i965/vs: Fix access of attribute arrays.
[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 struct brw_context *brw = c->func.brw;
73 struct intel_context *intel = &brw->intel;
74 GLuint nr_interp_regs = 0;
75 GLuint i = 0;
76 GLuint j;
77
78 for (j = 0; j < c->grf_limit; j++)
79 c->pass2_grf[j].nextuse = BRW_WM_MAX_INSN;
80
81 for (j = 0; j < (c->nr_payload_regs + 1) / 2; j++)
82 prealloc_reg(c, &c->payload.depth[j], i++);
83
84 for (j = 0; j < c->nr_creg; j++)
85 prealloc_reg(c, &c->creg[j], i++);
86
87 if (intel->gen >= 6) {
88 for (unsigned int j = 0; j < FRAG_ATTRIB_MAX; j++) {
89 if (brw->fragment_program->Base.InputsRead & BITFIELD64_BIT(j)) {
90 nr_interp_regs++;
91 prealloc_reg(c, &c->payload.input_interp[j], i++);
92 }
93 }
94 } else {
95 for (j = 0; j < VERT_RESULT_MAX; j++) {
96 if (c->key.vp_outputs_written & BITFIELD64_BIT(j)) {
97 int fp_index;
98
99 if (j >= VERT_RESULT_VAR0)
100 fp_index = j - (VERT_RESULT_VAR0 - FRAG_ATTRIB_VAR0);
101 else if (j <= VERT_RESULT_TEX7)
102 fp_index = j;
103 else
104 fp_index = -1;
105
106 nr_interp_regs++;
107 if (fp_index >= 0)
108 prealloc_reg(c, &c->payload.input_interp[fp_index], i++);
109 }
110 }
111 assert(nr_interp_regs >= 1);
112 }
113
114
115 c->prog_data.first_curbe_grf = ALIGN(c->nr_payload_regs, 2);
116 c->prog_data.urb_read_length = nr_interp_regs * 2;
117 c->prog_data.curb_read_length = c->nr_creg * 2;
118
119 c->max_wm_grf = i * 2;
120 }
121
122
123 /* Update the nextuse value for each register in our file.
124 */
125 static void update_register_usage(struct brw_wm_compile *c,
126 GLuint thisinsn)
127 {
128 GLuint i;
129
130 for (i = 1; i < c->grf_limit; i++) {
131 struct brw_wm_grf *grf = &c->pass2_grf[i];
132
133 /* Only search those which can change:
134 */
135 if (grf->nextuse < thisinsn) {
136 const struct brw_wm_ref *ref = grf->value->lastuse;
137
138 /* Has last use of value been passed?
139 */
140 if (ref->insn < thisinsn) {
141 grf->value->resident = 0;
142 grf->value = 0;
143 grf->nextuse = BRW_WM_MAX_INSN;
144 }
145 else {
146 /* Else loop through chain to update:
147 */
148 while (ref->prevuse && ref->prevuse->insn >= thisinsn)
149 ref = ref->prevuse;
150
151 grf->nextuse = ref->insn;
152 }
153 }
154 }
155 }
156
157
158 static void spill_value(struct brw_wm_compile *c,
159 struct brw_wm_value *value)
160 {
161 /* Allocate a spill slot. Note that allocations start from 0x40 -
162 * the first slot is reserved to mean "undef" in brw_wm_emit.c
163 */
164 if (!value->spill_slot) {
165 c->last_scratch += 0x40;
166 value->spill_slot = c->last_scratch;
167 }
168
169 /* The spill will be done in brw_wm_emit.c immediately after the
170 * value is calculated, so we can just take this reg without any
171 * further work.
172 */
173 value->resident->value = NULL;
174 value->resident->nextuse = BRW_WM_MAX_INSN;
175 value->resident = NULL;
176 }
177
178
179
180 /* Search for contiguous region with the most distant nearest
181 * member. Free regs count as very distant.
182 *
183 * TODO: implement spill-to-reg so that we can rearrange discontigous
184 * free regs and then spill the oldest non-free regs in sequence.
185 * This would mean inserting instructions in this pass.
186 */
187 static GLuint search_contiguous_regs(struct brw_wm_compile *c,
188 GLuint nr,
189 GLuint thisinsn)
190 {
191 struct brw_wm_grf *grf = c->pass2_grf;
192 GLuint furthest = 0;
193 GLuint reg = 0;
194 GLuint i, j;
195
196 /* Start search at 1: r0 is special and can't be used or spilled.
197 */
198 for (i = 1; i < c->grf_limit && furthest < BRW_WM_MAX_INSN; i++) {
199 GLuint group_nextuse = BRW_WM_MAX_INSN;
200
201 for (j = 0; j < nr; j++) {
202 if (grf[i+j].nextuse < group_nextuse)
203 group_nextuse = grf[i+j].nextuse;
204 }
205
206 if (group_nextuse > furthest) {
207 furthest = group_nextuse;
208 reg = i;
209 }
210 }
211
212 assert(furthest != thisinsn);
213
214 /* Any non-empty regs will need to be spilled:
215 */
216 for (j = 0; j < nr; j++)
217 if (grf[reg+j].value)
218 spill_value(c, grf[reg+j].value);
219
220 return reg;
221 }
222
223
224 static void alloc_contiguous_dest(struct brw_wm_compile *c,
225 struct brw_wm_value *dst[],
226 GLuint nr,
227 GLuint thisinsn)
228 {
229 GLuint reg = search_contiguous_regs(c, nr, thisinsn);
230 GLuint i;
231
232 for (i = 0; i < nr; i++) {
233 if (!dst[i]) {
234 /* Need to grab a dummy value in TEX case. Don't introduce
235 * it into the tracking scheme.
236 */
237 dst[i] = &c->vreg[c->nr_vreg++];
238 }
239 else {
240 assert(!dst[i]->resident);
241 assert(c->pass2_grf[reg+i].nextuse != thisinsn);
242
243 c->pass2_grf[reg+i].value = dst[i];
244 c->pass2_grf[reg+i].nextuse = thisinsn;
245
246 dst[i]->resident = &c->pass2_grf[reg+i];
247 }
248
249 dst[i]->hw_reg = brw_vec8_grf((reg+i)*2, 0);
250 }
251
252 if ((reg+nr)*2 > c->max_wm_grf)
253 c->max_wm_grf = (reg+nr) * 2;
254 }
255
256
257 static void load_args(struct brw_wm_compile *c,
258 struct brw_wm_instruction *inst)
259 {
260 GLuint thisinsn = inst - c->instruction;
261 GLuint i,j;
262
263 for (i = 0; i < 3; i++) {
264 for (j = 0; j < 4; j++) {
265 struct brw_wm_ref *ref = inst->src[i][j];
266
267 if (ref) {
268 if (!ref->value->resident) {
269 /* Need to bring the value in from scratch space. The code for
270 * this will be done in brw_wm_emit.c, here we just do the
271 * register allocation and mark the ref as requiring a fill.
272 */
273 GLuint reg = search_contiguous_regs(c, 1, thisinsn);
274
275 c->pass2_grf[reg].value = ref->value;
276 c->pass2_grf[reg].nextuse = thisinsn;
277
278 ref->value->resident = &c->pass2_grf[reg];
279
280 /* Note that a fill is required:
281 */
282 ref->unspill_reg = reg*2;
283 }
284
285 /* Adjust the hw_reg to point at the value's current location:
286 */
287 assert(ref->value == ref->value->resident->value);
288 ref->hw_reg.nr += (ref->value->resident - c->pass2_grf) * 2;
289 }
290 }
291 }
292 }
293
294
295
296 /* Step 3: Work forwards once again. Perform register allocations,
297 * taking into account instructions like TEX which require contiguous
298 * result registers. Where necessary spill registers to scratch space
299 * and reload later.
300 */
301 void brw_wm_pass2( struct brw_wm_compile *c )
302 {
303 GLuint insn;
304 GLuint i;
305
306 init_registers(c);
307
308 for (insn = 0; insn < c->nr_insns; insn++) {
309 struct brw_wm_instruction *inst = &c->instruction[insn];
310
311 /* Update registers' nextuse values:
312 */
313 update_register_usage(c, insn);
314
315 /* May need to unspill some args.
316 */
317 load_args(c, inst);
318
319 /* Allocate registers to hold results:
320 */
321 switch (inst->opcode) {
322 case OPCODE_TEX:
323 case OPCODE_TXB:
324 case OPCODE_TXP:
325 alloc_contiguous_dest(c, inst->dst, 4, insn);
326 break;
327
328 default:
329 for (i = 0; i < 4; i++) {
330 if (inst->writemask & (1<<i)) {
331 assert(inst->dst[i]);
332 alloc_contiguous_dest(c, &inst->dst[i], 1, insn);
333 }
334 }
335 break;
336 }
337
338 if (TEST_DST_SPILLS && inst->opcode != WM_PIXELXY) {
339 for (i = 0; i < 4; i++)
340 if (inst->dst[i])
341 spill_value(c, inst->dst[i]);
342 }
343 }
344
345 if (unlikely(INTEL_DEBUG & DEBUG_WM)) {
346 brw_wm_print_program(c, "pass2");
347 }
348
349 c->state = PASS2_DONE;
350
351 if (unlikely(INTEL_DEBUG & DEBUG_WM)) {
352 brw_wm_print_program(c, "pass2/done");
353 }
354 }