1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
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
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
|
/*
* Copyright (C) 2020 Collabora, Ltd.
*
* Permission is hereby granted, free of charge, to any person obtaining a
* copy of this software and associated documentation files (the "Software"),
* to deal in the Software without restriction, including without limitation
* the rights to use, copy, modify, merge, publish, distribute, sublicense,
* and/or sell copies of the Software, and to permit persons to whom the
* Software is furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice (including the next
* paragraph) shall be included in all copies or substantial portions of the
* Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
* THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
#include "compiler.h"
/* NIR creates vectors as vecN ops, which we represent by a synthetic
* BI_COMBINE instruction, e.g.:
*
* v = combine x, y, z, w
*
* These combines need to be lowered by the pass in this file. Fix a given
* source at component c.
*
* First suppose the source is SSA. If it is also scalar, then we may rewrite
* the destination of the generating instruction (unique by SSA+scalar) to
* write to v.c, and rewrite each of its uses to swizzle out .c instead of .x
* (the original by scalar). If it is vector, there are two cases. If the
* component c is `x`, we are accessing v.x, and each of the succeeding
* components y, z... up to the last component of the vector are accessed
* sequentially, then we may perform the same rewrite. If this is not the case,
* rewriting would require more complex vector features, so we fallback on a
* move.
*
* Otherwise is the source is not SSA, we also fallback on a move. We could
* probably do better.
*/
static void
bi_combine_mov32(bi_context *ctx, bi_instruction *parent, unsigned comp, unsigned R)
{
bi_instruction move = {
.type = BI_MOV,
.dest = R,
.dest_type = nir_type_uint32,
.dest_offset = comp,
.src = { parent->src[comp] },
.src_types = { nir_type_uint32 },
.swizzle = { { parent->swizzle[comp][0] } }
};
bi_emit_before(ctx, parent, move);
}
static void
bi_combine_sel16(bi_context *ctx, bi_instruction *parent, unsigned comp, unsigned R)
{
bi_instruction sel = {
.type = BI_SELECT,
.dest = R,
.dest_type = nir_type_uint32,
.dest_offset = comp >> 1,
.src = { parent->src[comp], parent->src[comp + 1] },
.src_types = { nir_type_uint16, nir_type_uint16 },
.swizzle = { {
parent->swizzle[comp][0],
parent->swizzle[comp + 1][0],
} }
};
bi_emit_before(ctx, parent, sel);
}
/* Gets the instruction generating a given source. Combine lowering is
* accidentally O(n^2) right now because this function is O(n) instead of O(1).
* If this pass is slow, this cost can be avoided in favour for better
* bookkeeping. */
static bi_instruction *
bi_get_parent(bi_context *ctx, unsigned idx)
{
bi_foreach_instr_global(ctx, ins) {
if (ins->dest == idx)
return ins;
}
return NULL;
}
/* Rewrites uses of an index. Again, this could be O(n) to the program but is
* currently O(nc) to the program and number of combines, so the pass becomes
* effectively O(n^2). Better bookkeeping would bring down to linear if that's
* an issue. */
static void
bi_rewrite_uses(bi_context *ctx,
unsigned old, unsigned oldc,
unsigned new, unsigned newc)
{
bi_foreach_instr_global(ctx, ins) {
bi_foreach_src(ins, s) {
if (ins->src[s] != old) continue;
for (unsigned i = 0; i < 16; ++i)
ins->swizzle[s][i] += (newc - oldc);
ins->src[s] = new;
}
}
}
/* Checks if we have a nicely aligned vector prefix */
static bool
bi_is_aligned_vec32(bi_instruction *combine, unsigned s, bi_instruction *io,
unsigned *count)
{
/* We only support prefixes */
if (s != 0)
return false;
if (!(bi_class_props[io->type] & BI_VECTOR))
return false;
if (nir_alu_type_get_type_size(combine->dest_type) != 32)
return false;
if (nir_alu_type_get_type_size(io->dest_type) != 32)
return false;
unsigned components = io->vector_channels;
/* Are we contiguous like that? */
for (unsigned i = 0; i < components; ++i) {
if (combine->src[i] != io->dest)
return false;
if (combine->swizzle[i][0] != i)
return false;
}
/* We're good to go */
*count = components;
return true;
}
#if 0
/* Tries to lower a given source of a combine to an appropriate rewrite,
* returning true if successful, and false with no changes otherwise. */
static bool
bi_lower_combine_src(bi_context *ctx, bi_instruction *ins, unsigned s, unsigned R,
unsigned *vec_count)
{
unsigned src = ins->src[s];
/* We currently only handle SSA */
if (!src) return false;
if (src & (BIR_SPECIAL | PAN_IS_REG)) return false;
/* We are SSA. Lookup the generating instruction. */
unsigned bytes = nir_alu_type_get_type_size(ins->dest_type) / 8;
bi_instruction *parent = bi_get_parent(ctx, src,
0xF << (ins->swizzle[s][0] * bytes));
if (!parent) return false;
/* We have a parent instuction, sanity check the typesize */
unsigned pbytes = nir_alu_type_get_type_size(parent->dest_type) / 8;
if (pbytes != bytes) return false;
bool scalar = parent->vector_channels != 0;
if (!(scalar || bi_is_aligned_vec(ins, s, parent, vec_count))) return false;
if (!bi_shift_mask(parent, bytes * s)) return false;
bi_rewrite_uses(ctx, parent->dest, 0, R, s);
parent->dest = R;
return true;
}
#endif
void
bi_lower_combine(bi_context *ctx, bi_block *block)
{
bi_foreach_instr_in_block_safe(block, ins) {
if (ins->type != BI_COMBINE) continue;
unsigned R = bi_make_temp_reg(ctx);
unsigned sz = nir_alu_type_get_type_size(ins->dest_type);
bi_foreach_src(ins, s) {
/* We're done early for vec2/3 */
if (!ins->src[s])
continue;
#if 0
unsigned vec_count = 0;
if (bi_lower_combine_src(ctx, ins, s, R, &vec_count)) {
/* Skip vectored sources */
if (vec_count)
s += (vec_count - 1);
} else {
bi_insert_combine_mov(ctx, ins, s, R);
}
#endif
if (sz == 32)
bi_combine_mov32(ctx, ins, s, R);
else if (sz == 16) {
bi_combine_sel16(ctx, ins, s, R);
s++;
} else {
unreachable("Unknown COMBINE size");
}
}
bi_rewrite_uses(ctx, ins->dest, 0, R, 0);
bi_remove_instruction(ins);
}
}
|
