projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
i965/gen8: add debug code to show FS disasm with jump locations
[mesa.git] / src / mesa / drivers / dri / i965 / gen8_fs_generator.cpp
1 /*
2 * Copyright © 2010, 2011, 2012 Intel Corporation
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21 * IN THE SOFTWARE.
22 */
23
24 /** @file gen8_fs_generate.cpp
25 *
26 * Code generation for Gen8+ hardware.
27 */
28
29 extern "C" {
30 #include "main/macros.h"
31 #include "brw_context.h"
32 } /* extern "C" */
33
34 #include "brw_fs.h"
35 #include "brw_cfg.h"
36 #include "glsl/ir_print_visitor.h"
37
38 gen8_fs_generator::gen8_fs_generator(struct brw_context *brw,
39 struct brw_wm_compile *c,
40 struct gl_shader_program *shader_prog,
41 struct gl_fragment_program *fp,
42 bool dual_source_output)
43 : gen8_generator(brw, shader_prog, fp ? &fp->Base : NULL, c), c(c), fp(fp),
44 dual_source_output(dual_source_output)
45 {
46 }
47
48 gen8_fs_generator::~gen8_fs_generator()
49 {
50 }
51
52 void
53 gen8_fs_generator::mark_surface_used(unsigned surf_index)
54 {
55 assert(surf_index < BRW_MAX_SURFACES);
56
57 c->prog_data.base.binding_table.size_bytes =
58 MAX2(c->prog_data.base.binding_table.size_bytes, (surf_index + 1) * 4);
59 }
60
61 void
62 gen8_fs_generator::generate_fb_write(fs_inst *ir)
63 {
64 /* Disable the discard condition while setting up the header. */
65 default_state.predicate = BRW_PREDICATE_NONE;
66 default_state.predicate_inverse = false;
67 default_state.flag_subreg_nr = 0;
68
69 if (ir->header_present) {
70 /* The GPU will use the predicate on SENDC, unless the header is present.
71 */
72 if (fp && fp->UsesKill) {
73 gen8_instruction *mov =
74 MOV(retype(brw_vec1_grf(1, 7), BRW_REGISTER_TYPE_UW),
75 brw_flag_reg(0, 1));
76 gen8_set_mask_control(mov, BRW_MASK_DISABLE);
77 }
78
79 gen8_instruction *mov =
80 MOV_RAW(brw_message_reg(ir->base_mrf), brw_vec8_grf(0, 0));
81 gen8_set_exec_size(mov, BRW_EXECUTE_16);
82
83 if (ir->target > 0 && c->key.replicate_alpha) {
84 /* Set "Source0 Alpha Present to RenderTarget" bit in the header. */
85 OR(vec1(retype(brw_message_reg(ir->base_mrf), BRW_REGISTER_TYPE_UD)),
86 vec1(retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD)),
87 brw_imm_ud(1 << 11));
88 }
89
90 if (ir->target > 0) {
91 /* Set the render target index for choosing BLEND_STATE. */
92 MOV(retype(brw_vec1_reg(BRW_MESSAGE_REGISTER_FILE, ir->base_mrf, 2),
93 BRW_REGISTER_TYPE_UD),
94 brw_imm_ud(ir->target));
95 }
96 }
97
98 /* Set the predicate back to get the conditional write if necessary for
99 * discards.
100 */
101 default_state.predicate = ir->predicate;
102 default_state.predicate_inverse = ir->predicate_inverse;
103 default_state.flag_subreg_nr = ir->flag_subreg;
104
105 gen8_instruction *inst = next_inst(BRW_OPCODE_SENDC);
106 gen8_set_dst(brw, inst, retype(vec8(brw_null_reg()), BRW_REGISTER_TYPE_UW));
107 gen8_set_src0(brw, inst, brw_message_reg(ir->base_mrf));
108
109 /* Set up the "Message Specific Control" bits for the Data Port Message
110 * Descriptor. These are documented in the "Render Target Write" message's
111 * "Message Descriptor" documentation (vol5c.2).
112 */
113 uint32_t msg_type;
114 /* Set the Message Type */
115 if (this->dual_source_output)
116 msg_type = BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD8_DUAL_SOURCE_SUBSPAN01;
117 else if (dispatch_width == 16)
118 msg_type = BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD16_SINGLE_SOURCE;
119 else
120 msg_type = BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD8_SINGLE_SOURCE_SUBSPAN01;
121
122 uint32_t msg_control = msg_type;
123
124 /* "Last Render Target Select" must be set on all writes to the last of
125 * the render targets (if using MRT), or always for a single RT scenario.
126 */
127 if ((ir->target == c->key.nr_color_regions - 1) || !c->key.nr_color_regions)
128 msg_control |= (1 << 4); /* Last Render Target Select */
129
130 uint32_t surf_index =
131 c->prog_data.binding_table.render_target_start + ir->target;
132
133 gen8_set_dp_message(brw, inst,
134 GEN6_SFID_DATAPORT_RENDER_CACHE,
135 surf_index,
136 GEN6_DATAPORT_WRITE_MESSAGE_RENDER_TARGET_WRITE,
137 msg_control,
138 ir->mlen,
139 0,
140 ir->header_present,
141 ir->eot);
142
143 mark_surface_used(surf_index);
144 }
145
146 void
147 gen8_fs_generator::generate_linterp(fs_inst *inst,
148 struct brw_reg dst,
149 struct brw_reg *src)
150 {
151 struct brw_reg delta_x = src[0];
152 struct brw_reg delta_y = src[1];
153 struct brw_reg interp = src[2];
154
155 (void) delta_y;
156 assert(delta_y.nr == delta_x.nr + 1);
157 PLN(dst, interp, delta_x);
158 }
159
160 void
161 gen8_fs_generator::generate_tex(fs_inst *ir,
162 struct brw_reg dst,
163 struct brw_reg src)
164 {
165 int msg_type = -1;
166 int rlen = 4;
167 uint32_t simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD8;
168
169 assert(src.file == BRW_GENERAL_REGISTER_FILE);
170
171 if (dispatch_width == 16 && !ir->force_uncompressed && !ir->force_sechalf)
172 simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD16;
173
174 switch (ir->opcode) {
175 case SHADER_OPCODE_TEX:
176 if (ir->shadow_compare) {
177 msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_COMPARE;
178 } else {
179 msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE;
180 }
181 break;
182 case FS_OPCODE_TXB:
183 if (ir->shadow_compare) {
184 msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_BIAS_COMPARE;
185 } else {
186 msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_BIAS;
187 }
188 break;
189 case SHADER_OPCODE_TXL:
190 if (ir->shadow_compare) {
191 msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_LOD_COMPARE;
192 } else {
193 msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_LOD;
194 }
195 break;
196 case SHADER_OPCODE_TXS:
197 msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_RESINFO;
198 break;
199 case SHADER_OPCODE_TXD:
200 if (ir->shadow_compare) {
201 msg_type = HSW_SAMPLER_MESSAGE_SAMPLE_DERIV_COMPARE;
202 } else {
203 msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_DERIVS;
204 }
205 break;
206 case SHADER_OPCODE_TXF:
207 msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_LD;
208 break;
209 case SHADER_OPCODE_TXF_CMS:
210 msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_LD2DMS;
211 break;
212 case SHADER_OPCODE_TXF_UMS:
213 msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_LD2DSS;
214 break;
215 case SHADER_OPCODE_TXF_MCS:
216 msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_LD_MCS;
217 break;
218 case SHADER_OPCODE_LOD:
219 msg_type = GEN5_SAMPLER_MESSAGE_LOD;
220 break;
221 case SHADER_OPCODE_TG4:
222 if (ir->shadow_compare) {
223 assert(brw->gen >= 7);
224 msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4_C;
225 } else {
226 assert(brw->gen >= 6);
227 msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4;
228 }
229 break;
230 case SHADER_OPCODE_TG4_OFFSET:
231 assert(brw->gen >= 7);
232 if (ir->shadow_compare) {
233 msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4_PO_C;
234 } else {
235 msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4_PO;
236 }
237 break;
238 default:
239 assert(!"not reached");
240 break;
241 }
242 assert(msg_type != -1);
243
244 if (simd_mode == BRW_SAMPLER_SIMD_MODE_SIMD16) {
245 rlen = 8;
246 dst = vec16(dst);
247 }
248
249 if (ir->header_present) {
250 /* The send-from-GRF for SIMD16 texturing with a header has an extra
251 * hardware register allocated to it, which we need to skip over (since
252 * our coordinates in the payload are in the even-numbered registers,
253 * and the header comes right before the first one.
254 */
255 if (dispatch_width == 16)
256 src.nr++;
257
258 unsigned save_exec_size = default_state.exec_size;
259 default_state.exec_size = BRW_EXECUTE_8;
260
261 MOV_RAW(src, brw_vec8_grf(0, 0));
262
263 if (ir->texture_offset) {
264 /* Set the texel offset bits. */
265 MOV_RAW(retype(brw_vec1_grf(src.nr, 2), BRW_REGISTER_TYPE_UD),
266 brw_imm_ud(ir->texture_offset));
267 }
268
269 if (ir->sampler >= 16) {
270 /* The "Sampler Index" field can only store values between 0 and 15.
271 * However, we can add an offset to the "Sampler State Pointer"
272 * field, effectively selecting a different set of 16 samplers.
273 *
274 * The "Sampler State Pointer" needs to be aligned to a 32-byte
275 * offset, and each sampler state is only 16-bytes, so we can't
276 * exclusively use the offset - we have to use both.
277 */
278 gen8_instruction *add =
279 ADD(get_element_ud(src, 3),
280 get_element_ud(brw_vec8_grf(0, 0), 3),
281 brw_imm_ud(16 * (ir->sampler / 16) *
282 sizeof(gen7_sampler_state)));
283 gen8_set_mask_control(add, BRW_MASK_DISABLE);
284 }
285
286 default_state.exec_size = save_exec_size;
287 }
288
289 uint32_t surf_index =
290 c->prog_data.base.binding_table.texture_start + ir->sampler;
291
292 gen8_instruction *inst = next_inst(BRW_OPCODE_SEND);
293 gen8_set_dst(brw, inst, dst);
294 gen8_set_src0(brw, inst, src);
295 gen8_set_sampler_message(brw, inst,
296 surf_index,
297 ir->sampler % 16,
298 msg_type,
299 rlen,
300 ir->mlen,
301 ir->header_present,
302 simd_mode);
303
304 mark_surface_used(surf_index);
305 }
306
307
308 /* For OPCODE_DDX and OPCODE_DDY, per channel of output we've got input
309 * looking like:
310 *
311 * arg0: ss0.tl ss0.tr ss0.bl ss0.br ss1.tl ss1.tr ss1.bl ss1.br
312 *
313 * and we're trying to produce:
314 *
315 * DDX DDY
316 * dst: (ss0.tr - ss0.tl) (ss0.tl - ss0.bl)
317 * (ss0.tr - ss0.tl) (ss0.tr - ss0.br)
318 * (ss0.br - ss0.bl) (ss0.tl - ss0.bl)
319 * (ss0.br - ss0.bl) (ss0.tr - ss0.br)
320 * (ss1.tr - ss1.tl) (ss1.tl - ss1.bl)
321 * (ss1.tr - ss1.tl) (ss1.tr - ss1.br)
322 * (ss1.br - ss1.bl) (ss1.tl - ss1.bl)
323 * (ss1.br - ss1.bl) (ss1.tr - ss1.br)
324 *
325 * and add another set of two more subspans if in 16-pixel dispatch mode.
326 *
327 * For DDX, it ends up being easy: width = 2, horiz=0 gets us the same result
328 * for each pair, and vertstride = 2 jumps us 2 elements after processing a
329 * pair. But for DDY, it's harder, as we want to produce the pairs swizzled
330 * between each other. We could probably do it like ddx and swizzle the right
331 * order later, but bail for now and just produce
332 * ((ss0.tl - ss0.bl)x4 (ss1.tl - ss1.bl)x4)
333 */
334 void
335 gen8_fs_generator::generate_ddx(fs_inst *inst,
336 struct brw_reg dst,
337 struct brw_reg src)
338 {
339 unsigned vstride, width;
340
341 if (c->key.high_quality_derivatives) {
342 /* Produce accurate derivatives. */
343 vstride = BRW_VERTICAL_STRIDE_2;
344 width = BRW_WIDTH_2;
345 } else {
346 /* Replicate the derivative at the top-left pixel to other pixels. */
347 vstride = BRW_VERTICAL_STRIDE_4;
348 width = BRW_WIDTH_4;
349 }
350
351 struct brw_reg src0 = brw_reg(src.file, src.nr, 1,
352 BRW_REGISTER_TYPE_F,
353 vstride,
354 width,
355 BRW_HORIZONTAL_STRIDE_0,
356 BRW_SWIZZLE_XYZW, WRITEMASK_XYZW);
357 struct brw_reg src1 = brw_reg(src.file, src.nr, 0,
358 BRW_REGISTER_TYPE_F,
359 vstride,
360 width,
361 BRW_HORIZONTAL_STRIDE_0,
362 BRW_SWIZZLE_XYZW, WRITEMASK_XYZW);
363 ADD(dst, src0, negate(src1));
364 }
365
366 /* The negate_value boolean is used to negate the derivative computation for
367 * FBOs, since they place the origin at the upper left instead of the lower
368 * left.
369 */
370 void
371 gen8_fs_generator::generate_ddy(fs_inst *inst,
372 struct brw_reg dst,
373 struct brw_reg src,
374 bool negate_value)
375 {
376 unsigned hstride;
377 unsigned src0_swizzle;
378 unsigned src1_swizzle;
379 unsigned src1_subnr;
380
381 if (c->key.high_quality_derivatives) {
382 /* Produce accurate derivatives. */
383 hstride = BRW_HORIZONTAL_STRIDE_1;
384 src0_swizzle = BRW_SWIZZLE_XYXY;
385 src1_swizzle = BRW_SWIZZLE_ZWZW;
386 src1_subnr = 0;
387
388 default_state.access_mode = BRW_ALIGN_16;
389 } else {
390 /* Replicate the derivative at the top-left pixel to other pixels. */
391 hstride = BRW_HORIZONTAL_STRIDE_0;
392 src0_swizzle = BRW_SWIZZLE_XYZW;
393 src1_swizzle = BRW_SWIZZLE_XYZW;
394 src1_subnr = 2;
395 }
396
397 struct brw_reg src0 = brw_reg(src.file, src.nr, 0,
398 BRW_REGISTER_TYPE_F,
399 BRW_VERTICAL_STRIDE_4,
400 BRW_WIDTH_4,
401 hstride,
402 src0_swizzle, WRITEMASK_XYZW);
403 struct brw_reg src1 = brw_reg(src.file, src.nr, src1_subnr,
404 BRW_REGISTER_TYPE_F,
405 BRW_VERTICAL_STRIDE_4,
406 BRW_WIDTH_4,
407 hstride,
408 src1_swizzle, WRITEMASK_XYZW);
409
410 if (negate_value)
411 ADD(dst, src1, negate(src0));
412 else
413 ADD(dst, src0, negate(src1));
414
415 default_state.access_mode = BRW_ALIGN_1;
416 }
417
418 void
419 gen8_fs_generator::generate_scratch_write(fs_inst *ir, struct brw_reg src)
420 {
421 MOV(retype(brw_message_reg(ir->base_mrf + 1), BRW_REGISTER_TYPE_UD),
422 retype(src, BRW_REGISTER_TYPE_UD));
423
424 struct brw_reg mrf =
425 retype(brw_message_reg(ir->base_mrf), BRW_REGISTER_TYPE_UD);
426
427 const int num_regs = dispatch_width / 8;
428
429 uint32_t msg_control;
430 if (num_regs == 1)
431 msg_control = BRW_DATAPORT_OWORD_BLOCK_2_OWORDS;
432 else
433 msg_control = BRW_DATAPORT_OWORD_BLOCK_4_OWORDS;
434
435 /* Set up the message header. This is g0, with g0.2 filled with
436 * the offset. We don't want to leave our offset around in g0 or
437 * it'll screw up texture samples, so set it up inside the message
438 * reg.
439 */
440 unsigned save_exec_size = default_state.exec_size;
441 default_state.exec_size = BRW_EXECUTE_8;
442
443 MOV_RAW(mrf, retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD));
444 /* set message header global offset field (reg 0, element 2) */
445 MOV_RAW(get_element_ud(mrf, 2), brw_imm_ud(ir->offset / 16));
446
447 struct brw_reg dst;
448 if (dispatch_width == 16)
449 dst = retype(vec16(brw_null_reg()), BRW_REGISTER_TYPE_UW);
450 else
451 dst = retype(vec8(brw_null_reg()), BRW_REGISTER_TYPE_UW);
452
453 default_state.exec_size = BRW_EXECUTE_16;
454
455 gen8_instruction *send = next_inst(BRW_OPCODE_SEND);
456 gen8_set_dst(brw, send, dst);
457 gen8_set_src0(brw, send, mrf);
458 gen8_set_dp_message(brw, send, GEN7_SFID_DATAPORT_DATA_CACHE,
459 255, /* binding table index: stateless access */
460 GEN6_DATAPORT_WRITE_MESSAGE_OWORD_BLOCK_WRITE,
461 msg_control,
462 1 + num_regs, /* mlen */
463 0, /* rlen */
464 true, /* header present */
465 false); /* EOT */
466
467 default_state.exec_size = save_exec_size;
468 }
469
470 void
471 gen8_fs_generator::generate_scratch_read(fs_inst *ir, struct brw_reg dst)
472 {
473 struct brw_reg mrf =
474 retype(brw_message_reg(ir->base_mrf), BRW_REGISTER_TYPE_UD);
475
476 const int num_regs = dispatch_width / 8;
477
478 uint32_t msg_control;
479 if (num_regs == 1)
480 msg_control = BRW_DATAPORT_OWORD_BLOCK_2_OWORDS;
481 else
482 msg_control = BRW_DATAPORT_OWORD_BLOCK_4_OWORDS;
483
484 unsigned save_exec_size = default_state.exec_size;
485 default_state.exec_size = BRW_EXECUTE_8;
486
487 MOV_RAW(mrf, retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD));
488 /* set message header global offset field (reg 0, element 2) */
489 MOV_RAW(get_element_ud(mrf, 2), brw_imm_ud(ir->offset / 16));
490
491 gen8_instruction *send = next_inst(BRW_OPCODE_SEND);
492 gen8_set_dst(brw, send, retype(dst, BRW_REGISTER_TYPE_UW));
493 gen8_set_src0(brw, send, mrf);
494 gen8_set_dp_message(brw, send, GEN7_SFID_DATAPORT_DATA_CACHE,
495 255, /* binding table index: stateless access */
496 BRW_DATAPORT_READ_MESSAGE_OWORD_BLOCK_READ,
497 msg_control,
498 1, /* mlen */
499 num_regs, /* rlen */
500 true, /* header present */
501 false); /* EOT */
502
503 default_state.exec_size = save_exec_size;
504 }
505
506 void
507 gen8_fs_generator::generate_scratch_read_gen7(fs_inst *ir, struct brw_reg dst)
508 {
509 unsigned save_exec_size = default_state.exec_size;
510 gen8_instruction *send = next_inst(BRW_OPCODE_SEND);
511
512 int num_regs = dispatch_width / 8;
513
514 /* According to the docs, offset is "A 12-bit HWord offset into the memory
515 * Immediate Memory buffer as specified by binding table 0xFF." An HWORD
516 * is 32 bytes, which happens to be the size of a register.
517 */
518 int offset = ir->offset / REG_SIZE;
519
520 /* The HW requires that the header is present; this is to get the g0.5
521 * scratch offset.
522 */
523 gen8_set_src0(brw, send, brw_vec8_grf(0, 0));
524 gen8_set_dst(brw, send, retype(dst, BRW_REGISTER_TYPE_UW));
525 gen8_set_dp_scratch_message(brw, send,
526 false, /* scratch read */
527 false, /* OWords */
528 false, /* invalidate after read */
529 num_regs,
530 offset,
531 1, /* mlen - just g0 */
532 num_regs, /* rlen */
533 true, /* header present */
534 false); /* EOT */
535
536 default_state.exec_size = save_exec_size;
537 }
538
539 void
540 gen8_fs_generator::generate_uniform_pull_constant_load(fs_inst *inst,
541 struct brw_reg dst,
542 struct brw_reg index,
543 struct brw_reg offset)
544 {
545 assert(inst->mlen == 0);
546
547 assert(index.file == BRW_IMMEDIATE_VALUE &&
548 index.type == BRW_REGISTER_TYPE_UD);
549 uint32_t surf_index = index.dw1.ud;
550
551 assert(offset.file == BRW_GENERAL_REGISTER_FILE);
552 /* Reference only the dword we need lest we anger validate_reg() with
553 * reg.width > reg.execszie.
554 */
555 offset = brw_vec1_grf(offset.nr, 0);
556
557 gen8_instruction *send = next_inst(BRW_OPCODE_SEND);
558 gen8_set_mask_control(send, BRW_MASK_DISABLE);
559
560 /* We use the SIMD4x2 mode because we want to end up with 4 constants in
561 * the destination loaded consecutively from the same offset (which appears
562 * in the first component, and the rest are ignored).
563 */
564 dst.width = BRW_WIDTH_4;
565 gen8_set_dst(brw, send, dst);
566 gen8_set_src0(brw, send, offset);
567 gen8_set_sampler_message(brw, send,
568 surf_index,
569 0, /* The LD message ignores the sampler unit. */
570 GEN5_SAMPLER_MESSAGE_SAMPLE_LD,
571 1, /* rlen */
572 1, /* mlen */
573 false, /* no header */
574 BRW_SAMPLER_SIMD_MODE_SIMD4X2);
575
576 mark_surface_used(surf_index);
577 }
578
579 void
580 gen8_fs_generator::generate_varying_pull_constant_load(fs_inst *ir,
581 struct brw_reg dst,
582 struct brw_reg index,
583 struct brw_reg offset)
584 {
585 /* Varying-offset pull constant loads are treated as a normal expression on
586 * gen7, so the fact that it's a send message is hidden at the IR level.
587 */
588 assert(!ir->header_present);
589 assert(!ir->mlen);
590
591 assert(index.file == BRW_IMMEDIATE_VALUE &&
592 index.type == BRW_REGISTER_TYPE_UD);
593 uint32_t surf_index = index.dw1.ud;
594
595 uint32_t simd_mode, rlen, mlen;
596 if (dispatch_width == 16) {
597 mlen = 2;
598 rlen = 8;
599 simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD16;
600 } else {
601 mlen = 1;
602 rlen = 4;
603 simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD8;
604 }
605
606 gen8_instruction *send = next_inst(BRW_OPCODE_SEND);
607 gen8_set_dst(brw, send, dst);
608 gen8_set_src0(brw, send, offset);
609 gen8_set_sampler_message(brw, send,
610 surf_index,
611 0, /* The LD message ignore the sampler unit. */
612 GEN5_SAMPLER_MESSAGE_SAMPLE_LD,
613 rlen, /* rlen */
614 mlen, /* mlen */
615 false, /* no header */
616 simd_mode);
617
618 mark_surface_used(surf_index);
619 }
620
621 /**
622 * Cause the current pixel/sample mask (from R1.7 bits 15:0) to be transferred
623 * into the flags register (f0.0).
624 */
625 void
626 gen8_fs_generator::generate_mov_dispatch_to_flags(fs_inst *ir)
627 {
628 struct brw_reg flags = brw_flag_reg(0, ir->flag_subreg);
629 struct brw_reg dispatch_mask =
630 retype(brw_vec1_grf(1, 7), BRW_REGISTER_TYPE_UW);
631
632 gen8_instruction *mov = MOV(flags, dispatch_mask);
633 gen8_set_mask_control(mov, BRW_MASK_DISABLE);
634 }
635
636 void
637 gen8_fs_generator::generate_discard_jump(fs_inst *ir)
638 {
639 /* This HALT will be patched up at FB write time to point UIP at the end of
640 * the program, and at brw_uip_jip() JIP will be set to the end of the
641 * current block (or the program).
642 */
643 discard_halt_patches.push_tail(new(mem_ctx) ip_record(nr_inst));
644
645 HALT();
646 }
647
648 void
649 gen8_fs_generator::patch_discard_jumps_to_fb_writes()
650 {
651 if (discard_halt_patches.is_empty())
652 return;
653
654 /* There is a somewhat strange undocumented requirement of using
655 * HALT, according to the simulator. If some channel has HALTed to
656 * a particular UIP, then by the end of the program, every channel
657 * must have HALTed to that UIP. Furthermore, the tracking is a
658 * stack, so you can't do the final halt of a UIP after starting
659 * halting to a new UIP.
660 *
661 * Symptoms of not emitting this instruction on actual hardware
662 * included GPU hangs and sparkly rendering on the piglit discard
663 * tests.
664 */
665 gen8_instruction *last_halt = HALT();
666 gen8_set_uip(last_halt, 16);
667 gen8_set_jip(last_halt, 16);
668
669 int ip = nr_inst;
670
671 foreach_list(node, &discard_halt_patches) {
672 ip_record *patch_ip = (ip_record *) node;
673 gen8_instruction *patch = &store[patch_ip->ip];
674 assert(gen8_opcode(patch) == BRW_OPCODE_HALT);
675
676 /* HALT takes an instruction distance from the pre-incremented IP. */
677 gen8_set_uip(patch, (ip - patch_ip->ip) * 16);
678 }
679
680 this->discard_halt_patches.make_empty();
681 }
682
683 /**
684 * Sets the first dword of a vgrf for simd4x2 uniform pull constant
685 * sampler LD messages.
686 *
687 * We don't want to bake it into the send message's code generation because
688 * that means we don't get a chance to schedule the instruction.
689 */
690 void
691 gen8_fs_generator::generate_set_simd4x2_offset(fs_inst *ir,
692 struct brw_reg dst,
693 struct brw_reg value)
694 {
695 assert(value.file == BRW_IMMEDIATE_VALUE);
696 MOV_RAW(retype(brw_vec1_reg(dst.file, dst.nr, 0), value.type), value);
697 }
698
699 /**
700 * Sets vstride=16, width=8, hstride=2 or vstride=0, width=1, hstride=0
701 * (when mask is passed as a uniform) of register mask before moving it
702 * to register dst.
703 */
704 void
705 gen8_fs_generator::generate_set_omask(fs_inst *inst,
706 struct brw_reg dst,
707 struct brw_reg mask)
708 {
709 assert(dst.type == BRW_REGISTER_TYPE_UW);
710
711 if (dispatch_width == 16)
712 dst = vec16(dst);
713
714 if (mask.vstride == BRW_VERTICAL_STRIDE_8 &&
715 mask.width == BRW_WIDTH_8 &&
716 mask.hstride == BRW_HORIZONTAL_STRIDE_1) {
717 mask = stride(mask, 16, 8, 2);
718 } else {
719 assert(mask.vstride == BRW_VERTICAL_STRIDE_0 &&
720 mask.width == BRW_WIDTH_1 &&
721 mask.hstride == BRW_HORIZONTAL_STRIDE_0);
722 }
723
724 unsigned save_exec_size = default_state.exec_size;
725 default_state.exec_size = BRW_EXECUTE_8;
726
727 gen8_instruction *mov = MOV(dst, retype(mask, dst.type));
728 gen8_set_mask_control(mov, BRW_MASK_DISABLE);
729
730 default_state.exec_size = save_exec_size;
731 }
732
733 /**
734 * Do a special ADD with vstride=1, width=4, hstride=0 for src1.
735 */
736 void
737 gen8_fs_generator::generate_set_sample_id(fs_inst *ir,
738 struct brw_reg dst,
739 struct brw_reg src0,
740 struct brw_reg src1)
741 {
742 assert(dst.type == BRW_REGISTER_TYPE_D || dst.type == BRW_REGISTER_TYPE_UD);
743 assert(src0.type == BRW_REGISTER_TYPE_D || src0.type == BRW_REGISTER_TYPE_UD);
744
745 struct brw_reg reg = retype(stride(src1, 1, 4, 0), BRW_REGISTER_TYPE_UW);
746
747 unsigned save_exec_size = default_state.exec_size;
748 default_state.exec_size = BRW_EXECUTE_8;
749
750 gen8_instruction *add = ADD(dst, src0, reg);
751 gen8_set_mask_control(add, BRW_MASK_DISABLE);
752 if (dispatch_width == 16) {
753 add = ADD(offset(dst, 1), offset(src0, 1), suboffset(reg, 2));
754 gen8_set_mask_control(add, BRW_MASK_DISABLE);
755 }
756
757 default_state.exec_size = save_exec_size;
758 }
759
760 /**
761 * Change the register's data type from UD to HF, doubling the strides in order
762 * to compensate for halving the data type width.
763 */
764 static struct brw_reg
765 ud_reg_to_hf(struct brw_reg r)
766 {
767 assert(r.type == BRW_REGISTER_TYPE_UD);
768 r.type = BRW_REGISTER_TYPE_HF;
769
770 /* The BRW_*_STRIDE enums are defined so that incrementing the field
771 * doubles the real stride.
772 */
773 if (r.hstride != 0)
774 ++r.hstride;
775 if (r.vstride != 0)
776 ++r.vstride;
777
778 return r;
779 }
780
781 void
782 gen8_fs_generator::generate_pack_half_2x16_split(fs_inst *inst,
783 struct brw_reg dst,
784 struct brw_reg x,
785 struct brw_reg y)
786 {
787 assert(dst.type == BRW_REGISTER_TYPE_UD);
788 assert(x.type == BRW_REGISTER_TYPE_F);
789 assert(y.type == BRW_REGISTER_TYPE_F);
790
791 struct brw_reg dst_hf = ud_reg_to_hf(dst);
792
793 /* Give each 32-bit channel of dst the form below , where "." means
794 * unchanged.
795 * 0x....hhhh
796 */
797 MOV(dst_hf, y);
798
799 /* Now the form:
800 * 0xhhhh0000
801 */
802 SHL(dst, dst, brw_imm_ud(16u));
803
804 /* And, finally the form of packHalf2x16's output:
805 * 0xhhhhllll
806 */
807 MOV(dst_hf, x);
808 }
809
810 void
811 gen8_fs_generator::generate_unpack_half_2x16_split(fs_inst *inst,
812 struct brw_reg dst,
813 struct brw_reg src)
814 {
815 assert(dst.type == BRW_REGISTER_TYPE_F);
816 assert(src.type == BRW_REGISTER_TYPE_UD);
817
818 struct brw_reg src_hf = ud_reg_to_hf(src);
819
820 /* Each channel of src has the form of unpackHalf2x16's input: 0xhhhhllll.
821 * For the Y case, we wish to access only the upper word; therefore
822 * a 16-bit subregister offset is needed.
823 */
824 assert(inst->opcode == FS_OPCODE_UNPACK_HALF_2x16_SPLIT_X ||
825 inst->opcode == FS_OPCODE_UNPACK_HALF_2x16_SPLIT_Y);
826 if (inst->opcode == FS_OPCODE_UNPACK_HALF_2x16_SPLIT_Y)
827 src_hf.subnr += 2;
828
829 MOV(dst, src_hf);
830 }
831
832 void
833 gen8_fs_generator::generate_code(exec_list *instructions)
834 {
835 int last_native_inst_offset = next_inst_offset;
836 const char *last_annotation_string = NULL;
837 const void *last_annotation_ir = NULL;
838
839 if (unlikely(INTEL_DEBUG & DEBUG_WM)) {
840 if (prog) {
841 fprintf(stderr,
842 "Native code for %s fragment shader %d (SIMD%d dispatch):\n",
843 shader_prog->Label ? shader_prog->Label : "unnamed",
844 shader_prog->Name, dispatch_width);
845 } else if (fp) {
846 fprintf(stderr,
847 "Native code for fragment program %d (SIMD%d dispatch):\n",
848 prog->Id, dispatch_width);
849 } else {
850 fprintf(stderr, "Native code for blorp program (SIMD%d dispatch):\n",
851 dispatch_width);
852 }
853 }
854
855 cfg_t *cfg = NULL;
856 if (unlikely(INTEL_DEBUG & DEBUG_WM))
857 cfg = new(mem_ctx) cfg_t(instructions);
858
859 foreach_list(node, instructions) {
860 fs_inst *ir = (fs_inst *) node;
861 struct brw_reg src[3], dst;
862
863 if (unlikely(INTEL_DEBUG & DEBUG_WM)) {
864 foreach_list(node, &cfg->block_list) {
865 bblock_link *link = (bblock_link *)node;
866 bblock_t *block = link->block;
867
868 if (block->start == ir) {
869 fprintf(stderr, " START B%d", block->block_num);
870 foreach_list(predecessor_node, &block->parents) {
871 bblock_link *predecessor_link =
872 (bblock_link *)predecessor_node;
873 bblock_t *predecessor_block = predecessor_link->block;
874 fprintf(stderr, " <-B%d", predecessor_block->block_num);
875 }
876 fprintf(stderr, "\n");
877 }
878 }
879
880 if (last_annotation_ir != ir->ir) {
881 last_annotation_ir = ir->ir;
882 if (last_annotation_ir) {
883 fprintf(stderr, " ");
884 if (prog) {
885 ((ir_instruction *) ir->ir)->fprint(stderr);
886 } else if (prog) {
887 const prog_instruction *fpi;
888 fpi = (const prog_instruction *) ir->ir;
889 fprintf(stderr, "%d: ", (int)(fpi - prog->Instructions));
890 _mesa_fprint_instruction_opt(stderr,
891 fpi,
892 0, PROG_PRINT_DEBUG, NULL);
893 }
894 fprintf(stderr, "\n");
895 }
896 }
897 if (last_annotation_string != ir->annotation) {
898 last_annotation_string = ir->annotation;
899 if (last_annotation_string)
900 fprintf(stderr, " %s\n", last_annotation_string);
901 }
902 }
903
904 for (unsigned int i = 0; i < 3; i++) {
905 src[i] = brw_reg_from_fs_reg(&ir->src[i]);
906
907 /* The accumulator result appears to get used for the
908 * conditional modifier generation. When negating a UD
909 * value, there is a 33rd bit generated for the sign in the
910 * accumulator value, so now you can't check, for example,
911 * equality with a 32-bit value. See piglit fs-op-neg-uvec4.
912 */
913 assert(!ir->conditional_mod ||
914 ir->src[i].type != BRW_REGISTER_TYPE_UD ||
915 !ir->src[i].negate);
916 }
917 dst = brw_reg_from_fs_reg(&ir->dst);
918
919 default_state.conditional_mod = ir->conditional_mod;
920 default_state.predicate = ir->predicate;
921 default_state.predicate_inverse = ir->predicate_inverse;
922 default_state.saturate = ir->saturate;
923 default_state.flag_subreg_nr = ir->flag_subreg;
924
925 if (dispatch_width == 16 && !ir->force_uncompressed)
926 default_state.exec_size = BRW_EXECUTE_16;
927 else
928 default_state.exec_size = BRW_EXECUTE_8;
929
930 /* fs_inst::force_sechalf is only used for original Gen4 code, so we
931 * don't handle it. Add qtr_control to default_state if that changes.
932 */
933 assert(!ir->force_sechalf);
934
935 switch (ir->opcode) {
936 case BRW_OPCODE_MOV:
937 MOV(dst, src[0]);
938 break;
939 case BRW_OPCODE_ADD:
940 ADD(dst, src[0], src[1]);
941 break;
942 case BRW_OPCODE_MUL:
943 MUL(dst, src[0], src[1]);
944 break;
945 case BRW_OPCODE_MACH:
946 MACH(dst, src[0], src[1]);
947 break;
948
949 case BRW_OPCODE_MAD:
950 default_state.access_mode = BRW_ALIGN_16;
951 MAD(dst, src[0], src[1], src[2]);
952 default_state.access_mode = BRW_ALIGN_1;
953 break;
954
955 case BRW_OPCODE_LRP:
956 default_state.access_mode = BRW_ALIGN_16;
957 LRP(dst, src[0], src[1], src[2]);
958 default_state.access_mode = BRW_ALIGN_1;
959 break;
960
961
962 case BRW_OPCODE_FRC:
963 FRC(dst, src[0]);
964 break;
965 case BRW_OPCODE_RNDD:
966 RNDD(dst, src[0]);
967 break;
968 case BRW_OPCODE_RNDE:
969 RNDE(dst, src[0]);
970 break;
971 case BRW_OPCODE_RNDZ:
972 RNDZ(dst, src[0]);
973 break;
974
975 case BRW_OPCODE_AND:
976 AND(dst, src[0], src[1]);
977 break;
978 case BRW_OPCODE_OR:
979 OR(dst, src[0], src[1]);
980 break;
981 case BRW_OPCODE_XOR:
982 XOR(dst, src[0], src[1]);
983 break;
984 case BRW_OPCODE_NOT:
985 NOT(dst, src[0]);
986 break;
987 case BRW_OPCODE_ASR:
988 ASR(dst, src[0], src[1]);
989 break;
990 case BRW_OPCODE_SHR:
991 SHR(dst, src[0], src[1]);
992 break;
993 case BRW_OPCODE_SHL:
994 SHL(dst, src[0], src[1]);
995 break;
996
997 case BRW_OPCODE_F32TO16:
998 MOV(retype(dst, BRW_REGISTER_TYPE_HF), src[0]);
999 break;
1000 case BRW_OPCODE_F16TO32:
1001 MOV(dst, retype(src[0], BRW_REGISTER_TYPE_HF));
1002 break;
1003
1004 case BRW_OPCODE_CMP:
1005 CMP(dst, ir->conditional_mod, src[0], src[1]);
1006 break;
1007 case BRW_OPCODE_SEL:
1008 SEL(dst, src[0], src[1]);
1009 break;
1010
1011 case BRW_OPCODE_BFREV:
1012 /* BFREV only supports UD type for src and dst. */
1013 BFREV(retype(dst, BRW_REGISTER_TYPE_UD),
1014 retype(src[0], BRW_REGISTER_TYPE_UD));
1015 break;
1016
1017 case BRW_OPCODE_FBH:
1018 /* FBH only supports UD type for dst. */
1019 FBH(retype(dst, BRW_REGISTER_TYPE_UD), src[0]);
1020 break;
1021
1022 case BRW_OPCODE_FBL:
1023 /* FBL only supports UD type for dst. */
1024 FBL(retype(dst, BRW_REGISTER_TYPE_UD), src[0]);
1025 break;
1026
1027 case BRW_OPCODE_CBIT:
1028 /* CBIT only supports UD type for dst. */
1029 CBIT(retype(dst, BRW_REGISTER_TYPE_UD), src[0]);
1030 break;
1031
1032 case BRW_OPCODE_ADDC:
1033 ADDC(dst, src[0], src[1]);
1034 break;
1035
1036 case BRW_OPCODE_SUBB:
1037 SUBB(dst, src[0], src[1]);
1038 break;
1039
1040 case BRW_OPCODE_BFE:
1041 default_state.access_mode = BRW_ALIGN_16;
1042 BFE(dst, src[0], src[1], src[2]);
1043 default_state.access_mode = BRW_ALIGN_1;
1044 break;
1045
1046 case BRW_OPCODE_BFI1:
1047 BFI1(dst, src[0], src[1]);
1048 break;
1049
1050 case BRW_OPCODE_BFI2:
1051 default_state.access_mode = BRW_ALIGN_16;
1052 BFI2(dst, src[0], src[1], src[2]);
1053 default_state.access_mode = BRW_ALIGN_1;
1054 break;
1055
1056 case BRW_OPCODE_IF:
1057 IF(BRW_PREDICATE_NORMAL);
1058 break;
1059
1060 case BRW_OPCODE_ELSE:
1061 ELSE();
1062 break;
1063
1064 case BRW_OPCODE_ENDIF:
1065 ENDIF();
1066 break;
1067
1068 case BRW_OPCODE_DO:
1069 DO();
1070 break;
1071
1072 case BRW_OPCODE_BREAK:
1073 BREAK();
1074 break;
1075
1076 case BRW_OPCODE_CONTINUE:
1077 CONTINUE();
1078 break;
1079
1080 case BRW_OPCODE_WHILE:
1081 WHILE();
1082 break;
1083
1084 case SHADER_OPCODE_RCP:
1085 MATH(BRW_MATH_FUNCTION_INV, dst, src[0]);
1086 break;
1087
1088 case SHADER_OPCODE_RSQ:
1089 MATH(BRW_MATH_FUNCTION_RSQ, dst, src[0]);
1090 break;
1091
1092 case SHADER_OPCODE_SQRT:
1093 MATH(BRW_MATH_FUNCTION_SQRT, dst, src[0]);
1094 break;
1095
1096 case SHADER_OPCODE_EXP2:
1097 MATH(BRW_MATH_FUNCTION_EXP, dst, src[0]);
1098 break;
1099
1100 case SHADER_OPCODE_LOG2:
1101 MATH(BRW_MATH_FUNCTION_LOG, dst, src[0]);
1102 break;
1103
1104 case SHADER_OPCODE_SIN:
1105 MATH(BRW_MATH_FUNCTION_SIN, dst, src[0]);
1106 break;
1107
1108 case SHADER_OPCODE_COS:
1109 MATH(BRW_MATH_FUNCTION_COS, dst, src[0]);
1110 break;
1111
1112 case SHADER_OPCODE_INT_QUOTIENT:
1113 MATH(BRW_MATH_FUNCTION_INT_DIV_QUOTIENT, dst, src[0], src[1]);
1114 break;
1115
1116 case SHADER_OPCODE_INT_REMAINDER:
1117 MATH(BRW_MATH_FUNCTION_INT_DIV_REMAINDER, dst, src[0], src[1]);
1118 break;
1119
1120 case SHADER_OPCODE_POW:
1121 MATH(BRW_MATH_FUNCTION_POW, dst, src[0], src[1]);
1122 break;
1123
1124 case FS_OPCODE_PIXEL_X:
1125 case FS_OPCODE_PIXEL_Y:
1126 assert(!"FS_OPCODE_PIXEL_X and FS_OPCODE_PIXEL_Y are only for Gen4-5.");
1127 break;
1128
1129 case FS_OPCODE_CINTERP:
1130 MOV(dst, src[0]);
1131 break;
1132 case FS_OPCODE_LINTERP:
1133 generate_linterp(ir, dst, src);
1134 break;
1135 case SHADER_OPCODE_TEX:
1136 case FS_OPCODE_TXB:
1137 case SHADER_OPCODE_TXD:
1138 case SHADER_OPCODE_TXF:
1139 case SHADER_OPCODE_TXF_CMS:
1140 case SHADER_OPCODE_TXF_UMS:
1141 case SHADER_OPCODE_TXF_MCS:
1142 case SHADER_OPCODE_TXL:
1143 case SHADER_OPCODE_TXS:
1144 case SHADER_OPCODE_LOD:
1145 case SHADER_OPCODE_TG4:
1146 case SHADER_OPCODE_TG4_OFFSET:
1147 generate_tex(ir, dst, src[0]);
1148 break;
1149
1150 case FS_OPCODE_DDX:
1151 generate_ddx(ir, dst, src[0]);
1152 break;
1153 case FS_OPCODE_DDY:
1154 /* Make sure fp->UsesDFdy flag got set (otherwise there's no
1155 * guarantee that c->key.render_to_fbo is set).
1156 */
1157 assert(fp->UsesDFdy);
1158 generate_ddy(ir, dst, src[0], c->key.render_to_fbo);
1159 break;
1160
1161 case SHADER_OPCODE_GEN4_SCRATCH_WRITE:
1162 generate_scratch_write(ir, src[0]);
1163 break;
1164
1165 case SHADER_OPCODE_GEN4_SCRATCH_READ:
1166 generate_scratch_read(ir, dst);
1167 break;
1168
1169 case SHADER_OPCODE_GEN7_SCRATCH_READ:
1170 generate_scratch_read_gen7(ir, dst);
1171 break;
1172
1173 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD_GEN7:
1174 generate_uniform_pull_constant_load(ir, dst, src[0], src[1]);
1175 break;
1176
1177 case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_GEN7:
1178 generate_varying_pull_constant_load(ir, dst, src[0], src[1]);
1179 break;
1180
1181 case FS_OPCODE_FB_WRITE:
1182 generate_fb_write(ir);
1183 break;
1184
1185 case FS_OPCODE_MOV_DISPATCH_TO_FLAGS:
1186 generate_mov_dispatch_to_flags(ir);
1187 break;
1188
1189 case FS_OPCODE_DISCARD_JUMP:
1190 generate_discard_jump(ir);
1191 break;
1192
1193 case SHADER_OPCODE_SHADER_TIME_ADD:
1194 assert(!"XXX: Missing Gen8 scalar support for INTEL_DEBUG=shader_time");
1195 break;
1196
1197 case SHADER_OPCODE_UNTYPED_ATOMIC:
1198 assert(!"XXX: Missing Gen8 scalar support for untyped atomics");
1199 break;
1200
1201 case SHADER_OPCODE_UNTYPED_SURFACE_READ:
1202 assert(!"XXX: Missing Gen8 scalar support for untyped surface reads");
1203 break;
1204
1205 case FS_OPCODE_SET_SIMD4X2_OFFSET:
1206 generate_set_simd4x2_offset(ir, dst, src[0]);
1207 break;
1208
1209 case FS_OPCODE_SET_OMASK:
1210 generate_set_omask(ir, dst, src[0]);
1211 break;
1212
1213 case FS_OPCODE_SET_SAMPLE_ID:
1214 generate_set_sample_id(ir, dst, src[0], src[1]);
1215 break;
1216
1217 case FS_OPCODE_PACK_HALF_2x16_SPLIT:
1218 generate_pack_half_2x16_split(ir, dst, src[0], src[1]);
1219 break;
1220
1221 case FS_OPCODE_UNPACK_HALF_2x16_SPLIT_X:
1222 case FS_OPCODE_UNPACK_HALF_2x16_SPLIT_Y:
1223 generate_unpack_half_2x16_split(ir, dst, src[0]);
1224 break;
1225
1226 case FS_OPCODE_PLACEHOLDER_HALT:
1227 /* This is the place where the final HALT needs to be inserted if
1228 * we've emitted any discards. If not, this will emit no code.
1229 */
1230 patch_discard_jumps_to_fb_writes();
1231 break;
1232
1233 default:
1234 if (ir->opcode < int(ARRAY_SIZE(opcode_descs))) {
1235 _mesa_problem(ctx, "Unsupported opcode `%s' in FS",
1236 opcode_descs[ir->opcode].name);
1237 } else {
1238 _mesa_problem(ctx, "Unsupported opcode %d in FS", ir->opcode);
1239 }
1240 abort();
1241 }
1242
1243 if (unlikely(INTEL_DEBUG & DEBUG_WM)) {
1244 disassemble(stderr, last_native_inst_offset, next_inst_offset);
1245
1246 foreach_list(node, &cfg->block_list) {
1247 bblock_link *link = (bblock_link *)node;
1248 bblock_t *block = link->block;
1249
1250 if (block->end == ir) {
1251 fprintf(stderr, " END B%d", block->block_num);
1252 foreach_list(successor_node, &block->children) {
1253 bblock_link *successor_link =
1254 (bblock_link *)successor_node;
1255 bblock_t *successor_block = successor_link->block;
1256 fprintf(stderr, " ->B%d", successor_block->block_num);
1257 }
1258 fprintf(stderr, "\n");
1259 }
1260 }
1261 }
1262
1263 last_native_inst_offset = next_inst_offset;
1264 }
1265
1266 if (unlikely(INTEL_DEBUG & DEBUG_WM)) {
1267 fprintf(stderr, "\n");
1268 }
1269
1270 patch_jump_targets();
1271
1272 /* OK, while the INTEL_DEBUG=fs above is very nice for debugging FS
1273 * emit issues, it doesn't get the jump distances into the output,
1274 * which is often something we want to debug. So this is here in
1275 * case you're doing that.
1276 */
1277 if (0 && unlikely(INTEL_DEBUG & DEBUG_WM)) {
1278 disassemble(stderr, 0, next_inst_offset);
1279 }
1280 }
1281
1282 const unsigned *
1283 gen8_fs_generator::generate_assembly(exec_list *simd8_instructions,
1284 exec_list *simd16_instructions,
1285 unsigned *assembly_size)
1286 {
1287 assert(simd8_instructions || simd16_instructions);
1288
1289 if (simd8_instructions) {
1290 dispatch_width = 8;
1291 generate_code(simd8_instructions);
1292 }
1293
1294 if (simd16_instructions) {
1295 /* Align to a 64-byte boundary. */
1296 while ((nr_inst * sizeof(gen8_instruction)) % 64)
1297 NOP();
1298
1299 /* Save off the start of this SIMD16 program */
1300 c->prog_data.prog_offset_16 = nr_inst * sizeof(gen8_instruction);
1301
1302 dispatch_width = 16;
1303 generate_code(simd16_instructions);
1304 }
1305
1306 *assembly_size = next_inst_offset;
1307 return (const unsigned *) store;
1308 }