2 * mini-ia64.c: IA64 backend for the Mono code generator
5 * Zoltan Varga (vargaz@gmail.com)
7 * (C) 2003 Ximian, Inc.
15 #ifdef __INTEL_COMPILER
16 #include <ia64intrin.h>
19 #include <mono/metadata/appdomain.h>
20 #include <mono/metadata/debug-helpers.h>
21 #include <mono/metadata/threads.h>
22 #include <mono/metadata/profiler-private.h>
23 #include <mono/utils/mono-math.h>
26 #include "mini-ia64.h"
28 #include "jit-icalls.h"
31 #define ALIGN_TO(val,align) ((((guint64)val) + ((align) - 1)) & ~((align) - 1))
33 #define IS_IMM32(val) ((((guint64)val) >> 32) == 0)
36 * IA64 register usage:
37 * - local registers are used for global register allocation
38 * - r8..r11, r14..r30 is used for local register allocation
39 * - r31 is a scratch register used within opcode implementations
40 * - FIXME: Use out registers as well
41 * - the first three locals are used for saving ar.pfst, b0, and sp
42 * - compare instructions allways set p6 and p7
46 * There are a lot of places where generated code is disassembled/patched.
47 * The automatic bundling of instructions done by the code generation macros
48 * could complicate things, so it is best to call
49 * ia64_codegen_set_one_ins_per_bundle () at those places.
52 #define ARGS_OFFSET 16
54 #define GP_SCRATCH_REG 31
55 #define GP_SCRATCH_REG2 30
56 #define FP_SCRATCH_REG 32
57 #define FP_SCRATCH_REG2 33
59 #define LOOP_ALIGNMENT 8
60 #define bb_is_loop_start(bb) ((bb)->loop_body_start && (bb)->nesting)
62 static const char* gregs [] = {
63 "r0", "r1", "r2", "r3", "r4", "r5", "r6", "r7", "r8", "r9",
64 "r10", "r11", "r12", "r13", "r14", "r15", "r16", "r17", "r18", "r19",
65 "r20", "r21", "r22", "r23", "r24", "r25", "r26", "r27", "r28", "r29",
66 "r30", "r31", "r32", "r33", "r34", "r35", "r36", "r37", "r38", "r39",
67 "r40", "r41", "r42", "r43", "r44", "r45", "r46", "r47", "r48", "r49",
68 "r50", "r51", "r52", "r53", "r54", "r55", "r56", "r57", "r58", "r59",
69 "r60", "r61", "r62", "r63", "r64", "r65", "r66", "r67", "r68", "r69",
70 "r70", "r71", "r72", "r73", "r74", "r75", "r76", "r77", "r78", "r79",
71 "r80", "r81", "r82", "r83", "r84", "r85", "r86", "r87", "r88", "r89",
72 "r90", "r91", "r92", "r93", "r94", "r95", "r96", "r97", "r98", "r99",
73 "r100", "r101", "r102", "r103", "r104", "r105", "r106", "r107", "r108", "r109",
74 "r110", "r111", "r112", "r113", "r114", "r115", "r116", "r117", "r118", "r119",
75 "r120", "r121", "r122", "r123", "r124", "r125", "r126", "r127"
79 mono_arch_regname (int reg)
87 static const char* fregs [] = {
88 "f0", "f1", "f2", "f3", "f4", "f5", "f6", "f7", "f8", "f9",
89 "f10", "f11", "f12", "f13", "f14", "f15", "f16", "f17", "f18", "f19",
90 "f20", "f21", "f22", "f23", "f24", "f25", "f26", "f27", "f28", "f29",
91 "f30", "f31", "f32", "f33", "f34", "f35", "f36", "f37", "f38", "f39",
92 "f40", "f41", "f42", "f43", "f44", "f45", "f46", "f47", "f48", "f49",
93 "f50", "f51", "f52", "f53", "f54", "f55", "f56", "f57", "f58", "f59",
94 "f60", "f61", "f62", "f63", "f64", "f65", "f66", "f67", "f68", "f69",
95 "f70", "f71", "f72", "f73", "f74", "f75", "f76", "f77", "f78", "f79",
96 "f80", "f81", "f82", "f83", "f84", "f85", "f86", "f87", "f88", "f89",
97 "f90", "f91", "f92", "f93", "f94", "f95", "f96", "f97", "f98", "f99",
98 "f100", "f101", "f102", "f103", "f104", "f105", "f106", "f107", "f108", "f109",
99 "f110", "f111", "f112", "f113", "f114", "f115", "f116", "f117", "f118", "f119",
100 "f120", "f121", "f122", "f123", "f124", "f125", "f126", "f127"
104 mono_arch_fregname (int reg)
112 G_GNUC_UNUSED static void
117 G_GNUC_UNUSED static gboolean
120 static int count = 0;
123 if (count == atoi (getenv ("COUNT"))) {
127 if (count > atoi (getenv ("COUNT"))) {
135 debug_ins_sched (void)
138 return debug_count ();
148 return debug_count ();
155 ia64_patch (unsigned char* code, gpointer target);
162 ArgValuetypeAddrInIReg,
180 /* Only if storage == ArgAggregate */
190 gboolean need_stack_align;
191 gboolean vtype_retaddr;
192 /* The index of the vret arg in the argument list */
199 #define DEBUG(a) if (cfg->verbose_level > 1) a
204 add_general (guint32 *gr, guint32 *stack_size, ArgInfo *ainfo)
206 ainfo->offset = *stack_size;
208 if (*gr >= PARAM_REGS) {
209 ainfo->storage = ArgOnStack;
210 (*stack_size) += sizeof (gpointer);
213 ainfo->storage = ArgInIReg;
219 #define FLOAT_PARAM_REGS 8
222 add_float (guint32 *gr, guint32 *fr, guint32 *stack_size, ArgInfo *ainfo, gboolean is_double)
224 ainfo->offset = *stack_size;
226 if (*gr >= PARAM_REGS) {
227 ainfo->storage = ArgOnStack;
228 (*stack_size) += sizeof (gpointer);
231 ainfo->storage = is_double ? ArgInFloatReg : ArgInFloatRegR4;
232 ainfo->reg = 8 + *fr;
239 add_valuetype (MonoGenericSharingContext *gsctx, MonoMethodSignature *sig, ArgInfo *ainfo, MonoType *type,
241 guint32 *gr, guint32 *fr, guint32 *stack_size)
245 MonoMarshalType *info;
246 gboolean is_hfa = TRUE;
247 guint32 hfa_type = 0;
249 klass = mono_class_from_mono_type (type);
250 if (type->type == MONO_TYPE_TYPEDBYREF)
251 size = 3 * sizeof (gpointer);
252 else if (sig->pinvoke)
253 size = mono_type_native_stack_size (&klass->byval_arg, NULL);
255 size = mini_type_stack_size (gsctx, &klass->byval_arg, NULL);
257 if (!sig->pinvoke || (size == 0)) {
258 /* Allways pass in memory */
259 ainfo->offset = *stack_size;
260 *stack_size += ALIGN_TO (size, 8);
261 ainfo->storage = ArgOnStack;
266 /* Determine whenever it is a HFA (Homogeneous Floating Point Aggregate) */
267 info = mono_marshal_load_type_info (klass);
269 for (i = 0; i < info->num_fields; ++i) {
270 guint32 ftype = info->fields [i].field->type->type;
271 if (!(info->fields [i].field->type->byref) &&
272 ((ftype == MONO_TYPE_R4) || (ftype == MONO_TYPE_R8))) {
275 else if (hfa_type != ftype)
284 ainfo->storage = ArgAggregate;
285 ainfo->atype = AggregateNormal;
288 ainfo->atype = hfa_type == MONO_TYPE_R4 ? AggregateSingleHFA : AggregateDoubleHFA;
290 if (info->num_fields <= 8) {
292 ainfo->nregs = info->num_fields;
293 ainfo->nslots = ainfo->nregs;
299 if ((*fr) + info->num_fields > 8)
302 ainfo->reg = 8 + (*fr);
303 ainfo->nregs = info->num_fields;
304 ainfo->nslots = ainfo->nregs;
305 (*fr) += info->num_fields;
306 if (ainfo->atype == AggregateSingleHFA) {
308 * FIXME: Have to keep track of the parameter slot number, which is
309 * not the same as *gr.
311 (*gr) += ALIGN_TO (info->num_fields, 2) / 2;
313 (*gr) += info->num_fields;
319 /* This also handles returning of TypedByRef used by some icalls */
322 ainfo->reg = IA64_R8;
323 ainfo->nregs = (size + 7) / 8;
324 ainfo->nslots = ainfo->nregs;
331 ainfo->offset = *stack_size;
332 ainfo->nslots = (size + 7) / 8;
334 if (((*gr) + ainfo->nslots) <= 8) {
335 /* Fits entirely in registers */
336 ainfo->nregs = ainfo->nslots;
337 (*gr) += ainfo->nregs;
341 ainfo->nregs = 8 - (*gr);
343 (*stack_size) += (ainfo->nslots - ainfo->nregs) * 8;
349 * Obtain information about a call according to the calling convention.
350 * For IA64, see the "Itanium Software Conventions and Runtime Architecture
351 * Gude" document for more information.
354 get_call_info (MonoCompile *cfg, MonoMemPool *mp, MonoMethodSignature *sig, gboolean is_pinvoke)
356 guint32 i, gr, fr, pstart;
358 int n = sig->hasthis + sig->param_count;
359 guint32 stack_size = 0;
361 MonoGenericSharingContext *gsctx = cfg ? cfg->generic_sharing_context : NULL;
364 cinfo = mono_mempool_alloc0 (mp, sizeof (CallInfo) + (sizeof (ArgInfo) * n));
366 cinfo = g_malloc0 (sizeof (CallInfo) + (sizeof (ArgInfo) * n));
373 ret_type = mono_type_get_underlying_type (sig->ret);
374 ret_type = mini_get_basic_type_from_generic (gsctx, ret_type);
375 switch (ret_type->type) {
376 case MONO_TYPE_BOOLEAN:
387 case MONO_TYPE_FNPTR:
388 case MONO_TYPE_CLASS:
389 case MONO_TYPE_OBJECT:
390 case MONO_TYPE_SZARRAY:
391 case MONO_TYPE_ARRAY:
392 case MONO_TYPE_STRING:
393 cinfo->ret.storage = ArgInIReg;
394 cinfo->ret.reg = IA64_R8;
398 cinfo->ret.storage = ArgInIReg;
399 cinfo->ret.reg = IA64_R8;
403 cinfo->ret.storage = ArgInFloatReg;
406 case MONO_TYPE_GENERICINST:
407 if (!mono_type_generic_inst_is_valuetype (ret_type)) {
408 cinfo->ret.storage = ArgInIReg;
409 cinfo->ret.reg = IA64_R8;
413 case MONO_TYPE_VALUETYPE:
414 case MONO_TYPE_TYPEDBYREF: {
415 guint32 tmp_gr = 0, tmp_fr = 0, tmp_stacksize = 0;
417 if (sig->ret->byref) {
418 /* This seems to happen with ldfld wrappers */
419 cinfo->ret.storage = ArgInIReg;
421 add_valuetype (gsctx, sig, &cinfo->ret, sig->ret, TRUE, &tmp_gr, &tmp_fr, &tmp_stacksize);
422 if (cinfo->ret.storage == ArgOnStack) {
423 /* The caller passes the address where the value is stored */
424 cinfo->vtype_retaddr = TRUE;
430 cinfo->ret.storage = ArgNone;
433 g_error ("Can't handle as return value 0x%x", sig->ret->type);
439 * To simplify get_this_arg_reg () and LLVM integration, emit the vret arg after
440 * the first argument, allowing 'this' to be always passed in the first arg reg.
441 * Also do this if the first argument is a reference type, since virtual calls
442 * are sometimes made using calli without sig->hasthis set, like in the delegate
445 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]))))) {
447 add_general (&gr, &stack_size, cinfo->args + 0);
449 add_general (&gr, &stack_size, &cinfo->args [sig->hasthis + 0]);
452 add_general (&gr, &stack_size, &cinfo->ret);
453 if (cinfo->ret.storage == ArgInIReg)
454 cinfo->ret.storage = ArgValuetypeAddrInIReg;
455 cinfo->vret_arg_index = 1;
459 add_general (&gr, &stack_size, cinfo->args + 0);
461 if (cinfo->vtype_retaddr) {
462 add_general (&gr, &stack_size, &cinfo->ret);
463 if (cinfo->ret.storage == ArgInIReg)
464 cinfo->ret.storage = ArgValuetypeAddrInIReg;
468 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG) && (n == 0)) {
470 fr = FLOAT_PARAM_REGS;
472 /* Emit the signature cookie just before the implicit arguments */
473 add_general (&gr, &stack_size, &cinfo->sig_cookie);
476 for (i = pstart; i < sig->param_count; ++i) {
477 ArgInfo *ainfo = &cinfo->args [sig->hasthis + i];
480 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG) && (i == sig->sentinelpos)) {
481 /* We allways pass the sig cookie on the stack for simplicity */
483 * Prevent implicit arguments + the sig cookie from being passed
487 fr = FLOAT_PARAM_REGS;
489 /* Emit the signature cookie just before the implicit arguments */
490 add_general (&gr, &stack_size, &cinfo->sig_cookie);
493 if (sig->params [i]->byref) {
494 add_general (&gr, &stack_size, ainfo);
497 ptype = mono_type_get_underlying_type (sig->params [i]);
498 ptype = mini_get_basic_type_from_generic (gsctx, ptype);
499 switch (ptype->type) {
500 case MONO_TYPE_BOOLEAN:
503 add_general (&gr, &stack_size, ainfo);
508 add_general (&gr, &stack_size, ainfo);
512 add_general (&gr, &stack_size, ainfo);
517 case MONO_TYPE_FNPTR:
518 case MONO_TYPE_CLASS:
519 case MONO_TYPE_OBJECT:
520 case MONO_TYPE_STRING:
521 case MONO_TYPE_SZARRAY:
522 case MONO_TYPE_ARRAY:
523 add_general (&gr, &stack_size, ainfo);
525 case MONO_TYPE_GENERICINST:
526 if (!mono_type_generic_inst_is_valuetype (ptype)) {
527 add_general (&gr, &stack_size, ainfo);
531 case MONO_TYPE_VALUETYPE:
532 case MONO_TYPE_TYPEDBYREF:
534 /* We allways pass valuetypes on the stack */
535 add_valuetype (gsctx, sig, ainfo, sig->params [i], FALSE, &gr, &fr, &stack_size);
539 add_general (&gr, &stack_size, ainfo);
542 add_float (&gr, &fr, &stack_size, ainfo, FALSE);
545 add_float (&gr, &fr, &stack_size, ainfo, TRUE);
548 g_assert_not_reached ();
552 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG) && (n > 0) && (sig->sentinelpos == sig->param_count)) {
554 fr = FLOAT_PARAM_REGS;
556 /* Emit the signature cookie just before the implicit arguments */
557 add_general (&gr, &stack_size, &cinfo->sig_cookie);
560 cinfo->stack_usage = stack_size;
561 cinfo->reg_usage = gr;
562 cinfo->freg_usage = fr;
567 * mono_arch_get_argument_info:
568 * @csig: a method signature
569 * @param_count: the number of parameters to consider
570 * @arg_info: an array to store the result infos
572 * Gathers information on parameters such as size, alignment and
573 * padding. arg_info should be large enought to hold param_count + 1 entries.
575 * Returns the size of the argument area on the stack.
578 mono_arch_get_argument_info (MonoMethodSignature *csig, int param_count, MonoJitArgumentInfo *arg_info)
581 CallInfo *cinfo = get_call_info (NULL, NULL, csig, FALSE);
582 guint32 args_size = cinfo->stack_usage;
584 /* The arguments are saved to a stack area in mono_arch_instrument_prolog */
586 arg_info [0].offset = 0;
589 for (k = 0; k < param_count; k++) {
590 arg_info [k + 1].offset = ((k + csig->hasthis) * 8);
592 arg_info [k + 1].size = 0;
601 * Initialize the cpu to execute managed code.
604 mono_arch_cpu_init (void)
609 * Initialize architecture specific code.
612 mono_arch_init (void)
617 * Cleanup architecture specific code.
620 mono_arch_cleanup (void)
625 * This function returns the optimizations supported on this cpu.
628 mono_arch_cpu_optimizations (guint32 *exclude_mask)
636 * This function test for all SIMD functions supported.
638 * Returns a bitmask corresponding to all supported versions.
642 mono_arch_cpu_enumerate_simd_versions (void)
644 /* SIMD is currently unimplemented */
649 mono_arch_get_allocatable_int_vars (MonoCompile *cfg)
653 MonoMethodSignature *sig;
654 MonoMethodHeader *header;
657 header = cfg->header;
659 sig = mono_method_signature (cfg->method);
661 cinfo = get_call_info (cfg, cfg->mempool, sig, FALSE);
663 for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
664 MonoInst *ins = cfg->args [i];
666 ArgInfo *ainfo = &cinfo->args [i];
668 if (ins->flags & (MONO_INST_IS_DEAD|MONO_INST_VOLATILE|MONO_INST_INDIRECT))
671 if (ainfo->storage == ArgInIReg) {
672 /* The input registers are non-volatile */
673 ins->opcode = OP_REGVAR;
674 ins->dreg = 32 + ainfo->reg;
678 for (i = 0; i < cfg->num_varinfo; i++) {
679 MonoInst *ins = cfg->varinfo [i];
680 MonoMethodVar *vmv = MONO_VARINFO (cfg, i);
683 if (vmv->range.first_use.abs_pos >= vmv->range.last_use.abs_pos)
686 if ((ins->flags & (MONO_INST_IS_DEAD|MONO_INST_VOLATILE|MONO_INST_INDIRECT)) ||
687 (ins->opcode != OP_LOCAL && ins->opcode != OP_ARG))
690 if (mono_is_regsize_var (ins->inst_vtype)) {
691 g_assert (MONO_VARINFO (cfg, i)->reg == -1);
692 g_assert (i == vmv->idx);
693 vars = g_list_prepend (vars, vmv);
697 vars = mono_varlist_sort (cfg, vars, 0);
703 mono_ia64_alloc_stacked_registers (MonoCompile *cfg)
706 guint32 reserved_regs;
707 MonoMethodHeader *header;
709 if (cfg->arch.reg_local0 > 0)
713 cinfo = get_call_info (cfg, cfg->mempool, mono_method_signature (cfg->method), FALSE);
715 header = cfg->header;
717 /* Some registers are reserved for use by the prolog/epilog */
718 reserved_regs = header->num_clauses ? 4 : 3;
720 if ((mono_jit_trace_calls != NULL && mono_trace_eval (cfg->method)) ||
721 (cfg->prof_options & MONO_PROFILE_ENTER_LEAVE)) {
722 /* One registers is needed by instrument_epilog to save the return value */
724 if (cinfo->reg_usage < 2)
725 /* Number of arguments passed to function call in instrument_prolog */
726 cinfo->reg_usage = 2;
729 cfg->arch.reg_in0 = 32;
730 cfg->arch.reg_local0 = cfg->arch.reg_in0 + cinfo->reg_usage + reserved_regs;
731 cfg->arch.reg_out0 = cfg->arch.reg_local0 + 16;
733 cfg->arch.reg_saved_ar_pfs = cfg->arch.reg_local0 - 1;
734 cfg->arch.reg_saved_b0 = cfg->arch.reg_local0 - 2;
735 cfg->arch.reg_fp = cfg->arch.reg_local0 - 3;
738 * Frames without handlers save sp to fp, frames with handlers save it into
739 * a dedicated register.
741 if (header->num_clauses)
742 cfg->arch.reg_saved_sp = cfg->arch.reg_local0 - 4;
744 cfg->arch.reg_saved_sp = cfg->arch.reg_fp;
746 if ((mono_jit_trace_calls != NULL && mono_trace_eval (cfg->method)) ||
747 (cfg->prof_options & MONO_PROFILE_ENTER_LEAVE)) {
748 cfg->arch.reg_saved_return_val = cfg->arch.reg_local0 - reserved_regs;
752 * Need to allocate at least 2 out register for use by OP_THROW / the system
753 * exception throwing code.
755 cfg->arch.n_out_regs = MAX (cfg->arch.n_out_regs, 2);
759 mono_arch_get_global_int_regs (MonoCompile *cfg)
764 mono_ia64_alloc_stacked_registers (cfg);
766 for (i = cfg->arch.reg_local0; i < cfg->arch.reg_out0; ++i) {
769 regs = g_list_prepend (regs, (gpointer)(gssize)(i));
776 * mono_arch_regalloc_cost:
778 * Return the cost, in number of memory references, of the action of
779 * allocating the variable VMV into a register during global register
783 mono_arch_regalloc_cost (MonoCompile *cfg, MonoMethodVar *vmv)
785 /* FIXME: Increase costs linearly to avoid using all local registers */
791 mono_arch_allocate_vars (MonoCompile *cfg)
793 MonoMethodSignature *sig;
794 MonoMethodHeader *header;
797 guint32 locals_stack_size, locals_stack_align;
801 header = cfg->header;
803 sig = mono_method_signature (cfg->method);
805 cinfo = get_call_info (cfg, cfg->mempool, sig, FALSE);
808 * Determine whenever the frame pointer can be eliminated.
809 * FIXME: Remove some of the restrictions.
811 cfg->arch.omit_fp = TRUE;
813 if (!debug_omit_fp ())
814 cfg->arch.omit_fp = FALSE;
816 if (cfg->flags & MONO_CFG_HAS_ALLOCA)
817 cfg->arch.omit_fp = FALSE;
818 if (header->num_clauses)
819 cfg->arch.omit_fp = FALSE;
821 cfg->arch.omit_fp = FALSE;
822 if ((sig->ret->type != MONO_TYPE_VOID) && (cinfo->ret.storage == ArgAggregate))
823 cfg->arch.omit_fp = FALSE;
824 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG))
825 cfg->arch.omit_fp = FALSE;
826 for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
827 ArgInfo *ainfo = &cinfo->args [i];
829 if (ainfo->storage == ArgOnStack) {
831 * The stack offset can only be determined when the frame
834 cfg->arch.omit_fp = FALSE;
838 mono_ia64_alloc_stacked_registers (cfg);
841 * We use the ABI calling conventions for managed code as well.
842 * Exception: valuetypes are never passed or returned in registers.
845 if (cfg->arch.omit_fp) {
846 cfg->flags |= MONO_CFG_HAS_SPILLUP;
847 cfg->frame_reg = IA64_SP;
848 offset = ARGS_OFFSET;
851 /* Locals are allocated backwards from %fp */
852 cfg->frame_reg = cfg->arch.reg_fp;
856 if (cfg->method->save_lmf) {
860 if (sig->ret->type != MONO_TYPE_VOID) {
861 switch (cinfo->ret.storage) {
863 cfg->ret->opcode = OP_REGVAR;
864 cfg->ret->inst_c0 = cinfo->ret.reg;
867 cfg->ret->opcode = OP_REGVAR;
868 cfg->ret->inst_c0 = cinfo->ret.reg;
870 case ArgValuetypeAddrInIReg:
871 cfg->vret_addr->opcode = OP_REGVAR;
872 cfg->vret_addr->dreg = cfg->arch.reg_in0 + cinfo->ret.reg;
875 /* Allocate a local to hold the result, the epilog will copy it to the correct place */
876 if (cfg->arch.omit_fp)
877 g_assert_not_reached ();
878 offset = ALIGN_TO (offset, 8);
879 offset += cinfo->ret.nslots * 8;
880 cfg->ret->opcode = OP_REGOFFSET;
881 cfg->ret->inst_basereg = cfg->frame_reg;
882 cfg->ret->inst_offset = - offset;
885 g_assert_not_reached ();
887 cfg->ret->dreg = cfg->ret->inst_c0;
890 /* Allocate locals */
891 offsets = mono_allocate_stack_slots (cfg, cfg->arch.omit_fp ? FALSE : TRUE, &locals_stack_size, &locals_stack_align);
892 if (locals_stack_align) {
893 offset = ALIGN_TO (offset, locals_stack_align);
895 for (i = cfg->locals_start; i < cfg->num_varinfo; i++) {
896 if (offsets [i] != -1) {
897 MonoInst *inst = cfg->varinfo [i];
898 inst->opcode = OP_REGOFFSET;
899 inst->inst_basereg = cfg->frame_reg;
900 if (cfg->arch.omit_fp)
901 inst->inst_offset = (offset + offsets [i]);
903 inst->inst_offset = - (offset + offsets [i]);
904 // printf ("allocated local %d to ", i); mono_print_tree_nl (inst);
907 offset += locals_stack_size;
909 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG)) {
910 if (cfg->arch.omit_fp)
911 g_assert_not_reached ();
912 g_assert (cinfo->sig_cookie.storage == ArgOnStack);
913 cfg->sig_cookie = cinfo->sig_cookie.offset + ARGS_OFFSET;
916 for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
917 inst = cfg->args [i];
918 if (inst->opcode != OP_REGVAR) {
919 ArgInfo *ainfo = &cinfo->args [i];
920 gboolean inreg = TRUE;
923 if (sig->hasthis && (i == 0))
924 arg_type = &mono_defaults.object_class->byval_arg;
926 arg_type = sig->params [i - sig->hasthis];
928 /* FIXME: VOLATILE is only set if the liveness pass runs */
929 if (inst->flags & (MONO_INST_VOLATILE|MONO_INST_INDIRECT))
932 inst->opcode = OP_REGOFFSET;
934 switch (ainfo->storage) {
936 inst->opcode = OP_REGVAR;
937 inst->dreg = cfg->arch.reg_in0 + ainfo->reg;
940 case ArgInFloatRegR4:
942 * Since float regs are volatile, we save the arguments to
943 * the stack in the prolog.
948 if (cfg->arch.omit_fp)
949 g_assert_not_reached ();
950 inst->opcode = OP_REGOFFSET;
951 inst->inst_basereg = cfg->frame_reg;
952 inst->inst_offset = ARGS_OFFSET + ainfo->offset;
961 if (!inreg && (ainfo->storage != ArgOnStack)) {
964 inst->opcode = OP_REGOFFSET;
965 inst->inst_basereg = cfg->frame_reg;
966 /* These arguments are saved to the stack in the prolog */
967 switch (ainfo->storage) {
969 if (ainfo->atype == AggregateSingleHFA)
970 size = ainfo->nslots * 4;
972 size = ainfo->nslots * 8;
975 size = sizeof (gpointer);
979 offset = ALIGN_TO (offset, sizeof (gpointer));
981 if (cfg->arch.omit_fp) {
982 inst->inst_offset = offset;
986 inst->inst_offset = - offset;
993 * FIXME: This doesn't work because some variables are allocated during local
997 if (cfg->arch.omit_fp && offset == 16)
1001 cfg->stack_offset = offset;
1005 mono_arch_create_vars (MonoCompile *cfg)
1007 MonoMethodSignature *sig;
1010 sig = mono_method_signature (cfg->method);
1012 cinfo = get_call_info (cfg, cfg->mempool, sig, FALSE);
1014 if (cinfo->ret.storage == ArgAggregate)
1015 cfg->ret_var_is_local = TRUE;
1016 if (cinfo->ret.storage == ArgValuetypeAddrInIReg) {
1017 cfg->vret_addr = mono_compile_create_var (cfg, &mono_defaults.int_class->byval_arg, OP_ARG);
1018 if (G_UNLIKELY (cfg->verbose_level > 1)) {
1019 printf ("vret_addr = ");
1020 mono_print_ins (cfg->vret_addr);
1026 add_outarg_reg (MonoCompile *cfg, MonoCallInst *call, ArgStorage storage, int reg, MonoInst *tree)
1030 MONO_INST_NEW (cfg, arg, OP_NOP);
1031 arg->sreg1 = tree->dreg;
1035 arg->opcode = OP_MOVE;
1036 arg->dreg = mono_alloc_ireg (cfg);
1038 mono_call_inst_add_outarg_reg (cfg, call, arg->dreg, reg, FALSE);
1041 arg->opcode = OP_FMOVE;
1042 arg->dreg = mono_alloc_freg (cfg);
1044 mono_call_inst_add_outarg_reg (cfg, call, arg->dreg, reg, TRUE);
1046 case ArgInFloatRegR4:
1047 arg->opcode = OP_FCONV_TO_R4;
1048 arg->dreg = mono_alloc_freg (cfg);
1050 mono_call_inst_add_outarg_reg (cfg, call, arg->dreg, reg, TRUE);
1053 g_assert_not_reached ();
1056 MONO_ADD_INS (cfg->cbb, arg);
1060 emit_sig_cookie (MonoCompile *cfg, MonoCallInst *call, CallInfo *cinfo)
1062 MonoMethodSignature *tmp_sig;
1064 /* Emit the signature cookie just before the implicit arguments */
1066 /* FIXME: Add support for signature tokens to AOT */
1067 cfg->disable_aot = TRUE;
1069 g_assert (cinfo->sig_cookie.storage == ArgOnStack);
1072 * mono_ArgIterator_Setup assumes the signature cookie is
1073 * passed first and all the arguments which were before it are
1074 * passed on the stack after the signature. So compensate by
1075 * passing a different signature.
1077 tmp_sig = mono_metadata_signature_dup (call->signature);
1078 tmp_sig->param_count -= call->signature->sentinelpos;
1079 tmp_sig->sentinelpos = 0;
1080 memcpy (tmp_sig->params, call->signature->params + call->signature->sentinelpos, tmp_sig->param_count * sizeof (MonoType*));
1082 MONO_INST_NEW (cfg, sig_arg, OP_ICONST);
1083 sig_arg->dreg = mono_alloc_ireg (cfg);
1084 sig_arg->inst_p0 = tmp_sig;
1085 MONO_ADD_INS (cfg->cbb, sig_arg);
1087 MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STOREI8_MEMBASE_REG, IA64_SP, 16 + cinfo->sig_cookie.offset, sig_arg->dreg);
1091 mono_arch_emit_call (MonoCompile *cfg, MonoCallInst *call)
1094 MonoMethodSignature *sig;
1095 int i, n, stack_size;
1101 mono_ia64_alloc_stacked_registers (cfg);
1103 sig = call->signature;
1104 n = sig->param_count + sig->hasthis;
1106 cinfo = get_call_info (cfg, cfg->mempool, sig, sig->pinvoke);
1108 if (cinfo->ret.storage == ArgAggregate) {
1113 * The valuetype is in registers after the call, need to be copied
1114 * to the stack. Save the address to a local here, so the call
1115 * instruction can access it.
1117 local = mono_compile_create_var (cfg, &mono_defaults.int_class->byval_arg, OP_LOCAL);
1118 local->flags |= MONO_INST_VOLATILE;
1119 cfg->arch.ret_var_addr_local = local;
1121 MONO_INST_NEW (cfg, vtarg, OP_MOVE);
1122 vtarg->sreg1 = call->vret_var->dreg;
1123 vtarg->dreg = local->dreg;
1124 MONO_ADD_INS (cfg->cbb, vtarg);
1127 if (cinfo->ret.storage == ArgValuetypeAddrInIReg) {
1128 add_outarg_reg (cfg, call, ArgInIReg, cfg->arch.reg_out0 + cinfo->ret.reg, call->vret_var);
1131 for (i = 0; i < n; ++i) {
1134 ainfo = cinfo->args + i;
1136 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG) && (i == sig->sentinelpos)) {
1137 /* Emit the signature cookie just before the implicit arguments */
1138 emit_sig_cookie (cfg, call, cinfo);
1141 in = call->args [i];
1143 if (sig->hasthis && (i == 0))
1144 arg_type = &mono_defaults.object_class->byval_arg;
1146 arg_type = sig->params [i - sig->hasthis];
1148 if ((i >= sig->hasthis) && (MONO_TYPE_ISSTRUCT(arg_type))) {
1152 if (arg_type->type == MONO_TYPE_TYPEDBYREF) {
1153 size = sizeof (MonoTypedRef);
1154 align = sizeof (gpointer);
1156 else if (sig->pinvoke)
1157 size = mono_type_native_stack_size (&in->klass->byval_arg, &align);
1160 * Other backends use mono_type_stack_size (), but that
1161 * aligns the size to 8, which is larger than the size of
1162 * the source, leading to reads of invalid memory if the
1163 * source is at the end of address space.
1165 size = mono_class_value_size (in->klass, &align);
1171 MONO_INST_NEW (cfg, arg, OP_OUTARG_VT);
1172 arg->sreg1 = in->dreg;
1173 arg->klass = in->klass;
1174 arg->backend.size = size;
1175 arg->inst_p0 = call;
1176 arg->inst_p1 = mono_mempool_alloc (cfg->mempool, sizeof (ArgInfo));
1177 memcpy (arg->inst_p1, ainfo, sizeof (ArgInfo));
1179 MONO_ADD_INS (cfg->cbb, arg);
1183 switch (ainfo->storage) {
1185 add_outarg_reg (cfg, call, ainfo->storage, cfg->arch.reg_out0 + ainfo->reg, in);
1188 case ArgInFloatRegR4:
1189 add_outarg_reg (cfg, call, ainfo->storage, ainfo->reg, in);
1192 if (arg_type->type == MONO_TYPE_R4 && !arg_type->byref)
1193 MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STORER4_MEMBASE_REG, IA64_SP, 16 + ainfo->offset, in->dreg);
1194 else if (arg_type->type == MONO_TYPE_R8 && !arg_type->byref)
1195 MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STORER8_MEMBASE_REG, IA64_SP, 16 + ainfo->offset, in->dreg);
1197 MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STOREI8_MEMBASE_REG, IA64_SP, 16 + ainfo->offset, in->dreg);
1200 g_assert_not_reached ();
1205 /* Handle the case where there are no implicit arguments */
1206 if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG) && (n == sig->sentinelpos)) {
1207 emit_sig_cookie (cfg, call, cinfo);
1210 call->stack_usage = cinfo->stack_usage;
1211 cfg->arch.n_out_regs = MAX (cfg->arch.n_out_regs, cinfo->reg_usage);
1215 mono_arch_emit_outarg_vt (MonoCompile *cfg, MonoInst *ins, MonoInst *src)
1217 MonoCallInst *call = (MonoCallInst*)ins->inst_p0;
1218 ArgInfo *ainfo = (ArgInfo*)ins->inst_p1;
1219 int size = ins->backend.size;
1221 if (ainfo->storage == ArgAggregate) {
1222 MonoInst *load, *store;
1226 * Part of the structure is passed in registers.
1228 for (i = 0; i < ainfo->nregs; ++i) {
1229 slot = ainfo->reg + i;
1231 if (ainfo->atype == AggregateSingleHFA) {
1232 MONO_INST_NEW (cfg, load, OP_LOADR4_MEMBASE);
1233 load->inst_basereg = src->dreg;
1234 load->inst_offset = i * 4;
1235 load->dreg = mono_alloc_freg (cfg);
1237 mono_call_inst_add_outarg_reg (cfg, call, load->dreg, ainfo->reg + i, TRUE);
1238 } else if (ainfo->atype == AggregateDoubleHFA) {
1239 MONO_INST_NEW (cfg, load, OP_LOADR8_MEMBASE);
1240 load->inst_basereg = src->dreg;
1241 load->inst_offset = i * 8;
1242 load->dreg = mono_alloc_freg (cfg);
1244 mono_call_inst_add_outarg_reg (cfg, call, load->dreg, ainfo->reg + i, TRUE);
1246 MONO_INST_NEW (cfg, load, OP_LOADI8_MEMBASE);
1247 load->inst_basereg = src->dreg;
1248 load->inst_offset = i * 8;
1249 load->dreg = mono_alloc_ireg (cfg);
1251 mono_call_inst_add_outarg_reg (cfg, call, load->dreg, cfg->arch.reg_out0 + ainfo->reg + i, FALSE);
1253 MONO_ADD_INS (cfg->cbb, load);
1257 * Part of the structure is passed on the stack.
1259 for (i = ainfo->nregs; i < ainfo->nslots; ++i) {
1260 slot = ainfo->reg + i;
1262 MONO_INST_NEW (cfg, load, OP_LOADI8_MEMBASE);
1263 load->inst_basereg = src->dreg;
1264 load->inst_offset = i * sizeof (gpointer);
1265 load->dreg = mono_alloc_preg (cfg);
1266 MONO_ADD_INS (cfg->cbb, load);
1268 MONO_INST_NEW (cfg, store, OP_STOREI8_MEMBASE_REG);
1269 store->sreg1 = load->dreg;
1270 store->inst_destbasereg = IA64_SP;
1271 store->inst_offset = 16 + ainfo->offset + (slot - 8) * 8;
1272 MONO_ADD_INS (cfg->cbb, store);
1275 mini_emit_memcpy (cfg, IA64_SP, 16 + ainfo->offset, src->dreg, 0, size, 4);
1280 mono_arch_emit_setret (MonoCompile *cfg, MonoMethod *method, MonoInst *val)
1282 CallInfo *cinfo = get_call_info (cfg, cfg->mempool, mono_method_signature (method), FALSE);
1284 switch (cinfo->ret.storage) {
1286 MONO_EMIT_NEW_UNALU (cfg, OP_MOVE, cfg->ret->dreg, val->dreg);
1289 MONO_EMIT_NEW_UNALU (cfg, OP_FMOVE, cfg->ret->dreg, val->dreg);
1292 g_assert_not_reached ();
1297 mono_arch_peephole_pass_1 (MonoCompile *cfg, MonoBasicBlock *bb)
1302 mono_arch_peephole_pass_2 (MonoCompile *cfg, MonoBasicBlock *bb)
1304 MonoInst *ins, *n, *last_ins = NULL;
1307 MONO_BB_FOR_EACH_INS_SAFE (bb, n, ins) {
1308 switch (ins->opcode) {
1316 if (ins->dreg == ins->sreg1) {
1317 MONO_DELETE_INS (bb, ins);
1323 * OP_MOVE sreg, dreg
1324 * OP_MOVE dreg, sreg
1326 if (last_ins && last_ins->opcode == OP_MOVE &&
1327 ins->sreg1 == last_ins->dreg &&
1328 ins->dreg == last_ins->sreg1) {
1329 MONO_DELETE_INS (bb, ins);
1335 /* remove unnecessary multiplication with 1 */
1336 if (ins->inst_imm == 1) {
1337 if (ins->dreg != ins->sreg1) {
1338 ins->opcode = OP_MOVE;
1340 MONO_DELETE_INS (bb, ins);
1350 bb->last_ins = last_ins;
1353 int cond_to_ia64_cmp [][3] = {
1354 {OP_IA64_CMP_EQ, OP_IA64_CMP4_EQ, OP_IA64_FCMP_EQ},
1355 {OP_IA64_CMP_NE, OP_IA64_CMP4_NE, OP_IA64_FCMP_NE},
1356 {OP_IA64_CMP_LE, OP_IA64_CMP4_LE, OP_IA64_FCMP_LE},
1357 {OP_IA64_CMP_GE, OP_IA64_CMP4_GE, OP_IA64_FCMP_GE},
1358 {OP_IA64_CMP_LT, OP_IA64_CMP4_LT, OP_IA64_FCMP_LT},
1359 {OP_IA64_CMP_GT, OP_IA64_CMP4_GT, OP_IA64_FCMP_GT},
1360 {OP_IA64_CMP_LE_UN, OP_IA64_CMP4_LE_UN, OP_IA64_FCMP_LE_UN},
1361 {OP_IA64_CMP_GE_UN, OP_IA64_CMP4_GE_UN, OP_IA64_FCMP_GE_UN},
1362 {OP_IA64_CMP_LT_UN, OP_IA64_CMP4_LT_UN, OP_IA64_FCMP_LT_UN},
1363 {OP_IA64_CMP_GT_UN, OP_IA64_CMP4_GT_UN, OP_IA64_FCMP_GT_UN}
1367 opcode_to_ia64_cmp (int opcode, int cmp_opcode)
1369 return cond_to_ia64_cmp [mono_opcode_to_cond (opcode)][mono_opcode_to_type (opcode, cmp_opcode)];
1372 int cond_to_ia64_cmp_imm [][3] = {
1373 {OP_IA64_CMP_EQ_IMM, OP_IA64_CMP4_EQ_IMM, 0},
1374 {OP_IA64_CMP_NE_IMM, OP_IA64_CMP4_NE_IMM, 0},
1375 {OP_IA64_CMP_GE_IMM, OP_IA64_CMP4_GE_IMM, 0},
1376 {OP_IA64_CMP_LE_IMM, OP_IA64_CMP4_LE_IMM, 0},
1377 {OP_IA64_CMP_GT_IMM, OP_IA64_CMP4_GT_IMM, 0},
1378 {OP_IA64_CMP_LT_IMM, OP_IA64_CMP4_LT_IMM, 0},
1379 {OP_IA64_CMP_GE_UN_IMM, OP_IA64_CMP4_GE_UN_IMM, 0},
1380 {OP_IA64_CMP_LE_UN_IMM, OP_IA64_CMP4_LE_UN_IMM, 0},
1381 {OP_IA64_CMP_GT_UN_IMM, OP_IA64_CMP4_GT_UN_IMM, 0},
1382 {OP_IA64_CMP_LT_UN_IMM, OP_IA64_CMP4_LT_UN_IMM, 0},
1386 opcode_to_ia64_cmp_imm (int opcode, int cmp_opcode)
1388 /* The condition needs to be reversed */
1389 return cond_to_ia64_cmp_imm [mono_opcode_to_cond (opcode)][mono_opcode_to_type (opcode, cmp_opcode)];
1392 #define NEW_INS(cfg,dest,op) do { \
1393 (dest) = mono_mempool_alloc0 ((cfg)->mempool, sizeof (MonoInst)); \
1394 (dest)->opcode = (op); \
1395 mono_bblock_insert_after_ins (bb, last_ins, (dest)); \
1396 last_ins = (dest); \
1400 * mono_arch_lowering_pass:
1402 * Converts complex opcodes into simpler ones so that each IR instruction
1403 * corresponds to one machine instruction.
1406 mono_arch_lowering_pass (MonoCompile *cfg, MonoBasicBlock *bb)
1408 MonoInst *ins, *n, *next, *temp, *temp2, *temp3, *last_ins = NULL;
1411 MONO_BB_FOR_EACH_INS_SAFE (bb, n, ins) {
1412 switch (ins->opcode) {
1413 case OP_STOREI1_MEMBASE_IMM:
1414 case OP_STOREI2_MEMBASE_IMM:
1415 case OP_STOREI4_MEMBASE_IMM:
1416 case OP_STOREI8_MEMBASE_IMM:
1417 case OP_STORE_MEMBASE_IMM:
1418 /* There are no store_membase instructions on ia64 */
1419 if (ins->inst_offset == 0) {
1421 } else if (ia64_is_imm14 (ins->inst_offset)) {
1422 NEW_INS (cfg, temp2, OP_ADD_IMM);
1423 temp2->sreg1 = ins->inst_destbasereg;
1424 temp2->inst_imm = ins->inst_offset;
1425 temp2->dreg = mono_alloc_ireg (cfg);
1428 NEW_INS (cfg, temp, OP_I8CONST);
1429 temp->inst_c0 = ins->inst_offset;
1430 temp->dreg = mono_alloc_ireg (cfg);
1432 NEW_INS (cfg, temp2, OP_LADD);
1433 temp2->sreg1 = ins->inst_destbasereg;
1434 temp2->sreg2 = temp->dreg;
1435 temp2->dreg = mono_alloc_ireg (cfg);
1438 switch (ins->opcode) {
1439 case OP_STOREI1_MEMBASE_IMM:
1440 ins->opcode = OP_STOREI1_MEMBASE_REG;
1442 case OP_STOREI2_MEMBASE_IMM:
1443 ins->opcode = OP_STOREI2_MEMBASE_REG;
1445 case OP_STOREI4_MEMBASE_IMM:
1446 ins->opcode = OP_STOREI4_MEMBASE_REG;
1448 case OP_STOREI8_MEMBASE_IMM:
1449 case OP_STORE_MEMBASE_IMM:
1450 ins->opcode = OP_STOREI8_MEMBASE_REG;
1453 g_assert_not_reached ();
1456 if (ins->inst_imm == 0)
1457 ins->sreg1 = IA64_R0;
1459 NEW_INS (cfg, temp3, OP_I8CONST);
1460 temp3->inst_c0 = ins->inst_imm;
1461 temp3->dreg = mono_alloc_ireg (cfg);
1462 ins->sreg1 = temp3->dreg;
1465 ins->inst_offset = 0;
1467 ins->inst_destbasereg = temp2->dreg;
1469 case OP_STOREI1_MEMBASE_REG:
1470 case OP_STOREI2_MEMBASE_REG:
1471 case OP_STOREI4_MEMBASE_REG:
1472 case OP_STOREI8_MEMBASE_REG:
1473 case OP_STORER4_MEMBASE_REG:
1474 case OP_STORER8_MEMBASE_REG:
1475 case OP_STORE_MEMBASE_REG:
1476 /* There are no store_membase instructions on ia64 */
1477 if (ins->inst_offset == 0) {
1480 else if (ia64_is_imm14 (ins->inst_offset)) {
1481 NEW_INS (cfg, temp2, OP_ADD_IMM);
1482 temp2->sreg1 = ins->inst_destbasereg;
1483 temp2->inst_imm = ins->inst_offset;
1484 temp2->dreg = mono_alloc_ireg (cfg);
1487 NEW_INS (cfg, temp, OP_I8CONST);
1488 temp->inst_c0 = ins->inst_offset;
1489 temp->dreg = mono_alloc_ireg (cfg);
1490 NEW_INS (cfg, temp2, OP_LADD);
1491 temp2->sreg1 = ins->inst_destbasereg;
1492 temp2->sreg2 = temp->dreg;
1493 temp2->dreg = mono_alloc_ireg (cfg);
1496 ins->inst_offset = 0;
1497 ins->inst_destbasereg = temp2->dreg;
1499 case OP_LOADI1_MEMBASE:
1500 case OP_LOADU1_MEMBASE:
1501 case OP_LOADI2_MEMBASE:
1502 case OP_LOADU2_MEMBASE:
1503 case OP_LOADI4_MEMBASE:
1504 case OP_LOADU4_MEMBASE:
1505 case OP_LOADI8_MEMBASE:
1506 case OP_LOAD_MEMBASE:
1507 case OP_LOADR4_MEMBASE:
1508 case OP_LOADR8_MEMBASE:
1509 case OP_ATOMIC_EXCHANGE_I4:
1510 case OP_ATOMIC_EXCHANGE_I8:
1511 case OP_ATOMIC_ADD_NEW_I4:
1512 case OP_ATOMIC_ADD_NEW_I8:
1513 case OP_ATOMIC_ADD_IMM_NEW_I4:
1514 case OP_ATOMIC_ADD_IMM_NEW_I8:
1515 /* There are no membase instructions on ia64 */
1516 if (ins->inst_offset == 0) {
1519 else if (ia64_is_imm14 (ins->inst_offset)) {
1520 NEW_INS (cfg, temp2, OP_ADD_IMM);
1521 temp2->sreg1 = ins->inst_basereg;
1522 temp2->inst_imm = ins->inst_offset;
1523 temp2->dreg = mono_alloc_ireg (cfg);
1526 NEW_INS (cfg, temp, OP_I8CONST);
1527 temp->inst_c0 = ins->inst_offset;
1528 temp->dreg = mono_alloc_ireg (cfg);
1529 NEW_INS (cfg, temp2, OP_LADD);
1530 temp2->sreg1 = ins->inst_basereg;
1531 temp2->sreg2 = temp->dreg;
1532 temp2->dreg = mono_alloc_ireg (cfg);
1535 ins->inst_offset = 0;
1536 ins->inst_basereg = temp2->dreg;
1556 case OP_ISHR_UN_IMM:
1557 case OP_LSHR_UN_IMM: {
1558 gboolean is_imm = FALSE;
1559 gboolean switched = FALSE;
1561 if (ins->opcode == OP_AND_IMM && ins->inst_imm == 255) {
1562 ins->opcode = OP_ZEXT_I1;
1566 switch (ins->opcode) {
1570 is_imm = ia64_is_imm14 (ins->inst_imm);
1575 is_imm = ia64_is_imm14 (- (ins->inst_imm));
1577 /* A = B - IMM -> A = B + (-IMM) */
1578 ins->inst_imm = - ins->inst_imm;
1579 ins->opcode = OP_IADD_IMM;
1590 is_imm = ia64_is_imm8 (ins->inst_imm);
1599 case OP_ISHR_UN_IMM:
1600 case OP_LSHR_UN_IMM:
1601 is_imm = (ins->inst_imm >= 0) && (ins->inst_imm < 64);
1609 ins->sreg2 = ins->sreg1;
1613 ins->opcode = mono_op_imm_to_op (ins->opcode);
1615 if (ins->inst_imm == 0)
1616 ins->sreg2 = IA64_R0;
1618 NEW_INS (cfg, temp, OP_I8CONST);
1619 temp->inst_c0 = ins->inst_imm;
1620 temp->dreg = mono_alloc_ireg (cfg);
1621 ins->sreg2 = temp->dreg;
1625 case OP_COMPARE_IMM:
1626 case OP_ICOMPARE_IMM:
1627 case OP_LCOMPARE_IMM: {
1628 /* Instead of compare+b<cond>, ia64 has compare<cond>+br */
1634 /* Branch opts can eliminate the branch */
1635 if (!next || (!(MONO_IS_COND_BRANCH_OP (next) || MONO_IS_COND_EXC (next) || MONO_IS_SETCC (next)))) {
1641 * The compare_imm instructions have switched up arguments, and
1642 * some of them take an imm between -127 and 128.
1645 cond = mono_opcode_to_cond (next->opcode);
1646 if ((cond == CMP_LT) || (cond == CMP_GE))
1647 imm = ia64_is_imm8 (ins->inst_imm - 1);
1648 else if ((cond == CMP_LT_UN) || (cond == CMP_GE_UN))
1649 imm = ia64_is_imm8 (ins->inst_imm - 1) && (ins->inst_imm > 0);
1651 imm = ia64_is_imm8 (ins->inst_imm);
1654 ins->opcode = opcode_to_ia64_cmp_imm (next->opcode, ins->opcode);
1655 ins->sreg2 = ins->sreg1;
1658 ins->opcode = opcode_to_ia64_cmp (next->opcode, ins->opcode);
1660 if (ins->inst_imm == 0)
1661 ins->sreg2 = IA64_R0;
1663 NEW_INS (cfg, temp, OP_I8CONST);
1664 temp->inst_c0 = ins->inst_imm;
1665 temp->dreg = mono_alloc_ireg (cfg);
1666 ins->sreg2 = temp->dreg;
1670 if (MONO_IS_COND_BRANCH_OP (next)) {
1671 next->opcode = OP_IA64_BR_COND;
1672 next->inst_target_bb = next->inst_true_bb;
1673 } else if (MONO_IS_COND_EXC (next)) {
1674 next->opcode = OP_IA64_COND_EXC;
1675 } else if (MONO_IS_SETCC (next)) {
1676 next->opcode = OP_IA64_CSET;
1678 printf ("%s\n", mono_inst_name (next->opcode));
1688 /* Instead of compare+b<cond>, ia64 has compare<cond>+br */
1692 /* Branch opts can eliminate the branch */
1693 if (!next || (!(MONO_IS_COND_BRANCH_OP (next) || MONO_IS_COND_EXC (next) || MONO_IS_SETCC (next)))) {
1698 ins->opcode = opcode_to_ia64_cmp (next->opcode, ins->opcode);
1700 if (MONO_IS_COND_BRANCH_OP (next)) {
1701 next->opcode = OP_IA64_BR_COND;
1702 next->inst_target_bb = next->inst_true_bb;
1703 } else if (MONO_IS_COND_EXC (next)) {
1704 next->opcode = OP_IA64_COND_EXC;
1705 } else if (MONO_IS_SETCC (next)) {
1706 next->opcode = OP_IA64_CSET;
1708 printf ("%s\n", mono_inst_name (next->opcode));
1719 /* The front end removes the fcompare, so introduce it again */
1720 NEW_INS (cfg, temp, opcode_to_ia64_cmp (ins->opcode, OP_FCOMPARE));
1721 temp->sreg1 = ins->sreg1;
1722 temp->sreg2 = ins->sreg2;
1724 ins->opcode = OP_IA64_CSET;
1725 MONO_INST_NULLIFY_SREGS (ins);
1731 gboolean found = FALSE;
1732 int shl_op = ins->opcode == OP_IMUL_IMM ? OP_ISHL_IMM : OP_SHL_IMM;
1734 /* First the easy cases */
1735 if (ins->inst_imm == 1) {
1736 ins->opcode = OP_MOVE;
1739 for (i = 1; i < 64; ++i)
1740 if (ins->inst_imm == (((gint64)1) << i)) {
1741 ins->opcode = shl_op;
1747 /* This could be optimized */
1750 for (i = 0; i < 64; ++i) {
1751 if (ins->inst_imm & (((gint64)1) << i)) {
1752 NEW_INS (cfg, temp, shl_op);
1753 temp->dreg = mono_alloc_ireg (cfg);
1754 temp->sreg1 = ins->sreg1;
1758 sum_reg = temp->dreg;
1760 NEW_INS (cfg, temp2, OP_LADD);
1761 temp2->dreg = mono_alloc_ireg (cfg);
1762 temp2->sreg1 = sum_reg;
1763 temp2->sreg2 = temp->dreg;
1764 sum_reg = temp2->dreg;
1768 ins->opcode = OP_MOVE;
1769 ins->sreg1 = sum_reg;
1773 case OP_LCONV_TO_OVF_U4:
1774 NEW_INS (cfg, temp, OP_IA64_CMP4_LT);
1775 temp->sreg1 = ins->sreg1;
1776 temp->sreg2 = IA64_R0;
1778 NEW_INS (cfg, temp, OP_IA64_COND_EXC);
1779 temp->inst_p1 = (char*)"OverflowException";
1781 ins->opcode = OP_MOVE;
1783 case OP_LCONV_TO_OVF_I4_UN:
1784 NEW_INS (cfg, temp, OP_ICONST);
1785 temp->inst_c0 = 0x7fffffff;
1786 temp->dreg = mono_alloc_ireg (cfg);
1788 NEW_INS (cfg, temp2, OP_IA64_CMP4_GT_UN);
1789 temp2->sreg1 = ins->sreg1;
1790 temp2->sreg2 = temp->dreg;
1792 NEW_INS (cfg, temp, OP_IA64_COND_EXC);
1793 temp->inst_p1 = (char*)"OverflowException";
1795 ins->opcode = OP_MOVE;
1797 case OP_FCONV_TO_I4:
1798 case OP_FCONV_TO_I2:
1799 case OP_FCONV_TO_U2:
1800 case OP_FCONV_TO_I1:
1801 case OP_FCONV_TO_U1:
1802 NEW_INS (cfg, temp, OP_FCONV_TO_I8);
1803 temp->sreg1 = ins->sreg1;
1804 temp->dreg = ins->dreg;
1806 switch (ins->opcode) {
1807 case OP_FCONV_TO_I4:
1808 ins->opcode = OP_SEXT_I4;
1810 case OP_FCONV_TO_I2:
1811 ins->opcode = OP_SEXT_I2;
1813 case OP_FCONV_TO_U2:
1814 ins->opcode = OP_ZEXT_I4;
1816 case OP_FCONV_TO_I1:
1817 ins->opcode = OP_SEXT_I1;
1819 case OP_FCONV_TO_U1:
1820 ins->opcode = OP_ZEXT_I1;
1823 g_assert_not_reached ();
1825 ins->sreg1 = ins->dreg;
1833 bb->last_ins = last_ins;
1835 bb->max_vreg = cfg->next_vreg;
1839 * emit_load_volatile_arguments:
1841 * Load volatile arguments from the stack to the original input registers.
1842 * Required before a tail call.
1844 static Ia64CodegenState
1845 emit_load_volatile_arguments (MonoCompile *cfg, Ia64CodegenState code)
1847 MonoMethod *method = cfg->method;
1848 MonoMethodSignature *sig;
1853 /* FIXME: Generate intermediate code instead */
1855 sig = mono_method_signature (method);
1857 cinfo = get_call_info (cfg, cfg->mempool, sig, FALSE);
1859 /* This is the opposite of the code in emit_prolog */
1860 for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
1861 ArgInfo *ainfo = cinfo->args + i;
1862 gint32 stack_offset;
1865 ins = cfg->args [i];
1867 if (sig->hasthis && (i == 0))
1868 arg_type = &mono_defaults.object_class->byval_arg;
1870 arg_type = sig->params [i - sig->hasthis];
1872 arg_type = mono_type_get_underlying_type (arg_type);
1874 stack_offset = ainfo->offset + ARGS_OFFSET;
1876 /* Save volatile arguments to the stack */
1877 if (ins->opcode != OP_REGVAR) {
1878 switch (ainfo->storage) {
1881 /* FIXME: big offsets */
1882 g_assert (ins->opcode == OP_REGOFFSET);
1883 ia64_adds_imm (code, GP_SCRATCH_REG, ins->inst_offset, ins->inst_basereg);
1884 if (arg_type->byref)
1885 ia64_ld8 (code, cfg->arch.reg_in0 + ainfo->reg, GP_SCRATCH_REG);
1887 switch (arg_type->type) {
1889 ia64_ldfs (code, ainfo->reg, GP_SCRATCH_REG);
1892 ia64_ldfd (code, ainfo->reg, GP_SCRATCH_REG);
1895 ia64_ld8 (code, cfg->arch.reg_in0 + ainfo->reg, GP_SCRATCH_REG);
1907 if (ins->opcode == OP_REGVAR) {
1908 /* Argument allocated to (non-volatile) register */
1909 switch (ainfo->storage) {
1911 if (ins->dreg != cfg->arch.reg_in0 + ainfo->reg)
1912 ia64_mov (code, cfg->arch.reg_in0 + ainfo->reg, ins->dreg);
1915 ia64_adds_imm (code, GP_SCRATCH_REG, 16 + ainfo->offset, cfg->frame_reg);
1916 ia64_st8 (code, GP_SCRATCH_REG, ins->dreg);
1927 static Ia64CodegenState
1928 emit_move_return_value (MonoCompile *cfg, MonoInst *ins, Ia64CodegenState code)
1933 /* Move return value to the target register */
1934 switch (ins->opcode) {
1936 case OP_VOIDCALL_REG:
1937 case OP_VOIDCALL_MEMBASE:
1941 case OP_CALL_MEMBASE:
1944 case OP_LCALL_MEMBASE:
1945 g_assert (ins->dreg == IA64_R8);
1949 case OP_FCALL_MEMBASE:
1950 g_assert (ins->dreg == 8);
1951 if (((MonoCallInst*)ins)->signature->ret->type == MONO_TYPE_R4)
1952 ia64_fnorm_d_sf (code, ins->dreg, ins->dreg, 0);
1956 case OP_VCALL_MEMBASE:
1959 case OP_VCALL2_MEMBASE: {
1962 cinfo = get_call_info (cfg, cfg->mempool, ((MonoCallInst*)ins)->signature, FALSE);
1963 storage = cinfo->ret.storage;
1965 if (storage == ArgAggregate) {
1966 MonoInst *local = (MonoInst*)cfg->arch.ret_var_addr_local;
1968 /* Load address of stack space allocated for the return value */
1969 ia64_movl (code, GP_SCRATCH_REG, local->inst_offset);
1970 ia64_add (code, GP_SCRATCH_REG, GP_SCRATCH_REG, local->inst_basereg);
1971 ia64_ld8 (code, GP_SCRATCH_REG, GP_SCRATCH_REG);
1973 for (i = 0; i < cinfo->ret.nregs; ++i) {
1974 switch (cinfo->ret.atype) {
1975 case AggregateNormal:
1976 ia64_st8_inc_imm_hint (code, GP_SCRATCH_REG, cinfo->ret.reg + i, 8, 0);
1978 case AggregateSingleHFA:
1979 ia64_stfs_inc_imm_hint (code, GP_SCRATCH_REG, cinfo->ret.reg + i, 4, 0);
1981 case AggregateDoubleHFA:
1982 ia64_stfd_inc_imm_hint (code, GP_SCRATCH_REG, cinfo->ret.reg + i, 8, 0);
1985 g_assert_not_reached ();
1992 g_assert_not_reached ();
1998 #define add_patch_info(cfg,code,patch_type,data) do { \
1999 mono_add_patch_info (cfg, code.buf + code.nins - cfg->native_code, patch_type, data); \
2002 #define emit_cond_system_exception(cfg,code,exc_name,predicate) do { \
2003 MonoInst *tins = mono_branch_optimize_exception_target (cfg, bb, exc_name); \
2005 add_patch_info (cfg, code, MONO_PATCH_INFO_EXC, exc_name); \
2007 add_patch_info (cfg, code, MONO_PATCH_INFO_BB, tins->inst_true_bb); \
2008 ia64_br_cond_pred (code, (predicate), 0); \
2011 static Ia64CodegenState
2012 emit_call (MonoCompile *cfg, Ia64CodegenState code, guint32 patch_type, gconstpointer data)
2014 add_patch_info (cfg, code, patch_type, data);
2016 if ((patch_type == MONO_PATCH_INFO_ABS) || (patch_type == MONO_PATCH_INFO_INTERNAL_METHOD)) {
2018 /* mono_arch_patch_callsite will patch this */
2019 /* mono_arch_nullify_class_init_trampoline will patch this */
2020 ia64_movl (code, GP_SCRATCH_REG, 0);
2021 ia64_ld8_inc_imm (code, GP_SCRATCH_REG2, GP_SCRATCH_REG, 8);
2022 ia64_mov_to_br (code, IA64_B6, GP_SCRATCH_REG2);
2023 ia64_ld8 (code, IA64_GP, GP_SCRATCH_REG);
2024 ia64_br_call_reg (code, IA64_B0, IA64_B6);
2027 /* Can't use a direct call since the displacement might be too small */
2028 /* mono_arch_patch_callsite will patch this */
2029 ia64_movl (code, GP_SCRATCH_REG, 0);
2030 ia64_mov_to_br (code, IA64_B6, GP_SCRATCH_REG);
2031 ia64_br_call_reg (code, IA64_B0, IA64_B6);
2037 #define bb_is_loop_start(bb) ((bb)->loop_body_start && (bb)->nesting)
2040 mono_arch_output_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
2045 Ia64CodegenState code;
2046 guint8 *code_start = cfg->native_code + cfg->code_len;
2047 MonoInst *last_ins = NULL;
2048 guint last_offset = 0;
2051 if (cfg->opt & MONO_OPT_LOOP) {
2055 if (cfg->verbose_level > 2)
2056 g_print ("Basic block %d starting at offset 0x%x\n", bb->block_num, bb->native_offset);
2058 cpos = bb->max_offset;
2060 if (cfg->prof_options & MONO_PROFILE_COVERAGE) {
2064 offset = code_start - cfg->native_code;
2066 ia64_codegen_init (code, code_start);
2069 if (strstr (cfg->method->name, "conv_ovf_i1") && (bb->block_num == 2))
2073 MONO_BB_FOR_EACH_INS (bb, ins) {
2074 offset = code.buf - cfg->native_code;
2076 max_len = ((int)(((guint8 *)ins_get_spec (ins->opcode))[MONO_INST_LEN])) + 128;
2078 while (offset + max_len + 16 > cfg->code_size) {
2079 ia64_codegen_close (code);
2081 offset = code.buf - cfg->native_code;
2083 cfg->code_size *= 2;
2084 cfg->native_code = g_realloc (cfg->native_code, cfg->code_size);
2085 code_start = cfg->native_code + offset;
2086 cfg->stat_code_reallocs++;
2088 ia64_codegen_init (code, code_start);
2091 mono_debug_record_line_number (cfg, ins, offset);
2093 switch (ins->opcode) {
2096 if (ia64_is_imm14 (ins->inst_c0))
2097 ia64_adds_imm (code, ins->dreg, ins->inst_c0, IA64_R0);
2099 ia64_movl (code, ins->dreg, ins->inst_c0);
2102 add_patch_info (cfg, code, (MonoJumpInfoType)ins->inst_i1, ins->inst_p0);
2103 ia64_movl (code, ins->dreg, 0);
2106 ia64_mov (code, ins->dreg, ins->sreg1);
2109 case OP_IA64_BR_COND: {
2111 if (ins->opcode == OP_IA64_BR_COND)
2113 if (ins->inst_target_bb->native_offset) {
2114 guint8 *pos = code.buf + code.nins;
2116 ia64_br_cond_pred (code, pred, 0);
2117 ia64_begin_bundle (code);
2118 ia64_patch (pos, cfg->native_code + ins->inst_target_bb->native_offset);
2120 add_patch_info (cfg, code, MONO_PATCH_INFO_BB, ins->inst_target_bb);
2121 ia64_br_cond_pred (code, pred, 0);
2126 ia64_begin_bundle (code);
2127 ins->inst_c0 = code.buf - cfg->native_code;
2130 case OP_RELAXED_NOP:
2132 case OP_DUMMY_STORE:
2133 case OP_NOT_REACHED:
2137 ia64_mov_to_br (code, IA64_B6, ins->sreg1);
2138 ia64_br_cond_reg (code, IA64_B6);
2142 ia64_add (code, ins->dreg, ins->sreg1, ins->sreg2);
2146 ia64_sub (code, ins->dreg, ins->sreg1, ins->sreg2);
2150 ia64_and (code, ins->dreg, ins->sreg1, ins->sreg2);
2154 ia64_or (code, ins->dreg, ins->sreg1, ins->sreg2);
2158 ia64_xor (code, ins->dreg, ins->sreg1, ins->sreg2);
2162 ia64_sub (code, ins->dreg, IA64_R0, ins->sreg1);
2166 ia64_andcm_imm (code, ins->dreg, -1, ins->sreg1);
2170 ia64_shl (code, ins->dreg, ins->sreg1, ins->sreg2);
2173 ia64_sxt4 (code, GP_SCRATCH_REG, ins->sreg1);
2174 ia64_shr (code, ins->dreg, GP_SCRATCH_REG, ins->sreg2);
2177 ia64_shr (code, ins->dreg, ins->sreg1, ins->sreg2);
2180 ia64_zxt4 (code, GP_SCRATCH_REG, ins->sreg1);
2181 ia64_shr_u (code, ins->dreg, GP_SCRATCH_REG, ins->sreg2);
2184 ia64_shr_u (code, ins->dreg, ins->sreg1, ins->sreg2);
2187 /* p6 and p7 is set if there is signed/unsigned overflow */
2189 /* Set p8-p9 == (sreg2 > 0) */
2190 ia64_cmp4_lt (code, 8, 9, IA64_R0, ins->sreg2);
2192 ia64_add (code, GP_SCRATCH_REG, ins->sreg1, ins->sreg2);
2194 /* (sreg2 > 0) && (res < ins->sreg1) => signed overflow */
2195 ia64_cmp4_lt_pred (code, 8, 6, 10, GP_SCRATCH_REG, ins->sreg1);
2196 /* (sreg2 <= 0) && (res > ins->sreg1) => signed overflow */
2197 ia64_cmp4_lt_pred (code, 9, 6, 10, ins->sreg1, GP_SCRATCH_REG);
2199 /* res <u sreg1 => unsigned overflow */
2200 ia64_cmp4_ltu (code, 7, 10, GP_SCRATCH_REG, ins->sreg1);
2202 /* FIXME: Predicate this since this is a side effect */
2203 ia64_mov (code, ins->dreg, GP_SCRATCH_REG);
2206 /* p6 and p7 is set if there is signed/unsigned overflow */
2208 /* Set p8-p9 == (sreg2 > 0) */
2209 ia64_cmp4_lt (code, 8, 9, IA64_R0, ins->sreg2);
2211 ia64_sub (code, GP_SCRATCH_REG, ins->sreg1, ins->sreg2);
2213 /* (sreg2 > 0) && (res > ins->sreg1) => signed overflow */
2214 ia64_cmp4_gt_pred (code, 8, 6, 10, GP_SCRATCH_REG, ins->sreg1);
2215 /* (sreg2 <= 0) && (res < ins->sreg1) => signed overflow */
2216 ia64_cmp4_lt_pred (code, 9, 6, 10, GP_SCRATCH_REG, ins->sreg1);
2218 /* sreg1 <u sreg2 => unsigned overflow */
2219 ia64_cmp4_ltu (code, 7, 10, ins->sreg1, ins->sreg2);
2221 /* FIXME: Predicate this since this is a side effect */
2222 ia64_mov (code, ins->dreg, GP_SCRATCH_REG);
2225 /* Same as OP_IADDCC */
2226 ia64_cmp_lt (code, 8, 9, IA64_R0, ins->sreg2);
2228 ia64_add (code, GP_SCRATCH_REG, ins->sreg1, ins->sreg2);
2230 ia64_cmp_lt_pred (code, 8, 6, 10, GP_SCRATCH_REG, ins->sreg1);
2231 ia64_cmp_lt_pred (code, 9, 6, 10, ins->sreg1, GP_SCRATCH_REG);
2233 ia64_cmp_ltu (code, 7, 10, GP_SCRATCH_REG, ins->sreg1);
2235 ia64_mov (code, ins->dreg, GP_SCRATCH_REG);
2238 /* Same as OP_ISUBCC */
2240 ia64_cmp_lt (code, 8, 9, IA64_R0, ins->sreg2);
2242 ia64_sub (code, GP_SCRATCH_REG, ins->sreg1, ins->sreg2);
2244 ia64_cmp_gt_pred (code, 8, 6, 10, GP_SCRATCH_REG, ins->sreg1);
2245 ia64_cmp_lt_pred (code, 9, 6, 10, GP_SCRATCH_REG, ins->sreg1);
2247 ia64_cmp_ltu (code, 7, 10, ins->sreg1, ins->sreg2);
2249 ia64_mov (code, ins->dreg, GP_SCRATCH_REG);
2254 ia64_adds_imm (code, ins->dreg, ins->inst_imm, ins->sreg1);
2259 ia64_and_imm (code, ins->dreg, ins->inst_imm, ins->sreg1);
2263 ia64_or_imm (code, ins->dreg, ins->inst_imm, ins->sreg1);
2267 ia64_xor_imm (code, ins->dreg, ins->inst_imm, ins->sreg1);
2272 ia64_shl_imm (code, ins->dreg, ins->sreg1, ins->inst_imm);
2276 ia64_shr_imm (code, ins->dreg, ins->sreg1, ins->inst_imm);
2279 g_assert (ins->inst_imm <= 64);
2280 ia64_extr (code, ins->dreg, ins->sreg1, ins->inst_imm, 32 - ins->inst_imm);
2282 case OP_ISHR_UN_IMM:
2283 ia64_zxt4 (code, GP_SCRATCH_REG, ins->sreg1);
2284 ia64_shr_u_imm (code, ins->dreg, GP_SCRATCH_REG, ins->inst_imm);
2286 case OP_LSHR_UN_IMM:
2287 ia64_shr_u_imm (code, ins->dreg, ins->sreg1, ins->inst_imm);
2290 /* Based on gcc code */
2291 ia64_setf_sig (code, FP_SCRATCH_REG, ins->sreg1);
2292 ia64_setf_sig (code, FP_SCRATCH_REG2, ins->sreg2);
2293 ia64_xmpy_l (code, FP_SCRATCH_REG, FP_SCRATCH_REG, FP_SCRATCH_REG2);
2294 ia64_getf_sig (code, ins->dreg, FP_SCRATCH_REG);
2297 case OP_STOREI1_MEMBASE_REG:
2298 ia64_st1_hint (code, ins->inst_destbasereg, ins->sreg1, 0);
2300 case OP_STOREI2_MEMBASE_REG:
2301 ia64_st2_hint (code, ins->inst_destbasereg, ins->sreg1, 0);
2303 case OP_STOREI4_MEMBASE_REG:
2304 ia64_st4_hint (code, ins->inst_destbasereg, ins->sreg1, 0);
2306 case OP_STOREI8_MEMBASE_REG:
2307 case OP_STORE_MEMBASE_REG:
2308 if (ins->inst_offset != 0) {
2309 /* This is generated by local regalloc */
2310 if (ia64_is_imm14 (ins->inst_offset)) {
2311 ia64_adds_imm (code, GP_SCRATCH_REG, ins->inst_offset, ins->inst_destbasereg);
2313 ia64_movl (code, GP_SCRATCH_REG, ins->inst_offset);
2314 ia64_add (code, GP_SCRATCH_REG, GP_SCRATCH_REG, ins->inst_destbasereg);
2316 ins->inst_destbasereg = GP_SCRATCH_REG;
2318 ia64_st8_hint (code, ins->inst_destbasereg, ins->sreg1, 0);
2321 case OP_IA64_STOREI1_MEMBASE_INC_REG:
2322 ia64_st1_inc_imm_hint (code, ins->inst_destbasereg, ins->sreg1, 1, 0);
2324 case OP_IA64_STOREI2_MEMBASE_INC_REG:
2325 ia64_st2_inc_imm_hint (code, ins->inst_destbasereg, ins->sreg1, 2, 0);
2327 case OP_IA64_STOREI4_MEMBASE_INC_REG:
2328 ia64_st4_inc_imm_hint (code, ins->inst_destbasereg, ins->sreg1, 4, 0);
2330 case OP_IA64_STOREI8_MEMBASE_INC_REG:
2331 ia64_st8_inc_imm_hint (code, ins->inst_destbasereg, ins->sreg1, 8, 0);
2334 case OP_LOADU1_MEMBASE:
2335 ia64_ld1 (code, ins->dreg, ins->inst_basereg);
2337 case OP_LOADU2_MEMBASE:
2338 ia64_ld2 (code, ins->dreg, ins->inst_basereg);
2340 case OP_LOADU4_MEMBASE:
2341 ia64_ld4 (code, ins->dreg, ins->inst_basereg);
2343 case OP_LOADI1_MEMBASE:
2344 ia64_ld1 (code, ins->dreg, ins->inst_basereg);
2345 ia64_sxt1 (code, ins->dreg, ins->dreg);
2347 case OP_LOADI2_MEMBASE:
2348 ia64_ld2 (code, ins->dreg, ins->inst_basereg);
2349 ia64_sxt2 (code, ins->dreg, ins->dreg);
2351 case OP_LOADI4_MEMBASE:
2352 ia64_ld4 (code, ins->dreg, ins->inst_basereg);
2353 ia64_sxt4 (code, ins->dreg, ins->dreg);
2355 case OP_LOAD_MEMBASE:
2356 case OP_LOADI8_MEMBASE:
2357 if (ins->inst_offset != 0) {
2358 /* This is generated by local regalloc */
2359 if (ia64_is_imm14 (ins->inst_offset)) {
2360 ia64_adds_imm (code, GP_SCRATCH_REG, ins->inst_offset, ins->inst_basereg);
2362 ia64_movl (code, GP_SCRATCH_REG, ins->inst_offset);
2363 ia64_add (code, GP_SCRATCH_REG, GP_SCRATCH_REG, ins->inst_basereg);
2365 ins->inst_basereg = GP_SCRATCH_REG;
2367 ia64_ld8 (code, ins->dreg, ins->inst_basereg);
2370 case OP_IA64_LOADU1_MEMBASE_INC:
2371 ia64_ld1_inc_imm_hint (code, ins->dreg, ins->inst_basereg, 1, 0);
2373 case OP_IA64_LOADU2_MEMBASE_INC:
2374 ia64_ld2_inc_imm_hint (code, ins->dreg, ins->inst_basereg, 2, 0);
2376 case OP_IA64_LOADU4_MEMBASE_INC:
2377 ia64_ld4_inc_imm_hint (code, ins->dreg, ins->inst_basereg, 4, 0);
2379 case OP_IA64_LOADI8_MEMBASE_INC:
2380 ia64_ld8_inc_imm_hint (code, ins->dreg, ins->inst_basereg, 8, 0);
2384 ia64_sxt1 (code, ins->dreg, ins->sreg1);
2387 ia64_sxt2 (code, ins->dreg, ins->sreg1);
2390 ia64_sxt4 (code, ins->dreg, ins->sreg1);
2393 ia64_zxt1 (code, ins->dreg, ins->sreg1);
2396 ia64_zxt2 (code, ins->dreg, ins->sreg1);
2399 ia64_zxt4 (code, ins->dreg, ins->sreg1);
2402 /* Compare opcodes */
2403 case OP_IA64_CMP4_EQ:
2404 ia64_cmp4_eq (code, 6, 7, ins->sreg1, ins->sreg2);
2406 case OP_IA64_CMP4_NE:
2407 ia64_cmp4_ne (code, 6, 7, ins->sreg1, ins->sreg2);
2409 case OP_IA64_CMP4_LE:
2410 ia64_cmp4_le (code, 6, 7, ins->sreg1, ins->sreg2);
2412 case OP_IA64_CMP4_LT:
2413 ia64_cmp4_lt (code, 6, 7, ins->sreg1, ins->sreg2);
2415 case OP_IA64_CMP4_GE:
2416 ia64_cmp4_ge (code, 6, 7, ins->sreg1, ins->sreg2);
2418 case OP_IA64_CMP4_GT:
2419 ia64_cmp4_gt (code, 6, 7, ins->sreg1, ins->sreg2);
2421 case OP_IA64_CMP4_LT_UN:
2422 ia64_cmp4_ltu (code, 6, 7, ins->sreg1, ins->sreg2);
2424 case OP_IA64_CMP4_LE_UN:
2425 ia64_cmp4_leu (code, 6, 7, ins->sreg1, ins->sreg2);
2427 case OP_IA64_CMP4_GT_UN:
2428 ia64_cmp4_gtu (code, 6, 7, ins->sreg1, ins->sreg2);
2430 case OP_IA64_CMP4_GE_UN:
2431 ia64_cmp4_geu (code, 6, 7, ins->sreg1, ins->sreg2);
2433 case OP_IA64_CMP_EQ:
2434 ia64_cmp_eq (code, 6, 7, ins->sreg1, ins->sreg2);
2436 case OP_IA64_CMP_NE:
2437 ia64_cmp_ne (code, 6, 7, ins->sreg1, ins->sreg2);
2439 case OP_IA64_CMP_LE:
2440 ia64_cmp_le (code, 6, 7, ins->sreg1, ins->sreg2);
2442 case OP_IA64_CMP_LT:
2443 ia64_cmp_lt (code, 6, 7, ins->sreg1, ins->sreg2);
2445 case OP_IA64_CMP_GE:
2446 ia64_cmp_ge (code, 6, 7, ins->sreg1, ins->sreg2);
2448 case OP_IA64_CMP_GT:
2449 ia64_cmp_gt (code, 6, 7, ins->sreg1, ins->sreg2);
2451 case OP_IA64_CMP_GT_UN:
2452 ia64_cmp_gtu (code, 6, 7, ins->sreg1, ins->sreg2);
2454 case OP_IA64_CMP_LT_UN:
2455 ia64_cmp_ltu (code, 6, 7, ins->sreg1, ins->sreg2);
2457 case OP_IA64_CMP_GE_UN:
2458 ia64_cmp_geu (code, 6, 7, ins->sreg1, ins->sreg2);
2460 case OP_IA64_CMP_LE_UN:
2461 ia64_cmp_leu (code, 6, 7, ins->sreg1, ins->sreg2);
2463 case OP_IA64_CMP4_EQ_IMM:
2464 ia64_cmp4_eq_imm (code, 6, 7, ins->inst_imm, ins->sreg2);
2466 case OP_IA64_CMP4_NE_IMM:
2467 ia64_cmp4_ne_imm (code, 6, 7, ins->inst_imm, ins->sreg2);
2469 case OP_IA64_CMP4_LE_IMM:
2470 ia64_cmp4_le_imm (code, 6, 7, ins->inst_imm, ins->sreg2);
2472 case OP_IA64_CMP4_LT_IMM:
2473 ia64_cmp4_lt_imm (code, 6, 7, ins->inst_imm, ins->sreg2);
2475 case OP_IA64_CMP4_GE_IMM:
2476 ia64_cmp4_ge_imm (code, 6, 7, ins->inst_imm, ins->sreg2);
2478 case OP_IA64_CMP4_GT_IMM:
2479 ia64_cmp4_gt_imm (code, 6, 7, ins->inst_imm, ins->sreg2);
2481 case OP_IA64_CMP4_LT_UN_IMM:
2482 ia64_cmp4_ltu_imm (code, 6, 7, ins->inst_imm, ins->sreg2);
2484 case OP_IA64_CMP4_LE_UN_IMM:
2485 ia64_cmp4_leu_imm (code, 6, 7, ins->inst_imm, ins->sreg2);
2487 case OP_IA64_CMP4_GT_UN_IMM:
2488 ia64_cmp4_gtu_imm (code, 6, 7, ins->inst_imm, ins->sreg2);
2490 case OP_IA64_CMP4_GE_UN_IMM:
2491 ia64_cmp4_geu_imm (code, 6, 7, ins->inst_imm, ins->sreg2);
2493 case OP_IA64_CMP_EQ_IMM:
2494 ia64_cmp_eq_imm (code, 6, 7, ins->inst_imm, ins->sreg2);
2496 case OP_IA64_CMP_NE_IMM:
2497 ia64_cmp_ne_imm (code, 6, 7, ins->inst_imm, ins->sreg2);
2499 case OP_IA64_CMP_LE_IMM:
2500 ia64_cmp_le_imm (code, 6, 7, ins->inst_imm, ins->sreg2);
2502 case OP_IA64_CMP_LT_IMM:
2503 ia64_cmp_lt_imm (code, 6, 7, ins->inst_imm, ins->sreg2);
2505 case OP_IA64_CMP_GE_IMM:
2506 ia64_cmp_ge_imm (code, 6, 7, ins->inst_imm, ins->sreg2);
2508 case OP_IA64_CMP_GT_IMM:
2509 ia64_cmp_gt_imm (code, 6, 7, ins->inst_imm, ins->sreg2);
2511 case OP_IA64_CMP_GT_UN_IMM:
2512 ia64_cmp_gtu_imm (code, 6, 7, ins->inst_imm, ins->sreg2);
2514 case OP_IA64_CMP_LT_UN_IMM:
2515 ia64_cmp_ltu_imm (code, 6, 7, ins->inst_imm, ins->sreg2);
2517 case OP_IA64_CMP_GE_UN_IMM:
2518 ia64_cmp_geu_imm (code, 6, 7, ins->inst_imm, ins->sreg2);
2520 case OP_IA64_CMP_LE_UN_IMM:
2521 ia64_cmp_leu_imm (code, 6, 7, ins->inst_imm, ins->sreg2);
2523 case OP_IA64_FCMP_EQ:
2524 ia64_fcmp_eq_sf (code, 6, 7, ins->sreg1, ins->sreg2, 0);
2526 case OP_IA64_FCMP_NE:
2527 ia64_fcmp_ne_sf (code, 6, 7, ins->sreg1, ins->sreg2, 0);
2529 case OP_IA64_FCMP_LT:
2530 ia64_fcmp_lt_sf (code, 6, 7, ins->sreg1, ins->sreg2, 0);
2532 case OP_IA64_FCMP_GT:
2533 ia64_fcmp_gt_sf (code, 6, 7, ins->sreg1, ins->sreg2, 0);
2535 case OP_IA64_FCMP_LE:
2536 ia64_fcmp_le_sf (code, 6, 7, ins->sreg1, ins->sreg2, 0);
2538 case OP_IA64_FCMP_GE:
2539 ia64_fcmp_ge_sf (code, 6, 7, ins->sreg1, ins->sreg2, 0);
2541 case OP_IA64_FCMP_GT_UN:
2542 ia64_fcmp_gt_sf (code, 6, 7, ins->sreg1, ins->sreg2, 0);
2543 ia64_fcmp_unord_sf_pred (code, 7, 6, 7, ins->sreg1, ins->sreg2, 0);
2545 case OP_IA64_FCMP_LT_UN:
2546 ia64_fcmp_lt_sf (code, 6, 7, ins->sreg1, ins->sreg2, 0);
2547 ia64_fcmp_unord_sf_pred (code, 7, 6, 7, ins->sreg1, ins->sreg2, 0);
2549 case OP_IA64_FCMP_GE_UN:
2550 ia64_fcmp_ge_sf (code, 6, 7, ins->sreg1, ins->sreg2, 0);
2551 ia64_fcmp_unord_sf_pred (code, 7, 6, 7, ins->sreg1, ins->sreg2, 0);
2553 case OP_IA64_FCMP_LE_UN:
2554 ia64_fcmp_le_sf (code, 6, 7, ins->sreg1, ins->sreg2, 0);
2555 ia64_fcmp_unord_sf_pred (code, 7, 6, 7, ins->sreg1, ins->sreg2, 0);
2558 case OP_COND_EXC_IOV:
2559 case OP_COND_EXC_OV:
2560 emit_cond_system_exception (cfg, code, "OverflowException", 6);
2562 case OP_COND_EXC_IC:
2564 emit_cond_system_exception (cfg, code, "OverflowException", 7);
2566 case OP_IA64_COND_EXC:
2567 emit_cond_system_exception (cfg, code, ins->inst_p1, 6);
2570 ia64_mov_pred (code, 7, ins->dreg, IA64_R0);
2571 ia64_no_stop (code);
2572 ia64_add1_pred (code, 6, ins->dreg, IA64_R0, IA64_R0);
2574 case OP_ICONV_TO_I1:
2575 case OP_LCONV_TO_I1:
2576 /* FIXME: Is this needed ? */
2577 ia64_sxt1 (code, ins->dreg, ins->sreg1);
2579 case OP_ICONV_TO_I2:
2580 case OP_LCONV_TO_I2:
2581 /* FIXME: Is this needed ? */
2582 ia64_sxt2 (code, ins->dreg, ins->sreg1);
2584 case OP_LCONV_TO_I4:
2585 /* FIXME: Is this needed ? */
2586 ia64_sxt4 (code, ins->dreg, ins->sreg1);
2588 case OP_ICONV_TO_U1:
2589 case OP_LCONV_TO_U1:
2590 /* FIXME: Is this needed */
2591 ia64_zxt1 (code, ins->dreg, ins->sreg1);
2593 case OP_ICONV_TO_U2:
2594 case OP_LCONV_TO_U2:
2595 /* FIXME: Is this needed */
2596 ia64_zxt2 (code, ins->dreg, ins->sreg1);
2598 case OP_LCONV_TO_U4:
2599 /* FIXME: Is this needed */
2600 ia64_zxt4 (code, ins->dreg, ins->sreg1);
2602 case OP_ICONV_TO_I8:
2604 case OP_LCONV_TO_I8:
2606 ia64_sxt4 (code, ins->dreg, ins->sreg1);
2608 case OP_LCONV_TO_U8:
2610 ia64_zxt4 (code, ins->dreg, ins->sreg1);
2617 double d = *(double *)ins->inst_p0;
2619 if ((d == 0.0) && (mono_signbit (d) == 0))
2620 ia64_fmov (code, ins->dreg, 0);
2622 ia64_fmov (code, ins->dreg, 1);
2624 add_patch_info (cfg, code, MONO_PATCH_INFO_R8, ins->inst_p0);
2625 ia64_movl (code, GP_SCRATCH_REG, 0);
2626 ia64_ldfd (code, ins->dreg, GP_SCRATCH_REG);
2631 float f = *(float *)ins->inst_p0;
2633 if ((f == 0.0) && (mono_signbit (f) == 0))
2634 ia64_fmov (code, ins->dreg, 0);
2636 ia64_fmov (code, ins->dreg, 1);
2638 add_patch_info (cfg, code, MONO_PATCH_INFO_R4, ins->inst_p0);
2639 ia64_movl (code, GP_SCRATCH_REG, 0);
2640 ia64_ldfs (code, ins->dreg, GP_SCRATCH_REG);
2645 ia64_fmov (code, ins->dreg, ins->sreg1);
2647 case OP_STORER8_MEMBASE_REG:
2648 if (ins->inst_offset != 0) {
2649 /* This is generated by local regalloc */
2650 if (ia64_is_imm14 (ins->inst_offset)) {
2651 ia64_adds_imm (code, GP_SCRATCH_REG, ins->inst_offset, ins->inst_destbasereg);
2653 ia64_movl (code, GP_SCRATCH_REG, ins->inst_offset);
2654 ia64_add (code, GP_SCRATCH_REG, GP_SCRATCH_REG, ins->inst_destbasereg);
2656 ins->inst_destbasereg = GP_SCRATCH_REG;
2658 ia64_stfd_hint (code, ins->inst_destbasereg, ins->sreg1, 0);
2660 case OP_STORER4_MEMBASE_REG:
2661 ia64_fnorm_s_sf (code, FP_SCRATCH_REG, ins->sreg1, 0);
2662 ia64_stfs_hint (code, ins->inst_destbasereg, FP_SCRATCH_REG, 0);
2664 case OP_LOADR8_MEMBASE:
2665 if (ins->inst_offset != 0) {
2666 /* This is generated by local regalloc */
2667 if (ia64_is_imm14 (ins->inst_offset)) {
2668 ia64_adds_imm (code, GP_SCRATCH_REG, ins->inst_offset, ins->inst_basereg);
2670 ia64_movl (code, GP_SCRATCH_REG, ins->inst_offset);
2671 ia64_add (code, GP_SCRATCH_REG, GP_SCRATCH_REG, ins->inst_basereg);
2673 ins->inst_basereg = GP_SCRATCH_REG;
2675 ia64_ldfd (code, ins->dreg, ins->inst_basereg);
2677 case OP_LOADR4_MEMBASE:
2678 ia64_ldfs (code, ins->dreg, ins->inst_basereg);
2679 ia64_fnorm_d_sf (code, ins->dreg, ins->dreg, 0);
2681 case OP_ICONV_TO_R4:
2682 case OP_LCONV_TO_R4:
2683 ia64_setf_sig (code, ins->dreg, ins->sreg1);
2684 ia64_fcvt_xf (code, ins->dreg, ins->dreg);
2685 ia64_fnorm_s_sf (code, ins->dreg, ins->dreg, 0);
2687 case OP_ICONV_TO_R8:
2688 case OP_LCONV_TO_R8:
2689 ia64_setf_sig (code, ins->dreg, ins->sreg1);
2690 ia64_fcvt_xf (code, ins->dreg, ins->dreg);
2691 ia64_fnorm_d_sf (code, ins->dreg, ins->dreg, 0);
2693 case OP_FCONV_TO_R4:
2694 ia64_fnorm_s_sf (code, ins->dreg, ins->sreg1, 0);
2696 case OP_FCONV_TO_I8:
2698 ia64_fcvt_fx_trunc_sf (code, FP_SCRATCH_REG, ins->sreg1, 0);
2699 ia64_getf_sig (code, ins->dreg, FP_SCRATCH_REG);
2702 ia64_fma_d_sf (code, ins->dreg, ins->sreg1, 1, ins->sreg2, 0);
2705 ia64_fms_d_sf (code, ins->dreg, ins->sreg1, 1, ins->sreg2, 0);
2708 ia64_fma_d_sf (code, ins->dreg, ins->sreg1, ins->sreg2, 0, 0);
2711 ia64_fmerge_ns (code, ins->dreg, ins->sreg1, ins->sreg1);
2715 ia64_fclass_m (code, 6, 7, ins->sreg1, 0x080);
2716 emit_cond_system_exception (cfg, code, "ArithmeticException", 6);
2718 ia64_fclass_m (code, 6, 7, ins->sreg1, 0x040);
2719 emit_cond_system_exception (cfg, code, "ArithmeticException", 6);
2720 /* Positive infinity */
2721 ia64_fclass_m (code, 6, 7, ins->sreg1, 0x021);
2722 emit_cond_system_exception (cfg, code, "ArithmeticException", 6);
2723 /* Negative infinity */
2724 ia64_fclass_m (code, 6, 7, ins->sreg1, 0x022);
2725 emit_cond_system_exception (cfg, code, "ArithmeticException", 6);
2730 /* ensure ins->sreg1 is not NULL */
2731 /* Can't use ld8 as this could be a vtype address */
2732 ia64_ld1 (code, GP_SCRATCH_REG, ins->sreg1);
2735 ia64_adds_imm (code, GP_SCRATCH_REG, cfg->sig_cookie, cfg->frame_reg);
2736 ia64_st8 (code, ins->sreg1, GP_SCRATCH_REG);
2744 call = (MonoCallInst*)ins;
2746 if (ins->flags & MONO_INST_HAS_METHOD)
2747 code = emit_call (cfg, code, MONO_PATCH_INFO_METHOD, call->method);
2749 code = emit_call (cfg, code, MONO_PATCH_INFO_ABS, call->fptr);
2751 code = emit_move_return_value (cfg, ins, code);
2759 case OP_VOIDCALL_REG: {
2760 MonoCallInst *call = (MonoCallInst*)ins;
2765 * mono_arch_get_this_arg_from_call () needs to find the this argument in a global
2768 cinfo = get_call_info (cfg, cfg->mempool, call->signature, FALSE);
2769 out_reg = cfg->arch.reg_out0;
2770 ia64_mov (code, IA64_R10, out_reg);
2773 ia64_mov (code, IA64_R8, ins->sreg1);
2774 ia64_ld8_inc_imm (code, GP_SCRATCH_REG2, IA64_R8, 8);
2775 ia64_mov_to_br (code, IA64_B6, GP_SCRATCH_REG2);
2776 ia64_ld8 (code, IA64_GP, IA64_R8);
2777 ia64_br_call_reg (code, IA64_B0, IA64_B6);
2779 code = emit_move_return_value (cfg, ins, code);
2782 case OP_FCALL_MEMBASE:
2783 case OP_LCALL_MEMBASE:
2784 case OP_VCALL_MEMBASE:
2785 case OP_VCALL2_MEMBASE:
2786 case OP_VOIDCALL_MEMBASE:
2787 case OP_CALL_MEMBASE: {
2788 MonoCallInst *call = (MonoCallInst*)ins;
2792 ia64_mov (code, IA64_R11, ins->sreg1);
2793 if (ia64_is_imm14 (ins->inst_offset))
2794 ia64_adds_imm (code, IA64_R8, ins->inst_offset, ins->sreg1);
2796 ia64_movl (code, GP_SCRATCH_REG, ins->inst_offset);
2797 ia64_add (code, IA64_R8, GP_SCRATCH_REG, ins->sreg1);
2800 if (call->method && ins->inst_offset < 0) {
2802 * This is a possible IMT call so save the IMT method in a global
2803 * register where mono_arch_find_imt_method () and its friends can
2806 ia64_movl (code, IA64_R9, call->method);
2810 * mono_arch_find_this_arg () needs to find the this argument in a global
2813 cinfo = get_call_info (cfg, cfg->mempool, call->signature, FALSE);
2814 out_reg = cfg->arch.reg_out0;
2815 ia64_mov (code, IA64_R10, out_reg);
2817 ia64_ld8 (code, GP_SCRATCH_REG, IA64_R8);
2819 ia64_mov_to_br (code, IA64_B6, GP_SCRATCH_REG);
2821 ia64_br_call_reg (code, IA64_B0, IA64_B6);
2823 code = emit_move_return_value (cfg, ins, code);
2828 * Keep in sync with the code in emit_epilog.
2831 if (cfg->prof_options & MONO_PROFILE_ENTER_LEAVE)
2834 g_assert (!cfg->method->save_lmf);
2836 /* Load arguments into their original registers */
2837 code = emit_load_volatile_arguments (cfg, code);
2839 if (cfg->arch.stack_alloc_size) {
2840 if (cfg->arch.omit_fp) {
2841 if (ia64_is_imm14 (cfg->arch.stack_alloc_size))
2842 ia64_adds_imm (code, IA64_SP, (cfg->arch.stack_alloc_size), IA64_SP);
2844 ia64_movl (code, GP_SCRATCH_REG, cfg->arch.stack_alloc_size);
2845 ia64_add (code, IA64_SP, GP_SCRATCH_REG, IA64_SP);
2849 ia64_mov (code, IA64_SP, cfg->arch.reg_saved_sp);
2851 ia64_mov_to_ar_i (code, IA64_PFS, cfg->arch.reg_saved_ar_pfs);
2852 ia64_mov_ret_to_br (code, IA64_B0, cfg->arch.reg_saved_b0);
2854 add_patch_info (cfg, code, MONO_PATCH_INFO_METHOD_JUMP, ins->inst_p0);
2855 ia64_movl (code, GP_SCRATCH_REG, 0);
2856 ia64_mov_to_br (code, IA64_B6, GP_SCRATCH_REG);
2857 ia64_br_cond_reg (code, IA64_B6);
2862 code = emit_call (cfg, code, MONO_PATCH_INFO_ABS, mono_break);
2868 /* FIXME: Sigaltstack support */
2870 /* keep alignment */
2871 ia64_adds_imm (code, GP_SCRATCH_REG, MONO_ARCH_LOCALLOC_ALIGNMENT - 1, ins->sreg1);
2872 ia64_movl (code, GP_SCRATCH_REG2, ~(MONO_ARCH_LOCALLOC_ALIGNMENT - 1));
2873 ia64_and (code, GP_SCRATCH_REG, GP_SCRATCH_REG, GP_SCRATCH_REG2);
2875 ia64_sub (code, IA64_SP, IA64_SP, GP_SCRATCH_REG);
2877 ia64_mov (code, ins->dreg, IA64_SP);
2879 /* An area at sp is reserved by the ABI for parameter passing */
2880 abi_offset = - ALIGN_TO (cfg->param_area + 16, MONO_ARCH_LOCALLOC_ALIGNMENT);
2881 if (ia64_is_adds_imm (abi_offset))
2882 ia64_adds_imm (code, IA64_SP, abi_offset, IA64_SP);
2884 ia64_movl (code, GP_SCRATCH_REG2, abi_offset);
2885 ia64_add (code, IA64_SP, IA64_SP, GP_SCRATCH_REG2);
2888 if (ins->flags & MONO_INST_INIT) {
2890 ia64_add (code, GP_SCRATCH_REG2, ins->dreg, GP_SCRATCH_REG);
2892 ia64_codegen_set_one_ins_per_bundle (code, TRUE);
2895 ia64_st8_inc_imm_hint (code, ins->dreg, IA64_R0, 8, 0);
2896 ia64_cmp_lt (code, 8, 9, ins->dreg, GP_SCRATCH_REG2);
2897 ia64_br_cond_pred (code, 8, -2);
2899 ia64_codegen_set_one_ins_per_bundle (code, FALSE);
2901 ia64_sub (code, ins->dreg, GP_SCRATCH_REG2, GP_SCRATCH_REG);
2906 case OP_LOCALLOC_IMM: {
2909 /* FIXME: Sigaltstack support */
2911 gssize size = ins->inst_imm;
2912 size = (size + (MONO_ARCH_FRAME_ALIGNMENT - 1)) & ~ (MONO_ARCH_FRAME_ALIGNMENT - 1);
2914 if (ia64_is_adds_imm (size))
2915 ia64_adds_imm (code, GP_SCRATCH_REG, size, IA64_R0);
2917 ia64_movl (code, GP_SCRATCH_REG, size);
2919 ia64_sub (code, IA64_SP, IA64_SP, GP_SCRATCH_REG);
2920 ia64_mov (code, ins->dreg, IA64_SP);
2922 /* An area at sp is reserved by the ABI for parameter passing */
2923 abi_offset = - ALIGN_TO (cfg->param_area + 16, MONO_ARCH_FRAME_ALIGNMENT);
2924 if (ia64_is_adds_imm (abi_offset))
2925 ia64_adds_imm (code, IA64_SP, abi_offset, IA64_SP);
2927 ia64_movl (code, GP_SCRATCH_REG2, abi_offset);
2928 ia64_add (code, IA64_SP, IA64_SP, GP_SCRATCH_REG2);
2931 if (ins->flags & MONO_INST_INIT) {
2933 ia64_add (code, GP_SCRATCH_REG2, ins->dreg, GP_SCRATCH_REG);
2935 ia64_codegen_set_one_ins_per_bundle (code, TRUE);
2938 ia64_st8_inc_imm_hint (code, ins->dreg, IA64_R0, 8, 0);
2939 ia64_cmp_lt (code, 8, 9, ins->dreg, GP_SCRATCH_REG2);
2940 ia64_br_cond_pred (code, 8, -2);
2942 ia64_codegen_set_one_ins_per_bundle (code, FALSE);
2944 ia64_sub (code, ins->dreg, GP_SCRATCH_REG2, GP_SCRATCH_REG);
2950 ia64_adds_imm (code, ins->dreg, ins->inst_offset, IA64_TP);
2951 ia64_ld8 (code, ins->dreg, ins->dreg);
2954 /* Synchronization */
2955 case OP_MEMORY_BARRIER:
2958 case OP_ATOMIC_ADD_IMM_NEW_I4:
2959 g_assert (ins->inst_offset == 0);
2960 ia64_fetchadd4_acq_hint (code, ins->dreg, ins->inst_basereg, ins->inst_imm, 0);
2961 ia64_adds_imm (code, ins->dreg, ins->inst_imm, ins->dreg);
2963 case OP_ATOMIC_ADD_IMM_NEW_I8:
2964 g_assert (ins->inst_offset == 0);
2965 ia64_fetchadd8_acq_hint (code, ins->dreg, ins->inst_basereg, ins->inst_imm, 0);
2966 ia64_adds_imm (code, ins->dreg, ins->inst_imm, ins->dreg);
2968 case OP_ATOMIC_EXCHANGE_I4:
2969 ia64_xchg4_hint (code, ins->dreg, ins->inst_basereg, ins->sreg2, 0);
2970 ia64_sxt4 (code, ins->dreg, ins->dreg);
2972 case OP_ATOMIC_EXCHANGE_I8:
2973 ia64_xchg8_hint (code, ins->dreg, ins->inst_basereg, ins->sreg2, 0);
2975 case OP_ATOMIC_ADD_NEW_I4: {
2976 guint8 *label, *buf;
2978 /* From libatomic_ops */
2981 ia64_begin_bundle (code);
2982 label = code.buf + code.nins;
2983 ia64_ld4_acq (code, GP_SCRATCH_REG, ins->sreg1);
2984 ia64_add (code, GP_SCRATCH_REG2, GP_SCRATCH_REG, ins->sreg2);
2985 ia64_mov_to_ar_m (code, IA64_CCV, GP_SCRATCH_REG);
2986 ia64_cmpxchg4_acq_hint (code, GP_SCRATCH_REG2, ins->sreg1, GP_SCRATCH_REG2, 0);
2987 ia64_cmp4_eq (code, 6, 7, GP_SCRATCH_REG, GP_SCRATCH_REG2);
2988 buf = code.buf + code.nins;
2989 ia64_br_cond_pred (code, 7, 0);
2990 ia64_begin_bundle (code);
2991 ia64_patch (buf, label);
2992 ia64_add (code, ins->dreg, GP_SCRATCH_REG, ins->sreg2);
2995 case OP_ATOMIC_ADD_NEW_I8: {
2996 guint8 *label, *buf;
2998 /* From libatomic_ops */
3001 ia64_begin_bundle (code);
3002 label = code.buf + code.nins;
3003 ia64_ld8_acq (code, GP_SCRATCH_REG, ins->sreg1);
3004 ia64_add (code, GP_SCRATCH_REG2, GP_SCRATCH_REG, ins->sreg2);
3005 ia64_mov_to_ar_m (code, IA64_CCV, GP_SCRATCH_REG);
3006 ia64_cmpxchg8_acq_hint (code, GP_SCRATCH_REG2, ins->sreg1, GP_SCRATCH_REG2, 0);
3007 ia64_cmp_eq (code, 6, 7, GP_SCRATCH_REG, GP_SCRATCH_REG2);
3008 buf = code.buf + code.nins;
3009 ia64_br_cond_pred (code, 7, 0);
3010 ia64_begin_bundle (code);
3011 ia64_patch (buf, label);
3012 ia64_add (code, ins->dreg, GP_SCRATCH_REG, ins->sreg2);
3016 /* Exception handling */
3017 case OP_CALL_HANDLER:
3019 * Using a call instruction would mess up the register stack, so
3020 * save the return address to a register and use a
3023 ia64_codegen_set_one_ins_per_bundle (code, TRUE);
3024 ia64_mov (code, IA64_R15, IA64_R0);
3025 ia64_mov_from_ip (code, GP_SCRATCH_REG);
3026 /* Add the length of OP_CALL_HANDLER */
3027 ia64_adds_imm (code, GP_SCRATCH_REG, 5 * 16, GP_SCRATCH_REG);
3028 add_patch_info (cfg, code, MONO_PATCH_INFO_BB, ins->inst_target_bb);
3029 ia64_movl (code, GP_SCRATCH_REG2, 0);
3030 ia64_mov_to_br (code, IA64_B6, GP_SCRATCH_REG2);
3031 ia64_br_cond_reg (code, IA64_B6);
3033 //mono_cfg_add_try_hole (cfg, ins->inst_eh_block, code, bb);
3034 ia64_codegen_set_one_ins_per_bundle (code, FALSE);
3036 case OP_START_HANDLER: {
3038 * We receive the return address in GP_SCRATCH_REG.
3040 MonoInst *spvar = mono_find_spvar_for_region (cfg, bb->region);
3043 * R15 determines our caller. It is used since it is writable using
3045 * R15 == 0 means we are called by OP_CALL_HANDLER or via resume_context ()
3046 * R15 != 0 means we are called by call_filter ().
3048 ia64_codegen_set_one_ins_per_bundle (code, TRUE);
3049 ia64_cmp_eq (code, 6, 7, IA64_R15, IA64_R0);
3051 ia64_br_cond_pred (code, 6, 6);
3054 * Called by call_filter:
3055 * Allocate a new stack frame, and set the fp register from the
3056 * value passed in by the caller.
3057 * We allocate a similar frame as is done by the prolog, so
3058 * if an exception is thrown while executing the filter, the
3059 * unwinder can unwind through the filter frame using the unwind
3060 * info for the prolog.
3062 ia64_alloc (code, cfg->arch.reg_saved_ar_pfs, cfg->arch.reg_local0 - cfg->arch.reg_in0, cfg->arch.reg_out0 - cfg->arch.reg_local0, cfg->arch.n_out_regs, 0);
3063 ia64_mov_from_br (code, cfg->arch.reg_saved_b0, IA64_B0);
3064 ia64_mov (code, cfg->arch.reg_saved_sp, IA64_SP);
3065 ia64_mov (code, cfg->frame_reg, IA64_R15);
3066 /* Signal to endfilter that we are called by call_filter */
3067 ia64_mov (code, GP_SCRATCH_REG, IA64_R0);
3069 /* Branch target: */
3070 if (ia64_is_imm14 (spvar->inst_offset))
3071 ia64_adds_imm (code, GP_SCRATCH_REG2, spvar->inst_offset, cfg->frame_reg);
3073 ia64_movl (code, GP_SCRATCH_REG2, spvar->inst_offset);
3074 ia64_add (code, GP_SCRATCH_REG2, cfg->frame_reg, GP_SCRATCH_REG2);
3077 /* Save the return address */
3078 ia64_st8_hint (code, GP_SCRATCH_REG2, GP_SCRATCH_REG, 0);
3079 ia64_codegen_set_one_ins_per_bundle (code, FALSE);
3084 case OP_ENDFILTER: {
3085 /* FIXME: Return the value in ENDFILTER */
3086 MonoInst *spvar = mono_find_spvar_for_region (cfg, bb->region);
3088 /* Load the return address */
3089 if (ia64_is_imm14 (spvar->inst_offset)) {
3090 ia64_adds_imm (code, GP_SCRATCH_REG, spvar->inst_offset, cfg->frame_reg);
3092 ia64_movl (code, GP_SCRATCH_REG, spvar->inst_offset);
3093 ia64_add (code, GP_SCRATCH_REG, cfg->frame_reg, GP_SCRATCH_REG);
3095 ia64_ld8_hint (code, GP_SCRATCH_REG, GP_SCRATCH_REG, 0);
3098 ia64_cmp_eq (code, 6, 7, GP_SCRATCH_REG, IA64_R0);
3099 ia64_br_cond_pred (code, 7, 4);
3101 /* Called by call_filter */
3103 ia64_mov_to_ar_i (code, IA64_PFS, cfg->arch.reg_saved_ar_pfs);
3104 ia64_mov_to_br (code, IA64_B0, cfg->arch.reg_saved_b0);
3105 ia64_br_ret_reg (code, IA64_B0);
3107 /* Called by CALL_HANDLER */
3108 ia64_mov_to_br (code, IA64_B6, GP_SCRATCH_REG);
3109 ia64_br_cond_reg (code, IA64_B6);
3113 ia64_mov (code, cfg->arch.reg_out0, ins->sreg1);
3114 code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD,
3115 (gpointer)"mono_arch_throw_exception");
3118 * This might be the last instruction in the method, so add a dummy
3119 * instruction so the unwinder will work.
3121 ia64_break_i (code, 0);
3124 ia64_mov (code, cfg->arch.reg_out0, ins->sreg1);
3125 code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD,
3126 (gpointer)"mono_arch_rethrow_exception");
3128 ia64_break_i (code, 0);
3132 g_warning ("unknown opcode %s in %s()\n", mono_inst_name (ins->opcode), __FUNCTION__);
3133 g_assert_not_reached ();
3136 if ((code.buf - cfg->native_code - offset) > max_len) {
3137 g_warning ("wrong maximal instruction length of instruction %s (expected %d, got %ld)",
3138 mono_inst_name (ins->opcode), max_len, code.buf - cfg->native_code - offset);
3139 g_assert_not_reached ();
3145 last_offset = offset;
3148 ia64_codegen_close (code);
3150 cfg->code_len = code.buf - cfg->native_code;
3154 mono_arch_register_lowlevel_calls (void)
3158 static Ia64InsType ins_types_in_template [32][3] = {
3159 {IA64_INS_TYPE_M, IA64_INS_TYPE_I, IA64_INS_TYPE_I},
3160 {IA64_INS_TYPE_M, IA64_INS_TYPE_I, IA64_INS_TYPE_I},
3161 {IA64_INS_TYPE_M, IA64_INS_TYPE_I, IA64_INS_TYPE_I},
3162 {IA64_INS_TYPE_M, IA64_INS_TYPE_I, IA64_INS_TYPE_I},
3163 {IA64_INS_TYPE_M, IA64_INS_TYPE_LX, IA64_INS_TYPE_LX},
3164 {IA64_INS_TYPE_M, IA64_INS_TYPE_LX, IA64_INS_TYPE_LX},
3167 {IA64_INS_TYPE_M, IA64_INS_TYPE_M, IA64_INS_TYPE_I},
3168 {IA64_INS_TYPE_M, IA64_INS_TYPE_M, IA64_INS_TYPE_I},
3169 {IA64_INS_TYPE_M, IA64_INS_TYPE_M, IA64_INS_TYPE_I},
3170 {IA64_INS_TYPE_M, IA64_INS_TYPE_M, IA64_INS_TYPE_I},
3171 {IA64_INS_TYPE_M, IA64_INS_TYPE_F, IA64_INS_TYPE_I},
3172 {IA64_INS_TYPE_M, IA64_INS_TYPE_F, IA64_INS_TYPE_I},
3173 {IA64_INS_TYPE_M, IA64_INS_TYPE_M, IA64_INS_TYPE_F},
3174 {IA64_INS_TYPE_M, IA64_INS_TYPE_M, IA64_INS_TYPE_F},
3175 {IA64_INS_TYPE_M, IA64_INS_TYPE_I, IA64_INS_TYPE_B},
3176 {IA64_INS_TYPE_M, IA64_INS_TYPE_I, IA64_INS_TYPE_B},
3177 {IA64_INS_TYPE_M, IA64_INS_TYPE_B, IA64_INS_TYPE_B},
3178 {IA64_INS_TYPE_M, IA64_INS_TYPE_B, IA64_INS_TYPE_B},
3181 {IA64_INS_TYPE_B, IA64_INS_TYPE_B, IA64_INS_TYPE_B},
3182 {IA64_INS_TYPE_B, IA64_INS_TYPE_B, IA64_INS_TYPE_B},
3183 {IA64_INS_TYPE_M, IA64_INS_TYPE_M, IA64_INS_TYPE_B},
3184 {IA64_INS_TYPE_M, IA64_INS_TYPE_M, IA64_INS_TYPE_B},
3187 {IA64_INS_TYPE_M, IA64_INS_TYPE_F, IA64_INS_TYPE_B},
3188 {IA64_INS_TYPE_M, IA64_INS_TYPE_F, IA64_INS_TYPE_B},
3193 static gboolean stops_in_template [32][3] = {
3194 { FALSE, FALSE, FALSE },
3195 { FALSE, FALSE, TRUE },
3196 { FALSE, TRUE, FALSE },
3197 { FALSE, TRUE, TRUE },
3198 { FALSE, FALSE, FALSE },
3199 { FALSE, FALSE, TRUE },
3200 { FALSE, FALSE, FALSE },
3201 { FALSE, FALSE, FALSE },
3203 { FALSE, FALSE, FALSE },
3204 { FALSE, FALSE, TRUE },
3205 { TRUE, FALSE, FALSE },
3206 { TRUE, FALSE, TRUE },
3207 { FALSE, FALSE, FALSE },
3208 { FALSE, FALSE, TRUE },
3209 { FALSE, FALSE, FALSE },
3210 { FALSE, FALSE, TRUE },
3212 { FALSE, FALSE, FALSE },
3213 { FALSE, FALSE, TRUE },
3214 { FALSE, FALSE, FALSE },
3215 { FALSE, FALSE, TRUE },
3216 { FALSE, FALSE, FALSE },
3217 { FALSE, FALSE, FALSE },
3218 { FALSE, FALSE, FALSE },
3219 { FALSE, FALSE, TRUE },
3221 { FALSE, FALSE, FALSE },
3222 { FALSE, FALSE, TRUE },
3223 { FALSE, FALSE, FALSE },
3224 { FALSE, FALSE, FALSE },
3225 { FALSE, FALSE, FALSE },
3226 { FALSE, FALSE, TRUE },
3227 { FALSE, FALSE, FALSE },
3228 { FALSE, FALSE, FALSE }
3231 static int last_stop_in_template [32] = {
3232 -1, 2, 1, 2, -1, 2, -1, -1,
3233 -1, 2, 0, 2, -1, 2, -1, 2,
3234 -1, 2, -1, 2, -1, -1, -1, 2,
3235 -1, 2, -1, -1, -1, 2, -1, -1
3238 static guint64 nops_for_ins_types [6] = {
3247 #define ITYPE_MATCH(itype1, itype2) (((itype1) == (itype2)) || (((itype2) == IA64_INS_TYPE_A) && (((itype1) == IA64_INS_TYPE_I) || ((itype1) == IA64_INS_TYPE_M))))
3254 #define DEBUG_INS_SCHED(a) do { a; } while (0)
3256 #define DEBUG_INS_SCHED(a)
3260 ia64_analyze_deps (Ia64CodegenState *code, int *deps_start, int *stops)
3262 int i, pos, ins_index, current_deps_start, current_ins_start, reg;
3263 guint8 *deps = code->dep_info;
3264 gboolean need_stop, no_stop;
3266 for (i = 0; i < code->nins; ++i)
3270 current_deps_start = 0;
3271 current_ins_start = 0;
3272 deps_start [ins_index] = current_ins_start;
3275 DEBUG_INS_SCHED (printf ("BEGIN.\n"));
3276 while (pos < code->dep_info_pos) {
3278 switch (deps [pos]) {
3279 case IA64_END_OF_INS:
3281 current_ins_start = pos + 2;
3282 deps_start [ins_index] = current_ins_start;
3284 DEBUG_INS_SCHED (printf ("(%d) END INS.\n", ins_index - 1));
3289 reg = deps [pos + 1];
3291 DEBUG_INS_SCHED (printf ("READ GR: %d\n", reg));
3292 for (i = current_deps_start; i < current_ins_start; i += 2)
3293 if (deps [i] == IA64_WRITE_GR && deps [i + 1] == reg)
3297 reg = code->dep_info [pos + 1];
3299 DEBUG_INS_SCHED (printf ("WRITE GR: %d\n", reg));
3300 for (i = current_deps_start; i < current_ins_start; i += 2)
3301 if (deps [i] == IA64_WRITE_GR && deps [i + 1] == reg)
3305 reg = deps [pos + 1];
3307 DEBUG_INS_SCHED (printf ("READ PR: %d\n", reg));
3308 for (i = current_deps_start; i < current_ins_start; i += 2)
3309 if (((deps [i] == IA64_WRITE_PR) || (deps [i] == IA64_WRITE_PR_FLOAT)) && deps [i + 1] == reg)
3312 case IA64_READ_PR_BRANCH:
3313 reg = deps [pos + 1];
3315 /* Writes to prs by non-float instructions are visible to branches */
3316 DEBUG_INS_SCHED (printf ("READ PR BRANCH: %d\n", reg));
3317 for (i = current_deps_start; i < current_ins_start; i += 2)
3318 if (deps [i] == IA64_WRITE_PR_FLOAT && deps [i + 1] == reg)
3322 reg = code->dep_info [pos + 1];
3324 DEBUG_INS_SCHED (printf ("WRITE PR: %d\n", reg));
3325 for (i = current_deps_start; i < current_ins_start; i += 2)
3326 if (((deps [i] == IA64_WRITE_PR) || (deps [i] == IA64_WRITE_PR_FLOAT)) && deps [i + 1] == reg)
3329 case IA64_WRITE_PR_FLOAT:
3330 reg = code->dep_info [pos + 1];
3332 DEBUG_INS_SCHED (printf ("WRITE PR FP: %d\n", reg));
3333 for (i = current_deps_start; i < current_ins_start; i += 2)
3334 if (((deps [i] == IA64_WRITE_GR) || (deps [i] == IA64_WRITE_PR_FLOAT)) && deps [i + 1] == reg)
3338 reg = deps [pos + 1];
3340 DEBUG_INS_SCHED (printf ("READ BR: %d\n", reg));
3341 for (i = current_deps_start; i < current_ins_start; i += 2)
3342 if (deps [i] == IA64_WRITE_BR && deps [i + 1] == reg)
3346 reg = code->dep_info [pos + 1];
3348 DEBUG_INS_SCHED (printf ("WRITE BR: %d\n", reg));
3349 for (i = current_deps_start; i < current_ins_start; i += 2)
3350 if (deps [i] == IA64_WRITE_BR && deps [i + 1] == reg)
3353 case IA64_READ_BR_BRANCH:
3354 reg = deps [pos + 1];
3356 /* Writes to brs are visible to branches */
3357 DEBUG_INS_SCHED (printf ("READ BR BRACH: %d\n", reg));
3360 reg = deps [pos + 1];
3362 DEBUG_INS_SCHED (printf ("READ BR: %d\n", reg));
3363 for (i = current_deps_start; i < current_ins_start; i += 2)
3364 if (deps [i] == IA64_WRITE_FR && deps [i + 1] == reg)
3368 reg = code->dep_info [pos + 1];
3370 DEBUG_INS_SCHED (printf ("WRITE BR: %d\n", reg));
3371 for (i = current_deps_start; i < current_ins_start; i += 2)
3372 if (deps [i] == IA64_WRITE_FR && deps [i + 1] == reg)
3376 reg = deps [pos + 1];
3378 DEBUG_INS_SCHED (printf ("READ AR: %d\n", reg));
3379 for (i = current_deps_start; i < current_ins_start; i += 2)
3380 if (deps [i] == IA64_WRITE_AR && deps [i + 1] == reg)
3384 reg = code->dep_info [pos + 1];
3386 DEBUG_INS_SCHED (printf ("WRITE AR: %d\n", reg));
3387 for (i = current_deps_start; i < current_ins_start; i += 2)
3388 if (deps [i] == IA64_WRITE_AR && deps [i + 1] == reg)
3393 * Explicitly indicate that a stop is not required. Useful for
3394 * example when two predicated instructions with negated predicates
3395 * write the same registers.
3400 g_assert_not_reached ();
3404 if (need_stop && !no_stop) {
3405 g_assert (ins_index > 0);
3406 stops [ins_index - 1] = 1;
3408 DEBUG_INS_SCHED (printf ("STOP\n"));
3409 current_deps_start = current_ins_start;
3411 /* Skip remaining deps for this instruction */
3412 while (deps [pos] != IA64_END_OF_INS)
3417 if (code->nins > 0) {
3418 /* No dependency info for the last instruction */
3419 stops [code->nins - 1] = 1;
3422 deps_start [code->nins] = code->dep_info_pos;
3426 ia64_real_emit_bundle (Ia64CodegenState *code, int *deps_start, int *stops, int n, guint64 template, guint64 ins1, guint64 ins2, guint64 ins3, guint8 nops)
3428 int stop_pos, i, deps_to_shift, dep_shift;
3430 g_assert (n <= code->nins);
3432 // if (n > 1) printf ("FOUND: %ld.\n", template);
3434 ia64_emit_bundle_template (code, template, ins1, ins2, ins3);
3436 stop_pos = last_stop_in_template [template] + 1;
3440 /* Compute the number of 'real' instructions before the stop */
3441 deps_to_shift = stop_pos;
3442 if (stop_pos >= 3 && (nops & (1 << 2)))
3444 if (stop_pos >= 2 && (nops & (1 << 1)))
3446 if (stop_pos >= 1 && (nops & (1 << 0)))
3450 * We have to keep some dependencies whose instructions have been shifted
3451 * out of the buffer. So nullify the end_of_ins markers in the dependency
3454 for (i = deps_start [deps_to_shift]; i < deps_start [n]; i += 2)
3455 if (code->dep_info [i] == IA64_END_OF_INS)
3456 code->dep_info [i] = IA64_NONE;
3458 g_assert (deps_start [deps_to_shift] <= code->dep_info_pos);
3459 memcpy (code->dep_info, &code->dep_info [deps_start [deps_to_shift]], code->dep_info_pos - deps_start [deps_to_shift]);
3460 code->dep_info_pos = code->dep_info_pos - deps_start [deps_to_shift];
3462 dep_shift = deps_start [deps_to_shift];
3463 for (i = 0; i < code->nins + 1 - n; ++i)
3464 deps_start [i] = deps_start [n + i] - dep_shift;
3466 /* Determine the exact positions of instructions with unwind ops */
3467 if (code->unw_op_count) {
3469 int curr_ins, curr_ins_pos;
3472 curr_ins_pos = ((code->buf - code->region_start - 16) / 16) * 3;
3473 for (i = 0; i < 3; ++i) {
3474 if (! (nops & (1 << i))) {
3475 ins_pos [curr_ins] = curr_ins_pos + i;
3480 for (i = code->unw_op_pos; i < code->unw_op_count; ++i) {
3481 if (code->unw_ops_pos [i] < n) {
3482 code->unw_ops [i].when = ins_pos [code->unw_ops_pos [i]];
3483 //printf ("UNW-OP: %d -> %d\n", code->unw_ops_pos [i], code->unw_ops [i].when);
3486 if (code->unw_op_pos < code->unw_op_count)
3487 code->unw_op_pos += n;
3490 if (n == code->nins) {
3495 memcpy (&code->instructions [0], &code->instructions [n], (code->nins - n) * sizeof (guint64));
3496 memcpy (&code->itypes [0], &code->itypes [n], (code->nins - n) * sizeof (int));
3497 memcpy (&stops [0], &stops [n], (code->nins - n) * sizeof (int));
3503 ia64_emit_bundle (Ia64CodegenState *code, gboolean flush)
3505 int i, ins_type, template, nins_to_emit;
3506 int deps_start [16];
3511 * We implement a simple scheduler which tries to put three instructions
3512 * per bundle, then two, then one.
3514 ia64_analyze_deps (code, deps_start, stops);
3516 if ((code->nins >= 3) && !code->one_ins_per_bundle) {
3517 /* Find a suitable template */
3518 for (template = 0; template < 32; ++template) {
3519 if (stops_in_template [template][0] != stops [0] ||
3520 stops_in_template [template][1] != stops [1] ||
3521 stops_in_template [template][2] != stops [2])
3525 for (i = 0; i < 3; ++i) {
3526 ins_type = ins_types_in_template [template][i];
3527 switch (code->itypes [i]) {
3528 case IA64_INS_TYPE_A:
3529 found &= (ins_type == IA64_INS_TYPE_I) || (ins_type == IA64_INS_TYPE_M);
3532 found &= (ins_type == code->itypes [i]);
3538 found = debug_ins_sched ();
3541 ia64_real_emit_bundle (code, deps_start, stops, 3, template, code->instructions [0], code->instructions [1], code->instructions [2], 0);
3547 if (code->nins < IA64_INS_BUFFER_SIZE && !flush)
3548 /* Wait for more instructions */
3551 /* If it didn't work out, try putting two instructions into one bundle */
3552 if ((code->nins >= 2) && !code->one_ins_per_bundle) {
3553 /* Try a nop at the end */
3554 for (template = 0; template < 32; ++template) {
3555 if (stops_in_template [template][0] != stops [0] ||
3556 ((stops_in_template [template][1] != stops [1]) &&
3557 (stops_in_template [template][2] != stops [1])))
3561 if (!ITYPE_MATCH (ins_types_in_template [template][0], code->itypes [0]) ||
3562 !ITYPE_MATCH (ins_types_in_template [template][1], code->itypes [1]))
3565 if (!debug_ins_sched ())
3568 ia64_real_emit_bundle (code, deps_start, stops, 2, template, code->instructions [0], code->instructions [1], nops_for_ins_types [ins_types_in_template [template][2]], 1 << 2);
3573 if (code->nins < IA64_INS_BUFFER_SIZE && !flush)
3574 /* Wait for more instructions */
3577 if ((code->nins >= 2) && !code->one_ins_per_bundle) {
3578 /* Try a nop in the middle */
3579 for (template = 0; template < 32; ++template) {
3580 if (((stops_in_template [template][0] != stops [0]) &&
3581 (stops_in_template [template][1] != stops [0])) ||
3582 stops_in_template [template][2] != stops [1])
3585 if (!ITYPE_MATCH (ins_types_in_template [template][0], code->itypes [0]) ||
3586 !ITYPE_MATCH (ins_types_in_template [template][2], code->itypes [1]))
3589 if (!debug_ins_sched ())
3592 ia64_real_emit_bundle (code, deps_start, stops, 2, template, code->instructions [0], nops_for_ins_types [ins_types_in_template [template][1]], code->instructions [1], 1 << 1);
3597 if ((code->nins >= 2) && flush && !code->one_ins_per_bundle) {
3598 /* Try a nop at the beginning */
3599 for (template = 0; template < 32; ++template) {
3600 if ((stops_in_template [template][1] != stops [0]) ||
3601 (stops_in_template [template][2] != stops [1]))
3604 if (!ITYPE_MATCH (ins_types_in_template [template][1], code->itypes [0]) ||
3605 !ITYPE_MATCH (ins_types_in_template [template][2], code->itypes [1]))
3608 if (!debug_ins_sched ())
3611 ia64_real_emit_bundle (code, deps_start, stops, 2, template, nops_for_ins_types [ins_types_in_template [template][0]], code->instructions [0], code->instructions [1], 1 << 0);
3616 if (code->nins < IA64_INS_BUFFER_SIZE && !flush)
3617 /* Wait for more instructions */
3621 nins_to_emit = code->nins;
3625 while (nins_to_emit > 0) {
3626 if (!debug_ins_sched ())
3628 switch (code->itypes [0]) {
3629 case IA64_INS_TYPE_A:
3631 ia64_real_emit_bundle (code, deps_start, stops, 1, IA64_TEMPLATE_MIIS, code->instructions [0], IA64_NOP_I, IA64_NOP_I, 0);
3633 ia64_real_emit_bundle (code, deps_start, stops, 1, IA64_TEMPLATE_MII, code->instructions [0], IA64_NOP_I, IA64_NOP_I, 0);
3635 case IA64_INS_TYPE_I:
3637 ia64_real_emit_bundle (code, deps_start, stops, 1, IA64_TEMPLATE_MIIS, IA64_NOP_M, code->instructions [0], IA64_NOP_I, 0);
3639 ia64_real_emit_bundle (code, deps_start, stops, 1, IA64_TEMPLATE_MII, IA64_NOP_M, code->instructions [0], IA64_NOP_I, 0);
3641 case IA64_INS_TYPE_M:
3643 ia64_real_emit_bundle (code, deps_start, stops, 1, IA64_TEMPLATE_MIIS, code->instructions [0], IA64_NOP_I, IA64_NOP_I, 0);
3645 ia64_real_emit_bundle (code, deps_start, stops, 1, IA64_TEMPLATE_MII, code->instructions [0], IA64_NOP_I, IA64_NOP_I, 0);
3647 case IA64_INS_TYPE_B:
3649 ia64_real_emit_bundle (code, deps_start, stops, 1, IA64_TEMPLATE_MIBS, IA64_NOP_M, IA64_NOP_I, code->instructions [0], 0);
3651 ia64_real_emit_bundle (code, deps_start, stops, 1, IA64_TEMPLATE_MIB, IA64_NOP_M, IA64_NOP_I, code->instructions [0], 0);
3653 case IA64_INS_TYPE_F:
3655 ia64_real_emit_bundle (code, deps_start, stops, 1, IA64_TEMPLATE_MFIS, IA64_NOP_M, code->instructions [0], IA64_NOP_I, 0);
3657 ia64_real_emit_bundle (code, deps_start, stops, 1, IA64_TEMPLATE_MFI, IA64_NOP_M, code->instructions [0], IA64_NOP_I, 0);
3659 case IA64_INS_TYPE_LX:
3660 if (stops [0] || stops [1])
3661 ia64_real_emit_bundle (code, deps_start, stops, 2, IA64_TEMPLATE_MLXS, IA64_NOP_M, code->instructions [0], code->instructions [1], 0);
3663 ia64_real_emit_bundle (code, deps_start, stops, 2, IA64_TEMPLATE_MLX, IA64_NOP_M, code->instructions [0], code->instructions [1], 0);
3667 g_assert_not_reached ();
3673 unw_dyn_region_info_t*
3674 mono_ia64_create_unwind_region (Ia64CodegenState *code)
3676 unw_dyn_region_info_t *r;
3678 g_assert (code->nins == 0);
3679 r = g_malloc0 (_U_dyn_region_info_size (code->unw_op_count));
3680 memcpy (&r->op, &code->unw_ops, sizeof (unw_dyn_op_t) * code->unw_op_count);
3681 r->op_count = code->unw_op_count;
3682 r->insn_count = ((code->buf - code->region_start) >> 4) * 3;
3683 code->unw_op_count = 0;
3684 code->unw_op_pos = 0;
3685 code->region_start = code->buf;
3691 ia64_patch (unsigned char* code, gpointer target)
3694 guint64 instructions [3];
3695 guint8 gen_buf [16];
3696 Ia64CodegenState gen;
3701 * code encodes both the position inside the buffer and code.nins when
3702 * the instruction was emitted.
3704 ins_to_skip = (guint64)code % 16;
3705 code = (unsigned char*)((guint64)code & ~15);
3708 * Search for the first instruction which is 'patchable', skipping
3709 * ins_to_skip instructions.
3714 template = ia64_bundle_template (code);
3715 instructions [0] = ia64_bundle_ins1 (code);
3716 instructions [1] = ia64_bundle_ins2 (code);
3717 instructions [2] = ia64_bundle_ins3 (code);
3719 ia64_codegen_init (gen, gen_buf);
3722 for (i = 0; i < 3; ++i) {
3723 guint64 ins = instructions [i];
3724 int opcode = ia64_ins_opcode (ins);
3726 if (ins == nops_for_ins_types [ins_types_in_template [template][i]])
3734 switch (ins_types_in_template [template][i]) {
3735 case IA64_INS_TYPE_A:
3736 case IA64_INS_TYPE_M:
3737 if ((opcode == 8) && (ia64_ins_x2a (ins) == 2) && (ia64_ins_ve (ins) == 0)) {
3739 ia64_adds_imm_pred (gen, ia64_ins_qp (ins), ia64_ins_r1 (ins), (guint64)target, ia64_ins_r3 (ins));
3740 instructions [i] = gen.instructions [0];
3746 case IA64_INS_TYPE_B:
3747 if ((opcode == 4) && (ia64_ins_btype (ins) == 0)) {
3749 gint64 disp = ((guint8*)target - code) >> 4;
3752 ia64_br_cond_hint_pred (gen, ia64_ins_qp (ins), disp, 0, 0, 0);
3754 instructions [i] = gen.instructions [0];
3757 else if (opcode == 5) {
3759 gint64 disp = ((guint8*)target - code) >> 4;
3762 ia64_br_call_hint_pred (gen, ia64_ins_qp (ins), ia64_ins_b1 (ins), disp, 0, 0, 0);
3763 instructions [i] = gen.instructions [0];
3769 case IA64_INS_TYPE_LX:
3773 if ((opcode == 6) && (ia64_ins_vc (ins) == 0)) {
3775 ia64_movl_pred (gen, ia64_ins_qp (ins), ia64_ins_r1 (ins), target);
3776 instructions [1] = gen.instructions [0];
3777 instructions [2] = gen.instructions [1];
3790 ia64_codegen_init (gen, code);
3791 ia64_emit_bundle_template (&gen, template, instructions [0], instructions [1], instructions [2]);
3801 mono_arch_patch_code (MonoMethod *method, MonoDomain *domain, guint8 *code, MonoJumpInfo *ji, MonoCodeManager *dyn_code_mp, gboolean run_cctors)
3803 MonoJumpInfo *patch_info;
3805 for (patch_info = ji; patch_info; patch_info = patch_info->next) {
3806 unsigned char *ip = patch_info->ip.i + code;
3807 const unsigned char *target;
3809 target = mono_resolve_patch_target (method, domain, code, patch_info, run_cctors);
3811 if (patch_info->type == MONO_PATCH_INFO_NONE)
3813 if (mono_compile_aot) {
3817 ia64_patch (ip, (gpointer)target);
3822 mono_arch_emit_prolog (MonoCompile *cfg)
3824 MonoMethod *method = cfg->method;
3825 MonoMethodSignature *sig;
3827 int alloc_size, pos, i;
3828 Ia64CodegenState code;
3831 sig = mono_method_signature (method);
3834 cinfo = get_call_info (cfg, cfg->mempool, sig, FALSE);
3836 cfg->code_size = MAX (cfg->header->code_size * 4, 512);
3838 if (mono_jit_trace_calls != NULL && mono_trace_eval (method))
3839 cfg->code_size += 1024;
3840 if (cfg->prof_options & MONO_PROFILE_ENTER_LEAVE)
3841 cfg->code_size += 1024;
3843 cfg->native_code = g_malloc (cfg->code_size);
3845 ia64_codegen_init (code, cfg->native_code);
3847 alloc_size = ALIGN_TO (cfg->stack_offset, MONO_ARCH_FRAME_ALIGNMENT);
3848 if (cfg->param_area)
3849 alloc_size += cfg->param_area;
3853 alloc_size = ALIGN_TO (alloc_size, MONO_ARCH_FRAME_ALIGNMENT);
3855 if (cfg->flags & MONO_CFG_HAS_ALLOCA)
3856 /* Force sp to be saved/restored */
3857 alloc_size += MONO_ARCH_FRAME_ALIGNMENT;
3859 cfg->arch.stack_alloc_size = alloc_size;
3863 if (method->save_lmf) {
3864 /* No LMF on IA64 */
3869 ia64_unw_save_reg (code, UNW_IA64_AR_PFS, UNW_IA64_GR + cfg->arch.reg_saved_ar_pfs);
3870 ia64_alloc (code, cfg->arch.reg_saved_ar_pfs, cfg->arch.reg_local0 - cfg->arch.reg_in0, cfg->arch.reg_out0 - cfg->arch.reg_local0, cfg->arch.n_out_regs, 0);
3871 ia64_unw_save_reg (code, UNW_IA64_RP, UNW_IA64_GR + cfg->arch.reg_saved_b0);
3872 ia64_mov_from_br (code, cfg->arch.reg_saved_b0, IA64_B0);
3874 if ((alloc_size || cinfo->stack_usage) && !cfg->arch.omit_fp) {
3875 ia64_unw_save_reg (code, UNW_IA64_SP, UNW_IA64_GR + cfg->arch.reg_saved_sp);
3876 ia64_mov (code, cfg->arch.reg_saved_sp, IA64_SP);
3877 if (cfg->frame_reg != cfg->arch.reg_saved_sp)
3878 ia64_mov (code, cfg->frame_reg, IA64_SP);
3882 #if defined(MONO_ARCH_SIGSEGV_ON_ALTSTACK)
3883 int pagesize = getpagesize ();
3885 if (alloc_size >= pagesize) {
3886 gint32 remaining_size = alloc_size;
3888 /* Generate stack touching code */
3889 ia64_mov (code, GP_SCRATCH_REG, IA64_SP);
3890 while (remaining_size >= pagesize) {
3891 ia64_movl (code, GP_SCRATCH_REG2, pagesize);
3892 ia64_sub (code, GP_SCRATCH_REG, GP_SCRATCH_REG, GP_SCRATCH_REG2);
3893 ia64_ld8 (code, GP_SCRATCH_REG2, GP_SCRATCH_REG);
3894 remaining_size -= pagesize;
3898 if (ia64_is_imm14 (-alloc_size)) {
3899 if (cfg->arch.omit_fp)
3900 ia64_unw_add (code, UNW_IA64_SP, (-alloc_size));
3901 ia64_adds_imm (code, IA64_SP, (-alloc_size), IA64_SP);
3904 ia64_movl (code, GP_SCRATCH_REG, -alloc_size);
3905 if (cfg->arch.omit_fp)
3906 ia64_unw_add (code, UNW_IA64_SP, (-alloc_size));
3907 ia64_add (code, IA64_SP, GP_SCRATCH_REG, IA64_SP);
3911 ia64_begin_bundle (code);
3913 /* Initialize unwind info */
3914 cfg->arch.r_pro = mono_ia64_create_unwind_region (&code);
3916 if (sig->ret->type != MONO_TYPE_VOID) {
3917 if ((cinfo->ret.storage == ArgInIReg) && (cfg->ret->opcode != OP_REGVAR)) {
3918 /* Save volatile arguments to the stack */
3923 /* Keep this in sync with emit_load_volatile_arguments */
3924 for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
3925 ArgInfo *ainfo = cinfo->args + i;
3926 gint32 stack_offset;
3929 inst = cfg->args [i];
3931 if (sig->hasthis && (i == 0))
3932 arg_type = &mono_defaults.object_class->byval_arg;
3934 arg_type = sig->params [i - sig->hasthis];
3936 arg_type = mono_type_get_underlying_type (arg_type);
3938 stack_offset = ainfo->offset + ARGS_OFFSET;
3941 * FIXME: Native code might pass non register sized integers
3942 * without initializing the upper bits.
3944 if (method->wrapper_type == MONO_WRAPPER_NATIVE_TO_MANAGED && !arg_type->byref && ainfo->storage == ArgInIReg) {
3945 int reg = cfg->arch.reg_in0 + ainfo->reg;
3947 switch (mono_type_to_load_membase (cfg, arg_type)) {
3948 case OP_LOADI1_MEMBASE:
3949 ia64_sxt1 (code, reg, reg);
3951 case OP_LOADU1_MEMBASE:
3952 ia64_zxt1 (code, reg, reg);
3954 case OP_LOADI2_MEMBASE:
3955 ia64_sxt2 (code, reg, reg);
3957 case OP_LOADU2_MEMBASE:
3958 ia64_zxt2 (code, reg, reg);
3965 /* Save volatile arguments to the stack */
3966 if (inst->opcode != OP_REGVAR) {
3967 switch (ainfo->storage) {
3970 case ArgInFloatRegR4:
3971 g_assert (inst->opcode == OP_REGOFFSET);
3972 if (ia64_is_adds_imm (inst->inst_offset))
3973 ia64_adds_imm (code, GP_SCRATCH_REG, inst->inst_offset, inst->inst_basereg);
3975 ia64_movl (code, GP_SCRATCH_REG2, inst->inst_offset);
3976 ia64_add (code, GP_SCRATCH_REG, GP_SCRATCH_REG, GP_SCRATCH_REG2);
3978 if (arg_type->byref)
3979 ia64_st8_hint (code, GP_SCRATCH_REG, cfg->arch.reg_in0 + ainfo->reg, 0);
3981 switch (arg_type->type) {
3983 ia64_stfs_hint (code, GP_SCRATCH_REG, ainfo->reg, 0);
3986 ia64_stfd_hint (code, GP_SCRATCH_REG, ainfo->reg, 0);
3989 ia64_st8_hint (code, GP_SCRATCH_REG, cfg->arch.reg_in0 + ainfo->reg, 0);
3997 if (ainfo->nslots != ainfo->nregs)
4000 g_assert (inst->opcode == OP_REGOFFSET);
4001 ia64_adds_imm (code, GP_SCRATCH_REG, inst->inst_offset, inst->inst_basereg);
4002 for (i = 0; i < ainfo->nregs; ++i) {
4003 switch (ainfo->atype) {
4004 case AggregateNormal:
4005 ia64_st8_inc_imm_hint (code, GP_SCRATCH_REG, cfg->arch.reg_in0 + ainfo->reg + i, sizeof (gpointer), 0);
4007 case AggregateSingleHFA:
4008 ia64_stfs_inc_imm_hint (code, GP_SCRATCH_REG, ainfo->reg + i, 4, 0);
4010 case AggregateDoubleHFA:
4011 ia64_stfd_inc_imm_hint (code, GP_SCRATCH_REG, ainfo->reg + i, sizeof (gpointer), 0);
4019 g_assert_not_reached ();
4023 if (inst->opcode == OP_REGVAR) {
4024 /* Argument allocated to (non-volatile) register */
4025 switch (ainfo->storage) {
4027 if (inst->dreg != cfg->arch.reg_in0 + ainfo->reg)
4028 ia64_mov (code, inst->dreg, cfg->arch.reg_in0 + ainfo->reg);
4031 ia64_adds_imm (code, GP_SCRATCH_REG, 16 + ainfo->offset, cfg->frame_reg);
4032 ia64_ld8 (code, inst->dreg, GP_SCRATCH_REG);
4040 if (method->save_lmf) {
4041 /* No LMF on IA64 */
4044 ia64_codegen_close (code);
4046 if (mono_jit_trace_calls != NULL && mono_trace_eval (method))
4047 code.buf = mono_arch_instrument_prolog (cfg, mono_trace_enter_method, code.buf, TRUE);
4049 cfg->code_len = code.buf - cfg->native_code;
4051 g_assert (cfg->code_len < cfg->code_size);
4053 cfg->arch.prolog_end_offset = cfg->code_len;
4059 mono_arch_emit_epilog (MonoCompile *cfg)
4061 MonoMethod *method = cfg->method;
4063 int max_epilog_size = 16 * 4;
4064 Ia64CodegenState code;
4069 if (mono_jit_trace_calls != NULL)
4070 max_epilog_size += 1024;
4072 cfg->arch.epilog_begin_offset = cfg->code_len;
4074 while (cfg->code_len + max_epilog_size > cfg->code_size) {
4075 cfg->code_size *= 2;
4076 cfg->native_code = g_realloc (cfg->native_code, cfg->code_size);
4077 cfg->stat_code_reallocs++;
4080 /* FIXME: Emit unwind info */
4082 buf = cfg->native_code + cfg->code_len;
4084 if (mono_jit_trace_calls != NULL && mono_trace_eval (method))
4085 buf = mono_arch_instrument_epilog (cfg, mono_trace_leave_method, buf, TRUE);
4087 ia64_codegen_init (code, buf);
4089 /* the code restoring the registers must be kept in sync with OP_JMP */
4092 if (method->save_lmf) {
4093 /* No LMF on IA64 */
4096 /* Load returned vtypes into registers if needed */
4097 cinfo = get_call_info (cfg, cfg->mempool, mono_method_signature (method), FALSE);
4098 ainfo = &cinfo->ret;
4099 switch (ainfo->storage) {
4101 if (ainfo->nslots != ainfo->nregs)
4104 g_assert (cfg->ret->opcode == OP_REGOFFSET);
4105 ia64_adds_imm (code, GP_SCRATCH_REG, cfg->ret->inst_offset, cfg->ret->inst_basereg);
4106 for (i = 0; i < ainfo->nregs; ++i) {
4107 switch (ainfo->atype) {
4108 case AggregateNormal:
4109 ia64_ld8_inc_imm_hint (code, ainfo->reg + i, GP_SCRATCH_REG, sizeof (gpointer), 0);
4111 case AggregateSingleHFA:
4112 ia64_ldfs_inc_imm_hint (code, ainfo->reg + i, GP_SCRATCH_REG, 4, 0);
4114 case AggregateDoubleHFA:
4115 ia64_ldfd_inc_imm_hint (code, ainfo->reg + i, GP_SCRATCH_REG, sizeof (gpointer), 0);
4118 g_assert_not_reached ();
4126 ia64_begin_bundle (code);
4128 code.region_start = cfg->native_code;
4130 /* Label the unwind state at the start of the exception throwing region */
4131 //ia64_unw_label_state (code, 1234);
4133 if (cfg->arch.stack_alloc_size) {
4134 if (cfg->arch.omit_fp) {
4135 if (ia64_is_imm14 (cfg->arch.stack_alloc_size)) {
4136 ia64_unw_pop_frames (code, 1);
4137 ia64_adds_imm (code, IA64_SP, (cfg->arch.stack_alloc_size), IA64_SP);
4139 ia64_movl (code, GP_SCRATCH_REG, cfg->arch.stack_alloc_size);
4140 ia64_unw_pop_frames (code, 1);
4141 ia64_add (code, IA64_SP, GP_SCRATCH_REG, IA64_SP);
4145 ia64_unw_pop_frames (code, 1);
4146 ia64_mov (code, IA64_SP, cfg->arch.reg_saved_sp);
4149 ia64_mov_to_ar_i (code, IA64_PFS, cfg->arch.reg_saved_ar_pfs);
4150 ia64_mov_ret_to_br (code, IA64_B0, cfg->arch.reg_saved_b0);
4151 ia64_br_ret_reg (code, IA64_B0);
4153 ia64_codegen_close (code);
4155 cfg->arch.r_epilog = mono_ia64_create_unwind_region (&code);
4156 cfg->arch.r_pro->next = cfg->arch.r_epilog;
4158 cfg->code_len = code.buf - cfg->native_code;
4160 g_assert (cfg->code_len < cfg->code_size);
4164 mono_arch_emit_exceptions (MonoCompile *cfg)
4166 MonoJumpInfo *patch_info;
4168 Ia64CodegenState code;
4169 gboolean empty = TRUE;
4170 //unw_dyn_region_info_t *r_exceptions;
4171 MonoClass *exc_classes [16];
4172 guint8 *exc_throw_start [16], *exc_throw_end [16];
4173 guint32 code_size = 0;
4175 /* Compute needed space */
4176 for (patch_info = cfg->patch_info; patch_info; patch_info = patch_info->next) {
4177 if (patch_info->type == MONO_PATCH_INFO_EXC)
4179 if (patch_info->type == MONO_PATCH_INFO_R8)
4180 code_size += 8 + 7; /* sizeof (double) + alignment */
4181 if (patch_info->type == MONO_PATCH_INFO_R4)
4182 code_size += 4 + 7; /* sizeof (float) + alignment */
4188 while (cfg->code_len + code_size > (cfg->code_size - 16)) {
4189 cfg->code_size *= 2;
4190 cfg->native_code = g_realloc (cfg->native_code, cfg->code_size);
4191 cfg->stat_code_reallocs++;
4194 ia64_codegen_init (code, cfg->native_code + cfg->code_len);
4196 /* The unwind state here is the same as before the epilog */
4197 //ia64_unw_copy_state (code, 1234);
4199 /* add code to raise exceptions */
4200 /* FIXME: Optimize this */
4202 for (patch_info = cfg->patch_info; patch_info; patch_info = patch_info->next) {
4203 switch (patch_info->type) {
4204 case MONO_PATCH_INFO_EXC: {
4205 MonoClass *exc_class;
4208 guint64 exc_token_index;
4210 exc_class = mono_class_from_name (mono_defaults.corlib, "System", patch_info->data.name);
4211 g_assert (exc_class);
4212 exc_token_index = mono_metadata_token_index (exc_class->type_token);
4213 throw_ip = cfg->native_code + patch_info->ip.i;
4215 ia64_begin_bundle (code);
4217 ia64_patch (cfg->native_code + patch_info->ip.i, code.buf);
4219 /* Find a throw sequence for the same exception class */
4220 for (i = 0; i < nthrows; ++i)
4221 if (exc_classes [i] == exc_class)
4225 gint64 offset = exc_throw_end [i] - 16 - throw_ip;
4227 if (ia64_is_adds_imm (offset))
4228 ia64_adds_imm (code, cfg->arch.reg_out0 + 1, offset, IA64_R0);
4230 ia64_movl (code, cfg->arch.reg_out0 + 1, offset);
4232 buf = code.buf + code.nins;
4233 ia64_br_cond_pred (code, 0, 0);
4234 ia64_begin_bundle (code);
4235 ia64_patch (buf, exc_throw_start [i]);
4237 patch_info->type = MONO_PATCH_INFO_NONE;
4242 ia64_movl (code, cfg->arch.reg_out0 + 1, 0);
4244 ia64_begin_bundle (code);
4247 exc_classes [nthrows] = exc_class;
4248 exc_throw_start [nthrows] = code.buf;
4252 if (ia64_is_adds_imm (exc_token_index))
4253 ia64_adds_imm (code, cfg->arch.reg_out0 + 0, exc_token_index, IA64_R0);
4255 ia64_movl (code, cfg->arch.reg_out0 + 0, exc_token_index);
4257 patch_info->data.name = "mono_arch_throw_corlib_exception";
4258 patch_info->type = MONO_PATCH_INFO_INTERNAL_METHOD;
4259 patch_info->ip.i = code.buf + code.nins - cfg->native_code;
4262 ia64_movl (code, GP_SCRATCH_REG, 0);
4263 ia64_ld8_inc_imm (code, GP_SCRATCH_REG2, GP_SCRATCH_REG, 8);
4264 ia64_mov_to_br (code, IA64_B6, GP_SCRATCH_REG2);
4265 ia64_ld8 (code, IA64_GP, GP_SCRATCH_REG);
4267 ia64_br_call_reg (code, IA64_B0, IA64_B6);
4269 /* Patch up the throw offset */
4270 ia64_begin_bundle (code);
4272 ia64_patch (buf, (gpointer)(code.buf - 16 - throw_ip));
4275 exc_throw_end [nthrows] = code.buf;
4289 /* The unwinder needs this to work */
4290 ia64_break_i (code, 0);
4292 ia64_codegen_close (code);
4295 //r_exceptions = mono_ia64_create_unwind_region (&code);
4296 //cfg->arch.r_epilog = r_exceptions;
4298 cfg->code_len = code.buf - cfg->native_code;
4300 g_assert (cfg->code_len < cfg->code_size);
4304 mono_arch_instrument_prolog (MonoCompile *cfg, void *func, void *p, gboolean enable_arguments)
4306 Ia64CodegenState code;
4307 CallInfo *cinfo = NULL;
4308 MonoMethodSignature *sig;
4310 int i, n, stack_area = 0;
4312 ia64_codegen_init (code, p);
4314 /* Keep this in sync with mono_arch_get_argument_info */
4316 if (enable_arguments) {
4317 /* Allocate a new area on the stack and save arguments there */
4318 sig = mono_method_signature (cfg->method);
4320 cinfo = get_call_info (cfg, cfg->mempool, sig, FALSE);
4322 n = sig->param_count + sig->hasthis;
4324 stack_area = ALIGN_TO (n * 8, 16);
4327 ia64_movl (code, GP_SCRATCH_REG, stack_area);
4329 ia64_sub (code, IA64_SP, IA64_SP, GP_SCRATCH_REG);
4331 /* FIXME: Allocate out registers */
4333 ia64_mov (code, cfg->arch.reg_out0 + 1, IA64_SP);
4335 /* Required by the ABI */
4336 ia64_adds_imm (code, IA64_SP, -16, IA64_SP);
4338 add_patch_info (cfg, code, MONO_PATCH_INFO_METHODCONST, cfg->method);
4339 ia64_movl (code, cfg->arch.reg_out0 + 0, 0);
4341 /* Save arguments to the stack */
4342 for (i = 0; i < n; ++i) {
4343 ins = cfg->args [i];
4345 if (ins->opcode == OP_REGVAR) {
4346 ia64_movl (code, GP_SCRATCH_REG, (i * 8));
4347 ia64_add (code, GP_SCRATCH_REG, cfg->arch.reg_out0 + 1, GP_SCRATCH_REG);
4348 ia64_st8 (code, GP_SCRATCH_REG, ins->dreg);
4351 ia64_movl (code, GP_SCRATCH_REG, ins->inst_offset);
4352 ia64_add (code, GP_SCRATCH_REG, ins->inst_basereg, GP_SCRATCH_REG);
4353 ia64_ld8 (code, GP_SCRATCH_REG2, GP_SCRATCH_REG);
4354 ia64_movl (code, GP_SCRATCH_REG, (i * 8));
4355 ia64_add (code, GP_SCRATCH_REG, cfg->arch.reg_out0 + 1, GP_SCRATCH_REG);
4356 ia64_st8 (code, GP_SCRATCH_REG, GP_SCRATCH_REG2);
4361 ia64_mov (code, cfg->arch.reg_out0 + 1, IA64_R0);
4364 ia64_mov (code, cfg->arch.reg_out0 + 1, IA64_R0);
4366 add_patch_info (cfg, code, MONO_PATCH_INFO_METHODCONST, cfg->method);
4367 ia64_movl (code, cfg->arch.reg_out0 + 0, 0);
4369 code = emit_call (cfg, code, MONO_PATCH_INFO_ABS, (gpointer)func);
4371 if (enable_arguments && stack_area) {
4372 ia64_movl (code, GP_SCRATCH_REG, stack_area);
4374 ia64_add (code, IA64_SP, IA64_SP, GP_SCRATCH_REG);
4376 ia64_adds_imm (code, IA64_SP, 16, IA64_SP);
4379 ia64_codegen_close (code);
4385 mono_arch_instrument_epilog_full (MonoCompile *cfg, void *func, void *p, gboolean enable_arguments, gboolean preserve_argument_registers)
4387 Ia64CodegenState code;
4388 CallInfo *cinfo = NULL;
4389 MonoMethod *method = cfg->method;
4390 MonoMethodSignature *sig = mono_method_signature (cfg->method);
4392 ia64_codegen_init (code, p);
4394 cinfo = get_call_info (cfg, cfg->mempool, sig, FALSE);
4396 /* Save return value + pass it to func */
4397 switch (cinfo->ret.storage) {
4401 ia64_mov (code, cfg->arch.reg_saved_return_val, cinfo->ret.reg);
4402 ia64_mov (code, cfg->arch.reg_out0 + 1, cinfo->ret.reg);
4405 ia64_adds_imm (code, IA64_SP, -16, IA64_SP);
4406 ia64_adds_imm (code, GP_SCRATCH_REG, 16, IA64_SP);
4407 ia64_stfd_hint (code, GP_SCRATCH_REG, cinfo->ret.reg, 0);
4408 ia64_fmov (code, 8 + 1, cinfo->ret.reg);
4410 case ArgValuetypeAddrInIReg:
4411 ia64_mov (code, cfg->arch.reg_out0 + 1, cfg->arch.reg_in0 + cinfo->ret.reg);
4420 add_patch_info (cfg, code, MONO_PATCH_INFO_METHODCONST, method);
4421 ia64_movl (code, cfg->arch.reg_out0 + 0, 0);
4422 code = emit_call (cfg, code, MONO_PATCH_INFO_ABS, (gpointer)func);
4424 /* Restore return value */
4425 switch (cinfo->ret.storage) {
4429 ia64_mov (code, cinfo->ret.reg, cfg->arch.reg_saved_return_val);
4432 ia64_adds_imm (code, GP_SCRATCH_REG, 16, IA64_SP);
4433 ia64_ldfd (code, cinfo->ret.reg, GP_SCRATCH_REG);
4435 case ArgValuetypeAddrInIReg:
4443 ia64_codegen_close (code);
4449 mono_arch_save_unwind_info (MonoCompile *cfg)
4453 /* FIXME: Unregister this for dynamic methods */
4455 di = g_malloc0 (sizeof (unw_dyn_info_t));
4456 di->start_ip = (unw_word_t) cfg->native_code;
4457 di->end_ip = (unw_word_t) cfg->native_code + cfg->code_len;
4459 di->format = UNW_INFO_FORMAT_DYNAMIC;
4460 di->u.pi.name_ptr = (unw_word_t)mono_method_full_name (cfg->method, TRUE);
4461 di->u.pi.regions = cfg->arch.r_pro;
4463 _U_dyn_register (di);
4467 unw_dyn_region_info_t *region = di->u.pi.regions;
4469 printf ("Unwind info for method %s:\n", mono_method_full_name (cfg->method, TRUE));
4471 printf (" [Region: %d]\n", region->insn_count);
4472 region = region->next;
4479 mono_arch_flush_icache (guint8 *code, gint size)
4481 guint8* p = (guint8*)((guint64)code & ~(0x3f));
4482 guint8* end = (guint8*)((guint64)code + size);
4484 #ifdef __INTEL_COMPILER
4485 /* icc doesn't define an fc.i instrinsic, but fc==fc.i on itanium 2 */
4492 __asm__ __volatile__ ("fc.i %0"::"r"(p));
4493 /* FIXME: This could be increased to 128 on some cpus */
4500 mono_arch_flush_register_windows (void)
4502 /* Not needed because of libunwind */
4506 mono_arch_is_inst_imm (gint64 imm)
4508 /* The lowering pass will take care of it */
4514 * Determine whenever the trap whose info is in SIGINFO is caused by
4518 mono_arch_is_int_overflow (void *sigctx, void *info)
4520 /* Division is emulated with explicit overflow checks */
4525 mono_arch_get_patch_offset (guint8 *code)
4533 mono_arch_get_delegate_method_ptr_addr (guint8* code, mgreg_t *regs)
4541 mono_arch_finish_init (void)
4546 mono_arch_free_jit_tls_data (MonoJitTlsData *tls)
4550 #ifdef MONO_ARCH_HAVE_IMT
4553 * LOCKING: called with the domain lock held
4556 mono_arch_build_imt_thunk (MonoVTable *vtable, MonoDomain *domain, MonoIMTCheckItem **imt_entries, int count,
4557 gpointer fail_tramp)
4561 guint8 *start, *buf;
4562 Ia64CodegenState code;
4565 buf = g_malloc0 (size);
4566 ia64_codegen_init (code, buf);
4568 /* IA64_R9 contains the IMT method */
4570 for (i = 0; i < count; ++i) {
4571 MonoIMTCheckItem *item = imt_entries [i];
4572 ia64_begin_bundle (code);
4573 item->code_target = (guint8*)code.buf + code.nins;
4574 if (item->is_equals) {
4575 gboolean fail_case = !item->check_target_idx && fail_tramp;
4577 if (item->check_target_idx || fail_case) {
4578 if (!item->compare_done || fail_case) {
4579 ia64_movl (code, GP_SCRATCH_REG, item->key);
4580 ia64_cmp_eq (code, 6, 7, IA64_R9, GP_SCRATCH_REG);
4582 item->jmp_code = (guint8*)code.buf + code.nins;
4583 ia64_br_cond_pred (code, 7, 0);
4585 if (item->has_target_code) {
4586 ia64_movl (code, GP_SCRATCH_REG, item->value.target_code);
4588 ia64_movl (code, GP_SCRATCH_REG, &(vtable->vtable [item->value.vtable_slot]));
4589 ia64_ld8 (code, GP_SCRATCH_REG, GP_SCRATCH_REG);
4591 ia64_mov_to_br (code, IA64_B6, GP_SCRATCH_REG);
4592 ia64_br_cond_reg (code, IA64_B6);
4595 ia64_begin_bundle (code);
4596 ia64_patch (item->jmp_code, (guint8*)code.buf + code.nins);
4597 ia64_movl (code, GP_SCRATCH_REG, fail_tramp);
4598 ia64_mov_to_br (code, IA64_B6, GP_SCRATCH_REG);
4599 ia64_br_cond_reg (code, IA64_B6);
4600 item->jmp_code = NULL;
4603 /* enable the commented code to assert on wrong method */
4604 #if ENABLE_WRONG_METHOD_CHECK
4605 g_assert_not_reached ();
4607 ia64_movl (code, GP_SCRATCH_REG, &(vtable->vtable [item->value.vtable_slot]));
4608 ia64_ld8 (code, GP_SCRATCH_REG, GP_SCRATCH_REG);
4609 ia64_mov_to_br (code, IA64_B6, GP_SCRATCH_REG);
4610 ia64_br_cond_reg (code, IA64_B6);
4611 #if ENABLE_WRONG_METHOD_CHECK
4612 g_assert_not_reached ();
4616 ia64_movl (code, GP_SCRATCH_REG, item->key);
4617 ia64_cmp_geu (code, 6, 7, IA64_R9, GP_SCRATCH_REG);
4618 item->jmp_code = (guint8*)code.buf + code.nins;
4619 ia64_br_cond_pred (code, 6, 0);
4622 /* patch the branches to get to the target items */
4623 for (i = 0; i < count; ++i) {
4624 MonoIMTCheckItem *item = imt_entries [i];
4625 if (item->jmp_code) {
4626 if (item->check_target_idx) {
4627 ia64_patch (item->jmp_code, imt_entries [item->check_target_idx]->code_target);
4632 ia64_codegen_close (code);
4633 g_assert (code.buf - buf <= size);
4635 size = code.buf - buf;
4637 start = mono_method_alloc_generic_virtual_thunk (domain, size + 16);
4638 start = (gpointer)ALIGN_TO (start, 16);
4640 start = mono_domain_code_reserve (domain, size);
4642 memcpy (start, buf, size);
4644 mono_arch_flush_icache (start, size);
4646 mono_stats.imt_thunks_size += size;
4652 mono_arch_find_imt_method (mgreg_t *regs, guint8 *code)
4654 return (MonoMethod*)regs [IA64_R9];
4658 mono_arch_emit_imt_argument (MonoCompile *cfg, MonoCallInst *call, MonoInst *imt_arg)
4660 /* Done by the implementation of the CALL_MEMBASE opcodes */
4665 mono_arch_get_this_arg_from_call (mgreg_t *regs, guint8 *code)
4667 return (gpointer)regs [IA64_R10];
4671 mono_arch_get_delegate_invoke_impl (MonoMethodSignature *sig, gboolean has_target)
4677 mono_arch_emit_inst_for_method (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature *fsig, MonoInst **args)
4679 MonoInst *ins = NULL;
4681 if (cmethod->klass->image == mono_defaults.corlib &&
4682 (strcmp (cmethod->klass->name_space, "System.Threading") == 0) &&
4683 (strcmp (cmethod->klass->name, "Interlocked") == 0)) {
4686 * We don't use the generic version in mini_emit_inst_for_method () since we
4687 * ia64 has atomic_add_imm opcodes.
4689 if (strcmp (cmethod->name, "Increment") == 0) {
4692 if (fsig->params [0]->type == MONO_TYPE_I4)
4693 opcode = OP_ATOMIC_ADD_IMM_NEW_I4;
4694 else if (fsig->params [0]->type == MONO_TYPE_I8)
4695 opcode = OP_ATOMIC_ADD_IMM_NEW_I8;
4697 g_assert_not_reached ();
4698 MONO_INST_NEW (cfg, ins, opcode);
4699 ins->dreg = mono_alloc_preg (cfg);
4701 ins->inst_basereg = args [0]->dreg;
4702 ins->inst_offset = 0;
4703 MONO_ADD_INS (cfg->cbb, ins);
4704 } else if (strcmp (cmethod->name, "Decrement") == 0) {
4707 if (fsig->params [0]->type == MONO_TYPE_I4)
4708 opcode = OP_ATOMIC_ADD_IMM_NEW_I4;
4709 else if (fsig->params [0]->type == MONO_TYPE_I8)
4710 opcode = OP_ATOMIC_ADD_IMM_NEW_I8;
4712 g_assert_not_reached ();
4713 MONO_INST_NEW (cfg, ins, opcode);
4714 ins->dreg = mono_alloc_preg (cfg);
4716 ins->inst_basereg = args [0]->dreg;
4717 ins->inst_offset = 0;
4718 MONO_ADD_INS (cfg->cbb, ins);
4719 } else if (strcmp (cmethod->name, "Add") == 0) {
4721 gboolean is_imm = FALSE;
4724 if ((args [1]->opcode == OP_ICONST) || (args [1]->opcode == OP_I8CONST)) {
4725 imm = (args [1]->opcode == OP_ICONST) ? args [1]->inst_c0 : args [1]->inst_l;
4727 is_imm = (imm == 1 || imm == 4 || imm == 8 || imm == 16 || imm == -1 || imm == -4 || imm == -8 || imm == -16);
4731 if (fsig->params [0]->type == MONO_TYPE_I4)
4732 opcode = OP_ATOMIC_ADD_IMM_NEW_I4;
4733 else if (fsig->params [0]->type == MONO_TYPE_I8)
4734 opcode = OP_ATOMIC_ADD_IMM_NEW_I8;
4736 g_assert_not_reached ();
4738 MONO_INST_NEW (cfg, ins, opcode);
4739 ins->dreg = mono_alloc_ireg (cfg);
4740 ins->inst_basereg = args [0]->dreg;
4741 ins->inst_offset = 0;
4742 ins->inst_imm = imm;
4743 ins->type = (opcode == OP_ATOMIC_ADD_IMM_NEW_I4) ? STACK_I4 : STACK_I8;
4745 if (fsig->params [0]->type == MONO_TYPE_I4)
4746 opcode = OP_ATOMIC_ADD_NEW_I4;
4747 else if (fsig->params [0]->type == MONO_TYPE_I8)
4748 opcode = OP_ATOMIC_ADD_NEW_I8;
4750 g_assert_not_reached ();
4752 MONO_INST_NEW (cfg, ins, opcode);
4753 ins->dreg = mono_alloc_ireg (cfg);
4754 ins->inst_basereg = args [0]->dreg;
4755 ins->inst_offset = 0;
4756 ins->sreg2 = args [1]->dreg;
4757 ins->type = (opcode == OP_ATOMIC_ADD_NEW_I4) ? STACK_I4 : STACK_I8;
4759 MONO_ADD_INS (cfg->cbb, ins);
4767 mono_arch_print_tree (MonoInst *tree, int arity)
4773 mono_arch_get_domain_intrinsic (MonoCompile* cfg)
4775 return mono_get_domain_intrinsic (cfg);
4779 mono_arch_context_get_int_reg (MonoContext *ctx, int reg)
4781 /* FIXME: implement */
4782 g_assert_not_reached ();