freedreno/ir3: fix register usage calculations
[mesa.git] / src / gallium / auxiliary / gallivm / lp_bld_tgsi.h
1 /**************************************************************************
2 *
3 * Copyright 2011-2012 Advanced Micro Devices, Inc.
4 * Copyright 2009 VMware, Inc.
5 * All Rights Reserved.
6 *
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * copy of this software and associated documentation files (the
9 * "Software"), to deal in the Software without restriction, including
10 * without limitation the rights to use, copy, modify, merge, publish,
11 * distribute, sub license, and/or sell copies of the Software, and to
12 * permit persons to whom the Software is furnished to do so, subject to
13 * the following conditions:
14 *
15 * The above copyright notice and this permission notice (including the
16 * next paragraph) shall be included in all copies or substantial portions
17 * of the Software.
18 *
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
20 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
21 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
22 * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
23 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
24 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
25 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
26 *
27 **************************************************************************/
28
29 /**
30 * @file
31 * TGSI to LLVM IR translation.
32 *
33 * @author Jose Fonseca <jfonseca@vmware.com>
34 * @author Tom Stellard <thomas.stellard@amd.com>
35 */
36
37 #ifndef LP_BLD_TGSI_H
38 #define LP_BLD_TGSI_H
39
40 #include "gallivm/lp_bld.h"
41 #include "gallivm/lp_bld_tgsi_action.h"
42 #include "gallivm/lp_bld_limits.h"
43 #include "gallivm/lp_bld_sample.h"
44 #include "lp_bld_type.h"
45 #include "pipe/p_compiler.h"
46 #include "pipe/p_state.h"
47 #include "tgsi/tgsi_exec.h"
48 #include "tgsi/tgsi_scan.h"
49 #include "tgsi/tgsi_info.h"
50
51 #define LP_CHAN_ALL ~0
52
53 #define LP_MAX_INSTRUCTIONS 256
54
55 struct tgsi_full_declaration;
56 struct tgsi_full_immediate;
57 struct tgsi_full_instruction;
58 struct tgsi_full_src_register;
59 struct tgsi_opcode_info;
60 struct tgsi_token;
61 struct tgsi_shader_info;
62 struct lp_build_mask_context;
63 struct gallivm_state;
64 struct lp_derivatives;
65 struct lp_build_tgsi_gs_iface;
66
67
68 enum lp_build_tex_modifier {
69 LP_BLD_TEX_MODIFIER_NONE = 0,
70 LP_BLD_TEX_MODIFIER_PROJECTED,
71 LP_BLD_TEX_MODIFIER_LOD_BIAS,
72 LP_BLD_TEX_MODIFIER_EXPLICIT_LOD,
73 LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV,
74 LP_BLD_TEX_MODIFIER_LOD_ZERO
75 };
76
77
78 /**
79 * Describe a channel of a register.
80 *
81 * The value can be a:
82 * - immediate value (i.e. derived from a IMM register)
83 * - CONST[n].x/y/z/w
84 * - IN[n].x/y/z/w
85 * - undetermined (when .file == TGSI_FILE_NULL)
86 *
87 * This is one of the analysis results, and is used to described
88 * the output color in terms of inputs.
89 */
90 struct lp_tgsi_channel_info
91 {
92 unsigned file:4; /* TGSI_FILE_* */
93 unsigned swizzle:3; /* PIPE_SWIZZLE_x */
94 union {
95 uint32_t index;
96 float value; /* for TGSI_FILE_IMMEDIATE */
97 } u;
98 };
99
100
101 /**
102 * Describe a texture sampler interpolator.
103 *
104 * The interpolation is described in terms of regular inputs.
105 */
106 struct lp_tgsi_texture_info
107 {
108 struct lp_tgsi_channel_info coord[4];
109 unsigned target:8; /* TGSI_TEXTURE_* */
110 unsigned sampler_unit:8; /* Sampler unit */
111 unsigned texture_unit:8; /* Texture unit */
112 unsigned modifier:8; /* LP_BLD_TEX_MODIFIER_* */
113 };
114
115
116 struct lp_tgsi_info
117 {
118 struct tgsi_shader_info base;
119
120 /*
121 * Whether any of the texture opcodes access a register file other than
122 * TGSI_FILE_INPUT.
123 *
124 * We could also handle TGSI_FILE_CONST/IMMEDIATE here, but there is little
125 * benefit.
126 */
127 unsigned indirect_textures:1;
128
129 /*
130 * Whether any of the texture (sample) ocpodes use different sampler
131 * and sampler view unit.
132 */
133 unsigned sampler_texture_units_different:1;
134
135 /*
136 * Whether any immediate values are outside the range of 0 and 1
137 */
138 unsigned unclamped_immediates:1;
139
140 /*
141 * Texture opcode description. Aimed at detecting and described direct
142 * texture opcodes.
143 */
144 unsigned num_texs;
145 struct lp_tgsi_texture_info tex[PIPE_MAX_SAMPLERS];
146
147 /*
148 * Output description. Aimed at detecting and describing simple blit
149 * shaders.
150 */
151 struct lp_tgsi_channel_info output[PIPE_MAX_SHADER_OUTPUTS][4];
152
153 /*
154 * Shortcut pointers into the above (for fragment shaders).
155 */
156 const struct lp_tgsi_channel_info *cbuf[PIPE_MAX_COLOR_BUFS];
157 };
158
159 /**
160 * Reference to system values.
161 */
162 struct lp_bld_tgsi_system_values {
163 LLVMValueRef instance_id;
164 LLVMValueRef vertex_id;
165 LLVMValueRef vertex_id_nobase;
166 LLVMValueRef prim_id;
167 LLVMValueRef basevertex;
168 };
169
170
171 /**
172 * Sampler code generation interface.
173 *
174 * Although texture sampling is a requirement for TGSI translation, it is
175 * a very different problem with several different approaches to it. This
176 * structure establishes an interface for texture sampling code generation, so
177 * that we can easily use different texture sampling strategies.
178 */
179 struct lp_build_sampler_soa
180 {
181 void
182 (*destroy)( struct lp_build_sampler_soa *sampler );
183
184 void
185 (*emit_fetch_texel)( const struct lp_build_sampler_soa *sampler,
186 struct gallivm_state *gallivm,
187 struct lp_type type,
188 boolean is_fetch,
189 unsigned texture_index,
190 unsigned sampler_index,
191 const LLVMValueRef *coords,
192 const LLVMValueRef *offsets,
193 const struct lp_derivatives *derivs,
194 LLVMValueRef lod_bias, /* optional */
195 LLVMValueRef explicit_lod, /* optional */
196 enum lp_sampler_lod_property,
197 LLVMValueRef *texel);
198
199 void
200 (*emit_size_query)( const struct lp_build_sampler_soa *sampler,
201 struct gallivm_state *gallivm,
202 struct lp_type type,
203 unsigned unit,
204 unsigned target,
205 boolean need_nr_mips,
206 enum lp_sampler_lod_property,
207 LLVMValueRef explicit_lod, /* optional */
208 LLVMValueRef *sizes_out);
209 };
210
211
212 struct lp_build_sampler_aos
213 {
214 LLVMValueRef
215 (*emit_fetch_texel)( struct lp_build_sampler_aos *sampler,
216 struct lp_build_context *bld,
217 unsigned target, /* TGSI_TEXTURE_* */
218 unsigned unit,
219 LLVMValueRef coords,
220 const struct lp_derivatives derivs,
221 enum lp_build_tex_modifier modifier);
222 };
223
224
225 void
226 lp_build_tgsi_info(const struct tgsi_token *tokens,
227 struct lp_tgsi_info *info);
228
229
230 void
231 lp_build_tgsi_soa(struct gallivm_state *gallivm,
232 const struct tgsi_token *tokens,
233 struct lp_type type,
234 struct lp_build_mask_context *mask,
235 LLVMValueRef consts_ptr,
236 LLVMValueRef const_sizes_ptr,
237 const struct lp_bld_tgsi_system_values *system_values,
238 const LLVMValueRef (*inputs)[4],
239 LLVMValueRef (*outputs)[4],
240 struct lp_build_sampler_soa *sampler,
241 const struct tgsi_shader_info *info,
242 const struct lp_build_tgsi_gs_iface *gs_iface);
243
244
245 void
246 lp_build_tgsi_aos(struct gallivm_state *gallivm,
247 const struct tgsi_token *tokens,
248 struct lp_type type,
249 const unsigned char swizzles[4],
250 LLVMValueRef consts_ptr,
251 const LLVMValueRef *inputs,
252 LLVMValueRef *outputs,
253 struct lp_build_sampler_aos *sampler,
254 const struct tgsi_shader_info *info);
255
256
257 enum lp_exec_mask_break_type {
258 LP_EXEC_MASK_BREAK_TYPE_LOOP,
259 LP_EXEC_MASK_BREAK_TYPE_SWITCH
260 };
261
262
263 struct lp_exec_mask {
264 struct lp_build_context *bld;
265
266 boolean has_mask;
267 boolean ret_in_main;
268
269 LLVMTypeRef int_vec_type;
270
271 LLVMValueRef exec_mask;
272
273 LLVMValueRef ret_mask;
274 LLVMValueRef cond_mask;
275 LLVMValueRef switch_mask; /* current switch exec mask */
276 LLVMValueRef cont_mask;
277 LLVMValueRef break_mask;
278
279 struct function_ctx {
280 int pc;
281 LLVMValueRef ret_mask;
282
283 LLVMValueRef cond_stack[LP_MAX_TGSI_NESTING];
284 int cond_stack_size;
285
286 /* keep track if break belongs to switch or loop */
287 enum lp_exec_mask_break_type break_type_stack[LP_MAX_TGSI_NESTING];
288 enum lp_exec_mask_break_type break_type;
289
290 struct {
291 LLVMValueRef switch_val;
292 LLVMValueRef switch_mask;
293 LLVMValueRef switch_mask_default;
294 boolean switch_in_default;
295 unsigned switch_pc;
296 } switch_stack[LP_MAX_TGSI_NESTING];
297 int switch_stack_size;
298 LLVMValueRef switch_val;
299 LLVMValueRef switch_mask_default; /* reverse of switch mask used for default */
300 boolean switch_in_default; /* if switch exec is currently in default */
301 unsigned switch_pc; /* when used points to default or endswitch-1 */
302
303 LLVMValueRef loop_limiter;
304 LLVMBasicBlockRef loop_block;
305 LLVMValueRef break_var;
306 struct {
307 LLVMBasicBlockRef loop_block;
308 LLVMValueRef cont_mask;
309 LLVMValueRef break_mask;
310 LLVMValueRef break_var;
311 } loop_stack[LP_MAX_TGSI_NESTING];
312 int loop_stack_size;
313
314 } *function_stack;
315 int function_stack_size;
316 };
317
318 struct lp_build_tgsi_inst_list
319 {
320 struct tgsi_full_instruction *instructions;
321 uint max_instructions;
322 uint num_instructions;
323 };
324
325 unsigned lp_bld_tgsi_list_init(struct lp_build_tgsi_context * bld_base);
326
327
328 unsigned lp_bld_tgsi_add_instruction(
329 struct lp_build_tgsi_context * bld_base,
330 const struct tgsi_full_instruction *inst_to_add);
331
332
333 struct lp_build_tgsi_context;
334
335
336 typedef LLVMValueRef (*lp_build_emit_fetch_fn)(struct lp_build_tgsi_context *,
337 const struct tgsi_full_src_register *,
338 enum tgsi_opcode_type,
339 unsigned);
340
341 struct lp_build_tgsi_context
342 {
343 struct lp_build_context base;
344
345 struct lp_build_context uint_bld;
346 struct lp_build_context int_bld;
347
348 /** This array stores functions that are used to transform TGSI opcodes to
349 * LLVM instructions.
350 */
351 struct lp_build_tgsi_action op_actions[TGSI_OPCODE_LAST];
352
353 /* TGSI_OPCODE_RSQ is defined as 1 / sqrt( abs(src0.x) ), rsq_action
354 * should compute 1 / sqrt (src0.x) */
355 struct lp_build_tgsi_action rsq_action;
356
357 struct lp_build_tgsi_action sqrt_action;
358
359 const struct tgsi_shader_info *info;
360
361 lp_build_emit_fetch_fn emit_fetch_funcs[TGSI_FILE_COUNT];
362
363 LLVMValueRef (*emit_swizzle)(struct lp_build_tgsi_context *,
364 LLVMValueRef, unsigned, unsigned, unsigned, unsigned);
365
366
367 void (*emit_debug)(struct lp_build_tgsi_context *,
368 const struct tgsi_full_instruction *,
369 const struct tgsi_opcode_info *);
370
371 void (*emit_store)(struct lp_build_tgsi_context *,
372 const struct tgsi_full_instruction *,
373 const struct tgsi_opcode_info *,
374 LLVMValueRef dst[4]);
375
376 void (*emit_declaration)(struct lp_build_tgsi_context *,
377 const struct tgsi_full_declaration *decl);
378
379 void (*emit_immediate)(struct lp_build_tgsi_context *,
380 const struct tgsi_full_immediate *imm);
381
382
383 /* Allow the user to store data in this structure rather than passing it
384 * to every function. */
385 void * userdata;
386
387 boolean soa;
388
389 int pc;
390
391 struct tgsi_full_instruction *instructions;
392 uint max_instructions;
393 uint num_instructions;
394
395 /** This function allows the user to insert some instructions at the
396 * beginning of the program. It is optional and does not need to be
397 * implemented.
398 */
399 void (*emit_prologue)(struct lp_build_tgsi_context*);
400
401 /** This function allows the user to insert some instructions at the end of
402 * the program. This callback is intended to be used for emitting
403 * instructions to handle the export for the output registers, but it can
404 * be used for any purpose. Implementing this function is optiona, but
405 * recommended.
406 */
407 void (*emit_epilogue)(struct lp_build_tgsi_context*);
408 };
409
410 struct lp_build_tgsi_gs_iface
411 {
412 LLVMValueRef (*fetch_input)(const struct lp_build_tgsi_gs_iface *gs_iface,
413 struct lp_build_tgsi_context * bld_base,
414 boolean is_vindex_indirect,
415 LLVMValueRef vertex_index,
416 boolean is_aindex_indirect,
417 LLVMValueRef attrib_index,
418 LLVMValueRef swizzle_index);
419 void (*emit_vertex)(const struct lp_build_tgsi_gs_iface *gs_iface,
420 struct lp_build_tgsi_context * bld_base,
421 LLVMValueRef (*outputs)[4],
422 LLVMValueRef emitted_vertices_vec);
423 void (*end_primitive)(const struct lp_build_tgsi_gs_iface *gs_iface,
424 struct lp_build_tgsi_context * bld_base,
425 LLVMValueRef verts_per_prim_vec,
426 LLVMValueRef emitted_prims_vec);
427 void (*gs_epilogue)(const struct lp_build_tgsi_gs_iface *gs_iface,
428 struct lp_build_tgsi_context * bld_base,
429 LLVMValueRef total_emitted_vertices_vec,
430 LLVMValueRef emitted_prims_vec);
431 };
432
433 struct lp_build_tgsi_soa_context
434 {
435 struct lp_build_tgsi_context bld_base;
436
437 /* Builder for scalar elements of shader's data type (float) */
438 struct lp_build_context elem_bld;
439
440 const struct lp_build_tgsi_gs_iface *gs_iface;
441 LLVMValueRef emitted_prims_vec_ptr;
442 LLVMValueRef total_emitted_vertices_vec_ptr;
443 LLVMValueRef emitted_vertices_vec_ptr;
444 LLVMValueRef max_output_vertices_vec;
445
446 LLVMValueRef consts_ptr;
447 LLVMValueRef const_sizes_ptr;
448 LLVMValueRef consts[LP_MAX_TGSI_CONST_BUFFERS];
449 LLVMValueRef consts_sizes[LP_MAX_TGSI_CONST_BUFFERS];
450 const LLVMValueRef (*inputs)[TGSI_NUM_CHANNELS];
451 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS];
452
453 const struct lp_build_sampler_soa *sampler;
454
455 struct tgsi_declaration_sampler_view sv[PIPE_MAX_SHADER_SAMPLER_VIEWS];
456
457 LLVMValueRef immediates[LP_MAX_INLINED_IMMEDIATES][TGSI_NUM_CHANNELS];
458 LLVMValueRef temps[LP_MAX_INLINED_TEMPS][TGSI_NUM_CHANNELS];
459 LLVMValueRef addr[LP_MAX_TGSI_ADDRS][TGSI_NUM_CHANNELS];
460 LLVMValueRef preds[LP_MAX_TGSI_PREDS][TGSI_NUM_CHANNELS];
461
462 /* We allocate/use this array of temps if (1 << TGSI_FILE_TEMPORARY) is
463 * set in the indirect_files field.
464 * The temps[] array above is unused then.
465 */
466 LLVMValueRef temps_array;
467
468 /* We allocate/use this array of output if (1 << TGSI_FILE_OUTPUT) is
469 * set in the indirect_files field.
470 * The outputs[] array above is unused then.
471 */
472 LLVMValueRef outputs_array;
473
474 /* We allocate/use this array of inputs if (1 << TGSI_FILE_INPUT) is
475 * set in the indirect_files field.
476 * The inputs[] array above is unused then.
477 */
478 LLVMValueRef inputs_array;
479
480 /* We allocate/use this array of temps if (1 << TGSI_FILE_IMMEDIATE) is
481 * set in the indirect_files field.
482 */
483 LLVMValueRef imms_array;
484
485
486 struct lp_bld_tgsi_system_values system_values;
487
488 /** bitmask indicating which register files are accessed indirectly */
489 unsigned indirect_files;
490
491 struct lp_build_mask_context *mask;
492 struct lp_exec_mask exec_mask;
493
494 uint num_immediates;
495 boolean use_immediates_array;
496 };
497
498 void
499 lp_emit_declaration_soa(
500 struct lp_build_tgsi_context *bld,
501 const struct tgsi_full_declaration *decl);
502
503 void lp_emit_immediate_soa(
504 struct lp_build_tgsi_context *bld_base,
505 const struct tgsi_full_immediate *imm);
506
507 boolean
508 lp_emit_instruction_soa(
509 struct lp_build_tgsi_soa_context *bld,
510 const struct tgsi_full_instruction *inst,
511 const struct tgsi_opcode_info *info);
512
513
514 LLVMValueRef
515 lp_get_temp_ptr_soa(
516 struct lp_build_tgsi_soa_context *bld,
517 unsigned index,
518 unsigned chan);
519
520 LLVMValueRef
521 lp_get_output_ptr(
522 struct lp_build_tgsi_soa_context *bld,
523 unsigned index,
524 unsigned chan);
525
526 struct lp_build_tgsi_aos_context
527 {
528 struct lp_build_tgsi_context bld_base;
529
530 /* Builder for integer masks and indices */
531 struct lp_build_context int_bld;
532
533 /*
534 * AoS swizzle used:
535 * - swizzles[0] = red index
536 * - swizzles[1] = green index
537 * - swizzles[2] = blue index
538 * - swizzles[3] = alpha index
539 */
540 unsigned char swizzles[4];
541 unsigned char inv_swizzles[4];
542
543 LLVMValueRef consts_ptr;
544 const LLVMValueRef *inputs;
545 LLVMValueRef *outputs;
546
547 struct lp_build_sampler_aos *sampler;
548
549 struct tgsi_declaration_sampler_view sv[PIPE_MAX_SHADER_SAMPLER_VIEWS];
550
551 LLVMValueRef immediates[LP_MAX_INLINED_IMMEDIATES];
552 LLVMValueRef temps[LP_MAX_INLINED_TEMPS];
553 LLVMValueRef addr[LP_MAX_TGSI_ADDRS];
554 LLVMValueRef preds[LP_MAX_TGSI_PREDS];
555
556 /* We allocate/use this array of temps if (1 << TGSI_FILE_TEMPORARY) is
557 * set in the indirect_files field.
558 * The temps[] array above is unused then.
559 */
560 LLVMValueRef temps_array;
561
562 /** bitmask indicating which register files are accessed indirectly */
563 unsigned indirect_files;
564
565 };
566
567 static INLINE struct lp_build_tgsi_soa_context *
568 lp_soa_context(struct lp_build_tgsi_context *bld_base)
569 {
570 return (struct lp_build_tgsi_soa_context *)bld_base;
571 }
572
573 static INLINE struct lp_build_tgsi_aos_context *
574 lp_aos_context(struct lp_build_tgsi_context *bld_base)
575 {
576 return (struct lp_build_tgsi_aos_context *)bld_base;
577 }
578
579 void
580 lp_emit_declaration_aos(
581 struct lp_build_tgsi_aos_context *bld,
582 const struct tgsi_full_declaration *decl);
583
584
585 boolean
586 lp_emit_instruction_aos(
587 struct lp_build_tgsi_aos_context *bld,
588 const struct tgsi_full_instruction *inst,
589 const struct tgsi_opcode_info *info,
590 int *pc);
591
592 void
593 lp_emit_store_aos(
594 struct lp_build_tgsi_aos_context *bld,
595 const struct tgsi_full_instruction *inst,
596 unsigned index,
597 LLVMValueRef value);
598
599 void lp_build_fetch_args(
600 struct lp_build_tgsi_context * bld_base,
601 struct lp_build_emit_data * emit_data);
602
603 LLVMValueRef
604 lp_build_tgsi_inst_llvm_aos(
605 struct lp_build_tgsi_context * bld_base,
606 const struct tgsi_full_instruction *inst);
607
608 void
609 lp_build_tgsi_intrinsic(
610 const struct lp_build_tgsi_action * action,
611 struct lp_build_tgsi_context * bld_base,
612 struct lp_build_emit_data * emit_data);
613
614 LLVMValueRef
615 lp_build_emit_llvm(
616 struct lp_build_tgsi_context *bld_base,
617 unsigned tgsi_opcode,
618 struct lp_build_emit_data * emit_data);
619
620 LLVMValueRef
621 lp_build_emit_llvm_unary(
622 struct lp_build_tgsi_context *bld_base,
623 unsigned tgsi_opcode,
624 LLVMValueRef arg0);
625
626 LLVMValueRef
627 lp_build_emit_llvm_binary(
628 struct lp_build_tgsi_context *bld_base,
629 unsigned tgsi_opcode,
630 LLVMValueRef arg0,
631 LLVMValueRef arg1);
632
633 LLVMValueRef
634 lp_build_emit_llvm_ternary(
635 struct lp_build_tgsi_context *bld_base,
636 unsigned tgsi_opcode,
637 LLVMValueRef arg0,
638 LLVMValueRef arg1,
639 LLVMValueRef arg2);
640
641 boolean
642 lp_build_tgsi_inst_llvm(
643 struct lp_build_tgsi_context * bld_base,
644 const struct tgsi_full_instruction *inst);
645
646 LLVMValueRef
647 lp_build_emit_fetch(
648 struct lp_build_tgsi_context *bld_base,
649 const struct tgsi_full_instruction *inst,
650 unsigned src_op,
651 const unsigned chan_index);
652
653
654 LLVMValueRef
655 lp_build_emit_fetch_texoffset(
656 struct lp_build_tgsi_context *bld_base,
657 const struct tgsi_full_instruction *inst,
658 unsigned tex_off_op,
659 const unsigned chan_index);
660
661 boolean
662 lp_build_tgsi_llvm(
663 struct lp_build_tgsi_context * bld_base,
664 const struct tgsi_token *tokens);
665
666 #endif /* LP_BLD_TGSI_H */