#include "mini.h"
#include <string.h>
#include <math.h>
-
-#ifndef PLATFORM_WIN32
+#ifdef HAVE_UNISTD_H
#include <unistd.h>
-#include <sys/mman.h>
#endif
#include <mono/metadata/appdomain.h>
#include "trace.h"
#include "mini-x86.h"
#include "inssel.h"
-#include "cpu-pentium.h"
+#include "cpu-x86.h"
/* On windows, these hold the key returned by TlsAlloc () */
static gint lmf_tls_offset = -1;
+static gint lmf_addr_tls_offset = -1;
static gint appdomain_tls_offset = -1;
static gint thread_tls_offset = -1;
+#ifdef MONO_XEN_OPT
+static gboolean optimize_for_xen = TRUE;
+#else
+#define optimize_for_xen 0
+#endif
+
+#ifdef PLATFORM_WIN32
+static gboolean is_win32 = TRUE;
+#else
+static gboolean is_win32 = FALSE;
+#endif
+
#define ALIGN_TO(val,align) ((((guint64)val) + ((align) - 1)) & ~((align) - 1))
+#define ARGS_OFFSET 8
+
#ifdef PLATFORM_WIN32
/* Under windows, the default pinvoke calling convention is stdcall */
#define CALLCONV_IS_STDCALL(sig) ((((sig)->call_convention) == MONO_CALL_STDCALL) || ((sig)->pinvoke && ((sig)->call_convention) == MONO_CALL_DEFAULT))
#define CALLCONV_IS_STDCALL(sig) (((sig)->call_convention) == MONO_CALL_STDCALL)
#endif
-#define SIGNAL_STACK_SIZE (64 * 1024)
-
#define NOT_IMPLEMENTED g_assert_not_reached ()
const char*
return "unknown";
}
+const char*
+mono_arch_fregname (int reg) {
+ return "unknown";
+}
+
typedef enum {
ArgInIReg,
ArgInFloatSSEReg,
guint32 reg_usage;
guint32 freg_usage;
gboolean need_stack_align;
+ guint32 stack_align_amount;
ArgInfo ret;
ArgInfo sig_cookie;
ArgInfo args [1];
static X86_Reg_No param_regs [] = { 0 };
-#ifdef PLATFORM_WIN32
+#if defined(PLATFORM_WIN32) || defined(__APPLE__) || defined(__FreeBSD__)
+#define SMALL_STRUCTS_IN_REGS
static X86_Reg_No return_regs [] = { X86_EAX, X86_EDX };
#endif
if (*gr >= FLOAT_PARAM_REGS) {
ainfo->storage = ArgOnStack;
- (*stack_size) += sizeof (gpointer);
+ (*stack_size) += is_double ? 8 : 4;
}
else {
/* A double register */
else
size = mono_type_stack_size (&klass->byval_arg, NULL);
-#ifdef PLATFORM_WIN32
+#ifdef SMALL_STRUCTS_IN_REGS
if (sig->pinvoke && is_return) {
MonoMarshalType *info;
* For x86 win32, see ???.
*/
static CallInfo*
-get_call_info (MonoMethodSignature *sig, gboolean is_pinvoke)
+get_call_info (MonoMemPool *mp, MonoMethodSignature *sig, gboolean is_pinvoke)
{
guint32 i, gr, fr;
MonoType *ret_type;
guint32 stack_size = 0;
CallInfo *cinfo;
- cinfo = g_malloc0 (sizeof (CallInfo) + (sizeof (ArgInfo) * n));
+ if (mp)
+ cinfo = mono_mempool_alloc0 (mp, sizeof (CallInfo) + (sizeof (ArgInfo) * n));
+ else
+ cinfo = g_malloc0 (sizeof (CallInfo) + (sizeof (ArgInfo) * n));
gr = 0;
fr = 0;
case MONO_TYPE_R8:
cinfo->ret.storage = ArgOnDoubleFpStack;
break;
+ case MONO_TYPE_GENERICINST:
+ if (!mono_type_generic_inst_is_valuetype (sig->ret)) {
+ cinfo->ret.storage = ArgInIReg;
+ cinfo->ret.reg = X86_EAX;
+ break;
+ }
+ /* Fall through */
case MONO_TYPE_VALUETYPE: {
guint32 tmp_gr = 0, tmp_fr = 0, tmp_stacksize = 0;
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:
add_valuetype (sig, ainfo, sig->params [i], FALSE, &gr, &fr, &stack_size);
break;
add_general (&gr, &stack_size, &cinfo->sig_cookie);
}
+#if defined(__APPLE__)
+ if ((stack_size % 16) != 0) {
+ cinfo->need_stack_align = TRUE;
+ stack_size += cinfo->stack_align_amount = 16-(stack_size % 16);
+ }
+#endif
+
cinfo->stack_usage = stack_size;
cinfo->reg_usage = gr;
cinfo->freg_usage = fr;
mono_arch_get_argument_info (MonoMethodSignature *csig, int param_count, MonoJitArgumentInfo *arg_info)
{
int k, frame_size = 0;
- int size, align, pad;
+ int size, pad;
+ guint32 align;
int offset = 8;
CallInfo *cinfo;
- cinfo = get_call_info (csig, FALSE);
+ cinfo = get_call_info (NULL, csig, FALSE);
if (MONO_TYPE_ISSTRUCT (csig->ret) && (cinfo->ret.storage == ArgOnStack)) {
frame_size += sizeof (gpointer);
if (csig->pinvoke)
size = mono_type_native_stack_size (csig->params [k], &align);
- else
- size = mono_type_stack_size (csig->params [k], &align);
+ else {
+ int ialign;
+ size = mono_type_stack_size (csig->params [k], &ialign);
+ align = ialign;
+ }
/* ignore alignment for now */
align = 1;
#endif
if (have_cpuid) {
/* Have to use the code manager to get around WinXP DEP */
- MonoCodeManager *codeman = mono_code_manager_new_dynamic ();
- CpuidFunc func;
- void *ptr = mono_code_manager_reserve (codeman, sizeof (cpuid_impl));
- memcpy (ptr, cpuid_impl, sizeof (cpuid_impl));
-
- func = (CpuidFunc)ptr;
+ static CpuidFunc func = NULL;
+ void *ptr;
+ if (!func) {
+ ptr = mono_global_codeman_reserve (sizeof (cpuid_impl));
+ memcpy (ptr, cpuid_impl, sizeof (cpuid_impl));
+ func = (CpuidFunc)ptr;
+ }
func (id, p_eax, p_ebx, p_ecx, p_edx);
- mono_code_manager_destroy (codeman);
-
/*
* We use this approach because of issues with gcc and pic code, see:
* http://gcc.gnu.org/cgi-bin/gnatsweb.pl?cmd=view%20audit-trail&database=gcc&pr=7329
__asm__ __volatile__ ("fldcw %0\n": : "m" (fpcw));
__asm__ __volatile__ ("fnstcw %0\n": "=m" (fpcw));
#else
- _control87 (_PC_64, MCW_PC);
+ _control87 (_PC_53, MCW_PC);
#endif
}
/* idiv REG */
switch (x86_modrm_rm (ip [1])) {
+ case X86_EAX:
+ reg = ctx.eax;
+ break;
case X86_ECX:
reg = ctx.ecx;
break;
+ case X86_EDX:
+ reg = ctx.edx;
+ break;
case X86_EBX:
reg = ctx.ebx;
break;
+ case X86_ESI:
+ reg = ctx.esi;
+ break;
+ case X86_EDI:
+ reg = ctx.edi;
+ break;
default:
g_assert_not_reached ();
reg = -1;
return FALSE;
}
-static gboolean
-is_regsize_var (MonoType *t) {
- if (t->byref)
- return TRUE;
- switch (mono_type_get_underlying_type (t)->type) {
- case MONO_TYPE_I4:
- case MONO_TYPE_U4:
- case MONO_TYPE_I:
- case MONO_TYPE_U:
- case MONO_TYPE_PTR:
- case MONO_TYPE_FNPTR:
- return TRUE;
- case MONO_TYPE_OBJECT:
- case MONO_TYPE_STRING:
- case MONO_TYPE_CLASS:
- case MONO_TYPE_SZARRAY:
- case MONO_TYPE_ARRAY:
- return TRUE;
- case MONO_TYPE_VALUETYPE:
- return FALSE;
- }
- return FALSE;
-}
-
GList *
mono_arch_get_allocatable_int_vars (MonoCompile *cfg)
{
/* we dont allocate I1 to registers because there is no simply way to sign extend
* 8bit quantities in caller saved registers on x86 */
- if (is_regsize_var (ins->inst_vtype) || (ins->inst_vtype->type == MONO_TYPE_BOOLEAN) ||
- (ins->inst_vtype->type == MONO_TYPE_U1) || (ins->inst_vtype->type == MONO_TYPE_U2)||
- (ins->inst_vtype->type == MONO_TYPE_I2) || (ins->inst_vtype->type == MONO_TYPE_CHAR)) {
+ if (mono_is_regsize_var (ins->inst_vtype) && (ins->inst_vtype->type != MONO_TYPE_I1)) {
g_assert (MONO_VARINFO (cfg, i)->reg == -1);
g_assert (i == vmv->idx);
vars = g_list_prepend (vars, vmv);
* The locals var stuff should most likely be split in another method.
*/
void
-mono_arch_allocate_vars (MonoCompile *m)
+mono_arch_allocate_vars (MonoCompile *cfg)
{
MonoMethodSignature *sig;
MonoMethodHeader *header;
MonoInst *inst;
guint32 locals_stack_size, locals_stack_align;
- int i, offset, curinst, size, align;
+ int i, offset;
gint32 *offsets;
CallInfo *cinfo;
- header = mono_method_get_header (m->method);
- sig = mono_method_signature (m->method);
-
- offset = 8;
- curinst = 0;
-
- cinfo = get_call_info (sig, FALSE);
-
- switch (cinfo->ret.storage) {
- case ArgOnStack:
- m->ret->opcode = OP_REGOFFSET;
- m->ret->inst_basereg = X86_EBP;
- m->ret->inst_offset = offset;
- offset += sizeof (gpointer);
- break;
- case ArgValuetypeInReg:
- break;
- case ArgInIReg:
- m->ret->opcode = OP_REGVAR;
- m->ret->inst_c0 = cinfo->ret.reg;
- break;
- case ArgNone:
- case ArgOnFloatFpStack:
- case ArgOnDoubleFpStack:
- break;
- default:
- g_assert_not_reached ();
- }
-
- if (sig->hasthis) {
- inst = m->varinfo [curinst];
- if (inst->opcode != OP_REGVAR) {
- inst->opcode = OP_REGOFFSET;
- inst->inst_basereg = X86_EBP;
- }
- inst->inst_offset = offset;
- offset += sizeof (gpointer);
- curinst++;
- }
-
- if (sig->call_convention == MONO_CALL_VARARG) {
- m->sig_cookie = offset;
- offset += sizeof (gpointer);
- }
+ header = mono_method_get_header (cfg->method);
+ sig = mono_method_signature (cfg->method);
- for (i = 0; i < sig->param_count; ++i) {
- inst = m->varinfo [curinst];
- if (inst->opcode != OP_REGVAR) {
- inst->opcode = OP_REGOFFSET;
- inst->inst_basereg = X86_EBP;
- }
- inst->inst_offset = offset;
- size = mono_type_size (sig->params [i], &align);
- size += 4 - 1;
- size &= ~(4 - 1);
- offset += size;
- curinst++;
- }
+ cinfo = get_call_info (cfg->mempool, sig, FALSE);
+ cfg->frame_reg = MONO_ARCH_BASEREG;
offset = 0;
- /* reserve space to save LMF and caller saved registers */
+ /* Reserve space to save LMF and caller saved registers */
- if (m->method->save_lmf) {
+ if (cfg->method->save_lmf) {
offset += sizeof (MonoLMF);
} else {
- if (m->used_int_regs & (1 << X86_EBX)) {
+ if (cfg->used_int_regs & (1 << X86_EBX)) {
offset += 4;
}
- if (m->used_int_regs & (1 << X86_EDI)) {
+ if (cfg->used_int_regs & (1 << X86_EDI)) {
offset += 4;
}
- if (m->used_int_regs & (1 << X86_ESI)) {
+ if (cfg->used_int_regs & (1 << X86_ESI)) {
offset += 4;
}
}
case ArgValuetypeInReg:
/* Allocate a local to hold the result, the epilog will copy it to the correct place */
offset += 8;
- m->ret->opcode = OP_REGOFFSET;
- m->ret->inst_basereg = X86_EBP;
- m->ret->inst_offset = - offset;
+ cfg->ret->opcode = OP_REGOFFSET;
+ cfg->ret->inst_basereg = X86_EBP;
+ cfg->ret->inst_offset = - offset;
break;
default:
break;
}
/* Allocate locals */
- offsets = mono_allocate_stack_slots (m, &locals_stack_size, &locals_stack_align);
+ offsets = mono_allocate_stack_slots (cfg, &locals_stack_size, &locals_stack_align);
if (locals_stack_align) {
offset += (locals_stack_align - 1);
offset &= ~(locals_stack_align - 1);
}
- for (i = m->locals_start; i < m->num_varinfo; i++) {
+ for (i = cfg->locals_start; i < cfg->num_varinfo; i++) {
if (offsets [i] != -1) {
- MonoInst *inst = m->varinfo [i];
+ MonoInst *inst = cfg->varinfo [i];
inst->opcode = OP_REGOFFSET;
inst->inst_basereg = X86_EBP;
inst->inst_offset = - (offset + offsets [i]);
//printf ("allocated local %d to ", i); mono_print_tree_nl (inst);
}
}
- g_free (offsets);
offset += locals_stack_size;
+
+ /*
+ * Allocate arguments+return value
+ */
+
+ switch (cinfo->ret.storage) {
+ case ArgOnStack:
+ cfg->ret->opcode = OP_REGOFFSET;
+ cfg->ret->inst_basereg = X86_EBP;
+ cfg->ret->inst_offset = cinfo->ret.offset + ARGS_OFFSET;
+ break;
+ case ArgValuetypeInReg:
+ break;
+ case ArgInIReg:
+ cfg->ret->opcode = OP_REGVAR;
+ cfg->ret->inst_c0 = cinfo->ret.reg;
+ break;
+ case ArgNone:
+ case ArgOnFloatFpStack:
+ case ArgOnDoubleFpStack:
+ break;
+ default:
+ g_assert_not_reached ();
+ }
+
+ if (sig->call_convention == MONO_CALL_VARARG) {
+ g_assert (cinfo->sig_cookie.storage == ArgOnStack);
+ cfg->sig_cookie = cinfo->sig_cookie.offset + ARGS_OFFSET;
+ }
+
+ for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
+ ArgInfo *ainfo = &cinfo->args [i];
+ inst = cfg->args [i];
+ if (inst->opcode != OP_REGVAR) {
+ inst->opcode = OP_REGOFFSET;
+ inst->inst_basereg = X86_EBP;
+ }
+ inst->inst_offset = ainfo->offset + ARGS_OFFSET;
+ }
+
offset += (MONO_ARCH_FRAME_ALIGNMENT - 1);
offset &= ~(MONO_ARCH_FRAME_ALIGNMENT - 1);
- g_free (cinfo);
-
- /* change sign? */
- m->stack_offset = -offset;
+ cfg->stack_offset = offset;
}
void
sig = mono_method_signature (cfg->method);
- cinfo = get_call_info (sig, FALSE);
+ cinfo = get_call_info (cfg->mempool, sig, FALSE);
if (cinfo->ret.storage == ArgValuetypeInReg)
cfg->ret_var_is_local = TRUE;
-
- g_free (cinfo);
}
/* Fixme: we need an alignment solution for enter_method and mono_arch_call_opcode,
* currently alignment in mono_arch_call_opcode is computed without arch_get_argument_info
*/
+static void
+emit_sig_cookie (MonoCompile *cfg, MonoCallInst *call)
+{
+ MonoInst *arg;
+ MonoMethodSignature *tmp_sig;
+ MonoInst *sig_arg;
+
+ /* FIXME: Add support for signature tokens to AOT */
+ cfg->disable_aot = TRUE;
+ MONO_INST_NEW (cfg, arg, OP_OUTARG);
+
+ /*
+ * mono_ArgIterator_Setup assumes the signature cookie is
+ * passed first and all the arguments which were before it are
+ * passed on the stack after the signature. So compensate by
+ * passing a different signature.
+ */
+ tmp_sig = mono_metadata_signature_dup (call->signature);
+ tmp_sig->param_count -= call->signature->sentinelpos;
+ tmp_sig->sentinelpos = 0;
+ memcpy (tmp_sig->params, call->signature->params + call->signature->sentinelpos, tmp_sig->param_count * sizeof (MonoType*));
+
+ MONO_INST_NEW (cfg, sig_arg, OP_ICONST);
+ sig_arg->inst_p0 = tmp_sig;
+
+ arg->inst_left = sig_arg;
+ arg->type = STACK_PTR;
+ /* prepend, so they get reversed */
+ arg->next = call->out_args;
+ call->out_args = arg;
+}
+
/*
* take the arguments and generate the arch-specific
* instructions to properly call the function in call.
* This includes pushing, moving arguments to the right register
* etc.
- * Issue: who does the spilling if needed, and when?
*/
MonoCallInst*
mono_arch_call_opcode (MonoCompile *cfg, MonoBasicBlock* bb, MonoCallInst *call, int is_virtual) {
MonoInst *arg, *in;
MonoMethodSignature *sig;
- int i, n, stack_size, type;
- MonoType *ptype;
+ int i, n;
CallInfo *cinfo;
+ int sentinelpos = 0;
- stack_size = 0;
- /* add the vararg cookie before the non-implicit args */
- if (call->signature->call_convention == MONO_CALL_VARARG) {
- MonoInst *sig_arg;
- /* FIXME: Add support for signature tokens to AOT */
- cfg->disable_aot = TRUE;
- MONO_INST_NEW (cfg, arg, OP_OUTARG);
- MONO_INST_NEW (cfg, sig_arg, OP_ICONST);
- sig_arg->inst_p0 = call->signature;
- arg->inst_left = sig_arg;
- arg->type = STACK_PTR;
- /* prepend, so they get reversed */
- arg->next = call->out_args;
- call->out_args = arg;
- stack_size += sizeof (gpointer);
- }
sig = call->signature;
n = sig->param_count + sig->hasthis;
- cinfo = get_call_info (sig, FALSE);
+ cinfo = get_call_info (cfg->mempool, sig, FALSE);
- if (sig->ret && MONO_TYPE_ISSTRUCT (sig->ret)) {
- if (cinfo->ret.storage == ArgOnStack)
- stack_size += sizeof (gpointer);
- }
+ if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG))
+ sentinelpos = sig->sentinelpos + (is_virtual ? 1 : 0);
for (i = 0; i < n; ++i) {
+ ArgInfo *ainfo = cinfo->args + i;
+
+ /* Emit the signature cookie just before the implicit arguments */
+ if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG) && (i == sentinelpos)) {
+ emit_sig_cookie (cfg, call);
+ }
+
if (is_virtual && i == 0) {
/* the argument will be attached to the call instrucion */
in = call->args [i];
- stack_size += 4;
} else {
+ MonoType *t;
+
+ if (i >= sig->hasthis)
+ t = sig->params [i - sig->hasthis];
+ else
+ t = &mono_defaults.int_class->byval_arg;
+ t = mono_type_get_underlying_type (t);
+
MONO_INST_NEW (cfg, arg, OP_OUTARG);
in = call->args [i];
arg->cil_code = in->cil_code;
/* prepend, so they get reversed */
arg->next = call->out_args;
call->out_args = arg;
- if (i >= sig->hasthis) {
- MonoType *t = sig->params [i - sig->hasthis];
- ptype = mono_type_get_underlying_type (t);
- if (t->byref)
- type = MONO_TYPE_U;
- else
- type = ptype->type;
- /* FIXME: validate arguments... */
- switch (type) {
- case MONO_TYPE_I:
- case MONO_TYPE_U:
- case MONO_TYPE_BOOLEAN:
- case MONO_TYPE_CHAR:
- case MONO_TYPE_I1:
- case MONO_TYPE_U1:
- case MONO_TYPE_I2:
- case MONO_TYPE_U2:
- case MONO_TYPE_I4:
- case MONO_TYPE_U4:
- case MONO_TYPE_STRING:
- case MONO_TYPE_CLASS:
- case MONO_TYPE_OBJECT:
- case MONO_TYPE_PTR:
- case MONO_TYPE_FNPTR:
- case MONO_TYPE_ARRAY:
- case MONO_TYPE_SZARRAY:
- stack_size += 4;
- break;
- case MONO_TYPE_I8:
- case MONO_TYPE_U8:
- stack_size += 8;
- break;
- case MONO_TYPE_R4:
- stack_size += 4;
- arg->opcode = OP_OUTARG_R4;
- break;
- case MONO_TYPE_R8:
- stack_size += 8;
- arg->opcode = OP_OUTARG_R8;
- break;
- case MONO_TYPE_VALUETYPE: {
- int size;
- if (sig->pinvoke)
- size = mono_type_native_stack_size (&in->klass->byval_arg, NULL);
- else
- size = mono_type_stack_size (&in->klass->byval_arg, NULL);
- stack_size += size;
- arg->opcode = OP_OUTARG_VT;
- arg->klass = in->klass;
- arg->unused = sig->pinvoke;
- arg->inst_imm = size;
- break;
+ if ((i >= sig->hasthis) && (MONO_TYPE_ISSTRUCT(t))) {
+ guint32 size, align;
+
+ if (t->type == MONO_TYPE_TYPEDBYREF) {
+ size = sizeof (MonoTypedRef);
+ align = sizeof (gpointer);
}
- case MONO_TYPE_TYPEDBYREF:
- stack_size += sizeof (MonoTypedRef);
- arg->opcode = OP_OUTARG_VT;
- arg->klass = in->klass;
- arg->unused = sig->pinvoke;
- arg->inst_imm = sizeof (MonoTypedRef);
+ else
+ if (sig->pinvoke)
+ size = mono_type_native_stack_size (&in->klass->byval_arg, &align);
+ else {
+ int ialign;
+ size = mono_type_stack_size (&in->klass->byval_arg, &ialign);
+ align = ialign;
+ }
+ arg->opcode = OP_OUTARG_VT;
+ arg->klass = in->klass;
+ arg->backend.is_pinvoke = sig->pinvoke;
+ arg->inst_imm = size;
+ }
+ else {
+ switch (ainfo->storage) {
+ case ArgOnStack:
+ arg->opcode = OP_OUTARG;
+ if (!t->byref) {
+ if (t->type == MONO_TYPE_R4)
+ arg->opcode = OP_OUTARG_R4;
+ else
+ if (t->type == MONO_TYPE_R8)
+ arg->opcode = OP_OUTARG_R8;
+ }
break;
default:
- g_error ("unknown type 0x%02x in mono_arch_call_opcode\n", type);
+ g_assert_not_reached ();
}
- } else {
- /* the this argument */
- stack_size += 4;
}
}
}
+ /* Handle the case where there are no implicit arguments */
+ if (!sig->pinvoke && (sig->call_convention == MONO_CALL_VARARG) && (n == sentinelpos)) {
+ emit_sig_cookie (cfg, call);
+ }
+
if (sig->ret && MONO_TYPE_ISSTRUCT (sig->ret)) {
if (cinfo->ret.storage == ArgValuetypeInReg) {
MonoInst *zero_inst;
else
/* if the function returns a struct, the called method already does a ret $0x4 */
if (sig->ret && MONO_TYPE_ISSTRUCT (sig->ret))
- stack_size -= 4;
+ cinfo->stack_usage -= 4;
}
+
+ call->stack_usage = cinfo->stack_usage;
- call->stack_usage = stack_size;
- g_free (cinfo);
-
- /*
- * should set more info in call, such as the stack space
- * used by the args that needs to be added back to esp
- */
+#if defined(__APPLE__)
+ if (cinfo->need_stack_align) {
+ MONO_INST_NEW (cfg, arg, OP_X86_OUTARG_ALIGN_STACK);
+ arg->inst_c0 = cinfo->stack_align_amount;
+ arg->next = call->out_args;
+ call->out_args = arg;
+ }
+#endif
return call;
}
{
guchar *code = p;
+#if __APPLE__
+ x86_alu_reg_imm (code, X86_SUB, X86_ESP, 8);
+#endif
+
/* if some args are passed in registers, we need to save them here */
x86_push_reg (code, X86_EBP);
mono_add_patch_info (cfg, code-cfg->native_code, MONO_PATCH_INFO_ABS, func);
x86_call_code (code, 0);
}
+#if __APPLE__
+ x86_alu_reg_imm (code, X86_ADD, X86_ESP, 16);
+#else
x86_alu_reg_imm (code, X86_ADD, X86_ESP, 8);
+#endif
return code;
}
case MONO_TYPE_R8:
save_mode = SAVE_FP;
break;
+ case MONO_TYPE_GENERICINST:
+ if (!mono_type_generic_inst_is_valuetype (mono_method_signature (method)->ret)) {
+ save_mode = SAVE_EAX;
+ break;
+ }
+ /* Fall through */
case MONO_TYPE_VALUETYPE:
save_mode = SAVE_STRUCT;
break;
} \
}
-/* emit an exception if condition is fail */
+/*
+ * Emit an exception if condition is fail and
+ * if possible do a directly branch to target
+ */
#define EMIT_COND_SYSTEM_EXCEPTION(cond,signed,exc_name) \
- do { \
- mono_add_patch_info (cfg, code - cfg->native_code, \
- MONO_PATCH_INFO_EXC, exc_name); \
- x86_branch32 (code, cond, 0, signed); \
+ do { \
+ MonoInst *tins = mono_branch_optimize_exception_target (cfg, bb, exc_name); \
+ if (tins == NULL) { \
+ mono_add_patch_info (cfg, code - cfg->native_code, \
+ MONO_PATCH_INFO_EXC, exc_name); \
+ x86_branch32 (code, cond, 0, signed); \
+ } else { \
+ EMIT_COND_BRANCH (tins, cond, signed); \
+ } \
} while (0);
#define EMIT_FPCOMPARE(code) do { \
static guint8*
emit_call (MonoCompile *cfg, guint8 *code, guint32 patch_type, gconstpointer data)
{
- if (cfg->compile_aot) {
- guint32 got_reg = X86_EAX;
-
- if (cfg->compile_aot) {
- /*
- * Since the patches are generated by the back end, there is
- * no way to generate a got_var at this point.
- */
- g_assert (cfg->got_var);
-
- if (cfg->got_var->opcode == OP_REGOFFSET)
- x86_mov_reg_membase (code, X86_EAX, cfg->got_var->inst_basereg, cfg->got_var->inst_offset, 4);
- else
- got_reg = cfg->got_var->dreg;
- }
-
- mono_add_patch_info (cfg, code - cfg->native_code, patch_type, data);
- x86_call_membase (code, got_reg, 0xf0f0f0f0);
- }
- else {
- mono_add_patch_info (cfg, code - cfg->native_code, patch_type, data);
- x86_call_code (code, 0);
- }
+ mono_add_patch_info (cfg, code - cfg->native_code, patch_type, data);
+ x86_call_code (code, 0);
return code;
}
-/* FIXME: Add more instructions */
-#define INST_IGNORES_CFLAGS(ins) (((ins)->opcode == CEE_BR) || ((ins)->opcode == OP_STORE_MEMBASE_IMM) || ((ins)->opcode == OP_STOREI4_MEMBASE_REG))
+#define INST_IGNORES_CFLAGS(opcode) (!(((opcode) == OP_ADC) || ((opcode) == OP_IADC) || ((opcode) == OP_ADC_IMM) || ((opcode) == OP_IADC_IMM) || ((opcode) == OP_SBB) || ((opcode) == OP_ISBB) || ((opcode) == OP_SBB_IMM) || ((opcode) == OP_ISBB_IMM)))
+/*
+ * peephole_pass_1:
+ *
+ * Perform peephole opts which should/can be performed before local regalloc
+ */
static void
-peephole_pass (MonoCompile *cfg, MonoBasicBlock *bb)
+peephole_pass_1 (MonoCompile *cfg, MonoBasicBlock *bb)
{
MonoInst *ins, *last_ins = NULL;
ins = bb->code;
while (ins) {
-
switch (ins->opcode) {
- case OP_ICONST:
- /* reg = 0 -> XOR (reg, reg) */
- /* XOR sets cflags on x86, so we cant do it always */
- if (ins->inst_c0 == 0 && ins->next && INST_IGNORES_CFLAGS (ins->next)) {
- ins->opcode = CEE_XOR;
- ins->sreg1 = ins->dreg;
- ins->sreg2 = ins->dreg;
- }
+ case OP_IADD_IMM:
+ case OP_ADD_IMM:
+ if ((ins->sreg1 < MONO_MAX_IREGS) && (ins->dreg >= MONO_MAX_IREGS)) {
+ /*
+ * X86_LEA is like ADD, but doesn't have the
+ * sreg1==dreg restriction.
+ */
+ ins->opcode = OP_X86_LEA_MEMBASE;
+ ins->inst_basereg = ins->sreg1;
+ } else if ((ins->inst_imm == 1) && (ins->dreg == ins->sreg1))
+ ins->opcode = OP_X86_INC_REG;
break;
- case OP_MUL_IMM:
- /* remove unnecessary multiplication with 1 */
- if (ins->inst_imm == 1) {
- if (ins->dreg != ins->sreg1) {
- ins->opcode = OP_MOVE;
- } else {
- last_ins->next = ins->next;
- ins = ins->next;
- continue;
- }
- }
+ case OP_SUB_IMM:
+ case OP_ISUB_IMM:
+ if ((ins->sreg1 < MONO_MAX_IREGS) && (ins->dreg >= MONO_MAX_IREGS)) {
+ ins->opcode = OP_X86_LEA_MEMBASE;
+ ins->inst_basereg = ins->sreg1;
+ ins->inst_imm = -ins->inst_imm;
+ } else if ((ins->inst_imm == 1) && (ins->dreg == ins->sreg1))
+ ins->opcode = OP_X86_DEC_REG;
break;
case OP_COMPARE_IMM:
+ case OP_ICOMPARE_IMM:
/* OP_COMPARE_IMM (reg, 0)
* -->
* OP_X86_TEST_NULL (reg)
case OP_LOADU1_MEMBASE:
case OP_LOADI1_MEMBASE:
/*
- * Note: if reg1 = reg2 the load op is removed
- *
* OP_STORE_MEMBASE_REG reg1, offset(basereg)
* OP_LOAD_MEMBASE offset(basereg), reg2
* -->
* OP_STORE_MEMBASE_REG reg1, offset(basereg)
- * OP_MOVE reg1, reg2
+ * CONV_I2/U2 reg1, reg2
*/
- if (last_ins && (last_ins->opcode == OP_STOREI1_MEMBASE_REG) &&
+ if (last_ins && X86_IS_BYTE_REG (last_ins->sreg1) &&
+ (last_ins->opcode == OP_STOREI1_MEMBASE_REG) &&
ins->inst_basereg == last_ins->inst_destbasereg &&
ins->inst_offset == last_ins->inst_offset) {
- if (ins->dreg == last_ins->sreg1) {
- last_ins->next = ins->next;
- ins = ins->next;
- continue;
- } else {
- //static int c = 0; printf ("MATCHX %s %d\n", cfg->method->name,c++);
- ins->opcode = OP_MOVE;
- ins->sreg1 = last_ins->sreg1;
- }
+ ins->opcode = (ins->opcode == OP_LOADI1_MEMBASE) ? CEE_CONV_I1 : CEE_CONV_U1;
+ ins->sreg1 = last_ins->sreg1;
}
break;
case OP_LOADU2_MEMBASE:
case OP_LOADI2_MEMBASE:
/*
- * Note: if reg1 = reg2 the load op is removed
- *
* OP_STORE_MEMBASE_REG reg1, offset(basereg)
* OP_LOAD_MEMBASE offset(basereg), reg2
* -->
* OP_STORE_MEMBASE_REG reg1, offset(basereg)
- * OP_MOVE reg1, reg2
+ * CONV_I2/U2 reg1, reg2
*/
if (last_ins && (last_ins->opcode == OP_STOREI2_MEMBASE_REG) &&
ins->inst_basereg == last_ins->inst_destbasereg &&
ins->inst_offset == last_ins->inst_offset) {
- if (ins->dreg == last_ins->sreg1) {
- last_ins->next = ins->next;
- ins = ins->next;
- continue;
- } else {
- //static int c = 0; printf ("MATCHX %s %d\n", cfg->method->name,c++);
- ins->opcode = OP_MOVE;
- ins->sreg1 = last_ins->sreg1;
- }
+ ins->opcode = (ins->opcode == OP_LOADI2_MEMBASE) ? CEE_CONV_I2 : CEE_CONV_U2;
+ ins->sreg1 = last_ins->sreg1;
}
break;
case CEE_CONV_I4:
case CEE_CONV_U4:
+ case OP_ICONV_TO_I4:
case OP_MOVE:
/*
* Removes:
bb->last_ins = last_ins;
}
-static const int
-branch_cc_table [] = {
- X86_CC_EQ, X86_CC_GE, X86_CC_GT, X86_CC_LE, X86_CC_LT,
- X86_CC_NE, X86_CC_GE, X86_CC_GT, X86_CC_LE, X86_CC_LT,
- X86_CC_O, X86_CC_NO, X86_CC_C, X86_CC_NC
-};
-
-#define DEBUG(a) if (cfg->verbose_level > 1) a
-//#define DEBUG(a)
-
-/*
- * returns the offset used by spillvar. It allocates a new
- * spill variable if necessary.
- */
-static int
-mono_spillvar_offset (MonoCompile *cfg, int spillvar)
+static void
+peephole_pass (MonoCompile *cfg, MonoBasicBlock *bb)
{
- MonoSpillInfo **si, *info;
- int i = 0;
-
- si = &cfg->spill_info;
-
- while (i <= spillvar) {
-
- if (!*si) {
- *si = info = mono_mempool_alloc (cfg->mempool, sizeof (MonoSpillInfo));
- info->next = NULL;
- cfg->stack_offset -= sizeof (gpointer);
- info->offset = cfg->stack_offset;
- }
-
- if (i == spillvar)
- return (*si)->offset;
-
- i++;
- si = &(*si)->next;
- }
-
- g_assert_not_reached ();
- return 0;
-}
+ MonoInst *ins, *last_ins = NULL;
+ ins = bb->code;
-/*
- * returns the offset used by spillvar. It allocates a new
- * spill float variable if necessary.
- * (same as mono_spillvar_offset but for float)
- */
-static int
-mono_spillvar_offset_float (MonoCompile *cfg, int spillvar)
-{
- MonoSpillInfo **si, *info;
- int i = 0;
+ while (ins) {
- si = &cfg->spill_info_float;
-
- while (i <= spillvar) {
+ switch (ins->opcode) {
+ case OP_ICONST:
+ /* reg = 0 -> XOR (reg, reg) */
+ /* XOR sets cflags on x86, so we cant do it always */
+ if (ins->inst_c0 == 0 && (!ins->next || (ins->next && INST_IGNORES_CFLAGS (ins->next->opcode)))) {
+ MonoInst *ins2;
- if (!*si) {
- *si = info = mono_mempool_alloc (cfg->mempool, sizeof (MonoSpillInfo));
- info->next = NULL;
- cfg->stack_offset -= sizeof (double);
- info->offset = cfg->stack_offset;
- }
+ ins->opcode = OP_IXOR;
+ ins->sreg1 = ins->dreg;
+ ins->sreg2 = ins->dreg;
- if (i == spillvar)
- return (*si)->offset;
-
- i++;
- si = &(*si)->next;
- }
-
- g_assert_not_reached ();
- return 0;
-}
-
-/*
- * Creates a store for spilled floating point items
- */
-static MonoInst*
-create_spilled_store_float (MonoCompile *cfg, int spill, int reg, MonoInst *ins)
-{
- MonoInst *store;
- MONO_INST_NEW (cfg, store, OP_STORER8_MEMBASE_REG);
- store->sreg1 = reg;
- store->inst_destbasereg = X86_EBP;
- store->inst_offset = mono_spillvar_offset_float (cfg, spill);
-
- DEBUG (g_print ("SPILLED FLOAT STORE (%d at 0x%08x(%%sp)) (from %d)\n", spill, store->inst_offset, reg));
- return store;
-}
-
-/*
- * Creates a load for spilled floating point items
- */
-static MonoInst*
-create_spilled_load_float (MonoCompile *cfg, int spill, int reg, MonoInst *ins)
-{
- MonoInst *load;
- MONO_INST_NEW (cfg, load, OP_LOADR8_SPILL_MEMBASE);
- load->dreg = reg;
- load->inst_basereg = X86_EBP;
- load->inst_offset = mono_spillvar_offset_float (cfg, spill);
-
- DEBUG (g_print ("SPILLED FLOAT LOAD (%d at 0x%08x(%%sp)) (from %d)\n", spill, load->inst_offset, reg));
- return load;
-}
-
-#define is_global_ireg(r) ((r) >= 0 && (r) < MONO_MAX_IREGS && !X86_IS_CALLEE ((r)))
-#define reg_is_freeable(r) ((r) >= 0 && (r) < MONO_MAX_IREGS && X86_IS_CALLEE ((r)))
-
-typedef struct {
- int born_in;
- int killed_in;
- int last_use;
- int prev_use;
- int flags; /* used to track fp spill/load */
-} RegTrack;
-
-static const char*const * ins_spec = pentium_desc;
-
-static void
-print_ins (int i, MonoInst *ins)
-{
- const char *spec = ins_spec [ins->opcode];
- g_print ("\t%-2d %s", i, mono_inst_name (ins->opcode));
- if (spec [MONO_INST_DEST]) {
- if (ins->dreg >= MONO_MAX_IREGS)
- g_print (" R%d <-", ins->dreg);
- else
- g_print (" %s <-", mono_arch_regname (ins->dreg));
- }
- if (spec [MONO_INST_SRC1]) {
- if (ins->sreg1 >= MONO_MAX_IREGS)
- g_print (" R%d", ins->sreg1);
- else
- g_print (" %s", mono_arch_regname (ins->sreg1));
- }
- if (spec [MONO_INST_SRC2]) {
- if (ins->sreg2 >= MONO_MAX_IREGS)
- g_print (" R%d", ins->sreg2);
- else
- g_print (" %s", mono_arch_regname (ins->sreg2));
- }
- if (spec [MONO_INST_CLOB])
- g_print (" clobbers: %c", spec [MONO_INST_CLOB]);
- g_print ("\n");
-}
-
-static void
-print_regtrack (RegTrack *t, int num)
-{
- int i;
- char buf [32];
- const char *r;
-
- for (i = 0; i < num; ++i) {
- if (!t [i].born_in)
- continue;
- if (i >= MONO_MAX_IREGS) {
- g_snprintf (buf, sizeof(buf), "R%d", i);
- r = buf;
- } else
- r = mono_arch_regname (i);
- g_print ("liveness: %s [%d - %d]\n", r, t [i].born_in, t[i].last_use);
- }
-}
-
-typedef struct InstList InstList;
-
-struct InstList {
- InstList *prev;
- InstList *next;
- MonoInst *data;
-};
-
-static inline InstList*
-inst_list_prepend (MonoMemPool *pool, InstList *list, MonoInst *data)
-{
- InstList *item = mono_mempool_alloc (pool, sizeof (InstList));
- item->data = data;
- item->prev = NULL;
- item->next = list;
- if (list)
- list->prev = item;
- return item;
-}
-
-/*
- * Force the spilling of the variable in the symbolic register 'reg'.
- */
-static int
-get_register_force_spilling (MonoCompile *cfg, InstList *item, MonoInst *ins, int reg)
-{
- MonoInst *load;
- int i, sel, spill;
-
- sel = cfg->rs->iassign [reg];
- /*i = cfg->rs->isymbolic [sel];
- g_assert (i == reg);*/
- i = reg;
- spill = ++cfg->spill_count;
- cfg->rs->iassign [i] = -spill - 1;
- mono_regstate_free_int (cfg->rs, sel);
- /* we need to create a spill var and insert a load to sel after the current instruction */
- MONO_INST_NEW (cfg, load, OP_LOAD_MEMBASE);
- load->dreg = sel;
- load->inst_basereg = X86_EBP;
- load->inst_offset = mono_spillvar_offset (cfg, spill);
- if (item->prev) {
- while (ins->next != item->prev->data)
- ins = ins->next;
- }
- load->next = ins->next;
- ins->next = load;
- DEBUG (g_print ("SPILLED LOAD (%d at 0x%08x(%%ebp)) R%d (freed %s)\n", spill, load->inst_offset, i, mono_arch_regname (sel)));
- i = mono_regstate_alloc_int (cfg->rs, 1 << sel);
- g_assert (i == sel);
-
- return sel;
-}
-
-static int
-get_register_spilling (MonoCompile *cfg, InstList *item, MonoInst *ins, guint32 regmask, int reg)
-{
- MonoInst *load;
- int i, sel, spill;
-
- DEBUG (g_print ("\tstart regmask to assign R%d: 0x%08x (R%d <- R%d R%d)\n", reg, regmask, ins->dreg, ins->sreg1, ins->sreg2));
- /* exclude the registers in the current instruction */
- if (reg != ins->sreg1 && (reg_is_freeable (ins->sreg1) || (ins->sreg1 >= MONO_MAX_IREGS && cfg->rs->iassign [ins->sreg1] >= 0))) {
- if (ins->sreg1 >= MONO_MAX_IREGS)
- regmask &= ~ (1 << cfg->rs->iassign [ins->sreg1]);
- else
- regmask &= ~ (1 << ins->sreg1);
- DEBUG (g_print ("\t\texcluding sreg1 %s\n", mono_arch_regname (ins->sreg1)));
- }
- if (reg != ins->sreg2 && (reg_is_freeable (ins->sreg2) || (ins->sreg2 >= MONO_MAX_IREGS && cfg->rs->iassign [ins->sreg2] >= 0))) {
- if (ins->sreg2 >= MONO_MAX_IREGS)
- regmask &= ~ (1 << cfg->rs->iassign [ins->sreg2]);
- else
- regmask &= ~ (1 << ins->sreg2);
- DEBUG (g_print ("\t\texcluding sreg2 %s %d\n", mono_arch_regname (ins->sreg2), ins->sreg2));
- }
- if (reg != ins->dreg && reg_is_freeable (ins->dreg)) {
- regmask &= ~ (1 << ins->dreg);
- DEBUG (g_print ("\t\texcluding dreg %s\n", mono_arch_regname (ins->dreg)));
- }
-
- DEBUG (g_print ("\t\tavailable regmask: 0x%08x\n", regmask));
- g_assert (regmask); /* need at least a register we can free */
- sel = -1;
- /* we should track prev_use and spill the register that's farther */
- for (i = 0; i < MONO_MAX_IREGS; ++i) {
- if (regmask & (1 << i)) {
- sel = i;
- DEBUG (g_print ("\t\tselected register %s has assignment %d\n", mono_arch_regname (sel), cfg->rs->iassign [sel]));
- break;
- }
- }
- i = cfg->rs->isymbolic [sel];
- spill = ++cfg->spill_count;
- cfg->rs->iassign [i] = -spill - 1;
- mono_regstate_free_int (cfg->rs, sel);
- /* we need to create a spill var and insert a load to sel after the current instruction */
- MONO_INST_NEW (cfg, load, OP_LOAD_MEMBASE);
- load->dreg = sel;
- load->inst_basereg = X86_EBP;
- load->inst_offset = mono_spillvar_offset (cfg, spill);
- if (item->prev) {
- while (ins->next != item->prev->data)
- ins = ins->next;
- }
- load->next = ins->next;
- ins->next = load;
- DEBUG (g_print ("\tSPILLED LOAD (%d at 0x%08x(%%ebp)) R%d (freed %s)\n", spill, load->inst_offset, i, mono_arch_regname (sel)));
- i = mono_regstate_alloc_int (cfg->rs, 1 << sel);
- g_assert (i == sel);
-
- return sel;
-}
-
-static MonoInst*
-create_copy_ins (MonoCompile *cfg, int dest, int src, MonoInst *ins)
-{
- MonoInst *copy;
- MONO_INST_NEW (cfg, copy, OP_MOVE);
- copy->dreg = dest;
- copy->sreg1 = src;
- if (ins) {
- copy->next = ins->next;
- ins->next = copy;
- }
- DEBUG (g_print ("\tforced copy from %s to %s\n", mono_arch_regname (src), mono_arch_regname (dest)));
- return copy;
-}
-
-static MonoInst*
-create_spilled_store (MonoCompile *cfg, int spill, int reg, int prev_reg, MonoInst *ins)
-{
- MonoInst *store;
- MONO_INST_NEW (cfg, store, OP_STORE_MEMBASE_REG);
- store->sreg1 = reg;
- store->inst_destbasereg = X86_EBP;
- store->inst_offset = mono_spillvar_offset (cfg, spill);
- if (ins) {
- store->next = ins->next;
- ins->next = store;
- }
- DEBUG (g_print ("\tSPILLED STORE (%d at 0x%08x(%%ebp)) R%d (from %s)\n", spill, store->inst_offset, prev_reg, mono_arch_regname (reg)));
- return store;
-}
-
-static void
-insert_before_ins (MonoInst *ins, InstList *item, MonoInst* to_insert)
-{
- MonoInst *prev;
- if (item->next) {
- prev = item->next->data;
-
- while (prev->next != ins)
- prev = prev->next;
- to_insert->next = ins;
- prev->next = to_insert;
- } else {
- to_insert->next = ins;
- }
- /*
- * needed otherwise in the next instruction we can add an ins to the
- * end and that would get past this instruction.
- */
- item->data = to_insert;
-}
-
-
-#if 0
-static int
-alloc_int_reg (MonoCompile *cfg, InstList *curinst, MonoInst *ins, int sym_reg, guint32 allow_mask)
-{
- int val = cfg->rs->iassign [sym_reg];
- if (val < 0) {
- int spill = 0;
- if (val < -1) {
- /* the register gets spilled after this inst */
- spill = -val -1;
- }
- val = mono_regstate_alloc_int (cfg->rs, allow_mask);
- if (val < 0)
- val = get_register_spilling (cfg, curinst, ins, allow_mask, sym_reg);
- cfg->rs->iassign [sym_reg] = val;
- /* add option to store before the instruction for src registers */
- if (spill)
- create_spilled_store (cfg, spill, val, sym_reg, ins);
- }
- cfg->rs->isymbolic [val] = sym_reg;
- return val;
-}
-#endif
-
-/* flags used in reginfo->flags */
-enum {
- MONO_X86_FP_NEEDS_LOAD_SPILL = 1 << 0,
- MONO_X86_FP_NEEDS_SPILL = 1 << 1,
- MONO_X86_FP_NEEDS_LOAD = 1 << 2,
- MONO_X86_REG_NOT_ECX = 1 << 3,
- MONO_X86_REG_EAX = 1 << 4,
- MONO_X86_REG_EDX = 1 << 5,
- MONO_X86_REG_ECX = 1 << 6
-};
-
-static int
-mono_x86_alloc_int_reg (MonoCompile *cfg, InstList *tmp, MonoInst *ins, guint32 dest_mask, int sym_reg, int flags)
-{
- int val;
- int test_mask = dest_mask;
-
- if (flags & MONO_X86_REG_EAX)
- test_mask &= (1 << X86_EAX);
- else if (flags & MONO_X86_REG_EDX)
- test_mask &= (1 << X86_EDX);
- else if (flags & MONO_X86_REG_ECX)
- test_mask &= (1 << X86_ECX);
- else if (flags & MONO_X86_REG_NOT_ECX)
- test_mask &= ~ (1 << X86_ECX);
-
- val = mono_regstate_alloc_int (cfg->rs, test_mask);
- if (val >= 0 && test_mask != dest_mask)
- DEBUG(g_print ("\tUsed flag to allocate reg %s for R%u\n", mono_arch_regname (val), sym_reg));
-
- if (val < 0 && (flags & MONO_X86_REG_NOT_ECX)) {
- DEBUG(g_print ("\tFailed to allocate flag suggested mask (%u) but exluding ECX\n", test_mask));
- val = mono_regstate_alloc_int (cfg->rs, (dest_mask & (~1 << X86_ECX)));
- }
-
- if (val < 0) {
- val = mono_regstate_alloc_int (cfg->rs, dest_mask);
- if (val < 0)
- val = get_register_spilling (cfg, tmp, ins, dest_mask, sym_reg);
- }
-
- return val;
-}
-
-static inline void
-assign_ireg (MonoRegState *rs, int reg, int hreg)
-{
- g_assert (reg >= MONO_MAX_IREGS);
- g_assert (hreg < MONO_MAX_IREGS);
- g_assert (! is_global_ireg (hreg));
-
- rs->iassign [reg] = hreg;
- rs->isymbolic [hreg] = reg;
- rs->ifree_mask &= ~ (1 << hreg);
-}
-
-/*#include "cprop.c"*/
-
-/*
- * Local register allocation.
- * We first scan the list of instructions and we save the liveness info of
- * each register (when the register is first used, when it's value is set etc.).
- * We also reverse the list of instructions (in the InstList list) because assigning
- * registers backwards allows for more tricks to be used.
- */
-void
-mono_arch_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
-{
- MonoInst *ins;
- MonoRegState *rs = cfg->rs;
- int i, val, fpcount;
- RegTrack *reginfo, *reginfof;
- RegTrack *reginfo1, *reginfo2, *reginfod;
- InstList *tmp, *reversed = NULL;
- const char *spec;
- guint32 src1_mask, src2_mask, dest_mask;
- GList *fspill_list = NULL;
- int fspill = 0;
-
- if (!bb->code)
- return;
- rs->next_vireg = bb->max_ireg;
- rs->next_vfreg = bb->max_freg;
- mono_regstate_assign (rs);
- reginfo = g_malloc0 (sizeof (RegTrack) * rs->next_vireg);
- reginfof = g_malloc0 (sizeof (RegTrack) * rs->next_vfreg);
- rs->ifree_mask = X86_CALLEE_REGS;
-
- ins = bb->code;
-
- /*if (cfg->opt & MONO_OPT_COPYPROP)
- local_copy_prop (cfg, ins);*/
-
- i = 1;
- fpcount = 0;
- DEBUG (g_print ("LOCAL regalloc: basic block: %d\n", bb->block_num));
- /* forward pass on the instructions to collect register liveness info */
- while (ins) {
- spec = ins_spec [ins->opcode];
-
- DEBUG (print_ins (i, ins));
-
- if (spec [MONO_INST_SRC1]) {
- if (spec [MONO_INST_SRC1] == 'f') {
- GList *spill;
- reginfo1 = reginfof;
-
- spill = g_list_first (fspill_list);
- if (spill && fpcount < MONO_MAX_FREGS) {
- reginfo1 [ins->sreg1].flags |= MONO_X86_FP_NEEDS_LOAD;
- fspill_list = g_list_remove (fspill_list, spill->data);
- } else
- fpcount--;
- }
- else
- reginfo1 = reginfo;
- reginfo1 [ins->sreg1].prev_use = reginfo1 [ins->sreg1].last_use;
- reginfo1 [ins->sreg1].last_use = i;
- if (spec [MONO_INST_SRC1] == 'L') {
- /* The virtual register is allocated sequentially */
- reginfo1 [ins->sreg1 + 1].prev_use = reginfo1 [ins->sreg1 + 1].last_use;
- reginfo1 [ins->sreg1 + 1].last_use = i;
- if (reginfo1 [ins->sreg1 + 1].born_in == 0 || reginfo1 [ins->sreg1 + 1].born_in > i)
- reginfo1 [ins->sreg1 + 1].born_in = i;
-
- reginfo1 [ins->sreg1].flags |= MONO_X86_REG_EAX;
- reginfo1 [ins->sreg1 + 1].flags |= MONO_X86_REG_EDX;
- }
- } else {
- ins->sreg1 = -1;
- }
- if (spec [MONO_INST_SRC2]) {
- if (spec [MONO_INST_SRC2] == 'f') {
- GList *spill;
- reginfo2 = reginfof;
- spill = g_list_first (fspill_list);
- if (spill) {
- reginfo2 [ins->sreg2].flags |= MONO_X86_FP_NEEDS_LOAD;
- fspill_list = g_list_remove (fspill_list, spill->data);
- if (fpcount >= MONO_MAX_FREGS) {
- fspill++;
- fspill_list = g_list_prepend (fspill_list, GINT_TO_POINTER(fspill));
- reginfo2 [ins->sreg2].flags |= MONO_X86_FP_NEEDS_LOAD_SPILL;
- }
- } else
- fpcount--;
- }
- else
- reginfo2 = reginfo;
- reginfo2 [ins->sreg2].prev_use = reginfo2 [ins->sreg2].last_use;
- reginfo2 [ins->sreg2].last_use = i;
- if (spec [MONO_INST_SRC2] == 'L') {
- /* The virtual register is allocated sequentially */
- reginfo2 [ins->sreg2 + 1].prev_use = reginfo2 [ins->sreg2 + 1].last_use;
- reginfo2 [ins->sreg2 + 1].last_use = i;
- if (reginfo2 [ins->sreg2 + 1].born_in == 0 || reginfo2 [ins->sreg2 + 1].born_in > i)
- reginfo2 [ins->sreg2 + 1].born_in = i;
- }
- if (spec [MONO_INST_CLOB] == 's') {
- reginfo2 [ins->sreg1].flags |= MONO_X86_REG_NOT_ECX;
- reginfo2 [ins->sreg2].flags |= MONO_X86_REG_ECX;
- }
- } else {
- ins->sreg2 = -1;
- }
- if (spec [MONO_INST_DEST]) {
- if (spec [MONO_INST_DEST] == 'f') {
- reginfod = reginfof;
- if (fpcount >= MONO_MAX_FREGS) {
- reginfod [ins->dreg].flags |= MONO_X86_FP_NEEDS_SPILL;
- fspill++;
- fspill_list = g_list_prepend (fspill_list, GINT_TO_POINTER(fspill));
- fpcount--;
- }
- fpcount++;
- }
- else
- reginfod = reginfo;
- if (spec [MONO_INST_DEST] != 'b') /* it's not just a base register */
- reginfod [ins->dreg].killed_in = i;
- reginfod [ins->dreg].prev_use = reginfod [ins->dreg].last_use;
- reginfod [ins->dreg].last_use = i;
- if (reginfod [ins->dreg].born_in == 0 || reginfod [ins->dreg].born_in > i)
- reginfod [ins->dreg].born_in = i;
- if (spec [MONO_INST_DEST] == 'l' || spec [MONO_INST_DEST] == 'L') {
- /* The virtual register is allocated sequentially */
- reginfod [ins->dreg + 1].prev_use = reginfod [ins->dreg + 1].last_use;
- reginfod [ins->dreg + 1].last_use = i;
- if (reginfod [ins->dreg + 1].born_in == 0 || reginfod [ins->dreg + 1].born_in > i)
- reginfod [ins->dreg + 1].born_in = i;
-
- reginfod [ins->dreg].flags |= MONO_X86_REG_EAX;
- reginfod [ins->dreg + 1].flags |= MONO_X86_REG_EDX;
- }
- } else {
- ins->dreg = -1;
- }
-
- reversed = inst_list_prepend (cfg->mempool, reversed, ins);
- ++i;
- ins = ins->next;
- }
-
- // todo: check if we have anything left on fp stack, in verify mode?
- fspill = 0;
-
- DEBUG (print_regtrack (reginfo, rs->next_vireg));
- DEBUG (print_regtrack (reginfof, rs->next_vfreg));
- tmp = reversed;
- while (tmp) {
- int prev_dreg, prev_sreg1, prev_sreg2, clob_dreg;
- dest_mask = src1_mask = src2_mask = X86_CALLEE_REGS;
- --i;
- ins = tmp->data;
- spec = ins_spec [ins->opcode];
- prev_dreg = -1;
- clob_dreg = -1;
- DEBUG (g_print ("processing:"));
- DEBUG (print_ins (i, ins));
- if (spec [MONO_INST_CLOB] == 's') {
- /*
- * Shift opcodes, SREG2 must be RCX
- */
- if (rs->ifree_mask & (1 << X86_ECX)) {
- if (ins->sreg2 < MONO_MAX_IREGS) {
- /* Argument already in hard reg, need to copy */
- MonoInst *copy = create_copy_ins (cfg, X86_ECX, ins->sreg2, NULL);
- insert_before_ins (ins, tmp, copy);
- }
- else {
- DEBUG (g_print ("\tshortcut assignment of R%d to ECX\n", ins->sreg2));
- assign_ireg (rs, ins->sreg2, X86_ECX);
- }
- } else {
- int need_ecx_spill = TRUE;
- /*
- * we first check if src1/dreg is already assigned a register
- * and then we force a spill of the var assigned to ECX.
- */
- /* the destination register can't be ECX */
- dest_mask &= ~ (1 << X86_ECX);
- src1_mask &= ~ (1 << X86_ECX);
- val = rs->iassign [ins->dreg];
/*
- * the destination register is already assigned to ECX:
- * we need to allocate another register for it and then
- * copy from this to ECX.
+ * Convert succeeding STORE_MEMBASE_IMM 0 ins to STORE_MEMBASE_REG
+ * since it takes 3 bytes instead of 7.
*/
- if (val == X86_ECX && ins->dreg != ins->sreg2) {
- int new_dest;
- new_dest = mono_x86_alloc_int_reg (cfg, tmp, ins, dest_mask, ins->dreg, reginfo [ins->dreg].flags);
- g_assert (new_dest >= 0);
- DEBUG (g_print ("\tclob:s changing dreg R%d to %s from ECX\n", ins->dreg, mono_arch_regname (new_dest)));
-
- rs->isymbolic [new_dest] = ins->dreg;
- rs->iassign [ins->dreg] = new_dest;
- clob_dreg = ins->dreg;
- ins->dreg = new_dest;
- create_copy_ins (cfg, X86_ECX, new_dest, ins);
- need_ecx_spill = FALSE;
- /*DEBUG (g_print ("\tforced spill of R%d\n", ins->dreg));
- val = get_register_force_spilling (cfg, tmp, ins, ins->dreg);
- rs->iassign [ins->dreg] = val;
- rs->isymbolic [val] = prev_dreg;
- ins->dreg = val;*/
- }
- if (is_global_ireg (ins->sreg2)) {
- MonoInst *copy = create_copy_ins (cfg, X86_ECX, ins->sreg2, NULL);
- insert_before_ins (ins, tmp, copy);
- }
- else {
- val = rs->iassign [ins->sreg2];
- if (val >= 0 && val != X86_ECX) {
- MonoInst *move = create_copy_ins (cfg, X86_ECX, val, NULL);
- DEBUG (g_print ("\tmoved arg from R%d (%d) to ECX\n", val, ins->sreg2));
- move->next = ins;
- g_assert_not_reached ();
- /* FIXME: where is move connected to the instruction list? */
- //tmp->prev->data->next = move;
+ for (ins2 = ins->next; ins2; ins2 = ins2->next) {
+ if ((ins2->opcode == OP_STORE_MEMBASE_IMM) && (ins2->inst_imm == 0)) {
+ ins2->opcode = OP_STORE_MEMBASE_REG;
+ ins2->sreg1 = ins->dreg;
}
- else {
- if (val == X86_ECX)
- need_ecx_spill = FALSE;
+ else if ((ins2->opcode == OP_STOREI4_MEMBASE_IMM) && (ins2->inst_imm == 0)) {
+ ins2->opcode = OP_STOREI4_MEMBASE_REG;
+ ins2->sreg1 = ins->dreg;
}
+ else if ((ins2->opcode == OP_STOREI1_MEMBASE_IMM) || (ins2->opcode == OP_STOREI2_MEMBASE_IMM)) {
+ /* Continue iteration */
+ }
+ else
+ break;
}
- if (need_ecx_spill && !(rs->ifree_mask & (1 << X86_ECX))) {
- DEBUG (g_print ("\tforced spill of R%d\n", rs->isymbolic [X86_ECX]));
- get_register_force_spilling (cfg, tmp, ins, rs->isymbolic [X86_ECX]);
- mono_regstate_free_int (rs, X86_ECX);
- }
- if (!is_global_ireg (ins->sreg2))
- /* force-set sreg2 */
- assign_ireg (rs, ins->sreg2, X86_ECX);
}
- ins->sreg2 = X86_ECX;
- } else if (spec [MONO_INST_CLOB] == 'd') {
- /*
- * DIVISION/REMAINER
+ break;
+ case OP_IADD_IMM:
+ case OP_ADD_IMM:
+ if ((ins->inst_imm == 1) && (ins->dreg == ins->sreg1))
+ ins->opcode = OP_X86_INC_REG;
+ break;
+ case OP_ISUB_IMM:
+ case OP_SUB_IMM:
+ if ((ins->inst_imm == 1) && (ins->dreg == ins->sreg1))
+ ins->opcode = OP_X86_DEC_REG;
+ break;
+ case OP_X86_COMPARE_MEMBASE_IMM:
+ /*
+ * OP_STORE_MEMBASE_REG reg, offset(basereg)
+ * OP_X86_COMPARE_MEMBASE_IMM offset(basereg), imm
+ * -->
+ * OP_STORE_MEMBASE_REG reg, offset(basereg)
+ * OP_COMPARE_IMM reg, imm
+ *
+ * Note: if imm = 0 then OP_COMPARE_IMM replaced with OP_X86_TEST_NULL
*/
- int dest_reg = X86_EAX;
- int clob_reg = X86_EDX;
- if (spec [MONO_INST_DEST] == 'd') {
- dest_reg = X86_EDX; /* reminder */
- clob_reg = X86_EAX;
- }
- if (is_global_ireg (ins->dreg))
- val = ins->dreg;
- else
- val = rs->iassign [ins->dreg];
- if (0 && val >= 0 && val != dest_reg && !(rs->ifree_mask & (1 << dest_reg))) {
- DEBUG (g_print ("\tforced spill of R%d\n", rs->isymbolic [dest_reg]));
- get_register_force_spilling (cfg, tmp, ins, rs->isymbolic [dest_reg]);
- mono_regstate_free_int (rs, dest_reg);
- }
- if (val < 0) {
- if (val < -1) {
- /* the register gets spilled after this inst */
- int spill = -val -1;
- dest_mask = 1 << dest_reg;
- prev_dreg = ins->dreg;
- val = mono_regstate_alloc_int (rs, dest_mask);
- if (val < 0)
- val = get_register_spilling (cfg, tmp, ins, dest_mask, ins->dreg);
- rs->iassign [ins->dreg] = val;
- if (spill)
- create_spilled_store (cfg, spill, val, prev_dreg, ins);
- DEBUG (g_print ("\tassigned dreg %s to dest R%d\n", mono_arch_regname (val), ins->dreg));
- rs->isymbolic [val] = prev_dreg;
- ins->dreg = val;
- } else {
- DEBUG (g_print ("\tshortcut assignment of R%d to %s\n", ins->dreg, mono_arch_regname (dest_reg)));
- prev_dreg = ins->dreg;
- assign_ireg (rs, ins->dreg, dest_reg);
- ins->dreg = dest_reg;
- val = dest_reg;
- }
- }
-
- //DEBUG (g_print ("dest reg in div assigned: %s\n", mono_arch_regname (val)));
- if (val != dest_reg) { /* force a copy */
- create_copy_ins (cfg, val, dest_reg, ins);
- if (!(rs->ifree_mask & (1 << dest_reg)) && rs->isymbolic [dest_reg] >= MONO_MAX_IREGS) {
- DEBUG (g_print ("\tforced spill of R%d\n", rs->isymbolic [dest_reg]));
- get_register_force_spilling (cfg, tmp, ins, rs->isymbolic [dest_reg]);
- mono_regstate_free_int (rs, dest_reg);
- }
- }
- if (!(rs->ifree_mask & (1 << clob_reg)) && (clob_reg != val) && (rs->isymbolic [clob_reg] >= 8)) {
- DEBUG (g_print ("\tforced spill of clobbered reg R%d\n", rs->isymbolic [clob_reg]));
- get_register_force_spilling (cfg, tmp, ins, rs->isymbolic [clob_reg]);
- mono_regstate_free_int (rs, clob_reg);
- }
- src1_mask = 1 << X86_EAX;
- src2_mask = 1 << X86_ECX;
- } else if (spec [MONO_INST_DEST] == 'l') {
- int hreg;
- val = rs->iassign [ins->dreg];
- /* check special case when dreg have been moved from ecx (clob shift) */
- if (spec [MONO_INST_CLOB] == 's' && clob_dreg != -1)
- hreg = clob_dreg + 1;
- else
- hreg = ins->dreg + 1;
-
- /* base prev_dreg on fixed hreg, handle clob case */
- val = hreg - 1;
-
- if (val != rs->isymbolic [X86_EAX] && !(rs->ifree_mask & (1 << X86_EAX))) {
- DEBUG (g_print ("\t(long-low) forced spill of R%d\n", rs->isymbolic [X86_EAX]));
- get_register_force_spilling (cfg, tmp, ins, rs->isymbolic [X86_EAX]);
- mono_regstate_free_int (rs, X86_EAX);
- }
- if (hreg != rs->isymbolic [X86_EDX] && !(rs->ifree_mask & (1 << X86_EDX))) {
- DEBUG (g_print ("\t(long-high) forced spill of R%d\n", rs->isymbolic [X86_EDX]));
- get_register_force_spilling (cfg, tmp, ins, rs->isymbolic [X86_EDX]);
- mono_regstate_free_int (rs, X86_EDX);
- }
- } else if (spec [MONO_INST_CLOB] == 'b') {
- /*
- * x86_set_reg instructions, dreg needs to be EAX..EDX
- */
- dest_mask = (1 << X86_EAX) | (1 << X86_EBX) | (1 << X86_ECX) | (1 << X86_EDX);
- if ((ins->dreg < MONO_MAX_IREGS) && (! (dest_mask & (1 << ins->dreg)))) {
- /*
- * ins->dreg is already a hard reg, need to allocate another
- * suitable hard reg and make a copy.
- */
- int new_dest = mono_x86_alloc_int_reg (cfg, tmp, ins, dest_mask, ins->dreg, reginfo [ins->dreg].flags);
- g_assert (new_dest >= 0);
-
- create_copy_ins (cfg, ins->dreg, new_dest, ins);
- DEBUG (g_print ("\tclob:b changing dreg R%d to %s\n", ins->dreg, mono_arch_regname (new_dest)));
- ins->dreg = new_dest;
-
- /* The hard reg is no longer needed */
- mono_regstate_free_int (rs, new_dest);
- }
- }
-
- /*
- * TRACK DREG
- */
- if (spec [MONO_INST_DEST] == 'f') {
- if (reginfof [ins->dreg].flags & MONO_X86_FP_NEEDS_SPILL) {
- GList *spill_node;
- MonoInst *store;
- spill_node = g_list_first (fspill_list);
- g_assert (spill_node);
-
- store = create_spilled_store_float (cfg, GPOINTER_TO_INT (spill_node->data), ins->dreg, ins);
- insert_before_ins (ins, tmp, store);
- fspill_list = g_list_remove (fspill_list, spill_node->data);
- fspill--;
- }
- } else if (spec [MONO_INST_DEST] == 'L') {
- int hreg;
- val = rs->iassign [ins->dreg];
- /* check special case when dreg have been moved from ecx (clob shift) */
- if (spec [MONO_INST_CLOB] == 's' && clob_dreg != -1)
- hreg = clob_dreg + 1;
- else
- hreg = ins->dreg + 1;
-
- /* base prev_dreg on fixed hreg, handle clob case */
- prev_dreg = hreg - 1;
+ if (last_ins && (last_ins->opcode == OP_STOREI4_MEMBASE_REG) &&
+ ins->inst_basereg == last_ins->inst_destbasereg &&
+ ins->inst_offset == last_ins->inst_offset) {
+ ins->opcode = OP_COMPARE_IMM;
+ ins->sreg1 = last_ins->sreg1;
- if (val < 0) {
- int spill = 0;
- if (val < -1) {
- /* the register gets spilled after this inst */
- spill = -val -1;
+ /* check if we can remove cmp reg,0 with test null */
+ if (!ins->inst_imm)
+ ins->opcode = OP_X86_TEST_NULL;
}
- val = mono_x86_alloc_int_reg (cfg, tmp, ins, dest_mask, ins->dreg, reginfo [ins->dreg].flags);
- rs->iassign [ins->dreg] = val;
- if (spill)
- create_spilled_store (cfg, spill, val, prev_dreg, ins);
- }
- DEBUG (g_print ("\tassigned dreg (long) %s to dest R%d\n", mono_arch_regname (val), hreg - 1));
-
- rs->isymbolic [val] = hreg - 1;
- ins->dreg = val;
-
- val = rs->iassign [hreg];
- if (val < 0) {
- int spill = 0;
- if (val < -1) {
- /* the register gets spilled after this inst */
- spill = -val -1;
+ break;
+ case OP_LOAD_MEMBASE:
+ case OP_LOADI4_MEMBASE:
+ /*
+ * Note: if reg1 = reg2 the load op is removed
+ *
+ * OP_STORE_MEMBASE_REG reg1, offset(basereg)
+ * OP_LOAD_MEMBASE offset(basereg), reg2
+ * -->
+ * OP_STORE_MEMBASE_REG reg1, offset(basereg)
+ * OP_MOVE reg1, reg2
+ */
+ if (last_ins && (last_ins->opcode == OP_STOREI4_MEMBASE_REG
+ || last_ins->opcode == OP_STORE_MEMBASE_REG) &&
+ ins->inst_basereg == last_ins->inst_destbasereg &&
+ ins->inst_offset == last_ins->inst_offset) {
+ if (ins->dreg == last_ins->sreg1) {
+ last_ins->next = ins->next;
+ ins = ins->next;
+ continue;
+ } else {
+ //static int c = 0; printf ("MATCHX %s %d\n", cfg->method->name,c++);
+ ins->opcode = OP_MOVE;
+ ins->sreg1 = last_ins->sreg1;
}
- val = mono_x86_alloc_int_reg (cfg, tmp, ins, dest_mask, hreg, reginfo [hreg].flags);
- rs->iassign [hreg] = val;
- if (spill)
- create_spilled_store (cfg, spill, val, hreg, ins);
- }
- DEBUG (g_print ("\tassigned hreg (long-high) %s to dest R%d\n", mono_arch_regname (val), hreg));
- rs->isymbolic [val] = hreg;
- /* save reg allocating into unused */
- ins->unused = val;
+ /*
+ * Note: reg1 must be different from the basereg in the second load
+ * Note: if reg1 = reg2 is equal then second load is removed
+ *
+ * OP_LOAD_MEMBASE offset(basereg), reg1
+ * OP_LOAD_MEMBASE offset(basereg), reg2
+ * -->
+ * OP_LOAD_MEMBASE offset(basereg), reg1
+ * OP_MOVE reg1, reg2
+ */
+ } if (last_ins && (last_ins->opcode == OP_LOADI4_MEMBASE
+ || last_ins->opcode == OP_LOAD_MEMBASE) &&
+ ins->inst_basereg != last_ins->dreg &&
+ ins->inst_basereg == last_ins->inst_basereg &&
+ ins->inst_offset == last_ins->inst_offset) {
- /* check if we can free our long reg */
- if (reg_is_freeable (val) && hreg >= 0 && reginfo [hreg].born_in >= i) {
- DEBUG (g_print ("\tfreeable %s (R%d) (born in %d)\n", mono_arch_regname (val), hreg, reginfo [hreg].born_in));
- mono_regstate_free_int (rs, val);
- }
- }
- else if (ins->dreg >= MONO_MAX_IREGS) {
- int hreg;
- val = rs->iassign [ins->dreg];
- if (spec [MONO_INST_DEST] == 'l') {
- /* check special case when dreg have been moved from ecx (clob shift) */
- if (spec [MONO_INST_CLOB] == 's' && clob_dreg != -1)
- hreg = clob_dreg + 1;
- else
- hreg = ins->dreg + 1;
-
- /* base prev_dreg on fixed hreg, handle clob case */
- prev_dreg = hreg - 1;
- } else
- prev_dreg = ins->dreg;
-
- if (val < 0) {
- int spill = 0;
- if (val < -1) {
- /* the register gets spilled after this inst */
- spill = -val -1;
- }
- val = mono_x86_alloc_int_reg (cfg, tmp, ins, dest_mask, ins->dreg, reginfo [ins->dreg].flags);
- rs->iassign [ins->dreg] = val;
- if (spill)
- create_spilled_store (cfg, spill, val, prev_dreg, ins);
- }
- DEBUG (g_print ("\tassigned dreg %s to dest R%d\n", mono_arch_regname (val), ins->dreg));
- rs->isymbolic [val] = prev_dreg;
- ins->dreg = val;
- /* handle cases where lreg needs to be eax:edx */
- if (spec [MONO_INST_DEST] == 'l') {
- /* check special case when dreg have been moved from ecx (clob shift) */
- int hreg = prev_dreg + 1;
- val = rs->iassign [hreg];
- if (val < 0) {
- int spill = 0;
- if (val < -1) {
- /* the register gets spilled after this inst */
- spill = -val -1;
- }
- val = mono_x86_alloc_int_reg (cfg, tmp, ins, dest_mask, hreg, reginfo [hreg].flags);
- rs->iassign [hreg] = val;
- if (spill)
- create_spilled_store (cfg, spill, val, hreg, ins);
- }
- DEBUG (g_print ("\tassigned hreg %s to dest R%d\n", mono_arch_regname (val), hreg));
- rs->isymbolic [val] = hreg;
- if (ins->dreg == X86_EAX) {
- if (val != X86_EDX)
- create_copy_ins (cfg, val, X86_EDX, ins);
- } else if (ins->dreg == X86_EDX) {
- if (val == X86_EAX) {
- /* swap */
- g_assert_not_reached ();
- } else {
- /* two forced copies */
- create_copy_ins (cfg, val, X86_EDX, ins);
- create_copy_ins (cfg, ins->dreg, X86_EAX, ins);
- }
+ if (ins->dreg == last_ins->dreg) {
+ last_ins->next = ins->next;
+ ins = ins->next;
+ continue;
} else {
- if (val == X86_EDX) {
- create_copy_ins (cfg, ins->dreg, X86_EAX, ins);
- } else {
- /* two forced copies */
- create_copy_ins (cfg, val, X86_EDX, ins);
- create_copy_ins (cfg, ins->dreg, X86_EAX, ins);
- }
- }
- if (reg_is_freeable (val) && hreg >= 0 && reginfo [hreg].born_in >= i) {
- DEBUG (g_print ("\tfreeable %s (R%d)\n", mono_arch_regname (val), hreg));
- mono_regstate_free_int (rs, val);
+ ins->opcode = OP_MOVE;
+ ins->sreg1 = last_ins->dreg;
}
- } else if (spec [MONO_INST_DEST] == 'a' && ins->dreg != X86_EAX && spec [MONO_INST_CLOB] != 'd') {
- /* this instruction only outputs to EAX, need to copy */
- create_copy_ins (cfg, ins->dreg, X86_EAX, ins);
- } else if (spec [MONO_INST_DEST] == 'd' && ins->dreg != X86_EDX && spec [MONO_INST_CLOB] != 'd') {
- create_copy_ins (cfg, ins->dreg, X86_EDX, ins);
- }
- }
- if (spec [MONO_INST_DEST] != 'f' && reg_is_freeable (ins->dreg) && prev_dreg >= 0 && reginfo [prev_dreg].born_in >= i) {
- DEBUG (g_print ("\tfreeable %s (R%d) (born in %d)\n", mono_arch_regname (ins->dreg), prev_dreg, reginfo [prev_dreg].born_in));
- mono_regstate_free_int (rs, ins->dreg);
- }
- /* put src1 in EAX if it needs to be */
- if (spec [MONO_INST_SRC1] == 'a') {
- if (!(rs->ifree_mask & (1 << X86_EAX))) {
- DEBUG (g_print ("\tforced spill of R%d\n", rs->isymbolic [X86_EAX]));
- get_register_force_spilling (cfg, tmp, ins, rs->isymbolic [X86_EAX]);
- mono_regstate_free_int (rs, X86_EAX);
- }
- if (ins->sreg1 < MONO_MAX_IREGS) {
- /* The argument is already in a hard reg, need to copy */
- MonoInst *copy = create_copy_ins (cfg, X86_EAX, ins->sreg1, NULL);
- insert_before_ins (ins, tmp, copy);
- }
- else
- /* force-set sreg1 */
- assign_ireg (rs, ins->sreg1, X86_EAX);
- ins->sreg1 = X86_EAX;
- }
- /*
- * TRACK SREG1
- */
- if (spec [MONO_INST_SRC1] == 'f') {
- if (reginfof [ins->sreg1].flags & MONO_X86_FP_NEEDS_LOAD) {
- MonoInst *load;
- MonoInst *store = NULL;
-
- if (reginfof [ins->sreg1].flags & MONO_X86_FP_NEEDS_LOAD_SPILL) {
- GList *spill_node;
- spill_node = g_list_first (fspill_list);
- g_assert (spill_node);
-
- store = create_spilled_store_float (cfg, GPOINTER_TO_INT (spill_node->data), ins->sreg1, ins);
- fspill_list = g_list_remove (fspill_list, spill_node->data);
- }
+ //g_assert_not_reached ();
- fspill++;
- fspill_list = g_list_prepend (fspill_list, GINT_TO_POINTER(fspill));
- load = create_spilled_load_float (cfg, fspill, ins->sreg1, ins);
- insert_before_ins (ins, tmp, load);
- if (store)
- insert_before_ins (load, tmp, store);
+#if 0
+ /*
+ * OP_STORE_MEMBASE_IMM imm, offset(basereg)
+ * OP_LOAD_MEMBASE offset(basereg), reg
+ * -->
+ * OP_STORE_MEMBASE_IMM imm, offset(basereg)
+ * OP_ICONST reg, imm
+ */
+ } else if (last_ins && (last_ins->opcode == OP_STOREI4_MEMBASE_IMM
+ || last_ins->opcode == OP_STORE_MEMBASE_IMM) &&
+ ins->inst_basereg == last_ins->inst_destbasereg &&
+ ins->inst_offset == last_ins->inst_offset) {
+ //static int c = 0; printf ("MATCHX %s %d\n", cfg->method->name,c++);
+ ins->opcode = OP_ICONST;
+ ins->inst_c0 = last_ins->inst_imm;
+ g_assert_not_reached (); // check this rule
+#endif
}
- } else if ((spec [MONO_INST_DEST] == 'L') && (spec [MONO_INST_SRC1] == 'L')) {
- /* force source to be same as dest */
- assign_ireg (rs, ins->sreg1, ins->dreg);
- assign_ireg (rs, ins->sreg1 + 1, ins->unused);
-
- DEBUG (g_print ("\tassigned sreg1 (long) %s to sreg1 R%d\n", mono_arch_regname (ins->dreg), ins->sreg1));
- DEBUG (g_print ("\tassigned sreg1 (long-high) %s to sreg1 R%d\n", mono_arch_regname (ins->unused), ins->sreg1 + 1));
-
- ins->sreg1 = ins->dreg;
+ break;
+ case OP_LOADU1_MEMBASE:
+ case OP_LOADI1_MEMBASE:
/*
- * No need for saving the reg, we know that src1=dest in this cases
- * ins->inst_c0 = ins->unused;
+ * OP_STORE_MEMBASE_REG reg1, offset(basereg)
+ * OP_LOAD_MEMBASE offset(basereg), reg2
+ * -->
+ * OP_STORE_MEMBASE_REG reg1, offset(basereg)
+ * CONV_I2/U2 reg1, reg2
*/
- }
- else if (ins->sreg1 >= MONO_MAX_IREGS) {
- val = rs->iassign [ins->sreg1];
- prev_sreg1 = ins->sreg1;
- if (val < 0) {
- int spill = 0;
- if (val < -1) {
- /* the register gets spilled after this inst */
- spill = -val -1;
- }
- if (0 && ins->opcode == OP_MOVE) {
- /*
- * small optimization: the dest register is already allocated
- * but the src one is not: we can simply assign the same register
- * here and peephole will get rid of the instruction later.
- * This optimization may interfere with the clobbering handling:
- * it removes a mov operation that will be added again to handle clobbering.
- * There are also some other issues that should with make testjit.
- */
- mono_regstate_alloc_int (rs, 1 << ins->dreg);
- val = rs->iassign [ins->sreg1] = ins->dreg;
- //g_assert (val >= 0);
- DEBUG (g_print ("\tfast assigned sreg1 %s to R%d\n", mono_arch_regname (val), ins->sreg1));
- } else {
- //g_assert (val == -1); /* source cannot be spilled */
- val = mono_x86_alloc_int_reg (cfg, tmp, ins, src1_mask, ins->sreg1, reginfo [ins->sreg1].flags);
- rs->iassign [ins->sreg1] = val;
- DEBUG (g_print ("\tassigned sreg1 %s to R%d\n", mono_arch_regname (val), ins->sreg1));
- }
- if (spill) {
- MonoInst *store = create_spilled_store (cfg, spill, val, prev_sreg1, NULL);
- insert_before_ins (ins, tmp, store);
- }
+ if (last_ins && X86_IS_BYTE_REG (last_ins->sreg1) &&
+ (last_ins->opcode == OP_STOREI1_MEMBASE_REG) &&
+ ins->inst_basereg == last_ins->inst_destbasereg &&
+ ins->inst_offset == last_ins->inst_offset) {
+ ins->opcode = (ins->opcode == OP_LOADI1_MEMBASE) ? CEE_CONV_I1 : CEE_CONV_U1;
+ ins->sreg1 = last_ins->sreg1;
}
- rs->isymbolic [val] = prev_sreg1;
- ins->sreg1 = val;
- } else {
- prev_sreg1 = -1;
- }
- /* handle clobbering of sreg1 */
- if ((spec [MONO_INST_CLOB] == '1' || spec [MONO_INST_CLOB] == 's') && ins->dreg != ins->sreg1) {
- MonoInst *sreg2_copy = NULL;
- MonoInst *copy = NULL;
-
- if (ins->dreg == ins->sreg2) {
- /*
- * copying sreg1 to dreg could clobber sreg2, so allocate a new
- * register for it.
- */
- int reg2 = 0;
-
- reg2 = mono_x86_alloc_int_reg (cfg, tmp, ins, dest_mask, ins->sreg2, 0);
-
- DEBUG (g_print ("\tneed to copy sreg2 %s to reg %s\n", mono_arch_regname (ins->sreg2), mono_arch_regname (reg2)));
- sreg2_copy = create_copy_ins (cfg, reg2, ins->sreg2, NULL);
- prev_sreg2 = ins->sreg2 = reg2;
-
- mono_regstate_free_int (rs, reg2);
+ break;
+ case OP_LOADU2_MEMBASE:
+ case OP_LOADI2_MEMBASE:
+ /*
+ * OP_STORE_MEMBASE_REG reg1, offset(basereg)
+ * OP_LOAD_MEMBASE offset(basereg), reg2
+ * -->
+ * OP_STORE_MEMBASE_REG reg1, offset(basereg)
+ * CONV_I2/U2 reg1, reg2
+ */
+ if (last_ins && (last_ins->opcode == OP_STOREI2_MEMBASE_REG) &&
+ ins->inst_basereg == last_ins->inst_destbasereg &&
+ ins->inst_offset == last_ins->inst_offset) {
+ ins->opcode = (ins->opcode == OP_LOADI2_MEMBASE) ? CEE_CONV_I2 : CEE_CONV_U2;
+ ins->sreg1 = last_ins->sreg1;
}
-
- copy = create_copy_ins (cfg, ins->dreg, ins->sreg1, NULL);
- DEBUG (g_print ("\tneed to copy sreg1 %s to dreg %s\n", mono_arch_regname (ins->sreg1), mono_arch_regname (ins->dreg)));
- insert_before_ins (ins, tmp, copy);
-
- if (sreg2_copy)
- insert_before_ins (copy, tmp, sreg2_copy);
-
+ break;
+ case CEE_CONV_I4:
+ case CEE_CONV_U4:
+ case OP_ICONV_TO_I4:
+ case OP_MOVE:
/*
- * Need to prevent sreg2 to be allocated to sreg1, since that
- * would screw up the previous copy.
+ * Removes:
+ *
+ * OP_MOVE reg, reg
*/
- src2_mask &= ~ (1 << ins->sreg1);
- /* we set sreg1 to dest as well */
- prev_sreg1 = ins->sreg1 = ins->dreg;
- src2_mask &= ~ (1 << ins->dreg);
- }
-
- /*
- * TRACK SREG2
- */
- if (spec [MONO_INST_SRC2] == 'f') {
- if (reginfof [ins->sreg2].flags & MONO_X86_FP_NEEDS_LOAD) {
- MonoInst *load;
- MonoInst *store = NULL;
-
- if (reginfof [ins->sreg2].flags & MONO_X86_FP_NEEDS_LOAD_SPILL) {
- GList *spill_node;
-
- spill_node = g_list_first (fspill_list);
- g_assert (spill_node);
- if (spec [MONO_INST_SRC1] == 'f' && (reginfof [ins->sreg1].flags & MONO_X86_FP_NEEDS_LOAD_SPILL))
- spill_node = g_list_next (spill_node);
-
- store = create_spilled_store_float (cfg, GPOINTER_TO_INT (spill_node->data), ins->sreg2, ins);
- fspill_list = g_list_remove (fspill_list, spill_node->data);
- }
-
- fspill++;
- fspill_list = g_list_prepend (fspill_list, GINT_TO_POINTER(fspill));
- load = create_spilled_load_float (cfg, fspill, ins->sreg2, ins);
- insert_before_ins (ins, tmp, load);
- if (store)
- insert_before_ins (load, tmp, store);
- }
- }
- else if (ins->sreg2 >= MONO_MAX_IREGS) {
- val = rs->iassign [ins->sreg2];
- prev_sreg2 = ins->sreg2;
- if (val < 0) {
- int spill = 0;
- if (val < -1) {
- /* the register gets spilled after this inst */
- spill = -val -1;
- }
- val = mono_x86_alloc_int_reg (cfg, tmp, ins, src2_mask, ins->sreg2, reginfo [ins->sreg2].flags);
- rs->iassign [ins->sreg2] = val;
- DEBUG (g_print ("\tassigned sreg2 %s to R%d\n", mono_arch_regname (val), ins->sreg2));
- if (spill)
- create_spilled_store (cfg, spill, val, prev_sreg2, ins);
- }
- rs->isymbolic [val] = prev_sreg2;
- ins->sreg2 = val;
- if (spec [MONO_INST_CLOB] == 's' && ins->sreg2 != X86_ECX) {
- DEBUG (g_print ("\tassigned sreg2 %s to R%d, but ECX is needed (R%d)\n", mono_arch_regname (val), ins->sreg2, rs->iassign [X86_ECX]));
+ if (ins->dreg == ins->sreg1) {
+ if (last_ins)
+ last_ins->next = ins->next;
+ ins = ins->next;
+ continue;
}
- } else {
- prev_sreg2 = -1;
- }
-
- if (spec [MONO_INST_CLOB] == 'c') {
- int j, s;
- guint32 clob_mask = X86_CALLEE_REGS;
- for (j = 0; j < MONO_MAX_IREGS; ++j) {
- s = 1 << j;
- if ((clob_mask & s) && !(rs->ifree_mask & s) && j != ins->sreg1) {
- //g_warning ("register %s busy at call site\n", mono_arch_regname (j));
- }
+ /*
+ * Removes:
+ *
+ * OP_MOVE sreg, dreg
+ * OP_MOVE dreg, sreg
+ */
+ if (last_ins && last_ins->opcode == OP_MOVE &&
+ ins->sreg1 == last_ins->dreg &&
+ ins->dreg == last_ins->sreg1) {
+ last_ins->next = ins->next;
+ ins = ins->next;
+ continue;
}
- }
- if (spec [MONO_INST_CLOB] == 'a') {
- guint32 clob_reg = X86_EAX;
- if (!(rs->ifree_mask & (1 << clob_reg)) && (rs->isymbolic [clob_reg] >= 8)) {
- DEBUG (g_print ("\tforced spill of clobbered reg R%d\n", rs->isymbolic [clob_reg]));
- get_register_force_spilling (cfg, tmp, ins, rs->isymbolic [clob_reg]);
- mono_regstate_free_int (rs, clob_reg);
+ break;
+ case OP_X86_PUSH_MEMBASE:
+ if (last_ins && (last_ins->opcode == OP_STOREI4_MEMBASE_REG ||
+ last_ins->opcode == OP_STORE_MEMBASE_REG) &&
+ ins->inst_basereg == last_ins->inst_destbasereg &&
+ ins->inst_offset == last_ins->inst_offset) {
+ ins->opcode = OP_X86_PUSH;
+ ins->sreg1 = last_ins->sreg1;
}
+ break;
}
- /*if (reg_is_freeable (ins->sreg1) && prev_sreg1 >= 0 && reginfo [prev_sreg1].born_in >= i) {
- DEBUG (g_print ("freeable %s\n", mono_arch_regname (ins->sreg1)));
- mono_regstate_free_int (rs, ins->sreg1);
- }
- if (reg_is_freeable (ins->sreg2) && prev_sreg2 >= 0 && reginfo [prev_sreg2].born_in >= i) {
- DEBUG (g_print ("freeable %s\n", mono_arch_regname (ins->sreg2)));
- mono_regstate_free_int (rs, ins->sreg2);
- }*/
-
- //DEBUG (print_ins (i, ins));
- /* this may result from a insert_before call */
- if (!tmp->next)
- bb->code = tmp->data;
- tmp = tmp->next;
+ last_ins = ins;
+ ins = ins->next;
}
+ bb->last_ins = last_ins;
+}
+
+static const int
+branch_cc_table [] = {
+ X86_CC_EQ, X86_CC_GE, X86_CC_GT, X86_CC_LE, X86_CC_LT,
+ X86_CC_NE, X86_CC_GE, X86_CC_GT, X86_CC_LE, X86_CC_LT,
+ X86_CC_O, X86_CC_NO, X86_CC_C, X86_CC_NC
+};
+
+/* Maps CMP_... constants to X86_CC_... constants */
+static const int
+cc_table [] = {
+ X86_CC_EQ, X86_CC_NE, X86_CC_LE, X86_CC_GE, X86_CC_LT, X86_CC_GT,
+ X86_CC_LE, X86_CC_GE, X86_CC_LT, X86_CC_GT
+};
+
+static const int
+cc_signed_table [] = {
+ TRUE, TRUE, TRUE, TRUE, TRUE, TRUE,
+ FALSE, FALSE, FALSE, FALSE
+};
+
+void
+mono_arch_local_regalloc (MonoCompile *cfg, MonoBasicBlock *bb)
+{
+ if (cfg->opt & MONO_OPT_PEEPHOLE)
+ peephole_pass_1 (cfg, bb);
- g_free (reginfo);
- g_free (reginfof);
- g_list_free (fspill_list);
+ mono_local_regalloc (cfg, bb);
}
static unsigned char*
case OP_VCALL:
case OP_VCALL_REG:
case OP_VCALL_MEMBASE:
- cinfo = get_call_info (((MonoCallInst*)ins)->signature, FALSE);
+ cinfo = get_call_info (cfg->mempool, ((MonoCallInst*)ins)->signature, FALSE);
if (cinfo->ret.storage == ArgValuetypeInReg) {
/* Pop the destination address from the stack */
x86_pop_reg (code, X86_ECX);
}
}
}
- g_free (cinfo);
default:
break;
}
return code;
}
+/*
+ * emit_tls_get:
+ * @code: buffer to store code to
+ * @dreg: hard register where to place the result
+ * @tls_offset: offset info
+ *
+ * emit_tls_get emits in @code the native code that puts in the dreg register
+ * the item in the thread local storage identified by tls_offset.
+ *
+ * Returns: a pointer to the end of the stored code
+ */
static guint8*
emit_tls_get (guint8* code, int dreg, int tls_offset)
{
x86_alu_membase_imm (code, X86_AND, dreg, 0x34, 0);
x86_mov_reg_membase (code, dreg, dreg, 3600 + (tls_offset * 4), 4);
#else
- x86_prefix (code, X86_GS_PREFIX);
- x86_mov_reg_mem (code, dreg, tls_offset, 4);
+ if (optimize_for_xen) {
+ x86_prefix (code, X86_GS_PREFIX);
+ x86_mov_reg_mem (code, dreg, 0, 4);
+ x86_mov_reg_membase (code, dreg, dreg, tls_offset, 4);
+ } else {
+ x86_prefix (code, X86_GS_PREFIX);
+ x86_mov_reg_mem (code, dreg, tls_offset, 4);
+ }
#endif
return code;
}
+/*
+ * emit_load_volatile_arguments:
+ *
+ * Load volatile arguments from the stack to the original input registers.
+ * Required before a tail call.
+ */
+static guint8*
+emit_load_volatile_arguments (MonoCompile *cfg, guint8 *code)
+{
+ MonoMethod *method = cfg->method;
+ MonoMethodSignature *sig;
+ MonoInst *inst;
+ CallInfo *cinfo;
+ guint32 i;
+
+ /* FIXME: Generate intermediate code instead */
+
+ sig = mono_method_signature (method);
+
+ cinfo = get_call_info (cfg->mempool, sig, FALSE);
+
+ /* This is the opposite of the code in emit_prolog */
+
+ for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
+ ArgInfo *ainfo = cinfo->args + i;
+ MonoType *arg_type;
+ inst = cfg->args [i];
+
+ if (sig->hasthis && (i == 0))
+ arg_type = &mono_defaults.object_class->byval_arg;
+ else
+ arg_type = sig->params [i - sig->hasthis];
+
+ /*
+ * On x86, the arguments are either in their original stack locations, or in
+ * global regs.
+ */
+ if (inst->opcode == OP_REGVAR) {
+ g_assert (ainfo->storage == ArgOnStack);
+
+ x86_mov_membase_reg (code, X86_EBP, inst->inst_offset, inst->dreg, 4);
+ }
+ }
+
+ return code;
+}
+
#define REAL_PRINT_REG(text,reg) \
mono_assert (reg >= 0); \
x86_push_reg (code, X86_EAX); \
offset = code - cfg->native_code;
+ mono_debug_open_block (cfg, bb, offset);
+
ins = bb->code;
while (ins) {
offset = code - cfg->native_code;
- max_len = ((guint8 *)ins_spec [ins->opcode])[MONO_INST_LEN];
+ max_len = ((guint8 *)ins_get_spec (ins->opcode))[MONO_INST_LEN];
if (offset > (cfg->code_size - max_len - 16)) {
cfg->code_size *= 2;
case OP_X86_SUB_MEMBASE:
x86_alu_reg_membase (code, X86_SUB, ins->sreg1, ins->sreg2, ins->inst_offset);
break;
+ case OP_X86_AND_MEMBASE_IMM:
+ x86_alu_membase_imm (code, X86_AND, ins->inst_basereg, ins->inst_offset, ins->inst_imm);
+ break;
+ case OP_X86_OR_MEMBASE_IMM:
+ x86_alu_membase_imm (code, X86_OR, ins->inst_basereg, ins->inst_offset, ins->inst_imm);
+ break;
+ case OP_X86_XOR_MEMBASE_IMM:
+ x86_alu_membase_imm (code, X86_XOR, ins->inst_basereg, ins->inst_offset, ins->inst_imm);
+ break;
case OP_X86_INC_MEMBASE:
x86_inc_membase (code, ins->inst_basereg, ins->inst_offset);
break;
case OP_X86_MUL_MEMBASE:
x86_imul_reg_membase (code, ins->sreg1, ins->sreg2, ins->inst_offset);
break;
- case CEE_BREAK:
+ case OP_BREAK:
x86_breakpoint (code);
break;
case OP_ADDCC:
x86_alu_reg_imm (code, X86_OR, ins->sreg1, ins->inst_imm);
break;
case CEE_XOR:
+ case OP_IXOR:
x86_alu_reg_reg (code, X86_XOR, ins->sreg1, ins->sreg2);
break;
case OP_XOR_IMM:
guint8 *jump_to_end;
/* handle shifts below 32 bits */
- x86_shld_reg (code, ins->unused, ins->sreg1);
+ x86_shld_reg (code, ins->backend.reg3, ins->sreg1);
x86_shift_reg (code, X86_SHL, ins->sreg1);
x86_test_reg_imm (code, X86_ECX, 32);
jump_to_end = code; x86_branch8 (code, X86_CC_EQ, 0, TRUE);
/* handle shift over 32 bit */
- x86_mov_reg_reg (code, ins->unused, ins->sreg1, 4);
+ x86_mov_reg_reg (code, ins->backend.reg3, ins->sreg1, 4);
x86_clear_reg (code, ins->sreg1);
x86_patch (jump_to_end, code);
guint8 *jump_to_end;
/* handle shifts below 32 bits */
- x86_shrd_reg (code, ins->sreg1, ins->unused);
- x86_shift_reg (code, X86_SAR, ins->unused);
+ x86_shrd_reg (code, ins->sreg1, ins->backend.reg3);
+ x86_shift_reg (code, X86_SAR, ins->backend.reg3);
x86_test_reg_imm (code, X86_ECX, 32);
jump_to_end = code; x86_branch8 (code, X86_CC_EQ, 0, FALSE);
/* handle shifts over 31 bits */
- x86_mov_reg_reg (code, ins->sreg1, ins->unused, 4);
- x86_shift_reg_imm (code, X86_SAR, ins->unused, 31);
+ x86_mov_reg_reg (code, ins->sreg1, ins->backend.reg3, 4);
+ x86_shift_reg_imm (code, X86_SAR, ins->backend.reg3, 31);
x86_patch (jump_to_end, code);
}
guint8 *jump_to_end;
/* handle shifts below 32 bits */
- x86_shrd_reg (code, ins->sreg1, ins->unused);
- x86_shift_reg (code, X86_SHR, ins->unused);
+ x86_shrd_reg (code, ins->sreg1, ins->backend.reg3);
+ x86_shift_reg (code, X86_SHR, ins->backend.reg3);
x86_test_reg_imm (code, X86_ECX, 32);
jump_to_end = code; x86_branch8 (code, X86_CC_EQ, 0, FALSE);
/* handle shifts over 31 bits */
- x86_mov_reg_reg (code, ins->sreg1, ins->unused, 4);
- x86_shift_reg_imm (code, X86_SHR, ins->unused, 31);
+ x86_mov_reg_reg (code, ins->sreg1, ins->backend.reg3, 4);
+ x86_clear_reg (code, ins->backend.reg3);
x86_patch (jump_to_end, code);
}
break;
case OP_LSHL_IMM:
if (ins->inst_imm >= 32) {
- x86_mov_reg_reg (code, ins->unused, ins->sreg1, 4);
+ x86_mov_reg_reg (code, ins->backend.reg3, ins->sreg1, 4);
x86_clear_reg (code, ins->sreg1);
- x86_shift_reg_imm (code, X86_SHL, ins->unused, ins->inst_imm - 32);
+ x86_shift_reg_imm (code, X86_SHL, ins->backend.reg3, ins->inst_imm - 32);
} else {
- x86_shld_reg_imm (code, ins->unused, ins->sreg1, ins->inst_imm);
+ x86_shld_reg_imm (code, ins->backend.reg3, ins->sreg1, ins->inst_imm);
x86_shift_reg_imm (code, X86_SHL, ins->sreg1, ins->inst_imm);
}
break;
case OP_LSHR_IMM:
if (ins->inst_imm >= 32) {
- x86_mov_reg_reg (code, ins->sreg1, ins->unused, 4);
- x86_shift_reg_imm (code, X86_SAR, ins->unused, 0x1f);
+ x86_mov_reg_reg (code, ins->sreg1, ins->backend.reg3, 4);
+ x86_shift_reg_imm (code, X86_SAR, ins->backend.reg3, 0x1f);
x86_shift_reg_imm (code, X86_SAR, ins->sreg1, ins->inst_imm - 32);
} else {
- x86_shrd_reg_imm (code, ins->sreg1, ins->unused, ins->inst_imm);
- x86_shift_reg_imm (code, X86_SAR, ins->unused, ins->inst_imm);
+ x86_shrd_reg_imm (code, ins->sreg1, ins->backend.reg3, ins->inst_imm);
+ x86_shift_reg_imm (code, X86_SAR, ins->backend.reg3, ins->inst_imm);
}
break;
case OP_LSHR_UN_IMM:
if (ins->inst_imm >= 32) {
- x86_mov_reg_reg (code, ins->sreg1, ins->unused, 4);
- x86_clear_reg (code, ins->unused);
+ x86_mov_reg_reg (code, ins->sreg1, ins->backend.reg3, 4);
+ x86_clear_reg (code, ins->backend.reg3);
x86_shift_reg_imm (code, X86_SHR, ins->sreg1, ins->inst_imm - 32);
} else {
- x86_shrd_reg_imm (code, ins->sreg1, ins->unused, ins->inst_imm);
- x86_shift_reg_imm (code, X86_SHR, ins->unused, ins->inst_imm);
+ x86_shrd_reg_imm (code, ins->sreg1, ins->backend.reg3, ins->inst_imm);
+ x86_shift_reg_imm (code, X86_SHR, ins->backend.reg3, ins->inst_imm);
}
break;
case CEE_NOT:
x86_imul_reg_reg (code, ins->sreg1, ins->sreg2);
break;
case OP_MUL_IMM:
- x86_imul_reg_reg_imm (code, ins->dreg, ins->sreg1, ins->inst_imm);
+ switch (ins->inst_imm) {
+ case 2:
+ /* MOV r1, r2 */
+ /* ADD r1, r1 */
+ if (ins->dreg != ins->sreg1)
+ x86_mov_reg_reg (code, ins->dreg, ins->sreg1, 4);
+ x86_alu_reg_reg (code, X86_ADD, ins->dreg, ins->dreg);
+ break;
+ case 3:
+ /* LEA r1, [r2 + r2*2] */
+ x86_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 1);
+ break;
+ case 5:
+ /* LEA r1, [r2 + r2*4] */
+ x86_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 2);
+ break;
+ case 6:
+ /* LEA r1, [r2 + r2*2] */
+ /* ADD r1, r1 */
+ x86_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 1);
+ x86_alu_reg_reg (code, X86_ADD, ins->dreg, ins->dreg);
+ break;
+ case 9:
+ /* LEA r1, [r2 + r2*8] */
+ x86_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 3);
+ break;
+ case 10:
+ /* LEA r1, [r2 + r2*4] */
+ /* ADD r1, r1 */
+ x86_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 2);
+ x86_alu_reg_reg (code, X86_ADD, ins->dreg, ins->dreg);
+ break;
+ case 12:
+ /* LEA r1, [r2 + r2*2] */
+ /* SHL r1, 2 */
+ x86_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 1);
+ x86_shift_reg_imm (code, X86_SHL, ins->dreg, 2);
+ break;
+ case 25:
+ /* LEA r1, [r2 + r2*4] */
+ /* LEA r1, [r1 + r1*4] */
+ x86_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 2);
+ x86_lea_memindex (code, ins->dreg, ins->dreg, 0, ins->dreg, 2);
+ break;
+ case 100:
+ /* LEA r1, [r2 + r2*4] */
+ /* SHL r1, 2 */
+ /* LEA r1, [r1 + r1*4] */
+ x86_lea_memindex (code, ins->dreg, ins->sreg1, 0, ins->sreg1, 2);
+ x86_shift_reg_imm (code, X86_SHL, ins->dreg, 2);
+ x86_lea_memindex (code, ins->dreg, ins->dreg, 0, ins->dreg, 2);
+ break;
+ default:
+ x86_imul_reg_reg_imm (code, ins->dreg, ins->sreg1, ins->inst_imm);
+ break;
+ }
break;
case CEE_MUL_OVF:
x86_imul_reg_reg (code, ins->sreg1, ins->sreg2);
break;
case CEE_CONV_U4:
g_assert_not_reached ();
- case CEE_JMP: {
+ case OP_JMP: {
/*
* Note: this 'frame destruction' logic is useful for tail calls, too.
* Keep in sync with the code in emit_epilog.
g_assert (!cfg->method->save_lmf);
+ code = emit_load_volatile_arguments (cfg, code);
+
if (cfg->used_int_regs & (1 << X86_EBX))
pos -= 4;
if (cfg->used_int_regs & (1 << X86_EDI))
x86_pop_reg (code, X86_EDI);
break;
case OP_X86_LEA:
- x86_lea_memindex (code, ins->dreg, ins->sreg1, ins->inst_imm, ins->sreg2, ins->unused);
+ x86_lea_memindex (code, ins->dreg, ins->sreg1, ins->inst_imm, ins->sreg2, ins->backend.shift_amount);
break;
case OP_X86_LEA_MEMBASE:
x86_lea_membase (code, ins->dreg, ins->sreg1, ins->inst_imm);
break;
case OP_LOCALLOC:
/* keep alignment */
- x86_alu_reg_imm (code, X86_ADD, ins->sreg1, MONO_ARCH_FRAME_ALIGNMENT - 1);
- x86_alu_reg_imm (code, X86_AND, ins->sreg1, ~(MONO_ARCH_FRAME_ALIGNMENT - 1));
+ x86_alu_reg_imm (code, X86_ADD, ins->sreg1, MONO_ARCH_LOCALLOC_ALIGNMENT - 1);
+ x86_alu_reg_imm (code, X86_AND, ins->sreg1, ~(MONO_ARCH_LOCALLOC_ALIGNMENT - 1));
code = mono_emit_stack_alloc (code, ins);
x86_mov_reg_reg (code, ins->dreg, X86_ESP, 4);
break;
case CEE_RET:
x86_ret (code);
break;
- case CEE_THROW: {
+ case OP_THROW: {
x86_push_reg (code, ins->sreg1);
code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD,
(gpointer)"mono_arch_throw_exception");
break;
}
case OP_CALL_HANDLER:
+ /* Align stack */
+#ifdef __APPLE__
+ x86_alu_reg_imm (code, X86_SUB, X86_ESP, 12);
+#endif
mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_BB, ins->inst_target_bb);
x86_call_imm (code, 0);
+#ifdef __APPLE__
+ x86_alu_reg_imm (code, X86_ADD, X86_ESP, 12);
+#endif
break;
case OP_LABEL:
ins->inst_c0 = code - cfg->native_code;
break;
- case CEE_BR:
+ case OP_BR:
//g_print ("target: %p, next: %p, curr: %p, last: %p\n", ins->inst_target_bb, bb->next_bb, ins, bb->last_ins);
//if ((ins->inst_target_bb == bb->next_bb) && ins == bb->last_ins)
//break;
x86_jump_reg (code, ins->sreg1);
break;
case OP_CEQ:
- x86_set_reg (code, X86_CC_EQ, ins->dreg, TRUE);
- x86_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
- break;
case OP_CLT:
- x86_set_reg (code, X86_CC_LT, ins->dreg, TRUE);
- x86_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
- break;
case OP_CLT_UN:
- x86_set_reg (code, X86_CC_LT, ins->dreg, FALSE);
- x86_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
- break;
case OP_CGT:
- x86_set_reg (code, X86_CC_GT, ins->dreg, TRUE);
- x86_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
- break;
case OP_CGT_UN:
- x86_set_reg (code, X86_CC_GT, ins->dreg, FALSE);
- x86_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
- break;
case OP_CNE:
- x86_set_reg (code, X86_CC_NE, ins->dreg, TRUE);
+ x86_set_reg (code, cc_table [mono_opcode_to_cond (ins->opcode)], ins->dreg, cc_signed_table [mono_opcode_to_cond (ins->opcode)]);
x86_widen_reg (code, ins->dreg, ins->dreg, FALSE, FALSE);
break;
case OP_COND_EXC_EQ:
case OP_COND_EXC_GE_UN:
case OP_COND_EXC_LE:
case OP_COND_EXC_LE_UN:
+ EMIT_COND_SYSTEM_EXCEPTION (cc_table [mono_opcode_to_cond (ins->opcode)], cc_signed_table [mono_opcode_to_cond (ins->opcode)], ins->inst_p1);
+ break;
case OP_COND_EXC_OV:
case OP_COND_EXC_NO:
case OP_COND_EXC_C:
case OP_COND_EXC_NC:
- EMIT_COND_SYSTEM_EXCEPTION (branch_cc_table [ins->opcode - OP_COND_EXC_EQ],
- (ins->opcode < OP_COND_EXC_NE_UN), ins->inst_p1);
+ EMIT_COND_SYSTEM_EXCEPTION (branch_cc_table [ins->opcode - OP_COND_EXC_EQ], (ins->opcode < OP_COND_EXC_NE_UN), ins->inst_p1);
break;
case CEE_BEQ:
case CEE_BNE_UN:
case CEE_BGE_UN:
case CEE_BLE:
case CEE_BLE_UN:
- EMIT_COND_BRANCH (ins, branch_cc_table [ins->opcode - CEE_BEQ], (ins->opcode < CEE_BNE_UN));
+ EMIT_COND_BRANCH (ins, cc_table [mono_opcode_to_cond (ins->opcode)], cc_signed_table [mono_opcode_to_cond (ins->opcode)]);
break;
/* floating point opcodes */
x86_alu_reg_imm (code, X86_SUB, X86_ESP, 8);
x86_fist_pop_membase (code, X86_ESP, 0, TRUE);
x86_pop_reg (code, ins->dreg);
- x86_pop_reg (code, ins->unused);
+ x86_pop_reg (code, ins->backend.reg3);
x86_fldcw_membase (code, X86_ESP, 0);
x86_alu_reg_imm (code, X86_ADD, X86_ESP, 4);
break;
}
case OP_LCONV_TO_OVF_I: {
guint8 *br [3], *label [1];
+ MonoInst *tins;
/*
* Valid ints: 0xffffffff:8000000 to 00000000:0x7f000000
label [0] = code;
/* throw exception */
- mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_EXC, "OverflowException");
- x86_jump32 (code, 0);
+ tins = mono_branch_optimize_exception_target (cfg, bb, "OverflowException");
+ if (tins) {
+ mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_BB, tins->inst_true_bb);
+ if ((cfg->opt & MONO_OPT_BRANCH) && x86_is_imm8 (tins->inst_true_bb->max_offset - cpos))
+ x86_jump8 (code, 0);
+ else
+ x86_jump32 (code, 0);
+ } else {
+ mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_EXC, "OverflowException");
+ x86_jump32 (code, 0);
+ }
+
x86_patch (br [0], code);
/* our top bit is set, check that top word is 0xfffffff */
case OP_FBGT:
case OP_FBGT_UN:
if (cfg->opt & MONO_OPT_FCMOV) {
- EMIT_COND_BRANCH (ins, X86_CC_LT, FALSE);
+ if (ins->opcode == OP_FBGT) {
+ guchar *br1;
+
+ /* skip branch if C1=1 */
+ br1 = code;
+ x86_branch8 (code, X86_CC_P, 0, FALSE);
+ /* branch if (C0 | C3) = 1 */
+ EMIT_COND_BRANCH (ins, X86_CC_LT, FALSE);
+ x86_patch (br1, code);
+ } else {
+ EMIT_COND_BRANCH (ins, X86_CC_LT, FALSE);
+ }
break;
}
x86_alu_reg_imm (code, X86_CMP, X86_EAX, X86_FP_C0);
x86_alu_reg_imm (code, X86_CMP, X86_EAX, X86_FP_C0);
EMIT_COND_BRANCH (ins, X86_CC_NE, FALSE);
break;
- case CEE_CKFINITE: {
+ case OP_CKFINITE: {
x86_push_reg (code, X86_EAX);
x86_fxam (code);
x86_fnstsw (code);
code = emit_tls_get (code, ins->dreg, ins->inst_offset);
break;
}
+ case OP_MEMORY_BARRIER: {
+ /* Not needed on x86 */
+ break;
+ }
case OP_ATOMIC_ADD_I4: {
int dreg = ins->dreg;
x86_push_reg (code, X86_EBP);
x86_mov_reg_reg (code, X86_EBP, X86_ESP, 4);
- alloc_size = - cfg->stack_offset;
+ alloc_size = cfg->stack_offset;
pos = 0;
if (method->wrapper_type == MONO_WRAPPER_NATIVE_TO_MANAGED) {
guint8 *buf;
code = emit_tls_get ( code, X86_EAX, lmf_tls_offset);
-#ifdef PLATFORM_WIN32
- /* The TLS key actually contains a pointer to the MonoJitTlsData structure */
- /* FIXME: Add a separate key for LMF to avoid this */
- x86_alu_reg_imm (code, X86_ADD, X86_EAX, G_STRUCT_OFFSET (MonoJitTlsData, lmf));
-#endif
x86_test_reg_reg (code, X86_EAX, X86_EAX);
buf = code;
x86_branch8 (code, X86_CC_NE, 0, 0);
code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD, (gpointer)"mono_jit_thread_attach");
x86_alu_reg_imm (code, X86_ADD, X86_ESP, 4);
x86_patch (buf, code);
- }
- else {
+#ifdef PLATFORM_WIN32
+ /* The TLS key actually contains a pointer to the MonoJitTlsData structure */
+ /* FIXME: Add a separate key for LMF to avoid this */
+ x86_alu_reg_imm (code, X86_ADD, X86_EAX, G_STRUCT_OFFSET (MonoJitTlsData, lmf));
+#endif
+ } else {
g_assert (!cfg->compile_aot);
x86_push_imm (code, cfg->domain);
code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD, (gpointer)"mono_jit_thread_attach");
x86_push_reg (code, X86_EDI);
x86_push_reg (code, X86_EBX);
- /* save method info */
- x86_push_imm (code, method);
+ if ((lmf_tls_offset != -1) && !is_win32 && !optimize_for_xen) {
+ /*
+ * Optimized version which uses the mono_lmf TLS variable instead of indirection
+ * through the mono_lmf_addr TLS variable.
+ */
+ /* %eax = previous_lmf */
+ x86_prefix (code, X86_GS_PREFIX);
+ x86_mov_reg_mem (code, X86_EAX, lmf_tls_offset, 4);
+ /* skip method_info + lmf */
+ x86_alu_reg_imm (code, X86_SUB, X86_ESP, 8);
+ /* push previous_lmf */
+ x86_push_reg (code, X86_EAX);
+ /* new lmf = ESP */
+ x86_prefix (code, X86_GS_PREFIX);
+ x86_mov_mem_reg (code, lmf_tls_offset, X86_ESP, 4);
+ } else {
+ /* get the address of lmf for the current thread */
+ /*
+ * This is performance critical so we try to use some tricks to make
+ * it fast.
+ */
- /* get the address of lmf for the current thread */
- /*
- * This is performance critical so we try to use some tricks to make
- * it fast.
- */
- if (lmf_tls_offset != -1) {
- /* Load lmf quicky using the GS register */
- code = emit_tls_get (code, X86_EAX, lmf_tls_offset);
+ if (lmf_addr_tls_offset != -1) {
+ /* Load lmf quicky using the GS register */
+ code = emit_tls_get (code, X86_EAX, lmf_addr_tls_offset);
#ifdef PLATFORM_WIN32
- /* The TLS key actually contains a pointer to the MonoJitTlsData structure */
- /* FIXME: Add a separate key for LMF to avoid this */
- x86_alu_reg_imm (code, X86_ADD, X86_EAX, G_STRUCT_OFFSET (MonoJitTlsData, lmf));
+ /* The TLS key actually contains a pointer to the MonoJitTlsData structure */
+ /* FIXME: Add a separate key for LMF to avoid this */
+ x86_alu_reg_imm (code, X86_ADD, X86_EAX, G_STRUCT_OFFSET (MonoJitTlsData, lmf));
#endif
- }
- else {
- if (cfg->compile_aot) {
- /* The GOT var does not exist yet */
- x86_call_imm (code, 0);
- mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_GOT_OFFSET, NULL);
- x86_pop_reg (code, X86_EAX);
- x86_alu_reg_imm (code, X86_ADD, X86_EAX, 0);
- mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_INTERNAL_METHOD, (gpointer)"mono_get_lmf_addr");
- x86_call_membase (code, X86_EAX, 0xf0f0f0f0);
- }
- else
+ } else {
code = emit_call (cfg, code, MONO_PATCH_INFO_INTERNAL_METHOD, (gpointer)"mono_get_lmf_addr");
- }
+ }
+
+ /* Skip method info */
+ x86_alu_reg_imm (code, X86_SUB, X86_ESP, 4);
- /* push lmf */
- x86_push_reg (code, X86_EAX);
- /* push *lfm (previous_lmf) */
- x86_push_membase (code, X86_EAX, 0);
- /* *(lmf) = ESP */
- x86_mov_membase_reg (code, X86_EAX, 0, X86_ESP, 4);
+ /* push lmf */
+ x86_push_reg (code, X86_EAX);
+ /* push *lfm (previous_lmf) */
+ x86_push_membase (code, X86_EAX, 0);
+ /* *(lmf) = ESP */
+ x86_mov_membase_reg (code, X86_EAX, 0, X86_ESP, 4);
+ }
} else {
if (cfg->used_int_regs & (1 << X86_EBX)) {
alloc_size -= pos;
+#if __APPLE__
+ /* the original alloc_size is already aligned: there is %ebp and retip pushed, so realign */
+ {
+ int tot = alloc_size + pos + 4 + 4; /* ret ip + ebp */
+ if (tot & 4) {
+ tot += 4;
+ alloc_size += 4;
+ }
+ if (tot & 8) {
+ alloc_size += 8;
+ }
+ }
+#endif
+
if (alloc_size) {
/* See mono_emit_stack_alloc */
#if defined(PLATFORM_WIN32) || defined(MONO_ARCH_SIGSEGV_ON_ALTSTACK)
#endif
}
+#if __APPLE_
+ /* check the stack is aligned */
+ x86_mov_reg_reg (code, X86_EDX, X86_ESP, 4);
+ x86_alu_reg_imm (code, X86_AND, X86_EDX, 15);
+ x86_alu_reg_imm (code, X86_CMP, X86_EDX, 0);
+ x86_branch_disp (code, X86_CC_EQ, 3, FALSE);
+ x86_breakpoint (code);
+#endif
+
/* compute max_offset in order to use short forward jumps */
max_offset = 0;
if (cfg->opt & MONO_OPT_BRANCH) {
if (ins->opcode == OP_LABEL)
ins->inst_c1 = max_offset;
- max_offset += ((guint8 *)ins_spec [ins->opcode])[MONO_INST_LEN];
+ max_offset += ((guint8 *)ins_get_spec (ins->opcode))[MONO_INST_LEN];
ins = ins->next;
}
}
pos = 0;
for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
- inst = cfg->varinfo [pos];
+ inst = cfg->args [pos];
if (inst->opcode == OP_REGVAR) {
x86_mov_reg_membase (code, inst->dreg, X86_EBP, inst->inst_offset, 4);
if (cfg->verbose_level > 2)
if (mono_jit_trace_calls != NULL && mono_trace_eval (method))
code = mono_arch_instrument_epilog (cfg, mono_trace_leave_method, code, TRUE);
- /* the code restoring the registers must be kept in sync with CEE_JMP */
+ /* the code restoring the registers must be kept in sync with OP_JMP */
pos = 0;
if (method->save_lmf) {
gint32 prev_lmf_reg;
+ gint32 lmf_offset = -sizeof (MonoLMF);
- /* Find a spare register */
- switch (sig->ret->type) {
- case MONO_TYPE_I8:
- case MONO_TYPE_U8:
- prev_lmf_reg = X86_EDI;
- cfg->used_int_regs |= (1 << X86_EDI);
- break;
- default:
- prev_lmf_reg = X86_EDX;
- break;
- }
+ if ((lmf_tls_offset != -1) && !is_win32 && !optimize_for_xen) {
+ /*
+ * Optimized version which uses the mono_lmf TLS variable instead of indirection
+ * through the mono_lmf_addr TLS variable.
+ */
+ /* reg = previous_lmf */
+ x86_mov_reg_membase (code, X86_ECX, X86_EBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, previous_lmf), 4);
+
+ /* lmf = previous_lmf */
+ x86_prefix (code, X86_GS_PREFIX);
+ x86_mov_mem_reg (code, lmf_tls_offset, X86_ECX, 4);
+ } else {
+ /* Find a spare register */
+ switch (sig->ret->type) {
+ case MONO_TYPE_I8:
+ case MONO_TYPE_U8:
+ prev_lmf_reg = X86_EDI;
+ cfg->used_int_regs |= (1 << X86_EDI);
+ break;
+ default:
+ prev_lmf_reg = X86_EDX;
+ break;
+ }
- /* reg = previous_lmf */
- x86_mov_reg_membase (code, prev_lmf_reg, X86_EBP, -32, 4);
+ /* reg = previous_lmf */
+ x86_mov_reg_membase (code, prev_lmf_reg, X86_EBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, previous_lmf), 4);
- /* ecx = lmf */
- x86_mov_reg_membase (code, X86_ECX, X86_EBP, -28, 4);
+ /* ecx = lmf */
+ x86_mov_reg_membase (code, X86_ECX, X86_EBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, lmf_addr), 4);
- /* *(lmf) = previous_lmf */
- x86_mov_membase_reg (code, X86_ECX, 0, prev_lmf_reg, 4);
+ /* *(lmf) = previous_lmf */
+ x86_mov_membase_reg (code, X86_ECX, 0, prev_lmf_reg, 4);
+ }
/* restore caller saved regs */
if (cfg->used_int_regs & (1 << X86_EBX)) {
- x86_mov_reg_membase (code, X86_EBX, X86_EBP, -20, 4);
+ x86_mov_reg_membase (code, X86_EBX, X86_EBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, ebx), 4);
}
if (cfg->used_int_regs & (1 << X86_EDI)) {
- x86_mov_reg_membase (code, X86_EDI, X86_EBP, -16, 4);
+ x86_mov_reg_membase (code, X86_EDI, X86_EBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, edi), 4);
}
if (cfg->used_int_regs & (1 << X86_ESI)) {
- x86_mov_reg_membase (code, X86_ESI, X86_EBP, -12, 4);
+ x86_mov_reg_membase (code, X86_ESI, X86_EBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, esi), 4);
}
/* EBP is restored by LEAVE */
}
/* Load returned vtypes into registers if needed */
- cinfo = get_call_info (sig, FALSE);
+ cinfo = get_call_info (cfg->mempool, sig, FALSE);
if (cinfo->ret.storage == ArgValuetypeInReg) {
for (quad = 0; quad < 2; quad ++) {
switch (cinfo->ret.pair_storage [quad]) {
else
x86_ret (code);
- g_free (cinfo);
-
cfg->code_len = code - cfg->native_code;
g_assert (cfg->code_len < cfg->code_size);
patch_info->type = MONO_PATCH_INFO_NONE;
}
else {
- guint32 got_reg = X86_EAX;
guint32 size;
/* Compute size of code following the push <OFFSET> */
- if (cfg->compile_aot) {
- size = 5 + 6;
- if (!cfg->got_var)
- size += 32;
- else if (cfg->got_var->opcode == OP_REGOFFSET)
- size += 6;
- }
- else
- size = 5 + 5;
+ size = 5 + 5;
if ((code - cfg->native_code) - throw_ip < 126 - size) {
/* Use the shorter form */
exc_throw_start [nthrows] = code;
}
- if (cfg->compile_aot) {
- /*
- * Since the patches are generated by the back end, there is * no way to generate a got_var at this point.
- */
- if (!cfg->got_var) {
- x86_call_imm (code, 0);
- mono_add_patch_info (cfg, code - cfg->native_code, MONO_PATCH_INFO_GOT_OFFSET, NULL);
- x86_pop_reg (code, X86_EAX);
- x86_alu_reg_imm (code, X86_ADD, X86_EAX, 0);
- }
- else {
- if (cfg->got_var->opcode == OP_REGOFFSET)
- x86_mov_reg_membase (code, X86_EAX, cfg->got_var->inst_basereg, cfg->got_var->inst_offset, 4);
- else
- got_reg = cfg->got_var->dreg;
- }
- }
-
- x86_push_imm (code, exc_class->type_token);
+ x86_push_imm (code, exc_class->type_token - MONO_TOKEN_TYPE_DEF);
patch_info->data.name = "mono_arch_throw_corlib_exception";
patch_info->type = MONO_PATCH_INFO_INTERNAL_METHOD;
patch_info->ip.i = code - cfg->native_code;
- if (cfg->compile_aot)
- x86_call_membase (code, got_reg, 0xf0f0f0f0);
- else
- x86_call_code (code, 0);
+ x86_call_code (code, 0);
x86_push_imm (buf, (code - cfg->native_code) - throw_ip);
while (buf < buf2)
x86_nop (buf);
{
}
-#ifdef MONO_ARCH_SIGSEGV_ON_ALTSTACK
-
-static void
-setup_stack (MonoJitTlsData *tls)
-{
- pthread_t self = pthread_self();
- pthread_attr_t attr;
- size_t stsize = 0;
- struct sigaltstack sa;
- guint8 *staddr = NULL;
- guint8 *current = (guint8*)&staddr;
-
- if (mono_running_on_valgrind ())
- return;
-
- /* Determine stack boundaries */
- pthread_attr_init( &attr );
-#ifdef HAVE_PTHREAD_GETATTR_NP
- pthread_getattr_np( self, &attr );
-#else
-#ifdef HAVE_PTHREAD_ATTR_GET_NP
- pthread_attr_get_np( self, &attr );
-#elif defined(sun)
- pthread_attr_getstacksize( &attr, &stsize );
-#else
-#error "Not implemented"
-#endif
-#endif
-#ifndef sun
- pthread_attr_getstack( &attr, (void**)&staddr, &stsize );
-#endif
-
- g_assert (staddr);
-
- g_assert ((current > staddr) && (current < staddr + stsize));
-
- tls->end_of_stack = staddr + stsize;
-
- /*
- * threads created by nptl does not seem to have a guard page, and
- * since the main thread is not created by us, we can't even set one.
- * Increasing stsize fools the SIGSEGV signal handler into thinking this
- * is a stack overflow exception.
- */
- tls->stack_size = stsize + getpagesize ();
-
- /* Setup an alternate signal stack */
- tls->signal_stack = mmap (0, SIGNAL_STACK_SIZE, PROT_READ|PROT_WRITE|PROT_EXEC, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0);
- tls->signal_stack_size = SIGNAL_STACK_SIZE;
-
- g_assert (tls->signal_stack);
-
- sa.ss_sp = tls->signal_stack;
- sa.ss_size = SIGNAL_STACK_SIZE;
- sa.ss_flags = SS_ONSTACK;
- sigaltstack (&sa, NULL);
-}
-
-#endif
-
/*
* Support for fast access to the thread-local lmf structure using the GS
* segment register on NPTL + kernel 2.6.x.
if (thread_tls_offset >= 64)
thread_tls_offset = -1;
#else
+#if MONO_XEN_OPT
+ optimize_for_xen = access ("/proc/xen", F_OK) == 0;
+#endif
tls_offset_inited = TRUE;
appdomain_tls_offset = mono_domain_get_tls_offset ();
lmf_tls_offset = mono_get_lmf_tls_offset ();
+ lmf_addr_tls_offset = mono_get_lmf_addr_tls_offset ();
thread_tls_offset = mono_thread_get_tls_offset ();
#endif
}
}
-
-#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
mono_arch_emit_this_vret_args (MonoCompile *cfg, MonoCallInst *inst, int this_reg, int this_type, int vt_reg)
{
+ MonoCallInst *call = (MonoCallInst*)inst;
+ CallInfo *cinfo = get_call_info (cfg->mempool, inst->signature, FALSE);
/* add the this argument */
if (this_reg != -1) {
- MonoInst *this;
- MONO_INST_NEW (cfg, this, OP_OUTARG);
- this->type = this_type;
- this->sreg1 = this_reg;
- mono_bblock_add_inst (cfg->cbb, this);
+ if (cinfo->args [0].storage == ArgInIReg) {
+ MonoInst *this;
+ MONO_INST_NEW (cfg, this, OP_MOVE);
+ this->type = this_type;
+ this->sreg1 = this_reg;
+ this->dreg = mono_regstate_next_int (cfg->rs);
+ mono_bblock_add_inst (cfg->cbb, this);
+
+ mono_call_inst_add_outarg_reg (cfg, call, this->dreg, cinfo->args [0].reg, FALSE);
+ }
+ else {
+ MonoInst *this;
+ MONO_INST_NEW (cfg, this, OP_OUTARG);
+ this->type = this_type;
+ this->sreg1 = this_reg;
+ mono_bblock_add_inst (cfg->cbb, this);
+ }
}
if (vt_reg != -1) {
- CallInfo * cinfo = get_call_info (inst->signature, FALSE);
MonoInst *vtarg;
if (cinfo->ret.storage == ArgValuetypeInReg) {
vtarg->sreg1 = vt_reg;
mono_bblock_add_inst (cfg->cbb, vtarg);
}
- else {
+ else if (cinfo->ret.storage == ArgInIReg) {
+ /* The return address is passed in a register */
+ MONO_INST_NEW (cfg, vtarg, OP_MOVE);
+ vtarg->sreg1 = vt_reg;
+ vtarg->dreg = mono_regstate_next_int (cfg->rs);
+ mono_bblock_add_inst (cfg->cbb, vtarg);
+
+ mono_call_inst_add_outarg_reg (cfg, call, vtarg->dreg, cinfo->ret.reg, FALSE);
+ } else {
MonoInst *vtarg;
MONO_INST_NEW (cfg, vtarg, OP_OUTARG);
vtarg->type = STACK_MP;
vtarg->sreg1 = vt_reg;
mono_bblock_add_inst (cfg->cbb, vtarg);
}
-
- g_free (cinfo);
}
}
-
MonoInst*
mono_arch_get_inst_for_method (MonoCompile *cfg, MonoMethod *cmethod, MonoMethodSignature *fsig, MonoInst **args)
{
ins->inst_i1 = args [1];
}
#endif
+ } else 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)) {
ins->inst_i0 = args [0];
ins->inst_i1 = args [1];
} else if (strcmp (cmethod->name, "Add") == 0 && fsig->params [0]->type == MONO_TYPE_I4) {
- MONO_INST_NEW (cfg, ins, OP_ATOMIC_ADD_I4);
+ MONO_INST_NEW (cfg, ins, OP_ATOMIC_ADD_NEW_I4);
ins->inst_i0 = args [0];
ins->inst_i1 = args [1];
* 0xff m=2,o=2 imm32
*/
code -= 6;
- if ((code [1] != 0xe8) && (code [3] == 0xff) && ((code [4] & 0x18) == 0x10) && ((code [4] >> 6) == 1)) {
+
+ /*
+ * A given byte sequence can match more than case here, so we have to be
+ * really careful about the ordering of the cases. Longer sequences
+ * come first.
+ */
+ if ((code [-2] == 0x8b) && (x86_modrm_mod (code [-1]) == 0x2) && (code [4] == 0xff) && (x86_modrm_reg (code [5]) == 0x2) && (x86_modrm_mod (code [5]) == 0x0)) {
+ /*
+ * This is an interface call
+ * 8b 80 0c e8 ff ff mov 0xffffe80c(%eax),%eax
+ * ff 10 call *(%eax)
+ */
+ reg = x86_modrm_rm (code [5]);
+ disp = 0;
+ } else if ((code [1] != 0xe8) && (code [3] == 0xff) && ((code [4] & 0x18) == 0x10) && ((code [4] >> 6) == 1)) {
reg = code [4] & 0x07;
disp = (signed char)code [5];
} else {
return NULL;
} else if ((code [4] == 0xff) && (((code [5] >> 6) & 0x3) == 0) && (((code [5] >> 3) & 0x7) == 2)) {
/*
- * This is a interface call: should check the above code can't catch it earlier
+ * This is a interface call
* 8b 40 30 mov 0x30(%eax),%eax
* ff 10 call *(%eax)
*/
return (gpointer*)(((gint32)(regs [reg])) + disp);
}
-gpointer*
-mono_arch_get_delegate_method_ptr_addr (guint8* code, gpointer *regs)
+gpointer
+mono_arch_get_this_arg_from_call (MonoMethodSignature *sig, gssize *regs, guint8 *code)
{
- guint8 reg = 0;
- gint32 disp = 0;
+ guint32 esp = regs [X86_ESP];
+ CallInfo *cinfo;
+ gpointer res;
- code -= 7;
- if ((code [0] == 0x8b) && (x86_modrm_mod (code [1]) == 3) && (x86_modrm_reg (code [1]) == X86_EAX) && (code [2] == 0x8b) && (code [3] == 0x40) && (code [5] == 0xff) && (code [6] == 0xd0)) {
- reg = x86_modrm_rm (code [1]);
- disp = code [4];
+ cinfo = get_call_info (NULL, sig, FALSE);
- if (reg == X86_EAX)
- return NULL;
- else
- return (gpointer*)(((gint32)(regs [reg])) + disp);
+ /*
+ * The stack looks like:
+ * <other args>
+ * <this=delegate>
+ * <possible vtype return address>
+ * <return addr>
+ * <4 pointers pushed by mono_arch_create_trampoline_code ()>
+ */
+ res = (((MonoObject**)esp) [5 + (cinfo->args [0].offset / 4)]);
+ g_free (cinfo);
+ return res;
+}
+
+gpointer
+mono_arch_get_delegate_invoke_impl (MonoMethodSignature *sig, gboolean has_target)
+{
+ guint8 *code, *start;
+ MonoDomain *domain = mono_domain_get ();
+
+ /* FIXME: Support more cases */
+ if (MONO_TYPE_ISSTRUCT (sig->ret))
+ return NULL;
+
+ /*
+ * The stack contains:
+ * <delegate>
+ * <return addr>
+ */
+
+ if (has_target) {
+ mono_domain_lock (domain);
+ start = code = mono_code_manager_reserve (domain->code_mp, 64);
+ mono_domain_unlock (domain);
+
+ /* Replace the this argument with the target */
+ x86_mov_reg_membase (code, X86_EAX, X86_ESP, 4, 4);
+ x86_mov_reg_membase (code, X86_ECX, X86_EAX, G_STRUCT_OFFSET (MonoDelegate, target), 4);
+ x86_mov_membase_reg (code, X86_ESP, 4, X86_ECX, 4);
+ x86_jump_membase (code, X86_EAX, G_STRUCT_OFFSET (MonoDelegate, method_ptr));
+
+ g_assert ((code - start) < 64);
+ } else {
+ if (sig->param_count == 0) {
+ mono_domain_lock (domain);
+ start = code = mono_code_manager_reserve (domain->code_mp, 32 + (sig->param_count * 8));
+ mono_domain_unlock (domain);
+
+ x86_mov_reg_membase (code, X86_EAX, X86_ESP, 4, 4);
+ x86_jump_membase (code, X86_EAX, G_STRUCT_OFFSET (MonoDelegate, method_ptr));
+ } else {
+ /*
+ * The code below does not work in the presence of exceptions, since it
+ * creates a new frame.
+ */
+ start = NULL;
+#if 0
+ for (i = 0; i < sig->param_count; ++i)
+ if (!mono_is_regsize_var (sig->params [i]))
+ return NULL;
+
+ mono_domain_lock (domain);
+ start = code = mono_code_manager_reserve (domain->code_mp, 32 + (sig->param_count * 8));
+ mono_domain_unlock (domain);
+
+ /* Load this == delegate */
+ x86_mov_reg_membase (code, X86_EAX, X86_ESP, 4, 4);
+
+ /* Push arguments in opposite order, taking changes in ESP into account */
+ for (i = 0; i < sig->param_count; ++i)
+ x86_push_membase (code, X86_ESP, 4 + (sig->param_count * 4));
+
+ /* Call the delegate */
+ x86_call_membase (code, X86_EAX, G_STRUCT_OFFSET (MonoDelegate, method_ptr));
+ if (sig->param_count > 0)
+ x86_alu_reg_imm (code, X86_ADD, X86_ESP, sig->param_count * 4);
+ x86_ret (code);
+#endif
+ }
}
- return NULL;
+ return start;
}
projects / mono.git / blobdiff