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 /* The LMF var is allocated normally */
1774 if (cfg->arch.omit_fp)
1775 cfg->arch.reg_save_area_offset = offset;
1776 /* Reserve space for caller saved registers */
1777 for (i = 0; i < AMD64_NREG; ++i)
1778 if (AMD64_IS_CALLEE_SAVED_REG (i) && (cfg->used_int_regs & (1 << i))) {
1779 offset += sizeof(mgreg_t);
1783 if (sig->ret->type != MONO_TYPE_VOID) {
1784 switch (cinfo->ret.storage) {
1786 case ArgInFloatSSEReg:
1787 case ArgInDoubleSSEReg:
1788 if ((MONO_TYPE_ISSTRUCT (sig->ret) && !mono_class_from_mono_type (sig->ret)->enumtype) || (sig->ret->type == MONO_TYPE_TYPEDBYREF)) {
1789 if (cfg->globalra) {
1790 cfg->vret_addr->opcode = OP_REGVAR;
1791 cfg->vret_addr->inst_c0 = cinfo->ret.reg;
1793 /* The register is volatile */
1794 cfg->vret_addr->opcode = OP_REGOFFSET;
1795 cfg->vret_addr->inst_basereg = cfg->frame_reg;
1796 if (cfg->arch.omit_fp) {
1797 cfg->vret_addr->inst_offset = offset;
1801 cfg->vret_addr->inst_offset = -offset;
1803 if (G_UNLIKELY (cfg->verbose_level > 1)) {
1804 printf ("vret_addr =");
1805 mono_print_ins (cfg->vret_addr);
1810 cfg->ret->opcode = OP_REGVAR;
1811 cfg->ret->inst_c0 = cinfo->ret.reg;
1814 case ArgValuetypeInReg:
1815 /* Allocate a local to hold the result, the epilog will copy it to the correct place */
1816 cfg->ret->opcode = OP_REGOFFSET;
1817 cfg->ret->inst_basereg = cfg->frame_reg;
1818 if (cfg->arch.omit_fp) {
1819 cfg->ret->inst_offset = offset;
1823 cfg->ret->inst_offset = - offset;
1827 g_assert_not_reached ();
1830 cfg->ret->dreg = cfg->ret->inst_c0;
1833 /* Allocate locals */
1834 if (!cfg->globalra) {
1835 offsets = mono_allocate_stack_slots (cfg, cfg->arch.omit_fp ? FALSE: TRUE, &locals_stack_size, &locals_stack_align);
1836 if (locals_stack_size > MONO_ARCH_MAX_FRAME_SIZE) {
1837 char *mname = mono_method_full_name (cfg->method, TRUE);
1838 cfg->exception_type = MONO_EXCEPTION_INVALID_PROGRAM;
1839 cfg->exception_message = g_strdup_printf ("Method %s stack is too big.", mname);
1844 if (locals_stack_align) {
1845 offset += (locals_stack_align - 1);
1846 offset &= ~(locals_stack_align - 1);
1848 if (cfg->arch.omit_fp) {
1849 cfg->locals_min_stack_offset = offset;
1850 cfg->locals_max_stack_offset = offset + locals_stack_size;
1852 cfg->locals_min_stack_offset = - (offset + locals_stack_size);
1853 cfg->locals_max_stack_offset = - offset;
1856 for (i = cfg->locals_start; i < cfg->num_varinfo; i++) {
1857 if (offsets [i] != -1) {
1858 MonoInst *ins = cfg->varinfo [i];
1859 ins->opcode = OP_REGOFFSET;
1860 ins->inst_basereg = cfg->frame_reg;
1861 if (cfg->arch.omit_fp)
1862 ins->inst_offset = (offset + offsets [i]);
1864 ins->inst_offset = - (offset + offsets [i]);
1865 //printf ("allocated local %d to ", i); mono_print_tree_nl (ins);
1868 offset += locals_stack_size;
1871 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG)) {
1872 g_assert (!cfg->arch.omit_fp);
1873 g_assert (cinfo->sig_cookie.storage == ArgOnStack);
1874 cfg->sig_cookie = cinfo->sig_cookie.offset + ARGS_OFFSET;
1877 for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
1878 ins = cfg->args [i];
1879 if (ins->opcode != OP_REGVAR) {
1880 ArgInfo *ainfo = &cinfo->args [i];
1881 gboolean inreg = TRUE;
1884 if (sig->hasthis && (i == 0))
1885 arg_type = &mono_defaults.object_class->byval_arg;
1887 arg_type = sig->params [i - sig->hasthis];
1889 if (cfg->globalra) {
1890 /* The new allocator needs info about the original locations of the arguments */
1891 switch (ainfo->storage) {
1893 case ArgInFloatSSEReg:
1894 case ArgInDoubleSSEReg:
1895 ins->opcode = OP_REGVAR;
1896 ins->inst_c0 = ainfo->reg;
1899 g_assert (!cfg->arch.omit_fp);
1900 ins->opcode = OP_REGOFFSET;
1901 ins->inst_basereg = cfg->frame_reg;
1902 ins->inst_offset = ainfo->offset + ARGS_OFFSET;
1904 case ArgValuetypeInReg:
1905 ins->opcode = OP_REGOFFSET;
1906 ins->inst_basereg = cfg->frame_reg;
1907 /* These arguments are saved to the stack in the prolog */
1908 offset = ALIGN_TO (offset, sizeof(mgreg_t));
1909 if (cfg->arch.omit_fp) {
1910 ins->inst_offset = offset;
1911 offset += (ainfo->storage == ArgValuetypeInReg) ? ainfo->nregs * sizeof (mgreg_t) : sizeof (mgreg_t);
1913 offset += (ainfo->storage == ArgValuetypeInReg) ? ainfo->nregs * sizeof (mgreg_t) : sizeof (mgreg_t);
1914 ins->inst_offset = - offset;
1918 g_assert_not_reached ();
1924 /* FIXME: Allocate volatile arguments to registers */
1925 if (ins->flags & (MONO_INST_VOLATILE|MONO_INST_INDIRECT))
1929 * Under AMD64, all registers used to pass arguments to functions
1930 * are volatile across calls.
1931 * FIXME: Optimize this.
1933 if ((ainfo->storage == ArgInIReg) || (ainfo->storage == ArgInFloatSSEReg) || (ainfo->storage == ArgInDoubleSSEReg) || (ainfo->storage == ArgValuetypeInReg))
1936 ins->opcode = OP_REGOFFSET;
1938 switch (ainfo->storage) {
1940 case ArgInFloatSSEReg:
1941 case ArgInDoubleSSEReg:
1943 ins->opcode = OP_REGVAR;
1944 ins->dreg = ainfo->reg;
1948 g_assert (!cfg->arch.omit_fp);
1949 ins->opcode = OP_REGOFFSET;
1950 ins->inst_basereg = cfg->frame_reg;
1951 ins->inst_offset = ainfo->offset + ARGS_OFFSET;
1953 case ArgValuetypeInReg:
1955 case ArgValuetypeAddrInIReg: {
1957 g_assert (!cfg->arch.omit_fp);
1959 MONO_INST_NEW (cfg, indir, 0);
1960 indir->opcode = OP_REGOFFSET;
1961 if (ainfo->pair_storage [0] == ArgInIReg) {
1962 indir->inst_basereg = cfg->frame_reg;
1963 offset = ALIGN_TO (offset, sizeof (gpointer));
1964 offset += (sizeof (gpointer));
1965 indir->inst_offset = - offset;
1968 indir->inst_basereg = cfg->frame_reg;
1969 indir->inst_offset = ainfo->offset + ARGS_OFFSET;
1972 ins->opcode = OP_VTARG_ADDR;
1973 ins->inst_left = indir;
1981 if (!inreg && (ainfo->storage != ArgOnStack) && (ainfo->storage != ArgValuetypeAddrInIReg)) {
1982 ins->opcode = OP_REGOFFSET;
1983 ins->inst_basereg = cfg->frame_reg;
1984 /* These arguments are saved to the stack in the prolog */
1985 offset = ALIGN_TO (offset, sizeof(mgreg_t));
1986 if (cfg->arch.omit_fp) {
1987 ins->inst_offset = offset;
1988 offset += (ainfo->storage == ArgValuetypeInReg) ? ainfo->nregs * sizeof (mgreg_t) : sizeof (mgreg_t);
1989 // Arguments are yet supported by the stack map creation code
1990 //cfg->locals_max_stack_offset = MAX (cfg->locals_max_stack_offset, offset);
1992 offset += (ainfo->storage == ArgValuetypeInReg) ? ainfo->nregs * sizeof (mgreg_t) : sizeof (mgreg_t);
1993 ins->inst_offset = - offset;
1994 //cfg->locals_min_stack_offset = MIN (cfg->locals_min_stack_offset, offset);
2000 cfg->stack_offset = offset;
2004 mono_arch_create_vars (MonoCompile *cfg)
2006 MonoMethodSignature *sig;
2009 sig = mono_method_signature (cfg->method);
2011 if (!cfg->arch.cinfo)
2012 cfg->arch.cinfo = get_call_info (cfg->generic_sharing_context, cfg->mempool, sig);
2013 cinfo = cfg->arch.cinfo;
2015 if (cinfo->ret.storage == ArgValuetypeInReg)
2016 cfg->ret_var_is_local = TRUE;
2018 if ((cinfo->ret.storage != ArgValuetypeInReg) && MONO_TYPE_ISSTRUCT (sig->ret)) {
2019 cfg->vret_addr = mono_compile_create_var (cfg, &mono_defaults.int_class->byval_arg, OP_ARG);
2020 if (G_UNLIKELY (cfg->verbose_level > 1)) {
2021 printf ("vret_addr = ");
2022 mono_print_ins (cfg->vret_addr);
2026 if (cfg->gen_seq_points) {
2029 ins = mono_compile_create_var (cfg, &mono_defaults.int_class->byval_arg, OP_LOCAL);
2030 ins->flags |= MONO_INST_VOLATILE;
2031 cfg->arch.ss_trigger_page_var = ins;
2034 #ifdef MONO_AMD64_NO_PUSHES
2036 * When this is set, we pass arguments on the stack by moves, and by allocating
2037 * a bigger stack frame, instead of pushes.
2038 * Pushes complicate exception handling because the arguments on the stack have
2039 * to be popped each time a frame is unwound. They also make fp elimination
2041 * FIXME: This doesn't work inside filter/finally clauses, since those execute
2042 * on a new frame which doesn't include a param area.
2044 cfg->arch.no_pushes = TRUE;
2047 if (cfg->method->save_lmf) {
2048 MonoInst *lmf_var = mono_compile_create_var (cfg, &mono_defaults.int_class->byval_arg, OP_LOCAL);
2049 lmf_var->flags |= MONO_INST_VOLATILE;
2050 lmf_var->flags |= MONO_INST_LMF;
2051 cfg->arch.lmf_var = lmf_var;
2056 add_outarg_reg (MonoCompile *cfg, MonoCallInst *call, ArgStorage storage, int reg, MonoInst *tree)
2062 MONO_INST_NEW (cfg, ins, OP_MOVE);
2063 ins->dreg = mono_alloc_ireg_copy (cfg, tree->dreg);
2064 ins->sreg1 = tree->dreg;
2065 MONO_ADD_INS (cfg->cbb, ins);
2066 mono_call_inst_add_outarg_reg (cfg, call, ins->dreg, reg, FALSE);
2068 case ArgInFloatSSEReg:
2069 MONO_INST_NEW (cfg, ins, OP_AMD64_SET_XMMREG_R4);
2070 ins->dreg = mono_alloc_freg (cfg);
2071 ins->sreg1 = tree->dreg;
2072 MONO_ADD_INS (cfg->cbb, ins);
2074 mono_call_inst_add_outarg_reg (cfg, call, ins->dreg, reg, TRUE);
2076 case ArgInDoubleSSEReg:
2077 MONO_INST_NEW (cfg, ins, OP_FMOVE);
2078 ins->dreg = mono_alloc_freg (cfg);
2079 ins->sreg1 = tree->dreg;
2080 MONO_ADD_INS (cfg->cbb, ins);
2082 mono_call_inst_add_outarg_reg (cfg, call, ins->dreg, reg, TRUE);
2086 g_assert_not_reached ();
2091 arg_storage_to_load_membase (ArgStorage storage)
2095 #if defined(__mono_ilp32__)
2096 return OP_LOADI8_MEMBASE;
2098 return OP_LOAD_MEMBASE;
2100 case ArgInDoubleSSEReg:
2101 return OP_LOADR8_MEMBASE;
2102 case ArgInFloatSSEReg:
2103 return OP_LOADR4_MEMBASE;
2105 g_assert_not_reached ();
2112 emit_sig_cookie (MonoCompile *cfg, MonoCallInst *call, CallInfo *cinfo)
2115 MonoMethodSignature *tmp_sig;
2118 if (call->tail_call)
2121 /* FIXME: Add support for signature tokens to AOT */
2122 cfg->disable_aot = TRUE;
2124 g_assert (cinfo->sig_cookie.storage == ArgOnStack);
2127 * mono_ArgIterator_Setup assumes the signature cookie is
2128 * passed first and all the arguments which were before it are
2129 * passed on the stack after the signature. So compensate by
2130 * passing a different signature.
2132 tmp_sig = mono_metadata_signature_dup_full (cfg->method->klass->image, call->signature);
2133 tmp_sig->param_count -= call->signature->sentinelpos;
2134 tmp_sig->sentinelpos = 0;
2135 memcpy (tmp_sig->params, call->signature->params + call->signature->sentinelpos, tmp_sig->param_count * sizeof (MonoType*));
2137 MONO_INST_NEW (cfg, sig_arg, OP_ICONST);
2138 sig_arg->dreg = mono_alloc_ireg (cfg);
2139 sig_arg->inst_p0 = tmp_sig;
2140 MONO_ADD_INS (cfg->cbb, sig_arg);
2142 if (cfg->arch.no_pushes) {
2143 MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STORE_MEMBASE_REG, AMD64_RSP, cinfo->sig_cookie.offset, sig_arg->dreg);
2145 MONO_INST_NEW (cfg, arg, OP_X86_PUSH);
2146 arg->sreg1 = sig_arg->dreg;
2147 MONO_ADD_INS (cfg->cbb, arg);
2151 static inline LLVMArgStorage
2152 arg_storage_to_llvm_arg_storage (MonoCompile *cfg, ArgStorage storage)
2156 return LLVMArgInIReg;
2160 g_assert_not_reached ();
2167 mono_arch_get_llvm_call_info (MonoCompile *cfg, MonoMethodSignature *sig)
2173 LLVMCallInfo *linfo;
2176 n = sig->param_count + sig->hasthis;
2178 cinfo = get_call_info (cfg->generic_sharing_context, cfg->mempool, sig);
2180 linfo = mono_mempool_alloc0 (cfg->mempool, sizeof (LLVMCallInfo) + (sizeof (LLVMArgInfo) * n));
2183 * LLVM always uses the native ABI while we use our own ABI, the
2184 * only difference is the handling of vtypes:
2185 * - we only pass/receive them in registers in some cases, and only
2186 * in 1 or 2 integer registers.
2188 if (cinfo->ret.storage == ArgValuetypeInReg) {
2190 cfg->exception_message = g_strdup ("pinvoke + vtypes");
2191 cfg->disable_llvm = TRUE;
2195 linfo->ret.storage = LLVMArgVtypeInReg;
2196 for (j = 0; j < 2; ++j)
2197 linfo->ret.pair_storage [j] = arg_storage_to_llvm_arg_storage (cfg, cinfo->ret.pair_storage [j]);
2200 if (MONO_TYPE_ISSTRUCT (sig->ret) && cinfo->ret.storage == ArgInIReg) {
2201 /* Vtype returned using a hidden argument */
2202 linfo->ret.storage = LLVMArgVtypeRetAddr;
2203 linfo->vret_arg_index = cinfo->vret_arg_index;
2206 for (i = 0; i < n; ++i) {
2207 ainfo = cinfo->args + i;
2209 if (i >= sig->hasthis)
2210 t = sig->params [i - sig->hasthis];
2212 t = &mono_defaults.int_class->byval_arg;
2214 linfo->args [i].storage = LLVMArgNone;
2216 switch (ainfo->storage) {
2218 linfo->args [i].storage = LLVMArgInIReg;
2220 case ArgInDoubleSSEReg:
2221 case ArgInFloatSSEReg:
2222 linfo->args [i].storage = LLVMArgInFPReg;
2225 if (MONO_TYPE_ISSTRUCT (t)) {
2226 linfo->args [i].storage = LLVMArgVtypeByVal;
2228 linfo->args [i].storage = LLVMArgInIReg;
2230 if (t->type == MONO_TYPE_R4)
2231 linfo->args [i].storage = LLVMArgInFPReg;
2232 else if (t->type == MONO_TYPE_R8)
2233 linfo->args [i].storage = LLVMArgInFPReg;
2237 case ArgValuetypeInReg:
2239 cfg->exception_message = g_strdup ("pinvoke + vtypes");
2240 cfg->disable_llvm = TRUE;
2244 linfo->args [i].storage = LLVMArgVtypeInReg;
2245 for (j = 0; j < 2; ++j)
2246 linfo->args [i].pair_storage [j] = arg_storage_to_llvm_arg_storage (cfg, ainfo->pair_storage [j]);
2249 cfg->exception_message = g_strdup ("ainfo->storage");
2250 cfg->disable_llvm = TRUE;
2260 mono_arch_emit_call (MonoCompile *cfg, MonoCallInst *call)
2263 MonoMethodSignature *sig;
2264 int i, n, stack_size;
2270 sig = call->signature;
2271 n = sig->param_count + sig->hasthis;
2273 cinfo = get_call_info (cfg->generic_sharing_context, cfg->mempool, sig);
2275 if (COMPILE_LLVM (cfg)) {
2276 /* We shouldn't be called in the llvm case */
2277 cfg->disable_llvm = TRUE;
2281 if (cinfo->need_stack_align) {
2282 if (!cfg->arch.no_pushes)
2283 MONO_EMIT_NEW_BIALU_IMM (cfg, OP_SUB_IMM, X86_ESP, X86_ESP, 8);
2287 * Emit all arguments which are passed on the stack to prevent register
2288 * allocation problems.
2290 if (cfg->arch.no_pushes) {
2291 for (i = 0; i < n; ++i) {
2293 ainfo = cinfo->args + i;
2295 in = call->args [i];
2297 if (sig->hasthis && i == 0)
2298 t = &mono_defaults.object_class->byval_arg;
2300 t = sig->params [i - sig->hasthis];
2302 if (ainfo->storage == ArgOnStack && !MONO_TYPE_ISSTRUCT (t) && !call->tail_call) {
2304 if (t->type == MONO_TYPE_R4)
2305 MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STORER4_MEMBASE_REG, AMD64_RSP, ainfo->offset, in->dreg);
2306 else if (t->type == MONO_TYPE_R8)
2307 MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STORER8_MEMBASE_REG, AMD64_RSP, ainfo->offset, in->dreg);
2309 MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STORE_MEMBASE_REG, AMD64_RSP, ainfo->offset, in->dreg);
2311 MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STORE_MEMBASE_REG, AMD64_RSP, ainfo->offset, in->dreg);
2313 if (cfg->compute_gc_maps) {
2316 EMIT_NEW_GC_PARAM_SLOT_LIVENESS_DEF (cfg, def, ainfo->offset, t);
2323 * Emit all parameters passed in registers in non-reverse order for better readability
2324 * and to help the optimization in emit_prolog ().
2326 for (i = 0; i < n; ++i) {
2327 ainfo = cinfo->args + i;
2329 in = call->args [i];
2331 if (ainfo->storage == ArgInIReg)
2332 add_outarg_reg (cfg, call, ainfo->storage, ainfo->reg, in);
2335 for (i = n - 1; i >= 0; --i) {
2336 ainfo = cinfo->args + i;
2338 in = call->args [i];
2340 switch (ainfo->storage) {
2344 case ArgInFloatSSEReg:
2345 case ArgInDoubleSSEReg:
2346 add_outarg_reg (cfg, call, ainfo->storage, ainfo->reg, in);
2349 case ArgValuetypeInReg:
2350 case ArgValuetypeAddrInIReg:
2351 if (ainfo->storage == ArgOnStack && call->tail_call) {
2352 MonoInst *call_inst = (MonoInst*)call;
2353 cfg->args [i]->flags |= MONO_INST_VOLATILE;
2354 EMIT_NEW_ARGSTORE (cfg, call_inst, i, in);
2355 } else if ((i >= sig->hasthis) && (MONO_TYPE_ISSTRUCT(sig->params [i - sig->hasthis]))) {
2359 if (sig->params [i - sig->hasthis]->type == MONO_TYPE_TYPEDBYREF) {
2360 size = sizeof (MonoTypedRef);
2361 align = sizeof (gpointer);
2365 size = mono_type_native_stack_size (&in->klass->byval_arg, &align);
2368 * Other backends use mono_type_stack_size (), but that
2369 * aligns the size to 8, which is larger than the size of
2370 * the source, leading to reads of invalid memory if the
2371 * source is at the end of address space.
2373 size = mono_class_value_size (in->klass, &align);
2376 g_assert (in->klass);
2378 if (ainfo->storage == ArgOnStack && size >= 10000) {
2379 /* Avoid asserts in emit_memcpy () */
2380 cfg->exception_type = MONO_EXCEPTION_INVALID_PROGRAM;
2381 cfg->exception_message = g_strdup_printf ("Passing an argument of size '%d'.", size);
2382 /* Continue normally */
2386 MONO_INST_NEW (cfg, arg, OP_OUTARG_VT);
2387 arg->sreg1 = in->dreg;
2388 arg->klass = in->klass;
2389 arg->backend.size = size;
2390 arg->inst_p0 = call;
2391 arg->inst_p1 = mono_mempool_alloc (cfg->mempool, sizeof (ArgInfo));
2392 memcpy (arg->inst_p1, ainfo, sizeof (ArgInfo));
2394 MONO_ADD_INS (cfg->cbb, arg);
2397 if (cfg->arch.no_pushes) {
2400 MONO_INST_NEW (cfg, arg, OP_X86_PUSH);
2401 arg->sreg1 = in->dreg;
2402 if (!sig->params [i - sig->hasthis]->byref) {
2403 if (sig->params [i - sig->hasthis]->type == MONO_TYPE_R4) {
2404 MONO_EMIT_NEW_BIALU_IMM (cfg, OP_SUB_IMM, X86_ESP, X86_ESP, 8);
2405 arg->opcode = OP_STORER4_MEMBASE_REG;
2406 arg->inst_destbasereg = X86_ESP;
2407 arg->inst_offset = 0;
2408 } else if (sig->params [i - sig->hasthis]->type == MONO_TYPE_R8) {
2409 MONO_EMIT_NEW_BIALU_IMM (cfg, OP_SUB_IMM, X86_ESP, X86_ESP, 8);
2410 arg->opcode = OP_STORER8_MEMBASE_REG;
2411 arg->inst_destbasereg = X86_ESP;
2412 arg->inst_offset = 0;
2415 MONO_ADD_INS (cfg->cbb, arg);
2420 g_assert_not_reached ();
2423 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG) && (i == sig->sentinelpos))
2424 /* Emit the signature cookie just before the implicit arguments */
2425 emit_sig_cookie (cfg, call, cinfo);
2428 /* Handle the case where there are no implicit arguments */
2429 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG) && (n == sig->sentinelpos))
2430 emit_sig_cookie (cfg, call, cinfo);
2432 if (sig->ret && MONO_TYPE_ISSTRUCT (sig->ret)) {
2435 if (cinfo->ret.storage == ArgValuetypeInReg) {
2436 if (cinfo->ret.pair_storage [0] == ArgInIReg && cinfo->ret.pair_storage [1] == ArgNone) {
2438 * Tell the JIT to use a more efficient calling convention: call using
2439 * OP_CALL, compute the result location after the call, and save the
2442 call->vret_in_reg = TRUE;
2444 * Nullify the instruction computing the vret addr to enable
2445 * future optimizations.
2448 NULLIFY_INS (call->vret_var);
2450 if (call->tail_call)
2453 * The valuetype is in RAX:RDX after the call, need to be copied to
2454 * the stack. Push the address here, so the call instruction can
2457 if (!cfg->arch.vret_addr_loc) {
2458 cfg->arch.vret_addr_loc = mono_compile_create_var (cfg, &mono_defaults.int_class->byval_arg, OP_LOCAL);
2459 /* Prevent it from being register allocated or optimized away */
2460 ((MonoInst*)cfg->arch.vret_addr_loc)->flags |= MONO_INST_VOLATILE;
2463 MONO_EMIT_NEW_UNALU (cfg, OP_MOVE, ((MonoInst*)cfg->arch.vret_addr_loc)->dreg, call->vret_var->dreg);
2467 MONO_INST_NEW (cfg, vtarg, OP_MOVE);
2468 vtarg->sreg1 = call->vret_var->dreg;
2469 vtarg->dreg = mono_alloc_preg (cfg);
2470 MONO_ADD_INS (cfg->cbb, vtarg);
2472 mono_call_inst_add_outarg_reg (cfg, call, vtarg->dreg, cinfo->ret.reg, FALSE);
2477 if (call->inst.opcode != OP_JMP && OP_TAILCALL != call->inst.opcode) {
2478 MONO_EMIT_NEW_BIALU_IMM (cfg, OP_SUB_IMM, X86_ESP, X86_ESP, 0x20);
2482 if (cfg->method->save_lmf) {
2483 MONO_INST_NEW (cfg, arg, OP_AMD64_SAVE_SP_TO_LMF);
2484 MONO_ADD_INS (cfg->cbb, arg);
2487 call->stack_usage = cinfo->stack_usage;
2491 mono_arch_emit_outarg_vt (MonoCompile *cfg, MonoInst *ins, MonoInst *src)
2494 MonoCallInst *call = (MonoCallInst*)ins->inst_p0;
2495 ArgInfo *ainfo = (ArgInfo*)ins->inst_p1;
2496 int size = ins->backend.size;
2498 if (ainfo->storage == ArgValuetypeInReg) {
2502 for (part = 0; part < 2; ++part) {
2503 if (ainfo->pair_storage [part] == ArgNone)
2506 MONO_INST_NEW (cfg, load, arg_storage_to_load_membase (ainfo->pair_storage [part]));
2507 load->inst_basereg = src->dreg;
2508 load->inst_offset = part * sizeof(mgreg_t);
2510 switch (ainfo->pair_storage [part]) {
2512 load->dreg = mono_alloc_ireg (cfg);
2514 case ArgInDoubleSSEReg:
2515 case ArgInFloatSSEReg:
2516 load->dreg = mono_alloc_freg (cfg);
2519 g_assert_not_reached ();
2521 MONO_ADD_INS (cfg->cbb, load);
2523 add_outarg_reg (cfg, call, ainfo->pair_storage [part], ainfo->pair_regs [part], load);
2525 } else if (ainfo->storage == ArgValuetypeAddrInIReg) {
2526 MonoInst *vtaddr, *load;
2527 vtaddr = mono_compile_create_var (cfg, &ins->klass->byval_arg, OP_LOCAL);
2529 g_assert (!cfg->arch.no_pushes);
2531 MONO_INST_NEW (cfg, load, OP_LDADDR);
2532 load->inst_p0 = vtaddr;
2533 vtaddr->flags |= MONO_INST_INDIRECT;
2534 load->type = STACK_MP;
2535 load->klass = vtaddr->klass;
2536 load->dreg = mono_alloc_ireg (cfg);
2537 MONO_ADD_INS (cfg->cbb, load);
2538 mini_emit_memcpy (cfg, load->dreg, 0, src->dreg, 0, size, 4);
2540 if (ainfo->pair_storage [0] == ArgInIReg) {
2541 MONO_INST_NEW (cfg, arg, OP_X86_LEA_MEMBASE);
2542 arg->dreg = mono_alloc_ireg (cfg);
2543 arg->sreg1 = load->dreg;
2545 MONO_ADD_INS (cfg->cbb, arg);
2546 mono_call_inst_add_outarg_reg (cfg, call, arg->dreg, ainfo->pair_regs [0], FALSE);
2548 MONO_INST_NEW (cfg, arg, OP_X86_PUSH);
2549 arg->sreg1 = load->dreg;
2550 MONO_ADD_INS (cfg->cbb, arg);
2554 if (cfg->arch.no_pushes) {
2555 int dreg = mono_alloc_ireg (cfg);
2557 MONO_EMIT_NEW_LOAD_MEMBASE (cfg, dreg, src->dreg, 0);
2558 MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STORE_MEMBASE_REG, AMD64_RSP, ainfo->offset, dreg);
2560 /* Can't use this for < 8 since it does an 8 byte memory load */
2561 MONO_INST_NEW (cfg, arg, OP_X86_PUSH_MEMBASE);
2562 arg->inst_basereg = src->dreg;
2563 arg->inst_offset = 0;
2564 MONO_ADD_INS (cfg->cbb, arg);
2566 } else if (size <= 40) {
2567 if (cfg->arch.no_pushes) {
2568 mini_emit_memcpy (cfg, AMD64_RSP, ainfo->offset, src->dreg, 0, size, 4);
2570 MONO_EMIT_NEW_BIALU_IMM (cfg, OP_SUB_IMM, X86_ESP, X86_ESP, ALIGN_TO (size, 8));
2571 mini_emit_memcpy (cfg, X86_ESP, 0, src->dreg, 0, size, 4);
2574 if (cfg->arch.no_pushes) {
2575 // FIXME: Code growth
2576 mini_emit_memcpy (cfg, AMD64_RSP, ainfo->offset, src->dreg, 0, size, 4);
2578 MONO_INST_NEW (cfg, arg, OP_X86_PUSH_OBJ);
2579 arg->inst_basereg = src->dreg;
2580 arg->inst_offset = 0;
2581 arg->inst_imm = size;
2582 MONO_ADD_INS (cfg->cbb, arg);
2586 if (cfg->compute_gc_maps) {
2588 EMIT_NEW_GC_PARAM_SLOT_LIVENESS_DEF (cfg, def, ainfo->offset, &ins->klass->byval_arg);
2594 mono_arch_emit_setret (MonoCompile *cfg, MonoMethod *method, MonoInst *val)
2596 MonoType *ret = mini_type_get_underlying_type (NULL, mono_method_signature (method)->ret);
2598 if (ret->type == MONO_TYPE_R4) {
2599 if (COMPILE_LLVM (cfg))
2600 MONO_EMIT_NEW_UNALU (cfg, OP_FMOVE, cfg->ret->dreg, val->dreg);
2602 MONO_EMIT_NEW_UNALU (cfg, OP_AMD64_SET_XMMREG_R4, cfg->ret->dreg, val->dreg);
2604 } else if (ret->type == MONO_TYPE_R8) {
2605 MONO_EMIT_NEW_UNALU (cfg, OP_FMOVE, cfg->ret->dreg, val->dreg);
2609 MONO_EMIT_NEW_UNALU (cfg, OP_MOVE, cfg->ret->dreg, val->dreg);
2612 #endif /* DISABLE_JIT */
2614 #define EMIT_COND_BRANCH(ins,cond,sign) \
2615 if (ins->inst_true_bb->native_offset) { \
2616 x86_branch (code, cond, cfg->native_code + ins->inst_true_bb->native_offset, sign); \
2618 mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_BB, ins->inst_true_bb); \
2619 if ((cfg->opt & MONO_OPT_BRANCH) && \
2620 x86_is_imm8 (ins->inst_true_bb->max_offset - offset)) \
2621 x86_branch8 (code, cond, 0, sign); \
2623 x86_branch32 (code, cond, 0, sign); \
2627 MonoMethodSignature *sig;
2632 mgreg_t regs [PARAM_REGS];
2638 dyn_call_supported (MonoMethodSignature *sig, CallInfo *cinfo)
2646 switch (cinfo->ret.storage) {
2650 case ArgValuetypeInReg: {
2651 ArgInfo *ainfo = &cinfo->ret;
2653 if (ainfo->pair_storage [0] != ArgNone && ainfo->pair_storage [0] != ArgInIReg)
2655 if (ainfo->pair_storage [1] != ArgNone && ainfo->pair_storage [1] != ArgInIReg)
2663 for (i = 0; i < cinfo->nargs; ++i) {
2664 ArgInfo *ainfo = &cinfo->args [i];
2665 switch (ainfo->storage) {
2668 case ArgValuetypeInReg:
2669 if (ainfo->pair_storage [0] != ArgNone && ainfo->pair_storage [0] != ArgInIReg)
2671 if (ainfo->pair_storage [1] != ArgNone && ainfo->pair_storage [1] != ArgInIReg)
2683 * mono_arch_dyn_call_prepare:
2685 * Return a pointer to an arch-specific structure which contains information
2686 * needed by mono_arch_get_dyn_call_args (). Return NULL if OP_DYN_CALL is not
2687 * supported for SIG.
2688 * This function is equivalent to ffi_prep_cif in libffi.
2691 mono_arch_dyn_call_prepare (MonoMethodSignature *sig)
2693 ArchDynCallInfo *info;
2696 cinfo = get_call_info (NULL, NULL, sig);
2698 if (!dyn_call_supported (sig, cinfo)) {
2703 info = g_new0 (ArchDynCallInfo, 1);
2704 // FIXME: Preprocess the info to speed up get_dyn_call_args ().
2706 info->cinfo = cinfo;
2708 return (MonoDynCallInfo*)info;
2712 * mono_arch_dyn_call_free:
2714 * Free a MonoDynCallInfo structure.
2717 mono_arch_dyn_call_free (MonoDynCallInfo *info)
2719 ArchDynCallInfo *ainfo = (ArchDynCallInfo*)info;
2721 g_free (ainfo->cinfo);
2725 #if !defined(__native_client__)
2726 #define PTR_TO_GREG(ptr) (mgreg_t)(ptr)
2727 #define GREG_TO_PTR(greg) (gpointer)(greg)
2729 /* Correctly handle casts to/from 32-bit pointers without compiler warnings */
2730 #define PTR_TO_GREG(ptr) (mgreg_t)(uintptr_t)(ptr)
2731 #define GREG_TO_PTR(greg) (gpointer)(guint32)(greg)
2735 * mono_arch_get_start_dyn_call:
2737 * Convert the arguments ARGS to a format which can be passed to OP_DYN_CALL, and
2738 * store the result into BUF.
2739 * ARGS should be an array of pointers pointing to the arguments.
2740 * RET should point to a memory buffer large enought to hold the result of the
2742 * This function should be as fast as possible, any work which does not depend
2743 * on the actual values of the arguments should be done in
2744 * mono_arch_dyn_call_prepare ().
2745 * start_dyn_call + OP_DYN_CALL + finish_dyn_call is equivalent to ffi_call in
2749 mono_arch_start_dyn_call (MonoDynCallInfo *info, gpointer **args, guint8 *ret, guint8 *buf, int buf_len)
2751 ArchDynCallInfo *dinfo = (ArchDynCallInfo*)info;
2752 DynCallArgs *p = (DynCallArgs*)buf;
2753 int arg_index, greg, i, pindex;
2754 MonoMethodSignature *sig = dinfo->sig;
2756 g_assert (buf_len >= sizeof (DynCallArgs));
2765 if (sig->hasthis || dinfo->cinfo->vret_arg_index == 1) {
2766 p->regs [greg ++] = PTR_TO_GREG(*(args [arg_index ++]));
2771 if (dinfo->cinfo->vtype_retaddr)
2772 p->regs [greg ++] = PTR_TO_GREG(ret);
2774 for (i = pindex; i < sig->param_count; i++) {
2775 MonoType *t = mono_type_get_underlying_type (sig->params [i]);
2776 gpointer *arg = args [arg_index ++];
2779 p->regs [greg ++] = PTR_TO_GREG(*(arg));
2784 case MONO_TYPE_STRING:
2785 case MONO_TYPE_CLASS:
2786 case MONO_TYPE_ARRAY:
2787 case MONO_TYPE_SZARRAY:
2788 case MONO_TYPE_OBJECT:
2792 #if !defined(__mono_ilp32__)
2796 g_assert (dinfo->cinfo->args [i + sig->hasthis].reg == param_regs [greg]);
2797 p->regs [greg ++] = PTR_TO_GREG(*(arg));
2799 #if defined(__mono_ilp32__)
2802 g_assert (dinfo->cinfo->args [i + sig->hasthis].reg == param_regs [greg]);
2803 p->regs [greg ++] = *(guint64*)(arg);
2806 case MONO_TYPE_BOOLEAN:
2808 p->regs [greg ++] = *(guint8*)(arg);
2811 p->regs [greg ++] = *(gint8*)(arg);
2814 p->regs [greg ++] = *(gint16*)(arg);
2817 case MONO_TYPE_CHAR:
2818 p->regs [greg ++] = *(guint16*)(arg);
2821 p->regs [greg ++] = *(gint32*)(arg);
2824 p->regs [greg ++] = *(guint32*)(arg);
2826 case MONO_TYPE_GENERICINST:
2827 if (MONO_TYPE_IS_REFERENCE (t)) {
2828 p->regs [greg ++] = PTR_TO_GREG(*(arg));
2833 case MONO_TYPE_VALUETYPE: {
2834 ArgInfo *ainfo = &dinfo->cinfo->args [i + sig->hasthis];
2836 g_assert (ainfo->storage == ArgValuetypeInReg);
2837 if (ainfo->pair_storage [0] != ArgNone) {
2838 g_assert (ainfo->pair_storage [0] == ArgInIReg);
2839 p->regs [greg ++] = ((mgreg_t*)(arg))[0];
2841 if (ainfo->pair_storage [1] != ArgNone) {
2842 g_assert (ainfo->pair_storage [1] == ArgInIReg);
2843 p->regs [greg ++] = ((mgreg_t*)(arg))[1];
2848 g_assert_not_reached ();
2852 g_assert (greg <= PARAM_REGS);
2856 * mono_arch_finish_dyn_call:
2858 * Store the result of a dyn call into the return value buffer passed to
2859 * start_dyn_call ().
2860 * This function should be as fast as possible, any work which does not depend
2861 * on the actual values of the arguments should be done in
2862 * mono_arch_dyn_call_prepare ().
2865 mono_arch_finish_dyn_call (MonoDynCallInfo *info, guint8 *buf)
2867 ArchDynCallInfo *dinfo = (ArchDynCallInfo*)info;
2868 MonoMethodSignature *sig = dinfo->sig;
2869 guint8 *ret = ((DynCallArgs*)buf)->ret;
2870 mgreg_t res = ((DynCallArgs*)buf)->res;
2872 switch (mono_type_get_underlying_type (sig->ret)->type) {
2873 case MONO_TYPE_VOID:
2874 *(gpointer*)ret = NULL;
2876 case MONO_TYPE_STRING:
2877 case MONO_TYPE_CLASS:
2878 case MONO_TYPE_ARRAY:
2879 case MONO_TYPE_SZARRAY:
2880 case MONO_TYPE_OBJECT:
2884 *(gpointer*)ret = GREG_TO_PTR(res);
2890 case MONO_TYPE_BOOLEAN:
2891 *(guint8*)ret = res;
2894 *(gint16*)ret = res;
2897 case MONO_TYPE_CHAR:
2898 *(guint16*)ret = res;
2901 *(gint32*)ret = res;
2904 *(guint32*)ret = res;
2907 *(gint64*)ret = res;
2910 *(guint64*)ret = res;
2912 case MONO_TYPE_GENERICINST:
2913 if (MONO_TYPE_IS_REFERENCE (sig->ret)) {
2914 *(gpointer*)ret = GREG_TO_PTR(res);
2919 case MONO_TYPE_VALUETYPE:
2920 if (dinfo->cinfo->vtype_retaddr) {
2923 ArgInfo *ainfo = &dinfo->cinfo->ret;
2925 g_assert (ainfo->storage == ArgValuetypeInReg);
2927 if (ainfo->pair_storage [0] != ArgNone) {
2928 g_assert (ainfo->pair_storage [0] == ArgInIReg);
2929 ((mgreg_t*)ret)[0] = res;
2932 g_assert (ainfo->pair_storage [1] == ArgNone);
2936 g_assert_not_reached ();
2940 /* emit an exception if condition is fail */
2941 #define EMIT_COND_SYSTEM_EXCEPTION(cond,signed,exc_name) \
2943 MonoInst *tins = mono_branch_optimize_exception_target (cfg, bb, exc_name); \
2944 if (tins == NULL) { \
2945 mono_add_patch_info (cfg, code - cfg->native_code, \
2946 MONO_PATCH_INFO_EXC, exc_name); \
2947 x86_branch32 (code, cond, 0, signed); \
2949 EMIT_COND_BRANCH (tins, cond, signed); \
2953 #define EMIT_FPCOMPARE(code) do { \
2954 amd64_fcompp (code); \
2955 amd64_fnstsw (code); \
2958 #define EMIT_SSE2_FPFUNC(code, op, dreg, sreg1) do { \
2959 amd64_movsd_membase_reg (code, AMD64_RSP, -8, (sreg1)); \
2960 amd64_fld_membase (code, AMD64_RSP, -8, TRUE); \
2961 amd64_ ##op (code); \
2962 amd64_fst_membase (code, AMD64_RSP, -8, TRUE, TRUE); \
2963 amd64_movsd_reg_membase (code, (dreg), AMD64_RSP, -8); \
2967 emit_call_body (MonoCompile *cfg, guint8 *code, guint32 patch_type, gconstpointer data)
2969 gboolean no_patch = FALSE;
2972 * FIXME: Add support for thunks
2975 gboolean near_call = FALSE;
2978 * Indirect calls are expensive so try to make a near call if possible.
2979 * The caller memory is allocated by the code manager so it is
2980 * guaranteed to be at a 32 bit offset.
2983 if (patch_type != MONO_PATCH_INFO_ABS) {
2984 /* The target is in memory allocated using the code manager */
2987 if ((patch_type == MONO_PATCH_INFO_METHOD) || (patch_type == MONO_PATCH_INFO_METHOD_JUMP)) {
2988 if (((MonoMethod*)data)->klass->image->aot_module)
2989 /* The callee might be an AOT method */
2991 if (((MonoMethod*)data)->dynamic)
2992 /* The target is in malloc-ed memory */
2996 if (patch_type == MONO_PATCH_INFO_INTERNAL_METHOD) {
2998 * The call might go directly to a native function without
3001 MonoJitICallInfo *mi = mono_find_jit_icall_by_name (data);
3003 gconstpointer target = mono_icall_get_wrapper (mi);
3004 if ((((guint64)target) >> 32) != 0)
3010 if (cfg->abs_patches && g_hash_table_lookup (cfg->abs_patches, data)) {
3012 * This is not really an optimization, but required because the
3013 * generic class init trampolines use R11 to pass the vtable.
3017 MonoJitICallInfo *info = mono_find_jit_icall_by_addr (data);
3019 if ((cfg->method->wrapper_type == MONO_WRAPPER_MANAGED_TO_NATIVE) &&
3020 strstr (cfg->method->name, info->name)) {
3021 /* A call to the wrapped function */
3022 if ((((guint64)data) >> 32) == 0)
3026 else if (info->func == info->wrapper) {
3028 if ((((guint64)info->func) >> 32) == 0)
3032 /* See the comment in mono_codegen () */
3033 if ((info->name [0] != 'v') || (strstr (info->name, "ves_array_new_va_") == NULL && strstr (info->name, "ves_array_element_address_") == NULL))
3037 else if ((((guint64)data) >> 32) == 0) {
3044 if (cfg->method->dynamic)
3045 /* These methods are allocated using malloc */
3048 #ifdef MONO_ARCH_NOMAP32BIT
3052 /* The 64bit XEN kernel does not honour the MAP_32BIT flag. (#522894) */
3053 if (optimize_for_xen)
3056 if (cfg->compile_aot) {
3063 * Align the call displacement to an address divisible by 4 so it does
3064 * not span cache lines. This is required for code patching to work on SMP
3067 if (!no_patch && ((guint32)(code + 1 - cfg->native_code) % 4) != 0) {
3068 guint32 pad_size = 4 - ((guint32)(code + 1 - cfg->native_code) % 4);
3069 amd64_padding (code, pad_size);
3071 mono_add_patch_info (cfg, code - cfg->native_code, patch_type, data);
3072 amd64_call_code (code, 0);
3075 mono_add_patch_info (cfg, code - cfg->native_code, patch_type, data);
3076 amd64_set_reg_template (code, GP_SCRATCH_REG);
3077 amd64_call_reg (code, GP_SCRATCH_REG);
3084 static inline guint8*
3085 emit_call (MonoCompile *cfg, guint8 *code, guint32 patch_type, gconstpointer data, gboolean win64_adjust_stack)
3088 if (win64_adjust_stack)
3089 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, 32);
3091 code = emit_call_body (cfg, code, patch_type, data);
3093 if (win64_adjust_stack)
3094 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, 32);
3101 store_membase_imm_to_store_membase_reg (int opcode)
3104 case OP_STORE_MEMBASE_IMM:
3105 return OP_STORE_MEMBASE_REG;
3106 case OP_STOREI4_MEMBASE_IMM:
3107 return OP_STOREI4_MEMBASE_REG;
3108 case OP_STOREI8_MEMBASE_IMM:
3109 return OP_STOREI8_MEMBASE_REG;
3117 #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)))
3120 * mono_arch_peephole_pass_1:
3122 * Perform peephole opts which should/can be performed before local regalloc
3125 mono_arch_peephole_pass_1 (MonoCompile *cfg, MonoBasicBlock *bb)
3129 MONO_BB_FOR_EACH_INS_SAFE (bb, n, ins) {
3130 MonoInst *last_ins = ins->prev;
3132 switch (ins->opcode) {
3136 if ((ins->sreg1 < MONO_MAX_IREGS) && (ins->dreg >= MONO_MAX_IREGS) && (ins->inst_imm > 0)) {
3138 * X86_LEA is like ADD, but doesn't have the
3139 * sreg1==dreg restriction. inst_imm > 0 is needed since LEA sign-extends
3140 * its operand to 64 bit.
3142 ins->opcode = OP_X86_LEA_MEMBASE;
3143 ins->inst_basereg = ins->sreg1;
3148 if ((ins->sreg1 == ins->sreg2) && (ins->sreg1 == ins->dreg)) {
3152 * Replace STORE_MEMBASE_IMM 0 with STORE_MEMBASE_REG since
3153 * the latter has length 2-3 instead of 6 (reverse constant
3154 * propagation). These instruction sequences are very common
3155 * in the initlocals bblock.
3157 for (ins2 = ins->next; ins2; ins2 = ins2->next) {
3158 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)) {
3159 ins2->opcode = store_membase_imm_to_store_membase_reg (ins2->opcode);
3160 ins2->sreg1 = ins->dreg;
3161 } 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)) {
3163 } else if (((ins2->opcode == OP_ICONST) || (ins2->opcode == OP_I8CONST)) && (ins2->dreg == ins->dreg) && (ins2->inst_c0 == 0)) {
3172 case OP_COMPARE_IMM:
3173 case OP_LCOMPARE_IMM:
3174 /* OP_COMPARE_IMM (reg, 0)
3176 * OP_AMD64_TEST_NULL (reg)
3179 ins->opcode = OP_AMD64_TEST_NULL;
3181 case OP_ICOMPARE_IMM:
3183 ins->opcode = OP_X86_TEST_NULL;
3185 case OP_AMD64_ICOMPARE_MEMBASE_IMM:
3187 * OP_STORE_MEMBASE_REG reg, offset(basereg)
3188 * OP_X86_COMPARE_MEMBASE_IMM offset(basereg), imm
3190 * OP_STORE_MEMBASE_REG reg, offset(basereg)
3191 * OP_COMPARE_IMM reg, imm
3193 * Note: if imm = 0 then OP_COMPARE_IMM replaced with OP_X86_TEST_NULL
3195 if (last_ins && (last_ins->opcode == OP_STOREI4_MEMBASE_REG) &&
3196 ins->inst_basereg == last_ins->inst_destbasereg &&
3197 ins->inst_offset == last_ins->inst_offset) {
3198 ins->opcode = OP_ICOMPARE_IMM;
3199 ins->sreg1 = last_ins->sreg1;
3201 /* check if we can remove cmp reg,0 with test null */
3203 ins->opcode = OP_X86_TEST_NULL;
3209 mono_peephole_ins (bb, ins);
3214 mono_arch_peephole_pass_2 (MonoCompile *cfg, MonoBasicBlock *bb)
3218 MONO_BB_FOR_EACH_INS_SAFE (bb, n, ins) {
3219 switch (ins->opcode) {
3222 /* reg = 0 -> XOR (reg, reg) */
3223 /* XOR sets cflags on x86, so we cant do it always */
3224 if (ins->inst_c0 == 0 && (!ins->next || (ins->next && INST_IGNORES_CFLAGS (ins->next->opcode)))) {
3225 ins->opcode = OP_LXOR;
3226 ins->sreg1 = ins->dreg;
3227 ins->sreg2 = ins->dreg;
3235 * Use IXOR to avoid a rex prefix if possible. The cpu will sign extend the
3236 * 0 result into 64 bits.
3238 if ((ins->sreg1 == ins->sreg2) && (ins->sreg1 == ins->dreg)) {
3239 ins->opcode = OP_IXOR;
3243 if ((ins->sreg1 == ins->sreg2) && (ins->sreg1 == ins->dreg)) {
3247 * Replace STORE_MEMBASE_IMM 0 with STORE_MEMBASE_REG since
3248 * the latter has length 2-3 instead of 6 (reverse constant
3249 * propagation). These instruction sequences are very common
3250 * in the initlocals bblock.
3252 for (ins2 = ins->next; ins2; ins2 = ins2->next) {
3253 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)) {
3254 ins2->opcode = store_membase_imm_to_store_membase_reg (ins2->opcode);
3255 ins2->sreg1 = ins->dreg;
3256 } 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)) {
3258 } else if (((ins2->opcode == OP_ICONST) || (ins2->opcode == OP_I8CONST)) && (ins2->dreg == ins->dreg) && (ins2->inst_c0 == 0)) {
3268 if ((ins->inst_imm == 1) && (ins->dreg == ins->sreg1))
3269 ins->opcode = OP_X86_INC_REG;
3272 if ((ins->inst_imm == 1) && (ins->dreg == ins->sreg1))
3273 ins->opcode = OP_X86_DEC_REG;
3277 mono_peephole_ins (bb, ins);
3281 #define NEW_INS(cfg,ins,dest,op) do { \
3282 MONO_INST_NEW ((cfg), (dest), (op)); \
3283 (dest)->cil_code = (ins)->cil_code; \
3284 mono_bblock_insert_before_ins (bb, ins, (dest)); \
3288 * mono_arch_lowering_pass:
3290 * Converts complex opcodes into simpler ones so that each IR instruction
3291 * corresponds to one machine instruction.
3294 mono_arch_lowering_pass (MonoCompile *cfg, MonoBasicBlock *bb)
3296 MonoInst *ins, *n, *temp;
3299 * FIXME: Need to add more instructions, but the current machine
3300 * description can't model some parts of the composite instructions like
3303 MONO_BB_FOR_EACH_INS_SAFE (bb, n, ins) {
3304 switch (ins->opcode) {
3308 case OP_IDIV_UN_IMM:
3309 case OP_IREM_UN_IMM:
3310 mono_decompose_op_imm (cfg, bb, ins);
3313 /* Keep the opcode if we can implement it efficiently */
3314 if (!((ins->inst_imm > 0) && (mono_is_power_of_two (ins->inst_imm) != -1)))
3315 mono_decompose_op_imm (cfg, bb, ins);
3317 case OP_COMPARE_IMM:
3318 case OP_LCOMPARE_IMM:
3319 if (!amd64_is_imm32 (ins->inst_imm)) {
3320 NEW_INS (cfg, ins, temp, OP_I8CONST);
3321 temp->inst_c0 = ins->inst_imm;
3322 temp->dreg = mono_alloc_ireg (cfg);
3323 ins->opcode = OP_COMPARE;
3324 ins->sreg2 = temp->dreg;
3327 #ifndef __mono_ilp32__
3328 case OP_LOAD_MEMBASE:
3330 case OP_LOADI8_MEMBASE:
3331 #ifndef __native_client_codegen__
3332 /* Don't generate memindex opcodes (to simplify */
3333 /* read sandboxing) */
3334 if (!amd64_is_imm32 (ins->inst_offset)) {
3335 NEW_INS (cfg, ins, temp, OP_I8CONST);
3336 temp->inst_c0 = ins->inst_offset;
3337 temp->dreg = mono_alloc_ireg (cfg);
3338 ins->opcode = OP_AMD64_LOADI8_MEMINDEX;
3339 ins->inst_indexreg = temp->dreg;
3343 #ifndef __mono_ilp32__
3344 case OP_STORE_MEMBASE_IMM:
3346 case OP_STOREI8_MEMBASE_IMM:
3347 if (!amd64_is_imm32 (ins->inst_imm)) {
3348 NEW_INS (cfg, ins, temp, OP_I8CONST);
3349 temp->inst_c0 = ins->inst_imm;
3350 temp->dreg = mono_alloc_ireg (cfg);
3351 ins->opcode = OP_STOREI8_MEMBASE_REG;
3352 ins->sreg1 = temp->dreg;
3355 #ifdef MONO_ARCH_SIMD_INTRINSICS
3356 case OP_EXPAND_I1: {
3357 int temp_reg1 = mono_alloc_ireg (cfg);
3358 int temp_reg2 = mono_alloc_ireg (cfg);
3359 int original_reg = ins->sreg1;
3361 NEW_INS (cfg, ins, temp, OP_ICONV_TO_U1);
3362 temp->sreg1 = original_reg;
3363 temp->dreg = temp_reg1;
3365 NEW_INS (cfg, ins, temp, OP_SHL_IMM);
3366 temp->sreg1 = temp_reg1;
3367 temp->dreg = temp_reg2;
3370 NEW_INS (cfg, ins, temp, OP_LOR);
3371 temp->sreg1 = temp->dreg = temp_reg2;
3372 temp->sreg2 = temp_reg1;
3374 ins->opcode = OP_EXPAND_I2;
3375 ins->sreg1 = temp_reg2;
3384 bb->max_vreg = cfg->next_vreg;
3388 branch_cc_table [] = {
3389 X86_CC_EQ, X86_CC_GE, X86_CC_GT, X86_CC_LE, X86_CC_LT,
3390 X86_CC_NE, X86_CC_GE, X86_CC_GT, X86_CC_LE, X86_CC_LT,
3391 X86_CC_O, X86_CC_NO, X86_CC_C, X86_CC_NC
3394 /* Maps CMP_... constants to X86_CC_... constants */
3397 X86_CC_EQ, X86_CC_NE, X86_CC_LE, X86_CC_GE, X86_CC_LT, X86_CC_GT,
3398 X86_CC_LE, X86_CC_GE, X86_CC_LT, X86_CC_GT
3402 cc_signed_table [] = {
3403 TRUE, TRUE, TRUE, TRUE, TRUE, TRUE,
3404 FALSE, FALSE, FALSE, FALSE
3407 /*#include "cprop.c"*/
3409 static unsigned char*
3410 emit_float_to_int (MonoCompile *cfg, guchar *code, int dreg, int sreg, int size, gboolean is_signed)
3412 amd64_sse_cvttsd2si_reg_reg (code, dreg, sreg);
3415 amd64_widen_reg (code, dreg, dreg, is_signed, FALSE);
3417 amd64_widen_reg (code, dreg, dreg, is_signed, TRUE);
3421 static unsigned char*
3422 mono_emit_stack_alloc (MonoCompile *cfg, guchar *code, MonoInst* tree)
3424 int sreg = tree->sreg1;
3425 int need_touch = FALSE;
3427 #if defined(HOST_WIN32) || defined(MONO_ARCH_SIGSEGV_ON_ALTSTACK)
3428 if (!tree->flags & MONO_INST_INIT)
3437 * If requested stack size is larger than one page,
3438 * perform stack-touch operation
3441 * Generate stack probe code.
3442 * Under Windows, it is necessary to allocate one page at a time,
3443 * "touching" stack after each successful sub-allocation. This is
3444 * because of the way stack growth is implemented - there is a
3445 * guard page before the lowest stack page that is currently commited.
3446 * Stack normally grows sequentially so OS traps access to the
3447 * guard page and commits more pages when needed.
3449 amd64_test_reg_imm (code, sreg, ~0xFFF);
3450 br[0] = code; x86_branch8 (code, X86_CC_Z, 0, FALSE);
3452 br[2] = code; /* loop */
3453 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, 0x1000);
3454 amd64_test_membase_reg (code, AMD64_RSP, 0, AMD64_RSP);
3455 amd64_alu_reg_imm (code, X86_SUB, sreg, 0x1000);
3456 amd64_alu_reg_imm (code, X86_CMP, sreg, 0x1000);
3457 br[3] = code; x86_branch8 (code, X86_CC_AE, 0, FALSE);
3458 amd64_patch (br[3], br[2]);
3459 amd64_test_reg_reg (code, sreg, sreg);
3460 br[4] = code; x86_branch8 (code, X86_CC_Z, 0, FALSE);
3461 amd64_alu_reg_reg (code, X86_SUB, AMD64_RSP, sreg);
3463 br[1] = code; x86_jump8 (code, 0);
3465 amd64_patch (br[0], code);
3466 amd64_alu_reg_reg (code, X86_SUB, AMD64_RSP, sreg);
3467 amd64_patch (br[1], code);
3468 amd64_patch (br[4], code);
3471 amd64_alu_reg_reg (code, X86_SUB, AMD64_RSP, tree->sreg1);
3473 if (tree->flags & MONO_INST_INIT) {
3475 if (tree->dreg != AMD64_RAX && sreg != AMD64_RAX) {
3476 amd64_push_reg (code, AMD64_RAX);
3479 if (tree->dreg != AMD64_RCX && sreg != AMD64_RCX) {
3480 amd64_push_reg (code, AMD64_RCX);
3483 if (tree->dreg != AMD64_RDI && sreg != AMD64_RDI) {
3484 amd64_push_reg (code, AMD64_RDI);
3488 amd64_shift_reg_imm (code, X86_SHR, sreg, 3);
3489 if (sreg != AMD64_RCX)
3490 amd64_mov_reg_reg (code, AMD64_RCX, sreg, 8);
3491 amd64_alu_reg_reg (code, X86_XOR, AMD64_RAX, AMD64_RAX);
3493 amd64_lea_membase (code, AMD64_RDI, AMD64_RSP, offset);
3494 if (cfg->param_area && cfg->arch.no_pushes)
3495 amd64_alu_reg_imm (code, X86_ADD, AMD64_RDI, cfg->param_area);
3497 #if defined(__default_codegen__)
3498 amd64_prefix (code, X86_REP_PREFIX);
3500 #elif defined(__native_client_codegen__)
3501 /* NaCl stos pseudo-instruction */
3502 amd64_codegen_pre(code);
3503 /* First, clear the upper 32 bits of RDI (mov %edi, %edi) */
3504 amd64_mov_reg_reg (code, AMD64_RDI, AMD64_RDI, 4);
3505 /* Add %r15 to %rdi using lea, condition flags unaffected. */
3506 amd64_lea_memindex_size (code, AMD64_RDI, AMD64_R15, 0, AMD64_RDI, 0, 8);
3507 amd64_prefix (code, X86_REP_PREFIX);
3509 amd64_codegen_post(code);
3510 #endif /* __native_client_codegen__ */
3512 if (tree->dreg != AMD64_RDI && sreg != AMD64_RDI)
3513 amd64_pop_reg (code, AMD64_RDI);
3514 if (tree->dreg != AMD64_RCX && sreg != AMD64_RCX)
3515 amd64_pop_reg (code, AMD64_RCX);
3516 if (tree->dreg != AMD64_RAX && sreg != AMD64_RAX)
3517 amd64_pop_reg (code, AMD64_RAX);
3523 emit_move_return_value (MonoCompile *cfg, MonoInst *ins, guint8 *code)
3528 /* Move return value to the target register */
3529 /* FIXME: do this in the local reg allocator */
3530 switch (ins->opcode) {
3533 case OP_CALL_MEMBASE:
3536 case OP_LCALL_MEMBASE:
3537 g_assert (ins->dreg == AMD64_RAX);
3541 case OP_FCALL_MEMBASE:
3542 if (((MonoCallInst*)ins)->signature->ret->type == MONO_TYPE_R4) {
3543 amd64_sse_cvtss2sd_reg_reg (code, ins->dreg, AMD64_XMM0);
3546 if (ins->dreg != AMD64_XMM0)
3547 amd64_sse_movsd_reg_reg (code, ins->dreg, AMD64_XMM0);
3552 case OP_VCALL_MEMBASE:
3555 case OP_VCALL2_MEMBASE:
3556 cinfo = get_call_info (cfg->generic_sharing_context, cfg->mempool, ((MonoCallInst*)ins)->signature);
3557 if (cinfo->ret.storage == ArgValuetypeInReg) {
3558 MonoInst *loc = cfg->arch.vret_addr_loc;
3560 /* Load the destination address */
3561 g_assert (loc->opcode == OP_REGOFFSET);
3562 amd64_mov_reg_membase (code, AMD64_RCX, loc->inst_basereg, loc->inst_offset, sizeof(gpointer));
3564 for (quad = 0; quad < 2; quad ++) {
3565 switch (cinfo->ret.pair_storage [quad]) {
3567 amd64_mov_membase_reg (code, AMD64_RCX, (quad * sizeof(mgreg_t)), cinfo->ret.pair_regs [quad], sizeof(mgreg_t));
3569 case ArgInFloatSSEReg:
3570 amd64_movss_membase_reg (code, AMD64_RCX, (quad * 8), cinfo->ret.pair_regs [quad]);
3572 case ArgInDoubleSSEReg:
3573 amd64_movsd_membase_reg (code, AMD64_RCX, (quad * 8), cinfo->ret.pair_regs [quad]);
3588 #endif /* DISABLE_JIT */
3591 static int tls_gs_offset;
3595 mono_amd64_have_tls_get (void)
3598 static gboolean have_tls_get = FALSE;
3599 static gboolean inited = FALSE;
3602 return have_tls_get;
3604 guint8 *ins = (guint8*)pthread_getspecific;
3607 * We're looking for these two instructions:
3609 * mov %gs:[offset](,%rdi,8),%rax
3612 have_tls_get = ins [0] == 0x65 &&
3624 tls_gs_offset = ins[5];
3626 return have_tls_get;
3633 * mono_amd64_emit_tls_get:
3634 * @code: buffer to store code to
3635 * @dreg: hard register where to place the result
3636 * @tls_offset: offset info
3638 * mono_amd64_emit_tls_get emits in @code the native code that puts in
3639 * the dreg register the item in the thread local storage identified
3642 * Returns: a pointer to the end of the stored code
3645 mono_amd64_emit_tls_get (guint8* code, int dreg, int tls_offset)
3648 g_assert (tls_offset < 64);
3649 x86_prefix (code, X86_GS_PREFIX);
3650 amd64_mov_reg_mem (code, dreg, (tls_offset * 8) + 0x1480, 8);
3651 #elif defined(__APPLE__)
3652 x86_prefix (code, X86_GS_PREFIX);
3653 amd64_mov_reg_mem (code, dreg, tls_gs_offset + (tls_offset * 8), 8);
3655 if (optimize_for_xen) {
3656 x86_prefix (code, X86_FS_PREFIX);
3657 amd64_mov_reg_mem (code, dreg, 0, 8);
3658 amd64_mov_reg_membase (code, dreg, dreg, tls_offset, 8);
3660 x86_prefix (code, X86_FS_PREFIX);
3661 amd64_mov_reg_mem (code, dreg, tls_offset, 8);
3670 * Emit code to initialize an LMF structure at LMF_OFFSET.
3673 emit_setup_lmf (MonoCompile *cfg, guint8 *code, gint32 lmf_offset, int cfa_offset)
3678 * The ip field is not set, the exception handling code will obtain it from the stack location pointed to by the sp field.
3681 * sp is saved right before calls but we need to save it here too so
3682 * async stack walks would work.
3684 amd64_mov_membase_reg (code, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rsp), AMD64_RSP, 8);
3685 /* Skip method (only needed for trampoline LMF frames) */
3686 /* Save callee saved regs */
3687 for (i = 0; i < MONO_MAX_IREGS; ++i) {
3691 case AMD64_RBX: offset = G_STRUCT_OFFSET (MonoLMF, rbx); break;
3692 case AMD64_RBP: offset = G_STRUCT_OFFSET (MonoLMF, rbp); break;
3693 case AMD64_R12: offset = G_STRUCT_OFFSET (MonoLMF, r12); break;
3694 case AMD64_R13: offset = G_STRUCT_OFFSET (MonoLMF, r13); break;
3695 case AMD64_R14: offset = G_STRUCT_OFFSET (MonoLMF, r14); break;
3696 #ifndef __native_client_codegen__
3697 case AMD64_R15: offset = G_STRUCT_OFFSET (MonoLMF, r15); break;
3700 case AMD64_RDI: offset = G_STRUCT_OFFSET (MonoLMF, rdi); break;
3701 case AMD64_RSI: offset = G_STRUCT_OFFSET (MonoLMF, rsi); break;
3709 amd64_mov_membase_reg (code, cfg->frame_reg, lmf_offset + offset, i, 8);
3710 if ((cfg->arch.omit_fp || (i != AMD64_RBP)) && cfa_offset != -1)
3711 mono_emit_unwind_op_offset (cfg, code, i, - (cfa_offset - (lmf_offset + offset)));
3715 /* These can't contain refs */
3716 mini_gc_set_slot_type_from_fp (cfg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, previous_lmf), SLOT_NOREF);
3717 mini_gc_set_slot_type_from_fp (cfg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, lmf_addr), SLOT_NOREF);
3718 mini_gc_set_slot_type_from_fp (cfg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, method), SLOT_NOREF);
3719 mini_gc_set_slot_type_from_fp (cfg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rip), SLOT_NOREF);
3720 mini_gc_set_slot_type_from_fp (cfg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rsp), SLOT_NOREF);
3722 /* These are handled automatically by the stack marking code */
3723 mini_gc_set_slot_type_from_fp (cfg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rbx), SLOT_NOREF);
3724 mini_gc_set_slot_type_from_fp (cfg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rbp), SLOT_NOREF);
3725 mini_gc_set_slot_type_from_fp (cfg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r12), SLOT_NOREF);
3726 mini_gc_set_slot_type_from_fp (cfg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r13), SLOT_NOREF);
3727 mini_gc_set_slot_type_from_fp (cfg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r14), SLOT_NOREF);
3728 mini_gc_set_slot_type_from_fp (cfg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r15), SLOT_NOREF);
3730 mini_gc_set_slot_type_from_fp (cfg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rdi), SLOT_NOREF);
3731 mini_gc_set_slot_type_from_fp (cfg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rsi), SLOT_NOREF);
3740 * Emit code to push an LMF structure on the LMF stack.
3743 emit_save_lmf (MonoCompile *cfg, guint8 *code, gint32 lmf_offset, gboolean *args_clobbered)
3745 if ((lmf_tls_offset != -1) && !optimize_for_xen) {
3747 * Optimized version which uses the mono_lmf TLS variable instead of
3748 * indirection through the mono_lmf_addr TLS variable.
3750 /* %rax = previous_lmf */
3751 x86_prefix (code, X86_FS_PREFIX);
3752 amd64_mov_reg_mem (code, AMD64_RAX, lmf_tls_offset, 8);
3754 /* Save previous_lmf */
3755 amd64_mov_membase_reg (code, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, previous_lmf), AMD64_RAX, 8);
3757 if (lmf_offset == 0) {
3758 x86_prefix (code, X86_FS_PREFIX);
3759 amd64_mov_mem_reg (code, lmf_tls_offset, cfg->frame_reg, 8);
3761 amd64_lea_membase (code, AMD64_R11, cfg->frame_reg, lmf_offset);
3762 x86_prefix (code, X86_FS_PREFIX);
3763 amd64_mov_mem_reg (code, lmf_tls_offset, AMD64_R11, 8);
3766 if (lmf_addr_tls_offset != -1) {
3767 /* Load lmf quicky using the FS register */
3768 code = mono_amd64_emit_tls_get (code, AMD64_RAX, lmf_addr_tls_offset);
3770 /* The TLS key actually contains a pointer to the MonoJitTlsData structure */
3771 /* FIXME: Add a separate key for LMF to avoid this */
3772 amd64_alu_reg_imm (code, X86_ADD, AMD64_RAX, G_STRUCT_OFFSET (MonoJitTlsData, lmf));
3777 * The call might clobber argument registers, but they are already
3778 * saved to the stack/global regs.
3781 *args_clobbered = TRUE;
3782 code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD,
3783 (gpointer)"mono_get_lmf_addr", TRUE);
3787 amd64_mov_membase_reg (code, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, lmf_addr), AMD64_RAX, sizeof(gpointer));
3788 /* Save previous_lmf */
3789 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RAX, 0, sizeof(gpointer));
3790 amd64_mov_membase_reg (code, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, previous_lmf), AMD64_R11, sizeof(gpointer));
3792 amd64_lea_membase (code, AMD64_R11, cfg->frame_reg, lmf_offset);
3793 amd64_mov_membase_reg (code, AMD64_RAX, 0, AMD64_R11, sizeof(gpointer));
3802 * Emit code to pop an LMF structure from the LMF stack.
3805 emit_restore_lmf (MonoCompile *cfg, guint8 *code, gint32 lmf_offset)
3807 if ((lmf_tls_offset != -1) && !optimize_for_xen) {
3809 * Optimized version which uses the mono_lmf TLS variable instead of indirection
3810 * through the mono_lmf_addr TLS variable.
3812 /* reg = previous_lmf */
3813 amd64_mov_reg_membase (code, AMD64_R11, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, previous_lmf), sizeof(gpointer));
3814 x86_prefix (code, X86_FS_PREFIX);
3815 amd64_mov_mem_reg (code, lmf_tls_offset, AMD64_R11, 8);
3817 /* Restore previous lmf */
3818 amd64_mov_reg_membase (code, AMD64_RCX, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, previous_lmf), sizeof(gpointer));
3819 amd64_mov_reg_membase (code, AMD64_R11, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, lmf_addr), sizeof(gpointer));
3820 amd64_mov_membase_reg (code, AMD64_R11, 0, AMD64_RCX, sizeof(gpointer));
3826 #define REAL_PRINT_REG(text,reg) \
3827 mono_assert (reg >= 0); \
3828 amd64_push_reg (code, AMD64_RAX); \
3829 amd64_push_reg (code, AMD64_RDX); \
3830 amd64_push_reg (code, AMD64_RCX); \
3831 amd64_push_reg (code, reg); \
3832 amd64_push_imm (code, reg); \
3833 amd64_push_imm (code, text " %d %p\n"); \
3834 amd64_mov_reg_imm (code, AMD64_RAX, printf); \
3835 amd64_call_reg (code, AMD64_RAX); \
3836 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, 3*4); \
3837 amd64_pop_reg (code, AMD64_RCX); \
3838 amd64_pop_reg (code, AMD64_RDX); \
3839 amd64_pop_reg (code, AMD64_RAX);
3841 /* benchmark and set based on cpu */
3842 #define LOOP_ALIGNMENT 8
3843 #define bb_is_loop_start(bb) ((bb)->loop_body_start && (bb)->nesting)
3847 #if defined(__native_client__) || defined(__native_client_codegen__)
3850 #ifdef __native_client_gc__
3851 __nacl_suspend_thread_if_needed();
3857 mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
3862 guint8 *code = cfg->native_code + cfg->code_len;
3863 MonoInst *last_ins = NULL;
3864 guint last_offset = 0;
3867 /* Fix max_offset estimate for each successor bb */
3868 if (cfg->opt & MONO_OPT_BRANCH) {
3869 int current_offset = cfg->code_len;
3870 MonoBasicBlock *current_bb;
3871 for (current_bb = bb; current_bb != NULL; current_bb = current_bb->next_bb) {
3872 current_bb->max_offset = current_offset;
3873 current_offset += current_bb->max_length;
3877 if (cfg->opt & MONO_OPT_LOOP) {
3878 int pad, align = LOOP_ALIGNMENT;
3879 /* set alignment depending on cpu */
3880 if (bb_is_loop_start (bb) && (pad = (cfg->code_len & (align - 1)))) {
3882 /*g_print ("adding %d pad at %x to loop in %s\n", pad, cfg->code_len, cfg->method->name);*/
3883 amd64_padding (code, pad);
3884 cfg->code_len += pad;
3885 bb->native_offset = cfg->code_len;
3889 #if defined(__native_client_codegen__)
3890 /* For Native Client, all indirect call/jump targets must be */
3891 /* 32-byte aligned. Exception handler blocks are jumped to */
3892 /* indirectly as well. */
3893 gboolean bb_needs_alignment = (bb->flags & BB_INDIRECT_JUMP_TARGET) ||
3894 (bb->flags & BB_EXCEPTION_HANDLER);
3896 if ( bb_needs_alignment && ((cfg->code_len & kNaClAlignmentMask) != 0)) {
3897 int pad = kNaClAlignment - (cfg->code_len & kNaClAlignmentMask);
3898 if (pad != kNaClAlignment) code = mono_arch_nacl_pad(code, pad);
3899 cfg->code_len += pad;
3900 bb->native_offset = cfg->code_len;
3902 #endif /*__native_client_codegen__*/
3904 if (cfg->verbose_level > 2)
3905 g_print ("Basic block %d starting at offset 0x%x\n", bb->block_num, bb->native_offset);
3907 if (cfg->prof_options & MONO_PROFILE_COVERAGE) {
3908 MonoProfileCoverageInfo *cov = cfg->coverage_info;
3909 g_assert (!cfg->compile_aot);
3911 cov->data [bb->dfn].cil_code = bb->cil_code;
3912 amd64_mov_reg_imm (code, AMD64_R11, (guint64)&cov->data [bb->dfn].count);
3913 /* this is not thread save, but good enough */
3914 amd64_inc_membase (code, AMD64_R11, 0);
3917 offset = code - cfg->native_code;
3919 mono_debug_open_block (cfg, bb, offset);
3921 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)
3922 x86_breakpoint (code);
3924 MONO_BB_FOR_EACH_INS (bb, ins) {
3925 offset = code - cfg->native_code;
3927 max_len = ((guint8 *)ins_get_spec (ins->opcode))[MONO_INST_LEN];
3929 #define EXTRA_CODE_SPACE (NACL_SIZE (16, 16 + kNaClAlignment))
3931 if (G_UNLIKELY (offset > (cfg->code_size - max_len - EXTRA_CODE_SPACE))) {
3932 cfg->code_size *= 2;
3933 cfg->native_code = mono_realloc_native_code(cfg);
3934 code = cfg->native_code + offset;
3935 cfg->stat_code_reallocs++;
3938 if (cfg->debug_info)
3939 mono_debug_record_line_number (cfg, ins, offset);
3941 switch (ins->opcode) {
3943 amd64_mul_reg (code, ins->sreg2, TRUE);
3946 amd64_mul_reg (code, ins->sreg2, FALSE);
3948 case OP_X86_SETEQ_MEMBASE:
3949 amd64_set_membase (code, X86_CC_EQ, ins->inst_basereg, ins->inst_offset, TRUE);
3951 case OP_STOREI1_MEMBASE_IMM:
3952 amd64_mov_membase_imm (code, ins->inst_destbasereg, ins->inst_offset, ins->inst_imm, 1);
3954 case OP_STOREI2_MEMBASE_IMM:
3955 amd64_mov_membase_imm (code, ins->inst_destbasereg, ins->inst_offset, ins->inst_imm, 2);
3957 case OP_STOREI4_MEMBASE_IMM:
3958 amd64_mov_membase_imm (code, ins->inst_destbasereg, ins->inst_offset, ins->inst_imm, 4);
3960 case OP_STOREI1_MEMBASE_REG:
3961 amd64_mov_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, ins->sreg1, 1);
3963 case OP_STOREI2_MEMBASE_REG:
3964 amd64_mov_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, ins->sreg1, 2);
3966 /* In AMD64 NaCl, pointers are 4 bytes, */
3967 /* so STORE_* != STOREI8_*. Likewise below. */
3968 case OP_STORE_MEMBASE_REG:
3969 amd64_mov_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, ins->sreg1, sizeof(gpointer));
3971 case OP_STOREI8_MEMBASE_REG:
3972 amd64_mov_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, ins->sreg1, 8);
3974 case OP_STOREI4_MEMBASE_REG:
3975 amd64_mov_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, ins->sreg1, 4);
3977 case OP_STORE_MEMBASE_IMM:
3978 #ifndef __native_client_codegen__
3979 /* In NaCl, this could be a PCONST type, which could */
3980 /* mean a pointer type was copied directly into the */
3981 /* lower 32-bits of inst_imm, so for InvalidPtr==-1 */
3982 /* the value would be 0x00000000FFFFFFFF which is */
3983 /* not proper for an imm32 unless you cast it. */
3984 g_assert (amd64_is_imm32 (ins->inst_imm));
3986 amd64_mov_membase_imm (code, ins->inst_destbasereg, ins->inst_offset, (gint32)ins->inst_imm, sizeof(gpointer));
3988 case OP_STOREI8_MEMBASE_IMM:
3989 g_assert (amd64_is_imm32 (ins->inst_imm));
3990 amd64_mov_membase_imm (code, ins->inst_destbasereg, ins->inst_offset, ins->inst_imm, 8);
3993 #ifdef __mono_ilp32__
3994 /* In ILP32, pointers are 4 bytes, so separate these */
3995 /* cases, use literal 8 below where we really want 8 */
3996 amd64_mov_reg_imm (code, ins->dreg, ins->inst_imm);
3997 amd64_mov_reg_membase (code, ins->dreg, ins->dreg, 0, sizeof(gpointer));
4001 // FIXME: Decompose this earlier
4002 if (amd64_is_imm32 (ins->inst_imm))
4003 amd64_mov_reg_mem (code, ins->dreg, ins->inst_imm, 8);
4005 amd64_mov_reg_imm (code, ins->dreg, ins->inst_imm);
4006 amd64_mov_reg_membase (code, ins->dreg, ins->dreg, 0, 8);
4010 amd64_mov_reg_imm (code, ins->dreg, ins->inst_imm);
4011 amd64_movsxd_reg_membase (code, ins->dreg, ins->dreg, 0);
4014 // FIXME: Decompose this earlier
4015 if (amd64_is_imm32 (ins->inst_imm))
4016 amd64_mov_reg_mem (code, ins->dreg, ins->inst_imm, 4);
4018 amd64_mov_reg_imm (code, ins->dreg, ins->inst_imm);
4019 amd64_mov_reg_membase (code, ins->dreg, ins->dreg, 0, 4);
4023 amd64_mov_reg_imm (code, ins->dreg, ins->inst_imm);
4024 amd64_widen_membase (code, ins->dreg, ins->dreg, 0, FALSE, FALSE);
4027 /* For NaCl, pointers are 4 bytes, so separate these */
4028 /* cases, use literal 8 below where we really want 8 */
4029 amd64_mov_reg_imm (code, ins->dreg, ins->inst_imm);
4030 amd64_widen_membase (code, ins->dreg, ins->dreg, 0, FALSE, TRUE);
4032 case OP_LOAD_MEMBASE:
4033 g_assert (amd64_is_imm32 (ins->inst_offset));
4034 amd64_mov_reg_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, sizeof(gpointer));
4036 case OP_LOADI8_MEMBASE:
4037 /* Use literal 8 instead of sizeof pointer or */
4038 /* register, we really want 8 for this opcode */
4039 g_assert (amd64_is_imm32 (ins->inst_offset));
4040 amd64_mov_reg_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, 8);
4042 case OP_LOADI4_MEMBASE:
4043 amd64_movsxd_reg_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset);
4045 case OP_LOADU4_MEMBASE:
4046 amd64_mov_reg_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, 4);
4048 case OP_LOADU1_MEMBASE:
4049 /* The cpu zero extends the result into 64 bits */
4050 amd64_widen_membase_size (code, ins->dreg, ins->inst_basereg, ins->inst_offset, FALSE, FALSE, 4);
4052 case OP_LOADI1_MEMBASE:
4053 amd64_widen_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, TRUE, FALSE);
4055 case OP_LOADU2_MEMBASE:
4056 /* The cpu zero extends the result into 64 bits */
4057 amd64_widen_membase_size (code, ins->dreg, ins->inst_basereg, ins->inst_offset, FALSE, TRUE, 4);
4059 case OP_LOADI2_MEMBASE:
4060 amd64_widen_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset, TRUE, TRUE);
4062 case OP_AMD64_LOADI8_MEMINDEX:
4063 amd64_mov_reg_memindex_size (code, ins->dreg, ins->inst_basereg, 0, ins->inst_indexreg, 0, 8);
4065 case OP_LCONV_TO_I1:
4066 case OP_ICONV_TO_I1:
4068 amd64_widen_reg (code, ins->dreg, ins->sreg1, TRUE, FALSE);
4070 case OP_LCONV_TO_I2:
4071 case OP_ICONV_TO_I2:
4073 amd64_widen_reg (code, ins->dreg, ins->sreg1, TRUE, TRUE);
4075 case OP_LCONV_TO_U1:
4076 case OP_ICONV_TO_U1:
4077 amd64_widen_reg (code, ins->dreg, ins->sreg1, FALSE, FALSE);
4079 case OP_LCONV_TO_U2:
4080 case OP_ICONV_TO_U2:
4081 amd64_widen_reg (code, ins->dreg, ins->sreg1, FALSE, TRUE);
4084 /* Clean out the upper word */
4085 amd64_mov_reg_reg_size (code, ins->dreg, ins->sreg1, 4);
4088 amd64_movsxd_reg_reg (code, ins->dreg, ins->sreg1);
4092 amd64_alu_reg_reg (code, X86_CMP, ins->sreg1, ins->sreg2);
4094 case OP_COMPARE_IMM:
4095 case OP_LCOMPARE_IMM:
4096 g_assert (amd64_is_imm32 (ins->inst_imm));
4097 amd64_alu_reg_imm (code, X86_CMP, ins->sreg1, ins->inst_imm);
4099 case OP_X86_COMPARE_REG_MEMBASE:
4100 amd64_alu_reg_membase (code, X86_CMP, ins->sreg1, ins->sreg2, ins->inst_offset);
4102 case OP_X86_TEST_NULL:
4103 amd64_test_reg_reg_size (code, ins->sreg1, ins->sreg1, 4);
4105 case OP_AMD64_TEST_NULL:
4106 amd64_test_reg_reg (code, ins->sreg1, ins->sreg1);
4109 case OP_X86_ADD_REG_MEMBASE:
4110 amd64_alu_reg_membase_size (code, X86_ADD, ins->sreg1, ins->sreg2, ins->inst_offset, 4);
4112 case OP_X86_SUB_REG_MEMBASE:
4113 amd64_alu_reg_membase_size (code, X86_SUB, ins->sreg1, ins->sreg2, ins->inst_offset, 4);
4115 case OP_X86_AND_REG_MEMBASE:
4116 amd64_alu_reg_membase_size (code, X86_AND, ins->sreg1, ins->sreg2, ins->inst_offset, 4);
4118 case OP_X86_OR_REG_MEMBASE:
4119 amd64_alu_reg_membase_size (code, X86_OR, ins->sreg1, ins->sreg2, ins->inst_offset, 4);
4121 case OP_X86_XOR_REG_MEMBASE:
4122 amd64_alu_reg_membase_size (code, X86_XOR, ins->sreg1, ins->sreg2, ins->inst_offset, 4);
4125 case OP_X86_ADD_MEMBASE_IMM:
4126 /* FIXME: Make a 64 version too */
4127 amd64_alu_membase_imm_size (code, X86_ADD, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 4);
4129 case OP_X86_SUB_MEMBASE_IMM:
4130 g_assert (amd64_is_imm32 (ins->inst_imm));
4131 amd64_alu_membase_imm_size (code, X86_SUB, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 4);
4133 case OP_X86_AND_MEMBASE_IMM:
4134 g_assert (amd64_is_imm32 (ins->inst_imm));
4135 amd64_alu_membase_imm_size (code, X86_AND, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 4);
4137 case OP_X86_OR_MEMBASE_IMM:
4138 g_assert (amd64_is_imm32 (ins->inst_imm));
4139 amd64_alu_membase_imm_size (code, X86_OR, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 4);
4141 case OP_X86_XOR_MEMBASE_IMM:
4142 g_assert (amd64_is_imm32 (ins->inst_imm));
4143 amd64_alu_membase_imm_size (code, X86_XOR, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 4);
4145 case OP_X86_ADD_MEMBASE_REG:
4146 amd64_alu_membase_reg_size (code, X86_ADD, ins->inst_basereg, ins->inst_offset, ins->sreg2, 4);
4148 case OP_X86_SUB_MEMBASE_REG:
4149 amd64_alu_membase_reg_size (code, X86_SUB, ins->inst_basereg, ins->inst_offset, ins->sreg2, 4);
4151 case OP_X86_AND_MEMBASE_REG:
4152 amd64_alu_membase_reg_size (code, X86_AND, ins->inst_basereg, ins->inst_offset, ins->sreg2, 4);
4154 case OP_X86_OR_MEMBASE_REG:
4155 amd64_alu_membase_reg_size (code, X86_OR, ins->inst_basereg, ins->inst_offset, ins->sreg2, 4);
4157 case OP_X86_XOR_MEMBASE_REG:
4158 amd64_alu_membase_reg_size (code, X86_XOR, ins->inst_basereg, ins->inst_offset, ins->sreg2, 4);
4160 case OP_X86_INC_MEMBASE:
4161 amd64_inc_membase_size (code, ins->inst_basereg, ins->inst_offset, 4);
4163 case OP_X86_INC_REG:
4164 amd64_inc_reg_size (code, ins->dreg, 4);
4166 case OP_X86_DEC_MEMBASE:
4167 amd64_dec_membase_size (code, ins->inst_basereg, ins->inst_offset, 4);
4169 case OP_X86_DEC_REG:
4170 amd64_dec_reg_size (code, ins->dreg, 4);
4172 case OP_X86_MUL_REG_MEMBASE:
4173 case OP_X86_MUL_MEMBASE_REG:
4174 amd64_imul_reg_membase_size (code, ins->sreg1, ins->sreg2, ins->inst_offset, 4);
4176 case OP_AMD64_ICOMPARE_MEMBASE_REG:
4177 amd64_alu_membase_reg_size (code, X86_CMP, ins->inst_basereg, ins->inst_offset, ins->sreg2, 4);
4179 case OP_AMD64_ICOMPARE_MEMBASE_IMM:
4180 amd64_alu_membase_imm_size (code, X86_CMP, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 4);
4182 case OP_AMD64_COMPARE_MEMBASE_REG:
4183 amd64_alu_membase_reg_size (code, X86_CMP, ins->inst_basereg, ins->inst_offset, ins->sreg2, 8);
4185 case OP_AMD64_COMPARE_MEMBASE_IMM:
4186 g_assert (amd64_is_imm32 (ins->inst_imm));
4187 amd64_alu_membase_imm_size (code, X86_CMP, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 8);
4189 case OP_X86_COMPARE_MEMBASE8_IMM:
4190 amd64_alu_membase8_imm_size (code, X86_CMP, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 4);
4192 case OP_AMD64_ICOMPARE_REG_MEMBASE:
4193 amd64_alu_reg_membase_size (code, X86_CMP, ins->sreg1, ins->sreg2, ins->inst_offset, 4);
4195 case OP_AMD64_COMPARE_REG_MEMBASE:
4196 amd64_alu_reg_membase_size (code, X86_CMP, ins->sreg1, ins->sreg2, ins->inst_offset, 8);
4199 case OP_AMD64_ADD_REG_MEMBASE:
4200 amd64_alu_reg_membase_size (code, X86_ADD, ins->sreg1, ins->sreg2, ins->inst_offset, 8);
4202 case OP_AMD64_SUB_REG_MEMBASE:
4203 amd64_alu_reg_membase_size (code, X86_SUB, ins->sreg1, ins->sreg2, ins->inst_offset, 8);
4205 case OP_AMD64_AND_REG_MEMBASE:
4206 amd64_alu_reg_membase_size (code, X86_AND, ins->sreg1, ins->sreg2, ins->inst_offset, 8);
4208 case OP_AMD64_OR_REG_MEMBASE:
4209 amd64_alu_reg_membase_size (code, X86_OR, ins->sreg1, ins->sreg2, ins->inst_offset, 8);
4211 case OP_AMD64_XOR_REG_MEMBASE:
4212 amd64_alu_reg_membase_size (code, X86_XOR, ins->sreg1, ins->sreg2, ins->inst_offset, 8);
4215 case OP_AMD64_ADD_MEMBASE_REG:
4216 amd64_alu_membase_reg_size (code, X86_ADD, ins->inst_basereg, ins->inst_offset, ins->sreg2, 8);
4218 case OP_AMD64_SUB_MEMBASE_REG:
4219 amd64_alu_membase_reg_size (code, X86_SUB, ins->inst_basereg, ins->inst_offset, ins->sreg2, 8);
4221 case OP_AMD64_AND_MEMBASE_REG:
4222 amd64_alu_membase_reg_size (code, X86_AND, ins->inst_basereg, ins->inst_offset, ins->sreg2, 8);
4224 case OP_AMD64_OR_MEMBASE_REG:
4225 amd64_alu_membase_reg_size (code, X86_OR, ins->inst_basereg, ins->inst_offset, ins->sreg2, 8);
4227 case OP_AMD64_XOR_MEMBASE_REG:
4228 amd64_alu_membase_reg_size (code, X86_XOR, ins->inst_basereg, ins->inst_offset, ins->sreg2, 8);
4231 case OP_AMD64_ADD_MEMBASE_IMM:
4232 g_assert (amd64_is_imm32 (ins->inst_imm));
4233 amd64_alu_membase_imm_size (code, X86_ADD, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 8);
4235 case OP_AMD64_SUB_MEMBASE_IMM:
4236 g_assert (amd64_is_imm32 (ins->inst_imm));
4237 amd64_alu_membase_imm_size (code, X86_SUB, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 8);
4239 case OP_AMD64_AND_MEMBASE_IMM:
4240 g_assert (amd64_is_imm32 (ins->inst_imm));
4241 amd64_alu_membase_imm_size (code, X86_AND, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 8);
4243 case OP_AMD64_OR_MEMBASE_IMM:
4244 g_assert (amd64_is_imm32 (ins->inst_imm));
4245 amd64_alu_membase_imm_size (code, X86_OR, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 8);
4247 case OP_AMD64_XOR_MEMBASE_IMM:
4248 g_assert (amd64_is_imm32 (ins->inst_imm));
4249 amd64_alu_membase_imm_size (code, X86_XOR, ins->inst_basereg, ins->inst_offset, ins->inst_imm, 8);
4253 amd64_breakpoint (code);
4255 case OP_RELAXED_NOP:
4256 x86_prefix (code, X86_REP_PREFIX);
4264 case OP_DUMMY_STORE:
4265 case OP_NOT_REACHED:
4268 case OP_SEQ_POINT: {
4271 if (cfg->compile_aot)
4275 * Read from the single stepping trigger page. This will cause a
4276 * SIGSEGV when single stepping is enabled.
4277 * We do this _before_ the breakpoint, so single stepping after
4278 * a breakpoint is hit will step to the next IL offset.
4280 if (ins->flags & MONO_INST_SINGLE_STEP_LOC) {
4281 if (((guint64)ss_trigger_page >> 32) == 0)
4282 amd64_mov_reg_mem (code, AMD64_R11, (guint64)ss_trigger_page, 4);
4284 MonoInst *var = cfg->arch.ss_trigger_page_var;
4286 amd64_mov_reg_membase (code, AMD64_R11, var->inst_basereg, var->inst_offset, 8);
4287 amd64_alu_membase_imm_size (code, X86_CMP, AMD64_R11, 0, 0, 4);
4292 * This is the address which is saved in seq points,
4293 * get_ip_for_single_step () / get_ip_for_breakpoint () needs to compute this
4294 * from the address of the instruction causing the fault.
4296 mono_add_seq_point (cfg, bb, ins, code - cfg->native_code);
4299 * A placeholder for a possible breakpoint inserted by
4300 * mono_arch_set_breakpoint ().
4302 for (i = 0; i < breakpoint_size; ++i)
4308 amd64_alu_reg_reg (code, X86_ADD, ins->sreg1, ins->sreg2);
4311 amd64_alu_reg_reg (code, X86_ADC, ins->sreg1, ins->sreg2);
4315 g_assert (amd64_is_imm32 (ins->inst_imm));
4316 amd64_alu_reg_imm (code, X86_ADD, ins->dreg, ins->inst_imm);
4319 g_assert (amd64_is_imm32 (ins->inst_imm));
4320 amd64_alu_reg_imm (code, X86_ADC, ins->dreg, ins->inst_imm);
4324 amd64_alu_reg_reg (code, X86_SUB, ins->sreg1, ins->sreg2);
4327 amd64_alu_reg_reg (code, X86_SBB, ins->sreg1, ins->sreg2);
4331 g_assert (amd64_is_imm32 (ins->inst_imm));
4332 amd64_alu_reg_imm (code, X86_SUB, ins->dreg, ins->inst_imm);
4335 g_assert (amd64_is_imm32 (ins->inst_imm));
4336 amd64_alu_reg_imm (code, X86_SBB, ins->dreg, ins->inst_imm);
4339 amd64_alu_reg_reg (code, X86_AND, ins->sreg1, ins->sreg2);
4343 g_assert (amd64_is_imm32 (ins->inst_imm));
4344 amd64_alu_reg_imm (code, X86_AND, ins->sreg1, ins->inst_imm);
4347 amd64_imul_reg_reg (code, ins->sreg1, ins->sreg2);
4352 guint32 size = (ins->opcode == OP_IMUL_IMM) ? 4 : 8;
4354 switch (ins->inst_imm) {
4358 if (ins->dreg != ins->sreg1)
4359 amd64_mov_reg_reg (code, ins->dreg, ins->sreg1, size);
4360 amd64_alu_reg_reg (code, X86_ADD, ins->dreg, ins->dreg);
4363 /* LEA r1, [r2 + r2*2] */
4364 amd64_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 1);
4367 /* LEA r1, [r2 + r2*4] */
4368 amd64_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 2);
4371 /* LEA r1, [r2 + r2*2] */
4373 amd64_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 1);
4374 amd64_alu_reg_reg (code, X86_ADD, ins->dreg, ins->dreg);
4377 /* LEA r1, [r2 + r2*8] */
4378 amd64_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 3);
4381 /* LEA r1, [r2 + r2*4] */
4383 amd64_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 2);
4384 amd64_alu_reg_reg (code, X86_ADD, ins->dreg, ins->dreg);
4387 /* LEA r1, [r2 + r2*2] */
4389 amd64_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 1);
4390 amd64_shift_reg_imm (code, X86_SHL, ins->dreg, 2);
4393 /* LEA r1, [r2 + r2*4] */
4394 /* LEA r1, [r1 + r1*4] */
4395 amd64_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 2);
4396 amd64_lea_memindex (code, ins->dreg, ins->dreg, 0, ins->dreg, 2);
4399 /* LEA r1, [r2 + r2*4] */
4401 /* LEA r1, [r1 + r1*4] */
4402 amd64_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 2);
4403 amd64_shift_reg_imm (code, X86_SHL, ins->dreg, 2);
4404 amd64_lea_memindex (code, ins->dreg, ins->dreg, 0, ins->dreg, 2);
4407 amd64_imul_reg_reg_imm_size (code, ins->dreg, ins->sreg1, ins->inst_imm, size);
4414 /* Regalloc magic makes the div/rem cases the same */
4415 if (ins->sreg2 == AMD64_RDX) {
4416 amd64_mov_membase_reg (code, AMD64_RSP, -8, AMD64_RDX, 8);
4418 amd64_div_membase (code, AMD64_RSP, -8, TRUE);
4421 amd64_div_reg (code, ins->sreg2, TRUE);
4426 if (ins->sreg2 == AMD64_RDX) {
4427 amd64_mov_membase_reg (code, AMD64_RSP, -8, AMD64_RDX, 8);
4428 amd64_alu_reg_reg (code, X86_XOR, AMD64_RDX, AMD64_RDX);
4429 amd64_div_membase (code, AMD64_RSP, -8, FALSE);
4431 amd64_alu_reg_reg (code, X86_XOR, AMD64_RDX, AMD64_RDX);
4432 amd64_div_reg (code, ins->sreg2, FALSE);
4437 if (ins->sreg2 == AMD64_RDX) {
4438 amd64_mov_membase_reg (code, AMD64_RSP, -8, AMD64_RDX, 8);
4439 amd64_cdq_size (code, 4);
4440 amd64_div_membase_size (code, AMD64_RSP, -8, TRUE, 4);
4442 amd64_cdq_size (code, 4);
4443 amd64_div_reg_size (code, ins->sreg2, TRUE, 4);
4448 if (ins->sreg2 == AMD64_RDX) {
4449 amd64_mov_membase_reg (code, AMD64_RSP, -8, AMD64_RDX, 8);
4450 amd64_alu_reg_reg (code, X86_XOR, AMD64_RDX, AMD64_RDX);
4451 amd64_div_membase_size (code, AMD64_RSP, -8, FALSE, 4);
4453 amd64_alu_reg_reg (code, X86_XOR, AMD64_RDX, AMD64_RDX);
4454 amd64_div_reg_size (code, ins->sreg2, FALSE, 4);
4458 int power = mono_is_power_of_two (ins->inst_imm);
4460 g_assert (ins->sreg1 == X86_EAX);
4461 g_assert (ins->dreg == X86_EAX);
4462 g_assert (power >= 0);
4465 amd64_mov_reg_imm (code, ins->dreg, 0);
4469 /* Based on gcc code */
4471 /* Add compensation for negative dividents */
4472 amd64_mov_reg_reg_size (code, AMD64_RDX, AMD64_RAX, 4);
4474 amd64_shift_reg_imm_size (code, X86_SAR, AMD64_RDX, 31, 4);
4475 amd64_shift_reg_imm_size (code, X86_SHR, AMD64_RDX, 32 - power, 4);
4476 amd64_alu_reg_reg_size (code, X86_ADD, AMD64_RAX, AMD64_RDX, 4);
4477 /* Compute remainder */
4478 amd64_alu_reg_imm_size (code, X86_AND, AMD64_RAX, (1 << power) - 1, 4);
4479 /* Remove compensation */
4480 amd64_alu_reg_reg_size (code, X86_SUB, AMD64_RAX, AMD64_RDX, 4);
4484 amd64_imul_reg_reg (code, ins->sreg1, ins->sreg2);
4485 EMIT_COND_SYSTEM_EXCEPTION (X86_CC_O, FALSE, "OverflowException");
4488 amd64_alu_reg_reg (code, X86_OR, ins->sreg1, ins->sreg2);
4492 g_assert (amd64_is_imm32 (ins->inst_imm));
4493 amd64_alu_reg_imm (code, X86_OR, ins->sreg1, ins->inst_imm);
4496 amd64_alu_reg_reg (code, X86_XOR, ins->sreg1, ins->sreg2);
4500 g_assert (amd64_is_imm32 (ins->inst_imm));
4501 amd64_alu_reg_imm (code, X86_XOR, ins->sreg1, ins->inst_imm);
4504 g_assert (ins->sreg2 == AMD64_RCX);
4505 amd64_shift_reg (code, X86_SHL, ins->dreg);
4508 g_assert (ins->sreg2 == AMD64_RCX);
4509 amd64_shift_reg (code, X86_SAR, ins->dreg);
4512 g_assert (amd64_is_imm32 (ins->inst_imm));
4513 amd64_shift_reg_imm_size (code, X86_SAR, ins->dreg, ins->inst_imm, 4);
4516 g_assert (amd64_is_imm32 (ins->inst_imm));
4517 amd64_shift_reg_imm (code, X86_SAR, ins->dreg, ins->inst_imm);
4520 g_assert (amd64_is_imm32 (ins->inst_imm));
4521 amd64_shift_reg_imm_size (code, X86_SHR, ins->dreg, ins->inst_imm, 4);
4523 case OP_LSHR_UN_IMM:
4524 g_assert (amd64_is_imm32 (ins->inst_imm));
4525 amd64_shift_reg_imm (code, X86_SHR, ins->dreg, ins->inst_imm);
4528 g_assert (ins->sreg2 == AMD64_RCX);
4529 amd64_shift_reg (code, X86_SHR, ins->dreg);
4532 g_assert (amd64_is_imm32 (ins->inst_imm));
4533 amd64_shift_reg_imm_size (code, X86_SHL, ins->dreg, ins->inst_imm, 4);
4536 g_assert (amd64_is_imm32 (ins->inst_imm));
4537 amd64_shift_reg_imm (code, X86_SHL, ins->dreg, ins->inst_imm);
4542 amd64_alu_reg_reg_size (code, X86_ADD, ins->sreg1, ins->sreg2, 4);
4545 amd64_alu_reg_reg_size (code, X86_ADC, ins->sreg1, ins->sreg2, 4);
4548 amd64_alu_reg_imm_size (code, X86_ADD, ins->dreg, ins->inst_imm, 4);
4551 amd64_alu_reg_imm_size (code, X86_ADC, ins->dreg, ins->inst_imm, 4);
4555 amd64_alu_reg_reg_size (code, X86_SUB, ins->sreg1, ins->sreg2, 4);
4558 amd64_alu_reg_reg_size (code, X86_SBB, ins->sreg1, ins->sreg2, 4);
4561 amd64_alu_reg_imm_size (code, X86_SUB, ins->dreg, ins->inst_imm, 4);
4564 amd64_alu_reg_imm_size (code, X86_SBB, ins->dreg, ins->inst_imm, 4);
4567 amd64_alu_reg_reg_size (code, X86_AND, ins->sreg1, ins->sreg2, 4);
4570 amd64_alu_reg_imm_size (code, X86_AND, ins->sreg1, ins->inst_imm, 4);
4573 amd64_alu_reg_reg_size (code, X86_OR, ins->sreg1, ins->sreg2, 4);
4576 amd64_alu_reg_imm_size (code, X86_OR, ins->sreg1, ins->inst_imm, 4);
4579 amd64_alu_reg_reg_size (code, X86_XOR, ins->sreg1, ins->sreg2, 4);
4582 amd64_alu_reg_imm_size (code, X86_XOR, ins->sreg1, ins->inst_imm, 4);
4585 amd64_neg_reg_size (code, ins->sreg1, 4);
4588 amd64_not_reg_size (code, ins->sreg1, 4);
4591 g_assert (ins->sreg2 == AMD64_RCX);
4592 amd64_shift_reg_size (code, X86_SHL, ins->dreg, 4);
4595 g_assert (ins->sreg2 == AMD64_RCX);
4596 amd64_shift_reg_size (code, X86_SAR, ins->dreg, 4);
4599 amd64_shift_reg_imm_size (code, X86_SAR, ins->dreg, ins->inst_imm, 4);
4601 case OP_ISHR_UN_IMM:
4602 amd64_shift_reg_imm_size (code, X86_SHR, ins->dreg, ins->inst_imm, 4);
4605 g_assert (ins->sreg2 == AMD64_RCX);
4606 amd64_shift_reg_size (code, X86_SHR, ins->dreg, 4);
4609 amd64_shift_reg_imm_size (code, X86_SHL, ins->dreg, ins->inst_imm, 4);
4612 amd64_imul_reg_reg_size (code, ins->sreg1, ins->sreg2, 4);
4615 amd64_imul_reg_reg_size (code, ins->sreg1, ins->sreg2, 4);
4616 EMIT_COND_SYSTEM_EXCEPTION (X86_CC_O, FALSE, "OverflowException");
4618 case OP_IMUL_OVF_UN:
4619 case OP_LMUL_OVF_UN: {
4620 /* the mul operation and the exception check should most likely be split */
4621 int non_eax_reg, saved_eax = FALSE, saved_edx = FALSE;
4622 int size = (ins->opcode == OP_IMUL_OVF_UN) ? 4 : 8;
4623 /*g_assert (ins->sreg2 == X86_EAX);
4624 g_assert (ins->dreg == X86_EAX);*/
4625 if (ins->sreg2 == X86_EAX) {
4626 non_eax_reg = ins->sreg1;
4627 } else if (ins->sreg1 == X86_EAX) {
4628 non_eax_reg = ins->sreg2;
4630 /* no need to save since we're going to store to it anyway */
4631 if (ins->dreg != X86_EAX) {
4633 amd64_push_reg (code, X86_EAX);
4635 amd64_mov_reg_reg (code, X86_EAX, ins->sreg1, size);
4636 non_eax_reg = ins->sreg2;
4638 if (ins->dreg == X86_EDX) {
4641 amd64_push_reg (code, X86_EAX);
4645 amd64_push_reg (code, X86_EDX);
4647 amd64_mul_reg_size (code, non_eax_reg, FALSE, size);
4648 /* save before the check since pop and mov don't change the flags */
4649 if (ins->dreg != X86_EAX)
4650 amd64_mov_reg_reg (code, ins->dreg, X86_EAX, size);
4652 amd64_pop_reg (code, X86_EDX);
4654 amd64_pop_reg (code, X86_EAX);
4655 EMIT_COND_SYSTEM_EXCEPTION (X86_CC_O, FALSE, "OverflowException");
4659 amd64_alu_reg_reg_size (code, X86_CMP, ins->sreg1, ins->sreg2, 4);
4661 case OP_ICOMPARE_IMM:
4662 amd64_alu_reg_imm_size (code, X86_CMP, ins->sreg1, ins->inst_imm, 4);
4684 EMIT_COND_BRANCH (ins, cc_table [mono_opcode_to_cond (ins->opcode)], cc_signed_table [mono_opcode_to_cond (ins->opcode)]);
4692 case OP_CMOV_INE_UN:
4693 case OP_CMOV_IGE_UN:
4694 case OP_CMOV_IGT_UN:
4695 case OP_CMOV_ILE_UN:
4696 case OP_CMOV_ILT_UN:
4702 case OP_CMOV_LNE_UN:
4703 case OP_CMOV_LGE_UN:
4704 case OP_CMOV_LGT_UN:
4705 case OP_CMOV_LLE_UN:
4706 case OP_CMOV_LLT_UN:
4707 g_assert (ins->dreg == ins->sreg1);
4708 /* This needs to operate on 64 bit values */
4709 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);
4713 amd64_not_reg (code, ins->sreg1);
4716 amd64_neg_reg (code, ins->sreg1);
4721 if ((((guint64)ins->inst_c0) >> 32) == 0)
4722 amd64_mov_reg_imm_size (code, ins->dreg, ins->inst_c0, 4);
4724 amd64_mov_reg_imm_size (code, ins->dreg, ins->inst_c0, 8);
4727 mono_add_patch_info (cfg, offset, (MonoJumpInfoType)ins->inst_i1, ins->inst_p0);
4728 amd64_mov_reg_membase (code, ins->dreg, AMD64_RIP, 0, sizeof(gpointer));
4731 mono_add_patch_info (cfg, offset, (MonoJumpInfoType)ins->inst_i1, ins->inst_p0);
4732 amd64_mov_reg_imm_size (code, ins->dreg, 0, 8);
4735 amd64_mov_reg_reg (code, ins->dreg, ins->sreg1, sizeof(mgreg_t));
4737 case OP_AMD64_SET_XMMREG_R4: {
4738 amd64_sse_cvtsd2ss_reg_reg (code, ins->dreg, ins->sreg1);
4741 case OP_AMD64_SET_XMMREG_R8: {
4742 if (ins->dreg != ins->sreg1)
4743 amd64_sse_movsd_reg_reg (code, ins->dreg, ins->sreg1);
4747 MonoCallInst *call = (MonoCallInst*)ins;
4750 /* FIXME: no tracing support... */
4751 if (cfg->prof_options & MONO_PROFILE_ENTER_LEAVE)
4752 code = mono_arch_instrument_epilog_full (cfg, mono_profiler_method_leave, code, FALSE, FALSE);
4754 g_assert (!cfg->method->save_lmf);
4756 if (cfg->arch.omit_fp) {
4757 guint32 save_offset = 0;
4758 /* Pop callee-saved registers */
4759 for (i = 0; i < AMD64_NREG; ++i)
4760 if (AMD64_IS_CALLEE_SAVED_REG (i) && (cfg->used_int_regs & (1 << i))) {
4761 amd64_mov_reg_membase (code, i, AMD64_RSP, save_offset, 8);
4764 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, cfg->arch.stack_alloc_size);
4767 if (call->stack_usage)
4771 for (i = 0; i < AMD64_NREG; ++i)
4772 if (AMD64_IS_CALLEE_SAVED_REG (i) && (cfg->used_int_regs & (1 << i)))
4773 pos -= sizeof(mgreg_t);
4775 /* Restore callee-saved registers */
4776 for (i = AMD64_NREG - 1; i > 0; --i) {
4777 if (AMD64_IS_CALLEE_SAVED_REG (i) && (cfg->used_int_regs & (1 << i))) {
4778 amd64_mov_reg_membase (code, i, AMD64_RBP, pos, sizeof(mgreg_t));
4779 pos += sizeof(mgreg_t);
4783 /* Copy arguments on the stack to our argument area */
4784 for (i = 0; i < call->stack_usage; i += sizeof(mgreg_t)) {
4785 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RSP, i, sizeof(mgreg_t));
4786 amd64_mov_membase_reg (code, AMD64_RBP, 16 + i, AMD64_RAX, sizeof(mgreg_t));
4790 amd64_lea_membase (code, AMD64_RSP, AMD64_RBP, pos);
4795 offset = code - cfg->native_code;
4796 mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_METHOD_JUMP, ins->inst_p0);
4797 if (cfg->compile_aot)
4798 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RIP, 0, 8);
4800 amd64_set_reg_template (code, AMD64_R11);
4801 amd64_jump_reg (code, AMD64_R11);
4802 ins->flags |= MONO_INST_GC_CALLSITE;
4803 ins->backend.pc_offset = code - cfg->native_code;
4807 /* ensure ins->sreg1 is not NULL */
4808 amd64_alu_membase_imm_size (code, X86_CMP, ins->sreg1, 0, 0, 4);
4811 amd64_lea_membase (code, AMD64_R11, cfg->frame_reg, cfg->sig_cookie);
4812 amd64_mov_membase_reg (code, ins->sreg1, 0, AMD64_R11, sizeof(gpointer));
4821 call = (MonoCallInst*)ins;
4823 * The AMD64 ABI forces callers to know about varargs.
4825 if ((call->signature->call_convention == MONO_CALL_VARARG) && (call->signature->pinvoke))
4826 amd64_alu_reg_reg (code, X86_XOR, AMD64_RAX, AMD64_RAX);
4827 else if ((cfg->method->wrapper_type == MONO_WRAPPER_MANAGED_TO_NATIVE) && (cfg->method->klass->image != mono_defaults.corlib)) {
4829 * Since the unmanaged calling convention doesn't contain a
4830 * 'vararg' entry, we have to treat every pinvoke call as a
4831 * potential vararg call.
4835 for (i = 0; i < AMD64_XMM_NREG; ++i)
4836 if (call->used_fregs & (1 << i))
4839 amd64_alu_reg_reg (code, X86_XOR, AMD64_RAX, AMD64_RAX);
4841 amd64_mov_reg_imm (code, AMD64_RAX, nregs);
4844 if (ins->flags & MONO_INST_HAS_METHOD)
4845 code = emit_call (cfg, code, MONO_PATCH_INFO_METHOD, call->method, FALSE);
4847 code = emit_call (cfg, code, MONO_PATCH_INFO_ABS, call->fptr, FALSE);
4848 ins->flags |= MONO_INST_GC_CALLSITE;
4849 ins->backend.pc_offset = code - cfg->native_code;
4850 if (call->stack_usage && !CALLCONV_IS_STDCALL (call->signature->call_convention) && !cfg->arch.no_pushes)
4851 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, call->stack_usage);
4852 code = emit_move_return_value (cfg, ins, code);
4858 case OP_VOIDCALL_REG:
4860 call = (MonoCallInst*)ins;
4862 if (AMD64_IS_ARGUMENT_REG (ins->sreg1)) {
4863 amd64_mov_reg_reg (code, AMD64_R11, ins->sreg1, 8);
4864 ins->sreg1 = AMD64_R11;
4868 * The AMD64 ABI forces callers to know about varargs.
4870 if ((call->signature->call_convention == MONO_CALL_VARARG) && (call->signature->pinvoke)) {
4871 if (ins->sreg1 == AMD64_RAX) {
4872 amd64_mov_reg_reg (code, AMD64_R11, AMD64_RAX, 8);
4873 ins->sreg1 = AMD64_R11;
4875 amd64_alu_reg_reg (code, X86_XOR, AMD64_RAX, AMD64_RAX);
4876 } else if ((cfg->method->wrapper_type == MONO_WRAPPER_MANAGED_TO_NATIVE) && (cfg->method->klass->image != mono_defaults.corlib)) {
4878 * Since the unmanaged calling convention doesn't contain a
4879 * 'vararg' entry, we have to treat every pinvoke call as a
4880 * potential vararg call.
4884 for (i = 0; i < AMD64_XMM_NREG; ++i)
4885 if (call->used_fregs & (1 << i))
4887 if (ins->sreg1 == AMD64_RAX) {
4888 amd64_mov_reg_reg (code, AMD64_R11, AMD64_RAX, 8);
4889 ins->sreg1 = AMD64_R11;
4892 amd64_alu_reg_reg (code, X86_XOR, AMD64_RAX, AMD64_RAX);
4894 amd64_mov_reg_imm (code, AMD64_RAX, nregs);
4897 amd64_call_reg (code, ins->sreg1);
4898 ins->flags |= MONO_INST_GC_CALLSITE;
4899 ins->backend.pc_offset = code - cfg->native_code;
4900 if (call->stack_usage && !CALLCONV_IS_STDCALL (call->signature->call_convention) && !cfg->arch.no_pushes)
4901 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, call->stack_usage);
4902 code = emit_move_return_value (cfg, ins, code);
4904 case OP_FCALL_MEMBASE:
4905 case OP_LCALL_MEMBASE:
4906 case OP_VCALL_MEMBASE:
4907 case OP_VCALL2_MEMBASE:
4908 case OP_VOIDCALL_MEMBASE:
4909 case OP_CALL_MEMBASE:
4910 call = (MonoCallInst*)ins;
4912 amd64_call_membase (code, ins->sreg1, ins->inst_offset);
4913 ins->flags |= MONO_INST_GC_CALLSITE;
4914 ins->backend.pc_offset = code - cfg->native_code;
4915 if (call->stack_usage && !CALLCONV_IS_STDCALL (call->signature->call_convention) && !cfg->arch.no_pushes)
4916 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, call->stack_usage);
4917 code = emit_move_return_value (cfg, ins, code);
4921 MonoInst *var = cfg->dyn_call_var;
4923 g_assert (var->opcode == OP_REGOFFSET);
4925 /* r11 = args buffer filled by mono_arch_get_dyn_call_args () */
4926 amd64_mov_reg_reg (code, AMD64_R11, ins->sreg1, 8);
4928 amd64_mov_reg_reg (code, AMD64_R10, ins->sreg2, 8);
4930 /* Save args buffer */
4931 amd64_mov_membase_reg (code, var->inst_basereg, var->inst_offset, AMD64_R11, 8);
4933 /* Set argument registers */
4934 for (i = 0; i < PARAM_REGS; ++i)
4935 amd64_mov_reg_membase (code, param_regs [i], AMD64_R11, i * sizeof(mgreg_t), sizeof(mgreg_t));
4938 amd64_call_reg (code, AMD64_R10);
4940 ins->flags |= MONO_INST_GC_CALLSITE;
4941 ins->backend.pc_offset = code - cfg->native_code;
4944 amd64_mov_reg_membase (code, AMD64_R11, var->inst_basereg, var->inst_offset, 8);
4945 amd64_mov_membase_reg (code, AMD64_R11, G_STRUCT_OFFSET (DynCallArgs, res), AMD64_RAX, 8);
4948 case OP_AMD64_SAVE_SP_TO_LMF: {
4949 MonoInst *lmf_var = cfg->arch.lmf_var;
4950 amd64_mov_membase_reg (code, cfg->frame_reg, lmf_var->inst_offset + G_STRUCT_OFFSET (MonoLMF, rsp), AMD64_RSP, 8);
4954 g_assert (!cfg->arch.no_pushes);
4955 amd64_push_reg (code, ins->sreg1);
4957 case OP_X86_PUSH_IMM:
4958 g_assert (!cfg->arch.no_pushes);
4959 g_assert (amd64_is_imm32 (ins->inst_imm));
4960 amd64_push_imm (code, ins->inst_imm);
4962 case OP_X86_PUSH_MEMBASE:
4963 g_assert (!cfg->arch.no_pushes);
4964 amd64_push_membase (code, ins->inst_basereg, ins->inst_offset);
4966 case OP_X86_PUSH_OBJ: {
4967 int size = ALIGN_TO (ins->inst_imm, 8);
4969 g_assert (!cfg->arch.no_pushes);
4971 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, size);
4972 amd64_push_reg (code, AMD64_RDI);
4973 amd64_push_reg (code, AMD64_RSI);
4974 amd64_push_reg (code, AMD64_RCX);
4975 if (ins->inst_offset)
4976 amd64_lea_membase (code, AMD64_RSI, ins->inst_basereg, ins->inst_offset);
4978 amd64_mov_reg_reg (code, AMD64_RSI, ins->inst_basereg, 8);
4979 amd64_lea_membase (code, AMD64_RDI, AMD64_RSP, (3 * 8));
4980 amd64_mov_reg_imm (code, AMD64_RCX, (size >> 3));
4982 amd64_prefix (code, X86_REP_PREFIX);
4984 amd64_pop_reg (code, AMD64_RCX);
4985 amd64_pop_reg (code, AMD64_RSI);
4986 amd64_pop_reg (code, AMD64_RDI);
4990 amd64_lea_memindex (code, ins->dreg, ins->sreg1, ins->inst_imm, ins->sreg2, ins->backend.shift_amount);
4992 case OP_X86_LEA_MEMBASE:
4993 amd64_lea_membase (code, ins->dreg, ins->sreg1, ins->inst_imm);
4996 amd64_xchg_reg_reg (code, ins->sreg1, ins->sreg2, 4);
4999 /* keep alignment */
5000 amd64_alu_reg_imm (code, X86_ADD, ins->sreg1, MONO_ARCH_FRAME_ALIGNMENT - 1);
5001 amd64_alu_reg_imm (code, X86_AND, ins->sreg1, ~(MONO_ARCH_FRAME_ALIGNMENT - 1));
5002 code = mono_emit_stack_alloc (cfg, code, ins);
5003 amd64_mov_reg_reg (code, ins->dreg, AMD64_RSP, 8);
5004 if (cfg->param_area && cfg->arch.no_pushes)
5005 amd64_alu_reg_imm (code, X86_ADD, ins->dreg, cfg->param_area);
5007 case OP_LOCALLOC_IMM: {
5008 guint32 size = ins->inst_imm;
5009 size = (size + (MONO_ARCH_FRAME_ALIGNMENT - 1)) & ~ (MONO_ARCH_FRAME_ALIGNMENT - 1);
5011 if (ins->flags & MONO_INST_INIT) {
5015 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, size);
5016 amd64_alu_reg_reg (code, X86_XOR, ins->dreg, ins->dreg);
5018 for (i = 0; i < size; i += 8)
5019 amd64_mov_membase_reg (code, AMD64_RSP, i, ins->dreg, 8);
5020 amd64_mov_reg_reg (code, ins->dreg, AMD64_RSP, 8);
5022 amd64_mov_reg_imm (code, ins->dreg, size);
5023 ins->sreg1 = ins->dreg;
5025 code = mono_emit_stack_alloc (cfg, code, ins);
5026 amd64_mov_reg_reg (code, ins->dreg, AMD64_RSP, 8);
5029 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, size);
5030 amd64_mov_reg_reg (code, ins->dreg, AMD64_RSP, 8);
5032 if (cfg->param_area && cfg->arch.no_pushes)
5033 amd64_alu_reg_imm (code, X86_ADD, ins->dreg, cfg->param_area);
5037 amd64_mov_reg_reg (code, AMD64_ARG_REG1, ins->sreg1, 8);
5038 code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD,
5039 (gpointer)"mono_arch_throw_exception", FALSE);
5040 ins->flags |= MONO_INST_GC_CALLSITE;
5041 ins->backend.pc_offset = code - cfg->native_code;
5045 amd64_mov_reg_reg (code, AMD64_ARG_REG1, ins->sreg1, 8);
5046 code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD,
5047 (gpointer)"mono_arch_rethrow_exception", FALSE);
5048 ins->flags |= MONO_INST_GC_CALLSITE;
5049 ins->backend.pc_offset = code - cfg->native_code;
5052 case OP_CALL_HANDLER:
5054 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, 8);
5055 mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_BB, ins->inst_target_bb);
5056 amd64_call_imm (code, 0);
5057 mono_cfg_add_try_hole (cfg, ins->inst_eh_block, code, bb);
5058 /* Restore stack alignment */
5059 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, 8);
5061 case OP_START_HANDLER: {
5062 /* Even though we're saving RSP, use sizeof */
5063 /* gpointer because spvar is of type IntPtr */
5064 /* see: mono_create_spvar_for_region */
5065 MonoInst *spvar = mono_find_spvar_for_region (cfg, bb->region);
5066 amd64_mov_membase_reg (code, spvar->inst_basereg, spvar->inst_offset, AMD64_RSP, sizeof(gpointer));
5068 if ((MONO_BBLOCK_IS_IN_REGION (bb, MONO_REGION_FINALLY) ||
5069 MONO_BBLOCK_IS_IN_REGION (bb, MONO_REGION_FINALLY)) &&
5070 cfg->param_area && cfg->arch.no_pushes) {
5071 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, ALIGN_TO (cfg->param_area, MONO_ARCH_FRAME_ALIGNMENT));
5075 case OP_ENDFINALLY: {
5076 MonoInst *spvar = mono_find_spvar_for_region (cfg, bb->region);
5077 amd64_mov_reg_membase (code, AMD64_RSP, spvar->inst_basereg, spvar->inst_offset, sizeof(gpointer));
5081 case OP_ENDFILTER: {
5082 MonoInst *spvar = mono_find_spvar_for_region (cfg, bb->region);
5083 amd64_mov_reg_membase (code, AMD64_RSP, spvar->inst_basereg, spvar->inst_offset, sizeof(gpointer));
5084 /* The local allocator will put the result into RAX */
5090 ins->inst_c0 = code - cfg->native_code;
5093 //g_print ("target: %p, next: %p, curr: %p, last: %p\n", ins->inst_target_bb, bb->next_bb, ins, bb->last_ins);
5094 //if ((ins->inst_target_bb == bb->next_bb) && ins == bb->last_ins)
5096 if (ins->inst_target_bb->native_offset) {
5097 amd64_jump_code (code, cfg->native_code + ins->inst_target_bb->native_offset);
5099 mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_BB, ins->inst_target_bb);
5100 if ((cfg->opt & MONO_OPT_BRANCH) &&
5101 x86_is_imm8 (ins->inst_target_bb->max_offset - offset))
5102 x86_jump8 (code, 0);
5104 x86_jump32 (code, 0);
5108 amd64_jump_reg (code, ins->sreg1);
5125 amd64_set_reg (code, cc_table [mono_opcode_to_cond (ins->opcode)], ins->dreg, cc_signed_table [mono_opcode_to_cond (ins->opcode)]);
5126 amd64_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
5128 case OP_COND_EXC_EQ:
5129 case OP_COND_EXC_NE_UN:
5130 case OP_COND_EXC_LT:
5131 case OP_COND_EXC_LT_UN:
5132 case OP_COND_EXC_GT:
5133 case OP_COND_EXC_GT_UN:
5134 case OP_COND_EXC_GE:
5135 case OP_COND_EXC_GE_UN:
5136 case OP_COND_EXC_LE:
5137 case OP_COND_EXC_LE_UN:
5138 case OP_COND_EXC_IEQ:
5139 case OP_COND_EXC_INE_UN:
5140 case OP_COND_EXC_ILT:
5141 case OP_COND_EXC_ILT_UN:
5142 case OP_COND_EXC_IGT:
5143 case OP_COND_EXC_IGT_UN:
5144 case OP_COND_EXC_IGE:
5145 case OP_COND_EXC_IGE_UN:
5146 case OP_COND_EXC_ILE:
5147 case OP_COND_EXC_ILE_UN:
5148 EMIT_COND_SYSTEM_EXCEPTION (cc_table [mono_opcode_to_cond (ins->opcode)], cc_signed_table [mono_opcode_to_cond (ins->opcode)], ins->inst_p1);
5150 case OP_COND_EXC_OV:
5151 case OP_COND_EXC_NO:
5153 case OP_COND_EXC_NC:
5154 EMIT_COND_SYSTEM_EXCEPTION (branch_cc_table [ins->opcode - OP_COND_EXC_EQ],
5155 (ins->opcode < OP_COND_EXC_NE_UN), ins->inst_p1);
5157 case OP_COND_EXC_IOV:
5158 case OP_COND_EXC_INO:
5159 case OP_COND_EXC_IC:
5160 case OP_COND_EXC_INC:
5161 EMIT_COND_SYSTEM_EXCEPTION (branch_cc_table [ins->opcode - OP_COND_EXC_IEQ],
5162 (ins->opcode < OP_COND_EXC_INE_UN), ins->inst_p1);
5165 /* floating point opcodes */
5167 double d = *(double *)ins->inst_p0;
5169 if ((d == 0.0) && (mono_signbit (d) == 0)) {
5170 amd64_sse_xorpd_reg_reg (code, ins->dreg, ins->dreg);
5173 mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_R8, ins->inst_p0);
5174 amd64_sse_movsd_reg_membase (code, ins->dreg, AMD64_RIP, 0);
5179 float f = *(float *)ins->inst_p0;
5181 if ((f == 0.0) && (mono_signbit (f) == 0)) {
5182 amd64_sse_xorpd_reg_reg (code, ins->dreg, ins->dreg);
5185 mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_R4, ins->inst_p0);
5186 amd64_sse_movss_reg_membase (code, ins->dreg, AMD64_RIP, 0);
5187 amd64_sse_cvtss2sd_reg_reg (code, ins->dreg, ins->dreg);
5191 case OP_STORER8_MEMBASE_REG:
5192 amd64_sse_movsd_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, ins->sreg1);
5194 case OP_LOADR8_MEMBASE:
5195 amd64_sse_movsd_reg_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset);
5197 case OP_STORER4_MEMBASE_REG:
5198 /* This requires a double->single conversion */
5199 amd64_sse_cvtsd2ss_reg_reg (code, AMD64_XMM15, ins->sreg1);
5200 amd64_sse_movss_membase_reg (code, ins->inst_destbasereg, ins->inst_offset, AMD64_XMM15);
5202 case OP_LOADR4_MEMBASE:
5203 amd64_sse_movss_reg_membase (code, ins->dreg, ins->inst_basereg, ins->inst_offset);
5204 amd64_sse_cvtss2sd_reg_reg (code, ins->dreg, ins->dreg);
5206 case OP_ICONV_TO_R4: /* FIXME: change precision */
5207 case OP_ICONV_TO_R8:
5208 amd64_sse_cvtsi2sd_reg_reg_size (code, ins->dreg, ins->sreg1, 4);
5210 case OP_LCONV_TO_R4: /* FIXME: change precision */
5211 case OP_LCONV_TO_R8:
5212 amd64_sse_cvtsi2sd_reg_reg (code, ins->dreg, ins->sreg1);
5214 case OP_FCONV_TO_R4:
5215 /* FIXME: nothing to do ?? */
5217 case OP_FCONV_TO_I1:
5218 code = emit_float_to_int (cfg, code, ins->dreg, ins->sreg1, 1, TRUE);
5220 case OP_FCONV_TO_U1:
5221 code = emit_float_to_int (cfg, code, ins->dreg, ins->sreg1, 1, FALSE);
5223 case OP_FCONV_TO_I2:
5224 code = emit_float_to_int (cfg, code, ins->dreg, ins->sreg1, 2, TRUE);
5226 case OP_FCONV_TO_U2:
5227 code = emit_float_to_int (cfg, code, ins->dreg, ins->sreg1, 2, FALSE);
5229 case OP_FCONV_TO_U4:
5230 code = emit_float_to_int (cfg, code, ins->dreg, ins->sreg1, 4, FALSE);
5232 case OP_FCONV_TO_I4:
5234 code = emit_float_to_int (cfg, code, ins->dreg, ins->sreg1, 4, TRUE);
5236 case OP_FCONV_TO_I8:
5237 code = emit_float_to_int (cfg, code, ins->dreg, ins->sreg1, 8, TRUE);
5239 case OP_LCONV_TO_R_UN: {
5242 /* Based on gcc code */
5243 amd64_test_reg_reg (code, ins->sreg1, ins->sreg1);
5244 br [0] = code; x86_branch8 (code, X86_CC_S, 0, TRUE);
5247 amd64_sse_cvtsi2sd_reg_reg (code, ins->dreg, ins->sreg1);
5248 br [1] = code; x86_jump8 (code, 0);
5249 amd64_patch (br [0], code);
5252 /* Save to the red zone */
5253 amd64_mov_membase_reg (code, AMD64_RSP, -8, AMD64_RAX, 8);
5254 amd64_mov_membase_reg (code, AMD64_RSP, -16, AMD64_RCX, 8);
5255 amd64_mov_reg_reg (code, AMD64_RCX, ins->sreg1, 8);
5256 amd64_mov_reg_reg (code, AMD64_RAX, ins->sreg1, 8);
5257 amd64_alu_reg_imm (code, X86_AND, AMD64_RCX, 1);
5258 amd64_shift_reg_imm (code, X86_SHR, AMD64_RAX, 1);
5259 amd64_alu_reg_imm (code, X86_OR, AMD64_RAX, AMD64_RCX);
5260 amd64_sse_cvtsi2sd_reg_reg (code, ins->dreg, AMD64_RAX);
5261 amd64_sse_addsd_reg_reg (code, ins->dreg, ins->dreg);
5263 amd64_mov_reg_membase (code, AMD64_RCX, AMD64_RSP, -16, 8);
5264 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RSP, -8, 8);
5265 amd64_patch (br [1], code);
5268 case OP_LCONV_TO_OVF_U4:
5269 amd64_alu_reg_imm (code, X86_CMP, ins->sreg1, 0);
5270 EMIT_COND_SYSTEM_EXCEPTION (X86_CC_LT, TRUE, "OverflowException");
5271 amd64_mov_reg_reg (code, ins->dreg, ins->sreg1, 8);
5273 case OP_LCONV_TO_OVF_I4_UN:
5274 amd64_alu_reg_imm (code, X86_CMP, ins->sreg1, 0x7fffffff);
5275 EMIT_COND_SYSTEM_EXCEPTION (X86_CC_GT, FALSE, "OverflowException");
5276 amd64_mov_reg_reg (code, ins->dreg, ins->sreg1, 8);
5279 if (ins->dreg != ins->sreg1)
5280 amd64_sse_movsd_reg_reg (code, ins->dreg, ins->sreg1);
5283 amd64_sse_addsd_reg_reg (code, ins->dreg, ins->sreg2);
5286 amd64_sse_subsd_reg_reg (code, ins->dreg, ins->sreg2);
5289 amd64_sse_mulsd_reg_reg (code, ins->dreg, ins->sreg2);
5292 amd64_sse_divsd_reg_reg (code, ins->dreg, ins->sreg2);
5295 static double r8_0 = -0.0;
5297 g_assert (ins->sreg1 == ins->dreg);
5299 mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_R8, &r8_0);
5300 amd64_sse_xorpd_reg_membase (code, ins->dreg, AMD64_RIP, 0);
5304 EMIT_SSE2_FPFUNC (code, fsin, ins->dreg, ins->sreg1);
5307 EMIT_SSE2_FPFUNC (code, fcos, ins->dreg, ins->sreg1);
5310 static guint64 d = 0x7fffffffffffffffUL;
5312 g_assert (ins->sreg1 == ins->dreg);
5314 mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_R8, &d);
5315 amd64_sse_andpd_reg_membase (code, ins->dreg, AMD64_RIP, 0);
5319 EMIT_SSE2_FPFUNC (code, fsqrt, ins->dreg, ins->sreg1);
5322 g_assert (cfg->opt & MONO_OPT_CMOV);
5323 g_assert (ins->dreg == ins->sreg1);
5324 amd64_alu_reg_reg_size (code, X86_CMP, ins->sreg1, ins->sreg2, 4);
5325 amd64_cmov_reg_size (code, X86_CC_GT, TRUE, ins->dreg, ins->sreg2, 4);
5328 g_assert (cfg->opt & MONO_OPT_CMOV);
5329 g_assert (ins->dreg == ins->sreg1);
5330 amd64_alu_reg_reg_size (code, X86_CMP, ins->sreg1, ins->sreg2, 4);
5331 amd64_cmov_reg_size (code, X86_CC_GT, FALSE, ins->dreg, ins->sreg2, 4);
5334 g_assert (cfg->opt & MONO_OPT_CMOV);
5335 g_assert (ins->dreg == ins->sreg1);
5336 amd64_alu_reg_reg_size (code, X86_CMP, ins->sreg1, ins->sreg2, 4);
5337 amd64_cmov_reg_size (code, X86_CC_LT, TRUE, ins->dreg, ins->sreg2, 4);
5340 g_assert (cfg->opt & MONO_OPT_CMOV);
5341 g_assert (ins->dreg == ins->sreg1);
5342 amd64_alu_reg_reg_size (code, X86_CMP, ins->sreg1, ins->sreg2, 4);
5343 amd64_cmov_reg_size (code, X86_CC_LT, FALSE, ins->dreg, ins->sreg2, 4);
5346 g_assert (cfg->opt & MONO_OPT_CMOV);
5347 g_assert (ins->dreg == ins->sreg1);
5348 amd64_alu_reg_reg (code, X86_CMP, ins->sreg1, ins->sreg2);
5349 amd64_cmov_reg (code, X86_CC_GT, TRUE, ins->dreg, ins->sreg2);
5352 g_assert (cfg->opt & MONO_OPT_CMOV);
5353 g_assert (ins->dreg == ins->sreg1);
5354 amd64_alu_reg_reg (code, X86_CMP, ins->sreg1, ins->sreg2);
5355 amd64_cmov_reg (code, X86_CC_GT, FALSE, ins->dreg, ins->sreg2);
5358 g_assert (cfg->opt & MONO_OPT_CMOV);
5359 g_assert (ins->dreg == ins->sreg1);
5360 amd64_alu_reg_reg (code, X86_CMP, ins->sreg1, ins->sreg2);
5361 amd64_cmov_reg (code, X86_CC_LT, TRUE, ins->dreg, ins->sreg2);
5364 g_assert (cfg->opt & MONO_OPT_CMOV);
5365 g_assert (ins->dreg == ins->sreg1);
5366 amd64_alu_reg_reg (code, X86_CMP, ins->sreg1, ins->sreg2);
5367 amd64_cmov_reg (code, X86_CC_LT, FALSE, ins->dreg, ins->sreg2);
5373 * The two arguments are swapped because the fbranch instructions
5374 * depend on this for the non-sse case to work.
5376 amd64_sse_comisd_reg_reg (code, ins->sreg2, ins->sreg1);
5379 /* zeroing the register at the start results in
5380 * shorter and faster code (we can also remove the widening op)
5382 guchar *unordered_check;
5383 amd64_alu_reg_reg (code, X86_XOR, ins->dreg, ins->dreg);
5384 amd64_sse_comisd_reg_reg (code, ins->sreg1, ins->sreg2);
5385 unordered_check = code;
5386 x86_branch8 (code, X86_CC_P, 0, FALSE);
5387 amd64_set_reg (code, X86_CC_EQ, ins->dreg, FALSE);
5388 amd64_patch (unordered_check, code);
5393 /* zeroing the register at the start results in
5394 * shorter and faster code (we can also remove the widening op)
5396 amd64_alu_reg_reg (code, X86_XOR, ins->dreg, ins->dreg);
5397 amd64_sse_comisd_reg_reg (code, ins->sreg2, ins->sreg1);
5398 if (ins->opcode == OP_FCLT_UN) {
5399 guchar *unordered_check = code;
5400 guchar *jump_to_end;
5401 x86_branch8 (code, X86_CC_P, 0, FALSE);
5402 amd64_set_reg (code, X86_CC_GT, ins->dreg, FALSE);
5404 x86_jump8 (code, 0);
5405 amd64_patch (unordered_check, code);
5406 amd64_inc_reg (code, ins->dreg);
5407 amd64_patch (jump_to_end, code);
5409 amd64_set_reg (code, X86_CC_GT, ins->dreg, FALSE);
5414 /* zeroing the register at the start results in
5415 * shorter and faster code (we can also remove the widening op)
5417 guchar *unordered_check;
5418 amd64_alu_reg_reg (code, X86_XOR, ins->dreg, ins->dreg);
5419 amd64_sse_comisd_reg_reg (code, ins->sreg2, ins->sreg1);
5420 if (ins->opcode == OP_FCGT) {
5421 unordered_check = code;
5422 x86_branch8 (code, X86_CC_P, 0, FALSE);
5423 amd64_set_reg (code, X86_CC_LT, ins->dreg, FALSE);
5424 amd64_patch (unordered_check, code);
5426 amd64_set_reg (code, X86_CC_LT, ins->dreg, FALSE);
5430 case OP_FCLT_MEMBASE:
5431 case OP_FCGT_MEMBASE:
5432 case OP_FCLT_UN_MEMBASE:
5433 case OP_FCGT_UN_MEMBASE:
5434 case OP_FCEQ_MEMBASE: {
5435 guchar *unordered_check, *jump_to_end;
5438 amd64_alu_reg_reg (code, X86_XOR, ins->dreg, ins->dreg);
5439 amd64_sse_comisd_reg_membase (code, ins->sreg1, ins->sreg2, ins->inst_offset);
5441 switch (ins->opcode) {
5442 case OP_FCEQ_MEMBASE:
5443 x86_cond = X86_CC_EQ;
5445 case OP_FCLT_MEMBASE:
5446 case OP_FCLT_UN_MEMBASE:
5447 x86_cond = X86_CC_LT;
5449 case OP_FCGT_MEMBASE:
5450 case OP_FCGT_UN_MEMBASE:
5451 x86_cond = X86_CC_GT;
5454 g_assert_not_reached ();
5457 unordered_check = code;
5458 x86_branch8 (code, X86_CC_P, 0, FALSE);
5459 amd64_set_reg (code, x86_cond, ins->dreg, FALSE);
5461 switch (ins->opcode) {
5462 case OP_FCEQ_MEMBASE:
5463 case OP_FCLT_MEMBASE:
5464 case OP_FCGT_MEMBASE:
5465 amd64_patch (unordered_check, code);
5467 case OP_FCLT_UN_MEMBASE:
5468 case OP_FCGT_UN_MEMBASE:
5470 x86_jump8 (code, 0);
5471 amd64_patch (unordered_check, code);
5472 amd64_inc_reg (code, ins->dreg);
5473 amd64_patch (jump_to_end, code);
5481 guchar *jump = code;
5482 x86_branch8 (code, X86_CC_P, 0, TRUE);
5483 EMIT_COND_BRANCH (ins, X86_CC_EQ, FALSE);
5484 amd64_patch (jump, code);
5488 /* Branch if C013 != 100 */
5489 /* branch if !ZF or (PF|CF) */
5490 EMIT_COND_BRANCH (ins, X86_CC_NE, FALSE);
5491 EMIT_COND_BRANCH (ins, X86_CC_P, FALSE);
5492 EMIT_COND_BRANCH (ins, X86_CC_B, FALSE);
5495 EMIT_COND_BRANCH (ins, X86_CC_GT, FALSE);
5498 EMIT_COND_BRANCH (ins, X86_CC_P, FALSE);
5499 EMIT_COND_BRANCH (ins, X86_CC_GT, FALSE);
5503 if (ins->opcode == OP_FBGT) {
5506 /* skip branch if C1=1 */
5508 x86_branch8 (code, X86_CC_P, 0, FALSE);
5509 /* branch if (C0 | C3) = 1 */
5510 EMIT_COND_BRANCH (ins, X86_CC_LT, FALSE);
5511 amd64_patch (br1, code);
5514 EMIT_COND_BRANCH (ins, X86_CC_LT, FALSE);
5518 /* Branch if C013 == 100 or 001 */
5521 /* skip branch if C1=1 */
5523 x86_branch8 (code, X86_CC_P, 0, FALSE);
5524 /* branch if (C0 | C3) = 1 */
5525 EMIT_COND_BRANCH (ins, X86_CC_BE, FALSE);
5526 amd64_patch (br1, code);
5530 /* Branch if C013 == 000 */
5531 EMIT_COND_BRANCH (ins, X86_CC_LE, FALSE);
5534 /* Branch if C013=000 or 100 */
5537 /* skip branch if C1=1 */
5539 x86_branch8 (code, X86_CC_P, 0, FALSE);
5540 /* branch if C0=0 */
5541 EMIT_COND_BRANCH (ins, X86_CC_NB, FALSE);
5542 amd64_patch (br1, code);
5546 /* Branch if C013 != 001 */
5547 EMIT_COND_BRANCH (ins, X86_CC_P, FALSE);
5548 EMIT_COND_BRANCH (ins, X86_CC_GE, FALSE);
5551 /* Transfer value to the fp stack */
5552 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, 16);
5553 amd64_movsd_membase_reg (code, AMD64_RSP, 0, ins->sreg1);
5554 amd64_fld_membase (code, AMD64_RSP, 0, TRUE);
5556 amd64_push_reg (code, AMD64_RAX);
5558 amd64_fnstsw (code);
5559 amd64_alu_reg_imm (code, X86_AND, AMD64_RAX, 0x4100);
5560 amd64_alu_reg_imm (code, X86_CMP, AMD64_RAX, X86_FP_C0);
5561 amd64_pop_reg (code, AMD64_RAX);
5562 amd64_fstp (code, 0);
5563 EMIT_COND_SYSTEM_EXCEPTION (X86_CC_EQ, FALSE, "ArithmeticException");
5564 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, 16);
5567 code = mono_amd64_emit_tls_get (code, ins->dreg, ins->inst_offset);
5570 case OP_MEMORY_BARRIER: {
5571 /* http://blogs.sun.com/dave/resource/NHM-Pipeline-Blog-V2.txt */
5572 x86_prefix (code, X86_LOCK_PREFIX);
5573 amd64_alu_membase_imm (code, X86_ADD, AMD64_RSP, 0, 0);
5576 case OP_ATOMIC_ADD_I4:
5577 case OP_ATOMIC_ADD_I8: {
5578 int dreg = ins->dreg;
5579 guint32 size = (ins->opcode == OP_ATOMIC_ADD_I4) ? 4 : 8;
5581 if (dreg == ins->inst_basereg)
5584 if (dreg != ins->sreg2)
5585 amd64_mov_reg_reg (code, ins->dreg, ins->sreg2, size);
5587 x86_prefix (code, X86_LOCK_PREFIX);
5588 amd64_xadd_membase_reg (code, ins->inst_basereg, ins->inst_offset, dreg, size);
5590 if (dreg != ins->dreg)
5591 amd64_mov_reg_reg (code, ins->dreg, dreg, size);
5595 case OP_ATOMIC_ADD_NEW_I4:
5596 case OP_ATOMIC_ADD_NEW_I8: {
5597 int dreg = ins->dreg;
5598 guint32 size = (ins->opcode == OP_ATOMIC_ADD_NEW_I4) ? 4 : 8;
5600 if ((dreg == ins->sreg2) || (dreg == ins->inst_basereg))
5603 amd64_mov_reg_reg (code, dreg, ins->sreg2, size);
5604 amd64_prefix (code, X86_LOCK_PREFIX);
5605 amd64_xadd_membase_reg (code, ins->inst_basereg, ins->inst_offset, dreg, size);
5606 /* dreg contains the old value, add with sreg2 value */
5607 amd64_alu_reg_reg_size (code, X86_ADD, dreg, ins->sreg2, size);
5609 if (ins->dreg != dreg)
5610 amd64_mov_reg_reg (code, ins->dreg, dreg, size);
5614 case OP_ATOMIC_EXCHANGE_I4:
5615 case OP_ATOMIC_EXCHANGE_I8: {
5617 int sreg2 = ins->sreg2;
5618 int breg = ins->inst_basereg;
5620 gboolean need_push = FALSE, rdx_pushed = FALSE;
5622 if (ins->opcode == OP_ATOMIC_EXCHANGE_I8)
5628 * See http://msdn.microsoft.com/en-us/magazine/cc302329.aspx for
5629 * an explanation of how this works.
5632 /* cmpxchg uses eax as comperand, need to make sure we can use it
5633 * hack to overcome limits in x86 reg allocator
5634 * (req: dreg == eax and sreg2 != eax and breg != eax)
5636 g_assert (ins->dreg == AMD64_RAX);
5638 if (breg == AMD64_RAX && ins->sreg2 == AMD64_RAX)
5639 /* Highly unlikely, but possible */
5642 /* The pushes invalidate rsp */
5643 if ((breg == AMD64_RAX) || need_push) {
5644 amd64_mov_reg_reg (code, AMD64_R11, breg, 8);
5648 /* We need the EAX reg for the comparand */
5649 if (ins->sreg2 == AMD64_RAX) {
5650 if (breg != AMD64_R11) {
5651 amd64_mov_reg_reg (code, AMD64_R11, AMD64_RAX, 8);
5654 g_assert (need_push);
5655 amd64_push_reg (code, AMD64_RDX);
5656 amd64_mov_reg_reg (code, AMD64_RDX, AMD64_RAX, size);
5662 amd64_mov_reg_membase (code, AMD64_RAX, breg, ins->inst_offset, size);
5664 br [0] = code; amd64_prefix (code, X86_LOCK_PREFIX);
5665 amd64_cmpxchg_membase_reg_size (code, breg, ins->inst_offset, sreg2, size);
5666 br [1] = code; amd64_branch8 (code, X86_CC_NE, -1, FALSE);
5667 amd64_patch (br [1], br [0]);
5670 amd64_pop_reg (code, AMD64_RDX);
5674 case OP_ATOMIC_CAS_I4:
5675 case OP_ATOMIC_CAS_I8: {
5678 if (ins->opcode == OP_ATOMIC_CAS_I8)
5684 * See http://msdn.microsoft.com/en-us/magazine/cc302329.aspx for
5685 * an explanation of how this works.
5687 g_assert (ins->sreg3 == AMD64_RAX);
5688 g_assert (ins->sreg1 != AMD64_RAX);
5689 g_assert (ins->sreg1 != ins->sreg2);
5691 amd64_prefix (code, X86_LOCK_PREFIX);
5692 amd64_cmpxchg_membase_reg_size (code, ins->sreg1, ins->inst_offset, ins->sreg2, size);
5694 if (ins->dreg != AMD64_RAX)
5695 amd64_mov_reg_reg (code, ins->dreg, AMD64_RAX, size);
5698 case OP_CARD_TABLE_WBARRIER: {
5699 int ptr = ins->sreg1;
5700 int value = ins->sreg2;
5702 int nursery_shift, card_table_shift;
5703 gpointer card_table_mask;
5704 size_t nursery_size;
5706 gpointer card_table = mono_gc_get_card_table (&card_table_shift, &card_table_mask);
5707 guint64 nursery_start = (guint64)mono_gc_get_nursery (&nursery_shift, &nursery_size);
5709 /*If either point to the stack we can simply avoid the WB. This happens due to
5710 * optimizations revealing a stack store that was not visible when op_cardtable was emited.
5712 if (ins->sreg1 == AMD64_RSP || ins->sreg2 == AMD64_RSP)
5716 * We need one register we can clobber, we choose EDX and make sreg1
5717 * fixed EAX to work around limitations in the local register allocator.
5718 * sreg2 might get allocated to EDX, but that is not a problem since
5719 * we use it before clobbering EDX.
5721 g_assert (ins->sreg1 == AMD64_RAX);
5724 * This is the code we produce:
5727 * edx >>= nursery_shift
5728 * cmp edx, (nursery_start >> nursery_shift)
5731 * edx >>= card_table_shift
5737 if (value != AMD64_RDX)
5738 amd64_mov_reg_reg (code, AMD64_RDX, value, 8);
5739 amd64_shift_reg_imm (code, X86_SHR, AMD64_RDX, nursery_shift);
5740 amd64_alu_reg_imm (code, X86_CMP, AMD64_RDX, nursery_start >> nursery_shift);
5741 br = code; x86_branch8 (code, X86_CC_NE, -1, FALSE);
5742 amd64_mov_reg_reg (code, AMD64_RDX, ptr, 8);
5743 amd64_shift_reg_imm (code, X86_SHR, AMD64_RDX, card_table_shift);
5744 if (card_table_mask)
5745 amd64_alu_reg_imm (code, X86_AND, AMD64_RDX, (guint32)(guint64)card_table_mask);
5747 mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_GC_CARD_TABLE_ADDR, card_table);
5748 amd64_alu_reg_membase (code, X86_ADD, AMD64_RDX, AMD64_RIP, 0);
5750 amd64_mov_membase_imm (code, AMD64_RDX, 0, 1, 1);
5751 x86_patch (br, code);
5754 #ifdef MONO_ARCH_SIMD_INTRINSICS
5755 /* TODO: Some of these IR opcodes are marked as no clobber when they indeed do. */
5757 amd64_sse_addps_reg_reg (code, ins->sreg1, ins->sreg2);
5760 amd64_sse_divps_reg_reg (code, ins->sreg1, ins->sreg2);
5763 amd64_sse_mulps_reg_reg (code, ins->sreg1, ins->sreg2);
5766 amd64_sse_subps_reg_reg (code, ins->sreg1, ins->sreg2);
5769 amd64_sse_maxps_reg_reg (code, ins->sreg1, ins->sreg2);
5772 amd64_sse_minps_reg_reg (code, ins->sreg1, ins->sreg2);
5775 g_assert (ins->inst_c0 >= 0 && ins->inst_c0 <= 7);
5776 amd64_sse_cmpps_reg_reg_imm (code, ins->sreg1, ins->sreg2, ins->inst_c0);
5779 amd64_sse_andps_reg_reg (code, ins->sreg1, ins->sreg2);
5782 amd64_sse_andnps_reg_reg (code, ins->sreg1, ins->sreg2);
5785 amd64_sse_orps_reg_reg (code, ins->sreg1, ins->sreg2);
5788 amd64_sse_xorps_reg_reg (code, ins->sreg1, ins->sreg2);
5791 amd64_sse_sqrtps_reg_reg (code, ins->dreg, ins->sreg1);
5794 amd64_sse_rsqrtps_reg_reg (code, ins->dreg, ins->sreg1);
5797 amd64_sse_rcpps_reg_reg (code, ins->dreg, ins->sreg1);
5800 amd64_sse_addsubps_reg_reg (code, ins->sreg1, ins->sreg2);
5803 amd64_sse_haddps_reg_reg (code, ins->sreg1, ins->sreg2);
5806 amd64_sse_hsubps_reg_reg (code, ins->sreg1, ins->sreg2);
5809 amd64_sse_movshdup_reg_reg (code, ins->dreg, ins->sreg1);
5812 amd64_sse_movsldup_reg_reg (code, ins->dreg, ins->sreg1);
5815 case OP_PSHUFLEW_HIGH:
5816 g_assert (ins->inst_c0 >= 0 && ins->inst_c0 <= 0xFF);
5817 amd64_sse_pshufhw_reg_reg_imm (code, ins->dreg, ins->sreg1, ins->inst_c0);
5819 case OP_PSHUFLEW_LOW:
5820 g_assert (ins->inst_c0 >= 0 && ins->inst_c0 <= 0xFF);
5821 amd64_sse_pshuflw_reg_reg_imm (code, ins->dreg, ins->sreg1, ins->inst_c0);
5824 g_assert (ins->inst_c0 >= 0 && ins->inst_c0 <= 0xFF);
5825 amd64_sse_pshufd_reg_reg_imm (code, ins->dreg, ins->sreg1, ins->inst_c0);
5828 g_assert (ins->inst_c0 >= 0 && ins->inst_c0 <= 0xFF);
5829 amd64_sse_shufps_reg_reg_imm (code, ins->sreg1, ins->sreg2, ins->inst_c0);
5832 g_assert (ins->inst_c0 >= 0 && ins->inst_c0 <= 0x3);
5833 amd64_sse_shufpd_reg_reg_imm (code, ins->sreg1, ins->sreg2, ins->inst_c0);
5837 amd64_sse_addpd_reg_reg (code, ins->sreg1, ins->sreg2);
5840 amd64_sse_divpd_reg_reg (code, ins->sreg1, ins->sreg2);
5843 amd64_sse_mulpd_reg_reg (code, ins->sreg1, ins->sreg2);
5846 amd64_sse_subpd_reg_reg (code, ins->sreg1, ins->sreg2);
5849 amd64_sse_maxpd_reg_reg (code, ins->sreg1, ins->sreg2);
5852 amd64_sse_minpd_reg_reg (code, ins->sreg1, ins->sreg2);
5855 g_assert (ins->inst_c0 >= 0 && ins->inst_c0 <= 7);
5856 amd64_sse_cmppd_reg_reg_imm (code, ins->sreg1, ins->sreg2, ins->inst_c0);
5859 amd64_sse_andpd_reg_reg (code, ins->sreg1, ins->sreg2);
5862 amd64_sse_andnpd_reg_reg (code, ins->sreg1, ins->sreg2);
5865 amd64_sse_orpd_reg_reg (code, ins->sreg1, ins->sreg2);
5868 amd64_sse_xorpd_reg_reg (code, ins->sreg1, ins->sreg2);
5871 amd64_sse_sqrtpd_reg_reg (code, ins->dreg, ins->sreg1);
5874 amd64_sse_addsubpd_reg_reg (code, ins->sreg1, ins->sreg2);
5877 amd64_sse_haddpd_reg_reg (code, ins->sreg1, ins->sreg2);
5880 amd64_sse_hsubpd_reg_reg (code, ins->sreg1, ins->sreg2);
5883 amd64_sse_movddup_reg_reg (code, ins->dreg, ins->sreg1);
5886 case OP_EXTRACT_MASK:
5887 amd64_sse_pmovmskb_reg_reg (code, ins->dreg, ins->sreg1);
5891 amd64_sse_pand_reg_reg (code, ins->sreg1, ins->sreg2);
5894 amd64_sse_por_reg_reg (code, ins->sreg1, ins->sreg2);
5897 amd64_sse_pxor_reg_reg (code, ins->sreg1, ins->sreg2);
5901 amd64_sse_paddb_reg_reg (code, ins->sreg1, ins->sreg2);
5904 amd64_sse_paddw_reg_reg (code, ins->sreg1, ins->sreg2);
5907 amd64_sse_paddd_reg_reg (code, ins->sreg1, ins->sreg2);
5910 amd64_sse_paddq_reg_reg (code, ins->sreg1, ins->sreg2);
5914 amd64_sse_psubb_reg_reg (code, ins->sreg1, ins->sreg2);
5917 amd64_sse_psubw_reg_reg (code, ins->sreg1, ins->sreg2);
5920 amd64_sse_psubd_reg_reg (code, ins->sreg1, ins->sreg2);
5923 amd64_sse_psubq_reg_reg (code, ins->sreg1, ins->sreg2);
5927 amd64_sse_pmaxub_reg_reg (code, ins->sreg1, ins->sreg2);
5930 amd64_sse_pmaxuw_reg_reg (code, ins->sreg1, ins->sreg2);
5933 amd64_sse_pmaxud_reg_reg (code, ins->sreg1, ins->sreg2);
5937 amd64_sse_pmaxsb_reg_reg (code, ins->sreg1, ins->sreg2);
5940 amd64_sse_pmaxsw_reg_reg (code, ins->sreg1, ins->sreg2);
5943 amd64_sse_pmaxsd_reg_reg (code, ins->sreg1, ins->sreg2);
5947 amd64_sse_pavgb_reg_reg (code, ins->sreg1, ins->sreg2);
5950 amd64_sse_pavgw_reg_reg (code, ins->sreg1, ins->sreg2);
5954 amd64_sse_pminub_reg_reg (code, ins->sreg1, ins->sreg2);
5957 amd64_sse_pminuw_reg_reg (code, ins->sreg1, ins->sreg2);
5960 amd64_sse_pminud_reg_reg (code, ins->sreg1, ins->sreg2);
5964 amd64_sse_pminsb_reg_reg (code, ins->sreg1, ins->sreg2);
5967 amd64_sse_pminsw_reg_reg (code, ins->sreg1, ins->sreg2);
5970 amd64_sse_pminsd_reg_reg (code, ins->sreg1, ins->sreg2);
5974 amd64_sse_pcmpeqb_reg_reg (code, ins->sreg1, ins->sreg2);
5977 amd64_sse_pcmpeqw_reg_reg (code, ins->sreg1, ins->sreg2);
5980 amd64_sse_pcmpeqd_reg_reg (code, ins->sreg1, ins->sreg2);
5983 amd64_sse_pcmpeqq_reg_reg (code, ins->sreg1, ins->sreg2);
5987 amd64_sse_pcmpgtb_reg_reg (code, ins->sreg1, ins->sreg2);
5990 amd64_sse_pcmpgtw_reg_reg (code, ins->sreg1, ins->sreg2);
5993 amd64_sse_pcmpgtd_reg_reg (code, ins->sreg1, ins->sreg2);
5996 amd64_sse_pcmpgtq_reg_reg (code, ins->sreg1, ins->sreg2);
5999 case OP_PSUM_ABS_DIFF:
6000 amd64_sse_psadbw_reg_reg (code, ins->sreg1, ins->sreg2);
6003 case OP_UNPACK_LOWB:
6004 amd64_sse_punpcklbw_reg_reg (code, ins->sreg1, ins->sreg2);
6006 case OP_UNPACK_LOWW:
6007 amd64_sse_punpcklwd_reg_reg (code, ins->sreg1, ins->sreg2);
6009 case OP_UNPACK_LOWD:
6010 amd64_sse_punpckldq_reg_reg (code, ins->sreg1, ins->sreg2);
6012 case OP_UNPACK_LOWQ:
6013 amd64_sse_punpcklqdq_reg_reg (code, ins->sreg1, ins->sreg2);
6015 case OP_UNPACK_LOWPS:
6016 amd64_sse_unpcklps_reg_reg (code, ins->sreg1, ins->sreg2);
6018 case OP_UNPACK_LOWPD:
6019 amd64_sse_unpcklpd_reg_reg (code, ins->sreg1, ins->sreg2);
6022 case OP_UNPACK_HIGHB:
6023 amd64_sse_punpckhbw_reg_reg (code, ins->sreg1, ins->sreg2);
6025 case OP_UNPACK_HIGHW:
6026 amd64_sse_punpckhwd_reg_reg (code, ins->sreg1, ins->sreg2);
6028 case OP_UNPACK_HIGHD:
6029 amd64_sse_punpckhdq_reg_reg (code, ins->sreg1, ins->sreg2);
6031 case OP_UNPACK_HIGHQ:
6032 amd64_sse_punpckhqdq_reg_reg (code, ins->sreg1, ins->sreg2);
6034 case OP_UNPACK_HIGHPS:
6035 amd64_sse_unpckhps_reg_reg (code, ins->sreg1, ins->sreg2);
6037 case OP_UNPACK_HIGHPD:
6038 amd64_sse_unpckhpd_reg_reg (code, ins->sreg1, ins->sreg2);
6042 amd64_sse_packsswb_reg_reg (code, ins->sreg1, ins->sreg2);
6045 amd64_sse_packssdw_reg_reg (code, ins->sreg1, ins->sreg2);
6048 amd64_sse_packuswb_reg_reg (code, ins->sreg1, ins->sreg2);
6051 amd64_sse_packusdw_reg_reg (code, ins->sreg1, ins->sreg2);
6054 case OP_PADDB_SAT_UN:
6055 amd64_sse_paddusb_reg_reg (code, ins->sreg1, ins->sreg2);
6057 case OP_PSUBB_SAT_UN:
6058 amd64_sse_psubusb_reg_reg (code, ins->sreg1, ins->sreg2);
6060 case OP_PADDW_SAT_UN:
6061 amd64_sse_paddusw_reg_reg (code, ins->sreg1, ins->sreg2);
6063 case OP_PSUBW_SAT_UN:
6064 amd64_sse_psubusw_reg_reg (code, ins->sreg1, ins->sreg2);
6068 amd64_sse_paddsb_reg_reg (code, ins->sreg1, ins->sreg2);
6071 amd64_sse_psubsb_reg_reg (code, ins->sreg1, ins->sreg2);
6074 amd64_sse_paddsw_reg_reg (code, ins->sreg1, ins->sreg2);
6077 amd64_sse_psubsw_reg_reg (code, ins->sreg1, ins->sreg2);
6081 amd64_sse_pmullw_reg_reg (code, ins->sreg1, ins->sreg2);
6084 amd64_sse_pmulld_reg_reg (code, ins->sreg1, ins->sreg2);
6087 amd64_sse_pmuludq_reg_reg (code, ins->sreg1, ins->sreg2);
6089 case OP_PMULW_HIGH_UN:
6090 amd64_sse_pmulhuw_reg_reg (code, ins->sreg1, ins->sreg2);
6093 amd64_sse_pmulhw_reg_reg (code, ins->sreg1, ins->sreg2);
6097 amd64_sse_psrlw_reg_imm (code, ins->dreg, ins->inst_imm);
6100 amd64_sse_psrlw_reg_reg (code, ins->dreg, ins->sreg2);
6104 amd64_sse_psraw_reg_imm (code, ins->dreg, ins->inst_imm);
6107 amd64_sse_psraw_reg_reg (code, ins->dreg, ins->sreg2);
6111 amd64_sse_psllw_reg_imm (code, ins->dreg, ins->inst_imm);
6114 amd64_sse_psllw_reg_reg (code, ins->dreg, ins->sreg2);
6118 amd64_sse_psrld_reg_imm (code, ins->dreg, ins->inst_imm);
6121 amd64_sse_psrld_reg_reg (code, ins->dreg, ins->sreg2);
6125 amd64_sse_psrad_reg_imm (code, ins->dreg, ins->inst_imm);
6128 amd64_sse_psrad_reg_reg (code, ins->dreg, ins->sreg2);
6132 amd64_sse_pslld_reg_imm (code, ins->dreg, ins->inst_imm);
6135 amd64_sse_pslld_reg_reg (code, ins->dreg, ins->sreg2);
6139 amd64_sse_psrlq_reg_imm (code, ins->dreg, ins->inst_imm);
6142 amd64_sse_psrlq_reg_reg (code, ins->dreg, ins->sreg2);
6145 /*TODO: This is appart of the sse spec but not added
6147 amd64_sse_psraq_reg_imm (code, ins->dreg, ins->inst_imm);
6150 amd64_sse_psraq_reg_reg (code, ins->dreg, ins->sreg2);
6155 amd64_sse_psllq_reg_imm (code, ins->dreg, ins->inst_imm);
6158 amd64_sse_psllq_reg_reg (code, ins->dreg, ins->sreg2);
6161 amd64_sse_cvtdq2pd_reg_reg (code, ins->dreg, ins->sreg1);
6164 amd64_sse_cvtdq2ps_reg_reg (code, ins->dreg, ins->sreg1);
6167 amd64_sse_cvtpd2dq_reg_reg (code, ins->dreg, ins->sreg1);
6170 amd64_sse_cvtpd2ps_reg_reg (code, ins->dreg, ins->sreg1);
6173 amd64_sse_cvtps2dq_reg_reg (code, ins->dreg, ins->sreg1);
6176 amd64_sse_cvtps2pd_reg_reg (code, ins->dreg, ins->sreg1);
6179 amd64_sse_cvttpd2dq_reg_reg (code, ins->dreg, ins->sreg1);
6182 amd64_sse_cvttps2dq_reg_reg (code, ins->dreg, ins->sreg1);
6186 amd64_movd_xreg_reg_size (code, ins->dreg, ins->sreg1, 4);
6189 amd64_movd_reg_xreg_size (code, ins->dreg, ins->sreg1, 4);
6193 amd64_movhlps_reg_reg (code, AMD64_XMM15, ins->sreg1);
6194 amd64_movd_reg_xreg_size (code, ins->dreg, AMD64_XMM15, 8);
6196 amd64_movd_reg_xreg_size (code, ins->dreg, ins->sreg1, 8);
6201 amd64_movd_reg_xreg_size (code, ins->dreg, ins->sreg1, 4);
6203 amd64_shift_reg_imm (code, X86_SHR, ins->dreg, ins->inst_c0 * 8);
6204 amd64_widen_reg (code, ins->dreg, ins->dreg, ins->opcode == OP_EXTRACT_I1, FALSE);
6208 /*amd64_movd_reg_xreg_size (code, ins->dreg, ins->sreg1, 4);
6210 amd64_shift_reg_imm_size (code, X86_SHR, ins->dreg, 16, 4);*/
6211 amd64_sse_pextrw_reg_reg_imm (code, ins->dreg, ins->sreg1, ins->inst_c0);
6212 amd64_widen_reg_size (code, ins->dreg, ins->dreg, ins->opcode == OP_EXTRACT_I2, TRUE, 4);
6216 amd64_movhlps_reg_reg (code, ins->dreg, ins->sreg1);
6218 amd64_sse_movsd_reg_reg (code, ins->dreg, ins->sreg1);
6221 amd64_sse_pinsrw_reg_reg_imm (code, ins->sreg1, ins->sreg2, ins->inst_c0);
6223 case OP_EXTRACTX_U2:
6224 amd64_sse_pextrw_reg_reg_imm (code, ins->dreg, ins->sreg1, ins->inst_c0);
6226 case OP_INSERTX_U1_SLOW:
6227 /*sreg1 is the extracted ireg (scratch)
6228 /sreg2 is the to be inserted ireg (scratch)
6229 /dreg is the xreg to receive the value*/
6231 /*clear the bits from the extracted word*/
6232 amd64_alu_reg_imm (code, X86_AND, ins->sreg1, ins->inst_c0 & 1 ? 0x00FF : 0xFF00);
6233 /*shift the value to insert if needed*/
6234 if (ins->inst_c0 & 1)
6235 amd64_shift_reg_imm_size (code, X86_SHL, ins->sreg2, 8, 4);
6236 /*join them together*/
6237 amd64_alu_reg_reg (code, X86_OR, ins->sreg1, ins->sreg2);
6238 amd64_sse_pinsrw_reg_reg_imm (code, ins->dreg, ins->sreg1, ins->inst_c0 / 2);
6240 case OP_INSERTX_I4_SLOW:
6241 amd64_sse_pinsrw_reg_reg_imm (code, ins->dreg, ins->sreg2, ins->inst_c0 * 2);
6242 amd64_shift_reg_imm (code, X86_SHR, ins->sreg2, 16);
6243 amd64_sse_pinsrw_reg_reg_imm (code, ins->dreg, ins->sreg2, ins->inst_c0 * 2 + 1);
6245 case OP_INSERTX_I8_SLOW:
6246 amd64_movd_xreg_reg_size(code, AMD64_XMM15, ins->sreg2, 8);
6248 amd64_movlhps_reg_reg (code, ins->dreg, AMD64_XMM15);
6250 amd64_sse_movsd_reg_reg (code, ins->dreg, AMD64_XMM15);
6253 case OP_INSERTX_R4_SLOW:
6254 switch (ins->inst_c0) {
6256 amd64_sse_cvtsd2ss_reg_reg (code, ins->dreg, ins->sreg2);
6259 amd64_sse_pshufd_reg_reg_imm (code, ins->dreg, ins->dreg, mono_simd_shuffle_mask(1, 0, 2, 3));
6260 amd64_sse_cvtsd2ss_reg_reg (code, ins->dreg, ins->sreg2);
6261 amd64_sse_pshufd_reg_reg_imm (code, ins->dreg, ins->dreg, mono_simd_shuffle_mask(1, 0, 2, 3));
6264 amd64_sse_pshufd_reg_reg_imm (code, ins->dreg, ins->dreg, mono_simd_shuffle_mask(2, 1, 0, 3));
6265 amd64_sse_cvtsd2ss_reg_reg (code, ins->dreg, ins->sreg2);
6266 amd64_sse_pshufd_reg_reg_imm (code, ins->dreg, ins->dreg, mono_simd_shuffle_mask(2, 1, 0, 3));
6269 amd64_sse_pshufd_reg_reg_imm (code, ins->dreg, ins->dreg, mono_simd_shuffle_mask(3, 1, 2, 0));
6270 amd64_sse_cvtsd2ss_reg_reg (code, ins->dreg, ins->sreg2);
6271 amd64_sse_pshufd_reg_reg_imm (code, ins->dreg, ins->dreg, mono_simd_shuffle_mask(3, 1, 2, 0));
6275 case OP_INSERTX_R8_SLOW:
6277 amd64_movlhps_reg_reg (code, ins->dreg, ins->sreg2);
6279 amd64_sse_movsd_reg_reg (code, ins->dreg, ins->sreg2);
6281 case OP_STOREX_MEMBASE_REG:
6282 case OP_STOREX_MEMBASE:
6283 amd64_sse_movups_membase_reg (code, ins->dreg, ins->inst_offset, ins->sreg1);
6285 case OP_LOADX_MEMBASE:
6286 amd64_sse_movups_reg_membase (code, ins->dreg, ins->sreg1, ins->inst_offset);
6288 case OP_LOADX_ALIGNED_MEMBASE:
6289 amd64_sse_movaps_reg_membase (code, ins->dreg, ins->sreg1, ins->inst_offset);
6291 case OP_STOREX_ALIGNED_MEMBASE_REG:
6292 amd64_sse_movaps_membase_reg (code, ins->dreg, ins->inst_offset, ins->sreg1);
6294 case OP_STOREX_NTA_MEMBASE_REG:
6295 amd64_sse_movntps_reg_membase (code, ins->dreg, ins->sreg1, ins->inst_offset);
6297 case OP_PREFETCH_MEMBASE:
6298 amd64_sse_prefetch_reg_membase (code, ins->backend.arg_info, ins->sreg1, ins->inst_offset);
6302 /*FIXME the peephole pass should have killed this*/
6303 if (ins->dreg != ins->sreg1)
6304 amd64_sse_movaps_reg_reg (code, ins->dreg, ins->sreg1);
6307 amd64_sse_pxor_reg_reg (code, ins->dreg, ins->dreg);
6309 case OP_ICONV_TO_R8_RAW:
6310 amd64_movd_xreg_reg_size (code, ins->dreg, ins->sreg1, 4);
6311 amd64_sse_cvtss2sd_reg_reg (code, ins->dreg, ins->dreg);
6314 case OP_FCONV_TO_R8_X:
6315 amd64_sse_movsd_reg_reg (code, ins->dreg, ins->sreg1);
6318 case OP_XCONV_R8_TO_I4:
6319 amd64_sse_cvttsd2si_reg_xreg_size (code, ins->dreg, ins->sreg1, 4);
6320 switch (ins->backend.source_opcode) {
6321 case OP_FCONV_TO_I1:
6322 amd64_widen_reg (code, ins->dreg, ins->dreg, TRUE, FALSE);
6324 case OP_FCONV_TO_U1:
6325 amd64_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
6327 case OP_FCONV_TO_I2:
6328 amd64_widen_reg (code, ins->dreg, ins->dreg, TRUE, TRUE);
6330 case OP_FCONV_TO_U2:
6331 amd64_widen_reg (code, ins->dreg, ins->dreg, FALSE, TRUE);
6337 amd64_sse_pinsrw_reg_reg_imm (code, ins->dreg, ins->sreg1, 0);
6338 amd64_sse_pinsrw_reg_reg_imm (code, ins->dreg, ins->sreg1, 1);
6339 amd64_sse_pshufd_reg_reg_imm (code, ins->dreg, ins->dreg, 0);
6342 amd64_movd_xreg_reg_size (code, ins->dreg, ins->sreg1, 4);
6343 amd64_sse_pshufd_reg_reg_imm (code, ins->dreg, ins->dreg, 0);
6346 amd64_movd_xreg_reg_size (code, ins->dreg, ins->sreg1, 8);
6347 amd64_sse_pshufd_reg_reg_imm (code, ins->dreg, ins->dreg, 0x44);
6350 amd64_sse_movsd_reg_reg (code, ins->dreg, ins->sreg1);
6351 amd64_sse_cvtsd2ss_reg_reg (code, ins->dreg, ins->dreg);
6352 amd64_sse_pshufd_reg_reg_imm (code, ins->dreg, ins->dreg, 0);
6355 amd64_sse_movsd_reg_reg (code, ins->dreg, ins->sreg1);
6356 amd64_sse_pshufd_reg_reg_imm (code, ins->dreg, ins->dreg, 0x44);
6359 case OP_LIVERANGE_START: {
6360 if (cfg->verbose_level > 1)
6361 printf ("R%d START=0x%x\n", MONO_VARINFO (cfg, ins->inst_c0)->vreg, (int)(code - cfg->native_code));
6362 MONO_VARINFO (cfg, ins->inst_c0)->live_range_start = code - cfg->native_code;
6365 case OP_LIVERANGE_END: {
6366 if (cfg->verbose_level > 1)
6367 printf ("R%d END=0x%x\n", MONO_VARINFO (cfg, ins->inst_c0)->vreg, (int)(code - cfg->native_code));
6368 MONO_VARINFO (cfg, ins->inst_c0)->live_range_end = code - cfg->native_code;
6371 case OP_NACL_GC_SAFE_POINT: {
6372 #if defined(__native_client_codegen__)
6373 code = emit_call (cfg, code, MONO_PATCH_INFO_ABS, (gpointer)mono_nacl_gc, TRUE);
6377 case OP_GC_LIVENESS_DEF:
6378 case OP_GC_LIVENESS_USE:
6379 case OP_GC_PARAM_SLOT_LIVENESS_DEF:
6380 ins->backend.pc_offset = code - cfg->native_code;
6382 case OP_GC_SPILL_SLOT_LIVENESS_DEF:
6383 ins->backend.pc_offset = code - cfg->native_code;
6384 bb->spill_slot_defs = g_slist_prepend_mempool (cfg->mempool, bb->spill_slot_defs, ins);
6387 g_warning ("unknown opcode %s in %s()\n", mono_inst_name (ins->opcode), __FUNCTION__);
6388 g_assert_not_reached ();
6391 if ((code - cfg->native_code - offset) > max_len) {
6392 #if !defined(__native_client_codegen__)
6393 g_warning ("wrong maximal instruction length of instruction %s (expected %d, got %ld)",
6394 mono_inst_name (ins->opcode), max_len, code - cfg->native_code - offset);
6395 g_assert_not_reached ();
6400 last_offset = offset;
6403 cfg->code_len = code - cfg->native_code;
6406 #endif /* DISABLE_JIT */
6409 mono_arch_register_lowlevel_calls (void)
6411 /* The signature doesn't matter */
6412 mono_register_jit_icall (mono_amd64_throw_exception, "mono_amd64_throw_exception", mono_create_icall_signature ("void"), TRUE);
6416 mono_arch_patch_code (MonoMethod *method, MonoDomain *domain, guint8 *code, MonoJumpInfo *ji, MonoCodeManager *dyn_code_mp, gboolean run_cctors)
6418 MonoJumpInfo *patch_info;
6419 gboolean compile_aot = !run_cctors;
6421 for (patch_info = ji; patch_info; patch_info = patch_info->next) {
6422 unsigned char *ip = patch_info->ip.i + code;
6423 unsigned char *target;
6425 target = mono_resolve_patch_target (method, domain, code, patch_info, run_cctors);
6428 switch (patch_info->type) {
6429 case MONO_PATCH_INFO_BB:
6430 case MONO_PATCH_INFO_LABEL:
6433 /* No need to patch these */
6438 switch (patch_info->type) {
6439 case MONO_PATCH_INFO_NONE:
6441 case MONO_PATCH_INFO_METHOD_REL:
6442 case MONO_PATCH_INFO_R8:
6443 case MONO_PATCH_INFO_R4:
6444 g_assert_not_reached ();
6446 case MONO_PATCH_INFO_BB:
6453 * Debug code to help track down problems where the target of a near call is
6456 if (amd64_is_near_call (ip)) {
6457 gint64 disp = (guint8*)target - (guint8*)ip;
6459 if (!amd64_is_imm32 (disp)) {
6460 printf ("TYPE: %d\n", patch_info->type);
6461 switch (patch_info->type) {
6462 case MONO_PATCH_INFO_INTERNAL_METHOD:
6463 printf ("V: %s\n", patch_info->data.name);
6465 case MONO_PATCH_INFO_METHOD_JUMP:
6466 case MONO_PATCH_INFO_METHOD:
6467 printf ("V: %s\n", patch_info->data.method->name);
6475 amd64_patch (ip, (gpointer)target);
6482 get_max_epilog_size (MonoCompile *cfg)
6484 int max_epilog_size = 16;
6486 if (cfg->method->save_lmf)
6487 max_epilog_size += 256;
6489 if (mono_jit_trace_calls != NULL)
6490 max_epilog_size += 50;
6492 if (cfg->prof_options & MONO_PROFILE_ENTER_LEAVE)
6493 max_epilog_size += 50;
6495 max_epilog_size += (AMD64_NREG * 2);
6497 return max_epilog_size;
6501 * This macro is used for testing whenever the unwinder works correctly at every point
6502 * where an async exception can happen.
6504 /* This will generate a SIGSEGV at the given point in the code */
6505 #define async_exc_point(code) do { \
6506 if (mono_inject_async_exc_method && mono_method_desc_full_match (mono_inject_async_exc_method, cfg->method)) { \
6507 if (cfg->arch.async_point_count == mono_inject_async_exc_pos) \
6508 amd64_mov_reg_mem (code, AMD64_RAX, 0, 4); \
6509 cfg->arch.async_point_count ++; \
6514 mono_arch_emit_prolog (MonoCompile *cfg)
6516 MonoMethod *method = cfg->method;
6518 MonoMethodSignature *sig;
6520 int alloc_size, pos, i, cfa_offset, quad, max_epilog_size;
6523 MonoInst *lmf_var = cfg->arch.lmf_var;
6524 gboolean args_clobbered = FALSE;
6525 gboolean trace = FALSE;
6526 #ifdef __native_client_codegen__
6527 guint alignment_check;
6530 cfg->code_size = MAX (cfg->header->code_size * 4, 10240);
6532 #if defined(__default_codegen__)
6533 code = cfg->native_code = g_malloc (cfg->code_size);
6534 #elif defined(__native_client_codegen__)
6535 /* native_code_alloc is not 32-byte aligned, native_code is. */
6536 cfg->native_code_alloc = g_malloc (cfg->code_size + kNaClAlignment);
6538 /* Align native_code to next nearest kNaclAlignment byte. */
6539 cfg->native_code = (uintptr_t)cfg->native_code_alloc + kNaClAlignment;
6540 cfg->native_code = (uintptr_t)cfg->native_code & ~kNaClAlignmentMask;
6542 code = cfg->native_code;
6544 alignment_check = (guint)cfg->native_code & kNaClAlignmentMask;
6545 g_assert (alignment_check == 0);
6548 if (mono_jit_trace_calls != NULL && mono_trace_eval (method))
6551 /* Amount of stack space allocated by register saving code */
6554 /* Offset between RSP and the CFA */
6558 * The prolog consists of the following parts:
6560 * - push rbp, mov rbp, rsp
6561 * - save callee saved regs using pushes
6563 * - save rgctx if needed
6564 * - save lmf if needed
6567 * - save rgctx if needed
6568 * - save lmf if needed
6569 * - save callee saved regs using moves
6574 mono_emit_unwind_op_def_cfa (cfg, code, AMD64_RSP, 8);
6575 // IP saved at CFA - 8
6576 mono_emit_unwind_op_offset (cfg, code, AMD64_RIP, -cfa_offset);
6577 async_exc_point (code);
6578 mini_gc_set_slot_type_from_cfa (cfg, -cfa_offset, SLOT_NOREF);
6580 if (!cfg->arch.omit_fp) {
6581 amd64_push_reg (code, AMD64_RBP);
6583 mono_emit_unwind_op_def_cfa_offset (cfg, code, cfa_offset);
6584 mono_emit_unwind_op_offset (cfg, code, AMD64_RBP, - cfa_offset);
6585 async_exc_point (code);
6587 mono_arch_unwindinfo_add_push_nonvol (&cfg->arch.unwindinfo, cfg->native_code, code, AMD64_RBP);
6589 /* These are handled automatically by the stack marking code */
6590 mini_gc_set_slot_type_from_cfa (cfg, -cfa_offset, SLOT_NOREF);
6592 amd64_mov_reg_reg (code, AMD64_RBP, AMD64_RSP, sizeof(mgreg_t));
6593 mono_emit_unwind_op_def_cfa_reg (cfg, code, AMD64_RBP);
6594 async_exc_point (code);
6596 mono_arch_unwindinfo_add_set_fpreg (&cfg->arch.unwindinfo, cfg->native_code, code, AMD64_RBP);
6600 /* Save callee saved registers */
6601 if (!cfg->arch.omit_fp && !method->save_lmf) {
6602 int offset = cfa_offset;
6604 for (i = 0; i < AMD64_NREG; ++i)
6605 if (AMD64_IS_CALLEE_SAVED_REG (i) && (cfg->used_int_regs & (1 << i))) {
6606 amd64_push_reg (code, i);
6607 pos += 8; /* AMD64 push inst is always 8 bytes, no way to change it */
6609 mono_emit_unwind_op_offset (cfg, code, i, - offset);
6610 async_exc_point (code);
6612 /* These are handled automatically by the stack marking code */
6613 mini_gc_set_slot_type_from_cfa (cfg, - offset, SLOT_NOREF);
6617 /* The param area is always at offset 0 from sp */
6618 /* This needs to be allocated here, since it has to come after the spill area */
6619 if (cfg->arch.no_pushes && cfg->param_area) {
6620 if (cfg->arch.omit_fp)
6622 g_assert_not_reached ();
6623 cfg->stack_offset += ALIGN_TO (cfg->param_area, sizeof(mgreg_t));
6626 if (cfg->arch.omit_fp) {
6628 * On enter, the stack is misaligned by the pushing of the return
6629 * address. It is either made aligned by the pushing of %rbp, or by
6632 alloc_size = ALIGN_TO (cfg->stack_offset, 8);
6633 if ((alloc_size % 16) == 0) {
6635 /* Mark the padding slot as NOREF */
6636 mini_gc_set_slot_type_from_cfa (cfg, -cfa_offset - sizeof (mgreg_t), SLOT_NOREF);
6639 alloc_size = ALIGN_TO (cfg->stack_offset, MONO_ARCH_FRAME_ALIGNMENT);
6640 if (cfg->stack_offset != alloc_size) {
6641 /* Mark the padding slot as NOREF */
6642 mini_gc_set_slot_type_from_fp (cfg, -alloc_size + cfg->param_area, SLOT_NOREF);
6644 cfg->arch.sp_fp_offset = alloc_size;
6648 cfg->arch.stack_alloc_size = alloc_size;
6650 /* Allocate stack frame */
6652 /* See mono_emit_stack_alloc */
6653 #if defined(HOST_WIN32) || defined(MONO_ARCH_SIGSEGV_ON_ALTSTACK)
6654 guint32 remaining_size = alloc_size;
6655 /*FIXME handle unbounded code expansion, we should use a loop in case of more than X interactions*/
6656 guint32 required_code_size = ((remaining_size / 0x1000) + 1) * 10; /*10 is the max size of amd64_alu_reg_imm + amd64_test_membase_reg*/
6657 guint32 offset = code - cfg->native_code;
6658 if (G_UNLIKELY (required_code_size >= (cfg->code_size - offset))) {
6659 while (required_code_size >= (cfg->code_size - offset))
6660 cfg->code_size *= 2;
6661 cfg->native_code = mono_realloc_native_code (cfg);
6662 code = cfg->native_code + offset;
6663 cfg->stat_code_reallocs++;
6666 while (remaining_size >= 0x1000) {
6667 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, 0x1000);
6668 if (cfg->arch.omit_fp) {
6669 cfa_offset += 0x1000;
6670 mono_emit_unwind_op_def_cfa_offset (cfg, code, cfa_offset);
6672 async_exc_point (code);
6674 if (cfg->arch.omit_fp)
6675 mono_arch_unwindinfo_add_alloc_stack (&cfg->arch.unwindinfo, cfg->native_code, code, 0x1000);
6678 amd64_test_membase_reg (code, AMD64_RSP, 0, AMD64_RSP);
6679 remaining_size -= 0x1000;
6681 if (remaining_size) {
6682 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, remaining_size);
6683 if (cfg->arch.omit_fp) {
6684 cfa_offset += remaining_size;
6685 mono_emit_unwind_op_def_cfa_offset (cfg, code, cfa_offset);
6686 async_exc_point (code);
6689 if (cfg->arch.omit_fp)
6690 mono_arch_unwindinfo_add_alloc_stack (&cfg->arch.unwindinfo, cfg->native_code, code, remaining_size);
6694 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, alloc_size);
6695 if (cfg->arch.omit_fp) {
6696 cfa_offset += alloc_size;
6697 mono_emit_unwind_op_def_cfa_offset (cfg, code, cfa_offset);
6698 async_exc_point (code);
6703 /* Stack alignment check */
6706 amd64_mov_reg_reg (code, AMD64_RAX, AMD64_RSP, 8);
6707 amd64_alu_reg_imm (code, X86_AND, AMD64_RAX, 0xf);
6708 amd64_alu_reg_imm (code, X86_CMP, AMD64_RAX, 0);
6709 x86_branch8 (code, X86_CC_EQ, 2, FALSE);
6710 amd64_breakpoint (code);
6714 #ifndef TARGET_WIN32
6715 if (mini_get_debug_options ()->init_stacks) {
6716 /* Fill the stack frame with a dummy value to force deterministic behavior */
6718 /* Save registers to the red zone */
6719 amd64_mov_membase_reg (code, AMD64_RSP, -8, AMD64_RDI, 8);
6720 amd64_mov_membase_reg (code, AMD64_RSP, -16, AMD64_RCX, 8);
6722 amd64_mov_reg_imm (code, AMD64_RAX, 0x2a2a2a2a2a2a2a2a);
6723 amd64_mov_reg_imm (code, AMD64_RCX, alloc_size / 8);
6724 amd64_mov_reg_reg (code, AMD64_RDI, AMD64_RSP, 8);
6727 #if defined(__default_codegen__)
6728 amd64_prefix (code, X86_REP_PREFIX);
6730 #elif defined(__native_client_codegen__)
6731 /* NaCl stos pseudo-instruction */
6732 amd64_codegen_pre (code);
6733 /* First, clear the upper 32 bits of RDI (mov %edi, %edi) */
6734 amd64_mov_reg_reg (code, AMD64_RDI, AMD64_RDI, 4);
6735 /* Add %r15 to %rdi using lea, condition flags unaffected. */
6736 amd64_lea_memindex_size (code, AMD64_RDI, AMD64_R15, 0, AMD64_RDI, 0, 8);
6737 amd64_prefix (code, X86_REP_PREFIX);
6739 amd64_codegen_post (code);
6740 #endif /* __native_client_codegen__ */
6742 amd64_mov_reg_membase (code, AMD64_RDI, AMD64_RSP, -8, 8);
6743 amd64_mov_reg_membase (code, AMD64_RCX, AMD64_RSP, -16, 8);
6748 if (method->save_lmf) {
6749 code = emit_setup_lmf (cfg, code, lmf_var->inst_offset, cfa_offset);
6752 /* Save callee saved registers */
6753 if (cfg->arch.omit_fp && !method->save_lmf) {
6754 gint32 save_area_offset = cfg->arch.reg_save_area_offset;
6756 /* Save caller saved registers after sp is adjusted */
6757 /* The registers are saved at the bottom of the frame */
6758 /* FIXME: Optimize this so the regs are saved at the end of the frame in increasing order */
6759 for (i = 0; i < AMD64_NREG; ++i)
6760 if (AMD64_IS_CALLEE_SAVED_REG (i) && (cfg->used_int_regs & (1 << i))) {
6761 amd64_mov_membase_reg (code, AMD64_RSP, save_area_offset, i, 8);
6762 mono_emit_unwind_op_offset (cfg, code, i, - (cfa_offset - save_area_offset));
6764 /* These are handled automatically by the stack marking code */
6765 mini_gc_set_slot_type_from_cfa (cfg, - (cfa_offset - save_area_offset), SLOT_NOREF);
6767 save_area_offset += 8;
6768 async_exc_point (code);
6772 /* store runtime generic context */
6773 if (cfg->rgctx_var) {
6774 g_assert (cfg->rgctx_var->opcode == OP_REGOFFSET &&
6775 (cfg->rgctx_var->inst_basereg == AMD64_RBP || cfg->rgctx_var->inst_basereg == AMD64_RSP));
6777 amd64_mov_membase_reg (code, cfg->rgctx_var->inst_basereg, cfg->rgctx_var->inst_offset, MONO_ARCH_RGCTX_REG, sizeof(gpointer));
6780 /* compute max_length in order to use short forward jumps */
6781 max_epilog_size = get_max_epilog_size (cfg);
6782 if (cfg->opt & MONO_OPT_BRANCH) {
6783 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
6787 if (cfg->prof_options & MONO_PROFILE_COVERAGE)
6789 /* max alignment for loops */
6790 if ((cfg->opt & MONO_OPT_LOOP) && bb_is_loop_start (bb))
6791 max_length += LOOP_ALIGNMENT;
6792 #ifdef __native_client_codegen__
6793 /* max alignment for native client */
6794 max_length += kNaClAlignment;
6797 MONO_BB_FOR_EACH_INS (bb, ins) {
6798 #ifdef __native_client_codegen__
6800 int space_in_block = kNaClAlignment -
6801 ((max_length + cfg->code_len) & kNaClAlignmentMask);
6802 int max_len = ((guint8 *)ins_get_spec (ins->opcode))[MONO_INST_LEN];
6803 if (space_in_block < max_len && max_len < kNaClAlignment) {
6804 max_length += space_in_block;
6807 #endif /*__native_client_codegen__*/
6808 max_length += ((guint8 *)ins_get_spec (ins->opcode))[MONO_INST_LEN];
6811 /* Take prolog and epilog instrumentation into account */
6812 if (bb == cfg->bb_entry || bb == cfg->bb_exit)
6813 max_length += max_epilog_size;
6815 bb->max_length = max_length;
6819 sig = mono_method_signature (method);
6822 cinfo = cfg->arch.cinfo;
6824 if (sig->ret->type != MONO_TYPE_VOID) {
6825 /* Save volatile arguments to the stack */
6826 if (cfg->vret_addr && (cfg->vret_addr->opcode != OP_REGVAR))
6827 amd64_mov_membase_reg (code, cfg->vret_addr->inst_basereg, cfg->vret_addr->inst_offset, cinfo->ret.reg, 8);
6830 /* Keep this in sync with emit_load_volatile_arguments */
6831 for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
6832 ArgInfo *ainfo = cinfo->args + i;
6833 gint32 stack_offset;
6836 ins = cfg->args [i];
6838 if ((ins->flags & MONO_INST_IS_DEAD) && !trace)
6839 /* Unused arguments */
6842 if (sig->hasthis && (i == 0))
6843 arg_type = &mono_defaults.object_class->byval_arg;
6845 arg_type = sig->params [i - sig->hasthis];
6847 stack_offset = ainfo->offset + ARGS_OFFSET;
6849 if (cfg->globalra) {
6850 /* All the other moves are done by the register allocator */
6851 switch (ainfo->storage) {
6852 case ArgInFloatSSEReg:
6853 amd64_sse_cvtss2sd_reg_reg (code, ainfo->reg, ainfo->reg);
6855 case ArgValuetypeInReg:
6856 for (quad = 0; quad < 2; quad ++) {
6857 switch (ainfo->pair_storage [quad]) {
6859 amd64_mov_membase_reg (code, ins->inst_basereg, ins->inst_offset + (quad * sizeof(mgreg_t)), ainfo->pair_regs [quad], sizeof(mgreg_t));
6861 case ArgInFloatSSEReg:
6862 amd64_movss_membase_reg (code, ins->inst_basereg, ins->inst_offset + (quad * sizeof(mgreg_t)), ainfo->pair_regs [quad]);
6864 case ArgInDoubleSSEReg:
6865 amd64_movsd_membase_reg (code, ins->inst_basereg, ins->inst_offset + (quad * sizeof(mgreg_t)), ainfo->pair_regs [quad]);
6870 g_assert_not_reached ();
6881 /* Save volatile arguments to the stack */
6882 if (ins->opcode != OP_REGVAR) {
6883 switch (ainfo->storage) {
6889 if (stack_offset & 0x1)
6891 else if (stack_offset & 0x2)
6893 else if (stack_offset & 0x4)
6898 amd64_mov_membase_reg (code, ins->inst_basereg, ins->inst_offset, ainfo->reg, size);
6901 case ArgInFloatSSEReg:
6902 amd64_movss_membase_reg (code, ins->inst_basereg, ins->inst_offset, ainfo->reg);
6904 case ArgInDoubleSSEReg:
6905 amd64_movsd_membase_reg (code, ins->inst_basereg, ins->inst_offset, ainfo->reg);
6907 case ArgValuetypeInReg:
6908 for (quad = 0; quad < 2; quad ++) {
6909 switch (ainfo->pair_storage [quad]) {
6911 amd64_mov_membase_reg (code, ins->inst_basereg, ins->inst_offset + (quad * sizeof(mgreg_t)), ainfo->pair_regs [quad], sizeof(mgreg_t));
6913 case ArgInFloatSSEReg:
6914 amd64_movss_membase_reg (code, ins->inst_basereg, ins->inst_offset + (quad * sizeof(mgreg_t)), ainfo->pair_regs [quad]);
6916 case ArgInDoubleSSEReg:
6917 amd64_movsd_membase_reg (code, ins->inst_basereg, ins->inst_offset + (quad * sizeof(mgreg_t)), ainfo->pair_regs [quad]);
6922 g_assert_not_reached ();
6926 case ArgValuetypeAddrInIReg:
6927 if (ainfo->pair_storage [0] == ArgInIReg)
6928 amd64_mov_membase_reg (code, ins->inst_left->inst_basereg, ins->inst_left->inst_offset, ainfo->pair_regs [0], sizeof (gpointer));
6934 /* Argument allocated to (non-volatile) register */
6935 switch (ainfo->storage) {
6937 amd64_mov_reg_reg (code, ins->dreg, ainfo->reg, 8);
6940 amd64_mov_reg_membase (code, ins->dreg, AMD64_RBP, ARGS_OFFSET + ainfo->offset, 8);
6943 g_assert_not_reached ();
6948 /* Might need to attach the thread to the JIT or change the domain for the callback */
6949 if (method->wrapper_type == MONO_WRAPPER_NATIVE_TO_MANAGED) {
6950 guint64 domain = (guint64)cfg->domain;
6952 args_clobbered = TRUE;
6955 * The call might clobber argument registers, but they are already
6956 * saved to the stack/global regs.
6958 if (appdomain_tls_offset != -1 && lmf_tls_offset != -1) {
6959 guint8 *buf, *no_domain_branch;
6961 code = mono_amd64_emit_tls_get (code, AMD64_RAX, appdomain_tls_offset);
6962 if (cfg->compile_aot) {
6963 /* AOT code is only used in the root domain */
6964 amd64_mov_reg_imm (code, AMD64_ARG_REG1, 0);
6966 if ((domain >> 32) == 0)
6967 amd64_mov_reg_imm_size (code, AMD64_ARG_REG1, domain, 4);
6969 amd64_mov_reg_imm_size (code, AMD64_ARG_REG1, domain, 8);
6971 amd64_alu_reg_reg (code, X86_CMP, AMD64_RAX, AMD64_ARG_REG1);
6972 no_domain_branch = code;
6973 x86_branch8 (code, X86_CC_NE, 0, 0);
6974 code = mono_amd64_emit_tls_get ( code, AMD64_RAX, lmf_addr_tls_offset);
6975 amd64_test_reg_reg (code, AMD64_RAX, AMD64_RAX);
6977 x86_branch8 (code, X86_CC_NE, 0, 0);
6978 amd64_patch (no_domain_branch, code);
6979 code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD,
6980 (gpointer)"mono_jit_thread_attach", TRUE);
6981 amd64_patch (buf, code);
6983 /* The TLS key actually contains a pointer to the MonoJitTlsData structure */
6984 /* FIXME: Add a separate key for LMF to avoid this */
6985 amd64_alu_reg_imm (code, X86_ADD, AMD64_RAX, G_STRUCT_OFFSET (MonoJitTlsData, lmf));
6988 g_assert (!cfg->compile_aot);
6989 if (cfg->compile_aot) {
6990 /* AOT code is only used in the root domain */
6991 amd64_mov_reg_imm (code, AMD64_ARG_REG1, 0);
6993 if ((domain >> 32) == 0)
6994 amd64_mov_reg_imm_size (code, AMD64_ARG_REG1, domain, 4);
6996 amd64_mov_reg_imm_size (code, AMD64_ARG_REG1, domain, 8);
6998 code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD,
6999 (gpointer)"mono_jit_thread_attach", TRUE);
7003 if (method->save_lmf) {
7004 code = emit_save_lmf (cfg, code, lmf_var->inst_offset, &args_clobbered);
7008 args_clobbered = TRUE;
7009 code = mono_arch_instrument_prolog (cfg, mono_trace_enter_method, code, TRUE);
7012 if (cfg->prof_options & MONO_PROFILE_ENTER_LEAVE)
7013 args_clobbered = TRUE;
7016 * Optimize the common case of the first bblock making a call with the same
7017 * arguments as the method. This works because the arguments are still in their
7018 * original argument registers.
7019 * FIXME: Generalize this
7021 if (!args_clobbered) {
7022 MonoBasicBlock *first_bb = cfg->bb_entry;
7025 next = mono_bb_first_ins (first_bb);
7026 if (!next && first_bb->next_bb) {
7027 first_bb = first_bb->next_bb;
7028 next = mono_bb_first_ins (first_bb);
7031 if (first_bb->in_count > 1)
7034 for (i = 0; next && i < sig->param_count + sig->hasthis; ++i) {
7035 ArgInfo *ainfo = cinfo->args + i;
7036 gboolean match = FALSE;
7038 ins = cfg->args [i];
7039 if (ins->opcode != OP_REGVAR) {
7040 switch (ainfo->storage) {
7042 if (((next->opcode == OP_LOAD_MEMBASE) || (next->opcode == OP_LOADI4_MEMBASE)) && next->inst_basereg == ins->inst_basereg && next->inst_offset == ins->inst_offset) {
7043 if (next->dreg == ainfo->reg) {
7047 next->opcode = OP_MOVE;
7048 next->sreg1 = ainfo->reg;
7049 /* Only continue if the instruction doesn't change argument regs */
7050 if (next->dreg == ainfo->reg || next->dreg == AMD64_RAX)
7060 /* Argument allocated to (non-volatile) register */
7061 switch (ainfo->storage) {
7063 if (next->opcode == OP_MOVE && next->sreg1 == ins->dreg && next->dreg == ainfo->reg) {
7075 //next = mono_inst_list_next (&next->node, &first_bb->ins_list);
7082 /* Initialize ss_trigger_page_var */
7083 if (cfg->arch.ss_trigger_page_var) {
7084 MonoInst *var = cfg->arch.ss_trigger_page_var;
7086 g_assert (!cfg->compile_aot);
7087 g_assert (var->opcode == OP_REGOFFSET);
7089 amd64_mov_reg_imm (code, AMD64_R11, (guint64)ss_trigger_page);
7090 amd64_mov_membase_reg (code, var->inst_basereg, var->inst_offset, AMD64_R11, 8);
7093 cfg->code_len = code - cfg->native_code;
7095 g_assert (cfg->code_len < cfg->code_size);
7101 mono_arch_emit_epilog (MonoCompile *cfg)
7103 MonoMethod *method = cfg->method;
7106 int max_epilog_size;
7108 gint32 lmf_offset = cfg->arch.lmf_var ? ((MonoInst*)cfg->arch.lmf_var)->inst_offset : -1;
7110 max_epilog_size = get_max_epilog_size (cfg);
7112 while (cfg->code_len + max_epilog_size > (cfg->code_size - 16)) {
7113 cfg->code_size *= 2;
7114 cfg->native_code = mono_realloc_native_code (cfg);
7115 cfg->stat_code_reallocs++;
7118 code = cfg->native_code + cfg->code_len;
7120 if (mono_jit_trace_calls != NULL && mono_trace_eval (method))
7121 code = mono_arch_instrument_epilog (cfg, mono_trace_leave_method, code, TRUE);
7123 /* the code restoring the registers must be kept in sync with OP_JMP */
7126 if (method->save_lmf) {
7127 /* check if we need to restore protection of the stack after a stack overflow */
7128 if (mono_get_jit_tls_offset () != -1) {
7130 code = mono_amd64_emit_tls_get (code, AMD64_RCX, mono_get_jit_tls_offset ());
7131 /* we load the value in a separate instruction: this mechanism may be
7132 * used later as a safer way to do thread interruption
7134 amd64_mov_reg_membase (code, AMD64_RCX, AMD64_RCX, G_STRUCT_OFFSET (MonoJitTlsData, restore_stack_prot), 8);
7135 x86_alu_reg_imm (code, X86_CMP, X86_ECX, 0);
7137 x86_branch8 (code, X86_CC_Z, 0, FALSE);
7138 /* note that the call trampoline will preserve eax/edx */
7139 x86_call_reg (code, X86_ECX);
7140 x86_patch (patch, code);
7142 /* FIXME: maybe save the jit tls in the prolog */
7145 code = emit_restore_lmf (cfg, code, lmf_offset);
7147 /* Restore caller saved regs */
7148 if (cfg->used_int_regs & (1 << AMD64_RBP)) {
7149 amd64_mov_reg_membase (code, AMD64_RBP, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rbp), 8);
7151 if (cfg->used_int_regs & (1 << AMD64_RBX)) {
7152 amd64_mov_reg_membase (code, AMD64_RBX, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rbx), 8);
7154 if (cfg->used_int_regs & (1 << AMD64_R12)) {
7155 amd64_mov_reg_membase (code, AMD64_R12, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r12), 8);
7157 if (cfg->used_int_regs & (1 << AMD64_R13)) {
7158 amd64_mov_reg_membase (code, AMD64_R13, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r13), 8);
7160 if (cfg->used_int_regs & (1 << AMD64_R14)) {
7161 amd64_mov_reg_membase (code, AMD64_R14, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r14), 8);
7163 if (cfg->used_int_regs & (1 << AMD64_R15)) {
7164 #if defined(__default_codegen__)
7165 amd64_mov_reg_membase (code, AMD64_R15, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r15), 8);
7166 #elif defined(__native_client_codegen__)
7167 g_assert_not_reached();
7171 if (cfg->used_int_regs & (1 << AMD64_RDI)) {
7172 amd64_mov_reg_membase (code, AMD64_RDI, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rdi), 8);
7174 if (cfg->used_int_regs & (1 << AMD64_RSI)) {
7175 amd64_mov_reg_membase (code, AMD64_RSI, cfg->frame_reg, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rsi), 8);
7180 if (cfg->arch.omit_fp) {
7181 gint32 save_area_offset = cfg->arch.reg_save_area_offset;
7183 for (i = 0; i < AMD64_NREG; ++i)
7184 if (AMD64_IS_CALLEE_SAVED_REG (i) && (cfg->used_int_regs & (1 << i))) {
7185 amd64_mov_reg_membase (code, i, AMD64_RSP, save_area_offset, 8);
7186 save_area_offset += 8;
7190 for (i = 0; i < AMD64_NREG; ++i)
7191 if (AMD64_IS_CALLEE_SAVED_REG (i) && (cfg->used_int_regs & (1 << i)))
7192 pos -= sizeof(mgreg_t);
7195 if (pos == - sizeof(mgreg_t)) {
7196 /* Only one register, so avoid lea */
7197 for (i = AMD64_NREG - 1; i > 0; --i)
7198 if (AMD64_IS_CALLEE_SAVED_REG (i) && (cfg->used_int_regs & (1 << i))) {
7199 amd64_mov_reg_membase (code, i, AMD64_RBP, pos, 8);
7203 amd64_lea_membase (code, AMD64_RSP, AMD64_RBP, pos);
7205 /* Pop registers in reverse order */
7206 for (i = AMD64_NREG - 1; i > 0; --i)
7207 if (AMD64_IS_CALLEE_SAVED_REG (i) && (cfg->used_int_regs & (1 << i))) {
7208 amd64_pop_reg (code, i);
7215 /* Load returned vtypes into registers if needed */
7216 cinfo = cfg->arch.cinfo;
7217 if (cinfo->ret.storage == ArgValuetypeInReg) {
7218 ArgInfo *ainfo = &cinfo->ret;
7219 MonoInst *inst = cfg->ret;
7221 for (quad = 0; quad < 2; quad ++) {
7222 switch (ainfo->pair_storage [quad]) {
7224 amd64_mov_reg_membase (code, ainfo->pair_regs [quad], inst->inst_basereg, inst->inst_offset + (quad * sizeof(mgreg_t)), sizeof(mgreg_t));
7226 case ArgInFloatSSEReg:
7227 amd64_movss_reg_membase (code, ainfo->pair_regs [quad], inst->inst_basereg, inst->inst_offset + (quad * sizeof(mgreg_t)));
7229 case ArgInDoubleSSEReg:
7230 amd64_movsd_reg_membase (code, ainfo->pair_regs [quad], inst->inst_basereg, inst->inst_offset + (quad * sizeof(mgreg_t)));
7235 g_assert_not_reached ();
7240 if (cfg->arch.omit_fp) {
7241 if (cfg->arch.stack_alloc_size)
7242 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, cfg->arch.stack_alloc_size);
7246 async_exc_point (code);
7249 cfg->code_len = code - cfg->native_code;
7251 g_assert (cfg->code_len < cfg->code_size);
7255 mono_arch_emit_exceptions (MonoCompile *cfg)
7257 MonoJumpInfo *patch_info;
7260 MonoClass *exc_classes [16];
7261 guint8 *exc_throw_start [16], *exc_throw_end [16];
7262 guint32 code_size = 0;
7264 /* Compute needed space */
7265 for (patch_info = cfg->patch_info; patch_info; patch_info = patch_info->next) {
7266 if (patch_info->type == MONO_PATCH_INFO_EXC)
7268 if (patch_info->type == MONO_PATCH_INFO_R8)
7269 code_size += 8 + 15; /* sizeof (double) + alignment */
7270 if (patch_info->type == MONO_PATCH_INFO_R4)
7271 code_size += 4 + 15; /* sizeof (float) + alignment */
7272 if (patch_info->type == MONO_PATCH_INFO_GC_CARD_TABLE_ADDR)
7273 code_size += 8 + 7; /*sizeof (void*) + alignment */
7276 #ifdef __native_client_codegen__
7277 /* Give us extra room on Native Client. This could be */
7278 /* more carefully calculated, but bundle alignment makes */
7279 /* it much trickier, so *2 like other places is good. */
7283 while (cfg->code_len + code_size > (cfg->code_size - 16)) {
7284 cfg->code_size *= 2;
7285 cfg->native_code = mono_realloc_native_code (cfg);
7286 cfg->stat_code_reallocs++;
7289 code = cfg->native_code + cfg->code_len;
7291 /* add code to raise exceptions */
7293 for (patch_info = cfg->patch_info; patch_info; patch_info = patch_info->next) {
7294 switch (patch_info->type) {
7295 case MONO_PATCH_INFO_EXC: {
7296 MonoClass *exc_class;
7300 amd64_patch (patch_info->ip.i + cfg->native_code, code);
7302 exc_class = mono_class_from_name (mono_defaults.corlib, "System", patch_info->data.name);
7303 g_assert (exc_class);
7304 throw_ip = patch_info->ip.i;
7306 //x86_breakpoint (code);
7307 /* Find a throw sequence for the same exception class */
7308 for (i = 0; i < nthrows; ++i)
7309 if (exc_classes [i] == exc_class)
7312 amd64_mov_reg_imm (code, AMD64_ARG_REG2, (exc_throw_end [i] - cfg->native_code) - throw_ip);
7313 x86_jump_code (code, exc_throw_start [i]);
7314 patch_info->type = MONO_PATCH_INFO_NONE;
7318 amd64_mov_reg_imm_size (code, AMD64_ARG_REG2, 0xf0f0f0f0, 4);
7322 exc_classes [nthrows] = exc_class;
7323 exc_throw_start [nthrows] = code;
7325 amd64_mov_reg_imm (code, AMD64_ARG_REG1, exc_class->type_token - MONO_TOKEN_TYPE_DEF);
7327 patch_info->type = MONO_PATCH_INFO_NONE;
7329 code = emit_call_body (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD, "mono_arch_throw_corlib_exception");
7331 amd64_mov_reg_imm (buf, AMD64_ARG_REG2, (code - cfg->native_code) - throw_ip);
7336 exc_throw_end [nthrows] = code;
7346 g_assert(code < cfg->native_code + cfg->code_size);
7349 /* Handle relocations with RIP relative addressing */
7350 for (patch_info = cfg->patch_info; patch_info; patch_info = patch_info->next) {
7351 gboolean remove = FALSE;
7352 guint8 *orig_code = code;
7354 switch (patch_info->type) {
7355 case MONO_PATCH_INFO_R8:
7356 case MONO_PATCH_INFO_R4: {
7357 guint8 *pos, *patch_pos;
7360 /* The SSE opcodes require a 16 byte alignment */
7361 #if defined(__default_codegen__)
7362 code = (guint8*)ALIGN_TO (code, 16);
7363 #elif defined(__native_client_codegen__)
7365 /* Pad this out with HLT instructions */
7366 /* or we can get garbage bytes emitted */
7367 /* which will fail validation */
7368 guint8 *aligned_code;
7369 /* extra align to make room for */
7370 /* mov/push below */
7371 int extra_align = patch_info->type == MONO_PATCH_INFO_R8 ? 2 : 1;
7372 aligned_code = (guint8*)ALIGN_TO (code + extra_align, 16);
7373 /* The technique of hiding data in an */
7374 /* instruction has a problem here: we */
7375 /* need the data aligned to a 16-byte */
7376 /* boundary but the instruction cannot */
7377 /* cross the bundle boundary. so only */
7378 /* odd multiples of 16 can be used */
7379 if ((intptr_t)aligned_code % kNaClAlignment == 0) {
7382 while (code < aligned_code) {
7383 *(code++) = 0xf4; /* hlt */
7388 pos = cfg->native_code + patch_info->ip.i;
7389 if (IS_REX (pos [1])) {
7390 patch_pos = pos + 5;
7391 target_pos = code - pos - 9;
7394 patch_pos = pos + 4;
7395 target_pos = code - pos - 8;
7398 if (patch_info->type == MONO_PATCH_INFO_R8) {
7399 #ifdef __native_client_codegen__
7400 /* Hide 64-bit data in a */
7401 /* "mov imm64, r11" instruction. */
7402 /* write it before the start of */
7404 *(code-2) = 0x49; /* prefix */
7405 *(code-1) = 0xbb; /* mov X, %r11 */
7407 *(double*)code = *(double*)patch_info->data.target;
7408 code += sizeof (double);
7410 #ifdef __native_client_codegen__
7411 /* Hide 32-bit data in a */
7412 /* "push imm32" instruction. */
7413 *(code-1) = 0x68; /* push */
7415 *(float*)code = *(float*)patch_info->data.target;
7416 code += sizeof (float);
7419 *(guint32*)(patch_pos) = target_pos;
7424 case MONO_PATCH_INFO_GC_CARD_TABLE_ADDR: {
7427 if (cfg->compile_aot)
7430 /*loading is faster against aligned addresses.*/
7431 code = (guint8*)ALIGN_TO (code, 8);
7432 memset (orig_code, 0, code - orig_code);
7434 pos = cfg->native_code + patch_info->ip.i;
7436 /*alu_op [rex] modr/m imm32 - 7 or 8 bytes */
7437 if (IS_REX (pos [1]))
7438 *(guint32*)(pos + 4) = (guint8*)code - pos - 8;
7440 *(guint32*)(pos + 3) = (guint8*)code - pos - 7;
7442 *(gpointer*)code = (gpointer)patch_info->data.target;
7443 code += sizeof (gpointer);
7453 if (patch_info == cfg->patch_info)
7454 cfg->patch_info = patch_info->next;
7458 for (tmp = cfg->patch_info; tmp->next != patch_info; tmp = tmp->next)
7460 tmp->next = patch_info->next;
7463 g_assert (code < cfg->native_code + cfg->code_size);
7466 cfg->code_len = code - cfg->native_code;
7468 g_assert (cfg->code_len < cfg->code_size);
7472 #endif /* DISABLE_JIT */
7475 mono_arch_instrument_prolog (MonoCompile *cfg, void *func, void *p, gboolean enable_arguments)
7478 CallInfo *cinfo = NULL;
7479 MonoMethodSignature *sig;
7481 int i, n, stack_area = 0;
7483 /* Keep this in sync with mono_arch_get_argument_info */
7485 if (enable_arguments) {
7486 /* Allocate a new area on the stack and save arguments there */
7487 sig = mono_method_signature (cfg->method);
7489 cinfo = get_call_info (cfg->generic_sharing_context, cfg->mempool, sig);
7491 n = sig->param_count + sig->hasthis;
7493 stack_area = ALIGN_TO (n * 8, 16);
7495 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, stack_area);
7497 for (i = 0; i < n; ++i) {
7498 inst = cfg->args [i];
7500 if (inst->opcode == OP_REGVAR)
7501 amd64_mov_membase_reg (code, AMD64_RSP, (i * 8), inst->dreg, 8);
7503 amd64_mov_reg_membase (code, AMD64_R11, inst->inst_basereg, inst->inst_offset, 8);
7504 amd64_mov_membase_reg (code, AMD64_RSP, (i * 8), AMD64_R11, 8);
7509 mono_add_patch_info (cfg, code-cfg->native_code, MONO_PATCH_INFO_METHODCONST, cfg->method);
7510 amd64_set_reg_template (code, AMD64_ARG_REG1);
7511 amd64_mov_reg_reg (code, AMD64_ARG_REG2, AMD64_RSP, 8);
7512 code = emit_call (cfg, code, MONO_PATCH_INFO_ABS, (gpointer)func, TRUE);
7514 if (enable_arguments)
7515 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, stack_area);
7529 mono_arch_instrument_epilog_full (MonoCompile *cfg, void *func, void *p, gboolean enable_arguments, gboolean preserve_argument_registers)
7532 int save_mode = SAVE_NONE;
7533 MonoMethod *method = cfg->method;
7534 MonoType *ret_type = mini_type_get_underlying_type (NULL, mono_method_signature (method)->ret);
7536 switch (ret_type->type) {
7537 case MONO_TYPE_VOID:
7538 /* special case string .ctor icall */
7539 if (strcmp (".ctor", method->name) && method->klass == mono_defaults.string_class)
7540 save_mode = SAVE_EAX;
7542 save_mode = SAVE_NONE;
7546 save_mode = SAVE_EAX;
7550 save_mode = SAVE_XMM;
7552 case MONO_TYPE_GENERICINST:
7553 if (!mono_type_generic_inst_is_valuetype (ret_type)) {
7554 save_mode = SAVE_EAX;
7558 case MONO_TYPE_VALUETYPE:
7559 save_mode = SAVE_STRUCT;
7562 save_mode = SAVE_EAX;
7566 /* Save the result and copy it into the proper argument register */
7567 switch (save_mode) {
7569 amd64_push_reg (code, AMD64_RAX);
7571 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, 8);
7572 if (enable_arguments)
7573 amd64_mov_reg_reg (code, AMD64_ARG_REG2, AMD64_RAX, 8);
7577 if (enable_arguments)
7578 amd64_mov_reg_imm (code, AMD64_ARG_REG2, 0);
7581 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, 8);
7582 amd64_movsd_membase_reg (code, AMD64_RSP, 0, AMD64_XMM0);
7584 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, 8);
7586 * The result is already in the proper argument register so no copying
7593 g_assert_not_reached ();
7596 /* Set %al since this is a varargs call */
7597 if (save_mode == SAVE_XMM)
7598 amd64_mov_reg_imm (code, AMD64_RAX, 1);
7600 amd64_mov_reg_imm (code, AMD64_RAX, 0);
7602 if (preserve_argument_registers) {
7603 amd64_push_reg (code, MONO_AMD64_ARG_REG1);
7604 amd64_push_reg (code, MONO_AMD64_ARG_REG2);
7607 mono_add_patch_info (cfg, code-cfg->native_code, MONO_PATCH_INFO_METHODCONST, method);
7608 amd64_set_reg_template (code, AMD64_ARG_REG1);
7609 code = emit_call (cfg, code, MONO_PATCH_INFO_ABS, (gpointer)func, TRUE);
7611 if (preserve_argument_registers) {
7612 amd64_pop_reg (code, MONO_AMD64_ARG_REG2);
7613 amd64_pop_reg (code, MONO_AMD64_ARG_REG1);
7616 /* Restore result */
7617 switch (save_mode) {
7619 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, 8);
7620 amd64_pop_reg (code, AMD64_RAX);
7626 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, 8);
7627 amd64_movsd_reg_membase (code, AMD64_XMM0, AMD64_RSP, 0);
7628 amd64_alu_reg_imm (code, X86_ADD, AMD64_RSP, 8);
7633 g_assert_not_reached ();
7640 mono_arch_flush_icache (guint8 *code, gint size)
7646 mono_arch_flush_register_windows (void)
7651 mono_arch_is_inst_imm (gint64 imm)
7653 return amd64_is_imm32 (imm);
7657 * Determine whenever the trap whose info is in SIGINFO is caused by
7661 mono_arch_is_int_overflow (void *sigctx, void *info)
7668 mono_arch_sigctx_to_monoctx (sigctx, &ctx);
7670 rip = (guint8*)ctx.rip;
7672 if (IS_REX (rip [0])) {
7673 reg = amd64_rex_b (rip [0]);
7679 if ((rip [0] == 0xf7) && (x86_modrm_mod (rip [1]) == 0x3) && (x86_modrm_reg (rip [1]) == 0x7)) {
7681 reg += x86_modrm_rm (rip [1]);
7721 g_assert_not_reached ();
7733 mono_arch_get_patch_offset (guint8 *code)
7739 * mono_breakpoint_clean_code:
7741 * Copy @size bytes from @code - @offset to the buffer @buf. If the debugger inserted software
7742 * breakpoints in the original code, they are removed in the copy.
7744 * Returns TRUE if no sw breakpoint was present.
7747 mono_breakpoint_clean_code (guint8 *method_start, guint8 *code, int offset, guint8 *buf, int size)
7750 gboolean can_write = TRUE;
7752 * If method_start is non-NULL we need to perform bound checks, since we access memory
7753 * at code - offset we could go before the start of the method and end up in a different
7754 * page of memory that is not mapped or read incorrect data anyway. We zero-fill the bytes
7757 if (!method_start || code - offset >= method_start) {
7758 memcpy (buf, code - offset, size);
7760 int diff = code - method_start;
7761 memset (buf, 0, size);
7762 memcpy (buf + offset - diff, method_start, diff + size - offset);
7765 for (i = 0; i < MONO_BREAKPOINT_ARRAY_SIZE; ++i) {
7766 int idx = mono_breakpoint_info_index [i];
7770 ptr = mono_breakpoint_info [idx].address;
7771 if (ptr >= code && ptr < code + size) {
7772 guint8 saved_byte = mono_breakpoint_info [idx].saved_byte;
7774 /*g_print ("patching %p with 0x%02x (was: 0x%02x)\n", ptr, saved_byte, buf [ptr - code]);*/
7775 buf [ptr - code] = saved_byte;
7781 #if defined(__native_client_codegen__)
7782 /* For membase calls, we want the base register. for Native Client, */
7783 /* all indirect calls have the following sequence with the given sizes: */
7784 /* mov %eXX,%eXX [2-3] */
7785 /* mov disp(%r15,%rXX,scale),%r11d [4-8] */
7786 /* and $0xffffffffffffffe0,%r11d [4] */
7787 /* add %r15,%r11 [3] */
7788 /* callq *%r11 [3] */
7791 /* Determine if code points to a NaCl call-through-register sequence, */
7792 /* (i.e., the last 3 instructions listed above) */
7794 is_nacl_call_reg_sequence(guint8* code)
7796 const char *sequence = "\x41\x83\xe3\xe0" /* and */
7797 "\x4d\x03\xdf" /* add */
7798 "\x41\xff\xd3"; /* call */
7799 return memcmp(code, sequence, 10) == 0;
7802 /* Determine if code points to the first opcode of the mov membase component */
7803 /* of an indirect call sequence (i.e. the first 2 instructions listed above) */
7804 /* (there could be a REX prefix before the opcode but it is ignored) */
7806 is_nacl_indirect_call_membase_sequence(guint8* code)
7808 /* Check for mov opcode, reg-reg addressing mode (mod = 3), */
7809 return code[0] == 0x8b && amd64_modrm_mod(code[1]) == 3 &&
7810 /* and that src reg = dest reg */
7811 amd64_modrm_reg(code[1]) == amd64_modrm_rm(code[1]) &&
7812 /* Check that next inst is mov, uses SIB byte (rm = 4), */
7814 code[3] == 0x8b && amd64_modrm_rm(code[4]) == 4 &&
7815 /* and has dst of r11 and base of r15 */
7816 (amd64_modrm_reg(code[4]) + amd64_rex_r(code[2])) == AMD64_R11 &&
7817 (amd64_sib_base(code[5]) + amd64_rex_b(code[2])) == AMD64_R15;
7819 #endif /* __native_client_codegen__ */
7822 mono_arch_get_this_arg_reg (guint8 *code)
7824 return AMD64_ARG_REG1;
7828 mono_arch_get_this_arg_from_call (mgreg_t *regs, guint8 *code)
7830 return (gpointer)regs [mono_arch_get_this_arg_reg (code)];
7833 #define MAX_ARCH_DELEGATE_PARAMS 10
7836 get_delegate_invoke_impl (gboolean has_target, guint32 param_count, guint32 *code_len)
7838 guint8 *code, *start;
7842 start = code = mono_global_codeman_reserve (64);
7844 /* Replace the this argument with the target */
7845 amd64_mov_reg_reg (code, AMD64_RAX, AMD64_ARG_REG1, 8);
7846 amd64_mov_reg_membase (code, AMD64_ARG_REG1, AMD64_RAX, G_STRUCT_OFFSET (MonoDelegate, target), 8);
7847 amd64_jump_membase (code, AMD64_RAX, G_STRUCT_OFFSET (MonoDelegate, method_ptr));
7849 g_assert ((code - start) < 64);
7851 start = code = mono_global_codeman_reserve (64);
7853 if (param_count == 0) {
7854 amd64_jump_membase (code, AMD64_ARG_REG1, G_STRUCT_OFFSET (MonoDelegate, method_ptr));
7856 /* We have to shift the arguments left */
7857 amd64_mov_reg_reg (code, AMD64_RAX, AMD64_ARG_REG1, 8);
7858 for (i = 0; i < param_count; ++i) {
7861 amd64_mov_reg_reg (code, param_regs [i], param_regs [i + 1], 8);
7863 amd64_mov_reg_membase (code, param_regs [i], AMD64_RSP, 0x28, 8);
7865 amd64_mov_reg_reg (code, param_regs [i], param_regs [i + 1], 8);
7869 amd64_jump_membase (code, AMD64_RAX, G_STRUCT_OFFSET (MonoDelegate, method_ptr));
7871 g_assert ((code - start) < 64);
7874 nacl_global_codeman_validate(&start, 64, &code);
7876 mono_debug_add_delegate_trampoline (start, code - start);
7879 *code_len = code - start;
7882 if (mono_jit_map_is_enabled ()) {
7885 buff = (char*)"delegate_invoke_has_target";
7887 buff = g_strdup_printf ("delegate_invoke_no_target_%d", param_count);
7888 mono_emit_jit_tramp (start, code - start, buff);
7897 * mono_arch_get_delegate_invoke_impls:
7899 * Return a list of MonoTrampInfo structures for the delegate invoke impl
7903 mono_arch_get_delegate_invoke_impls (void)
7910 code = get_delegate_invoke_impl (TRUE, 0, &code_len);
7911 res = g_slist_prepend (res, mono_tramp_info_create (g_strdup ("delegate_invoke_impl_has_target"), code, code_len, NULL, NULL));
7913 for (i = 0; i < MAX_ARCH_DELEGATE_PARAMS; ++i) {
7914 code = get_delegate_invoke_impl (FALSE, i, &code_len);
7915 res = g_slist_prepend (res, mono_tramp_info_create (g_strdup_printf ("delegate_invoke_impl_target_%d", i), code, code_len, NULL, NULL));
7922 mono_arch_get_delegate_invoke_impl (MonoMethodSignature *sig, gboolean has_target)
7924 guint8 *code, *start;
7927 if (sig->param_count > MAX_ARCH_DELEGATE_PARAMS)
7930 /* FIXME: Support more cases */
7931 if (MONO_TYPE_ISSTRUCT (sig->ret))
7935 static guint8* cached = NULL;
7941 start = mono_aot_get_trampoline ("delegate_invoke_impl_has_target");
7943 start = get_delegate_invoke_impl (TRUE, 0, NULL);
7945 mono_memory_barrier ();
7949 static guint8* cache [MAX_ARCH_DELEGATE_PARAMS + 1] = {NULL};
7950 for (i = 0; i < sig->param_count; ++i)
7951 if (!mono_is_regsize_var (sig->params [i]))
7953 if (sig->param_count > 4)
7956 code = cache [sig->param_count];
7960 if (mono_aot_only) {
7961 char *name = g_strdup_printf ("delegate_invoke_impl_target_%d", sig->param_count);
7962 start = mono_aot_get_trampoline (name);
7965 start = get_delegate_invoke_impl (FALSE, sig->param_count, NULL);
7968 mono_memory_barrier ();
7970 cache [sig->param_count] = start;
7977 * Support for fast access to the thread-local lmf structure using the GS
7978 * segment register on NPTL + kernel 2.6.x.
7981 static gboolean tls_offset_inited = FALSE;
7984 mono_arch_setup_jit_tls_data (MonoJitTlsData *tls)
7986 if (!tls_offset_inited) {
7989 * We need to init this multiple times, since when we are first called, the key might not
7990 * be initialized yet.
7992 appdomain_tls_offset = mono_domain_get_tls_key ();
7993 lmf_tls_offset = mono_get_jit_tls_key ();
7994 lmf_addr_tls_offset = mono_get_jit_tls_key ();
7996 /* Only 64 tls entries can be accessed using inline code */
7997 if (appdomain_tls_offset >= 64)
7998 appdomain_tls_offset = -1;
7999 if (lmf_tls_offset >= 64)
8000 lmf_tls_offset = -1;
8002 tls_offset_inited = TRUE;
8004 optimize_for_xen = access ("/proc/xen", F_OK) == 0;
8006 appdomain_tls_offset = mono_domain_get_tls_offset ();
8007 lmf_tls_offset = mono_get_lmf_tls_offset ();
8008 lmf_addr_tls_offset = mono_get_lmf_addr_tls_offset ();
8014 mono_arch_free_jit_tls_data (MonoJitTlsData *tls)
8018 #ifdef MONO_ARCH_HAVE_IMT
8020 #if defined(__default_codegen__)
8021 #define CMP_SIZE (6 + 1)
8022 #define CMP_REG_REG_SIZE (4 + 1)
8023 #define BR_SMALL_SIZE 2
8024 #define BR_LARGE_SIZE 6
8025 #define MOV_REG_IMM_SIZE 10
8026 #define MOV_REG_IMM_32BIT_SIZE 6
8027 #define JUMP_REG_SIZE (2 + 1)
8028 #elif defined(__native_client_codegen__)
8029 /* NaCl N-byte instructions can be padded up to N-1 bytes */
8030 #define CMP_SIZE ((6 + 1) * 2 - 1)
8031 #define CMP_REG_REG_SIZE ((4 + 1) * 2 - 1)
8032 #define BR_SMALL_SIZE (2 * 2 - 1)
8033 #define BR_LARGE_SIZE (6 * 2 - 1)
8034 #define MOV_REG_IMM_SIZE (10 * 2 - 1)
8035 #define MOV_REG_IMM_32BIT_SIZE (6 * 2 - 1)
8036 /* Jump reg for NaCl adds a mask (+4) and add (+3) */
8037 #define JUMP_REG_SIZE ((2 + 1 + 4 + 3) * 2 - 1)
8038 /* Jump membase's size is large and unpredictable */
8039 /* in native client, just pad it out a whole bundle. */
8040 #define JUMP_MEMBASE_SIZE (kNaClAlignment)
8044 imt_branch_distance (MonoIMTCheckItem **imt_entries, int start, int target)
8046 int i, distance = 0;
8047 for (i = start; i < target; ++i)
8048 distance += imt_entries [i]->chunk_size;
8053 * LOCKING: called with the domain lock held
8056 mono_arch_build_imt_thunk (MonoVTable *vtable, MonoDomain *domain, MonoIMTCheckItem **imt_entries, int count,
8057 gpointer fail_tramp)
8061 guint8 *code, *start;
8062 gboolean vtable_is_32bit = ((gsize)(vtable) == (gsize)(int)(gsize)(vtable));
8064 for (i = 0; i < count; ++i) {
8065 MonoIMTCheckItem *item = imt_entries [i];
8066 if (item->is_equals) {
8067 if (item->check_target_idx) {
8068 if (!item->compare_done) {
8069 if (amd64_is_imm32 (item->key))
8070 item->chunk_size += CMP_SIZE;
8072 item->chunk_size += MOV_REG_IMM_SIZE + CMP_REG_REG_SIZE;
8074 if (item->has_target_code) {
8075 item->chunk_size += MOV_REG_IMM_SIZE;
8077 if (vtable_is_32bit)
8078 item->chunk_size += MOV_REG_IMM_32BIT_SIZE;
8080 item->chunk_size += MOV_REG_IMM_SIZE;
8081 #ifdef __native_client_codegen__
8082 item->chunk_size += JUMP_MEMBASE_SIZE;
8085 item->chunk_size += BR_SMALL_SIZE + JUMP_REG_SIZE;
8088 item->chunk_size += MOV_REG_IMM_SIZE * 3 + CMP_REG_REG_SIZE +
8089 BR_SMALL_SIZE + JUMP_REG_SIZE * 2;
8091 if (vtable_is_32bit)
8092 item->chunk_size += MOV_REG_IMM_32BIT_SIZE;
8094 item->chunk_size += MOV_REG_IMM_SIZE;
8095 item->chunk_size += JUMP_REG_SIZE;
8096 /* with assert below:
8097 * item->chunk_size += CMP_SIZE + BR_SMALL_SIZE + 1;
8099 #ifdef __native_client_codegen__
8100 item->chunk_size += JUMP_MEMBASE_SIZE;
8105 if (amd64_is_imm32 (item->key))
8106 item->chunk_size += CMP_SIZE;
8108 item->chunk_size += MOV_REG_IMM_SIZE + CMP_REG_REG_SIZE;
8109 item->chunk_size += BR_LARGE_SIZE;
8110 imt_entries [item->check_target_idx]->compare_done = TRUE;
8112 size += item->chunk_size;
8114 #if defined(__native_client__) && defined(__native_client_codegen__)
8115 /* In Native Client, we don't re-use thunks, allocate from the */
8116 /* normal code manager paths. */
8117 code = mono_domain_code_reserve (domain, size);
8120 code = mono_method_alloc_generic_virtual_thunk (domain, size);
8122 code = mono_domain_code_reserve (domain, size);
8125 for (i = 0; i < count; ++i) {
8126 MonoIMTCheckItem *item = imt_entries [i];
8127 item->code_target = code;
8128 if (item->is_equals) {
8129 gboolean fail_case = !item->check_target_idx && fail_tramp;
8131 if (item->check_target_idx || fail_case) {
8132 if (!item->compare_done || fail_case) {
8133 if (amd64_is_imm32 (item->key))
8134 amd64_alu_reg_imm (code, X86_CMP, MONO_ARCH_IMT_REG, (guint32)(gssize)item->key);
8136 amd64_mov_reg_imm (code, MONO_ARCH_IMT_SCRATCH_REG, item->key);
8137 amd64_alu_reg_reg (code, X86_CMP, MONO_ARCH_IMT_REG, MONO_ARCH_IMT_SCRATCH_REG);
8140 item->jmp_code = code;
8141 amd64_branch8 (code, X86_CC_NE, 0, FALSE);
8142 if (item->has_target_code) {
8143 amd64_mov_reg_imm (code, MONO_ARCH_IMT_SCRATCH_REG, item->value.target_code);
8144 amd64_jump_reg (code, MONO_ARCH_IMT_SCRATCH_REG);
8146 amd64_mov_reg_imm (code, MONO_ARCH_IMT_SCRATCH_REG, & (vtable->vtable [item->value.vtable_slot]));
8147 amd64_jump_membase (code, MONO_ARCH_IMT_SCRATCH_REG, 0);
8151 amd64_patch (item->jmp_code, code);
8152 amd64_mov_reg_imm (code, MONO_ARCH_IMT_SCRATCH_REG, fail_tramp);
8153 amd64_jump_reg (code, MONO_ARCH_IMT_SCRATCH_REG);
8154 item->jmp_code = NULL;
8157 /* enable the commented code to assert on wrong method */
8159 if (amd64_is_imm32 (item->key))
8160 amd64_alu_reg_imm (code, X86_CMP, MONO_ARCH_IMT_REG, (guint32)(gssize)item->key);
8162 amd64_mov_reg_imm (code, MONO_ARCH_IMT_SCRATCH_REG, item->key);
8163 amd64_alu_reg_reg (code, X86_CMP, MONO_ARCH_IMT_REG, MONO_ARCH_IMT_SCRATCH_REG);
8165 item->jmp_code = code;
8166 amd64_branch8 (code, X86_CC_NE, 0, FALSE);
8167 /* See the comment below about R10 */
8168 amd64_mov_reg_imm (code, MONO_ARCH_IMT_SCRATCH_REG, & (vtable->vtable [item->value.vtable_slot]));
8169 amd64_jump_membase (code, MONO_ARCH_IMT_SCRATCH_REG, 0);
8170 amd64_patch (item->jmp_code, code);
8171 amd64_breakpoint (code);
8172 item->jmp_code = NULL;
8174 /* We're using R10 (MONO_ARCH_IMT_SCRATCH_REG) here because R11 (MONO_ARCH_IMT_REG)
8175 needs to be preserved. R10 needs
8176 to be preserved for calls which
8177 require a runtime generic context,
8178 but interface calls don't. */
8179 amd64_mov_reg_imm (code, MONO_ARCH_IMT_SCRATCH_REG, & (vtable->vtable [item->value.vtable_slot]));
8180 amd64_jump_membase (code, MONO_ARCH_IMT_SCRATCH_REG, 0);
8184 if (amd64_is_imm32 (item->key))
8185 amd64_alu_reg_imm (code, X86_CMP, MONO_ARCH_IMT_REG, (guint32)(gssize)item->key);
8187 amd64_mov_reg_imm (code, MONO_ARCH_IMT_SCRATCH_REG, item->key);
8188 amd64_alu_reg_reg (code, X86_CMP, MONO_ARCH_IMT_REG, MONO_ARCH_IMT_SCRATCH_REG);
8190 item->jmp_code = code;
8191 if (x86_is_imm8 (imt_branch_distance (imt_entries, i, item->check_target_idx)))
8192 x86_branch8 (code, X86_CC_GE, 0, FALSE);
8194 x86_branch32 (code, X86_CC_GE, 0, FALSE);
8196 g_assert (code - item->code_target <= item->chunk_size);
8198 /* patch the branches to get to the target items */
8199 for (i = 0; i < count; ++i) {
8200 MonoIMTCheckItem *item = imt_entries [i];
8201 if (item->jmp_code) {
8202 if (item->check_target_idx) {
8203 amd64_patch (item->jmp_code, imt_entries [item->check_target_idx]->code_target);
8209 mono_stats.imt_thunks_size += code - start;
8210 g_assert (code - start <= size);
8212 nacl_domain_code_validate(domain, &start, size, &code);
8218 mono_arch_find_imt_method (mgreg_t *regs, guint8 *code)
8220 return (MonoMethod*)regs [MONO_ARCH_IMT_REG];
8225 mono_arch_find_static_call_vtable (mgreg_t *regs, guint8 *code)
8227 return (MonoVTable*) regs [MONO_ARCH_RGCTX_REG];
8231 mono_arch_get_cie_program (void)
8235 mono_add_unwind_op_def_cfa (l, (guint8*)NULL, (guint8*)NULL, AMD64_RSP, 8);
8236 mono_add_unwind_op_offset (l, (guint8*)NULL, (guint8*)NULL, AMD64_RIP, -8);
8242 mono_arch_emit_inst_for_method (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature *fsig, MonoInst **args)
8244 MonoInst *ins = NULL;
8247 if (cmethod->klass == mono_defaults.math_class) {
8248 if (strcmp (cmethod->name, "Sin") == 0) {
8250 } else if (strcmp (cmethod->name, "Cos") == 0) {
8252 } else if (strcmp (cmethod->name, "Sqrt") == 0) {
8254 } else if (strcmp (cmethod->name, "Abs") == 0 && fsig->params [0]->type == MONO_TYPE_R8) {
8259 MONO_INST_NEW (cfg, ins, opcode);
8260 ins->type = STACK_R8;
8261 ins->dreg = mono_alloc_freg (cfg);
8262 ins->sreg1 = args [0]->dreg;
8263 MONO_ADD_INS (cfg->cbb, ins);
8267 if (cfg->opt & MONO_OPT_CMOV) {
8268 if (strcmp (cmethod->name, "Min") == 0) {
8269 if (fsig->params [0]->type == MONO_TYPE_I4)
8271 if (fsig->params [0]->type == MONO_TYPE_U4)
8272 opcode = OP_IMIN_UN;
8273 else if (fsig->params [0]->type == MONO_TYPE_I8)
8275 else if (fsig->params [0]->type == MONO_TYPE_U8)
8276 opcode = OP_LMIN_UN;
8277 } else if (strcmp (cmethod->name, "Max") == 0) {
8278 if (fsig->params [0]->type == MONO_TYPE_I4)
8280 if (fsig->params [0]->type == MONO_TYPE_U4)
8281 opcode = OP_IMAX_UN;
8282 else if (fsig->params [0]->type == MONO_TYPE_I8)
8284 else if (fsig->params [0]->type == MONO_TYPE_U8)
8285 opcode = OP_LMAX_UN;
8290 MONO_INST_NEW (cfg, ins, opcode);
8291 ins->type = fsig->params [0]->type == MONO_TYPE_I4 ? STACK_I4 : STACK_I8;
8292 ins->dreg = mono_alloc_ireg (cfg);
8293 ins->sreg1 = args [0]->dreg;
8294 ins->sreg2 = args [1]->dreg;
8295 MONO_ADD_INS (cfg->cbb, ins);
8299 /* OP_FREM is not IEEE compatible */
8300 else if (strcmp (cmethod->name, "IEEERemainder") == 0) {
8301 MONO_INST_NEW (cfg, ins, OP_FREM);
8302 ins->inst_i0 = args [0];
8303 ins->inst_i1 = args [1];
8309 * Can't implement CompareExchange methods this way since they have
8317 mono_arch_print_tree (MonoInst *tree, int arity)
8322 MonoInst* mono_arch_get_domain_intrinsic (MonoCompile* cfg)
8326 if (appdomain_tls_offset == -1)
8329 MONO_INST_NEW (cfg, ins, OP_TLS_GET);
8330 ins->inst_offset = appdomain_tls_offset;
8334 #define _CTX_REG(ctx,fld,i) ((&ctx->fld)[i])
8337 mono_arch_context_get_int_reg (MonoContext *ctx, int reg)
8340 case AMD64_RCX: return ctx->rcx;
8341 case AMD64_RDX: return ctx->rdx;
8342 case AMD64_RBX: return ctx->rbx;
8343 case AMD64_RBP: return ctx->rbp;
8344 case AMD64_RSP: return ctx->rsp;
8347 return _CTX_REG (ctx, rax, reg);
8349 return _CTX_REG (ctx, r12, reg - 12);
8351 g_assert_not_reached ();
8356 mono_arch_context_set_int_reg (MonoContext *ctx, int reg, mgreg_t val)
8376 _CTX_REG (ctx, rax, reg) = val;
8378 _CTX_REG (ctx, r12, reg - 12) = val;
8380 g_assert_not_reached ();
8384 /*MONO_ARCH_HAVE_HANDLER_BLOCK_GUARD*/
8386 mono_arch_install_handler_block_guard (MonoJitInfo *ji, MonoJitExceptionInfo *clause, MonoContext *ctx, gpointer new_value)
8389 gpointer *sp, old_value;
8391 const unsigned char *handler;
8393 /*Decode the first instruction to figure out where did we store the spvar*/
8394 /*Our jit MUST generate the following:
8397 Which is encoded as: REX.W 0x89 mod_rm
8398 mod_rm (rsp, rbp, imm) which can be: (imm will never be zero)
8399 mod (reg + imm8): 01 reg(rsp): 100 rm(rbp): 101 -> 01100101 (0x65)
8400 mod (reg + imm32): 10 reg(rsp): 100 rm(rbp): 101 -> 10100101 (0xA5)
8402 FIXME can we generate frameless methods on this case?
8405 handler = clause->handler_start;
8408 if (*handler != 0x48)
8413 if (*handler != 0x89)
8417 if (*handler == 0x65)
8418 offset = *(signed char*)(handler + 1);
8419 else if (*handler == 0xA5)
8420 offset = *(int*)(handler + 1);
8425 bp = MONO_CONTEXT_GET_BP (ctx);
8426 sp = *(gpointer*)(bp + offset);
8429 if (old_value < ji->code_start || (char*)old_value > ((char*)ji->code_start + ji->code_size))
8438 * mono_arch_emit_load_aotconst:
8440 * Emit code to load the contents of the GOT slot identified by TRAMP_TYPE and
8441 * TARGET from the mscorlib GOT in full-aot code.
8442 * On AMD64, the result is placed into R11.
8445 mono_arch_emit_load_aotconst (guint8 *start, guint8 *code, MonoJumpInfo **ji, int tramp_type, gconstpointer target)
8447 *ji = mono_patch_info_list_prepend (*ji, code - start, tramp_type, target);
8448 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RIP, 0, 8);
8454 * mono_arch_get_trampolines:
8456 * Return a list of MonoTrampInfo structures describing arch specific trampolines
8460 mono_arch_get_trampolines (gboolean aot)
8462 return mono_amd64_get_exception_trampolines (aot);
8465 /* Soft Debug support */
8466 #ifdef MONO_ARCH_SOFT_DEBUG_SUPPORTED
8469 * mono_arch_set_breakpoint:
8471 * Set a breakpoint at the native code corresponding to JI at NATIVE_OFFSET.
8472 * The location should contain code emitted by OP_SEQ_POINT.
8475 mono_arch_set_breakpoint (MonoJitInfo *ji, guint8 *ip)
8478 guint8 *orig_code = code;
8481 * In production, we will use int3 (has to fix the size in the md
8482 * file). But that could confuse gdb, so during development, we emit a SIGSEGV
8485 g_assert (code [0] == 0x90);
8486 if (breakpoint_size == 8) {
8487 amd64_mov_reg_mem (code, AMD64_R11, (guint64)bp_trigger_page, 4);
8489 amd64_mov_reg_imm_size (code, AMD64_R11, (guint64)bp_trigger_page, 8);
8490 amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, 0, 4);
8493 g_assert (code - orig_code == breakpoint_size);
8497 * mono_arch_clear_breakpoint:
8499 * Clear the breakpoint at IP.
8502 mono_arch_clear_breakpoint (MonoJitInfo *ji, guint8 *ip)
8507 for (i = 0; i < breakpoint_size; ++i)
8512 mono_arch_is_breakpoint_event (void *info, void *sigctx)
8515 EXCEPTION_RECORD* einfo = (EXCEPTION_RECORD*)info;
8518 siginfo_t* sinfo = (siginfo_t*) info;
8519 /* Sometimes the address is off by 4 */
8520 if (sinfo->si_addr >= bp_trigger_page && (guint8*)sinfo->si_addr <= (guint8*)bp_trigger_page + 128)
8528 * mono_arch_get_ip_for_breakpoint:
8530 * Convert the ip in CTX to the address where a breakpoint was placed.
8533 mono_arch_get_ip_for_breakpoint (MonoJitInfo *ji, MonoContext *ctx)
8535 guint8 *ip = MONO_CONTEXT_GET_IP (ctx);
8537 /* ip points to the instruction causing the fault */
8538 ip -= (breakpoint_size - breakpoint_fault_size);
8544 * mono_arch_skip_breakpoint:
8546 * Modify CTX so the ip is placed after the breakpoint instruction, so when
8547 * we resume, the instruction is not executed again.
8550 mono_arch_skip_breakpoint (MonoContext *ctx)
8552 MONO_CONTEXT_SET_IP (ctx, (guint8*)MONO_CONTEXT_GET_IP (ctx) + breakpoint_fault_size);
8556 * mono_arch_start_single_stepping:
8558 * Start single stepping.
8561 mono_arch_start_single_stepping (void)
8563 mono_mprotect (ss_trigger_page, mono_pagesize (), 0);
8567 * mono_arch_stop_single_stepping:
8569 * Stop single stepping.
8572 mono_arch_stop_single_stepping (void)
8574 mono_mprotect (ss_trigger_page, mono_pagesize (), MONO_MMAP_READ);
8578 * mono_arch_is_single_step_event:
8580 * Return whenever the machine state in SIGCTX corresponds to a single
8584 mono_arch_is_single_step_event (void *info, void *sigctx)
8587 EXCEPTION_RECORD* einfo = (EXCEPTION_RECORD*)info;
8590 siginfo_t* sinfo = (siginfo_t*) info;
8591 /* Sometimes the address is off by 4 */
8592 if (sinfo->si_addr >= ss_trigger_page && (guint8*)sinfo->si_addr <= (guint8*)ss_trigger_page + 128)
8600 * mono_arch_get_ip_for_single_step:
8602 * Convert the ip in CTX to the address stored in seq_points.
8605 mono_arch_get_ip_for_single_step (MonoJitInfo *ji, MonoContext *ctx)
8607 guint8 *ip = MONO_CONTEXT_GET_IP (ctx);
8609 ip += single_step_fault_size;
8615 * mono_arch_skip_single_step:
8617 * Modify CTX so the ip is placed after the single step trigger instruction,
8618 * we resume, the instruction is not executed again.
8621 mono_arch_skip_single_step (MonoContext *ctx)
8623 MONO_CONTEXT_SET_IP (ctx, (guint8*)MONO_CONTEXT_GET_IP (ctx) + single_step_fault_size);
8627 * mono_arch_create_seq_point_info:
8629 * Return a pointer to a data structure which is used by the sequence
8630 * point implementation in AOTed code.
8633 mono_arch_get_seq_point_info (MonoDomain *domain, guint8 *code)