = *stack_size;
1012 *stack_size += ALIGN_TO (info->native_size, 8);
1013 ainfo->storage = is_return ? ArgValuetypeAddrInIReg : ArgOnStack;
1019 args [0] = ARG_CLASS_NO_CLASS;
1020 args [1] = ARG_CLASS_NO_CLASS;
1021 for (quad = 0; quad < nquads; ++quad) {
1024 ArgumentClass class1;
1027 class1 = ARG_CLASS_MEMORY;
1029 class1 = ARG_CLASS_NO_CLASS;
1030 for (i = 0; i < nfields; ++i) {
1031 size = mono_marshal_type_size (fields [i].field->type,
1033 &align, TRUE, klass->unicode);
1034 if ((fields [i].offset < 8) && (fields [i].offset + size) > 8) {
1035 /* Unaligned field */
1039 /* Skip fields in other quad */
1040 if ((quad == 0) && (fields [i].offset >= 8))
1042 if ((quad == 1) && (fields [i].offset < 8))
1045 /* How far into this quad this data extends.*/
1046 /* (8 is size of quad) */
1047 quadsize [quad] = fields [i].offset + size - (quad * 8);
1049 class1 = merge_argument_class_from_type (fields [i].field->type, class1);
1051 g_assert (class1 != ARG_CLASS_NO_CLASS);
1052 args [quad] = class1;
1058 /* Post merger cleanup */
1059 if ((args [0] == ARG_CLASS_MEMORY) || (args [1] == ARG_CLASS_MEMORY))
1060 args [0] = args [1] = ARG_CLASS_MEMORY;
1062 /* Allocate registers */
1067 while (quadsize [0] != 1 && quadsize [0] != 2 && quadsize [0] != 4 && quadsize [0] != 8)
1069 while (quadsize [1] != 1 && quadsize [1] != 2 && quadsize [1] != 4 && quadsize [1] != 8)
1072 ainfo->storage = ArgValuetypeInReg;
1073 ainfo->pair_storage [0] = ainfo->pair_storage [1] = ArgNone;
1074 g_assert (quadsize [0] <= 8);
1075 g_assert (quadsize [1] <= 8);
1076 ainfo->pair_size [0] = quadsize [0];
1077 ainfo->pair_size [1] = quadsize [1];
1078 ainfo->nregs = nquads;
1079 for (quad = 0; quad < nquads; ++quad) {
1080 switch (args [quad]) {
1081 case ARG_CLASS_INTEGER:
1082 if (*gr >= PARAM_REGS)
1083 args [quad] = ARG_CLASS_MEMORY;
1085 ainfo->pair_storage [quad] = ArgInIReg;
1087 ainfo->pair_regs [quad] = return_regs [*gr];
1089 ainfo->pair_regs [quad] = param_regs [*gr];
1094 if (*fr >= FLOAT_PARAM_REGS)
1095 args [quad] = ARG_CLASS_MEMORY;
1097 if (quadsize[quad] <= 4)
1098 ainfo->pair_storage [quad] = ArgInFloatSSEReg;
1099 else ainfo->pair_storage [quad] = ArgInDoubleSSEReg;
1100 ainfo->pair_regs [quad] = *fr;
1104 case ARG_CLASS_MEMORY:
1107 g_assert_not_reached ();
1111 if ((args [0] == ARG_CLASS_MEMORY) || (args [1] == ARG_CLASS_MEMORY)) {
1112 /* Revert possible register assignments */
1116 ainfo->offset = *stack_size;
1118 *stack_size += ALIGN_TO (info->native_size, 8);
1120 *stack_size += nquads * sizeof(mgreg_t);
1121 ainfo->storage = is_return ? ArgValuetypeAddrInIReg : ArgOnStack;
1124 #endif /* !TARGET_WIN32 */
1130 * Obtain information about a call according to the calling convention.
1131 * For AMD64, see the "System V ABI, x86-64 Architecture Processor Supplement
1132 * Draft Version 0.23" document for more information.
1135 get_call_info (MonoMemPool *mp, MonoMethodSignature *sig)
1137 guint32 i, gr, fr, pstart;
1139 int n = sig->hasthis + sig->param_count;
1140 guint32 stack_size = 0;
1142 gboolean is_pinvoke = sig->pinvoke;
1145 cinfo = mono_mempool_alloc0 (mp, sizeof (CallInfo) + (sizeof (ArgInfo) * n));
1147 cinfo = g_malloc0 (sizeof (CallInfo) + (sizeof (ArgInfo) * n));
1155 /* Reserve space where the callee can save the argument registers */
1156 stack_size = 4 * sizeof (mgreg_t);
1160 ret_type = mini_get_underlying_type (sig->ret);
1161 switch (ret_type->type) {
1171 case MONO_TYPE_FNPTR:
1172 case MONO_TYPE_CLASS:
1173 case MONO_TYPE_OBJECT:
1174 case MONO_TYPE_SZARRAY:
1175 case MONO_TYPE_ARRAY:
1176 case MONO_TYPE_STRING:
1177 cinfo->ret.storage = ArgInIReg;
1178 cinfo->ret.reg = AMD64_RAX;
1182 cinfo->ret.storage = ArgInIReg;
1183 cinfo->ret.reg = AMD64_RAX;
1186 cinfo->ret.storage = ArgInFloatSSEReg;
1187 cinfo->ret.reg = AMD64_XMM0;
1190 cinfo->ret.storage = ArgInDoubleSSEReg;
1191 cinfo->ret.reg = AMD64_XMM0;
1193 case MONO_TYPE_GENERICINST:
1194 if (!mono_type_generic_inst_is_valuetype (ret_type)) {
1195 cinfo->ret.storage = ArgInIReg;
1196 cinfo->ret.reg = AMD64_RAX;
1199 if (mini_is_gsharedvt_type (ret_type)) {
1200 cinfo->ret.storage = ArgValuetypeAddrInIReg;
1204 case MONO_TYPE_VALUETYPE:
1205 case MONO_TYPE_TYPEDBYREF: {
1206 guint32 tmp_gr = 0, tmp_fr = 0, tmp_stacksize = 0;
1208 add_valuetype (sig, &cinfo->ret, ret_type, TRUE, &tmp_gr, &tmp_fr, &tmp_stacksize);
1209 g_assert (cinfo->ret.storage != ArgInIReg);
1213 case MONO_TYPE_MVAR:
1214 g_assert (mini_is_gsharedvt_type (ret_type));
1215 cinfo->ret.storage = ArgValuetypeAddrInIReg;
1217 case MONO_TYPE_VOID:
1220 g_error ("Can't handle as return value 0x%x", ret_type->type);
1225 * To simplify get_this_arg_reg () and LLVM integration, emit the vret arg after
1226 * the first argument, allowing 'this' to be always passed in the first arg reg.
1227 * Also do this if the first argument is a reference type, since virtual calls
1228 * are sometimes made using calli without sig->hasthis set, like in the delegate
1231 if (cinfo->ret.storage == ArgValuetypeAddrInIReg && !is_pinvoke && (sig->hasthis || (sig->param_count > 0 && MONO_TYPE_IS_REFERENCE (mini_get_underlying_type (sig->params [0]))))) {
1233 add_general (&gr, &stack_size, cinfo->args + 0);
1235 add_general (&gr, &stack_size, &cinfo->args [sig->hasthis + 0]);
1238 add_general (&gr, &stack_size, &cinfo->ret);
1239 cinfo->ret.storage = ArgValuetypeAddrInIReg;
1240 cinfo->vret_arg_index = 1;
1244 add_general (&gr, &stack_size, cinfo->args + 0);
1246 if (cinfo->ret.storage == ArgValuetypeAddrInIReg) {
1247 add_general (&gr, &stack_size, &cinfo->ret);
1248 cinfo->ret.storage = ArgValuetypeAddrInIReg;
1252 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG) && (n == 0)) {
1254 fr = FLOAT_PARAM_REGS;
1256 /* Emit the signature cookie just before the implicit arguments */
1257 add_general (&gr, &stack_size, &cinfo->sig_cookie);
1260 for (i = pstart; i < sig->param_count; ++i) {
1261 ArgInfo *ainfo = &cinfo->args [sig->hasthis + i];
1265 /* The float param registers and other param registers must be the same index on Windows x64.*/
1272 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG) && (i == sig->sentinelpos)) {
1273 /* We allways pass the sig cookie on the stack for simplicity */
1275 * Prevent implicit arguments + the sig cookie from being passed
1279 fr = FLOAT_PARAM_REGS;
1281 /* Emit the signature cookie just before the implicit arguments */
1282 add_general (&gr, &stack_size, &cinfo->sig_cookie);
1285 ptype = mini_get_underlying_type (sig->params [i]);
1286 switch (ptype->type) {
1289 add_general (&gr, &stack_size, ainfo);
1293 add_general (&gr, &stack_size, ainfo);
1297 add_general (&gr, &stack_size, ainfo);
1302 case MONO_TYPE_FNPTR:
1303 case MONO_TYPE_CLASS:
1304 case MONO_TYPE_OBJECT:
1305 case MONO_TYPE_STRING:
1306 case MONO_TYPE_SZARRAY:
1307 case MONO_TYPE_ARRAY:
1308 add_general (&gr, &stack_size, ainfo);
1310 case MONO_TYPE_GENERICINST:
1311 if (!mono_type_generic_inst_is_valuetype (ptype)) {
1312 add_general (&gr, &stack_size, ainfo);
1315 if (mini_is_gsharedvt_type (ptype)) {
1316 /* gsharedvt arguments are passed by ref */
1317 add_general (&gr, &stack_size, ainfo);
1318 if (ainfo->storage == ArgInIReg)
1319 ainfo->storage = ArgGSharedVtInReg;
1321 ainfo->storage = ArgGSharedVtOnStack;
1325 case MONO_TYPE_VALUETYPE:
1326 case MONO_TYPE_TYPEDBYREF:
1327 add_valuetype (sig, ainfo, sig->params [i], FALSE, &gr, &fr, &stack_size);
1332 add_general (&gr, &stack_size, ainfo);
1335 add_float (&fr, &stack_size, ainfo, FALSE);
1338 add_float (&fr, &stack_size, ainfo, TRUE);
1341 case MONO_TYPE_MVAR:
1342 /* gsharedvt arguments are passed by ref */
1343 g_assert (mini_is_gsharedvt_type (ptype));
1344 add_general (&gr, &stack_size, ainfo);
1345 if (ainfo->storage == ArgInIReg)
1346 ainfo->storage = ArgGSharedVtInReg;
1348 ainfo->storage = ArgGSharedVtOnStack;
1351 g_assert_not_reached ();
1355 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG) && (n > 0) && (sig->sentinelpos == sig->param_count)) {
1357 fr = FLOAT_PARAM_REGS;
1359 /* Emit the signature cookie just before the implicit arguments */
1360 add_general (&gr, &stack_size, &cinfo->sig_cookie);
1363 cinfo->stack_usage = stack_size;
1364 cinfo->reg_usage = gr;
1365 cinfo->freg_usage = fr;
1370 * mono_arch_get_argument_info:
1371 * @csig: a method signature
1372 * @param_count: the number of parameters to consider
1373 * @arg_info: an array to store the result infos
1375 * Gathers information on parameters such as size, alignment and
1376 * padding. arg_info should be large enought to hold param_count + 1 entries.
1378 * Returns the size of the argument area on the stack.
1381 mono_arch_get_argument_info (MonoMethodSignature *csig, int param_count, MonoJitArgumentInfo *arg_info)
1384 CallInfo *cinfo = get_call_info (NULL, csig);
1385 guint32 args_size = cinfo->stack_usage;
1387 /* The arguments are saved to a stack area in mono_arch_instrument_prolog */
1388 if (csig->hasthis) {
1389 arg_info [0].offset = 0;
1392 for (k = 0; k < param_count; k++) {
1393 arg_info [k + 1].offset = ((k + csig->hasthis) * 8);
1395 arg_info [k + 1].size = 0;
1404 mono_arch_tail_call_supported (MonoCompile *cfg, MonoMethodSignature *caller_sig, MonoMethodSignature *callee_sig)
1408 MonoType *callee_ret;
1410 c1 = get_call_info (NULL, caller_sig);
1411 c2 = get_call_info (NULL, callee_sig);
1412 res = c1->stack_usage >= c2->stack_usage;
1413 callee_ret = mini_get_underlying_type (callee_sig->ret);
1414 if (callee_ret && MONO_TYPE_ISSTRUCT (callee_ret) && c2->ret.storage != ArgValuetypeInReg)
1415 /* An address on the callee's stack is passed as the first argument */
1425 * Initialize the cpu to execute managed code.
1428 mono_arch_cpu_init (void)
1433 /* spec compliance requires running with double precision */
1434 __asm__ __volatile__ ("fnstcw %0\n": "=m" (fpcw));
1435 fpcw &= ~X86_FPCW_PRECC_MASK;
1436 fpcw |= X86_FPCW_PREC_DOUBLE;
1437 __asm__ __volatile__ ("fldcw %0\n": : "m" (fpcw));
1438 __asm__ __volatile__ ("fnstcw %0\n": "=m" (fpcw));
1440 /* TODO: This is crashing on Win64 right now.
1441 * _control87 (_PC_53, MCW_PC);
1447 * Initialize architecture specific code.
1450 mono_arch_init (void)
1452 mono_mutex_init_recursive (&mini_arch_mutex);
1453 #if defined(__native_client_codegen__)
1454 mono_native_tls_alloc (&nacl_instruction_depth, NULL);
1455 mono_native_tls_set_value (nacl_instruction_depth, (gpointer)0);
1456 mono_native_tls_alloc (&nacl_rex_tag, NULL);
1457 mono_native_tls_alloc (&nacl_legacy_prefix_tag, NULL);
1460 mono_aot_register_jit_icall ("mono_amd64_throw_exception", mono_amd64_throw_exception);
1461 mono_aot_register_jit_icall ("mono_amd64_throw_corlib_exception", mono_amd64_throw_corlib_exception);
1462 mono_aot_register_jit_icall ("mono_amd64_resume_unwind", mono_amd64_resume_unwind);
1463 mono_aot_register_jit_icall ("mono_amd64_get_original_ip", mono_amd64_get_original_ip);
1466 bp_trampoline = mini_get_breakpoint_trampoline ();
1470 * Cleanup architecture specific code.
1473 mono_arch_cleanup (void)
1475 mono_mutex_destroy (&mini_arch_mutex);
1476 #if defined(__native_client_codegen__)
1477 mono_native_tls_free (nacl_instruction_depth);
1478 mono_native_tls_free (nacl_rex_tag);
1479 mono_native_tls_free (nacl_legacy_prefix_tag);
1484 * This function returns the optimizations supported on this cpu.
1487 mono_arch_cpu_optimizations (guint32 *exclude_mask)
1493 if (mono_hwcap_x86_has_cmov) {
1494 opts |= MONO_OPT_CMOV;
1496 if (mono_hwcap_x86_has_fcmov)
1497 opts |= MONO_OPT_FCMOV;
1499 *exclude_mask |= MONO_OPT_FCMOV;
1501 *exclude_mask |= MONO_OPT_CMOV;
1508 * This function test for all SSE functions supported.
1510 * Returns a bitmask corresponding to all supported versions.
1514 mono_arch_cpu_enumerate_simd_versions (void)
1516 guint32 sse_opts = 0;
1518 if (mono_hwcap_x86_has_sse1)
1519 sse_opts |= SIMD_VERSION_SSE1;
1521 if (mono_hwcap_x86_has_sse2)
1522 sse_opts |= SIMD_VERSION_SSE2;
1524 if (mono_hwcap_x86_has_sse3)
1525 sse_opts |= SIMD_VERSION_SSE3;
1527 if (mono_hwcap_x86_has_ssse3)
1528 sse_opts |= SIMD_VERSION_SSSE3;
1530 if (mono_hwcap_x86_has_sse41)
1531 sse_opts |= SIMD_VERSION_SSE41;
1533 if (mono_hwcap_x86_has_sse42)
1534 sse_opts |= SIMD_VERSION_SSE42;
1536 if (mono_hwcap_x86_has_sse4a)
1537 sse_opts |= SIMD_VERSION_SSE4a;
1545 mono_arch_get_allocatable_int_vars (MonoCompile *cfg)
1550 for (i = 0; i < cfg->num_varinfo; i++) {
1551 MonoInst *ins = cfg->varinfo [i];
1552 MonoMethodVar *vmv = MONO_VARINFO (cfg, i);
1555 if (vmv->range.first_use.abs_pos >= vmv->range.last_use.abs_pos)
1558 if ((ins->flags & (MONO_INST_IS_DEAD|MONO_INST_VOLATILE|MONO_INST_INDIRECT)) ||
1559 (ins->opcode != OP_LOCAL && ins->opcode != OP_ARG))
1562 if (mono_is_regsize_var (ins->inst_vtype)) {
1563 g_assert (MONO_VARINFO (cfg, i)->reg == -1);
1564 g_assert (i == vmv->idx);
1565 vars = g_list_prepend (vars, vmv);
1569 vars = mono_varlist_sort (cfg, vars, 0);
1575 * mono_arch_compute_omit_fp:
1577 * Determine whenever the frame pointer can be eliminated.
1580 mono_arch_compute_omit_fp (MonoCompile *cfg)
1582 MonoMethodSignature *sig;
1583 MonoMethodHeader *header;
1587 if (cfg->arch.omit_fp_computed)
1590 header = cfg->header;
1592 sig = mono_method_signature (cfg->method);
1594 if (!cfg->arch.cinfo)
1595 cfg->arch.cinfo = get_call_info (cfg->mempool, sig);
1596 cinfo = cfg->arch.cinfo;
1599 * FIXME: Remove some of the restrictions.
1601 cfg->arch.omit_fp = TRUE;
1602 cfg->arch.omit_fp_computed = TRUE;
1604 #ifdef __native_client_codegen__
1605 /* NaCl modules may not change the value of RBP, so it cannot be */
1606 /* used as a normal register, but it can be used as a frame pointer*/
1607 cfg->disable_omit_fp = TRUE;
1608 cfg->arch.omit_fp = FALSE;
1611 if (cfg->disable_omit_fp)
1612 cfg->arch.omit_fp = FALSE;
1614 if (!debug_omit_fp ())
1615 cfg->arch.omit_fp = FALSE;
1617 if (cfg->method->save_lmf)
1618 cfg->arch.omit_fp = FALSE;
1620 if (cfg->flags & MONO_CFG_HAS_ALLOCA)
1621 cfg->arch.omit_fp = FALSE;
1622 if (header->num_clauses)
1623 cfg->arch.omit_fp = FALSE;
1624 if (cfg->param_area)
1625 cfg->arch.omit_fp = FALSE;
1626 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG))
1627 cfg->arch.omit_fp = FALSE;
1628 if ((mono_jit_trace_calls != NULL && mono_trace_eval (cfg->method)) ||
1629 (cfg->prof_options & MONO_PROFILE_ENTER_LEAVE))
1630 cfg->arch.omit_fp = FALSE;
1631 for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
1632 ArgInfo *ainfo = &cinfo->args [i];
1634 if (ainfo->storage == ArgOnStack) {
1636 * The stack offset can only be determined when the frame
1639 cfg->arch.omit_fp = FALSE;
1644 for (i = cfg->locals_start; i < cfg->num_varinfo; i++) {
1645 MonoInst *ins = cfg->varinfo [i];
1648 locals_size += mono_type_size (ins->inst_vtype, &ialign);
1653 mono_arch_get_global_int_regs (MonoCompile *cfg)
1657 mono_arch_compute_omit_fp (cfg);
1659 if (cfg->arch.omit_fp)
1660 regs = g_list_prepend (regs, (gpointer)AMD64_RBP);
1662 /* We use the callee saved registers for global allocation */
1663 regs = g_list_prepend (regs, (gpointer)AMD64_RBX);
1664 regs = g_list_prepend (regs, (gpointer)AMD64_R12);
1665 regs = g_list_prepend (regs, (gpointer)AMD64_R13);
1666 regs = g_list_prepend (regs, (gpointer)AMD64_R14);
1667 #ifndef __native_client_codegen__
1668 regs = g_list_prepend (regs, (gpointer)AMD64_R15);
1671 regs = g_list_prepend (regs, (gpointer)AMD64_RDI);
1672 regs = g_list_prepend (regs, (gpointer)AMD64_RSI);
1679 mono_arch_get_global_fp_regs (MonoCompile *cfg)
1684 /* All XMM registers */
1685 for (i = 0; i < 16; ++i)
1686 regs = g_list_prepend (regs, GINT_TO_POINTER (i));
1692 mono_arch_get_iregs_clobbered_by_call (MonoCallInst *call)
1694 static GList *r = NULL;
1699 regs = g_list_prepend (regs, (gpointer)AMD64_RBP);
1700 regs = g_list_prepend (regs, (gpointer)AMD64_RBX);
1701 regs = g_list_prepend (regs, (gpointer)AMD64_R12);
1702 regs = g_list_prepend (regs, (gpointer)AMD64_R13);
1703 regs = g_list_prepend (regs, (gpointer)AMD64_R14);
1704 #ifndef __native_client_codegen__
1705 regs = g_list_prepend (regs, (gpointer)AMD64_R15);
1708 regs = g_list_prepend (regs, (gpointer)AMD64_R10);
1709 regs = g_list_prepend (regs, (gpointer)AMD64_R9);
1710 regs = g_list_prepend (regs, (gpointer)AMD64_R8);
1711 regs = g_list_prepend (regs, (gpointer)AMD64_RDI);
1712 regs = g_list_prepend (regs, (gpointer)AMD64_RSI);
1713 regs = g_list_prepend (regs, (gpointer)AMD64_RDX);
1714 regs = g_list_prepend (regs, (gpointer)AMD64_RCX);
1715 regs = g_list_prepend (regs, (gpointer)AMD64_RAX);
1717 InterlockedCompareExchangePointer ((gpointer*)&r, regs, NULL);
1724 mono_arch_get_fregs_clobbered_by_call (MonoCallInst *call)
1727 static GList *r = NULL;
1732 for (i = 0; i < AMD64_XMM_NREG; ++i)
1733 regs = g_list_prepend (regs, GINT_TO_POINTER (MONO_MAX_IREGS + i));
1735 InterlockedCompareExchangePointer ((gpointer*)&r, regs, NULL);
1742 * mono_arch_regalloc_cost:
1744 * Return the cost, in number of memory references, of the action of
1745 * allocating the variable VMV into a register during global register
1749 mono_arch_regalloc_cost (MonoCompile *cfg, MonoMethodVar *vmv)
1751 MonoInst *ins = cfg->varinfo [vmv->idx];
1753 if (cfg->method->save_lmf)
1754 /* The register is already saved */
1755 /* substract 1 for the invisible store in the prolog */
1756 return (ins->opcode == OP_ARG) ? 0 : 1;
1759 return (ins->opcode == OP_ARG) ? 1 : 2;
1763 * mono_arch_fill_argument_info:
1765 * Populate cfg->args, cfg->ret and cfg->vret_addr with information about the arguments
1769 mono_arch_fill_argument_info (MonoCompile *cfg)
1772 MonoMethodSignature *sig;
1777 sig = mono_method_signature (cfg->method);
1779 cinfo = cfg->arch.cinfo;
1780 sig_ret = mini_get_underlying_type (sig->ret);
1783 * Contrary to mono_arch_allocate_vars (), the information should describe
1784 * where the arguments are at the beginning of the method, not where they can be
1785 * accessed during the execution of the method. The later makes no sense for the
1786 * global register allocator, since a variable can be in more than one location.
1788 switch (cinfo->ret.storage) {
1790 case ArgInFloatSSEReg:
1791 case ArgInDoubleSSEReg:
1792 cfg->ret->opcode = OP_REGVAR;
1793 cfg->ret->inst_c0 = cinfo->ret.reg;
1795 case ArgValuetypeInReg:
1796 cfg->ret->opcode = OP_REGOFFSET;
1797 cfg->ret->inst_basereg = -1;
1798 cfg->ret->inst_offset = -1;
1803 g_assert_not_reached ();
1806 for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
1807 ArgInfo *ainfo = &cinfo->args [i];
1809 ins = cfg->args [i];
1811 switch (ainfo->storage) {
1813 case ArgInFloatSSEReg:
1814 case ArgInDoubleSSEReg:
1815 ins->opcode = OP_REGVAR;
1816 ins->inst_c0 = ainfo->reg;
1819 ins->opcode = OP_REGOFFSET;
1820 ins->inst_basereg = -1;
1821 ins->inst_offset = -1;
1823 case ArgValuetypeInReg:
1825 ins->opcode = OP_NOP;
1828 g_assert_not_reached ();
1834 mono_arch_allocate_vars (MonoCompile *cfg)
1837 MonoMethodSignature *sig;
1840 guint32 locals_stack_size, locals_stack_align;
1844 sig = mono_method_signature (cfg->method);
1846 cinfo = cfg->arch.cinfo;
1847 sig_ret = mini_get_underlying_type (sig->ret);
1849 mono_arch_compute_omit_fp (cfg);
1852 * We use the ABI calling conventions for managed code as well.
1853 * Exception: valuetypes are only sometimes passed or returned in registers.
1857 * The stack looks like this:
1858 * <incoming arguments passed on the stack>
1860 * <lmf/caller saved registers>
1863 * <localloc area> -> grows dynamically
1867 if (cfg->arch.omit_fp) {
1868 cfg->flags |= MONO_CFG_HAS_SPILLUP;
1869 cfg->frame_reg = AMD64_RSP;
1872 /* Locals are allocated backwards from %fp */
1873 cfg->frame_reg = AMD64_RBP;
1877 cfg->arch.saved_iregs = cfg->used_int_regs;
1878 if (cfg->method->save_lmf)
1879 /* Save all callee-saved registers normally, and restore them when unwinding through an LMF */
1880 cfg->arch.saved_iregs |= (1 << AMD64_RBX) | (1 << AMD64_R12) | (1 << AMD64_R13) | (1 << AMD64_R14) | (1 << AMD64_R15);
1882 if (cfg->arch.omit_fp)
1883 cfg->arch.reg_save_area_offset = offset;
1884 /* Reserve space for callee saved registers */
1885 for (i = 0; i < AMD64_NREG; ++i)
1886 if (AMD64_IS_CALLEE_SAVED_REG (i) && (cfg->arch.saved_iregs & (1 << i))) {
1887 offset += sizeof(mgreg_t);
1889 if (!cfg->arch.omit_fp)
1890 cfg->arch.reg_save_area_offset = -offset;
1892 if (sig_ret->type != MONO_TYPE_VOID) {
1893 switch (cinfo->ret.storage) {
1895 case ArgInFloatSSEReg:
1896 case ArgInDoubleSSEReg:
1897 cfg->ret->opcode = OP_REGVAR;
1898 cfg->ret->inst_c0 = cinfo->ret.reg;
1900 case ArgValuetypeAddrInIReg:
1901 /* The register is volatile */
1902 cfg->vret_addr->opcode = OP_REGOFFSET;
1903 cfg->vret_addr->inst_basereg = cfg->frame_reg;
1904 if (cfg->arch.omit_fp) {
1905 cfg->vret_addr->inst_offset = offset;
1909 cfg->vret_addr->inst_offset = -offset;
1911 if (G_UNLIKELY (cfg->verbose_level > 1)) {
1912 printf ("vret_addr =");
1913 mono_print_ins (cfg->vret_addr);
1916 case ArgValuetypeInReg:
1917 /* Allocate a local to hold the result, the epilog will copy it to the correct place */
1918 cfg->ret->opcode = OP_REGOFFSET;
1919 cfg->ret->inst_basereg = cfg->frame_reg;
1920 if (cfg->arch.omit_fp) {
1921 cfg->ret->inst_offset = offset;
1922 offset += cinfo->ret.pair_storage [1] == ArgNone ? 8 : 16;
1924 offset += cinfo->ret.pair_storage [1] == ArgNone ? 8 : 16;
1925 cfg->ret->inst_offset = - offset;
1929 g_assert_not_reached ();
1931 cfg->ret->dreg = cfg->ret->inst_c0;
1934 /* Allocate locals */
1935 offsets = mono_allocate_stack_slots (cfg, cfg->arch.omit_fp ? FALSE: TRUE, &locals_stack_size, &locals_stack_align);
1936 if (locals_stack_size > MONO_ARCH_MAX_FRAME_SIZE) {
1937 char *mname = mono_method_full_name (cfg->method, TRUE);
1938 cfg->exception_type = MONO_EXCEPTION_INVALID_PROGRAM;
1939 cfg->exception_message = g_strdup_printf ("Method %s stack is too big.", mname);
1944 if (locals_stack_align) {
1945 offset += (locals_stack_align - 1);
1946 offset &= ~(locals_stack_align - 1);
1948 if (cfg->arch.omit_fp) {
1949 cfg->locals_min_stack_offset = offset;
1950 cfg->locals_max_stack_offset = offset + locals_stack_size;
1952 cfg->locals_min_stack_offset = - (offset + locals_stack_size);
1953 cfg->locals_max_stack_offset = - offset;
1956 for (i = cfg->locals_start; i < cfg->num_varinfo; i++) {
1957 if (offsets [i] != -1) {
1958 MonoInst *ins = cfg->varinfo [i];
1959 ins->opcode = OP_REGOFFSET;
1960 ins->inst_basereg = cfg->frame_reg;
1961 if (cfg->arch.omit_fp)
1962 ins->inst_offset = (offset + offsets [i]);
1964 ins->inst_offset = - (offset + offsets [i]);
1965 //printf ("allocated local %d to ", i); mono_print_tree_nl (ins);
1968 offset += locals_stack_size;
1970 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG)) {
1971 g_assert (!cfg->arch.omit_fp);
1972 g_assert (cinfo->sig_cookie.storage == ArgOnStack);
1973 cfg->sig_cookie = cinfo->sig_cookie.offset + ARGS_OFFSET;
1976 for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
1977 ins = cfg->args [i];
1978 if (ins->opcode != OP_REGVAR) {
1979 ArgInfo *ainfo = &cinfo->args [i];
1980 gboolean inreg = TRUE;
1982 /* FIXME: Allocate volatile arguments to registers */
1983 if (ins->flags & (MONO_INST_VOLATILE|MONO_INST_INDIRECT))
1987 * Under AMD64, all registers used to pass arguments to functions
1988 * are volatile across calls.
1989 * FIXME: Optimize this.
1991 if ((ainfo->storage == ArgInIReg) || (ainfo->storage == ArgInFloatSSEReg) || (ainfo->storage == ArgInDoubleSSEReg) || (ainfo->storage == ArgValuetypeInReg) || (ainfo->storage == ArgGSharedVtInReg))
1994 ins->opcode = OP_REGOFFSET;
1996 switch (ainfo->storage) {
1998 case ArgInFloatSSEReg:
1999 case ArgInDoubleSSEReg:
2000 case ArgGSharedVtInReg:
2002 ins->opcode = OP_REGVAR;
2003 ins->dreg = ainfo->reg;
2007 case ArgGSharedVtOnStack:
2008 g_assert (!cfg->arch.omit_fp);
2009 ins->opcode = OP_REGOFFSET;
2010 ins->inst_basereg = cfg->frame_reg;
2011 ins->inst_offset = ainfo->offset + ARGS_OFFSET;
2013 case ArgValuetypeInReg:
2015 case ArgValuetypeAddrInIReg: {
2017 g_assert (!cfg->arch.omit_fp);
2019 MONO_INST_NEW (cfg, indir, 0);
2020 indir->opcode = OP_REGOFFSET;
2021 if (ainfo->pair_storage [0] == ArgInIReg) {
2022 indir->inst_basereg = cfg->frame_reg;
2023 offset = ALIGN_TO (offset, sizeof (gpointer));
2024 offset += (sizeof (gpointer));
2025 indir->inst_offset = - offset;
2028 indir->inst_basereg = cfg->frame_reg;
2029 indir->inst_offset = ainfo->offset + ARGS_OFFSET;
2032 ins->opcode = OP_VTARG_ADDR;
2033 ins->inst_left = indir;
2041 if (!inreg && (ainfo->storage != ArgOnStack) && (ainfo->storage != ArgValuetypeAddrInIReg)) {
2042 ins->opcode = OP_REGOFFSET;
2043 ins->inst_basereg = cfg->frame_reg;
2044 /* These arguments are saved to the stack in the prolog */
2045 offset = ALIGN_TO (offset, sizeof(mgreg_t));
2046 if (cfg->arch.omit_fp) {
2047 ins->inst_offset = offset;
2048 offset += (ainfo->storage == ArgValuetypeInReg) ? ainfo->nregs * sizeof (mgreg_t) : sizeof (mgreg_t);
2049 // Arguments are yet supported by the stack map creation code
2050 //cfg->locals_max_stack_offset = MAX (cfg->locals_max_stack_offset, offset);
2052 offset += (ainfo->storage == ArgValuetypeInReg) ? ainfo->nregs * sizeof (mgreg_t) : sizeof (mgreg_t);
2053 ins->inst_offset = - offset;
2054 //cfg->locals_min_stack_offset = MIN (cfg->locals_min_stack_offset, offset);
2060 cfg->stack_offset = offset;
2064 mono_arch_create_vars (MonoCompile *cfg)
2066 MonoMethodSignature *sig;
2070 sig = mono_method_signature (cfg->method);
2072 if (!cfg->arch.cinfo)
2073 cfg->arch.cinfo = get_call_info (cfg->mempool, sig);
2074 cinfo = cfg->arch.cinfo;
2076 if (cinfo->ret.storage == ArgValuetypeInReg)
2077 cfg->ret_var_is_local = TRUE;
2079 sig_ret = mini_get_underlying_type (sig->ret);
2080 if (cinfo->ret.storage == ArgValuetypeAddrInIReg) {
2081 cfg->vret_addr = mono_compile_create_var (cfg, &mono_defaults.int_class->byval_arg, OP_ARG);
2082 if (G_UNLIKELY (cfg->verbose_level > 1)) {
2083 printf ("vret_addr = ");
2084 mono_print_ins (cfg->vret_addr);
2088 if (cfg->gen_sdb_seq_points) {
2091 if (cfg->compile_aot) {
2092 MonoInst *ins = mono_compile_create_var (cfg, &mono_defaults.int_class->byval_arg, OP_LOCAL);
2093 ins->flags |= MONO_INST_VOLATILE;
2094 cfg->arch.seq_point_info_var = ins;
2096 ins = mono_compile_create_var (cfg, &mono_defaults.int_class->byval_arg, OP_LOCAL);
2097 ins->flags |= MONO_INST_VOLATILE;
2098 cfg->arch.ss_tramp_var = ins;
2100 ins = mono_compile_create_var (cfg, &mono_defaults.int_class->byval_arg, OP_LOCAL);
2101 ins->flags |= MONO_INST_VOLATILE;
2102 cfg->arch.bp_tramp_var = ins;
2105 if (cfg->method->save_lmf)
2106 cfg->create_lmf_var = TRUE;
2108 if (cfg->method->save_lmf) {
2110 #if !defined(TARGET_WIN32)
2111 if (mono_get_lmf_tls_offset () != -1 && !optimize_for_xen)
2112 cfg->lmf_ir_mono_lmf = TRUE;
2118 add_outarg_reg (MonoCompile *cfg, MonoCallInst *call, ArgStorage storage, int reg, MonoInst *tree)
2124 MONO_INST_NEW (cfg, ins, OP_MOVE);
2125 ins->dreg = mono_alloc_ireg_copy (cfg, tree->dreg);
2126 ins->sreg1 = tree->dreg;
2127 MONO_ADD_INS (cfg->cbb, ins);
2128 mono_call_inst_add_outarg_reg (cfg, call, ins->dreg, reg, FALSE);
2130 case ArgInFloatSSEReg:
2131 MONO_INST_NEW (cfg, ins, OP_AMD64_SET_XMMREG_R4);
2132 ins->dreg = mono_alloc_freg (cfg);
2133 ins->sreg1 = tree->dreg;
2134 MONO_ADD_INS (cfg->cbb, ins);
2136 mono_call_inst_add_outarg_reg (cfg, call, ins->dreg, reg, TRUE);
2138 case ArgInDoubleSSEReg:
2139 MONO_INST_NEW (cfg, ins, OP_FMOVE);
2140 ins->dreg = mono_alloc_freg (cfg);
2141 ins->sreg1 = tree->dreg;
2142 MONO_ADD_INS (cfg->cbb, ins);
2144 mono_call_inst_add_outarg_reg (cfg, call, ins->dreg, reg, TRUE);
2148 g_assert_not_reached ();
2153 arg_storage_to_load_membase (ArgStorage storage)
2157 #if defined(__mono_ilp32__)
2158 return OP_LOADI8_MEMBASE;
2160 return OP_LOAD_MEMBASE;
2162 case ArgInDoubleSSEReg:
2163 return OP_LOADR8_MEMBASE;
2164 case ArgInFloatSSEReg:
2165 return OP_LOADR4_MEMBASE;
2167 g_assert_not_reached ();
2174 emit_sig_cookie (MonoCompile *cfg, MonoCallInst *call, CallInfo *cinfo)
2176 MonoMethodSignature *tmp_sig;
2179 if (call->tail_call)
2182 g_assert (cinfo->sig_cookie.storage == ArgOnStack);
2185 * mono_ArgIterator_Setup assumes the signature cookie is
2186 * passed first and all the arguments which were before it are
2187 * passed on the stack after the signature. So compensate by
2188 * passing a different signature.
2190 tmp_sig = mono_metadata_signature_dup_full (cfg->method->klass->image, call->signature);
2191 tmp_sig->param_count -= call->signature->sentinelpos;
2192 tmp_sig->sentinelpos = 0;
2193 memcpy (tmp_sig->params, call->signature->params + call->signature->sentinelpos, tmp_sig->param_count * sizeof (MonoType*));
2195 sig_reg = mono_alloc_ireg (cfg);
2196 MONO_EMIT_NEW_SIGNATURECONST (cfg, sig_reg, tmp_sig);
2198 MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STORE_MEMBASE_REG, AMD64_RSP, cinfo->sig_cookie.offset, sig_reg);
2202 static inline LLVMArgStorage
2203 arg_storage_to_llvm_arg_storage (MonoCompile *cfg, ArgStorage storage)
2207 return LLVMArgInIReg;
2210 case ArgGSharedVtInReg:
2211 case ArgGSharedVtOnStack:
2212 return LLVMArgGSharedVt;
2214 g_assert_not_reached ();
2220 mono_arch_get_llvm_call_info (MonoCompile *cfg, MonoMethodSignature *sig)
2226 LLVMCallInfo *linfo;
2227 MonoType *t, *sig_ret;
2229 n = sig->param_count + sig->hasthis;
2230 sig_ret = mini_get_underlying_type (sig->ret);
2232 cinfo = get_call_info (cfg->mempool, sig);
2234 linfo = mono_mempool_alloc0 (cfg->mempool, sizeof (LLVMCallInfo) + (sizeof (LLVMArgInfo) * n));
2237 * LLVM always uses the native ABI while we use our own ABI, the
2238 * only difference is the handling of vtypes:
2239 * - we only pass/receive them in registers in some cases, and only
2240 * in 1 or 2 integer registers.
2242 switch (cinfo->ret.storage) {
2244 linfo->ret.storage = LLVMArgNone;
2247 case ArgInFloatSSEReg:
2248 case ArgInDoubleSSEReg:
2249 linfo->ret.storage = LLVMArgNormal;
2251 case ArgValuetypeInReg:
2253 cfg->exception_message = g_strdup ("pinvoke + vtypes");
2254 cfg->disable_llvm = TRUE;
2258 linfo->ret.storage = LLVMArgVtypeInReg;
2259 for (j = 0; j < 2; ++j)
2260 linfo->ret.pair_storage [j] = arg_storage_to_llvm_arg_storage (cfg, cinfo->ret.pair_storage [j]);
2262 case ArgValuetypeAddrInIReg:
2263 /* Vtype returned using a hidden argument */
2264 linfo->ret.storage = LLVMArgVtypeRetAddr;
2265 linfo->vret_arg_index = cinfo->vret_arg_index;
2268 g_assert_not_reached ();
2272 for (i = 0; i < n; ++i) {
2273 ainfo = cinfo->args + i;
2275 if (i >= sig->hasthis)
2276 t = sig->params [i - sig->hasthis];
2278 t = &mono_defaults.int_class->byval_arg;
2280 linfo->args [i].storage = LLVMArgNone;
2282 switch (ainfo->storage) {
2284 linfo->args [i].storage = LLVMArgNormal;
2286 case ArgInDoubleSSEReg:
2287 case ArgInFloatSSEReg:
2288 linfo->args [i].storage = LLVMArgNormal;
2291 if (MONO_TYPE_ISSTRUCT (t))
2292 linfo->args [i].storage = LLVMArgVtypeByVal;
2294 linfo->args [i].storage = LLVMArgNormal;
2296 case ArgValuetypeInReg:
2298 cfg->exception_message = g_strdup ("pinvoke + vtypes");
2299 cfg->disable_llvm = TRUE;
2303 linfo->args [i].storage = LLVMArgVtypeInReg;
2304 for (j = 0; j < 2; ++j)
2305 linfo->args [i].pair_storage [j] = arg_storage_to_llvm_arg_storage (cfg, ainfo->pair_storage [j]);
2307 case ArgGSharedVtInReg:
2308 case ArgGSharedVtOnStack:
2309 linfo->args [i].storage = LLVMArgGSharedVt;
2312 cfg->exception_message = g_strdup ("ainfo->storage");
2313 cfg->disable_llvm = TRUE;
2323 mono_arch_emit_call (MonoCompile *cfg, MonoCallInst *call)
2326 MonoMethodSignature *sig;
2332 sig = call->signature;
2333 n = sig->param_count + sig->hasthis;
2335 cinfo = get_call_info (cfg->mempool, sig);
2339 if (COMPILE_LLVM (cfg)) {
2340 /* We shouldn't be called in the llvm case */
2341 cfg->disable_llvm = TRUE;
2346 * Emit all arguments which are passed on the stack to prevent register
2347 * allocation problems.
2349 for (i = 0; i < n; ++i) {
2351 ainfo = cinfo->args + i;
2353 in = call->args [i];
2355 if (sig->hasthis && i == 0)
2356 t = &mono_defaults.object_class->byval_arg;
2358 t = sig->params [i - sig->hasthis];
2360 t = mini_get_underlying_type (t);
2361 if (ainfo->storage == ArgOnStack && !MONO_TYPE_ISSTRUCT (t) && !call->tail_call) {
2363 if (t->type == MONO_TYPE_R4)
2364 MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STORER4_MEMBASE_REG, AMD64_RSP, ainfo->offset, in->dreg);
2365 else if (t->type == MONO_TYPE_R8)
2366 MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STORER8_MEMBASE_REG, AMD64_RSP, ainfo->offset, in->dreg);
2368 MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STORE_MEMBASE_REG, AMD64_RSP, ainfo->offset, in->dreg);
2370 MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STORE_MEMBASE_REG, AMD64_RSP, ainfo->offset, in->dreg);
2372 if (cfg->compute_gc_maps) {
2375 EMIT_NEW_GC_PARAM_SLOT_LIVENESS_DEF (cfg, def, ainfo->offset, t);
2381 * Emit all parameters passed in registers in non-reverse order for better readability
2382 * and to help the optimization in emit_prolog ().
2384 for (i = 0; i < n; ++i) {
2385 ainfo = cinfo->args + i;
2387 in = call->args [i];
2389 if (ainfo->storage == ArgInIReg)
2390 add_outarg_reg (cfg, call, ainfo->storage, ainfo->reg, in);
2393 for (i = n - 1; i >= 0; --i) {
2396 ainfo = cinfo->args + i;
2398 in = call->args [i];
2400 if (sig->hasthis && i == 0)
2401 t = &mono_defaults.object_class->byval_arg;
2403 t = sig->params [i - sig->hasthis];
2404 t = mini_get_underlying_type (t);
2406 switch (ainfo->storage) {
2410 case ArgInFloatSSEReg:
2411 case ArgInDoubleSSEReg:
2412 add_outarg_reg (cfg, call, ainfo->storage, ainfo->reg, in);
2415 case ArgValuetypeInReg:
2416 case ArgValuetypeAddrInIReg:
2417 case ArgGSharedVtInReg:
2418 case ArgGSharedVtOnStack: {
2419 if (ainfo->storage == ArgOnStack && !MONO_TYPE_ISSTRUCT (t) && !call->tail_call)
2420 /* Already emitted above */
2422 if (ainfo->storage == ArgOnStack && call->tail_call) {
2423 MonoInst *call_inst = (MonoInst*)call;
2424 cfg->args [i]->flags |= MONO_INST_VOLATILE;
2425 EMIT_NEW_ARGSTORE (cfg, call_inst, i, in);
2433 size = mono_type_native_stack_size (t, &align);
2436 * Other backends use mono_type_stack_size (), but that
2437 * aligns the size to 8, which is larger than the size of
2438 * the source, leading to reads of invalid memory if the
2439 * source is at the end of address space.
2441 size = mono_class_value_size (mono_class_from_mono_type (t), &align);
2444 if (size >= 10000) {
2445 /* Avoid asserts in emit_memcpy () */
2446 cfg->exception_type = MONO_EXCEPTION_INVALID_PROGRAM;
2447 cfg->exception_message = g_strdup_printf ("Passing an argument of size '%d'.", size);
2448 /* Continue normally */
2452 MONO_INST_NEW (cfg, arg, OP_OUTARG_VT);
2453 arg->sreg1 = in->dreg;
2454 arg->klass = mono_class_from_mono_type (t);
2455 arg->backend.size = size;
2456 arg->inst_p0 = call;
2457 arg->inst_p1 = mono_mempool_alloc (cfg->mempool, sizeof (ArgInfo));
2458 memcpy (arg->inst_p1, ainfo, sizeof (ArgInfo));
2460 MONO_ADD_INS (cfg->cbb, arg);
2465 g_assert_not_reached ();
2468 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG) && (i == sig->sentinelpos))
2469 /* Emit the signature cookie just before the implicit arguments */
2470 emit_sig_cookie (cfg, call, cinfo);
2473 /* Handle the case where there are no implicit arguments */
2474 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG) && (n == sig->sentinelpos))
2475 emit_sig_cookie (cfg, call, cinfo);
2477 switch (cinfo->ret.storage) {
2478 case ArgValuetypeInReg:
2479 if (cinfo->ret.pair_storage [0] == ArgInIReg && cinfo->ret.pair_storage [1] == ArgNone) {
2481 * Tell the JIT to use a more efficient calling convention: call using
2482 * OP_CALL, compute the result location after the call, and save the
2485 call->vret_in_reg = TRUE;
2487 * Nullify the instruction computing the vret addr to enable
2488 * future optimizations.
2491 NULLIFY_INS (call->vret_var);
2493 if (call->tail_call)
2496 * The valuetype is in RAX:RDX after the call, need to be copied to
2497 * the stack. Push the address here, so the call instruction can
2500 if (!cfg->arch.vret_addr_loc) {
2501 cfg->arch.vret_addr_loc = mono_compile_create_var (cfg, &mono_defaults.int_class->byval_arg, OP_LOCAL);
2502 /* Prevent it from being register allocated or optimized away */
2503 ((MonoInst*)cfg->arch.vret_addr_loc)->flags |= MONO_INST_VOLATILE;
2506 MONO_EMIT_NEW_UNALU (cfg, OP_MOVE, ((MonoInst*)cfg->arch.vret_addr_loc)->dreg, call->vret_var->dreg);
2509 case ArgValuetypeAddrInIReg: {
2511 MONO_INST_NEW (cfg, vtarg, OP_MOVE);
2512 vtarg->sreg1 = call->vret_var->dreg;
2513 vtarg->dreg = mono_alloc_preg (cfg);
2514 MONO_ADD_INS (cfg->cbb, vtarg);
2516 mono_call_inst_add_outarg_reg (cfg, call, vtarg->dreg, cinfo->ret.reg, FALSE);
2523 if (cfg->method->save_lmf) {
2524 MONO_INST_NEW (cfg, arg, OP_AMD64_SAVE_SP_TO_LMF);
2525 MONO_ADD_INS (cfg->cbb, arg);
2528 call->stack_usage = cinfo->stack_usage;
2532 mono_arch_emit_outarg_vt (MonoCompile *cfg, MonoInst *ins, MonoInst *src)
2535 MonoCallInst *call = (MonoCallInst*)ins->inst_p0;
2536 ArgInfo *ainfo = (ArgInfo*)ins->inst_p1;
2537 int size = ins->backend.size;
2539 switch (ainfo->storage) {
2540 case ArgValuetypeInReg: {
2544 for (part = 0; part < 2; ++part) {
2545 if (ainfo->pair_storage [part] == ArgNone)
2548 MONO_INST_NEW (cfg, load, arg_storage_to_load_membase (ainfo->pair_storage [part]));
2549 load->inst_basereg = src->dreg;
2550 load->inst_offset = part * sizeof(mgreg_t);
2552 switch (ainfo->pair_storage [part]) {
2554 load->dreg = mono_alloc_ireg (cfg);
2556 case ArgInDoubleSSEReg:
2557 case ArgInFloatSSEReg:
2558 load->dreg = mono_alloc_freg (cfg);
2561 g_assert_not_reached ();
2563 MONO_ADD_INS (cfg->cbb, load);
2565 add_outarg_reg (cfg, call, ainfo->pair_storage [part], ainfo->pair_regs [part], load);
2569 case ArgValuetypeAddrInIReg: {
2570 MonoInst *vtaddr, *load;
2571 vtaddr = mono_compile_create_var (cfg, &ins->klass->byval_arg, OP_LOCAL);
2573 MONO_INST_NEW (cfg, load, OP_LDADDR);
2574 cfg->has_indirection = TRUE;
2575 load->inst_p0 = vtaddr;
2576 vtaddr->flags |= MONO_INST_INDIRECT;
2577 load->type = STACK_MP;
2578 load->klass = vtaddr->klass;
2579 load->dreg = mono_alloc_ireg (cfg);
2580 MONO_ADD_INS (cfg->cbb, load);
2581 mini_emit_memcpy (cfg, load->dreg, 0, src->dreg, 0, size, 4);
2583 if (ainfo->pair_storage [0] == ArgInIReg) {
2584 MONO_INST_NEW (cfg, arg, OP_X86_LEA_MEMBASE);
2585 arg->dreg = mono_alloc_ireg (cfg);
2586 arg->sreg1 = load->dreg;
2588 MONO_ADD_INS (cfg->cbb, arg);
2589 mono_call_inst_add_outarg_reg (cfg, call, arg->dreg, ainfo->pair_regs [0], FALSE);
2591 MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STORE_MEMBASE_REG, AMD64_RSP, ainfo->offset, load->dreg);
2595 case ArgGSharedVtInReg:
2597 mono_call_inst_add_outarg_reg (cfg, call, src->dreg, ainfo->reg, FALSE);
2599 case ArgGSharedVtOnStack:
2600 g_assert_not_reached ();
2604 int dreg = mono_alloc_ireg (cfg);
2606 MONO_EMIT_NEW_LOAD_MEMBASE (cfg, dreg, src->dreg, 0);
2607 MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STORE_MEMBASE_REG, AMD64_RSP, ainfo->offset, dreg);
2608 } else if (size <= 40) {
2609 mini_emit_memcpy (cfg, AMD64_RSP, ainfo->offset, src->dreg, 0, size, 4);
2611 // FIXME: Code growth
2612 mini_emit_memcpy (cfg, AMD64_RSP, ainfo->offset, src->dreg, 0, size, 4);
2615 if (cfg->compute_gc_maps) {
2617 EMIT_NEW_GC_PARAM_SLOT_LIVENESS_DEF (cfg, def, ainfo->offset, &ins->klass->byval_arg);
2623 mono_arch_emit_setret (MonoCompile *cfg, MonoMethod *method, MonoInst *val)
2625 MonoType *ret = mini_get_underlying_type (mono_method_signature (method)->ret);
2627 if (ret->type == MONO_TYPE_R4) {
2628 if (COMPILE_LLVM (cfg))
2629 MONO_EMIT_NEW_UNALU (cfg, OP_FMOVE, cfg->ret->dreg, val->dreg);
2631 MONO_EMIT_NEW_UNALU (cfg, OP_AMD64_SET_XMMREG_R4, cfg->ret->dreg, val->dreg);
2633 } else if (ret->type == MONO_TYPE_R8) {
2634 MONO_EMIT_NEW_UNALU (cfg, OP_FMOVE, cfg->ret->dreg, val->dreg);
2638 MONO_EMIT_NEW_UNALU (cfg, OP_MOVE, cfg->ret->dreg, val->dreg);
2641 #endif /* DISABLE_JIT */
2643 #define EMIT_COND_BRANCH(ins,cond,sign) \
2644 if (ins->inst_true_bb->native_offset) { \
2645 x86_branch (code, cond, cfg->native_code + ins->inst_true_bb->native_offset, sign); \
2647 mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_BB, ins->inst_true_bb); \
2648 if ((cfg->opt & MONO_OPT_BRANCH) && \
2649 x86_is_imm8 (ins->inst_true_bb->max_offset - offset)) \
2650 x86_branch8 (code, cond, 0, sign); \
2652 x86_branch32 (code, cond, 0, sign); \
2656 MonoMethodSignature *sig;
2661 dyn_call_supported (MonoMethodSignature *sig, CallInfo *cinfo)
2669 switch (cinfo->ret.storage) {
2673 case ArgValuetypeInReg: {
2674 ArgInfo *ainfo = &cinfo->ret;
2676 if (ainfo->pair_storage [0] != ArgNone && ainfo->pair_storage [0] != ArgInIReg)
2678 if (ainfo->pair_storage [1] != ArgNone && ainfo->pair_storage [1] != ArgInIReg)
2686 for (i = 0; i < cinfo->nargs; ++i) {
2687 ArgInfo *ainfo = &cinfo->args [i];
2688 switch (ainfo->storage) {
2691 case ArgValuetypeInReg:
2692 if (ainfo->pair_storage [0] != ArgNone && ainfo->pair_storage [0] != ArgInIReg)
2694 if (ainfo->pair_storage [1] != ArgNone && ainfo->pair_storage [1] != ArgInIReg)
2706 * mono_arch_dyn_call_prepare:
2708 * Return a pointer to an arch-specific structure which contains information
2709 * needed by mono_arch_get_dyn_call_args (). Return NULL if OP_DYN_CALL is not
2710 * supported for SIG.
2711 * This function is equivalent to ffi_prep_cif in libffi.
2714 mono_arch_dyn_call_prepare (MonoMethodSignature *sig)
2716 ArchDynCallInfo *info;
2719 cinfo = get_call_info (NULL, sig);
2721 if (!dyn_call_supported (sig, cinfo)) {
2726 info = g_new0 (ArchDynCallInfo, 1);
2727 // FIXME: Preprocess the info to speed up get_dyn_call_args ().
2729 info->cinfo = cinfo;
2731 return (MonoDynCallInfo*)info;
2735 * mono_arch_dyn_call_free:
2737 * Free a MonoDynCallInfo structure.
2740 mono_arch_dyn_call_free (MonoDynCallInfo *info)
2742 ArchDynCallInfo *ainfo = (ArchDynCallInfo*)info;
2744 g_free (ainfo->cinfo);
2748 #if !defined(__native_client__)
2749 #define PTR_TO_GREG(ptr) (mgreg_t)(ptr)
2750 #define GREG_TO_PTR(greg) (gpointer)(greg)
2752 /* Correctly handle casts to/from 32-bit pointers without compiler warnings */
2753 #define PTR_TO_GREG(ptr) (mgreg_t)(uintptr_t)(ptr)
2754 #define GREG_TO_PTR(greg) (gpointer)(guint32)(greg)
2758 * mono_arch_get_start_dyn_call:
2760 * Convert the arguments ARGS to a format which can be passed to OP_DYN_CALL, and
2761 * store the result into BUF.
2762 * ARGS should be an array of pointers pointing to the arguments.
2763 * RET should point to a memory buffer large enought to hold the result of the
2765 * This function should be as fast as possible, any work which does not depend
2766 * on the actual values of the arguments should be done in
2767 * mono_arch_dyn_call_prepare ().
2768 * start_dyn_call + OP_DYN_CALL + finish_dyn_call is equivalent to ffi_call in
2772 mono_arch_start_dyn_call (MonoDynCallInfo *info, gpointer **args, guint8 *ret, guint8 *buf, int buf_len)
2774 ArchDynCallInfo *dinfo = (ArchDynCallInfo*)info;
2775 DynCallArgs *p = (DynCallArgs*)buf;
2776 int arg_index, greg, i, pindex;
2777 MonoMethodSignature *sig = dinfo->sig;
2779 g_assert (buf_len >= sizeof (DynCallArgs));
2788 if (sig->hasthis || dinfo->cinfo->vret_arg_index == 1) {
2789 p->regs [greg ++] = PTR_TO_GREG(*(args [arg_index ++]));
2794 if (dinfo->cinfo->ret.storage == ArgValuetypeAddrInIReg)
2795 p->regs [greg ++] = PTR_TO_GREG(ret);
2797 for (i = pindex; i < sig->param_count; i++) {
2798 MonoType *t = mini_get_underlying_type (sig->params [i]);
2799 gpointer *arg = args [arg_index ++];
2802 p->regs [greg ++] = PTR_TO_GREG(*(arg));
2807 case MONO_TYPE_STRING:
2808 case MONO_TYPE_CLASS:
2809 case MONO_TYPE_ARRAY:
2810 case MONO_TYPE_SZARRAY:
2811 case MONO_TYPE_OBJECT:
2815 #if !defined(__mono_ilp32__)
2819 g_assert (dinfo->cinfo->args [i + sig->hasthis].reg == param_regs [greg]);
2820 p->regs [greg ++] = PTR_TO_GREG(*(arg));
2822 #if defined(__mono_ilp32__)
2825 g_assert (dinfo->cinfo->args [i + sig->hasthis].reg == param_regs [greg]);
2826 p->regs [greg ++] = *(guint64*)(arg);
2830 p->regs [greg ++] = *(guint8*)(arg);
2833 p->regs [greg ++] = *(gint8*)(arg);
2836 p->regs [greg ++] = *(gint16*)(arg);
2839 p->regs [greg ++] = *(guint16*)(arg);
2842 p->regs [greg ++] = *(gint32*)(arg);
2845 p->regs [greg ++] = *(guint32*)(arg);
2847 case MONO_TYPE_GENERICINST:
2848 if (MONO_TYPE_IS_REFERENCE (t)) {
2849 p->regs [greg ++] = PTR_TO_GREG(*(arg));
2854 case MONO_TYPE_VALUETYPE: {
2855 ArgInfo *ainfo = &dinfo->cinfo->args [i + sig->hasthis];
2857 g_assert (ainfo->storage == ArgValuetypeInReg);
2858 if (ainfo->pair_storage [0] != ArgNone) {
2859 g_assert (ainfo->pair_storage [0] == ArgInIReg);
2860 p->regs [greg ++] = ((mgreg_t*)(arg))[0];
2862 if (ainfo->pair_storage [1] != ArgNone) {
2863 g_assert (ainfo->pair_storage [1] == ArgInIReg);
2864 p->regs [greg ++] = ((mgreg_t*)(arg))[1];
2869 g_assert_not_reached ();
2873 g_assert (greg <= PARAM_REGS);
2877 * mono_arch_finish_dyn_call:
2879 * Store the result of a dyn call into the return value buffer passed to
2880 * start_dyn_call ().
2881 * This function should be as fast as possible, any work which does not depend
2882 * on the actual values of the arguments should be done in
2883 * mono_arch_dyn_call_prepare ().
2886 mono_arch_finish_dyn_call (MonoDynCallInfo *info, guint8 *buf)
2888 ArchDynCallInfo *dinfo = (ArchDynCallInfo*)info;
2889 MonoMethodSignature *sig = dinfo->sig;
2890 guint8 *ret = ((DynCallArgs*)buf)->ret;
2891 mgreg_t res = ((DynCallArgs*)buf)->res;
2892 MonoType *sig_ret = mini_get_underlying_type (sig->ret);
2894 switch (sig_ret->type) {
2895 case MONO_TYPE_VOID:
2896 *(gpointer*)ret = NULL;
2898 case MONO_TYPE_STRING:
2899 case MONO_TYPE_CLASS:
2900 case MONO_TYPE_ARRAY:
2901 case MONO_TYPE_SZARRAY:
2902 case MONO_TYPE_OBJECT:
2906 *(gpointer*)ret = GREG_TO_PTR(res);
2912 *(guint8*)ret = res;
2915 *(gint16*)ret = res;
2918 *(guint16*)ret = res;
2921 *(gint32*)ret = res;
2924 *(guint32*)ret = res;
2927 *(gint64*)ret = res;
2930 *(guint64*)ret = res;
2932 case MONO_TYPE_GENERICINST:
2933 if (MONO_TYPE_IS_REFERENCE (sig_ret)) {
2934 *(gpointer*)ret = GREG_TO_PTR(res);
2939 case MONO_TYPE_VALUETYPE:
2940 if (dinfo->cinfo->ret.storage == ArgValuetypeAddrInIReg) {
2943 ArgInfo *ainfo = &dinfo->cinfo->ret;
2945 g_assert (ainfo->storage == ArgValuetypeInReg);
2947 if (ainfo->pair_storage [0] != ArgNone) {
2948 g_assert (ainfo->pair_storage [0] == ArgInIReg);
2949 ((mgreg_t*)ret)[0] = res;
2952 g_assert (ainfo->pair_storage [1] == ArgNone);
2956 g_assert_not_reached ();
2960 /* emit an exception if condition is fail */
2961 #define EMIT_COND_SYSTEM_EXCEPTION(cond,signed,exc_name) \
2963 MonoInst *tins = mono_branch_optimize_exception_target (cfg, bb, exc_name); \
2964 if (tins == NULL) { \
2965 mono_add_patch_info (cfg, code - cfg->native_code, \
2966 MONO_PATCH_INFO_EXC, exc_name); \
2967 x86_branch32 (code, cond, 0, signed); \
2969 EMIT_COND_BRANCH (tins, cond, signed); \
2973 #define EMIT_FPCOMPARE(code) do { \
2974 amd64_fcompp (code); \
2975 amd64_fnstsw (code); \
2978 #define EMIT_SSE2_FPFUNC(code, op, dreg, sreg1) do { \
2979 amd64_movsd_membase_reg (code, AMD64_RSP, -8, (sreg1)); \
2980 amd64_fld_membase (code, AMD64_RSP, -8, TRUE); \
2981 amd64_ ##op (code); \
2982 amd64_fst_membase (code, AMD64_RSP, -8, TRUE, TRUE); \
2983 amd64_movsd_reg_membase (code, (dreg), AMD64_RSP, -8); \
2987 emit_call_body (MonoCompile *cfg, guint8 *code, guint32 patch_type, gconstpointer data)
2989 gboolean no_patch = FALSE;
2992 * FIXME: Add support for thunks
2995 gboolean near_call = FALSE;
2998 * Indirect calls are expensive so try to make a near call if possible.
2999 * The caller memory is allocated by the code manager so it is
3000 * guaranteed to be at a 32 bit offset.
3003 if (patch_type != MONO_PATCH_INFO_ABS) {
3004 /* The target is in memory allocated using the code manager */
3007 if ((patch_type == MONO_PATCH_INFO_METHOD) || (patch_type == MONO_PATCH_INFO_METHOD_JUMP)) {
3008 if (((MonoMethod*)data)->klass->image->aot_module)
3009 /* The callee might be an AOT method */
3011 if (((MonoMethod*)data)->dynamic)
3012 /* The target is in malloc-ed memory */
3016 if (patch_type == MONO_PATCH_INFO_INTERNAL_METHOD) {
3018 * The call might go directly to a native function without
3021 MonoJitICallInfo *mi = mono_find_jit_icall_by_name (data);
3023 gconstpointer target = mono_icall_get_wrapper (mi);
3024 if ((((guint64)target) >> 32) != 0)
3030 MonoJumpInfo *jinfo = NULL;
3032 if (cfg->abs_patches)
3033 jinfo = g_hash_table_lookup (cfg->abs_patches, data);
3035 if (jinfo->type == MONO_PATCH_INFO_JIT_ICALL_ADDR) {
3036 MonoJitICallInfo *mi = mono_find_jit_icall_by_name (jinfo->data.name);
3037 if (mi && (((guint64)mi->func) >> 32) == 0)
3042 * This is not really an optimization, but required because the
3043 * generic class init trampolines use R11 to pass the vtable.
3048 MonoJitICallInfo *info = mono_find_jit_icall_by_addr (data);
3050 if (info->func == info->wrapper) {
3052 if ((((guint64)info->func) >> 32) == 0)
3056 /* See the comment in mono_codegen () */
3057 if ((info->name [0] != 'v') || (strstr (info->name, "ves_array_new_va_") == NULL && strstr (info->name, "ves_array_element_address_") == NULL))
3061 else if ((((guint64)data) >> 32) == 0) {
3068 if (cfg->method->dynamic)
3069 /* These methods are allocated using malloc */
3072 #ifdef MONO_ARCH_NOMAP32BIT
3075 #if defined(__native_client__)
3076 /* Always use near_call == TRUE for Native Client */
3079 /* The 64bit XEN kernel does not honour the MAP_32BIT flag. (#522894) */
3080 if (optimize_for_xen)
3083 if (cfg->compile_aot) {
3090 * Align the call displacement to an address divisible by 4 so it does
3091 * not span cache lines. This is required for code patching to work on SMP
3094 if (!no_patch && ((guint32)(code + 1 - cfg->native_code) % 4) != 0) {
3095 guint32 pad_size = 4 - ((guint32)(code + 1 - cfg->native_code) % 4);
3096 amd64_padding (code, pad_size);
3098 mono_add_patch_info (cfg, code - cfg->native_code, patch_type, data);
3099 amd64_call_code (code, 0);
3102 mono_add_patch_info (cfg, code - cfg->native_code, patch_type, data);
3103 amd64_set_reg_template (code, GP_SCRATCH_REG);
3104 amd64_call_reg (code, GP_SCRATCH_REG);
3111 static inline guint8*
3112 emit_call (MonoCompile *cfg, guint8 *code, guint32 patch_type, gconstpointer data, gboolean win64_adjust_stack)
3115 if (win64_adjust_stack)
3116 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, 32);
3118 code = emit_call_body (cfg, code, patch_type, data);
3120 if (win64_adjust_stack)
3121 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, 32);
3128 store_membase_imm_to_store_membase_reg (int opcode)
3131 case OP_STORE_MEMBASE_IMM:
3132 return OP_STORE_MEMBASE_REG;
3133 case OP_STOREI4_MEMBASE_IMM:
3134 return OP_STOREI4_MEMBASE_REG;
3135 case OP_STOREI8_MEMBASE_IMM:
3136 return OP_STOREI8_MEMBASE_REG;
3144 #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)))
3147 * mono_arch_peephole_pass_1:
3149 * Perform peephole opts which should/can be performed before local regalloc
3152 mono_arch_peephole_pass_1 (MonoCompile *cfg, MonoBasicBlock *bb)
3156 MONO_BB_FOR_EACH_INS_SAFE (bb, n, ins) {
3157 MonoInst *last_ins = mono_inst_prev (ins, FILTER_IL_SEQ_POINT);
3159 switch (ins->opcode) {
3163 if ((ins->sreg1 < MONO_MAX_IREGS) && (ins->dreg >= MONO_MAX_IREGS) && (ins->inst_imm > 0)) {
3165 * X86_LEA is like ADD, but doesn't have the
3166 * sreg1==dreg restriction. inst_imm > 0 is needed since LEA sign-extends
3167 * its operand to 64 bit.
3169 ins->opcode = OP_X86_LEA_MEMBASE;
3170 ins->inst_basereg = ins->sreg1;
3175 if ((ins->sreg1 == ins->sreg2) && (ins->sreg1 == ins->dreg)) {
3179 * Replace STORE_MEMBASE_IMM 0 with STORE_MEMBASE_REG since
3180 * the latter has length 2-3 instead of 6 (reverse constant
3181 * propagation). These instruction sequences are very common
3182 * in the initlocals bblock.
3184 for (ins2 = ins->next; ins2; ins2 = ins2->next) {
3185 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)) {
3186 ins2->opcode = store_membase_imm_to_store_membase_reg (ins2->opcode);
3187 ins2->sreg1 = ins->dreg;
3188 } 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)) {
3190 } else if (((ins2->opcode == OP_ICONST) || (ins2->opcode == OP_I8CONST)) && (ins2->dreg == ins->dreg) && (ins2->inst_c0 == 0)) {
3193 } else if (ins2->opcode == OP_IL_SEQ_POINT) {
3201 case OP_COMPARE_IMM:
3202 case OP_LCOMPARE_IMM:
3203 /* OP_COMPARE_IMM (reg, 0)
3205 * OP_AMD64_TEST_NULL (reg)
3208 ins->opcode = OP_AMD64_TEST_NULL;
3210 case OP_ICOMPARE_IMM:
3212 ins->opcode = OP_X86_TEST_NULL;
3214 case OP_AMD64_ICOMPARE_MEMBASE_IMM:
3216 * OP_STORE_MEMBASE_REG reg, offset(basereg)
3217 * OP_X86_COMPARE_MEMBASE_IMM offset(basereg), imm
3219 * OP_STORE_MEMBASE_REG reg, offset(basereg)
3220 * OP_COMPARE_IMM reg, imm
3222 * Note: if imm = 0 then OP_COMPARE_IMM replaced with OP_X86_TEST_NULL
3224 if (last_ins && (last_ins->opcode == OP_STOREI4_MEMBASE_REG) &&
3225 ins->inst_basereg == last_ins->inst_destbasereg &&
3226 ins->inst_offset == last_ins->inst_offset) {
3227 ins->opcode = OP_ICOMPARE_IMM;
3228 ins->sreg1 = last_ins->sreg1;
3230 /* check if we can remove cmp reg,0 with test null */
3232 ins->opcode = OP_X86_TEST_NULL;
3238 mono_peephole_ins (bb, ins);
3243 mono_arch_peephole_pass_2 (MonoCompile *cfg, MonoBasicBlock *bb)
3247 MONO_BB_FOR_EACH_INS_SAFE (bb, n, ins) {
3248 switch (ins->opcode) {
3251 MonoInst *next = mono_inst_next (ins, FILTER_IL_SEQ_POINT);
3252 /* reg = 0 -> XOR (reg, reg) */
3253 /* XOR sets cflags on x86, so we cant do it always */
3254 if (ins->inst_c0 == 0 && (!next || (next && INST_IGNORES_CFLAGS (next->opcode)))) {
3255 ins->opcode = OP_LXOR;
3256 ins->sreg1 = ins->dreg;
3257 ins->sreg2 = ins->dreg;
3265 * Use IXOR to avoid a rex prefix if possible. The cpu will sign extend the
3266 * 0 result into 64 bits.
3268 if ((ins->sreg1 == ins->sreg2) && (ins->sreg1 == ins->dreg)) {
3269 ins->opcode = OP_IXOR;
3273 if ((ins->sreg1 == ins->sreg2) && (ins->sreg1 == ins->dreg)) {
3277 * Replace STORE_MEMBASE_IMM 0 with STORE_MEMBASE_REG since
3278 * the latter has length 2-3 instead of 6 (reverse constant
3279 * propagation). These instruction sequences are very common
3280 * in the initlocals bblock.
3282 for (ins2 = ins->next; ins2; ins2 = ins2->next) {
3283 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)) {
3284 ins2->opcode = store_membase_imm_to_store_membase_reg (ins2->opcode);
3285 ins2->sreg1 = ins->dreg;
3286 } 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)) {
3288 } else if (((ins2->opcode == OP_ICONST) || (ins2->opcode == OP_I8CONST)) && (ins2->dreg == ins->dreg) && (ins2->inst_c0 == 0)) {
3291 } else if (ins2->opcode == OP_IL_SEQ_POINT) {
3300 if ((ins->inst_imm == 1) && (ins->dreg == ins->sreg1))
3301 ins->opcode = OP_X86_INC_REG;
3304 if ((ins->inst_imm == 1) && (ins->dreg == ins->sreg1))
3305 ins->opcode = OP_X86_DEC_REG;
3309 mono_peephole_ins (bb, ins);
3313 #define NEW_INS(cfg,ins,dest,op) do { \
3314 MONO_INST_NEW ((cfg), (dest), (op)); \
3315 (dest)->cil_code = (ins)->cil_code; \
3316 mono_bblock_insert_before_ins (bb, ins, (dest)); \
3320 * mono_arch_lowering_pass:
3322 * Converts complex opcodes into simpler ones so that each IR instruction
3323 * corresponds to one machine instruction.
3326 mono_arch_lowering_pass (MonoCompile *cfg, MonoBasicBlock *bb)
3328 MonoInst *ins, *n, *temp;
3331 * FIXME: Need to add more instructions, but the current machine
3332 * description can't model some parts of the composite instructions like
3335 MONO_BB_FOR_EACH_INS_SAFE (bb, n, ins) {
3336 switch (ins->opcode) {
3340 case OP_IDIV_UN_IMM:
3341 case OP_IREM_UN_IMM:
3344 mono_decompose_op_imm (cfg, bb, ins);
3346 case OP_COMPARE_IMM:
3347 case OP_LCOMPARE_IMM:
3348 if (!amd64_use_imm32 (ins->inst_imm)) {
3349 NEW_INS (cfg, ins, temp, OP_I8CONST);
3350 temp->inst_c0 = ins->inst_imm;
3351 temp->dreg = mono_alloc_ireg (cfg);
3352 ins->opcode = OP_COMPARE;
3353 ins->sreg2 = temp->dreg;
3356 #ifndef __mono_ilp32__
3357 case OP_LOAD_MEMBASE:
3359 case OP_LOADI8_MEMBASE:
3360 #ifndef __native_client_codegen__
3361 /* Don't generate memindex opcodes (to simplify */
3362 /* read sandboxing) */
3363 if (!amd64_use_imm32 (ins->inst_offset)) {
3364 NEW_INS (cfg, ins, temp, OP_I8CONST);
3365 temp->inst_c0 = ins->inst_offset;
3366 temp->dreg = mono_alloc_ireg (cfg);
3367 ins->opcode = OP_AMD64_LOADI8_MEMINDEX;
3368 ins->inst_indexreg = temp->dreg;
3372 #ifndef __mono_ilp32__
3373 case OP_STORE_MEMBASE_IMM:
3375 case OP_STOREI8_MEMBASE_IMM:
3376 if (!amd64_use_imm32 (ins->inst_imm)) {
3377 NEW_INS (cfg, ins, temp, OP_I8CONST);
3378 temp->inst_c0 = ins->inst_imm;
3379 temp->dreg = mono_alloc_ireg (cfg);
3380 ins->opcode = OP_STOREI8_MEMBASE_REG;
3381 ins->sreg1 = temp->dreg;
3384 #ifdef MONO_ARCH_SIMD_INTRINSICS
3385 case OP_EXPAND_I1: {
3386 int temp_reg1 = mono_alloc_ireg (cfg);
3387 int temp_reg2 = mono_alloc_ireg (cfg);
3388 int original_reg = ins->sreg1;
3390 NEW_INS (cfg, ins, temp, OP_ICONV_TO_U1);
3391 temp->sreg1 = original_reg;
3392 temp->dreg = temp_reg1;
3394 NEW_INS (cfg, ins, temp, OP_SHL_IMM);
3395 temp->sreg1 = temp_reg1;
3396 temp->dreg = temp_reg2;
3399 NEW_INS (cfg, ins, temp, OP_LOR);
3400 temp->sreg1 = temp->dreg = temp_reg2;
3401 temp->sreg2 = temp_reg1;
3403 ins->opcode = OP_EXPAND_I2;
3404 ins->sreg1 = temp_reg2;
3413 bb->max_vreg = cfg->next_vreg;
3417 branch_cc_table [] = {
3418 X86_CC_EQ, X86_CC_GE, X86_CC_GT, X86_CC_LE, X86_CC_LT,
3419 X86_CC_NE, X86_CC_GE, X86_CC_GT, X86_CC_LE, X86_CC_LT,
3420 X86_CC_O, X86_CC_NO, X86_CC_C, X86_CC_NC
3423 /* Maps CMP_... constants to X86_CC_... constants */
3426 X86_CC_EQ, X86_CC_NE, X86_CC_LE, X86_CC_GE, X86_CC_LT, X86_CC_GT,
3427 X86_CC_LE, X86_CC_GE, X86_CC_LT, X86_CC_GT
3431 cc_signed_table [] = {
3432 TRUE, TRUE, TRUE, TRUE, TRUE, TRUE,
3433 FALSE, FALSE, FALSE, FALSE
3436 /*#include "cprop.c"*/
3438 static unsigned char*
3439 emit_float_to_int (MonoCompile *cfg, guchar *code, int dreg, int sreg, int size, gboolean is_signed)
3442 amd64_sse_cvttsd2si_reg_reg (code, dreg, sreg);
3444 amd64_sse_cvttsd2si_reg_reg_size (code, dreg, sreg, 4);
3447 amd64_widen_reg (code, dreg, dreg, is_signed, FALSE);
3449 amd64_widen_reg (code, dreg, dreg, is_signed, TRUE);
3453 static unsigned char*
3454 mono_emit_stack_alloc (MonoCompile *cfg, guchar *code, MonoInst* tree)
3456 int sreg = tree->sreg1;
3457 int need_touch = FALSE;
3459 #if defined(TARGET_WIN32)
3461 #elif defined(MONO_ARCH_SIGSEGV_ON_ALTSTACK)
3462 if (!tree->flags & MONO_INST_INIT)
3471 * If requested stack size is larger than one page,
3472 * perform stack-touch operation
3475 * Generate stack probe code.
3476 * Under Windows, it is necessary to allocate one page at a time,
3477 * "touching" stack after each successful sub-allocation. This is
3478 * because of the way stack growth is implemented - there is a
3479 * guard page before the lowest stack page that is currently commited.
3480 * Stack normally grows sequentially so OS traps access to the
3481 * guard page and commits more pages when needed.
3483 amd64_test_reg_imm (code, sreg, ~0xFFF);
3484 br[0] = code; x86_branch8 (code, X86_CC_Z, 0, FALSE);
3486 br[2] = code; /* loop */
3487 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, 0x1000);
3488 amd64_test_membase_reg (code, AMD64_RSP, 0, AMD64_RSP);
3489 amd64_alu_reg_imm (code, X86_SUB, sreg, 0x1000);
3490 amd64_alu_reg_imm (code, X86_CMP, sreg, 0x1000);
3491 br[3] = code; x86_branch8 (code, X86_CC_AE, 0, FALSE);
3492 amd64_patch (br[3], br[2]);
3493 amd64_test_reg_reg (code, sreg, sreg);
3494 br[4] = code; x86_branch8 (code, X86_CC_Z, 0, FALSE);
3495 amd64_alu_reg_reg (code, X86_SUB, AMD64_RSP, sreg);
3497 br[1] = code; x86_jump8 (code, 0);
3499 amd64_patch (br[0], code);
3500 amd64_alu_reg_reg (code, X86_SUB, AMD64_RSP, sreg);
3501 amd64_patch (br[1], code);
3502 amd64_patch (br[4], code);
3505 amd64_alu_reg_reg (code, X86_SUB, AMD64_RSP, tree->sreg1);
3507 if (tree->flags & MONO_INST_INIT) {
3509 if (tree->dreg != AMD64_RAX && sreg != AMD64_RAX) {
3510 amd64_push_reg (code, AMD64_RAX);
3513 if (tree->dreg != AMD64_RCX && sreg != AMD64_RCX) {
3514 amd64_push_reg (code, AMD64_RCX);
3517 if (tree->dreg != AMD64_RDI && sreg != AMD64_RDI) {
3518 amd64_push_reg (code, AMD64_RDI);
3522 amd64_shift_reg_imm (code, X86_SHR, sreg, 3);
3523 if (sreg != AMD64_RCX)
3524 amd64_mov_reg_reg (code, AMD64_RCX, sreg, 8);
3525 amd64_alu_reg_reg (code, X86_XOR, AMD64_RAX, AMD64_RAX);
3527 amd64_lea_membase (code, AMD64_RDI, AMD64_RSP, offset);
3528 if (cfg->param_area)
3529 amd64_alu_reg_imm (code, X86_ADD, AMD64_RDI, cfg->param_area);
3531 #if defined(__default_codegen__)
3532 amd64_prefix (code, X86_REP_PREFIX);
3534 #elif defined(__native_client_codegen__)
3535 /* NaCl stos pseudo-instruction */
3536 amd64_codegen_pre(code);
3537 /* First, clear the upper 32 bits of RDI (mov %edi, %edi) */
3538 amd64_mov_reg_reg (code, AMD64_RDI, AMD64_RDI, 4);
3539 /* Add %r15 to %rdi using lea, condition flags unaffected. */
3540 amd64_lea_memindex_size (code, AMD64_RDI, AMD64_R15, 0, AMD64_RDI, 0, 8);
3541 amd64_prefix (code, X86_REP_PREFIX);
3543 amd64_codegen_post(code);
3544 #endif /* __native_client_codegen__ */
3546 if (tree->dreg != AMD64_RDI && sreg != AMD64_RDI)
3547 amd64_pop_reg (code, AMD64_RDI);
3548 if (tree->dreg != AMD64_RCX && sreg != AMD64_RCX)
3549 amd64_pop_reg (code, AMD64_RCX);
3550 if (tree->dreg != AMD64_RAX && sreg != AMD64_RAX)
3551 amd64_pop_reg (code, AMD64_RAX);
3557 emit_move_return_value (MonoCompile *cfg, MonoInst *ins, guint8 *code)
3562 /* Move return value to the target register */
3563 /* FIXME: do this in the local reg allocator */
3564 switch (ins->opcode) {
3567 case OP_CALL_MEMBASE:
3570 case OP_LCALL_MEMBASE:
3571 g_assert (ins->dreg == AMD64_RAX);
3575 case OP_FCALL_MEMBASE: {
3576 MonoType *rtype = mini_get_underlying_type (((MonoCallInst*)ins)->signature->ret);
3577 if (rtype->type == MONO_TYPE_R4) {
3578 amd64_sse_cvtss2sd_reg_reg (code, ins->dreg, AMD64_XMM0);
3581 if (ins->dreg != AMD64_XMM0)
3582 amd64_sse_movsd_reg_reg (code, ins->dreg, AMD64_XMM0);
3588 case OP_RCALL_MEMBASE:
3589 if (ins->dreg != AMD64_XMM0)
3590 amd64_sse_movss_reg_reg (code, ins->dreg, AMD64_XMM0);
3594 case OP_VCALL_MEMBASE:
3597 case OP_VCALL2_MEMBASE:
3598 cinfo = get_call_info (cfg->mempool, ((MonoCallInst*)ins)->signature);
3599 if (cinfo->ret.storage == ArgValuetypeInReg) {
3600 MonoInst *loc = cfg->arch.vret_addr_loc;
3602 /* Load the destination address */
3603 g_assert (loc->opcode == OP_REGOFFSET);
3604 amd64_mov_reg_membase (code, AMD64_RCX, loc->inst_basereg, loc->inst_offset, sizeof(gpointer));
3606 for (quad = 0; quad < 2; quad ++) {
3607 switch (cinfo->ret.pair_storage [quad]) {
3609 amd64_mov_membase_reg (code, AMD64_RCX, (quad * sizeof(mgreg_t)), cinfo->ret.pair_regs [quad], sizeof(mgreg_t));
3611 case ArgInFloatSSEReg:
3612 amd64_movss_membase_reg (code, AMD64_RCX, (quad * 8), cinfo->ret.pair_regs [quad]);
3614 case ArgInDoubleSSEReg:
3615 amd64_movsd_membase_reg (code, AMD64_RCX, (quad * 8), cinfo->ret.pair_regs [quad]);
3630 #endif /* DISABLE_JIT */
3633 static int tls_gs_offset;
3637 mono_amd64_have_tls_get (void)
3640 static gboolean have_tls_get = FALSE;
3641 static gboolean inited = FALSE;
3644 return have_tls_get;
3646 #if MONO_HAVE_FAST_TLS
3647 guint8 *ins = (guint8*)pthread_getspecific;
3650 * We're looking for these two instructions:
3652 * mov %gs:[offset](,%rdi,8),%rax
3655 have_tls_get = ins [0] == 0x65 &&
3665 tls_gs_offset = ins[5];
3668 * Apple now loads a different version of pthread_getspecific when launched from Xcode
3669 * For that version we're looking for these instructions:
3673 * mov %gs:[offset](,%rdi,8),%rax
3677 if (!have_tls_get) {
3678 have_tls_get = ins [0] == 0x55 &&
3693 tls_gs_offset = ins[9];
3699 return have_tls_get;
3700 #elif defined(TARGET_ANDROID)
3708 mono_amd64_get_tls_gs_offset (void)
3711 return tls_gs_offset;
3713 g_assert_not_reached ();
3719 * mono_amd64_emit_tls_get:
3720 * @code: buffer to store code to
3721 * @dreg: hard register where to place the result
3722 * @tls_offset: offset info
3724 * mono_amd64_emit_tls_get emits in @code the native code that puts in
3725 * the dreg register the item in the thread local storage identified
3728 * Returns: a pointer to the end of the stored code
3731 mono_amd64_emit_tls_get (guint8* code, int dreg, int tls_offset)
3734 if (tls_offset < 64) {
3735 x86_prefix (code, X86_GS_PREFIX);
3736 amd64_mov_reg_mem (code, dreg, (tls_offset * 8) + 0x1480, 8);
3740 g_assert (tls_offset < 0x440);
3741 /* Load TEB->TlsExpansionSlots */
3742 x86_prefix (code, X86_GS_PREFIX);
3743 amd64_mov_reg_mem (code, dreg, 0x1780, 8);
3744 amd64_test_reg_reg (code, dreg, dreg);
3746 amd64_branch (code, X86_CC_EQ, code, TRUE);
3747 amd64_mov_reg_membase (code, dreg, dreg, (tls_offset * 8) - 0x200, 8);
3748 amd64_patch (buf [0], code);
3750 #elif defined(__APPLE__)
3751 x86_prefix (code, X86_GS_PREFIX);
3752 amd64_mov_reg_mem (code, dreg, tls_gs_offset + (tls_offset * 8), 8);
3754 if (optimize_for_xen) {
3755 x86_prefix (code, X86_FS_PREFIX);
3756 amd64_mov_reg_mem (code, dreg, 0, 8);
3757 amd64_mov_reg_membase (code, dreg, dreg, tls_offset, 8);
3759 x86_prefix (code, X86_FS_PREFIX);
3760 amd64_mov_reg_mem (code, dreg, tls_offset, 8);
3767 emit_tls_get_reg (guint8* code, int dreg, int offset_reg)
3769 /* offset_reg contains a value translated by mono_arch_translate_tls_offset () */
3771 if (dreg != offset_reg)
3772 amd64_mov_reg_reg (code, dreg, offset_reg, sizeof (mgreg_t));
3773 amd64_prefix (code, X86_GS_PREFIX);
3774 amd64_mov_reg_membase (code, dreg, dreg, 0, sizeof (mgreg_t));
3775 #elif defined(__linux__)
3778 if (dreg == offset_reg) {
3779 /* Use a temporary reg by saving it to the redzone */
3780 tmpreg = dreg == AMD64_RAX ? AMD64_RCX : AMD64_RAX;
3781 amd64_mov_membase_reg (code, AMD64_RSP, -8, tmpreg, 8);
3782 amd64_mov_reg_reg (code, tmpreg, offset_reg, sizeof (gpointer));
3783 offset_reg = tmpreg;
3785 x86_prefix (code, X86_FS_PREFIX);
3786 amd64_mov_reg_mem (code, dreg, 0, 8);
3787 amd64_mov_reg_memindex (code, dreg, dreg, 0, offset_reg, 0, 8);
3789 amd64_mov_reg_membase (code, tmpreg, AMD64_RSP, -8, 8);
3791 g_assert_not_reached ();
3797 amd64_emit_tls_set (guint8 *code, int sreg, int tls_offset)
3800 g_assert_not_reached ();
3801 #elif defined(__APPLE__)
3802 x86_prefix (code, X86_GS_PREFIX);
3803 amd64_mov_mem_reg (code, tls_gs_offset + (tls_offset * 8), sreg, 8);
3805 g_assert (!optimize_for_xen);
3806 x86_prefix (code, X86_FS_PREFIX);
3807 amd64_mov_mem_reg (code, tls_offset, sreg, 8);
3813 amd64_emit_tls_set_reg (guint8 *code, int sreg, int offset_reg)
3815 /* offset_reg contains a value translated by mono_arch_translate_tls_offset () */
3817 g_assert_not_reached ();
3818 #elif defined(__APPLE__)
3819 x86_prefix (code, X86_GS_PREFIX);
3820 amd64_mov_membase_reg (code, offset_reg, 0, sreg, 8);
3822 x86_prefix (code, X86_FS_PREFIX);
3823 amd64_mov_membase_reg (code, offset_reg, 0, sreg, 8);
3829 * mono_arch_translate_tls_offset:
3831 * Translate the TLS offset OFFSET computed by MONO_THREAD_VAR_OFFSET () into a format usable by OP_TLS_GET_REG/OP_TLS_SET_REG.
3834 mono_arch_translate_tls_offset (int offset)
3837 return tls_gs_offset + (offset * 8);
3846 * Emit code to initialize an LMF structure at LMF_OFFSET.
3849 emit_setup_lmf (MonoCompile *cfg, guint8 *code, gint32 lmf_offset, int cfa_offset)
3852 * The ip field is not set, the exception handling code will obtain it from the stack location pointed to by the sp field.
3855 * sp is saved right before calls but we need to save it here too so
3856 * async stack walks would work.
3858 amd64_mov_membase_reg (code, cfg->frame_reg, lmf_offset + MONO_STRUCT_OFFSET (MonoLMF, rsp), AMD64_RSP, 8);
3860 amd64_mov_membase_reg (code, cfg->frame_reg, lmf_offset + MONO_STRUCT_OFFSET (MonoLMF, rbp), AMD64_RBP, 8);
3861 if (cfg->arch.omit_fp && cfa_offset != -1)
3862 mono_emit_unwind_op_offset (cfg, code, AMD64_RBP, - (cfa_offset - (lmf_offset + MONO_STRUCT_OFFSET (MonoLMF, rbp))));
3864 /* These can't contain refs */
3865 mini_gc_set_slot_type_from_fp (cfg, lmf_offset + MONO_STRUCT_OFFSET (MonoLMF, previous_lmf), SLOT_NOREF);
3866 mini_gc_set_slot_type_from_fp (cfg, lmf_offset + MONO_STRUCT_OFFSET (MonoLMF, rip), SLOT_NOREF);
3867 mini_gc_set_slot_type_from_fp (cfg, lmf_offset + MONO_STRUCT_OFFSET (MonoLMF, rsp), SLOT_NOREF);
3868 /* These are handled automatically by the stack marking code */
3869 mini_gc_set_slot_type_from_fp (cfg, lmf_offset + MONO_STRUCT_OFFSET (MonoLMF, rbp), SLOT_NOREF);
3874 #define REAL_PRINT_REG(text,reg) \
3875 mono_assert (reg >= 0); \
3876 amd64_push_reg (code, AMD64_RAX); \
3877 amd64_push_reg (code, AMD64_RDX); \
3878 amd64_push_reg (code, AMD64_RCX); \
3879 amd64_push_reg (code, reg); \
3880 amd64_push_imm (code, reg); \
3881 amd64_push_imm (code, text " %d %p\n"); \
3882 amd64_mov_reg_imm (code, AMD64_RAX, printf); \
3883 amd64_call_reg (code, AMD64_RAX); \
3884 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, 3*4); \
3885 amd64_pop_reg (code, AMD64_RCX); \
3886 amd64_pop_reg (code, AMD64_RDX); \
3887 amd64_pop_reg (code, AMD64_RAX);
3889 /* benchmark and set based on cpu */
3890 #define LOOP_ALIGNMENT 8
3891 #define bb_is_loop_start(bb) ((bb)->loop_body_start && (bb)->nesting)
3895 mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
3900 guint8 *code = cfg->native_code + cfg->code_len;
3903 /* Fix max_offset estimate for each successor bb */
3904 if (cfg->opt & MONO_OPT_BRANCH) {
3905 int current_offset = cfg->code_len;
3906 MonoBasicBlock *current_bb;
3907 for (current_bb = bb; current_bb != NULL; current_bb = current_bb->next_bb) {
3908 current_bb->max_offset = current_offset;
3909 current_offset += current_bb->max_length;
3913 if (cfg->opt & MONO_OPT_LOOP) {
3914 int pad, align = LOOP_ALIGNMENT;
3915 /* set alignment depending on cpu */
3916 if (bb_is_loop_start (bb) && (pad = (cfg->code_len & (align - 1)))) {
3918 /*g_print ("adding %d pad at %x to loop in %s\n", pad, cfg->code_len, cfg->method->name);*/
3919 amd64_padding (code, pad);
3920 cfg->code_len += pad;
3921 bb->native_offset = cfg->code_len;
3925 #if defined(__native_client_codegen__)
3926 /* For Native Client, all indirect call/jump targets must be */
3927 /* 32-byte aligned. Exception handler blocks are jumped to */
3928 /* indirectly as well. */
3929 gboolean bb_needs_alignment = (bb->flags & BB_INDIRECT_JUMP_TARGET) ||
3930 (bb->flags & BB_EXCEPTION_HANDLER);
3932 if ( bb_needs_alignment && ((cfg->code_len & kNaClAlignmentMask) != 0)) {
3933 int pad = kNaClAlignment - (cfg->code_len & kNaClAlignmentMask);
3934 if (pad != kNaClAlignment) code = mono_arch_nacl_pad(code, pad);
3935 cfg->code_len += pad;
3936 bb->native_offset = cfg->code_len;
3938 #endif /*__native_client_codegen__*/
3940 if (cfg->verbose_level > 2)
3941 g_print ("Basic block %d starting at offset 0x%x\n", bb->block_num, bb->native_offset);
3943 if ((cfg->prof_options & MONO_PROFILE_COVERAGE) && cfg->coverage_info) {
3944 MonoProfileCoverageInfo *cov = cfg->coverage_info;
3945 g_assert (!cfg->compile_aot);
3947 cov->data [bb->dfn].cil_code = bb->cil_code;
3948 amd64_mov_reg_imm (code, AMD64_R11, (guint64)&cov->data [bb->dfn].count);
3949 /* this is not thread save, but good enough */
3950 amd64_inc_membase (code, AMD64_R11, 0);
3953 offset = code - cfg->native_code;
3955 mono_debug_open_block (cfg, bb, offset);
3957 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)
3958 x86_breakpoint (code);
3960 MONO_BB_FOR_EACH_INS (bb, ins) {
3961 offset = code - cfg->native_code;
3963 max_len = ((guint8 *)ins_get_spec (ins->opcode))[MONO_INST_LEN];
3965 #define EXTRA_CODE_SPACE (NACL_SIZE (16, 16 + kNaClAlignment))
3967 if (G_UNLIKELY (offset > (cfg->code_size - max_len - EXTRA_CODE_SPACE))) {
3968 cfg->code_size *= 2;
3969 cfg->native_code = mono_realloc_native_code(cfg);
3970 code = cfg->native_code + offset;
3971 cfg->stat_code_reallocs++;
3974 if (cfg->debug_info)
3975 mono_debug_record_line_number (cfg, ins, offset);
3977 switch (ins->opcode) {
3979 amd64_mul_reg (code, ins->sreg2, TRUE);
3982 amd64_mul_reg (code, ins->sreg2, FALSE);
3984 case OP_X86_SETEQ_MEMBASE:
3985 amd64_set_membase (code, X86_CC_EQ, ins->inst_basereg, ins->inst_offset, TRUE);
3987 case OP_STOREI1_MEMBASE_IMM:
3988 amd64_mov_membase_imm (code, ins->inst_destbasereg, ins->inst_offset, ins->inst_imm, 1);
3990 case OP_STOREI2_MEMBASE_IMM:
3991 amd64_mov_membase_imm (code, ins->inst_destbasereg, ins->inst_offset, ins->inst_imm, 2);
3993 case OP_STOREI4_MEMBASE_IMM:
3994 amd64_mov_membase_imm (code, ins->inst_destbasereg, ins->inst_offset, ins->inst_imm, 4);
3996 case OP_STOREI1_MEMBASE_REG:
3997 amd64_mov_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, ins->sreg1, 1);
3999 case OP_STOREI2_MEMBASE_REG:
4000 amd64_mov_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, ins->sreg1, 2);
4002 /* In AMD64 NaCl, pointers are 4 bytes, */
4003 /* so STORE_* != STOREI8_*. Likewise below. */
4004 case OP_STORE_MEMBASE_REG:
4005 amd64_mov_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, ins->sreg1, sizeof(gpointer));
4007 case OP_STOREI8_MEMBASE_REG:
4008 amd64_mov_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, ins->sreg1, 8);
4010 case OP_STOREI4_MEMBASE_REG:
4011 amd64_mov_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, ins->sreg1, 4);
4013 case OP_STORE_MEMBASE_IMM:
4014 #ifndef __native_client_codegen__
4015 /* In NaCl, this could be a PCONST type, which could */
4016 /* mean a pointer type was copied directly into the */
4017 /* lower 32-bits of inst_imm, so for InvalidPtr==-1 */
4018 /* the value would be 0x00000000FFFFFFFF which is */
4019 /* not proper for an imm32 unless you cast it. */
4020 g_assert (amd64_is_imm32 (ins->inst_imm));
4022 amd64_mov_membase_imm (code, ins->inst_destbasereg, ins->inst_offset, (gint32)ins->inst_imm, sizeof(gpointer));
4024 case OP_STOREI8_MEMBASE_IMM:
4025 g_assert (amd64_is_imm32 (ins->inst_imm));
4026 amd64_mov_membase_imm (code, ins->inst_destbasereg, ins->inst_offset, ins->inst_imm, 8);
4029 #ifdef __mono_ilp32__
4030 /* In ILP32, pointers are 4 bytes, so separate these */
4031 /* cases, use literal 8 below where we really want 8 */
4032 amd64_mov_reg_imm (code, ins->dreg, ins->inst_imm);
4033 amd64_mov_reg_membase (code, ins->dreg, ins->dreg, 0, sizeof(gpointer));
4037 // FIXME: Decompose this earlier
4038 if (amd64_use_imm32 (ins->inst_imm))
4039 amd64_mov_reg_mem (code, ins->dreg, ins->inst_imm, 8);
4041 amd64_mov_reg_imm_size (code, ins->dreg, ins->inst_imm, sizeof(gpointer));
4042 amd64_mov_reg_membase (code, ins->dreg, ins->dreg, 0, 8);
4046 amd64_mov_reg_imm (code, ins->dreg, ins->inst_imm);
4047 amd64_movsxd_reg_membase (code, ins->dreg, ins->dreg, 0);
4050 // FIXME: Decompose this earlier
4051 if (amd64_use_imm32 (ins->inst_imm))
4052 amd64_mov_reg_mem (code, ins->dreg, ins->inst_imm, 4);
4054 amd64_mov_reg_imm_size (code, ins->dreg, ins->inst_imm, sizeof(gpointer));
4055 amd64_mov_reg_membase (code, ins->dreg, ins->dreg, 0, 4);
4059 amd64_mov_reg_imm (code, ins->dreg, ins->inst_imm);
4060 amd64_widen_membase (code, ins->dreg, ins->dreg, 0, FALSE, FALSE);
4063 /* For NaCl, pointers are 4 bytes, so separate these */
4064 /* cases, use literal 8 below where we really want 8 */
4065 amd64_mov_reg_imm (code, ins->dreg, ins->inst_imm);
4066 amd64_widen_membase (code, ins->dreg, ins->dreg, 0, FALSE, TRUE);
4068 case OP_LOAD_MEMBASE:
4069 g_assert (amd64_is_imm32 (ins->inst_offset));
4070 amd64_mov_reg_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, sizeof(gpointer));
4072 case OP_LOADI8_MEMBASE:
4073 /* Use literal 8 instead of sizeof pointer or */
4074 /* register, we really want 8 for this opcode */
4075 g_assert (amd64_is_imm32 (ins->inst_offset));
4076 amd64_mov_reg_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, 8);
4078 case OP_LOADI4_MEMBASE:
4079 amd64_movsxd_reg_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset);
4081 case OP_LOADU4_MEMBASE:
4082 amd64_mov_reg_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, 4);
4084 case OP_LOADU1_MEMBASE:
4085 /* The cpu zero extends the result into 64 bits */
4086 amd64_widen_membase_size (code, ins->dreg, ins->inst_basereg, ins->inst_offset, FALSE, FALSE, 4);
4088 case OP_LOADI1_MEMBASE:
4089 amd64_widen_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, TRUE, FALSE);
4091 case OP_LOADU2_MEMBASE:
4092 /* The cpu zero extends the result into 64 bits */
4093 amd64_widen_membase_size (code, ins->dreg, ins->inst_basereg, ins->inst_offset, FALSE, TRUE, 4);
4095 case OP_LOADI2_MEMBASE:
4096 amd64_widen_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, TRUE, TRUE);
4098 case OP_AMD64_LOADI8_MEMINDEX:
4099 amd64_mov_reg_memindex_size (code, ins->dreg, ins->inst_basereg, 0, ins->inst_indexreg, 0, 8);
4101 case OP_LCONV_TO_I1:
4102 case OP_ICONV_TO_I1:
4104 amd64_widen_reg (code, ins->dreg, ins->sreg1, TRUE, FALSE);
4106 case OP_LCONV_TO_I2:
4107 case OP_ICONV_TO_I2:
4109 amd64_widen_reg (code, ins->dreg, ins->sreg1, TRUE, TRUE);
4111 case OP_LCONV_TO_U1:
4112 case OP_ICONV_TO_U1:
4113 amd64_widen_reg (code, ins->dreg, ins->sreg1, FALSE, FALSE);
4115 case OP_LCONV_TO_U2:
4116 case OP_ICONV_TO_U2:
4117 amd64_widen_reg (code, ins->dreg, ins->sreg1, FALSE, TRUE);
4120 /* Clean out the upper word */
4121 amd64_mov_reg_reg_size (code, ins->dreg, ins->sreg1, 4);
4124 amd64_movsxd_reg_reg (code, ins->dreg, ins->sreg1);
4128 amd64_alu_reg_reg (code, X86_CMP, ins->sreg1, ins->sreg2);
4130 case OP_COMPARE_IMM:
4131 #if defined(__mono_ilp32__)
4132 /* Comparison of pointer immediates should be 4 bytes to avoid sign-extend problems */
4133 g_assert (amd64_is_imm32 (ins->inst_imm));
4134 amd64_alu_reg_imm_size (code, X86_CMP, ins->sreg1, ins->inst_imm, 4);
4137 case OP_LCOMPARE_IMM:
4138 g_assert (amd64_is_imm32 (ins->inst_imm));
4139 amd64_alu_reg_imm (code, X86_CMP, ins->sreg1, ins->inst_imm);
4141 case OP_X86_COMPARE_REG_MEMBASE:
4142 amd64_alu_reg_membase (code, X86_CMP, ins->sreg1, ins->sreg2, ins->inst_offset);
4144 case OP_X86_TEST_NULL:
4145 amd64_test_reg_reg_size (code, ins->sreg1, ins->sreg1, 4);
4147 case OP_AMD64_TEST_NULL:
4148 amd64_test_reg_reg (code, ins->sreg1, ins->sreg1);
4151 case OP_X86_ADD_REG_MEMBASE:
4152 amd64_alu_reg_membase_size (code, X86_ADD, ins->sreg1, ins->sreg2, ins->inst_offset, 4);
4154 case OP_X86_SUB_REG_MEMBASE:
4155 amd64_alu_reg_membase_size (code, X86_SUB, ins->sreg1, ins->sreg2, ins->inst_offset, 4);
4157 case OP_X86_AND_REG_MEMBASE:
4158 amd64_alu_reg_membase_size (code, X86_AND, ins->sreg1, ins->sreg2, ins->inst_offset, 4);
4160 case OP_X86_OR_REG_MEMBASE:
4161 amd64_alu_reg_membase_size (code, X86_OR, ins->sreg1, ins->sreg2, ins->inst_offset, 4);
4163 case OP_X86_XOR_REG_MEMBASE:
4164 amd64_alu_reg_membase_size (code, X86_XOR, ins->sreg1, ins->sreg2, ins->inst_offset, 4);
4167 case OP_X86_ADD_MEMBASE_IMM:
4168 /* FIXME: Make a 64 version too */
4169 amd64_alu_membase_imm_size (code, X86_ADD, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 4);
4171 case OP_X86_SUB_MEMBASE_IMM:
4172 g_assert (amd64_is_imm32 (ins->inst_imm));
4173 amd64_alu_membase_imm_size (code, X86_SUB, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 4);
4175 case OP_X86_AND_MEMBASE_IMM:
4176 g_assert (amd64_is_imm32 (ins->inst_imm));
4177 amd64_alu_membase_imm_size (code, X86_AND, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 4);
4179 case OP_X86_OR_MEMBASE_IMM:
4180 g_assert (amd64_is_imm32 (ins->inst_imm));
4181 amd64_alu_membase_imm_size (code, X86_OR, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 4);
4183 case OP_X86_XOR_MEMBASE_IMM:
4184 g_assert (amd64_is_imm32 (ins->inst_imm));
4185 amd64_alu_membase_imm_size (code, X86_XOR, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 4);
4187 case OP_X86_ADD_MEMBASE_REG:
4188 amd64_alu_membase_reg_size (code, X86_ADD, ins->inst_basereg, ins->inst_offset, ins->sreg2, 4);
4190 case OP_X86_SUB_MEMBASE_REG:
4191 amd64_alu_membase_reg_size (code, X86_SUB, ins->inst_basereg, ins->inst_offset, ins->sreg2, 4);
4193 case OP_X86_AND_MEMBASE_REG:
4194 amd64_alu_membase_reg_size (code, X86_AND, ins->inst_basereg, ins->inst_offset, ins->sreg2, 4);
4196 case OP_X86_OR_MEMBASE_REG:
4197 amd64_alu_membase_reg_size (code, X86_OR, ins->inst_basereg, ins->inst_offset, ins->sreg2, 4);
4199 case OP_X86_XOR_MEMBASE_REG:
4200 amd64_alu_membase_reg_size (code, X86_XOR, ins->inst_basereg, ins->inst_offset, ins->sreg2, 4);
4202 case OP_X86_INC_MEMBASE:
4203 amd64_inc_membase_size (code, ins->inst_basereg, ins->inst_offset, 4);
4205 case OP_X86_INC_REG:
4206 amd64_inc_reg_size (code, ins->dreg, 4);
4208 case OP_X86_DEC_MEMBASE:
4209 amd64_dec_membase_size (code, ins->inst_basereg, ins->inst_offset, 4);
4211 case OP_X86_DEC_REG:
4212 amd64_dec_reg_size (code, ins->dreg, 4);
4214 case OP_X86_MUL_REG_MEMBASE:
4215 case OP_X86_MUL_MEMBASE_REG:
4216 amd64_imul_reg_membase_size (code, ins->sreg1, ins->sreg2, ins->inst_offset, 4);
4218 case OP_AMD64_ICOMPARE_MEMBASE_REG:
4219 amd64_alu_membase_reg_size (code, X86_CMP, ins->inst_basereg, ins->inst_offset, ins->sreg2, 4);
4221 case OP_AMD64_ICOMPARE_MEMBASE_IMM:
4222 amd64_alu_membase_imm_size (code, X86_CMP, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 4);
4224 case OP_AMD64_COMPARE_MEMBASE_REG:
4225 amd64_alu_membase_reg_size (code, X86_CMP, ins->inst_basereg, ins->inst_offset, ins->sreg2, 8);
4227 case OP_AMD64_COMPARE_MEMBASE_IMM:
4228 g_assert (amd64_is_imm32 (ins->inst_imm));
4229 amd64_alu_membase_imm_size (code, X86_CMP, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 8);
4231 case OP_X86_COMPARE_MEMBASE8_IMM:
4232 amd64_alu_membase8_imm_size (code, X86_CMP, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 4);
4234 case OP_AMD64_ICOMPARE_REG_MEMBASE:
4235 amd64_alu_reg_membase_size (code, X86_CMP, ins->sreg1, ins->sreg2, ins->inst_offset, 4);
4237 case OP_AMD64_COMPARE_REG_MEMBASE:
4238 amd64_alu_reg_membase_size (code, X86_CMP, ins->sreg1, ins->sreg2, ins->inst_offset, 8);
4241 case OP_AMD64_ADD_REG_MEMBASE:
4242 amd64_alu_reg_membase_size (code, X86_ADD, ins->sreg1, ins->sreg2, ins->inst_offset, 8);
4244 case OP_AMD64_SUB_REG_MEMBASE:
4245 amd64_alu_reg_membase_size (code, X86_SUB, ins->sreg1, ins->sreg2, ins->inst_offset, 8);
4247 case OP_AMD64_AND_REG_MEMBASE:
4248 amd64_alu_reg_membase_size (code, X86_AND, ins->sreg1, ins->sreg2, ins->inst_offset, 8);
4250 case OP_AMD64_OR_REG_MEMBASE:
4251 amd64_alu_reg_membase_size (code, X86_OR, ins->sreg1, ins->sreg2, ins->inst_offset, 8);
4253 case OP_AMD64_XOR_REG_MEMBASE:
4254 amd64_alu_reg_membase_size (code, X86_XOR, ins->sreg1, ins->sreg2, ins->inst_offset, 8);
4257 case OP_AMD64_ADD_MEMBASE_REG:
4258 amd64_alu_membase_reg_size (code, X86_ADD, ins->inst_basereg, ins->inst_offset, ins->sreg2, 8);
4260 case OP_AMD64_SUB_MEMBASE_REG:
4261 amd64_alu_membase_reg_size (code, X86_SUB, ins->inst_basereg, ins->inst_offset, ins->sreg2, 8);
4263 case OP_AMD64_AND_MEMBASE_REG:
4264 amd64_alu_membase_reg_size (code, X86_AND, ins->inst_basereg, ins->inst_offset, ins->sreg2, 8);
4266 case OP_AMD64_OR_MEMBASE_REG:
4267 amd64_alu_membase_reg_size (code, X86_OR, ins->inst_basereg, ins->inst_offset, ins->sreg2, 8);
4269 case OP_AMD64_XOR_MEMBASE_REG:
4270 amd64_alu_membase_reg_size (code, X86_XOR, ins->inst_basereg, ins->inst_offset, ins->sreg2, 8);
4273 case OP_AMD64_ADD_MEMBASE_IMM:
4274 g_assert (amd64_is_imm32 (ins->inst_imm));
4275 amd64_alu_membase_imm_size (code, X86_ADD, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 8);
4277 case OP_AMD64_SUB_MEMBASE_IMM:
4278 g_assert (amd64_is_imm32 (ins->inst_imm));
4279 amd64_alu_membase_imm_size (code, X86_SUB, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 8);
4281 case OP_AMD64_AND_MEMBASE_IMM:
4282 g_assert (amd64_is_imm32 (ins->inst_imm));
4283 amd64_alu_membase_imm_size (code, X86_AND, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 8);
4285 case OP_AMD64_OR_MEMBASE_IMM:
4286 g_assert (amd64_is_imm32 (ins->inst_imm));
4287 amd64_alu_membase_imm_size (code, X86_OR, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 8);
4289 case OP_AMD64_XOR_MEMBASE_IMM:
4290 g_assert (amd64_is_imm32 (ins->inst_imm));
4291 amd64_alu_membase_imm_size (code, X86_XOR, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 8);
4295 amd64_breakpoint (code);
4297 case OP_RELAXED_NOP:
4298 x86_prefix (code, X86_REP_PREFIX);
4306 case OP_DUMMY_STORE:
4307 case OP_DUMMY_ICONST:
4308 case OP_DUMMY_R8CONST:
4309 case OP_NOT_REACHED:
4312 case OP_IL_SEQ_POINT:
4313 mono_add_seq_point (cfg, bb, ins, code - cfg->native_code);
4315 case OP_SEQ_POINT: {
4316 if (ins->flags & MONO_INST_SINGLE_STEP_LOC) {
4317 MonoInst *var = cfg->arch.ss_tramp_var;
4320 /* Load ss_tramp_var */
4321 /* This is equal to &ss_trampoline */
4322 amd64_mov_reg_membase (code, AMD64_R11, var->inst_basereg, var->inst_offset, 8);
4323 /* Load the trampoline address */
4324 amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, 0, 8);
4325 /* Call it if it is non-null */
4326 amd64_test_reg_reg (code, AMD64_R11, AMD64_R11);
4328 amd64_branch8 (code, X86_CC_Z, 0, FALSE);
4329 amd64_call_reg (code, AMD64_R11);
4330 amd64_patch (label, code);
4334 * This is the address which is saved in seq points,
4336 mono_add_seq_point (cfg, bb, ins, code - cfg->native_code);
4338 if (cfg->compile_aot) {
4339 guint32 offset = code - cfg->native_code;
4341 MonoInst *info_var = cfg->arch.seq_point_info_var;
4345 amd64_mov_reg_membase (code, AMD64_R11, info_var->inst_basereg, info_var->inst_offset, 8);
4346 val = ((offset) * sizeof (guint8*)) + MONO_STRUCT_OFFSET (SeqPointInfo, bp_addrs);
4347 /* Load the info->bp_addrs [offset], which is either NULL or the address of the breakpoint trampoline */
4348 amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, val, 8);
4349 amd64_test_reg_reg (code, AMD64_R11, AMD64_R11);
4351 amd64_branch8 (code, X86_CC_Z, 0, FALSE);
4352 /* Call the trampoline */
4353 amd64_call_reg (code, AMD64_R11);
4354 amd64_patch (label, code);
4356 MonoInst *var = cfg->arch.bp_tramp_var;
4360 * Emit a test+branch against a constant, the constant will be overwritten
4361 * by mono_arch_set_breakpoint () to cause the test to fail.
4363 amd64_mov_reg_imm (code, AMD64_R11, 0);
4364 amd64_test_reg_reg (code, AMD64_R11, AMD64_R11);
4366 amd64_branch8 (code, X86_CC_Z, 0, FALSE);
4369 g_assert (var->opcode == OP_REGOFFSET);
4370 /* Load bp_tramp_var */
4371 /* This is equal to &bp_trampoline */
4372 amd64_mov_reg_membase (code, AMD64_R11, var->inst_basereg, var->inst_offset, 8);
4373 /* Call the trampoline */
4374 amd64_call_membase (code, AMD64_R11, 0);
4375 amd64_patch (label, code);
4378 * Add an additional nop so skipping the bp doesn't cause the ip to point
4379 * to another IL offset.
4387 amd64_alu_reg_reg (code, X86_ADD, ins->sreg1, ins->sreg2);
4390 amd64_alu_reg_reg (code, X86_ADC, ins->sreg1, ins->sreg2);
4394 g_assert (amd64_is_imm32 (ins->inst_imm));
4395 amd64_alu_reg_imm (code, X86_ADD, ins->dreg, ins->inst_imm);
4398 g_assert (amd64_is_imm32 (ins->inst_imm));
4399 amd64_alu_reg_imm (code, X86_ADC, ins->dreg, ins->inst_imm);
4404 amd64_alu_reg_reg (code, X86_SUB, ins->sreg1, ins->sreg2);
4407 amd64_alu_reg_reg (code, X86_SBB, ins->sreg1, ins->sreg2);
4411 g_assert (amd64_is_imm32 (ins->inst_imm));
4412 amd64_alu_reg_imm (code, X86_SUB, ins->dreg, ins->inst_imm);
4415 g_assert (amd64_is_imm32 (ins->inst_imm));
4416 amd64_alu_reg_imm (code, X86_SBB, ins->dreg, ins->inst_imm);
4419 amd64_alu_reg_reg (code, X86_AND, ins->sreg1, ins->sreg2);
4423 g_assert (amd64_is_imm32 (ins->inst_imm));
4424 amd64_alu_reg_imm (code, X86_AND, ins->sreg1, ins->inst_imm);
4427 amd64_imul_reg_reg (code, ins->sreg1, ins->sreg2);
4432 guint32 size = (ins->opcode == OP_IMUL_IMM) ? 4 : 8;
4434 switch (ins->inst_imm) {
4438 if (ins->dreg != ins->sreg1)
4439 amd64_mov_reg_reg (code, ins->dreg, ins->sreg1, size);
4440 amd64_alu_reg_reg (code, X86_ADD, ins->dreg, ins->dreg);
4443 /* LEA r1, [r2 + r2*2] */
4444 amd64_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 1);
4447 /* LEA r1, [r2 + r2*4] */
4448 amd64_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 2);
4451 /* LEA r1, [r2 + r2*2] */
4453 amd64_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 1);
4454 amd64_alu_reg_reg (code, X86_ADD, ins->dreg, ins->dreg);
4457 /* LEA r1, [r2 + r2*8] */
4458 amd64_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 3);
4461 /* LEA r1, [r2 + r2*4] */
4463 amd64_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 2);
4464 amd64_alu_reg_reg (code, X86_ADD, ins->dreg, ins->dreg);
4467 /* LEA r1, [r2 + r2*2] */
4469 amd64_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 1);
4470 amd64_shift_reg_imm (code, X86_SHL, ins->dreg, 2);
4473 /* LEA r1, [r2 + r2*4] */
4474 /* LEA r1, [r1 + r1*4] */
4475 amd64_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 2);
4476 amd64_lea_memindex (code, ins->dreg, ins->dreg, 0, ins->dreg, 2);
4479 /* LEA r1, [r2 + r2*4] */
4481 /* LEA r1, [r1 + r1*4] */
4482 amd64_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 2);
4483 amd64_shift_reg_imm (code, X86_SHL, ins->dreg, 2);
4484 amd64_lea_memindex (code, ins->dreg, ins->dreg, 0, ins->dreg, 2);
4487 amd64_imul_reg_reg_imm_size (code, ins->dreg, ins->sreg1, ins->inst_imm, size);
4494 #if defined( __native_client_codegen__ )
4495 amd64_alu_reg_imm (code, X86_CMP, ins->sreg2, 0);
4496 EMIT_COND_SYSTEM_EXCEPTION (X86_CC_EQ, TRUE, "DivideByZeroException");
4498 /* Regalloc magic makes the div/rem cases the same */
4499 if (ins->sreg2 == AMD64_RDX) {
4500 amd64_mov_membase_reg (code, AMD64_RSP, -8, AMD64_RDX, 8);
4502 amd64_div_membase (code, AMD64_RSP, -8, TRUE);
4505 amd64_div_reg (code, ins->sreg2, TRUE);
4510 #if defined( __native_client_codegen__ )
4511 amd64_alu_reg_imm (code, X86_CMP, ins->sreg2, 0);
4512 EMIT_COND_SYSTEM_EXCEPTION (X86_CC_EQ, TRUE, "DivideByZeroException");
4514 if (ins->sreg2 == AMD64_RDX) {
4515 amd64_mov_membase_reg (code, AMD64_RSP, -8, AMD64_RDX, 8);
4516 amd64_alu_reg_reg (code, X86_XOR, AMD64_RDX, AMD64_RDX);
4517 amd64_div_membase (code, AMD64_RSP, -8, FALSE);
4519 amd64_alu_reg_reg (code, X86_XOR, AMD64_RDX, AMD64_RDX);
4520 amd64_div_reg (code, ins->sreg2, FALSE);
4525 #if defined( __native_client_codegen__ )
4526 amd64_alu_reg_imm (code, X86_CMP, ins->sreg2, 0);
4527 EMIT_COND_SYSTEM_EXCEPTION (X86_CC_EQ, TRUE, "DivideByZeroException");
4529 if (ins->sreg2 == AMD64_RDX) {
4530 amd64_mov_membase_reg (code, AMD64_RSP, -8, AMD64_RDX, 8);
4531 amd64_cdq_size (code, 4);
4532 amd64_div_membase_size (code, AMD64_RSP, -8, TRUE, 4);
4534 amd64_cdq_size (code, 4);
4535 amd64_div_reg_size (code, ins->sreg2, TRUE, 4);
4540 #if defined( __native_client_codegen__ )
4541 amd64_alu_reg_imm_size (code, X86_CMP, ins->sreg2, 0, 4);
4542 EMIT_COND_SYSTEM_EXCEPTION (X86_CC_EQ, TRUE, "DivideByZeroException");
4544 if (ins->sreg2 == AMD64_RDX) {
4545 amd64_mov_membase_reg (code, AMD64_RSP, -8, AMD64_RDX, 8);
4546 amd64_alu_reg_reg (code, X86_XOR, AMD64_RDX, AMD64_RDX);
4547 amd64_div_membase_size (code, AMD64_RSP, -8, FALSE, 4);
4549 amd64_alu_reg_reg (code, X86_XOR, AMD64_RDX, AMD64_RDX);
4550 amd64_div_reg_size (code, ins->sreg2, FALSE, 4);
4554 amd64_imul_reg_reg (code, ins->sreg1, ins->sreg2);
4555 EMIT_COND_SYSTEM_EXCEPTION (X86_CC_O, FALSE, "OverflowException");
4558 amd64_alu_reg_reg (code, X86_OR, ins->sreg1, ins->sreg2);
4562 g_assert (amd64_is_imm32 (ins->inst_imm));
4563 amd64_alu_reg_imm (code, X86_OR, ins->sreg1, ins->inst_imm);
4566 amd64_alu_reg_reg (code, X86_XOR, ins->sreg1, ins->sreg2);
4570 g_assert (amd64_is_imm32 (ins->inst_imm));
4571 amd64_alu_reg_imm (code, X86_XOR, ins->sreg1, ins->inst_imm);
4574 g_assert (ins->sreg2 == AMD64_RCX);
4575 amd64_shift_reg (code, X86_SHL, ins->dreg);
4578 g_assert (ins->sreg2 == AMD64_RCX);
4579 amd64_shift_reg (code, X86_SAR, ins->dreg);
4583 g_assert (amd64_is_imm32 (ins->inst_imm));
4584 amd64_shift_reg_imm (code, X86_SAR, ins->dreg, ins->inst_imm);
4587 g_assert (amd64_is_imm32 (ins->inst_imm));
4588 amd64_shift_reg_imm_size (code, X86_SHR, ins->dreg, ins->inst_imm, 4);
4590 case OP_LSHR_UN_IMM:
4591 g_assert (amd64_is_imm32 (ins->inst_imm));
4592 amd64_shift_reg_imm (code, X86_SHR, ins->dreg, ins->inst_imm);
4595 g_assert (ins->sreg2 == AMD64_RCX);
4596 amd64_shift_reg (code, X86_SHR, ins->dreg);
4600 g_assert (amd64_is_imm32 (ins->inst_imm));
4601 amd64_shift_reg_imm (code, X86_SHL, ins->dreg, ins->inst_imm);
4606 amd64_alu_reg_reg_size (code, X86_ADD, ins->sreg1, ins->sreg2, 4);
4609 amd64_alu_reg_reg_size (code, X86_ADC, ins->sreg1, ins->sreg2, 4);
4612 amd64_alu_reg_imm_size (code, X86_ADD, ins->dreg, ins->inst_imm, 4);
4615 amd64_alu_reg_imm_size (code, X86_ADC, ins->dreg, ins->inst_imm, 4);
4619 amd64_alu_reg_reg_size (code, X86_SUB, ins->sreg1, ins->sreg2, 4);
4622 amd64_alu_reg_reg_size (code, X86_SBB, ins->sreg1, ins->sreg2, 4);
4625 amd64_alu_reg_imm_size (code, X86_SUB, ins->dreg, ins->inst_imm, 4);
4628 amd64_alu_reg_imm_size (code, X86_SBB, ins->dreg, ins->inst_imm, 4);
4631 amd64_alu_reg_reg_size (code, X86_AND, ins->sreg1, ins->sreg2, 4);
4634 amd64_alu_reg_imm_size (code, X86_AND, ins->sreg1, ins->inst_imm, 4);
4637 amd64_alu_reg_reg_size (code, X86_OR, ins->sreg1, ins->sreg2, 4);
4640 amd64_alu_reg_imm_size (code, X86_OR, ins->sreg1, ins->inst_imm, 4);
4643 amd64_alu_reg_reg_size (code, X86_XOR, ins->sreg1, ins->sreg2, 4);
4646 amd64_alu_reg_imm_size (code, X86_XOR, ins->sreg1, ins->inst_imm, 4);
4649 amd64_neg_reg_size (code, ins->sreg1, 4);
4652 amd64_not_reg_size (code, ins->sreg1, 4);
4655 g_assert (ins->sreg2 == AMD64_RCX);
4656 amd64_shift_reg_size (code, X86_SHL, ins->dreg, 4);
4659 g_assert (ins->sreg2 == AMD64_RCX);
4660 amd64_shift_reg_size (code, X86_SAR, ins->dreg, 4);
4663 amd64_shift_reg_imm_size (code, X86_SAR, ins->dreg, ins->inst_imm, 4);
4665 case OP_ISHR_UN_IMM:
4666 amd64_shift_reg_imm_size (code, X86_SHR, ins->dreg, ins->inst_imm, 4);
4669 g_assert (ins->sreg2 == AMD64_RCX);
4670 amd64_shift_reg_size (code, X86_SHR, ins->dreg, 4);
4673 amd64_shift_reg_imm_size (code, X86_SHL, ins->dreg, ins->inst_imm, 4);
4676 amd64_imul_reg_reg_size (code, ins->sreg1, ins->sreg2, 4);
4679 amd64_imul_reg_reg_size (code, ins->sreg1, ins->sreg2, 4);
4680 EMIT_COND_SYSTEM_EXCEPTION (X86_CC_O, FALSE, "OverflowException");
4682 case OP_IMUL_OVF_UN:
4683 case OP_LMUL_OVF_UN: {
4684 /* the mul operation and the exception check should most likely be split */
4685 int non_eax_reg, saved_eax = FALSE, saved_edx = FALSE;
4686 int size = (ins->opcode == OP_IMUL_OVF_UN) ? 4 : 8;
4687 /*g_assert (ins->sreg2 == X86_EAX);
4688 g_assert (ins->dreg == X86_EAX);*/
4689 if (ins->sreg2 == X86_EAX) {
4690 non_eax_reg = ins->sreg1;
4691 } else if (ins->sreg1 == X86_EAX) {
4692 non_eax_reg = ins->sreg2;
4694 /* no need to save since we're going to store to it anyway */
4695 if (ins->dreg != X86_EAX) {
4697 amd64_push_reg (code, X86_EAX);
4699 amd64_mov_reg_reg (code, X86_EAX, ins->sreg1, size);
4700 non_eax_reg = ins->sreg2;
4702 if (ins->dreg == X86_EDX) {
4705 amd64_push_reg (code, X86_EAX);
4709 amd64_push_reg (code, X86_EDX);
4711 amd64_mul_reg_size (code, non_eax_reg, FALSE, size);
4712 /* save before the check since pop and mov don't change the flags */
4713 if (ins->dreg != X86_EAX)
4714 amd64_mov_reg_reg (code, ins->dreg, X86_EAX, size);
4716 amd64_pop_reg (code, X86_EDX);
4718 amd64_pop_reg (code, X86_EAX);
4719 EMIT_COND_SYSTEM_EXCEPTION (X86_CC_O, FALSE, "OverflowException");
4723 amd64_alu_reg_reg_size (code, X86_CMP, ins->sreg1, ins->sreg2, 4);
4725 case OP_ICOMPARE_IMM:
4726 amd64_alu_reg_imm_size (code, X86_CMP, ins->sreg1, ins->inst_imm, 4);
4748 EMIT_COND_BRANCH (ins, cc_table [mono_opcode_to_cond (ins->opcode)], cc_signed_table [mono_opcode_to_cond (ins->opcode)]);
4756 case OP_CMOV_INE_UN:
4757 case OP_CMOV_IGE_UN:
4758 case OP_CMOV_IGT_UN:
4759 case OP_CMOV_ILE_UN:
4760 case OP_CMOV_ILT_UN:
4766 case OP_CMOV_LNE_UN:
4767 case OP_CMOV_LGE_UN:
4768 case OP_CMOV_LGT_UN:
4769 case OP_CMOV_LLE_UN:
4770 case OP_CMOV_LLT_UN:
4771 g_assert (ins->dreg == ins->sreg1);
4772 /* This needs to operate on 64 bit values */
4773 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);
4777 amd64_not_reg (code, ins->sreg1);
4780 amd64_neg_reg (code, ins->sreg1);
4785 if ((((guint64)ins->inst_c0) >> 32) == 0 && !mini_get_debug_options()->single_imm_size)
4786 amd64_mov_reg_imm_size (code, ins->dreg, ins->inst_c0, 4);
4788 amd64_mov_reg_imm_size (code, ins->dreg, ins->inst_c0, 8);
4791 mono_add_patch_info (cfg, offset, (MonoJumpInfoType)ins->inst_i1, ins->inst_p0);
4792 amd64_mov_reg_membase (code, ins->dreg, AMD64_RIP, 0, sizeof(gpointer));
4795 mono_add_patch_info (cfg, offset, (MonoJumpInfoType)ins->inst_i1, ins->inst_p0);
4796 amd64_mov_reg_imm_size (code, ins->dreg, 0, 8);
4799 if (ins->dreg != ins->sreg1)
4800 amd64_mov_reg_reg (code, ins->dreg, ins->sreg1, sizeof(mgreg_t));
4802 case OP_AMD64_SET_XMMREG_R4: {
4804 if (ins->dreg != ins->sreg1)
4805 amd64_sse_movss_reg_reg (code, ins->dreg, ins->sreg1);
4807 amd64_sse_cvtsd2ss_reg_reg (code, ins->dreg, ins->sreg1);
4811 case OP_AMD64_SET_XMMREG_R8: {
4812 if (ins->dreg != ins->sreg1)
4813 amd64_sse_movsd_reg_reg (code, ins->dreg, ins->sreg1);
4817 MonoCallInst *call = (MonoCallInst*)ins;
4818 int i, save_area_offset;
4820 g_assert (!cfg->method->save_lmf);
4822 /* Restore callee saved registers */
4823 save_area_offset = cfg->arch.reg_save_area_offset;
4824 for (i = 0; i < AMD64_NREG; ++i)
4825 if (AMD64_IS_CALLEE_SAVED_REG (i) && (cfg->used_int_regs & (1 << i))) {
4826 amd64_mov_reg_membase (code, i, cfg->frame_reg, save_area_offset, 8);
4827 save_area_offset += 8;
4830 if (cfg->arch.omit_fp) {
4831 if (cfg->arch.stack_alloc_size)
4832 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, cfg->arch.stack_alloc_size);
4834 if (call->stack_usage)
4837 /* Copy arguments on the stack to our argument area */
4838 for (i = 0; i < call->stack_usage; i += sizeof(mgreg_t)) {
4839 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RSP, i, sizeof(mgreg_t));
4840 amd64_mov_membase_reg (code, AMD64_RBP, 16 + i, AMD64_RAX, sizeof(mgreg_t));
4846 offset = code - cfg->native_code;
4847 mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_METHOD_JUMP, call->method);
4848 if (cfg->compile_aot)
4849 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RIP, 0, 8);
4851 amd64_set_reg_template (code, AMD64_R11);
4852 amd64_jump_reg (code, AMD64_R11);
4853 ins->flags |= MONO_INST_GC_CALLSITE;
4854 ins->backend.pc_offset = code - cfg->native_code;
4858 /* ensure ins->sreg1 is not NULL */
4859 amd64_alu_membase_imm_size (code, X86_CMP, ins->sreg1, 0, 0, 4);
4862 amd64_lea_membase (code, AMD64_R11, cfg->frame_reg, cfg->sig_cookie);
4863 amd64_mov_membase_reg (code, ins->sreg1, 0, AMD64_R11, sizeof(gpointer));
4873 call = (MonoCallInst*)ins;
4875 * The AMD64 ABI forces callers to know about varargs.
4877 if ((call->signature->call_convention == MONO_CALL_VARARG) && (call->signature->pinvoke))
4878 amd64_alu_reg_reg (code, X86_XOR, AMD64_RAX, AMD64_RAX);
4879 else if ((cfg->method->wrapper_type == MONO_WRAPPER_MANAGED_TO_NATIVE) && (cfg->method->klass->image != mono_defaults.corlib)) {
4881 * Since the unmanaged calling convention doesn't contain a
4882 * 'vararg' entry, we have to treat every pinvoke call as a
4883 * potential vararg call.
4887 for (i = 0; i < AMD64_XMM_NREG; ++i)
4888 if (call->used_fregs & (1 << i))
4891 amd64_alu_reg_reg (code, X86_XOR, AMD64_RAX, AMD64_RAX);
4893 amd64_mov_reg_imm (code, AMD64_RAX, nregs);
4896 if (ins->flags & MONO_INST_HAS_METHOD)
4897 code = emit_call (cfg, code, MONO_PATCH_INFO_METHOD, call->method, FALSE);
4899 code = emit_call (cfg, code, MONO_PATCH_INFO_ABS, call->fptr, FALSE);
4900 ins->flags |= MONO_INST_GC_CALLSITE;
4901 ins->backend.pc_offset = code - cfg->native_code;
4902 code = emit_move_return_value (cfg, ins, code);
4909 case OP_VOIDCALL_REG:
4911 call = (MonoCallInst*)ins;
4913 if (AMD64_IS_ARGUMENT_REG (ins->sreg1)) {
4914 amd64_mov_reg_reg (code, AMD64_R11, ins->sreg1, 8);
4915 ins->sreg1 = AMD64_R11;
4919 * The AMD64 ABI forces callers to know about varargs.
4921 if ((call->signature->call_convention == MONO_CALL_VARARG) && (call->signature->pinvoke)) {
4922 if (ins->sreg1 == AMD64_RAX) {
4923 amd64_mov_reg_reg (code, AMD64_R11, AMD64_RAX, 8);
4924 ins->sreg1 = AMD64_R11;
4926 amd64_alu_reg_reg (code, X86_XOR, AMD64_RAX, AMD64_RAX);
4927 } else if ((cfg->method->wrapper_type == MONO_WRAPPER_MANAGED_TO_NATIVE) && (cfg->method->klass->image != mono_defaults.corlib)) {
4929 * Since the unmanaged calling convention doesn't contain a
4930 * 'vararg' entry, we have to treat every pinvoke call as a
4931 * potential vararg call.
4935 for (i = 0; i < AMD64_XMM_NREG; ++i)
4936 if (call->used_fregs & (1 << i))
4938 if (ins->sreg1 == AMD64_RAX) {
4939 amd64_mov_reg_reg (code, AMD64_R11, AMD64_RAX, 8);
4940 ins->sreg1 = AMD64_R11;
4943 amd64_alu_reg_reg (code, X86_XOR, AMD64_RAX, AMD64_RAX);
4945 amd64_mov_reg_imm (code, AMD64_RAX, nregs);
4948 amd64_call_reg (code, ins->sreg1);
4949 ins->flags |= MONO_INST_GC_CALLSITE;
4950 ins->backend.pc_offset = code - cfg->native_code;
4951 code = emit_move_return_value (cfg, ins, code);
4953 case OP_FCALL_MEMBASE:
4954 case OP_RCALL_MEMBASE:
4955 case OP_LCALL_MEMBASE:
4956 case OP_VCALL_MEMBASE:
4957 case OP_VCALL2_MEMBASE:
4958 case OP_VOIDCALL_MEMBASE:
4959 case OP_CALL_MEMBASE:
4960 call = (MonoCallInst*)ins;
4962 amd64_call_membase (code, ins->sreg1, ins->inst_offset);
4963 ins->flags |= MONO_INST_GC_CALLSITE;
4964 ins->backend.pc_offset = code - cfg->native_code;
4965 code = emit_move_return_value (cfg, ins, code);
4969 MonoInst *var = cfg->dyn_call_var;
4971 g_assert (var->opcode == OP_REGOFFSET);
4973 /* r11 = args buffer filled by mono_arch_get_dyn_call_args () */
4974 amd64_mov_reg_reg (code, AMD64_R11, ins->sreg1, 8);
4976 amd64_mov_reg_reg (code, AMD64_R10, ins->sreg2, 8);
4978 /* Save args buffer */
4979 amd64_mov_membase_reg (code, var->inst_basereg, var->inst_offset, AMD64_R11, 8);
4981 /* Set argument registers */
4982 for (i = 0; i < PARAM_REGS; ++i)
4983 amd64_mov_reg_membase (code, param_regs [i], AMD64_R11, i * sizeof(mgreg_t), sizeof(mgreg_t));
4986 amd64_call_reg (code, AMD64_R10);
4988 ins->flags |= MONO_INST_GC_CALLSITE;
4989 ins->backend.pc_offset = code - cfg->native_code;
4992 amd64_mov_reg_membase (code, AMD64_R11, var->inst_basereg, var->inst_offset, 8);
4993 amd64_mov_membase_reg (code, AMD64_R11, MONO_STRUCT_OFFSET (DynCallArgs, res), AMD64_RAX, 8);
4996 case OP_AMD64_SAVE_SP_TO_LMF: {
4997 MonoInst *lmf_var = cfg->lmf_var;
4998 amd64_mov_membase_reg (code, lmf_var->inst_basereg, lmf_var->inst_offset + MONO_STRUCT_OFFSET (MonoLMF, rsp), AMD64_RSP, 8);
5002 g_assert_not_reached ();
5003 amd64_push_reg (code, ins->sreg1);
5005 case OP_X86_PUSH_IMM:
5006 g_assert_not_reached ();
5007 g_assert (amd64_is_imm32 (ins->inst_imm));
5008 amd64_push_imm (code, ins->inst_imm);
5010 case OP_X86_PUSH_MEMBASE:
5011 g_assert_not_reached ();
5012 amd64_push_membase (code, ins->inst_basereg, ins->inst_offset);
5014 case OP_X86_PUSH_OBJ: {
5015 int size = ALIGN_TO (ins->inst_imm, 8);
5017 g_assert_not_reached ();
5019 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, size);
5020 amd64_push_reg (code, AMD64_RDI);
5021 amd64_push_reg (code, AMD64_RSI);
5022 amd64_push_reg (code, AMD64_RCX);
5023 if (ins->inst_offset)
5024 amd64_lea_membase (code, AMD64_RSI, ins->inst_basereg, ins->inst_offset);
5026 amd64_mov_reg_reg (code, AMD64_RSI, ins->inst_basereg, 8);
5027 amd64_lea_membase (code, AMD64_RDI, AMD64_RSP, (3 * 8));
5028 amd64_mov_reg_imm (code, AMD64_RCX, (size >> 3));
5030 amd64_prefix (code, X86_REP_PREFIX);
5032 amd64_pop_reg (code, AMD64_RCX);
5033 amd64_pop_reg (code, AMD64_RSI);
5034 amd64_pop_reg (code, AMD64_RDI);
5037 case OP_GENERIC_CLASS_INIT: {
5038 static int byte_offset = -1;
5039 static guint8 bitmask;
5042 g_assert (ins->sreg1 == MONO_AMD64_ARG_REG1);
5044 if (byte_offset < 0)
5045 mono_marshal_find_bitfield_offset (MonoVTable, initialized, &byte_offset, &bitmask);
5047 amd64_test_membase_imm_size (code, ins->sreg1, byte_offset, bitmask, 1);
5049 amd64_branch8 (code, X86_CC_NZ, -1, 1);
5051 code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD, "mono_generic_class_init", FALSE);
5052 ins->flags |= MONO_INST_GC_CALLSITE;
5053 ins->backend.pc_offset = code - cfg->native_code;
5055 x86_patch (jump, code);
5060 amd64_lea_memindex (code, ins->dreg, ins->sreg1, ins->inst_imm, ins->sreg2, ins->backend.shift_amount);
5062 case OP_X86_LEA_MEMBASE:
5063 amd64_lea_membase (code, ins->dreg, ins->sreg1, ins->inst_imm);
5066 amd64_xchg_reg_reg (code, ins->sreg1, ins->sreg2, 4);
5069 /* keep alignment */
5070 amd64_alu_reg_imm (code, X86_ADD, ins->sreg1, MONO_ARCH_FRAME_ALIGNMENT - 1);
5071 amd64_alu_reg_imm (code, X86_AND, ins->sreg1, ~(MONO_ARCH_FRAME_ALIGNMENT - 1));
5072 code = mono_emit_stack_alloc (cfg, code, ins);
5073 amd64_mov_reg_reg (code, ins->dreg, AMD64_RSP, 8);
5074 if (cfg->param_area)
5075 amd64_alu_reg_imm (code, X86_ADD, ins->dreg, cfg->param_area);
5077 case OP_LOCALLOC_IMM: {
5078 guint32 size = ins->inst_imm;
5079 size = (size + (MONO_ARCH_FRAME_ALIGNMENT - 1)) & ~ (MONO_ARCH_FRAME_ALIGNMENT - 1);
5081 if (ins->flags & MONO_INST_INIT) {
5085 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, size);
5086 amd64_alu_reg_reg (code, X86_XOR, ins->dreg, ins->dreg);
5088 for (i = 0; i < size; i += 8)
5089 amd64_mov_membase_reg (code, AMD64_RSP, i, ins->dreg, 8);
5090 amd64_mov_reg_reg (code, ins->dreg, AMD64_RSP, 8);
5092 amd64_mov_reg_imm (code, ins->dreg, size);
5093 ins->sreg1 = ins->dreg;
5095 code = mono_emit_stack_alloc (cfg, code, ins);
5096 amd64_mov_reg_reg (code, ins->dreg, AMD64_RSP, 8);
5099 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, size);
5100 amd64_mov_reg_reg (code, ins->dreg, AMD64_RSP, 8);
5102 if (cfg->param_area)
5103 amd64_alu_reg_imm (code, X86_ADD, ins->dreg, cfg->param_area);
5107 amd64_mov_reg_reg (code, AMD64_ARG_REG1, ins->sreg1, 8);
5108 code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD,
5109 (gpointer)"mono_arch_throw_exception", FALSE);
5110 ins->flags |= MONO_INST_GC_CALLSITE;
5111 ins->backend.pc_offset = code - cfg->native_code;
5115 amd64_mov_reg_reg (code, AMD64_ARG_REG1, ins->sreg1, 8);
5116 code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD,
5117 (gpointer)"mono_arch_rethrow_exception", FALSE);
5118 ins->flags |= MONO_INST_GC_CALLSITE;
5119 ins->backend.pc_offset = code - cfg->native_code;
5122 case OP_CALL_HANDLER:
5124 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, 8);
5125 mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_BB, ins->inst_target_bb);
5126 amd64_call_imm (code, 0);
5127 mono_cfg_add_try_hole (cfg, ins->inst_eh_block, code, bb);
5128 /* Restore stack alignment */
5129 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, 8);
5131 case OP_START_HANDLER: {
5132 /* Even though we're saving RSP, use sizeof */
5133 /* gpointer because spvar is of type IntPtr */
5134 /* see: mono_create_spvar_for_region */
5135 MonoInst *spvar = mono_find_spvar_for_region (cfg, bb->region);
5136 amd64_mov_membase_reg (code, spvar->inst_basereg, spvar->inst_offset, AMD64_RSP, sizeof(gpointer));
5138 if ((MONO_BBLOCK_IS_IN_REGION (bb, MONO_REGION_FINALLY) ||
5139 MONO_BBLOCK_IS_IN_REGION (bb, MONO_REGION_FINALLY)) &&
5141 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, ALIGN_TO (cfg->param_area, MONO_ARCH_FRAME_ALIGNMENT));
5145 case OP_ENDFINALLY: {
5146 MonoInst *spvar = mono_find_spvar_for_region (cfg, bb->region);
5147 amd64_mov_reg_membase (code, AMD64_RSP, spvar->inst_basereg, spvar->inst_offset, sizeof(gpointer));
5151 case OP_ENDFILTER: {
5152 MonoInst *spvar = mono_find_spvar_for_region (cfg, bb->region);
5153 amd64_mov_reg_membase (code, AMD64_RSP, spvar->inst_basereg, spvar->inst_offset, sizeof(gpointer));
5154 /* The local allocator will put the result into RAX */
5159 if (ins->dreg != AMD64_RAX)
5160 amd64_mov_reg_reg (code, ins->dreg, AMD64_RAX, sizeof (gpointer));
5163 ins->inst_c0 = code - cfg->native_code;
5166 //g_print ("target: %p, next: %p, curr: %p, last: %p\n", ins->inst_target_bb, bb->next_bb, ins, bb->last_ins);
5167 //if ((ins->inst_target_bb == bb->next_bb) && ins == bb->last_ins)
5169 if (ins->inst_target_bb->native_offset) {
5170 amd64_jump_code (code, cfg->native_code + ins->inst_target_bb->native_offset);
5172 mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_BB, ins->inst_target_bb);
5173 if ((cfg->opt & MONO_OPT_BRANCH) &&
5174 x86_is_imm8 (ins->inst_target_bb->max_offset - offset))
5175 x86_jump8 (code, 0);
5177 x86_jump32 (code, 0);
5181 amd64_jump_reg (code, ins->sreg1);
5204 amd64_set_reg (code, cc_table [mono_opcode_to_cond (ins->opcode)], ins->dreg, cc_signed_table [mono_opcode_to_cond (ins->opcode)]);
5205 amd64_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
5207 case OP_COND_EXC_EQ:
5208 case OP_COND_EXC_NE_UN:
5209 case OP_COND_EXC_LT:
5210 case OP_COND_EXC_LT_UN:
5211 case OP_COND_EXC_GT:
5212 case OP_COND_EXC_GT_UN:
5213 case OP_COND_EXC_GE:
5214 case OP_COND_EXC_GE_UN:
5215 case OP_COND_EXC_LE:
5216 case OP_COND_EXC_LE_UN:
5217 case OP_COND_EXC_IEQ:
5218 case OP_COND_EXC_INE_UN:
5219 case OP_COND_EXC_ILT:
5220 case OP_COND_EXC_ILT_UN:
5221 case OP_COND_EXC_IGT:
5222 case OP_COND_EXC_IGT_UN:
5223 case OP_COND_EXC_IGE:
5224 case OP_COND_EXC_IGE_UN:
5225 case OP_COND_EXC_ILE:
5226 case OP_COND_EXC_ILE_UN:
5227 EMIT_COND_SYSTEM_EXCEPTION (cc_table [mono_opcode_to_cond (ins->opcode)], cc_signed_table [mono_opcode_to_cond (ins->opcode)], ins->inst_p1);
5229 case OP_COND_EXC_OV:
5230 case OP_COND_EXC_NO:
5232 case OP_COND_EXC_NC:
5233 EMIT_COND_SYSTEM_EXCEPTION (branch_cc_table [ins->opcode - OP_COND_EXC_EQ],
5234 (ins->opcode < OP_COND_EXC_NE_UN), ins->inst_p1);
5236 case OP_COND_EXC_IOV:
5237 case OP_COND_EXC_INO:
5238 case OP_COND_EXC_IC:
5239 case OP_COND_EXC_INC:
5240 EMIT_COND_SYSTEM_EXCEPTION (branch_cc_table [ins->opcode - OP_COND_EXC_IEQ],
5241 (ins->opcode < OP_COND_EXC_INE_UN), ins->inst_p1);
5244 /* floating point opcodes */
5246 double d = *(double *)ins->inst_p0;
5248 if ((d == 0.0) && (mono_signbit (d) == 0)) {
5249 amd64_sse_xorpd_reg_reg (code, ins->dreg, ins->dreg);
5252 mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_R8, ins->inst_p0);
5253 amd64_sse_movsd_reg_membase (code, ins->dreg, AMD64_RIP, 0);
5258 float f = *(float *)ins->inst_p0;
5260 if ((f == 0.0) && (mono_signbit (f) == 0)) {
5262 amd64_sse_xorps_reg_reg (code, ins->dreg, ins->dreg);
5264 amd64_sse_xorpd_reg_reg (code, ins->dreg, ins->dreg);
5267 mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_R4, ins->inst_p0);
5268 amd64_sse_movss_reg_membase (code, ins->dreg, AMD64_RIP, 0);
5270 amd64_sse_cvtss2sd_reg_reg (code, ins->dreg, ins->dreg);
5274 case OP_STORER8_MEMBASE_REG:
5275 amd64_sse_movsd_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, ins->sreg1);
5277 case OP_LOADR8_MEMBASE:
5278 amd64_sse_movsd_reg_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset);
5280 case OP_STORER4_MEMBASE_REG:
5282 amd64_sse_movss_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, ins->sreg1);
5284 /* This requires a double->single conversion */
5285 amd64_sse_cvtsd2ss_reg_reg (code, MONO_ARCH_FP_SCRATCH_REG, ins->sreg1);
5286 amd64_sse_movss_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, MONO_ARCH_FP_SCRATCH_REG);
5289 case OP_LOADR4_MEMBASE:
5291 amd64_sse_movss_reg_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset);
5293 amd64_sse_movss_reg_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset);
5294 amd64_sse_cvtss2sd_reg_reg (code, ins->dreg, ins->dreg);
5297 case OP_ICONV_TO_R4:
5299 amd64_sse_cvtsi2ss_reg_reg_size (code, ins->dreg, ins->sreg1, 4);
5301 amd64_sse_cvtsi2ss_reg_reg_size (code, ins->dreg, ins->sreg1, 4);
5302 amd64_sse_cvtss2sd_reg_reg (code, ins->dreg, ins->dreg);
5305 case OP_ICONV_TO_R8:
5306 amd64_sse_cvtsi2sd_reg_reg_size (code, ins->dreg, ins->sreg1, 4);
5308 case OP_LCONV_TO_R4:
5310 amd64_sse_cvtsi2ss_reg_reg (code, ins->dreg, ins->sreg1);
5312 amd64_sse_cvtsi2ss_reg_reg (code, ins->dreg, ins->sreg1);
5313 amd64_sse_cvtss2sd_reg_reg (code, ins->dreg, ins->dreg);
5316 case OP_LCONV_TO_R8:
5317 amd64_sse_cvtsi2sd_reg_reg (code, ins->dreg, ins->sreg1);
5319 case OP_FCONV_TO_R4:
5321 amd64_sse_cvtsd2ss_reg_reg (code, ins->dreg, ins->sreg1);
5323 amd64_sse_cvtsd2ss_reg_reg (code, ins->dreg, ins->sreg1);
5324 amd64_sse_cvtss2sd_reg_reg (code, ins->dreg, ins->dreg);
5327 case OP_FCONV_TO_I1:
5328 code = emit_float_to_int (cfg, code, ins->dreg, ins->sreg1, 1, TRUE);
5330 case OP_FCONV_TO_U1:
5331 code = emit_float_to_int (cfg, code, ins->dreg, ins->sreg1, 1, FALSE);
5333 case OP_FCONV_TO_I2:
5334 code = emit_float_to_int (cfg, code, ins->dreg, ins->sreg1, 2, TRUE);
5336 case OP_FCONV_TO_U2:
5337 code = emit_float_to_int (cfg, code, ins->dreg, ins->sreg1, 2, FALSE);
5339 case OP_FCONV_TO_U4:
5340 code = emit_float_to_int (cfg, code, ins->dreg, ins->sreg1, 4, FALSE);
5342 case OP_FCONV_TO_I4:
5344 code = emit_float_to_int (cfg, code, ins->dreg, ins->sreg1, 4, TRUE);
5346 case OP_FCONV_TO_I8:
5347 code = emit_float_to_int (cfg, code, ins->dreg, ins->sreg1, 8, TRUE);
5350 case OP_RCONV_TO_I1:
5351 amd64_sse_cvtss2si_reg_reg_size (code, ins->dreg, ins->sreg1, 4);
5352 amd64_widen_reg (code, ins->dreg, ins->dreg, TRUE, FALSE);
5354 case OP_RCONV_TO_U1:
5355 amd64_sse_cvtss2si_reg_reg_size (code, ins->dreg, ins->sreg1, 4);
5356 amd64_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
5358 case OP_RCONV_TO_I2:
5359 amd64_sse_cvtss2si_reg_reg_size (code, ins->dreg, ins->sreg1, 4);
5360 amd64_widen_reg (code, ins->dreg, ins->dreg, TRUE, TRUE);
5362 case OP_RCONV_TO_U2:
5363 amd64_sse_cvtss2si_reg_reg_size (code, ins->dreg, ins->sreg1, 4);
5364 amd64_widen_reg (code, ins->dreg, ins->dreg, FALSE, TRUE);
5366 case OP_RCONV_TO_I4:
5367 amd64_sse_cvtss2si_reg_reg_size (code, ins->dreg, ins->sreg1, 4);
5369 case OP_RCONV_TO_U4:
5370 amd64_sse_cvtss2si_reg_reg_size (code, ins->dreg, ins->sreg1, 4);
5372 case OP_RCONV_TO_I8:
5373 amd64_sse_cvtss2si_reg_reg_size (code, ins->dreg, ins->sreg1, 8);
5375 case OP_RCONV_TO_R8:
5376 amd64_sse_cvtss2sd_reg_reg (code, ins->dreg, ins->sreg1);
5378 case OP_RCONV_TO_R4:
5379 if (ins->dreg != ins->sreg1)
5380 amd64_sse_movss_reg_reg (code, ins->dreg, ins->sreg1);
5383 case OP_LCONV_TO_R_UN: {
5386 /* Based on gcc code */
5387 amd64_test_reg_reg (code, ins->sreg1, ins->sreg1);
5388 br [0] = code; x86_branch8 (code, X86_CC_S, 0, TRUE);
5391 amd64_sse_cvtsi2sd_reg_reg (code, ins->dreg, ins->sreg1);
5392 br [1] = code; x86_jump8 (code, 0);
5393 amd64_patch (br [0], code);
5396 /* Save to the red zone */
5397 amd64_mov_membase_reg (code, AMD64_RSP, -8, AMD64_RAX, 8);
5398 amd64_mov_membase_reg (code, AMD64_RSP, -16, AMD64_RCX, 8);
5399 amd64_mov_reg_reg (code, AMD64_RCX, ins->sreg1, 8);
5400 amd64_mov_reg_reg (code, AMD64_RAX, ins->sreg1, 8);
5401 amd64_alu_reg_imm (code, X86_AND, AMD64_RCX, 1);
5402 amd64_shift_reg_imm (code, X86_SHR, AMD64_RAX, 1);
5403 amd64_alu_reg_imm (code, X86_OR, AMD64_RAX, AMD64_RCX);
5404 amd64_sse_cvtsi2sd_reg_reg (code, ins->dreg, AMD64_RAX);
5405 amd64_sse_addsd_reg_reg (code, ins->dreg, ins->dreg);
5407 amd64_mov_reg_membase (code, AMD64_RCX, AMD64_RSP, -16, 8);
5408 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RSP, -8, 8);
5409 amd64_patch (br [1], code);
5412 case OP_LCONV_TO_OVF_U4:
5413 amd64_alu_reg_imm (code, X86_CMP, ins->sreg1, 0);
5414 EMIT_COND_SYSTEM_EXCEPTION (X86_CC_LT, TRUE, "OverflowException");
5415 amd64_mov_reg_reg (code, ins->dreg, ins->sreg1, 8);
5417 case OP_LCONV_TO_OVF_I4_UN:
5418 amd64_alu_reg_imm (code, X86_CMP, ins->sreg1, 0x7fffffff);
5419 EMIT_COND_SYSTEM_EXCEPTION (X86_CC_GT, FALSE, "OverflowException");
5420 amd64_mov_reg_reg (code, ins->dreg, ins->sreg1, 8);
5423 if (ins->dreg != ins->sreg1)
5424 amd64_sse_movsd_reg_reg (code, ins->dreg, ins->sreg1);
5427 if (ins->dreg != ins->sreg1)
5428 amd64_sse_movss_reg_reg (code, ins->dreg, ins->sreg1);
5430 case OP_MOVE_F_TO_I4:
5432 amd64_movd_reg_xreg_size (code, ins->dreg, ins->sreg1, 8);
5434 amd64_sse_cvtsd2ss_reg_reg (code, MONO_ARCH_FP_SCRATCH_REG, ins->sreg1);
5435 amd64_movd_reg_xreg_size (code, ins->dreg, MONO_ARCH_FP_SCRATCH_REG, 8);
5438 case OP_MOVE_I4_TO_F:
5439 amd64_movd_xreg_reg_size (code, ins->dreg, ins->sreg1, 8);
5441 amd64_sse_cvtss2sd_reg_reg (code, ins->dreg, ins->dreg);
5443 case OP_MOVE_F_TO_I8:
5444 amd64_movd_reg_xreg_size (code, ins->dreg, ins->sreg1, 8);
5446 case OP_MOVE_I8_TO_F:
5447 amd64_movd_xreg_reg_size (code, ins->dreg, ins->sreg1, 8);
5450 amd64_sse_addsd_reg_reg (code, ins->dreg, ins->sreg2);
5453 amd64_sse_subsd_reg_reg (code, ins->dreg, ins->sreg2);
5456 amd64_sse_mulsd_reg_reg (code, ins->dreg, ins->sreg2);
5459 amd64_sse_divsd_reg_reg (code, ins->dreg, ins->sreg2);
5462 static double r8_0 = -0.0;
5464 g_assert (ins->sreg1 == ins->dreg);
5466 mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_R8, &r8_0);
5467 amd64_sse_xorpd_reg_membase (code, ins->dreg, AMD64_RIP, 0);
5471 EMIT_SSE2_FPFUNC (code, fsin, ins->dreg, ins->sreg1);
5474 EMIT_SSE2_FPFUNC (code, fcos, ins->dreg, ins->sreg1);
5477 static guint64 d = 0x7fffffffffffffffUL;
5479 g_assert (ins->sreg1 == ins->dreg);
5481 mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_R8, &d);
5482 amd64_sse_andpd_reg_membase (code, ins->dreg, AMD64_RIP, 0);
5486 EMIT_SSE2_FPFUNC (code, fsqrt, ins->dreg, ins->sreg1);
5490 amd64_sse_addss_reg_reg (code, ins->dreg, ins->sreg2);
5493 amd64_sse_subss_reg_reg (code, ins->dreg, ins->sreg2);
5496 amd64_sse_mulss_reg_reg (code, ins->dreg, ins->sreg2);
5499 amd64_sse_divss_reg_reg (code, ins->dreg, ins->sreg2);
5502 static float r4_0 = -0.0;
5504 g_assert (ins->sreg1 == ins->dreg);
5506 mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_R4, &r4_0);
5507 amd64_sse_movss_reg_membase (code, MONO_ARCH_FP_SCRATCH_REG, AMD64_RIP, 0);
5508 amd64_sse_xorps_reg_reg (code, ins->dreg, MONO_ARCH_FP_SCRATCH_REG);
5513 g_assert (cfg->opt & MONO_OPT_CMOV);
5514 g_assert (ins->dreg == ins->sreg1);
5515 amd64_alu_reg_reg_size (code, X86_CMP, ins->sreg1, ins->sreg2, 4);
5516 amd64_cmov_reg_size (code, X86_CC_GT, TRUE, ins->dreg, ins->sreg2, 4);
5519 g_assert (cfg->opt & MONO_OPT_CMOV);
5520 g_assert (ins->dreg == ins->sreg1);
5521 amd64_alu_reg_reg_size (code, X86_CMP, ins->sreg1, ins->sreg2, 4);
5522 amd64_cmov_reg_size (code, X86_CC_GT, FALSE, ins->dreg, ins->sreg2, 4);
5525 g_assert (cfg->opt & MONO_OPT_CMOV);
5526 g_assert (ins->dreg == ins->sreg1);
5527 amd64_alu_reg_reg_size (code, X86_CMP, ins->sreg1, ins->sreg2, 4);
5528 amd64_cmov_reg_size (code, X86_CC_LT, TRUE, ins->dreg, ins->sreg2, 4);
5531 g_assert (cfg->opt & MONO_OPT_CMOV);
5532 g_assert (ins->dreg == ins->sreg1);
5533 amd64_alu_reg_reg_size (code, X86_CMP, ins->sreg1, ins->sreg2, 4);
5534 amd64_cmov_reg_size (code, X86_CC_LT, FALSE, ins->dreg, ins->sreg2, 4);
5537 g_assert (cfg->opt & MONO_OPT_CMOV);
5538 g_assert (ins->dreg == ins->sreg1);
5539 amd64_alu_reg_reg (code, X86_CMP, ins->sreg1, ins->sreg2);
5540 amd64_cmov_reg (code, X86_CC_GT, TRUE, ins->dreg, ins->sreg2);
5543 g_assert (cfg->opt & MONO_OPT_CMOV);
5544 g_assert (ins->dreg == ins->sreg1);
5545 amd64_alu_reg_reg (code, X86_CMP, ins->sreg1, ins->sreg2);
5546 amd64_cmov_reg (code, X86_CC_GT, FALSE, ins->dreg, ins->sreg2);
5549 g_assert (cfg->opt & MONO_OPT_CMOV);
5550 g_assert (ins->dreg == ins->sreg1);
5551 amd64_alu_reg_reg (code, X86_CMP, ins->sreg1, ins->sreg2);
5552 amd64_cmov_reg (code, X86_CC_LT, TRUE, ins->dreg, ins->sreg2);
5555 g_assert (cfg->opt & MONO_OPT_CMOV);
5556 g_assert (ins->dreg == ins->sreg1);
5557 amd64_alu_reg_reg (code, X86_CMP, ins->sreg1, ins->sreg2);
5558 amd64_cmov_reg (code, X86_CC_LT, FALSE, ins->dreg, ins->sreg2);
5564 * The two arguments are swapped because the fbranch instructions
5565 * depend on this for the non-sse case to work.
5567 amd64_sse_comisd_reg_reg (code, ins->sreg2, ins->sreg1);
5571 * FIXME: Get rid of this.
5572 * The two arguments are swapped because the fbranch instructions
5573 * depend on this for the non-sse case to work.
5575 amd64_sse_comiss_reg_reg (code, ins->sreg2, ins->sreg1);
5579 /* zeroing the register at the start results in
5580 * shorter and faster code (we can also remove the widening op)
5582 guchar *unordered_check;
5584 amd64_alu_reg_reg (code, X86_XOR, ins->dreg, ins->dreg);
5585 amd64_sse_comisd_reg_reg (code, ins->sreg1, ins->sreg2);
5586 unordered_check = code;
5587 x86_branch8 (code, X86_CC_P, 0, FALSE);
5589 if (ins->opcode == OP_FCEQ) {
5590 amd64_set_reg (code, X86_CC_EQ, ins->dreg, FALSE);
5591 amd64_patch (unordered_check, code);
5593 guchar *jump_to_end;
5594 amd64_set_reg (code, X86_CC_NE, ins->dreg, FALSE);
5596 x86_jump8 (code, 0);
5597 amd64_patch (unordered_check, code);
5598 amd64_inc_reg (code, ins->dreg);
5599 amd64_patch (jump_to_end, code);
5605 /* zeroing the register at the start results in
5606 * shorter and faster code (we can also remove the widening op)
5608 amd64_alu_reg_reg (code, X86_XOR, ins->dreg, ins->dreg);
5609 amd64_sse_comisd_reg_reg (code, ins->sreg2, ins->sreg1);
5610 if (ins->opcode == OP_FCLT_UN) {
5611 guchar *unordered_check = code;
5612 guchar *jump_to_end;
5613 x86_branch8 (code, X86_CC_P, 0, FALSE);
5614 amd64_set_reg (code, X86_CC_GT, ins->dreg, FALSE);
5616 x86_jump8 (code, 0);
5617 amd64_patch (unordered_check, code);
5618 amd64_inc_reg (code, ins->dreg);
5619 amd64_patch (jump_to_end, code);
5621 amd64_set_reg (code, X86_CC_GT, ins->dreg, FALSE);
5626 guchar *unordered_check;
5627 amd64_alu_reg_reg (code, X86_XOR, ins->dreg, ins->dreg);
5628 amd64_sse_comisd_reg_reg (code, ins->sreg2, ins->sreg1);
5629 unordered_check = code;
5630 x86_branch8 (code, X86_CC_P, 0, FALSE);
5631 amd64_set_reg (code, X86_CC_NB, ins->dreg, FALSE);
5632 amd64_patch (unordered_check, code);
5637 /* zeroing the register at the start results in
5638 * shorter and faster code (we can also remove the widening op)
5640 guchar *unordered_check;
5642 amd64_alu_reg_reg (code, X86_XOR, ins->dreg, ins->dreg);
5643 amd64_sse_comisd_reg_reg (code, ins->sreg2, ins->sreg1);
5644 if (ins->opcode == OP_FCGT) {
5645 unordered_check = code;
5646 x86_branch8 (code, X86_CC_P, 0, FALSE);
5647 amd64_set_reg (code, X86_CC_LT, ins->dreg, FALSE);
5648 amd64_patch (unordered_check, code);
5650 amd64_set_reg (code, X86_CC_LT, ins->dreg, FALSE);
5655 guchar *unordered_check;
5656 amd64_alu_reg_reg (code, X86_XOR, ins->dreg, ins->dreg);
5657 amd64_sse_comisd_reg_reg (code, ins->sreg2, ins->sreg1);
5658 unordered_check = code;
5659 x86_branch8 (code, X86_CC_P, 0, FALSE);
5660 amd64_set_reg (code, X86_CC_NA, ins->dreg, FALSE);
5661 amd64_patch (unordered_check, code);
5671 gboolean unordered = FALSE;
5673 amd64_alu_reg_reg (code, X86_XOR, ins->dreg, ins->dreg);
5674 amd64_sse_comiss_reg_reg (code, ins->sreg2, ins->sreg1);
5676 switch (ins->opcode) {
5678 x86_cond = X86_CC_EQ;
5681 x86_cond = X86_CC_LT;
5684 x86_cond = X86_CC_GT;
5687 x86_cond = X86_CC_GT;
5691 x86_cond = X86_CC_LT;
5695 g_assert_not_reached ();
5700 guchar *unordered_check;
5701 guchar *jump_to_end;
5703 unordered_check = code;
5704 x86_branch8 (code, X86_CC_P, 0, FALSE);
5705 amd64_set_reg (code, x86_cond, ins->dreg, FALSE);
5707 x86_jump8 (code, 0);
5708 amd64_patch (unordered_check, code);
5709 amd64_inc_reg (code, ins->dreg);
5710 amd64_patch (jump_to_end, code);
5712 amd64_set_reg (code, x86_cond, ins->dreg, FALSE);
5716 case OP_FCLT_MEMBASE:
5717 case OP_FCGT_MEMBASE:
5718 case OP_FCLT_UN_MEMBASE:
5719 case OP_FCGT_UN_MEMBASE:
5720 case OP_FCEQ_MEMBASE: {
5721 guchar *unordered_check, *jump_to_end;
5724 amd64_alu_reg_reg (code, X86_XOR, ins->dreg, ins->dreg);
5725 amd64_sse_comisd_reg_membase (code, ins->sreg1, ins->sreg2, ins->inst_offset);
5727 switch (ins->opcode) {
5728 case OP_FCEQ_MEMBASE:
5729 x86_cond = X86_CC_EQ;
5731 case OP_FCLT_MEMBASE:
5732 case OP_FCLT_UN_MEMBASE:
5733 x86_cond = X86_CC_LT;
5735 case OP_FCGT_MEMBASE:
5736 case OP_FCGT_UN_MEMBASE:
5737 x86_cond = X86_CC_GT;
5740 g_assert_not_reached ();
5743 unordered_check = code;
5744 x86_branch8 (code, X86_CC_P, 0, FALSE);
5745 amd64_set_reg (code, x86_cond, ins->dreg, FALSE);
5747 switch (ins->opcode) {
5748 case OP_FCEQ_MEMBASE:
5749 case OP_FCLT_MEMBASE:
5750 case OP_FCGT_MEMBASE:
5751 amd64_patch (unordered_check, code);
5753 case OP_FCLT_UN_MEMBASE:
5754 case OP_FCGT_UN_MEMBASE:
5756 x86_jump8 (code, 0);
5757 amd64_patch (unordered_check, code);
5758 amd64_inc_reg (code, ins->dreg);
5759 amd64_patch (jump_to_end, code);
5767 guchar *jump = code;
5768 x86_branch8 (code, X86_CC_P, 0, TRUE);
5769 EMIT_COND_BRANCH (ins, X86_CC_EQ, FALSE);
5770 amd64_patch (jump, code);
5774 /* Branch if C013 != 100 */
5775 /* branch if !ZF or (PF|CF) */
5776 EMIT_COND_BRANCH (ins, X86_CC_NE, FALSE);
5777 EMIT_COND_BRANCH (ins, X86_CC_P, FALSE);
5778 EMIT_COND_BRANCH (ins, X86_CC_B, FALSE);
5781 EMIT_COND_BRANCH (ins, X86_CC_GT, FALSE);
5784 EMIT_COND_BRANCH (ins, X86_CC_P, FALSE);
5785 EMIT_COND_BRANCH (ins, X86_CC_GT, FALSE);
5789 if (ins->opcode == OP_FBGT) {
5792 /* skip branch if C1=1 */
5794 x86_branch8 (code, X86_CC_P, 0, FALSE);
5795 /* branch if (C0 | C3) = 1 */
5796 EMIT_COND_BRANCH (ins, X86_CC_LT, FALSE);
5797 amd64_patch (br1, code);
5800 EMIT_COND_BRANCH (ins, X86_CC_LT, FALSE);
5804 /* Branch if C013 == 100 or 001 */
5807 /* skip branch if C1=1 */
5809 x86_branch8 (code, X86_CC_P, 0, FALSE);
5810 /* branch if (C0 | C3) = 1 */
5811 EMIT_COND_BRANCH (ins, X86_CC_BE, FALSE);
5812 amd64_patch (br1, code);
5816 /* Branch if C013 == 000 */
5817 EMIT_COND_BRANCH (ins, X86_CC_LE, FALSE);
5820 /* Branch if C013=000 or 100 */
5823 /* skip branch if C1=1 */
5825 x86_branch8 (code, X86_CC_P, 0, FALSE);
5826 /* branch if C0=0 */
5827 EMIT_COND_BRANCH (ins, X86_CC_NB, FALSE);
5828 amd64_patch (br1, code);
5832 /* Branch if C013 != 001 */
5833 EMIT_COND_BRANCH (ins, X86_CC_P, FALSE);
5834 EMIT_COND_BRANCH (ins, X86_CC_GE, FALSE);
5837 /* Transfer value to the fp stack */
5838 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, 16);
5839 amd64_movsd_membase_reg (code, AMD64_RSP, 0, ins->sreg1);
5840 amd64_fld_membase (code, AMD64_RSP, 0, TRUE);
5842 amd64_push_reg (code, AMD64_RAX);
5844 amd64_fnstsw (code);
5845 amd64_alu_reg_imm (code, X86_AND, AMD64_RAX, 0x4100);
5846 amd64_alu_reg_imm (code, X86_CMP, AMD64_RAX, X86_FP_C0);
5847 amd64_pop_reg (code, AMD64_RAX);
5848 amd64_fstp (code, 0);
5849 EMIT_COND_SYSTEM_EXCEPTION (X86_CC_EQ, FALSE, "ArithmeticException");
5850 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, 16);
5853 code = mono_amd64_emit_tls_get (code, ins->dreg, ins->inst_offset);
5856 case OP_TLS_GET_REG:
5857 code = emit_tls_get_reg (code, ins->dreg, ins->sreg1);
5860 code = amd64_emit_tls_set (code, ins->sreg1, ins->inst_offset);
5863 case OP_TLS_SET_REG: {
5864 code = amd64_emit_tls_set_reg (code, ins->sreg1, ins->sreg2);
5867 case OP_MEMORY_BARRIER: {
5868 if (ins->backend.memory_barrier_kind == MONO_MEMORY_BARRIER_SEQ)
5872 case OP_ATOMIC_ADD_I4:
5873 case OP_ATOMIC_ADD_I8: {
5874 int dreg = ins->dreg;
5875 guint32 size = (ins->opcode == OP_ATOMIC_ADD_I4) ? 4 : 8;
5877 if ((dreg == ins->sreg2) || (dreg == ins->inst_basereg))
5880 amd64_mov_reg_reg (code, dreg, ins->sreg2, size);
5881 amd64_prefix (code, X86_LOCK_PREFIX);
5882 amd64_xadd_membase_reg (code, ins->inst_basereg, ins->inst_offset, dreg, size);
5883 /* dreg contains the old value, add with sreg2 value */
5884 amd64_alu_reg_reg_size (code, X86_ADD, dreg, ins->sreg2, size);
5886 if (ins->dreg != dreg)
5887 amd64_mov_reg_reg (code, ins->dreg, dreg, size);
5891 case OP_ATOMIC_EXCHANGE_I4:
5892 case OP_ATOMIC_EXCHANGE_I8: {
5893 guint32 size = ins->opcode == OP_ATOMIC_EXCHANGE_I4 ? 4 : 8;
5895 /* LOCK prefix is implied. */
5896 amd64_mov_reg_reg (code, GP_SCRATCH_REG, ins->sreg2, size);
5897 amd64_xchg_membase_reg_size (code, ins->sreg1, ins->inst_offset, GP_SCRATCH_REG, size);
5898 amd64_mov_reg_reg (code, ins->dreg, GP_SCRATCH_REG, size);
5901 case OP_ATOMIC_CAS_I4:
5902 case OP_ATOMIC_CAS_I8: {
5905 if (ins->opcode == OP_ATOMIC_CAS_I8)
5911 * See http://msdn.microsoft.com/en-us/magazine/cc302329.aspx for
5912 * an explanation of how this works.
5914 g_assert (ins->sreg3 == AMD64_RAX);
5915 g_assert (ins->sreg1 != AMD64_RAX);
5916 g_assert (ins->sreg1 != ins->sreg2);
5918 amd64_prefix (code, X86_LOCK_PREFIX);
5919 amd64_cmpxchg_membase_reg_size (code, ins->sreg1, ins->inst_offset, ins->sreg2, size);
5921 if (ins->dreg != AMD64_RAX)
5922 amd64_mov_reg_reg (code, ins->dreg, AMD64_RAX, size);
5925 case OP_ATOMIC_LOAD_I1: {
5926 amd64_widen_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, TRUE, FALSE);
5929 case OP_ATOMIC_LOAD_U1: {
5930 amd64_widen_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, FALSE, FALSE);
5933 case OP_ATOMIC_LOAD_I2: {
5934 amd64_widen_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, TRUE, TRUE);
5937 case OP_ATOMIC_LOAD_U2: {
5938 amd64_widen_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, FALSE, TRUE);
5941 case OP_ATOMIC_LOAD_I4: {
5942 amd64_movsxd_reg_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset);
5945 case OP_ATOMIC_LOAD_U4:
5946 case OP_ATOMIC_LOAD_I8:
5947 case OP_ATOMIC_LOAD_U8: {
5948 amd64_mov_reg_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, ins->opcode == OP_ATOMIC_LOAD_U4 ? 4 : 8);
5951 case OP_ATOMIC_LOAD_R4: {
5952 amd64_sse_movss_reg_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset);
5953 amd64_sse_cvtss2sd_reg_reg (code, ins->dreg, ins->dreg);
5956 case OP_ATOMIC_LOAD_R8: {
5957 amd64_sse_movsd_reg_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset);
5960 case OP_ATOMIC_STORE_I1:
5961 case OP_ATOMIC_STORE_U1:
5962 case OP_ATOMIC_STORE_I2:
5963 case OP_ATOMIC_STORE_U2:
5964 case OP_ATOMIC_STORE_I4:
5965 case OP_ATOMIC_STORE_U4:
5966 case OP_ATOMIC_STORE_I8:
5967 case OP_ATOMIC_STORE_U8: {
5970 switch (ins->opcode) {
5971 case OP_ATOMIC_STORE_I1:
5972 case OP_ATOMIC_STORE_U1:
5975 case OP_ATOMIC_STORE_I2:
5976 case OP_ATOMIC_STORE_U2:
5979 case OP_ATOMIC_STORE_I4:
5980 case OP_ATOMIC_STORE_U4:
5983 case OP_ATOMIC_STORE_I8:
5984 case OP_ATOMIC_STORE_U8:
5989 amd64_mov_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, ins->sreg1, size);
5991 if (ins->backend.memory_barrier_kind == MONO_MEMORY_BARRIER_SEQ)
5995 case OP_ATOMIC_STORE_R4: {
5996 amd64_sse_cvtsd2ss_reg_reg (code, MONO_ARCH_FP_SCRATCH_REG, ins->sreg1);
5997 amd64_sse_movss_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, MONO_ARCH_FP_SCRATCH_REG);
5999 if (ins->backend.memory_barrier_kind == MONO_MEMORY_BARRIER_SEQ)
6003 case OP_ATOMIC_STORE_R8: {
6006 amd64_sse_movsd_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, ins->sreg1);
6010 if (ins->backend.memory_barrier_kind == MONO_MEMORY_BARRIER_SEQ)
6014 case OP_CARD_TABLE_WBARRIER: {
6015 int ptr = ins->sreg1;
6016 int value = ins->sreg2;
6018 int nursery_shift, card_table_shift;
6019 gpointer card_table_mask;
6020 size_t nursery_size;
6022 gpointer card_table = mono_gc_get_card_table (&card_table_shift, &card_table_mask);
6023 guint64 nursery_start = (guint64)mono_gc_get_nursery (&nursery_shift, &nursery_size);
6024 guint64 shifted_nursery_start = nursery_start >> nursery_shift;
6026 /*If either point to the stack we can simply avoid the WB. This happens due to
6027 * optimizations revealing a stack store that was not visible when op_cardtable was emited.
6029 if (ins->sreg1 == AMD64_RSP || ins->sreg2 == AMD64_RSP)
6033 * We need one register we can clobber, we choose EDX and make sreg1
6034 * fixed EAX to work around limitations in the local register allocator.
6035 * sreg2 might get allocated to EDX, but that is not a problem since
6036 * we use it before clobbering EDX.
6038 g_assert (ins->sreg1 == AMD64_RAX);
6041 * This is the code we produce:
6044 * edx >>= nursery_shift
6045 * cmp edx, (nursery_start >> nursery_shift)
6048 * edx >>= card_table_shift
6054 if (mono_gc_card_table_nursery_check ()) {
6055 if (value != AMD64_RDX)
6056 amd64_mov_reg_reg (code, AMD64_RDX, value, 8);
6057 amd64_shift_reg_imm (code, X86_SHR, AMD64_RDX, nursery_shift);
6058 if (shifted_nursery_start >> 31) {
6060 * The value we need to compare against is 64 bits, so we need
6061 * another spare register. We use RBX, which we save and
6064 amd64_mov_membase_reg (code, AMD64_RSP, -8, AMD64_RBX, 8);
6065 amd64_mov_reg_imm (code, AMD64_RBX, shifted_nursery_start);
6066 amd64_alu_reg_reg (code, X86_CMP, AMD64_RDX, AMD64_RBX);
6067 amd64_mov_reg_membase (code, AMD64_RBX, AMD64_RSP, -8, 8);
6069 amd64_alu_reg_imm (code, X86_CMP, AMD64_RDX, shifted_nursery_start);
6071 br = code; x86_branch8 (code, X86_CC_NE, -1, FALSE);
6073 amd64_mov_reg_reg (code, AMD64_RDX, ptr, 8);
6074 amd64_shift_reg_imm (code, X86_SHR, AMD64_RDX, card_table_shift);
6075 if (card_table_mask)
6076 amd64_alu_reg_imm (code, X86_AND, AMD64_RDX, (guint32)(guint64)card_table_mask);
6078 mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_GC_CARD_TABLE_ADDR, card_table);
6079 amd64_alu_reg_membase (code, X86_ADD, AMD64_RDX, AMD64_RIP, 0);
6081 amd64_mov_membase_imm (code, AMD64_RDX, 0, 1, 1);
6083 if (mono_gc_card_table_nursery_check ())
6084 x86_patch (br, code);
6087 #ifdef MONO_ARCH_SIMD_INTRINSICS
6088 /* TODO: Some of these IR opcodes are marked as no clobber when they indeed do. */
6090 amd64_sse_addps_reg_reg (code, ins->sreg1, ins->sreg2);
6093 amd64_sse_divps_reg_reg (code, ins->sreg1, ins->sreg2);
6096 amd64_sse_mulps_reg_reg (code, ins->sreg1, ins->sreg2);
6099 amd64_sse_subps_reg_reg (code, ins->sreg1, ins->sreg2);
6102 amd64_sse_maxps_reg_reg (code, ins->sreg1, ins->sreg2);
6105 amd64_sse_minps_reg_reg (code, ins->sreg1, ins->sreg2);
6108 g_assert (ins->inst_c0 >= 0 && ins->inst_c0 <= 7);
6109 amd64_sse_cmpps_reg_reg_imm (code, ins->sreg1, ins->sreg2, ins->inst_c0);
6112 amd64_sse_andps_reg_reg (code, ins->sreg1, ins->sreg2);
6115 amd64_sse_andnps_reg_reg (code, ins->sreg1, ins->sreg2);
6118 amd64_sse_orps_reg_reg (code, ins->sreg1, ins->sreg2);
6121 amd64_sse_xorps_reg_reg (code, ins->sreg1, ins->sreg2);
6124 amd64_sse_sqrtps_reg_reg (code, ins->dreg, ins->sreg1);
6127 amd64_sse_rsqrtps_reg_reg (code, ins->dreg, ins->sreg1);
6130 amd64_sse_rcpps_reg_reg (code, ins->dreg, ins->sreg1);
6133 amd64_sse_addsubps_reg_reg (code, ins->sreg1, ins->sreg2);
6136 amd64_sse_haddps_reg_reg (code, ins->sreg1, ins->sreg2);
6139 amd64_sse_hsubps_reg_reg (code, ins->sreg1, ins->sreg2);
6142 amd64_sse_movshdup_reg_reg (code, ins->dreg, ins->sreg1);
6145 amd64_sse_movsldup_reg_reg (code, ins->dreg, ins->sreg1);
6148 case OP_PSHUFLEW_HIGH:
6149 g_assert (ins->inst_c0 >= 0 && ins->inst_c0 <= 0xFF);
6150 amd64_sse_pshufhw_reg_reg_imm (code, ins->dreg, ins->sreg1, ins->inst_c0);
6152 case OP_PSHUFLEW_LOW:
6153 g_assert (ins->inst_c0 >= 0 && ins->inst_c0 <= 0xFF);
6154 amd64_sse_pshuflw_reg_reg_imm (code, ins->dreg, ins->sreg1, ins->inst_c0);
6157 g_assert (ins->inst_c0 >= 0 && ins->inst_c0 <= 0xFF);
6158 amd64_sse_pshufd_reg_reg_imm (code, ins->dreg, ins->sreg1, ins->inst_c0);
6161 g_assert (ins->inst_c0 >= 0 && ins->inst_c0 <= 0xFF);
6162 amd64_sse_shufps_reg_reg_imm (code, ins->sreg1, ins->sreg2, ins->inst_c0);
6165 g_assert (ins->inst_c0 >= 0 && ins->inst_c0 <= 0x3);
6166 amd64_sse_shufpd_reg_reg_imm (code, ins->sreg1, ins->sreg2, ins->inst_c0);
6170 amd64_sse_addpd_reg_reg (code, ins->sreg1, ins->sreg2);
6173 amd64_sse_divpd_reg_reg (code, ins->sreg1, ins->sreg2);
6176 amd64_sse_mulpd_reg_reg (code, ins->sreg1, ins->sreg2);
6179 amd64_sse_subpd_reg_reg (code, ins->sreg1, ins->sreg2);
6182 amd64_sse_maxpd_reg_reg (code, ins->sreg1, ins->sreg2);
6185 amd64_sse_minpd_reg_reg (code, ins->sreg1, ins->sreg2);
6188 g_assert (ins->inst_c0 >= 0 && ins->inst_c0 <= 7);
6189 amd64_sse_cmppd_reg_reg_imm (code, ins->sreg1, ins->sreg2, ins->inst_c0);
6192 amd64_sse_andpd_reg_reg (code, ins->sreg1, ins->sreg2);
6195 amd64_sse_andnpd_reg_reg (code, ins->sreg1, ins->sreg2);
6198 amd64_sse_orpd_reg_reg (code, ins->sreg1, ins->sreg2);
6201 amd64_sse_xorpd_reg_reg (code, ins->sreg1, ins->sreg2);
6204 amd64_sse_sqrtpd_reg_reg (code, ins->dreg, ins->sreg1);
6207 amd64_sse_addsubpd_reg_reg (code, ins->sreg1, ins->sreg2);
6210 amd64_sse_haddpd_reg_reg (code, ins->sreg1, ins->sreg2);
6213 amd64_sse_hsubpd_reg_reg (code, ins->sreg1, ins->sreg2);
6216 amd64_sse_movddup_reg_reg (code, ins->dreg, ins->sreg1);
6219 case OP_EXTRACT_MASK:
6220 amd64_sse_pmovmskb_reg_reg (code, ins->dreg, ins->sreg1);
6224 amd64_sse_pand_reg_reg (code, ins->sreg1, ins->sreg2);
6227 amd64_sse_por_reg_reg (code, ins->sreg1, ins->sreg2);
6230 amd64_sse_pxor_reg_reg (code, ins->sreg1, ins->sreg2);
6234 amd64_sse_paddb_reg_reg (code, ins->sreg1, ins->sreg2);
6237 amd64_sse_paddw_reg_reg (code, ins->sreg1, ins->sreg2);
6240 amd64_sse_paddd_reg_reg (code, ins->sreg1, ins->sreg2);
6243 amd64_sse_paddq_reg_reg (code, ins->sreg1, ins->sreg2);
6247 amd64_sse_psubb_reg_reg (code, ins->sreg1, ins->sreg2);
6250 amd64_sse_psubw_reg_reg (code, ins->sreg1, ins->sreg2);
6253 amd64_sse_psubd_reg_reg (code, ins->sreg1, ins->sreg2);
6256 amd64_sse_psubq_reg_reg (code, ins->sreg1, ins->sreg2);
6260 amd64_sse_pmaxub_reg_reg (code, ins->sreg1, ins->sreg2);
6263 amd64_sse_pmaxuw_reg_reg (code, ins->sreg1, ins->sreg2);
6266 amd64_sse_pmaxud_reg_reg (code, ins->sreg1, ins->sreg2);
6270 amd64_sse_pmaxsb_reg_reg (code, ins->sreg1, ins->sreg2);
6273 amd64_sse_pmaxsw_reg_reg (code, ins->sreg1, ins->sreg2);
6276 amd64_sse_pmaxsd_reg_reg (code, ins->sreg1, ins->sreg2);
6280 amd64_sse_pavgb_reg_reg (code, ins->sreg1, ins->sreg2);
6283 amd64_sse_pavgw_reg_reg (code, ins->sreg1, ins->sreg2);
6287 amd64_sse_pminub_reg_reg (code, ins->sreg1, ins->sreg2);
6290 amd64_sse_pminuw_reg_reg (code, ins->sreg1, ins->sreg2);
6293 amd64_sse_pminud_reg_reg (code, ins->sreg1, ins->sreg2);
6297 amd64_sse_pminsb_reg_reg (code, ins->sreg1, ins->sreg2);
6300 amd64_sse_pminsw_reg_reg (code, ins->sreg1, ins->sreg2);
6303 amd64_sse_pminsd_reg_reg (code, ins->sreg1, ins->sreg2);
6307 amd64_sse_pcmpeqb_reg_reg (code, ins->sreg1, ins->sreg2);
6310 amd64_sse_pcmpeqw_reg_reg (code, ins->sreg1, ins->sreg2);
6313 amd64_sse_pcmpeqd_reg_reg (code, ins->sreg1, ins->sreg2);
6316 amd64_sse_pcmpeqq_reg_reg (code, ins->sreg1, ins->sreg2);
6320 amd64_sse_pcmpgtb_reg_reg (code, ins->sreg1, ins->sreg2);
6323 amd64_sse_pcmpgtw_reg_reg (code, ins->sreg1, ins->sreg2);
6326 amd64_sse_pcmpgtd_reg_reg (code, ins->sreg1, ins->sreg2);
6329 amd64_sse_pcmpgtq_reg_reg (code, ins->sreg1, ins->sreg2);
6332 case OP_PSUM_ABS_DIFF:
6333 amd64_sse_psadbw_reg_reg (code, ins->sreg1, ins->sreg2);
6336 case OP_UNPACK_LOWB:
6337 amd64_sse_punpcklbw_reg_reg (code, ins->sreg1, ins->sreg2);
6339 case OP_UNPACK_LOWW:
6340 amd64_sse_punpcklwd_reg_reg (code, ins->sreg1, ins->sreg2);
6342 case OP_UNPACK_LOWD:
6343 amd64_sse_punpckldq_reg_reg (code, ins->sreg1, ins->sreg2);
6345 case OP_UNPACK_LOWQ:
6346 amd64_sse_punpcklqdq_reg_reg (code, ins->sreg1, ins->sreg2);
6348 case OP_UNPACK_LOWPS:
6349 amd64_sse_unpcklps_reg_reg (code, ins->sreg1, ins->sreg2);
6351 case OP_UNPACK_LOWPD:
6352 amd64_sse_unpcklpd_reg_reg (code, ins->sreg1, ins->sreg2);
6355 case OP_UNPACK_HIGHB:
6356 amd64_sse_punpckhbw_reg_reg (code, ins->sreg1, ins->sreg2);
6358 case OP_UNPACK_HIGHW:
6359 amd64_sse_punpckhwd_reg_reg (code, ins->sreg1, ins->sreg2);
6361 case OP_UNPACK_HIGHD:
6362 amd64_sse_punpckhdq_reg_reg (code, ins->sreg1, ins->sreg2);
6364 case OP_UNPACK_HIGHQ:
6365 amd64_sse_punpckhqdq_reg_reg (code, ins->sreg1, ins->sreg2);
6367 case OP_UNPACK_HIGHPS:
6368 amd64_sse_unpckhps_reg_reg (code, ins->sreg1, ins->sreg2);
6370 case OP_UNPACK_HIGHPD:
6371 amd64_sse_unpckhpd_reg_reg (code, ins->sreg1, ins->sreg2);
6375 amd64_sse_packsswb_reg_reg (code, ins->sreg1, ins->sreg2);
6378 amd64_sse_packssdw_reg_reg (code, ins->sreg1, ins->sreg2);
6381 amd64_sse_packuswb_reg_reg (code, ins->sreg1, ins->sreg2);
6384 amd64_sse_packusdw_reg_reg (code, ins->sreg1, ins->sreg2);
6387 case OP_PADDB_SAT_UN:
6388 amd64_sse_paddusb_reg_reg (code, ins->sreg1, ins->sreg2);
6390 case OP_PSUBB_SAT_UN:
6391 amd64_sse_psubusb_reg_reg (code, ins->sreg1, ins->sreg2);
6393 case OP_PADDW_SAT_UN:
6394 amd64_sse_paddusw_reg_reg (code, ins->sreg1, ins->sreg2);
6396 case OP_PSUBW_SAT_UN:
6397 amd64_sse_psubusw_reg_reg (code, ins->sreg1, ins->sreg2);
6401 amd64_sse_paddsb_reg_reg (code, ins->sreg1, ins->sreg2);
6404 amd64_sse_psubsb_reg_reg (code, ins->sreg1, ins->sreg2);
6407 amd64_sse_paddsw_reg_reg (code, ins->sreg1, ins->sreg2);
6410 amd64_sse_psubsw_reg_reg (code, ins->sreg1, ins->sreg2);
6414 amd64_sse_pmullw_reg_reg (code, ins->sreg1, ins->sreg2);
6417 amd64_sse_pmulld_reg_reg (code, ins->sreg1, ins->sreg2);
6420 amd64_sse_pmuludq_reg_reg (code, ins->sreg1, ins->sreg2);
6422 case OP_PMULW_HIGH_UN:
6423 amd64_sse_pmulhuw_reg_reg (code, ins->sreg1, ins->sreg2);
6426 amd64_sse_pmulhw_reg_reg (code, ins->sreg1, ins->sreg2);
6430 amd64_sse_psrlw_reg_imm (code, ins->dreg, ins->inst_imm);
6433 amd64_sse_psrlw_reg_reg (code, ins->dreg, ins->sreg2);
6437 amd64_sse_psraw_reg_imm (code, ins->dreg, ins->inst_imm);
6440 amd64_sse_psraw_reg_reg (code, ins->dreg, ins->sreg2);
6444 amd64_sse_psllw_reg_imm (code, ins->dreg, ins->inst_imm);
6447 amd64_sse_psllw_reg_reg (code, ins->dreg, ins->sreg2);
6451 amd64_sse_psrld_reg_imm (code, ins->dreg, ins->inst_imm);
6454 amd64_sse_psrld_reg_reg (code, ins->dreg, ins->sreg2);
6458 amd64_sse_psrad_reg_imm (code, ins->dreg, ins->inst_imm);
6461 amd64_sse_psrad_reg_reg (code, ins->dreg, ins->sreg2);
6465 amd64_sse_pslld_reg_imm (code, ins->dreg, ins->inst_imm);
6468 amd64_sse_pslld_reg_reg (code, ins->dreg, ins->sreg2);
6472 amd64_sse_psrlq_reg_imm (code, ins->dreg, ins->inst_imm);
6475 amd64_sse_psrlq_reg_reg (code, ins->dreg, ins->sreg2);
6478 /*TODO: This is appart of the sse spec but not added
6480 amd64_sse_psraq_reg_imm (code, ins->dreg, ins->inst_imm);
6483 amd64_sse_psraq_reg_reg (code, ins->dreg, ins->sreg2);
6488 amd64_sse_psllq_reg_imm (code, ins->dreg, ins->inst_imm);
6491 amd64_sse_psllq_reg_reg (code, ins->dreg, ins->sreg2);
6494 amd64_sse_cvtdq2pd_reg_reg (code, ins->dreg, ins->sreg1);
6497 amd64_sse_cvtdq2ps_reg_reg (code, ins->dreg, ins->sreg1);
6500 amd64_sse_cvtpd2dq_reg_reg (code, ins->dreg, ins->sreg1);
6503 amd64_sse_cvtpd2ps_reg_reg (code, ins->dreg, ins->sreg1);
6506 amd64_sse_cvtps2dq_reg_reg (code, ins->dreg, ins->sreg1);
6509 amd64_sse_cvtps2pd_reg_reg (code, ins->dreg, ins->sreg1);
6512 amd64_sse_cvttpd2dq_reg_reg (code, ins->dreg, ins->sreg1);
6515 amd64_sse_cvttps2dq_reg_reg (code, ins->dreg, ins->sreg1);
6519 amd64_movd_xreg_reg_size (code, ins->dreg, ins->sreg1, 4);
6522 amd64_movd_reg_xreg_size (code, ins->dreg, ins->sreg1, 4);
6526 amd64_movhlps_reg_reg (code, MONO_ARCH_FP_SCRATCH_REG, ins->sreg1);
6527 amd64_movd_reg_xreg_size (code, ins->dreg, MONO_ARCH_FP_SCRATCH_REG, 8);
6529 amd64_movd_reg_xreg_size (code, ins->dreg, ins->sreg1, 8);
6534 amd64_movd_reg_xreg_size (code, ins->dreg, ins->sreg1, 4);
6536 amd64_shift_reg_imm (code, X86_SHR, ins->dreg, ins->inst_c0 * 8);
6537 amd64_widen_reg (code, ins->dreg, ins->dreg, ins->opcode == OP_EXTRACT_I1, FALSE);
6541 /*amd64_movd_reg_xreg_size (code, ins->dreg, ins->sreg1, 4);
6543 amd64_shift_reg_imm_size (code, X86_SHR, ins->dreg, 16, 4);*/
6544 amd64_sse_pextrw_reg_reg_imm (code, ins->dreg, ins->sreg1, ins->inst_c0);
6545 amd64_widen_reg_size (code, ins->dreg, ins->dreg, ins->opcode == OP_EXTRACT_I2, TRUE, 4);
6549 amd64_movhlps_reg_reg (code, ins->dreg, ins->sreg1);
6551 amd64_sse_movsd_reg_reg (code, ins->dreg, ins->sreg1);
6554 amd64_sse_pinsrw_reg_reg_imm (code, ins->sreg1, ins->sreg2, ins->inst_c0);
6556 case OP_EXTRACTX_U2:
6557 amd64_sse_pextrw_reg_reg_imm (code, ins->dreg, ins->sreg1, ins->inst_c0);
6559 case OP_INSERTX_U1_SLOW:
6560 /*sreg1 is the extracted ireg (scratch)
6561 /sreg2 is the to be inserted ireg (scratch)
6562 /dreg is the xreg to receive the value*/
6564 /*clear the bits from the extracted word*/
6565 amd64_alu_reg_imm (code, X86_AND, ins->sreg1, ins->inst_c0 & 1 ? 0x00FF : 0xFF00);
6566 /*shift the value to insert if needed*/
6567 if (ins->inst_c0 & 1)
6568 amd64_shift_reg_imm_size (code, X86_SHL, ins->sreg2, 8, 4);
6569 /*join them together*/
6570 amd64_alu_reg_reg (code, X86_OR, ins->sreg1, ins->sreg2);
6571 amd64_sse_pinsrw_reg_reg_imm (code, ins->dreg, ins->sreg1, ins->inst_c0 / 2);
6573 case OP_INSERTX_I4_SLOW:
6574 amd64_sse_pinsrw_reg_reg_imm (code, ins->dreg, ins->sreg2, ins->inst_c0 * 2);
6575 amd64_shift_reg_imm (code, X86_SHR, ins->sreg2, 16);
6576 amd64_sse_pinsrw_reg_reg_imm (code, ins->dreg, ins->sreg2, ins->inst_c0 * 2 + 1);
6578 case OP_INSERTX_I8_SLOW:
6579 amd64_movd_xreg_reg_size(code, MONO_ARCH_FP_SCRATCH_REG, ins->sreg2, 8);
6581 amd64_movlhps_reg_reg (code, ins->dreg, MONO_ARCH_FP_SCRATCH_REG);
6583 amd64_sse_movsd_reg_reg (code, ins->dreg, MONO_ARCH_FP_SCRATCH_REG);
6586 case OP_INSERTX_R4_SLOW:
6587 switch (ins->inst_c0) {
6590 amd64_sse_movss_reg_reg (code, ins->dreg, ins->sreg2);
6592 amd64_sse_cvtsd2ss_reg_reg (code, ins->dreg, ins->sreg2);
6595 amd64_sse_pshufd_reg_reg_imm (code, ins->dreg, ins->dreg, mono_simd_shuffle_mask(1, 0, 2, 3));
6597 amd64_sse_movss_reg_reg (code, ins->dreg, ins->sreg2);
6599 amd64_sse_cvtsd2ss_reg_reg (code, ins->dreg, ins->sreg2);
6600 amd64_sse_pshufd_reg_reg_imm (code, ins->dreg, ins->dreg, mono_simd_shuffle_mask(1, 0, 2, 3));
6603 amd64_sse_pshufd_reg_reg_imm (code, ins->dreg, ins->dreg, mono_simd_shuffle_mask(2, 1, 0, 3));
6605 amd64_sse_movss_reg_reg (code, ins->dreg, ins->sreg2);
6607 amd64_sse_cvtsd2ss_reg_reg (code, ins->dreg, ins->sreg2);
6608 amd64_sse_pshufd_reg_reg_imm (code, ins->dreg, ins->dreg, mono_simd_shuffle_mask(2, 1, 0, 3));
6611 amd64_sse_pshufd_reg_reg_imm (code, ins->dreg, ins->dreg, mono_simd_shuffle_mask(3, 1, 2, 0));
6613 amd64_sse_movss_reg_reg (code, ins->dreg, ins->sreg2);
6615 amd64_sse_cvtsd2ss_reg_reg (code, ins->dreg, ins->sreg2);
6616 amd64_sse_pshufd_reg_reg_imm (code, ins->dreg, ins->dreg, mono_simd_shuffle_mask(3, 1, 2, 0));
6620 case OP_INSERTX_R8_SLOW:
6622 amd64_movlhps_reg_reg (code, ins->dreg, ins->sreg2);
6624 amd64_sse_movsd_reg_reg (code, ins->dreg, ins->sreg2);
6626 case OP_STOREX_MEMBASE_REG:
6627 case OP_STOREX_MEMBASE:
6628 amd64_sse_movups_membase_reg (code, ins->dreg, ins->inst_offset, ins->sreg1);
6630 case OP_LOADX_MEMBASE:
6631 amd64_sse_movups_reg_membase (code, ins->dreg, ins->sreg1, ins->inst_offset);
6633 case OP_LOADX_ALIGNED_MEMBASE:
6634 amd64_sse_movaps_reg_membase (code, ins->dreg, ins->sreg1, ins->inst_offset);
6636 case OP_STOREX_ALIGNED_MEMBASE_REG:
6637 amd64_sse_movaps_membase_reg (code, ins->dreg, ins->inst_offset, ins->sreg1);
6639 case OP_STOREX_NTA_MEMBASE_REG:
6640 amd64_sse_movntps_reg_membase (code, ins->dreg, ins->sreg1, ins->inst_offset);
6642 case OP_PREFETCH_MEMBASE:
6643 amd64_sse_prefetch_reg_membase (code, ins->backend.arg_info, ins->sreg1, ins->inst_offset);
6647 /*FIXME the peephole pass should have killed this*/
6648 if (ins->dreg != ins->sreg1)
6649 amd64_sse_movaps_reg_reg (code, ins->dreg, ins->sreg1);
6652 amd64_sse_pxor_reg_reg (code, ins->dreg, ins->dreg);
6654 case OP_ICONV_TO_R4_RAW:
6655 amd64_movd_xreg_reg_size (code, ins->dreg, ins->sreg1, 4);
6658 case OP_FCONV_TO_R8_X:
6659 amd64_sse_movsd_reg_reg (code, ins->dreg, ins->sreg1);
6662 case OP_XCONV_R8_TO_I4:
6663 amd64_sse_cvttsd2si_reg_xreg_size (code, ins->dreg, ins->sreg1, 4);
6664 switch (ins->backend.source_opcode) {
6665 case OP_FCONV_TO_I1:
6666 amd64_widen_reg (code, ins->dreg, ins->dreg, TRUE, FALSE);
6668 case OP_FCONV_TO_U1:
6669 amd64_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
6671 case OP_FCONV_TO_I2:
6672 amd64_widen_reg (code, ins->dreg, ins->dreg, TRUE, TRUE);
6674 case OP_FCONV_TO_U2:
6675 amd64_widen_reg (code, ins->dreg, ins->dreg, FALSE, TRUE);
6681 amd64_sse_pinsrw_reg_reg_imm (code, ins->dreg, ins->sreg1, 0);
6682 amd64_sse_pinsrw_reg_reg_imm (code, ins->dreg, ins->sreg1, 1);
6683 amd64_sse_pshufd_reg_reg_imm (code, ins->dreg, ins->dreg, 0);
6686 amd64_movd_xreg_reg_size (code, ins->dreg, ins->sreg1, 4);
6687 amd64_sse_pshufd_reg_reg_imm (code, ins->dreg, ins->dreg, 0);
6690 amd64_movd_xreg_reg_size (code, ins->dreg, ins->sreg1, 8);
6691 amd64_sse_pshufd_reg_reg_imm (code, ins->dreg, ins->dreg, 0x44);
6695 amd64_sse_movsd_reg_reg (code, ins->dreg, ins->sreg1);
6697 amd64_sse_movsd_reg_reg (code, ins->dreg, ins->sreg1);
6698 amd64_sse_cvtsd2ss_reg_reg (code, ins->dreg, ins->dreg);
6700 amd64_sse_pshufd_reg_reg_imm (code, ins->dreg, ins->dreg, 0);
6703 amd64_sse_movsd_reg_reg (code, ins->dreg, ins->sreg1);
6704 amd64_sse_pshufd_reg_reg_imm (code, ins->dreg, ins->dreg, 0x44);
6707 case OP_LIVERANGE_START: {
6708 if (cfg->verbose_level > 1)
6709 printf ("R%d START=0x%x\n", MONO_VARINFO (cfg, ins->inst_c0)->vreg, (int)(code - cfg->native_code));
6710 MONO_VARINFO (cfg, ins->inst_c0)->live_range_start = code - cfg->native_code;
6713 case OP_LIVERANGE_END: {
6714 if (cfg->verbose_level > 1)
6715 printf ("R%d END=0x%x\n", MONO_VARINFO (cfg, ins->inst_c0)->vreg, (int)(code - cfg->native_code));
6716 MONO_VARINFO (cfg, ins->inst_c0)->live_range_end = code - cfg->native_code;
6719 case OP_GC_SAFE_POINT: {
6720 const char *polling_func = NULL;
6721 int compare_val = 0;
6724 #if defined(__native_client_codegen__) && defined(__native_client_gc__)
6725 polling_func = "mono_nacl_gc";
6726 compare_val = 0xFFFFFFFF;
6728 g_assert (mono_threads_is_coop_enabled ());
6729 polling_func = "mono_threads_state_poll";
6733 amd64_test_membase_imm_size (code, ins->sreg1, 0, compare_val, 4);
6734 br[0] = code; x86_branch8 (code, X86_CC_EQ, 0, FALSE);
6735 code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD, polling_func, FALSE);
6736 amd64_patch (br[0], code);
6740 case OP_GC_LIVENESS_DEF:
6741 case OP_GC_LIVENESS_USE:
6742 case OP_GC_PARAM_SLOT_LIVENESS_DEF:
6743 ins->backend.pc_offset = code - cfg->native_code;
6745 case OP_GC_SPILL_SLOT_LIVENESS_DEF:
6746 ins->backend.pc_offset = code - cfg->native_code;
6747 bb->spill_slot_defs = g_slist_prepend_mempool (cfg->mempool, bb->spill_slot_defs, ins);
6750 g_warning ("unknown opcode %s in %s()\n", mono_inst_name (ins->opcode), __FUNCTION__);
6751 g_assert_not_reached ();
6754 if ((code - cfg->native_code - offset) > max_len) {
6755 #if !defined(__native_client_codegen__)
6756 g_warning ("wrong maximal instruction length of instruction %s (expected %d, got %ld)",
6757 mono_inst_name (ins->opcode), max_len, code - cfg->native_code - offset);
6758 g_assert_not_reached ();
6763 cfg->code_len = code - cfg->native_code;
6766 #endif /* DISABLE_JIT */
6769 mono_arch_register_lowlevel_calls (void)
6771 /* The signature doesn't matter */
6772 mono_register_jit_icall (mono_amd64_throw_exception, "mono_amd64_throw_exception", mono_create_icall_signature ("void"), TRUE);
6776 mono_arch_patch_code_new (MonoCompile *cfg, MonoDomain *domain, guint8 *code, MonoJumpInfo *ji, gpointer target)
6778 unsigned char *ip = ji->ip.i + code;
6781 * Debug code to help track down problems where the target of a near call is
6784 if (amd64_is_near_call (ip)) {
6785 gint64 disp = (guint8*)target - (guint8*)ip;
6787 if (!amd64_is_imm32 (disp)) {
6788 printf ("TYPE: %d\n", ji->type);
6790 case MONO_PATCH_INFO_INTERNAL_METHOD:
6791 printf ("V: %s\n", ji->data.name);
6793 case MONO_PATCH_INFO_METHOD_JUMP:
6794 case MONO_PATCH_INFO_METHOD:
6795 printf ("V: %s\n", ji->data.method->name);
6803 amd64_patch (ip, (gpointer)target);
6809 get_max_epilog_size (MonoCompile *cfg)
6811 int max_epilog_size = 16;
6813 if (cfg->method->save_lmf)
6814 max_epilog_size += 256;
6816 if (mono_jit_trace_calls != NULL)
6817 max_epilog_size += 50;
6819 if (cfg->prof_options & MONO_PROFILE_ENTER_LEAVE)
6820 max_epilog_size += 50;
6822 max_epilog_size += (AMD64_NREG * 2);
6824 return max_epilog_size;
6828 * This macro is used for testing whenever the unwinder works correctly at every point
6829 * where an async exception can happen.
6831 /* This will generate a SIGSEGV at the given point in the code */
6832 #define async_exc_point(code) do { \
6833 if (mono_inject_async_exc_method && mono_method_desc_full_match (mono_inject_async_exc_method, cfg->method)) { \
6834 if (cfg->arch.async_point_count == mono_inject_async_exc_pos) \
6835 amd64_mov_reg_mem (code, AMD64_RAX, 0, 4); \
6836 cfg->arch.async_point_count ++; \
6841 mono_arch_emit_prolog (MonoCompile *cfg)
6843 MonoMethod *method = cfg->method;
6845 MonoMethodSignature *sig;
6847 int alloc_size, pos, i, cfa_offset, quad, max_epilog_size, save_area_offset;
6850 MonoInst *lmf_var = cfg->lmf_var;
6851 gboolean args_clobbered = FALSE;
6852 gboolean trace = FALSE;
6853 #ifdef __native_client_codegen__
6854 guint alignment_check;
6857 cfg->code_size = MAX (cfg->header->code_size * 4, 1024);
6859 #if defined(__default_codegen__)
6860 code = cfg->native_code = g_malloc (cfg->code_size);
6861 #elif defined(__native_client_codegen__)
6862 /* native_code_alloc is not 32-byte aligned, native_code is. */
6863 cfg->native_code_alloc = g_malloc (cfg->code_size + kNaClAlignment);
6865 /* Align native_code to next nearest kNaclAlignment byte. */
6866 cfg->native_code = (uintptr_t)cfg->native_code_alloc + kNaClAlignment;
6867 cfg->native_code = (uintptr_t)cfg->native_code & ~kNaClAlignmentMask;
6869 code = cfg->native_code;
6871 alignment_check = (guint)cfg->native_code & kNaClAlignmentMask;
6872 g_assert (alignment_check == 0);
6875 if (mono_jit_trace_calls != NULL && mono_trace_eval (method))
6878 /* Amount of stack space allocated by register saving code */
6881 /* Offset between RSP and the CFA */
6885 * The prolog consists of the following parts:
6887 * - push rbp, mov rbp, rsp
6888 * - save callee saved regs using pushes
6890 * - save rgctx if needed
6891 * - save lmf if needed
6894 * - save rgctx if needed
6895 * - save lmf if needed
6896 * - save callee saved regs using moves
6901 mono_emit_unwind_op_def_cfa (cfg, code, AMD64_RSP, 8);
6902 // IP saved at CFA - 8
6903 mono_emit_unwind_op_offset (cfg, code, AMD64_RIP, -cfa_offset);
6904 async_exc_point (code);
6905 mini_gc_set_slot_type_from_cfa (cfg, -cfa_offset, SLOT_NOREF);
6907 if (!cfg->arch.omit_fp) {
6908 amd64_push_reg (code, AMD64_RBP);
6910 mono_emit_unwind_op_def_cfa_offset (cfg, code, cfa_offset);
6911 mono_emit_unwind_op_offset (cfg, code, AMD64_RBP, - cfa_offset);
6912 async_exc_point (code);
6914 mono_arch_unwindinfo_add_push_nonvol (&cfg->arch.unwindinfo, cfg->native_code, code, AMD64_RBP);
6916 /* These are handled automatically by the stack marking code */
6917 mini_gc_set_slot_type_from_cfa (cfg, -cfa_offset, SLOT_NOREF);
6919 amd64_mov_reg_reg (code, AMD64_RBP, AMD64_RSP, sizeof(mgreg_t));
6920 mono_emit_unwind_op_def_cfa_reg (cfg, code, AMD64_RBP);
6921 async_exc_point (code);
6923 mono_arch_unwindinfo_add_set_fpreg (&cfg->arch.unwindinfo, cfg->native_code, code, AMD64_RBP);
6927 /* The param area is always at offset 0 from sp */
6928 /* This needs to be allocated here, since it has to come after the spill area */
6929 if (cfg->param_area) {
6930 if (cfg->arch.omit_fp)
6932 g_assert_not_reached ();
6933 cfg->stack_offset += ALIGN_TO (cfg->param_area, sizeof(mgreg_t));
6936 if (cfg->arch.omit_fp) {
6938 * On enter, the stack is misaligned by the pushing of the return
6939 * address. It is either made aligned by the pushing of %rbp, or by
6942 alloc_size = ALIGN_TO (cfg->stack_offset, 8);
6943 if ((alloc_size % 16) == 0) {
6945 /* Mark the padding slot as NOREF */
6946 mini_gc_set_slot_type_from_cfa (cfg, -cfa_offset - sizeof (mgreg_t), SLOT_NOREF);
6949 alloc_size = ALIGN_TO (cfg->stack_offset, MONO_ARCH_FRAME_ALIGNMENT);
6950 if (cfg->stack_offset != alloc_size) {
6951 /* Mark the padding slot as NOREF */
6952 mini_gc_set_slot_type_from_fp (cfg, -alloc_size + cfg->param_area, SLOT_NOREF);
6954 cfg->arch.sp_fp_offset = alloc_size;
6958 cfg->arch.stack_alloc_size = alloc_size;
6960 /* Allocate stack frame */
6962 /* See mono_emit_stack_alloc */
6963 #if defined(TARGET_WIN32) || defined(MONO_ARCH_SIGSEGV_ON_ALTSTACK)
6964 guint32 remaining_size = alloc_size;
6965 /*FIXME handle unbounded code expansion, we should use a loop in case of more than X interactions*/
6966 guint32 required_code_size = ((remaining_size / 0x1000) + 1) * 10; /*10 is the max size of amd64_alu_reg_imm + amd64_test_membase_reg*/
6967 guint32 offset = code - cfg->native_code;
6968 if (G_UNLIKELY (required_code_size >= (cfg->code_size - offset))) {
6969 while (required_code_size >= (cfg->code_size - offset))
6970 cfg->code_size *= 2;
6971 cfg->native_code = mono_realloc_native_code (cfg);
6972 code = cfg->native_code + offset;
6973 cfg->stat_code_reallocs++;
6976 while (remaining_size >= 0x1000) {
6977 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, 0x1000);
6978 if (cfg->arch.omit_fp) {
6979 cfa_offset += 0x1000;
6980 mono_emit_unwind_op_def_cfa_offset (cfg, code, cfa_offset);
6982 async_exc_point (code);
6984 if (cfg->arch.omit_fp)
6985 mono_arch_unwindinfo_add_alloc_stack (&cfg->arch.unwindinfo, cfg->native_code, code, 0x1000);
6988 amd64_test_membase_reg (code, AMD64_RSP, 0, AMD64_RSP);
6989 remaining_size -= 0x1000;
6991 if (remaining_size) {
6992 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, remaining_size);
6993 if (cfg->arch.omit_fp) {
6994 cfa_offset += remaining_size;
6995 mono_emit_unwind_op_def_cfa_offset (cfg, code, cfa_offset);
6996 async_exc_point (code);
6999 if (cfg->arch.omit_fp)
7000 mono_arch_unwindinfo_add_alloc_stack (&cfg->arch.unwindinfo, cfg->native_code, code, remaining_size);
7004 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, alloc_size);
7005 if (cfg->arch.omit_fp) {
7006 cfa_offset += alloc_size;
7007 mono_emit_unwind_op_def_cfa_offset (cfg, code, cfa_offset);
7008 async_exc_point (code);
7013 /* Stack alignment check */
7016 amd64_mov_reg_reg (code, AMD64_RAX, AMD64_RSP, 8);
7017 amd64_alu_reg_imm (code, X86_AND, AMD64_RAX, 0xf);
7018 amd64_alu_reg_imm (code, X86_CMP, AMD64_RAX, 0);
7019 x86_branch8 (code, X86_CC_EQ, 2, FALSE);
7020 amd64_breakpoint (code);
7024 if (mini_get_debug_options ()->init_stacks) {
7025 /* Fill the stack frame with a dummy value to force deterministic behavior */
7027 /* Save registers to the red zone */
7028 amd64_mov_membase_reg (code, AMD64_RSP, -8, AMD64_RDI, 8);
7029 amd64_mov_membase_reg (code, AMD64_RSP, -16, AMD64_RCX, 8);
7031 amd64_mov_reg_imm (code, AMD64_RAX, 0x2a2a2a2a2a2a2a2a);
7032 amd64_mov_reg_imm (code, AMD64_RCX, alloc_size / 8);
7033 amd64_mov_reg_reg (code, AMD64_RDI, AMD64_RSP, 8);
7036 #if defined(__default_codegen__)
7037 amd64_prefix (code, X86_REP_PREFIX);
7039 #elif defined(__native_client_codegen__)
7040 /* NaCl stos pseudo-instruction */
7041 amd64_codegen_pre (code);
7042 /* First, clear the upper 32 bits of RDI (mov %edi, %edi) */
7043 amd64_mov_reg_reg (code, AMD64_RDI, AMD64_RDI, 4);
7044 /* Add %r15 to %rdi using lea, condition flags unaffected. */
7045 amd64_lea_memindex_size (code, AMD64_RDI, AMD64_R15, 0, AMD64_RDI, 0, 8);
7046 amd64_prefix (code, X86_REP_PREFIX);
7048 amd64_codegen_post (code);
7049 #endif /* __native_client_codegen__ */
7051 amd64_mov_reg_membase (code, AMD64_RDI, AMD64_RSP, -8, 8);
7052 amd64_mov_reg_membase (code, AMD64_RCX, AMD64_RSP, -16, 8);
7056 if (method->save_lmf)
7057 code = emit_setup_lmf (cfg, code, lmf_var->inst_offset, cfa_offset);
7059 /* Save callee saved registers */
7060 if (cfg->arch.omit_fp) {
7061 save_area_offset = cfg->arch.reg_save_area_offset;
7062 /* Save caller saved registers after sp is adjusted */
7063 /* The registers are saved at the bottom of the frame */
7064 /* FIXME: Optimize this so the regs are saved at the end of the frame in increasing order */
7066 /* The registers are saved just below the saved rbp */
7067 save_area_offset = cfg->arch.reg_save_area_offset;
7070 for (i = 0; i < AMD64_NREG; ++i) {
7071 if (AMD64_IS_CALLEE_SAVED_REG (i) && (cfg->arch.saved_iregs & (1 << i))) {
7072 amd64_mov_membase_reg (code, cfg->frame_reg, save_area_offset, i, 8);
7074 if (cfg->arch.omit_fp) {
7075 mono_emit_unwind_op_offset (cfg, code, i, - (cfa_offset - save_area_offset));
7076 /* These are handled automatically by the stack marking code */
7077 mini_gc_set_slot_type_from_cfa (cfg, - (cfa_offset - save_area_offset), SLOT_NOREF);
7079 mono_emit_unwind_op_offset (cfg, code, i, - (-save_area_offset + (2 * 8)));
7083 save_area_offset += 8;
7084 async_exc_point (code);
7088 /* store runtime generic context */
7089 if (cfg->rgctx_var) {
7090 g_assert (cfg->rgctx_var->opcode == OP_REGOFFSET &&
7091 (cfg->rgctx_var->inst_basereg == AMD64_RBP || cfg->rgctx_var->inst_basereg == AMD64_RSP));
7093 amd64_mov_membase_reg (code, cfg->rgctx_var->inst_basereg, cfg->rgctx_var->inst_offset, MONO_ARCH_RGCTX_REG, sizeof(gpointer));
7095 mono_add_var_location (cfg, cfg->rgctx_var, TRUE, MONO_ARCH_RGCTX_REG, 0, 0, code - cfg->native_code);
7096 mono_add_var_location (cfg, cfg->rgctx_var, FALSE, cfg->rgctx_var->inst_basereg, cfg->rgctx_var->inst_offset, code - cfg->native_code, 0);
7099 /* compute max_length in order to use short forward jumps */
7100 max_epilog_size = get_max_epilog_size (cfg);
7101 if (cfg->opt & MONO_OPT_BRANCH) {
7102 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
7106 if (cfg->prof_options & MONO_PROFILE_COVERAGE)
7108 /* max alignment for loops */
7109 if ((cfg->opt & MONO_OPT_LOOP) && bb_is_loop_start (bb))
7110 max_length += LOOP_ALIGNMENT;
7111 #ifdef __native_client_codegen__
7112 /* max alignment for native client */
7113 max_length += kNaClAlignment;
7116 MONO_BB_FOR_EACH_INS (bb, ins) {
7117 #ifdef __native_client_codegen__
7119 int space_in_block = kNaClAlignment -
7120 ((max_length + cfg->code_len) & kNaClAlignmentMask);
7121 int max_len = ((guint8 *)ins_get_spec (ins->opcode))[MONO_INST_LEN];
7122 if (space_in_block < max_len && max_len < kNaClAlignment) {
7123 max_length += space_in_block;
7126 #endif /*__native_client_codegen__*/
7127 max_length += ((guint8 *)ins_get_spec (ins->opcode))[MONO_INST_LEN];
7130 /* Take prolog and epilog instrumentation into account */
7131 if (bb == cfg->bb_entry || bb == cfg->bb_exit)
7132 max_length += max_epilog_size;
7134 bb->max_length = max_length;
7138 sig = mono_method_signature (method);
7141 cinfo = cfg->arch.cinfo;
7143 if (sig->ret->type != MONO_TYPE_VOID) {
7144 /* Save volatile arguments to the stack */
7145 if (cfg->vret_addr && (cfg->vret_addr->opcode != OP_REGVAR))
7146 amd64_mov_membase_reg (code, cfg->vret_addr->inst_basereg, cfg->vret_addr->inst_offset, cinfo->ret.reg, 8);
7149 /* Keep this in sync with emit_load_volatile_arguments */
7150 for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
7151 ArgInfo *ainfo = cinfo->args + i;
7153 ins = cfg->args [i];
7155 if ((ins->flags & MONO_INST_IS_DEAD) && !trace)
7156 /* Unused arguments */
7159 /* Save volatile arguments to the stack */
7160 if (ins->opcode != OP_REGVAR) {
7161 switch (ainfo->storage) {
7167 if (stack_offset & 0x1)
7169 else if (stack_offset & 0x2)
7171 else if (stack_offset & 0x4)
7176 amd64_mov_membase_reg (code, ins->inst_basereg, ins->inst_offset, ainfo->reg, size);
7179 * Save the original location of 'this',
7180 * get_generic_info_from_stack_frame () needs this to properly look up
7181 * the argument value during the handling of async exceptions.
7183 if (ins == cfg->args [0]) {
7184 mono_add_var_location (cfg, ins, TRUE, ainfo->reg, 0, 0, code - cfg->native_code);
7185 mono_add_var_location (cfg, ins, FALSE, ins->inst_basereg, ins->inst_offset, code - cfg->native_code, 0);
7189 case ArgInFloatSSEReg:
7190 amd64_movss_membase_reg (code, ins->inst_basereg, ins->inst_offset, ainfo->reg);
7192 case ArgInDoubleSSEReg:
7193 amd64_movsd_membase_reg (code, ins->inst_basereg, ins->inst_offset, ainfo->reg);
7195 case ArgValuetypeInReg:
7196 for (quad = 0; quad < 2; quad ++) {
7197 switch (ainfo->pair_storage [quad]) {
7199 amd64_mov_membase_reg (code, ins->inst_basereg, ins->inst_offset + (quad * sizeof(mgreg_t)), ainfo->pair_regs [quad], sizeof(mgreg_t));
7201 case ArgInFloatSSEReg:
7202 amd64_movss_membase_reg (code, ins->inst_basereg, ins->inst_offset + (quad * sizeof(mgreg_t)), ainfo->pair_regs [quad]);
7204 case ArgInDoubleSSEReg:
7205 amd64_movsd_membase_reg (code, ins->inst_basereg, ins->inst_offset + (quad * sizeof(mgreg_t)), ainfo->pair_regs [quad]);
7210 g_assert_not_reached ();
7214 case ArgValuetypeAddrInIReg:
7215 if (ainfo->pair_storage [0] == ArgInIReg)
7216 amd64_mov_membase_reg (code, ins->inst_left->inst_basereg, ins->inst_left->inst_offset, ainfo->pair_regs [0], sizeof (gpointer));
7222 /* Argument allocated to (non-volatile) register */
7223 switch (ainfo->storage) {
7225 amd64_mov_reg_reg (code, ins->dreg, ainfo->reg, 8);
7228 amd64_mov_reg_membase (code, ins->dreg, AMD64_RBP, ARGS_OFFSET + ainfo->offset, 8);
7231 g_assert_not_reached ();
7234 if (ins == cfg->args [0]) {
7235 mono_add_var_location (cfg, ins, TRUE, ainfo->reg, 0, 0, code - cfg->native_code);
7236 mono_add_var_location (cfg, ins, TRUE, ins->dreg, 0, code - cfg->native_code, 0);
7241 if (cfg->method->save_lmf)
7242 args_clobbered = TRUE;
7245 args_clobbered = TRUE;
7246 code = mono_arch_instrument_prolog (cfg, mono_trace_enter_method, code, TRUE);
7249 if (cfg->prof_options & MONO_PROFILE_ENTER_LEAVE)
7250 args_clobbered = TRUE;
7253 * Optimize the common case of the first bblock making a call with the same
7254 * arguments as the method. This works because the arguments are still in their
7255 * original argument registers.
7256 * FIXME: Generalize this
7258 if (!args_clobbered) {
7259 MonoBasicBlock *first_bb = cfg->bb_entry;
7261 int filter = FILTER_IL_SEQ_POINT;
7263 next = mono_bb_first_inst (first_bb, filter);
7264 if (!next && first_bb->next_bb) {
7265 first_bb = first_bb->next_bb;
7266 next = mono_bb_first_inst (first_bb, filter);
7269 if (first_bb->in_count > 1)
7272 for (i = 0; next && i < sig->param_count + sig->hasthis; ++i) {
7273 ArgInfo *ainfo = cinfo->args + i;
7274 gboolean match = FALSE;
7276 ins = cfg->args [i];
7277 if (ins->opcode != OP_REGVAR) {
7278 switch (ainfo->storage) {
7280 if (((next->opcode == OP_LOAD_MEMBASE) || (next->opcode == OP_LOADI4_MEMBASE)) && next->inst_basereg == ins->inst_basereg && next->inst_offset == ins->inst_offset) {
7281 if (next->dreg == ainfo->reg) {
7285 next->opcode = OP_MOVE;
7286 next->sreg1 = ainfo->reg;
7287 /* Only continue if the instruction doesn't change argument regs */
7288 if (next->dreg == ainfo->reg || next->dreg == AMD64_RAX)
7298 /* Argument allocated to (non-volatile) register */
7299 switch (ainfo->storage) {
7301 if (next->opcode == OP_MOVE && next->sreg1 == ins->dreg && next->dreg == ainfo->reg) {
7312 next = mono_inst_next (next, filter);
7313 //next = mono_inst_list_next (&next->node, &first_bb->ins_list);
7320 if (cfg->gen_sdb_seq_points) {
7321 MonoInst *info_var = cfg->arch.seq_point_info_var;
7323 /* Initialize seq_point_info_var */
7324 if (cfg->compile_aot) {
7325 /* Initialize the variable from a GOT slot */
7326 /* Same as OP_AOTCONST */
7327 mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_SEQ_POINT_INFO, cfg->method);
7328 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RIP, 0, sizeof(gpointer));
7329 g_assert (info_var->opcode == OP_REGOFFSET);
7330 amd64_mov_membase_reg (code, info_var->inst_basereg, info_var->inst_offset, AMD64_R11, 8);
7333 if (cfg->compile_aot) {
7334 /* Initialize ss_tramp_var */
7335 ins = cfg->arch.ss_tramp_var;
7336 g_assert (ins->opcode == OP_REGOFFSET);
7338 amd64_mov_reg_membase (code, AMD64_R11, info_var->inst_basereg, info_var->inst_offset, 8);
7339 amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, MONO_STRUCT_OFFSET (SeqPointInfo, ss_tramp_addr), 8);
7340 amd64_mov_membase_reg (code, ins->inst_basereg, ins->inst_offset, AMD64_R11, 8);
7342 /* Initialize ss_tramp_var */
7343 ins = cfg->arch.ss_tramp_var;
7344 g_assert (ins->opcode == OP_REGOFFSET);
7346 amd64_mov_reg_imm (code, AMD64_R11, (guint64)&ss_trampoline);
7347 amd64_mov_membase_reg (code, ins->inst_basereg, ins->inst_offset, AMD64_R11, 8);
7349 /* Initialize bp_tramp_var */
7350 ins = cfg->arch.bp_tramp_var;
7351 g_assert (ins->opcode == OP_REGOFFSET);
7353 amd64_mov_reg_imm (code, AMD64_R11, (guint64)&bp_trampoline);
7354 amd64_mov_membase_reg (code, ins->inst_basereg, ins->inst_offset, AMD64_R11, 8);
7358 cfg->code_len = code - cfg->native_code;
7360 g_assert (cfg->code_len < cfg->code_size);
7366 mono_arch_emit_epilog (MonoCompile *cfg)
7368 MonoMethod *method = cfg->method;
7371 int max_epilog_size;
7373 gint32 lmf_offset = cfg->lmf_var ? ((MonoInst*)cfg->lmf_var)->inst_offset : -1;
7374 gint32 save_area_offset = cfg->arch.reg_save_area_offset;
7376 max_epilog_size = get_max_epilog_size (cfg);
7378 while (cfg->code_len + max_epilog_size > (cfg->code_size - 16)) {
7379 cfg->code_size *= 2;
7380 cfg->native_code = mono_realloc_native_code (cfg);
7381 cfg->stat_code_reallocs++;
7383 code = cfg->native_code + cfg->code_len;
7385 cfg->has_unwind_info_for_epilog = TRUE;
7387 /* Mark the start of the epilog */
7388 mono_emit_unwind_op_mark_loc (cfg, code, 0);
7390 /* Save the uwind state which is needed by the out-of-line code */
7391 mono_emit_unwind_op_remember_state (cfg, code);
7393 if (mono_jit_trace_calls != NULL && mono_trace_eval (method))
7394 code = mono_arch_instrument_epilog (cfg, mono_trace_leave_method, code, TRUE);
7396 /* the code restoring the registers must be kept in sync with OP_TAILCALL */
7398 if (method->save_lmf) {
7399 /* check if we need to restore protection of the stack after a stack overflow */
7400 if (!cfg->compile_aot && mono_get_jit_tls_offset () != -1) {
7402 code = mono_amd64_emit_tls_get (code, AMD64_RCX, mono_get_jit_tls_offset ());
7403 /* we load the value in a separate instruction: this mechanism may be
7404 * used later as a safer way to do thread interruption
7406 amd64_mov_reg_membase (code, AMD64_RCX, AMD64_RCX, MONO_STRUCT_OFFSET (MonoJitTlsData, restore_stack_prot), 8);
7407 x86_alu_reg_imm (code, X86_CMP, X86_ECX, 0);
7409 x86_branch8 (code, X86_CC_Z, 0, FALSE);
7410 /* note that the call trampoline will preserve eax/edx */
7411 x86_call_reg (code, X86_ECX);
7412 x86_patch (patch, code);
7414 /* FIXME: maybe save the jit tls in the prolog */
7416 if (cfg->used_int_regs & (1 << AMD64_RBP)) {
7417 amd64_mov_reg_membase (code, AMD64_RBP, cfg->frame_reg, lmf_offset + MONO_STRUCT_OFFSET (MonoLMF, rbp), 8);
7421 /* Restore callee saved regs */
7422 for (i = 0; i < AMD64_NREG; ++i) {
7423 if (AMD64_IS_CALLEE_SAVED_REG (i) && (cfg->arch.saved_iregs & (1 << i))) {
7424 /* Restore only used_int_regs, not arch.saved_iregs */
7425 if (cfg->used_int_regs & (1 << i)) {
7426 amd64_mov_reg_membase (code, i, cfg->frame_reg, save_area_offset, 8);
7427 mono_emit_unwind_op_same_value (cfg, code, i);
7428 async_exc_point (code);
7430 save_area_offset += 8;
7434 /* Load returned vtypes into registers if needed */
7435 cinfo = cfg->arch.cinfo;
7436 if (cinfo->ret.storage == ArgValuetypeInReg) {
7437 ArgInfo *ainfo = &cinfo->ret;
7438 MonoInst *inst = cfg->ret;
7440 for (quad = 0; quad < 2; quad ++) {
7441 switch (ainfo->pair_storage [quad]) {
7443 amd64_mov_reg_membase (code, ainfo->pair_regs [quad], inst->inst_basereg, inst->inst_offset + (quad * sizeof(mgreg_t)), ainfo->pair_size [quad]);
7445 case ArgInFloatSSEReg:
7446 amd64_movss_reg_membase (code, ainfo->pair_regs [quad], inst->inst_basereg, inst->inst_offset + (quad * sizeof(mgreg_t)));
7448 case ArgInDoubleSSEReg:
7449 amd64_movsd_reg_membase (code, ainfo->pair_regs [quad], inst->inst_basereg, inst->inst_offset + (quad * sizeof(mgreg_t)));
7454 g_assert_not_reached ();
7459 if (cfg->arch.omit_fp) {
7460 if (cfg->arch.stack_alloc_size) {
7461 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, cfg->arch.stack_alloc_size);
7465 mono_emit_unwind_op_same_value (cfg, code, AMD64_RBP);
7467 mono_emit_unwind_op_def_cfa (cfg, code, AMD64_RSP, 8);
7468 async_exc_point (code);
7471 /* Restore the unwind state to be the same as before the epilog */
7472 mono_emit_unwind_op_restore_state (cfg, code);
7474 cfg->code_len = code - cfg->native_code;
7476 g_assert (cfg->code_len < cfg->code_size);
7480 mono_arch_emit_exceptions (MonoCompile *cfg)
7482 MonoJumpInfo *patch_info;
7485 MonoClass *exc_classes [16];
7486 guint8 *exc_throw_start [16], *exc_throw_end [16];
7487 guint32 code_size = 0;
7489 /* Compute needed space */
7490 for (patch_info = cfg->patch_info; patch_info; patch_info = patch_info->next) {
7491 if (patch_info->type == MONO_PATCH_INFO_EXC)
7493 if (patch_info->type == MONO_PATCH_INFO_R8)
7494 code_size += 8 + 15; /* sizeof (double) + alignment */
7495 if (patch_info->type == MONO_PATCH_INFO_R4)
7496 code_size += 4 + 15; /* sizeof (float) + alignment */
7497 if (patch_info->type == MONO_PATCH_INFO_GC_CARD_TABLE_ADDR)
7498 code_size += 8 + 7; /*sizeof (void*) + alignment */
7501 #ifdef __native_client_codegen__
7502 /* Give us extra room on Native Client. This could be */
7503 /* more carefully calculated, but bundle alignment makes */
7504 /* it much trickier, so *2 like other places is good. */
7508 while (cfg->code_len + code_size > (cfg->code_size - 16)) {
7509 cfg->code_size *= 2;
7510 cfg->native_code = mono_realloc_native_code (cfg);
7511 cfg->stat_code_reallocs++;
7514 code = cfg->native_code + cfg->code_len;
7516 /* add code to raise exceptions */
7518 for (patch_info = cfg->patch_info; patch_info; patch_info = patch_info->next) {
7519 switch (patch_info->type) {
7520 case MONO_PATCH_INFO_EXC: {
7521 MonoClass *exc_class;
7525 amd64_patch (patch_info->ip.i + cfg->native_code, code);
7527 exc_class = mono_class_from_name (mono_defaults.corlib, "System", patch_info->data.name);
7528 g_assert (exc_class);
7529 throw_ip = patch_info->ip.i;
7531 //x86_breakpoint (code);
7532 /* Find a throw sequence for the same exception class */
7533 for (i = 0; i < nthrows; ++i)
7534 if (exc_classes [i] == exc_class)
7537 amd64_mov_reg_imm (code, AMD64_ARG_REG2, (exc_throw_end [i] - cfg->native_code) - throw_ip);
7538 x86_jump_code (code, exc_throw_start [i]);
7539 patch_info->type = MONO_PATCH_INFO_NONE;
7543 amd64_mov_reg_imm_size (code, AMD64_ARG_REG2, 0xf0f0f0f0, 4);
7547 exc_classes [nthrows] = exc_class;
7548 exc_throw_start [nthrows] = code;
7550 amd64_mov_reg_imm (code, AMD64_ARG_REG1, exc_class->type_token - MONO_TOKEN_TYPE_DEF);
7552 patch_info->type = MONO_PATCH_INFO_NONE;
7554 code = emit_call_body (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD, "mono_arch_throw_corlib_exception");
7556 amd64_mov_reg_imm (buf, AMD64_ARG_REG2, (code - cfg->native_code) - throw_ip);
7561 exc_throw_end [nthrows] = code;
7571 g_assert(code < cfg->native_code + cfg->code_size);
7574 /* Handle relocations with RIP relative addressing */
7575 for (patch_info = cfg->patch_info; patch_info; patch_info = patch_info->next) {
7576 gboolean remove = FALSE;
7577 guint8 *orig_code = code;
7579 switch (patch_info->type) {
7580 case MONO_PATCH_INFO_R8:
7581 case MONO_PATCH_INFO_R4: {
7582 guint8 *pos, *patch_pos;
7585 /* The SSE opcodes require a 16 byte alignment */
7586 #if defined(__default_codegen__)
7587 code = (guint8*)ALIGN_TO (code, 16);
7588 #elif defined(__native_client_codegen__)
7590 /* Pad this out with HLT instructions */
7591 /* or we can get garbage bytes emitted */
7592 /* which will fail validation */
7593 guint8 *aligned_code;
7594 /* extra align to make room for */
7595 /* mov/push below */
7596 int extra_align = patch_info->type == MONO_PATCH_INFO_R8 ? 2 : 1;
7597 aligned_code = (guint8*)ALIGN_TO (code + extra_align, 16);
7598 /* The technique of hiding data in an */
7599 /* instruction has a problem here: we */
7600 /* need the data aligned to a 16-byte */
7601 /* boundary but the instruction cannot */
7602 /* cross the bundle boundary. so only */
7603 /* odd multiples of 16 can be used */
7604 if ((intptr_t)aligned_code % kNaClAlignment == 0) {
7607 while (code < aligned_code) {
7608 *(code++) = 0xf4; /* hlt */
7613 pos = cfg->native_code + patch_info->ip.i;
7614 if (IS_REX (pos [1])) {
7615 patch_pos = pos + 5;
7616 target_pos = code - pos - 9;
7619 patch_pos = pos + 4;
7620 target_pos = code - pos - 8;
7623 if (patch_info->type == MONO_PATCH_INFO_R8) {
7624 #ifdef __native_client_codegen__
7625 /* Hide 64-bit data in a */
7626 /* "mov imm64, r11" instruction. */
7627 /* write it before the start of */
7629 *(code-2) = 0x49; /* prefix */
7630 *(code-1) = 0xbb; /* mov X, %r11 */
7632 *(double*)code = *(double*)patch_info->data.target;
7633 code += sizeof (double);
7635 #ifdef __native_client_codegen__
7636 /* Hide 32-bit data in a */
7637 /* "push imm32" instruction. */
7638 *(code-1) = 0x68; /* push */
7640 *(float*)code = *(float*)patch_info->data.target;
7641 code += sizeof (float);
7644 *(guint32*)(patch_pos) = target_pos;
7649 case MONO_PATCH_INFO_GC_CARD_TABLE_ADDR: {
7652 if (cfg->compile_aot)
7655 /*loading is faster against aligned addresses.*/
7656 code = (guint8*)ALIGN_TO (code, 8);
7657 memset (orig_code, 0, code - orig_code);
7659 pos = cfg->native_code + patch_info->ip.i;
7661 /*alu_op [rex] modr/m imm32 - 7 or 8 bytes */
7662 if (IS_REX (pos [1]))
7663 *(guint32*)(pos + 4) = (guint8*)code - pos - 8;
7665 *(guint32*)(pos + 3) = (guint8*)code - pos - 7;
7667 *(gpointer*)code = (gpointer)patch_info->data.target;
7668 code += sizeof (gpointer);
7678 if (patch_info == cfg->patch_info)
7679 cfg->patch_info = patch_info->next;
7683 for (tmp = cfg->patch_info; tmp->next != patch_info; tmp = tmp->next)
7685 tmp->next = patch_info->next;
7688 g_assert (code < cfg->native_code + cfg->code_size);
7691 cfg->code_len = code - cfg->native_code;
7693 g_assert (cfg->code_len < cfg->code_size);
7697 #endif /* DISABLE_JIT */
7700 mono_arch_instrument_prolog (MonoCompile *cfg, void *func, void *p, gboolean enable_arguments)
7703 MonoMethodSignature *sig;
7705 int i, n, stack_area = 0;
7707 /* Keep this in sync with mono_arch_get_argument_info */
7709 if (enable_arguments) {
7710 /* Allocate a new area on the stack and save arguments there */
7711 sig = mono_method_signature (cfg->method);
7713 n = sig->param_count + sig->hasthis;
7715 stack_area = ALIGN_TO (n * 8, 16);
7717 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, stack_area);
7719 for (i = 0; i < n; ++i) {
7720 inst = cfg->args [i];
7722 if (inst->opcode == OP_REGVAR)
7723 amd64_mov_membase_reg (code, AMD64_RSP, (i * 8), inst->dreg, 8);
7725 amd64_mov_reg_membase (code, AMD64_R11, inst->inst_basereg, inst->inst_offset, 8);
7726 amd64_mov_membase_reg (code, AMD64_RSP, (i * 8), AMD64_R11, 8);
7731 mono_add_patch_info (cfg, code-cfg->native_code, MONO_PATCH_INFO_METHODCONST, cfg->method);
7732 amd64_set_reg_template (code, AMD64_ARG_REG1);
7733 amd64_mov_reg_reg (code, AMD64_ARG_REG2, AMD64_RSP, 8);
7734 code = emit_call (cfg, code, MONO_PATCH_INFO_ABS, (gpointer)func, TRUE);
7736 if (enable_arguments)
7737 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, stack_area);
7751 mono_arch_instrument_epilog_full (MonoCompile *cfg, void *func, void *p, gboolean enable_arguments, gboolean preserve_argument_registers)
7754 int save_mode = SAVE_NONE;
7755 MonoMethod *method = cfg->method;
7756 MonoType *ret_type = mini_get_underlying_type (mono_method_signature (method)->ret);
7759 switch (ret_type->type) {
7760 case MONO_TYPE_VOID:
7761 /* special case string .ctor icall */
7762 if (strcmp (".ctor", method->name) && method->klass == mono_defaults.string_class)
7763 save_mode = SAVE_EAX;
7765 save_mode = SAVE_NONE;
7769 save_mode = SAVE_EAX;
7773 save_mode = SAVE_XMM;
7775 case MONO_TYPE_GENERICINST:
7776 if (!mono_type_generic_inst_is_valuetype (ret_type)) {
7777 save_mode = SAVE_EAX;
7781 case MONO_TYPE_VALUETYPE:
7782 save_mode = SAVE_STRUCT;
7785 save_mode = SAVE_EAX;
7789 /* Save the result and copy it into the proper argument register */
7790 switch (save_mode) {
7792 amd64_push_reg (code, AMD64_RAX);
7794 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, 8);
7795 if (enable_arguments)
7796 amd64_mov_reg_reg (code, AMD64_ARG_REG2, AMD64_RAX, 8);
7800 if (enable_arguments)
7801 amd64_mov_reg_imm (code, AMD64_ARG_REG2, 0);
7804 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, 8);
7805 amd64_movsd_membase_reg (code, AMD64_RSP, 0, AMD64_XMM0);
7807 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, 8);
7809 * The result is already in the proper argument register so no copying
7816 g_assert_not_reached ();
7819 /* Set %al since this is a varargs call */
7820 if (save_mode == SAVE_XMM)
7821 amd64_mov_reg_imm (code, AMD64_RAX, 1);
7823 amd64_mov_reg_imm (code, AMD64_RAX, 0);
7825 if (preserve_argument_registers) {
7826 for (i = 0; i < PARAM_REGS; ++i)
7827 amd64_push_reg (code, param_regs [i]);
7830 mono_add_patch_info (cfg, code-cfg->native_code, MONO_PATCH_INFO_METHODCONST, method);
7831 amd64_set_reg_template (code, AMD64_ARG_REG1);
7832 code = emit_call (cfg, code, MONO_PATCH_INFO_ABS, (gpointer)func, TRUE);
7834 if (preserve_argument_registers) {
7835 for (i = PARAM_REGS - 1; i >= 0; --i)
7836 amd64_pop_reg (code, param_regs [i]);
7839 /* Restore result */
7840 switch (save_mode) {
7842 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, 8);
7843 amd64_pop_reg (code, AMD64_RAX);
7849 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, 8);
7850 amd64_movsd_reg_membase (code, AMD64_XMM0, AMD64_RSP, 0);
7851 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, 8);
7856 g_assert_not_reached ();
7863 mono_arch_flush_icache (guint8 *code, gint size)
7869 mono_arch_flush_register_windows (void)
7874 mono_arch_is_inst_imm (gint64 imm)
7876 return amd64_use_imm32 (imm);
7880 * Determine whenever the trap whose info is in SIGINFO is caused by
7884 mono_arch_is_int_overflow (void *sigctx, void *info)
7891 mono_sigctx_to_monoctx (sigctx, &ctx);
7893 rip = (guint8*)ctx.gregs [AMD64_RIP];
7895 if (IS_REX (rip [0])) {
7896 reg = amd64_rex_b (rip [0]);
7902 if ((rip [0] == 0xf7) && (x86_modrm_mod (rip [1]) == 0x3) && (x86_modrm_reg (rip [1]) == 0x7)) {
7904 reg += x86_modrm_rm (rip [1]);
7906 value = ctx.gregs [reg];
7916 mono_arch_get_patch_offset (guint8 *code)
7922 * mono_breakpoint_clean_code:
7924 * Copy @size bytes from @code - @offset to the buffer @buf. If the debugger inserted software
7925 * breakpoints in the original code, they are removed in the copy.
7927 * Returns TRUE if no sw breakpoint was present.
7930 mono_breakpoint_clean_code (guint8 *method_start, guint8 *code, int offset, guint8 *buf, int size)
7933 * If method_start is non-NULL we need to perform bound checks, since we access memory
7934 * at code - offset we could go before the start of the method and end up in a different
7935 * page of memory that is not mapped or read incorrect data anyway. We zero-fill the bytes
7938 if (!method_start || code - offset >= method_start) {
7939 memcpy (buf, code - offset, size);
7941 int diff = code - method_start;
7942 memset (buf, 0, size);
7943 memcpy (buf + offset - diff, method_start, diff + size - offset);
7948 #if defined(__native_client_codegen__)
7949 /* For membase calls, we want the base register. for Native Client, */
7950 /* all indirect calls have the following sequence with the given sizes: */
7951 /* mov %eXX,%eXX [2-3] */
7952 /* mov disp(%r15,%rXX,scale),%r11d [4-8] */
7953 /* and $0xffffffffffffffe0,%r11d [4] */
7954 /* add %r15,%r11 [3] */
7955 /* callq *%r11 [3] */
7958 /* Determine if code points to a NaCl call-through-register sequence, */
7959 /* (i.e., the last 3 instructions listed above) */
7961 is_nacl_call_reg_sequence(guint8* code)
7963 const char *sequence = "\x41\x83\xe3\xe0" /* and */
7964 "\x4d\x03\xdf" /* add */
7965 "\x41\xff\xd3"; /* call */
7966 return memcmp(code, sequence, 10) == 0;
7969 /* Determine if code points to the first opcode of the mov membase component */
7970 /* of an indirect call sequence (i.e. the first 2 instructions listed above) */
7971 /* (there could be a REX prefix before the opcode but it is ignored) */
7973 is_nacl_indirect_call_membase_sequence(guint8* code)
7975 /* Check for mov opcode, reg-reg addressing mode (mod = 3), */
7976 return code[0] == 0x8b && amd64_modrm_mod(code[1]) == 3 &&
7977 /* and that src reg = dest reg */
7978 amd64_modrm_reg(code[1]) == amd64_modrm_rm(code[1]) &&
7979 /* Check that next inst is mov, uses SIB byte (rm = 4), */
7981 code[3] == 0x8b && amd64_modrm_rm(code[4]) == 4 &&
7982 /* and has dst of r11 and base of r15 */
7983 (amd64_modrm_reg(code[4]) + amd64_rex_r(code[2])) == AMD64_R11 &&
7984 (amd64_sib_base(code[5]) + amd64_rex_b(code[2])) == AMD64_R15;
7986 #endif /* __native_client_codegen__ */
7989 mono_arch_get_this_arg_reg (guint8 *code)
7991 return AMD64_ARG_REG1;
7995 mono_arch_get_this_arg_from_call (mgreg_t *regs, guint8 *code)
7997 return (gpointer)regs [mono_arch_get_this_arg_reg (code)];
8000 #define MAX_ARCH_DELEGATE_PARAMS 10
8003 get_delegate_invoke_impl (MonoTrampInfo **info, gboolean has_target, guint32 param_count)
8005 guint8 *code, *start;
8006 GSList *unwind_ops = NULL;
8009 unwind_ops = mono_arch_get_cie_program ();
8012 start = code = mono_global_codeman_reserve (64);
8014 /* Replace the this argument with the target */
8015 amd64_mov_reg_reg (code, AMD64_RAX, AMD64_ARG_REG1, 8);
8016 amd64_mov_reg_membase (code, AMD64_ARG_REG1, AMD64_RAX, MONO_STRUCT_OFFSET (MonoDelegate, target), 8);
8017 amd64_jump_membase (code, AMD64_RAX, MONO_STRUCT_OFFSET (MonoDelegate, method_ptr));
8019 g_assert ((code - start) < 64);
8021 start = code = mono_global_codeman_reserve (64);
8023 if (param_count == 0) {
8024 amd64_jump_membase (code, AMD64_ARG_REG1, MONO_STRUCT_OFFSET (MonoDelegate, method_ptr));
8026 /* We have to shift the arguments left */
8027 amd64_mov_reg_reg (code, AMD64_RAX, AMD64_ARG_REG1, 8);
8028 for (i = 0; i < param_count; ++i) {
8031 amd64_mov_reg_reg (code, param_regs [i], param_regs [i + 1], 8);
8033 amd64_mov_reg_membase (code, param_regs [i], AMD64_RSP, 0x28, 8);
8035 amd64_mov_reg_reg (code, param_regs [i], param_regs [i + 1], 8);
8039 amd64_jump_membase (code, AMD64_RAX, MONO_STRUCT_OFFSET (MonoDelegate, method_ptr));
8041 g_assert ((code - start) < 64);
8044 nacl_global_codeman_validate (&start, 64, &code);
8045 mono_arch_flush_icache (start, code - start);
8048 *info = mono_tramp_info_create ("delegate_invoke_impl_has_target", start, code - start, NULL, unwind_ops);
8050 char *name = g_strdup_printf ("delegate_invoke_impl_target_%d", param_count);
8051 *info = mono_tramp_info_create (name, start, code - start, NULL, unwind_ops);
8055 if (mono_jit_map_is_enabled ()) {
8058 buff = (char*)"delegate_invoke_has_target";
8060 buff = g_strdup_printf ("delegate_invoke_no_target_%d", param_count);
8061 mono_emit_jit_tramp (start, code - start, buff);
8065 mono_profiler_code_buffer_new (start, code - start, MONO_PROFILER_CODE_BUFFER_DELEGATE_INVOKE, NULL);
8070 #define MAX_VIRTUAL_DELEGATE_OFFSET 32
8073 get_delegate_virtual_invoke_impl (MonoTrampInfo **info, gboolean load_imt_reg, int offset)
8075 guint8 *code, *start;
8080 if (offset / (int)sizeof (gpointer) > MAX_VIRTUAL_DELEGATE_OFFSET)
8083 start = code = mono_global_codeman_reserve (size);
8085 unwind_ops = mono_arch_get_cie_program ();
8087 /* Replace the this argument with the target */
8088 amd64_mov_reg_reg (code, AMD64_RAX, AMD64_ARG_REG1, 8);
8089 amd64_mov_reg_membase (code, AMD64_ARG_REG1, AMD64_RAX, MONO_STRUCT_OFFSET (MonoDelegate, target), 8);
8092 /* Load the IMT reg */
8093 amd64_mov_reg_membase (code, MONO_ARCH_IMT_REG, AMD64_RAX, MONO_STRUCT_OFFSET (MonoDelegate, method), 8);
8096 /* Load the vtable */
8097 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_ARG_REG1, MONO_STRUCT_OFFSET (MonoObject, vtable), 8);
8098 amd64_jump_membase (code, AMD64_RAX, offset);
8099 mono_profiler_code_buffer_new (start, code - start, MONO_PROFILER_CODE_BUFFER_DELEGATE_INVOKE, NULL);
8102 tramp_name = g_strdup_printf ("delegate_virtual_invoke_imt_%d", - offset / sizeof (gpointer));
8104 tramp_name = g_strdup_printf ("delegate_virtual_invoke_%d", offset / sizeof (gpointer));
8105 *info = mono_tramp_info_create (tramp_name, start, code - start, NULL, unwind_ops);
8106 g_free (tramp_name);
8112 * mono_arch_get_delegate_invoke_impls:
8114 * Return a list of MonoTrampInfo structures for the delegate invoke impl
8118 mono_arch_get_delegate_invoke_impls (void)
8121 MonoTrampInfo *info;
8124 get_delegate_invoke_impl (&info, TRUE, 0);
8125 res = g_slist_prepend (res, info);
8127 for (i = 0; i <= MAX_ARCH_DELEGATE_PARAMS; ++i) {
8128 get_delegate_invoke_impl (&info, FALSE, i);
8129 res = g_slist_prepend (res, info);
8132 for (i = 0; i <= MAX_VIRTUAL_DELEGATE_OFFSET; ++i) {
8133 get_delegate_virtual_invoke_impl (&info, TRUE, - i * SIZEOF_VOID_P);
8134 res = g_slist_prepend (res, info);
8136 get_delegate_virtual_invoke_impl (&info, FALSE, i * SIZEOF_VOID_P);
8137 res = g_slist_prepend (res, info);
8144 mono_arch_get_delegate_invoke_impl (MonoMethodSignature *sig, gboolean has_target)
8146 guint8 *code, *start;
8149 if (sig->param_count > MAX_ARCH_DELEGATE_PARAMS)
8152 /* FIXME: Support more cases */
8153 if (MONO_TYPE_ISSTRUCT (mini_get_underlying_type (sig->ret)))
8157 static guint8* cached = NULL;
8162 if (mono_aot_only) {
8163 start = mono_aot_get_trampoline ("delegate_invoke_impl_has_target");
8165 MonoTrampInfo *info;
8166 start = get_delegate_invoke_impl (&info, TRUE, 0);
8167 mono_tramp_info_register (info, NULL);
8170 mono_memory_barrier ();
8174 static guint8* cache [MAX_ARCH_DELEGATE_PARAMS + 1] = {NULL};
8175 for (i = 0; i < sig->param_count; ++i)
8176 if (!mono_is_regsize_var (sig->params [i]))
8178 if (sig->param_count > 4)
8181 code = cache [sig->param_count];
8185 if (mono_aot_only) {
8186 char *name = g_strdup_printf ("delegate_invoke_impl_target_%d", sig->param_count);
8187 start = mono_aot_get_trampoline (name);
8190 MonoTrampInfo *info;
8191 start = get_delegate_invoke_impl (&info, FALSE, sig->param_count);
8192 mono_tramp_info_register (info, NULL);
8195 mono_memory_barrier ();
8197 cache [sig->param_count] = start;
8204 mono_arch_get_delegate_virtual_invoke_impl (MonoMethodSignature *sig, MonoMethod *method, int offset, gboolean load_imt_reg)
8206 MonoTrampInfo *info;
8209 code = get_delegate_virtual_invoke_impl (&info, load_imt_reg, offset);
8211 mono_tramp_info_register (info, NULL);
8216 mono_arch_finish_init (void)
8218 #if !defined(HOST_WIN32) && defined(MONO_XEN_OPT)
8219 optimize_for_xen = access ("/proc/xen", F_OK) == 0;
8224 mono_arch_free_jit_tls_data (MonoJitTlsData *tls)
8228 #if defined(__default_codegen__)
8229 #define CMP_SIZE (6 + 1)
8230 #define CMP_REG_REG_SIZE (4 + 1)
8231 #define BR_SMALL_SIZE 2
8232 #define BR_LARGE_SIZE 6
8233 #define MOV_REG_IMM_SIZE 10
8234 #define MOV_REG_IMM_32BIT_SIZE 6
8235 #define JUMP_REG_SIZE (2 + 1)
8236 #elif defined(__native_client_codegen__)
8237 /* NaCl N-byte instructions can be padded up to N-1 bytes */
8238 #define CMP_SIZE ((6 + 1) * 2 - 1)
8239 #define CMP_REG_REG_SIZE ((4 + 1) * 2 - 1)
8240 #define BR_SMALL_SIZE (2 * 2 - 1)
8241 #define BR_LARGE_SIZE (6 * 2 - 1)
8242 #define MOV_REG_IMM_SIZE (10 * 2 - 1)
8243 #define MOV_REG_IMM_32BIT_SIZE (6 * 2 - 1)
8244 /* Jump reg for NaCl adds a mask (+4) and add (+3) */
8245 #define JUMP_REG_SIZE ((2 + 1 + 4 + 3) * 2 - 1)
8246 /* Jump membase's size is large and unpredictable */
8247 /* in native client, just pad it out a whole bundle. */
8248 #define JUMP_MEMBASE_SIZE (kNaClAlignment)
8252 imt_branch_distance (MonoIMTCheckItem **imt_entries, int start, int target)
8254 int i, distance = 0;
8255 for (i = start; i < target; ++i)
8256 distance += imt_entries [i]->chunk_size;
8261 * LOCKING: called with the domain lock held
8264 mono_arch_build_imt_thunk (MonoVTable *vtable, MonoDomain *domain, MonoIMTCheckItem **imt_entries, int count,
8265 gpointer fail_tramp)
8269 guint8 *code, *start;
8270 gboolean vtable_is_32bit = ((gsize)(vtable) == (gsize)(int)(gsize)(vtable));
8273 for (i = 0; i < count; ++i) {
8274 MonoIMTCheckItem *item = imt_entries [i];
8275 if (item->is_equals) {
8276 if (item->check_target_idx) {
8277 if (!item->compare_done) {
8278 if (amd64_use_imm32 ((gint64)item->key))
8279 item->chunk_size += CMP_SIZE;
8281 item->chunk_size += MOV_REG_IMM_SIZE + CMP_REG_REG_SIZE;
8283 if (item->has_target_code) {
8284 item->chunk_size += MOV_REG_IMM_SIZE;
8286 if (vtable_is_32bit)
8287 item->chunk_size += MOV_REG_IMM_32BIT_SIZE;
8289 item->chunk_size += MOV_REG_IMM_SIZE;
8290 #ifdef __native_client_codegen__
8291 item->chunk_size += JUMP_MEMBASE_SIZE;
8294 item->chunk_size += BR_SMALL_SIZE + JUMP_REG_SIZE;
8297 item->chunk_size += MOV_REG_IMM_SIZE * 3 + CMP_REG_REG_SIZE +
8298 BR_SMALL_SIZE + JUMP_REG_SIZE * 2;
8300 if (vtable_is_32bit)
8301 item->chunk_size += MOV_REG_IMM_32BIT_SIZE;
8303 item->chunk_size += MOV_REG_IMM_SIZE;
8304 item->chunk_size += JUMP_REG_SIZE;
8305 /* with assert below:
8306 * item->chunk_size += CMP_SIZE + BR_SMALL_SIZE + 1;
8308 #ifdef __native_client_codegen__
8309 item->chunk_size += JUMP_MEMBASE_SIZE;
8314 if (amd64_use_imm32 ((gint64)item->key))
8315 item->chunk_size += CMP_SIZE;
8317 item->chunk_size += MOV_REG_IMM_SIZE + CMP_REG_REG_SIZE;
8318 item->chunk_size += BR_LARGE_SIZE;
8319 imt_entries [item->check_target_idx]->compare_done = TRUE;
8321 size += item->chunk_size;
8323 #if defined(__native_client__) && defined(__native_client_codegen__)
8324 /* In Native Client, we don't re-use thunks, allocate from the */
8325 /* normal code manager paths. */
8326 code = mono_domain_code_reserve (domain, size);
8329 code = mono_method_alloc_generic_virtual_thunk (domain, size);
8331 code = mono_domain_code_reserve (domain, size);
8335 unwind_ops = mono_arch_get_cie_program ();
8337 for (i = 0; i < count; ++i) {
8338 MonoIMTCheckItem *item = imt_entries [i];
8339 item->code_target = code;
8340 if (item->is_equals) {
8341 gboolean fail_case = !item->check_target_idx && fail_tramp;
8343 if (item->check_target_idx || fail_case) {
8344 if (!item->compare_done || fail_case) {
8345 if (amd64_use_imm32 ((gint64)item->key))
8346 amd64_alu_reg_imm_size (code, X86_CMP, MONO_ARCH_IMT_REG, (guint32)(gssize)item->key, sizeof(gpointer));
8348 amd64_mov_reg_imm_size (code, MONO_ARCH_IMT_SCRATCH_REG, item->key, sizeof(gpointer));
8349 amd64_alu_reg_reg (code, X86_CMP, MONO_ARCH_IMT_REG, MONO_ARCH_IMT_SCRATCH_REG);
8352 item->jmp_code = code;
8353 amd64_branch8 (code, X86_CC_NE, 0, FALSE);
8354 if (item->has_target_code) {
8355 amd64_mov_reg_imm (code, MONO_ARCH_IMT_SCRATCH_REG, item->value.target_code);
8356 amd64_jump_reg (code, MONO_ARCH_IMT_SCRATCH_REG);
8358 amd64_mov_reg_imm (code, MONO_ARCH_IMT_SCRATCH_REG, & (vtable->vtable [item->value.vtable_slot]));
8359 amd64_jump_membase (code, MONO_ARCH_IMT_SCRATCH_REG, 0);
8363 amd64_patch (item->jmp_code, code);
8364 amd64_mov_reg_imm (code, MONO_ARCH_IMT_SCRATCH_REG, fail_tramp);
8365 amd64_jump_reg (code, MONO_ARCH_IMT_SCRATCH_REG);
8366 item->jmp_code = NULL;
8369 /* enable the commented code to assert on wrong method */
8371 if (amd64_is_imm32 (item->key))
8372 amd64_alu_reg_imm_size (code, X86_CMP, MONO_ARCH_IMT_REG, (guint32)(gssize)item->key, sizeof(gpointer));
8374 amd64_mov_reg_imm (code, MONO_ARCH_IMT_SCRATCH_REG, item->key);
8375 amd64_alu_reg_reg (code, X86_CMP, MONO_ARCH_IMT_REG, MONO_ARCH_IMT_SCRATCH_REG);
8377 item->jmp_code = code;
8378 amd64_branch8 (code, X86_CC_NE, 0, FALSE);
8379 /* See the comment below about R10 */
8380 amd64_mov_reg_imm (code, MONO_ARCH_IMT_SCRATCH_REG, & (vtable->vtable [item->value.vtable_slot]));
8381 amd64_jump_membase (code, MONO_ARCH_IMT_SCRATCH_REG, 0);
8382 amd64_patch (item->jmp_code, code);
8383 amd64_breakpoint (code);
8384 item->jmp_code = NULL;
8386 /* We're using R10 (MONO_ARCH_IMT_SCRATCH_REG) here because R11 (MONO_ARCH_IMT_REG)
8387 needs to be preserved. R10 needs
8388 to be preserved for calls which
8389 require a runtime generic context,
8390 but interface calls don't. */
8391 amd64_mov_reg_imm (code, MONO_ARCH_IMT_SCRATCH_REG, & (vtable->vtable [item->value.vtable_slot]));
8392 amd64_jump_membase (code, MONO_ARCH_IMT_SCRATCH_REG, 0);
8396 if (amd64_use_imm32 ((gint64)item->key))
8397 amd64_alu_reg_imm_size (code, X86_CMP, MONO_ARCH_IMT_REG, (guint32)(gssize)item->key, sizeof (gpointer));
8399 amd64_mov_reg_imm_size (code, MONO_ARCH_IMT_SCRATCH_REG, item->key, sizeof (gpointer));
8400 amd64_alu_reg_reg (code, X86_CMP, MONO_ARCH_IMT_REG, MONO_ARCH_IMT_SCRATCH_REG);
8402 item->jmp_code = code;
8403 if (x86_is_imm8 (imt_branch_distance (imt_entries, i, item->check_target_idx)))
8404 x86_branch8 (code, X86_CC_GE, 0, FALSE);
8406 x86_branch32 (code, X86_CC_GE, 0, FALSE);
8408 g_assert (code - item->code_target <= item->chunk_size);
8410 /* patch the branches to get to the target items */
8411 for (i = 0; i < count; ++i) {
8412 MonoIMTCheckItem *item = imt_entries [i];
8413 if (item->jmp_code) {
8414 if (item->check_target_idx) {
8415 amd64_patch (item->jmp_code, imt_entries [item->check_target_idx]->code_target);
8421 mono_stats.imt_thunks_size += code - start;
8422 g_assert (code - start <= size);
8424 nacl_domain_code_validate(domain, &start, size, &code);
8425 mono_profiler_code_buffer_new (start, code - start, MONO_PROFILER_CODE_BUFFER_IMT_TRAMPOLINE, NULL);
8427 mono_tramp_info_register (mono_tramp_info_create (NULL, start, code - start, NULL, unwind_ops), domain);
8433 mono_arch_find_imt_method (mgreg_t *regs, guint8 *code)
8435 return (MonoMethod*)regs [MONO_ARCH_IMT_REG];
8439 mono_arch_find_static_call_vtable (mgreg_t *regs, guint8 *code)
8441 return (MonoVTable*) regs [MONO_ARCH_RGCTX_REG];
8445 mono_arch_get_cie_program (void)
8449 mono_add_unwind_op_def_cfa (l, (guint8*)NULL, (guint8*)NULL, AMD64_RSP, 8);
8450 mono_add_unwind_op_offset (l, (guint8*)NULL, (guint8*)NULL, AMD64_RIP, -8);
8458 mono_arch_emit_inst_for_method (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature *fsig, MonoInst **args)
8460 MonoInst *ins = NULL;
8463 if (cmethod->klass == mono_defaults.math_class) {
8464 if (strcmp (cmethod->name, "Sin") == 0) {
8466 } else if (strcmp (cmethod->name, "Cos") == 0) {
8468 } else if (strcmp (cmethod->name, "Sqrt") == 0) {
8470 } else if (strcmp (cmethod->name, "Abs") == 0 && fsig->params [0]->type == MONO_TYPE_R8) {
8474 if (opcode && fsig->param_count == 1) {
8475 MONO_INST_NEW (cfg, ins, opcode);
8476 ins->type = STACK_R8;
8477 ins->dreg = mono_alloc_freg (cfg);
8478 ins->sreg1 = args [0]->dreg;
8479 MONO_ADD_INS (cfg->cbb, ins);
8483 if (cfg->opt & MONO_OPT_CMOV) {
8484 if (strcmp (cmethod->name, "Min") == 0) {
8485 if (fsig->params [0]->type == MONO_TYPE_I4)
8487 if (fsig->params [0]->type == MONO_TYPE_U4)
8488 opcode = OP_IMIN_UN;
8489 else if (fsig->params [0]->type == MONO_TYPE_I8)
8491 else if (fsig->params [0]->type == MONO_TYPE_U8)
8492 opcode = OP_LMIN_UN;
8493 } else if (strcmp (cmethod->name, "Max") == 0) {
8494 if (fsig->params [0]->type == MONO_TYPE_I4)
8496 if (fsig->params [0]->type == MONO_TYPE_U4)
8497 opcode = OP_IMAX_UN;
8498 else if (fsig->params [0]->type == MONO_TYPE_I8)
8500 else if (fsig->params [0]->type == MONO_TYPE_U8)
8501 opcode = OP_LMAX_UN;
8505 if (opcode && fsig->param_count == 2) {
8506 MONO_INST_NEW (cfg, ins, opcode);
8507 ins->type = fsig->params [0]->type == MONO_TYPE_I4 ? STACK_I4 : STACK_I8;
8508 ins->dreg = mono_alloc_ireg (cfg);
8509 ins->sreg1 = args [0]->dreg;
8510 ins->sreg2 = args [1]->dreg;
8511 MONO_ADD_INS (cfg->cbb, ins);
8515 /* OP_FREM is not IEEE compatible */
8516 else if (strcmp (cmethod->name, "IEEERemainder") == 0 && fsig->param_count == 2) {
8517 MONO_INST_NEW (cfg, ins, OP_FREM);
8518 ins->inst_i0 = args [0];
8519 ins->inst_i1 = args [1];
8529 mono_arch_print_tree (MonoInst *tree, int arity)
8535 mono_arch_context_get_int_reg (MonoContext *ctx, int reg)
8537 return ctx->gregs [reg];
8541 mono_arch_context_set_int_reg (MonoContext *ctx, int reg, mgreg_t val)
8543 ctx->gregs [reg] = val;
8547 mono_arch_install_handler_block_guard (MonoJitInfo *ji, MonoJitExceptionInfo *clause, MonoContext *ctx, gpointer new_value)
8549 gpointer *sp, old_value;
8553 bp = MONO_CONTEXT_GET_BP (ctx);
8554 sp = *(gpointer*)(bp + clause->exvar_offset);
8557 if (old_value < ji->code_start || (char*)old_value > ((char*)ji->code_start + ji->code_size))
8566 * mono_arch_emit_load_aotconst:
8568 * Emit code to load the contents of the GOT slot identified by TRAMP_TYPE and
8569 * TARGET from the mscorlib GOT in full-aot code.
8570 * On AMD64, the result is placed into R11.
8573 mono_arch_emit_load_aotconst (guint8 *start, guint8 *code, MonoJumpInfo **ji, int tramp_type, gconstpointer target)
8575 *ji = mono_patch_info_list_prepend (*ji, code - start, tramp_type, target);
8576 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RIP, 0, 8);
8582 * mono_arch_get_trampolines:
8584 * Return a list of MonoTrampInfo structures describing arch specific trampolines
8588 mono_arch_get_trampolines (gboolean aot)
8590 return mono_amd64_get_exception_trampolines (aot);
8593 /* Soft Debug support */
8594 #ifdef MONO_ARCH_SOFT_DEBUG_SUPPORTED
8597 * mono_arch_set_breakpoint:
8599 * Set a breakpoint at the native code corresponding to JI at NATIVE_OFFSET.
8600 * The location should contain code emitted by OP_SEQ_POINT.
8603 mono_arch_set_breakpoint (MonoJitInfo *ji, guint8 *ip)
8608 guint32 native_offset = ip - (guint8*)ji->code_start;
8609 SeqPointInfo *info = mono_arch_get_seq_point_info (mono_domain_get (), ji->code_start);
8611 g_assert (info->bp_addrs [native_offset] == 0);
8612 info->bp_addrs [native_offset] = mini_get_breakpoint_trampoline ();
8614 /* ip points to a mov r11, 0 */
8615 g_assert (code [0] == 0x41);
8616 g_assert (code [1] == 0xbb);
8617 amd64_mov_reg_imm (code, AMD64_R11, 1);
8622 * mono_arch_clear_breakpoint:
8624 * Clear the breakpoint at IP.
8627 mono_arch_clear_breakpoint (MonoJitInfo *ji, guint8 *ip)
8632 guint32 native_offset = ip - (guint8*)ji->code_start;
8633 SeqPointInfo *info = mono_arch_get_seq_point_info (mono_domain_get (), ji->code_start);
8635 info->bp_addrs [native_offset] = NULL;
8637 amd64_mov_reg_imm (code, AMD64_R11, 0);
8642 mono_arch_is_breakpoint_event (void *info, void *sigctx)
8644 /* We use soft breakpoints on amd64 */
8649 * mono_arch_skip_breakpoint:
8651 * Modify CTX so the ip is placed after the breakpoint instruction, so when
8652 * we resume, the instruction is not executed again.
8655 mono_arch_skip_breakpoint (MonoContext *ctx, MonoJitInfo *ji)
8657 g_assert_not_reached ();
8661 * mono_arch_start_single_stepping:
8663 * Start single stepping.
8666 mono_arch_start_single_stepping (void)
8668 ss_trampoline = mini_get_single_step_trampoline ();
8672 * mono_arch_stop_single_stepping:
8674 * Stop single stepping.
8677 mono_arch_stop_single_stepping (void)
8679 ss_trampoline = NULL;
8683 * mono_arch_is_single_step_event:
8685 * Return whenever the machine state in SIGCTX corresponds to a single
8689 mono_arch_is_single_step_event (void *info, void *sigctx)
8691 /* We use soft breakpoints on amd64 */
8696 * mono_arch_skip_single_step:
8698 * Modify CTX so the ip is placed after the single step trigger instruction,
8699 * we resume, the instruction is not executed again.
8702 mono_arch_skip_single_step (MonoContext *ctx)
8704 g_assert_not_reached ();
8708 * mono_arch_create_seq_point_info:
8710 * Return a pointer to a data structure which is used by the sequence
8711 * point implementation in AOTed code.
8714 mono_arch_get_seq_point_info (MonoDomain *domain, guint8 *code)
8719 // FIXME: Add a free function
8721 mono_domain_lock (domain);
8722 info = g_hash_table_lookup (domain_jit_info (domain)->arch_seq_points,
8724 mono_domain_unlock (domain);
8727 ji = mono_jit_info_table_find (domain, (char*)code);
8730 // FIXME: Optimize the size
8731 info = g_malloc0 (sizeof (SeqPointInfo) + (ji->code_size * sizeof (gpointer)));
8733 info->ss_tramp_addr = &ss_trampoline;
8735 mono_domain_lock (domain);
8736 g_hash_table_insert (domain_jit_info (domain)->arch_seq_points,
8738 mono_domain_unlock (domain);
8745 mono_arch_init_lmf_ext (MonoLMFExt *ext, gpointer prev_lmf)
8747 ext->lmf.previous_lmf = prev_lmf;
8748 /* Mark that this is a MonoLMFExt */
8749 ext->lmf.previous_lmf = (gpointer)(((gssize)ext->lmf.previous_lmf) | 2);
8750 ext->lmf.rsp = (gssize)ext;
8756 mono_arch_opcode_supported (int opcode)
8759 case OP_ATOMIC_ADD_I4:
8760 case OP_ATOMIC_ADD_I8:
8761 case OP_ATOMIC_EXCHANGE_I4:
8762 case OP_ATOMIC_EXCHANGE_I8:
8763 case OP_ATOMIC_CAS_I4:
8764 case OP_ATOMIC_CAS_I8:
8765 case OP_ATOMIC_LOAD_I1:
8766 case OP_ATOMIC_LOAD_I2:
8767 case OP_ATOMIC_LOAD_I4:
8768 case OP_ATOMIC_LOAD_I8:
8769 case OP_ATOMIC_LOAD_U1:
8770 case OP_ATOMIC_LOAD_U2:
8771 case OP_ATOMIC_LOAD_U4:
8772 case OP_ATOMIC_LOAD_U8:
8773 case OP_ATOMIC_LOAD_R4:
8774 case OP_ATOMIC_LOAD_R8:
8775 case OP_ATOMIC_STORE_I1:
8776 case OP_ATOMIC_STORE_I2:
8777 case OP_ATOMIC_STORE_I4:
8778 case OP_ATOMIC_STORE_I8:
8779 case OP_ATOMIC_STORE_U1:
8780 case OP_ATOMIC_STORE_U2:
8781 case OP_ATOMIC_STORE_U4:
8782 case OP_ATOMIC_STORE_U8:
8783 case OP_ATOMIC_STORE_R4:
8784 case OP_ATOMIC_STORE_R8:
8791 #if defined(ENABLE_GSHAREDVT)
8793 #include "../../../mono-extensions/mono/mini/mini-amd64-gsharedvt.c"
8795 #endif /* !MONOTOUCH */