Merge branch 'lp-offset-twoside'
[mesa.git] / src / mesa / drivers / dri / i965 / brw_wm_pass2.c
1 /*
2 Copyright (C) Intel Corp. 2006. All Rights Reserved.
3 Intel funded Tungsten Graphics (http://www.tungstengraphics.com) to
4 develop this 3D driver.
5
6 Permission is hereby granted, free of charge, to any person obtaining
7 a copy of this software and associated documentation files (the
8 "Software"), to deal in the Software without restriction, including
9 without limitation the rights to use, copy, modify, merge, publish,
10 distribute, sublicense, and/or sell copies of the Software, and to
11 permit persons to whom the Software is furnished to do so, subject to
12 the following conditions:
13
14 The above copyright notice and this permission notice (including the
15 next paragraph) shall be included in all copies or substantial
16 portions of the Software.
17
18 THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
19 EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20 MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
21 IN NO EVENT SHALL THE COPYRIGHT OWNER(S) AND/OR ITS SUPPLIERS BE
22 LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
23 OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
24 WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
25
26 **********************************************************************/
27 /*
28 * Authors:
29 * Keith Whitwell <keith@tungstengraphics.com>
30 */
31
32
33 #include "brw_context.h"
34 #include "brw_wm.h"
35
36
37 /* Use these to force spilling so that that functionality can be
38 * tested with known-good examples rather than having to construct new
39 * tests.
40 */
41 #define TEST_PAYLOAD_SPILLS 0
42 #define TEST_DST_SPILLS 0
43
44 static void spill_value(struct brw_wm_compile *c,
45 struct brw_wm_value *value);
46
47 static void prealloc_reg(struct brw_wm_compile *c,
48 struct brw_wm_value *value,
49 GLuint reg)
50 {
51 if (value->lastuse) {
52 /* Set nextuse to zero, it will be corrected by
53 * update_register_usage().
54 */
55 c->pass2_grf[reg].value = value;
56 c->pass2_grf[reg].nextuse = 0;
57
58 value->resident = &c->pass2_grf[reg];
59 value->hw_reg = brw_vec8_grf(reg*2, 0);
60
61 if (TEST_PAYLOAD_SPILLS)
62 spill_value(c, value);
63 }
64 }
65
66
67 /* Initialize all the register values. Do the initial setup
68 * calculations for interpolants.
69 */
70 static void init_registers( struct brw_wm_compile *c )
71 {
72 GLuint nr_interp_regs = 0;
73 GLuint i = 0;
74 GLuint j;
75
76 for (j = 0; j < c->grf_limit; j++)
77 c->pass2_grf[j].nextuse = BRW_WM_MAX_INSN;
78
79 for (j = 0; j < (c->key.nr_payload_regs + 1) / 2; j++)
80 prealloc_reg(c, &c->payload.depth[j], i++);
81
82 for (j = 0; j < c->nr_creg; j++)
83 prealloc_reg(c, &c->creg[j], i++);
84
85 for (j = 0; j < VERT_RESULT_MAX; j++) {
86 if (c->key.vp_outputs_written & BITFIELD64_BIT(j)) {
87 int fp_index;
88
89 if (j >= VERT_RESULT_VAR0)
90 fp_index = j - (VERT_RESULT_VAR0 - FRAG_ATTRIB_VAR0);
91 else if (j <= VERT_RESULT_TEX7)
92 fp_index = j;
93 else
94 fp_index = -1;
95
96 nr_interp_regs++;
97 if (fp_index >= 0)
98 prealloc_reg(c, &c->payload.input_interp[fp_index], i++);
99 }
100 }
101
102 assert(nr_interp_regs >= 1);
103
104 c->prog_data.first_curbe_grf = ALIGN(c->key.nr_payload_regs, 2);
105 c->prog_data.urb_read_length = nr_interp_regs * 2;
106 c->prog_data.curb_read_length = c->nr_creg * 2;
107
108 c->max_wm_grf = i * 2;
109 }
110
111
112 /* Update the nextuse value for each register in our file.
113 */
114 static void update_register_usage(struct brw_wm_compile *c,
115 GLuint thisinsn)
116 {
117 GLuint i;
118
119 for (i = 1; i < c->grf_limit; i++) {
120 struct brw_wm_grf *grf = &c->pass2_grf[i];
121
122 /* Only search those which can change:
123 */
124 if (grf->nextuse < thisinsn) {
125 const struct brw_wm_ref *ref = grf->value->lastuse;
126
127 /* Has last use of value been passed?
128 */
129 if (ref->insn < thisinsn) {
130 grf->value->resident = 0;
131 grf->value = 0;
132 grf->nextuse = BRW_WM_MAX_INSN;
133 }
134 else {
135 /* Else loop through chain to update:
136 */
137 while (ref->prevuse && ref->prevuse->insn >= thisinsn)
138 ref = ref->prevuse;
139
140 grf->nextuse = ref->insn;
141 }
142 }
143 }
144 }
145
146
147 static void spill_value(struct brw_wm_compile *c,
148 struct brw_wm_value *value)
149 {
150 /* Allocate a spill slot. Note that allocations start from 0x40 -
151 * the first slot is reserved to mean "undef" in brw_wm_emit.c
152 */
153 if (!value->spill_slot) {
154 c->last_scratch += 0x40;
155 value->spill_slot = c->last_scratch;
156 }
157
158 /* The spill will be done in brw_wm_emit.c immediately after the
159 * value is calculated, so we can just take this reg without any
160 * further work.
161 */
162 value->resident->value = NULL;
163 value->resident->nextuse = BRW_WM_MAX_INSN;
164 value->resident = NULL;
165 }
166
167
168
169 /* Search for contiguous region with the most distant nearest
170 * member. Free regs count as very distant.
171 *
172 * TODO: implement spill-to-reg so that we can rearrange discontigous
173 * free regs and then spill the oldest non-free regs in sequence.
174 * This would mean inserting instructions in this pass.
175 */
176 static GLuint search_contiguous_regs(struct brw_wm_compile *c,
177 GLuint nr,
178 GLuint thisinsn)
179 {
180 struct brw_wm_grf *grf = c->pass2_grf;
181 GLuint furthest = 0;
182 GLuint reg = 0;
183 GLuint i, j;
184
185 /* Start search at 1: r0 is special and can't be used or spilled.
186 */
187 for (i = 1; i < c->grf_limit && furthest < BRW_WM_MAX_INSN; i++) {
188 GLuint group_nextuse = BRW_WM_MAX_INSN;
189
190 for (j = 0; j < nr; j++) {
191 if (grf[i+j].nextuse < group_nextuse)
192 group_nextuse = grf[i+j].nextuse;
193 }
194
195 if (group_nextuse > furthest) {
196 furthest = group_nextuse;
197 reg = i;
198 }
199 }
200
201 assert(furthest != thisinsn);
202
203 /* Any non-empty regs will need to be spilled:
204 */
205 for (j = 0; j < nr; j++)
206 if (grf[reg+j].value)
207 spill_value(c, grf[reg+j].value);
208
209 return reg;
210 }
211
212
213 static void alloc_contiguous_dest(struct brw_wm_compile *c,
214 struct brw_wm_value *dst[],
215 GLuint nr,
216 GLuint thisinsn)
217 {
218 GLuint reg = search_contiguous_regs(c, nr, thisinsn);
219 GLuint i;
220
221 for (i = 0; i < nr; i++) {
222 if (!dst[i]) {
223 /* Need to grab a dummy value in TEX case. Don't introduce
224 * it into the tracking scheme.
225 */
226 dst[i] = &c->vreg[c->nr_vreg++];
227 }
228 else {
229 assert(!dst[i]->resident);
230 assert(c->pass2_grf[reg+i].nextuse != thisinsn);
231
232 c->pass2_grf[reg+i].value = dst[i];
233 c->pass2_grf[reg+i].nextuse = thisinsn;
234
235 dst[i]->resident = &c->pass2_grf[reg+i];
236 }
237
238 dst[i]->hw_reg = brw_vec8_grf((reg+i)*2, 0);
239 }
240
241 if ((reg+nr)*2 > c->max_wm_grf)
242 c->max_wm_grf = (reg+nr) * 2;
243 }
244
245
246 static void load_args(struct brw_wm_compile *c,
247 struct brw_wm_instruction *inst)
248 {
249 GLuint thisinsn = inst - c->instruction;
250 GLuint i,j;
251
252 for (i = 0; i < 3; i++) {
253 for (j = 0; j < 4; j++) {
254 struct brw_wm_ref *ref = inst->src[i][j];
255
256 if (ref) {
257 if (!ref->value->resident) {
258 /* Need to bring the value in from scratch space. The code for
259 * this will be done in brw_wm_emit.c, here we just do the
260 * register allocation and mark the ref as requiring a fill.
261 */
262 GLuint reg = search_contiguous_regs(c, 1, thisinsn);
263
264 c->pass2_grf[reg].value = ref->value;
265 c->pass2_grf[reg].nextuse = thisinsn;
266
267 ref->value->resident = &c->pass2_grf[reg];
268
269 /* Note that a fill is required:
270 */
271 ref->unspill_reg = reg*2;
272 }
273
274 /* Adjust the hw_reg to point at the value's current location:
275 */
276 assert(ref->value == ref->value->resident->value);
277 ref->hw_reg.nr += (ref->value->resident - c->pass2_grf) * 2;
278 }
279 }
280 }
281 }
282
283
284
285 /* Step 3: Work forwards once again. Perform register allocations,
286 * taking into account instructions like TEX which require contiguous
287 * result registers. Where necessary spill registers to scratch space
288 * and reload later.
289 */
290 void brw_wm_pass2( struct brw_wm_compile *c )
291 {
292 GLuint insn;
293 GLuint i;
294
295 init_registers(c);
296
297 for (insn = 0; insn < c->nr_insns; insn++) {
298 struct brw_wm_instruction *inst = &c->instruction[insn];
299
300 /* Update registers' nextuse values:
301 */
302 update_register_usage(c, insn);
303
304 /* May need to unspill some args.
305 */
306 load_args(c, inst);
307
308 /* Allocate registers to hold results:
309 */
310 switch (inst->opcode) {
311 case OPCODE_TEX:
312 case OPCODE_TXB:
313 case OPCODE_TXP:
314 alloc_contiguous_dest(c, inst->dst, 4, insn);
315 break;
316
317 default:
318 for (i = 0; i < 4; i++) {
319 if (inst->writemask & (1<<i)) {
320 assert(inst->dst[i]);
321 alloc_contiguous_dest(c, &inst->dst[i], 1, insn);
322 }
323 }
324 break;
325 }
326
327 if (TEST_DST_SPILLS && inst->opcode != WM_PIXELXY) {
328 for (i = 0; i < 4; i++)
329 if (inst->dst[i])
330 spill_value(c, inst->dst[i]);
331 }
332 }
333
334 if (unlikely(INTEL_DEBUG & DEBUG_WM)) {
335 brw_wm_print_program(c, "pass2");
336 }
337
338 c->state = PASS2_DONE;
339
340 if (unlikely(INTEL_DEBUG & DEBUG_WM)) {
341 brw_wm_print_program(c, "pass2/done");
342 }
343 }