amd64_finish.c 8.23 KB
Newer Older
1
2
/*
 * This file is part of libFirm.
3
 * Copyright (C) 2014 University of Karlsruhe.
4
5
6
7
8
9
10
 */

/**
 * @file
 * @brief   This file implements functions to finalize the irg for emit.
 */
#include "amd64_finish.h"
Matthias Braun's avatar
Matthias Braun committed
11
12
#include "amd64_new_nodes.h"
#include "amd64_nodes_attr.h"
13
#include "amd64_transform.h"
14
15
16
17
#include "bearch.h"
#include "benode.h"
#include "besched.h"
#include "debug.h"
Matthias Braun's avatar
Matthias Braun committed
18
#include "panic.h"
19
#include "gen_amd64_new_nodes.h"
20
#include "gen_amd64_regalloc_if.h"
21
#include "irgwalk.h"
Matthias Braun's avatar
Matthias Braun committed
22
#include "util.h"
23
#include "irgmod.h"
24
25
26
27
28
29

DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)

/**
 * Returns the index of the first "same" register.
 */
Matthias Braun's avatar
Matthias Braun committed
30
static unsigned get_first_same(arch_register_req_t const *const req)
31
32
{
	unsigned const other = req->other_same;
Matthias Braun's avatar
Matthias Braun committed
33
	for (unsigned i = 0; i != 32; ++i) {
34
35
36
37
38
39
		if (other & (1U << i))
			return i;
	}
	panic("same position not found");
}

40
41
42
43
44
static bool is_commutative(const ir_node *node)
{
	return arch_get_irn_flags(node) & amd64_arch_irn_flag_commutative_binop;
}

Matthias Braun's avatar
Matthias Braun committed
45
46
47
static bool try_swap_inputs(ir_node *node)
{
	/* commutative operation, just switch the inputs */
48
49
	if (is_commutative(node)) {
		assert(get_amd64_attr_const(node)->op_mode == AMD64_OP_REG_REG);
Matthias Braun's avatar
Matthias Braun committed
50
51
52
53
54
55
56
57
58
59
		/* TODO: support Cmp input swapping */
		ir_node *in0 = get_irn_n(node, 0);
		ir_node *in1 = get_irn_n(node, 1);
		set_irn_n(node, 0, in1);
		set_irn_n(node, 1, in0);
		return true;
	}
	return false;
}

60
61
62
63
64
65
66
67
68
69
/**
  * Transforms a Sub to a Neg + Add, which subsequently allows swapping
  * of the inputs. The swapping is also (implicitly) done here.
  */
static void transform_sub_to_neg_add(ir_node *node,
                                     const arch_register_t *out_reg)
{
	ir_node  *block = get_nodes_block(node);
	dbg_info *dbgi  = get_irn_dbg_info(node);

70
71
72
	ir_graph *irg = get_irn_irg(node);
	ir_node  *in1 = get_irn_n(node, 0);
	ir_node  *in2 = get_irn_n(node, 1);
73

74
	const arch_register_t *in2_reg = arch_get_irn_register(in2);
75

76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
	const amd64_binop_addr_attr_t *attr = get_amd64_binop_addr_attr(node);
	ir_node *add, *add_res;

	if (is_amd64_xSubs(node)) {
		ir_mode *mode = get_irn_mode(in2);
		ir_tarval *tv;
		if (get_mode_size_bits(mode) <= 32)
			tv = create_sign_tv(mode_Iu);
		else
			tv = create_sign_tv(mode_Lu);

		ir_entity *sign_bit_const = create_float_const_entity(irg, tv);

		/* xXorp needs a 128-bit memory address since it's a vector instruction.
		 * Setting the alignment works in this case, but is not the nice way.
		 * TODO: Create and set correct mode/type for the entity.
		 */
		set_entity_alignment(sign_bit_const, 16);

		amd64_binop_addr_attr_t xor_attr;
		memset(&xor_attr, 0, sizeof(xor_attr));
		xor_attr.base.addr.base_input       = NO_INPUT;
		xor_attr.base.addr.index_input      = NO_INPUT;
		xor_attr.base.addr.immediate.entity = sign_bit_const;
		xor_attr.base.base.op_mode          = AMD64_OP_ADDR_REG;

		ir_node *xor_in[] = { in2 };
		ir_node *xor = new_bd_amd64_xXorp(dbgi, block, ARRAY_SIZE(xor_in),
		                                  xor_in, &xor_attr);
		ir_node *neg = new_r_Proj(xor, mode_D, pn_amd64_xXorp_res);

		sched_add_before(node, xor);
		arch_set_irn_register(neg, in2_reg);

		ir_node *in[] = { neg, in1 };
		add     = new_bd_amd64_xAdds(dbgi, block, ARRAY_SIZE(in), in, attr);
		add_res = new_r_Proj(add, mode_D, pn_amd64_xAdds_res);
	} else {
		assert(is_amd64_Sub(node));
115
116
		ir_node *neg = new_bd_amd64_Neg(dbgi, block, in2, attr->base.insn_mode);
		arch_set_irn_register_out(neg, pn_amd64_Neg_res, out_reg);
117
		sched_add_before(node, neg);
118
119
120
		ir_node *neg_res
			= new_r_Proj(neg, amd64_reg_classes[CLASS_amd64_gp].mode,
			             pn_amd64_Neg_res);
121

122
		ir_node *in[] = { neg_res, in1 };
123
124
125
		add     = new_bd_amd64_Add(dbgi, block, ARRAY_SIZE(in), in, attr);
		add_res = new_r_Proj(add, mode_Lu, pn_amd64_Add_res);
	}
126
127
128
129
130
131
132
133
134
	arch_set_irn_register(add_res, out_reg);

	/* exchange the add and the sub */
	edges_reroute(node, add);
	sched_replace(node, add);

	kill_node(node);
}

Matthias Braun's avatar
Matthias Braun committed
135
136
static ir_node *amd64_turn_back_am(ir_node *node)
{
Matthias Braun's avatar
Matthias Braun committed
137
138
139
	dbg_info          *dbgi  = get_irn_dbg_info(node);
	ir_node           *block = get_nodes_block(node);
	amd64_addr_attr_t *attr  = get_amd64_addr_attr(node);
Matthias Braun's avatar
Matthias Braun committed
140

Matthias Braun's avatar
Matthias Braun committed
141
	amd64_addr_t new_addr = attr->addr;
Matthias Braun's avatar
Matthias Braun committed
142
143
	ir_node *load_in[3];
	int      load_arity = 0;
Matthias Braun's avatar
Matthias Braun committed
144
145
146
147
	if (attr->addr.base_input != NO_INPUT
	 && attr->addr.base_input != RIP_INPUT) {
		new_addr.base_input = load_arity;
		load_in[load_arity++] = get_irn_n(node, attr->addr.base_input);
Matthias Braun's avatar
Matthias Braun committed
148
	}
Matthias Braun's avatar
Matthias Braun committed
149
150
151
	if (attr->addr.index_input != NO_INPUT) {
		new_addr.index_input = load_arity;
		load_in[load_arity++] = get_irn_n(node, attr->addr.index_input);
Matthias Braun's avatar
Matthias Braun committed
152
	}
Matthias Braun's avatar
Matthias Braun committed
153
154
155
	assert(attr->addr.mem_input != NO_INPUT);
	new_addr.mem_input = load_arity;
	load_in[load_arity++] = get_irn_n(node, attr->addr.mem_input);
Matthias Braun's avatar
Matthias Braun committed
156

Matthias Braun's avatar
Matthias Braun committed
157
	ir_node *load = new_bd_amd64_Mov(dbgi, block, load_arity, load_in,
158
	                                 attr->insn_mode, AMD64_OP_ADDR, new_addr);
Matthias Braun's avatar
Matthias Braun committed
159
	ir_node *load_res = new_r_Proj(load, mode_Lu, pn_amd64_Mov_res);
Matthias Braun's avatar
Matthias Braun committed
160
161

	/* change operation */
Matthias Braun's avatar
Matthias Braun committed
162
163
	const amd64_binop_addr_attr_t *binop_attr
		= (const amd64_binop_addr_attr_t*)attr;
Matthias Braun's avatar
Matthias Braun committed
164
	ir_node *new_in[2];
Matthias Braun's avatar
Matthias Braun committed
165
	new_in[0] = get_irn_n(node, binop_attr->u.reg_input);
Matthias Braun's avatar
Matthias Braun committed
166
167
	new_in[1] = load_res;
	set_irn_in(node, ARRAY_SIZE(new_in), new_in);
Matthias Braun's avatar
Matthias Braun committed
168
169
170
	attr->base.op_mode     = AMD64_OP_REG_REG;
	attr->addr.base_input  = NO_INPUT;
	attr->addr.index_input = NO_INPUT;
Matthias Braun's avatar
Matthias Braun committed
171
172
173
174
175
176

	/* rewire mem-proj */
	foreach_out_edge(node, edge) {
		ir_node *out = get_edge_src_irn(edge);
		if (get_irn_mode(out) == mode_M) {
			set_Proj_pred(out, load);
177
			set_Proj_num(out, pn_amd64_Mov_M);
Matthias Braun's avatar
Matthias Braun committed
178
179
180
181
182
183
184
185
186
			break;
		}
	}

	if (sched_is_scheduled(node))
		sched_add_before(node, load);
	return load_res;
}

187
188
189
190
191
192
193
/**
 * Insert copies for all amd64 nodes where the should_be_same requirement is
 * not fulfilled.
 */
static void assure_should_be_same_requirements(ir_node *const node)
{
	/* Check all OUT requirements, if there is a should_be_same. */
Christoph Mallon's avatar
Christoph Mallon committed
194
	be_foreach_out(node, i) {
Matthias Braun's avatar
Matthias Braun committed
195
196
197
198
199
200
201
202
203
204
205
		arch_register_req_t const *const req
			= arch_get_irn_register_req_out(node, i);
		if (!arch_register_req_is(req, should_be_same))
			continue;
		unsigned               const same_pos = get_first_same(req);
		ir_node               *const in_node  = get_irn_n(node, same_pos);
		arch_register_t const *const in_reg   = arch_get_irn_register(in_node);
		arch_register_t const *const out_reg
			= arch_get_irn_register_out(node, i);
		if (in_reg == out_reg)
			continue;
206

Matthias Braun's avatar
Matthias Braun committed
207
208
209
210
211
		/* test if any other input is using the out register */
		for (int i2 = 0, arity = get_irn_arity(node); i2 < arity; ++i2) {
			const arch_register_t *reg = arch_get_irn_register_in(node, i2);
			if (reg == out_reg && (unsigned)i2 != same_pos) {
				if (!is_amd64_irn(node))
212
					panic("cannot fulfill should_be_same on non-amd64 node");
Matthias Braun's avatar
Matthias Braun committed
213
214
				/* see what role this register has */
				const amd64_attr_t *attr = get_amd64_attr_const(node);
Matthias Braun's avatar
Matthias Braun committed
215
216
217
				if (attr->op_mode == AMD64_OP_ADDR
				 || attr->op_mode == AMD64_OP_REG
				 || attr->op_mode == AMD64_OP_REG_IMM) {
Matthias Braun's avatar
Matthias Braun committed
218
					panic("unexpected op_mode");
Matthias Braun's avatar
Matthias Braun committed
219
				} else if (attr->op_mode == AMD64_OP_REG_REG) {
Matthias Braun's avatar
Matthias Braun committed
220
221
222
223
swap:;
					bool res = try_swap_inputs(node);
					if (res)
						return;
224

225
					if (is_amd64_Sub(node) || is_amd64_xSubs(node)) {
226
227
228
						transform_sub_to_neg_add(node, out_reg);
						return;
					}
Matthias Braun's avatar
Matthias Braun committed
229
230
					panic("couldn't swap inputs of %+F", node);
				} else {
Matthias Braun's avatar
Matthias Braun committed
231
					assert(attr->op_mode == AMD64_OP_ADDR_REG);
Matthias Braun's avatar
Matthias Braun committed
232
233
234
235
236
237
238
239
					/* extract load into an own instruction */
					ir_node *res = amd64_turn_back_am(node);
					arch_set_irn_register(res, out_reg);
					goto swap;
				}
			}
		}

Matthias Braun's avatar
Matthias Braun committed
240
241
		ir_node *const block = get_nodes_block(node);
		ir_node *const copy  = be_new_Copy(block, in_node);
242

Matthias Braun's avatar
Matthias Braun committed
243
244
245
246
247
248
249
250
		/* Destination is the out register. */
		arch_set_irn_register(copy, out_reg);
		/* Insert copy before the node into the schedule. */
		sched_add_before(node, copy);
		/* Set copy as in. */
		set_irn_n(node, same_pos, copy);

		DBG((dbg, LEVEL_1, "created copy %+F for should be same argument at input %d of %+F\n", copy, same_pos, node));
251
252
253
254
255
256
257
258
259
260
261
	}
}

/**
 * Block walker: finishes a block.
 */
static void amd64_finish_irg_walker(ir_node *const block, void *const env)
{
	(void) env;

	/* Insert should_be_same copies. */
262
263
	sched_foreach_safe(block, irn) {
		if (is_amd64_irn(irn))
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
			assure_should_be_same_requirements(irn);
	}
}

/**
 * Add Copy nodes for not fulfilled should_be_same constraints.
 */
void amd64_finish_irg(ir_graph *const irg)
{
	irg_block_walk_graph(irg, 0, amd64_finish_irg_walker, 0);
}

void amd64_init_finish(void)
{
	FIRM_DBG_REGISTER(dbg, "firm.be.amd64.finish");
}