#include <unistd.h>
#include <sys/mman.h>
+#ifdef __INTEL_COMPILER
+#include <ia64intrin.h>
+#endif
+
#include <mono/metadata/appdomain.h>
#include <mono/metadata/debug-helpers.h>
#include <mono/metadata/threads.h>
#include "inssel.h"
#include "cpu-ia64.h"
-static gint lmf_tls_offset = -1;
static gint appdomain_tls_offset = -1;
static gint thread_tls_offset = -1;
* - compare instructions allways set p6 and p7
*/
-#define SIGNAL_STACK_SIZE (64 * 1024)
+/*
+ * There are a lot of places where generated code is disassembled/patched.
+ * The automatic bundling of instructions done by the code generation macros
+ * could complicate things, so it is best to call
+ * ia64_codegen_set_one_ins_per_bundle () at those places.
+ */
#define ARGS_OFFSET 16
return "unknown";
}
+G_GNUC_UNUSED static void
+break_count (void)
+{
+}
+
+G_GNUC_UNUSED static gboolean
+debug_count (void)
+{
+ static int count = 0;
+ count ++;
+
+ if (count == atoi (getenv ("COUNT"))) {
+ break_count ();
+ }
+
+ if (count > atoi (getenv ("COUNT"))) {
+ return FALSE;
+ }
+
+ return TRUE;
+}
+
+static gboolean
+debug_ins_sched (void)
+{
+#if 0
+ return debug_count ();
+#else
+ return TRUE;
+#endif
+}
+
+static gboolean
+debug_omit_fp (void)
+{
+#if 0
+ return debug_count ();
+#else
+ return TRUE;
+#endif
+}
+
+static void
+ia64_patch (unsigned char* code, gpointer target);
+
typedef enum {
ArgInIReg,
ArgInFloatReg,
ArgOnStack,
ArgValuetypeAddrInIReg,
ArgAggregate,
- ArgAggregateReturnInReg,
- ArgNone /* only in pair_storage */
+ ArgSingleHFA,
+ ArgDoubleHFA,
+ ArgNone
} ArgStorage;
+typedef enum {
+ AggregateNormal,
+ AggregateSingleHFA,
+ AggregateDoubleHFA
+} AggregateType;
+
typedef struct {
gint16 offset;
gint8 reg;
ArgStorage storage;
- /* Only if storage == ArgAggregate/ArgAggregateReturnInReg */
- int nregs;
-
/* Only if storage == ArgAggregate */
- int nslots;
+ int nregs, nslots;
+ AggregateType atype;
} ArgInfo;
typedef struct {
g_assert (info);
for (i = 0; i < info->num_fields; ++i) {
guint32 ftype = info->fields [i].field->type->type;
- if ((info->fields [i].field->type->byref) &&
+ if (!(info->fields [i].field->type->byref) &&
((ftype == MONO_TYPE_R4) || (ftype == MONO_TYPE_R8))) {
if (hfa_type == 0)
hfa_type = ftype;
else if (hfa_type != ftype)
is_hfa = FALSE;
}
+ else
+ is_hfa = FALSE;
}
if (hfa_type == 0)
is_hfa = FALSE;
- if (is_hfa)
- NOT_IMPLEMENTED;
+ ainfo->storage = ArgAggregate;
+ ainfo->atype = AggregateNormal;
+
+ if (is_hfa) {
+ ainfo->atype = hfa_type == MONO_TYPE_R4 ? AggregateSingleHFA : AggregateDoubleHFA;
+ if (is_return) {
+ if (info->num_fields <= 8) {
+ ainfo->reg = 8;
+ ainfo->nregs = info->num_fields;
+ ainfo->nslots = ainfo->nregs;
+ return;
+ }
+ /* Fall through */
+ }
+ else {
+ if ((*fr) + info->num_fields > 8)
+ NOT_IMPLEMENTED;
+
+ ainfo->reg = 8 + (*fr);
+ ainfo->nregs = info->num_fields;
+ ainfo->nslots = ainfo->nregs;
+ (*fr) += info->num_fields;
+ return;
+ }
+ }
/* This also handles returning of TypedByRef used by some icalls */
if (is_return) {
if (size <= 32) {
- ainfo->storage = ArgAggregateReturnInReg;
ainfo->reg = IA64_R8;
ainfo->nregs = (size + 7) / 8;
+ ainfo->nslots = ainfo->nregs;
return;
}
NOT_IMPLEMENTED;
}
- ainfo->storage = ArgAggregate;
ainfo->reg = (*gr);
ainfo->offset = *stack_size;
ainfo->nslots = (size + 7) / 8;
cinfo->ret.storage = ArgInFloatReg;
cinfo->ret.reg = 8;
break;
+ case MONO_TYPE_GENERICINST:
+ if (!mono_type_generic_inst_is_valuetype (sig->ret)) {
+ cinfo->ret.storage = ArgInIReg;
+ cinfo->ret.reg = IA64_R8;
+ break;
+ }
+ /* Fall through */
case MONO_TYPE_VALUETYPE:
case MONO_TYPE_TYPEDBYREF: {
guint32 tmp_gr = 0, tmp_fr = 0, tmp_stacksize = 0;
break;
}
case MONO_TYPE_VOID:
+ cinfo->ret.storage = ArgNone;
break;
default:
g_error ("Can't handle as return value 0x%x", sig->ret->type);
case MONO_TYPE_ARRAY:
add_general (&gr, &stack_size, ainfo);
break;
+ case MONO_TYPE_GENERICINST:
+ if (!mono_type_generic_inst_is_valuetype (sig->params [i])) {
+ add_general (&gr, &stack_size, ainfo);
+ break;
+ }
+ /* Fall through */
case MONO_TYPE_VALUETYPE:
case MONO_TYPE_TYPEDBYREF:
/* FIXME: */
int
mono_arch_get_argument_info (MonoMethodSignature *csig, int param_count, MonoJitArgumentInfo *arg_info)
{
- g_assert_not_reached ();
+ int k;
+ CallInfo *cinfo = get_call_info (csig, FALSE);
+ guint32 args_size = cinfo->stack_usage;
- return 0;
+ /* The arguments are saved to a stack area in mono_arch_instrument_prolog */
+ if (csig->hasthis) {
+ arg_info [0].offset = 0;
+ }
+
+ for (k = 0; k < param_count; k++) {
+ arg_info [k + 1].offset = ((k + csig->hasthis) * 8);
+ /* FIXME: */
+ arg_info [k + 1].size = 0;
+ }
+
+ g_free (cinfo);
+
+ return args_size;
}
/*
case MONO_TYPE_U:
case MONO_TYPE_PTR:
case MONO_TYPE_FNPTR:
+ case MONO_TYPE_BOOLEAN:
return TRUE;
case MONO_TYPE_OBJECT:
case MONO_TYPE_STRING:
case MONO_TYPE_SZARRAY:
case MONO_TYPE_ARRAY:
return TRUE;
+ case MONO_TYPE_GENERICINST:
+ if (!mono_type_generic_inst_is_valuetype (t))
+ return TRUE;
+ return FALSE;
case MONO_TYPE_VALUETYPE:
return FALSE;
}
{
GList *vars = NULL;
int i;
+ MonoMethodSignature *sig;
+ MonoMethodHeader *header;
+ CallInfo *cinfo;
+
+ header = mono_method_get_header (cfg->method);
+
+ sig = mono_method_signature (cfg->method);
+
+ cinfo = get_call_info (sig, FALSE);
+
+ for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
+ MonoInst *ins = cfg->varinfo [i];
+
+ ArgInfo *ainfo = &cinfo->args [i];
+
+ if (ins->flags & (MONO_INST_IS_DEAD|MONO_INST_VOLATILE|MONO_INST_INDIRECT))
+ continue;
+
+ if (ainfo->storage == ArgInIReg) {
+ /* The input registers are non-volatile */
+ ins->opcode = OP_REGVAR;
+ ins->dreg = 32 + ainfo->reg;
+ }
+ }
for (i = 0; i < cfg->num_varinfo; i++) {
MonoInst *ins = cfg->varinfo [i];
mono_ia64_alloc_stacked_registers (MonoCompile *cfg)
{
CallInfo *cinfo;
+ guint32 reserved_regs;
+ MonoMethodHeader *header;
if (cfg->arch.reg_local0 > 0)
/* Already done */
cinfo = get_call_info (mono_method_signature (cfg->method), FALSE);
- /* Three registers are reserved for use by the prolog/epilog */
+ header = mono_method_get_header (cfg->method);
+
+ /* Some registers are reserved for use by the prolog/epilog */
+ reserved_regs = header->num_clauses ? 4 : 3;
+
+ if ((mono_jit_trace_calls != NULL && mono_trace_eval (cfg->method)) ||
+ (cfg->prof_options & MONO_PROFILE_ENTER_LEAVE)) {
+ /* One registers is needed by instrument_epilog to save the return value */
+ reserved_regs ++;
+ if (cinfo->reg_usage < 2)
+ /* Number of arguments passed to function call in instrument_prolog */
+ cinfo->reg_usage = 2;
+ }
+
cfg->arch.reg_in0 = 32;
- cfg->arch.reg_local0 = cfg->arch.reg_in0 + cinfo->reg_usage + 3;
- cfg->arch.reg_out0 = cfg->arch.reg_local0 + 8;
+ cfg->arch.reg_local0 = cfg->arch.reg_in0 + cinfo->reg_usage + reserved_regs;
+ cfg->arch.reg_out0 = cfg->arch.reg_local0 + 16;
cfg->arch.reg_saved_ar_pfs = cfg->arch.reg_local0 - 1;
cfg->arch.reg_saved_b0 = cfg->arch.reg_local0 - 2;
- cfg->arch.reg_saved_sp = cfg->arch.reg_local0 - 3;
+ cfg->arch.reg_fp = cfg->arch.reg_local0 - 3;
+
+ /*
+ * Frames without handlers save sp to fp, frames with handlers save it into
+ * a dedicated register.
+ */
+ if (header->num_clauses)
+ cfg->arch.reg_saved_sp = cfg->arch.reg_local0 - 4;
+ else
+ cfg->arch.reg_saved_sp = cfg->arch.reg_fp;
+
+ if ((mono_jit_trace_calls != NULL && mono_trace_eval (cfg->method)) ||
+ (cfg->prof_options & MONO_PROFILE_ENTER_LEAVE)) {
+ cfg->arch.reg_saved_return_val = cfg->arch.reg_local0 - reserved_regs;
+ }
/*
* Need to allocate at least 2 out register for use by CEE_THROW / the system
gint32 *offsets;
CallInfo *cinfo;
- mono_ia64_alloc_stacked_registers (cfg);
-
header = mono_method_get_header (cfg->method);
sig = mono_method_signature (cfg->method);
cinfo = get_call_info (sig, FALSE);
+ /*
+ * Determine whenever the frame pointer can be eliminated.
+ * FIXME: Remove some of the restrictions.
+ */
+ cfg->arch.omit_fp = TRUE;
+
+ if (!debug_omit_fp ())
+ cfg->arch.omit_fp = FALSE;
+
+ if (cfg->flags & MONO_CFG_HAS_ALLOCA)
+ cfg->arch.omit_fp = FALSE;
+ if (header->num_clauses)
+ cfg->arch.omit_fp = FALSE;
+ if (cfg->param_area)
+ cfg->arch.omit_fp = FALSE;
+ for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
+ ArgInfo *ainfo = &cinfo->args [i];
+
+ if (ainfo->storage == ArgOnStack) {
+ /*
+ * The stack offset can only be determined when the frame
+ * size is known.
+ */
+ cfg->arch.omit_fp = FALSE;
+ }
+ }
+
+ mono_ia64_alloc_stacked_registers (cfg);
+
/*
* We use the ABI calling conventions for managed code as well.
* Exception: valuetypes are never passed or returned in registers.
*/
- /* Locals are allocated backwards from %fp */
- cfg->frame_reg = cfg->arch.reg_saved_sp;
- offset = 0;
+ if (cfg->arch.omit_fp) {
+ cfg->frame_reg = IA64_SP;
+ offset = ARGS_OFFSET;
+ }
+ else {
+ /* Locals are allocated backwards from %fp */
+ cfg->frame_reg = cfg->arch.reg_fp;
+ offset = 0;
+ }
if (cfg->method->save_lmf) {
- /* FIXME: */
-#if 0
- /* Reserve stack space for saving LMF + argument regs */
- offset += sizeof (MonoLMF);
- cfg->arch.lmf_offset = offset;
-#endif
+ /* No LMF on IA64 */
}
if (sig->ret->type != MONO_TYPE_VOID) {
cfg->ret->opcode = OP_REGVAR;
cfg->ret->inst_c0 = cfg->arch.reg_in0 + cinfo->ret.reg;
break;
+ case ArgAggregate:
+ /* Allocate a local to hold the result, the epilog will copy it to the correct place */
+ if (cfg->arch.omit_fp)
+ g_assert_not_reached ();
+ offset = ALIGN_TO (offset, 8);
+ offset += cinfo->ret.nslots * 8;
+ cfg->ret->opcode = OP_REGOFFSET;
+ cfg->ret->inst_basereg = cfg->frame_reg;
+ cfg->ret->inst_offset = - offset;
+ break;
default:
g_assert_not_reached ();
}
}
/* Allocate locals */
- offsets = mono_allocate_stack_slots (cfg, &locals_stack_size, &locals_stack_align);
+ offsets = mono_allocate_stack_slots_full (cfg, cfg->arch.omit_fp ? FALSE : TRUE, &locals_stack_size, &locals_stack_align);
if (locals_stack_align) {
offset = ALIGN_TO (offset, locals_stack_align);
}
MonoInst *inst = cfg->varinfo [i];
inst->opcode = OP_REGOFFSET;
inst->inst_basereg = cfg->frame_reg;
- inst->inst_offset = - (offset + offsets [i]);
+ if (cfg->arch.omit_fp)
+ inst->inst_offset = (offset + offsets [i]);
+ else
+ inst->inst_offset = - (offset + offsets [i]);
// printf ("allocated local %d to ", i); mono_print_tree_nl (inst);
}
}
offset += locals_stack_size;
if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG)) {
+ if (cfg->arch.omit_fp)
+ g_assert_not_reached ();
g_assert (cinfo->sig_cookie.storage == ArgOnStack);
cfg->sig_cookie = cinfo->sig_cookie.offset + ARGS_OFFSET;
}
inreg = FALSE;
break;
case ArgOnStack:
+ if (cfg->arch.omit_fp)
+ g_assert_not_reached ();
inst->opcode = OP_REGOFFSET;
inst->inst_basereg = cfg->frame_reg;
inst->inst_offset = ARGS_OFFSET + ainfo->offset;
break;
case ArgAggregate:
+ inreg = FALSE;
break;
default:
NOT_IMPLEMENTED;
inst->opcode = OP_REGOFFSET;
inst->inst_basereg = cfg->frame_reg;
/* These arguments are saved to the stack in the prolog */
- if (ainfo->storage == ArgAggregate) {
- NOT_IMPLEMENTED;
- offset += 2 * sizeof (gpointer);
- }
- else
+ switch (ainfo->storage) {
+ case ArgAggregate:
+ if (ainfo->atype == AggregateSingleHFA)
+ offset += ainfo->nslots * 4;
+ else
+ offset += ainfo->nslots * 8;
+ break;
+ default:
offset += sizeof (gpointer);
+ break;
+ }
offset = ALIGN_TO (offset, sizeof (gpointer));
- inst->inst_offset = - offset;
+ if (cfg->arch.omit_fp)
+ inst->inst_offset = offset;
+ else
+ inst->inst_offset = - offset;
}
}
}
+ if (cfg->arch.omit_fp && offset == 16)
+ offset = 0;
+
cfg->stack_offset = offset;
g_free (cinfo);
void
mono_arch_create_vars (MonoCompile *cfg)
{
- g_assert_not_reached ();
+ MonoMethodSignature *sig;
+ CallInfo *cinfo;
+
+ sig = mono_method_signature (cfg->method);
+
+ cinfo = get_call_info (sig, FALSE);
+
+ if (cinfo->ret.storage == ArgAggregate)
+ cfg->ret_var_is_local = TRUE;
+
+ g_free (cinfo);
}
static void
cinfo = get_call_info (sig, sig->pinvoke);
- if (cinfo->ret.storage == ArgAggregateReturnInReg) {
+ if (cinfo->ret.storage == ArgAggregate) {
/* The code in emit_this_vret_arg needs a local */
cfg->arch.ret_var_addr_local = mono_compile_create_var (cfg, &mono_defaults.int_class->byval_arg, OP_LOCAL);
((MonoInst*)cfg->arch.ret_var_addr_local)->flags |= MONO_INST_VOLATILE;
if ((i >= sig->hasthis) && (MONO_TYPE_ISSTRUCT(arg_type))) {
MonoInst *stack_addr;
- gint align;
+ guint32 align;
guint32 size;
if (arg_type->type == MONO_TYPE_TYPEDBYREF) {
else
if (sig->pinvoke)
size = mono_type_native_stack_size (&in->klass->byval_arg, &align);
- else
- size = mono_type_stack_size (&in->klass->byval_arg, &align);
+ else {
+ /*
+ * Other backends use mono_type_stack_size (), but that
+ * aligns the size to 8, which is larger than the size of
+ * the source, leading to reads of invalid memory if the
+ * source is at the end of address space.
+ */
+ size = mono_class_value_size (in->klass, &align);
+ }
- /*
- * FIXME: The destination is 'size' long, but the source might
- * be smaller.
- */
if (ainfo->storage == ArgAggregate) {
MonoInst *vtaddr, *load, *load2, *offset_ins, *set_reg;
int slot;
* Part of the structure is passed in registers.
*/
for (i = 0; i < ainfo->nregs; ++i) {
+ int offset, load_op, dest_reg, arg_storage;
+
slot = ainfo->reg + i;
+
+ if (ainfo->atype == AggregateSingleHFA) {
+ load_op = CEE_LDIND_R4;
+ offset = i * 4;
+ dest_reg = ainfo->reg + i;
+ arg_storage = ArgInFloatReg;
+ } else if (ainfo->atype == AggregateDoubleHFA) {
+ load_op = CEE_LDIND_R8;
+ offset = i * 8;
+ dest_reg = ainfo->reg + i;
+ arg_storage = ArgInFloatReg;
+ } else {
+ load_op = CEE_LDIND_I;
+ offset = i * 8;
+ dest_reg = cfg->arch.reg_out0 + ainfo->reg + i;
+ arg_storage = ArgInIReg;
+ }
MONO_INST_NEW (cfg, load, CEE_LDIND_I);
load->ssa_op = MONO_SSA_LOAD;
load->inst_i0 = (cfg)->varinfo [vtaddr->inst_c0];
- NEW_ICONST (cfg, offset_ins, (i * sizeof (gpointer)));
+ NEW_ICONST (cfg, offset_ins, offset);
MONO_INST_NEW (cfg, load2, CEE_ADD);
load2->inst_left = load;
load2->inst_right = offset_ins;
- MONO_INST_NEW (cfg, load, CEE_LDIND_I);
+ MONO_INST_NEW (cfg, load, load_op);
load->inst_left = load2;
if (i == 0)
set_reg = arg;
else
MONO_INST_NEW (cfg, set_reg, OP_OUTARG_REG);
- add_outarg_reg (cfg, call, set_reg, ArgInIReg, cfg->arch.reg_out0 + slot, load);
-
+ add_outarg_reg (cfg, call, set_reg, arg_storage, dest_reg, load);
if (set_reg != call->out_args) {
set_reg->next = call->out_args;
call->out_args = set_reg;
}
break;
case OP_MUL_IMM:
+ case OP_IMUL_IMM:
/* remove unnecessary multiplication with 1 */
if (ins->inst_imm == 1) {
if (ins->dreg != ins->sreg1) {
case OP_STOREI8_MEMBASE_IMM:
case OP_STORE_MEMBASE_IMM:
/* There are no store_membase instructions on ia64 */
- NEW_INS (cfg, temp, OP_I8CONST);
- temp->inst_c0 = ins->inst_offset;
- temp->dreg = mono_regstate_next_int (cfg->rs);
- NEW_INS (cfg, temp2, CEE_ADD);
- temp2->sreg1 = ins->inst_destbasereg;
- temp2->sreg2 = temp->dreg;
- temp2->dreg = mono_regstate_next_int (cfg->rs);
+ if (ia64_is_imm14 (ins->inst_offset)) {
+ NEW_INS (cfg, temp2, OP_ADD_IMM);
+ temp2->sreg1 = ins->inst_destbasereg;
+ temp2->inst_imm = ins->inst_offset;
+ temp2->dreg = mono_regstate_next_int (cfg->rs);
+ }
+ else {
+ NEW_INS (cfg, temp, OP_I8CONST);
+ temp->inst_c0 = ins->inst_offset;
+ temp->dreg = mono_regstate_next_int (cfg->rs);
+ NEW_INS (cfg, temp2, CEE_ADD);
+ temp2->sreg1 = ins->inst_destbasereg;
+ temp2->sreg2 = temp->dreg;
+ temp2->dreg = mono_regstate_next_int (cfg->rs);
+ }
switch (ins->opcode) {
case OP_STOREI1_MEMBASE_IMM:
ins->inst_offset = 0;
ins->inst_basereg = temp2->dreg;
break;
- case OP_FCALL_MEMBASE:
- case OP_LCALL_MEMBASE:
- case OP_VCALL_MEMBASE:
- case OP_VOIDCALL_MEMBASE:
- case OP_CALL_MEMBASE:
+ case OP_IA64_FETCHADD4_IMM:
+ case OP_IA64_FETCHADD8_IMM:
+ case OP_ATOMIC_EXCHANGE_I4:
+ case OP_ATOMIC_EXCHANGE_I8:
/* There are no membase instructions on ia64 */
if (ia64_is_imm14 (ins->inst_offset)) {
NEW_INS (cfg, temp2, OP_ADD_IMM);
- temp2->sreg1 = ins->sreg1;
+ temp2->sreg1 = ins->inst_basereg;
temp2->inst_imm = ins->inst_offset;
temp2->dreg = mono_regstate_next_int (cfg->rs);
}
temp->inst_c0 = ins->inst_offset;
temp->dreg = mono_regstate_next_int (cfg->rs);
NEW_INS (cfg, temp2, CEE_ADD);
- temp2->sreg1 = ins->sreg1;
+ temp2->sreg1 = ins->inst_basereg;
temp2->sreg2 = temp->dreg;
temp2->dreg = mono_regstate_next_int (cfg->rs);
}
-
- NEW_INS (cfg, temp, OP_LOADI8_MEMBASE);
- temp->sreg1 = temp2->dreg;
- temp->dreg = mono_regstate_next_int (cfg->rs);
-
- ins->sreg1 = temp->dreg;
-
- switch (ins->opcode) {
- case OP_FCALL_MEMBASE:
- ins->opcode = OP_FCALL_REG;
- break;
- case OP_LCALL_MEMBASE:
- ins->opcode = OP_LCALL_REG;
- break;
- case OP_VCALL_MEMBASE:
- ins->opcode = OP_VCALL_REG;
- break;
- case OP_VOIDCALL_MEMBASE:
- ins->opcode = OP_VOIDCALL_REG;
- break;
- case OP_CALL_MEMBASE:
- ins->opcode = OP_CALL_REG;
- break;
- default:
- g_assert_not_reached ();
- }
+ ins->inst_offset = 0;
+ ins->inst_basereg = temp2->dreg;
break;
case OP_ADD_IMM:
case OP_IADD_IMM:
gboolean is_imm = FALSE;
gboolean switched = FALSE;
+ if (ins->opcode == OP_AND_IMM && ins->inst_imm == 255) {
+ ins->opcode = OP_ZEXT_I1;
+ break;
+ }
+
switch (ins->opcode) {
case OP_ADD_IMM:
case OP_IADD_IMM:
}
break;
}
- case OP_MUL_IMM: {
- /* This should be emulated, but rules in inssel.brg generate it */
+ case OP_MUL_IMM:
+ case OP_LMUL_IMM:
+ case OP_IMUL_IMM: {
int i, sum_reg;
+ gboolean found = FALSE;
+ int shl_op = ins->opcode == OP_IMUL_IMM ? OP_ISHL_IMM : OP_SHL_IMM;
/* First the easy cases */
if (ins->inst_imm == 1) {
}
for (i = 1; i < 64; ++i)
if (ins->inst_imm == (((gint64)1) << i)) {
- ins->opcode = OP_SHL_IMM;
+ ins->opcode = shl_op;
ins->inst_imm = i;
+ found = TRUE;
break;
}
/* This could be optimized */
- if (ins->opcode == OP_MUL_IMM) {
+ if (!found) {
sum_reg = 0;
for (i = 0; i < 64; ++i) {
if (ins->inst_imm & (((gint64)1) << i)) {
- NEW_INS (cfg, temp, OP_SHL_IMM);
+ NEW_INS (cfg, temp, shl_op);
temp->dreg = mono_regstate_next_int (cfg->rs);
temp->sreg1 = ins->sreg1;
temp->inst_imm = i;
break;
case OP_VCALL:
case OP_VCALL_REG:
- case OP_VCALL_MEMBASE:
+ case OP_VCALL_MEMBASE: {
+ ArgStorage storage;
+
cinfo = get_call_info (((MonoCallInst*)ins)->signature, FALSE);
- if (cinfo->ret.storage == ArgAggregateReturnInReg) {
+ storage = cinfo->ret.storage;
+
+ if (storage == ArgAggregate) {
MonoInst *local = (MonoInst*)cfg->arch.ret_var_addr_local;
/* Load address of stack space allocated for the return value */
ia64_ld8 (code, GP_SCRATCH_REG, GP_SCRATCH_REG);
for (i = 0; i < cinfo->ret.nregs; ++i) {
- ia64_st8_inc_imm_hint (code, GP_SCRATCH_REG, cinfo->ret.reg + i, 8, 0);
+ switch (cinfo->ret.atype) {
+ case AggregateNormal:
+ ia64_st8_inc_imm_hint (code, GP_SCRATCH_REG, cinfo->ret.reg + i, 8, 0);
+ break;
+ case AggregateSingleHFA:
+ ia64_stfs_inc_imm_hint (code, GP_SCRATCH_REG, cinfo->ret.reg + i, 4, 0);
+ break;
+ case AggregateDoubleHFA:
+ ia64_stfd_inc_imm_hint (code, GP_SCRATCH_REG, cinfo->ret.reg + i, 8, 0);
+ break;
+ default:
+ g_assert_not_reached ();
+ }
}
}
g_free (cinfo);
break;
+ }
default:
g_assert_not_reached ();
}
return code;
}
+#define add_patch_info(cfg,code,patch_type,data) do { \
+ mono_add_patch_info (cfg, code.buf + code.nins - cfg->native_code, patch_type, data); \
+} while (0)
+
+#define emit_cond_system_exception(cfg,code,exc_name,predicate) do { \
+ MonoInst *tins = mono_branch_optimize_exception_target (cfg, bb, exc_name); \
+ if (tins == NULL) \
+ add_patch_info (cfg, code, MONO_PATCH_INFO_EXC, exc_name); \
+ else \
+ add_patch_info (cfg, code, MONO_PATCH_INFO_BB, tins->inst_true_bb); \
+ ia64_br_cond_pred (code, (predicate), 0); \
+} while (0)
+
static Ia64CodegenState
emit_call (MonoCompile *cfg, Ia64CodegenState code, guint32 patch_type, gconstpointer data)
{
- mono_add_patch_info (cfg, code.buf - cfg->native_code, patch_type, data);
+ add_patch_info (cfg, code, patch_type, data);
if ((patch_type == MONO_PATCH_INFO_ABS) || (patch_type == MONO_PATCH_INFO_INTERNAL_METHOD)) {
/* Indirect call */
+ /* mono_arch_patch_callsite will patch this */
+ /* mono_arch_nullify_class_init_trampoline will patch this */
ia64_movl (code, GP_SCRATCH_REG, 0);
ia64_ld8_inc_imm (code, GP_SCRATCH_REG2, GP_SCRATCH_REG, 8);
ia64_mov_to_br (code, IA64_B6, GP_SCRATCH_REG2);
}
else {
/* Can't use a direct call since the displacement might be too small */
+ /* mono_arch_patch_callsite will patch this */
ia64_movl (code, GP_SCRATCH_REG, 0);
ia64_mov_to_br (code, IA64_B6, GP_SCRATCH_REG);
ia64_br_call_reg (code, IA64_B0, IA64_B6);
ia64_codegen_init (code, code_start);
+#if 0
+ if (strstr (cfg->method->name, "conv_ovf_i1") && (bb->block_num == 2))
+ break_count ();
+#endif
+
ins = bb->code;
while (ins) {
offset = code.buf - cfg->native_code;
if (ins->inst_i0->inst_c0) {
NOT_IMPLEMENTED;
} else {
- ia64_begin_bundle (code);
- mono_add_patch_info (cfg, code.buf - cfg->native_code, MONO_PATCH_INFO_LABEL, ins->inst_i0);
+ add_patch_info (cfg, code, MONO_PATCH_INFO_LABEL, ins->inst_i0);
ia64_br_cond_pred (code, pred, 0);
}
} else {
if (ins->inst_target_bb->native_offset) {
- gint64 disp = ((gint64)ins->inst_target_bb->native_offset - offset) >> 4;
- ia64_br_cond_pred (code, pred, disp);
- } else {
+ guint8 *pos = code.buf + code.nins;
+
+ ia64_br_cond_pred (code, pred, 0);
ia64_begin_bundle (code);
- mono_add_patch_info (cfg, code.buf - cfg->native_code, MONO_PATCH_INFO_BB, ins->inst_target_bb);
+ ia64_patch (pos, cfg->native_code + ins->inst_target_bb->native_offset);
+ } else {
+ add_patch_info (cfg, code, MONO_PATCH_INFO_BB, ins->inst_target_bb);
ia64_br_cond_pred (code, pred, 0);
}
}
case OP_STORE_MEMBASE_REG:
ia64_st8_hint (code, ins->inst_destbasereg, ins->sreg1, 0);
break;
+
+ case OP_IA64_STOREI1_MEMBASE_INC_REG:
+ ia64_st1_inc_imm_hint (code, ins->inst_destbasereg, ins->sreg1, 1, 0);
+ break;
+ case OP_IA64_STOREI2_MEMBASE_INC_REG:
+ ia64_st2_inc_imm_hint (code, ins->inst_destbasereg, ins->sreg1, 2, 0);
+ break;
+ case OP_IA64_STOREI4_MEMBASE_INC_REG:
+ ia64_st4_inc_imm_hint (code, ins->inst_destbasereg, ins->sreg1, 4, 0);
+ break;
+ case OP_IA64_STOREI8_MEMBASE_INC_REG:
+ ia64_st8_inc_imm_hint (code, ins->inst_destbasereg, ins->sreg1, 8, 0);
+ break;
+
case OP_LOADU1_MEMBASE:
ia64_ld1 (code, ins->dreg, ins->inst_basereg);
break;
case OP_LOADI8_MEMBASE:
ia64_ld8 (code, ins->dreg, ins->inst_basereg);
break;
+
+ case OP_IA64_LOADU1_MEMBASE_INC:
+ ia64_ld1_inc_imm_hint (code, ins->dreg, ins->inst_basereg, 1, 0);
+ break;
+ case OP_IA64_LOADU2_MEMBASE_INC:
+ ia64_ld2_inc_imm_hint (code, ins->dreg, ins->inst_basereg, 2, 0);
+ break;
+ case OP_IA64_LOADU4_MEMBASE_INC:
+ ia64_ld4_inc_imm_hint (code, ins->dreg, ins->inst_basereg, 4, 0);
+ break;
+ case OP_IA64_LOADI8_MEMBASE_INC:
+ ia64_ld8_inc_imm_hint (code, ins->dreg, ins->inst_basereg, 8, 0);
+ break;
+
case OP_SEXT_I1:
ia64_sxt1 (code, ins->dreg, ins->sreg1);
break;
case OP_COND_EXC_IOV:
case OP_COND_EXC_OV:
- mono_add_patch_info (cfg, code.buf - cfg->native_code,
- MONO_PATCH_INFO_EXC, "OverflowException");
- ia64_br_cond_pred (code, 6, 0);
+ emit_cond_system_exception (cfg, code, "OverflowException", 6);
break;
case OP_COND_EXC_IC:
case OP_COND_EXC_C:
- mono_add_patch_info (cfg, code.buf - cfg->native_code,
- MONO_PATCH_INFO_EXC, "OverflowException");
- ia64_br_cond_pred (code, 7, 0);
+ emit_cond_system_exception (cfg, code, "OverflowException", 7);
break;
case OP_IA64_COND_EXC:
- mono_add_patch_info (cfg, code.buf - cfg->native_code,
- MONO_PATCH_INFO_EXC, ins->inst_p1);
- ia64_br_cond_pred (code, 6, 0);
+ emit_cond_system_exception (cfg, code, ins->inst_p1, 6);
break;
case OP_IA64_CSET:
- /* FIXME: Do this with one instruction ? */
- ia64_mov (code, ins->dreg, IA64_R0);
+ ia64_mov_pred (code, 7, ins->dreg, IA64_R0);
+ ia64_no_stop (code);
ia64_add1_pred (code, 6, ins->dreg, IA64_R0, IA64_R0);
break;
case CEE_CONV_I1:
else if (d == 1.0)
ia64_fmov (code, ins->dreg, 1);
else {
- mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_R8, ins->inst_p0);
+ add_patch_info (cfg, code, MONO_PATCH_INFO_R8, ins->inst_p0);
ia64_movl (code, GP_SCRATCH_REG, 0);
ia64_ldfd (code, ins->dreg, GP_SCRATCH_REG);
}
else if (f == 1.0)
ia64_fmov (code, ins->dreg, 1);
else {
- mono_add_patch_info (cfg, offset, MONO_PATCH_INFO_R4, ins->inst_p0);
+ add_patch_info (cfg, code, MONO_PATCH_INFO_R4, ins->inst_p0);
ia64_movl (code, GP_SCRATCH_REG, 0);
ia64_ldfs (code, ins->dreg, GP_SCRATCH_REG);
}
ia64_stfd_hint (code, ins->inst_destbasereg, ins->sreg1, 0);
break;
case OP_STORER4_MEMBASE_REG:
- ia64_stfs_hint (code, ins->inst_destbasereg, ins->sreg1, 0);
+ ia64_fnorm_s_sf (code, FP_SCRATCH_REG, ins->sreg1, 0);
+ ia64_stfs_hint (code, ins->inst_destbasereg, FP_SCRATCH_REG, 0);
break;
case OP_LOADR8_MEMBASE:
ia64_ldfd (code, ins->dreg, ins->inst_basereg);
break;
case OP_LOADR4_MEMBASE:
ia64_ldfs (code, ins->dreg, ins->inst_basereg);
+ ia64_fnorm_d_sf (code, ins->dreg, ins->dreg, 0);
break;
case CEE_CONV_R4:
ia64_setf_sig (code, ins->dreg, ins->sreg1);
case CEE_CKFINITE:
/* Quiet NaN */
ia64_fclass_m (code, 6, 7, ins->sreg1, 0x080);
- mono_add_patch_info (cfg, code.buf - cfg->native_code,
- MONO_PATCH_INFO_EXC, "ArithmeticException");
- ia64_br_cond_pred (code, 6, 0);
+ emit_cond_system_exception (cfg, code, "ArithmeticException", 6);
/* Signaling NaN */
ia64_fclass_m (code, 6, 7, ins->sreg1, 0x040);
- mono_add_patch_info (cfg, code.buf - cfg->native_code,
- MONO_PATCH_INFO_EXC, "ArithmeticException");
- ia64_br_cond_pred (code, 6, 0);
+ emit_cond_system_exception (cfg, code, "ArithmeticException", 6);
/* Positive infinity */
ia64_fclass_m (code, 6, 7, ins->sreg1, 0x021);
- mono_add_patch_info (cfg, code.buf - cfg->native_code,
- MONO_PATCH_INFO_EXC, "ArithmeticException");
- ia64_br_cond_pred (code, 6, 0);
+ emit_cond_system_exception (cfg, code, "ArithmeticException", 6);
/* Negative infinity */
ia64_fclass_m (code, 6, 7, ins->sreg1, 0x022);
- mono_add_patch_info (cfg, code.buf - cfg->native_code,
- MONO_PATCH_INFO_EXC, "ArithmeticException");
- ia64_br_cond_pred (code, 6, 0);
+ emit_cond_system_exception (cfg, code, "ArithmeticException", 6);
break;
/* Calls */
case OP_VOIDCALL_REG:
call = (MonoCallInst*)ins;
- if (call->virtual) {
- /* Keep this in synch with get_vcall_slot_addr */
-
- /* This is a virtual call */
- ia64_mov_to_br (code, IA64_B6, ins->sreg1);
-
- /*
- * This nop will tell get_vcall_slot_addr that this is a virtual
- * call.
- */
- ia64_nop_i (code, 0x12345);
- }
- else {
- /* Indirect call */
- ia64_mov (code, GP_SCRATCH_REG, ins->sreg1);
- ia64_ld8_inc_imm (code, GP_SCRATCH_REG2, GP_SCRATCH_REG, 8);
- ia64_mov_to_br (code, IA64_B6, GP_SCRATCH_REG2);
- ia64_ld8 (code, IA64_GP, GP_SCRATCH_REG);
- }
+ /* Indirect call */
+ /*
+ * mono_arch_patch_delegate_trampoline will patch this, this is why R8 is
+ * used.
+ */
+ ia64_mov (code, IA64_R8, ins->sreg1);
+ ia64_ld8_inc_imm (code, GP_SCRATCH_REG2, IA64_R8, 8);
+ ia64_mov_to_br (code, IA64_B6, GP_SCRATCH_REG2);
+ ia64_ld8 (code, IA64_GP, IA64_R8);
ia64_br_call_reg (code, IA64_B0, IA64_B6);
code = emit_move_return_value (cfg, ins, code);
break;
- case CEE_JMP: {
- /*
- * Keep in sync with the code in emit_epilog.
+
+ case OP_FCALL_MEMBASE:
+ case OP_LCALL_MEMBASE:
+ case OP_VCALL_MEMBASE:
+ case OP_VOIDCALL_MEMBASE:
+ case OP_CALL_MEMBASE:
+ /*
+ * There are no membase instructions on ia64, but we can't
+ * lower this since get_vcall_slot_addr () needs to decode it.
*/
- if (cfg->prof_options & MONO_PROFILE_ENTER_LEAVE)
- NOT_IMPLEMENTED;
+ /* Keep this in synch with get_vcall_slot_addr */
+ if (ia64_is_imm14 (ins->inst_offset))
+ ia64_adds_imm (code, IA64_R8, ins->inst_offset, ins->sreg1);
+ else {
+ ia64_movl (code, GP_SCRATCH_REG, ins->inst_offset);
+ ia64_add (code, IA64_R8, GP_SCRATCH_REG, ins->sreg1);
+ }
+
+ ia64_begin_bundle (code);
+ ia64_codegen_set_one_ins_per_bundle (code, TRUE);
+
+ ia64_ld8 (code, GP_SCRATCH_REG, IA64_R8);
+
+ ia64_mov_to_br (code, IA64_B6, GP_SCRATCH_REG);
+
+ /*
+ * This nop will tell get_vcall_slot_addr that this is a virtual
+ * call.
+ */
+ ia64_nop_i (code, 0x12345);
+
+ ia64_br_call_reg (code, IA64_B0, IA64_B6);
+
+ ia64_codegen_set_one_ins_per_bundle (code, FALSE);
+
+ code = emit_move_return_value (cfg, ins, code);
+ break;
+ case CEE_JMP: {
+ /*
+ * Keep in sync with the code in emit_epilog.
+ */
+
+ if (cfg->prof_options & MONO_PROFILE_ENTER_LEAVE)
+ NOT_IMPLEMENTED;
g_assert (!cfg->method->save_lmf);
/* Load arguments into their original registers */
code = emit_load_volatile_arguments (cfg, code);
- if (cfg->arch.stack_alloc_size)
- ia64_mov (code, IA64_SP, cfg->arch.reg_saved_sp);
+ if (cfg->arch.stack_alloc_size) {
+ if (cfg->arch.omit_fp) {
+ if (ia64_is_imm14 (cfg->arch.stack_alloc_size))
+ ia64_adds_imm (code, IA64_SP, (cfg->arch.stack_alloc_size), IA64_SP);
+ else {
+ ia64_movl (code, GP_SCRATCH_REG, cfg->arch.stack_alloc_size);
+ ia64_add (code, IA64_SP, GP_SCRATCH_REG, IA64_SP);
+ }
+ }
+ else
+ ia64_mov (code, IA64_SP, cfg->arch.reg_saved_sp);
+ }
ia64_mov_to_ar_i (code, IA64_PFS, cfg->arch.reg_saved_ar_pfs);
ia64_mov_ret_to_br (code, IA64_B0, cfg->arch.reg_saved_b0);
- mono_add_patch_info (cfg, code.buf - cfg->native_code, MONO_PATCH_INFO_METHOD_JUMP, ins->inst_p0);
+ add_patch_info (cfg, code, MONO_PATCH_INFO_METHOD_JUMP, ins->inst_p0);
ia64_movl (code, GP_SCRATCH_REG, 0);
ia64_mov_to_br (code, IA64_B6, GP_SCRATCH_REG);
ia64_br_cond_reg (code, IA64_B6);
break;
}
+ case CEE_BREAK:
+ code = emit_call (cfg, code, MONO_PATCH_INFO_ABS, mono_arch_break);
+ break;
case OP_LOCALLOC: {
gint32 abi_offset;
+ /* FIXME: Sigaltstack support */
+
/* keep alignment */
ia64_adds_imm (code, GP_SCRATCH_REG, MONO_ARCH_FRAME_ALIGNMENT - 1, ins->sreg1);
ia64_movl (code, GP_SCRATCH_REG2, ~(MONO_ARCH_FRAME_ALIGNMENT - 1));
/* Upper limit */
ia64_add (code, GP_SCRATCH_REG2, ins->dreg, GP_SCRATCH_REG);
+ ia64_codegen_set_one_ins_per_bundle (code, TRUE);
+
/* Init loop */
ia64_st8_inc_imm_hint (code, ins->dreg, IA64_R0, 8, 0);
ia64_cmp_lt (code, 8, 9, ins->dreg, GP_SCRATCH_REG2);
ia64_br_cond_pred (code, 8, -2);
+ ia64_codegen_set_one_ins_per_bundle (code, FALSE);
+
ia64_sub (code, ins->dreg, GP_SCRATCH_REG2, GP_SCRATCH_REG);
}
ia64_ld8 (code, ins->dreg, ins->dreg);
break;
+ /* Synchronization */
+ case OP_MEMORY_BARRIER:
+ ia64_mf (code);
+ break;
+ case OP_IA64_FETCHADD4_IMM:
+ g_assert (ins->inst_offset == 0);
+ ia64_fetchadd4_acq_hint (code, ins->dreg, ins->inst_basereg, ins->inst_imm, 0);
+ ia64_adds_imm (code, ins->dreg, ins->inst_imm, ins->dreg);
+ break;
+ case OP_IA64_FETCHADD8_IMM:
+ g_assert (ins->inst_offset == 0);
+ ia64_fetchadd8_acq_hint (code, ins->dreg, ins->inst_basereg, ins->inst_imm, 0);
+ ia64_adds_imm (code, ins->dreg, ins->inst_imm, ins->dreg);
+ break;
+ case OP_ATOMIC_EXCHANGE_I4:
+ ia64_xchg4_hint (code, ins->dreg, ins->inst_basereg, ins->sreg2, 0);
+ ia64_sxt4 (code, ins->dreg, ins->dreg);
+ break;
+ case OP_ATOMIC_EXCHANGE_I8:
+ ia64_xchg8_hint (code, ins->dreg, ins->inst_basereg, ins->sreg2, 0);
+ break;
+
/* Exception handling */
case OP_CALL_HANDLER:
/*
* save the return address to a register and use a
* branch.
*/
+ ia64_codegen_set_one_ins_per_bundle (code, TRUE);
ia64_mov (code, IA64_R15, IA64_R0);
ia64_mov_from_ip (code, GP_SCRATCH_REG);
/* Add the length of OP_CALL_HANDLER */
ia64_adds_imm (code, GP_SCRATCH_REG, 5 * 16, GP_SCRATCH_REG);
- mono_add_patch_info (cfg, code.buf - cfg->native_code, MONO_PATCH_INFO_BB, ins->inst_target_bb);
+ add_patch_info (cfg, code, MONO_PATCH_INFO_BB, ins->inst_target_bb);
ia64_movl (code, GP_SCRATCH_REG2, 0);
ia64_mov_to_br (code, IA64_B6, GP_SCRATCH_REG2);
ia64_br_cond_reg (code, IA64_B6);
+ ia64_codegen_set_one_ins_per_bundle (code, FALSE);
break;
case OP_START_HANDLER: {
/*
MonoInst *spvar = mono_find_spvar_for_region (cfg, bb->region);
/*
- * We might be called by the exception handling code, in which case the
- * the register stack is not set up correctly. So do it now.
+ * R15 determines our caller. It is used since it is writable using
+ * libunwind.
+ * R15 == 0 means we are called by OP_CALL_HANDLER or via resume_context ()
+ * R15 != 0 means we are called by call_filter ().
*/
- ia64_alloc (code, GP_SCRATCH_REG2, cfg->arch.reg_local0 - cfg->arch.reg_in0, cfg->arch.reg_out0 - cfg->arch.reg_local0, cfg->arch.n_out_regs, 0);
-
- /* Set the fp register from the value passed in by the caller */
- /* R15 is used since it is writable using libunwind */
- /* R15 == 0 means we are called by OP_CALL_HANDLER or via resume_context () */
+ ia64_codegen_set_one_ins_per_bundle (code, TRUE);
ia64_cmp_eq (code, 6, 7, IA64_R15, IA64_R0);
- ia64_add_pred (code, 7, cfg->frame_reg, IA64_R0, IA64_R15);
+ ia64_br_cond_pred (code, 6, 6);
+
+ /*
+ * Called by call_filter:
+ * Allocate a new stack frame, and set the fp register from the
+ * value passed in by the caller.
+ * We allocate a similar frame as is done by the prolog, so
+ * if an exception is thrown while executing the filter, the
+ * unwinder can unwind through the filter frame using the unwind
+ * info for the prolog.
+ */
+ 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);
+ ia64_mov_from_br (code, cfg->arch.reg_saved_b0, IA64_B0);
+ ia64_mov (code, cfg->arch.reg_saved_sp, IA64_SP);
+ ia64_mov (code, cfg->frame_reg, IA64_R15);
+ /* Signal to endfilter that we are called by call_filter */
+ ia64_mov (code, GP_SCRATCH_REG, IA64_R0);
+
+ /* Save the return address */
ia64_adds_imm (code, GP_SCRATCH_REG2, spvar->inst_offset, cfg->frame_reg);
ia64_st8_hint (code, GP_SCRATCH_REG2, GP_SCRATCH_REG, 0);
+ ia64_codegen_set_one_ins_per_bundle (code, FALSE);
break;
}
- case CEE_ENDFINALLY: {
- MonoInst *spvar = mono_find_spvar_for_region (cfg, bb->region);
- ia64_adds_imm (code, GP_SCRATCH_REG, spvar->inst_offset, cfg->frame_reg);
- ia64_ld8_hint (code, GP_SCRATCH_REG, GP_SCRATCH_REG, 0);
- ia64_mov_to_br (code, IA64_B6, GP_SCRATCH_REG);
- ia64_br_cond_reg (code, IA64_B6);
- break;
- }
+ case CEE_ENDFINALLY:
case OP_ENDFILTER: {
- /* FIXME: Return the value */
+ /* FIXME: Return the value in ENDFILTER */
MonoInst *spvar = mono_find_spvar_for_region (cfg, bb->region);
+
+ /* Load the return address */
ia64_adds_imm (code, GP_SCRATCH_REG, spvar->inst_offset, cfg->frame_reg);
ia64_ld8_hint (code, GP_SCRATCH_REG, GP_SCRATCH_REG, 0);
+
+ /* Test caller */
+ ia64_cmp_eq (code, 6, 7, GP_SCRATCH_REG, IA64_R0);
+ ia64_br_cond_pred (code, 7, 4);
+
+ /* Called by call_filter */
+ /* Pop frame */
+ ia64_mov_to_ar_i (code, IA64_PFS, cfg->arch.reg_saved_ar_pfs);
+ ia64_mov_to_br (code, IA64_B0, cfg->arch.reg_saved_b0);
+ ia64_br_ret_reg (code, IA64_B0);
+
+ /* Called by CALL_HANDLER */
ia64_mov_to_br (code, IA64_B6, GP_SCRATCH_REG);
ia64_br_cond_reg (code, IA64_B6);
break;
{ FALSE, FALSE, FALSE }
};
+static int last_stop_in_template [32] = {
+ -1, 2, 1, 2, -1, 2, -1, -1,
+ -1, 2, 0, 2, -1, 2, -1, 2,
+ -1, 2, -1, 2, -1, -1, -1, 2,
+ -1, 2, -1, -1, -1, 2, -1, -1
+};
+
+static guint64 nops_for_ins_types [6] = {
+ IA64_NOP_I,
+ IA64_NOP_I,
+ IA64_NOP_M,
+ IA64_NOP_F,
+ IA64_NOP_B,
+ IA64_NOP_X
+};
+
+#define ITYPE_MATCH(itype1, itype2) (((itype1) == (itype2)) || (((itype2) == IA64_INS_TYPE_A) && (((itype1) == IA64_INS_TYPE_I) || ((itype1) == IA64_INS_TYPE_M))))
+
+/*
+ * Debugging support
+ */
+
+#if 0
+#define DEBUG_INS_SCHED(a) do { a; } while (0)
+#else
+#define DEBUG_INS_SCHED(a)
+#endif
+
+static void
+ia64_analyze_deps (Ia64CodegenState *code, int *deps_start, int *stops)
+{
+ int i, pos, ins_index, current_deps_start, current_ins_start, reg;
+ guint8 *deps = code->dep_info;
+ gboolean need_stop, no_stop;
+
+ for (i = 0; i < code->nins; ++i)
+ stops [i] = FALSE;
+
+ ins_index = 0;
+ current_deps_start = 0;
+ current_ins_start = 0;
+ deps_start [ins_index] = current_ins_start;
+ pos = 0;
+ no_stop = FALSE;
+ DEBUG_INS_SCHED (printf ("BEGIN.\n"));
+ while (pos < code->dep_info_pos) {
+ need_stop = FALSE;
+ switch (deps [pos]) {
+ case IA64_END_OF_INS:
+ ins_index ++;
+ current_ins_start = pos + 2;
+ deps_start [ins_index] = current_ins_start;
+ no_stop = FALSE;
+ DEBUG_INS_SCHED (printf ("(%d) END INS.\n", ins_index - 1));
+ break;
+ case IA64_NONE:
+ break;
+ case IA64_READ_GR:
+ reg = deps [pos + 1];
+
+ DEBUG_INS_SCHED (printf ("READ GR: %d\n", reg));
+ for (i = current_deps_start; i < current_ins_start; i += 2)
+ if (deps [i] == IA64_WRITE_GR && deps [i + 1] == reg)
+ need_stop = TRUE;
+ break;
+ case IA64_WRITE_GR:
+ reg = code->dep_info [pos + 1];
+
+ DEBUG_INS_SCHED (printf ("WRITE GR: %d\n", reg));
+ for (i = current_deps_start; i < current_ins_start; i += 2)
+ if (deps [i] == IA64_WRITE_GR && deps [i + 1] == reg)
+ need_stop = TRUE;
+ break;
+ case IA64_READ_PR:
+ reg = deps [pos + 1];
+
+ DEBUG_INS_SCHED (printf ("READ PR: %d\n", reg));
+ for (i = current_deps_start; i < current_ins_start; i += 2)
+ if (((deps [i] == IA64_WRITE_PR) || (deps [i] == IA64_WRITE_PR_FLOAT)) && deps [i + 1] == reg)
+ need_stop = TRUE;
+ break;
+ case IA64_READ_PR_BRANCH:
+ reg = deps [pos + 1];
+
+ /* Writes to prs by non-float instructions are visible to branches */
+ DEBUG_INS_SCHED (printf ("READ PR BRANCH: %d\n", reg));
+ for (i = current_deps_start; i < current_ins_start; i += 2)
+ if (deps [i] == IA64_WRITE_PR_FLOAT && deps [i + 1] == reg)
+ need_stop = TRUE;
+ break;
+ case IA64_WRITE_PR:
+ reg = code->dep_info [pos + 1];
+
+ DEBUG_INS_SCHED (printf ("WRITE PR: %d\n", reg));
+ for (i = current_deps_start; i < current_ins_start; i += 2)
+ if (((deps [i] == IA64_WRITE_PR) || (deps [i] == IA64_WRITE_PR_FLOAT)) && deps [i + 1] == reg)
+ need_stop = TRUE;
+ break;
+ case IA64_WRITE_PR_FLOAT:
+ reg = code->dep_info [pos + 1];
+
+ DEBUG_INS_SCHED (printf ("WRITE PR FP: %d\n", reg));
+ for (i = current_deps_start; i < current_ins_start; i += 2)
+ if (((deps [i] == IA64_WRITE_GR) || (deps [i] == IA64_WRITE_PR_FLOAT)) && deps [i + 1] == reg)
+ need_stop = TRUE;
+ break;
+ case IA64_READ_BR:
+ reg = deps [pos + 1];
+
+ DEBUG_INS_SCHED (printf ("READ BR: %d\n", reg));
+ for (i = current_deps_start; i < current_ins_start; i += 2)
+ if (deps [i] == IA64_WRITE_BR && deps [i + 1] == reg)
+ need_stop = TRUE;
+ break;
+ case IA64_WRITE_BR:
+ reg = code->dep_info [pos + 1];
+
+ DEBUG_INS_SCHED (printf ("WRITE BR: %d\n", reg));
+ for (i = current_deps_start; i < current_ins_start; i += 2)
+ if (deps [i] == IA64_WRITE_BR && deps [i + 1] == reg)
+ need_stop = TRUE;
+ break;
+ case IA64_READ_BR_BRANCH:
+ reg = deps [pos + 1];
+
+ /* Writes to brs are visible to branches */
+ DEBUG_INS_SCHED (printf ("READ BR BRACH: %d\n", reg));
+ break;
+ case IA64_READ_FR:
+ reg = deps [pos + 1];
+
+ DEBUG_INS_SCHED (printf ("READ BR: %d\n", reg));
+ for (i = current_deps_start; i < current_ins_start; i += 2)
+ if (deps [i] == IA64_WRITE_FR && deps [i + 1] == reg)
+ need_stop = TRUE;
+ break;
+ case IA64_WRITE_FR:
+ reg = code->dep_info [pos + 1];
+
+ DEBUG_INS_SCHED (printf ("WRITE BR: %d\n", reg));
+ for (i = current_deps_start; i < current_ins_start; i += 2)
+ if (deps [i] == IA64_WRITE_FR && deps [i + 1] == reg)
+ need_stop = TRUE;
+ break;
+ case IA64_READ_AR:
+ reg = deps [pos + 1];
+
+ DEBUG_INS_SCHED (printf ("READ AR: %d\n", reg));
+ for (i = current_deps_start; i < current_ins_start; i += 2)
+ if (deps [i] == IA64_WRITE_AR && deps [i + 1] == reg)
+ need_stop = TRUE;
+ break;
+ case IA64_WRITE_AR:
+ reg = code->dep_info [pos + 1];
+
+ DEBUG_INS_SCHED (printf ("WRITE AR: %d\n", reg));
+ for (i = current_deps_start; i < current_ins_start; i += 2)
+ if (deps [i] == IA64_WRITE_AR && deps [i + 1] == reg)
+ need_stop = TRUE;
+ break;
+ case IA64_NO_STOP:
+ /*
+ * Explicitly indicate that a stop is not required. Useful for
+ * example when two predicated instructions with negated predicates
+ * write the same registers.
+ */
+ no_stop = TRUE;
+ break;
+ default:
+ g_assert_not_reached ();
+ }
+ pos += 2;
+
+ if (need_stop && !no_stop) {
+ g_assert (ins_index > 0);
+ stops [ins_index - 1] = 1;
+
+ DEBUG_INS_SCHED (printf ("STOP\n"));
+ current_deps_start = current_ins_start;
+
+ /* Skip remaining deps for this instruction */
+ while (deps [pos] != IA64_END_OF_INS)
+ pos += 2;
+ }
+ }
+
+ if (code->nins > 0) {
+ /* No dependency info for the last instruction */
+ stops [code->nins - 1] = 1;
+ }
+
+ deps_start [code->nins] = code->dep_info_pos;
+}
+
+static void
+ia64_real_emit_bundle (Ia64CodegenState *code, int *deps_start, int *stops, int n, guint64 template, guint64 ins1, guint64 ins2, guint64 ins3, guint8 nops)
+{
+ int stop_pos, i, deps_to_shift, dep_shift;
+
+ g_assert (n <= code->nins);
+
+ // if (n > 1) printf ("FOUND: %ld.\n", template);
+
+ ia64_emit_bundle_template (code, template, ins1, ins2, ins3);
+
+ stop_pos = last_stop_in_template [template] + 1;
+ if (stop_pos > n)
+ stop_pos = n;
+
+ /* Compute the number of 'real' instructions before the stop */
+ deps_to_shift = stop_pos;
+ if (stop_pos >= 3 && (nops & (1 << 2)))
+ deps_to_shift --;
+ if (stop_pos >= 2 && (nops & (1 << 1)))
+ deps_to_shift --;
+ if (stop_pos >= 1 && (nops & (1 << 0)))
+ deps_to_shift --;
+
+ /*
+ * We have to keep some dependencies whose instructions have been shifted
+ * out of the buffer. So nullify the end_of_ins markers in the dependency
+ * array.
+ */
+ for (i = deps_start [deps_to_shift]; i < deps_start [n]; i += 2)
+ if (code->dep_info [i] == IA64_END_OF_INS)
+ code->dep_info [i] = IA64_NONE;
+
+ g_assert (deps_start [deps_to_shift] <= code->dep_info_pos);
+ memcpy (code->dep_info, &code->dep_info [deps_start [deps_to_shift]], code->dep_info_pos - deps_start [deps_to_shift]);
+ code->dep_info_pos = code->dep_info_pos - deps_start [deps_to_shift];
+
+ dep_shift = deps_start [deps_to_shift];
+ for (i = 0; i < code->nins + 1 - n; ++i)
+ deps_start [i] = deps_start [n + i] - dep_shift;
+
+ /* Determine the exact positions of instructions with unwind ops */
+ if (code->unw_op_count) {
+ int ins_pos [16];
+ int curr_ins, curr_ins_pos;
+
+ curr_ins = 0;
+ curr_ins_pos = ((code->buf - code->region_start - 16) / 16) * 3;
+ for (i = 0; i < 3; ++i) {
+ if (! (nops & (1 << i))) {
+ ins_pos [curr_ins] = curr_ins_pos + i;
+ curr_ins ++;
+ }
+ }
+
+ for (i = code->unw_op_pos; i < code->unw_op_count; ++i) {
+ if (code->unw_ops_pos [i] < n) {
+ code->unw_ops [i].when = ins_pos [code->unw_ops_pos [i]];
+ //printf ("UNW-OP: %d -> %d\n", code->unw_ops_pos [i], code->unw_ops [i].when);
+ }
+ }
+ if (code->unw_op_pos < code->unw_op_count)
+ code->unw_op_pos += n;
+ }
+
+ if (n == code->nins) {
+ code->template = 0;
+ code->nins = 0;
+ }
+ else {
+ memcpy (&code->instructions [0], &code->instructions [n], (code->nins - n) * sizeof (guint64));
+ memcpy (&code->itypes [0], &code->itypes [n], (code->nins - n) * sizeof (int));
+ memcpy (&stops [0], &stops [n], (code->nins - n) * sizeof (int));
+ code->nins -= n;
+ }
+}
+
void
ia64_emit_bundle (Ia64CodegenState *code, gboolean flush)
{
- int i, j, ins_type, template;
+ int i, ins_type, template, nins_to_emit;
+ int deps_start [16];
+ int stops [16];
+ gboolean found;
- if (!code->automatic) {
- if (code->nins == 0)
- return;
+ /*
+ * We implement a simple scheduler which tries to put three instructions
+ * per bundle, then two, then one.
+ */
+ ia64_analyze_deps (code, deps_start, stops);
+
+ if ((code->nins >= 3) && !code->one_ins_per_bundle) {
+ /* Find a suitable template */
+ for (template = 0; template < 32; ++template) {
+ if (stops_in_template [template][0] != stops [0] ||
+ stops_in_template [template][1] != stops [1] ||
+ stops_in_template [template][2] != stops [2])
+ continue;
- g_assert (code->nins == 3);
+ found = TRUE;
+ for (i = 0; i < 3; ++i) {
+ ins_type = ins_types_in_template [template][i];
+ switch (code->itypes [i]) {
+ case IA64_INS_TYPE_A:
+ found &= (ins_type == IA64_INS_TYPE_I) || (ins_type == IA64_INS_TYPE_M);
+ break;
+ default:
+ found &= (ins_type == code->itypes [i]);
+ break;
+ }
+ }
- /* Verify template is correct */
- template = code->template;
- for (j = 0; j < 3; ++j) {
- if (code->stops [j])
- g_assert (stops_in_template [template]);
+ if (found)
+ found = debug_ins_sched ();
- ins_type = ins_types_in_template [template][j];
- switch (code->itypes [j]) {
- case IA64_INS_TYPE_A:
- g_assert ((ins_type == IA64_INS_TYPE_I) || (ins_type == IA64_INS_TYPE_M));
- break;
- case IA64_INS_TYPE_LX:
- g_assert (j == 1);
- g_assert (ins_type == IA64_INS_TYPE_LX);
- j ++;
+ if (found) {
+ ia64_real_emit_bundle (code, deps_start, stops, 3, template, code->instructions [0], code->instructions [1], code->instructions [2], 0);
break;
- default:
- g_assert (ins_type == code->itypes [j]);
}
}
+ }
- ia64_emit_bundle_template (code, template, code->instructions [0], code->instructions [1], code->instructions [2]);
- code->template = 0;
- code->nins = 0;
+ if (code->nins < IA64_INS_BUFFER_SIZE && !flush)
+ /* Wait for more instructions */
return;
+
+ /* If it didn't work out, try putting two instructions into one bundle */
+ if ((code->nins >= 2) && !code->one_ins_per_bundle) {
+ /* Try a nop at the end */
+ for (template = 0; template < 32; ++template) {
+ if (stops_in_template [template][0] != stops [0] ||
+ ((stops_in_template [template][1] != stops [1]) &&
+ (stops_in_template [template][2] != stops [1])))
+
+ continue;
+
+ if (!ITYPE_MATCH (ins_types_in_template [template][0], code->itypes [0]) ||
+ !ITYPE_MATCH (ins_types_in_template [template][1], code->itypes [1]))
+ continue;
+
+ if (!debug_ins_sched ())
+ continue;
+
+ 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);
+ break;
+ }
}
- for (i = 0; i < code->nins; ++i) {
- switch (code->itypes [i]) {
+ if (code->nins < IA64_INS_BUFFER_SIZE && !flush)
+ /* Wait for more instructions */
+ return;
+
+ if ((code->nins >= 2) && !code->one_ins_per_bundle) {
+ /* Try a nop in the middle */
+ for (template = 0; template < 32; ++template) {
+ if (((stops_in_template [template][0] != stops [0]) &&
+ (stops_in_template [template][1] != stops [0])) ||
+ stops_in_template [template][2] != stops [1])
+ continue;
+
+ if (!ITYPE_MATCH (ins_types_in_template [template][0], code->itypes [0]) ||
+ !ITYPE_MATCH (ins_types_in_template [template][2], code->itypes [1]))
+ continue;
+
+ if (!debug_ins_sched ())
+ continue;
+
+ 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);
+ break;
+ }
+ }
+
+ if ((code->nins >= 2) && flush && !code->one_ins_per_bundle) {
+ /* Try a nop at the beginning */
+ for (template = 0; template < 32; ++template) {
+ if ((stops_in_template [template][1] != stops [0]) ||
+ (stops_in_template [template][2] != stops [1]))
+ continue;
+
+ if (!ITYPE_MATCH (ins_types_in_template [template][1], code->itypes [0]) ||
+ !ITYPE_MATCH (ins_types_in_template [template][2], code->itypes [1]))
+ continue;
+
+ if (!debug_ins_sched ())
+ continue;
+
+ 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);
+ break;
+ }
+ }
+
+ if (code->nins < IA64_INS_BUFFER_SIZE && !flush)
+ /* Wait for more instructions */
+ return;
+
+ if (flush)
+ nins_to_emit = code->nins;
+ else
+ nins_to_emit = 1;
+
+ while (nins_to_emit > 0) {
+ if (!debug_ins_sched ())
+ stops [0] = 1;
+ switch (code->itypes [0]) {
case IA64_INS_TYPE_A:
- ia64_emit_bundle_template (code, IA64_TEMPLATE_MIIS, code->instructions [i], IA64_NOP_I, IA64_NOP_I);
+ if (stops [0])
+ ia64_real_emit_bundle (code, deps_start, stops, 1, IA64_TEMPLATE_MIIS, code->instructions [0], IA64_NOP_I, IA64_NOP_I, 0);
+ else
+ ia64_real_emit_bundle (code, deps_start, stops, 1, IA64_TEMPLATE_MII, code->instructions [0], IA64_NOP_I, IA64_NOP_I, 0);
break;
case IA64_INS_TYPE_I:
- ia64_emit_bundle_template (code, IA64_TEMPLATE_MIIS, IA64_NOP_M, code->instructions [i], IA64_NOP_I);
+ if (stops [0])
+ ia64_real_emit_bundle (code, deps_start, stops, 1, IA64_TEMPLATE_MIIS, IA64_NOP_M, code->instructions [0], IA64_NOP_I, 0);
+ else
+ ia64_real_emit_bundle (code, deps_start, stops, 1, IA64_TEMPLATE_MII, IA64_NOP_M, code->instructions [0], IA64_NOP_I, 0);
break;
case IA64_INS_TYPE_M:
- ia64_emit_bundle_template (code, IA64_TEMPLATE_MIIS, code->instructions [i], IA64_NOP_I, IA64_NOP_I);
+ if (stops [0])
+ ia64_real_emit_bundle (code, deps_start, stops, 1, IA64_TEMPLATE_MIIS, code->instructions [0], IA64_NOP_I, IA64_NOP_I, 0);
+ else
+ ia64_real_emit_bundle (code, deps_start, stops, 1, IA64_TEMPLATE_MII, code->instructions [0], IA64_NOP_I, IA64_NOP_I, 0);
break;
case IA64_INS_TYPE_B:
- ia64_emit_bundle_template (code, IA64_TEMPLATE_MIBS, IA64_NOP_M, IA64_NOP_I, code->instructions [i]);
+ if (stops [0])
+ ia64_real_emit_bundle (code, deps_start, stops, 1, IA64_TEMPLATE_MIBS, IA64_NOP_M, IA64_NOP_I, code->instructions [0], 0);
+ else
+ ia64_real_emit_bundle (code, deps_start, stops, 1, IA64_TEMPLATE_MIB, IA64_NOP_M, IA64_NOP_I, code->instructions [0], 0);
break;
case IA64_INS_TYPE_F:
- ia64_emit_bundle_template (code, IA64_TEMPLATE_MFIS, IA64_NOP_M, code->instructions [i], IA64_NOP_I);
+ if (stops [0])
+ ia64_real_emit_bundle (code, deps_start, stops, 1, IA64_TEMPLATE_MFIS, IA64_NOP_M, code->instructions [0], IA64_NOP_I, 0);
+ else
+ ia64_real_emit_bundle (code, deps_start, stops, 1, IA64_TEMPLATE_MFI, IA64_NOP_M, code->instructions [0], IA64_NOP_I, 0);
break;
case IA64_INS_TYPE_LX:
- ia64_emit_bundle_template (code, IA64_TEMPLATE_MLXS, IA64_NOP_M, code->instructions [i], code->instructions [i + 1]);
- i ++;
+ if (stops [0] || stops [1])
+ ia64_real_emit_bundle (code, deps_start, stops, 2, IA64_TEMPLATE_MLXS, IA64_NOP_M, code->instructions [0], code->instructions [1], 0);
+ else
+ ia64_real_emit_bundle (code, deps_start, stops, 2, IA64_TEMPLATE_MLX, IA64_NOP_M, code->instructions [0], code->instructions [1], 0);
+ nins_to_emit --;
break;
default:
g_assert_not_reached ();
}
+ nins_to_emit --;
}
+}
- code->nins = 0;
+unw_dyn_region_info_t*
+mono_ia64_create_unwind_region (Ia64CodegenState *code)
+{
+ unw_dyn_region_info_t *r;
+
+ g_assert (code->nins == 0);
+ r = g_malloc0 (_U_dyn_region_info_size (code->unw_op_count));
+ memcpy (&r->op, &code->unw_ops, sizeof (unw_dyn_op_t) * code->unw_op_count);
+ r->op_count = code->unw_op_count;
+ r->insn_count = ((code->buf - code->region_start) >> 4) * 3;
+ code->unw_op_count = 0;
+ code->unw_op_pos = 0;
+ code->region_start = code->buf;
+
+ return r;
}
static void
guint64 instructions [3];
guint8 gen_buf [16];
Ia64CodegenState gen;
+ int ins_to_skip;
+ gboolean found;
+
+ /*
+ * code encodes both the position inside the buffer and code.nins when
+ * the instruction was emitted.
+ */
+ ins_to_skip = (guint64)code % 16;
+ code = (unsigned char*)((guint64)code & ~15);
+
+ /*
+ * Search for the first instruction which is 'patchable', skipping
+ * ins_to_skip instructions.
+ */
+
+ while (TRUE) {
template = ia64_bundle_template (code);
instructions [0] = ia64_bundle_ins1 (code);
ia64_codegen_init (gen, gen_buf);
+ found = FALSE;
for (i = 0; i < 3; ++i) {
guint64 ins = instructions [i];
int opcode = ia64_ins_opcode (ins);
- /* Skip nops */
- gboolean nop = FALSE;
- switch (ins_types_in_template [template][i]) {
- case IA64_INS_TYPE_I:
- nop = (ins == IA64_NOP_I);
- break;
- case IA64_INS_TYPE_M:
- nop = (ins == IA64_NOP_M);
- break;
- case IA64_INS_TYPE_LX:
- break;
- default:
- break;
- }
+ if (ins == nops_for_ins_types [ins_types_in_template [template][i]])
+ continue;
- if (nop)
+ if (ins_to_skip) {
+ ins_to_skip --;
continue;
+ }
switch (ins_types_in_template [template][i]) {
case IA64_INS_TYPE_A:
/* adds */
ia64_adds_imm_pred (gen, ia64_ins_qp (ins), ia64_ins_r1 (ins), (guint64)target, ia64_ins_r3 (ins));
instructions [i] = gen.instructions [0];
+ found = TRUE;
}
else
NOT_IMPLEMENTED;
ia64_br_cond_hint_pred (gen, ia64_ins_qp (ins), disp, 0, 0, 0);
instructions [i] = gen.instructions [0];
+ found = TRUE;
}
else if (opcode == 5) {
/* br.call */
/* FIXME: hints */
ia64_br_call_hint_pred (gen, ia64_ins_qp (ins), ia64_ins_b1 (ins), disp, 0, 0, 0);
instructions [i] = gen.instructions [0];
+ found = TRUE;
}
else
NOT_IMPLEMENTED;
ia64_movl_pred (gen, ia64_ins_qp (ins), ia64_ins_r1 (ins), target);
instructions [1] = gen.instructions [0];
instructions [2] = gen.instructions [1];
+ found = TRUE;
}
else
NOT_IMPLEMENTED;
default:
NOT_IMPLEMENTED;
}
+
+ if (found) {
+ /* Rewrite code */
+ ia64_codegen_init (gen, code);
+ ia64_emit_bundle_template (&gen, template, instructions [0], instructions [1], instructions [2]);
+ return;
+ }
}
- /* Rewrite code */
- ia64_codegen_init (gen, code);
- ia64_emit_bundle_template (&gen, template, instructions [0], instructions [1], instructions [2]);
+ code += 16;
+ }
}
void
target = mono_resolve_patch_target (method, domain, code, patch_info, run_cctors);
+ if (patch_info->type == MONO_PATCH_INFO_NONE)
+ continue;
if (mono_compile_aot) {
NOT_IMPLEMENTED;
}
int alloc_size, pos, i;
Ia64CodegenState code;
CallInfo *cinfo;
- unw_dyn_region_info_t *r_pro;
- int unw_op_count;
-
+
sig = mono_method_signature (method);
pos = 0;
cinfo = get_call_info (sig, FALSE);
cfg->code_size = MAX (((MonoMethodNormal *)method)->header->code_size * 4, 512);
+
+ if (mono_jit_trace_calls != NULL && mono_trace_eval (method))
+ cfg->code_size += 1024;
+ if (cfg->prof_options & MONO_PROFILE_ENTER_LEAVE)
+ cfg->code_size += 1024;
+
cfg->native_code = g_malloc (cfg->code_size);
ia64_codegen_init (code, cfg->native_code);
- ia64_codegen_set_automatic (code, FALSE);
alloc_size = ALIGN_TO (cfg->stack_offset, MONO_ARCH_FRAME_ALIGNMENT);
if (cfg->param_area)
pos = 0;
if (method->save_lmf) {
- /* FIXME: */
+ /* No LMF on IA64 */
}
alloc_size -= pos;
- /* Initialize unwind info */
- r_pro = g_malloc0 (_U_dyn_region_info_size (3));
- unw_op_count = 0;
-
- ia64_begin_bundle_template (code, IA64_TEMPLATE_MIIS);
+ ia64_unw_save_reg (code, UNW_IA64_AR_PFS, UNW_IA64_GR + cfg->arch.reg_saved_ar_pfs);
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);
+ ia64_unw_save_reg (code, UNW_IA64_RP, UNW_IA64_GR + cfg->arch.reg_saved_b0);
ia64_mov_from_br (code, cfg->arch.reg_saved_b0, IA64_B0);
- _U_dyn_op_save_reg (&r_pro->op[unw_op_count++], _U_QP_TRUE, /* when=*/ 0,
- /* reg=*/ UNW_IA64_AR_PFS, /* dst=*/ UNW_IA64_GR + cfg->arch.reg_saved_ar_pfs);
- _U_dyn_op_save_reg (&r_pro->op[unw_op_count++], _U_QP_TRUE, /* when=*/ 1,
- /* reg=*/ UNW_IA64_RP, /* dst=*/ UNW_IA64_GR + cfg->arch.reg_saved_b0);
-
- if (alloc_size || cinfo->stack_usage) {
- ia64_mov (code, cfg->frame_reg, IA64_SP);
- _U_dyn_op_save_reg (&r_pro->op[unw_op_count++], _U_QP_TRUE, /* when=*/ 2,
- /* reg=*/ UNW_IA64_SP, /* dst=*/ UNW_IA64_GR + cfg->frame_reg);
+ if ((alloc_size || cinfo->stack_usage) && !cfg->arch.omit_fp) {
+ ia64_unw_save_reg (code, UNW_IA64_SP, UNW_IA64_GR + cfg->arch.reg_saved_sp);
+ ia64_mov (code, cfg->arch.reg_saved_sp, IA64_SP);
+ if (cfg->frame_reg != cfg->arch.reg_saved_sp)
+ ia64_mov (code, cfg->frame_reg, IA64_SP);
}
- else
- ia64_nop_i (code, 0);
- ia64_stop (code);
- ia64_end_bundle (code);
-
- /* Finish unwind info */
- r_pro->op_count = unw_op_count;
- r_pro->insn_count = (code.buf - cfg->native_code) >> 4;
-
- cfg->arch.r_pro = r_pro;
if (alloc_size) {
- /* See mono_emit_stack_alloc */
-#if defined(MONO_ARCH_SIGSEGV_ON_ALTSTACK)
- NOT_IMPLEMENTED;
-#else
+ int pagesize = getpagesize ();
+#if defined(MONO_ARCH_SIGSEGV_ON_ALTSTACK)
+ if (alloc_size >= pagesize) {
+ gint32 remaining_size = alloc_size;
+
+ /* Generate stack touching code */
+ ia64_mov (code, GP_SCRATCH_REG, IA64_SP);
+ while (remaining_size >= pagesize) {
+ ia64_movl (code, GP_SCRATCH_REG2, pagesize);
+ ia64_sub (code, GP_SCRATCH_REG, GP_SCRATCH_REG, GP_SCRATCH_REG2);
+ ia64_ld8 (code, GP_SCRATCH_REG2, GP_SCRATCH_REG);
+ remaining_size -= pagesize;
+ }
+ }
+#endif
if (ia64_is_imm14 (-alloc_size)) {
- ia64_begin_bundle_template (code, IA64_TEMPLATE_MISI);
- ia64_nop_m (code, 0);
- ia64_mov (code, cfg->arch.reg_saved_sp, IA64_SP); ia64_stop (code);
+ if (cfg->arch.omit_fp)
+ ia64_unw_add (code, UNW_IA64_SP, (-alloc_size));
ia64_adds_imm (code, IA64_SP, (-alloc_size), IA64_SP);
- ia64_end_bundle (code);
}
else {
- ia64_begin_bundle_template (code, IA64_TEMPLATE_MLXS);
- ia64_mov (code, cfg->arch.reg_saved_sp, IA64_SP);
- ia64_movl (code, GP_SCRATCH_REG, -alloc_size); ia64_stop (code);
- ia64_begin_bundle_template (code, IA64_TEMPLATE_MIIS);
+ ia64_movl (code, GP_SCRATCH_REG, -alloc_size);
+ if (cfg->arch.omit_fp)
+ ia64_unw_add (code, UNW_IA64_SP, (-alloc_size));
ia64_add (code, IA64_SP, GP_SCRATCH_REG, IA64_SP);
- ia64_nop_i (code, 0);
- ia64_nop_i (code, 0); ia64_stop (code);
- ia64_end_bundle (code);
}
-#endif
}
- ia64_codegen_set_automatic (code, TRUE);
+
+ ia64_begin_bundle (code);
+
+ /* Initialize unwind info */
+ cfg->arch.r_pro = mono_ia64_create_unwind_region (&code);
if (sig->ret->type != MONO_TYPE_VOID) {
if ((cinfo->ret.storage == ArgInIReg) && (cfg->ret->opcode != OP_REGVAR)) {
break;
case ArgOnStack:
break;
+ case ArgAggregate:
+ if (ainfo->nslots != ainfo->nregs)
+ NOT_IMPLEMENTED;
+
+ g_assert (inst->opcode == OP_REGOFFSET);
+ ia64_adds_imm (code, GP_SCRATCH_REG, inst->inst_offset, inst->inst_basereg);
+ for (i = 0; i < ainfo->nregs; ++i) {
+ switch (ainfo->atype) {
+ case AggregateNormal:
+ ia64_st8_inc_imm_hint (code, GP_SCRATCH_REG, cfg->arch.reg_in0 + ainfo->reg + i, sizeof (gpointer), 0);
+ break;
+ case AggregateSingleHFA:
+ ia64_stfs_inc_imm_hint (code, GP_SCRATCH_REG, ainfo->reg + i, 4, 0);
+ break;
+ case AggregateDoubleHFA:
+ ia64_stfd_inc_imm_hint (code, GP_SCRATCH_REG, ainfo->reg + i, sizeof (gpointer), 0);
+ break;
+ default:
+ NOT_IMPLEMENTED;
+ }
+ }
+ break;
default:
- NOT_IMPLEMENTED;
+ g_assert_not_reached ();
}
}
}
if (method->save_lmf) {
- /* FIXME: */
+ /* No LMF on IA64 */
}
+ if (strstr (cfg->method->name, "end_invoke_int_IAsyncResult"))
+ code = emit_call (cfg, code, MONO_PATCH_INFO_ABS, mono_arch_break);
+
ia64_codegen_close (code);
g_free (cinfo);
mono_arch_emit_epilog (MonoCompile *cfg)
{
MonoMethod *method = cfg->method;
- int pos;
+ int i, pos;
int max_epilog_size = 16 * 4;
Ia64CodegenState code;
guint8 *buf;
CallInfo *cinfo;
+ ArgInfo *ainfo;
+
+ if (mono_jit_trace_calls != NULL)
+ max_epilog_size += 1024;
cfg->arch.epilog_begin_offset = cfg->code_len;
- while (cfg->code_len + max_epilog_size > (cfg->code_size - 16)) {
+ while (cfg->code_len + max_epilog_size > cfg->code_size) {
cfg->code_size *= 2;
cfg->native_code = g_realloc (cfg->native_code, cfg->code_size);
mono_jit_stats.code_reallocs++;
pos = 0;
if (method->save_lmf) {
- /* FIXME: */
+ /* No LMF on IA64 */
}
/* Load returned vtypes into registers if needed */
cinfo = get_call_info (mono_method_signature (method), FALSE);
- if (cinfo->ret.storage == ArgAggregate) {
- NOT_IMPLEMENTED;
+ ainfo = &cinfo->ret;
+ switch (ainfo->storage) {
+ case ArgAggregate:
+ if (ainfo->nslots != ainfo->nregs)
+ NOT_IMPLEMENTED;
+
+ g_assert (cfg->ret->opcode == OP_REGOFFSET);
+ ia64_adds_imm (code, GP_SCRATCH_REG, cfg->ret->inst_offset, cfg->ret->inst_basereg);
+ for (i = 0; i < ainfo->nregs; ++i) {
+ switch (ainfo->atype) {
+ case AggregateNormal:
+ ia64_ld8_inc_imm_hint (code, ainfo->reg + i, GP_SCRATCH_REG, sizeof (gpointer), 0);
+ break;
+ case AggregateSingleHFA:
+ ia64_ldfs_inc_imm_hint (code, ainfo->reg + i, GP_SCRATCH_REG, 4, 0);
+ break;
+ case AggregateDoubleHFA:
+ ia64_ldfd_inc_imm_hint (code, ainfo->reg + i, GP_SCRATCH_REG, sizeof (gpointer), 0);
+ break;
+ default:
+ g_assert_not_reached ();
+ }
+ }
+ break;
+ default:
+ break;
}
g_free (cinfo);
- ia64_end_bundle (code);
- ia64_codegen_set_automatic (code, FALSE);
+ ia64_begin_bundle (code);
- ia64_begin_bundle_template (code, IA64_TEMPLATE_MIIS);
- if (cfg->arch.stack_alloc_size)
- ia64_mov (code, IA64_SP, cfg->arch.reg_saved_sp);
- else
- ia64_nop_m (code, 0);
- ia64_mov_to_ar_i (code, IA64_PFS, cfg->arch.reg_saved_ar_pfs);
- ia64_mov_ret_to_br (code, IA64_B0, cfg->arch.reg_saved_b0); ia64_stop (code);
- ia64_end_bundle (code);
+ code.region_start = cfg->native_code;
- ia64_begin_bundle_template (code, IA64_TEMPLATE_BBBS);
+ /* Label the unwind state at the start of the exception throwing region */
+ //ia64_unw_label_state (code, 1234);
+
+ if (cfg->arch.stack_alloc_size) {
+ if (cfg->arch.omit_fp) {
+ if (ia64_is_imm14 (cfg->arch.stack_alloc_size)) {
+ ia64_unw_pop_frames (code, 1);
+ ia64_adds_imm (code, IA64_SP, (cfg->arch.stack_alloc_size), IA64_SP);
+ } else {
+ ia64_movl (code, GP_SCRATCH_REG, cfg->arch.stack_alloc_size);
+ ia64_unw_pop_frames (code, 1);
+ ia64_add (code, IA64_SP, GP_SCRATCH_REG, IA64_SP);
+ }
+ }
+ else {
+ ia64_unw_pop_frames (code, 1);
+ ia64_mov (code, IA64_SP, cfg->arch.reg_saved_sp);
+ }
+ }
+ ia64_mov_to_ar_i (code, IA64_PFS, cfg->arch.reg_saved_ar_pfs);
+ ia64_mov_ret_to_br (code, IA64_B0, cfg->arch.reg_saved_b0);
ia64_br_ret_reg (code, IA64_B0);
- ia64_nop_b (code, 0);
- ia64_nop_b (code, 0); ia64_stop (code);
- ia64_end_bundle (code);
- ia64_codegen_set_automatic (code, TRUE);
ia64_codegen_close (code);
+ cfg->arch.r_epilog = mono_ia64_create_unwind_region (&code);
+ cfg->arch.r_pro->next = cfg->arch.r_epilog;
+
cfg->code_len = code.buf - cfg->native_code;
g_assert (cfg->code_len < cfg->code_size);
mono_arch_emit_exceptions (MonoCompile *cfg)
{
MonoJumpInfo *patch_info;
- int nthrows;
+ int i, nthrows;
Ia64CodegenState code;
gboolean empty = TRUE;
- /*
+ //unw_dyn_region_info_t *r_exceptions;
MonoClass *exc_classes [16];
guint8 *exc_throw_start [16], *exc_throw_end [16];
- */
guint32 code_size = 0;
/* Compute needed space */
code_size += 4 + 7; /* sizeof (float) + alignment */
}
+ if (code_size == 0)
+ return;
+
while (cfg->code_len + code_size > (cfg->code_size - 16)) {
cfg->code_size *= 2;
cfg->native_code = g_realloc (cfg->native_code, cfg->code_size);
ia64_codegen_init (code, cfg->native_code + cfg->code_len);
+ /* The unwind state here is the same as before the epilog */
+ //ia64_unw_copy_state (code, 1234);
+
/* add code to raise exceptions */
/* FIXME: Optimize this */
nthrows = 0;
MonoClass *exc_class;
guint8* throw_ip;
guint8* buf;
+ guint64 exc_token_index;
exc_class = mono_class_from_name (mono_defaults.corlib, "System", patch_info->data.name);
g_assert (exc_class);
+ exc_token_index = mono_metadata_token_index (exc_class->type_token);
throw_ip = cfg->native_code + patch_info->ip.i;
+ ia64_begin_bundle (code);
+
ia64_patch (cfg->native_code + patch_info->ip.i, code.buf);
- ia64_movl (code, cfg->arch.reg_out0 + 0, exc_class->type_token);
+ /* Find a throw sequence for the same exception class */
+ for (i = 0; i < nthrows; ++i)
+ if (exc_classes [i] == exc_class)
+ break;
- ia64_begin_bundle (code);
+ if (i < nthrows) {
+ gint64 offset = exc_throw_end [i] - 16 - throw_ip;
- patch_info->data.name = "mono_arch_throw_corlib_exception";
- patch_info->type = MONO_PATCH_INFO_INTERNAL_METHOD;
- patch_info->ip.i = code.buf - cfg->native_code;
+ if (ia64_is_adds_imm (offset))
+ ia64_adds_imm (code, cfg->arch.reg_out0 + 1, offset, IA64_R0);
+ else
+ ia64_movl (code, cfg->arch.reg_out0 + 1, offset);
- /* Indirect call */
- ia64_movl (code, GP_SCRATCH_REG, 0);
- ia64_ld8_inc_imm (code, GP_SCRATCH_REG2, GP_SCRATCH_REG, 8);
- ia64_mov_to_br (code, IA64_B6, GP_SCRATCH_REG2);
- ia64_ld8 (code, IA64_GP, GP_SCRATCH_REG);
+ buf = code.buf + code.nins;
+ ia64_br_cond_pred (code, 0, 0);
+ ia64_begin_bundle (code);
+ ia64_patch (buf, exc_throw_start [i]);
+
+ patch_info->type = MONO_PATCH_INFO_NONE;
+ }
+ else {
+ /* Arg1 */
+ buf = code.buf;
+ ia64_movl (code, cfg->arch.reg_out0 + 1, 0);
- /* Compute the offset */
- buf = code.buf + 32;
- ia64_movl (code, cfg->arch.reg_out0 + 1, buf - throw_ip);
+ ia64_begin_bundle (code);
- ia64_br_call_reg (code, IA64_B0, IA64_B6);
+ if (nthrows < 16) {
+ exc_classes [nthrows] = exc_class;
+ exc_throw_start [nthrows] = code.buf;
+ }
+
+ /* Arg2 */
+ if (ia64_is_adds_imm (exc_token_index))
+ ia64_adds_imm (code, cfg->arch.reg_out0 + 0, exc_token_index, IA64_R0);
+ else
+ ia64_movl (code, cfg->arch.reg_out0 + 0, exc_token_index);
+
+ patch_info->data.name = "mono_arch_throw_corlib_exception";
+ patch_info->type = MONO_PATCH_INFO_INTERNAL_METHOD;
+ patch_info->ip.i = code.buf + code.nins - cfg->native_code;
+
+ /* Indirect call */
+ ia64_movl (code, GP_SCRATCH_REG, 0);
+ ia64_ld8_inc_imm (code, GP_SCRATCH_REG2, GP_SCRATCH_REG, 8);
+ ia64_mov_to_br (code, IA64_B6, GP_SCRATCH_REG2);
+ ia64_ld8 (code, IA64_GP, GP_SCRATCH_REG);
+
+ ia64_br_call_reg (code, IA64_B0, IA64_B6);
+
+ /* Patch up the throw offset */
+ ia64_begin_bundle (code);
+
+ ia64_patch (buf, (gpointer)(code.buf - 16 - throw_ip));
+
+ if (nthrows < 16) {
+ exc_throw_end [nthrows] = code.buf;
+ nthrows ++;
+ }
+ }
empty = FALSE;
break;
ia64_codegen_close (code);
+ /* FIXME: */
+ //r_exceptions = mono_ia64_create_unwind_region (&code);
+ //cfg->arch.r_epilog = r_exceptions;
+
cfg->code_len = code.buf - cfg->native_code;
g_assert (cfg->code_len < cfg->code_size);
void*
mono_arch_instrument_prolog (MonoCompile *cfg, void *func, void *p, gboolean enable_arguments)
{
- NOT_IMPLEMENTED;
+ Ia64CodegenState code;
+ CallInfo *cinfo = NULL;
+ MonoMethodSignature *sig;
+ MonoInst *ins;
+ int i, n, stack_area = 0;
- return NULL;
+ ia64_codegen_init (code, p);
+
+ /* Keep this in sync with mono_arch_get_argument_info */
+
+ if (enable_arguments) {
+ /* Allocate a new area on the stack and save arguments there */
+ sig = mono_method_signature (cfg->method);
+
+ cinfo = get_call_info (sig, FALSE);
+
+ n = sig->param_count + sig->hasthis;
+
+ stack_area = ALIGN_TO (n * 8, 16);
+
+ if (n) {
+ ia64_movl (code, GP_SCRATCH_REG, stack_area);
+
+ ia64_sub (code, IA64_SP, IA64_SP, GP_SCRATCH_REG);
+
+ /* FIXME: Allocate out registers */
+
+ ia64_mov (code, cfg->arch.reg_out0 + 1, IA64_SP);
+
+ /* Required by the ABI */
+ ia64_adds_imm (code, IA64_SP, -16, IA64_SP);
+
+ add_patch_info (cfg, code, MONO_PATCH_INFO_METHODCONST, cfg->method);
+ ia64_movl (code, cfg->arch.reg_out0 + 0, 0);
+
+ /* Save arguments to the stack */
+ for (i = 0; i < n; ++i) {
+ ins = cfg->varinfo [i];
+
+ if (ins->opcode == OP_REGVAR) {
+ ia64_movl (code, GP_SCRATCH_REG, (i * 8));
+ ia64_add (code, GP_SCRATCH_REG, cfg->arch.reg_out0 + 1, GP_SCRATCH_REG);
+ ia64_st8 (code, GP_SCRATCH_REG, ins->dreg);
+ }
+ else {
+ ia64_movl (code, GP_SCRATCH_REG, ins->inst_offset);
+ ia64_add (code, GP_SCRATCH_REG, ins->inst_basereg, GP_SCRATCH_REG);
+ ia64_ld8 (code, GP_SCRATCH_REG2, GP_SCRATCH_REG);
+ ia64_movl (code, GP_SCRATCH_REG, (i * 8));
+ ia64_add (code, GP_SCRATCH_REG, cfg->arch.reg_out0 + 1, GP_SCRATCH_REG);
+ ia64_st8 (code, GP_SCRATCH_REG, GP_SCRATCH_REG2);
+ }
+ }
+ }
+ else
+ ia64_mov (code, cfg->arch.reg_out0 + 1, IA64_R0);
+ }
+ else
+ ia64_mov (code, cfg->arch.reg_out0 + 1, IA64_R0);
+
+ add_patch_info (cfg, code, MONO_PATCH_INFO_METHODCONST, cfg->method);
+ ia64_movl (code, cfg->arch.reg_out0 + 0, 0);
+
+ code = emit_call (cfg, code, MONO_PATCH_INFO_ABS, (gpointer)func);
+
+ if (enable_arguments && stack_area) {
+ ia64_movl (code, GP_SCRATCH_REG, stack_area);
+
+ ia64_add (code, IA64_SP, IA64_SP, GP_SCRATCH_REG);
+
+ ia64_adds_imm (code, IA64_SP, 16, IA64_SP);
+
+ g_free (cinfo);
+ }
+
+ ia64_codegen_close (code);
+
+ return code.buf;
}
void*
mono_arch_instrument_epilog (MonoCompile *cfg, void *func, void *p, gboolean enable_arguments)
{
- NOT_IMPLEMENTED;
+ Ia64CodegenState code;
+ CallInfo *cinfo = NULL;
+ MonoMethod *method = cfg->method;
+ MonoMethodSignature *sig = mono_method_signature (cfg->method);
- return NULL;
+ ia64_codegen_init (code, p);
+
+ cinfo = get_call_info (sig, FALSE);
+
+ /* Save return value + pass it to func */
+ switch (cinfo->ret.storage) {
+ case ArgNone:
+ break;
+ case ArgInIReg:
+ ia64_mov (code, cfg->arch.reg_saved_return_val, cinfo->ret.reg);
+ ia64_mov (code, cfg->arch.reg_out0 + 1, cinfo->ret.reg);
+ break;
+ case ArgInFloatReg:
+ ia64_adds_imm (code, IA64_SP, -16, IA64_SP);
+ ia64_adds_imm (code, GP_SCRATCH_REG, 16, IA64_SP);
+ ia64_stfd_hint (code, GP_SCRATCH_REG, cinfo->ret.reg, 0);
+ ia64_fmov (code, 8 + 1, cinfo->ret.reg);
+ break;
+ case ArgValuetypeAddrInIReg:
+ ia64_mov (code, cfg->arch.reg_out0 + 1, cfg->arch.reg_in0 + cinfo->ret.reg);
+ break;
+ case ArgAggregate:
+ NOT_IMPLEMENTED;
+ break;
+ default:
+ break;
+ }
+
+ g_free (cinfo);
+
+ add_patch_info (cfg, code, MONO_PATCH_INFO_METHODCONST, method);
+ ia64_movl (code, cfg->arch.reg_out0 + 0, 0);
+ code = emit_call (cfg, code, MONO_PATCH_INFO_ABS, (gpointer)func);
+
+ /* Restore return value */
+ switch (cinfo->ret.storage) {
+ case ArgNone:
+ break;
+ case ArgInIReg:
+ ia64_mov (code, cinfo->ret.reg, cfg->arch.reg_saved_return_val);
+ break;
+ case ArgInFloatReg:
+ ia64_adds_imm (code, GP_SCRATCH_REG, 16, IA64_SP);
+ ia64_ldfd (code, cinfo->ret.reg, GP_SCRATCH_REG);
+ break;
+ case ArgValuetypeAddrInIReg:
+ break;
+ case ArgAggregate:
+ break;
+ default:
+ break;
+ }
+
+ ia64_codegen_close (code);
+
+ return code.buf;
}
void
di->u.pi.regions = cfg->arch.r_pro;
_U_dyn_register (di);
+
+ /*
+ {
+ unw_dyn_region_info_t *region = di->u.pi.regions;
+
+ printf ("Unwind info for method %s:\n", mono_method_full_name (cfg->method, TRUE));
+ while (region) {
+ printf (" [Region: %d]\n", region->insn_count);
+ region = region->next;
+ }
+ }
+ */
}
void
guint8* p = (guint8*)((guint64)code & ~(0x3f));
guint8* end = (guint8*)((guint64)code + size);
+#ifdef __INTEL_COMPILER
+ /* icc doesn't define an fc.i instrinsic, but fc==fc.i on itanium 2 */
+ while (p < end) {
+ __fc ((guint64)p);
+ p += 32;
+ }
+#else
while (p < end) {
__asm__ __volatile__ ("fc.i %0"::"r"(p));
+ /* FIXME: This could be increased to 128 on some cpus */
p += 32;
}
+#endif
}
void
gpointer*
mono_arch_get_vcall_slot_addr (guint8* code, gpointer *regs)
{
- guint8 *bundle1 = code - 48;
- guint8 *bundle2 = code - 32;
- guint8 *bundle3 = code - 16;
- guint64 ins11 = ia64_bundle_ins1 (bundle1);
- guint64 ins12 = ia64_bundle_ins2 (bundle1);
- guint64 ins13 = ia64_bundle_ins3 (bundle1);
+ guint8 *bundle2 = code - 48;
+ guint8 *bundle3 = code - 32;
+ guint8 *bundle4 = code - 16;
guint64 ins21 = ia64_bundle_ins1 (bundle2);
guint64 ins22 = ia64_bundle_ins2 (bundle2);
guint64 ins23 = ia64_bundle_ins3 (bundle2);
guint64 ins31 = ia64_bundle_ins1 (bundle3);
guint64 ins32 = ia64_bundle_ins2 (bundle3);
guint64 ins33 = ia64_bundle_ins3 (bundle3);
+ guint64 ins41 = ia64_bundle_ins1 (bundle4);
+ guint64 ins42 = ia64_bundle_ins2 (bundle4);
+ guint64 ins43 = ia64_bundle_ins3 (bundle4);
int reg;
/*
* Virtual calls are made with:
- * [MII] nop.m 0x0
- * mov.sptk b6=r8,0x2000000000f32a80
+ *
+ * [MII] ld8 r31=[r8]
+ * nop.i 0x0
* nop.i 0x0;;
* [MII] nop.m 0x0
+ * mov.sptk b6=r31,0x2000000000f32a80
+ * nop.i 0x0
+ * [MII] nop.m 0x0
* nop.i 0x123456
* nop.i 0x0
* [MIB] nop.m 0x0
* br.call.sptk.few b0=b6;;
*/
- if ((ia64_bundle_template (bundle2) == IA64_TEMPLATE_MIIS) &&
- (ia64_bundle_template (bundle3) == IA64_TEMPLATE_MIBS) &&
- (ins21 == IA64_NOP_M) &&
- (ia64_ins_opcode (ins22) == 0) && (ia64_ins_x3 (ins22) == 0) && (ia64_ins_x6 (ins22) == 0x1) && (ia64_ins_y (ins22) == 0) &&
- (ins23 == IA64_NOP_I) &&
- (ins31 == IA64_NOP_M) &&
- (ins32 == IA64_NOP_I) &&
- (ia64_ins_opcode (ins33) == 1) && (ia64_ins_b1 (ins33) == 0) && (ia64_ins_b2 (ins33) == 6) &&
- ((ins22 >> 6) & 0xfffff) == 0x12345) {
- g_assert (ins11 == IA64_NOP_M);
- g_assert (ins13 == IA64_NOP_I);
- g_assert (ia64_ins_opcode (ins12) == 0);
- g_assert (ia64_ins_x3 (ins12) == 7);
- g_assert (ia64_ins_x (ins12) == 0);
- g_assert (ia64_ins_b1 (ins12) == IA64_B6);
-
- reg = ia64_ins_r2 (ins12);
+ if (((ia64_bundle_template (bundle3) == IA64_TEMPLATE_MII) ||
+ (ia64_bundle_template (bundle3) == IA64_TEMPLATE_MIIS)) &&
+ (ia64_bundle_template (bundle4) == IA64_TEMPLATE_MIBS) &&
+ (ins31 == IA64_NOP_M) &&
+ (ia64_ins_opcode (ins32) == 0) && (ia64_ins_x3 (ins32) == 0) && (ia64_ins_x6 (ins32) == 0x1) && (ia64_ins_y (ins32) == 0) &&
+ (ins33 == IA64_NOP_I) &&
+ (ins41 == IA64_NOP_M) &&
+ (ins42 == IA64_NOP_I) &&
+ (ia64_ins_opcode (ins43) == 1) && (ia64_ins_b1 (ins43) == 0) && (ia64_ins_b2 (ins43) == 6) &&
+ ((ins32 >> 6) & 0xfffff) == 0x12345) {
+ g_assert (ins21 == IA64_NOP_M);
+ g_assert (ins23 == IA64_NOP_I);
+ g_assert (ia64_ins_opcode (ins22) == 0);
+ g_assert (ia64_ins_x3 (ins22) == 7);
+ g_assert (ia64_ins_x (ins22) == 0);
+ g_assert (ia64_ins_b1 (ins22) == IA64_B6);
+
+ reg = IA64_R8;
/*
* Must be a scratch register, since only those are saved by the trampoline
static gboolean tls_offset_inited = FALSE;
-#ifdef MONO_ARCH_SIGSEGV_ON_ALTSTACK
-
-static void
-setup_stack (MonoJitTlsData *tls)
-{
- NOT_IMPLEMENTED;
-}
-
-#endif
-
void
mono_arch_setup_jit_tls_data (MonoJitTlsData *tls)
{
tls_offset_inited = TRUE;
appdomain_tls_offset = mono_domain_get_tls_offset ();
- lmf_tls_offset = mono_get_lmf_tls_offset ();
thread_tls_offset = mono_thread_get_tls_offset ();
}
-
-#ifdef MONO_ARCH_SIGSEGV_ON_ALTSTACK
- setup_stack (tls);
-#endif
}
void
mono_arch_free_jit_tls_data (MonoJitTlsData *tls)
{
-#ifdef MONO_ARCH_SIGSEGV_ON_ALTSTACK
- struct sigaltstack sa;
-
- sa.ss_sp = tls->signal_stack;
- sa.ss_size = SIGNAL_STACK_SIZE;
- sa.ss_flags = SS_DISABLE;
- sigaltstack (&sa, NULL);
-
- if (tls->signal_stack)
- munmap (tls->signal_stack, SIGNAL_STACK_SIZE);
-#endif
}
void
CallInfo * cinfo = get_call_info (inst->signature, FALSE);
MonoInst *vtarg;
- if (cinfo->ret.storage == ArgAggregateReturnInReg) {
+ if (cinfo->ret.storage == ArgAggregate) {
MonoInst *local = (MonoInst*)cfg->arch.ret_var_addr_local;
/*
- * The valuetype is in r8..r11 after the call, need to be copied to
- * the stack. Save the address to a local here, so the call instruction
- * can access it.
+ * The valuetype is in registers after the call, need to be copied
+ * to the stack. Save the address to a local here, so the call
+ * instruction can access it.
*/
g_assert (local->opcode == OP_REGOFFSET);
MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STOREI8_MEMBASE_REG, local->inst_basereg, local->inst_offset, vt_reg);
vtarg->dreg = mono_regstate_next_int (cfg->rs);
mono_bblock_add_inst (cfg->cbb, vtarg);
- mono_call_inst_add_outarg_reg (call, vtarg->dreg, out_reg, FALSE);
+ mono_call_inst_add_outarg_reg (cfg, call, vtarg->dreg, out_reg, FALSE);
out_reg ++;
}
this->dreg = mono_regstate_next_int (cfg->rs);
mono_bblock_add_inst (cfg->cbb, this);
- mono_call_inst_add_outarg_reg (call, this->dreg, out_reg, FALSE);
+ mono_call_inst_add_outarg_reg (cfg, call, this->dreg, out_reg, FALSE);
}
}
{
MonoInst *ins = NULL;
- /* FIXME: */
- return NULL;
-
- if (cmethod->klass == mono_defaults.math_class) {
- if (strcmp (cmethod->name, "Sin") == 0) {
- MONO_INST_NEW (cfg, ins, OP_SIN);
- ins->inst_i0 = args [0];
- } else if (strcmp (cmethod->name, "Cos") == 0) {
- MONO_INST_NEW (cfg, ins, OP_COS);
- ins->inst_i0 = args [0];
- } else if (strcmp (cmethod->name, "Tan") == 0) {
- return ins;
- MONO_INST_NEW (cfg, ins, OP_TAN);
- ins->inst_i0 = args [0];
- } else if (strcmp (cmethod->name, "Atan") == 0) {
- return ins;
- MONO_INST_NEW (cfg, ins, OP_ATAN);
- ins->inst_i0 = args [0];
- } else if (strcmp (cmethod->name, "Sqrt") == 0) {
- MONO_INST_NEW (cfg, ins, OP_SQRT);
- ins->inst_i0 = args [0];
- } else if (strcmp (cmethod->name, "Abs") == 0 && fsig->params [0]->type == MONO_TYPE_R8) {
- MONO_INST_NEW (cfg, ins, OP_ABS);
- ins->inst_i0 = args [0];
- }
-#if 0
- /* OP_FREM is not IEEE compatible */
- else if (strcmp (cmethod->name, "IEEERemainder") == 0) {
- MONO_INST_NEW (cfg, ins, OP_FREM);
- ins->inst_i0 = args [0];
- ins->inst_i1 = args [1];
- }
-#endif
+ if (cmethod->klass == mono_defaults.thread_class &&
+ strcmp (cmethod->name, "MemoryBarrier") == 0) {
+ MONO_INST_NEW (cfg, ins, OP_MEMORY_BARRIER);
} else if(cmethod->klass->image == mono_defaults.corlib &&
(strcmp (cmethod->klass->name_space, "System.Threading") == 0) &&
(strcmp (cmethod->klass->name, "Interlocked") == 0)) {
g_assert_not_reached ();
MONO_INST_NEW (cfg, ins, opcode);
MONO_INST_NEW (cfg, ins_iconst, OP_ICONST);
- ins_iconst->inst_c0 = 1;
+ ins_iconst->inst_imm = 1;
ins->inst_i0 = args [0];
ins->inst_i1 = ins_iconst;
g_assert_not_reached ();
MONO_INST_NEW (cfg, ins, opcode);
MONO_INST_NEW (cfg, ins_iconst, OP_ICONST);
- ins_iconst->inst_c0 = -1;
+ ins_iconst->inst_imm = -1;
ins->inst_i0 = args [0];
ins->inst_i1 = ins_iconst;
- } else if (strcmp (cmethod->name, "Add") == 0) {
- guint32 opcode;
-
- if (fsig->params [0]->type == MONO_TYPE_I4)
- opcode = OP_ATOMIC_ADD_I4;
- else if (fsig->params [0]->type == MONO_TYPE_I8)
- opcode = OP_ATOMIC_ADD_I8;
- else
- g_assert_not_reached ();
-
- MONO_INST_NEW (cfg, ins, opcode);
-
- ins->inst_i0 = args [0];
- ins->inst_i1 = args [1];
+ /* FIXME: */
} else if (strcmp (cmethod->name, "Exchange") == 0) {
guint32 opcode;
MONO_INST_NEW (cfg, ins, CEE_LDIND_I8);
ins->inst_i0 = args [0];
}
-
- /*
- * Can't implement CompareExchange methods this way since they have
- * three arguments.
- */
}
return ins;