#include <mono/utils/mono-tls.h>
#include <mono/utils/mono-hwcap.h>
#include <mono/utils/mono-threads.h>
+#include <mono/utils/unlocked.h>
#include "trace.h"
#include "ir-emit.h"
case MONO_TYPE_U4:
case MONO_TYPE_I:
case MONO_TYPE_U:
- case MONO_TYPE_STRING:
case MONO_TYPE_OBJECT:
- case MONO_TYPE_CLASS:
- case MONO_TYPE_SZARRAY:
case MONO_TYPE_PTR:
case MONO_TYPE_FNPTR:
- case MONO_TYPE_ARRAY:
case MONO_TYPE_I8:
case MONO_TYPE_U8:
class2 = ARG_CLASS_INTEGER;
case MONO_TYPE_U:
case MONO_TYPE_PTR:
case MONO_TYPE_FNPTR:
- case MONO_TYPE_CLASS:
case MONO_TYPE_OBJECT:
- case MONO_TYPE_SZARRAY:
- case MONO_TYPE_ARRAY:
- case MONO_TYPE_STRING:
cinfo->ret.storage = ArgInIReg;
cinfo->ret.reg = AMD64_RAX;
break;
case MONO_TYPE_I1:
case MONO_TYPE_U1:
add_general (&gr, &stack_size, ainfo);
+ ainfo->byte_arg_size = 1;
break;
case MONO_TYPE_I2:
case MONO_TYPE_U2:
add_general (&gr, &stack_size, ainfo);
+ ainfo->byte_arg_size = 2;
break;
case MONO_TYPE_I4:
case MONO_TYPE_U4:
add_general (&gr, &stack_size, ainfo);
+ ainfo->byte_arg_size = 4;
break;
case MONO_TYPE_I:
case MONO_TYPE_U:
case MONO_TYPE_PTR:
case MONO_TYPE_FNPTR:
- case MONO_TYPE_CLASS:
case MONO_TYPE_OBJECT:
- case MONO_TYPE_STRING:
- case MONO_TYPE_SZARRAY:
- case MONO_TYPE_ARRAY:
add_general (&gr, &stack_size, ainfo);
break;
case MONO_TYPE_GENERICINST:
mono_aot_register_jit_icall ("mono_amd64_throw_corlib_exception", mono_amd64_throw_corlib_exception);
mono_aot_register_jit_icall ("mono_amd64_resume_unwind", mono_amd64_resume_unwind);
mono_aot_register_jit_icall ("mono_amd64_get_original_ip", mono_amd64_get_original_ip);
- mono_aot_register_jit_icall ("mono_amd64_handler_block_trampoline_helper", mono_amd64_handler_block_trampoline_helper);
#if defined(MONO_ARCH_GSHAREDVT_SUPPORTED)
mono_aot_register_jit_icall ("mono_amd64_start_gsharedvt_call", mono_amd64_start_gsharedvt_call);
cfg->arch.omit_fp = FALSE;
if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG))
cfg->arch.omit_fp = FALSE;
- if ((mono_jit_trace_calls != NULL && mono_trace_eval (cfg->method)) ||
- (cfg->prof_options & MONO_PROFILE_ENTER_LEAVE))
+ if ((mono_jit_trace_calls != NULL && mono_trace_eval (cfg->method)))
cfg->arch.omit_fp = FALSE;
for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
ArgInfo *ainfo = &cinfo->args [i];
/* Allocate locals */
offsets = mono_allocate_stack_slots (cfg, cfg->arch.omit_fp ? FALSE: TRUE, &locals_stack_size, &locals_stack_align);
- if (locals_stack_size > MONO_ARCH_MAX_FRAME_SIZE) {
- char *mname = mono_method_full_name (cfg->method, TRUE);
- mono_cfg_set_exception_invalid_program (cfg, g_strdup_printf ("Method %s stack is too big.", mname));
- g_free (mname);
- return;
- }
-
if (locals_stack_align) {
offset += (locals_stack_align - 1);
offset &= ~(locals_stack_align - 1);
load->klass = vtaddr->klass;
load->dreg = mono_alloc_ireg (cfg);
MONO_ADD_INS (cfg->cbb, load);
- mini_emit_memcpy (cfg, load->dreg, 0, src->dreg, 0, size, 4);
+ mini_emit_memcpy (cfg, load->dreg, 0, src->dreg, 0, size, SIZEOF_VOID_P);
if (ainfo->pair_storage [0] == ArgInIReg) {
MONO_INST_NEW (cfg, arg, OP_X86_LEA_MEMBASE);
MONO_EMIT_NEW_LOAD_MEMBASE (cfg, dreg, src->dreg, 0);
MONO_EMIT_NEW_STORE_MEMBASE (cfg, OP_STORE_MEMBASE_REG, AMD64_RSP, ainfo->offset, dreg);
} else if (size <= 40) {
- mini_emit_memcpy (cfg, AMD64_RSP, ainfo->offset, src->dreg, 0, size, 4);
+ mini_emit_memcpy (cfg, AMD64_RSP, ainfo->offset, src->dreg, 0, size, SIZEOF_VOID_P);
} else {
// FIXME: Code growth
- mini_emit_memcpy (cfg, AMD64_RSP, ainfo->offset, src->dreg, 0, size, 4);
+ mini_emit_memcpy (cfg, AMD64_RSP, ainfo->offset, src->dreg, 0, size, SIZEOF_VOID_P);
}
if (cfg->compute_gc_maps) {
typedef struct {
MonoMethodSignature *sig;
CallInfo *cinfo;
+ int nstack_args;
} ArchDynCallInfo;
static gboolean
case ArgInFloatSSEReg:
case ArgInDoubleSSEReg:
case ArgValuetypeInReg:
- break;
case ArgOnStack:
- if (!(ainfo->offset + (ainfo->arg_size / 8) <= DYN_CALL_STACK_ARGS))
- return FALSE;
break;
default:
return FALSE;
{
ArchDynCallInfo *info;
CallInfo *cinfo;
+ int i;
cinfo = get_call_info (NULL, sig);
// FIXME: Preprocess the info to speed up get_dyn_call_args ().
info->sig = sig;
info->cinfo = cinfo;
+ info->nstack_args = 0;
+
+ for (i = 0; i < cinfo->nargs; ++i) {
+ ArgInfo *ainfo = &cinfo->args [i];
+ switch (ainfo->storage) {
+ case ArgOnStack:
+ info->nstack_args = MAX (info->nstack_args, ainfo->offset + (ainfo->arg_size / 8));
+ break;
+ default:
+ break;
+ }
+ }
+ /* Align to 16 bytes */
+ if (info->nstack_args & 1)
+ info->nstack_args ++;
return (MonoDynCallInfo*)info;
}
g_free (ainfo);
}
+int
+mono_arch_dyn_call_get_buf_size (MonoDynCallInfo *info)
+{
+ ArchDynCallInfo *ainfo = (ArchDynCallInfo*)info;
+
+ /* Extend the 'regs' field dynamically */
+ return sizeof (DynCallArgs) + (ainfo->nstack_args * sizeof (mgreg_t));
+}
+
#define PTR_TO_GREG(ptr) (mgreg_t)(ptr)
#define GREG_TO_PTR(greg) (gpointer)(greg)
* libffi.
*/
void
-mono_arch_start_dyn_call (MonoDynCallInfo *info, gpointer **args, guint8 *ret, guint8 *buf, int buf_len)
+mono_arch_start_dyn_call (MonoDynCallInfo *info, gpointer **args, guint8 *ret, guint8 *buf)
{
ArchDynCallInfo *dinfo = (ArchDynCallInfo*)info;
DynCallArgs *p = (DynCallArgs*)buf;
param_reg_to_index_inited = 1;
}
- g_assert (buf_len >= sizeof (DynCallArgs));
-
p->res = 0;
p->ret = ret;
+ p->nstack_args = dinfo->nstack_args;
arg_index = 0;
greg = 0;
}
switch (t->type) {
- case MONO_TYPE_STRING:
- case MONO_TYPE_CLASS:
- case MONO_TYPE_ARRAY:
- case MONO_TYPE_SZARRAY:
case MONO_TYPE_OBJECT:
case MONO_TYPE_PTR:
case MONO_TYPE_I:
case MONO_TYPE_VOID:
*(gpointer*)ret = NULL;
break;
- case MONO_TYPE_STRING:
- case MONO_TYPE_CLASS:
- case MONO_TYPE_ARRAY:
- case MONO_TYPE_SZARRAY:
case MONO_TYPE_OBJECT:
case MONO_TYPE_I:
case MONO_TYPE_U:
amd64_call_code (code, 0);
}
else {
+ if (!no_patch && ((guint32)(code + 2 - cfg->native_code) % 8) != 0) {
+ guint32 pad_size = 8 - ((guint32)(code + 2 - cfg->native_code) % 8);
+ amd64_padding (code, pad_size);
+ g_assert ((guint64)(code + 2 - cfg->native_code) % 8 == 0);
+ }
mono_add_patch_info (cfg, code - cfg->native_code, patch_type, data);
amd64_set_reg_template (code, GP_SCRATCH_REG);
amd64_call_reg (code, GP_SCRATCH_REG);
#if defined(TARGET_WIN32)
need_touch = TRUE;
#elif defined(MONO_ARCH_SIGSEGV_ON_ALTSTACK)
- if (!tree->flags & MONO_INST_INIT)
+ if (!(tree->flags & MONO_INST_INIT))
need_touch = TRUE;
#endif
if (cfg->verbose_level > 2)
g_print ("Basic block %d starting at offset 0x%x\n", bb->block_num, bb->native_offset);
- if ((cfg->prof_options & MONO_PROFILE_COVERAGE) && cfg->coverage_info) {
- MonoProfileCoverageInfo *cov = cfg->coverage_info;
- g_assert (!cfg->compile_aot);
-
- cov->data [bb->dfn].cil_code = bb->cil_code;
- amd64_mov_reg_imm (code, AMD64_R11, (guint64)&cov->data [bb->dfn].count);
- /* this is not thread save, but good enough */
- amd64_inc_membase (code, AMD64_R11, 0);
- }
-
offset = code - cfg->native_code;
mono_debug_open_block (cfg, bb, offset);
#define EXTRA_CODE_SPACE (16)
- if (G_UNLIKELY (offset > (cfg->code_size - max_len - EXTRA_CODE_SPACE))) {
+ if (G_UNLIKELY ((offset + max_len + EXTRA_CODE_SPACE) > cfg->code_size)) {
cfg->code_size *= 2;
cfg->native_code = (unsigned char *)mono_realloc_native_code(cfg);
code = cfg->native_code + offset;
g_assert (!cfg->method->save_lmf);
+ /* the size of the tailcall op depends on signature, let's check for enough
+ * space in the code buffer here again */
+ max_len += AMD64_NREG * 4 + call->stack_usage * 15 + EXTRA_CODE_SPACE;
+
+ if (G_UNLIKELY (offset + max_len > cfg->code_size)) {
+ cfg->code_size *= 2;
+ cfg->native_code = (unsigned char *) mono_realloc_native_code(cfg);
+ code = cfg->native_code + offset;
+ cfg->stat_code_reallocs++;
+ }
+
/* Restore callee saved registers */
save_area_offset = cfg->arch.reg_save_area_offset;
for (i = 0; i < AMD64_NREG; ++i)
#endif
}
- offset = code - cfg->native_code;
mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_METHOD_JUMP, call->method);
if (cfg->compile_aot)
amd64_mov_reg_membase (code, AMD64_R11, AMD64_RIP, 0, 8);
code = emit_move_return_value (cfg, ins, code);
break;
case OP_DYN_CALL: {
- int i;
+ int i, limit_reg, index_reg, src_reg, dst_reg;
MonoInst *var = cfg->dyn_call_var;
guint8 *label;
+ guint8 *buf [16];
g_assert (var->opcode == OP_REGOFFSET);
amd64_sse_movsd_reg_membase (code, i, AMD64_R11, MONO_STRUCT_OFFSET (DynCallArgs, fregs) + (i * sizeof (double)));
amd64_patch (label, code);
+ /* Allocate param area */
+ /* This doesn't need to be freed since OP_DYN_CALL is never called in a loop */
+ amd64_mov_reg_membase (code, AMD64_RAX, AMD64_R11, MONO_STRUCT_OFFSET (DynCallArgs, nstack_args), 8);
+ amd64_shift_reg_imm (code, X86_SHL, AMD64_RAX, 3);
+ amd64_alu_reg_reg (code, X86_SUB, AMD64_RSP, AMD64_RAX);
/* Set stack args */
- for (i = 0; i < DYN_CALL_STACK_ARGS; ++i) {
- amd64_mov_reg_membase (code, AMD64_RAX, AMD64_R11, MONO_STRUCT_OFFSET (DynCallArgs, regs) + ((PARAM_REGS + i) * sizeof(mgreg_t)), sizeof(mgreg_t));
- amd64_mov_membase_reg (code, AMD64_RSP, i * sizeof (mgreg_t), AMD64_RAX, sizeof (mgreg_t));
- }
+ /* rax/rcx/rdx/r8/r9 is scratch */
+ limit_reg = AMD64_RAX;
+ index_reg = AMD64_RCX;
+ src_reg = AMD64_R8;
+ dst_reg = AMD64_R9;
+ amd64_mov_reg_membase (code, limit_reg, AMD64_R11, MONO_STRUCT_OFFSET (DynCallArgs, nstack_args), 8);
+ amd64_mov_reg_imm (code, index_reg, 0);
+ amd64_lea_membase (code, src_reg, AMD64_R11, MONO_STRUCT_OFFSET (DynCallArgs, regs) + ((PARAM_REGS) * sizeof(mgreg_t)));
+ amd64_mov_reg_reg (code, dst_reg, AMD64_RSP, 8);
+ buf [0] = code;
+ x86_jump8 (code, 0);
+ buf [1] = code;
+ amd64_mov_reg_membase (code, AMD64_RDX, src_reg, 0, 8);
+ amd64_mov_membase_reg (code, dst_reg, 0, AMD64_RDX, 8);
+ amd64_alu_reg_imm (code, X86_ADD, index_reg, 1);
+ amd64_alu_reg_imm (code, X86_ADD, src_reg, 8);
+ amd64_alu_reg_imm (code, X86_ADD, dst_reg, 8);
+ amd64_patch (buf [0], code);
+ amd64_alu_reg_reg (code, X86_CMP, index_reg, limit_reg);
+ buf [2] = code;
+ x86_branch8 (code, X86_CC_LT, 0, FALSE);
+ amd64_patch (buf [2], buf [1]);
/* Set argument registers */
for (i = 0; i < PARAM_REGS; ++i)
- amd64_mov_reg_membase (code, param_regs [i], AMD64_R11, i * sizeof(mgreg_t), sizeof(mgreg_t));
+ amd64_mov_reg_membase (code, param_regs [i], AMD64_R11, MONO_STRUCT_OFFSET (DynCallArgs, regs) + (i * sizeof(mgreg_t)), sizeof(mgreg_t));
/* Make the call */
amd64_call_reg (code, AMD64_R10);
amd64_mov_membase_reg (code, spvar->inst_basereg, spvar->inst_offset, AMD64_RSP, sizeof(gpointer));
if ((MONO_BBLOCK_IS_IN_REGION (bb, MONO_REGION_FINALLY) ||
- MONO_BBLOCK_IS_IN_REGION (bb, MONO_REGION_FINALLY)) &&
+ MONO_BBLOCK_IS_IN_REGION (bb, MONO_REGION_FILTER)) &&
cfg->param_area) {
amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, ALIGN_TO (cfg->param_area, MONO_ARCH_FRAME_ALIGNMENT));
}
case OP_GET_LAST_ERROR:
emit_get_last_error(code, ins->dreg);
break;
+ case OP_FILL_PROF_CALL_CTX:
+ for (int i = 0; i < AMD64_NREG; i++)
+ if (AMD64_IS_CALLEE_SAVED_REG (i) || i == AMD64_RSP)
+ amd64_mov_membase_reg (code, ins->sreg1, MONO_STRUCT_OFFSET (MonoContext, gregs) + i * sizeof (mgreg_t), i, sizeof (mgreg_t));
+ break;
default:
g_warning ("unknown opcode %s in %s()\n", mono_inst_name (ins->opcode), __FUNCTION__);
g_assert_not_reached ();
if (mono_jit_trace_calls != NULL)
max_epilog_size += 50;
- if (cfg->prof_options & MONO_PROFILE_ENTER_LEAVE)
- max_epilog_size += 50;
-
max_epilog_size += (AMD64_NREG * 2);
return max_epilog_size;
/* See mono_emit_stack_alloc */
#if defined(MONO_ARCH_SIGSEGV_ON_ALTSTACK)
guint32 remaining_size = alloc_size;
- /*FIXME handle unbounded code expansion, we should use a loop in case of more than X interactions*/
+
+ /* Use a loop for large sizes */
+ if (remaining_size > 10 * 0x1000) {
+ amd64_mov_reg_imm (code, X86_EAX, remaining_size / 0x1000);
+ guint8 *label = code;
+ amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, 0x1000);
+ amd64_test_membase_reg (code, AMD64_RSP, 0, AMD64_RSP);
+ amd64_alu_reg_imm (code, X86_SUB, AMD64_RAX, 1);
+ amd64_alu_reg_imm (code, X86_CMP, AMD64_RAX, 0);
+ guint8 *label2 = code;
+ x86_branch8 (code, X86_CC_NE, 0, FALSE);
+ amd64_patch (label2, label);
+ if (cfg->arch.omit_fp) {
+ cfa_offset += (remaining_size / 0x1000) * 0x1000;
+ mono_emit_unwind_op_def_cfa_offset (cfg, code, cfa_offset);
+ }
+
+ remaining_size = remaining_size % 0x1000;
+ }
+
guint32 required_code_size = ((remaining_size / 0x1000) + 1) * 11; /*11 is the max size of amd64_alu_reg_imm + amd64_test_membase_reg*/
guint32 offset = code - cfg->native_code;
if (G_UNLIKELY (required_code_size >= (cfg->code_size - offset))) {
MonoInst *ins;
int max_length = 0;
- if (cfg->prof_options & MONO_PROFILE_COVERAGE)
- max_length += 6;
/* max alignment for loops */
if ((cfg->opt & MONO_OPT_LOOP) && bb_is_loop_start (bb))
max_length += LOOP_ALIGNMENT;
code = (guint8 *)mono_arch_instrument_prolog (cfg, mono_trace_enter_method, code, TRUE);
}
- if (cfg->prof_options & MONO_PROFILE_ENTER_LEAVE)
- args_clobbered = TRUE;
-
/*
* Optimize the common case of the first bblock making a call with the same
* arguments as the method. This works because the arguments are still in their
} else {
/* FIXME: maybe save the jit tls in the prolog */
}
- if (cfg->used_int_regs & (1 << AMD64_RBP)) {
+ if (cfg->used_int_regs & (1 << AMD64_RBP))
amd64_mov_reg_membase (code, AMD64_RBP, cfg->frame_reg, lmf_offset + MONO_STRUCT_OFFSET (MonoLMF, rbp), 8);
- }
+ if (cfg->arch.omit_fp)
+ /*
+ * emit_setup_lmf () marks RBP as saved, we have to mark it as same value here before clearing up the stack
+ * since its stack slot will become invalid.
+ */
+ mono_emit_unwind_op_same_value (cfg, code, AMD64_RBP);
}
/* Restore callee saved regs */
if (AMD64_IS_CALLEE_SAVED_REG (i) && (cfg->arch.saved_iregs & (1 << i))) {
/* Restore only used_int_regs, not arch.saved_iregs */
#if defined(MONO_SUPPORT_TASKLETS)
- int restore_reg=1;
+ int restore_reg = 1;
#else
- int restore_reg=(cfg->used_int_regs & (1 << i));
+ int restore_reg = (cfg->used_int_regs & (1 << i));
#endif
if (restore_reg) {
amd64_mov_reg_membase (code, i, cfg->frame_reg, save_area_offset, 8);
if (!has_target)
g_free (buff);
}
- mono_profiler_code_buffer_new (start, code - start, MONO_PROFILER_CODE_BUFFER_DELEGATE_INVOKE, NULL);
+ MONO_PROFILER_RAISE (jit_code_buffer, (start, code - start, MONO_PROFILER_CODE_BUFFER_DELEGATE_INVOKE, NULL));
return start;
}
/* Load the vtable */
amd64_mov_reg_membase (code, AMD64_RAX, AMD64_ARG_REG1, MONO_STRUCT_OFFSET (MonoObject, vtable), 8);
amd64_jump_membase (code, AMD64_RAX, offset);
- mono_profiler_code_buffer_new (start, code - start, MONO_PROFILER_CODE_BUFFER_DELEGATE_INVOKE, NULL);
+ MONO_PROFILER_RAISE (jit_code_buffer, (start, code - start, MONO_PROFILER_CODE_BUFFER_DELEGATE_INVOKE, NULL));
tramp_name = mono_get_delegate_virtual_invoke_impl_name (load_imt_reg, offset);
*info = mono_tramp_info_create (tramp_name, start, code - start, NULL, unwind_ops);
}
if (!fail_tramp)
- mono_stats.imt_trampolines_size += code - start;
+ UnlockedAdd (&mono_stats.imt_trampolines_size, code - start);
g_assert (code - start <= size);
g_assert_checked (mono_arch_unwindinfo_validate_size (unwind_ops, MONO_TRAMPOLINE_UNWINDINFO_SIZE(0)));
- mono_profiler_code_buffer_new (start, code - start, MONO_PROFILER_CODE_BUFFER_IMT_TRAMPOLINE, NULL);
+ MONO_PROFILER_RAISE (jit_code_buffer, (start, code - start, MONO_PROFILER_CODE_BUFFER_IMT_TRAMPOLINE, NULL));
mono_tramp_info_register (mono_tramp_info_create (NULL, start, code - start, NULL, unwind_ops), domain);
}
#endif
-gboolean
-mono_arch_print_tree (MonoInst *tree, int arity)
-{
- return 0;
-}
-
mgreg_t
mono_arch_context_get_int_reg (MonoContext *ctx, int reg)
{
ctx->gregs [reg] = val;
}
-gpointer
-mono_arch_install_handler_block_guard (MonoJitInfo *ji, MonoJitExceptionInfo *clause, MonoContext *ctx, gpointer new_value)
-{
- gpointer *sp, old_value;
- char *bp;
-
- /*Load the spvar*/
- bp = (char *)MONO_CONTEXT_GET_BP (ctx);
- sp = (gpointer *)*(gpointer*)(bp + clause->exvar_offset);
-
- old_value = *sp;
- if (old_value < ji->code_start || (char*)old_value > ((char*)ji->code_start + ji->code_size))
- return old_value;
-
- *sp = new_value;
-
- return old_value;
-}
-
/*
* mono_arch_emit_load_aotconst:
*
projects / mono.git / blobdiff