#define MONO_EMIT_NEW_AMD64_ICOMPARE_MEMBASE_REG(cfg,basereg,offset,operand) do { \
		MonoInst *inst; \
		inst = mono_mempool_alloc0 ((cfg)->mempool, sizeof (MonoInst));	\
		inst->opcode = OP_AMD64_ICOMPARE_MEMBASE_REG; \
		inst->inst_basereg = basereg; \
		inst->inst_offset = offset; \
		inst->sreg2 = operand; \
		mono_bblock_add_inst (cfg->cbb, inst); \
	} while (0)

#define MONO_EMIT_NEW_AMD64_ICOMPARE_MEMBASE_IMM(cfg,basereg,offset,operand) do { \
		MonoInst *inst; \
		inst = mono_mempool_alloc0 ((cfg)->mempool, sizeof (MonoInst));	\
		inst->opcode = OP_AMD64_ICOMPARE_MEMBASE_IMM; \
		inst->inst_basereg = basereg; \
		inst->inst_offset = offset; \
		inst->inst_imm = operand; \
		mono_bblock_add_inst (cfg->cbb, inst); \
	} while (0)

/* override the arch independant versions with fast x86 versions */

#undef MONO_EMIT_BOUNDS_CHECK
#undef MONO_EMIT_BOUNDS_CHECK_IMM

#define MONO_EMIT_BOUNDS_CHECK(cfg, array_reg, array_type, array_length_field, index_reg) do { \
		if (! (state->tree->flags & MONO_INST_NORANGECHECK)) { \
			MONO_EMIT_NEW_AMD64_ICOMPARE_MEMBASE_REG (cfg, array_reg, G_STRUCT_OFFSET (array_type, array_length_field), index_reg); \
			MONO_EMIT_NEW_COND_EXC (cfg, LE_UN, "IndexOutOfRangeException"); \
		} \
	} while (0)

#define MONO_EMIT_BOUNDS_CHECK_IMM(cfg, array_reg, array_type, array_length_field, index_imm) do { \
		if (! (state->tree->flags & MONO_INST_NORANGECHECK)) { \
			MONO_EMIT_NEW_AMD64_ICOMPARE_MEMBASE_IMM (cfg, array_reg, G_STRUCT_OFFSET (array_type, array_length_field), index_imm); \
			MONO_EMIT_NEW_COND_EXC (cfg, LE_UN, "IndexOutOfRangeException"); \
		} \
	} while (0)

int cbranch_to_fcbranch (int opcode);
int bialu_to_bialu_imm (int opcode);
int ceq_to_fceq (int opcode);
int call_reg_to_call_membase (int opcode);

%%

#
# inssel-amd64.brg: burg file for special AMD64 instructions
#
# Author:
#   Dietmar Maurer (dietmar@ximian.com)
#   Paolo Molaro (lupus@ximian.com)
#
# (C) 2002 Ximian, Inc.
#

stmt: OP_START_HANDLER {
	MonoInst *spvar = mono_find_spvar_for_region (s, s->cbb->region);
	MONO_EMIT_NEW_STORE_MEMBASE (s, OP_STORE_MEMBASE_REG, spvar->inst_basereg, spvar->inst_offset, X86_ESP);
}

stmt: CEE_ENDFINALLY {
	MonoInst *spvar = mono_find_spvar_for_region (s, s->cbb->region);
	MONO_EMIT_NEW_LOAD_MEMBASE (s, X86_ESP, spvar->inst_basereg, spvar->inst_offset); 
	tree->opcode = CEE_RET;
	mono_bblock_add_inst (s->cbb, tree);
}

stmt: OP_ENDFILTER (reg) {
	MonoInst *spvar = mono_find_spvar_for_region (s, s->cbb->region);
	MONO_EMIT_NEW_UNALU (s, OP_MOVE, X86_EAX, state->left->reg1);
	MONO_EMIT_NEW_LOAD_MEMBASE (s, X86_ESP, spvar->inst_basereg, spvar->inst_offset); 
	tree->opcode = CEE_RET;
	mono_bblock_add_inst (s->cbb, tree);
}

freg: OP_LCONV_TO_R8 (reg) {
	tree->sreg1 = state->left->reg1;
	tree->dreg = state->reg1;
	mono_bblock_add_inst (s->cbb, tree);
}

freg: OP_LCONV_TO_R4 (reg) {
	tree->sreg1 = state->left->reg1;
	tree->dreg = state->reg1;
	mono_bblock_add_inst (s->cbb, tree);
}

cflags: OP_COMPARE (CEE_LDIND_I4 (base), reg) {
	tree->opcode = OP_AMD64_ICOMPARE_MEMBASE_REG;
	tree->inst_basereg = state->left->left->tree->inst_basereg;
	tree->inst_offset = state->left->left->tree->inst_offset;
	tree->sreg2 = state->right->reg1;
	mono_bblock_add_inst (s->cbb, tree);
}

cflags: OP_COMPARE (CEE_LDIND_I4 (base), OP_ICONST) {
	tree->opcode = OP_AMD64_ICOMPARE_MEMBASE_IMM;
	tree->inst_basereg = state->left->left->tree->inst_basereg;
	tree->inst_offset = state->left->left->tree->inst_offset;
	tree->inst_imm = state->right->tree->inst_c0;
	mono_bblock_add_inst (s->cbb, tree);
}

cflags: OP_COMPARE (reg, CEE_LDIND_I4 (base)) {
	tree->opcode = OP_AMD64_ICOMPARE_REG_MEMBASE;
	tree->sreg2 = state->right->left->tree->inst_basereg;
	tree->inst_offset = state->right->left->tree->inst_offset;
	tree->sreg1 = state->left->reg1;
	mono_bblock_add_inst (s->cbb, tree);
}

stmt: CEE_STIND_I1 (base, OP_CEQ (cflags)) {
	tree->opcode = OP_X86_SETEQ_MEMBASE;
	tree->inst_offset = state->left->tree->inst_offset;
	tree->inst_basereg = state->left->tree->inst_basereg;
	mono_bblock_add_inst (s->cbb, tree);
}

reg: OP_LOCALLOC (OP_ICONST) {
	if (tree->flags & MONO_INST_INIT) {
		/* microcoded in mini-x86.c */
		tree->sreg1 = mono_regstate_next_int (s->rs);
		tree->dreg = state->reg1;
		MONO_EMIT_NEW_ICONST (s, tree->sreg1, state->left->tree->inst_c0);
		mono_bblock_add_inst (s->cbb, tree);
	} else {
		guint32 size = state->left->tree->inst_c0;
		size = (size + (MONO_ARCH_FRAME_ALIGNMENT - 1)) & ~ (MONO_ARCH_FRAME_ALIGNMENT - 1);
		MONO_EMIT_NEW_BIALU_IMM (s, OP_SUB_IMM, X86_ESP, X86_ESP, size);
		MONO_EMIT_UNALU (s, tree, OP_MOVE, state->reg1, X86_ESP);
	}
}

reg: OP_LOCALLOC (reg) {
	tree->sreg1 = state->left->tree->dreg;
	tree->dreg = state->reg1;
	mono_bblock_add_inst (s->cbb, tree);
}

stmt: OP_SETRET (reg) {
	tree->opcode = OP_MOVE;
	tree->sreg1 = state->left->reg1;
	tree->dreg = X86_EAX;
	mono_bblock_add_inst (s->cbb, tree);
}

stmt: OP_SETRET (reg) {
	MONO_EMIT_NEW_UNALU (s, OP_MOVE, X86_EDX, state->left->reg2);
	tree->opcode = OP_MOVE;
	tree->sreg1 = state->left->reg1;
	tree->dreg = X86_EAX;
	mono_bblock_add_inst (s->cbb, tree);
}

reg: CEE_LDIND_REF (OP_REGVAR),
reg: CEE_LDIND_I (OP_REGVAR),
reg: CEE_LDIND_I4 (OP_REGVAR),
reg: CEE_LDIND_U4 (OP_REGVAR) "0" {
	/* This rule might not work on all archs */
	state->reg1 = state->left->tree->dreg;
	tree->dreg = state->reg1;
}

stmt: OP_SETRET (CEE_LDIND_REF (OP_REGVAR)),
stmt: OP_SETRET (CEE_LDIND_I4 (OP_REGVAR)),
stmt: OP_SETRET (CEE_LDIND_U4 (OP_REGVAR)),
stmt: OP_SETRET (CEE_LDIND_I (OP_REGVAR)) {
	tree->opcode = OP_MOVE;
	tree->sreg1 = state->left->left->tree->dreg;
	tree->dreg = X86_EAX;
	mono_bblock_add_inst (s->cbb, tree);
}

stmt: OP_SETRET (freg) {
	if (mono_method_signature (s->method)->ret->type == MONO_TYPE_R4)
		tree->opcode = OP_AMD64_SET_XMMREG_R4;
	else
		tree->opcode = OP_AMD64_SET_XMMREG_R8;
	tree->sreg1 = state->left->reg1;
	tree->dreg = 0; /* %xmm0 */
	mono_bblock_add_inst (s->cbb, tree);	
	/* nothing to do */
}

stmt: OP_SETRET (OP_ICONST) {
	if (state->left->tree->inst_c0 == 0) {
		MONO_EMIT_BIALU (s, tree, CEE_XOR, AMD64_RAX, AMD64_RAX, AMD64_RAX);
	}
	else {
		tree->opcode = OP_ICONST;
		tree->inst_c0 = state->left->tree->inst_c0;
		tree->dreg = X86_EAX;
		mono_bblock_add_inst (s->cbb, tree);
	}
}

stmt: OP_OUTARG (reg) {
	tree->opcode = OP_X86_PUSH;
	tree->sreg1 = state->left->reg1;
	mono_bblock_add_inst (s->cbb, tree);
}

stmt: OP_OUTARG_REG (reg) {	
	MonoCallInst *call = (MonoCallInst*)tree->inst_right;

	tree->opcode = OP_MOVE;
	tree->sreg1 = state->left->reg1;
	tree->dreg = mono_regstate_next_int (s->rs);
	mono_bblock_add_inst (s->cbb, tree);

	mono_call_inst_add_outarg_reg (s, call, tree->dreg, tree->unused, FALSE);
}

# we need to reduce this code duplication with some burg syntax extension
stmt: OP_OUTARG (CEE_LDIND_REF (OP_REGVAR)) {
	tree->opcode = OP_X86_PUSH;
	tree->sreg1 = state->left->left->tree->dreg;
	mono_bblock_add_inst (s->cbb, tree);
}

stmt: OP_OUTARG (CEE_LDIND_I4 (OP_REGVAR)) {
	tree->opcode = OP_X86_PUSH;
	tree->sreg1 = state->left->left->tree->dreg;
	mono_bblock_add_inst (s->cbb, tree);
}

stmt: OP_OUTARG (CEE_LDIND_U4 (OP_REGVAR)) {
	tree->opcode = OP_X86_PUSH;
	tree->sreg1 = state->left->left->tree->dreg;
	mono_bblock_add_inst (s->cbb, tree);
}

stmt: OP_OUTARG (CEE_LDIND_I (OP_REGVAR)) {
	tree->opcode = OP_X86_PUSH;
	tree->sreg1 = state->left->left->tree->dreg;
	mono_bblock_add_inst (s->cbb, tree);
}

stmt: OP_OUTARG (CEE_LDIND_I (base)) {
	tree->opcode = OP_X86_PUSH_MEMBASE;
	tree->inst_basereg = state->left->left->tree->inst_basereg;
	tree->inst_offset = state->left->left->tree->inst_offset;
	mono_bblock_add_inst (s->cbb, tree);
}

stmt: OP_OUTARG (CEE_LDIND_REF (base)) {
	tree->opcode = OP_X86_PUSH_MEMBASE;
	tree->inst_basereg = state->left->left->tree->inst_basereg;
	tree->inst_offset = state->left->left->tree->inst_offset;
	mono_bblock_add_inst (s->cbb, tree);
}

stmt: OP_OUTARG (CEE_LDIND_REF (OP_REGVAR)) {
	tree->opcode = OP_X86_PUSH;
	tree->sreg1 = state->left->left->tree->dreg;
	mono_bblock_add_inst (s->cbb, tree);
}

stmt: OP_OUTARG (CEE_LDOBJ (reg)) {
	tree->opcode = OP_X86_PUSH;
	tree->sreg1 = state->left->reg1;
	mono_bblock_add_inst (s->cbb, tree);
}

stmt: OP_OUTARG (freg) {
	MONO_EMIT_NEW_BIALU_IMM (s, OP_SUB_IMM, X86_ESP, X86_ESP, 8);
	tree->opcode = OP_STORER8_MEMBASE_REG;
	tree->sreg1 = state->left->reg1;
	tree->inst_destbasereg = X86_ESP;
	tree->inst_offset = 0;
	mono_bblock_add_inst (s->cbb, tree);
}

stmt: OP_OUTARG_R4 (freg) {
	MONO_EMIT_NEW_BIALU_IMM (s, OP_SUB_IMM, X86_ESP, X86_ESP, 8);
	tree->opcode = OP_STORER4_MEMBASE_REG;
	tree->sreg1 = state->left->reg1;
	tree->inst_destbasereg = X86_ESP;
	tree->inst_offset = 0;
	mono_bblock_add_inst (s->cbb, tree);
}

stmt: OP_OUTARG_R8 (freg) {
	MONO_EMIT_NEW_BIALU_IMM (s, OP_SUB_IMM, X86_ESP, X86_ESP, 8);
	tree->opcode = OP_STORER8_MEMBASE_REG;
	tree->sreg1 = state->left->reg1;
	tree->inst_destbasereg = X86_ESP;
	tree->inst_offset = 0;
	mono_bblock_add_inst (s->cbb, tree);
}

stmt: OP_AMD64_OUTARG_XMMREG_R4 (freg) {
	MonoCallInst *call = (MonoCallInst*)tree->inst_right;

	tree->opcode = OP_AMD64_SET_XMMREG_R4;
	tree->sreg1 = state->left->reg1;
	tree->dreg = mono_regstate_next_float (s->rs);
	mono_bblock_add_inst (s->cbb, tree);

	mono_call_inst_add_outarg_reg (s, call, tree->dreg, tree->unused, TRUE);
}

stmt: OP_AMD64_OUTARG_XMMREG_R8 (freg) {
	MonoCallInst *call = (MonoCallInst*)tree->inst_right;

	tree->opcode = OP_AMD64_SET_XMMREG_R8;
	tree->sreg1 = state->left->reg1;
	tree->dreg = mono_regstate_next_float (s->rs);
	mono_bblock_add_inst (s->cbb, tree);

	mono_call_inst_add_outarg_reg (s, call, tree->dreg, tree->unused, TRUE);
}

stmt: OP_OUTARG_VT (CEE_LDOBJ (base)) {
	MonoInst *vt = state->left->left->tree;
	//g_print ("vt size: %d at R%d + %d\n", tree->inst_imm, vt->inst_basereg, vt->inst_offset);

	if (!tree->inst_imm)
		return;

	if (tree->inst_imm == 8) {
		/* Can't use this for < 8 since it does an 8 byte memory load */
		tree->opcode = OP_X86_PUSH_MEMBASE;
		tree->inst_basereg = vt->inst_basereg;
		tree->inst_offset = vt->inst_offset;
		mono_bblock_add_inst (s->cbb, tree);
	} else if (tree->inst_imm <= 20) {
		int sz = tree->inst_imm;
		sz += 7;
		sz &= ~7;
		MONO_EMIT_NEW_BIALU_IMM (s, OP_SUB_IMM, X86_ESP, X86_ESP, sz);
		mini_emit_memcpy (s, X86_ESP, 0, vt->inst_basereg, vt->inst_offset, tree->inst_imm, 0);
	} else {
		tree->opcode = OP_X86_PUSH_OBJ;
		tree->inst_basereg = vt->inst_basereg;
		tree->inst_offset = vt->inst_offset;
		mono_bblock_add_inst (s->cbb, tree);
	}
}

stmt: OP_OUTARG_VT (OP_ICONST) {
	tree->opcode = OP_X86_PUSH_IMM;
	tree->inst_imm = state->left->tree->inst_c0;
	mono_bblock_add_inst (s->cbb, tree);
}

stmt: OP_OUTARG_VT (reg) {
	tree->opcode = OP_X86_PUSH;
	tree->sreg1 = state->left->tree->dreg;
	mono_bblock_add_inst (s->cbb, tree);
}

stmt: OP_AMD64_OUTARG_ALIGN_STACK {
	MONO_EMIT_NEW_BIALU_IMM (s, OP_SUB_IMM, X86_ESP, X86_ESP, 8);
}	

base: OP_INARG_VT (base) {
	MONO_EMIT_NEW_LOAD_MEMBASE (s, state->reg1, state->left->tree->inst_basereg, 
					state->left->tree->inst_offset);
}

reg: OP_LDADDR (OP_INARG_VT (base)) {
	MONO_EMIT_NEW_LOAD_MEMBASE (s, state->reg1, state->left->left->tree->inst_basereg, 
					state->left->left->tree->inst_offset);
}

reg: CEE_LDOBJ (OP_INARG_VT (base)) {
	MONO_EMIT_NEW_LOAD_MEMBASE (s, state->reg1, state->left->left->tree->inst_basereg, 
					state->left->left->tree->inst_offset);
}

reg: OP_LDADDR (OP_REGOFFSET) "1" {
	if (state->left->tree->inst_offset) {
		tree->opcode = OP_X86_LEA_MEMBASE;
		tree->sreg1 = state->left->tree->inst_basereg;
		tree->inst_imm = state->left->tree->inst_offset;
		tree->dreg = state->reg1;
	} else {
		tree->opcode = OP_MOVE;
		tree->sreg1 = state->left->tree->inst_basereg;
		tree->dreg = state->reg1;
	}
	mono_bblock_add_inst (s->cbb, tree);
}

reg: CEE_LDOBJ (OP_REGOFFSET) "1" {
	if (state->left->tree->inst_offset) {
		tree->opcode = OP_X86_LEA_MEMBASE;
		tree->sreg1 = state->left->tree->inst_basereg;
		tree->inst_imm = state->left->tree->inst_offset;
		tree->dreg = state->reg1;
	} else {
		tree->opcode = OP_MOVE;
		tree->sreg1 = state->left->tree->inst_basereg;
		tree->dreg = state->reg1;
	}
	mono_bblock_add_inst (s->cbb, tree);
}

reg: CEE_LDELEMA (reg, reg) "15" {
	guint32 size = mono_class_array_element_size (tree->klass);
	
	MONO_EMIT_BOUNDS_CHECK (s, state->left->reg1, MonoArray, max_length, state->right->reg1);

	if (size == 1 || size == 2 || size == 4 || size == 8) {
		static const int fast_log2 [] = { 1, 0, 1, -1, 2, -1, -1, -1, 3 };
		int reg;

		/* The array reg is 64 bits but the index reg is only 32 */
		tree->dreg = mono_regstate_next_float (s->rs);
		reg = mono_regstate_next_int (s->rs);
		MONO_EMIT_NEW_UNALU (s, OP_SEXT_I4, reg, state->right->reg1);

		tree->opcode = OP_X86_LEA;
		tree->dreg = state->reg1;
		tree->sreg1 = state->left->reg1;
		tree->sreg2 = reg;
		tree->inst_imm = G_STRUCT_OFFSET (MonoArray, vector);
		tree->unused = fast_log2 [size];
		mono_bblock_add_inst (s->cbb, tree);
	} else {
		int mult_reg = mono_regstate_next_int (s->rs);
		int add_reg = mono_regstate_next_int (s->rs);
		MONO_EMIT_NEW_BIALU_IMM (s, OP_MUL_IMM, mult_reg, state->right->reg1, size);
		MONO_EMIT_NEW_BIALU (s, CEE_ADD, add_reg, mult_reg, state->left->reg1);
		MONO_EMIT_NEW_BIALU_IMM (s, OP_ADD_IMM, state->reg1, add_reg, G_STRUCT_OFFSET (MonoArray, vector));
	}
}

stmt: CEE_STIND_R8 (OP_REGVAR, freg) {
	/* nothing to do: the value is already on the FP stack */
}

stmt: CEE_STIND_I4 (base, CEE_ADD (CEE_LDIND_I4 (base), OP_ICONST)) {
	int con = state->right->right->tree->inst_c0;	

	if (con == 1) {
		tree->opcode = OP_X86_INC_MEMBASE;
	} else {
		tree->opcode = OP_X86_ADD_MEMBASE_IMM;
		tree->inst_imm = con;
	}

	tree->inst_basereg = state->left->tree->inst_basereg;
	tree->inst_offset = state->left->tree->inst_offset;
	mono_bblock_add_inst (s->cbb, tree);
} cost {
	MBTREE_TYPE *t1 = state->right->left->left->tree;
	MBTREE_TYPE *t2 = state->left->tree;
	MBCOND (t1->inst_basereg == t2->inst_basereg &&
		t1->inst_offset == t2->inst_offset);
	return 2;
}

stmt: CEE_STIND_I4 (base, CEE_SUB (CEE_LDIND_I4 (base), OP_ICONST)) {
	int con = state->right->right->tree->inst_c0;	

	if (con == 1) {
		tree->opcode = OP_X86_DEC_MEMBASE;
	} else {
		tree->opcode = OP_X86_SUB_MEMBASE_IMM;
		tree->inst_imm = con;
	}

	tree->inst_basereg = state->left->tree->inst_basereg;
	tree->inst_offset = state->left->tree->inst_offset;
	mono_bblock_add_inst (s->cbb, tree);
} cost {
	MBTREE_TYPE *t1 = state->right->left->left->tree;
	MBTREE_TYPE *t2 = state->left->tree;
	MBCOND (t1->inst_basereg == t2->inst_basereg &&
		t1->inst_offset == t2->inst_offset);
	return 2;
}

stmt: CEE_STIND_I4 (OP_REGVAR, CEE_SUB (CEE_LDIND_I4 (OP_REGVAR), OP_ICONST)) {
	int con = state->right->right->tree->inst_c0;	
	int dreg = state->left->tree->dreg;
	int sreg = state->right->left->left->tree->dreg;

	if (con == 1) {
		if (dreg != sreg)
			MONO_EMIT_NEW_UNALU (s, OP_MOVE, dreg, sreg);
		tree->opcode = OP_X86_DEC_REG;
		tree->dreg = tree->sreg1 = dreg;
	} else if (con == -1) {
		if (dreg != sreg)
			MONO_EMIT_NEW_UNALU (s, OP_MOVE, dreg, sreg);
		tree->opcode = OP_X86_INC_REG;
		tree->dreg = tree->sreg1 = dreg;
	} else {
		tree->opcode = OP_SUB_IMM;
		tree->inst_imm = con;
		tree->sreg1 = sreg;
		tree->dreg = dreg;
	}
	mono_bblock_add_inst (s->cbb, tree);
}

stmt: CEE_STIND_I (OP_REGVAR, CEE_ADD (CEE_LDIND_I (OP_REGVAR), OP_ICONST)),
stmt: CEE_STIND_I4 (OP_REGVAR, CEE_ADD (CEE_LDIND_I4 (OP_REGVAR), OP_ICONST)) {
	int con = state->right->right->tree->inst_c0;
	int dreg = state->left->tree->dreg;
	int sreg = state->right->left->left->tree->dreg;

	if (con == 1) {
		if (dreg != sreg)
			MONO_EMIT_NEW_UNALU (s, OP_MOVE, dreg, sreg);
		tree->opcode = OP_X86_INC_REG;
		tree->dreg = tree->sreg1 = dreg;
	} else if (con == -1) {
		if (dreg != sreg)
			MONO_EMIT_NEW_UNALU (s, OP_MOVE, dreg, sreg);
		tree->opcode = OP_X86_DEC_REG;
		tree->dreg = tree->sreg1 = dreg;
	} else {
		tree->opcode = OP_ADD_IMM;
		tree->inst_imm = con;
		tree->sreg1 = sreg;
		tree->dreg = dreg;
	}
	mono_bblock_add_inst (s->cbb, tree);
}

reg: CEE_LDIND_I2 (OP_REGVAR) {
	MONO_EMIT_UNALU (s, tree, OP_SEXT_I2, state->reg1, state->left->tree->dreg);
}

# The XOR rule
stmt: CEE_STIND_I8 (OP_REGVAR, OP_ICONST),
stmt: CEE_STIND_I4 (OP_REGVAR, OP_ICONST),
stmt: CEE_STIND_I2 (OP_REGVAR, OP_ICONST),
stmt: CEE_STIND_I1 (OP_REGVAR, OP_ICONST),
stmt: CEE_STIND_REF (OP_REGVAR, OP_ICONST),
stmt: CEE_STIND_I (OP_REGVAR, OP_ICONST),
stmt: CEE_STIND_I8 (OP_REGVAR, OP_I8CONST),
stmt: CEE_STIND_I4 (OP_REGVAR, OP_I8CONST),
stmt: CEE_STIND_I2 (OP_REGVAR, OP_I8CONST),
stmt: CEE_STIND_I1 (OP_REGVAR, OP_I8CONST),
stmt: CEE_STIND_REF (OP_REGVAR, OP_I8CONST),
stmt: CEE_STIND_I (OP_REGVAR, OP_I8CONST) {
	int r = state->left->tree->dreg;
	MONO_EMIT_BIALU (s, tree, CEE_XOR, r, r, r);
} cost {
	MBCOND (!state->right->tree->inst_c0);
	
	return 0;
}

# on x86, fp compare overwrites EAX, so we must
# either improve the local register allocator or
# emit coarse opcodes which saves EAX for us.

reg: OP_CEQ (OP_COMPARE (freg, freg)),
reg: OP_CLT (OP_COMPARE (freg, freg)),
reg: OP_CGT (OP_COMPARE (freg, freg)),
reg: OP_CLT_UN (OP_COMPARE (freg, freg)),
reg: OP_CGT_UN (OP_COMPARE (freg, freg)) {
	MONO_EMIT_BIALU (s, tree, ceq_to_fceq (tree->opcode), state->reg1, state->left->left->reg1,
			 state->left->right->reg1);
}

# fpcflags overwrites EAX, but this does not matter for statements
# because we are the last operation in the tree.
 
stmt: CEE_BNE_UN (fpcflags),
stmt: CEE_BEQ (fpcflags),
stmt: CEE_BLT (fpcflags),
stmt: CEE_BLT_UN (fpcflags),
stmt: CEE_BGT (fpcflags),
stmt: CEE_BGT_UN (fpcflags),
stmt: CEE_BGE  (fpcflags),
stmt: CEE_BGE_UN (fpcflags),
stmt: CEE_BLE  (fpcflags),
stmt: CEE_BLE_UN (fpcflags) {
	tree->opcode = cbranch_to_fcbranch (tree->opcode);
	mono_bblock_add_inst (s->cbb, tree);
}

stmt: CEE_POP (freg) "0" {
	/* we need to pop the value from the x86 FP stack */
	MONO_EMIT_UNALU (s, tree, OP_X86_FPOP, -1, state->left->reg1);	
}     

# override the rules in inssel-float.brg that work for machines with FP registers 

freg: OP_FCONV_TO_R8 (freg) "0" {
	tree->opcode = OP_FMOVE;
	tree->sreg1 = state->left->reg1;
	tree->dreg = state->reg1;
	mono_bblock_add_inst (s->cbb, tree);
}

freg: OP_FCONV_TO_R4 (freg) "0" {
	tree->opcode = OP_FMOVE;
	tree->sreg1 = state->left->reg1;
	tree->dreg = state->reg1;
	mono_bblock_add_inst (s->cbb, tree);
}

reg: CEE_ADD(reg, CEE_LDIND_I4 (base)) {
	MonoInst *base = state->right->left->tree;

	MONO_EMIT_BIALU_MEMBASE (cfg, tree, OP_X86_ADD_MEMBASE, state->reg1, state->left->reg1, base->inst_basereg, base->inst_offset);
} 

reg: CEE_SUB(reg, CEE_LDIND_I4 (base)) {
	MonoInst *base = state->right->left->tree;

	MONO_EMIT_BIALU_MEMBASE (cfg, tree, OP_X86_SUB_MEMBASE, state->reg1, state->left->reg1, base->inst_basereg, base->inst_offset);
} 

reg: CEE_MUL(reg, CEE_LDIND_I4 (base)) {
	MonoInst *base = state->right->left->tree;

	MONO_EMIT_BIALU_MEMBASE (cfg, tree, OP_X86_MUL_MEMBASE, state->reg1, state->left->reg1, base->inst_basereg, base->inst_offset);
} 

reg: OP_LSHL (reg, reg),
reg: OP_LSHR (reg, reg),
reg: OP_LSHR_UN (reg, reg),
reg: OP_LMUL (reg, reg),
reg: OP_LDIV (reg, reg),
reg: OP_LDIV_UN (reg, reg),
reg: OP_LREM (reg, reg),
reg: OP_LREM_UN (reg, reg),
reg: OP_LMUL_OVF (reg, reg),
reg: OP_LMUL_OVF_UN (reg, reg) "0" {
	MONO_EMIT_BIALU (s, tree, tree->opcode, state->reg1, state->left->reg1, state->right->reg1);
}

reg: OP_LMUL (reg, OP_I8CONST),
reg: OP_LSHL (reg, OP_ICONST),
reg: OP_LSHR (reg, OP_ICONST),
reg: OP_LSHR_UN (reg, OP_ICONST) {
	MONO_EMIT_BIALU_IMM (s, tree, bialu_to_bialu_imm (tree->opcode), state->reg1, state->left->reg1, state->right->tree->inst_c0);
} cost {
	MBCOND (mono_arch_is_inst_imm (state->right->tree->inst_c0));
	return 0;
}

reg: OP_ATOMIC_ADD_NEW_I4 (base, reg),
reg: OP_ATOMIC_ADD_NEW_I8 (base, reg),
reg: OP_ATOMIC_ADD_I4 (base, reg),
reg: OP_ATOMIC_ADD_I8 (base, reg),
reg: OP_ATOMIC_EXCHANGE_I4 (base, reg),
reg: OP_ATOMIC_EXCHANGE_I8 (base, reg) {
	tree->opcode = tree->opcode;
	tree->dreg = state->reg1;
	tree->sreg2 = state->right->reg1;
	tree->inst_basereg = state->left->tree->inst_basereg; 
	tree->inst_offset = state->left->tree->inst_offset; 
    
	mono_bblock_add_inst (s->cbb, tree);
}

# Optimized call instructions
# mono_arch_patch_delegate_trampoline depends on these
reg: OP_CALL_REG (CEE_LDIND_I (base)),
freg: OP_FCALL_REG (CEE_LDIND_I (base)),
reg: OP_LCALL_REG (CEE_LDIND_I (base)) {
	tree->opcode = call_reg_to_call_membase (tree->opcode);
	tree->inst_basereg = state->left->left->tree->inst_basereg;
	tree->inst_offset = state->left->left->tree->inst_offset;
	tree->dreg = state->reg1;
	mono_bblock_add_inst (s->cbb, tree);
}

stmt: OP_VOIDCALL_REG (CEE_LDIND_I (base)) {
	tree->opcode = call_reg_to_call_membase (tree->opcode);
	tree->inst_basereg = state->left->left->tree->inst_basereg;
	tree->inst_offset = state->left->left->tree->inst_offset;
	mono_bblock_add_inst (s->cbb, tree);
}

stmt: OP_VCALL_REG (CEE_LDIND_I (base), reg) {
	mono_arch_emit_this_vret_args (s, (MonoCallInst*)tree, -1, -1, state->right->reg1);
	
	tree->opcode = call_reg_to_call_membase (tree->opcode);
	tree->inst_basereg = state->left->left->tree->inst_basereg;
	tree->inst_offset = state->left->left->tree->inst_offset;
	tree->dreg = state->reg1;
	mono_bblock_add_inst (s->cbb, tree);
}

%%

int
bialu_to_bialu_imm (int opcode)
{
	switch (opcode) {
	case OP_LMUL:
		return OP_LMUL_IMM;
	case OP_LSHL:
		return OP_LSHL_IMM;
	case OP_LSHR:
		return OP_LSHR_IMM;
	case OP_LSHR_UN:
		return OP_LSHR_UN_IMM;
	default:
		g_assert_not_reached ();
	}

	return -1;
}

int
cbranch_to_fcbranch (int opcode)
{
	switch (opcode) {
	case CEE_BNE_UN:
		return OP_FBNE_UN;
	case CEE_BEQ:
		return OP_FBEQ;
	case CEE_BLT:
		return OP_FBLT;
	case CEE_BLT_UN:
		return OP_FBLT_UN;
	case CEE_BGT:
		return OP_FBGT;
	case CEE_BGT_UN:
		return OP_FBGT_UN;
	case CEE_BGE:
		return OP_FBGE;
	case CEE_BGE_UN:
		return OP_FBGE_UN;
	case CEE_BLE:
		return OP_FBLE;
	case CEE_BLE_UN:
		return OP_FBLE_UN;
	default:
		g_assert_not_reached ();
	}

	return -1;
}

int
ceq_to_fceq (int opcode)
{
	switch (opcode) {
	case OP_CEQ:
		return OP_FCEQ;
	case OP_CLT:
		return OP_FCLT;
	case OP_CGT:
		return OP_FCGT;
	case OP_CLT_UN:
		return OP_FCLT_UN;
	case OP_CGT_UN:
		return OP_FCGT_UN;
	default:
		g_assert_not_reached ();
	}

	return -1;
}

int
call_reg_to_call_membase (int opcode)
{
	switch (opcode) {
	case OP_CALL_REG:
		return OP_CALL_MEMBASE;
	case OP_FCALL_REG:
		return OP_FCALL_MEMBASE;
	case OP_VCALL_REG:
		return OP_VCALL_MEMBASE;
	case OP_LCALL_REG:
		return OP_LCALL_MEMBASE;
	case OP_VOIDCALL_REG:
		return OP_VOIDCALL_MEMBASE;
	default:
		g_assert_not_reached ();
	}

	return -1;
}