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.
21 #include <mono/metadata/appdomain.h>
22 #include <mono/metadata/debug-helpers.h>
23 #include <mono/metadata/threads.h>
24 #include <mono/metadata/profiler-private.h>
25 #include <mono/metadata/mono-debug.h>
26 #include <mono/metadata/gc-internal.h>
27 #include <mono/utils/mono-math.h>
28 #include <mono/utils/mono-mmap.h>
32 #include "mini-amd64.h"
33 #include "cpu-amd64.h"
34 #include "debugger-agent.h"
37 static gint lmf_tls_offset = -1;
38 static gint lmf_addr_tls_offset = -1;
39 static gint appdomain_tls_offset = -1;
42 static gboolean optimize_for_xen = TRUE;
44 #define optimize_for_xen 0
47 #define ALIGN_TO(val,align) ((((guint64)val) + ((align) - 1)) & ~((align) - 1))
49 #define IS_IMM32(val) ((((guint64)val) >> 32) == 0)
51 #define IS_REX(inst) (((inst) >= 0x40) && ((inst) <= 0x4f))
54 /* Under windows, the calling convention is never stdcall */
55 #define CALLCONV_IS_STDCALL(call_conv) (FALSE)
57 #define CALLCONV_IS_STDCALL(call_conv) ((call_conv) == MONO_CALL_STDCALL)
60 /* This mutex protects architecture specific caches */
61 #define mono_mini_arch_lock() EnterCriticalSection (&mini_arch_mutex)
62 #define mono_mini_arch_unlock() LeaveCriticalSection (&mini_arch_mutex)
63 static CRITICAL_SECTION mini_arch_mutex;
66 mono_breakpoint_info [MONO_BREAKPOINT_ARRAY_SIZE];
69 * The code generated for sequence points reads from this location, which is
70 * made read-only when single stepping is enabled.
72 static gpointer ss_trigger_page;
74 /* Enabled breakpoints read from this trigger page */
75 static gpointer bp_trigger_page;
77 /* The size of the breakpoint sequence */
78 static int breakpoint_size;
80 /* The size of the breakpoint instruction causing the actual fault */
81 static int breakpoint_fault_size;
83 /* The size of the single step instruction causing the actual fault */
84 static int single_step_fault_size;
87 /* On Win64 always reserve first 32 bytes for first four arguments */
88 #define ARGS_OFFSET 48
90 #define ARGS_OFFSET 16
92 #define GP_SCRATCH_REG AMD64_R11
95 * AMD64 register usage:
96 * - callee saved registers are used for global register allocation
97 * - %r11 is used for materializing 64 bit constants in opcodes
98 * - the rest is used for local allocation
102 * Floating point comparison results:
112 mono_arch_regname (int reg)
115 case AMD64_RAX: return "%rax";
116 case AMD64_RBX: return "%rbx";
117 case AMD64_RCX: return "%rcx";
118 case AMD64_RDX: return "%rdx";
119 case AMD64_RSP: return "%rsp";
120 case AMD64_RBP: return "%rbp";
121 case AMD64_RDI: return "%rdi";
122 case AMD64_RSI: return "%rsi";
123 case AMD64_R8: return "%r8";
124 case AMD64_R9: return "%r9";
125 case AMD64_R10: return "%r10";
126 case AMD64_R11: return "%r11";
127 case AMD64_R12: return "%r12";
128 case AMD64_R13: return "%r13";
129 case AMD64_R14: return "%r14";
130 case AMD64_R15: return "%r15";
135 static const char * packed_xmmregs [] = {
136 "p:xmm0", "p:xmm1", "p:xmm2", "p:xmm3", "p:xmm4", "p:xmm5", "p:xmm6", "p:xmm7", "p:xmm8",
137 "p:xmm9", "p:xmm10", "p:xmm11", "p:xmm12", "p:xmm13", "p:xmm14", "p:xmm15"
140 static const char * single_xmmregs [] = {
141 "s:xmm0", "s:xmm1", "s:xmm2", "s:xmm3", "s:xmm4", "s:xmm5", "s:xmm6", "s:xmm7", "s:xmm8",
142 "s:xmm9", "s:xmm10", "s:xmm11", "s:xmm12", "s:xmm13", "s:xmm14", "s:xmm15"
146 mono_arch_fregname (int reg)
148 if (reg < AMD64_XMM_NREG)
149 return single_xmmregs [reg];
155 mono_arch_xregname (int reg)
157 if (reg < AMD64_XMM_NREG)
158 return packed_xmmregs [reg];
163 G_GNUC_UNUSED static void
168 G_GNUC_UNUSED static gboolean
171 static int count = 0;
174 if (!getenv ("COUNT"))
177 if (count == atoi (getenv ("COUNT"))) {
181 if (count > atoi (getenv ("COUNT"))) {
192 return debug_count ();
198 static inline gboolean
199 amd64_is_near_call (guint8 *code)
202 if ((code [0] >= 0x40) && (code [0] <= 0x4f))
205 return code [0] == 0xe8;
208 #ifdef __native_client_codegen__
210 /* Keep track of instruction "depth", that is, the level of sub-instruction */
211 /* for any given instruction. For instance, amd64_call_reg resolves to */
212 /* amd64_call_reg_internal, which uses amd64_alu_* macros, etc. */
213 /* We only want to force bundle alignment for the top level instruction, */
214 /* so NaCl pseudo-instructions can be implemented with sub instructions. */
215 static guint32 nacl_instruction_depth;
217 static guint32 nacl_rex_tag;
218 static guint32 nacl_legacy_prefix_tag;
221 amd64_nacl_clear_legacy_prefix_tag ()
223 TlsSetValue (nacl_legacy_prefix_tag, NULL);
227 amd64_nacl_tag_legacy_prefix (guint8* code)
229 if (TlsGetValue (nacl_legacy_prefix_tag) == NULL)
230 TlsSetValue (nacl_legacy_prefix_tag, code);
234 amd64_nacl_tag_rex (guint8* code)
236 TlsSetValue (nacl_rex_tag, code);
240 amd64_nacl_get_legacy_prefix_tag ()
242 return (guint8*)TlsGetValue (nacl_legacy_prefix_tag);
246 amd64_nacl_get_rex_tag ()
248 return (guint8*)TlsGetValue (nacl_rex_tag);
251 /* Increment the instruction "depth" described above */
253 amd64_nacl_instruction_pre ()
255 intptr_t depth = (intptr_t) TlsGetValue (nacl_instruction_depth);
257 TlsSetValue (nacl_instruction_depth, (gpointer)depth);
260 /* amd64_nacl_instruction_post: Decrement instruction "depth", force bundle */
261 /* alignment if depth == 0 (top level instruction) */
262 /* IN: start, end pointers to instruction beginning and end */
263 /* OUT: start, end pointers to beginning and end after possible alignment */
264 /* GLOBALS: nacl_instruction_depth defined above */
266 amd64_nacl_instruction_post (guint8 **start, guint8 **end)
268 intptr_t depth = (intptr_t) TlsGetValue(nacl_instruction_depth);
270 TlsSetValue (nacl_instruction_depth, (void*)depth);
272 g_assert ( depth >= 0 );
274 uintptr_t space_in_block;
276 guint8 *prefix = amd64_nacl_get_legacy_prefix_tag ();
277 /* if legacy prefix is present, and if it was emitted before */
278 /* the start of the instruction sequence, adjust the start */
279 if (prefix != NULL && prefix < *start) {
280 g_assert (*start - prefix <= 3);/* only 3 are allowed */
283 space_in_block = kNaClAlignment - ((uintptr_t)(*start) & kNaClAlignmentMask);
284 instlen = (uintptr_t)(*end - *start);
285 /* Only check for instructions which are less than */
286 /* kNaClAlignment. The only instructions that should ever */
287 /* be that long are call sequences, which are already */
288 /* padded out to align the return to the next bundle. */
289 if (instlen > space_in_block && instlen < kNaClAlignment) {
290 const size_t MAX_NACL_INST_LENGTH = kNaClAlignment;
291 guint8 copy_of_instruction[MAX_NACL_INST_LENGTH];
292 const size_t length = (size_t)((*end)-(*start));
293 g_assert (length < MAX_NACL_INST_LENGTH);
295 memcpy (copy_of_instruction, *start, length);
296 *start = mono_arch_nacl_pad (*start, space_in_block);
297 memcpy (*start, copy_of_instruction, length);
298 *end = *start + length;
300 amd64_nacl_clear_legacy_prefix_tag ();
301 amd64_nacl_tag_rex (NULL);
305 /* amd64_nacl_membase_handler: ensure all access to memory of the form */
306 /* OFFSET(%rXX) is sandboxed. For allowable base registers %rip, %rbp, */
307 /* %rsp, and %r15, emit the membase as usual. For all other registers, */
308 /* make sure the upper 32-bits are cleared, and use that register in the */
309 /* index field of a new address of this form: OFFSET(%r15,%eXX,1) */
311 /* pointer to current instruction stream (in the */
312 /* middle of an instruction, after opcode is emitted) */
313 /* basereg/offset/dreg */
314 /* operands of normal membase address */
316 /* pointer to the end of the membase/memindex emit */
317 /* GLOBALS: nacl_rex_tag */
318 /* position in instruction stream that rex prefix was emitted */
319 /* nacl_legacy_prefix_tag */
320 /* (possibly NULL) position in instruction of legacy x86 prefix */
322 amd64_nacl_membase_handler (guint8** code, gint8 basereg, gint32 offset, gint8 dreg)
324 gint8 true_basereg = basereg;
326 /* Cache these values, they might change */
327 /* as new instructions are emitted below. */
328 guint8* rex_tag = amd64_nacl_get_rex_tag ();
329 guint8* legacy_prefix_tag = amd64_nacl_get_legacy_prefix_tag ();
331 /* 'basereg' is given masked to 0x7 at this point, so check */
332 /* the rex prefix to see if this is an extended register. */
333 if ((rex_tag != NULL) && IS_REX(*rex_tag) && (*rex_tag & AMD64_REX_B)) {
337 #define X86_LEA_OPCODE (0x8D)
339 if (!amd64_is_valid_nacl_base (true_basereg) && (*(*code-1) != X86_LEA_OPCODE)) {
340 guint8* old_instruction_start;
342 /* This will hold the 'mov %eXX, %eXX' that clears the upper */
343 /* 32-bits of the old base register (new index register) */
345 guint8* buf_ptr = buf;
348 g_assert (rex_tag != NULL);
350 if (IS_REX(*rex_tag)) {
351 /* The old rex.B should be the new rex.X */
352 if (*rex_tag & AMD64_REX_B) {
353 *rex_tag |= AMD64_REX_X;
355 /* Since our new base is %r15 set rex.B */
356 *rex_tag |= AMD64_REX_B;
358 /* Shift the instruction by one byte */
359 /* so we can insert a rex prefix */
360 memmove (rex_tag + 1, rex_tag, (size_t)(*code - rex_tag));
362 /* New rex prefix only needs rex.B for %r15 base */
363 *rex_tag = AMD64_REX(AMD64_REX_B);
366 if (legacy_prefix_tag) {
367 old_instruction_start = legacy_prefix_tag;
369 old_instruction_start = rex_tag;
372 /* Clears the upper 32-bits of the previous base register */
373 amd64_mov_reg_reg_size (buf_ptr, true_basereg, true_basereg, 4);
374 insert_len = buf_ptr - buf;
376 /* Move the old instruction forward to make */
377 /* room for 'mov' stored in 'buf_ptr' */
378 memmove (old_instruction_start + insert_len, old_instruction_start, (size_t)(*code - old_instruction_start));
380 memcpy (old_instruction_start, buf, insert_len);
382 /* Sandboxed replacement for the normal membase_emit */
383 x86_memindex_emit (*code, dreg, AMD64_R15, offset, basereg, 0);
386 /* Normal default behavior, emit membase memory location */
387 x86_membase_emit_body (*code, dreg, basereg, offset);
392 static inline unsigned char*
393 amd64_skip_nops (unsigned char* code)
398 if ( code[0] == 0x90) {
402 if ( code[0] == 0x66 && code[1] == 0x90) {
406 if (code[0] == 0x0f && code[1] == 0x1f
407 && code[2] == 0x00) {
411 if (code[0] == 0x0f && code[1] == 0x1f
412 && code[2] == 0x40 && code[3] == 0x00) {
416 if (code[0] == 0x0f && code[1] == 0x1f
417 && code[2] == 0x44 && code[3] == 0x00
418 && code[4] == 0x00) {
422 if (code[0] == 0x66 && code[1] == 0x0f
423 && code[2] == 0x1f && code[3] == 0x44
424 && code[4] == 0x00 && code[5] == 0x00) {
428 if (code[0] == 0x0f && code[1] == 0x1f
429 && code[2] == 0x80 && code[3] == 0x00
430 && code[4] == 0x00 && code[5] == 0x00
431 && code[6] == 0x00) {
435 if (code[0] == 0x0f && code[1] == 0x1f
436 && code[2] == 0x84 && code[3] == 0x00
437 && code[4] == 0x00 && code[5] == 0x00
438 && code[6] == 0x00 && code[7] == 0x00) {
447 mono_arch_nacl_skip_nops (guint8* code)
449 return amd64_skip_nops(code);
452 #endif /*__native_client_codegen__*/
455 amd64_patch (unsigned char* code, gpointer target)
459 #ifdef __native_client_codegen__
460 code = amd64_skip_nops (code);
462 #if defined(__native_client_codegen__) && defined(__native_client__)
463 if (nacl_is_code_address (code)) {
464 /* For tail calls, code is patched after being installed */
465 /* but not through the normal "patch callsite" method. */
466 unsigned char buf[kNaClAlignment];
467 unsigned char *aligned_code = (uintptr_t)code & ~kNaClAlignmentMask;
469 memcpy (buf, aligned_code, kNaClAlignment);
470 /* Patch a temp buffer of bundle size, */
471 /* then install to actual location. */
472 amd64_patch (buf + ((uintptr_t)code - (uintptr_t)aligned_code), target);
473 ret = nacl_dyncode_modify (aligned_code, buf, kNaClAlignment);
477 target = nacl_modify_patch_target (target);
481 if ((code [0] >= 0x40) && (code [0] <= 0x4f)) {
486 if ((code [0] & 0xf8) == 0xb8) {
487 /* amd64_set_reg_template */
488 *(guint64*)(code + 1) = (guint64)target;
490 else if ((code [0] == 0x8b) && rex && x86_modrm_mod (code [1]) == 0 && x86_modrm_rm (code [1]) == 5) {
491 /* mov 0(%rip), %dreg */
492 *(guint32*)(code + 2) = (guint32)(guint64)target - 7;
494 else if ((code [0] == 0xff) && (code [1] == 0x15)) {
495 /* call *<OFFSET>(%rip) */
496 *(guint32*)(code + 2) = ((guint32)(guint64)target) - 7;
498 else if ((code [0] == 0xe8)) {
500 gint64 disp = (guint8*)target - (guint8*)code;
501 g_assert (amd64_is_imm32 (disp));
502 x86_patch (code, (unsigned char*)target);
505 x86_patch (code, (unsigned char*)target);
509 mono_amd64_patch (unsigned char* code, gpointer target)
511 amd64_patch (code, target);
520 ArgValuetypeAddrInIReg,
521 ArgNone /* only in pair_storage */
529 /* Only if storage == ArgValuetypeInReg */
530 ArgStorage pair_storage [2];
540 gboolean need_stack_align;
541 gboolean vtype_retaddr;
542 /* The index of the vret arg in the argument list */
549 #define DEBUG(a) if (cfg->verbose_level > 1) a
554 static AMD64_Reg_No param_regs [] = { AMD64_RCX, AMD64_RDX, AMD64_R8, AMD64_R9 };
556 static AMD64_Reg_No return_regs [] = { AMD64_RAX, AMD64_RDX };
560 static AMD64_Reg_No param_regs [] = { AMD64_RDI, AMD64_RSI, AMD64_RDX, AMD64_RCX, AMD64_R8, AMD64_R9 };
562 static AMD64_Reg_No return_regs [] = { AMD64_RAX, AMD64_RDX };
566 add_general (guint32 *gr, guint32 *stack_size, ArgInfo *ainfo)
568 ainfo->offset = *stack_size;
570 if (*gr >= PARAM_REGS) {
571 ainfo->storage = ArgOnStack;
572 /* Since the same stack slot size is used for all arg */
573 /* types, it needs to be big enough to hold them all */
574 (*stack_size) += sizeof(mgreg_t);
577 ainfo->storage = ArgInIReg;
578 ainfo->reg = param_regs [*gr];
584 #define FLOAT_PARAM_REGS 4
586 #define FLOAT_PARAM_REGS 8
590 add_float (guint32 *gr, guint32 *stack_size, ArgInfo *ainfo, gboolean is_double)
592 ainfo->offset = *stack_size;
594 if (*gr >= FLOAT_PARAM_REGS) {
595 ainfo->storage = ArgOnStack;
596 /* Since the same stack slot size is used for both float */
597 /* types, it needs to be big enough to hold them both */
598 (*stack_size) += sizeof(mgreg_t);
601 /* A double register */
603 ainfo->storage = ArgInDoubleSSEReg;
605 ainfo->storage = ArgInFloatSSEReg;
611 typedef enum ArgumentClass {
619 merge_argument_class_from_type (MonoType *type, ArgumentClass class1)
621 ArgumentClass class2 = ARG_CLASS_NO_CLASS;
624 ptype = mini_type_get_underlying_type (NULL, type);
625 switch (ptype->type) {
626 case MONO_TYPE_BOOLEAN:
636 case MONO_TYPE_STRING:
637 case MONO_TYPE_OBJECT:
638 case MONO_TYPE_CLASS:
639 case MONO_TYPE_SZARRAY:
641 case MONO_TYPE_FNPTR:
642 case MONO_TYPE_ARRAY:
645 class2 = ARG_CLASS_INTEGER;
650 class2 = ARG_CLASS_INTEGER;
652 class2 = ARG_CLASS_SSE;
656 case MONO_TYPE_TYPEDBYREF:
657 g_assert_not_reached ();
659 case MONO_TYPE_GENERICINST:
660 if (!mono_type_generic_inst_is_valuetype (ptype)) {
661 class2 = ARG_CLASS_INTEGER;
665 case MONO_TYPE_VALUETYPE: {
666 MonoMarshalType *info = mono_marshal_load_type_info (ptype->data.klass);
669 for (i = 0; i < info->num_fields; ++i) {
671 class2 = merge_argument_class_from_type (info->fields [i].field->type, class2);
676 g_assert_not_reached ();
680 if (class1 == class2)
682 else if (class1 == ARG_CLASS_NO_CLASS)
684 else if ((class1 == ARG_CLASS_MEMORY) || (class2 == ARG_CLASS_MEMORY))
685 class1 = ARG_CLASS_MEMORY;
686 else if ((class1 == ARG_CLASS_INTEGER) || (class2 == ARG_CLASS_INTEGER))
687 class1 = ARG_CLASS_INTEGER;
689 class1 = ARG_CLASS_SSE;
693 #ifdef __native_client_codegen__
694 const guint kNaClAlignment = kNaClAlignmentAMD64;
695 const guint kNaClAlignmentMask = kNaClAlignmentMaskAMD64;
697 /* Default alignment for Native Client is 32-byte. */
698 gint8 nacl_align_byte = -32; /* signed version of 0xe0 */
700 /* mono_arch_nacl_pad: Add pad bytes of alignment instructions at code, */
701 /* Check that alignment doesn't cross an alignment boundary. */
703 mono_arch_nacl_pad(guint8 *code, int pad)
705 const int kMaxPadding = 8; /* see amd64-codegen.h:amd64_padding_size() */
707 if (pad == 0) return code;
708 /* assertion: alignment cannot cross a block boundary */
709 g_assert (((uintptr_t)code & (~kNaClAlignmentMask)) ==
710 (((uintptr_t)code + pad - 1) & (~kNaClAlignmentMask)));
711 while (pad >= kMaxPadding) {
712 amd64_padding (code, kMaxPadding);
715 if (pad != 0) amd64_padding (code, pad);
721 add_valuetype (MonoGenericSharingContext *gsctx, MonoMethodSignature *sig, ArgInfo *ainfo, MonoType *type,
723 guint32 *gr, guint32 *fr, guint32 *stack_size)
725 guint32 size, quad, nquads, i;
726 /* Keep track of the size used in each quad so we can */
727 /* use the right size when copying args/return vars. */
728 guint32 quadsize [2] = {8, 8};
729 ArgumentClass args [2];
730 MonoMarshalType *info = NULL;
732 MonoGenericSharingContext tmp_gsctx;
733 gboolean pass_on_stack = FALSE;
736 * The gsctx currently contains no data, it is only used for checking whenever
737 * open types are allowed, some callers like mono_arch_get_argument_info ()
738 * don't pass it to us, so work around that.
743 klass = mono_class_from_mono_type (type);
744 size = mini_type_stack_size_full (gsctx, &klass->byval_arg, NULL, sig->pinvoke);
746 if (!sig->pinvoke && !disable_vtypes_in_regs && ((is_return && (size == 8)) || (!is_return && (size <= 16)))) {
747 /* We pass and return vtypes of size 8 in a register */
748 } else if (!sig->pinvoke || (size == 0) || (size > 16)) {
749 pass_on_stack = TRUE;
753 pass_on_stack = TRUE;
757 /* If this struct can't be split up naturally into 8-byte */
758 /* chunks (registers), pass it on the stack. */
759 if (sig->pinvoke && !pass_on_stack) {
763 info = mono_marshal_load_type_info (klass);
765 for (i = 0; i < info->num_fields; ++i) {
766 field_size = mono_marshal_type_size (info->fields [i].field->type,
767 info->fields [i].mspec,
768 &align, TRUE, klass->unicode);
769 if ((info->fields [i].offset < 8) && (info->fields [i].offset + field_size) > 8) {
770 pass_on_stack = TRUE;
777 /* Allways pass in memory */
778 ainfo->offset = *stack_size;
779 *stack_size += ALIGN_TO (size, 8);
780 ainfo->storage = ArgOnStack;
785 /* FIXME: Handle structs smaller than 8 bytes */
786 //if ((size % 8) != 0)
795 /* Always pass in 1 or 2 integer registers */
796 args [0] = ARG_CLASS_INTEGER;
797 args [1] = ARG_CLASS_INTEGER;
798 /* Only the simplest cases are supported */
799 if (is_return && nquads != 1) {
800 args [0] = ARG_CLASS_MEMORY;
801 args [1] = ARG_CLASS_MEMORY;
805 * Implement the algorithm from section 3.2.3 of the X86_64 ABI.
806 * The X87 and SSEUP stuff is left out since there are no such types in
809 info = mono_marshal_load_type_info (klass);
813 if (info->native_size > 16) {
814 ainfo->offset = *stack_size;
815 *stack_size += ALIGN_TO (info->native_size, 8);
816 ainfo->storage = ArgOnStack;
821 switch (info->native_size) {
822 case 1: case 2: case 4: case 8:
826 ainfo->storage = ArgOnStack;
827 ainfo->offset = *stack_size;
828 *stack_size += ALIGN_TO (info->native_size, 8);
831 ainfo->storage = ArgValuetypeAddrInIReg;
833 if (*gr < PARAM_REGS) {
834 ainfo->pair_storage [0] = ArgInIReg;
835 ainfo->pair_regs [0] = param_regs [*gr];
839 ainfo->pair_storage [0] = ArgOnStack;
840 ainfo->offset = *stack_size;
849 args [0] = ARG_CLASS_NO_CLASS;
850 args [1] = ARG_CLASS_NO_CLASS;
851 for (quad = 0; quad < nquads; ++quad) {
854 ArgumentClass class1;
856 if (info->num_fields == 0)
857 class1 = ARG_CLASS_MEMORY;
859 class1 = ARG_CLASS_NO_CLASS;
860 for (i = 0; i < info->num_fields; ++i) {
861 size = mono_marshal_type_size (info->fields [i].field->type,
862 info->fields [i].mspec,
863 &align, TRUE, klass->unicode);
864 if ((info->fields [i].offset < 8) && (info->fields [i].offset + size) > 8) {
865 /* Unaligned field */
869 /* Skip fields in other quad */
870 if ((quad == 0) && (info->fields [i].offset >= 8))
872 if ((quad == 1) && (info->fields [i].offset < 8))
875 /* How far into this quad this data extends.*/
876 /* (8 is size of quad) */
877 quadsize [quad] = info->fields [i].offset + size - (quad * 8);
879 class1 = merge_argument_class_from_type (info->fields [i].field->type, class1);
881 g_assert (class1 != ARG_CLASS_NO_CLASS);
882 args [quad] = class1;
886 /* Post merger cleanup */
887 if ((args [0] == ARG_CLASS_MEMORY) || (args [1] == ARG_CLASS_MEMORY))
888 args [0] = args [1] = ARG_CLASS_MEMORY;
890 /* Allocate registers */
895 ainfo->storage = ArgValuetypeInReg;
896 ainfo->pair_storage [0] = ainfo->pair_storage [1] = ArgNone;
897 ainfo->nregs = nquads;
898 for (quad = 0; quad < nquads; ++quad) {
899 switch (args [quad]) {
900 case ARG_CLASS_INTEGER:
901 if (*gr >= PARAM_REGS)
902 args [quad] = ARG_CLASS_MEMORY;
904 ainfo->pair_storage [quad] = ArgInIReg;
906 ainfo->pair_regs [quad] = return_regs [*gr];
908 ainfo->pair_regs [quad] = param_regs [*gr];
913 if (*fr >= FLOAT_PARAM_REGS)
914 args [quad] = ARG_CLASS_MEMORY;
916 if (quadsize[quad] <= 4)
917 ainfo->pair_storage [quad] = ArgInFloatSSEReg;
918 else ainfo->pair_storage [quad] = ArgInDoubleSSEReg;
919 ainfo->pair_regs [quad] = *fr;
923 case ARG_CLASS_MEMORY:
926 g_assert_not_reached ();
930 if ((args [0] == ARG_CLASS_MEMORY) || (args [1] == ARG_CLASS_MEMORY)) {
931 /* Revert possible register assignments */
935 ainfo->offset = *stack_size;
937 *stack_size += ALIGN_TO (info->native_size, 8);
939 *stack_size += nquads * sizeof(mgreg_t);
940 ainfo->storage = ArgOnStack;
948 * Obtain information about a call according to the calling convention.
949 * For AMD64, see the "System V ABI, x86-64 Architecture Processor Supplement
950 * Draft Version 0.23" document for more information.
953 get_call_info (MonoGenericSharingContext *gsctx, MonoMemPool *mp, MonoMethodSignature *sig)
955 guint32 i, gr, fr, pstart;
957 int n = sig->hasthis + sig->param_count;
958 guint32 stack_size = 0;
960 gboolean is_pinvoke = sig->pinvoke;
963 cinfo = mono_mempool_alloc0 (mp, sizeof (CallInfo) + (sizeof (ArgInfo) * n));
965 cinfo = g_malloc0 (sizeof (CallInfo) + (sizeof (ArgInfo) * n));
974 ret_type = mini_type_get_underlying_type (gsctx, sig->ret);
975 switch (ret_type->type) {
976 case MONO_TYPE_BOOLEAN:
987 case MONO_TYPE_FNPTR:
988 case MONO_TYPE_CLASS:
989 case MONO_TYPE_OBJECT:
990 case MONO_TYPE_SZARRAY:
991 case MONO_TYPE_ARRAY:
992 case MONO_TYPE_STRING:
993 cinfo->ret.storage = ArgInIReg;
994 cinfo->ret.reg = AMD64_RAX;
998 cinfo->ret.storage = ArgInIReg;
999 cinfo->ret.reg = AMD64_RAX;
1002 cinfo->ret.storage = ArgInFloatSSEReg;
1003 cinfo->ret.reg = AMD64_XMM0;
1006 cinfo->ret.storage = ArgInDoubleSSEReg;
1007 cinfo->ret.reg = AMD64_XMM0;
1009 case MONO_TYPE_GENERICINST:
1010 if (!mono_type_generic_inst_is_valuetype (ret_type)) {
1011 cinfo->ret.storage = ArgInIReg;
1012 cinfo->ret.reg = AMD64_RAX;
1016 case MONO_TYPE_VALUETYPE: {
1017 guint32 tmp_gr = 0, tmp_fr = 0, tmp_stacksize = 0;
1019 add_valuetype (gsctx, sig, &cinfo->ret, sig->ret, TRUE, &tmp_gr, &tmp_fr, &tmp_stacksize);
1020 if (cinfo->ret.storage == ArgOnStack) {
1021 cinfo->vtype_retaddr = TRUE;
1022 /* The caller passes the address where the value is stored */
1026 case MONO_TYPE_TYPEDBYREF:
1027 /* Same as a valuetype with size 24 */
1028 cinfo->vtype_retaddr = TRUE;
1030 case MONO_TYPE_VOID:
1033 g_error ("Can't handle as return value 0x%x", sig->ret->type);
1039 * To simplify get_this_arg_reg () and LLVM integration, emit the vret arg after
1040 * the first argument, allowing 'this' to be always passed in the first arg reg.
1041 * Also do this if the first argument is a reference type, since virtual calls
1042 * are sometimes made using calli without sig->hasthis set, like in the delegate
1045 if (cinfo->vtype_retaddr && !is_pinvoke && (sig->hasthis || (sig->param_count > 0 && MONO_TYPE_IS_REFERENCE (mini_type_get_underlying_type (gsctx, sig->params [0]))))) {
1047 add_general (&gr, &stack_size, cinfo->args + 0);
1049 add_general (&gr, &stack_size, &cinfo->args [sig->hasthis + 0]);
1052 add_general (&gr, &stack_size, &cinfo->ret);
1053 cinfo->vret_arg_index = 1;
1057 add_general (&gr, &stack_size, cinfo->args + 0);
1059 if (cinfo->vtype_retaddr)
1060 add_general (&gr, &stack_size, &cinfo->ret);
1063 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG) && (n == 0)) {
1065 fr = FLOAT_PARAM_REGS;
1067 /* Emit the signature cookie just before the implicit arguments */
1068 add_general (&gr, &stack_size, &cinfo->sig_cookie);
1071 for (i = pstart; i < sig->param_count; ++i) {
1072 ArgInfo *ainfo = &cinfo->args [sig->hasthis + i];
1076 /* The float param registers and other param registers must be the same index on Windows x64.*/
1083 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG) && (i == sig->sentinelpos)) {
1084 /* We allways pass the sig cookie on the stack for simplicity */
1086 * Prevent implicit arguments + the sig cookie from being passed
1090 fr = FLOAT_PARAM_REGS;
1092 /* Emit the signature cookie just before the implicit arguments */
1093 add_general (&gr, &stack_size, &cinfo->sig_cookie);
1096 ptype = mini_type_get_underlying_type (gsctx, sig->params [i]);
1097 switch (ptype->type) {
1098 case MONO_TYPE_BOOLEAN:
1101 add_general (&gr, &stack_size, ainfo);
1105 case MONO_TYPE_CHAR:
1106 add_general (&gr, &stack_size, ainfo);
1110 add_general (&gr, &stack_size, ainfo);
1115 case MONO_TYPE_FNPTR:
1116 case MONO_TYPE_CLASS:
1117 case MONO_TYPE_OBJECT:
1118 case MONO_TYPE_STRING:
1119 case MONO_TYPE_SZARRAY:
1120 case MONO_TYPE_ARRAY:
1121 add_general (&gr, &stack_size, ainfo);
1123 case MONO_TYPE_GENERICINST:
1124 if (!mono_type_generic_inst_is_valuetype (ptype)) {
1125 add_general (&gr, &stack_size, ainfo);
1129 case MONO_TYPE_VALUETYPE:
1130 add_valuetype (gsctx, sig, ainfo, sig->params [i], FALSE, &gr, &fr, &stack_size);
1132 case MONO_TYPE_TYPEDBYREF:
1134 add_valuetype (gsctx, sig, ainfo, sig->params [i], FALSE, &gr, &fr, &stack_size);
1136 stack_size += sizeof (MonoTypedRef);
1137 ainfo->storage = ArgOnStack;
1142 add_general (&gr, &stack_size, ainfo);
1145 add_float (&fr, &stack_size, ainfo, FALSE);
1148 add_float (&fr, &stack_size, ainfo, TRUE);
1151 g_assert_not_reached ();
1155 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG) && (n > 0) && (sig->sentinelpos == sig->param_count)) {
1157 fr = FLOAT_PARAM_REGS;
1159 /* Emit the signature cookie just before the implicit arguments */
1160 add_general (&gr, &stack_size, &cinfo->sig_cookie);
1164 // There always is 32 bytes reserved on the stack when calling on Winx64
1168 #ifndef MONO_AMD64_NO_PUSHES
1169 if (stack_size & 0x8) {
1170 /* The AMD64 ABI requires each stack frame to be 16 byte aligned */
1171 cinfo->need_stack_align = TRUE;
1176 cinfo->stack_usage = stack_size;
1177 cinfo->reg_usage = gr;
1178 cinfo->freg_usage = fr;
1183 * mono_arch_get_argument_info:
1184 * @csig: a method signature
1185 * @param_count: the number of parameters to consider
1186 * @arg_info: an array to store the result infos
1188 * Gathers information on parameters such as size, alignment and
1189 * padding. arg_info should be large enought to hold param_count + 1 entries.
1191 * Returns the size of the argument area on the stack.
1194 mono_arch_get_argument_info (MonoMethodSignature *csig, int param_count, MonoJitArgumentInfo *arg_info)
1197 CallInfo *cinfo = get_call_info (NULL, NULL, csig);
1198 guint32 args_size = cinfo->stack_usage;
1200 /* The arguments are saved to a stack area in mono_arch_instrument_prolog */
1201 if (csig->hasthis) {
1202 arg_info [0].offset = 0;
1205 for (k = 0; k < param_count; k++) {
1206 arg_info [k + 1].offset = ((k + csig->hasthis) * 8);
1208 arg_info [k + 1].size = 0;
1217 mono_amd64_tail_call_supported (MonoMethodSignature *caller_sig, MonoMethodSignature *callee_sig)
1222 c1 = get_call_info (NULL, NULL, caller_sig);
1223 c2 = get_call_info (NULL, NULL, callee_sig);
1224 res = c1->stack_usage >= c2->stack_usage;
1225 if (callee_sig->ret && MONO_TYPE_ISSTRUCT (callee_sig->ret) && c2->ret.storage != ArgValuetypeInReg)
1226 /* An address on the callee's stack is passed as the first argument */
1236 cpuid (int id, int* p_eax, int* p_ebx, int* p_ecx, int* p_edx)
1238 #if defined(MONO_CROSS_COMPILE)
1242 __asm__ __volatile__ ("cpuid"
1243 : "=a" (*p_eax), "=b" (*p_ebx), "=c" (*p_ecx), "=d" (*p_edx)
1258 * Initialize the cpu to execute managed code.
1261 mono_arch_cpu_init (void)
1266 /* spec compliance requires running with double precision */
1267 __asm__ __volatile__ ("fnstcw %0\n": "=m" (fpcw));
1268 fpcw &= ~X86_FPCW_PRECC_MASK;
1269 fpcw |= X86_FPCW_PREC_DOUBLE;
1270 __asm__ __volatile__ ("fldcw %0\n": : "m" (fpcw));
1271 __asm__ __volatile__ ("fnstcw %0\n": "=m" (fpcw));
1273 /* TODO: This is crashing on Win64 right now.
1274 * _control87 (_PC_53, MCW_PC);
1280 * Initialize architecture specific code.
1283 mono_arch_init (void)
1287 InitializeCriticalSection (&mini_arch_mutex);
1288 #if defined(__native_client_codegen__)
1289 nacl_instruction_depth = TlsAlloc ();
1290 TlsSetValue (nacl_instruction_depth, (gpointer)0);
1291 nacl_rex_tag = TlsAlloc ();
1292 nacl_legacy_prefix_tag = TlsAlloc ();
1295 #ifdef MONO_ARCH_NOMAP32BIT
1296 flags = MONO_MMAP_READ;
1297 /* amd64_mov_reg_imm () + amd64_mov_reg_membase () */
1298 breakpoint_size = 13;
1299 breakpoint_fault_size = 3;
1300 /* amd64_alu_membase_imm_size (code, X86_CMP, AMD64_R11, 0, 0, 4); */
1301 single_step_fault_size = 5;
1303 flags = MONO_MMAP_READ|MONO_MMAP_32BIT;
1304 /* amd64_mov_reg_mem () */
1305 breakpoint_size = 8;
1306 breakpoint_fault_size = 8;
1307 single_step_fault_size = 8;
1310 ss_trigger_page = mono_valloc (NULL, mono_pagesize (), flags);
1311 bp_trigger_page = mono_valloc (NULL, mono_pagesize (), flags);
1312 mono_mprotect (bp_trigger_page, mono_pagesize (), 0);
1314 mono_aot_register_jit_icall ("mono_amd64_throw_exception", mono_amd64_throw_exception);
1315 mono_aot_register_jit_icall ("mono_amd64_throw_corlib_exception", mono_amd64_throw_corlib_exception);
1316 mono_aot_register_jit_icall ("mono_amd64_get_original_ip", mono_amd64_get_original_ip);
1320 * Cleanup architecture specific code.
1323 mono_arch_cleanup (void)
1325 DeleteCriticalSection (&mini_arch_mutex);
1326 #if defined(__native_client_codegen__)
1327 TlsFree (nacl_instruction_depth);
1328 TlsFree (nacl_rex_tag);
1329 TlsFree (nacl_legacy_prefix_tag);
1334 * This function returns the optimizations supported on this cpu.
1337 mono_arch_cpu_optimizazions (guint32 *exclude_mask)
1339 int eax, ebx, ecx, edx;
1343 /* Feature Flags function, flags returned in EDX. */
1344 if (cpuid (1, &eax, &ebx, &ecx, &edx)) {
1345 if (edx & (1 << 15)) {
1346 opts |= MONO_OPT_CMOV;
1348 opts |= MONO_OPT_FCMOV;
1350 *exclude_mask |= MONO_OPT_FCMOV;
1352 *exclude_mask |= MONO_OPT_CMOV;
1359 * This function test for all SSE functions supported.
1361 * Returns a bitmask corresponding to all supported versions.
1365 mono_arch_cpu_enumerate_simd_versions (void)
1367 int eax, ebx, ecx, edx;
1368 guint32 sse_opts = 0;
1370 if (cpuid (1, &eax, &ebx, &ecx, &edx)) {
1371 if (edx & (1 << 25))
1372 sse_opts |= SIMD_VERSION_SSE1;
1373 if (edx & (1 << 26))
1374 sse_opts |= SIMD_VERSION_SSE2;
1376 sse_opts |= SIMD_VERSION_SSE3;
1378 sse_opts |= SIMD_VERSION_SSSE3;
1379 if (ecx & (1 << 19))
1380 sse_opts |= SIMD_VERSION_SSE41;
1381 if (ecx & (1 << 20))
1382 sse_opts |= SIMD_VERSION_SSE42;
1385 /* Yes, all this needs to be done to check for sse4a.
1386 See: "Amd: CPUID Specification"
1388 if (cpuid (0x80000000, &eax, &ebx, &ecx, &edx)) {
1389 /* eax greater or equal than 0x80000001, ebx = 'htuA', ecx = DMAc', edx = 'itne'*/
1390 if ((((unsigned int) eax) >= 0x80000001) && (ebx == 0x68747541) && (ecx == 0x444D4163) && (edx == 0x69746E65)) {
1391 cpuid (0x80000001, &eax, &ebx, &ecx, &edx);
1393 sse_opts |= SIMD_VERSION_SSE4a;
1403 mono_arch_get_allocatable_int_vars (MonoCompile *cfg)
1408 for (i = 0; i < cfg->num_varinfo; i++) {
1409 MonoInst *ins = cfg->varinfo [i];
1410 MonoMethodVar *vmv = MONO_VARINFO (cfg, i);
1413 if (vmv->range.first_use.abs_pos >= vmv->range.last_use.abs_pos)
1416 if ((ins->flags & (MONO_INST_IS_DEAD|MONO_INST_VOLATILE|MONO_INST_INDIRECT)) ||
1417 (ins->opcode != OP_LOCAL && ins->opcode != OP_ARG))
1420 if (mono_is_regsize_var (ins->inst_vtype)) {
1421 g_assert (MONO_VARINFO (cfg, i)->reg == -1);
1422 g_assert (i == vmv->idx);
1423 vars = g_list_prepend (vars, vmv);
1427 vars = mono_varlist_sort (cfg, vars, 0);
1433 * mono_arch_compute_omit_fp:
1435 * Determine whenever the frame pointer can be eliminated.
1438 mono_arch_compute_omit_fp (MonoCompile *cfg)
1440 MonoMethodSignature *sig;
1441 MonoMethodHeader *header;
1445 if (cfg->arch.omit_fp_computed)
1448 header = cfg->header;
1450 sig = mono_method_signature (cfg->method);
1452 if (!cfg->arch.cinfo)
1453 cfg->arch.cinfo = get_call_info (cfg->generic_sharing_context, cfg->mempool, sig);
1454 cinfo = cfg->arch.cinfo;
1457 * FIXME: Remove some of the restrictions.
1459 cfg->arch.omit_fp = TRUE;
1460 cfg->arch.omit_fp_computed = TRUE;
1462 #ifdef __native_client_codegen__
1463 /* NaCl modules may not change the value of RBP, so it cannot be */
1464 /* used as a normal register, but it can be used as a frame pointer*/
1465 cfg->disable_omit_fp = TRUE;
1466 cfg->arch.omit_fp = FALSE;
1469 if (cfg->disable_omit_fp)
1470 cfg->arch.omit_fp = FALSE;
1472 if (!debug_omit_fp ())
1473 cfg->arch.omit_fp = FALSE;
1475 if (cfg->method->save_lmf)
1476 cfg->arch.omit_fp = FALSE;
1478 if (cfg->flags & MONO_CFG_HAS_ALLOCA)
1479 cfg->arch.omit_fp = FALSE;
1480 if (header->num_clauses)
1481 cfg->arch.omit_fp = FALSE;
1482 if (cfg->param_area)
1483 cfg->arch.omit_fp = FALSE;
1484 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG))
1485 cfg->arch.omit_fp = FALSE;
1486 if ((mono_jit_trace_calls != NULL && mono_trace_eval (cfg->method)) ||
1487 (cfg->prof_options & MONO_PROFILE_ENTER_LEAVE))
1488 cfg->arch.omit_fp = FALSE;
1489 for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
1490 ArgInfo *ainfo = &cinfo->args [i];
1492 if (ainfo->storage == ArgOnStack) {
1494 * The stack offset can only be determined when the frame
1497 cfg->arch.omit_fp = FALSE;
1502 for (i = cfg->locals_start; i < cfg->num_varinfo; i++) {
1503 MonoInst *ins = cfg->varinfo [i];
1506 locals_size += mono_type_size (ins->inst_vtype, &ialign);
1511 mono_arch_get_global_int_regs (MonoCompile *cfg)
1515 mono_arch_compute_omit_fp (cfg);
1517 if (cfg->globalra) {
1518 if (cfg->arch.omit_fp)
1519 regs = g_list_prepend (regs, (gpointer)AMD64_RBP);
1521 regs = g_list_prepend (regs, (gpointer)AMD64_RBX);
1522 regs = g_list_prepend (regs, (gpointer)AMD64_R12);
1523 regs = g_list_prepend (regs, (gpointer)AMD64_R13);
1524 regs = g_list_prepend (regs, (gpointer)AMD64_R14);
1525 #ifndef __native_client_codegen__
1526 regs = g_list_prepend (regs, (gpointer)AMD64_R15);
1529 regs = g_list_prepend (regs, (gpointer)AMD64_R10);
1530 regs = g_list_prepend (regs, (gpointer)AMD64_R9);
1531 regs = g_list_prepend (regs, (gpointer)AMD64_R8);
1532 regs = g_list_prepend (regs, (gpointer)AMD64_RDI);
1533 regs = g_list_prepend (regs, (gpointer)AMD64_RSI);
1534 regs = g_list_prepend (regs, (gpointer)AMD64_RDX);
1535 regs = g_list_prepend (regs, (gpointer)AMD64_RCX);
1536 regs = g_list_prepend (regs, (gpointer)AMD64_RAX);
1538 if (cfg->arch.omit_fp)
1539 regs = g_list_prepend (regs, (gpointer)AMD64_RBP);
1541 /* We use the callee saved registers for global allocation */
1542 regs = g_list_prepend (regs, (gpointer)AMD64_RBX);
1543 regs = g_list_prepend (regs, (gpointer)AMD64_R12);
1544 regs = g_list_prepend (regs, (gpointer)AMD64_R13);
1545 regs = g_list_prepend (regs, (gpointer)AMD64_R14);
1546 #ifndef __native_client_codegen__
1547 regs = g_list_prepend (regs, (gpointer)AMD64_R15);
1550 regs = g_list_prepend (regs, (gpointer)AMD64_RDI);
1551 regs = g_list_prepend (regs, (gpointer)AMD64_RSI);
1559 mono_arch_get_global_fp_regs (MonoCompile *cfg)
1564 /* All XMM registers */
1565 for (i = 0; i < 16; ++i)
1566 regs = g_list_prepend (regs, GINT_TO_POINTER (i));
1572 mono_arch_get_iregs_clobbered_by_call (MonoCallInst *call)
1574 static GList *r = NULL;
1579 regs = g_list_prepend (regs, (gpointer)AMD64_RBP);
1580 regs = g_list_prepend (regs, (gpointer)AMD64_RBX);
1581 regs = g_list_prepend (regs, (gpointer)AMD64_R12);
1582 regs = g_list_prepend (regs, (gpointer)AMD64_R13);
1583 regs = g_list_prepend (regs, (gpointer)AMD64_R14);
1584 #ifndef __native_client_codegen__
1585 regs = g_list_prepend (regs, (gpointer)AMD64_R15);
1588 regs = g_list_prepend (regs, (gpointer)AMD64_R10);
1589 regs = g_list_prepend (regs, (gpointer)AMD64_R9);
1590 regs = g_list_prepend (regs, (gpointer)AMD64_R8);
1591 regs = g_list_prepend (regs, (gpointer)AMD64_RDI);
1592 regs = g_list_prepend (regs, (gpointer)AMD64_RSI);
1593 regs = g_list_prepend (regs, (gpointer)AMD64_RDX);
1594 regs = g_list_prepend (regs, (gpointer)AMD64_RCX);
1595 regs = g_list_prepend (regs, (gpointer)AMD64_RAX);
1597 InterlockedCompareExchangePointer ((gpointer*)&r, regs, NULL);
1604 mono_arch_get_fregs_clobbered_by_call (MonoCallInst *call)
1607 static GList *r = NULL;
1612 for (i = 0; i < AMD64_XMM_NREG; ++i)
1613 regs = g_list_prepend (regs, GINT_TO_POINTER (MONO_MAX_IREGS + i));
1615 InterlockedCompareExchangePointer ((gpointer*)&r, regs, NULL);
1622 * mono_arch_regalloc_cost:
1624 * Return the cost, in number of memory references, of the action of
1625 * allocating the variable VMV into a register during global register
1629 mono_arch_regalloc_cost (MonoCompile *cfg, MonoMethodVar *vmv)
1631 MonoInst *ins = cfg->varinfo [vmv->idx];
1633 if (cfg->method->save_lmf)
1634 /* The register is already saved */
1635 /* substract 1 for the invisible store in the prolog */
1636 return (ins->opcode == OP_ARG) ? 0 : 1;
1639 return (ins->opcode == OP_ARG) ? 1 : 2;
1643 * mono_arch_fill_argument_info:
1645 * Populate cfg->args, cfg->ret and cfg->vret_addr with information about the arguments
1649 mono_arch_fill_argument_info (MonoCompile *cfg)
1651 MonoMethodSignature *sig;
1652 MonoMethodHeader *header;
1657 header = cfg->header;
1659 sig = mono_method_signature (cfg->method);
1661 cinfo = cfg->arch.cinfo;
1664 * Contrary to mono_arch_allocate_vars (), the information should describe
1665 * where the arguments are at the beginning of the method, not where they can be
1666 * accessed during the execution of the method. The later makes no sense for the
1667 * global register allocator, since a variable can be in more than one location.
1669 if (sig->ret->type != MONO_TYPE_VOID) {
1670 switch (cinfo->ret.storage) {
1672 case ArgInFloatSSEReg:
1673 case ArgInDoubleSSEReg:
1674 if ((MONO_TYPE_ISSTRUCT (sig->ret) && !mono_class_from_mono_type (sig->ret)->enumtype) || (sig->ret->type == MONO_TYPE_TYPEDBYREF)) {
1675 cfg->vret_addr->opcode = OP_REGVAR;
1676 cfg->vret_addr->inst_c0 = cinfo->ret.reg;
1679 cfg->ret->opcode = OP_REGVAR;
1680 cfg->ret->inst_c0 = cinfo->ret.reg;
1683 case ArgValuetypeInReg:
1684 cfg->ret->opcode = OP_REGOFFSET;
1685 cfg->ret->inst_basereg = -1;
1686 cfg->ret->inst_offset = -1;
1689 g_assert_not_reached ();
1693 for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
1694 ArgInfo *ainfo = &cinfo->args [i];
1697 ins = cfg->args [i];
1699 if (sig->hasthis && (i == 0))
1700 arg_type = &mono_defaults.object_class->byval_arg;
1702 arg_type = sig->params [i - sig->hasthis];
1704 switch (ainfo->storage) {
1706 case ArgInFloatSSEReg:
1707 case ArgInDoubleSSEReg:
1708 ins->opcode = OP_REGVAR;
1709 ins->inst_c0 = ainfo->reg;
1712 ins->opcode = OP_REGOFFSET;
1713 ins->inst_basereg = -1;
1714 ins->inst_offset = -1;
1716 case ArgValuetypeInReg:
1718 ins->opcode = OP_NOP;
1721 g_assert_not_reached ();
1727 mono_arch_allocate_vars (MonoCompile *cfg)
1729 MonoMethodSignature *sig;
1730 MonoMethodHeader *header;
1733 guint32 locals_stack_size, locals_stack_align;
1737 header = cfg->header;
1739 sig = mono_method_signature (cfg->method);
1741 cinfo = cfg->arch.cinfo;
1743 mono_arch_compute_omit_fp (cfg);
1746 * We use the ABI calling conventions for managed code as well.
1747 * Exception: valuetypes are only sometimes passed or returned in registers.
1751 * The stack looks like this:
1752 * <incoming arguments passed on the stack>
1754 * <lmf/caller saved registers>
1757 * <localloc area> -> grows dynamically
1761 if (cfg->arch.omit_fp) {
1762 cfg->flags |= MONO_CFG_HAS_SPILLUP;
1763 cfg->frame_reg = AMD64_RSP;
1766 /* Locals are allocated backwards from %fp */
1767 cfg->frame_reg = AMD64_RBP;
1771 if (cfg->method->save_lmf) {
1772 /* Reserve stack space for saving LMF */
1773 if (cfg->arch.omit_fp) {
1774 cfg->arch.lmf_offset = offset;
1775 offset += sizeof (MonoLMF);
1778 offset += sizeof (MonoLMF);
1779 cfg->arch.lmf_offset = -offset;
1782 if (cfg->arch.omit_fp)
1783 cfg->arch.reg_save_area_offset = offset;
1784 /* Reserve space for caller saved registers */
1785 for (i = 0; i < AMD64_NREG; ++i)
1786 if (AMD64_IS_CALLEE_SAVED_REG (i) && (cfg->used_int_regs & (1 << i))) {
1787 offset += sizeof(mgreg_t);
1791 if (sig->ret->type != MONO_TYPE_VOID) {
1792 switch (cinfo->ret.storage) {
1794 case ArgInFloatSSEReg:
1795 case ArgInDoubleSSEReg:
1796 if ((MONO_TYPE_ISSTRUCT (sig->ret) && !mono_class_from_mono_type (sig->ret)->enumtype) || (sig->ret->type == MONO_TYPE_TYPEDBYREF)) {
1797 if (cfg->globalra) {
1798 cfg->vret_addr->opcode = OP_REGVAR;
1799 cfg->vret_addr->inst_c0 = cinfo->ret.reg;
1801 /* The register is volatile */
1802 cfg->vret_addr->opcode = OP_REGOFFSET;
1803 cfg->vret_addr->inst_basereg = cfg->frame_reg;
1804 if (cfg->arch.omit_fp) {
1805 cfg->vret_addr->inst_offset = offset;
1809 cfg->vret_addr->inst_offset = -offset;
1811 if (G_UNLIKELY (cfg->verbose_level > 1)) {
1812 printf ("vret_addr =");
1813 mono_print_ins (cfg->vret_addr);
1818 cfg->ret->opcode = OP_REGVAR;
1819 cfg->ret->inst_c0 = cinfo->ret.reg;
1822 case ArgValuetypeInReg:
1823 /* Allocate a local to hold the result, the epilog will copy it to the correct place */
1824 cfg->ret->opcode = OP_REGOFFSET;
1825 cfg->ret->inst_basereg = cfg->frame_reg;
1826 if (cfg->arch.omit_fp) {
1827 cfg->ret->inst_offset = offset;
1831 cfg->ret->inst_offset = - offset;
1835 g_assert_not_reached ();
1838 cfg->ret->dreg = cfg->ret->inst_c0;
1841 /* Allocate locals */
1842 if (!cfg->globalra) {
1843 offsets = mono_allocate_stack_slots_full (cfg, cfg->arch.omit_fp ? FALSE: TRUE, &locals_stack_size, &locals_stack_align);
1844 if (locals_stack_size > MONO_ARCH_MAX_FRAME_SIZE) {
1845 char *mname = mono_method_full_name (cfg->method, TRUE);
1846 cfg->exception_type = MONO_EXCEPTION_INVALID_PROGRAM;
1847 cfg->exception_message = g_strdup_printf ("Method %s stack is too big.", mname);
1852 if (locals_stack_align) {
1853 offset += (locals_stack_align - 1);
1854 offset &= ~(locals_stack_align - 1);
1856 if (cfg->arch.omit_fp) {
1857 cfg->locals_min_stack_offset = offset;
1858 cfg->locals_max_stack_offset = offset + locals_stack_size;
1860 cfg->locals_min_stack_offset = - (offset + locals_stack_size);
1861 cfg->locals_max_stack_offset = - offset;
1864 for (i = cfg->locals_start; i < cfg->num_varinfo; i++) {
1865 if (offsets [i] != -1) {
1866 MonoInst *ins = cfg->varinfo [i];
1867 ins->opcode = OP_REGOFFSET;
1868 ins->inst_basereg = cfg->frame_reg;
1869 if (cfg->arch.omit_fp)
1870 ins->inst_offset = (offset + offsets [i]);
1872 ins->inst_offset = - (offset + offsets [i]);
1873 //printf ("allocated local %d to ", i); mono_print_tree_nl (ins);
1876 offset += locals_stack_size;
1879 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG)) {
1880 g_assert (!cfg->arch.omit_fp);
1881 g_assert (cinfo->sig_cookie.storage == ArgOnStack);
1882 cfg->sig_cookie = cinfo->sig_cookie.offset + ARGS_OFFSET;
1885 for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
1886 ins = cfg->args [i];
1887 if (ins->opcode != OP_REGVAR) {
1888 ArgInfo *ainfo = &cinfo->args [i];
1889 gboolean inreg = TRUE;
1892 if (sig->hasthis && (i == 0))
1893 arg_type = &mono_defaults.object_class->byval_arg;
1895 arg_type = sig->params [i - sig->hasthis];
1897 if (cfg->globalra) {
1898 /* The new allocator needs info about the original locations of the arguments */
1899 switch (ainfo->storage) {
1901 case ArgInFloatSSEReg:
1902 case ArgInDoubleSSEReg:
1903 ins->opcode = OP_REGVAR;
1904 ins->inst_c0 = ainfo->reg;
1907 g_assert (!cfg->arch.omit_fp);
1908 ins->opcode = OP_REGOFFSET;
1909 ins->inst_basereg = cfg->frame_reg;
1910 ins->inst_offset = ainfo->offset + ARGS_OFFSET;
1912 case ArgValuetypeInReg:
1913 ins->opcode = OP_REGOFFSET;
1914 ins->inst_basereg = cfg->frame_reg;
1915 /* These arguments are saved to the stack in the prolog */
1916 offset = ALIGN_TO (offset, sizeof(mgreg_t));
1917 if (cfg->arch.omit_fp) {
1918 ins->inst_offset = offset;
1919 offset += (ainfo->storage == ArgValuetypeInReg) ? ainfo->nregs * sizeof (mgreg_t) : sizeof (mgreg_t);
1921 offset += (ainfo->storage == ArgValuetypeInReg) ? ainfo->nregs * sizeof (mgreg_t) : sizeof (mgreg_t);
1922 ins->inst_offset = - offset;
1926 g_assert_not_reached ();
1932 /* FIXME: Allocate volatile arguments to registers */
1933 if (ins->flags & (MONO_INST_VOLATILE|MONO_INST_INDIRECT))
1937 * Under AMD64, all registers used to pass arguments to functions
1938 * are volatile across calls.
1939 * FIXME: Optimize this.
1941 if ((ainfo->storage == ArgInIReg) || (ainfo->storage == ArgInFloatSSEReg) || (ainfo->storage == ArgInDoubleSSEReg) || (ainfo->storage == ArgValuetypeInReg))
1944 ins->opcode = OP_REGOFFSET;
1946 switch (ainfo->storage) {
1948 case ArgInFloatSSEReg:
1949 case ArgInDoubleSSEReg:
1951 ins->opcode = OP_REGVAR;
1952 ins->dreg = ainfo->reg;
1956 g_assert (!cfg->arch.omit_fp);
1957 ins->opcode = OP_REGOFFSET;
1958 ins->inst_basereg = cfg->frame_reg;
1959 ins->inst_offset = ainfo->offset + ARGS_OFFSET;
1961 case ArgValuetypeInReg:
1963 case ArgValuetypeAddrInIReg: {
1965 g_assert (!cfg->arch.omit_fp);
1967 MONO_INST_NEW (cfg, indir, 0);
1968 indir->opcode = OP_REGOFFSET;
1969 if (ainfo->pair_storage [0] == ArgInIReg) {
1970 indir->inst_basereg = cfg->frame_reg;
1971 offset = ALIGN_TO (offset, sizeof (gpointer));
1972 offset += (sizeof (gpointer));
1973 indir->inst_offset = - offset;
1976 indir->inst_basereg = cfg->frame_reg;
1977 indir->inst_offset = ainfo->offset + ARGS_OFFSET;
1980 ins->opcode = OP_VTARG_ADDR;
1981 ins->inst_left = indir;
1989 if (!inreg && (ainfo->storage != ArgOnStack) && (ainfo->storage != ArgValuetypeAddrInIReg)) {
1990 ins->opcode = OP_REGOFFSET;
1991 ins->inst_basereg = cfg->frame_reg;
1992 /* These arguments are saved to the stack in the prolog */
1993 offset = ALIGN_TO (offset, sizeof(mgreg_t));
1994 if (cfg->arch.omit_fp) {
1995 ins->inst_offset = offset;
1996 offset += (ainfo->storage == ArgValuetypeInReg) ? ainfo->nregs * sizeof (mgreg_t) : sizeof (mgreg_t);
1997 // Arguments are yet supported by the stack map creation code
1998 //cfg->locals_max_stack_offset = MAX (cfg->locals_max_stack_offset, offset);
2000 offset += (ainfo->storage == ArgValuetypeInReg) ? ainfo->nregs * sizeof (mgreg_t) : sizeof (mgreg_t);
2001 ins->inst_offset = - offset;
2002 //cfg->locals_min_stack_offset = MIN (cfg->locals_min_stack_offset, offset);
2008 cfg->stack_offset = offset;
2012 mono_arch_create_vars (MonoCompile *cfg)
2014 MonoMethodSignature *sig;
2017 sig = mono_method_signature (cfg->method);
2019 if (!cfg->arch.cinfo)
2020 cfg->arch.cinfo = get_call_info (cfg->generic_sharing_context, cfg->mempool, sig);
2021 cinfo = cfg->arch.cinfo;
2023 if (cinfo->ret.storage == ArgValuetypeInReg)
2024 cfg->ret_var_is_local = TRUE;
2026 if ((cinfo->ret.storage != ArgValuetypeInReg) && MONO_TYPE_ISSTRUCT (sig->ret)) {
2027 cfg->vret_addr = mono_compile_create_var (cfg, &mono_defaults.int_class->byval_arg, OP_ARG);
2028 if (G_UNLIKELY (cfg->verbose_level > 1)) {
2029 printf ("vret_addr = ");
2030 mono_print_ins (cfg->vret_addr);
2034 if (cfg->gen_seq_points) {
2037 ins = mono_compile_create_var (cfg, &mono_defaults.int_class->byval_arg, OP_LOCAL);
2038 ins->flags |= MONO_INST_VOLATILE;
2039 cfg->arch.ss_trigger_page_var = ins;
2042 #ifdef MONO_AMD64_NO_PUSHES
2044 * When this is set, we pass arguments on the stack by moves, and by allocating
2045 * a bigger stack frame, instead of pushes.
2046 * Pushes complicate exception handling because the arguments on the stack have
2047 * to be popped each time a frame is unwound. They also make fp elimination
2049 * FIXME: This doesn't work inside filter/finally clauses, since those execute
2050 * on a new frame which doesn't include a param area.
2052 cfg->arch.no_pushes = TRUE;
2057 add_outarg_reg (MonoCompile *cfg, MonoCallInst *call, ArgStorage storage, int reg, MonoInst *tree)
2063 MONO_INST_NEW (cfg, ins, OP_MOVE);
2064 ins->dreg = mono_alloc_ireg_copy (cfg, tree->dreg);
2065 ins->sreg1 = tree->dreg;
2066 MONO_ADD_INS (cfg->cbb, ins);
2067 mono_call_inst_add_outarg_reg (cfg, call, ins->dreg, reg, FALSE);
2069 case ArgInFloatSSEReg:
2070 MONO_INST_NEW (cfg, ins, OP_AMD64_SET_XMMREG_R4);
2071 ins->dreg = mono_alloc_freg (cfg);
2072 ins->sreg1 = tree->dreg;
2073 MONO_ADD_INS (cfg->cbb, ins);
2075 mono_call_inst_add_outarg_reg (cfg, call, ins->dreg, reg, TRUE);
2077 case ArgInDoubleSSEReg:
2078 MONO_INST_NEW (cfg, ins, OP_FMOVE);
2079 ins->dreg = mono_alloc_freg (cfg);
2080 ins->sreg1 = tree->dreg;
2081 MONO_ADD_INS (cfg->cbb, ins);
2083 mono_call_inst_add_outarg_reg (cfg, call, ins->dreg, reg, TRUE);
2087 g_assert_not_reached ();
2092 arg_storage_to_load_membase (ArgStorage storage)
2096 #if defined(__mono_ilp32__)
2097 return OP_LOADI8_MEMBASE;
2099 return OP_LOAD_MEMBASE;
2101 case ArgInDoubleSSEReg:
2102 return OP_LOADR8_MEMBASE;
2103 case ArgInFloatSSEReg:
2104 return OP_LOADR4_MEMBASE;
2106 g_assert_not_reached ();
2113 emit_sig_cookie (MonoCompile *cfg, MonoCallInst *call, CallInfo *cinfo)
2116 MonoMethodSignature *tmp_sig;
2119 if (call->tail_call)
2122 /* FIXME: Add support for signature tokens to AOT */
2123 cfg->disable_aot = TRUE;
2125 g_assert (cinfo->sig_cookie.storage == ArgOnStack);
2128 * mono_ArgIterator_Setup assumes the signature cookie is
2129 * passed first and all the arguments which were before it are
2130 * passed on the stack after the signature. So compensate by
2131 * passing a different signature.
2133 tmp_sig = mono_metadata_signature_dup_full (cfg->method->klass->image, call->signature);
2134 tmp_sig->param_count -= call->signature->sentinelpos;
2135 tmp_sig->sentinelpos = 0;
2136 memcpy (tmp_sig->params, call->signature->params + call->signature->sentinelpos, tmp_sig->param_count * sizeof (MonoType*));
2138 MONO_INST_NEW (cfg, sig_arg, OP_ICONST);
2139 sig_arg->dreg = mono_alloc_ireg (cfg);
2140 sig_arg->inst_p0 = tmp_sig;
2141 MONO_ADD_INS (cfg->cbb, sig_arg);
2143 if (cfg->arch.no_pushes) {
2144 MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STORE_MEMBASE_REG, AMD64_RSP, cinfo->sig_cookie.offset, sig_arg->dreg);
2146 MONO_INST_NEW (cfg, arg, OP_X86_PUSH);
2147 arg->sreg1 = sig_arg->dreg;
2148 MONO_ADD_INS (cfg->cbb, arg);
2152 static inline LLVMArgStorage
2153 arg_storage_to_llvm_arg_storage (MonoCompile *cfg, ArgStorage storage)
2157 return LLVMArgInIReg;
2161 g_assert_not_reached ();
2168 mono_arch_get_llvm_call_info (MonoCompile *cfg, MonoMethodSignature *sig)
2174 LLVMCallInfo *linfo;
2177 n = sig->param_count + sig->hasthis;
2179 cinfo = get_call_info (cfg->generic_sharing_context, cfg->mempool, sig);
2181 linfo = mono_mempool_alloc0 (cfg->mempool, sizeof (LLVMCallInfo) + (sizeof (LLVMArgInfo) * n));
2184 * LLVM always uses the native ABI while we use our own ABI, the
2185 * only difference is the handling of vtypes:
2186 * - we only pass/receive them in registers in some cases, and only
2187 * in 1 or 2 integer registers.
2189 if (cinfo->ret.storage == ArgValuetypeInReg) {
2191 cfg->exception_message = g_strdup ("pinvoke + vtypes");
2192 cfg->disable_llvm = TRUE;
2196 linfo->ret.storage = LLVMArgVtypeInReg;
2197 for (j = 0; j < 2; ++j)
2198 linfo->ret.pair_storage [j] = arg_storage_to_llvm_arg_storage (cfg, cinfo->ret.pair_storage [j]);
2201 if (MONO_TYPE_ISSTRUCT (sig->ret) && cinfo->ret.storage == ArgInIReg) {
2202 /* Vtype returned using a hidden argument */
2203 linfo->ret.storage = LLVMArgVtypeRetAddr;
2204 linfo->vret_arg_index = cinfo->vret_arg_index;
2207 for (i = 0; i < n; ++i) {
2208 ainfo = cinfo->args + i;
2210 if (i >= sig->hasthis)
2211 t = sig->params [i - sig->hasthis];
2213 t = &mono_defaults.int_class->byval_arg;
2215 linfo->args [i].storage = LLVMArgNone;
2217 switch (ainfo->storage) {
2219 linfo->args [i].storage = LLVMArgInIReg;
2221 case ArgInDoubleSSEReg:
2222 case ArgInFloatSSEReg:
2223 linfo->args [i].storage = LLVMArgInFPReg;
2226 if (MONO_TYPE_ISSTRUCT (t)) {
2227 linfo->args [i].storage = LLVMArgVtypeByVal;
2229 linfo->args [i].storage = LLVMArgInIReg;
2231 if (t->type == MONO_TYPE_R4)
2232 linfo->args [i].storage = LLVMArgInFPReg;
2233 else if (t->type == MONO_TYPE_R8)
2234 linfo->args [i].storage = LLVMArgInFPReg;
2238 case ArgValuetypeInReg:
2240 cfg->exception_message = g_strdup ("pinvoke + vtypes");
2241 cfg->disable_llvm = TRUE;
2245 linfo->args [i].storage = LLVMArgVtypeInReg;
2246 for (j = 0; j < 2; ++j)
2247 linfo->args [i].pair_storage [j] = arg_storage_to_llvm_arg_storage (cfg, ainfo->pair_storage [j]);
2250 cfg->exception_message = g_strdup ("ainfo->storage");
2251 cfg->disable_llvm = TRUE;
2261 mono_arch_emit_call (MonoCompile *cfg, MonoCallInst *call)
2264 MonoMethodSignature *sig;
2265 int i, n, stack_size;
2271 sig = call->signature;
2272 n = sig->param_count + sig->hasthis;
2274 cinfo = get_call_info (cfg->generic_sharing_context, cfg->mempool, sig);
2276 if (COMPILE_LLVM (cfg)) {
2277 /* We shouldn't be called in the llvm case */
2278 cfg->disable_llvm = TRUE;
2282 if (cinfo->need_stack_align) {
2283 if (!cfg->arch.no_pushes)
2284 MONO_EMIT_NEW_BIALU_IMM (cfg, OP_SUB_IMM, X86_ESP, X86_ESP, 8);
2288 * Emit all arguments which are passed on the stack to prevent register
2289 * allocation problems.
2291 if (cfg->arch.no_pushes) {
2292 for (i = 0; i < n; ++i) {
2294 ainfo = cinfo->args + i;
2296 in = call->args [i];
2298 if (sig->hasthis && i == 0)
2299 t = &mono_defaults.object_class->byval_arg;
2301 t = sig->params [i - sig->hasthis];
2303 if (ainfo->storage == ArgOnStack && !MONO_TYPE_ISSTRUCT (t) && !call->tail_call) {
2305 if (t->type == MONO_TYPE_R4)
2306 MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STORER4_MEMBASE_REG, AMD64_RSP, ainfo->offset, in->dreg);
2307 else if (t->type == MONO_TYPE_R8)
2308 MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STORER8_MEMBASE_REG, AMD64_RSP, ainfo->offset, in->dreg);
2310 MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STORE_MEMBASE_REG, AMD64_RSP, ainfo->offset, in->dreg);
2312 MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STORE_MEMBASE_REG, AMD64_RSP, ainfo->offset, in->dreg);
2314 if (cfg->compute_gc_maps) {
2317 EMIT_NEW_GC_PARAM_SLOT_LIVENESS_DEF (cfg, def, ainfo->offset, t);
2324 * Emit all parameters passed in registers in non-reverse order for better readability
2325 * and to help the optimization in emit_prolog ().
2327 for (i = 0; i < n; ++i) {
2328 ainfo = cinfo->args + i;
2330 in = call->args [i];
2332 if (ainfo->storage == ArgInIReg)
2333 add_outarg_reg (cfg, call, ainfo->storage, ainfo->reg, in);
2336 for (i = n - 1; i >= 0; --i) {
2337 ainfo = cinfo->args + i;
2339 in = call->args [i];
2341 switch (ainfo->storage) {
2345 case ArgInFloatSSEReg:
2346 case ArgInDoubleSSEReg:
2347 add_outarg_reg (cfg, call, ainfo->storage, ainfo->reg, in);
2350 case ArgValuetypeInReg:
2351 case ArgValuetypeAddrInIReg:
2352 if (ainfo->storage == ArgOnStack && call->tail_call) {
2353 MonoInst *call_inst = (MonoInst*)call;
2354 cfg->args [i]->flags |= MONO_INST_VOLATILE;
2355 EMIT_NEW_ARGSTORE (cfg, call_inst, i, in);
2356 } else if ((i >= sig->hasthis) && (MONO_TYPE_ISSTRUCT(sig->params [i - sig->hasthis]))) {
2360 if (sig->params [i - sig->hasthis]->type == MONO_TYPE_TYPEDBYREF) {
2361 size = sizeof (MonoTypedRef);
2362 align = sizeof (gpointer);
2366 size = mono_type_native_stack_size (&in->klass->byval_arg, &align);
2369 * Other backends use mono_type_stack_size (), but that
2370 * aligns the size to 8, which is larger than the size of
2371 * the source, leading to reads of invalid memory if the
2372 * source is at the end of address space.
2374 size = mono_class_value_size (in->klass, &align);
2377 g_assert (in->klass);
2379 if (ainfo->storage == ArgOnStack && size >= 10000) {
2380 /* Avoid asserts in emit_memcpy () */
2381 cfg->exception_type = MONO_EXCEPTION_INVALID_PROGRAM;
2382 cfg->exception_message = g_strdup_printf ("Passing an argument of size '%d'.", size);
2383 /* Continue normally */
2387 MONO_INST_NEW (cfg, arg, OP_OUTARG_VT);
2388 arg->sreg1 = in->dreg;
2389 arg->klass = in->klass;
2390 arg->backend.size = size;
2391 arg->inst_p0 = call;
2392 arg->inst_p1 = mono_mempool_alloc (cfg->mempool, sizeof (ArgInfo));
2393 memcpy (arg->inst_p1, ainfo, sizeof (ArgInfo));
2395 MONO_ADD_INS (cfg->cbb, arg);
2398 if (cfg->arch.no_pushes) {
2401 MONO_INST_NEW (cfg, arg, OP_X86_PUSH);
2402 arg->sreg1 = in->dreg;
2403 if (!sig->params [i - sig->hasthis]->byref) {
2404 if (sig->params [i - sig->hasthis]->type == MONO_TYPE_R4) {
2405 MONO_EMIT_NEW_BIALU_IMM (cfg, OP_SUB_IMM, X86_ESP, X86_ESP, 8);
2406 arg->opcode = OP_STORER4_MEMBASE_REG;
2407 arg->inst_destbasereg = X86_ESP;
2408 arg->inst_offset = 0;
2409 } else if (sig->params [i - sig->hasthis]->type == MONO_TYPE_R8) {
2410 MONO_EMIT_NEW_BIALU_IMM (cfg, OP_SUB_IMM, X86_ESP, X86_ESP, 8);
2411 arg->opcode = OP_STORER8_MEMBASE_REG;
2412 arg->inst_destbasereg = X86_ESP;
2413 arg->inst_offset = 0;
2416 MONO_ADD_INS (cfg->cbb, arg);
2421 g_assert_not_reached ();
2424 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG) && (i == sig->sentinelpos))
2425 /* Emit the signature cookie just before the implicit arguments */
2426 emit_sig_cookie (cfg, call, cinfo);
2429 /* Handle the case where there are no implicit arguments */
2430 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG) && (n == sig->sentinelpos))
2431 emit_sig_cookie (cfg, call, cinfo);
2433 if (sig->ret && MONO_TYPE_ISSTRUCT (sig->ret)) {
2436 if (cinfo->ret.storage == ArgValuetypeInReg) {
2437 if (cinfo->ret.pair_storage [0] == ArgInIReg && cinfo->ret.pair_storage [1] == ArgNone) {
2439 * Tell the JIT to use a more efficient calling convention: call using
2440 * OP_CALL, compute the result location after the call, and save the
2443 call->vret_in_reg = TRUE;
2445 * Nullify the instruction computing the vret addr to enable
2446 * future optimizations.
2449 NULLIFY_INS (call->vret_var);
2451 if (call->tail_call)
2454 * The valuetype is in RAX:RDX after the call, need to be copied to
2455 * the stack. Push the address here, so the call instruction can
2458 if (!cfg->arch.vret_addr_loc) {
2459 cfg->arch.vret_addr_loc = mono_compile_create_var (cfg, &mono_defaults.int_class->byval_arg, OP_LOCAL);
2460 /* Prevent it from being register allocated or optimized away */
2461 ((MonoInst*)cfg->arch.vret_addr_loc)->flags |= MONO_INST_VOLATILE;
2464 MONO_EMIT_NEW_UNALU (cfg, OP_MOVE, ((MonoInst*)cfg->arch.vret_addr_loc)->dreg, call->vret_var->dreg);
2468 MONO_INST_NEW (cfg, vtarg, OP_MOVE);
2469 vtarg->sreg1 = call->vret_var->dreg;
2470 vtarg->dreg = mono_alloc_preg (cfg);
2471 MONO_ADD_INS (cfg->cbb, vtarg);
2473 mono_call_inst_add_outarg_reg (cfg, call, vtarg->dreg, cinfo->ret.reg, FALSE);
2478 if (call->inst.opcode != OP_JMP && OP_TAILCALL != call->inst.opcode) {
2479 MONO_EMIT_NEW_BIALU_IMM (cfg, OP_SUB_IMM, X86_ESP, X86_ESP, 0x20);
2483 if (cfg->method->save_lmf) {
2484 MONO_INST_NEW (cfg, arg, OP_AMD64_SAVE_SP_TO_LMF);
2485 MONO_ADD_INS (cfg->cbb, arg);
2488 call->stack_usage = cinfo->stack_usage;
2492 mono_arch_emit_outarg_vt (MonoCompile *cfg, MonoInst *ins, MonoInst *src)
2495 MonoCallInst *call = (MonoCallInst*)ins->inst_p0;
2496 ArgInfo *ainfo = (ArgInfo*)ins->inst_p1;
2497 int size = ins->backend.size;
2499 if (ainfo->storage == ArgValuetypeInReg) {
2503 for (part = 0; part < 2; ++part) {
2504 if (ainfo->pair_storage [part] == ArgNone)
2507 MONO_INST_NEW (cfg, load, arg_storage_to_load_membase (ainfo->pair_storage [part]));
2508 load->inst_basereg = src->dreg;
2509 load->inst_offset = part * sizeof(mgreg_t);
2511 switch (ainfo->pair_storage [part]) {
2513 load->dreg = mono_alloc_ireg (cfg);
2515 case ArgInDoubleSSEReg:
2516 case ArgInFloatSSEReg:
2517 load->dreg = mono_alloc_freg (cfg);
2520 g_assert_not_reached ();
2522 MONO_ADD_INS (cfg->cbb, load);
2524 add_outarg_reg (cfg, call, ainfo->pair_storage [part], ainfo->pair_regs [part], load);
2526 } else if (ainfo->storage == ArgValuetypeAddrInIReg) {
2527 MonoInst *vtaddr, *load;
2528 vtaddr = mono_compile_create_var (cfg, &ins->klass->byval_arg, OP_LOCAL);
2530 g_assert (!cfg->arch.no_pushes);
2532 MONO_INST_NEW (cfg, load, OP_LDADDR);
2533 load->inst_p0 = vtaddr;
2534 vtaddr->flags |= MONO_INST_INDIRECT;
2535 load->type = STACK_MP;
2536 load->klass = vtaddr->klass;
2537 load->dreg = mono_alloc_ireg (cfg);
2538 MONO_ADD_INS (cfg->cbb, load);
2539 mini_emit_memcpy (cfg, load->dreg, 0, src->dreg, 0, size, 4);
2541 if (ainfo->pair_storage [0] == ArgInIReg) {
2542 MONO_INST_NEW (cfg, arg, OP_X86_LEA_MEMBASE);
2543 arg->dreg = mono_alloc_ireg (cfg);
2544 arg->sreg1 = load->dreg;
2546 MONO_ADD_INS (cfg->cbb, arg);
2547 mono_call_inst_add_outarg_reg (cfg, call, arg->dreg, ainfo->pair_regs [0], FALSE);
2549 MONO_INST_NEW (cfg, arg, OP_X86_PUSH);
2550 arg->sreg1 = load->dreg;
2551 MONO_ADD_INS (cfg->cbb, arg);
2555 if (cfg->arch.no_pushes) {
2556 int dreg = mono_alloc_ireg (cfg);
2558 MONO_EMIT_NEW_LOAD_MEMBASE (cfg, dreg, src->dreg, 0);
2559 MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STORE_MEMBASE_REG, AMD64_RSP, ainfo->offset, dreg);
2561 /* Can't use this for < 8 since it does an 8 byte memory load */
2562 MONO_INST_NEW (cfg, arg, OP_X86_PUSH_MEMBASE);
2563 arg->inst_basereg = src->dreg;
2564 arg->inst_offset = 0;
2565 MONO_ADD_INS (cfg->cbb, arg);
2567 } else if (size <= 40) {
2568 if (cfg->arch.no_pushes) {
2569 mini_emit_memcpy (cfg, AMD64_RSP, ainfo->offset, src->dreg, 0, size, 4);
2571 MONO_EMIT_NEW_BIALU_IMM (cfg, OP_SUB_IMM, X86_ESP, X86_ESP, ALIGN_TO (size, 8));
2572 mini_emit_memcpy (cfg, X86_ESP, 0, src->dreg, 0, size, 4);
2575 if (cfg->arch.no_pushes) {
2576 // FIXME: Code growth
2577 mini_emit_memcpy (cfg, AMD64_RSP, ainfo->offset, src->dreg, 0, size, 4);
2579 MONO_INST_NEW (cfg, arg, OP_X86_PUSH_OBJ);
2580 arg->inst_basereg = src->dreg;
2581 arg->inst_offset = 0;
2582 arg->inst_imm = size;
2583 MONO_ADD_INS (cfg->cbb, arg);
2587 if (cfg->compute_gc_maps) {
2589 EMIT_NEW_GC_PARAM_SLOT_LIVENESS_DEF (cfg, def, ainfo->offset, &ins->klass->byval_arg);
2595 mono_arch_emit_setret (MonoCompile *cfg, MonoMethod *method, MonoInst *val)
2597 MonoType *ret = mini_type_get_underlying_type (NULL, mono_method_signature (method)->ret);
2599 if (ret->type == MONO_TYPE_R4) {
2600 if (COMPILE_LLVM (cfg))
2601 MONO_EMIT_NEW_UNALU (cfg, OP_FMOVE, cfg->ret->dreg, val->dreg);
2603 MONO_EMIT_NEW_UNALU (cfg, OP_AMD64_SET_XMMREG_R4, cfg->ret->dreg, val->dreg);
2605 } else if (ret->type == MONO_TYPE_R8) {
2606 MONO_EMIT_NEW_UNALU (cfg, OP_FMOVE, cfg->ret->dreg, val->dreg);
2610 MONO_EMIT_NEW_UNALU (cfg, OP_MOVE, cfg->ret->dreg, val->dreg);
2613 #endif /* DISABLE_JIT */
2615 #define EMIT_COND_BRANCH(ins,cond,sign) \
2616 if (ins->inst_true_bb->native_offset) { \
2617 x86_branch (code, cond, cfg->native_code + ins->inst_true_bb->native_offset, sign); \
2619 mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_BB, ins->inst_true_bb); \
2620 if ((cfg->opt & MONO_OPT_BRANCH) && \
2621 x86_is_imm8 (ins->inst_true_bb->max_offset - offset)) \
2622 x86_branch8 (code, cond, 0, sign); \
2624 x86_branch32 (code, cond, 0, sign); \
2628 MonoMethodSignature *sig;
2633 mgreg_t regs [PARAM_REGS];
2639 dyn_call_supported (MonoMethodSignature *sig, CallInfo *cinfo)
2647 switch (cinfo->ret.storage) {
2651 case ArgValuetypeInReg: {
2652 ArgInfo *ainfo = &cinfo->ret;
2654 if (ainfo->pair_storage [0] != ArgNone && ainfo->pair_storage [0] != ArgInIReg)
2656 if (ainfo->pair_storage [1] != ArgNone && ainfo->pair_storage [1] != ArgInIReg)
2664 for (i = 0; i < cinfo->nargs; ++i) {
2665 ArgInfo *ainfo = &cinfo->args [i];
2666 switch (ainfo->storage) {
2669 case ArgValuetypeInReg:
2670 if (ainfo->pair_storage [0] != ArgNone && ainfo->pair_storage [0] != ArgInIReg)
2672 if (ainfo->pair_storage [1] != ArgNone && ainfo->pair_storage [1] != ArgInIReg)
2684 * mono_arch_dyn_call_prepare:
2686 * Return a pointer to an arch-specific structure which contains information
2687 * needed by mono_arch_get_dyn_call_args (). Return NULL if OP_DYN_CALL is not
2688 * supported for SIG.
2689 * This function is equivalent to ffi_prep_cif in libffi.
2692 mono_arch_dyn_call_prepare (MonoMethodSignature *sig)
2694 ArchDynCallInfo *info;
2697 cinfo = get_call_info (NULL, NULL, sig);
2699 if (!dyn_call_supported (sig, cinfo)) {
2704 info = g_new0 (ArchDynCallInfo, 1);
2705 // FIXME: Preprocess the info to speed up get_dyn_call_args ().
2707 info->cinfo = cinfo;
2709 return (MonoDynCallInfo*)info;
2713 * mono_arch_dyn_call_free:
2715 * Free a MonoDynCallInfo structure.
2718 mono_arch_dyn_call_free (MonoDynCallInfo *info)
2720 ArchDynCallInfo *ainfo = (ArchDynCallInfo*)info;
2722 g_free (ainfo->cinfo);
2726 #if !defined(__native_client__)
2727 #define PTR_TO_GREG(ptr) (mgreg_t)(ptr)
2728 #define GREG_TO_PTR(greg) (gpointer)(greg)
2730 /* Correctly handle casts to/from 32-bit pointers without compiler warnings */
2731 #define PTR_TO_GREG(ptr) (mgreg_t)(uintptr_t)(ptr)
2732 #define GREG_TO_PTR(greg) (gpointer)(guint32)(greg)
2736 * mono_arch_get_start_dyn_call:
2738 * Convert the arguments ARGS to a format which can be passed to OP_DYN_CALL, and
2739 * store the result into BUF.
2740 * ARGS should be an array of pointers pointing to the arguments.
2741 * RET should point to a memory buffer large enought to hold the result of the
2743 * This function should be as fast as possible, any work which does not depend
2744 * on the actual values of the arguments should be done in
2745 * mono_arch_dyn_call_prepare ().
2746 * start_dyn_call + OP_DYN_CALL + finish_dyn_call is equivalent to ffi_call in
2750 mono_arch_start_dyn_call (MonoDynCallInfo *info, gpointer **args, guint8 *ret, guint8 *buf, int buf_len)
2752 ArchDynCallInfo *dinfo = (ArchDynCallInfo*)info;
2753 DynCallArgs *p = (DynCallArgs*)buf;
2754 int arg_index, greg, i, pindex;
2755 MonoMethodSignature *sig = dinfo->sig;
2757 g_assert (buf_len >= sizeof (DynCallArgs));
2766 if (sig->hasthis || dinfo->cinfo->vret_arg_index == 1) {
2767 p->regs [greg ++] = PTR_TO_GREG(*(args [arg_index ++]));
2772 if (dinfo->cinfo->vtype_retaddr)
2773 p->regs [greg ++] = PTR_TO_GREG(ret);
2775 for (i = pindex; i < sig->param_count; i++) {
2776 MonoType *t = mono_type_get_underlying_type (sig->params [i]);
2777 gpointer *arg = args [arg_index ++];
2780 p->regs [greg ++] = PTR_TO_GREG(*(arg));
2785 case MONO_TYPE_STRING:
2786 case MONO_TYPE_CLASS:
2787 case MONO_TYPE_ARRAY:
2788 case MONO_TYPE_SZARRAY:
2789 case MONO_TYPE_OBJECT:
2793 #if !defined(__mono_ilp32__)
2797 g_assert (dinfo->cinfo->args [i + sig->hasthis].reg == param_regs [greg]);
2798 p->regs [greg ++] = PTR_TO_GREG(*(arg));
2800 #if defined(__mono_ilp32__)
2803 g_assert (dinfo->cinfo->args [i + sig->hasthis].reg == param_regs [greg]);
2804 p->regs [greg ++] = *(guint64*)(arg);
2807 case MONO_TYPE_BOOLEAN:
2809 p->regs [greg ++] = *(guint8*)(arg);
2812 p->regs [greg ++] = *(gint8*)(arg);
2815 p->regs [greg ++] = *(gint16*)(arg);
2818 case MONO_TYPE_CHAR:
2819 p->regs [greg ++] = *(guint16*)(arg);
2822 p->regs [greg ++] = *(gint32*)(arg);
2825 p->regs [greg ++] = *(guint32*)(arg);
2827 case MONO_TYPE_GENERICINST:
2828 if (MONO_TYPE_IS_REFERENCE (t)) {
2829 p->regs [greg ++] = PTR_TO_GREG(*(arg));
2834 case MONO_TYPE_VALUETYPE: {
2835 ArgInfo *ainfo = &dinfo->cinfo->args [i + sig->hasthis];
2837 g_assert (ainfo->storage == ArgValuetypeInReg);
2838 if (ainfo->pair_storage [0] != ArgNone) {
2839 g_assert (ainfo->pair_storage [0] == ArgInIReg);
2840 p->regs [greg ++] = ((mgreg_t*)(arg))[0];
2842 if (ainfo->pair_storage [1] != ArgNone) {
2843 g_assert (ainfo->pair_storage [1] == ArgInIReg);
2844 p->regs [greg ++] = ((mgreg_t*)(arg))[1];
2849 g_assert_not_reached ();
2853 g_assert (greg <= PARAM_REGS);
2857 * mono_arch_finish_dyn_call:
2859 * Store the result of a dyn call into the return value buffer passed to
2860 * start_dyn_call ().
2861 * This function should be as fast as possible, any work which does not depend
2862 * on the actual values of the arguments should be done in
2863 * mono_arch_dyn_call_prepare ().
2866 mono_arch_finish_dyn_call (MonoDynCallInfo *info, guint8 *buf)
2868 ArchDynCallInfo *dinfo = (ArchDynCallInfo*)info;
2869 MonoMethodSignature *sig = dinfo->sig;
2870 guint8 *ret = ((DynCallArgs*)buf)->ret;
2871 mgreg_t res = ((DynCallArgs*)buf)->res;
2873 switch (mono_type_get_underlying_type (sig->ret)->type) {
2874 case MONO_TYPE_VOID:
2875 *(gpointer*)ret = NULL;
2877 case MONO_TYPE_STRING:
2878 case MONO_TYPE_CLASS:
2879 case MONO_TYPE_ARRAY:
2880 case MONO_TYPE_SZARRAY:
2881 case MONO_TYPE_OBJECT:
2885 *(gpointer*)ret = GREG_TO_PTR(res);
2891 case MONO_TYPE_BOOLEAN:
2892 *(guint8*)ret = res;
2895 *(gint16*)ret = res;
2898 case MONO_TYPE_CHAR:
2899 *(guint16*)ret = res;
2902 *(gint32*)ret = res;
2905 *(guint32*)ret = res;
2908 *(gint64*)ret = res;
2911 *(guint64*)ret = res;
2913 case MONO_TYPE_GENERICINST:
2914 if (MONO_TYPE_IS_REFERENCE (sig->ret)) {
2915 *(gpointer*)ret = GREG_TO_PTR(res);
2920 case MONO_TYPE_VALUETYPE:
2921 if (dinfo->cinfo->vtype_retaddr) {
2924 ArgInfo *ainfo = &dinfo->cinfo->ret;
2926 g_assert (ainfo->storage == ArgValuetypeInReg);
2928 if (ainfo->pair_storage [0] != ArgNone) {
2929 g_assert (ainfo->pair_storage [0] == ArgInIReg);
2930 ((mgreg_t*)ret)[0] = res;
2933 g_assert (ainfo->pair_storage [1] == ArgNone);
2937 g_assert_not_reached ();
2941 /* emit an exception if condition is fail */
2942 #define EMIT_COND_SYSTEM_EXCEPTION(cond,signed,exc_name) \
2944 MonoInst *tins = mono_branch_optimize_exception_target (cfg, bb, exc_name); \
2945 if (tins == NULL) { \
2946 mono_add_patch_info (cfg, code - cfg->native_code, \
2947 MONO_PATCH_INFO_EXC, exc_name); \
2948 x86_branch32 (code, cond, 0, signed); \
2950 EMIT_COND_BRANCH (tins, cond, signed); \
2954 #define EMIT_FPCOMPARE(code) do { \
2955 amd64_fcompp (code); \
2956 amd64_fnstsw (code); \
2959 #define EMIT_SSE2_FPFUNC(code, op, dreg, sreg1) do { \
2960 amd64_movsd_membase_reg (code, AMD64_RSP, -8, (sreg1)); \
2961 amd64_fld_membase (code, AMD64_RSP, -8, TRUE); \
2962 amd64_ ##op (code); \
2963 amd64_fst_membase (code, AMD64_RSP, -8, TRUE, TRUE); \
2964 amd64_movsd_reg_membase (code, (dreg), AMD64_RSP, -8); \
2968 emit_call_body (MonoCompile *cfg, guint8 *code, guint32 patch_type, gconstpointer data)
2970 gboolean no_patch = FALSE;
2973 * FIXME: Add support for thunks
2976 gboolean near_call = FALSE;
2979 * Indirect calls are expensive so try to make a near call if possible.
2980 * The caller memory is allocated by the code manager so it is
2981 * guaranteed to be at a 32 bit offset.
2984 if (patch_type != MONO_PATCH_INFO_ABS) {
2985 /* The target is in memory allocated using the code manager */
2988 if ((patch_type == MONO_PATCH_INFO_METHOD) || (patch_type == MONO_PATCH_INFO_METHOD_JUMP)) {
2989 if (((MonoMethod*)data)->klass->image->aot_module)
2990 /* The callee might be an AOT method */
2992 if (((MonoMethod*)data)->dynamic)
2993 /* The target is in malloc-ed memory */
2997 if (patch_type == MONO_PATCH_INFO_INTERNAL_METHOD) {
2999 * The call might go directly to a native function without
3002 MonoJitICallInfo *mi = mono_find_jit_icall_by_name (data);
3004 gconstpointer target = mono_icall_get_wrapper (mi);
3005 if ((((guint64)target) >> 32) != 0)
3011 if (cfg->abs_patches && g_hash_table_lookup (cfg->abs_patches, data)) {
3013 * This is not really an optimization, but required because the
3014 * generic class init trampolines use R11 to pass the vtable.
3018 MonoJitICallInfo *info = mono_find_jit_icall_by_addr (data);
3020 if ((cfg->method->wrapper_type == MONO_WRAPPER_MANAGED_TO_NATIVE) &&
3021 strstr (cfg->method->name, info->name)) {
3022 /* A call to the wrapped function */
3023 if ((((guint64)data) >> 32) == 0)
3027 else if (info->func == info->wrapper) {
3029 if ((((guint64)info->func) >> 32) == 0)
3033 /* See the comment in mono_codegen () */
3034 if ((info->name [0] != 'v') || (strstr (info->name, "ves_array_new_va_") == NULL && strstr (info->name, "ves_array_element_address_") == NULL))
3038 else if ((((guint64)data) >> 32) == 0) {
3045 if (cfg->method->dynamic)
3046 /* These methods are allocated using malloc */
3049 #ifdef MONO_ARCH_NOMAP32BIT
3053 /* The 64bit XEN kernel does not honour the MAP_32BIT flag. (#522894) */
3054 if (optimize_for_xen)
3057 if (cfg->compile_aot) {
3064 * Align the call displacement to an address divisible by 4 so it does
3065 * not span cache lines. This is required for code patching to work on SMP
3068 if (!no_patch && ((guint32)(code + 1 - cfg->native_code) % 4) != 0) {
3069 guint32 pad_size = 4 - ((guint32)(code + 1 - cfg->native_code) % 4);
3070 amd64_padding (code, pad_size);
3072 mono_add_patch_info (cfg, code - cfg->native_code, patch_type, data);
3073 amd64_call_code (code, 0);
3076 mono_add_patch_info (cfg, code - cfg->native_code, patch_type, data);
3077 amd64_set_reg_template (code, GP_SCRATCH_REG);
3078 amd64_call_reg (code, GP_SCRATCH_REG);
3085 static inline guint8*
3086 emit_call (MonoCompile *cfg, guint8 *code, guint32 patch_type, gconstpointer data, gboolean win64_adjust_stack)
3089 if (win64_adjust_stack)
3090 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, 32);
3092 code = emit_call_body (cfg, code, patch_type, data);
3094 if (win64_adjust_stack)
3095 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, 32);
3102 store_membase_imm_to_store_membase_reg (int opcode)
3105 case OP_STORE_MEMBASE_IMM:
3106 return OP_STORE_MEMBASE_REG;
3107 case OP_STOREI4_MEMBASE_IMM:
3108 return OP_STOREI4_MEMBASE_REG;
3109 case OP_STOREI8_MEMBASE_IMM:
3110 return OP_STOREI8_MEMBASE_REG;
3118 #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)))
3121 * mono_arch_peephole_pass_1:
3123 * Perform peephole opts which should/can be performed before local regalloc
3126 mono_arch_peephole_pass_1 (MonoCompile *cfg, MonoBasicBlock *bb)
3130 MONO_BB_FOR_EACH_INS_SAFE (bb, n, ins) {
3131 MonoInst *last_ins = ins->prev;
3133 switch (ins->opcode) {
3137 if ((ins->sreg1 < MONO_MAX_IREGS) && (ins->dreg >= MONO_MAX_IREGS) && (ins->inst_imm > 0)) {
3139 * X86_LEA is like ADD, but doesn't have the
3140 * sreg1==dreg restriction. inst_imm > 0 is needed since LEA sign-extends
3141 * its operand to 64 bit.
3143 ins->opcode = OP_X86_LEA_MEMBASE;
3144 ins->inst_basereg = ins->sreg1;
3149 if ((ins->sreg1 == ins->sreg2) && (ins->sreg1 == ins->dreg)) {
3153 * Replace STORE_MEMBASE_IMM 0 with STORE_MEMBASE_REG since
3154 * the latter has length 2-3 instead of 6 (reverse constant
3155 * propagation). These instruction sequences are very common
3156 * in the initlocals bblock.
3158 for (ins2 = ins->next; ins2; ins2 = ins2->next) {
3159 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)) {
3160 ins2->opcode = store_membase_imm_to_store_membase_reg (ins2->opcode);
3161 ins2->sreg1 = ins->dreg;
3162 } 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)) {
3164 } else if (((ins2->opcode == OP_ICONST) || (ins2->opcode == OP_I8CONST)) && (ins2->dreg == ins->dreg) && (ins2->inst_c0 == 0)) {
3173 case OP_COMPARE_IMM:
3174 case OP_LCOMPARE_IMM:
3175 /* OP_COMPARE_IMM (reg, 0)
3177 * OP_AMD64_TEST_NULL (reg)
3180 ins->opcode = OP_AMD64_TEST_NULL;
3182 case OP_ICOMPARE_IMM:
3184 ins->opcode = OP_X86_TEST_NULL;
3186 case OP_AMD64_ICOMPARE_MEMBASE_IMM:
3188 * OP_STORE_MEMBASE_REG reg, offset(basereg)
3189 * OP_X86_COMPARE_MEMBASE_IMM offset(basereg), imm
3191 * OP_STORE_MEMBASE_REG reg, offset(basereg)
3192 * OP_COMPARE_IMM reg, imm
3194 * Note: if imm = 0 then OP_COMPARE_IMM replaced with OP_X86_TEST_NULL
3196 if (last_ins && (last_ins->opcode == OP_STOREI4_MEMBASE_REG) &&
3197 ins->inst_basereg == last_ins->inst_destbasereg &&
3198 ins->inst_offset == last_ins->inst_offset) {
3199 ins->opcode = OP_ICOMPARE_IMM;
3200 ins->sreg1 = last_ins->sreg1;
3202 /* check if we can remove cmp reg,0 with test null */
3204 ins->opcode = OP_X86_TEST_NULL;
3210 mono_peephole_ins (bb, ins);
3215 mono_arch_peephole_pass_2 (MonoCompile *cfg, MonoBasicBlock *bb)
3219 MONO_BB_FOR_EACH_INS_SAFE (bb, n, ins) {
3220 switch (ins->opcode) {
3223 /* reg = 0 -> XOR (reg, reg) */
3224 /* XOR sets cflags on x86, so we cant do it always */
3225 if (ins->inst_c0 == 0 && (!ins->next || (ins->next && INST_IGNORES_CFLAGS (ins->next->opcode)))) {
3226 ins->opcode = OP_LXOR;
3227 ins->sreg1 = ins->dreg;
3228 ins->sreg2 = ins->dreg;
3236 * Use IXOR to avoid a rex prefix if possible. The cpu will sign extend the
3237 * 0 result into 64 bits.
3239 if ((ins->sreg1 == ins->sreg2) && (ins->sreg1 == ins->dreg)) {
3240 ins->opcode = OP_IXOR;
3244 if ((ins->sreg1 == ins->sreg2) && (ins->sreg1 == ins->dreg)) {
3248 * Replace STORE_MEMBASE_IMM 0 with STORE_MEMBASE_REG since
3249 * the latter has length 2-3 instead of 6 (reverse constant
3250 * propagation). These instruction sequences are very common
3251 * in the initlocals bblock.
3253 for (ins2 = ins->next; ins2; ins2 = ins2->next) {
3254 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)) {
3255 ins2->opcode = store_membase_imm_to_store_membase_reg (ins2->opcode);
3256 ins2->sreg1 = ins->dreg;
3257 } 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)) {
3259 } else if (((ins2->opcode == OP_ICONST) || (ins2->opcode == OP_I8CONST)) && (ins2->dreg == ins->dreg) && (ins2->inst_c0 == 0)) {
3269 if ((ins->inst_imm == 1) && (ins->dreg == ins->sreg1))
3270 ins->opcode = OP_X86_INC_REG;
3273 if ((ins->inst_imm == 1) && (ins->dreg == ins->sreg1))
3274 ins->opcode = OP_X86_DEC_REG;
3278 mono_peephole_ins (bb, ins);
3282 #define NEW_INS(cfg,ins,dest,op) do { \
3283 MONO_INST_NEW ((cfg), (dest), (op)); \
3284 (dest)->cil_code = (ins)->cil_code; \
3285 mono_bblock_insert_before_ins (bb, ins, (dest)); \
3289 * mono_arch_lowering_pass:
3291 * Converts complex opcodes into simpler ones so that each IR instruction
3292 * corresponds to one machine instruction.
3295 mono_arch_lowering_pass (MonoCompile *cfg, MonoBasicBlock *bb)
3297 MonoInst *ins, *n, *temp;
3300 * FIXME: Need to add more instructions, but the current machine
3301 * description can't model some parts of the composite instructions like
3304 MONO_BB_FOR_EACH_INS_SAFE (bb, n, ins) {
3305 switch (ins->opcode) {
3309 case OP_IDIV_UN_IMM:
3310 case OP_IREM_UN_IMM:
3311 mono_decompose_op_imm (cfg, bb, ins);
3314 /* Keep the opcode if we can implement it efficiently */
3315 if (!((ins->inst_imm > 0) && (mono_is_power_of_two (ins->inst_imm) != -1)))
3316 mono_decompose_op_imm (cfg, bb, ins);
3318 case OP_COMPARE_IMM:
3319 case OP_LCOMPARE_IMM:
3320 if (!amd64_is_imm32 (ins->inst_imm)) {
3321 NEW_INS (cfg, ins, temp, OP_I8CONST);
3322 temp->inst_c0 = ins->inst_imm;
3323 temp->dreg = mono_alloc_ireg (cfg);
3324 ins->opcode = OP_COMPARE;
3325 ins->sreg2 = temp->dreg;
3328 #ifndef __mono_ilp32__
3329 case OP_LOAD_MEMBASE:
3331 case OP_LOADI8_MEMBASE:
3332 #ifndef __native_client_codegen__
3333 /* Don't generate memindex opcodes (to simplify */
3334 /* read sandboxing) */
3335 if (!amd64_is_imm32 (ins->inst_offset)) {
3336 NEW_INS (cfg, ins, temp, OP_I8CONST);
3337 temp->inst_c0 = ins->inst_offset;
3338 temp->dreg = mono_alloc_ireg (cfg);
3339 ins->opcode = OP_AMD64_LOADI8_MEMINDEX;
3340 ins->inst_indexreg = temp->dreg;
3344 #ifndef __mono_ilp32__
3345 case OP_STORE_MEMBASE_IMM:
3347 case OP_STOREI8_MEMBASE_IMM:
3348 if (!amd64_is_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_STOREI8_MEMBASE_REG;
3353 ins->sreg1 = temp->dreg;
3356 #ifdef MONO_ARCH_SIMD_INTRINSICS
3357 case OP_EXPAND_I1: {
3358 int temp_reg1 = mono_alloc_ireg (cfg);
3359 int temp_reg2 = mono_alloc_ireg (cfg);
3360 int original_reg = ins->sreg1;
3362 NEW_INS (cfg, ins, temp, OP_ICONV_TO_U1);
3363 temp->sreg1 = original_reg;
3364 temp->dreg = temp_reg1;
3366 NEW_INS (cfg, ins, temp, OP_SHL_IMM);
3367 temp->sreg1 = temp_reg1;
3368 temp->dreg = temp_reg2;
3371 NEW_INS (cfg, ins, temp, OP_LOR);
3372 temp->sreg1 = temp->dreg = temp_reg2;
3373 temp->sreg2 = temp_reg1;
3375 ins->opcode = OP_EXPAND_I2;
3376 ins->sreg1 = temp_reg2;
3385 bb->max_vreg = cfg->next_vreg;
3389 branch_cc_table [] = {
3390 X86_CC_EQ, X86_CC_GE, X86_CC_GT, X86_CC_LE, X86_CC_LT,
3391 X86_CC_NE, X86_CC_GE, X86_CC_GT, X86_CC_LE, X86_CC_LT,
3392 X86_CC_O, X86_CC_NO, X86_CC_C, X86_CC_NC
3395 /* Maps CMP_... constants to X86_CC_... constants */
3398 X86_CC_EQ, X86_CC_NE, X86_CC_LE, X86_CC_GE, X86_CC_LT, X86_CC_GT,
3399 X86_CC_LE, X86_CC_GE, X86_CC_LT, X86_CC_GT
3403 cc_signed_table [] = {
3404 TRUE, TRUE, TRUE, TRUE, TRUE, TRUE,
3405 FALSE, FALSE, FALSE, FALSE
3408 /*#include "cprop.c"*/
3410 static unsigned char*
3411 emit_float_to_int (MonoCompile *cfg, guchar *code, int dreg, int sreg, int size, gboolean is_signed)
3413 amd64_sse_cvttsd2si_reg_reg (code, dreg, sreg);
3416 amd64_widen_reg (code, dreg, dreg, is_signed, FALSE);
3418 amd64_widen_reg (code, dreg, dreg, is_signed, TRUE);
3422 static unsigned char*
3423 mono_emit_stack_alloc (MonoCompile *cfg, guchar *code, MonoInst* tree)
3425 int sreg = tree->sreg1;
3426 int need_touch = FALSE;
3428 #if defined(HOST_WIN32) || defined(MONO_ARCH_SIGSEGV_ON_ALTSTACK)
3429 if (!tree->flags & MONO_INST_INIT)
3438 * If requested stack size is larger than one page,
3439 * perform stack-touch operation
3442 * Generate stack probe code.
3443 * Under Windows, it is necessary to allocate one page at a time,
3444 * "touching" stack after each successful sub-allocation. This is
3445 * because of the way stack growth is implemented - there is a
3446 * guard page before the lowest stack page that is currently commited.
3447 * Stack normally grows sequentially so OS traps access to the
3448 * guard page and commits more pages when needed.
3450 amd64_test_reg_imm (code, sreg, ~0xFFF);
3451 br[0] = code; x86_branch8 (code, X86_CC_Z, 0, FALSE);
3453 br[2] = code; /* loop */
3454 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, 0x1000);
3455 amd64_test_membase_reg (code, AMD64_RSP, 0, AMD64_RSP);
3456 amd64_alu_reg_imm (code, X86_SUB, sreg, 0x1000);
3457 amd64_alu_reg_imm (code, X86_CMP, sreg, 0x1000);
3458 br[3] = code; x86_branch8 (code, X86_CC_AE, 0, FALSE);
3459 amd64_patch (br[3], br[2]);
3460 amd64_test_reg_reg (code, sreg, sreg);
3461 br[4] = code; x86_branch8 (code, X86_CC_Z, 0, FALSE);
3462 amd64_alu_reg_reg (code, X86_SUB, AMD64_RSP, sreg);
3464 br[1] = code; x86_jump8 (code, 0);
3466 amd64_patch (br[0], code);
3467 amd64_alu_reg_reg (code, X86_SUB, AMD64_RSP, sreg);
3468 amd64_patch (br[1], code);
3469 amd64_patch (br[4], code);
3472 amd64_alu_reg_reg (code, X86_SUB, AMD64_RSP, tree->sreg1);
3474 if (tree->flags & MONO_INST_INIT) {
3476 if (tree->dreg != AMD64_RAX && sreg != AMD64_RAX) {
3477 amd64_push_reg (code, AMD64_RAX);
3480 if (tree->dreg != AMD64_RCX && sreg != AMD64_RCX) {
3481 amd64_push_reg (code, AMD64_RCX);
3484 if (tree->dreg != AMD64_RDI && sreg != AMD64_RDI) {
3485 amd64_push_reg (code, AMD64_RDI);
3489 amd64_shift_reg_imm (code, X86_SHR, sreg, 3);
3490 if (sreg != AMD64_RCX)
3491 amd64_mov_reg_reg (code, AMD64_RCX, sreg, 8);
3492 amd64_alu_reg_reg (code, X86_XOR, AMD64_RAX, AMD64_RAX);
3494 amd64_lea_membase (code, AMD64_RDI, AMD64_RSP, offset);
3495 if (cfg->param_area && cfg->arch.no_pushes)
3496 amd64_alu_reg_imm (code, X86_ADD, AMD64_RDI, cfg->param_area);
3498 #if defined(__default_codegen__)
3499 amd64_prefix (code, X86_REP_PREFIX);
3501 #elif defined(__native_client_codegen__)
3502 /* NaCl stos pseudo-instruction */
3503 amd64_codegen_pre(code);
3504 /* First, clear the upper 32 bits of RDI (mov %edi, %edi) */
3505 amd64_mov_reg_reg (code, AMD64_RDI, AMD64_RDI, 4);
3506 /* Add %r15 to %rdi using lea, condition flags unaffected. */
3507 amd64_lea_memindex_size (code, AMD64_RDI, AMD64_R15, 0, AMD64_RDI, 0, 8);
3508 amd64_prefix (code, X86_REP_PREFIX);
3510 amd64_codegen_post(code);
3511 #endif /* __native_client_codegen__ */
3513 if (tree->dreg != AMD64_RDI && sreg != AMD64_RDI)
3514 amd64_pop_reg (code, AMD64_RDI);
3515 if (tree->dreg != AMD64_RCX && sreg != AMD64_RCX)
3516 amd64_pop_reg (code, AMD64_RCX);
3517 if (tree->dreg != AMD64_RAX && sreg != AMD64_RAX)
3518 amd64_pop_reg (code, AMD64_RAX);
3524 emit_move_return_value (MonoCompile *cfg, MonoInst *ins, guint8 *code)
3529 /* Move return value to the target register */
3530 /* FIXME: do this in the local reg allocator */
3531 switch (ins->opcode) {
3534 case OP_CALL_MEMBASE:
3537 case OP_LCALL_MEMBASE:
3538 g_assert (ins->dreg == AMD64_RAX);
3542 case OP_FCALL_MEMBASE:
3543 if (((MonoCallInst*)ins)->signature->ret->type == MONO_TYPE_R4) {
3544 amd64_sse_cvtss2sd_reg_reg (code, ins->dreg, AMD64_XMM0);
3547 if (ins->dreg != AMD64_XMM0)
3548 amd64_sse_movsd_reg_reg (code, ins->dreg, AMD64_XMM0);
3553 case OP_VCALL_MEMBASE:
3556 case OP_VCALL2_MEMBASE:
3557 cinfo = get_call_info (cfg->generic_sharing_context, cfg->mempool, ((MonoCallInst*)ins)->signature);
3558 if (cinfo->ret.storage == ArgValuetypeInReg) {
3559 MonoInst *loc = cfg->arch.vret_addr_loc;
3561 /* Load the destination address */
3562 g_assert (loc->opcode == OP_REGOFFSET);
3563 amd64_mov_reg_membase (code, AMD64_RCX, loc->inst_basereg, loc->inst_offset, sizeof(gpointer));
3565 for (quad = 0; quad < 2; quad ++) {
3566 switch (cinfo->ret.pair_storage [quad]) {
3568 amd64_mov_membase_reg (code, AMD64_RCX, (quad * sizeof(mgreg_t)), cinfo->ret.pair_regs [quad], sizeof(mgreg_t));
3570 case ArgInFloatSSEReg:
3571 amd64_movss_membase_reg (code, AMD64_RCX, (quad * 8), cinfo->ret.pair_regs [quad]);
3573 case ArgInDoubleSSEReg:
3574 amd64_movsd_membase_reg (code, AMD64_RCX, (quad * 8), cinfo->ret.pair_regs [quad]);
3589 #endif /* DISABLE_JIT */
3592 mono_amd64_have_tls_get (void)
3595 static gboolean have_tls_get = FALSE;
3596 static gboolean inited = FALSE;
3599 return have_tls_get;
3601 guint8 *ins = (guint8*)pthread_getspecific;
3604 * We're looking for these two instructions:
3606 * mov %gs:0x60(,%rdi,8),%rax
3609 have_tls_get = ins [0] == 0x65 &&
3622 return have_tls_get;
3629 * mono_amd64_emit_tls_get:
3630 * @code: buffer to store code to
3631 * @dreg: hard register where to place the result
3632 * @tls_offset: offset info
3634 * mono_amd64_emit_tls_get emits in @code the native code that puts in
3635 * the dreg register the item in the thread local storage identified
3638 * Returns: a pointer to the end of the stored code
3641 mono_amd64_emit_tls_get (guint8* code, int dreg, int tls_offset)
3644 g_assert (tls_offset < 64);
3645 x86_prefix (code, X86_GS_PREFIX);
3646 amd64_mov_reg_mem (code, dreg, (tls_offset * 8) + 0x1480, 8);
3647 #elif defined(__APPLE__)
3648 x86_prefix (code, X86_GS_PREFIX);
3649 amd64_mov_reg_mem (code, dreg, 0x60 + tls_offset * 8, 8);
3651 if (optimize_for_xen) {
3652 x86_prefix (code, X86_FS_PREFIX);
3653 amd64_mov_reg_mem (code, dreg, 0, 8);
3654 amd64_mov_reg_membase (code, dreg, dreg, tls_offset, 8);
3656 x86_prefix (code, X86_FS_PREFIX);
3657 amd64_mov_reg_mem (code, dreg, tls_offset, 8);
3663 #define REAL_PRINT_REG(text,reg) \
3664 mono_assert (reg >= 0); \
3665 amd64_push_reg (code, AMD64_RAX); \
3666 amd64_push_reg (code, AMD64_RDX); \
3667 amd64_push_reg (code, AMD64_RCX); \
3668 amd64_push_reg (code, reg); \
3669 amd64_push_imm (code, reg); \
3670 amd64_push_imm (code, text " %d %p\n"); \
3671 amd64_mov_reg_imm (code, AMD64_RAX, printf); \
3672 amd64_call_reg (code, AMD64_RAX); \
3673 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, 3*4); \
3674 amd64_pop_reg (code, AMD64_RCX); \
3675 amd64_pop_reg (code, AMD64_RDX); \
3676 amd64_pop_reg (code, AMD64_RAX);
3678 /* benchmark and set based on cpu */
3679 #define LOOP_ALIGNMENT 8
3680 #define bb_is_loop_start(bb) ((bb)->loop_body_start && (bb)->nesting)
3684 #if defined(__native_client__) || defined(__native_client_codegen__)
3687 #ifdef __native_client_gc__
3688 __nacl_suspend_thread_if_needed();
3694 mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
3699 guint8 *code = cfg->native_code + cfg->code_len;
3700 MonoInst *last_ins = NULL;
3701 guint last_offset = 0;
3704 /* Fix max_offset estimate for each successor bb */
3705 if (cfg->opt & MONO_OPT_BRANCH) {
3706 int current_offset = cfg->code_len;
3707 MonoBasicBlock *current_bb;
3708 for (current_bb = bb; current_bb != NULL; current_bb = current_bb->next_bb) {
3709 current_bb->max_offset = current_offset;
3710 current_offset += current_bb->max_length;
3714 if (cfg->opt & MONO_OPT_LOOP) {
3715 int pad, align = LOOP_ALIGNMENT;
3716 /* set alignment depending on cpu */
3717 if (bb_is_loop_start (bb) && (pad = (cfg->code_len & (align - 1)))) {
3719 /*g_print ("adding %d pad at %x to loop in %s\n", pad, cfg->code_len, cfg->method->name);*/
3720 amd64_padding (code, pad);
3721 cfg->code_len += pad;
3722 bb->native_offset = cfg->code_len;
3726 #if defined(__native_client_codegen__)
3727 /* For Native Client, all indirect call/jump targets must be */
3728 /* 32-byte aligned. Exception handler blocks are jumped to */
3729 /* indirectly as well. */
3730 gboolean bb_needs_alignment = (bb->flags & BB_INDIRECT_JUMP_TARGET) ||
3731 (bb->flags & BB_EXCEPTION_HANDLER);
3733 if ( bb_needs_alignment && ((cfg->code_len & kNaClAlignmentMask) != 0)) {
3734 int pad = kNaClAlignment - (cfg->code_len & kNaClAlignmentMask);
3735 if (pad != kNaClAlignment) code = mono_arch_nacl_pad(code, pad);
3736 cfg->code_len += pad;
3737 bb->native_offset = cfg->code_len;
3739 #endif /*__native_client_codegen__*/
3741 if (cfg->verbose_level > 2)
3742 g_print ("Basic block %d starting at offset 0x%x\n", bb->block_num, bb->native_offset);
3744 if (cfg->prof_options & MONO_PROFILE_COVERAGE) {
3745 MonoProfileCoverageInfo *cov = cfg->coverage_info;
3746 g_assert (!cfg->compile_aot);
3748 cov->data [bb->dfn].cil_code = bb->cil_code;
3749 amd64_mov_reg_imm (code, AMD64_R11, (guint64)&cov->data [bb->dfn].count);
3750 /* this is not thread save, but good enough */
3751 amd64_inc_membase (code, AMD64_R11, 0);
3754 offset = code - cfg->native_code;
3756 mono_debug_open_block (cfg, bb, offset);
3758 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)
3759 x86_breakpoint (code);
3761 MONO_BB_FOR_EACH_INS (bb, ins) {
3762 offset = code - cfg->native_code;
3764 max_len = ((guint8 *)ins_get_spec (ins->opcode))[MONO_INST_LEN];
3766 #define EXTRA_CODE_SPACE (NACL_SIZE (16, 16 + kNaClAlignment))
3768 if (G_UNLIKELY (offset > (cfg->code_size - max_len - EXTRA_CODE_SPACE))) {
3769 cfg->code_size *= 2;
3770 cfg->native_code = mono_realloc_native_code(cfg);
3771 code = cfg->native_code + offset;
3772 mono_jit_stats.code_reallocs++;
3775 if (cfg->debug_info)
3776 mono_debug_record_line_number (cfg, ins, offset);
3778 switch (ins->opcode) {
3780 amd64_mul_reg (code, ins->sreg2, TRUE);
3783 amd64_mul_reg (code, ins->sreg2, FALSE);
3785 case OP_X86_SETEQ_MEMBASE:
3786 amd64_set_membase (code, X86_CC_EQ, ins->inst_basereg, ins->inst_offset, TRUE);
3788 case OP_STOREI1_MEMBASE_IMM:
3789 amd64_mov_membase_imm (code, ins->inst_destbasereg, ins->inst_offset, ins->inst_imm, 1);
3791 case OP_STOREI2_MEMBASE_IMM:
3792 amd64_mov_membase_imm (code, ins->inst_destbasereg, ins->inst_offset, ins->inst_imm, 2);
3794 case OP_STOREI4_MEMBASE_IMM:
3795 amd64_mov_membase_imm (code, ins->inst_destbasereg, ins->inst_offset, ins->inst_imm, 4);
3797 case OP_STOREI1_MEMBASE_REG:
3798 amd64_mov_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, ins->sreg1, 1);
3800 case OP_STOREI2_MEMBASE_REG:
3801 amd64_mov_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, ins->sreg1, 2);
3803 /* In AMD64 NaCl, pointers are 4 bytes, */
3804 /* so STORE_* != STOREI8_*. Likewise below. */
3805 case OP_STORE_MEMBASE_REG:
3806 amd64_mov_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, ins->sreg1, sizeof(gpointer));
3808 case OP_STOREI8_MEMBASE_REG:
3809 amd64_mov_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, ins->sreg1, 8);
3811 case OP_STOREI4_MEMBASE_REG:
3812 amd64_mov_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, ins->sreg1, 4);
3814 case OP_STORE_MEMBASE_IMM:
3815 #ifndef __native_client_codegen__
3816 /* In NaCl, this could be a PCONST type, which could */
3817 /* mean a pointer type was copied directly into the */
3818 /* lower 32-bits of inst_imm, so for InvalidPtr==-1 */
3819 /* the value would be 0x00000000FFFFFFFF which is */
3820 /* not proper for an imm32 unless you cast it. */
3821 g_assert (amd64_is_imm32 (ins->inst_imm));
3823 amd64_mov_membase_imm (code, ins->inst_destbasereg, ins->inst_offset, (gint32)ins->inst_imm, sizeof(gpointer));
3825 case OP_STOREI8_MEMBASE_IMM:
3826 g_assert (amd64_is_imm32 (ins->inst_imm));
3827 amd64_mov_membase_imm (code, ins->inst_destbasereg, ins->inst_offset, ins->inst_imm, 8);
3830 #ifdef __mono_ilp32__
3831 /* In ILP32, pointers are 4 bytes, so separate these */
3832 /* cases, use literal 8 below where we really want 8 */
3833 amd64_mov_reg_imm (code, ins->dreg, ins->inst_imm);
3834 amd64_mov_reg_membase (code, ins->dreg, ins->dreg, 0, sizeof(gpointer));
3838 // FIXME: Decompose this earlier
3839 if (amd64_is_imm32 (ins->inst_imm))
3840 amd64_mov_reg_mem (code, ins->dreg, ins->inst_imm, 8);
3842 amd64_mov_reg_imm (code, ins->dreg, ins->inst_imm);
3843 amd64_mov_reg_membase (code, ins->dreg, ins->dreg, 0, 8);
3847 amd64_mov_reg_imm (code, ins->dreg, ins->inst_imm);
3848 amd64_movsxd_reg_membase (code, ins->dreg, ins->dreg, 0);
3851 // FIXME: Decompose this earlier
3852 if (amd64_is_imm32 (ins->inst_imm))
3853 amd64_mov_reg_mem (code, ins->dreg, ins->inst_imm, 4);
3855 amd64_mov_reg_imm (code, ins->dreg, ins->inst_imm);
3856 amd64_mov_reg_membase (code, ins->dreg, ins->dreg, 0, 4);
3860 amd64_mov_reg_imm (code, ins->dreg, ins->inst_imm);
3861 amd64_widen_membase (code, ins->dreg, ins->dreg, 0, FALSE, FALSE);
3864 /* For NaCl, pointers are 4 bytes, so separate these */
3865 /* cases, use literal 8 below where we really want 8 */
3866 amd64_mov_reg_imm (code, ins->dreg, ins->inst_imm);
3867 amd64_widen_membase (code, ins->dreg, ins->dreg, 0, FALSE, TRUE);
3869 case OP_LOAD_MEMBASE:
3870 g_assert (amd64_is_imm32 (ins->inst_offset));
3871 amd64_mov_reg_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, sizeof(gpointer));
3873 case OP_LOADI8_MEMBASE:
3874 /* Use literal 8 instead of sizeof pointer or */
3875 /* register, we really want 8 for this opcode */
3876 g_assert (amd64_is_imm32 (ins->inst_offset));
3877 amd64_mov_reg_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, 8);
3879 case OP_LOADI4_MEMBASE:
3880 amd64_movsxd_reg_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset);
3882 case OP_LOADU4_MEMBASE:
3883 amd64_mov_reg_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, 4);
3885 case OP_LOADU1_MEMBASE:
3886 /* The cpu zero extends the result into 64 bits */
3887 amd64_widen_membase_size (code, ins->dreg, ins->inst_basereg, ins->inst_offset, FALSE, FALSE, 4);
3889 case OP_LOADI1_MEMBASE:
3890 amd64_widen_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, TRUE, FALSE);
3892 case OP_LOADU2_MEMBASE:
3893 /* The cpu zero extends the result into 64 bits */
3894 amd64_widen_membase_size (code, ins->dreg, ins->inst_basereg, ins->inst_offset, FALSE, TRUE, 4);
3896 case OP_LOADI2_MEMBASE:
3897 amd64_widen_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, TRUE, TRUE);
3899 case OP_AMD64_LOADI8_MEMINDEX:
3900 amd64_mov_reg_memindex_size (code, ins->dreg, ins->inst_basereg, 0, ins->inst_indexreg, 0, 8);
3902 case OP_LCONV_TO_I1:
3903 case OP_ICONV_TO_I1:
3905 amd64_widen_reg (code, ins->dreg, ins->sreg1, TRUE, FALSE);
3907 case OP_LCONV_TO_I2:
3908 case OP_ICONV_TO_I2:
3910 amd64_widen_reg (code, ins->dreg, ins->sreg1, TRUE, TRUE);
3912 case OP_LCONV_TO_U1:
3913 case OP_ICONV_TO_U1:
3914 amd64_widen_reg (code, ins->dreg, ins->sreg1, FALSE, FALSE);
3916 case OP_LCONV_TO_U2:
3917 case OP_ICONV_TO_U2:
3918 amd64_widen_reg (code, ins->dreg, ins->sreg1, FALSE, TRUE);
3921 /* Clean out the upper word */
3922 amd64_mov_reg_reg_size (code, ins->dreg, ins->sreg1, 4);
3925 amd64_movsxd_reg_reg (code, ins->dreg, ins->sreg1);
3929 amd64_alu_reg_reg (code, X86_CMP, ins->sreg1, ins->sreg2);
3931 case OP_COMPARE_IMM:
3932 case OP_LCOMPARE_IMM:
3933 g_assert (amd64_is_imm32 (ins->inst_imm));
3934 amd64_alu_reg_imm (code, X86_CMP, ins->sreg1, ins->inst_imm);
3936 case OP_X86_COMPARE_REG_MEMBASE:
3937 amd64_alu_reg_membase (code, X86_CMP, ins->sreg1, ins->sreg2, ins->inst_offset);
3939 case OP_X86_TEST_NULL:
3940 amd64_test_reg_reg_size (code, ins->sreg1, ins->sreg1, 4);
3942 case OP_AMD64_TEST_NULL:
3943 amd64_test_reg_reg (code, ins->sreg1, ins->sreg1);
3946 case OP_X86_ADD_REG_MEMBASE:
3947 amd64_alu_reg_membase_size (code, X86_ADD, ins->sreg1, ins->sreg2, ins->inst_offset, 4);
3949 case OP_X86_SUB_REG_MEMBASE:
3950 amd64_alu_reg_membase_size (code, X86_SUB, ins->sreg1, ins->sreg2, ins->inst_offset, 4);
3952 case OP_X86_AND_REG_MEMBASE:
3953 amd64_alu_reg_membase_size (code, X86_AND, ins->sreg1, ins->sreg2, ins->inst_offset, 4);
3955 case OP_X86_OR_REG_MEMBASE:
3956 amd64_alu_reg_membase_size (code, X86_OR, ins->sreg1, ins->sreg2, ins->inst_offset, 4);
3958 case OP_X86_XOR_REG_MEMBASE:
3959 amd64_alu_reg_membase_size (code, X86_XOR, ins->sreg1, ins->sreg2, ins->inst_offset, 4);
3962 case OP_X86_ADD_MEMBASE_IMM:
3963 /* FIXME: Make a 64 version too */
3964 amd64_alu_membase_imm_size (code, X86_ADD, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 4);
3966 case OP_X86_SUB_MEMBASE_IMM:
3967 g_assert (amd64_is_imm32 (ins->inst_imm));
3968 amd64_alu_membase_imm_size (code, X86_SUB, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 4);
3970 case OP_X86_AND_MEMBASE_IMM:
3971 g_assert (amd64_is_imm32 (ins->inst_imm));
3972 amd64_alu_membase_imm_size (code, X86_AND, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 4);
3974 case OP_X86_OR_MEMBASE_IMM:
3975 g_assert (amd64_is_imm32 (ins->inst_imm));
3976 amd64_alu_membase_imm_size (code, X86_OR, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 4);
3978 case OP_X86_XOR_MEMBASE_IMM:
3979 g_assert (amd64_is_imm32 (ins->inst_imm));
3980 amd64_alu_membase_imm_size (code, X86_XOR, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 4);
3982 case OP_X86_ADD_MEMBASE_REG:
3983 amd64_alu_membase_reg_size (code, X86_ADD, ins->inst_basereg, ins->inst_offset, ins->sreg2, 4);
3985 case OP_X86_SUB_MEMBASE_REG:
3986 amd64_alu_membase_reg_size (code, X86_SUB, ins->inst_basereg, ins->inst_offset, ins->sreg2, 4);
3988 case OP_X86_AND_MEMBASE_REG:
3989 amd64_alu_membase_reg_size (code, X86_AND, ins->inst_basereg, ins->inst_offset, ins->sreg2, 4);
3991 case OP_X86_OR_MEMBASE_REG:
3992 amd64_alu_membase_reg_size (code, X86_OR, ins->inst_basereg, ins->inst_offset, ins->sreg2, 4);
3994 case OP_X86_XOR_MEMBASE_REG:
3995 amd64_alu_membase_reg_size (code, X86_XOR, ins->inst_basereg, ins->inst_offset, ins->sreg2, 4);
3997 case OP_X86_INC_MEMBASE:
3998 amd64_inc_membase_size (code, ins->inst_basereg, ins->inst_offset, 4);
4000 case OP_X86_INC_REG:
4001 amd64_inc_reg_size (code, ins->dreg, 4);
4003 case OP_X86_DEC_MEMBASE:
4004 amd64_dec_membase_size (code, ins->inst_basereg, ins->inst_offset, 4);
4006 case OP_X86_DEC_REG:
4007 amd64_dec_reg_size (code, ins->dreg, 4);
4009 case OP_X86_MUL_REG_MEMBASE:
4010 case OP_X86_MUL_MEMBASE_REG:
4011 amd64_imul_reg_membase_size (code, ins->sreg1, ins->sreg2, ins->inst_offset, 4);
4013 case OP_AMD64_ICOMPARE_MEMBASE_REG:
4014 amd64_alu_membase_reg_size (code, X86_CMP, ins->inst_basereg, ins->inst_offset, ins->sreg2, 4);
4016 case OP_AMD64_ICOMPARE_MEMBASE_IMM:
4017 amd64_alu_membase_imm_size (code, X86_CMP, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 4);
4019 case OP_AMD64_COMPARE_MEMBASE_REG:
4020 amd64_alu_membase_reg_size (code, X86_CMP, ins->inst_basereg, ins->inst_offset, ins->sreg2, 8);
4022 case OP_AMD64_COMPARE_MEMBASE_IMM:
4023 g_assert (amd64_is_imm32 (ins->inst_imm));
4024 amd64_alu_membase_imm_size (code, X86_CMP, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 8);
4026 case OP_X86_COMPARE_MEMBASE8_IMM:
4027 amd64_alu_membase8_imm_size (code, X86_CMP, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 4);
4029 case OP_AMD64_ICOMPARE_REG_MEMBASE:
4030 amd64_alu_reg_membase_size (code, X86_CMP, ins->sreg1, ins->sreg2, ins->inst_offset, 4);
4032 case OP_AMD64_COMPARE_REG_MEMBASE:
4033 amd64_alu_reg_membase_size (code, X86_CMP, ins->sreg1, ins->sreg2, ins->inst_offset, 8);
4036 case OP_AMD64_ADD_REG_MEMBASE:
4037 amd64_alu_reg_membase_size (code, X86_ADD, ins->sreg1, ins->sreg2, ins->inst_offset, 8);
4039 case OP_AMD64_SUB_REG_MEMBASE:
4040 amd64_alu_reg_membase_size (code, X86_SUB, ins->sreg1, ins->sreg2, ins->inst_offset, 8);
4042 case OP_AMD64_AND_REG_MEMBASE:
4043 amd64_alu_reg_membase_size (code, X86_AND, ins->sreg1, ins->sreg2, ins->inst_offset, 8);
4045 case OP_AMD64_OR_REG_MEMBASE:
4046 amd64_alu_reg_membase_size (code, X86_OR, ins->sreg1, ins->sreg2, ins->inst_offset, 8);
4048 case OP_AMD64_XOR_REG_MEMBASE:
4049 amd64_alu_reg_membase_size (code, X86_XOR, ins->sreg1, ins->sreg2, ins->inst_offset, 8);
4052 case OP_AMD64_ADD_MEMBASE_REG:
4053 amd64_alu_membase_reg_size (code, X86_ADD, ins->inst_basereg, ins->inst_offset, ins->sreg2, 8);
4055 case OP_AMD64_SUB_MEMBASE_REG:
4056 amd64_alu_membase_reg_size (code, X86_SUB, ins->inst_basereg, ins->inst_offset, ins->sreg2, 8);
4058 case OP_AMD64_AND_MEMBASE_REG:
4059 amd64_alu_membase_reg_size (code, X86_AND, ins->inst_basereg, ins->inst_offset, ins->sreg2, 8);
4061 case OP_AMD64_OR_MEMBASE_REG:
4062 amd64_alu_membase_reg_size (code, X86_OR, ins->inst_basereg, ins->inst_offset, ins->sreg2, 8);
4064 case OP_AMD64_XOR_MEMBASE_REG:
4065 amd64_alu_membase_reg_size (code, X86_XOR, ins->inst_basereg, ins->inst_offset, ins->sreg2, 8);
4068 case OP_AMD64_ADD_MEMBASE_IMM:
4069 g_assert (amd64_is_imm32 (ins->inst_imm));
4070 amd64_alu_membase_imm_size (code, X86_ADD, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 8);
4072 case OP_AMD64_SUB_MEMBASE_IMM:
4073 g_assert (amd64_is_imm32 (ins->inst_imm));
4074 amd64_alu_membase_imm_size (code, X86_SUB, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 8);
4076 case OP_AMD64_AND_MEMBASE_IMM:
4077 g_assert (amd64_is_imm32 (ins->inst_imm));
4078 amd64_alu_membase_imm_size (code, X86_AND, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 8);
4080 case OP_AMD64_OR_MEMBASE_IMM:
4081 g_assert (amd64_is_imm32 (ins->inst_imm));
4082 amd64_alu_membase_imm_size (code, X86_OR, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 8);
4084 case OP_AMD64_XOR_MEMBASE_IMM:
4085 g_assert (amd64_is_imm32 (ins->inst_imm));
4086 amd64_alu_membase_imm_size (code, X86_XOR, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 8);
4090 amd64_breakpoint (code);
4092 case OP_RELAXED_NOP:
4093 x86_prefix (code, X86_REP_PREFIX);
4101 case OP_DUMMY_STORE:
4102 case OP_NOT_REACHED:
4105 case OP_SEQ_POINT: {
4108 if (cfg->compile_aot)
4112 * Read from the single stepping trigger page. This will cause a
4113 * SIGSEGV when single stepping is enabled.
4114 * We do this _before_ the breakpoint, so single stepping after
4115 * a breakpoint is hit will step to the next IL offset.
4117 if (ins->flags & MONO_INST_SINGLE_STEP_LOC) {
4118 if (((guint64)ss_trigger_page >> 32) == 0)
4119 amd64_mov_reg_mem (code, AMD64_R11, (guint64)ss_trigger_page, 4);
4121 MonoInst *var = cfg->arch.ss_trigger_page_var;
4123 amd64_mov_reg_membase (code, AMD64_R11, var->inst_basereg, var->inst_offset, 8);
4124 amd64_alu_membase_imm_size (code, X86_CMP, AMD64_R11, 0, 0, 4);
4129 * This is the address which is saved in seq points,
4130 * get_ip_for_single_step () / get_ip_for_breakpoint () needs to compute this
4131 * from the address of the instruction causing the fault.
4133 mono_add_seq_point (cfg, bb, ins, code - cfg->native_code);
4136 * A placeholder for a possible breakpoint inserted by
4137 * mono_arch_set_breakpoint ().
4139 for (i = 0; i < breakpoint_size; ++i)
4145 amd64_alu_reg_reg (code, X86_ADD, ins->sreg1, ins->sreg2);
4148 amd64_alu_reg_reg (code, X86_ADC, ins->sreg1, ins->sreg2);
4152 g_assert (amd64_is_imm32 (ins->inst_imm));
4153 amd64_alu_reg_imm (code, X86_ADD, ins->dreg, ins->inst_imm);
4156 g_assert (amd64_is_imm32 (ins->inst_imm));
4157 amd64_alu_reg_imm (code, X86_ADC, ins->dreg, ins->inst_imm);
4161 amd64_alu_reg_reg (code, X86_SUB, ins->sreg1, ins->sreg2);
4164 amd64_alu_reg_reg (code, X86_SBB, ins->sreg1, ins->sreg2);
4168 g_assert (amd64_is_imm32 (ins->inst_imm));
4169 amd64_alu_reg_imm (code, X86_SUB, ins->dreg, ins->inst_imm);
4172 g_assert (amd64_is_imm32 (ins->inst_imm));
4173 amd64_alu_reg_imm (code, X86_SBB, ins->dreg, ins->inst_imm);
4176 amd64_alu_reg_reg (code, X86_AND, ins->sreg1, ins->sreg2);
4180 g_assert (amd64_is_imm32 (ins->inst_imm));
4181 amd64_alu_reg_imm (code, X86_AND, ins->sreg1, ins->inst_imm);
4184 amd64_imul_reg_reg (code, ins->sreg1, ins->sreg2);
4189 guint32 size = (ins->opcode == OP_IMUL_IMM) ? 4 : 8;
4191 switch (ins->inst_imm) {
4195 if (ins->dreg != ins->sreg1)
4196 amd64_mov_reg_reg (code, ins->dreg, ins->sreg1, size);
4197 amd64_alu_reg_reg (code, X86_ADD, ins->dreg, ins->dreg);
4200 /* LEA r1, [r2 + r2*2] */
4201 amd64_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 1);
4204 /* LEA r1, [r2 + r2*4] */
4205 amd64_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 2);
4208 /* LEA r1, [r2 + r2*2] */
4210 amd64_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 1);
4211 amd64_alu_reg_reg (code, X86_ADD, ins->dreg, ins->dreg);
4214 /* LEA r1, [r2 + r2*8] */
4215 amd64_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 3);
4218 /* LEA r1, [r2 + r2*4] */
4220 amd64_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 2);
4221 amd64_alu_reg_reg (code, X86_ADD, ins->dreg, ins->dreg);
4224 /* LEA r1, [r2 + r2*2] */
4226 amd64_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 1);
4227 amd64_shift_reg_imm (code, X86_SHL, ins->dreg, 2);
4230 /* LEA r1, [r2 + r2*4] */
4231 /* LEA r1, [r1 + r1*4] */
4232 amd64_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 2);
4233 amd64_lea_memindex (code, ins->dreg, ins->dreg, 0, ins->dreg, 2);
4236 /* LEA r1, [r2 + r2*4] */
4238 /* LEA r1, [r1 + r1*4] */
4239 amd64_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 2);
4240 amd64_shift_reg_imm (code, X86_SHL, ins->dreg, 2);
4241 amd64_lea_memindex (code, ins->dreg, ins->dreg, 0, ins->dreg, 2);
4244 amd64_imul_reg_reg_imm_size (code, ins->dreg, ins->sreg1, ins->inst_imm, size);
4251 /* Regalloc magic makes the div/rem cases the same */
4252 if (ins->sreg2 == AMD64_RDX) {
4253 amd64_mov_membase_reg (code, AMD64_RSP, -8, AMD64_RDX, 8);
4255 amd64_div_membase (code, AMD64_RSP, -8, TRUE);
4258 amd64_div_reg (code, ins->sreg2, TRUE);
4263 if (ins->sreg2 == AMD64_RDX) {
4264 amd64_mov_membase_reg (code, AMD64_RSP, -8, AMD64_RDX, 8);
4265 amd64_alu_reg_reg (code, X86_XOR, AMD64_RDX, AMD64_RDX);
4266 amd64_div_membase (code, AMD64_RSP, -8, FALSE);
4268 amd64_alu_reg_reg (code, X86_XOR, AMD64_RDX, AMD64_RDX);
4269 amd64_div_reg (code, ins->sreg2, FALSE);
4274 if (ins->sreg2 == AMD64_RDX) {
4275 amd64_mov_membase_reg (code, AMD64_RSP, -8, AMD64_RDX, 8);
4276 amd64_cdq_size (code, 4);
4277 amd64_div_membase_size (code, AMD64_RSP, -8, TRUE, 4);
4279 amd64_cdq_size (code, 4);
4280 amd64_div_reg_size (code, ins->sreg2, TRUE, 4);
4285 if (ins->sreg2 == AMD64_RDX) {
4286 amd64_mov_membase_reg (code, AMD64_RSP, -8, AMD64_RDX, 8);
4287 amd64_alu_reg_reg (code, X86_XOR, AMD64_RDX, AMD64_RDX);
4288 amd64_div_membase_size (code, AMD64_RSP, -8, FALSE, 4);
4290 amd64_alu_reg_reg (code, X86_XOR, AMD64_RDX, AMD64_RDX);
4291 amd64_div_reg_size (code, ins->sreg2, FALSE, 4);
4295 int power = mono_is_power_of_two (ins->inst_imm);
4297 g_assert (ins->sreg1 == X86_EAX);
4298 g_assert (ins->dreg == X86_EAX);
4299 g_assert (power >= 0);
4302 amd64_mov_reg_imm (code, ins->dreg, 0);
4306 /* Based on gcc code */
4308 /* Add compensation for negative dividents */
4309 amd64_mov_reg_reg_size (code, AMD64_RDX, AMD64_RAX, 4);
4311 amd64_shift_reg_imm_size (code, X86_SAR, AMD64_RDX, 31, 4);
4312 amd64_shift_reg_imm_size (code, X86_SHR, AMD64_RDX, 32 - power, 4);
4313 amd64_alu_reg_reg_size (code, X86_ADD, AMD64_RAX, AMD64_RDX, 4);
4314 /* Compute remainder */
4315 amd64_alu_reg_imm_size (code, X86_AND, AMD64_RAX, (1 << power) - 1, 4);
4316 /* Remove compensation */
4317 amd64_alu_reg_reg_size (code, X86_SUB, AMD64_RAX, AMD64_RDX, 4);
4321 amd64_imul_reg_reg (code, ins->sreg1, ins->sreg2);
4322 EMIT_COND_SYSTEM_EXCEPTION (X86_CC_O, FALSE, "OverflowException");
4325 amd64_alu_reg_reg (code, X86_OR, ins->sreg1, ins->sreg2);
4329 g_assert (amd64_is_imm32 (ins->inst_imm));
4330 amd64_alu_reg_imm (code, X86_OR, ins->sreg1, ins->inst_imm);
4333 amd64_alu_reg_reg (code, X86_XOR, ins->sreg1, ins->sreg2);
4337 g_assert (amd64_is_imm32 (ins->inst_imm));
4338 amd64_alu_reg_imm (code, X86_XOR, ins->sreg1, ins->inst_imm);
4341 g_assert (ins->sreg2 == AMD64_RCX);
4342 amd64_shift_reg (code, X86_SHL, ins->dreg);
4345 g_assert (ins->sreg2 == AMD64_RCX);
4346 amd64_shift_reg (code, X86_SAR, ins->dreg);
4349 g_assert (amd64_is_imm32 (ins->inst_imm));
4350 amd64_shift_reg_imm_size (code, X86_SAR, ins->dreg, ins->inst_imm, 4);
4353 g_assert (amd64_is_imm32 (ins->inst_imm));
4354 amd64_shift_reg_imm (code, X86_SAR, ins->dreg, ins->inst_imm);
4357 g_assert (amd64_is_imm32 (ins->inst_imm));
4358 amd64_shift_reg_imm_size (code, X86_SHR, ins->dreg, ins->inst_imm, 4);
4360 case OP_LSHR_UN_IMM:
4361 g_assert (amd64_is_imm32 (ins->inst_imm));
4362 amd64_shift_reg_imm (code, X86_SHR, ins->dreg, ins->inst_imm);
4365 g_assert (ins->sreg2 == AMD64_RCX);
4366 amd64_shift_reg (code, X86_SHR, ins->dreg);
4369 g_assert (amd64_is_imm32 (ins->inst_imm));
4370 amd64_shift_reg_imm_size (code, X86_SHL, ins->dreg, ins->inst_imm, 4);
4373 g_assert (amd64_is_imm32 (ins->inst_imm));
4374 amd64_shift_reg_imm (code, X86_SHL, ins->dreg, ins->inst_imm);
4379 amd64_alu_reg_reg_size (code, X86_ADD, ins->sreg1, ins->sreg2, 4);
4382 amd64_alu_reg_reg_size (code, X86_ADC, ins->sreg1, ins->sreg2, 4);
4385 amd64_alu_reg_imm_size (code, X86_ADD, ins->dreg, ins->inst_imm, 4);
4388 amd64_alu_reg_imm_size (code, X86_ADC, ins->dreg, ins->inst_imm, 4);
4392 amd64_alu_reg_reg_size (code, X86_SUB, ins->sreg1, ins->sreg2, 4);
4395 amd64_alu_reg_reg_size (code, X86_SBB, ins->sreg1, ins->sreg2, 4);
4398 amd64_alu_reg_imm_size (code, X86_SUB, ins->dreg, ins->inst_imm, 4);
4401 amd64_alu_reg_imm_size (code, X86_SBB, ins->dreg, ins->inst_imm, 4);
4404 amd64_alu_reg_reg_size (code, X86_AND, ins->sreg1, ins->sreg2, 4);
4407 amd64_alu_reg_imm_size (code, X86_AND, ins->sreg1, ins->inst_imm, 4);
4410 amd64_alu_reg_reg_size (code, X86_OR, ins->sreg1, ins->sreg2, 4);
4413 amd64_alu_reg_imm_size (code, X86_OR, ins->sreg1, ins->inst_imm, 4);
4416 amd64_alu_reg_reg_size (code, X86_XOR, ins->sreg1, ins->sreg2, 4);
4419 amd64_alu_reg_imm_size (code, X86_XOR, ins->sreg1, ins->inst_imm, 4);
4422 amd64_neg_reg_size (code, ins->sreg1, 4);
4425 amd64_not_reg_size (code, ins->sreg1, 4);
4428 g_assert (ins->sreg2 == AMD64_RCX);
4429 amd64_shift_reg_size (code, X86_SHL, ins->dreg, 4);
4432 g_assert (ins->sreg2 == AMD64_RCX);
4433 amd64_shift_reg_size (code, X86_SAR, ins->dreg, 4);
4436 amd64_shift_reg_imm_size (code, X86_SAR, ins->dreg, ins->inst_imm, 4);
4438 case OP_ISHR_UN_IMM:
4439 amd64_shift_reg_imm_size (code, X86_SHR, ins->dreg, ins->inst_imm, 4);
4442 g_assert (ins->sreg2 == AMD64_RCX);
4443 amd64_shift_reg_size (code, X86_SHR, ins->dreg, 4);
4446 amd64_shift_reg_imm_size (code, X86_SHL, ins->dreg, ins->inst_imm, 4);
4449 amd64_imul_reg_reg_size (code, ins->sreg1, ins->sreg2, 4);
4452 amd64_imul_reg_reg_size (code, ins->sreg1, ins->sreg2, 4);
4453 EMIT_COND_SYSTEM_EXCEPTION (X86_CC_O, FALSE, "OverflowException");
4455 case OP_IMUL_OVF_UN:
4456 case OP_LMUL_OVF_UN: {
4457 /* the mul operation and the exception check should most likely be split */
4458 int non_eax_reg, saved_eax = FALSE, saved_edx = FALSE;
4459 int size = (ins->opcode == OP_IMUL_OVF_UN) ? 4 : 8;
4460 /*g_assert (ins->sreg2 == X86_EAX);
4461 g_assert (ins->dreg == X86_EAX);*/
4462 if (ins->sreg2 == X86_EAX) {
4463 non_eax_reg = ins->sreg1;
4464 } else if (ins->sreg1 == X86_EAX) {
4465 non_eax_reg = ins->sreg2;
4467 /* no need to save since we're going to store to it anyway */
4468 if (ins->dreg != X86_EAX) {
4470 amd64_push_reg (code, X86_EAX);
4472 amd64_mov_reg_reg (code, X86_EAX, ins->sreg1, size);
4473 non_eax_reg = ins->sreg2;
4475 if (ins->dreg == X86_EDX) {
4478 amd64_push_reg (code, X86_EAX);
4482 amd64_push_reg (code, X86_EDX);
4484 amd64_mul_reg_size (code, non_eax_reg, FALSE, size);
4485 /* save before the check since pop and mov don't change the flags */
4486 if (ins->dreg != X86_EAX)
4487 amd64_mov_reg_reg (code, ins->dreg, X86_EAX, size);
4489 amd64_pop_reg (code, X86_EDX);
4491 amd64_pop_reg (code, X86_EAX);
4492 EMIT_COND_SYSTEM_EXCEPTION (X86_CC_O, FALSE, "OverflowException");
4496 amd64_alu_reg_reg_size (code, X86_CMP, ins->sreg1, ins->sreg2, 4);
4498 case OP_ICOMPARE_IMM:
4499 amd64_alu_reg_imm_size (code, X86_CMP, ins->sreg1, ins->inst_imm, 4);
4521 EMIT_COND_BRANCH (ins, cc_table [mono_opcode_to_cond (ins->opcode)], cc_signed_table [mono_opcode_to_cond (ins->opcode)]);
4529 case OP_CMOV_INE_UN:
4530 case OP_CMOV_IGE_UN:
4531 case OP_CMOV_IGT_UN:
4532 case OP_CMOV_ILE_UN:
4533 case OP_CMOV_ILT_UN:
4539 case OP_CMOV_LNE_UN:
4540 case OP_CMOV_LGE_UN:
4541 case OP_CMOV_LGT_UN:
4542 case OP_CMOV_LLE_UN:
4543 case OP_CMOV_LLT_UN:
4544 g_assert (ins->dreg == ins->sreg1);
4545 /* This needs to operate on 64 bit values */
4546 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);
4550 amd64_not_reg (code, ins->sreg1);
4553 amd64_neg_reg (code, ins->sreg1);
4558 if ((((guint64)ins->inst_c0) >> 32) == 0)
4559 amd64_mov_reg_imm_size (code, ins->dreg, ins->inst_c0, 4);
4561 amd64_mov_reg_imm_size (code, ins->dreg, ins->inst_c0, 8);
4564 mono_add_patch_info (cfg, offset, (MonoJumpInfoType)ins->inst_i1, ins->inst_p0);
4565 amd64_mov_reg_membase (code, ins->dreg, AMD64_RIP, 0, sizeof(gpointer));
4568 mono_add_patch_info (cfg, offset, (MonoJumpInfoType)ins->inst_i1, ins->inst_p0);
4569 amd64_mov_reg_imm_size (code, ins->dreg, 0, 8);
4572 amd64_mov_reg_reg (code, ins->dreg, ins->sreg1, sizeof(mgreg_t));
4574 case OP_AMD64_SET_XMMREG_R4: {
4575 amd64_sse_cvtsd2ss_reg_reg (code, ins->dreg, ins->sreg1);
4578 case OP_AMD64_SET_XMMREG_R8: {
4579 if (ins->dreg != ins->sreg1)
4580 amd64_sse_movsd_reg_reg (code, ins->dreg, ins->sreg1);
4584 MonoCallInst *call = (MonoCallInst*)ins;
4587 /* FIXME: no tracing support... */
4588 if (cfg->prof_options & MONO_PROFILE_ENTER_LEAVE)
4589 code = mono_arch_instrument_epilog_full (cfg, mono_profiler_method_leave, code, FALSE, FALSE);
4591 g_assert (!cfg->method->save_lmf);
4593 if (cfg->arch.omit_fp) {
4594 guint32 save_offset = 0;
4595 /* Pop callee-saved registers */
4596 for (i = 0; i < AMD64_NREG; ++i)
4597 if (AMD64_IS_CALLEE_SAVED_REG (i) && (cfg->used_int_regs & (1 << i))) {
4598 amd64_mov_reg_membase (code, i, AMD64_RSP, save_offset, 8);
4601 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, cfg->arch.stack_alloc_size);
4604 if (call->stack_usage)
4608 for (i = 0; i < AMD64_NREG; ++i)
4609 if (AMD64_IS_CALLEE_SAVED_REG (i) && (cfg->used_int_regs & (1 << i)))
4610 pos -= sizeof(mgreg_t);
4612 /* Restore callee-saved registers */
4613 for (i = AMD64_NREG - 1; i > 0; --i) {
4614 if (AMD64_IS_CALLEE_SAVED_REG (i) && (cfg->used_int_regs & (1 << i))) {
4615 amd64_mov_reg_membase (code, i, AMD64_RBP, pos, sizeof(mgreg_t));
4616 pos += sizeof(mgreg_t);
4620 /* Copy arguments on the stack to our argument area */
4621 for (i = 0; i < call->stack_usage; i += sizeof(mgreg_t)) {
4622 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RSP, i, sizeof(mgreg_t));
4623 amd64_mov_membase_reg (code, AMD64_RBP, 16 + i, AMD64_RAX, sizeof(mgreg_t));
4627 amd64_lea_membase (code, AMD64_RSP, AMD64_RBP, pos);
4632 offset = code - cfg->native_code;
4633 mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_METHOD_JUMP, ins->inst_p0);
4634 if (cfg->compile_aot)
4635 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RIP, 0, 8);
4637 amd64_set_reg_template (code, AMD64_R11);
4638 amd64_jump_reg (code, AMD64_R11);
4639 ins->flags |= MONO_INST_GC_CALLSITE;
4640 ins->backend.pc_offset = code - cfg->native_code;
4644 /* ensure ins->sreg1 is not NULL */
4645 amd64_alu_membase_imm_size (code, X86_CMP, ins->sreg1, 0, 0, 4);
4648 amd64_lea_membase (code, AMD64_R11, cfg->frame_reg, cfg->sig_cookie);
4649 amd64_mov_membase_reg (code, ins->sreg1, 0, AMD64_R11, sizeof(gpointer));
4658 call = (MonoCallInst*)ins;
4660 * The AMD64 ABI forces callers to know about varargs.
4662 if ((call->signature->call_convention == MONO_CALL_VARARG) && (call->signature->pinvoke))
4663 amd64_alu_reg_reg (code, X86_XOR, AMD64_RAX, AMD64_RAX);
4664 else if ((cfg->method->wrapper_type == MONO_WRAPPER_MANAGED_TO_NATIVE) && (cfg->method->klass->image != mono_defaults.corlib)) {
4666 * Since the unmanaged calling convention doesn't contain a
4667 * 'vararg' entry, we have to treat every pinvoke call as a
4668 * potential vararg call.
4672 for (i = 0; i < AMD64_XMM_NREG; ++i)
4673 if (call->used_fregs & (1 << i))
4676 amd64_alu_reg_reg (code, X86_XOR, AMD64_RAX, AMD64_RAX);
4678 amd64_mov_reg_imm (code, AMD64_RAX, nregs);
4681 if (ins->flags & MONO_INST_HAS_METHOD)
4682 code = emit_call (cfg, code, MONO_PATCH_INFO_METHOD, call->method, FALSE);
4684 code = emit_call (cfg, code, MONO_PATCH_INFO_ABS, call->fptr, FALSE);
4685 ins->flags |= MONO_INST_GC_CALLSITE;
4686 ins->backend.pc_offset = code - cfg->native_code;
4687 if (call->stack_usage && !CALLCONV_IS_STDCALL (call->signature->call_convention) && !cfg->arch.no_pushes)
4688 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, call->stack_usage);
4689 code = emit_move_return_value (cfg, ins, code);
4695 case OP_VOIDCALL_REG:
4697 call = (MonoCallInst*)ins;
4699 if (AMD64_IS_ARGUMENT_REG (ins->sreg1)) {
4700 amd64_mov_reg_reg (code, AMD64_R11, ins->sreg1, 8);
4701 ins->sreg1 = AMD64_R11;
4705 * The AMD64 ABI forces callers to know about varargs.
4707 if ((call->signature->call_convention == MONO_CALL_VARARG) && (call->signature->pinvoke)) {
4708 if (ins->sreg1 == AMD64_RAX) {
4709 amd64_mov_reg_reg (code, AMD64_R11, AMD64_RAX, 8);
4710 ins->sreg1 = AMD64_R11;
4712 amd64_alu_reg_reg (code, X86_XOR, AMD64_RAX, AMD64_RAX);
4713 } else if ((cfg->method->wrapper_type == MONO_WRAPPER_MANAGED_TO_NATIVE) && (cfg->method->klass->image != mono_defaults.corlib)) {
4715 * Since the unmanaged calling convention doesn't contain a
4716 * 'vararg' entry, we have to treat every pinvoke call as a
4717 * potential vararg call.
4721 for (i = 0; i < AMD64_XMM_NREG; ++i)
4722 if (call->used_fregs & (1 << i))
4724 if (ins->sreg1 == AMD64_RAX) {
4725 amd64_mov_reg_reg (code, AMD64_R11, AMD64_RAX, 8);
4726 ins->sreg1 = AMD64_R11;
4729 amd64_alu_reg_reg (code, X86_XOR, AMD64_RAX, AMD64_RAX);
4731 amd64_mov_reg_imm (code, AMD64_RAX, nregs);
4734 amd64_call_reg (code, ins->sreg1);
4735 ins->flags |= MONO_INST_GC_CALLSITE;
4736 ins->backend.pc_offset = code - cfg->native_code;
4737 if (call->stack_usage && !CALLCONV_IS_STDCALL (call->signature->call_convention) && !cfg->arch.no_pushes)
4738 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, call->stack_usage);
4739 code = emit_move_return_value (cfg, ins, code);
4741 case OP_FCALL_MEMBASE:
4742 case OP_LCALL_MEMBASE:
4743 case OP_VCALL_MEMBASE:
4744 case OP_VCALL2_MEMBASE:
4745 case OP_VOIDCALL_MEMBASE:
4746 case OP_CALL_MEMBASE:
4747 call = (MonoCallInst*)ins;
4749 amd64_call_membase (code, ins->sreg1, ins->inst_offset);
4750 ins->flags |= MONO_INST_GC_CALLSITE;
4751 ins->backend.pc_offset = code - cfg->native_code;
4752 if (call->stack_usage && !CALLCONV_IS_STDCALL (call->signature->call_convention) && !cfg->arch.no_pushes)
4753 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, call->stack_usage);
4754 code = emit_move_return_value (cfg, ins, code);
4758 MonoInst *var = cfg->dyn_call_var;
4760 g_assert (var->opcode == OP_REGOFFSET);
4762 /* r11 = args buffer filled by mono_arch_get_dyn_call_args () */
4763 amd64_mov_reg_reg (code, AMD64_R11, ins->sreg1, 8);
4765 amd64_mov_reg_reg (code, AMD64_R10, ins->sreg2, 8);
4767 /* Save args buffer */
4768 amd64_mov_membase_reg (code, var->inst_basereg, var->inst_offset, AMD64_R11, 8);
4770 /* Set argument registers */
4771 for (i = 0; i < PARAM_REGS; ++i)
4772 amd64_mov_reg_membase (code, param_regs [i], AMD64_R11, i * sizeof(mgreg_t), sizeof(mgreg_t));
4775 amd64_call_reg (code, AMD64_R10);
4777 ins->flags |= MONO_INST_GC_CALLSITE;
4778 ins->backend.pc_offset = code - cfg->native_code;
4781 amd64_mov_reg_membase (code, AMD64_R11, var->inst_basereg, var->inst_offset, 8);
4782 amd64_mov_membase_reg (code, AMD64_R11, G_STRUCT_OFFSET (DynCallArgs, res), AMD64_RAX, 8);
4785 case OP_AMD64_SAVE_SP_TO_LMF:
4786 amd64_mov_membase_reg (code, cfg->frame_reg, cfg->arch.lmf_offset + G_STRUCT_OFFSET (MonoLMF, rsp), AMD64_RSP, 8);
4789 g_assert (!cfg->arch.no_pushes);
4790 amd64_push_reg (code, ins->sreg1);
4792 case OP_X86_PUSH_IMM:
4793 g_assert (!cfg->arch.no_pushes);
4794 g_assert (amd64_is_imm32 (ins->inst_imm));
4795 amd64_push_imm (code, ins->inst_imm);
4797 case OP_X86_PUSH_MEMBASE:
4798 g_assert (!cfg->arch.no_pushes);
4799 amd64_push_membase (code, ins->inst_basereg, ins->inst_offset);
4801 case OP_X86_PUSH_OBJ: {
4802 int size = ALIGN_TO (ins->inst_imm, 8);
4804 g_assert (!cfg->arch.no_pushes);
4806 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, size);
4807 amd64_push_reg (code, AMD64_RDI);
4808 amd64_push_reg (code, AMD64_RSI);
4809 amd64_push_reg (code, AMD64_RCX);
4810 if (ins->inst_offset)
4811 amd64_lea_membase (code, AMD64_RSI, ins->inst_basereg, ins->inst_offset);
4813 amd64_mov_reg_reg (code, AMD64_RSI, ins->inst_basereg, 8);
4814 amd64_lea_membase (code, AMD64_RDI, AMD64_RSP, (3 * 8));
4815 amd64_mov_reg_imm (code, AMD64_RCX, (size >> 3));
4817 amd64_prefix (code, X86_REP_PREFIX);
4819 amd64_pop_reg (code, AMD64_RCX);
4820 amd64_pop_reg (code, AMD64_RSI);
4821 amd64_pop_reg (code, AMD64_RDI);
4825 amd64_lea_memindex (code, ins->dreg, ins->sreg1, ins->inst_imm, ins->sreg2, ins->backend.shift_amount);
4827 case OP_X86_LEA_MEMBASE:
4828 amd64_lea_membase (code, ins->dreg, ins->sreg1, ins->inst_imm);
4831 amd64_xchg_reg_reg (code, ins->sreg1, ins->sreg2, 4);
4834 /* keep alignment */
4835 amd64_alu_reg_imm (code, X86_ADD, ins->sreg1, MONO_ARCH_FRAME_ALIGNMENT - 1);
4836 amd64_alu_reg_imm (code, X86_AND, ins->sreg1, ~(MONO_ARCH_FRAME_ALIGNMENT - 1));
4837 code = mono_emit_stack_alloc (cfg, code, ins);
4838 amd64_mov_reg_reg (code, ins->dreg, AMD64_RSP, 8);
4839 if (cfg->param_area && cfg->arch.no_pushes)
4840 amd64_alu_reg_imm (code, X86_ADD, ins->dreg, cfg->param_area);
4842 case OP_LOCALLOC_IMM: {
4843 guint32 size = ins->inst_imm;
4844 size = (size + (MONO_ARCH_FRAME_ALIGNMENT - 1)) & ~ (MONO_ARCH_FRAME_ALIGNMENT - 1);
4846 if (ins->flags & MONO_INST_INIT) {
4850 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, size);
4851 amd64_alu_reg_reg (code, X86_XOR, ins->dreg, ins->dreg);
4853 for (i = 0; i < size; i += 8)
4854 amd64_mov_membase_reg (code, AMD64_RSP, i, ins->dreg, 8);
4855 amd64_mov_reg_reg (code, ins->dreg, AMD64_RSP, 8);
4857 amd64_mov_reg_imm (code, ins->dreg, size);
4858 ins->sreg1 = ins->dreg;
4860 code = mono_emit_stack_alloc (cfg, code, ins);
4861 amd64_mov_reg_reg (code, ins->dreg, AMD64_RSP, 8);
4864 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, size);
4865 amd64_mov_reg_reg (code, ins->dreg, AMD64_RSP, 8);
4867 if (cfg->param_area && cfg->arch.no_pushes)
4868 amd64_alu_reg_imm (code, X86_ADD, ins->dreg, cfg->param_area);
4872 amd64_mov_reg_reg (code, AMD64_ARG_REG1, ins->sreg1, 8);
4873 code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD,
4874 (gpointer)"mono_arch_throw_exception", FALSE);
4875 ins->flags |= MONO_INST_GC_CALLSITE;
4876 ins->backend.pc_offset = code - cfg->native_code;
4880 amd64_mov_reg_reg (code, AMD64_ARG_REG1, ins->sreg1, 8);
4881 code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD,
4882 (gpointer)"mono_arch_rethrow_exception", FALSE);
4883 ins->flags |= MONO_INST_GC_CALLSITE;
4884 ins->backend.pc_offset = code - cfg->native_code;
4887 case OP_CALL_HANDLER:
4889 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, 8);
4890 mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_BB, ins->inst_target_bb);
4891 amd64_call_imm (code, 0);
4892 mono_cfg_add_try_hole (cfg, ins->inst_eh_block, code, bb);
4893 /* Restore stack alignment */
4894 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, 8);
4896 case OP_START_HANDLER: {
4897 /* Even though we're saving RSP, use sizeof */
4898 /* gpointer because spvar is of type IntPtr */
4899 /* see: mono_create_spvar_for_region */
4900 MonoInst *spvar = mono_find_spvar_for_region (cfg, bb->region);
4901 amd64_mov_membase_reg (code, spvar->inst_basereg, spvar->inst_offset, AMD64_RSP, sizeof(gpointer));
4903 if ((MONO_BBLOCK_IS_IN_REGION (bb, MONO_REGION_FINALLY) ||
4904 MONO_BBLOCK_IS_IN_REGION (bb, MONO_REGION_FINALLY)) &&
4905 cfg->param_area && cfg->arch.no_pushes) {
4906 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, ALIGN_TO (cfg->param_area, MONO_ARCH_FRAME_ALIGNMENT));
4910 case OP_ENDFINALLY: {
4911 MonoInst *spvar = mono_find_spvar_for_region (cfg, bb->region);
4912 amd64_mov_reg_membase (code, AMD64_RSP, spvar->inst_basereg, spvar->inst_offset, sizeof(gpointer));
4916 case OP_ENDFILTER: {
4917 MonoInst *spvar = mono_find_spvar_for_region (cfg, bb->region);
4918 amd64_mov_reg_membase (code, AMD64_RSP, spvar->inst_basereg, spvar->inst_offset, sizeof(gpointer));
4919 /* The local allocator will put the result into RAX */
4925 ins->inst_c0 = code - cfg->native_code;
4928 //g_print ("target: %p, next: %p, curr: %p, last: %p\n", ins->inst_target_bb, bb->next_bb, ins, bb->last_ins);
4929 //if ((ins->inst_target_bb == bb->next_bb) && ins == bb->last_ins)
4931 if (ins->inst_target_bb->native_offset) {
4932 amd64_jump_code (code, cfg->native_code + ins->inst_target_bb->native_offset);
4934 mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_BB, ins->inst_target_bb);
4935 if ((cfg->opt & MONO_OPT_BRANCH) &&
4936 x86_is_imm8 (ins->inst_target_bb->max_offset - offset))
4937 x86_jump8 (code, 0);
4939 x86_jump32 (code, 0);
4943 amd64_jump_reg (code, ins->sreg1);
4960 amd64_set_reg (code, cc_table [mono_opcode_to_cond (ins->opcode)], ins->dreg, cc_signed_table [mono_opcode_to_cond (ins->opcode)]);
4961 amd64_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
4963 case OP_COND_EXC_EQ:
4964 case OP_COND_EXC_NE_UN:
4965 case OP_COND_EXC_LT:
4966 case OP_COND_EXC_LT_UN:
4967 case OP_COND_EXC_GT:
4968 case OP_COND_EXC_GT_UN:
4969 case OP_COND_EXC_GE:
4970 case OP_COND_EXC_GE_UN:
4971 case OP_COND_EXC_LE:
4972 case OP_COND_EXC_LE_UN:
4973 case OP_COND_EXC_IEQ:
4974 case OP_COND_EXC_INE_UN:
4975 case OP_COND_EXC_ILT:
4976 case OP_COND_EXC_ILT_UN:
4977 case OP_COND_EXC_IGT:
4978 case OP_COND_EXC_IGT_UN:
4979 case OP_COND_EXC_IGE:
4980 case OP_COND_EXC_IGE_UN:
4981 case OP_COND_EXC_ILE:
4982 case OP_COND_EXC_ILE_UN:
4983 EMIT_COND_SYSTEM_EXCEPTION (cc_table [mono_opcode_to_cond (ins->opcode)], cc_signed_table [mono_opcode_to_cond (ins->opcode)], ins->inst_p1);
4985 case OP_COND_EXC_OV:
4986 case OP_COND_EXC_NO:
4988 case OP_COND_EXC_NC:
4989 EMIT_COND_SYSTEM_EXCEPTION (branch_cc_table [ins->opcode - OP_COND_EXC_EQ],
4990 (ins->opcode < OP_COND_EXC_NE_UN), ins->inst_p1);
4992 case OP_COND_EXC_IOV:
4993 case OP_COND_EXC_INO:
4994 case OP_COND_EXC_IC:
4995 case OP_COND_EXC_INC:
4996 EMIT_COND_SYSTEM_EXCEPTION (branch_cc_table [ins->opcode - OP_COND_EXC_IEQ],
4997 (ins->opcode < OP_COND_EXC_INE_UN), ins->inst_p1);
5000 /* floating point opcodes */
5002 double d = *(double *)ins->inst_p0;
5004 if ((d == 0.0) && (mono_signbit (d) == 0)) {
5005 amd64_sse_xorpd_reg_reg (code, ins->dreg, ins->dreg);
5008 mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_R8, ins->inst_p0);
5009 amd64_sse_movsd_reg_membase (code, ins->dreg, AMD64_RIP, 0);
5014 float f = *(float *)ins->inst_p0;
5016 if ((f == 0.0) && (mono_signbit (f) == 0)) {
5017 amd64_sse_xorpd_reg_reg (code, ins->dreg, ins->dreg);
5020 mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_R4, ins->inst_p0);
5021 amd64_sse_movss_reg_membase (code, ins->dreg, AMD64_RIP, 0);
5022 amd64_sse_cvtss2sd_reg_reg (code, ins->dreg, ins->dreg);
5026 case OP_STORER8_MEMBASE_REG:
5027 amd64_sse_movsd_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, ins->sreg1);
5029 case OP_LOADR8_MEMBASE:
5030 amd64_sse_movsd_reg_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset);
5032 case OP_STORER4_MEMBASE_REG:
5033 /* This requires a double->single conversion */
5034 amd64_sse_cvtsd2ss_reg_reg (code, AMD64_XMM15, ins->sreg1);
5035 amd64_sse_movss_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, AMD64_XMM15);
5037 case OP_LOADR4_MEMBASE:
5038 amd64_sse_movss_reg_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset);
5039 amd64_sse_cvtss2sd_reg_reg (code, ins->dreg, ins->dreg);
5041 case OP_ICONV_TO_R4: /* FIXME: change precision */
5042 case OP_ICONV_TO_R8:
5043 amd64_sse_cvtsi2sd_reg_reg_size (code, ins->dreg, ins->sreg1, 4);
5045 case OP_LCONV_TO_R4: /* FIXME: change precision */
5046 case OP_LCONV_TO_R8:
5047 amd64_sse_cvtsi2sd_reg_reg (code, ins->dreg, ins->sreg1);
5049 case OP_FCONV_TO_R4:
5050 /* FIXME: nothing to do ?? */
5052 case OP_FCONV_TO_I1:
5053 code = emit_float_to_int (cfg, code, ins->dreg, ins->sreg1, 1, TRUE);
5055 case OP_FCONV_TO_U1:
5056 code = emit_float_to_int (cfg, code, ins->dreg, ins->sreg1, 1, FALSE);
5058 case OP_FCONV_TO_I2:
5059 code = emit_float_to_int (cfg, code, ins->dreg, ins->sreg1, 2, TRUE);
5061 case OP_FCONV_TO_U2:
5062 code = emit_float_to_int (cfg, code, ins->dreg, ins->sreg1, 2, FALSE);
5064 case OP_FCONV_TO_U4:
5065 code = emit_float_to_int (cfg, code, ins->dreg, ins->sreg1, 4, FALSE);
5067 case OP_FCONV_TO_I4:
5069 code = emit_float_to_int (cfg, code, ins->dreg, ins->sreg1, 4, TRUE);
5071 case OP_FCONV_TO_I8:
5072 code = emit_float_to_int (cfg, code, ins->dreg, ins->sreg1, 8, TRUE);
5074 case OP_LCONV_TO_R_UN: {
5077 /* Based on gcc code */
5078 amd64_test_reg_reg (code, ins->sreg1, ins->sreg1);
5079 br [0] = code; x86_branch8 (code, X86_CC_S, 0, TRUE);
5082 amd64_sse_cvtsi2sd_reg_reg (code, ins->dreg, ins->sreg1);
5083 br [1] = code; x86_jump8 (code, 0);
5084 amd64_patch (br [0], code);
5087 /* Save to the red zone */
5088 amd64_mov_membase_reg (code, AMD64_RSP, -8, AMD64_RAX, 8);
5089 amd64_mov_membase_reg (code, AMD64_RSP, -16, AMD64_RCX, 8);
5090 amd64_mov_reg_reg (code, AMD64_RCX, ins->sreg1, 8);
5091 amd64_mov_reg_reg (code, AMD64_RAX, ins->sreg1, 8);
5092 amd64_alu_reg_imm (code, X86_AND, AMD64_RCX, 1);
5093 amd64_shift_reg_imm (code, X86_SHR, AMD64_RAX, 1);
5094 amd64_alu_reg_imm (code, X86_OR, AMD64_RAX, AMD64_RCX);
5095 amd64_sse_cvtsi2sd_reg_reg (code, ins->dreg, AMD64_RAX);
5096 amd64_sse_addsd_reg_reg (code, ins->dreg, ins->dreg);
5098 amd64_mov_reg_membase (code, AMD64_RCX, AMD64_RSP, -16, 8);
5099 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RSP, -8, 8);
5100 amd64_patch (br [1], code);
5103 case OP_LCONV_TO_OVF_U4:
5104 amd64_alu_reg_imm (code, X86_CMP, ins->sreg1, 0);
5105 EMIT_COND_SYSTEM_EXCEPTION (X86_CC_LT, TRUE, "OverflowException");
5106 amd64_mov_reg_reg (code, ins->dreg, ins->sreg1, 8);
5108 case OP_LCONV_TO_OVF_I4_UN:
5109 amd64_alu_reg_imm (code, X86_CMP, ins->sreg1, 0x7fffffff);
5110 EMIT_COND_SYSTEM_EXCEPTION (X86_CC_GT, FALSE, "OverflowException");
5111 amd64_mov_reg_reg (code, ins->dreg, ins->sreg1, 8);
5114 if (ins->dreg != ins->sreg1)
5115 amd64_sse_movsd_reg_reg (code, ins->dreg, ins->sreg1);
5118 amd64_sse_addsd_reg_reg (code, ins->dreg, ins->sreg2);
5121 amd64_sse_subsd_reg_reg (code, ins->dreg, ins->sreg2);
5124 amd64_sse_mulsd_reg_reg (code, ins->dreg, ins->sreg2);
5127 amd64_sse_divsd_reg_reg (code, ins->dreg, ins->sreg2);
5130 static double r8_0 = -0.0;
5132 g_assert (ins->sreg1 == ins->dreg);
5134 mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_R8, &r8_0);
5135 amd64_sse_xorpd_reg_membase (code, ins->dreg, AMD64_RIP, 0);
5139 EMIT_SSE2_FPFUNC (code, fsin, ins->dreg, ins->sreg1);
5142 EMIT_SSE2_FPFUNC (code, fcos, ins->dreg, ins->sreg1);
5145 static guint64 d = 0x7fffffffffffffffUL;
5147 g_assert (ins->sreg1 == ins->dreg);
5149 mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_R8, &d);
5150 amd64_sse_andpd_reg_membase (code, ins->dreg, AMD64_RIP, 0);
5154 EMIT_SSE2_FPFUNC (code, fsqrt, ins->dreg, ins->sreg1);
5157 g_assert (cfg->opt & MONO_OPT_CMOV);
5158 g_assert (ins->dreg == ins->sreg1);
5159 amd64_alu_reg_reg_size (code, X86_CMP, ins->sreg1, ins->sreg2, 4);
5160 amd64_cmov_reg_size (code, X86_CC_GT, TRUE, ins->dreg, ins->sreg2, 4);
5163 g_assert (cfg->opt & MONO_OPT_CMOV);
5164 g_assert (ins->dreg == ins->sreg1);
5165 amd64_alu_reg_reg_size (code, X86_CMP, ins->sreg1, ins->sreg2, 4);
5166 amd64_cmov_reg_size (code, X86_CC_GT, FALSE, ins->dreg, ins->sreg2, 4);
5169 g_assert (cfg->opt & MONO_OPT_CMOV);
5170 g_assert (ins->dreg == ins->sreg1);
5171 amd64_alu_reg_reg_size (code, X86_CMP, ins->sreg1, ins->sreg2, 4);
5172 amd64_cmov_reg_size (code, X86_CC_LT, TRUE, ins->dreg, ins->sreg2, 4);
5175 g_assert (cfg->opt & MONO_OPT_CMOV);
5176 g_assert (ins->dreg == ins->sreg1);
5177 amd64_alu_reg_reg_size (code, X86_CMP, ins->sreg1, ins->sreg2, 4);
5178 amd64_cmov_reg_size (code, X86_CC_LT, FALSE, ins->dreg, ins->sreg2, 4);
5181 g_assert (cfg->opt & MONO_OPT_CMOV);
5182 g_assert (ins->dreg == ins->sreg1);
5183 amd64_alu_reg_reg (code, X86_CMP, ins->sreg1, ins->sreg2);
5184 amd64_cmov_reg (code, X86_CC_GT, TRUE, ins->dreg, ins->sreg2);
5187 g_assert (cfg->opt & MONO_OPT_CMOV);
5188 g_assert (ins->dreg == ins->sreg1);
5189 amd64_alu_reg_reg (code, X86_CMP, ins->sreg1, ins->sreg2);
5190 amd64_cmov_reg (code, X86_CC_GT, FALSE, ins->dreg, ins->sreg2);
5193 g_assert (cfg->opt & MONO_OPT_CMOV);
5194 g_assert (ins->dreg == ins->sreg1);
5195 amd64_alu_reg_reg (code, X86_CMP, ins->sreg1, ins->sreg2);
5196 amd64_cmov_reg (code, X86_CC_LT, TRUE, ins->dreg, ins->sreg2);
5199 g_assert (cfg->opt & MONO_OPT_CMOV);
5200 g_assert (ins->dreg == ins->sreg1);
5201 amd64_alu_reg_reg (code, X86_CMP, ins->sreg1, ins->sreg2);
5202 amd64_cmov_reg (code, X86_CC_LT, FALSE, ins->dreg, ins->sreg2);
5208 * The two arguments are swapped because the fbranch instructions
5209 * depend on this for the non-sse case to work.
5211 amd64_sse_comisd_reg_reg (code, ins->sreg2, ins->sreg1);
5214 /* zeroing the register at the start results in
5215 * shorter and faster code (we can also remove the widening op)
5217 guchar *unordered_check;
5218 amd64_alu_reg_reg (code, X86_XOR, ins->dreg, ins->dreg);
5219 amd64_sse_comisd_reg_reg (code, ins->sreg1, ins->sreg2);
5220 unordered_check = code;
5221 x86_branch8 (code, X86_CC_P, 0, FALSE);
5222 amd64_set_reg (code, X86_CC_EQ, ins->dreg, FALSE);
5223 amd64_patch (unordered_check, code);
5228 /* zeroing the register at the start results in
5229 * shorter and faster code (we can also remove the widening op)
5231 amd64_alu_reg_reg (code, X86_XOR, ins->dreg, ins->dreg);
5232 amd64_sse_comisd_reg_reg (code, ins->sreg2, ins->sreg1);
5233 if (ins->opcode == OP_FCLT_UN) {
5234 guchar *unordered_check = code;
5235 guchar *jump_to_end;
5236 x86_branch8 (code, X86_CC_P, 0, FALSE);
5237 amd64_set_reg (code, X86_CC_GT, ins->dreg, FALSE);
5239 x86_jump8 (code, 0);
5240 amd64_patch (unordered_check, code);
5241 amd64_inc_reg (code, ins->dreg);
5242 amd64_patch (jump_to_end, code);
5244 amd64_set_reg (code, X86_CC_GT, ins->dreg, FALSE);
5249 /* zeroing the register at the start results in
5250 * shorter and faster code (we can also remove the widening op)
5252 guchar *unordered_check;
5253 amd64_alu_reg_reg (code, X86_XOR, ins->dreg, ins->dreg);
5254 amd64_sse_comisd_reg_reg (code, ins->sreg2, ins->sreg1);
5255 if (ins->opcode == OP_FCGT) {
5256 unordered_check = code;
5257 x86_branch8 (code, X86_CC_P, 0, FALSE);
5258 amd64_set_reg (code, X86_CC_LT, ins->dreg, FALSE);
5259 amd64_patch (unordered_check, code);
5261 amd64_set_reg (code, X86_CC_LT, ins->dreg, FALSE);
5265 case OP_FCLT_MEMBASE:
5266 case OP_FCGT_MEMBASE:
5267 case OP_FCLT_UN_MEMBASE:
5268 case OP_FCGT_UN_MEMBASE:
5269 case OP_FCEQ_MEMBASE: {
5270 guchar *unordered_check, *jump_to_end;
5273 amd64_alu_reg_reg (code, X86_XOR, ins->dreg, ins->dreg);
5274 amd64_sse_comisd_reg_membase (code, ins->sreg1, ins->sreg2, ins->inst_offset);
5276 switch (ins->opcode) {
5277 case OP_FCEQ_MEMBASE:
5278 x86_cond = X86_CC_EQ;
5280 case OP_FCLT_MEMBASE:
5281 case OP_FCLT_UN_MEMBASE:
5282 x86_cond = X86_CC_LT;
5284 case OP_FCGT_MEMBASE:
5285 case OP_FCGT_UN_MEMBASE:
5286 x86_cond = X86_CC_GT;
5289 g_assert_not_reached ();
5292 unordered_check = code;
5293 x86_branch8 (code, X86_CC_P, 0, FALSE);
5294 amd64_set_reg (code, x86_cond, ins->dreg, FALSE);
5296 switch (ins->opcode) {
5297 case OP_FCEQ_MEMBASE:
5298 case OP_FCLT_MEMBASE:
5299 case OP_FCGT_MEMBASE:
5300 amd64_patch (unordered_check, code);
5302 case OP_FCLT_UN_MEMBASE:
5303 case OP_FCGT_UN_MEMBASE:
5305 x86_jump8 (code, 0);
5306 amd64_patch (unordered_check, code);
5307 amd64_inc_reg (code, ins->dreg);
5308 amd64_patch (jump_to_end, code);
5316 guchar *jump = code;
5317 x86_branch8 (code, X86_CC_P, 0, TRUE);
5318 EMIT_COND_BRANCH (ins, X86_CC_EQ, FALSE);
5319 amd64_patch (jump, code);
5323 /* Branch if C013 != 100 */
5324 /* branch if !ZF or (PF|CF) */
5325 EMIT_COND_BRANCH (ins, X86_CC_NE, FALSE);
5326 EMIT_COND_BRANCH (ins, X86_CC_P, FALSE);
5327 EMIT_COND_BRANCH (ins, X86_CC_B, FALSE);
5330 EMIT_COND_BRANCH (ins, X86_CC_GT, FALSE);
5333 EMIT_COND_BRANCH (ins, X86_CC_P, FALSE);
5334 EMIT_COND_BRANCH (ins, X86_CC_GT, FALSE);
5338 if (ins->opcode == OP_FBGT) {
5341 /* skip branch if C1=1 */
5343 x86_branch8 (code, X86_CC_P, 0, FALSE);
5344 /* branch if (C0 | C3) = 1 */
5345 EMIT_COND_BRANCH (ins, X86_CC_LT, FALSE);
5346 amd64_patch (br1, code);
5349 EMIT_COND_BRANCH (ins, X86_CC_LT, FALSE);
5353 /* Branch if C013 == 100 or 001 */
5356 /* skip branch if C1=1 */
5358 x86_branch8 (code, X86_CC_P, 0, FALSE);
5359 /* branch if (C0 | C3) = 1 */
5360 EMIT_COND_BRANCH (ins, X86_CC_BE, FALSE);
5361 amd64_patch (br1, code);
5365 /* Branch if C013 == 000 */
5366 EMIT_COND_BRANCH (ins, X86_CC_LE, FALSE);
5369 /* Branch if C013=000 or 100 */
5372 /* skip branch if C1=1 */
5374 x86_branch8 (code, X86_CC_P, 0, FALSE);
5375 /* branch if C0=0 */
5376 EMIT_COND_BRANCH (ins, X86_CC_NB, FALSE);
5377 amd64_patch (br1, code);
5381 /* Branch if C013 != 001 */
5382 EMIT_COND_BRANCH (ins, X86_CC_P, FALSE);
5383 EMIT_COND_BRANCH (ins, X86_CC_GE, FALSE);
5386 /* Transfer value to the fp stack */
5387 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, 16);
5388 amd64_movsd_membase_reg (code, AMD64_RSP, 0, ins->sreg1);
5389 amd64_fld_membase (code, AMD64_RSP, 0, TRUE);
5391 amd64_push_reg (code, AMD64_RAX);
5393 amd64_fnstsw (code);
5394 amd64_alu_reg_imm (code, X86_AND, AMD64_RAX, 0x4100);
5395 amd64_alu_reg_imm (code, X86_CMP, AMD64_RAX, X86_FP_C0);
5396 amd64_pop_reg (code, AMD64_RAX);
5397 amd64_fstp (code, 0);
5398 EMIT_COND_SYSTEM_EXCEPTION (X86_CC_EQ, FALSE, "ArithmeticException");
5399 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, 16);
5402 code = mono_amd64_emit_tls_get (code, ins->dreg, ins->inst_offset);
5405 case OP_MEMORY_BARRIER: {
5406 /* http://blogs.sun.com/dave/resource/NHM-Pipeline-Blog-V2.txt */
5407 x86_prefix (code, X86_LOCK_PREFIX);
5408 amd64_alu_membase_imm (code, X86_ADD, AMD64_RSP, 0, 0);
5411 case OP_ATOMIC_ADD_I4:
5412 case OP_ATOMIC_ADD_I8: {
5413 int dreg = ins->dreg;
5414 guint32 size = (ins->opcode == OP_ATOMIC_ADD_I4) ? 4 : 8;
5416 if (dreg == ins->inst_basereg)
5419 if (dreg != ins->sreg2)
5420 amd64_mov_reg_reg (code, ins->dreg, ins->sreg2, size);
5422 x86_prefix (code, X86_LOCK_PREFIX);
5423 amd64_xadd_membase_reg (code, ins->inst_basereg, ins->inst_offset, dreg, size);
5425 if (dreg != ins->dreg)
5426 amd64_mov_reg_reg (code, ins->dreg, dreg, size);
5430 case OP_ATOMIC_ADD_NEW_I4:
5431 case OP_ATOMIC_ADD_NEW_I8: {
5432 int dreg = ins->dreg;
5433 guint32 size = (ins->opcode == OP_ATOMIC_ADD_NEW_I4) ? 4 : 8;
5435 if ((dreg == ins->sreg2) || (dreg == ins->inst_basereg))
5438 amd64_mov_reg_reg (code, dreg, ins->sreg2, size);
5439 amd64_prefix (code, X86_LOCK_PREFIX);
5440 amd64_xadd_membase_reg (code, ins->inst_basereg, ins->inst_offset, dreg, size);
5441 /* dreg contains the old value, add with sreg2 value */
5442 amd64_alu_reg_reg_size (code, X86_ADD, dreg, ins->sreg2, size);
5444 if (ins->dreg != dreg)
5445 amd64_mov_reg_reg (code, ins->dreg, dreg, size);
5449 case OP_ATOMIC_EXCHANGE_I4:
5450 case OP_ATOMIC_EXCHANGE_I8: {
5452 int sreg2 = ins->sreg2;
5453 int breg = ins->inst_basereg;
5455 gboolean need_push = FALSE, rdx_pushed = FALSE;
5457 if (ins->opcode == OP_ATOMIC_EXCHANGE_I8)
5463 * See http://msdn.microsoft.com/en-us/magazine/cc302329.aspx for
5464 * an explanation of how this works.
5467 /* cmpxchg uses eax as comperand, need to make sure we can use it
5468 * hack to overcome limits in x86 reg allocator
5469 * (req: dreg == eax and sreg2 != eax and breg != eax)
5471 g_assert (ins->dreg == AMD64_RAX);
5473 if (breg == AMD64_RAX && ins->sreg2 == AMD64_RAX)
5474 /* Highly unlikely, but possible */
5477 /* The pushes invalidate rsp */
5478 if ((breg == AMD64_RAX) || need_push) {
5479 amd64_mov_reg_reg (code, AMD64_R11, breg, 8);
5483 /* We need the EAX reg for the comparand */
5484 if (ins->sreg2 == AMD64_RAX) {
5485 if (breg != AMD64_R11) {
5486 amd64_mov_reg_reg (code, AMD64_R11, AMD64_RAX, 8);
5489 g_assert (need_push);
5490 amd64_push_reg (code, AMD64_RDX);
5491 amd64_mov_reg_reg (code, AMD64_RDX, AMD64_RAX, size);
5497 amd64_mov_reg_membase (code, AMD64_RAX, breg, ins->inst_offset, size);
5499 br [0] = code; amd64_prefix (code, X86_LOCK_PREFIX);
5500 amd64_cmpxchg_membase_reg_size (code, breg, ins->inst_offset, sreg2, size);
5501 br [1] = code; amd64_branch8 (code, X86_CC_NE, -1, FALSE);
5502 amd64_patch (br [1], br [0]);
5505 amd64_pop_reg (code, AMD64_RDX);
5509 case OP_ATOMIC_CAS_I4:
5510 case OP_ATOMIC_CAS_I8: {
5513 if (ins->opcode == OP_ATOMIC_CAS_I8)
5519 * See http://msdn.microsoft.com/en-us/magazine/cc302329.aspx for
5520 * an explanation of how this works.
5522 g_assert (ins->sreg3 == AMD64_RAX);
5523 g_assert (ins->sreg1 != AMD64_RAX);
5524 g_assert (ins->sreg1 != ins->sreg2);
5526 amd64_prefix (code, X86_LOCK_PREFIX);
5527 amd64_cmpxchg_membase_reg_size (code, ins->sreg1, ins->inst_offset, ins->sreg2, size);
5529 if (ins->dreg != AMD64_RAX)
5530 amd64_mov_reg_reg (code, ins->dreg, AMD64_RAX, size);
5533 case OP_CARD_TABLE_WBARRIER: {
5534 int ptr = ins->sreg1;
5535 int value = ins->sreg2;
5537 int nursery_shift, card_table_shift;
5538 gpointer card_table_mask;
5539 size_t nursery_size;
5541 gpointer card_table = mono_gc_get_card_table (&card_table_shift, &card_table_mask);
5542 guint64 nursery_start = (guint64)mono_gc_get_nursery (&nursery_shift, &nursery_size);
5544 /*If either point to the stack we can simply avoid the WB. This happens due to
5545 * optimizations revealing a stack store that was not visible when op_cardtable was emited.
5547 if (ins->sreg1 == AMD64_RSP || ins->sreg2 == AMD64_RSP)
5551 * We need one register we can clobber, we choose EDX and make sreg1
5552 * fixed EAX to work around limitations in the local register allocator.
5553 * sreg2 might get allocated to EDX, but that is not a problem since
5554 * we use it before clobbering EDX.
5556 g_assert (ins->sreg1 == AMD64_RAX);
5559 * This is the code we produce:
5562 * edx >>= nursery_shift
5563 * cmp edx, (nursery_start >> nursery_shift)
5566 * edx >>= card_table_shift
5572 if (value != AMD64_RDX)
5573 amd64_mov_reg_reg (code, AMD64_RDX, value, 8);
5574 amd64_shift_reg_imm (code, X86_SHR, AMD64_RDX, nursery_shift);
5575 amd64_alu_reg_imm (code, X86_CMP, AMD64_RDX, nursery_start >> nursery_shift);
5576 br = code; x86_branch8 (code, X86_CC_NE, -1, FALSE);
5577 amd64_mov_reg_reg (code, AMD64_RDX, ptr, 8);
5578 amd64_shift_reg_imm (code, X86_SHR, AMD64_RDX, card_table_shift);
5579 if (card_table_mask)
5580 amd64_alu_reg_imm (code, X86_AND, AMD64_RDX, (guint32)(guint64)card_table_mask);
5582 mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_GC_CARD_TABLE_ADDR, card_table);
5583 amd64_alu_reg_membase (code, X86_ADD, AMD64_RDX, AMD64_RIP, 0);
5585 amd64_mov_membase_imm (code, AMD64_RDX, 0, 1, 1);
5586 x86_patch (br, code);
5589 #ifdef MONO_ARCH_SIMD_INTRINSICS
5590 /* TODO: Some of these IR opcodes are marked as no clobber when they indeed do. */
5592 amd64_sse_addps_reg_reg (code, ins->sreg1, ins->sreg2);
5595 amd64_sse_divps_reg_reg (code, ins->sreg1, ins->sreg2);
5598 amd64_sse_mulps_reg_reg (code, ins->sreg1, ins->sreg2);
5601 amd64_sse_subps_reg_reg (code, ins->sreg1, ins->sreg2);
5604 amd64_sse_maxps_reg_reg (code, ins->sreg1, ins->sreg2);
5607 amd64_sse_minps_reg_reg (code, ins->sreg1, ins->sreg2);
5610 g_assert (ins->inst_c0 >= 0 && ins->inst_c0 <= 7);
5611 amd64_sse_cmpps_reg_reg_imm (code, ins->sreg1, ins->sreg2, ins->inst_c0);
5614 amd64_sse_andps_reg_reg (code, ins->sreg1, ins->sreg2);
5617 amd64_sse_andnps_reg_reg (code, ins->sreg1, ins->sreg2);
5620 amd64_sse_orps_reg_reg (code, ins->sreg1, ins->sreg2);
5623 amd64_sse_xorps_reg_reg (code, ins->sreg1, ins->sreg2);
5626 amd64_sse_sqrtps_reg_reg (code, ins->dreg, ins->sreg1);
5629 amd64_sse_rsqrtps_reg_reg (code, ins->dreg, ins->sreg1);
5632 amd64_sse_rcpps_reg_reg (code, ins->dreg, ins->sreg1);
5635 amd64_sse_addsubps_reg_reg (code, ins->sreg1, ins->sreg2);
5638 amd64_sse_haddps_reg_reg (code, ins->sreg1, ins->sreg2);
5641 amd64_sse_hsubps_reg_reg (code, ins->sreg1, ins->sreg2);
5644 amd64_sse_movshdup_reg_reg (code, ins->dreg, ins->sreg1);
5647 amd64_sse_movsldup_reg_reg (code, ins->dreg, ins->sreg1);
5650 case OP_PSHUFLEW_HIGH:
5651 g_assert (ins->inst_c0 >= 0 && ins->inst_c0 <= 0xFF);
5652 amd64_sse_pshufhw_reg_reg_imm (code, ins->dreg, ins->sreg1, ins->inst_c0);
5654 case OP_PSHUFLEW_LOW:
5655 g_assert (ins->inst_c0 >= 0 && ins->inst_c0 <= 0xFF);
5656 amd64_sse_pshuflw_reg_reg_imm (code, ins->dreg, ins->sreg1, ins->inst_c0);
5659 g_assert (ins->inst_c0 >= 0 && ins->inst_c0 <= 0xFF);
5660 amd64_sse_pshufd_reg_reg_imm (code, ins->dreg, ins->sreg1, ins->inst_c0);
5663 g_assert (ins->inst_c0 >= 0 && ins->inst_c0 <= 0xFF);
5664 amd64_sse_shufps_reg_reg_imm (code, ins->sreg1, ins->sreg2, ins->inst_c0);
5667 g_assert (ins->inst_c0 >= 0 && ins->inst_c0 <= 0x3);
5668 amd64_sse_shufpd_reg_reg_imm (code, ins->sreg1, ins->sreg2, ins->inst_c0);
5672 amd64_sse_addpd_reg_reg (code, ins->sreg1, ins->sreg2);
5675 amd64_sse_divpd_reg_reg (code, ins->sreg1, ins->sreg2);
5678 amd64_sse_mulpd_reg_reg (code, ins->sreg1, ins->sreg2);
5681 amd64_sse_subpd_reg_reg (code, ins->sreg1, ins->sreg2);
5684 amd64_sse_maxpd_reg_reg (code, ins->sreg1, ins->sreg2);
5687 amd64_sse_minpd_reg_reg (code, ins->sreg1, ins->sreg2);
5690 g_assert (ins->inst_c0 >= 0 && ins->inst_c0 <= 7);
5691 amd64_sse_cmppd_reg_reg_imm (code, ins->sreg1, ins->sreg2, ins->inst_c0);
5694 amd64_sse_andpd_reg_reg (code, ins->sreg1, ins->sreg2);
5697 amd64_sse_andnpd_reg_reg (code, ins->sreg1, ins->sreg2);
5700 amd64_sse_orpd_reg_reg (code, ins->sreg1, ins->sreg2);
5703 amd64_sse_xorpd_reg_reg (code, ins->sreg1, ins->sreg2);
5706 amd64_sse_sqrtpd_reg_reg (code, ins->dreg, ins->sreg1);
5709 amd64_sse_addsubpd_reg_reg (code, ins->sreg1, ins->sreg2);
5712 amd64_sse_haddpd_reg_reg (code, ins->sreg1, ins->sreg2);
5715 amd64_sse_hsubpd_reg_reg (code, ins->sreg1, ins->sreg2);
5718 amd64_sse_movddup_reg_reg (code, ins->dreg, ins->sreg1);
5721 case OP_EXTRACT_MASK:
5722 amd64_sse_pmovmskb_reg_reg (code, ins->dreg, ins->sreg1);
5726 amd64_sse_pand_reg_reg (code, ins->sreg1, ins->sreg2);
5729 amd64_sse_por_reg_reg (code, ins->sreg1, ins->sreg2);
5732 amd64_sse_pxor_reg_reg (code, ins->sreg1, ins->sreg2);
5736 amd64_sse_paddb_reg_reg (code, ins->sreg1, ins->sreg2);
5739 amd64_sse_paddw_reg_reg (code, ins->sreg1, ins->sreg2);
5742 amd64_sse_paddd_reg_reg (code, ins->sreg1, ins->sreg2);
5745 amd64_sse_paddq_reg_reg (code, ins->sreg1, ins->sreg2);
5749 amd64_sse_psubb_reg_reg (code, ins->sreg1, ins->sreg2);
5752 amd64_sse_psubw_reg_reg (code, ins->sreg1, ins->sreg2);
5755 amd64_sse_psubd_reg_reg (code, ins->sreg1, ins->sreg2);
5758 amd64_sse_psubq_reg_reg (code, ins->sreg1, ins->sreg2);
5762 amd64_sse_pmaxub_reg_reg (code, ins->sreg1, ins->sreg2);
5765 amd64_sse_pmaxuw_reg_reg (code, ins->sreg1, ins->sreg2);
5768 amd64_sse_pmaxud_reg_reg (code, ins->sreg1, ins->sreg2);
5772 amd64_sse_pmaxsb_reg_reg (code, ins->sreg1, ins->sreg2);
5775 amd64_sse_pmaxsw_reg_reg (code, ins->sreg1, ins->sreg2);
5778 amd64_sse_pmaxsd_reg_reg (code, ins->sreg1, ins->sreg2);
5782 amd64_sse_pavgb_reg_reg (code, ins->sreg1, ins->sreg2);
5785 amd64_sse_pavgw_reg_reg (code, ins->sreg1, ins->sreg2);
5789 amd64_sse_pminub_reg_reg (code, ins->sreg1, ins->sreg2);
5792 amd64_sse_pminuw_reg_reg (code, ins->sreg1, ins->sreg2);
5795 amd64_sse_pminud_reg_reg (code, ins->sreg1, ins->sreg2);
5799 amd64_sse_pminsb_reg_reg (code, ins->sreg1, ins->sreg2);
5802 amd64_sse_pminsw_reg_reg (code, ins->sreg1, ins->sreg2);
5805 amd64_sse_pminsd_reg_reg (code, ins->sreg1, ins->sreg2);
5809 amd64_sse_pcmpeqb_reg_reg (code, ins->sreg1, ins->sreg2);
5812 amd64_sse_pcmpeqw_reg_reg (code, ins->sreg1, ins->sreg2);
5815 amd64_sse_pcmpeqd_reg_reg (code, ins->sreg1, ins->sreg2);
5818 amd64_sse_pcmpeqq_reg_reg (code, ins->sreg1, ins->sreg2);
5822 amd64_sse_pcmpgtb_reg_reg (code, ins->sreg1, ins->sreg2);
5825 amd64_sse_pcmpgtw_reg_reg (code, ins->sreg1, ins->sreg2);
5828 amd64_sse_pcmpgtd_reg_reg (code, ins->sreg1, ins->sreg2);
5831 amd64_sse_pcmpgtq_reg_reg (code, ins->sreg1, ins->sreg2);
5834 case OP_PSUM_ABS_DIFF:
5835 amd64_sse_psadbw_reg_reg (code, ins->sreg1, ins->sreg2);
5838 case OP_UNPACK_LOWB:
5839 amd64_sse_punpcklbw_reg_reg (code, ins->sreg1, ins->sreg2);
5841 case OP_UNPACK_LOWW:
5842 amd64_sse_punpcklwd_reg_reg (code, ins->sreg1, ins->sreg2);
5844 case OP_UNPACK_LOWD:
5845 amd64_sse_punpckldq_reg_reg (code, ins->sreg1, ins->sreg2);
5847 case OP_UNPACK_LOWQ:
5848 amd64_sse_punpcklqdq_reg_reg (code, ins->sreg1, ins->sreg2);
5850 case OP_UNPACK_LOWPS:
5851 amd64_sse_unpcklps_reg_reg (code, ins->sreg1, ins->sreg2);
5853 case OP_UNPACK_LOWPD:
5854 amd64_sse_unpcklpd_reg_reg (code, ins->sreg1, ins->sreg2);
5857 case OP_UNPACK_HIGHB:
5858 amd64_sse_punpckhbw_reg_reg (code, ins->sreg1, ins->sreg2);
5860 case OP_UNPACK_HIGHW:
5861 amd64_sse_punpckhwd_reg_reg (code, ins->sreg1, ins->sreg2);
5863 case OP_UNPACK_HIGHD:
5864 amd64_sse_punpckhdq_reg_reg (code, ins->sreg1, ins->sreg2);
5866 case OP_UNPACK_HIGHQ:
5867 amd64_sse_punpckhqdq_reg_reg (code, ins->sreg1, ins->sreg2);
5869 case OP_UNPACK_HIGHPS:
5870 amd64_sse_unpckhps_reg_reg (code, ins->sreg1, ins->sreg2);
5872 case OP_UNPACK_HIGHPD:
5873 amd64_sse_unpckhpd_reg_reg (code, ins->sreg1, ins->sreg2);
5877 amd64_sse_packsswb_reg_reg (code, ins->sreg1, ins->sreg2);
5880 amd64_sse_packssdw_reg_reg (code, ins->sreg1, ins->sreg2);
5883 amd64_sse_packuswb_reg_reg (code, ins->sreg1, ins->sreg2);
5886 amd64_sse_packusdw_reg_reg (code, ins->sreg1, ins->sreg2);
5889 case OP_PADDB_SAT_UN:
5890 amd64_sse_paddusb_reg_reg (code, ins->sreg1, ins->sreg2);
5892 case OP_PSUBB_SAT_UN:
5893 amd64_sse_psubusb_reg_reg (code, ins->sreg1, ins->sreg2);
5895 case OP_PADDW_SAT_UN:
5896 amd64_sse_paddusw_reg_reg (code, ins->sreg1, ins->sreg2);
5898 case OP_PSUBW_SAT_UN:
5899 amd64_sse_psubusw_reg_reg (code, ins->sreg1, ins->sreg2);
5903 amd64_sse_paddsb_reg_reg (code, ins->sreg1, ins->sreg2);
5906 amd64_sse_psubsb_reg_reg (code, ins->sreg1, ins->sreg2);
5909 amd64_sse_paddsw_reg_reg (code, ins->sreg1, ins->sreg2);
5912 amd64_sse_psubsw_reg_reg (code, ins->sreg1, ins->sreg2);
5916 amd64_sse_pmullw_reg_reg (code, ins->sreg1, ins->sreg2);
5919 amd64_sse_pmulld_reg_reg (code, ins->sreg1, ins->sreg2);
5922 amd64_sse_pmuludq_reg_reg (code, ins->sreg1, ins->sreg2);
5924 case OP_PMULW_HIGH_UN:
5925 amd64_sse_pmulhuw_reg_reg (code, ins->sreg1, ins->sreg2);
5928 amd64_sse_pmulhw_reg_reg (code, ins->sreg1, ins->sreg2);
5932 amd64_sse_psrlw_reg_imm (code, ins->dreg, ins->inst_imm);
5935 amd64_sse_psrlw_reg_reg (code, ins->dreg, ins->sreg2);
5939 amd64_sse_psraw_reg_imm (code, ins->dreg, ins->inst_imm);
5942 amd64_sse_psraw_reg_reg (code, ins->dreg, ins->sreg2);
5946 amd64_sse_psllw_reg_imm (code, ins->dreg, ins->inst_imm);
5949 amd64_sse_psllw_reg_reg (code, ins->dreg, ins->sreg2);
5953 amd64_sse_psrld_reg_imm (code, ins->dreg, ins->inst_imm);
5956 amd64_sse_psrld_reg_reg (code, ins->dreg, ins->sreg2);
5960 amd64_sse_psrad_reg_imm (code, ins->dreg, ins->inst_imm);
5963 amd64_sse_psrad_reg_reg (code, ins->dreg, ins->sreg2);
5967 amd64_sse_pslld_reg_imm (code, ins->dreg, ins->inst_imm);
5970 amd64_sse_pslld_reg_reg (code, ins->dreg, ins->sreg2);
5974 amd64_sse_psrlq_reg_imm (code, ins->dreg, ins->inst_imm);
5977 amd64_sse_psrlq_reg_reg (code, ins->dreg, ins->sreg2);
5980 /*TODO: This is appart of the sse spec but not added
5982 amd64_sse_psraq_reg_imm (code, ins->dreg, ins->inst_imm);
5985 amd64_sse_psraq_reg_reg (code, ins->dreg, ins->sreg2);
5990 amd64_sse_psllq_reg_imm (code, ins->dreg, ins->inst_imm);
5993 amd64_sse_psllq_reg_reg (code, ins->dreg, ins->sreg2);
5996 amd64_sse_cvtdq2pd_reg_reg (code, ins->dreg, ins->sreg1);
5999 amd64_sse_cvtdq2ps_reg_reg (code, ins->dreg, ins->sreg1);
6002 amd64_sse_cvtpd2dq_reg_reg (code, ins->dreg, ins->sreg1);
6005 amd64_sse_cvtpd2ps_reg_reg (code, ins->dreg, ins->sreg1);
6008 amd64_sse_cvtps2dq_reg_reg (code, ins->dreg, ins->sreg1);
6011 amd64_sse_cvtps2pd_reg_reg (code, ins->dreg, ins->sreg1);
6014 amd64_sse_cvttpd2dq_reg_reg (code, ins->dreg, ins->sreg1);
6017 amd64_sse_cvttps2dq_reg_reg (code, ins->dreg, ins->sreg1);
6021 amd64_movd_xreg_reg_size (code, ins->dreg, ins->sreg1, 4);
6024 amd64_movd_reg_xreg_size (code, ins->dreg, ins->sreg1, 4);
6028 amd64_movhlps_reg_reg (code, AMD64_XMM15, ins->sreg1);
6029 amd64_movd_reg_xreg_size (code, ins->dreg, AMD64_XMM15, 8);
6031 amd64_movd_reg_xreg_size (code, ins->dreg, ins->sreg1, 8);
6036 amd64_movd_reg_xreg_size (code, ins->dreg, ins->sreg1, 4);
6038 amd64_shift_reg_imm (code, X86_SHR, ins->dreg, ins->inst_c0 * 8);
6039 amd64_widen_reg (code, ins->dreg, ins->dreg, ins->opcode == OP_EXTRACT_I1, FALSE);
6043 /*amd64_movd_reg_xreg_size (code, ins->dreg, ins->sreg1, 4);
6045 amd64_shift_reg_imm_size (code, X86_SHR, ins->dreg, 16, 4);*/
6046 amd64_sse_pextrw_reg_reg_imm (code, ins->dreg, ins->sreg1, ins->inst_c0);
6047 amd64_widen_reg_size (code, ins->dreg, ins->dreg, ins->opcode == OP_EXTRACT_I2, TRUE, 4);
6051 amd64_movhlps_reg_reg (code, ins->dreg, ins->sreg1);
6053 amd64_sse_movsd_reg_reg (code, ins->dreg, ins->sreg1);
6056 amd64_sse_pinsrw_reg_reg_imm (code, ins->sreg1, ins->sreg2, ins->inst_c0);
6058 case OP_EXTRACTX_U2:
6059 amd64_sse_pextrw_reg_reg_imm (code, ins->dreg, ins->sreg1, ins->inst_c0);
6061 case OP_INSERTX_U1_SLOW:
6062 /*sreg1 is the extracted ireg (scratch)
6063 /sreg2 is the to be inserted ireg (scratch)
6064 /dreg is the xreg to receive the value*/
6066 /*clear the bits from the extracted word*/
6067 amd64_alu_reg_imm (code, X86_AND, ins->sreg1, ins->inst_c0 & 1 ? 0x00FF : 0xFF00);
6068 /*shift the value to insert if needed*/
6069 if (ins->inst_c0 & 1)
6070 amd64_shift_reg_imm_size (code, X86_SHL, ins->sreg2, 8, 4);
6071 /*join them together*/
6072 amd64_alu_reg_reg (code, X86_OR, ins->sreg1, ins->sreg2);
6073 amd64_sse_pinsrw_reg_reg_imm (code, ins->dreg, ins->sreg1, ins->inst_c0 / 2);
6075 case OP_INSERTX_I4_SLOW:
6076 amd64_sse_pinsrw_reg_reg_imm (code, ins->dreg, ins->sreg2, ins->inst_c0 * 2);
6077 amd64_shift_reg_imm (code, X86_SHR, ins->sreg2, 16);
6078 amd64_sse_pinsrw_reg_reg_imm (code, ins->dreg, ins->sreg2, ins->inst_c0 * 2 + 1);
6080 case OP_INSERTX_I8_SLOW:
6081 amd64_movd_xreg_reg_size(code, AMD64_XMM15, ins->sreg2, 8);
6083 amd64_movlhps_reg_reg (code, ins->dreg, AMD64_XMM15);
6085 amd64_sse_movsd_reg_reg (code, ins->dreg, AMD64_XMM15);
6088 case OP_INSERTX_R4_SLOW:
6089 switch (ins->inst_c0) {
6091 amd64_sse_cvtsd2ss_reg_reg (code, ins->dreg, ins->sreg2);
6094 amd64_sse_pshufd_reg_reg_imm (code, ins->dreg, ins->dreg, mono_simd_shuffle_mask(1, 0, 2, 3));
6095 amd64_sse_cvtsd2ss_reg_reg (code, ins->dreg, ins->sreg2);
6096 amd64_sse_pshufd_reg_reg_imm (code, ins->dreg, ins->dreg, mono_simd_shuffle_mask(1, 0, 2, 3));
6099 amd64_sse_pshufd_reg_reg_imm (code, ins->dreg, ins->dreg, mono_simd_shuffle_mask(2, 1, 0, 3));
6100 amd64_sse_cvtsd2ss_reg_reg (code, ins->dreg, ins->sreg2);
6101 amd64_sse_pshufd_reg_reg_imm (code, ins->dreg, ins->dreg, mono_simd_shuffle_mask(2, 1, 0, 3));
6104 amd64_sse_pshufd_reg_reg_imm (code, ins->dreg, ins->dreg, mono_simd_shuffle_mask(3, 1, 2, 0));
6105 amd64_sse_cvtsd2ss_reg_reg (code, ins->dreg, ins->sreg2);
6106 amd64_sse_pshufd_reg_reg_imm (code, ins->dreg, ins->dreg, mono_simd_shuffle_mask(3, 1, 2, 0));
6110 case OP_INSERTX_R8_SLOW:
6112 amd64_movlhps_reg_reg (code, ins->dreg, ins->sreg2);
6114 amd64_sse_movsd_reg_reg (code, ins->dreg, ins->sreg2);
6116 case OP_STOREX_MEMBASE_REG:
6117 case OP_STOREX_MEMBASE:
6118 amd64_sse_movups_membase_reg (code, ins->dreg, ins->inst_offset, ins->sreg1);
6120 case OP_LOADX_MEMBASE:
6121 amd64_sse_movups_reg_membase (code, ins->dreg, ins->sreg1, ins->inst_offset);
6123 case OP_LOADX_ALIGNED_MEMBASE:
6124 amd64_sse_movaps_reg_membase (code, ins->dreg, ins->sreg1, ins->inst_offset);
6126 case OP_STOREX_ALIGNED_MEMBASE_REG:
6127 amd64_sse_movaps_membase_reg (code, ins->dreg, ins->inst_offset, ins->sreg1);
6129 case OP_STOREX_NTA_MEMBASE_REG:
6130 amd64_sse_movntps_reg_membase (code, ins->dreg, ins->sreg1, ins->inst_offset);
6132 case OP_PREFETCH_MEMBASE:
6133 amd64_sse_prefetch_reg_membase (code, ins->backend.arg_info, ins->sreg1, ins->inst_offset);
6137 /*FIXME the peephole pass should have killed this*/
6138 if (ins->dreg != ins->sreg1)
6139 amd64_sse_movaps_reg_reg (code, ins->dreg, ins->sreg1);
6142 amd64_sse_pxor_reg_reg (code, ins->dreg, ins->dreg);
6144 case OP_ICONV_TO_R8_RAW:
6145 amd64_movd_xreg_reg_size (code, ins->dreg, ins->sreg1, 4);
6146 amd64_sse_cvtss2sd_reg_reg (code, ins->dreg, ins->dreg);
6149 case OP_FCONV_TO_R8_X:
6150 amd64_sse_movsd_reg_reg (code, ins->dreg, ins->sreg1);
6153 case OP_XCONV_R8_TO_I4:
6154 amd64_sse_cvttsd2si_reg_xreg_size (code, ins->dreg, ins->sreg1, 4);
6155 switch (ins->backend.source_opcode) {
6156 case OP_FCONV_TO_I1:
6157 amd64_widen_reg (code, ins->dreg, ins->dreg, TRUE, FALSE);
6159 case OP_FCONV_TO_U1:
6160 amd64_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
6162 case OP_FCONV_TO_I2:
6163 amd64_widen_reg (code, ins->dreg, ins->dreg, TRUE, TRUE);
6165 case OP_FCONV_TO_U2:
6166 amd64_widen_reg (code, ins->dreg, ins->dreg, FALSE, TRUE);
6172 amd64_sse_pinsrw_reg_reg_imm (code, ins->dreg, ins->sreg1, 0);
6173 amd64_sse_pinsrw_reg_reg_imm (code, ins->dreg, ins->sreg1, 1);
6174 amd64_sse_pshufd_reg_reg_imm (code, ins->dreg, ins->dreg, 0);
6177 amd64_movd_xreg_reg_size (code, ins->dreg, ins->sreg1, 4);
6178 amd64_sse_pshufd_reg_reg_imm (code, ins->dreg, ins->dreg, 0);
6181 amd64_movd_xreg_reg_size (code, ins->dreg, ins->sreg1, 8);
6182 amd64_sse_pshufd_reg_reg_imm (code, ins->dreg, ins->dreg, 0x44);
6185 amd64_sse_movsd_reg_reg (code, ins->dreg, ins->sreg1);
6186 amd64_sse_cvtsd2ss_reg_reg (code, ins->dreg, ins->dreg);
6187 amd64_sse_pshufd_reg_reg_imm (code, ins->dreg, ins->dreg, 0);
6190 amd64_sse_movsd_reg_reg (code, ins->dreg, ins->sreg1);
6191 amd64_sse_pshufd_reg_reg_imm (code, ins->dreg, ins->dreg, 0x44);
6194 case OP_LIVERANGE_START: {
6195 if (cfg->verbose_level > 1)
6196 printf ("R%d START=0x%x\n", MONO_VARINFO (cfg, ins->inst_c0)->vreg, (int)(code - cfg->native_code));
6197 MONO_VARINFO (cfg, ins->inst_c0)->live_range_start = code - cfg->native_code;
6200 case OP_LIVERANGE_END: {
6201 if (cfg->verbose_level > 1)
6202 printf ("R%d END=0x%x\n", MONO_VARINFO (cfg, ins->inst_c0)->vreg, (int)(code - cfg->native_code));
6203 MONO_VARINFO (cfg, ins->inst_c0)->live_range_end = code - cfg->native_code;
6206 case OP_NACL_GC_SAFE_POINT: {
6207 #if defined(__native_client_codegen__)
6208 code = emit_call (cfg, code, MONO_PATCH_INFO_ABS, (gpointer)mono_nacl_gc, TRUE);
6212 case OP_GC_LIVENESS_DEF:
6213 case OP_GC_LIVENESS_USE:
6214 case OP_GC_PARAM_SLOT_LIVENESS_DEF:
6215 ins->backend.pc_offset = code - cfg->native_code;
6217 case OP_GC_SPILL_SLOT_LIVENESS_DEF:
6218 ins->backend.pc_offset = code - cfg->native_code;
6219 bb->spill_slot_defs = g_slist_prepend_mempool (cfg->mempool, bb->spill_slot_defs, ins);
6222 g_warning ("unknown opcode %s in %s()\n", mono_inst_name (ins->opcode), __FUNCTION__);
6223 g_assert_not_reached ();
6226 if ((code - cfg->native_code - offset) > max_len) {
6227 #if !defined(__native_client_codegen__)
6228 g_warning ("wrong maximal instruction length of instruction %s (expected %d, got %ld)",
6229 mono_inst_name (ins->opcode), max_len, code - cfg->native_code - offset);
6230 g_assert_not_reached ();
6235 last_offset = offset;
6238 cfg->code_len = code - cfg->native_code;
6241 #endif /* DISABLE_JIT */
6244 mono_arch_register_lowlevel_calls (void)
6246 /* The signature doesn't matter */
6247 mono_register_jit_icall (mono_amd64_throw_exception, "mono_amd64_throw_exception", mono_create_icall_signature ("void"), TRUE);
6251 mono_arch_patch_code (MonoMethod *method, MonoDomain *domain, guint8 *code, MonoJumpInfo *ji, gboolean run_cctors)
6253 MonoJumpInfo *patch_info;
6254 gboolean compile_aot = !run_cctors;
6256 for (patch_info = ji; patch_info; patch_info = patch_info->next) {
6257 unsigned char *ip = patch_info->ip.i + code;
6258 unsigned char *target;
6260 target = mono_resolve_patch_target (method, domain, code, patch_info, run_cctors);
6263 switch (patch_info->type) {
6264 case MONO_PATCH_INFO_BB:
6265 case MONO_PATCH_INFO_LABEL:
6268 /* No need to patch these */
6273 switch (patch_info->type) {
6274 case MONO_PATCH_INFO_NONE:
6276 case MONO_PATCH_INFO_METHOD_REL:
6277 case MONO_PATCH_INFO_R8:
6278 case MONO_PATCH_INFO_R4:
6279 g_assert_not_reached ();
6281 case MONO_PATCH_INFO_BB:
6288 * Debug code to help track down problems where the target of a near call is
6291 if (amd64_is_near_call (ip)) {
6292 gint64 disp = (guint8*)target - (guint8*)ip;
6294 if (!amd64_is_imm32 (disp)) {
6295 printf ("TYPE: %d\n", patch_info->type);
6296 switch (patch_info->type) {
6297 case MONO_PATCH_INFO_INTERNAL_METHOD:
6298 printf ("V: %s\n", patch_info->data.name);
6300 case MONO_PATCH_INFO_METHOD_JUMP:
6301 case MONO_PATCH_INFO_METHOD:
6302 printf ("V: %s\n", patch_info->data.method->name);
6310 amd64_patch (ip, (gpointer)target);
6317 get_max_epilog_size (MonoCompile *cfg)
6319 int max_epilog_size = 16;
6321 if (cfg->method->save_lmf)
6322 max_epilog_size += 256;
6324 if (mono_jit_trace_calls != NULL)
6325 max_epilog_size += 50;
6327 if (cfg->prof_options & MONO_PROFILE_ENTER_LEAVE)
6328 max_epilog_size += 50;
6330 max_epilog_size += (AMD64_NREG * 2);
6332 return max_epilog_size;
6336 * This macro is used for testing whenever the unwinder works correctly at every point
6337 * where an async exception can happen.
6339 /* This will generate a SIGSEGV at the given point in the code */
6340 #define async_exc_point(code) do { \
6341 if (mono_inject_async_exc_method && mono_method_desc_full_match (mono_inject_async_exc_method, cfg->method)) { \
6342 if (cfg->arch.async_point_count == mono_inject_async_exc_pos) \
6343 amd64_mov_reg_mem (code, AMD64_RAX, 0, 4); \
6344 cfg->arch.async_point_count ++; \
6349 mono_arch_emit_prolog (MonoCompile *cfg)
6351 MonoMethod *method = cfg->method;
6353 MonoMethodSignature *sig;
6355 int alloc_size, pos, i, cfa_offset, quad, max_epilog_size;
6358 gint32 lmf_offset = cfg->arch.lmf_offset;
6359 gboolean args_clobbered = FALSE;
6360 gboolean trace = FALSE;
6361 #ifdef __native_client_codegen__
6362 guint alignment_check;
6365 cfg->code_size = MAX (cfg->header->code_size * 4, 10240);
6367 #if defined(__default_codegen__)
6368 code = cfg->native_code = g_malloc (cfg->code_size);
6369 #elif defined(__native_client_codegen__)
6370 /* native_code_alloc is not 32-byte aligned, native_code is. */
6371 cfg->native_code_alloc = g_malloc (cfg->code_size + kNaClAlignment);
6373 /* Align native_code to next nearest kNaclAlignment byte. */
6374 cfg->native_code = (uintptr_t)cfg->native_code_alloc + kNaClAlignment;
6375 cfg->native_code = (uintptr_t)cfg->native_code & ~kNaClAlignmentMask;
6377 code = cfg->native_code;
6379 alignment_check = (guint)cfg->native_code & kNaClAlignmentMask;
6380 g_assert (alignment_check == 0);
6383 if (mono_jit_trace_calls != NULL && mono_trace_eval (method))
6386 /* Amount of stack space allocated by register saving code */
6389 /* Offset between RSP and the CFA */
6393 * The prolog consists of the following parts:
6395 * - push rbp, mov rbp, rsp
6396 * - save callee saved regs using pushes
6398 * - save rgctx if needed
6399 * - save lmf if needed
6402 * - save rgctx if needed
6403 * - save lmf if needed
6404 * - save callee saved regs using moves
6409 mono_emit_unwind_op_def_cfa (cfg, code, AMD64_RSP, 8);
6410 // IP saved at CFA - 8
6411 mono_emit_unwind_op_offset (cfg, code, AMD64_RIP, -cfa_offset);
6412 async_exc_point (code);
6413 mini_gc_set_slot_type_from_cfa (cfg, -cfa_offset, SLOT_NOREF);
6415 if (!cfg->arch.omit_fp) {
6416 amd64_push_reg (code, AMD64_RBP);
6418 mono_emit_unwind_op_def_cfa_offset (cfg, code, cfa_offset);
6419 mono_emit_unwind_op_offset (cfg, code, AMD64_RBP, - cfa_offset);
6420 async_exc_point (code);
6422 mono_arch_unwindinfo_add_push_nonvol (&cfg->arch.unwindinfo, cfg->native_code, code, AMD64_RBP);
6424 /* These are handled automatically by the stack marking code */
6425 mini_gc_set_slot_type_from_cfa (cfg, -cfa_offset, SLOT_NOREF);
6427 amd64_mov_reg_reg (code, AMD64_RBP, AMD64_RSP, sizeof(mgreg_t));
6428 mono_emit_unwind_op_def_cfa_reg (cfg, code, AMD64_RBP);
6429 async_exc_point (code);
6431 mono_arch_unwindinfo_add_set_fpreg (&cfg->arch.unwindinfo, cfg->native_code, code, AMD64_RBP);
6435 /* Save callee saved registers */
6436 if (!cfg->arch.omit_fp && !method->save_lmf) {
6437 int offset = cfa_offset;
6439 for (i = 0; i < AMD64_NREG; ++i)
6440 if (AMD64_IS_CALLEE_SAVED_REG (i) && (cfg->used_int_regs & (1 << i))) {
6441 amd64_push_reg (code, i);
6442 pos += 8; /* AMD64 push inst is always 8 bytes, no way to change it */
6444 mono_emit_unwind_op_offset (cfg, code, i, - offset);
6445 async_exc_point (code);
6447 /* These are handled automatically by the stack marking code */
6448 mini_gc_set_slot_type_from_cfa (cfg, - offset, SLOT_NOREF);
6452 /* The param area is always at offset 0 from sp */
6453 /* This needs to be allocated here, since it has to come after the spill area */
6454 if (cfg->arch.no_pushes && cfg->param_area) {
6455 if (cfg->arch.omit_fp)
6457 g_assert_not_reached ();
6458 cfg->stack_offset += ALIGN_TO (cfg->param_area, sizeof(mgreg_t));
6461 if (cfg->arch.omit_fp) {
6463 * On enter, the stack is misaligned by the pushing of the return
6464 * address. It is either made aligned by the pushing of %rbp, or by
6467 alloc_size = ALIGN_TO (cfg->stack_offset, 8);
6468 if ((alloc_size % 16) == 0) {
6470 /* Mark the padding slot as NOREF */
6471 mini_gc_set_slot_type_from_cfa (cfg, -cfa_offset - sizeof (mgreg_t), SLOT_NOREF);
6474 alloc_size = ALIGN_TO (cfg->stack_offset, MONO_ARCH_FRAME_ALIGNMENT);
6475 if (cfg->stack_offset != alloc_size) {
6476 /* Mark the padding slot as NOREF */
6477 mini_gc_set_slot_type_from_fp (cfg, -alloc_size + cfg->param_area, SLOT_NOREF);
6479 cfg->arch.sp_fp_offset = alloc_size;
6483 cfg->arch.stack_alloc_size = alloc_size;
6485 /* Allocate stack frame */
6487 /* See mono_emit_stack_alloc */
6488 #if defined(HOST_WIN32) || defined(MONO_ARCH_SIGSEGV_ON_ALTSTACK)
6489 guint32 remaining_size = alloc_size;
6490 /*FIXME handle unbounded code expansion, we should use a loop in case of more than X interactions*/
6491 guint32 required_code_size = ((remaining_size / 0x1000) + 1) * 10; /*10 is the max size of amd64_alu_reg_imm + amd64_test_membase_reg*/
6492 guint32 offset = code - cfg->native_code;
6493 if (G_UNLIKELY (required_code_size >= (cfg->code_size - offset))) {
6494 while (required_code_size >= (cfg->code_size - offset))
6495 cfg->code_size *= 2;
6496 cfg->native_code = mono_realloc_native_code (cfg);
6497 code = cfg->native_code + offset;
6498 mono_jit_stats.code_reallocs++;
6501 while (remaining_size >= 0x1000) {
6502 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, 0x1000);
6503 if (cfg->arch.omit_fp) {
6504 cfa_offset += 0x1000;
6505 mono_emit_unwind_op_def_cfa_offset (cfg, code, cfa_offset);
6507 async_exc_point (code);
6509 if (cfg->arch.omit_fp)
6510 mono_arch_unwindinfo_add_alloc_stack (&cfg->arch.unwindinfo, cfg->native_code, code, 0x1000);
6513 amd64_test_membase_reg (code, AMD64_RSP, 0, AMD64_RSP);
6514 remaining_size -= 0x1000;
6516 if (remaining_size) {
6517 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, remaining_size);
6518 if (cfg->arch.omit_fp) {
6519 cfa_offset += remaining_size;
6520 mono_emit_unwind_op_def_cfa_offset (cfg, code, cfa_offset);
6521 async_exc_point (code);
6524 if (cfg->arch.omit_fp)
6525 mono_arch_unwindinfo_add_alloc_stack (&cfg->arch.unwindinfo, cfg->native_code, code, remaining_size);
6529 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, alloc_size);
6530 if (cfg->arch.omit_fp) {
6531 cfa_offset += alloc_size;
6532 mono_emit_unwind_op_def_cfa_offset (cfg, code, cfa_offset);
6533 async_exc_point (code);
6538 /* Stack alignment check */
6541 amd64_mov_reg_reg (code, AMD64_RAX, AMD64_RSP, 8);
6542 amd64_alu_reg_imm (code, X86_AND, AMD64_RAX, 0xf);
6543 amd64_alu_reg_imm (code, X86_CMP, AMD64_RAX, 0);
6544 x86_branch8 (code, X86_CC_EQ, 2, FALSE);
6545 amd64_breakpoint (code);
6549 #ifndef TARGET_WIN32
6550 if (mini_get_debug_options ()->init_stacks) {
6551 /* Fill the stack frame with a dummy value to force deterministic behavior */
6553 /* Save registers to the red zone */
6554 amd64_mov_membase_reg (code, AMD64_RSP, -8, AMD64_RDI, 8);
6555 amd64_mov_membase_reg (code, AMD64_RSP, -16, AMD64_RCX, 8);
6557 amd64_mov_reg_imm (code, AMD64_RAX, 0x2a2a2a2a2a2a2a2a);
6558 amd64_mov_reg_imm (code, AMD64_RCX, alloc_size / 8);
6559 amd64_mov_reg_reg (code, AMD64_RDI, AMD64_RSP, 8);
6562 #if defined(__default_codegen__)
6563 amd64_prefix (code, X86_REP_PREFIX);
6565 #elif defined(__native_client_codegen__)
6566 /* NaCl stos pseudo-instruction */
6567 amd64_codegen_pre (code);
6568 /* First, clear the upper 32 bits of RDI (mov %edi, %edi) */
6569 amd64_mov_reg_reg (code, AMD64_RDI, AMD64_RDI, 4);
6570 /* Add %r15 to %rdi using lea, condition flags unaffected. */
6571 amd64_lea_memindex_size (code, AMD64_RDI, AMD64_R15, 0, AMD64_RDI, 0, 8);
6572 amd64_prefix (code, X86_REP_PREFIX);
6574 amd64_codegen_post (code);
6575 #endif /* __native_client_codegen__ */
6577 amd64_mov_reg_membase (code, AMD64_RDI, AMD64_RSP, -8, 8);
6578 amd64_mov_reg_membase (code, AMD64_RCX, AMD64_RSP, -16, 8);
6583 if (method->save_lmf) {
6585 * The ip field is not set, the exception handling code will obtain it from the stack location pointed to by the sp field.
6588 * sp is saved right before calls but we need to save it here too so
6589 * async stack walks would work.
6591 amd64_mov_membase_reg (code, cfg->frame_reg, cfg->arch.lmf_offset + G_STRUCT_OFFSET (MonoLMF, rsp), AMD64_RSP, 8);
6592 /* Skip method (only needed for trampoline LMF frames) */
6593 /* Save callee saved regs */
6594 for (i = 0; i < MONO_MAX_IREGS; ++i) {
6598 case AMD64_RBX: offset = G_STRUCT_OFFSET (MonoLMF, rbx); break;
6599 case AMD64_RBP: offset = G_STRUCT_OFFSET (MonoLMF, rbp); break;
6600 case AMD64_R12: offset = G_STRUCT_OFFSET (MonoLMF, r12); break;
6601 case AMD64_R13: offset = G_STRUCT_OFFSET (MonoLMF, r13); break;
6602 case AMD64_R14: offset = G_STRUCT_OFFSET (MonoLMF, r14); break;
6603 #ifndef __native_client_codegen__
6604 case AMD64_R15: offset = G_STRUCT_OFFSET (MonoLMF, r15); break;
6607 case AMD64_RDI: offset = G_STRUCT_OFFSET (MonoLMF, rdi); break;
6608 case AMD64_RSI: offset = G_STRUCT_OFFSET (MonoLMF, rsi); break;
6616 amd64_mov_membase_reg (code, cfg->frame_reg, lmf_offset + offset, i, 8);
6617 if (cfg->arch.omit_fp || (i != AMD64_RBP))
6618 mono_emit_unwind_op_offset (cfg, code, i, - (cfa_offset - (lmf_offset + offset)));
6622 /* These can't contain refs */
6623 mini_gc_set_slot_type_from_fp (cfg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, previous_lmf), SLOT_NOREF);
6624 mini_gc_set_slot_type_from_fp (cfg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, lmf_addr), SLOT_NOREF);
6625 mini_gc_set_slot_type_from_fp (cfg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, method), SLOT_NOREF);
6626 mini_gc_set_slot_type_from_fp (cfg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rip), SLOT_NOREF);
6627 mini_gc_set_slot_type_from_fp (cfg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rsp), SLOT_NOREF);
6629 /* These are handled automatically by the stack marking code */
6630 mini_gc_set_slot_type_from_fp (cfg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rbx), SLOT_NOREF);
6631 mini_gc_set_slot_type_from_fp (cfg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rbp), SLOT_NOREF);
6632 mini_gc_set_slot_type_from_fp (cfg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r12), SLOT_NOREF);
6633 mini_gc_set_slot_type_from_fp (cfg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r13), SLOT_NOREF);
6634 mini_gc_set_slot_type_from_fp (cfg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r14), SLOT_NOREF);
6635 mini_gc_set_slot_type_from_fp (cfg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r15), SLOT_NOREF);
6637 mini_gc_set_slot_type_from_fp (cfg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rdi), SLOT_NOREF);
6638 mini_gc_set_slot_type_from_fp (cfg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rsi), SLOT_NOREF);
6643 /* Save callee saved registers */
6644 if (cfg->arch.omit_fp && !method->save_lmf) {
6645 gint32 save_area_offset = cfg->arch.reg_save_area_offset;
6647 /* Save caller saved registers after sp is adjusted */
6648 /* The registers are saved at the bottom of the frame */
6649 /* FIXME: Optimize this so the regs are saved at the end of the frame in increasing order */
6650 for (i = 0; i < AMD64_NREG; ++i)
6651 if (AMD64_IS_CALLEE_SAVED_REG (i) && (cfg->used_int_regs & (1 << i))) {
6652 amd64_mov_membase_reg (code, AMD64_RSP, save_area_offset, i, 8);
6653 mono_emit_unwind_op_offset (cfg, code, i, - (cfa_offset - save_area_offset));
6655 /* These are handled automatically by the stack marking code */
6656 mini_gc_set_slot_type_from_cfa (cfg, - (cfa_offset - save_area_offset), SLOT_NOREF);
6658 save_area_offset += 8;
6659 async_exc_point (code);
6663 /* store runtime generic context */
6664 if (cfg->rgctx_var) {
6665 g_assert (cfg->rgctx_var->opcode == OP_REGOFFSET &&
6666 (cfg->rgctx_var->inst_basereg == AMD64_RBP || cfg->rgctx_var->inst_basereg == AMD64_RSP));
6668 amd64_mov_membase_reg (code, cfg->rgctx_var->inst_basereg, cfg->rgctx_var->inst_offset, MONO_ARCH_RGCTX_REG, sizeof(gpointer));
6671 /* compute max_length in order to use short forward jumps */
6672 max_epilog_size = get_max_epilog_size (cfg);
6673 if (cfg->opt & MONO_OPT_BRANCH) {
6674 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
6678 if (cfg->prof_options & MONO_PROFILE_COVERAGE)
6680 /* max alignment for loops */
6681 if ((cfg->opt & MONO_OPT_LOOP) && bb_is_loop_start (bb))
6682 max_length += LOOP_ALIGNMENT;
6683 #ifdef __native_client_codegen__
6684 /* max alignment for native client */
6685 max_length += kNaClAlignment;
6688 MONO_BB_FOR_EACH_INS (bb, ins) {
6689 #ifdef __native_client_codegen__
6691 int space_in_block = kNaClAlignment -
6692 ((max_length + cfg->code_len) & kNaClAlignmentMask);
6693 int max_len = ((guint8 *)ins_get_spec (ins->opcode))[MONO_INST_LEN];
6694 if (space_in_block < max_len && max_len < kNaClAlignment) {
6695 max_length += space_in_block;
6698 #endif /*__native_client_codegen__*/
6699 max_length += ((guint8 *)ins_get_spec (ins->opcode))[MONO_INST_LEN];
6702 /* Take prolog and epilog instrumentation into account */
6703 if (bb == cfg->bb_entry || bb == cfg->bb_exit)
6704 max_length += max_epilog_size;
6706 bb->max_length = max_length;
6710 sig = mono_method_signature (method);
6713 cinfo = cfg->arch.cinfo;
6715 if (sig->ret->type != MONO_TYPE_VOID) {
6716 /* Save volatile arguments to the stack */
6717 if (cfg->vret_addr && (cfg->vret_addr->opcode != OP_REGVAR))
6718 amd64_mov_membase_reg (code, cfg->vret_addr->inst_basereg, cfg->vret_addr->inst_offset, cinfo->ret.reg, 8);
6721 /* Keep this in sync with emit_load_volatile_arguments */
6722 for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
6723 ArgInfo *ainfo = cinfo->args + i;
6724 gint32 stack_offset;
6727 ins = cfg->args [i];
6729 if ((ins->flags & MONO_INST_IS_DEAD) && !trace)
6730 /* Unused arguments */
6733 if (sig->hasthis && (i == 0))
6734 arg_type = &mono_defaults.object_class->byval_arg;
6736 arg_type = sig->params [i - sig->hasthis];
6738 stack_offset = ainfo->offset + ARGS_OFFSET;
6740 if (cfg->globalra) {
6741 /* All the other moves are done by the register allocator */
6742 switch (ainfo->storage) {
6743 case ArgInFloatSSEReg:
6744 amd64_sse_cvtss2sd_reg_reg (code, ainfo->reg, ainfo->reg);
6746 case ArgValuetypeInReg:
6747 for (quad = 0; quad < 2; quad ++) {
6748 switch (ainfo->pair_storage [quad]) {
6750 amd64_mov_membase_reg (code, ins->inst_basereg, ins->inst_offset + (quad * sizeof(mgreg_t)), ainfo->pair_regs [quad], sizeof(mgreg_t));
6752 case ArgInFloatSSEReg:
6753 amd64_movss_membase_reg (code, ins->inst_basereg, ins->inst_offset + (quad * sizeof(mgreg_t)), ainfo->pair_regs [quad]);
6755 case ArgInDoubleSSEReg:
6756 amd64_movsd_membase_reg (code, ins->inst_basereg, ins->inst_offset + (quad * sizeof(mgreg_t)), ainfo->pair_regs [quad]);
6761 g_assert_not_reached ();
6772 /* Save volatile arguments to the stack */
6773 if (ins->opcode != OP_REGVAR) {
6774 switch (ainfo->storage) {
6780 if (stack_offset & 0x1)
6782 else if (stack_offset & 0x2)
6784 else if (stack_offset & 0x4)
6789 amd64_mov_membase_reg (code, ins->inst_basereg, ins->inst_offset, ainfo->reg, size);
6792 case ArgInFloatSSEReg:
6793 amd64_movss_membase_reg (code, ins->inst_basereg, ins->inst_offset, ainfo->reg);
6795 case ArgInDoubleSSEReg:
6796 amd64_movsd_membase_reg (code, ins->inst_basereg, ins->inst_offset, ainfo->reg);
6798 case ArgValuetypeInReg:
6799 for (quad = 0; quad < 2; quad ++) {
6800 switch (ainfo->pair_storage [quad]) {
6802 amd64_mov_membase_reg (code, ins->inst_basereg, ins->inst_offset + (quad * sizeof(mgreg_t)), ainfo->pair_regs [quad], sizeof(mgreg_t));
6804 case ArgInFloatSSEReg:
6805 amd64_movss_membase_reg (code, ins->inst_basereg, ins->inst_offset + (quad * sizeof(mgreg_t)), ainfo->pair_regs [quad]);
6807 case ArgInDoubleSSEReg:
6808 amd64_movsd_membase_reg (code, ins->inst_basereg, ins->inst_offset + (quad * sizeof(mgreg_t)), ainfo->pair_regs [quad]);
6813 g_assert_not_reached ();
6817 case ArgValuetypeAddrInIReg:
6818 if (ainfo->pair_storage [0] == ArgInIReg)
6819 amd64_mov_membase_reg (code, ins->inst_left->inst_basereg, ins->inst_left->inst_offset, ainfo->pair_regs [0], sizeof (gpointer));
6825 /* Argument allocated to (non-volatile) register */
6826 switch (ainfo->storage) {
6828 amd64_mov_reg_reg (code, ins->dreg, ainfo->reg, 8);
6831 amd64_mov_reg_membase (code, ins->dreg, AMD64_RBP, ARGS_OFFSET + ainfo->offset, 8);
6834 g_assert_not_reached ();
6839 /* Might need to attach the thread to the JIT or change the domain for the callback */
6840 if (method->wrapper_type == MONO_WRAPPER_NATIVE_TO_MANAGED) {
6841 guint64 domain = (guint64)cfg->domain;
6843 args_clobbered = TRUE;
6846 * The call might clobber argument registers, but they are already
6847 * saved to the stack/global regs.
6849 if (appdomain_tls_offset != -1 && lmf_tls_offset != -1) {
6850 guint8 *buf, *no_domain_branch;
6852 code = mono_amd64_emit_tls_get (code, AMD64_RAX, appdomain_tls_offset);
6853 if (cfg->compile_aot) {
6854 /* AOT code is only used in the root domain */
6855 amd64_mov_reg_imm (code, AMD64_ARG_REG1, 0);
6857 if ((domain >> 32) == 0)
6858 amd64_mov_reg_imm_size (code, AMD64_ARG_REG1, domain, 4);
6860 amd64_mov_reg_imm_size (code, AMD64_ARG_REG1, domain, 8);
6862 amd64_alu_reg_reg (code, X86_CMP, AMD64_RAX, AMD64_ARG_REG1);
6863 no_domain_branch = code;
6864 x86_branch8 (code, X86_CC_NE, 0, 0);
6865 code = mono_amd64_emit_tls_get ( code, AMD64_RAX, lmf_addr_tls_offset);
6866 amd64_test_reg_reg (code, AMD64_RAX, AMD64_RAX);
6868 x86_branch8 (code, X86_CC_NE, 0, 0);
6869 amd64_patch (no_domain_branch, code);
6870 code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD,
6871 (gpointer)"mono_jit_thread_attach", TRUE);
6872 amd64_patch (buf, code);
6874 /* The TLS key actually contains a pointer to the MonoJitTlsData structure */
6875 /* FIXME: Add a separate key for LMF to avoid this */
6876 amd64_alu_reg_imm (code, X86_ADD, AMD64_RAX, G_STRUCT_OFFSET (MonoJitTlsData, lmf));
6879 g_assert (!cfg->compile_aot);
6880 if (cfg->compile_aot) {
6881 /* AOT code is only used in the root domain */
6882 amd64_mov_reg_imm (code, AMD64_ARG_REG1, 0);
6884 if ((domain >> 32) == 0)
6885 amd64_mov_reg_imm_size (code, AMD64_ARG_REG1, domain, 4);
6887 amd64_mov_reg_imm_size (code, AMD64_ARG_REG1, domain, 8);
6889 code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD,
6890 (gpointer)"mono_jit_thread_attach", TRUE);
6894 if (method->save_lmf) {
6895 if ((lmf_tls_offset != -1) && !optimize_for_xen) {
6897 * Optimized version which uses the mono_lmf TLS variable instead of
6898 * indirection through the mono_lmf_addr TLS variable.
6900 /* %rax = previous_lmf */
6901 x86_prefix (code, X86_FS_PREFIX);
6902 amd64_mov_reg_mem (code, AMD64_RAX, lmf_tls_offset, 8);
6904 /* Save previous_lmf */
6905 amd64_mov_membase_reg (code, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, previous_lmf), AMD64_RAX, 8);
6907 if (lmf_offset == 0) {
6908 x86_prefix (code, X86_FS_PREFIX);
6909 amd64_mov_mem_reg (code, lmf_tls_offset, cfg->frame_reg, 8);
6911 amd64_lea_membase (code, AMD64_R11, cfg->frame_reg, lmf_offset);
6912 x86_prefix (code, X86_FS_PREFIX);
6913 amd64_mov_mem_reg (code, lmf_tls_offset, AMD64_R11, 8);
6916 if (lmf_addr_tls_offset != -1) {
6917 /* Load lmf quicky using the FS register */
6918 code = mono_amd64_emit_tls_get (code, AMD64_RAX, lmf_addr_tls_offset);
6920 /* The TLS key actually contains a pointer to the MonoJitTlsData structure */
6921 /* FIXME: Add a separate key for LMF to avoid this */
6922 amd64_alu_reg_imm (code, X86_ADD, AMD64_RAX, G_STRUCT_OFFSET (MonoJitTlsData, lmf));
6927 * The call might clobber argument registers, but they are already
6928 * saved to the stack/global regs.
6930 args_clobbered = TRUE;
6931 code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD,
6932 (gpointer)"mono_get_lmf_addr", TRUE);
6936 amd64_mov_membase_reg (code, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, lmf_addr), AMD64_RAX, sizeof(gpointer));
6937 /* Save previous_lmf */
6938 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RAX, 0, sizeof(gpointer));
6939 amd64_mov_membase_reg (code, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, previous_lmf), AMD64_R11, sizeof(gpointer));
6941 amd64_lea_membase (code, AMD64_R11, cfg->frame_reg, lmf_offset);
6942 amd64_mov_membase_reg (code, AMD64_RAX, 0, AMD64_R11, sizeof(gpointer));
6947 args_clobbered = TRUE;
6948 code = mono_arch_instrument_prolog (cfg, mono_trace_enter_method, code, TRUE);
6951 if (cfg->prof_options & MONO_PROFILE_ENTER_LEAVE)
6952 args_clobbered = TRUE;
6955 * Optimize the common case of the first bblock making a call with the same
6956 * arguments as the method. This works because the arguments are still in their
6957 * original argument registers.
6958 * FIXME: Generalize this
6960 if (!args_clobbered) {
6961 MonoBasicBlock *first_bb = cfg->bb_entry;
6964 next = mono_bb_first_ins (first_bb);
6965 if (!next && first_bb->next_bb) {
6966 first_bb = first_bb->next_bb;
6967 next = mono_bb_first_ins (first_bb);
6970 if (first_bb->in_count > 1)
6973 for (i = 0; next && i < sig->param_count + sig->hasthis; ++i) {
6974 ArgInfo *ainfo = cinfo->args + i;
6975 gboolean match = FALSE;
6977 ins = cfg->args [i];
6978 if (ins->opcode != OP_REGVAR) {
6979 switch (ainfo->storage) {
6981 if (((next->opcode == OP_LOAD_MEMBASE) || (next->opcode == OP_LOADI4_MEMBASE)) && next->inst_basereg == ins->inst_basereg && next->inst_offset == ins->inst_offset) {
6982 if (next->dreg == ainfo->reg) {
6986 next->opcode = OP_MOVE;
6987 next->sreg1 = ainfo->reg;
6988 /* Only continue if the instruction doesn't change argument regs */
6989 if (next->dreg == ainfo->reg || next->dreg == AMD64_RAX)
6999 /* Argument allocated to (non-volatile) register */
7000 switch (ainfo->storage) {
7002 if (next->opcode == OP_MOVE && next->sreg1 == ins->dreg && next->dreg == ainfo->reg) {
7014 //next = mono_inst_list_next (&next->node, &first_bb->ins_list);
7021 /* Initialize ss_trigger_page_var */
7022 if (cfg->arch.ss_trigger_page_var) {
7023 MonoInst *var = cfg->arch.ss_trigger_page_var;
7025 g_assert (!cfg->compile_aot);
7026 g_assert (var->opcode == OP_REGOFFSET);
7028 amd64_mov_reg_imm (code, AMD64_R11, (guint64)ss_trigger_page);
7029 amd64_mov_membase_reg (code, var->inst_basereg, var->inst_offset, AMD64_R11, 8);
7032 cfg->code_len = code - cfg->native_code;
7034 g_assert (cfg->code_len < cfg->code_size);
7040 mono_arch_emit_epilog (MonoCompile *cfg)
7042 MonoMethod *method = cfg->method;
7045 int max_epilog_size;
7047 gint32 lmf_offset = cfg->arch.lmf_offset;
7049 max_epilog_size = get_max_epilog_size (cfg);
7051 while (cfg->code_len + max_epilog_size > (cfg->code_size - 16)) {
7052 cfg->code_size *= 2;
7053 cfg->native_code = mono_realloc_native_code (cfg);
7054 mono_jit_stats.code_reallocs++;
7057 code = cfg->native_code + cfg->code_len;
7059 if (mono_jit_trace_calls != NULL && mono_trace_eval (method))
7060 code = mono_arch_instrument_epilog (cfg, mono_trace_leave_method, code, TRUE);
7062 /* the code restoring the registers must be kept in sync with OP_JMP */
7065 if (method->save_lmf) {
7066 /* check if we need to restore protection of the stack after a stack overflow */
7067 if (mono_get_jit_tls_offset () != -1) {
7069 code = mono_amd64_emit_tls_get (code, X86_ECX, mono_get_jit_tls_offset ());
7070 /* we load the value in a separate instruction: this mechanism may be
7071 * used later as a safer way to do thread interruption
7073 amd64_mov_reg_membase (code, X86_ECX, X86_ECX, G_STRUCT_OFFSET (MonoJitTlsData, restore_stack_prot), 8);
7074 x86_alu_reg_imm (code, X86_CMP, X86_ECX, 0);
7076 x86_branch8 (code, X86_CC_Z, 0, FALSE);
7077 /* note that the call trampoline will preserve eax/edx */
7078 x86_call_reg (code, X86_ECX);
7079 x86_patch (patch, code);
7081 /* FIXME: maybe save the jit tls in the prolog */
7083 if ((lmf_tls_offset != -1) && !optimize_for_xen) {
7085 * Optimized version which uses the mono_lmf TLS variable instead of indirection
7086 * through the mono_lmf_addr TLS variable.
7088 /* reg = previous_lmf */
7089 amd64_mov_reg_membase (code, AMD64_R11, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, previous_lmf), sizeof(gpointer));
7090 x86_prefix (code, X86_FS_PREFIX);
7091 amd64_mov_mem_reg (code, lmf_tls_offset, AMD64_R11, 8);
7093 /* Restore previous lmf */
7094 amd64_mov_reg_membase (code, AMD64_RCX, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, previous_lmf), sizeof(gpointer));
7095 amd64_mov_reg_membase (code, AMD64_R11, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, lmf_addr), sizeof(gpointer));
7096 amd64_mov_membase_reg (code, AMD64_R11, 0, AMD64_RCX, sizeof(gpointer));
7099 /* Restore caller saved regs */
7100 if (cfg->used_int_regs & (1 << AMD64_RBP)) {
7101 amd64_mov_reg_membase (code, AMD64_RBP, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rbp), 8);
7103 if (cfg->used_int_regs & (1 << AMD64_RBX)) {
7104 amd64_mov_reg_membase (code, AMD64_RBX, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rbx), 8);
7106 if (cfg->used_int_regs & (1 << AMD64_R12)) {
7107 amd64_mov_reg_membase (code, AMD64_R12, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r12), 8);
7109 if (cfg->used_int_regs & (1 << AMD64_R13)) {
7110 amd64_mov_reg_membase (code, AMD64_R13, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r13), 8);
7112 if (cfg->used_int_regs & (1 << AMD64_R14)) {
7113 amd64_mov_reg_membase (code, AMD64_R14, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r14), 8);
7115 if (cfg->used_int_regs & (1 << AMD64_R15)) {
7116 #if defined(__default_codegen__)
7117 amd64_mov_reg_membase (code, AMD64_R15, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r15), 8);
7118 #elif defined(__native_client_codegen__)
7119 g_assert_not_reached();
7123 if (cfg->used_int_regs & (1 << AMD64_RDI)) {
7124 amd64_mov_reg_membase (code, AMD64_RDI, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rdi), 8);
7126 if (cfg->used_int_regs & (1 << AMD64_RSI)) {
7127 amd64_mov_reg_membase (code, AMD64_RSI, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rsi), 8);
7132 if (cfg->arch.omit_fp) {
7133 gint32 save_area_offset = cfg->arch.reg_save_area_offset;
7135 for (i = 0; i < AMD64_NREG; ++i)
7136 if (AMD64_IS_CALLEE_SAVED_REG (i) && (cfg->used_int_regs & (1 << i))) {
7137 amd64_mov_reg_membase (code, i, AMD64_RSP, save_area_offset, 8);
7138 save_area_offset += 8;
7142 for (i = 0; i < AMD64_NREG; ++i)
7143 if (AMD64_IS_CALLEE_SAVED_REG (i) && (cfg->used_int_regs & (1 << i)))
7144 pos -= sizeof(mgreg_t);
7147 if (pos == - sizeof(mgreg_t)) {
7148 /* Only one register, so avoid lea */
7149 for (i = AMD64_NREG - 1; i > 0; --i)
7150 if (AMD64_IS_CALLEE_SAVED_REG (i) && (cfg->used_int_regs & (1 << i))) {
7151 amd64_mov_reg_membase (code, i, AMD64_RBP, pos, 8);
7155 amd64_lea_membase (code, AMD64_RSP, AMD64_RBP, pos);
7157 /* Pop registers in reverse order */
7158 for (i = AMD64_NREG - 1; i > 0; --i)
7159 if (AMD64_IS_CALLEE_SAVED_REG (i) && (cfg->used_int_regs & (1 << i))) {
7160 amd64_pop_reg (code, i);
7167 /* Load returned vtypes into registers if needed */
7168 cinfo = cfg->arch.cinfo;
7169 if (cinfo->ret.storage == ArgValuetypeInReg) {
7170 ArgInfo *ainfo = &cinfo->ret;
7171 MonoInst *inst = cfg->ret;
7173 for (quad = 0; quad < 2; quad ++) {
7174 switch (ainfo->pair_storage [quad]) {
7176 amd64_mov_reg_membase (code, ainfo->pair_regs [quad], inst->inst_basereg, inst->inst_offset + (quad * sizeof(mgreg_t)), sizeof(mgreg_t));
7178 case ArgInFloatSSEReg:
7179 amd64_movss_reg_membase (code, ainfo->pair_regs [quad], inst->inst_basereg, inst->inst_offset + (quad * sizeof(mgreg_t)));
7181 case ArgInDoubleSSEReg:
7182 amd64_movsd_reg_membase (code, ainfo->pair_regs [quad], inst->inst_basereg, inst->inst_offset + (quad * sizeof(mgreg_t)));
7187 g_assert_not_reached ();
7192 if (cfg->arch.omit_fp) {
7193 if (cfg->arch.stack_alloc_size)
7194 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, cfg->arch.stack_alloc_size);
7198 async_exc_point (code);
7201 cfg->code_len = code - cfg->native_code;
7203 g_assert (cfg->code_len < cfg->code_size);
7207 mono_arch_emit_exceptions (MonoCompile *cfg)
7209 MonoJumpInfo *patch_info;
7212 MonoClass *exc_classes [16];
7213 guint8 *exc_throw_start [16], *exc_throw_end [16];
7214 guint32 code_size = 0;
7216 /* Compute needed space */
7217 for (patch_info = cfg->patch_info; patch_info; patch_info = patch_info->next) {
7218 if (patch_info->type == MONO_PATCH_INFO_EXC)
7220 if (patch_info->type == MONO_PATCH_INFO_R8)
7221 code_size += 8 + 15; /* sizeof (double) + alignment */
7222 if (patch_info->type == MONO_PATCH_INFO_R4)
7223 code_size += 4 + 15; /* sizeof (float) + alignment */
7224 if (patch_info->type == MONO_PATCH_INFO_GC_CARD_TABLE_ADDR)
7225 code_size += 8 + 7; /*sizeof (void*) + alignment */
7228 #ifdef __native_client_codegen__
7229 /* Give us extra room on Native Client. This could be */
7230 /* more carefully calculated, but bundle alignment makes */
7231 /* it much trickier, so *2 like other places is good. */
7235 while (cfg->code_len + code_size > (cfg->code_size - 16)) {
7236 cfg->code_size *= 2;
7237 cfg->native_code = mono_realloc_native_code (cfg);
7238 mono_jit_stats.code_reallocs++;
7241 code = cfg->native_code + cfg->code_len;
7243 /* add code to raise exceptions */
7245 for (patch_info = cfg->patch_info; patch_info; patch_info = patch_info->next) {
7246 switch (patch_info->type) {
7247 case MONO_PATCH_INFO_EXC: {
7248 MonoClass *exc_class;
7252 amd64_patch (patch_info->ip.i + cfg->native_code, code);
7254 exc_class = mono_class_from_name (mono_defaults.corlib, "System", patch_info->data.name);
7255 g_assert (exc_class);
7256 throw_ip = patch_info->ip.i;
7258 //x86_breakpoint (code);
7259 /* Find a throw sequence for the same exception class */
7260 for (i = 0; i < nthrows; ++i)
7261 if (exc_classes [i] == exc_class)
7264 amd64_mov_reg_imm (code, AMD64_ARG_REG2, (exc_throw_end [i] - cfg->native_code) - throw_ip);
7265 x86_jump_code (code, exc_throw_start [i]);
7266 patch_info->type = MONO_PATCH_INFO_NONE;
7270 amd64_mov_reg_imm_size (code, AMD64_ARG_REG2, 0xf0f0f0f0, 4);
7274 exc_classes [nthrows] = exc_class;
7275 exc_throw_start [nthrows] = code;
7277 amd64_mov_reg_imm (code, AMD64_ARG_REG1, exc_class->type_token - MONO_TOKEN_TYPE_DEF);
7279 patch_info->type = MONO_PATCH_INFO_NONE;
7281 code = emit_call_body (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD, "mono_arch_throw_corlib_exception");
7283 amd64_mov_reg_imm (buf, AMD64_ARG_REG2, (code - cfg->native_code) - throw_ip);
7288 exc_throw_end [nthrows] = code;
7298 g_assert(code < cfg->native_code + cfg->code_size);
7301 /* Handle relocations with RIP relative addressing */
7302 for (patch_info = cfg->patch_info; patch_info; patch_info = patch_info->next) {
7303 gboolean remove = FALSE;
7304 guint8 *orig_code = code;
7306 switch (patch_info->type) {
7307 case MONO_PATCH_INFO_R8:
7308 case MONO_PATCH_INFO_R4: {
7309 guint8 *pos, *patch_pos;
7312 /* The SSE opcodes require a 16 byte alignment */
7313 #if defined(__default_codegen__)
7314 code = (guint8*)ALIGN_TO (code, 16);
7315 #elif defined(__native_client_codegen__)
7317 /* Pad this out with HLT instructions */
7318 /* or we can get garbage bytes emitted */
7319 /* which will fail validation */
7320 guint8 *aligned_code;
7321 /* extra align to make room for */
7322 /* mov/push below */
7323 int extra_align = patch_info->type == MONO_PATCH_INFO_R8 ? 2 : 1;
7324 aligned_code = (guint8*)ALIGN_TO (code + extra_align, 16);
7325 /* The technique of hiding data in an */
7326 /* instruction has a problem here: we */
7327 /* need the data aligned to a 16-byte */
7328 /* boundary but the instruction cannot */
7329 /* cross the bundle boundary. so only */
7330 /* odd multiples of 16 can be used */
7331 if ((intptr_t)aligned_code % kNaClAlignment == 0) {
7334 while (code < aligned_code) {
7335 *(code++) = 0xf4; /* hlt */
7340 pos = cfg->native_code + patch_info->ip.i;
7341 if (IS_REX (pos [1])) {
7342 patch_pos = pos + 5;
7343 target_pos = code - pos - 9;
7346 patch_pos = pos + 4;
7347 target_pos = code - pos - 8;
7350 if (patch_info->type == MONO_PATCH_INFO_R8) {
7351 #ifdef __native_client_codegen__
7352 /* Hide 64-bit data in a */
7353 /* "mov imm64, r11" instruction. */
7354 /* write it before the start of */
7356 *(code-2) = 0x49; /* prefix */
7357 *(code-1) = 0xbb; /* mov X, %r11 */
7359 *(double*)code = *(double*)patch_info->data.target;
7360 code += sizeof (double);
7362 #ifdef __native_client_codegen__
7363 /* Hide 32-bit data in a */
7364 /* "push imm32" instruction. */
7365 *(code-1) = 0x68; /* push */
7367 *(float*)code = *(float*)patch_info->data.target;
7368 code += sizeof (float);
7371 *(guint32*)(patch_pos) = target_pos;
7376 case MONO_PATCH_INFO_GC_CARD_TABLE_ADDR: {
7379 if (cfg->compile_aot)
7382 /*loading is faster against aligned addresses.*/
7383 code = (guint8*)ALIGN_TO (code, 8);
7384 memset (orig_code, 0, code - orig_code);
7386 pos = cfg->native_code + patch_info->ip.i;
7388 /*alu_op [rex] modr/m imm32 - 7 or 8 bytes */
7389 if (IS_REX (pos [1]))
7390 *(guint32*)(pos + 4) = (guint8*)code - pos - 8;
7392 *(guint32*)(pos + 3) = (guint8*)code - pos - 7;
7394 *(gpointer*)code = (gpointer)patch_info->data.target;
7395 code += sizeof (gpointer);
7405 if (patch_info == cfg->patch_info)
7406 cfg->patch_info = patch_info->next;
7410 for (tmp = cfg->patch_info; tmp->next != patch_info; tmp = tmp->next)
7412 tmp->next = patch_info->next;
7415 g_assert (code < cfg->native_code + cfg->code_size);
7418 cfg->code_len = code - cfg->native_code;
7420 g_assert (cfg->code_len < cfg->code_size);
7424 #endif /* DISABLE_JIT */
7427 mono_arch_instrument_prolog (MonoCompile *cfg, void *func, void *p, gboolean enable_arguments)
7430 CallInfo *cinfo = NULL;
7431 MonoMethodSignature *sig;
7433 int i, n, stack_area = 0;
7435 /* Keep this in sync with mono_arch_get_argument_info */
7437 if (enable_arguments) {
7438 /* Allocate a new area on the stack and save arguments there */
7439 sig = mono_method_signature (cfg->method);
7441 cinfo = get_call_info (cfg->generic_sharing_context, cfg->mempool, sig);
7443 n = sig->param_count + sig->hasthis;
7445 stack_area = ALIGN_TO (n * 8, 16);
7447 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, stack_area);
7449 for (i = 0; i < n; ++i) {
7450 inst = cfg->args [i];
7452 if (inst->opcode == OP_REGVAR)
7453 amd64_mov_membase_reg (code, AMD64_RSP, (i * 8), inst->dreg, 8);
7455 amd64_mov_reg_membase (code, AMD64_R11, inst->inst_basereg, inst->inst_offset, 8);
7456 amd64_mov_membase_reg (code, AMD64_RSP, (i * 8), AMD64_R11, 8);
7461 mono_add_patch_info (cfg, code-cfg->native_code, MONO_PATCH_INFO_METHODCONST, cfg->method);
7462 amd64_set_reg_template (code, AMD64_ARG_REG1);
7463 amd64_mov_reg_reg (code, AMD64_ARG_REG2, AMD64_RSP, 8);
7464 code = emit_call (cfg, code, MONO_PATCH_INFO_ABS, (gpointer)func, TRUE);
7466 if (enable_arguments)
7467 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, stack_area);
7481 mono_arch_instrument_epilog_full (MonoCompile *cfg, void *func, void *p, gboolean enable_arguments, gboolean preserve_argument_registers)
7484 int save_mode = SAVE_NONE;
7485 MonoMethod *method = cfg->method;
7486 MonoType *ret_type = mini_type_get_underlying_type (NULL, mono_method_signature (method)->ret);
7488 switch (ret_type->type) {
7489 case MONO_TYPE_VOID:
7490 /* special case string .ctor icall */
7491 if (strcmp (".ctor", method->name) && method->klass == mono_defaults.string_class)
7492 save_mode = SAVE_EAX;
7494 save_mode = SAVE_NONE;
7498 save_mode = SAVE_EAX;
7502 save_mode = SAVE_XMM;
7504 case MONO_TYPE_GENERICINST:
7505 if (!mono_type_generic_inst_is_valuetype (ret_type)) {
7506 save_mode = SAVE_EAX;
7510 case MONO_TYPE_VALUETYPE:
7511 save_mode = SAVE_STRUCT;
7514 save_mode = SAVE_EAX;
7518 /* Save the result and copy it into the proper argument register */
7519 switch (save_mode) {
7521 amd64_push_reg (code, AMD64_RAX);
7523 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, 8);
7524 if (enable_arguments)
7525 amd64_mov_reg_reg (code, AMD64_ARG_REG2, AMD64_RAX, 8);
7529 if (enable_arguments)
7530 amd64_mov_reg_imm (code, AMD64_ARG_REG2, 0);
7533 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, 8);
7534 amd64_movsd_membase_reg (code, AMD64_RSP, 0, AMD64_XMM0);
7536 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, 8);
7538 * The result is already in the proper argument register so no copying
7545 g_assert_not_reached ();
7548 /* Set %al since this is a varargs call */
7549 if (save_mode == SAVE_XMM)
7550 amd64_mov_reg_imm (code, AMD64_RAX, 1);
7552 amd64_mov_reg_imm (code, AMD64_RAX, 0);
7554 if (preserve_argument_registers) {
7555 amd64_push_reg (code, MONO_AMD64_ARG_REG1);
7556 amd64_push_reg (code, MONO_AMD64_ARG_REG2);
7559 mono_add_patch_info (cfg, code-cfg->native_code, MONO_PATCH_INFO_METHODCONST, method);
7560 amd64_set_reg_template (code, AMD64_ARG_REG1);
7561 code = emit_call (cfg, code, MONO_PATCH_INFO_ABS, (gpointer)func, TRUE);
7563 if (preserve_argument_registers) {
7564 amd64_pop_reg (code, MONO_AMD64_ARG_REG2);
7565 amd64_pop_reg (code, MONO_AMD64_ARG_REG1);
7568 /* Restore result */
7569 switch (save_mode) {
7571 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, 8);
7572 amd64_pop_reg (code, AMD64_RAX);
7578 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, 8);
7579 amd64_movsd_reg_membase (code, AMD64_XMM0, AMD64_RSP, 0);
7580 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, 8);
7585 g_assert_not_reached ();
7592 mono_arch_flush_icache (guint8 *code, gint size)
7598 mono_arch_flush_register_windows (void)
7603 mono_arch_is_inst_imm (gint64 imm)
7605 return amd64_is_imm32 (imm);
7609 * Determine whenever the trap whose info is in SIGINFO is caused by
7613 mono_arch_is_int_overflow (void *sigctx, void *info)
7620 mono_arch_sigctx_to_monoctx (sigctx, &ctx);
7622 rip = (guint8*)ctx.rip;
7624 if (IS_REX (rip [0])) {
7625 reg = amd64_rex_b (rip [0]);
7631 if ((rip [0] == 0xf7) && (x86_modrm_mod (rip [1]) == 0x3) && (x86_modrm_reg (rip [1]) == 0x7)) {
7633 reg += x86_modrm_rm (rip [1]);
7673 g_assert_not_reached ();
7685 mono_arch_get_patch_offset (guint8 *code)
7691 * mono_breakpoint_clean_code:
7693 * Copy @size bytes from @code - @offset to the buffer @buf. If the debugger inserted software
7694 * breakpoints in the original code, they are removed in the copy.
7696 * Returns TRUE if no sw breakpoint was present.
7699 mono_breakpoint_clean_code (guint8 *method_start, guint8 *code, int offset, guint8 *buf, int size)
7702 gboolean can_write = TRUE;
7704 * If method_start is non-NULL we need to perform bound checks, since we access memory
7705 * at code - offset we could go before the start of the method and end up in a different
7706 * page of memory that is not mapped or read incorrect data anyway. We zero-fill the bytes
7709 if (!method_start || code - offset >= method_start) {
7710 memcpy (buf, code - offset, size);
7712 int diff = code - method_start;
7713 memset (buf, 0, size);
7714 memcpy (buf + offset - diff, method_start, diff + size - offset);
7717 for (i = 0; i < MONO_BREAKPOINT_ARRAY_SIZE; ++i) {
7718 int idx = mono_breakpoint_info_index [i];
7722 ptr = mono_breakpoint_info [idx].address;
7723 if (ptr >= code && ptr < code + size) {
7724 guint8 saved_byte = mono_breakpoint_info [idx].saved_byte;
7726 /*g_print ("patching %p with 0x%02x (was: 0x%02x)\n", ptr, saved_byte, buf [ptr - code]);*/
7727 buf [ptr - code] = saved_byte;
7733 #if defined(__native_client_codegen__)
7734 /* For membase calls, we want the base register. for Native Client, */
7735 /* all indirect calls have the following sequence with the given sizes: */
7736 /* mov %eXX,%eXX [2-3] */
7737 /* mov disp(%r15,%rXX,scale),%r11d [4-8] */
7738 /* and $0xffffffffffffffe0,%r11d [4] */
7739 /* add %r15,%r11 [3] */
7740 /* callq *%r11 [3] */
7743 /* Determine if code points to a NaCl call-through-register sequence, */
7744 /* (i.e., the last 3 instructions listed above) */
7746 is_nacl_call_reg_sequence(guint8* code)
7748 const char *sequence = "\x41\x83\xe3\xe0" /* and */
7749 "\x4d\x03\xdf" /* add */
7750 "\x41\xff\xd3"; /* call */
7751 return memcmp(code, sequence, 10) == 0;
7754 /* Determine if code points to the first opcode of the mov membase component */
7755 /* of an indirect call sequence (i.e. the first 2 instructions listed above) */
7756 /* (there could be a REX prefix before the opcode but it is ignored) */
7758 is_nacl_indirect_call_membase_sequence(guint8* code)
7760 /* Check for mov opcode, reg-reg addressing mode (mod = 3), */
7761 return code[0] == 0x8b && amd64_modrm_mod(code[1]) == 3 &&
7762 /* and that src reg = dest reg */
7763 amd64_modrm_reg(code[1]) == amd64_modrm_rm(code[1]) &&
7764 /* Check that next inst is mov, uses SIB byte (rm = 4), */
7766 code[3] == 0x8b && amd64_modrm_rm(code[4]) == 4 &&
7767 /* and has dst of r11 and base of r15 */
7768 (amd64_modrm_reg(code[4]) + amd64_rex_r(code[2])) == AMD64_R11 &&
7769 (amd64_sib_base(code[5]) + amd64_rex_b(code[2])) == AMD64_R15;
7771 #endif /* __native_client_codegen__ */
7774 mono_arch_get_this_arg_reg (guint8 *code)
7776 return AMD64_ARG_REG1;
7780 mono_arch_get_this_arg_from_call (mgreg_t *regs, guint8 *code)
7782 return (gpointer)regs [mono_arch_get_this_arg_reg (code)];
7785 #define MAX_ARCH_DELEGATE_PARAMS 10
7788 get_delegate_invoke_impl (gboolean has_target, guint32 param_count, guint32 *code_len)
7790 guint8 *code, *start;
7794 start = code = mono_global_codeman_reserve (64);
7796 /* Replace the this argument with the target */
7797 amd64_mov_reg_reg (code, AMD64_RAX, AMD64_ARG_REG1, 8);
7798 amd64_mov_reg_membase (code, AMD64_ARG_REG1, AMD64_RAX, G_STRUCT_OFFSET (MonoDelegate, target), 8);
7799 amd64_jump_membase (code, AMD64_RAX, G_STRUCT_OFFSET (MonoDelegate, method_ptr));
7801 g_assert ((code - start) < 64);
7803 start = code = mono_global_codeman_reserve (64);
7805 if (param_count == 0) {
7806 amd64_jump_membase (code, AMD64_ARG_REG1, G_STRUCT_OFFSET (MonoDelegate, method_ptr));
7808 /* We have to shift the arguments left */
7809 amd64_mov_reg_reg (code, AMD64_RAX, AMD64_ARG_REG1, 8);
7810 for (i = 0; i < param_count; ++i) {
7813 amd64_mov_reg_reg (code, param_regs [i], param_regs [i + 1], 8);
7815 amd64_mov_reg_membase (code, param_regs [i], AMD64_RSP, 0x28, 8);
7817 amd64_mov_reg_reg (code, param_regs [i], param_regs [i + 1], 8);
7821 amd64_jump_membase (code, AMD64_RAX, G_STRUCT_OFFSET (MonoDelegate, method_ptr));
7823 g_assert ((code - start) < 64);
7826 nacl_global_codeman_validate(&start, 64, &code);
7828 mono_debug_add_delegate_trampoline (start, code - start);
7831 *code_len = code - start;
7834 if (mono_jit_map_is_enabled ()) {
7837 buff = (char*)"delegate_invoke_has_target";
7839 buff = g_strdup_printf ("delegate_invoke_no_target_%d", param_count);
7840 mono_emit_jit_tramp (start, code - start, buff);
7849 * mono_arch_get_delegate_invoke_impls:
7851 * Return a list of MonoTrampInfo structures for the delegate invoke impl
7855 mono_arch_get_delegate_invoke_impls (void)
7862 code = get_delegate_invoke_impl (TRUE, 0, &code_len);
7863 res = g_slist_prepend (res, mono_tramp_info_create (g_strdup ("delegate_invoke_impl_has_target"), code, code_len, NULL, NULL));
7865 for (i = 0; i < MAX_ARCH_DELEGATE_PARAMS; ++i) {
7866 code = get_delegate_invoke_impl (FALSE, i, &code_len);
7867 res = g_slist_prepend (res, mono_tramp_info_create (g_strdup_printf ("delegate_invoke_impl_target_%d", i), code, code_len, NULL, NULL));
7874 mono_arch_get_delegate_invoke_impl (MonoMethodSignature *sig, gboolean has_target)
7876 guint8 *code, *start;
7879 if (sig->param_count > MAX_ARCH_DELEGATE_PARAMS)
7882 /* FIXME: Support more cases */
7883 if (MONO_TYPE_ISSTRUCT (sig->ret))
7887 static guint8* cached = NULL;
7893 start = mono_aot_get_trampoline ("delegate_invoke_impl_has_target");
7895 start = get_delegate_invoke_impl (TRUE, 0, NULL);
7897 mono_memory_barrier ();
7901 static guint8* cache [MAX_ARCH_DELEGATE_PARAMS + 1] = {NULL};
7902 for (i = 0; i < sig->param_count; ++i)
7903 if (!mono_is_regsize_var (sig->params [i]))
7905 if (sig->param_count > 4)
7908 code = cache [sig->param_count];
7912 if (mono_aot_only) {
7913 char *name = g_strdup_printf ("delegate_invoke_impl_target_%d", sig->param_count);
7914 start = mono_aot_get_trampoline (name);
7917 start = get_delegate_invoke_impl (FALSE, sig->param_count, NULL);
7920 mono_memory_barrier ();
7922 cache [sig->param_count] = start;
7929 * Support for fast access to the thread-local lmf structure using the GS
7930 * segment register on NPTL + kernel 2.6.x.
7933 static gboolean tls_offset_inited = FALSE;
7936 mono_arch_setup_jit_tls_data (MonoJitTlsData *tls)
7938 if (!tls_offset_inited) {
7941 * We need to init this multiple times, since when we are first called, the key might not
7942 * be initialized yet.
7944 appdomain_tls_offset = mono_domain_get_tls_key ();
7945 lmf_tls_offset = mono_get_jit_tls_key ();
7946 lmf_addr_tls_offset = mono_get_jit_tls_key ();
7948 /* Only 64 tls entries can be accessed using inline code */
7949 if (appdomain_tls_offset >= 64)
7950 appdomain_tls_offset = -1;
7951 if (lmf_tls_offset >= 64)
7952 lmf_tls_offset = -1;
7954 tls_offset_inited = TRUE;
7956 optimize_for_xen = access ("/proc/xen", F_OK) == 0;
7958 appdomain_tls_offset = mono_domain_get_tls_offset ();
7959 lmf_tls_offset = mono_get_lmf_tls_offset ();
7960 lmf_addr_tls_offset = mono_get_lmf_addr_tls_offset ();
7966 mono_arch_free_jit_tls_data (MonoJitTlsData *tls)
7970 #ifdef MONO_ARCH_HAVE_IMT
7972 #if defined(__default_codegen__)
7973 #define CMP_SIZE (6 + 1)
7974 #define CMP_REG_REG_SIZE (4 + 1)
7975 #define BR_SMALL_SIZE 2
7976 #define BR_LARGE_SIZE 6
7977 #define MOV_REG_IMM_SIZE 10
7978 #define MOV_REG_IMM_32BIT_SIZE 6
7979 #define JUMP_REG_SIZE (2 + 1)
7980 #elif defined(__native_client_codegen__)
7981 /* NaCl N-byte instructions can be padded up to N-1 bytes */
7982 #define CMP_SIZE ((6 + 1) * 2 - 1)
7983 #define CMP_REG_REG_SIZE ((4 + 1) * 2 - 1)
7984 #define BR_SMALL_SIZE (2 * 2 - 1)
7985 #define BR_LARGE_SIZE (6 * 2 - 1)
7986 #define MOV_REG_IMM_SIZE (10 * 2 - 1)
7987 #define MOV_REG_IMM_32BIT_SIZE (6 * 2 - 1)
7988 /* Jump reg for NaCl adds a mask (+4) and add (+3) */
7989 #define JUMP_REG_SIZE ((2 + 1 + 4 + 3) * 2 - 1)
7990 /* Jump membase's size is large and unpredictable */
7991 /* in native client, just pad it out a whole bundle. */
7992 #define JUMP_MEMBASE_SIZE (kNaClAlignment)
7996 imt_branch_distance (MonoIMTCheckItem **imt_entries, int start, int target)
7998 int i, distance = 0;
7999 for (i = start; i < target; ++i)
8000 distance += imt_entries [i]->chunk_size;
8005 * LOCKING: called with the domain lock held
8008 mono_arch_build_imt_thunk (MonoVTable *vtable, MonoDomain *domain, MonoIMTCheckItem **imt_entries, int count,
8009 gpointer fail_tramp)
8013 guint8 *code, *start;
8014 gboolean vtable_is_32bit = ((gsize)(vtable) == (gsize)(int)(gsize)(vtable));
8016 for (i = 0; i < count; ++i) {
8017 MonoIMTCheckItem *item = imt_entries [i];
8018 if (item->is_equals) {
8019 if (item->check_target_idx) {
8020 if (!item->compare_done) {
8021 if (amd64_is_imm32 (item->key))
8022 item->chunk_size += CMP_SIZE;
8024 item->chunk_size += MOV_REG_IMM_SIZE + CMP_REG_REG_SIZE;
8026 if (item->has_target_code) {
8027 item->chunk_size += MOV_REG_IMM_SIZE;
8029 if (vtable_is_32bit)
8030 item->chunk_size += MOV_REG_IMM_32BIT_SIZE;
8032 item->chunk_size += MOV_REG_IMM_SIZE;
8033 #ifdef __native_client_codegen__
8034 item->chunk_size += JUMP_MEMBASE_SIZE;
8037 item->chunk_size += BR_SMALL_SIZE + JUMP_REG_SIZE;
8040 item->chunk_size += MOV_REG_IMM_SIZE * 3 + CMP_REG_REG_SIZE +
8041 BR_SMALL_SIZE + JUMP_REG_SIZE * 2;
8043 if (vtable_is_32bit)
8044 item->chunk_size += MOV_REG_IMM_32BIT_SIZE;
8046 item->chunk_size += MOV_REG_IMM_SIZE;
8047 item->chunk_size += JUMP_REG_SIZE;
8048 /* with assert below:
8049 * item->chunk_size += CMP_SIZE + BR_SMALL_SIZE + 1;
8051 #ifdef __native_client_codegen__
8052 item->chunk_size += JUMP_MEMBASE_SIZE;
8057 if (amd64_is_imm32 (item->key))
8058 item->chunk_size += CMP_SIZE;
8060 item->chunk_size += MOV_REG_IMM_SIZE + CMP_REG_REG_SIZE;
8061 item->chunk_size += BR_LARGE_SIZE;
8062 imt_entries [item->check_target_idx]->compare_done = TRUE;
8064 size += item->chunk_size;
8066 #if defined(__native_client__) && defined(__native_client_codegen__)
8067 /* In Native Client, we don't re-use thunks, allocate from the */
8068 /* normal code manager paths. */
8069 code = mono_domain_code_reserve (domain, size);
8072 code = mono_method_alloc_generic_virtual_thunk (domain, size);
8074 code = mono_domain_code_reserve (domain, size);
8077 for (i = 0; i < count; ++i) {
8078 MonoIMTCheckItem *item = imt_entries [i];
8079 item->code_target = code;
8080 if (item->is_equals) {
8081 gboolean fail_case = !item->check_target_idx && fail_tramp;
8083 if (item->check_target_idx || fail_case) {
8084 if (!item->compare_done || fail_case) {
8085 if (amd64_is_imm32 (item->key))
8086 amd64_alu_reg_imm (code, X86_CMP, MONO_ARCH_IMT_REG, (guint32)(gssize)item->key);
8088 amd64_mov_reg_imm (code, MONO_ARCH_IMT_SCRATCH_REG, item->key);
8089 amd64_alu_reg_reg (code, X86_CMP, MONO_ARCH_IMT_REG, MONO_ARCH_IMT_SCRATCH_REG);
8092 item->jmp_code = code;
8093 amd64_branch8 (code, X86_CC_NE, 0, FALSE);
8094 if (item->has_target_code) {
8095 amd64_mov_reg_imm (code, MONO_ARCH_IMT_SCRATCH_REG, item->value.target_code);
8096 amd64_jump_reg (code, MONO_ARCH_IMT_SCRATCH_REG);
8098 amd64_mov_reg_imm (code, MONO_ARCH_IMT_SCRATCH_REG, & (vtable->vtable [item->value.vtable_slot]));
8099 amd64_jump_membase (code, MONO_ARCH_IMT_SCRATCH_REG, 0);
8103 amd64_patch (item->jmp_code, code);
8104 amd64_mov_reg_imm (code, MONO_ARCH_IMT_SCRATCH_REG, fail_tramp);
8105 amd64_jump_reg (code, MONO_ARCH_IMT_SCRATCH_REG);
8106 item->jmp_code = NULL;
8109 /* enable the commented code to assert on wrong method */
8111 if (amd64_is_imm32 (item->key))
8112 amd64_alu_reg_imm (code, X86_CMP, MONO_ARCH_IMT_REG, (guint32)(gssize)item->key);
8114 amd64_mov_reg_imm (code, MONO_ARCH_IMT_SCRATCH_REG, item->key);
8115 amd64_alu_reg_reg (code, X86_CMP, MONO_ARCH_IMT_REG, MONO_ARCH_IMT_SCRATCH_REG);
8117 item->jmp_code = code;
8118 amd64_branch8 (code, X86_CC_NE, 0, FALSE);
8119 /* See the comment below about R10 */
8120 amd64_mov_reg_imm (code, MONO_ARCH_IMT_SCRATCH_REG, & (vtable->vtable [item->value.vtable_slot]));
8121 amd64_jump_membase (code, MONO_ARCH_IMT_SCRATCH_REG, 0);
8122 amd64_patch (item->jmp_code, code);
8123 amd64_breakpoint (code);
8124 item->jmp_code = NULL;
8126 /* We're using R10 (MONO_ARCH_IMT_SCRATCH_REG) here because R11 (MONO_ARCH_IMT_REG)
8127 needs to be preserved. R10 needs
8128 to be preserved for calls which
8129 require a runtime generic context,
8130 but interface calls don't. */
8131 amd64_mov_reg_imm (code, MONO_ARCH_IMT_SCRATCH_REG, & (vtable->vtable [item->value.vtable_slot]));
8132 amd64_jump_membase (code, MONO_ARCH_IMT_SCRATCH_REG, 0);
8136 if (amd64_is_imm32 (item->key))
8137 amd64_alu_reg_imm (code, X86_CMP, MONO_ARCH_IMT_REG, (guint32)(gssize)item->key);
8139 amd64_mov_reg_imm (code, MONO_ARCH_IMT_SCRATCH_REG, item->key);
8140 amd64_alu_reg_reg (code, X86_CMP, MONO_ARCH_IMT_REG, MONO_ARCH_IMT_SCRATCH_REG);
8142 item->jmp_code = code;
8143 if (x86_is_imm8 (imt_branch_distance (imt_entries, i, item->check_target_idx)))
8144 x86_branch8 (code, X86_CC_GE, 0, FALSE);
8146 x86_branch32 (code, X86_CC_GE, 0, FALSE);
8148 g_assert (code - item->code_target <= item->chunk_size);
8150 /* patch the branches to get to the target items */
8151 for (i = 0; i < count; ++i) {
8152 MonoIMTCheckItem *item = imt_entries [i];
8153 if (item->jmp_code) {
8154 if (item->check_target_idx) {
8155 amd64_patch (item->jmp_code, imt_entries [item->check_target_idx]->code_target);
8161 mono_stats.imt_thunks_size += code - start;
8162 g_assert (code - start <= size);
8164 nacl_domain_code_validate(domain, &start, size, &code);
8170 mono_arch_find_imt_method (mgreg_t *regs, guint8 *code)
8172 return (MonoMethod*)regs [MONO_ARCH_IMT_REG];
8177 mono_arch_find_static_call_vtable (mgreg_t *regs, guint8 *code)
8179 return (MonoVTable*) regs [MONO_ARCH_RGCTX_REG];
8183 mono_arch_get_cie_program (void)
8187 mono_add_unwind_op_def_cfa (l, (guint8*)NULL, (guint8*)NULL, AMD64_RSP, 8);
8188 mono_add_unwind_op_offset (l, (guint8*)NULL, (guint8*)NULL, AMD64_RIP, -8);
8194 mono_arch_emit_inst_for_method (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature *fsig, MonoInst **args)
8196 MonoInst *ins = NULL;
8199 if (cmethod->klass == mono_defaults.math_class) {
8200 if (strcmp (cmethod->name, "Sin") == 0) {
8202 } else if (strcmp (cmethod->name, "Cos") == 0) {
8204 } else if (strcmp (cmethod->name, "Sqrt") == 0) {
8206 } else if (strcmp (cmethod->name, "Abs") == 0 && fsig->params [0]->type == MONO_TYPE_R8) {
8211 MONO_INST_NEW (cfg, ins, opcode);
8212 ins->type = STACK_R8;
8213 ins->dreg = mono_alloc_freg (cfg);
8214 ins->sreg1 = args [0]->dreg;
8215 MONO_ADD_INS (cfg->cbb, ins);
8219 if (cfg->opt & MONO_OPT_CMOV) {
8220 if (strcmp (cmethod->name, "Min") == 0) {
8221 if (fsig->params [0]->type == MONO_TYPE_I4)
8223 if (fsig->params [0]->type == MONO_TYPE_U4)
8224 opcode = OP_IMIN_UN;
8225 else if (fsig->params [0]->type == MONO_TYPE_I8)
8227 else if (fsig->params [0]->type == MONO_TYPE_U8)
8228 opcode = OP_LMIN_UN;
8229 } else if (strcmp (cmethod->name, "Max") == 0) {
8230 if (fsig->params [0]->type == MONO_TYPE_I4)
8232 if (fsig->params [0]->type == MONO_TYPE_U4)
8233 opcode = OP_IMAX_UN;
8234 else if (fsig->params [0]->type == MONO_TYPE_I8)
8236 else if (fsig->params [0]->type == MONO_TYPE_U8)
8237 opcode = OP_LMAX_UN;
8242 MONO_INST_NEW (cfg, ins, opcode);
8243 ins->type = fsig->params [0]->type == MONO_TYPE_I4 ? STACK_I4 : STACK_I8;
8244 ins->dreg = mono_alloc_ireg (cfg);
8245 ins->sreg1 = args [0]->dreg;
8246 ins->sreg2 = args [1]->dreg;
8247 MONO_ADD_INS (cfg->cbb, ins);
8251 /* OP_FREM is not IEEE compatible */
8252 else if (strcmp (cmethod->name, "IEEERemainder") == 0) {
8253 MONO_INST_NEW (cfg, ins, OP_FREM);
8254 ins->inst_i0 = args [0];
8255 ins->inst_i1 = args [1];
8261 * Can't implement CompareExchange methods this way since they have
8269 mono_arch_print_tree (MonoInst *tree, int arity)
8274 MonoInst* mono_arch_get_domain_intrinsic (MonoCompile* cfg)
8278 if (appdomain_tls_offset == -1)
8281 MONO_INST_NEW (cfg, ins, OP_TLS_GET);
8282 ins->inst_offset = appdomain_tls_offset;
8286 #define _CTX_REG(ctx,fld,i) ((gpointer)((&ctx->fld)[i]))
8289 mono_arch_context_get_int_reg (MonoContext *ctx, int reg)
8292 case AMD64_RCX: return (gpointer)ctx->rcx;
8293 case AMD64_RDX: return (gpointer)ctx->rdx;
8294 case AMD64_RBX: return (gpointer)ctx->rbx;
8295 case AMD64_RBP: return (gpointer)ctx->rbp;
8296 case AMD64_RSP: return (gpointer)ctx->rsp;
8299 return _CTX_REG (ctx, rax, reg);
8301 return _CTX_REG (ctx, r12, reg - 12);
8303 g_assert_not_reached ();
8307 /*MONO_ARCH_HAVE_HANDLER_BLOCK_GUARD*/
8309 mono_arch_install_handler_block_guard (MonoJitInfo *ji, MonoJitExceptionInfo *clause, MonoContext *ctx, gpointer new_value)
8312 gpointer *sp, old_value;
8314 const unsigned char *handler;
8316 /*Decode the first instruction to figure out where did we store the spvar*/
8317 /*Our jit MUST generate the following:
8320 Which is encoded as: REX.W 0x89 mod_rm
8321 mod_rm (rsp, rbp, imm) which can be: (imm will never be zero)
8322 mod (reg + imm8): 01 reg(rsp): 100 rm(rbp): 101 -> 01100101 (0x65)
8323 mod (reg + imm32): 10 reg(rsp): 100 rm(rbp): 101 -> 10100101 (0xA5)
8325 FIXME can we generate frameless methods on this case?
8328 handler = clause->handler_start;
8331 if (*handler != 0x48)
8336 if (*handler != 0x89)
8340 if (*handler == 0x65)
8341 offset = *(signed char*)(handler + 1);
8342 else if (*handler == 0xA5)
8343 offset = *(int*)(handler + 1);
8348 bp = MONO_CONTEXT_GET_BP (ctx);
8349 sp = *(gpointer*)(bp + offset);
8352 if (old_value < ji->code_start || (char*)old_value > ((char*)ji->code_start + ji->code_size))
8361 * mono_arch_emit_load_aotconst:
8363 * Emit code to load the contents of the GOT slot identified by TRAMP_TYPE and
8364 * TARGET from the mscorlib GOT in full-aot code.
8365 * On AMD64, the result is placed into R11.
8368 mono_arch_emit_load_aotconst (guint8 *start, guint8 *code, MonoJumpInfo **ji, int tramp_type, gconstpointer target)
8370 *ji = mono_patch_info_list_prepend (*ji, code - start, tramp_type, target);
8371 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RIP, 0, 8);
8377 * mono_arch_get_trampolines:
8379 * Return a list of MonoTrampInfo structures describing arch specific trampolines
8383 mono_arch_get_trampolines (gboolean aot)
8385 return mono_amd64_get_exception_trampolines (aot);
8388 /* Soft Debug support */
8389 #ifdef MONO_ARCH_SOFT_DEBUG_SUPPORTED
8392 * mono_arch_set_breakpoint:
8394 * Set a breakpoint at the native code corresponding to JI at NATIVE_OFFSET.
8395 * The location should contain code emitted by OP_SEQ_POINT.
8398 mono_arch_set_breakpoint (MonoJitInfo *ji, guint8 *ip)
8401 guint8 *orig_code = code;
8404 * In production, we will use int3 (has to fix the size in the md
8405 * file). But that could confuse gdb, so during development, we emit a SIGSEGV
8408 g_assert (code [0] == 0x90);
8409 if (breakpoint_size == 8) {
8410 amd64_mov_reg_mem (code, AMD64_R11, (guint64)bp_trigger_page, 4);
8412 amd64_mov_reg_imm_size (code, AMD64_R11, (guint64)bp_trigger_page, 8);
8413 amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, 0, 4);
8416 g_assert (code - orig_code == breakpoint_size);
8420 * mono_arch_clear_breakpoint:
8422 * Clear the breakpoint at IP.
8425 mono_arch_clear_breakpoint (MonoJitInfo *ji, guint8 *ip)
8430 for (i = 0; i < breakpoint_size; ++i)
8435 mono_arch_is_breakpoint_event (void *info, void *sigctx)
8438 EXCEPTION_RECORD* einfo = (EXCEPTION_RECORD*)info;
8441 siginfo_t* sinfo = (siginfo_t*) info;
8442 /* Sometimes the address is off by 4 */
8443 if (sinfo->si_addr >= bp_trigger_page && (guint8*)sinfo->si_addr <= (guint8*)bp_trigger_page + 128)
8451 * mono_arch_get_ip_for_breakpoint:
8453 * Convert the ip in CTX to the address where a breakpoint was placed.
8456 mono_arch_get_ip_for_breakpoint (MonoJitInfo *ji, MonoContext *ctx)
8458 guint8 *ip = MONO_CONTEXT_GET_IP (ctx);
8460 /* ip points to the instruction causing the fault */
8461 ip -= (breakpoint_size - breakpoint_fault_size);
8467 * mono_arch_skip_breakpoint:
8469 * Modify CTX so the ip is placed after the breakpoint instruction, so when
8470 * we resume, the instruction is not executed again.
8473 mono_arch_skip_breakpoint (MonoContext *ctx)
8475 MONO_CONTEXT_SET_IP (ctx, (guint8*)MONO_CONTEXT_GET_IP (ctx) + breakpoint_fault_size);
8479 * mono_arch_start_single_stepping:
8481 * Start single stepping.
8484 mono_arch_start_single_stepping (void)
8486 mono_mprotect (ss_trigger_page, mono_pagesize (), 0);
8490 * mono_arch_stop_single_stepping:
8492 * Stop single stepping.
8495 mono_arch_stop_single_stepping (void)
8497 mono_mprotect (ss_trigger_page, mono_pagesize (), MONO_MMAP_READ);
8501 * mono_arch_is_single_step_event:
8503 * Return whenever the machine state in SIGCTX corresponds to a single
8507 mono_arch_is_single_step_event (void *info, void *sigctx)
8510 EXCEPTION_RECORD* einfo = (EXCEPTION_RECORD*)info;
8513 siginfo_t* sinfo = (siginfo_t*) info;
8514 /* Sometimes the address is off by 4 */
8515 if (sinfo->si_addr >= ss_trigger_page && (guint8*)sinfo->si_addr <= (guint8*)ss_trigger_page + 128)
8523 * mono_arch_get_ip_for_single_step:
8525 * Convert the ip in CTX to the address stored in seq_points.
8528 mono_arch_get_ip_for_single_step (MonoJitInfo *ji, MonoContext *ctx)
8530 guint8 *ip = MONO_CONTEXT_GET_IP (ctx);
8532 ip += single_step_fault_size;
8538 * mono_arch_skip_single_step:
8540 * Modify CTX so the ip is placed after the single step trigger instruction,
8541 * we resume, the instruction is not executed again.
8544 mono_arch_skip_single_step (MonoContext *ctx)
8546 MONO_CONTEXT_SET_IP (ctx, (guint8*)MONO_CONTEXT_GET_IP (ctx) + single_step_fault_size);
8550 * mono_arch_create_seq_point_info:
8552 * Return a pointer to a data structure which is used by the sequence
8553 * point implementation in AOTed code.
8556 mono_arch_get_seq_point_info (MonoDomain *domain, guint8 *code)