1 /*
2 * Copyright (C) 2005-2007 Brian Paul All Rights Reserved.
3 * Copyright (C) 2008 VMware, Inc. All Rights Reserved.
4 * Copyright © 2010 Intel Corporation
5 * Copyright © 2011 Bryan Cain
6 *
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * copy of this software and associated documentation files (the "Software"),
9 * to deal in the Software without restriction, including without limitation
10 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
11 * and/or sell copies of the Software, and to permit persons to whom the
12 * Software is furnished to do so, subject to the following conditions:
13 *
14 * The above copyright notice and this permission notice (including the next
15 * paragraph) shall be included in all copies or substantial portions of the
16 * Software.
17 *
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
19 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
20 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
21 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
22 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
23 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
24 * DEALINGS IN THE SOFTWARE.
25 */
26
27 /**
28 * \file glsl_to_tgsi.cpp
29 *
30 * Translate GLSL IR to TGSI.
31 */
32
33 #include "st_glsl_to_tgsi.h"
34
35 #include "compiler/glsl/glsl_parser_extras.h"
36 #include "compiler/glsl/ir_optimization.h"
37 #include "compiler/glsl/program.h"
38
39 #include "main/errors.h"
40 #include "main/shaderobj.h"
41 #include "main/uniforms.h"
42 #include "main/shaderapi.h"
43 #include "main/shaderimage.h"
44 #include "program/prog_instruction.h"
45
46 #include "pipe/p_context.h"
47 #include "pipe/p_screen.h"
48 #include "tgsi/tgsi_ureg.h"
49 #include "tgsi/tgsi_info.h"
50 #include "util/u_math.h"
51 #include "util/u_memory.h"
52 #include "st_glsl_types.h"
53 #include "st_program.h"
54 #include "st_mesa_to_tgsi.h"
55 #include "st_format.h"
56 #include "st_nir.h"
57 #include "st_shader_cache.h"
58 #include "st_glsl_to_tgsi_temprename.h"
59
60 #include "util/hash_table.h"
61 #include <algorithm>
62
63 #define PROGRAM_ANY_CONST ((1 << PROGRAM_STATE_VAR) | \
64 (1 << PROGRAM_CONSTANT) | \
65 (1 << PROGRAM_UNIFORM))
66
67 #define MAX_GLSL_TEXTURE_OFFSET 4
68
is_precise(const ir_variable * ir)69 static unsigned is_precise(const ir_variable *ir)
70 {
71 if (!ir)
72 return 0;
73 return ir->data.precise || ir->data.invariant;
74 }
75
76 class variable_storage {
77 DECLARE_RZALLOC_CXX_OPERATORS(variable_storage)
78
79 public:
variable_storage(ir_variable * var,gl_register_file file,int index,unsigned array_id=0)80 variable_storage(ir_variable *var, gl_register_file file, int index,
81 unsigned array_id = 0)
82 : file(file), index(index), component(0), var(var), array_id(array_id)
83 {
84 assert(file != PROGRAM_ARRAY || array_id != 0);
85 }
86
87 gl_register_file file;
88 int index;
89
90 /* Explicit component location. This is given in terms of the GLSL-style
91 * swizzles where each double is a single component, i.e. for 64-bit types
92 * it can only be 0 or 1.
93 */
94 int component;
95 ir_variable *var; /* variable that maps to this, if any */
96 unsigned array_id;
97 };
98
99 class immediate_storage : public exec_node {
100 public:
immediate_storage(gl_constant_value * values,int size32,int type)101 immediate_storage(gl_constant_value *values, int size32, int type)
102 {
103 memcpy(this->values, values, size32 * sizeof(gl_constant_value));
104 this->size32 = size32;
105 this->type = type;
106 }
107
108 /* doubles are stored across 2 gl_constant_values */
109 gl_constant_value values[4];
110 int size32; /**< Number of 32-bit components (1-4) */
111 int type; /**< GL_DOUBLE, GL_FLOAT, GL_INT, GL_BOOL, or GL_UNSIGNED_INT */
112 };
113
114 static const st_src_reg undef_src = st_src_reg(PROGRAM_UNDEFINED, 0, GLSL_TYPE_ERROR);
115 static const st_dst_reg undef_dst = st_dst_reg(PROGRAM_UNDEFINED, SWIZZLE_NOOP, GLSL_TYPE_ERROR);
116
117 struct inout_decl {
118 unsigned mesa_index;
119 unsigned array_id; /* TGSI ArrayID; 1-based: 0 means not an array */
120 unsigned size;
121 unsigned interp_loc;
122 unsigned gs_out_streams;
123 enum glsl_interp_mode interp;
124 enum glsl_base_type base_type;
125 ubyte usage_mask; /* GLSL-style usage-mask, i.e. single bit per double */
126 };
127
128 static struct inout_decl *
find_inout_array(struct inout_decl * decls,unsigned count,unsigned array_id)129 find_inout_array(struct inout_decl *decls, unsigned count, unsigned array_id)
130 {
131 assert(array_id != 0);
132
133 for (unsigned i = 0; i < count; i++) {
134 struct inout_decl *decl = &decls[i];
135
136 if (array_id == decl->array_id) {
137 return decl;
138 }
139 }
140
141 return NULL;
142 }
143
144 static enum glsl_base_type
find_array_type(struct inout_decl * decls,unsigned count,unsigned array_id)145 find_array_type(struct inout_decl *decls, unsigned count, unsigned array_id)
146 {
147 if (!array_id)
148 return GLSL_TYPE_ERROR;
149 struct inout_decl *decl = find_inout_array(decls, count, array_id);
150 if (decl)
151 return decl->base_type;
152 return GLSL_TYPE_ERROR;
153 }
154
155 struct hwatomic_decl {
156 unsigned location;
157 unsigned binding;
158 unsigned size;
159 unsigned array_id;
160 };
161
162 struct glsl_to_tgsi_visitor : public ir_visitor {
163 public:
164 glsl_to_tgsi_visitor();
165 ~glsl_to_tgsi_visitor();
166
167 struct gl_context *ctx;
168 struct gl_program *prog;
169 struct gl_shader_program *shader_program;
170 struct gl_linked_shader *shader;
171 struct gl_shader_compiler_options *options;
172
173 int next_temp;
174
175 unsigned *array_sizes;
176 unsigned max_num_arrays;
177 unsigned next_array;
178
179 struct inout_decl inputs[4 * PIPE_MAX_SHADER_INPUTS];
180 unsigned num_inputs;
181 unsigned num_input_arrays;
182 struct inout_decl outputs[4 * PIPE_MAX_SHADER_OUTPUTS];
183 unsigned num_outputs;
184 unsigned num_output_arrays;
185
186 struct hwatomic_decl atomic_info[PIPE_MAX_HW_ATOMIC_BUFFERS];
187 unsigned num_atomics;
188 unsigned num_atomic_arrays;
189 int num_address_regs;
190 uint32_t samplers_used;
191 glsl_base_type sampler_types[PIPE_MAX_SAMPLERS];
192 enum tgsi_texture_type sampler_targets[PIPE_MAX_SAMPLERS];
193 int images_used;
194 int image_targets[PIPE_MAX_SHADER_IMAGES];
195 enum pipe_format image_formats[PIPE_MAX_SHADER_IMAGES];
196 bool indirect_addr_consts;
197 int wpos_transform_const;
198
199 bool native_integers;
200 bool have_sqrt;
201 bool have_fma;
202 bool use_shared_memory;
203 bool has_tex_txf_lz;
204 bool precise;
205 bool need_uarl;
206
207 variable_storage *find_variable_storage(ir_variable *var);
208
209 int add_constant(gl_register_file file, gl_constant_value values[8],
210 int size, int datatype, uint16_t *swizzle_out);
211
212 st_src_reg get_temp(const glsl_type *type);
213 void reladdr_to_temp(ir_instruction *ir, st_src_reg *reg, int *num_reladdr);
214
215 st_src_reg st_src_reg_for_double(double val);
216 st_src_reg st_src_reg_for_float(float val);
217 st_src_reg st_src_reg_for_int(int val);
218 st_src_reg st_src_reg_for_int64(int64_t val);
219 st_src_reg st_src_reg_for_type(enum glsl_base_type type, int val);
220
221 /**
222 * \name Visit methods
223 *
224 * As typical for the visitor pattern, there must be one \c visit method for
225 * each concrete subclass of \c ir_instruction. Virtual base classes within
226 * the hierarchy should not have \c visit methods.
227 */
228 /*@{*/
229 virtual void visit(ir_variable *);
230 virtual void visit(ir_loop *);
231 virtual void visit(ir_loop_jump *);
232 virtual void visit(ir_function_signature *);
233 virtual void visit(ir_function *);
234 virtual void visit(ir_expression *);
235 virtual void visit(ir_swizzle *);
236 virtual void visit(ir_dereference_variable *);
237 virtual void visit(ir_dereference_array *);
238 virtual void visit(ir_dereference_record *);
239 virtual void visit(ir_assignment *);
240 virtual void visit(ir_constant *);
241 virtual void visit(ir_call *);
242 virtual void visit(ir_return *);
243 virtual void visit(ir_discard *);
244 virtual void visit(ir_texture *);
245 virtual void visit(ir_if *);
246 virtual void visit(ir_emit_vertex *);
247 virtual void visit(ir_end_primitive *);
248 virtual void visit(ir_barrier *);
249 /*@}*/
250
251 void visit_expression(ir_expression *, st_src_reg *) ATTRIBUTE_NOINLINE;
252
253 void visit_atomic_counter_intrinsic(ir_call *);
254 void visit_ssbo_intrinsic(ir_call *);
255 void visit_membar_intrinsic(ir_call *);
256 void visit_shared_intrinsic(ir_call *);
257 void visit_image_intrinsic(ir_call *);
258 void visit_generic_intrinsic(ir_call *, unsigned op);
259
260 st_src_reg result;
261
262 /** List of variable_storage */
263 struct hash_table *variables;
264
265 /** List of immediate_storage */
266 exec_list immediates;
267 unsigned num_immediates;
268
269 /** List of glsl_to_tgsi_instruction */
270 exec_list instructions;
271
272 glsl_to_tgsi_instruction *emit_asm(ir_instruction *ir, unsigned op,
273 st_dst_reg dst = undef_dst,
274 st_src_reg src0 = undef_src,
275 st_src_reg src1 = undef_src,
276 st_src_reg src2 = undef_src,
277 st_src_reg src3 = undef_src);
278
279 glsl_to_tgsi_instruction *emit_asm(ir_instruction *ir, unsigned op,
280 st_dst_reg dst, st_dst_reg dst1,
281 st_src_reg src0 = undef_src,
282 st_src_reg src1 = undef_src,
283 st_src_reg src2 = undef_src,
284 st_src_reg src3 = undef_src);
285
286 unsigned get_opcode(unsigned op,
287 st_dst_reg dst,
288 st_src_reg src0, st_src_reg src1);
289
290 /**
291 * Emit the correct dot-product instruction for the type of arguments
292 */
293 glsl_to_tgsi_instruction *emit_dp(ir_instruction *ir,
294 st_dst_reg dst,
295 st_src_reg src0,
296 st_src_reg src1,
297 unsigned elements);
298
299 void emit_scalar(ir_instruction *ir, unsigned op,
300 st_dst_reg dst, st_src_reg src0);
301
302 void emit_scalar(ir_instruction *ir, unsigned op,
303 st_dst_reg dst, st_src_reg src0, st_src_reg src1);
304
305 void emit_arl(ir_instruction *ir, st_dst_reg dst, st_src_reg src0);
306
307 void get_deref_offsets(ir_dereference *ir,
308 unsigned *array_size,
309 unsigned *base,
310 uint16_t *index,
311 st_src_reg *reladdr,
312 bool opaque);
313 void calc_deref_offsets(ir_dereference *tail,
314 unsigned *array_elements,
315 uint16_t *index,
316 st_src_reg *indirect,
317 unsigned *location);
318 st_src_reg canonicalize_gather_offset(st_src_reg offset);
319
320 bool try_emit_mad(ir_expression *ir,
321 int mul_operand);
322 bool try_emit_mad_for_and_not(ir_expression *ir,
323 int mul_operand);
324
325 void emit_swz(ir_expression *ir);
326
327 bool process_move_condition(ir_rvalue *ir);
328
329 void simplify_cmp(void);
330
331 void rename_temp_registers(struct rename_reg_pair *renames);
332 void get_first_temp_read(int *first_reads);
333 void get_first_temp_write(int *first_writes);
334 void get_last_temp_read_first_temp_write(int *last_reads, int *first_writes);
335 void get_last_temp_write(int *last_writes);
336
337 void copy_propagate(void);
338 int eliminate_dead_code(void);
339
340 void merge_two_dsts(void);
341 void merge_registers(void);
342 void renumber_registers(void);
343
344 void emit_block_mov(ir_assignment *ir, const struct glsl_type *type,
345 st_dst_reg *l, st_src_reg *r,
346 st_src_reg *cond, bool cond_swap);
347
348 void *mem_ctx;
349 };
350
351 static st_dst_reg address_reg = st_dst_reg(PROGRAM_ADDRESS, WRITEMASK_X, GLSL_TYPE_FLOAT, 0);
352 static st_dst_reg address_reg2 = st_dst_reg(PROGRAM_ADDRESS, WRITEMASK_X, GLSL_TYPE_FLOAT, 1);
353 static st_dst_reg sampler_reladdr = st_dst_reg(PROGRAM_ADDRESS, WRITEMASK_X, GLSL_TYPE_FLOAT, 2);
354
355 static void
356 fail_link(struct gl_shader_program *prog, const char *fmt, ...) PRINTFLIKE(2, 3);
357
358 static void
fail_link(struct gl_shader_program * prog,const char * fmt,...)359 fail_link(struct gl_shader_program *prog, const char *fmt, ...)
360 {
361 va_list args;
362 va_start(args, fmt);
363 ralloc_vasprintf_append(&prog->data->InfoLog, fmt, args);
364 va_end(args);
365
366 prog->data->LinkStatus = linking_failure;
367 }
368
369 int
swizzle_for_size(int size)370 swizzle_for_size(int size)
371 {
372 static const int size_swizzles[4] = {
373 MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_X, SWIZZLE_X, SWIZZLE_X),
374 MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, SWIZZLE_Y, SWIZZLE_Y),
375 MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, SWIZZLE_Z, SWIZZLE_Z),
376 MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, SWIZZLE_Z, SWIZZLE_W),
377 };
378
379 assert((size >= 1) && (size <= 4));
380 return size_swizzles[size - 1];
381 }
382
383
384 glsl_to_tgsi_instruction *
emit_asm(ir_instruction * ir,unsigned op,st_dst_reg dst,st_dst_reg dst1,st_src_reg src0,st_src_reg src1,st_src_reg src2,st_src_reg src3)385 glsl_to_tgsi_visitor::emit_asm(ir_instruction *ir, unsigned op,
386 st_dst_reg dst, st_dst_reg dst1,
387 st_src_reg src0, st_src_reg src1,
388 st_src_reg src2, st_src_reg src3)
389 {
390 glsl_to_tgsi_instruction *inst = new(mem_ctx) glsl_to_tgsi_instruction();
391 int num_reladdr = 0, i, j;
392 bool dst_is_64bit[2];
393
394 op = get_opcode(op, dst, src0, src1);
395
396 /* If we have to do relative addressing, we want to load the ARL
397 * reg directly for one of the regs, and preload the other reladdr
398 * sources into temps.
399 */
400 num_reladdr += dst.reladdr != NULL || dst.reladdr2;
401 assert(!dst1.reladdr); /* should be lowered in earlier passes */
402 num_reladdr += src0.reladdr != NULL || src0.reladdr2 != NULL;
403 num_reladdr += src1.reladdr != NULL || src1.reladdr2 != NULL;
404 num_reladdr += src2.reladdr != NULL || src2.reladdr2 != NULL;
405 num_reladdr += src3.reladdr != NULL || src3.reladdr2 != NULL;
406
407 reladdr_to_temp(ir, &src3, &num_reladdr);
408 reladdr_to_temp(ir, &src2, &num_reladdr);
409 reladdr_to_temp(ir, &src1, &num_reladdr);
410 reladdr_to_temp(ir, &src0, &num_reladdr);
411
412 if (dst.reladdr || dst.reladdr2) {
413 if (dst.reladdr)
414 emit_arl(ir, address_reg, *dst.reladdr);
415 if (dst.reladdr2)
416 emit_arl(ir, address_reg2, *dst.reladdr2);
417 num_reladdr--;
418 }
419
420 assert(num_reladdr == 0);
421
422 /* inst->op has only 8 bits. */
423 STATIC_ASSERT(TGSI_OPCODE_LAST <= 255);
424
425 inst->op = op;
426 inst->precise = this->precise;
427 inst->info = tgsi_get_opcode_info(op);
428 inst->dst[0] = dst;
429 inst->dst[1] = dst1;
430 inst->src[0] = src0;
431 inst->src[1] = src1;
432 inst->src[2] = src2;
433 inst->src[3] = src3;
434 inst->is_64bit_expanded = false;
435 inst->ir = ir;
436 inst->dead_mask = 0;
437 inst->tex_offsets = NULL;
438 inst->tex_offset_num_offset = 0;
439 inst->saturate = 0;
440 inst->tex_shadow = 0;
441 /* default to float, for paths where this is not initialized
442 * (since 0==UINT which is likely wrong):
443 */
444 inst->tex_type = GLSL_TYPE_FLOAT;
445
446 /* Update indirect addressing status used by TGSI */
447 if (dst.reladdr || dst.reladdr2) {
448 switch(dst.file) {
449 case PROGRAM_STATE_VAR:
450 case PROGRAM_CONSTANT:
451 case PROGRAM_UNIFORM:
452 this->indirect_addr_consts = true;
453 break;
454 case PROGRAM_IMMEDIATE:
455 assert(!"immediates should not have indirect addressing");
456 break;
457 default:
458 break;
459 }
460 }
461 else {
462 for (i = 0; i < 4; i++) {
463 if(inst->src[i].reladdr) {
464 switch(inst->src[i].file) {
465 case PROGRAM_STATE_VAR:
466 case PROGRAM_CONSTANT:
467 case PROGRAM_UNIFORM:
468 this->indirect_addr_consts = true;
469 break;
470 case PROGRAM_IMMEDIATE:
471 assert(!"immediates should not have indirect addressing");
472 break;
473 default:
474 break;
475 }
476 }
477 }
478 }
479
480 /*
481 * This section contains the double processing.
482 * GLSL just represents doubles as single channel values,
483 * however most HW and TGSI represent doubles as pairs of register channels.
484 *
485 * so we have to fixup destination writemask/index and src swizzle/indexes.
486 * dest writemasks need to translate from single channel write mask
487 * to a dual-channel writemask, but also need to modify the index,
488 * if we are touching the Z,W fields in the pre-translated writemask.
489 *
490 * src channels have similiar index modifications along with swizzle
491 * changes to we pick the XY, ZW pairs from the correct index.
492 *
493 * GLSL [0].x -> TGSI [0].xy
494 * GLSL [0].y -> TGSI [0].zw
495 * GLSL [0].z -> TGSI [1].xy
496 * GLSL [0].w -> TGSI [1].zw
497 */
498 for (j = 0; j < 2; j++) {
499 dst_is_64bit[j] = glsl_base_type_is_64bit(inst->dst[j].type);
500 if (!dst_is_64bit[j] && inst->dst[j].file == PROGRAM_OUTPUT && inst->dst[j].type == GLSL_TYPE_ARRAY) {
501 enum glsl_base_type type = find_array_type(this->outputs, this->num_outputs, inst->dst[j].array_id);
502 if (glsl_base_type_is_64bit(type))
503 dst_is_64bit[j] = true;
504 }
505 }
506
507 if (dst_is_64bit[0] || dst_is_64bit[1] ||
508 glsl_base_type_is_64bit(inst->src[0].type)) {
509 glsl_to_tgsi_instruction *dinst = NULL;
510 int initial_src_swz[4], initial_src_idx[4];
511 int initial_dst_idx[2], initial_dst_writemask[2];
512 /* select the writemask for dst0 or dst1 */
513 unsigned writemask = inst->dst[1].file == PROGRAM_UNDEFINED ? inst->dst[0].writemask : inst->dst[1].writemask;
514
515 /* copy out the writemask, index and swizzles for all src/dsts. */
516 for (j = 0; j < 2; j++) {
517 initial_dst_writemask[j] = inst->dst[j].writemask;
518 initial_dst_idx[j] = inst->dst[j].index;
519 }
520
521 for (j = 0; j < 4; j++) {
522 initial_src_swz[j] = inst->src[j].swizzle;
523 initial_src_idx[j] = inst->src[j].index;
524 }
525
526 /*
527 * scan all the components in the dst writemask
528 * generate an instruction for each of them if required.
529 */
530 st_src_reg addr;
531 while (writemask) {
532
533 int i = u_bit_scan(&writemask);
534
535 /* before emitting the instruction, see if we have to adjust load / store
536 * address */
537 if (i > 1 && (inst->op == TGSI_OPCODE_LOAD || inst->op == TGSI_OPCODE_STORE) &&
538 addr.file == PROGRAM_UNDEFINED) {
539 /* We have to advance the buffer address by 16 */
540 addr = get_temp(glsl_type::uint_type);
541 emit_asm(ir, TGSI_OPCODE_UADD, st_dst_reg(addr),
542 inst->src[0], st_src_reg_for_int(16));
543 }
544
545 /* first time use previous instruction */
546 if (dinst == NULL) {
547 dinst = inst;
548 } else {
549 /* create a new instructions for subsequent attempts */
550 dinst = new(mem_ctx) glsl_to_tgsi_instruction();
551 *dinst = *inst;
552 dinst->next = NULL;
553 dinst->prev = NULL;
554 }
555 this->instructions.push_tail(dinst);
556 dinst->is_64bit_expanded = true;
557
558 /* modify the destination if we are splitting */
559 for (j = 0; j < 2; j++) {
560 if (dst_is_64bit[j]) {
561 dinst->dst[j].writemask = (i & 1) ? WRITEMASK_ZW : WRITEMASK_XY;
562 dinst->dst[j].index = initial_dst_idx[j];
563 if (i > 1) {
564 if (dinst->op == TGSI_OPCODE_LOAD || dinst->op == TGSI_OPCODE_STORE)
565 dinst->src[0] = addr;
566 if (dinst->op != TGSI_OPCODE_STORE)
567 dinst->dst[j].index++;
568 }
569 } else {
570 /* if we aren't writing to a double, just get the bit of the initial writemask
571 for this channel */
572 dinst->dst[j].writemask = initial_dst_writemask[j] & (1 << i);
573 }
574 }
575
576 /* modify the src registers */
577 for (j = 0; j < 4; j++) {
578 int swz = GET_SWZ(initial_src_swz[j], i);
579
580 if (glsl_base_type_is_64bit(dinst->src[j].type)) {
581 dinst->src[j].index = initial_src_idx[j];
582 if (swz > 1) {
583 dinst->src[j].double_reg2 = true;
584 dinst->src[j].index++;
585 }
586
587 if (swz & 1)
588 dinst->src[j].swizzle = MAKE_SWIZZLE4(SWIZZLE_Z, SWIZZLE_W, SWIZZLE_Z, SWIZZLE_W);
589 else
590 dinst->src[j].swizzle = MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, SWIZZLE_X, SWIZZLE_Y);
591
592 } else {
593 /* some opcodes are special case in what they use as sources
594 - [FUI]2D/[UI]2I64 is a float/[u]int src0, (D)LDEXP is integer src1 */
595 if (op == TGSI_OPCODE_F2D || op == TGSI_OPCODE_U2D || op == TGSI_OPCODE_I2D ||
596 op == TGSI_OPCODE_I2I64 || op == TGSI_OPCODE_U2I64 ||
597 op == TGSI_OPCODE_DLDEXP || op == TGSI_OPCODE_LDEXP ||
598 (op == TGSI_OPCODE_UCMP && dst_is_64bit[0])) {
599 dinst->src[j].swizzle = MAKE_SWIZZLE4(swz, swz, swz, swz);
600 }
601 }
602 }
603 }
604 inst = dinst;
605 } else {
606 this->instructions.push_tail(inst);
607 }
608
609
610 return inst;
611 }
612
613 glsl_to_tgsi_instruction *
emit_asm(ir_instruction * ir,unsigned op,st_dst_reg dst,st_src_reg src0,st_src_reg src1,st_src_reg src2,st_src_reg src3)614 glsl_to_tgsi_visitor::emit_asm(ir_instruction *ir, unsigned op,
615 st_dst_reg dst,
616 st_src_reg src0, st_src_reg src1,
617 st_src_reg src2, st_src_reg src3)
618 {
619 return emit_asm(ir, op, dst, undef_dst, src0, src1, src2, src3);
620 }
621
622 /**
623 * Determines whether to use an integer, unsigned integer, or float opcode
624 * based on the operands and input opcode, then emits the result.
625 */
626 unsigned
get_opcode(unsigned op,st_dst_reg dst,st_src_reg src0,st_src_reg src1)627 glsl_to_tgsi_visitor::get_opcode(unsigned op,
628 st_dst_reg dst,
629 st_src_reg src0, st_src_reg src1)
630 {
631 enum glsl_base_type type = GLSL_TYPE_FLOAT;
632
633 if (op == TGSI_OPCODE_MOV)
634 return op;
635
636 assert(src0.type != GLSL_TYPE_ARRAY);
637 assert(src0.type != GLSL_TYPE_STRUCT);
638 assert(src1.type != GLSL_TYPE_ARRAY);
639 assert(src1.type != GLSL_TYPE_STRUCT);
640
641 if (is_resource_instruction(op))
642 type = src1.type;
643 else if (src0.type == GLSL_TYPE_INT64 || src1.type == GLSL_TYPE_INT64)
644 type = GLSL_TYPE_INT64;
645 else if (src0.type == GLSL_TYPE_UINT64 || src1.type == GLSL_TYPE_UINT64)
646 type = GLSL_TYPE_UINT64;
647 else if (src0.type == GLSL_TYPE_DOUBLE || src1.type == GLSL_TYPE_DOUBLE)
648 type = GLSL_TYPE_DOUBLE;
649 else if (src0.type == GLSL_TYPE_FLOAT || src1.type == GLSL_TYPE_FLOAT)
650 type = GLSL_TYPE_FLOAT;
651 else if (native_integers)
652 type = src0.type == GLSL_TYPE_BOOL ? GLSL_TYPE_INT : src0.type;
653
654 #define case7(c, f, i, u, d, i64, ui64) \
655 case TGSI_OPCODE_##c: \
656 if (type == GLSL_TYPE_UINT64) \
657 op = TGSI_OPCODE_##ui64; \
658 else if (type == GLSL_TYPE_INT64) \
659 op = TGSI_OPCODE_##i64; \
660 else if (type == GLSL_TYPE_DOUBLE) \
661 op = TGSI_OPCODE_##d; \
662 else if (type == GLSL_TYPE_INT) \
663 op = TGSI_OPCODE_##i; \
664 else if (type == GLSL_TYPE_UINT) \
665 op = TGSI_OPCODE_##u; \
666 else \
667 op = TGSI_OPCODE_##f; \
668 break;
669
670 #define casecomp(c, f, i, u, d, i64, ui64) \
671 case TGSI_OPCODE_##c: \
672 if (type == GLSL_TYPE_INT64) \
673 op = TGSI_OPCODE_##i64; \
674 else if (type == GLSL_TYPE_UINT64) \
675 op = TGSI_OPCODE_##ui64; \
676 else if (type == GLSL_TYPE_DOUBLE) \
677 op = TGSI_OPCODE_##d; \
678 else if (type == GLSL_TYPE_INT || type == GLSL_TYPE_SUBROUTINE) \
679 op = TGSI_OPCODE_##i; \
680 else if (type == GLSL_TYPE_UINT) \
681 op = TGSI_OPCODE_##u; \
682 else if (native_integers) \
683 op = TGSI_OPCODE_##f; \
684 else \
685 op = TGSI_OPCODE_##c; \
686 break;
687
688 switch(op) {
689 /* Some instructions are initially selected without considering the type.
690 * This fixes the type:
691 *
692 * INIT FLOAT SINT UINT DOUBLE SINT64 UINT64
693 */
694 case7(ADD, ADD, UADD, UADD, DADD, U64ADD, U64ADD);
695 case7(CEIL, CEIL, LAST, LAST, DCEIL, LAST, LAST);
696 case7(DIV, DIV, IDIV, UDIV, DDIV, I64DIV, U64DIV);
697 case7(FMA, FMA, UMAD, UMAD, DFMA, LAST, LAST);
698 case7(FLR, FLR, LAST, LAST, DFLR, LAST, LAST);
699 case7(FRC, FRC, LAST, LAST, DFRAC, LAST, LAST);
700 case7(MUL, MUL, UMUL, UMUL, DMUL, U64MUL, U64MUL);
701 case7(MAD, MAD, UMAD, UMAD, DMAD, LAST, LAST);
702 case7(MAX, MAX, IMAX, UMAX, DMAX, I64MAX, U64MAX);
703 case7(MIN, MIN, IMIN, UMIN, DMIN, I64MIN, U64MIN);
704 case7(RCP, RCP, LAST, LAST, DRCP, LAST, LAST);
705 case7(ROUND, ROUND,LAST, LAST, DROUND, LAST, LAST);
706 case7(RSQ, RSQ, LAST, LAST, DRSQ, LAST, LAST);
707 case7(SQRT, SQRT, LAST, LAST, DSQRT, LAST, LAST);
708 case7(SSG, SSG, ISSG, ISSG, DSSG, I64SSG, I64SSG);
709 case7(TRUNC, TRUNC,LAST, LAST, DTRUNC, LAST, LAST);
710
711 case7(MOD, LAST, MOD, UMOD, LAST, I64MOD, U64MOD);
712 case7(SHL, LAST, SHL, SHL, LAST, U64SHL, U64SHL);
713 case7(IBFE, LAST, IBFE, UBFE, LAST, LAST, LAST);
714 case7(IMSB, LAST, IMSB, UMSB, LAST, LAST, LAST);
715 case7(IMUL_HI, LAST, IMUL_HI, UMUL_HI, LAST, LAST, LAST);
716 case7(ISHR, LAST, ISHR, USHR, LAST, I64SHR, U64SHR);
717 case7(ATOMIMAX,LAST, ATOMIMAX,ATOMUMAX,LAST, LAST, LAST);
718 case7(ATOMIMIN,LAST, ATOMIMIN,ATOMUMIN,LAST, LAST, LAST);
719
720 casecomp(SEQ, FSEQ, USEQ, USEQ, DSEQ, U64SEQ, U64SEQ);
721 casecomp(SNE, FSNE, USNE, USNE, DSNE, U64SNE, U64SNE);
722 casecomp(SGE, FSGE, ISGE, USGE, DSGE, I64SGE, U64SGE);
723 casecomp(SLT, FSLT, ISLT, USLT, DSLT, I64SLT, U64SLT);
724
725 default: break;
726 }
727
728 assert(op != TGSI_OPCODE_LAST);
729 return op;
730 }
731
732 glsl_to_tgsi_instruction *
emit_dp(ir_instruction * ir,st_dst_reg dst,st_src_reg src0,st_src_reg src1,unsigned elements)733 glsl_to_tgsi_visitor::emit_dp(ir_instruction *ir,
734 st_dst_reg dst, st_src_reg src0, st_src_reg src1,
735 unsigned elements)
736 {
737 static const unsigned dot_opcodes[] = {
738 TGSI_OPCODE_DP2, TGSI_OPCODE_DP3, TGSI_OPCODE_DP4
739 };
740
741 return emit_asm(ir, dot_opcodes[elements - 2], dst, src0, src1);
742 }
743
744 /**
745 * Emits TGSI scalar opcodes to produce unique answers across channels.
746 *
747 * Some TGSI opcodes are scalar-only, like ARB_fp/vp. The src X
748 * channel determines the result across all channels. So to do a vec4
749 * of this operation, we want to emit a scalar per source channel used
750 * to produce dest channels.
751 */
752 void
emit_scalar(ir_instruction * ir,unsigned op,st_dst_reg dst,st_src_reg orig_src0,st_src_reg orig_src1)753 glsl_to_tgsi_visitor::emit_scalar(ir_instruction *ir, unsigned op,
754 st_dst_reg dst,
755 st_src_reg orig_src0, st_src_reg orig_src1)
756 {
757 int i, j;
758 int done_mask = ~dst.writemask;
759
760 /* TGSI RCP is a scalar operation splatting results to all channels,
761 * like ARB_fp/vp. So emit as many RCPs as necessary to cover our
762 * dst channels.
763 */
764 for (i = 0; i < 4; i++) {
765 GLuint this_mask = (1 << i);
766 st_src_reg src0 = orig_src0;
767 st_src_reg src1 = orig_src1;
768
769 if (done_mask & this_mask)
770 continue;
771
772 GLuint src0_swiz = GET_SWZ(src0.swizzle, i);
773 GLuint src1_swiz = GET_SWZ(src1.swizzle, i);
774 for (j = i + 1; j < 4; j++) {
775 /* If there is another enabled component in the destination that is
776 * derived from the same inputs, generate its value on this pass as
777 * well.
778 */
779 if (!(done_mask & (1 << j)) &&
780 GET_SWZ(src0.swizzle, j) == src0_swiz &&
781 GET_SWZ(src1.swizzle, j) == src1_swiz) {
782 this_mask |= (1 << j);
783 }
784 }
785 src0.swizzle = MAKE_SWIZZLE4(src0_swiz, src0_swiz,
786 src0_swiz, src0_swiz);
787 src1.swizzle = MAKE_SWIZZLE4(src1_swiz, src1_swiz,
788 src1_swiz, src1_swiz);
789
790 dst.writemask = this_mask;
791 emit_asm(ir, op, dst, src0, src1);
792 done_mask |= this_mask;
793 }
794 }
795
796 void
emit_scalar(ir_instruction * ir,unsigned op,st_dst_reg dst,st_src_reg src0)797 glsl_to_tgsi_visitor::emit_scalar(ir_instruction *ir, unsigned op,
798 st_dst_reg dst, st_src_reg src0)
799 {
800 st_src_reg undef = undef_src;
801
802 undef.swizzle = SWIZZLE_XXXX;
803
804 emit_scalar(ir, op, dst, src0, undef);
805 }
806
807 void
emit_arl(ir_instruction * ir,st_dst_reg dst,st_src_reg src0)808 glsl_to_tgsi_visitor::emit_arl(ir_instruction *ir,
809 st_dst_reg dst, st_src_reg src0)
810 {
811 int op = TGSI_OPCODE_ARL;
812
813 if (src0.type == GLSL_TYPE_INT || src0.type == GLSL_TYPE_UINT) {
814 if (!this->need_uarl && src0.is_legal_tgsi_address_operand())
815 return;
816
817 op = TGSI_OPCODE_UARL;
818 }
819
820 assert(dst.file == PROGRAM_ADDRESS);
821 if (dst.index >= this->num_address_regs)
822 this->num_address_regs = dst.index + 1;
823
824 emit_asm(NULL, op, dst, src0);
825 }
826
827 int
add_constant(gl_register_file file,gl_constant_value values[8],int size,int datatype,uint16_t * swizzle_out)828 glsl_to_tgsi_visitor::add_constant(gl_register_file file,
829 gl_constant_value values[8], int size, int datatype,
830 uint16_t *swizzle_out)
831 {
832 if (file == PROGRAM_CONSTANT) {
833 GLuint swizzle = swizzle_out ? *swizzle_out : 0;
834 int result = _mesa_add_typed_unnamed_constant(this->prog->Parameters, values,
835 size, datatype, &swizzle);
836 if (swizzle_out)
837 *swizzle_out = swizzle;
838 return result;
839 }
840
841 assert(file == PROGRAM_IMMEDIATE);
842
843 int index = 0;
844 immediate_storage *entry;
845 int size32 = size * ((datatype == GL_DOUBLE ||
846 datatype == GL_INT64_ARB ||
847 datatype == GL_UNSIGNED_INT64_ARB)? 2 : 1);
848 int i;
849
850 /* Search immediate storage to see if we already have an identical
851 * immediate that we can use instead of adding a duplicate entry.
852 */
853 foreach_in_list(immediate_storage, entry, &this->immediates) {
854 immediate_storage *tmp = entry;
855
856 for (i = 0; i * 4 < size32; i++) {
857 int slot_size = MIN2(size32 - (i * 4), 4);
858 if (tmp->type != datatype || tmp->size32 != slot_size)
859 break;
860 if (memcmp(tmp->values, &values[i * 4],
861 slot_size * sizeof(gl_constant_value)))
862 break;
863
864 /* Everything matches, keep going until the full size is matched */
865 tmp = (immediate_storage *)tmp->next;
866 }
867
868 /* The full value matched */
869 if (i * 4 >= size32)
870 return index;
871
872 index++;
873 }
874
875 for (i = 0; i * 4 < size32; i++) {
876 int slot_size = MIN2(size32 - (i * 4), 4);
877 /* Add this immediate to the list. */
878 entry = new(mem_ctx) immediate_storage(&values[i * 4], slot_size, datatype);
879 this->immediates.push_tail(entry);
880 this->num_immediates++;
881 }
882 return index;
883 }
884
885 st_src_reg
st_src_reg_for_float(float val)886 glsl_to_tgsi_visitor::st_src_reg_for_float(float val)
887 {
888 st_src_reg src(PROGRAM_IMMEDIATE, -1, GLSL_TYPE_FLOAT);
889 union gl_constant_value uval;
890
891 uval.f = val;
892 src.index = add_constant(src.file, &uval, 1, GL_FLOAT, &src.swizzle);
893
894 return src;
895 }
896
897 st_src_reg
st_src_reg_for_double(double val)898 glsl_to_tgsi_visitor::st_src_reg_for_double(double val)
899 {
900 st_src_reg src(PROGRAM_IMMEDIATE, -1, GLSL_TYPE_DOUBLE);
901 union gl_constant_value uval[2];
902
903 memcpy(uval, &val, sizeof(uval));
904 src.index = add_constant(src.file, uval, 1, GL_DOUBLE, &src.swizzle);
905 src.swizzle = MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, SWIZZLE_X, SWIZZLE_Y);
906 return src;
907 }
908
909 st_src_reg
st_src_reg_for_int(int val)910 glsl_to_tgsi_visitor::st_src_reg_for_int(int val)
911 {
912 st_src_reg src(PROGRAM_IMMEDIATE, -1, GLSL_TYPE_INT);
913 union gl_constant_value uval;
914
915 assert(native_integers);
916
917 uval.i = val;
918 src.index = add_constant(src.file, &uval, 1, GL_INT, &src.swizzle);
919
920 return src;
921 }
922
923 st_src_reg
st_src_reg_for_int64(int64_t val)924 glsl_to_tgsi_visitor::st_src_reg_for_int64(int64_t val)
925 {
926 st_src_reg src(PROGRAM_IMMEDIATE, -1, GLSL_TYPE_INT64);
927 union gl_constant_value uval[2];
928
929 memcpy(uval, &val, sizeof(uval));
930 src.index = add_constant(src.file, uval, 1, GL_DOUBLE, &src.swizzle);
931 src.swizzle = MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, SWIZZLE_X, SWIZZLE_Y);
932
933 return src;
934 }
935
936 st_src_reg
st_src_reg_for_type(enum glsl_base_type type,int val)937 glsl_to_tgsi_visitor::st_src_reg_for_type(enum glsl_base_type type, int val)
938 {
939 if (native_integers)
940 return type == GLSL_TYPE_FLOAT ? st_src_reg_for_float(val) :
941 st_src_reg_for_int(val);
942 else
943 return st_src_reg_for_float(val);
944 }
945
946 static int
attrib_type_size(const struct glsl_type * type,bool is_vs_input)947 attrib_type_size(const struct glsl_type *type, bool is_vs_input)
948 {
949 return type->count_attribute_slots(is_vs_input);
950 }
951
952 static int
type_size(const struct glsl_type * type)953 type_size(const struct glsl_type *type)
954 {
955 return type->count_attribute_slots(false);
956 }
957
958 static void
add_buffer_to_load_and_stores(glsl_to_tgsi_instruction * inst,st_src_reg * buf,exec_list * instructions,ir_constant * access)959 add_buffer_to_load_and_stores(glsl_to_tgsi_instruction *inst, st_src_reg *buf,
960 exec_list *instructions, ir_constant *access)
961 {
962 /**
963 * emit_asm() might have actually split the op into pieces, e.g. for
964 * double stores. We have to go back and fix up all the generated ops.
965 */
966 unsigned op = inst->op;
967 do {
968 inst->resource = *buf;
969 if (access)
970 inst->buffer_access = access->value.u[0];
971
972 if (inst == instructions->get_head_raw())
973 break;
974 inst = (glsl_to_tgsi_instruction *)inst->get_prev();
975
976 if (inst->op == TGSI_OPCODE_UADD) {
977 if (inst == instructions->get_head_raw())
978 break;
979 inst = (glsl_to_tgsi_instruction *)inst->get_prev();
980 }
981 } while (inst->op == op && inst->resource.file == PROGRAM_UNDEFINED);
982 }
983
984 /**
985 * If the given GLSL type is an array or matrix or a structure containing
986 * an array/matrix member, return true. Else return false.
987 *
988 * This is used to determine which kind of temp storage (PROGRAM_TEMPORARY
989 * or PROGRAM_ARRAY) should be used for variables of this type. Anytime
990 * we have an array that might be indexed with a variable, we need to use
991 * the later storage type.
992 */
993 static bool
type_has_array_or_matrix(const glsl_type * type)994 type_has_array_or_matrix(const glsl_type *type)
995 {
996 if (type->is_array() || type->is_matrix())
997 return true;
998
999 if (type->is_record()) {
1000 for (unsigned i = 0; i < type->length; i++) {
1001 if (type_has_array_or_matrix(type->fields.structure[i].type)) {
1002 return true;
1003 }
1004 }
1005 }
1006
1007 return false;
1008 }
1009
1010
1011 /**
1012 * In the initial pass of codegen, we assign temporary numbers to
1013 * intermediate results. (not SSA -- variable assignments will reuse
1014 * storage).
1015 */
1016 st_src_reg
get_temp(const glsl_type * type)1017 glsl_to_tgsi_visitor::get_temp(const glsl_type *type)
1018 {
1019 st_src_reg src;
1020
1021 src.type = native_integers ? type->base_type : GLSL_TYPE_FLOAT;
1022 src.reladdr = NULL;
1023 src.negate = 0;
1024 src.abs = 0;
1025
1026 if (!options->EmitNoIndirectTemp && type_has_array_or_matrix(type)) {
1027 if (next_array >= max_num_arrays) {
1028 max_num_arrays += 32;
1029 array_sizes = (unsigned*)
1030 realloc(array_sizes, sizeof(array_sizes[0]) * max_num_arrays);
1031 }
1032
1033 src.file = PROGRAM_ARRAY;
1034 src.index = 0;
1035 src.array_id = next_array + 1;
1036 array_sizes[next_array] = type_size(type);
1037 ++next_array;
1038
1039 } else {
1040 src.file = PROGRAM_TEMPORARY;
1041 src.index = next_temp;
1042 next_temp += type_size(type);
1043 }
1044
1045 if (type->is_array() || type->is_record()) {
1046 src.swizzle = SWIZZLE_NOOP;
1047 } else {
1048 src.swizzle = swizzle_for_size(type->vector_elements);
1049 }
1050
1051 return src;
1052 }
1053
1054 variable_storage *
find_variable_storage(ir_variable * var)1055 glsl_to_tgsi_visitor::find_variable_storage(ir_variable *var)
1056 {
1057 struct hash_entry *entry;
1058
1059 entry = _mesa_hash_table_search(this->variables, var);
1060 if (!entry)
1061 return NULL;
1062
1063 return (variable_storage *)entry->data;
1064 }
1065
1066 void
visit(ir_variable * ir)1067 glsl_to_tgsi_visitor::visit(ir_variable *ir)
1068 {
1069 if (strcmp(ir->name, "gl_FragCoord") == 0) {
1070 this->prog->OriginUpperLeft = ir->data.origin_upper_left;
1071 this->prog->PixelCenterInteger = ir->data.pixel_center_integer;
1072 }
1073
1074 if (ir->data.mode == ir_var_uniform && strncmp(ir->name, "gl_", 3) == 0) {
1075 unsigned int i;
1076 const ir_state_slot *const slots = ir->get_state_slots();
1077 assert(slots != NULL);
1078
1079 /* Check if this statevar's setup in the STATE file exactly
1080 * matches how we'll want to reference it as a
1081 * struct/array/whatever. If not, then we need to move it into
1082 * temporary storage and hope that it'll get copy-propagated
1083 * out.
1084 */
1085 for (i = 0; i < ir->get_num_state_slots(); i++) {
1086 if (slots[i].swizzle != SWIZZLE_XYZW) {
1087 break;
1088 }
1089 }
1090
1091 variable_storage *storage;
1092 st_dst_reg dst;
1093 if (i == ir->get_num_state_slots()) {
1094 /* We'll set the index later. */
1095 storage = new(mem_ctx) variable_storage(ir, PROGRAM_STATE_VAR, -1);
1096
1097 _mesa_hash_table_insert(this->variables, ir, storage);
1098
1099 dst = undef_dst;
1100 } else {
1101 /* The variable_storage constructor allocates slots based on the size
1102 * of the type. However, this had better match the number of state
1103 * elements that we're going to copy into the new temporary.
1104 */
1105 assert((int) ir->get_num_state_slots() == type_size(ir->type));
1106
1107 dst = st_dst_reg(get_temp(ir->type));
1108
1109 storage = new(mem_ctx) variable_storage(ir, dst.file, dst.index,
1110 dst.array_id);
1111
1112 _mesa_hash_table_insert(this->variables, ir, storage);
1113 }
1114
1115
1116 for (unsigned int i = 0; i < ir->get_num_state_slots(); i++) {
1117 int index = _mesa_add_state_reference(this->prog->Parameters,
1118 (gl_state_index *)slots[i].tokens);
1119
1120 if (storage->file == PROGRAM_STATE_VAR) {
1121 if (storage->index == -1) {
1122 storage->index = index;
1123 } else {
1124 assert(index == storage->index + (int)i);
1125 }
1126 } else {
1127 /* We use GLSL_TYPE_FLOAT here regardless of the actual type of
1128 * the data being moved since MOV does not care about the type of
1129 * data it is moving, and we don't want to declare registers with
1130 * array or struct types.
1131 */
1132 st_src_reg src(PROGRAM_STATE_VAR, index, GLSL_TYPE_FLOAT);
1133 src.swizzle = slots[i].swizzle;
1134 emit_asm(ir, TGSI_OPCODE_MOV, dst, src);
1135 /* even a float takes up a whole vec4 reg in a struct/array. */
1136 dst.index++;
1137 }
1138 }
1139
1140 if (storage->file == PROGRAM_TEMPORARY &&
1141 dst.index != storage->index + (int) ir->get_num_state_slots()) {
1142 fail_link(this->shader_program,
1143 "failed to load builtin uniform `%s' (%d/%d regs loaded)\n",
1144 ir->name, dst.index - storage->index,
1145 type_size(ir->type));
1146 }
1147 }
1148 }
1149
1150 void
visit(ir_loop * ir)1151 glsl_to_tgsi_visitor::visit(ir_loop *ir)
1152 {
1153 emit_asm(NULL, TGSI_OPCODE_BGNLOOP);
1154
1155 visit_exec_list(&ir->body_instructions, this);
1156
1157 emit_asm(NULL, TGSI_OPCODE_ENDLOOP);
1158 }
1159
1160 void
visit(ir_loop_jump * ir)1161 glsl_to_tgsi_visitor::visit(ir_loop_jump *ir)
1162 {
1163 switch (ir->mode) {
1164 case ir_loop_jump::jump_break:
1165 emit_asm(NULL, TGSI_OPCODE_BRK);
1166 break;
1167 case ir_loop_jump::jump_continue:
1168 emit_asm(NULL, TGSI_OPCODE_CONT);
1169 break;
1170 }
1171 }
1172
1173
1174 void
visit(ir_function_signature * ir)1175 glsl_to_tgsi_visitor::visit(ir_function_signature *ir)
1176 {
1177 assert(0);
1178 (void)ir;
1179 }
1180
1181 void
visit(ir_function * ir)1182 glsl_to_tgsi_visitor::visit(ir_function *ir)
1183 {
1184 /* Ignore function bodies other than main() -- we shouldn't see calls to
1185 * them since they should all be inlined before we get to glsl_to_tgsi.
1186 */
1187 if (strcmp(ir->name, "main") == 0) {
1188 const ir_function_signature *sig;
1189 exec_list empty;
1190
1191 sig = ir->matching_signature(NULL, &empty, false);
1192
1193 assert(sig);
1194
1195 foreach_in_list(ir_instruction, ir, &sig->body) {
1196 ir->accept(this);
1197 }
1198 }
1199 }
1200
1201 bool
try_emit_mad(ir_expression * ir,int mul_operand)1202 glsl_to_tgsi_visitor::try_emit_mad(ir_expression *ir, int mul_operand)
1203 {
1204 int nonmul_operand = 1 - mul_operand;
1205 st_src_reg a, b, c;
1206 st_dst_reg result_dst;
1207
1208 ir_expression *expr = ir->operands[mul_operand]->as_expression();
1209 if (!expr || expr->operation != ir_binop_mul)
1210 return false;
1211
1212 expr->operands[0]->accept(this);
1213 a = this->result;
1214 expr->operands[1]->accept(this);
1215 b = this->result;
1216 ir->operands[nonmul_operand]->accept(this);
1217 c = this->result;
1218
1219 this->result = get_temp(ir->type);
1220 result_dst = st_dst_reg(this->result);
1221 result_dst.writemask = (1 << ir->type->vector_elements) - 1;
1222 emit_asm(ir, TGSI_OPCODE_MAD, result_dst, a, b, c);
1223
1224 return true;
1225 }
1226
1227 /**
1228 * Emit MAD(a, -b, a) instead of AND(a, NOT(b))
1229 *
1230 * The logic values are 1.0 for true and 0.0 for false. Logical-and is
1231 * implemented using multiplication, and logical-or is implemented using
1232 * addition. Logical-not can be implemented as (true - x), or (1.0 - x).
1233 * As result, the logical expression (a & !b) can be rewritten as:
1234 *
1235 * - a * !b
1236 * - a * (1 - b)
1237 * - (a * 1) - (a * b)
1238 * - a + -(a * b)
1239 * - a + (a * -b)
1240 *
1241 * This final expression can be implemented as a single MAD(a, -b, a)
1242 * instruction.
1243 */
1244 bool
try_emit_mad_for_and_not(ir_expression * ir,int try_operand)1245 glsl_to_tgsi_visitor::try_emit_mad_for_and_not(ir_expression *ir, int try_operand)
1246 {
1247 const int other_operand = 1 - try_operand;
1248 st_src_reg a, b;
1249
1250 ir_expression *expr = ir->operands[try_operand]->as_expression();
1251 if (!expr || expr->operation != ir_unop_logic_not)
1252 return false;
1253
1254 ir->operands[other_operand]->accept(this);
1255 a = this->result;
1256 expr->operands[0]->accept(this);
1257 b = this->result;
1258
1259 b.negate = ~b.negate;
1260
1261 this->result = get_temp(ir->type);
1262 emit_asm(ir, TGSI_OPCODE_MAD, st_dst_reg(this->result), a, b, a);
1263
1264 return true;
1265 }
1266
1267 void
reladdr_to_temp(ir_instruction * ir,st_src_reg * reg,int * num_reladdr)1268 glsl_to_tgsi_visitor::reladdr_to_temp(ir_instruction *ir,
1269 st_src_reg *reg, int *num_reladdr)
1270 {
1271 if (!reg->reladdr && !reg->reladdr2)
1272 return;
1273
1274 if (reg->reladdr) emit_arl(ir, address_reg, *reg->reladdr);
1275 if (reg->reladdr2) emit_arl(ir, address_reg2, *reg->reladdr2);
1276
1277 if (*num_reladdr != 1) {
1278 st_src_reg temp = get_temp(glsl_type::get_instance(reg->type, 4, 1));
1279
1280 emit_asm(ir, TGSI_OPCODE_MOV, st_dst_reg(temp), *reg);
1281 *reg = temp;
1282 }
1283
1284 (*num_reladdr)--;
1285 }
1286
1287 void
visit(ir_expression * ir)1288 glsl_to_tgsi_visitor::visit(ir_expression *ir)
1289 {
1290 st_src_reg op[ARRAY_SIZE(ir->operands)];
1291
1292 /* Quick peephole: Emit MAD(a, b, c) instead of ADD(MUL(a, b), c)
1293 */
1294 if (!this->precise && ir->operation == ir_binop_add) {
1295 if (try_emit_mad(ir, 1))
1296 return;
1297 if (try_emit_mad(ir, 0))
1298 return;
1299 }
1300
1301 /* Quick peephole: Emit OPCODE_MAD(-a, -b, a) instead of AND(a, NOT(b))
1302 */
1303 if (!native_integers && ir->operation == ir_binop_logic_and) {
1304 if (try_emit_mad_for_and_not(ir, 1))
1305 return;
1306 if (try_emit_mad_for_and_not(ir, 0))
1307 return;
1308 }
1309
1310 if (ir->operation == ir_quadop_vector)
1311 assert(!"ir_quadop_vector should have been lowered");
1312
1313 for (unsigned int operand = 0; operand < ir->num_operands; operand++) {
1314 this->result.file = PROGRAM_UNDEFINED;
1315 ir->operands[operand]->accept(this);
1316 if (this->result.file == PROGRAM_UNDEFINED) {
1317 printf("Failed to get tree for expression operand:\n");
1318 ir->operands[operand]->print();
1319 printf("\n");
1320 exit(1);
1321 }
1322 op[operand] = this->result;
1323
1324 /* Matrix expression operands should have been broken down to vector
1325 * operations already.
1326 */
1327 assert(!ir->operands[operand]->type->is_matrix());
1328 }
1329
1330 visit_expression(ir, op);
1331 }
1332
1333 /* The non-recursive part of the expression visitor lives in a separate
1334 * function and should be prevented from being inlined, to avoid a stack
1335 * explosion when deeply nested expressions are visited.
1336 */
1337 void
visit_expression(ir_expression * ir,st_src_reg * op)1338 glsl_to_tgsi_visitor::visit_expression(ir_expression* ir, st_src_reg *op)
1339 {
1340 st_src_reg result_src;
1341 st_dst_reg result_dst;
1342
1343 int vector_elements = ir->operands[0]->type->vector_elements;
1344 if (ir->operands[1] &&
1345 ir->operation != ir_binop_interpolate_at_offset &&
1346 ir->operation != ir_binop_interpolate_at_sample) {
1347 st_src_reg *swz_op = NULL;
1348 if (vector_elements > ir->operands[1]->type->vector_elements) {
1349 assert(ir->operands[1]->type->vector_elements == 1);
1350 swz_op = &op[1];
1351 } else if (vector_elements < ir->operands[1]->type->vector_elements) {
1352 assert(ir->operands[0]->type->vector_elements == 1);
1353 swz_op = &op[0];
1354 }
1355 if (swz_op) {
1356 uint16_t swizzle_x = GET_SWZ(swz_op->swizzle, 0);
1357 swz_op->swizzle = MAKE_SWIZZLE4(swizzle_x, swizzle_x,
1358 swizzle_x, swizzle_x);
1359 }
1360 vector_elements = MAX2(vector_elements,
1361 ir->operands[1]->type->vector_elements);
1362 }
1363 if (ir->operands[2] &&
1364 ir->operands[2]->type->vector_elements != vector_elements) {
1365 /* This can happen with ir_triop_lrp, i.e. glsl mix */
1366 assert(ir->operands[2]->type->vector_elements == 1);
1367 uint16_t swizzle_x = GET_SWZ(op[2].swizzle, 0);
1368 op[2].swizzle = MAKE_SWIZZLE4(swizzle_x, swizzle_x,
1369 swizzle_x, swizzle_x);
1370 }
1371
1372 this->result.file = PROGRAM_UNDEFINED;
1373
1374 /* Storage for our result. Ideally for an assignment we'd be using
1375 * the actual storage for the result here, instead.
1376 */
1377 result_src = get_temp(ir->type);
1378 /* convenience for the emit functions below. */
1379 result_dst = st_dst_reg(result_src);
1380 /* Limit writes to the channels that will be used by result_src later.
1381 * This does limit this temp's use as a temporary for multi-instruction
1382 * sequences.
1383 */
1384 result_dst.writemask = (1 << ir->type->vector_elements) - 1;
1385
1386 switch (ir->operation) {
1387 case ir_unop_logic_not:
1388 if (result_dst.type != GLSL_TYPE_FLOAT)
1389 emit_asm(ir, TGSI_OPCODE_NOT, result_dst, op[0]);
1390 else {
1391 /* Previously 'SEQ dst, src, 0.0' was used for this. However, many
1392 * older GPUs implement SEQ using multiple instructions (i915 uses two
1393 * SGE instructions and a MUL instruction). Since our logic values are
1394 * 0.0 and 1.0, 1-x also implements !x.
1395 */
1396 op[0].negate = ~op[0].negate;
1397 emit_asm(ir, TGSI_OPCODE_ADD, result_dst, op[0], st_src_reg_for_float(1.0));
1398 }
1399 break;
1400 case ir_unop_neg:
1401 if (result_dst.type == GLSL_TYPE_INT64 || result_dst.type == GLSL_TYPE_UINT64)
1402 emit_asm(ir, TGSI_OPCODE_I64NEG, result_dst, op[0]);
1403 else if (result_dst.type == GLSL_TYPE_INT || result_dst.type == GLSL_TYPE_UINT)
1404 emit_asm(ir, TGSI_OPCODE_INEG, result_dst, op[0]);
1405 else if (result_dst.type == GLSL_TYPE_DOUBLE)
1406 emit_asm(ir, TGSI_OPCODE_DNEG, result_dst, op[0]);
1407 else {
1408 op[0].negate = ~op[0].negate;
1409 result_src = op[0];
1410 }
1411 break;
1412 case ir_unop_subroutine_to_int:
1413 emit_asm(ir, TGSI_OPCODE_MOV, result_dst, op[0]);
1414 break;
1415 case ir_unop_abs:
1416 if (result_dst.type == GLSL_TYPE_FLOAT)
1417 emit_asm(ir, TGSI_OPCODE_MOV, result_dst, op[0].get_abs());
1418 else if (result_dst.type == GLSL_TYPE_DOUBLE)
1419 emit_asm(ir, TGSI_OPCODE_DABS, result_dst, op[0]);
1420 else if (result_dst.type == GLSL_TYPE_INT64 || result_dst.type == GLSL_TYPE_UINT64)
1421 emit_asm(ir, TGSI_OPCODE_I64ABS, result_dst, op[0]);
1422 else
1423 emit_asm(ir, TGSI_OPCODE_IABS, result_dst, op[0]);
1424 break;
1425 case ir_unop_sign:
1426 emit_asm(ir, TGSI_OPCODE_SSG, result_dst, op[0]);
1427 break;
1428 case ir_unop_rcp:
1429 emit_scalar(ir, TGSI_OPCODE_RCP, result_dst, op[0]);
1430 break;
1431
1432 case ir_unop_exp2:
1433 emit_scalar(ir, TGSI_OPCODE_EX2, result_dst, op[0]);
1434 break;
1435 case ir_unop_exp:
1436 assert(!"not reached: should be handled by exp_to_exp2");
1437 break;
1438 case ir_unop_log:
1439 assert(!"not reached: should be handled by log_to_log2");
1440 break;
1441 case ir_unop_log2:
1442 emit_scalar(ir, TGSI_OPCODE_LG2, result_dst, op[0]);
1443 break;
1444 case ir_unop_sin:
1445 emit_scalar(ir, TGSI_OPCODE_SIN, result_dst, op[0]);
1446 break;
1447 case ir_unop_cos:
1448 emit_scalar(ir, TGSI_OPCODE_COS, result_dst, op[0]);
1449 break;
1450 case ir_unop_saturate: {
1451 glsl_to_tgsi_instruction *inst;
1452 inst = emit_asm(ir, TGSI_OPCODE_MOV, result_dst, op[0]);
1453 inst->saturate = true;
1454 break;
1455 }
1456
1457 case ir_unop_dFdx:
1458 case ir_unop_dFdx_coarse:
1459 emit_asm(ir, TGSI_OPCODE_DDX, result_dst, op[0]);
1460 break;
1461 case ir_unop_dFdx_fine:
1462 emit_asm(ir, TGSI_OPCODE_DDX_FINE, result_dst, op[0]);
1463 break;
1464 case ir_unop_dFdy:
1465 case ir_unop_dFdy_coarse:
1466 case ir_unop_dFdy_fine:
1467 {
1468 /* The X component contains 1 or -1 depending on whether the framebuffer
1469 * is a FBO or the window system buffer, respectively.
1470 * It is then multiplied with the source operand of DDY.
1471 */
1472 static const gl_state_index transform_y_state[STATE_LENGTH]
1473 = { STATE_INTERNAL, STATE_FB_WPOS_Y_TRANSFORM };
1474
1475 unsigned transform_y_index =
1476 _mesa_add_state_reference(this->prog->Parameters,
1477 transform_y_state);
1478
1479 st_src_reg transform_y = st_src_reg(PROGRAM_STATE_VAR,
1480 transform_y_index,
1481 glsl_type::vec4_type);
1482 transform_y.swizzle = SWIZZLE_XXXX;
1483
1484 st_src_reg temp = get_temp(glsl_type::vec4_type);
1485
1486 emit_asm(ir, TGSI_OPCODE_MUL, st_dst_reg(temp), transform_y, op[0]);
1487 emit_asm(ir, ir->operation == ir_unop_dFdy_fine ?
1488 TGSI_OPCODE_DDY_FINE : TGSI_OPCODE_DDY, result_dst, temp);
1489 break;
1490 }
1491
1492 case ir_unop_frexp_sig:
1493 emit_asm(ir, TGSI_OPCODE_DFRACEXP, result_dst, undef_dst, op[0]);
1494 break;
1495
1496 case ir_unop_frexp_exp:
1497 emit_asm(ir, TGSI_OPCODE_DFRACEXP, undef_dst, result_dst, op[0]);
1498 break;
1499
1500 case ir_unop_noise: {
1501 /* At some point, a motivated person could add a better
1502 * implementation of noise. Currently not even the nvidia
1503 * binary drivers do anything more than this. In any case, the
1504 * place to do this is in the GL state tracker, not the poor
1505 * driver.
1506 */
1507 emit_asm(ir, TGSI_OPCODE_MOV, result_dst, st_src_reg_for_float(0.5));
1508 break;
1509 }
1510
1511 case ir_binop_add:
1512 emit_asm(ir, TGSI_OPCODE_ADD, result_dst, op[0], op[1]);
1513 break;
1514 case ir_binop_sub:
1515 op[1].negate = ~op[1].negate;
1516 emit_asm(ir, TGSI_OPCODE_ADD, result_dst, op[0], op[1]);
1517 break;
1518
1519 case ir_binop_mul:
1520 emit_asm(ir, TGSI_OPCODE_MUL, result_dst, op[0], op[1]);
1521 break;
1522 case ir_binop_div:
1523 emit_asm(ir, TGSI_OPCODE_DIV, result_dst, op[0], op[1]);
1524 break;
1525 case ir_binop_mod:
1526 if (result_dst.type == GLSL_TYPE_FLOAT)
1527 assert(!"ir_binop_mod should have been converted to b * fract(a/b)");
1528 else
1529 emit_asm(ir, TGSI_OPCODE_MOD, result_dst, op[0], op[1]);
1530 break;
1531
1532 case ir_binop_less:
1533 emit_asm(ir, TGSI_OPCODE_SLT, result_dst, op[0], op[1]);
1534 break;
1535 case ir_binop_gequal:
1536 emit_asm(ir, TGSI_OPCODE_SGE, result_dst, op[0], op[1]);
1537 break;
1538 case ir_binop_equal:
1539 emit_asm(ir, TGSI_OPCODE_SEQ, result_dst, op[0], op[1]);
1540 break;
1541 case ir_binop_nequal:
1542 emit_asm(ir, TGSI_OPCODE_SNE, result_dst, op[0], op[1]);
1543 break;
1544 case ir_binop_all_equal:
1545 /* "==" operator producing a scalar boolean. */
1546 if (ir->operands[0]->type->is_vector() ||
1547 ir->operands[1]->type->is_vector()) {
1548 st_src_reg temp = get_temp(native_integers ?
1549 glsl_type::uvec4_type :
1550 glsl_type::vec4_type);
1551
1552 if (native_integers) {
1553 st_dst_reg temp_dst = st_dst_reg(temp);
1554 st_src_reg temp1 = st_src_reg(temp), temp2 = st_src_reg(temp);
1555
1556 if (ir->operands[0]->type->is_boolean() &&
1557 ir->operands[1]->as_constant() &&
1558 ir->operands[1]->as_constant()->is_one()) {
1559 emit_asm(ir, TGSI_OPCODE_MOV, st_dst_reg(temp), op[0]);
1560 } else {
1561 emit_asm(ir, TGSI_OPCODE_SEQ, st_dst_reg(temp), op[0], op[1]);
1562 }
1563
1564 /* Emit 1-3 AND operations to combine the SEQ results. */
1565 switch (ir->operands[0]->type->vector_elements) {
1566 case 2:
1567 break;
1568 case 3:
1569 temp_dst.writemask = WRITEMASK_Y;
1570 temp1.swizzle = SWIZZLE_YYYY;
1571 temp2.swizzle = SWIZZLE_ZZZZ;
1572 emit_asm(ir, TGSI_OPCODE_AND, temp_dst, temp1, temp2);
1573 break;
1574 case 4:
1575 temp_dst.writemask = WRITEMASK_X;
1576 temp1.swizzle = SWIZZLE_XXXX;
1577 temp2.swizzle = SWIZZLE_YYYY;
1578 emit_asm(ir, TGSI_OPCODE_AND, temp_dst, temp1, temp2);
1579 temp_dst.writemask = WRITEMASK_Y;
1580 temp1.swizzle = SWIZZLE_ZZZZ;
1581 temp2.swizzle = SWIZZLE_WWWW;
1582 emit_asm(ir, TGSI_OPCODE_AND, temp_dst, temp1, temp2);
1583 }
1584
1585 temp1.swizzle = SWIZZLE_XXXX;
1586 temp2.swizzle = SWIZZLE_YYYY;
1587 emit_asm(ir, TGSI_OPCODE_AND, result_dst, temp1, temp2);
1588 } else {
1589 emit_asm(ir, TGSI_OPCODE_SNE, st_dst_reg(temp), op[0], op[1]);
1590
1591 /* After the dot-product, the value will be an integer on the
1592 * range [0,4]. Zero becomes 1.0, and positive values become zero.
1593 */
1594 emit_dp(ir, result_dst, temp, temp, vector_elements);
1595
1596 /* Negating the result of the dot-product gives values on the range
1597 * [-4, 0]. Zero becomes 1.0, and negative values become zero.
1598 * This is achieved using SGE.
1599 */
1600 st_src_reg sge_src = result_src;
1601 sge_src.negate = ~sge_src.negate;
1602 emit_asm(ir, TGSI_OPCODE_SGE, result_dst, sge_src, st_src_reg_for_float(0.0));
1603 }
1604 } else {
1605 emit_asm(ir, TGSI_OPCODE_SEQ, result_dst, op[0], op[1]);
1606 }
1607 break;
1608 case ir_binop_any_nequal:
1609 /* "!=" operator producing a scalar boolean. */
1610 if (ir->operands[0]->type->is_vector() ||
1611 ir->operands[1]->type->is_vector()) {
1612 st_src_reg temp = get_temp(native_integers ?
1613 glsl_type::uvec4_type :
1614 glsl_type::vec4_type);
1615 if (ir->operands[0]->type->is_boolean() &&
1616 ir->operands[1]->as_constant() &&
1617 ir->operands[1]->as_constant()->is_zero()) {
1618 emit_asm(ir, TGSI_OPCODE_MOV, st_dst_reg(temp), op[0]);
1619 } else {
1620 emit_asm(ir, TGSI_OPCODE_SNE, st_dst_reg(temp), op[0], op[1]);
1621 }
1622
1623 if (native_integers) {
1624 st_dst_reg temp_dst = st_dst_reg(temp);
1625 st_src_reg temp1 = st_src_reg(temp), temp2 = st_src_reg(temp);
1626
1627 /* Emit 1-3 OR operations to combine the SNE results. */
1628 switch (ir->operands[0]->type->vector_elements) {
1629 case 2:
1630 break;
1631 case 3:
1632 temp_dst.writemask = WRITEMASK_Y;
1633 temp1.swizzle = SWIZZLE_YYYY;
1634 temp2.swizzle = SWIZZLE_ZZZZ;
1635 emit_asm(ir, TGSI_OPCODE_OR, temp_dst, temp1, temp2);
1636 break;
1637 case 4:
1638 temp_dst.writemask = WRITEMASK_X;
1639 temp1.swizzle = SWIZZLE_XXXX;
1640 temp2.swizzle = SWIZZLE_YYYY;
1641 emit_asm(ir, TGSI_OPCODE_OR, temp_dst, temp1, temp2);
1642 temp_dst.writemask = WRITEMASK_Y;
1643 temp1.swizzle = SWIZZLE_ZZZZ;
1644 temp2.swizzle = SWIZZLE_WWWW;
1645 emit_asm(ir, TGSI_OPCODE_OR, temp_dst, temp1, temp2);
1646 }
1647
1648 temp1.swizzle = SWIZZLE_XXXX;
1649 temp2.swizzle = SWIZZLE_YYYY;
1650 emit_asm(ir, TGSI_OPCODE_OR, result_dst, temp1, temp2);
1651 } else {
1652 /* After the dot-product, the value will be an integer on the
1653 * range [0,4]. Zero stays zero, and positive values become 1.0.
1654 */
1655 glsl_to_tgsi_instruction *const dp =
1656 emit_dp(ir, result_dst, temp, temp, vector_elements);
1657 if (this->prog->Target == GL_FRAGMENT_PROGRAM_ARB) {
1658 /* The clamping to [0,1] can be done for free in the fragment
1659 * shader with a saturate.
1660 */
1661 dp->saturate = true;
1662 } else {
1663 /* Negating the result of the dot-product gives values on the range
1664 * [-4, 0]. Zero stays zero, and negative values become 1.0. This
1665 * achieved using SLT.
1666 */
1667 st_src_reg slt_src = result_src;
1668 slt_src.negate = ~slt_src.negate;
1669 emit_asm(ir, TGSI_OPCODE_SLT, result_dst, slt_src, st_src_reg_for_float(0.0));
1670 }
1671 }
1672 } else {
1673 emit_asm(ir, TGSI_OPCODE_SNE, result_dst, op[0], op[1]);
1674 }
1675 break;
1676
1677 case ir_binop_logic_xor:
1678 if (native_integers)
1679 emit_asm(ir, TGSI_OPCODE_XOR, result_dst, op[0], op[1]);
1680 else
1681 emit_asm(ir, TGSI_OPCODE_SNE, result_dst, op[0], op[1]);
1682 break;
1683
1684 case ir_binop_logic_or: {
1685 if (native_integers) {
1686 /* If integers are used as booleans, we can use an actual "or"
1687 * instruction.
1688 */
1689 assert(native_integers);
1690 emit_asm(ir, TGSI_OPCODE_OR, result_dst, op[0], op[1]);
1691 } else {
1692 /* After the addition, the value will be an integer on the
1693 * range [0,2]. Zero stays zero, and positive values become 1.0.
1694 */
1695 glsl_to_tgsi_instruction *add =
1696 emit_asm(ir, TGSI_OPCODE_ADD, result_dst, op[0], op[1]);
1697 if (this->prog->Target == GL_FRAGMENT_PROGRAM_ARB) {
1698 /* The clamping to [0,1] can be done for free in the fragment
1699 * shader with a saturate if floats are being used as boolean values.
1700 */
1701 add->saturate = true;
1702 } else {
1703 /* Negating the result of the addition gives values on the range
1704 * [-2, 0]. Zero stays zero, and negative values become 1.0. This
1705 * is achieved using SLT.
1706 */
1707 st_src_reg slt_src = result_src;
1708 slt_src.negate = ~slt_src.negate;
1709 emit_asm(ir, TGSI_OPCODE_SLT, result_dst, slt_src, st_src_reg_for_float(0.0));
1710 }
1711 }
1712 break;
1713 }
1714
1715 case ir_binop_logic_and:
1716 /* If native integers are disabled, the bool args are stored as float 0.0
1717 * or 1.0, so "mul" gives us "and". If they're enabled, just use the
1718 * actual AND opcode.
1719 */
1720 if (native_integers)
1721 emit_asm(ir, TGSI_OPCODE_AND, result_dst, op[0], op[1]);
1722 else
1723 emit_asm(ir, TGSI_OPCODE_MUL, result_dst, op[0], op[1]);
1724 break;
1725
1726 case ir_binop_dot:
1727 assert(ir->operands[0]->type->is_vector());
1728 assert(ir->operands[0]->type == ir->operands[1]->type);
1729 emit_dp(ir, result_dst, op[0], op[1],
1730 ir->operands[0]->type->vector_elements);
1731 break;
1732
1733 case ir_unop_sqrt:
1734 if (have_sqrt) {
1735 emit_scalar(ir, TGSI_OPCODE_SQRT, result_dst, op[0]);
1736 } else {
1737 /* This is the only instruction sequence that makes the game "Risen"
1738 * render correctly. ABS is not required for the game, but since GLSL
1739 * declares negative values as "undefined", allowing us to do whatever
1740 * we want, I choose to use ABS to match DX9 and pre-GLSL RSQ
1741 * behavior.
1742 */
1743 emit_scalar(ir, TGSI_OPCODE_RSQ, result_dst, op[0].get_abs());
1744 emit_scalar(ir, TGSI_OPCODE_RCP, result_dst, result_src);
1745 }
1746 break;
1747 case ir_unop_rsq:
1748 emit_scalar(ir, TGSI_OPCODE_RSQ, result_dst, op[0]);
1749 break;
1750 case ir_unop_i2f:
1751 if (native_integers) {
1752 emit_asm(ir, TGSI_OPCODE_I2F, result_dst, op[0]);
1753 break;
1754 }
1755 /* fallthrough to next case otherwise */
1756 case ir_unop_b2f:
1757 if (native_integers) {
1758 emit_asm(ir, TGSI_OPCODE_AND, result_dst, op[0], st_src_reg_for_float(1.0));
1759 break;
1760 }
1761 /* fallthrough to next case otherwise */
1762 case ir_unop_i2u:
1763 case ir_unop_u2i:
1764 case ir_unop_i642u64:
1765 case ir_unop_u642i64:
1766 /* Converting between signed and unsigned integers is a no-op. */
1767 result_src = op[0];
1768 result_src.type = result_dst.type;
1769 break;
1770 case ir_unop_b2i:
1771 if (native_integers) {
1772 /* Booleans are stored as integers using ~0 for true and 0 for false.
1773 * GLSL requires that int(bool) return 1 for true and 0 for false.
1774 * This conversion is done with AND, but it could be done with NEG.
1775 */
1776 emit_asm(ir, TGSI_OPCODE_AND, result_dst, op[0], st_src_reg_for_int(1));
1777 } else {
1778 /* Booleans and integers are both stored as floats when native
1779 * integers are disabled.
1780 */
1781 result_src = op[0];
1782 }
1783 break;
1784 case ir_unop_f2i:
1785 if (native_integers)
1786 emit_asm(ir, TGSI_OPCODE_F2I, result_dst, op[0]);
1787 else
1788 emit_asm(ir, TGSI_OPCODE_TRUNC, result_dst, op[0]);
1789 break;
1790 case ir_unop_f2u:
1791 if (native_integers)
1792 emit_asm(ir, TGSI_OPCODE_F2U, result_dst, op[0]);
1793 else
1794 emit_asm(ir, TGSI_OPCODE_TRUNC, result_dst, op[0]);
1795 break;
1796 case ir_unop_bitcast_f2i:
1797 case ir_unop_bitcast_f2u:
1798 /* Make sure we don't propagate the negate modifier to integer opcodes. */
1799 if (op[0].negate || op[0].abs)
1800 emit_asm(ir, TGSI_OPCODE_MOV, result_dst, op[0]);
1801 else
1802 result_src = op[0];
1803 result_src.type = ir->operation == ir_unop_bitcast_f2i ? GLSL_TYPE_INT :
1804 GLSL_TYPE_UINT;
1805 break;
1806 case ir_unop_bitcast_i2f:
1807 case ir_unop_bitcast_u2f:
1808 result_src = op[0];
1809 result_src.type = GLSL_TYPE_FLOAT;
1810 break;
1811 case ir_unop_f2b:
1812 emit_asm(ir, TGSI_OPCODE_SNE, result_dst, op[0], st_src_reg_for_float(0.0));
1813 break;
1814 case ir_unop_d2b:
1815 emit_asm(ir, TGSI_OPCODE_SNE, result_dst, op[0], st_src_reg_for_double(0.0));
1816 break;
1817 case ir_unop_i2b:
1818 if (native_integers)
1819 emit_asm(ir, TGSI_OPCODE_USNE, result_dst, op[0], st_src_reg_for_int(0));
1820 else
1821 emit_asm(ir, TGSI_OPCODE_SNE, result_dst, op[0], st_src_reg_for_float(0.0));
1822 break;
1823 case ir_unop_bitcast_u642d:
1824 case ir_unop_bitcast_i642d:
1825 result_src = op[0];
1826 result_src.type = GLSL_TYPE_DOUBLE;
1827 break;
1828 case ir_unop_bitcast_d2i64:
1829 result_src = op[0];
1830 result_src.type = GLSL_TYPE_INT64;
1831 break;
1832 case ir_unop_bitcast_d2u64:
1833 result_src = op[0];
1834 result_src.type = GLSL_TYPE_UINT64;
1835 break;
1836 case ir_unop_trunc:
1837 emit_asm(ir, TGSI_OPCODE_TRUNC, result_dst, op[0]);
1838 break;
1839 case ir_unop_ceil:
1840 emit_asm(ir, TGSI_OPCODE_CEIL, result_dst, op[0]);
1841 break;
1842 case ir_unop_floor:
1843 emit_asm(ir, TGSI_OPCODE_FLR, result_dst, op[0]);
1844 break;
1845 case ir_unop_round_even:
1846 emit_asm(ir, TGSI_OPCODE_ROUND, result_dst, op[0]);
1847 break;
1848 case ir_unop_fract:
1849 emit_asm(ir, TGSI_OPCODE_FRC, result_dst, op[0]);
1850 break;
1851
1852 case ir_binop_min:
1853 emit_asm(ir, TGSI_OPCODE_MIN, result_dst, op[0], op[1]);
1854 break;
1855 case ir_binop_max:
1856 emit_asm(ir, TGSI_OPCODE_MAX, result_dst, op[0], op[1]);
1857 break;
1858 case ir_binop_pow:
1859 emit_scalar(ir, TGSI_OPCODE_POW, result_dst, op[0], op[1]);
1860 break;
1861
1862 case ir_unop_bit_not:
1863 if (native_integers) {
1864 emit_asm(ir, TGSI_OPCODE_NOT, result_dst, op[0]);
1865 break;
1866 }
1867 case ir_unop_u2f:
1868 if (native_integers) {
1869 emit_asm(ir, TGSI_OPCODE_U2F, result_dst, op[0]);
1870 break;
1871 }
1872 case ir_binop_lshift:
1873 case ir_binop_rshift:
1874 if (native_integers) {
1875 unsigned opcode = ir->operation == ir_binop_lshift ? TGSI_OPCODE_SHL
1876 : TGSI_OPCODE_ISHR;
1877 st_src_reg count;
1878
1879 if (glsl_base_type_is_64bit(op[0].type)) {
1880 /* GLSL shift operations have 32-bit shift counts, but TGSI uses
1881 * 64 bits.
1882 */
1883 count = get_temp(glsl_type::u64vec(ir->operands[1]->type->components()));
1884 emit_asm(ir, TGSI_OPCODE_U2I64, st_dst_reg(count), op[1]);
1885 } else {
1886 count = op[1];
1887 }
1888
1889 emit_asm(ir, opcode, result_dst, op[0], count);
1890 break;
1891 }
1892 case ir_binop_bit_and:
1893 if (native_integers) {
1894 emit_asm(ir, TGSI_OPCODE_AND, result_dst, op[0], op[1]);
1895 break;
1896 }
1897 case ir_binop_bit_xor:
1898 if (native_integers) {
1899 emit_asm(ir, TGSI_OPCODE_XOR, result_dst, op[0], op[1]);
1900 break;
1901 }
1902 case ir_binop_bit_or:
1903 if (native_integers) {
1904 emit_asm(ir, TGSI_OPCODE_OR, result_dst, op[0], op[1]);
1905 break;
1906 }
1907
1908 assert(!"GLSL 1.30 features unsupported");
1909 break;
1910
1911 case ir_binop_ubo_load: {
1912 if (ctx->Const.UseSTD430AsDefaultPacking) {
1913 ir_rvalue *block = ir->operands[0];
1914 ir_rvalue *offset = ir->operands[1];
1915 ir_constant *const_block = block->as_constant();
1916
1917 st_src_reg cbuf(PROGRAM_CONSTANT,
1918 (const_block ? const_block->value.u[0] + 1 : 1),
1919 ir->type->base_type);
1920
1921 cbuf.has_index2 = true;
1922
1923 if (!const_block) {
1924 block->accept(this);
1925 cbuf.reladdr = ralloc(mem_ctx, st_src_reg);
1926 *cbuf.reladdr = this->result;
1927 emit_arl(ir, sampler_reladdr, this->result);
1928 }
1929
1930 /* Calculate the surface offset */
1931 offset->accept(this);
1932 st_src_reg off = this->result;
1933
1934 glsl_to_tgsi_instruction *inst =
1935 emit_asm(ir, TGSI_OPCODE_LOAD, result_dst, off);
1936
1937 if (result_dst.type == GLSL_TYPE_BOOL)
1938 emit_asm(ir, TGSI_OPCODE_USNE, result_dst, st_src_reg(result_dst),
1939 st_src_reg_for_int(0));
1940
1941 add_buffer_to_load_and_stores(inst, &cbuf, &this->instructions,
1942 NULL);
1943 } else {
1944 ir_constant *const_uniform_block = ir->operands[0]->as_constant();
1945 ir_constant *const_offset_ir = ir->operands[1]->as_constant();
1946 unsigned const_offset = const_offset_ir ?
1947 const_offset_ir->value.u[0] : 0;
1948 unsigned const_block = const_uniform_block ?
1949 const_uniform_block->value.u[0] + 1 : 1;
1950 st_src_reg index_reg = get_temp(glsl_type::uint_type);
1951 st_src_reg cbuf;
1952
1953 cbuf.type = ir->type->base_type;
1954 cbuf.file = PROGRAM_CONSTANT;
1955 cbuf.index = 0;
1956 cbuf.reladdr = NULL;
1957 cbuf.negate = 0;
1958 cbuf.abs = 0;
1959 cbuf.index2D = const_block;
1960
1961 assert(ir->type->is_vector() || ir->type->is_scalar());
1962
1963 if (const_offset_ir) {
1964 /* Constant index into constant buffer */
1965 cbuf.reladdr = NULL;
1966 cbuf.index = const_offset / 16;
1967 } else {
1968 ir_expression *offset_expr = ir->operands[1]->as_expression();
1969 st_src_reg offset = op[1];
1970
1971 /* The OpenGL spec is written in such a way that accesses with
1972 * non-constant offset are almost always vec4-aligned. The only
1973 * exception to this are members of structs in arrays of structs:
1974 * each struct in an array of structs is at least vec4-aligned,
1975 * but single-element and [ui]vec2 members of the struct may be at
1976 * an offset that is not a multiple of 16 bytes.
1977 *
1978 * Here, we extract that offset, relying on previous passes to
1979 * always generate offset expressions of the form
1980 * (+ expr constant_offset).
1981 *
1982 * Note that the std430 layout, which allows more cases of
1983 * alignment less than vec4 in arrays, is not supported for
1984 * uniform blocks, so we do not have to deal with it here.
1985 */
1986 if (offset_expr && offset_expr->operation == ir_binop_add) {
1987 const_offset_ir = offset_expr->operands[1]->as_constant();
1988 if (const_offset_ir) {
1989 const_offset = const_offset_ir->value.u[0];
1990 cbuf.index = const_offset / 16;
1991 offset_expr->operands[0]->accept(this);
1992 offset = this->result;
1993 }
1994 }
1995
1996 /* Relative/variable index into constant buffer */
1997 emit_asm(ir, TGSI_OPCODE_USHR, st_dst_reg(index_reg), offset,
1998 st_src_reg_for_int(4));
1999 cbuf.reladdr = ralloc(mem_ctx, st_src_reg);
2000 memcpy(cbuf.reladdr, &index_reg, sizeof(index_reg));
2001 }
2002
2003 if (const_uniform_block) {
2004 /* Constant constant buffer */
2005 cbuf.reladdr2 = NULL;
2006 } else {
2007 /* Relative/variable constant buffer */
2008 cbuf.reladdr2 = ralloc(mem_ctx, st_src_reg);
2009 memcpy(cbuf.reladdr2, &op[0], sizeof(st_src_reg));
2010 }
2011 cbuf.has_index2 = true;
2012
2013 cbuf.swizzle = swizzle_for_size(ir->type->vector_elements);
2014 if (glsl_base_type_is_64bit(cbuf.type))
2015 cbuf.swizzle += MAKE_SWIZZLE4(const_offset % 16 / 8,
2016 const_offset % 16 / 8,
2017 const_offset % 16 / 8,
2018 const_offset % 16 / 8);
2019 else
2020 cbuf.swizzle += MAKE_SWIZZLE4(const_offset % 16 / 4,
2021 const_offset % 16 / 4,
2022 const_offset % 16 / 4,
2023 const_offset % 16 / 4);
2024
2025 if (ir->type->is_boolean()) {
2026 emit_asm(ir, TGSI_OPCODE_USNE, result_dst, cbuf,
2027 st_src_reg_for_int(0));
2028 } else {
2029 emit_asm(ir, TGSI_OPCODE_MOV, result_dst, cbuf);
2030 }
2031 }
2032 break;
2033 }
2034 case ir_triop_lrp:
2035 /* note: we have to reorder the three args here */
2036 emit_asm(ir, TGSI_OPCODE_LRP, result_dst, op[2], op[1], op[0]);
2037 break;
2038 case ir_triop_csel:
2039 if (this->ctx->Const.NativeIntegers)
2040 emit_asm(ir, TGSI_OPCODE_UCMP, result_dst, op[0], op[1], op[2]);
2041 else {
2042 op[0].negate = ~op[0].negate;
2043 emit_asm(ir, TGSI_OPCODE_CMP, result_dst, op[0], op[1], op[2]);
2044 }
2045 break;
2046 case ir_triop_bitfield_extract:
2047 emit_asm(ir, TGSI_OPCODE_IBFE, result_dst, op[0], op[1], op[2]);
2048 break;
2049 case ir_quadop_bitfield_insert:
2050 emit_asm(ir, TGSI_OPCODE_BFI, result_dst, op[0], op[1], op[2], op[3]);
2051 break;
2052 case ir_unop_bitfield_reverse:
2053 emit_asm(ir, TGSI_OPCODE_BREV, result_dst, op[0]);
2054 break;
2055 case ir_unop_bit_count:
2056 emit_asm(ir, TGSI_OPCODE_POPC, result_dst, op[0]);
2057 break;
2058 case ir_unop_find_msb:
2059 emit_asm(ir, TGSI_OPCODE_IMSB, result_dst, op[0]);
2060 break;
2061 case ir_unop_find_lsb:
2062 emit_asm(ir, TGSI_OPCODE_LSB, result_dst, op[0]);
2063 break;
2064 case ir_binop_imul_high:
2065 emit_asm(ir, TGSI_OPCODE_IMUL_HI, result_dst, op[0], op[1]);
2066 break;
2067 case ir_triop_fma:
2068 /* In theory, MAD is incorrect here. */
2069 if (have_fma)
2070 emit_asm(ir, TGSI_OPCODE_FMA, result_dst, op[0], op[1], op[2]);
2071 else
2072 emit_asm(ir, TGSI_OPCODE_MAD, result_dst, op[0], op[1], op[2]);
2073 break;
2074 case ir_unop_interpolate_at_centroid:
2075 emit_asm(ir, TGSI_OPCODE_INTERP_CENTROID, result_dst, op[0]);
2076 break;
2077 case ir_binop_interpolate_at_offset: {
2078 /* The y coordinate needs to be flipped for the default fb */
2079 static const gl_state_index transform_y_state[STATE_LENGTH]
2080 = { STATE_INTERNAL, STATE_FB_WPOS_Y_TRANSFORM };
2081
2082 unsigned transform_y_index =
2083 _mesa_add_state_reference(this->prog->Parameters,
2084 transform_y_state);
2085
2086 st_src_reg transform_y = st_src_reg(PROGRAM_STATE_VAR,
2087 transform_y_index,
2088 glsl_type::vec4_type);
2089 transform_y.swizzle = SWIZZLE_XXXX;
2090
2091 st_src_reg temp = get_temp(glsl_type::vec2_type);
2092 st_dst_reg temp_dst = st_dst_reg(temp);
2093
2094 emit_asm(ir, TGSI_OPCODE_MOV, temp_dst, op[1]);
2095 temp_dst.writemask = WRITEMASK_Y;
2096 emit_asm(ir, TGSI_OPCODE_MUL, temp_dst, transform_y, op[1]);
2097 emit_asm(ir, TGSI_OPCODE_INTERP_OFFSET, result_dst, op[0], temp);
2098 break;
2099 }
2100 case ir_binop_interpolate_at_sample:
2101 emit_asm(ir, TGSI_OPCODE_INTERP_SAMPLE, result_dst, op[0], op[1]);
2102 break;
2103
2104 case ir_unop_d2f:
2105 emit_asm(ir, TGSI_OPCODE_D2F, result_dst, op[0]);
2106 break;
2107 case ir_unop_f2d:
2108 emit_asm(ir, TGSI_OPCODE_F2D, result_dst, op[0]);
2109 break;
2110 case ir_unop_d2i:
2111 emit_asm(ir, TGSI_OPCODE_D2I, result_dst, op[0]);
2112 break;
2113 case ir_unop_i2d:
2114 emit_asm(ir, TGSI_OPCODE_I2D, result_dst, op[0]);
2115 break;
2116 case ir_unop_d2u:
2117 emit_asm(ir, TGSI_OPCODE_D2U, result_dst, op[0]);
2118 break;
2119 case ir_unop_u2d:
2120 emit_asm(ir, TGSI_OPCODE_U2D, result_dst, op[0]);
2121 break;
2122 case ir_unop_unpack_double_2x32:
2123 case ir_unop_pack_double_2x32:
2124 case ir_unop_unpack_int_2x32:
2125 case ir_unop_pack_int_2x32:
2126 case ir_unop_unpack_uint_2x32:
2127 case ir_unop_pack_uint_2x32:
2128 case ir_unop_unpack_sampler_2x32:
2129 case ir_unop_pack_sampler_2x32:
2130 case ir_unop_unpack_image_2x32:
2131 case ir_unop_pack_image_2x32:
2132 emit_asm(ir, TGSI_OPCODE_MOV, result_dst, op[0]);
2133 break;
2134
2135 case ir_binop_ldexp:
2136 if (ir->operands[0]->type->is_double()) {
2137 emit_asm(ir, TGSI_OPCODE_DLDEXP, result_dst, op[0], op[1]);
2138 } else if (ir->operands[0]->type->is_float()) {
2139 emit_asm(ir, TGSI_OPCODE_LDEXP, result_dst, op[0], op[1]);
2140 } else {
2141 assert(!"Invalid ldexp for non-double opcode in glsl_to_tgsi_visitor::visit()");
2142 }
2143 break;
2144
2145 case ir_unop_pack_half_2x16:
2146 emit_asm(ir, TGSI_OPCODE_PK2H, result_dst, op[0]);
2147 break;
2148 case ir_unop_unpack_half_2x16:
2149 emit_asm(ir, TGSI_OPCODE_UP2H, result_dst, op[0]);
2150 break;
2151
2152 case ir_unop_get_buffer_size: {
2153 ir_constant *const_offset = ir->operands[0]->as_constant();
2154 int buf_base = ctx->st->has_hw_atomics ? 0 : ctx->Const.Program[shader->Stage].MaxAtomicBuffers;
2155 st_src_reg buffer(
2156 PROGRAM_BUFFER,
2157 buf_base + (const_offset ? const_offset->value.u[0] : 0),
2158 GLSL_TYPE_UINT);
2159 if (!const_offset) {
2160 buffer.reladdr = ralloc(mem_ctx, st_src_reg);
2161 *buffer.reladdr = op[0];
2162 emit_arl(ir, sampler_reladdr, op[0]);
2163 }
2164 emit_asm(ir, TGSI_OPCODE_RESQ, result_dst)->resource = buffer;
2165 break;
2166 }
2167
2168 case ir_unop_u2i64:
2169 case ir_unop_u2u64:
2170 case ir_unop_b2i64: {
2171 st_src_reg temp = get_temp(glsl_type::uvec4_type);
2172 st_dst_reg temp_dst = st_dst_reg(temp);
2173 unsigned orig_swz = op[0].swizzle;
2174 /*
2175 * To convert unsigned to 64-bit:
2176 * zero Y channel, copy X channel.
2177 */
2178 temp_dst.writemask = WRITEMASK_Y;
2179 if (vector_elements > 1)
2180 temp_dst.writemask |= WRITEMASK_W;
2181 emit_asm(ir, TGSI_OPCODE_MOV, temp_dst, st_src_reg_for_int(0));
2182 temp_dst.writemask = WRITEMASK_X;
2183 if (vector_elements > 1)
2184 temp_dst.writemask |= WRITEMASK_Z;
2185 op[0].swizzle = MAKE_SWIZZLE4(GET_SWZ(orig_swz, 0), GET_SWZ(orig_swz, 0),
2186 GET_SWZ(orig_swz, 1), GET_SWZ(orig_swz, 1));
2187 if (ir->operation == ir_unop_u2i64 || ir->operation == ir_unop_u2u64)
2188 emit_asm(ir, TGSI_OPCODE_MOV, temp_dst, op[0]);
2189 else
2190 emit_asm(ir, TGSI_OPCODE_AND, temp_dst, op[0], st_src_reg_for_int(1));
2191 result_src = temp;
2192 result_src.type = GLSL_TYPE_UINT64;
2193 if (vector_elements > 2) {
2194 /* Subtle: We rely on the fact that get_temp here returns the next
2195 * TGSI temporary register directly after the temp register used for
2196 * the first two components, so that the result gets picked up
2197 * automatically.
2198 */
2199 st_src_reg temp = get_temp(glsl_type::uvec4_type);
2200 st_dst_reg temp_dst = st_dst_reg(temp);
2201 temp_dst.writemask = WRITEMASK_Y;
2202 if (vector_elements > 3)
2203 temp_dst.writemask |= WRITEMASK_W;
2204 emit_asm(ir, TGSI_OPCODE_MOV, temp_dst, st_src_reg_for_int(0));
2205
2206 temp_dst.writemask = WRITEMASK_X;
2207 if (vector_elements > 3)
2208 temp_dst.writemask |= WRITEMASK_Z;
2209 op[0].swizzle = MAKE_SWIZZLE4(GET_SWZ(orig_swz, 2), GET_SWZ(orig_swz, 2),
2210 GET_SWZ(orig_swz, 3), GET_SWZ(orig_swz, 3));
2211 if (ir->operation == ir_unop_u2i64 || ir->operation == ir_unop_u2u64)
2212 emit_asm(ir, TGSI_OPCODE_MOV, temp_dst, op[0]);
2213 else
2214 emit_asm(ir, TGSI_OPCODE_AND, temp_dst, op[0], st_src_reg_for_int(1));
2215 }
2216 break;
2217 }
2218 case ir_unop_i642i:
2219 case ir_unop_u642i:
2220 case ir_unop_u642u:
2221 case ir_unop_i642u: {
2222 st_src_reg temp = get_temp(glsl_type::uvec4_type);
2223 st_dst_reg temp_dst = st_dst_reg(temp);
2224 unsigned orig_swz = op[0].swizzle;
2225 unsigned orig_idx = op[0].index;
2226 int el;
2227 temp_dst.writemask = WRITEMASK_X;
2228
2229 for (el = 0; el < vector_elements; el++) {
2230 unsigned swz = GET_SWZ(orig_swz, el);
2231 if (swz & 1)
2232 op[0].swizzle = MAKE_SWIZZLE4(SWIZZLE_Z, SWIZZLE_Z, SWIZZLE_Z, SWIZZLE_Z);
2233 else
2234 op[0].swizzle = MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_X, SWIZZLE_X, SWIZZLE_X);
2235 if (swz > 2)
2236 op[0].index = orig_idx + 1;
2237 op[0].type = GLSL_TYPE_UINT;
2238 temp_dst.writemask = WRITEMASK_X << el;
2239 emit_asm(ir, TGSI_OPCODE_MOV, temp_dst, op[0]);
2240 }
2241 result_src = temp;
2242 if (ir->operation == ir_unop_u642u || ir->operation == ir_unop_i642u)
2243 result_src.type = GLSL_TYPE_UINT;
2244 else
2245 result_src.type = GLSL_TYPE_INT;
2246 break;
2247 }
2248 case ir_unop_i642b:
2249 emit_asm(ir, TGSI_OPCODE_U64SNE, result_dst, op[0], st_src_reg_for_int64(0));
2250 break;
2251 case ir_unop_i642f:
2252 emit_asm(ir, TGSI_OPCODE_I642F, result_dst, op[0]);
2253 break;
2254 case ir_unop_u642f:
2255 emit_asm(ir, TGSI_OPCODE_U642F, result_dst, op[0]);
2256 break;
2257 case ir_unop_i642d:
2258 emit_asm(ir, TGSI_OPCODE_I642D, result_dst, op[0]);
2259 break;
2260 case ir_unop_u642d:
2261 emit_asm(ir, TGSI_OPCODE_U642D, result_dst, op[0]);
2262 break;
2263 case ir_unop_i2i64:
2264 emit_asm(ir, TGSI_OPCODE_I2I64, result_dst, op[0]);
2265 break;
2266 case ir_unop_f2i64:
2267 emit_asm(ir, TGSI_OPCODE_F2I64, result_dst, op[0]);
2268 break;
2269 case ir_unop_d2i64:
2270 emit_asm(ir, TGSI_OPCODE_D2I64, result_dst, op[0]);
2271 break;
2272 case ir_unop_i2u64:
2273 emit_asm(ir, TGSI_OPCODE_I2I64, result_dst, op[0]);
2274 break;
2275 case ir_unop_f2u64:
2276 emit_asm(ir, TGSI_OPCODE_F2U64, result_dst, op[0]);
2277 break;
2278 case ir_unop_d2u64:
2279 emit_asm(ir, TGSI_OPCODE_D2U64, result_dst, op[0]);
2280 break;
2281 /* these might be needed */
2282 case ir_unop_pack_snorm_2x16:
2283 case ir_unop_pack_unorm_2x16:
2284 case ir_unop_pack_snorm_4x8:
2285 case ir_unop_pack_unorm_4x8:
2286
2287 case ir_unop_unpack_snorm_2x16:
2288 case ir_unop_unpack_unorm_2x16:
2289 case ir_unop_unpack_snorm_4x8:
2290 case ir_unop_unpack_unorm_4x8:
2291
2292 case ir_quadop_vector:
2293 case ir_binop_vector_extract:
2294 case ir_triop_vector_insert:
2295 case ir_binop_carry:
2296 case ir_binop_borrow:
2297 case ir_unop_ssbo_unsized_array_length:
2298 /* This operation is not supported, or should have already been handled.
2299 */
2300 assert(!"Invalid ir opcode in glsl_to_tgsi_visitor::visit()");
2301 break;
2302 }
2303
2304 this->result = result_src;
2305 }
2306
2307
2308 void
visit(ir_swizzle * ir)2309 glsl_to_tgsi_visitor::visit(ir_swizzle *ir)
2310 {
2311 st_src_reg src;
2312 int i;
2313 int swizzle[4];
2314
2315 /* Note that this is only swizzles in expressions, not those on the left
2316 * hand side of an assignment, which do write masking. See ir_assignment
2317 * for that.
2318 */
2319
2320 ir->val->accept(this);
2321 src = this->result;
2322 assert(src.file != PROGRAM_UNDEFINED);
2323 assert(ir->type->vector_elements > 0);
2324
2325 for (i = 0; i < 4; i++) {
2326 if (i < ir->type->vector_elements) {
2327 switch (i) {
2328 case 0:
2329 swizzle[i] = GET_SWZ(src.swizzle, ir->mask.x);
2330 break;
2331 case 1:
2332 swizzle[i] = GET_SWZ(src.swizzle, ir->mask.y);
2333 break;
2334 case 2:
2335 swizzle[i] = GET_SWZ(src.swizzle, ir->mask.z);
2336 break;
2337 case 3:
2338 swizzle[i] = GET_SWZ(src.swizzle, ir->mask.w);
2339 break;
2340 }
2341 } else {
2342 /* If the type is smaller than a vec4, replicate the last
2343 * channel out.
2344 */
2345 swizzle[i] = swizzle[ir->type->vector_elements - 1];
2346 }
2347 }
2348
2349 src.swizzle = MAKE_SWIZZLE4(swizzle[0], swizzle[1], swizzle[2], swizzle[3]);
2350
2351 this->result = src;
2352 }
2353
2354 /* Test if the variable is an array. Note that geometry and
2355 * tessellation shader inputs are outputs are always arrays (except
2356 * for patch inputs), so only the array element type is considered.
2357 */
2358 static bool
is_inout_array(unsigned stage,ir_variable * var,bool * remove_array)2359 is_inout_array(unsigned stage, ir_variable *var, bool *remove_array)
2360 {
2361 const glsl_type *type = var->type;
2362
2363 *remove_array = false;
2364
2365 if ((stage == MESA_SHADER_VERTEX && var->data.mode == ir_var_shader_in) ||
2366 (stage == MESA_SHADER_FRAGMENT && var->data.mode == ir_var_shader_out))
2367 return false;
2368
2369 if (((stage == MESA_SHADER_GEOMETRY && var->data.mode == ir_var_shader_in) ||
2370 (stage == MESA_SHADER_TESS_EVAL && var->data.mode == ir_var_shader_in) ||
2371 stage == MESA_SHADER_TESS_CTRL) &&
2372 !var->data.patch) {
2373 if (!var->type->is_array())
2374 return false; /* a system value probably */
2375
2376 type = var->type->fields.array;
2377 *remove_array = true;
2378 }
2379
2380 return type->is_array() || type->is_matrix();
2381 }
2382
2383 static unsigned
st_translate_interp_loc(ir_variable * var)2384 st_translate_interp_loc(ir_variable *var)
2385 {
2386 if (var->data.centroid)
2387 return TGSI_INTERPOLATE_LOC_CENTROID;
2388 else if (var->data.sample)
2389 return TGSI_INTERPOLATE_LOC_SAMPLE;
2390 else
2391 return TGSI_INTERPOLATE_LOC_CENTER;
2392 }
2393
2394 void
visit(ir_dereference_variable * ir)2395 glsl_to_tgsi_visitor::visit(ir_dereference_variable *ir)
2396 {
2397 variable_storage *entry = find_variable_storage(ir->var);
2398 ir_variable *var = ir->var;
2399 bool remove_array;
2400
2401 if (!entry) {
2402 switch (var->data.mode) {
2403 case ir_var_uniform:
2404 entry = new(mem_ctx) variable_storage(var, PROGRAM_UNIFORM,
2405 var->data.param_index);
2406 _mesa_hash_table_insert(this->variables, var, entry);
2407 break;
2408 case ir_var_shader_in: {
2409 /* The linker assigns locations for varyings and attributes,
2410 * including deprecated builtins (like gl_Color), user-assign
2411 * generic attributes (glBindVertexLocation), and
2412 * user-defined varyings.
2413 */
2414 assert(var->data.location != -1);
2415
2416 const glsl_type *type_without_array = var->type->without_array();
2417 struct inout_decl *decl = &inputs[num_inputs];
2418 unsigned component = var->data.location_frac;
2419 unsigned num_components;
2420 num_inputs++;
2421
2422 if (type_without_array->is_64bit())
2423 component = component / 2;
2424 if (type_without_array->vector_elements)
2425 num_components = type_without_array->vector_elements;
2426 else
2427 num_components = 4;
2428
2429 decl->mesa_index = var->data.location;
2430 decl->interp = (glsl_interp_mode) var->data.interpolation;
2431 decl->interp_loc = st_translate_interp_loc(var);
2432 decl->base_type = type_without_array->base_type;
2433 decl->usage_mask = u_bit_consecutive(component, num_components);
2434
2435 if (is_inout_array(shader->Stage, var, &remove_array)) {
2436 decl->array_id = num_input_arrays + 1;
2437 num_input_arrays++;
2438 } else {
2439 decl->array_id = 0;
2440 }
2441
2442 if (remove_array)
2443 decl->size = type_size(var->type->fields.array);
2444 else
2445 decl->size = type_size(var->type);
2446
2447 entry = new(mem_ctx) variable_storage(var,
2448 PROGRAM_INPUT,
2449 decl->mesa_index,
2450 decl->array_id);
2451 entry->component = component;
2452
2453 _mesa_hash_table_insert(this->variables, var, entry);
2454
2455 break;
2456 }
2457 case ir_var_shader_out: {
2458 assert(var->data.location != -1);
2459
2460 const glsl_type *type_without_array = var->type->without_array();
2461 struct inout_decl *decl = &outputs[num_outputs];
2462 unsigned component = var->data.location_frac;
2463 unsigned num_components;
2464 num_outputs++;
2465
2466 if (type_without_array->is_64bit())
2467 component = component / 2;
2468 if (type_without_array->vector_elements)
2469 num_components = type_without_array->vector_elements;
2470 else
2471 num_components = 4;
2472
2473 decl->mesa_index = var->data.location + FRAG_RESULT_MAX * var->data.index;
2474 decl->base_type = type_without_array->base_type;
2475 decl->usage_mask = u_bit_consecutive(component, num_components);
2476 if (var->data.stream & (1u << 31)) {
2477 decl->gs_out_streams = var->data.stream & ~(1u << 31);
2478 } else {
2479 assert(var->data.stream < 4);
2480 decl->gs_out_streams = 0;
2481 for (unsigned i = 0; i < num_components; ++i)
2482 decl->gs_out_streams |= var->data.stream << (2 * (component + i));
2483 }
2484
2485 if (is_inout_array(shader->Stage, var, &remove_array)) {
2486 decl->array_id = num_output_arrays + 1;
2487 num_output_arrays++;
2488 } else {
2489 decl->array_id = 0;
2490 }
2491
2492 if (remove_array)
2493 decl->size = type_size(var->type->fields.array);
2494 else
2495 decl->size = type_size(var->type);
2496
2497 if (var->data.fb_fetch_output) {
2498 st_dst_reg dst = st_dst_reg(get_temp(var->type));
2499 st_src_reg src = st_src_reg(PROGRAM_OUTPUT, decl->mesa_index,
2500 var->type, component, decl->array_id);
2501 emit_asm(NULL, TGSI_OPCODE_FBFETCH, dst, src);
2502 entry = new(mem_ctx) variable_storage(var, dst.file, dst.index,
2503 dst.array_id);
2504 } else {
2505 entry = new(mem_ctx) variable_storage(var,
2506 PROGRAM_OUTPUT,
2507 decl->mesa_index,
2508 decl->array_id);
2509 }
2510 entry->component = component;
2511
2512 _mesa_hash_table_insert(this->variables, var, entry);
2513
2514 break;
2515 }
2516 case ir_var_system_value:
2517 entry = new(mem_ctx) variable_storage(var,
2518 PROGRAM_SYSTEM_VALUE,
2519 var->data.location);
2520 break;
2521 case ir_var_auto:
2522 case ir_var_temporary:
2523 st_src_reg src = get_temp(var->type);
2524
2525 entry = new(mem_ctx) variable_storage(var, src.file, src.index,
2526 src.array_id);
2527 _mesa_hash_table_insert(this->variables, var, entry);
2528
2529 break;
2530 }
2531
2532 if (!entry) {
2533 printf("Failed to make storage for %s\n", var->name);
2534 exit(1);
2535 }
2536 }
2537
2538 this->result = st_src_reg(entry->file, entry->index, var->type,
2539 entry->component, entry->array_id);
2540 if (this->shader->Stage == MESA_SHADER_VERTEX &&
2541 var->data.mode == ir_var_shader_in &&
2542 var->type->without_array()->is_double())
2543 this->result.is_double_vertex_input = true;
2544 if (!native_integers)
2545 this->result.type = GLSL_TYPE_FLOAT;
2546 }
2547
2548 static void
shrink_array_declarations(struct inout_decl * decls,unsigned count,GLbitfield64 * usage_mask,GLbitfield64 double_usage_mask,GLbitfield * patch_usage_mask)2549 shrink_array_declarations(struct inout_decl *decls, unsigned count,
2550 GLbitfield64* usage_mask,
2551 GLbitfield64 double_usage_mask,
2552 GLbitfield* patch_usage_mask)
2553 {
2554 unsigned i;
2555 int j;
2556
2557 /* Fix array declarations by removing unused array elements at both ends
2558 * of the arrays. For example, mat4[3] where only mat[1] is used.
2559 */
2560 for (i = 0; i < count; i++) {
2561 struct inout_decl *decl = &decls[i];
2562 if (!decl->array_id)
2563 continue;
2564
2565 /* Shrink the beginning. */
2566 for (j = 0; j < (int)decl->size; j++) {
2567 if (decl->mesa_index >= VARYING_SLOT_PATCH0) {
2568 if (*patch_usage_mask &
2569 BITFIELD64_BIT(decl->mesa_index - VARYING_SLOT_PATCH0 + j))
2570 break;
2571 }
2572 else {
2573 if (*usage_mask & BITFIELD64_BIT(decl->mesa_index+j))
2574 break;
2575 if (double_usage_mask & BITFIELD64_BIT(decl->mesa_index+j-1))
2576 break;
2577 }
2578
2579 decl->mesa_index++;
2580 decl->size--;
2581 j--;
2582 }
2583
2584 /* Shrink the end. */
2585 for (j = decl->size-1; j >= 0; j--) {
2586 if (decl->mesa_index >= VARYING_SLOT_PATCH0) {
2587 if (*patch_usage_mask &
2588 BITFIELD64_BIT(decl->mesa_index - VARYING_SLOT_PATCH0 + j))
2589 break;
2590 }
2591 else {
2592 if (*usage_mask & BITFIELD64_BIT(decl->mesa_index+j))
2593 break;
2594 if (double_usage_mask & BITFIELD64_BIT(decl->mesa_index+j-1))
2595 break;
2596 }
2597
2598 decl->size--;
2599 }
2600
2601 /* When not all entries of an array are accessed, we mark them as used
2602 * here anyway, to ensure that the input/output mapping logic doesn't get
2603 * confused.
2604 *
2605 * TODO This happens when an array isn't used via indirect access, which
2606 * some game ports do (at least eON-based). There is an optimization
2607 * opportunity here by replacing the array declaration with non-array
2608 * declarations of those slots that are actually used.
2609 */
2610 for (j = 1; j < (int)decl->size; ++j) {
2611 if (decl->mesa_index >= VARYING_SLOT_PATCH0)
2612 *patch_usage_mask |= BITFIELD64_BIT(decl->mesa_index - VARYING_SLOT_PATCH0 + j);
2613 else
2614 *usage_mask |= BITFIELD64_BIT(decl->mesa_index + j);
2615 }
2616 }
2617 }
2618
2619 void
visit(ir_dereference_array * ir)2620 glsl_to_tgsi_visitor::visit(ir_dereference_array *ir)
2621 {
2622 ir_constant *index;
2623 st_src_reg src;
2624 bool is_2D = false;
2625 ir_variable *var = ir->variable_referenced();
2626
2627 /* We only need the logic provided by st_glsl_storage_type_size()
2628 * for arrays of structs. Indirect sampler and image indexing is handled
2629 * elsewhere.
2630 */
2631 int element_size = ir->type->without_array()->is_record() ?
2632 st_glsl_storage_type_size(ir->type, var->data.bindless) :
2633 type_size(ir->type);
2634
2635 index = ir->array_index->constant_expression_value(ralloc_parent(ir));
2636
2637 ir->array->accept(this);
2638 src = this->result;
2639
2640 if (!src.has_index2) {
2641 switch (this->prog->Target) {
2642 case GL_TESS_CONTROL_PROGRAM_NV:
2643 is_2D = (src.file == PROGRAM_INPUT || src.file == PROGRAM_OUTPUT) &&
2644 !ir->variable_referenced()->data.patch;
2645 break;
2646 case GL_TESS_EVALUATION_PROGRAM_NV:
2647 is_2D = src.file == PROGRAM_INPUT &&
2648 !ir->variable_referenced()->data.patch;
2649 break;
2650 case GL_GEOMETRY_PROGRAM_NV:
2651 is_2D = src.file == PROGRAM_INPUT;
2652 break;
2653 }
2654 }
2655
2656 if (is_2D)
2657 element_size = 1;
2658
2659 if (index) {
2660
2661 if (this->prog->Target == GL_VERTEX_PROGRAM_ARB &&
2662 src.file == PROGRAM_INPUT)
2663 element_size = attrib_type_size(ir->type, true);
2664 if (is_2D) {
2665 src.index2D = index->value.i[0];
2666 src.has_index2 = true;
2667 } else
2668 src.index += index->value.i[0] * element_size;
2669 } else {
2670 /* Variable index array dereference. It eats the "vec4" of the
2671 * base of the array and an index that offsets the TGSI register
2672 * index.
2673 */
2674 ir->array_index->accept(this);
2675
2676 st_src_reg index_reg;
2677
2678 if (element_size == 1) {
2679 index_reg = this->result;
2680 } else {
2681 index_reg = get_temp(native_integers ?
2682 glsl_type::int_type : glsl_type::float_type);
2683
2684 emit_asm(ir, TGSI_OPCODE_MUL, st_dst_reg(index_reg),
2685 this->result, st_src_reg_for_type(index_reg.type, element_size));
2686 }
2687
2688 /* If there was already a relative address register involved, add the
2689 * new and the old together to get the new offset.
2690 */
2691 if (!is_2D && src.reladdr != NULL) {
2692 st_src_reg accum_reg = get_temp(native_integers ?
2693 glsl_type::int_type : glsl_type::float_type);
2694
2695 emit_asm(ir, TGSI_OPCODE_ADD, st_dst_reg(accum_reg),
2696 index_reg, *src.reladdr);
2697
2698 index_reg = accum_reg;
2699 }
2700
2701 if (is_2D) {
2702 src.reladdr2 = ralloc(mem_ctx, st_src_reg);
2703 memcpy(src.reladdr2, &index_reg, sizeof(index_reg));
2704 src.index2D = 0;
2705 src.has_index2 = true;
2706 } else {
2707 src.reladdr = ralloc(mem_ctx, st_src_reg);
2708 memcpy(src.reladdr, &index_reg, sizeof(index_reg));
2709 }
2710 }
2711
2712 /* Change the register type to the element type of the array. */
2713 src.type = ir->type->base_type;
2714
2715 this->result = src;
2716 }
2717
2718 void
visit(ir_dereference_record * ir)2719 glsl_to_tgsi_visitor::visit(ir_dereference_record *ir)
2720 {
2721 unsigned int i;
2722 const glsl_type *struct_type = ir->record->type;
2723 ir_variable *var = ir->record->variable_referenced();
2724 int offset = 0;
2725
2726 ir->record->accept(this);
2727
2728 assert(ir->field_idx >= 0);
2729 assert(var);
2730 for (i = 0; i < struct_type->length; i++) {
2731 if (i == (unsigned) ir->field_idx)
2732 break;
2733 const glsl_type *member_type = struct_type->fields.structure[i].type;
2734 offset += st_glsl_storage_type_size(member_type, var->data.bindless);
2735 }
2736
2737 /* If the type is smaller than a vec4, replicate the last channel out. */
2738 if (ir->type->is_scalar() || ir->type->is_vector())
2739 this->result.swizzle = swizzle_for_size(ir->type->vector_elements);
2740 else
2741 this->result.swizzle = SWIZZLE_NOOP;
2742
2743 this->result.index += offset;
2744 this->result.type = ir->type->base_type;
2745 }
2746
2747 /**
2748 * We want to be careful in assignment setup to hit the actual storage
2749 * instead of potentially using a temporary like we might with the
2750 * ir_dereference handler.
2751 */
2752 static st_dst_reg
get_assignment_lhs(ir_dereference * ir,glsl_to_tgsi_visitor * v,int * component)2753 get_assignment_lhs(ir_dereference *ir, glsl_to_tgsi_visitor *v, int *component)
2754 {
2755 /* The LHS must be a dereference. If the LHS is a variable indexed array
2756 * access of a vector, it must be separated into a series conditional moves
2757 * before reaching this point (see ir_vec_index_to_cond_assign).
2758 */
2759 assert(ir->as_dereference());
2760 ir_dereference_array *deref_array = ir->as_dereference_array();
2761 if (deref_array) {
2762 assert(!deref_array->array->type->is_vector());
2763 }
2764
2765 /* Use the rvalue deref handler for the most part. We write swizzles using
2766 * the writemask, but we do extract the base component for enhanced layouts
2767 * from the source swizzle.
2768 */
2769 ir->accept(v);
2770 *component = GET_SWZ(v->result.swizzle, 0);
2771 return st_dst_reg(v->result);
2772 }
2773
2774 /**
2775 * Process the condition of a conditional assignment
2776 *
2777 * Examines the condition of a conditional assignment to generate the optimal
2778 * first operand of a \c CMP instruction. If the condition is a relational
2779 * operator with 0 (e.g., \c ir_binop_less), the value being compared will be
2780 * used as the source for the \c CMP instruction. Otherwise the comparison
2781 * is processed to a boolean result, and the boolean result is used as the
2782 * operand to the CMP instruction.
2783 */
2784 bool
process_move_condition(ir_rvalue * ir)2785 glsl_to_tgsi_visitor::process_move_condition(ir_rvalue *ir)
2786 {
2787 ir_rvalue *src_ir = ir;
2788 bool negate = true;
2789 bool switch_order = false;
2790
2791 ir_expression *const expr = ir->as_expression();
2792
2793 if (native_integers) {
2794 if ((expr != NULL) && (expr->num_operands == 2)) {
2795 enum glsl_base_type type = expr->operands[0]->type->base_type;
2796 if (type == GLSL_TYPE_INT || type == GLSL_TYPE_UINT ||
2797 type == GLSL_TYPE_BOOL) {
2798 if (expr->operation == ir_binop_equal) {
2799 if (expr->operands[0]->is_zero()) {
2800 src_ir = expr->operands[1];
2801 switch_order = true;
2802 }
2803 else if (expr->operands[1]->is_zero()) {
2804 src_ir = expr->operands[0];
2805 switch_order = true;
2806 }
2807 }
2808 else if (expr->operation == ir_binop_nequal) {
2809 if (expr->operands[0]->is_zero()) {
2810 src_ir = expr->operands[1];
2811 }
2812 else if (expr->operands[1]->is_zero()) {
2813 src_ir = expr->operands[0];
2814 }
2815 }
2816 }
2817 }
2818
2819 src_ir->accept(this);
2820 return switch_order;
2821 }
2822
2823 if ((expr != NULL) && (expr->num_operands == 2)) {
2824 bool zero_on_left = false;
2825
2826 if (expr->operands[0]->is_zero()) {
2827 src_ir = expr->operands[1];
2828 zero_on_left = true;
2829 } else if (expr->operands[1]->is_zero()) {
2830 src_ir = expr->operands[0];
2831 zero_on_left = false;
2832 }
2833
2834 /* a is - 0 + - 0 +
2835 * (a < 0) T F F ( a < 0) T F F
2836 * (0 < a) F F T (-a < 0) F F T
2837 * (a >= 0) F T T ( a < 0) T F F (swap order of other operands)
2838 * (0 >= a) T T F (-a < 0) F F T (swap order of other operands)
2839 *
2840 * Note that exchanging the order of 0 and 'a' in the comparison simply
2841 * means that the value of 'a' should be negated.
2842 */
2843 if (src_ir != ir) {
2844 switch (expr->operation) {
2845 case ir_binop_less:
2846 switch_order = false;
2847 negate = zero_on_left;
2848 break;
2849
2850 case ir_binop_gequal:
2851 switch_order = true;
2852 negate = zero_on_left;
2853 break;
2854
2855 default:
2856 /* This isn't the right kind of comparison afterall, so make sure
2857 * the whole condition is visited.
2858 */
2859 src_ir = ir;
2860 break;
2861 }
2862 }
2863 }
2864
2865 src_ir->accept(this);
2866
2867 /* We use the TGSI_OPCODE_CMP (a < 0 ? b : c) for conditional moves, and the
2868 * condition we produced is 0.0 or 1.0. By flipping the sign, we can
2869 * choose which value TGSI_OPCODE_CMP produces without an extra instruction
2870 * computing the condition.
2871 */
2872 if (negate)
2873 this->result.negate = ~this->result.negate;
2874
2875 return switch_order;
2876 }
2877
2878 void
emit_block_mov(ir_assignment * ir,const struct glsl_type * type,st_dst_reg * l,st_src_reg * r,st_src_reg * cond,bool cond_swap)2879 glsl_to_tgsi_visitor::emit_block_mov(ir_assignment *ir, const struct glsl_type *type,
2880 st_dst_reg *l, st_src_reg *r,
2881 st_src_reg *cond, bool cond_swap)
2882 {
2883 if (type->is_record()) {
2884 for (unsigned int i = 0; i < type->length; i++) {
2885 emit_block_mov(ir, type->fields.structure[i].type, l, r,
2886 cond, cond_swap);
2887 }
2888 return;
2889 }
2890
2891 if (type->is_array()) {
2892 for (unsigned int i = 0; i < type->length; i++) {
2893 emit_block_mov(ir, type->fields.array, l, r, cond, cond_swap);
2894 }
2895 return;
2896 }
2897
2898 if (type->is_matrix()) {
2899 const struct glsl_type *vec_type;
2900
2901 vec_type = glsl_type::get_instance(type->is_double() ? GLSL_TYPE_DOUBLE : GLSL_TYPE_FLOAT,
2902 type->vector_elements, 1);
2903
2904 for (int i = 0; i < type->matrix_columns; i++) {
2905 emit_block_mov(ir, vec_type, l, r, cond, cond_swap);
2906 }
2907 return;
2908 }
2909
2910 assert(type->is_scalar() || type->is_vector());
2911
2912 l->type = type->base_type;
2913 r->type = type->base_type;
2914 if (cond) {
2915 st_src_reg l_src = st_src_reg(*l);
2916
2917 if (l_src.file == PROGRAM_OUTPUT &&
2918 this->prog->Target == GL_FRAGMENT_PROGRAM_ARB &&
2919 (l_src.index == FRAG_RESULT_DEPTH || l_src.index == FRAG_RESULT_STENCIL)) {
2920 /* This is a special case because the source swizzles will be shifted
2921 * later to account for the difference between GLSL (where they're
2922 * plain floats) and TGSI (where they're Z and Y components). */
2923 l_src.swizzle = SWIZZLE_XXXX;
2924 }
2925
2926 if (native_integers) {
2927 emit_asm(ir, TGSI_OPCODE_UCMP, *l, *cond,
2928 cond_swap ? l_src : *r,
2929 cond_swap ? *r : l_src);
2930 } else {
2931 emit_asm(ir, TGSI_OPCODE_CMP, *l, *cond,
2932 cond_swap ? l_src : *r,
2933 cond_swap ? *r : l_src);
2934 }
2935 } else {
2936 emit_asm(ir, TGSI_OPCODE_MOV, *l, *r);
2937 }
2938 l->index++;
2939 r->index++;
2940 if (type->is_dual_slot()) {
2941 l->index++;
2942 if (r->is_double_vertex_input == false)
2943 r->index++;
2944 }
2945 }
2946
2947 void
visit(ir_assignment * ir)2948 glsl_to_tgsi_visitor::visit(ir_assignment *ir)
2949 {
2950 int dst_component;
2951 st_dst_reg l;
2952 st_src_reg r;
2953
2954 /* all generated instructions need to be flaged as precise */
2955 this->precise = is_precise(ir->lhs->variable_referenced());
2956 ir->rhs->accept(this);
2957 r = this->result;
2958
2959 l = get_assignment_lhs(ir->lhs, this, &dst_component);
2960
2961 {
2962 int swizzles[4];
2963 int first_enabled_chan = 0;
2964 int rhs_chan = 0;
2965 ir_variable *variable = ir->lhs->variable_referenced();
2966
2967 if (shader->Stage == MESA_SHADER_FRAGMENT &&
2968 variable->data.mode == ir_var_shader_out &&
2969 (variable->data.location == FRAG_RESULT_DEPTH ||
2970 variable->data.location == FRAG_RESULT_STENCIL)) {
2971 assert(ir->lhs->type->is_scalar());
2972 assert(ir->write_mask == WRITEMASK_X);
2973
2974 if (variable->data.location == FRAG_RESULT_DEPTH)
2975 l.writemask = WRITEMASK_Z;
2976 else {
2977 assert(variable->data.location == FRAG_RESULT_STENCIL);
2978 l.writemask = WRITEMASK_Y;
2979 }
2980 } else if (ir->write_mask == 0) {
2981 assert(!ir->lhs->type->is_scalar() && !ir->lhs->type->is_vector());
2982
2983 unsigned num_elements = ir->lhs->type->without_array()->vector_elements;
2984
2985 if (num_elements) {
2986 l.writemask = u_bit_consecutive(0, num_elements);
2987 } else {
2988 /* The type is a struct or an array of (array of) structs. */
2989 l.writemask = WRITEMASK_XYZW;
2990 }
2991 } else {
2992 l.writemask = ir->write_mask;
2993 }
2994
2995 for (int i = 0; i < 4; i++) {
2996 if (l.writemask & (1 << i)) {
2997 first_enabled_chan = GET_SWZ(r.swizzle, i);
2998 break;
2999 }
3000 }
3001
3002 l.writemask = l.writemask << dst_component;
3003
3004 /* Swizzle a small RHS vector into the channels being written.
3005 *
3006 * glsl ir treats write_mask as dictating how many channels are
3007 * present on the RHS while TGSI treats write_mask as just
3008 * showing which channels of the vec4 RHS get written.
3009 */
3010 for (int i = 0; i < 4; i++) {
3011 if (l.writemask & (1 << i))
3012 swizzles[i] = GET_SWZ(r.swizzle, rhs_chan++);
3013 else
3014 swizzles[i] = first_enabled_chan;
3015 }
3016 r.swizzle = MAKE_SWIZZLE4(swizzles[0], swizzles[1],
3017 swizzles[2], swizzles[3]);
3018 }
3019
3020 assert(l.file != PROGRAM_UNDEFINED);
3021 assert(r.file != PROGRAM_UNDEFINED);
3022
3023 if (ir->condition) {
3024 const bool switch_order = this->process_move_condition(ir->condition);
3025 st_src_reg condition = this->result;
3026
3027 emit_block_mov(ir, ir->lhs->type, &l, &r, &condition, switch_order);
3028 } else if (ir->rhs->as_expression() &&
3029 this->instructions.get_tail() &&
3030 ir->rhs == ((glsl_to_tgsi_instruction *)this->instructions.get_tail())->ir &&
3031 !((glsl_to_tgsi_instruction *)this->instructions.get_tail())->is_64bit_expanded &&
3032 type_size(ir->lhs->type) == 1 &&
3033 l.writemask == ((glsl_to_tgsi_instruction *)this->instructions.get_tail())->dst[0].writemask) {
3034 /* To avoid emitting an extra MOV when assigning an expression to a
3035 * variable, emit the last instruction of the expression again, but
3036 * replace the destination register with the target of the assignment.
3037 * Dead code elimination will remove the original instruction.
3038 */
3039 glsl_to_tgsi_instruction *inst, *new_inst;
3040 inst = (glsl_to_tgsi_instruction *)this->instructions.get_tail();
3041 new_inst = emit_asm(ir, inst->op, l, inst->src[0], inst->src[1], inst->src[2], inst->src[3]);
3042 new_inst->saturate = inst->saturate;
3043 new_inst->resource = inst->resource;
3044 inst->dead_mask = inst->dst[0].writemask;
3045 } else {
3046 emit_block_mov(ir, ir->rhs->type, &l, &r, NULL, false);
3047 }
3048 this->precise = 0;
3049 }
3050
3051
3052 void
visit(ir_constant * ir)3053 glsl_to_tgsi_visitor::visit(ir_constant *ir)
3054 {
3055 st_src_reg src;
3056 GLdouble stack_vals[4] = { 0 };
3057 gl_constant_value *values = (gl_constant_value *) stack_vals;
3058 GLenum gl_type = GL_NONE;
3059 unsigned int i;
3060 static int in_array = 0;
3061 gl_register_file file = in_array ? PROGRAM_CONSTANT : PROGRAM_IMMEDIATE;
3062
3063 /* Unfortunately, 4 floats is all we can get into
3064 * _mesa_add_typed_unnamed_constant. So, make a temp to store an
3065 * aggregate constant and move each constant value into it. If we
3066 * get lucky, copy propagation will eliminate the extra moves.
3067 */
3068 if (ir->type->is_record()) {
3069 st_src_reg temp_base = get_temp(ir->type);
3070 st_dst_reg temp = st_dst_reg(temp_base);
3071
3072 for (i = 0; i < ir->type->length; i++) {
3073 ir_constant *const field_value = ir->get_record_field(i);
3074 int size = type_size(field_value->type);
3075
3076 assert(size > 0);
3077
3078 field_value->accept(this);
3079 src = this->result;
3080
3081 for (unsigned j = 0; j < (unsigned int)size; j++) {
3082 emit_asm(ir, TGSI_OPCODE_MOV, temp, src);
3083
3084 src.index++;
3085 temp.index++;
3086 }
3087 }
3088 this->result = temp_base;
3089 return;
3090 }
3091
3092 if (ir->type->is_array()) {
3093 st_src_reg temp_base = get_temp(ir->type);
3094 st_dst_reg temp = st_dst_reg(temp_base);
3095 int size = type_size(ir->type->fields.array);
3096
3097 assert(size > 0);
3098 in_array++;
3099
3100 for (i = 0; i < ir->type->length; i++) {
3101 ir->const_elements[i]->accept(this);
3102 src = this->result;
3103 for (int j = 0; j < size; j++) {
3104 emit_asm(ir, TGSI_OPCODE_MOV, temp, src);
3105
3106 src.index++;
3107 temp.index++;
3108 }
3109 }
3110 this->result = temp_base;
3111 in_array--;
3112 return;
3113 }
3114
3115 if (ir->type->is_matrix()) {
3116 st_src_reg mat = get_temp(ir->type);
3117 st_dst_reg mat_column = st_dst_reg(mat);
3118
3119 for (i = 0; i < ir->type->matrix_columns; i++) {
3120 switch (ir->type->base_type) {
3121 case GLSL_TYPE_FLOAT:
3122 values = (gl_constant_value *) &ir->value.f[i * ir->type->vector_elements];
3123
3124 src = st_src_reg(file, -1, ir->type->base_type);
3125 src.index = add_constant(file,
3126 values,
3127 ir->type->vector_elements,
3128 GL_FLOAT,
3129 &src.swizzle);
3130 emit_asm(ir, TGSI_OPCODE_MOV, mat_column, src);
3131 break;
3132 case GLSL_TYPE_DOUBLE:
3133 values = (gl_constant_value *) &ir->value.d[i * ir->type->vector_elements];
3134 src = st_src_reg(file, -1, ir->type->base_type);
3135 src.index = add_constant(file,
3136 values,
3137 ir->type->vector_elements,
3138 GL_DOUBLE,
3139 &src.swizzle);
3140 if (ir->type->vector_elements >= 2) {
3141 mat_column.writemask = WRITEMASK_XY;
3142 src.swizzle = MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, SWIZZLE_X, SWIZZLE_Y);
3143 emit_asm(ir, TGSI_OPCODE_MOV, mat_column, src);
3144 } else {
3145 mat_column.writemask = WRITEMASK_X;
3146 src.swizzle = MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_X, SWIZZLE_X, SWIZZLE_X);
3147 emit_asm(ir, TGSI_OPCODE_MOV, mat_column, src);
3148 }
3149 src.index++;
3150 if (ir->type->vector_elements > 2) {
3151 if (ir->type->vector_elements == 4) {
3152 mat_column.writemask = WRITEMASK_ZW;
3153 src.swizzle = MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, SWIZZLE_X, SWIZZLE_Y);
3154 emit_asm(ir, TGSI_OPCODE_MOV, mat_column, src);
3155 } else {
3156 mat_column.writemask = WRITEMASK_Z;
3157 src.swizzle = MAKE_SWIZZLE4(SWIZZLE_Y, SWIZZLE_Y, SWIZZLE_Y, SWIZZLE_Y);
3158 emit_asm(ir, TGSI_OPCODE_MOV, mat_column, src);
3159 mat_column.writemask = WRITEMASK_XYZW;
3160 src.swizzle = SWIZZLE_XYZW;
3161 }
3162 mat_column.index++;
3163 }
3164 break;
3165 default:
3166 unreachable("Illegal matrix constant type.\n");
3167 break;
3168 }
3169 mat_column.index++;
3170 }
3171 this->result = mat;
3172 return;
3173 }
3174
3175 switch (ir->type->base_type) {
3176 case GLSL_TYPE_FLOAT:
3177 gl_type = GL_FLOAT;
3178 for (i = 0; i < ir->type->vector_elements; i++) {
3179 values[i].f = ir->value.f[i];
3180 }
3181 break;
3182 case GLSL_TYPE_DOUBLE:
3183 gl_type = GL_DOUBLE;
3184 for (i = 0; i < ir->type->vector_elements; i++) {
3185 memcpy(&values[i * 2], &ir->value.d[i], sizeof(double));
3186 }
3187 break;
3188 case GLSL_TYPE_INT64:
3189 gl_type = GL_INT64_ARB;
3190 for (i = 0; i < ir->type->vector_elements; i++) {
3191 memcpy(&values[i * 2], &ir->value.d[i], sizeof(int64_t));
3192 }
3193 break;
3194 case GLSL_TYPE_UINT64:
3195 gl_type = GL_UNSIGNED_INT64_ARB;
3196 for (i = 0; i < ir->type->vector_elements; i++) {
3197 memcpy(&values[i * 2], &ir->value.d[i], sizeof(uint64_t));
3198 }
3199 break;
3200 case GLSL_TYPE_UINT:
3201 gl_type = native_integers ? GL_UNSIGNED_INT : GL_FLOAT;
3202 for (i = 0; i < ir->type->vector_elements; i++) {
3203 if (native_integers)
3204 values[i].u = ir->value.u[i];
3205 else
3206 values[i].f = ir->value.u[i];
3207 }
3208 break;
3209 case GLSL_TYPE_INT:
3210 gl_type = native_integers ? GL_INT : GL_FLOAT;
3211 for (i = 0; i < ir->type->vector_elements; i++) {
3212 if (native_integers)
3213 values[i].i = ir->value.i[i];
3214 else
3215 values[i].f = ir->value.i[i];
3216 }
3217 break;
3218 case GLSL_TYPE_BOOL:
3219 gl_type = native_integers ? GL_BOOL : GL_FLOAT;
3220 for (i = 0; i < ir->type->vector_elements; i++) {
3221 values[i].u = ir->value.b[i] ? ctx->Const.UniformBooleanTrue : 0;
3222 }
3223 break;
3224 default:
3225 assert(!"Non-float/uint/int/bool constant");
3226 }
3227
3228 this->result = st_src_reg(file, -1, ir->type);
3229 this->result.index = add_constant(file,
3230 values,
3231 ir->type->vector_elements,
3232 gl_type,
3233 &this->result.swizzle);
3234 }
3235
3236 void
visit_atomic_counter_intrinsic(ir_call * ir)3237 glsl_to_tgsi_visitor::visit_atomic_counter_intrinsic(ir_call *ir)
3238 {
3239 exec_node *param = ir->actual_parameters.get_head();
3240 ir_dereference *deref = static_cast<ir_dereference *>(param);
3241 ir_variable *location = deref->variable_referenced();
3242 bool has_hw_atomics = st_context(ctx)->has_hw_atomics;
3243 /* Calculate the surface offset */
3244 st_src_reg offset;
3245 unsigned array_size = 0, base = 0;
3246 uint16_t index = 0;
3247 st_src_reg resource;
3248
3249 get_deref_offsets(deref, &array_size, &base, &index, &offset, false);
3250
3251 if (has_hw_atomics) {
3252 variable_storage *entry = find_variable_storage(location);
3253 st_src_reg buffer(PROGRAM_HW_ATOMIC, 0, GLSL_TYPE_ATOMIC_UINT, location->data.binding);
3254
3255 if (!entry) {
3256 entry = new(mem_ctx) variable_storage(location, PROGRAM_HW_ATOMIC,
3257 num_atomics);
3258 _mesa_hash_table_insert(this->variables, location, entry);
3259
3260 atomic_info[num_atomics].location = location->data.location;
3261 atomic_info[num_atomics].binding = location->data.binding;
3262 atomic_info[num_atomics].size = location->type->arrays_of_arrays_size();
3263 if (atomic_info[num_atomics].size == 0)
3264 atomic_info[num_atomics].size = 1;
3265 atomic_info[num_atomics].array_id = 0;
3266 num_atomics++;
3267 }
3268
3269 if (offset.file != PROGRAM_UNDEFINED) {
3270 if (atomic_info[entry->index].array_id == 0) {
3271 num_atomic_arrays++;
3272 atomic_info[entry->index].array_id = num_atomic_arrays;
3273 }
3274 buffer.array_id = atomic_info[entry->index].array_id;
3275 }
3276
3277 buffer.index = index;
3278 buffer.index += location->data.offset / ATOMIC_COUNTER_SIZE;
3279 buffer.has_index2 = true;
3280
3281 if (offset.file != PROGRAM_UNDEFINED) {
3282 buffer.reladdr = ralloc(mem_ctx, st_src_reg);
3283 *buffer.reladdr = offset;
3284 emit_arl(ir, sampler_reladdr, offset);
3285 }
3286 offset = st_src_reg_for_int(0);
3287
3288 resource = buffer;
3289 } else {
3290 st_src_reg buffer(PROGRAM_BUFFER, location->data.binding,
3291 GLSL_TYPE_ATOMIC_UINT);
3292
3293 if (offset.file != PROGRAM_UNDEFINED) {
3294 emit_asm(ir, TGSI_OPCODE_MUL, st_dst_reg(offset),
3295 offset, st_src_reg_for_int(ATOMIC_COUNTER_SIZE));
3296 emit_asm(ir, TGSI_OPCODE_ADD, st_dst_reg(offset),
3297 offset, st_src_reg_for_int(location->data.offset + index * ATOMIC_COUNTER_SIZE));
3298 } else {
3299 offset = st_src_reg_for_int(location->data.offset + index * ATOMIC_COUNTER_SIZE);
3300 }
3301 resource = buffer;
3302 }
3303
3304 ir->return_deref->accept(this);
3305 st_dst_reg dst(this->result);
3306 dst.writemask = WRITEMASK_X;
3307
3308 glsl_to_tgsi_instruction *inst;
3309
3310 if (ir->callee->intrinsic_id == ir_intrinsic_atomic_counter_read) {
3311 inst = emit_asm(ir, TGSI_OPCODE_LOAD, dst, offset);
3312 } else if (ir->callee->intrinsic_id == ir_intrinsic_atomic_counter_increment) {
3313 inst = emit_asm(ir, TGSI_OPCODE_ATOMUADD, dst, offset,
3314 st_src_reg_for_int(1));
3315 } else if (ir->callee->intrinsic_id == ir_intrinsic_atomic_counter_predecrement) {
3316 inst = emit_asm(ir, TGSI_OPCODE_ATOMUADD, dst, offset,
3317 st_src_reg_for_int(-1));
3318 emit_asm(ir, TGSI_OPCODE_ADD, dst, this->result, st_src_reg_for_int(-1));
3319 } else {
3320 param = param->get_next();
3321 ir_rvalue *val = ((ir_instruction *)param)->as_rvalue();
3322 val->accept(this);
3323
3324 st_src_reg data = this->result, data2 = undef_src;
3325 unsigned opcode;
3326 switch (ir->callee->intrinsic_id) {
3327 case ir_intrinsic_atomic_counter_add:
3328 opcode = TGSI_OPCODE_ATOMUADD;
3329 break;
3330 case ir_intrinsic_atomic_counter_min:
3331 opcode = TGSI_OPCODE_ATOMIMIN;
3332 break;
3333 case ir_intrinsic_atomic_counter_max:
3334 opcode = TGSI_OPCODE_ATOMIMAX;
3335 break;
3336 case ir_intrinsic_atomic_counter_and:
3337 opcode = TGSI_OPCODE_ATOMAND;
3338 break;
3339 case ir_intrinsic_atomic_counter_or:
3340 opcode = TGSI_OPCODE_ATOMOR;
3341 break;
3342 case ir_intrinsic_atomic_counter_xor:
3343 opcode = TGSI_OPCODE_ATOMXOR;
3344 break;
3345 case ir_intrinsic_atomic_counter_exchange:
3346 opcode = TGSI_OPCODE_ATOMXCHG;
3347 break;
3348 case ir_intrinsic_atomic_counter_comp_swap: {
3349 opcode = TGSI_OPCODE_ATOMCAS;
3350 param = param->get_next();
3351 val = ((ir_instruction *)param)->as_rvalue();
3352 val->accept(this);
3353 data2 = this->result;
3354 break;
3355 }
3356 default:
3357 assert(!"Unexpected intrinsic");
3358 return;
3359 }
3360
3361 inst = emit_asm(ir, opcode, dst, offset, data, data2);
3362 }
3363
3364 inst->resource = resource;
3365 }
3366
3367 void
visit_ssbo_intrinsic(ir_call * ir)3368 glsl_to_tgsi_visitor::visit_ssbo_intrinsic(ir_call *ir)
3369 {
3370 exec_node *param = ir->actual_parameters.get_head();
3371
3372 ir_rvalue *block = ((ir_instruction *)param)->as_rvalue();
3373
3374 param = param->get_next();
3375 ir_rvalue *offset = ((ir_instruction *)param)->as_rvalue();
3376
3377 ir_constant *const_block = block->as_constant();
3378 int buf_base = st_context(ctx)->has_hw_atomics ? 0 : ctx->Const.Program[shader->Stage].MaxAtomicBuffers;
3379 st_src_reg buffer(
3380 PROGRAM_BUFFER,
3381 buf_base + (const_block ? const_block->value.u[0] : 0),
3382 GLSL_TYPE_UINT);
3383
3384 if (!const_block) {
3385 block->accept(this);
3386 buffer.reladdr = ralloc(mem_ctx, st_src_reg);
3387 *buffer.reladdr = this->result;
3388 emit_arl(ir, sampler_reladdr, this->result);
3389 }
3390
3391 /* Calculate the surface offset */
3392 offset->accept(this);
3393 st_src_reg off = this->result;
3394
3395 st_dst_reg dst = undef_dst;
3396 if (ir->return_deref) {
3397 ir->return_deref->accept(this);
3398 dst = st_dst_reg(this->result);
3399 dst.writemask = (1 << ir->return_deref->type->vector_elements) - 1;
3400 }
3401
3402 glsl_to_tgsi_instruction *inst;
3403
3404 if (ir->callee->intrinsic_id == ir_intrinsic_ssbo_load) {
3405 inst = emit_asm(ir, TGSI_OPCODE_LOAD, dst, off);
3406 if (dst.type == GLSL_TYPE_BOOL)
3407 emit_asm(ir, TGSI_OPCODE_USNE, dst, st_src_reg(dst), st_src_reg_for_int(0));
3408 } else if (ir->callee->intrinsic_id == ir_intrinsic_ssbo_store) {
3409 param = param->get_next();
3410 ir_rvalue *val = ((ir_instruction *)param)->as_rvalue();
3411 val->accept(this);
3412
3413 param = param->get_next();
3414 ir_constant *write_mask = ((ir_instruction *)param)->as_constant();
3415 assert(write_mask);
3416 dst.writemask = write_mask->value.u[0];
3417
3418 dst.type = this->result.type;
3419 inst = emit_asm(ir, TGSI_OPCODE_STORE, dst, off, this->result);
3420 } else {
3421 param = param->get_next();
3422 ir_rvalue *val = ((ir_instruction *)param)->as_rvalue();
3423 val->accept(this);
3424
3425 st_src_reg data = this->result, data2 = undef_src;
3426 unsigned opcode;
3427 switch (ir->callee->intrinsic_id) {
3428 case ir_intrinsic_ssbo_atomic_add:
3429 opcode = TGSI_OPCODE_ATOMUADD;
3430 break;
3431 case ir_intrinsic_ssbo_atomic_min:
3432 opcode = TGSI_OPCODE_ATOMIMIN;
3433 break;
3434 case ir_intrinsic_ssbo_atomic_max:
3435 opcode = TGSI_OPCODE_ATOMIMAX;
3436 break;
3437 case ir_intrinsic_ssbo_atomic_and:
3438 opcode = TGSI_OPCODE_ATOMAND;
3439 break;
3440 case ir_intrinsic_ssbo_atomic_or:
3441 opcode = TGSI_OPCODE_ATOMOR;
3442 break;
3443 case ir_intrinsic_ssbo_atomic_xor:
3444 opcode = TGSI_OPCODE_ATOMXOR;
3445 break;
3446 case ir_intrinsic_ssbo_atomic_exchange:
3447 opcode = TGSI_OPCODE_ATOMXCHG;
3448 break;
3449 case ir_intrinsic_ssbo_atomic_comp_swap:
3450 opcode = TGSI_OPCODE_ATOMCAS;
3451 param = param->get_next();
3452 val = ((ir_instruction *)param)->as_rvalue();
3453 val->accept(this);
3454 data2 = this->result;
3455 break;
3456 default:
3457 assert(!"Unexpected intrinsic");
3458 return;
3459 }
3460
3461 inst = emit_asm(ir, opcode, dst, off, data, data2);
3462 }
3463
3464 param = param->get_next();
3465 ir_constant *access = NULL;
3466 if (!param->is_tail_sentinel()) {
3467 access = ((ir_instruction *)param)->as_constant();
3468 assert(access);
3469 }
3470
3471 add_buffer_to_load_and_stores(inst, &buffer, &this->instructions, access);
3472 }
3473
3474 void
visit_membar_intrinsic(ir_call * ir)3475 glsl_to_tgsi_visitor::visit_membar_intrinsic(ir_call *ir)
3476 {
3477 switch (ir->callee->intrinsic_id) {
3478 case ir_intrinsic_memory_barrier:
3479 emit_asm(ir, TGSI_OPCODE_MEMBAR, undef_dst,
3480 st_src_reg_for_int(TGSI_MEMBAR_SHADER_BUFFER |
3481 TGSI_MEMBAR_ATOMIC_BUFFER |
3482 TGSI_MEMBAR_SHADER_IMAGE |
3483 TGSI_MEMBAR_SHARED));
3484 break;
3485 case ir_intrinsic_memory_barrier_atomic_counter:
3486 emit_asm(ir, TGSI_OPCODE_MEMBAR, undef_dst,
3487 st_src_reg_for_int(TGSI_MEMBAR_ATOMIC_BUFFER));
3488 break;
3489 case ir_intrinsic_memory_barrier_buffer:
3490 emit_asm(ir, TGSI_OPCODE_MEMBAR, undef_dst,
3491 st_src_reg_for_int(TGSI_MEMBAR_SHADER_BUFFER));
3492 break;
3493 case ir_intrinsic_memory_barrier_image:
3494 emit_asm(ir, TGSI_OPCODE_MEMBAR, undef_dst,
3495 st_src_reg_for_int(TGSI_MEMBAR_SHADER_IMAGE));
3496 break;
3497 case ir_intrinsic_memory_barrier_shared:
3498 emit_asm(ir, TGSI_OPCODE_MEMBAR, undef_dst,
3499 st_src_reg_for_int(TGSI_MEMBAR_SHARED));
3500 break;
3501 case ir_intrinsic_group_memory_barrier:
3502 emit_asm(ir, TGSI_OPCODE_MEMBAR, undef_dst,
3503 st_src_reg_for_int(TGSI_MEMBAR_SHADER_BUFFER |
3504 TGSI_MEMBAR_ATOMIC_BUFFER |
3505 TGSI_MEMBAR_SHADER_IMAGE |
3506 TGSI_MEMBAR_SHARED |
3507 TGSI_MEMBAR_THREAD_GROUP));
3508 break;
3509 default:
3510 assert(!"Unexpected memory barrier intrinsic");
3511 }
3512 }
3513
3514 void
visit_shared_intrinsic(ir_call * ir)3515 glsl_to_tgsi_visitor::visit_shared_intrinsic(ir_call *ir)
3516 {
3517 exec_node *param = ir->actual_parameters.get_head();
3518
3519 ir_rvalue *offset = ((ir_instruction *)param)->as_rvalue();
3520
3521 st_src_reg buffer(PROGRAM_MEMORY, 0, GLSL_TYPE_UINT);
3522
3523 /* Calculate the surface offset */
3524 offset->accept(this);
3525 st_src_reg off = this->result;
3526
3527 st_dst_reg dst = undef_dst;
3528 if (ir->return_deref) {
3529 ir->return_deref->accept(this);
3530 dst = st_dst_reg(this->result);
3531 dst.writemask = (1 << ir->return_deref->type->vector_elements) - 1;
3532 }
3533
3534 glsl_to_tgsi_instruction *inst;
3535
3536 if (ir->callee->intrinsic_id == ir_intrinsic_shared_load) {
3537 inst = emit_asm(ir, TGSI_OPCODE_LOAD, dst, off);
3538 inst->resource = buffer;
3539 } else if (ir->callee->intrinsic_id == ir_intrinsic_shared_store) {
3540 param = param->get_next();
3541 ir_rvalue *val = ((ir_instruction *)param)->as_rvalue();
3542 val->accept(this);
3543
3544 param = param->get_next();
3545 ir_constant *write_mask = ((ir_instruction *)param)->as_constant();
3546 assert(write_mask);
3547 dst.writemask = write_mask->value.u[0];
3548
3549 dst.type = this->result.type;
3550 inst = emit_asm(ir, TGSI_OPCODE_STORE, dst, off, this->result);
3551 inst->resource = buffer;
3552 } else {
3553 param = param->get_next();
3554 ir_rvalue *val = ((ir_instruction *)param)->as_rvalue();
3555 val->accept(this);
3556
3557 st_src_reg data = this->result, data2 = undef_src;
3558 unsigned opcode;
3559 switch (ir->callee->intrinsic_id) {
3560 case ir_intrinsic_shared_atomic_add:
3561 opcode = TGSI_OPCODE_ATOMUADD;
3562 break;
3563 case ir_intrinsic_shared_atomic_min:
3564 opcode = TGSI_OPCODE_ATOMIMIN;
3565 break;
3566 case ir_intrinsic_shared_atomic_max:
3567 opcode = TGSI_OPCODE_ATOMIMAX;
3568 break;
3569 case ir_intrinsic_shared_atomic_and:
3570 opcode = TGSI_OPCODE_ATOMAND;
3571 break;
3572 case ir_intrinsic_shared_atomic_or:
3573 opcode = TGSI_OPCODE_ATOMOR;
3574 break;
3575 case ir_intrinsic_shared_atomic_xor:
3576 opcode = TGSI_OPCODE_ATOMXOR;
3577 break;
3578 case ir_intrinsic_shared_atomic_exchange:
3579 opcode = TGSI_OPCODE_ATOMXCHG;
3580 break;
3581 case ir_intrinsic_shared_atomic_comp_swap:
3582 opcode = TGSI_OPCODE_ATOMCAS;
3583 param = param->get_next();
3584 val = ((ir_instruction *)param)->as_rvalue();
3585 val->accept(this);
3586 data2 = this->result;
3587 break;
3588 default:
3589 assert(!"Unexpected intrinsic");
3590 return;
3591 }
3592
3593 inst = emit_asm(ir, opcode, dst, off, data, data2);
3594 inst->resource = buffer;
3595 }
3596 }
3597
3598 static void
get_image_qualifiers(ir_dereference * ir,const glsl_type ** type,bool * memory_coherent,bool * memory_volatile,bool * memory_restrict,unsigned * image_format)3599 get_image_qualifiers(ir_dereference *ir, const glsl_type **type,
3600 bool *memory_coherent, bool *memory_volatile,
3601 bool *memory_restrict, unsigned *image_format)
3602 {
3603
3604 switch (ir->ir_type) {
3605 case ir_type_dereference_record: {
3606 ir_dereference_record *deref_record = ir->as_dereference_record();
3607 const glsl_type *struct_type = deref_record->record->type;
3608 int fild_idx = deref_record->field_idx;
3609
3610 *type = struct_type->fields.structure[fild_idx].type->without_array();
3611 *memory_coherent =
3612 struct_type->fields.structure[fild_idx].memory_coherent;
3613 *memory_volatile =
3614 struct_type->fields.structure[fild_idx].memory_volatile;
3615 *memory_restrict =
3616 struct_type->fields.structure[fild_idx].memory_restrict;
3617 *image_format =
3618 struct_type->fields.structure[fild_idx].image_format;
3619 break;
3620 }
3621
3622 case ir_type_dereference_array: {
3623 ir_dereference_array *deref_arr = ir->as_dereference_array();
3624 get_image_qualifiers((ir_dereference *)deref_arr->array, type,
3625 memory_coherent, memory_volatile, memory_restrict,
3626 image_format);
3627 break;
3628 }
3629
3630 case ir_type_dereference_variable: {
3631 ir_variable *var = ir->variable_referenced();
3632
3633 *type = var->type->without_array();
3634 *memory_coherent = var->data.memory_coherent;
3635 *memory_volatile = var->data.memory_volatile;
3636 *memory_restrict = var->data.memory_restrict;
3637 *image_format = var->data.image_format;
3638 break;
3639 }
3640
3641 default:
3642 break;
3643 }
3644 }
3645
3646 void
visit_image_intrinsic(ir_call * ir)3647 glsl_to_tgsi_visitor::visit_image_intrinsic(ir_call *ir)
3648 {
3649 exec_node *param = ir->actual_parameters.get_head();
3650
3651 ir_dereference *img = (ir_dereference *)param;
3652 const ir_variable *imgvar = img->variable_referenced();
3653 unsigned sampler_array_size = 1, sampler_base = 0;
3654 bool memory_coherent = false, memory_volatile = false, memory_restrict = false;
3655 unsigned image_format = 0;
3656 const glsl_type *type = NULL;
3657
3658 get_image_qualifiers(img, &type, &memory_coherent, &memory_volatile,
3659 &memory_restrict, &image_format);
3660
3661 st_src_reg reladdr;
3662 st_src_reg image(PROGRAM_IMAGE, 0, GLSL_TYPE_UINT);
3663 uint16_t index = 0;
3664 get_deref_offsets(img, &sampler_array_size, &sampler_base,
3665 &index, &reladdr, !imgvar->contains_bindless());
3666
3667 image.index = index;
3668 if (reladdr.file != PROGRAM_UNDEFINED) {
3669 image.reladdr = ralloc(mem_ctx, st_src_reg);
3670 *image.reladdr = reladdr;
3671 emit_arl(ir, sampler_reladdr, reladdr);
3672 }
3673
3674 st_dst_reg dst = undef_dst;
3675 if (ir->return_deref) {
3676 ir->return_deref->accept(this);
3677 dst = st_dst_reg(this->result);
3678 dst.writemask = (1 << ir->return_deref->type->vector_elements) - 1;
3679 }
3680
3681 glsl_to_tgsi_instruction *inst;
3682
3683 st_src_reg bindless;
3684 if (imgvar->contains_bindless()) {
3685 img->accept(this);
3686 bindless = this->result;
3687 }
3688
3689 if (ir->callee->intrinsic_id == ir_intrinsic_image_size) {
3690 dst.writemask = WRITEMASK_XYZ;
3691 inst = emit_asm(ir, TGSI_OPCODE_RESQ, dst);
3692 } else if (ir->callee->intrinsic_id == ir_intrinsic_image_samples) {
3693 st_src_reg res = get_temp(glsl_type::ivec4_type);
3694 st_dst_reg dstres = st_dst_reg(res);
3695 dstres.writemask = WRITEMASK_W;
3696 inst = emit_asm(ir, TGSI_OPCODE_RESQ, dstres);
3697 res.swizzle = SWIZZLE_WWWW;
3698 emit_asm(ir, TGSI_OPCODE_MOV, dst, res);
3699 } else {
3700 st_src_reg arg1 = undef_src, arg2 = undef_src;
3701 st_src_reg coord;
3702 st_dst_reg coord_dst;
3703 coord = get_temp(glsl_type::ivec4_type);
3704 coord_dst = st_dst_reg(coord);
3705 coord_dst.writemask = (1 << type->coordinate_components()) - 1;
3706 param = param->get_next();
3707 ((ir_dereference *)param)->accept(this);
3708 emit_asm(ir, TGSI_OPCODE_MOV, coord_dst, this->result);
3709 coord.swizzle = SWIZZLE_XXXX;
3710 switch (type->coordinate_components()) {
3711 case 4: assert(!"unexpected coord count");
3712 /* fallthrough */
3713 case 3: coord.swizzle |= SWIZZLE_Z << 6;
3714 /* fallthrough */
3715 case 2: coord.swizzle |= SWIZZLE_Y << 3;
3716 }
3717
3718 if (type->sampler_dimensionality == GLSL_SAMPLER_DIM_MS) {
3719 param = param->get_next();
3720 ((ir_dereference *)param)->accept(this);
3721 st_src_reg sample = this->result;
3722 sample.swizzle = SWIZZLE_XXXX;
3723 coord_dst.writemask = WRITEMASK_W;
3724 emit_asm(ir, TGSI_OPCODE_MOV, coord_dst, sample);
3725 coord.swizzle |= SWIZZLE_W << 9;
3726 }
3727
3728 param = param->get_next();
3729 if (!param->is_tail_sentinel()) {
3730 ((ir_dereference *)param)->accept(this);
3731 arg1 = this->result;
3732 param = param->get_next();
3733 }
3734
3735 if (!param->is_tail_sentinel()) {
3736 ((ir_dereference *)param)->accept(this);
3737 arg2 = this->result;
3738 param = param->get_next();
3739 }
3740
3741 assert(param->is_tail_sentinel());
3742
3743 unsigned opcode;
3744 switch (ir->callee->intrinsic_id) {
3745 case ir_intrinsic_image_load:
3746 opcode = TGSI_OPCODE_LOAD;
3747 break;
3748 case ir_intrinsic_image_store:
3749 opcode = TGSI_OPCODE_STORE;
3750 break;
3751 case ir_intrinsic_image_atomic_add:
3752 opcode = TGSI_OPCODE_ATOMUADD;
3753 break;
3754 case ir_intrinsic_image_atomic_min:
3755 opcode = TGSI_OPCODE_ATOMIMIN;
3756 break;
3757 case ir_intrinsic_image_atomic_max:
3758 opcode = TGSI_OPCODE_ATOMIMAX;
3759 break;
3760 case ir_intrinsic_image_atomic_and:
3761 opcode = TGSI_OPCODE_ATOMAND;
3762 break;
3763 case ir_intrinsic_image_atomic_or:
3764 opcode = TGSI_OPCODE_ATOMOR;
3765 break;
3766 case ir_intrinsic_image_atomic_xor:
3767 opcode = TGSI_OPCODE_ATOMXOR;
3768 break;
3769 case ir_intrinsic_image_atomic_exchange:
3770 opcode = TGSI_OPCODE_ATOMXCHG;
3771 break;
3772 case ir_intrinsic_image_atomic_comp_swap:
3773 opcode = TGSI_OPCODE_ATOMCAS;
3774 break;
3775 default:
3776 assert(!"Unexpected intrinsic");
3777 return;
3778 }
3779
3780 inst = emit_asm(ir, opcode, dst, coord, arg1, arg2);
3781 if (opcode == TGSI_OPCODE_STORE)
3782 inst->dst[0].writemask = WRITEMASK_XYZW;
3783 }
3784
3785 if (imgvar->contains_bindless()) {
3786 inst->resource = bindless;
3787 inst->resource.swizzle = MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y,
3788 SWIZZLE_X, SWIZZLE_Y);
3789 } else {
3790 inst->resource = image;
3791 inst->sampler_array_size = sampler_array_size;
3792 inst->sampler_base = sampler_base;
3793 }
3794
3795 inst->tex_target = type->sampler_index();
3796 inst->image_format = st_mesa_format_to_pipe_format(st_context(ctx),
3797 _mesa_get_shader_image_format(image_format));
3798
3799 if (memory_coherent)
3800 inst->buffer_access |= TGSI_MEMORY_COHERENT;
3801 if (memory_restrict)
3802 inst->buffer_access |= TGSI_MEMORY_RESTRICT;
3803 if (memory_volatile)
3804 inst->buffer_access |= TGSI_MEMORY_VOLATILE;
3805 }
3806
3807 void
visit_generic_intrinsic(ir_call * ir,unsigned op)3808 glsl_to_tgsi_visitor::visit_generic_intrinsic(ir_call *ir, unsigned op)
3809 {
3810 ir->return_deref->accept(this);
3811 st_dst_reg dst = st_dst_reg(this->result);
3812
3813 dst.writemask = u_bit_consecutive(0, ir->return_deref->var->type->vector_elements);
3814
3815 st_src_reg src[4] = { undef_src, undef_src, undef_src, undef_src };
3816 unsigned num_src = 0;
3817 foreach_in_list(ir_rvalue, param, &ir->actual_parameters) {
3818 assert(num_src < ARRAY_SIZE(src));
3819
3820 this->result.file = PROGRAM_UNDEFINED;
3821 param->accept(this);
3822 assert(this->result.file != PROGRAM_UNDEFINED);
3823
3824 src[num_src] = this->result;
3825 num_src++;
3826 }
3827
3828 emit_asm(ir, op, dst, src[0], src[1], src[2], src[3]);
3829 }
3830
3831 void
visit(ir_call * ir)3832 glsl_to_tgsi_visitor::visit(ir_call *ir)
3833 {
3834 ir_function_signature *sig = ir->callee;
3835
3836 /* Filter out intrinsics */
3837 switch (sig->intrinsic_id) {
3838 case ir_intrinsic_atomic_counter_read:
3839 case ir_intrinsic_atomic_counter_increment:
3840 case ir_intrinsic_atomic_counter_predecrement:
3841 case ir_intrinsic_atomic_counter_add:
3842 case ir_intrinsic_atomic_counter_min:
3843 case ir_intrinsic_atomic_counter_max:
3844 case ir_intrinsic_atomic_counter_and:
3845 case ir_intrinsic_atomic_counter_or:
3846 case ir_intrinsic_atomic_counter_xor:
3847 case ir_intrinsic_atomic_counter_exchange:
3848 case ir_intrinsic_atomic_counter_comp_swap:
3849 visit_atomic_counter_intrinsic(ir);
3850 return;
3851
3852 case ir_intrinsic_ssbo_load:
3853 case ir_intrinsic_ssbo_store:
3854 case ir_intrinsic_ssbo_atomic_add:
3855 case ir_intrinsic_ssbo_atomic_min:
3856 case ir_intrinsic_ssbo_atomic_max:
3857 case ir_intrinsic_ssbo_atomic_and:
3858 case ir_intrinsic_ssbo_atomic_or:
3859 case ir_intrinsic_ssbo_atomic_xor:
3860 case ir_intrinsic_ssbo_atomic_exchange:
3861 case ir_intrinsic_ssbo_atomic_comp_swap:
3862 visit_ssbo_intrinsic(ir);
3863 return;
3864
3865 case ir_intrinsic_memory_barrier:
3866 case ir_intrinsic_memory_barrier_atomic_counter:
3867 case ir_intrinsic_memory_barrier_buffer:
3868 case ir_intrinsic_memory_barrier_image:
3869 case ir_intrinsic_memory_barrier_shared:
3870 case ir_intrinsic_group_memory_barrier:
3871 visit_membar_intrinsic(ir);
3872 return;
3873
3874 case ir_intrinsic_shared_load:
3875 case ir_intrinsic_shared_store:
3876 case ir_intrinsic_shared_atomic_add:
3877 case ir_intrinsic_shared_atomic_min:
3878 case ir_intrinsic_shared_atomic_max:
3879 case ir_intrinsic_shared_atomic_and:
3880 case ir_intrinsic_shared_atomic_or:
3881 case ir_intrinsic_shared_atomic_xor:
3882 case ir_intrinsic_shared_atomic_exchange:
3883 case ir_intrinsic_shared_atomic_comp_swap:
3884 visit_shared_intrinsic(ir);
3885 return;
3886
3887 case ir_intrinsic_image_load:
3888 case ir_intrinsic_image_store:
3889 case ir_intrinsic_image_atomic_add:
3890 case ir_intrinsic_image_atomic_min:
3891 case ir_intrinsic_image_atomic_max:
3892 case ir_intrinsic_image_atomic_and:
3893 case ir_intrinsic_image_atomic_or:
3894 case ir_intrinsic_image_atomic_xor:
3895 case ir_intrinsic_image_atomic_exchange:
3896 case ir_intrinsic_image_atomic_comp_swap:
3897 case ir_intrinsic_image_size:
3898 case ir_intrinsic_image_samples:
3899 visit_image_intrinsic(ir);
3900 return;
3901
3902 case ir_intrinsic_shader_clock:
3903 visit_generic_intrinsic(ir, TGSI_OPCODE_CLOCK);
3904 return;
3905
3906 case ir_intrinsic_vote_all:
3907 visit_generic_intrinsic(ir, TGSI_OPCODE_VOTE_ALL);
3908 return;
3909 case ir_intrinsic_vote_any:
3910 visit_generic_intrinsic(ir, TGSI_OPCODE_VOTE_ANY);
3911 return;
3912 case ir_intrinsic_vote_eq:
3913 visit_generic_intrinsic(ir, TGSI_OPCODE_VOTE_EQ);
3914 return;
3915 case ir_intrinsic_ballot:
3916 visit_generic_intrinsic(ir, TGSI_OPCODE_BALLOT);
3917 return;
3918 case ir_intrinsic_read_first_invocation:
3919 visit_generic_intrinsic(ir, TGSI_OPCODE_READ_FIRST);
3920 return;
3921 case ir_intrinsic_read_invocation:
3922 visit_generic_intrinsic(ir, TGSI_OPCODE_READ_INVOC);
3923 return;
3924
3925 case ir_intrinsic_invalid:
3926 case ir_intrinsic_generic_load:
3927 case ir_intrinsic_generic_store:
3928 case ir_intrinsic_generic_atomic_add:
3929 case ir_intrinsic_generic_atomic_and:
3930 case ir_intrinsic_generic_atomic_or:
3931 case ir_intrinsic_generic_atomic_xor:
3932 case ir_intrinsic_generic_atomic_min:
3933 case ir_intrinsic_generic_atomic_max:
3934 case ir_intrinsic_generic_atomic_exchange:
3935 case ir_intrinsic_generic_atomic_comp_swap:
3936 unreachable("Invalid intrinsic");
3937 }
3938 }
3939
3940 void
calc_deref_offsets(ir_dereference * tail,unsigned * array_elements,uint16_t * index,st_src_reg * indirect,unsigned * location)3941 glsl_to_tgsi_visitor::calc_deref_offsets(ir_dereference *tail,
3942 unsigned *array_elements,
3943 uint16_t *index,
3944 st_src_reg *indirect,
3945 unsigned *location)
3946 {
3947 switch (tail->ir_type) {
3948 case ir_type_dereference_record: {
3949 ir_dereference_record *deref_record = tail->as_dereference_record();
3950 const glsl_type *struct_type = deref_record->record->type;
3951 int field_index = deref_record->field_idx;
3952
3953 calc_deref_offsets(deref_record->record->as_dereference(), array_elements, index, indirect, location);
3954
3955 assert(field_index >= 0);
3956 *location += struct_type->record_location_offset(field_index);
3957 break;
3958 }
3959
3960 case ir_type_dereference_array: {
3961 ir_dereference_array *deref_arr = tail->as_dereference_array();
3962
3963 void *mem_ctx = ralloc_parent(deref_arr);
3964 ir_constant *array_index =
3965 deref_arr->array_index->constant_expression_value(mem_ctx);
3966
3967 if (!array_index) {
3968 st_src_reg temp_reg;
3969 st_dst_reg temp_dst;
3970
3971 temp_reg = get_temp(glsl_type::uint_type);
3972 temp_dst = st_dst_reg(temp_reg);
3973 temp_dst.writemask = 1;
3974
3975 deref_arr->array_index->accept(this);
3976 if (*array_elements != 1)
3977 emit_asm(NULL, TGSI_OPCODE_MUL, temp_dst, this->result, st_src_reg_for_int(*array_elements));
3978 else
3979 emit_asm(NULL, TGSI_OPCODE_MOV, temp_dst, this->result);
3980
3981 if (indirect->file == PROGRAM_UNDEFINED)
3982 *indirect = temp_reg;
3983 else {
3984 temp_dst = st_dst_reg(*indirect);
3985 temp_dst.writemask = 1;
3986 emit_asm(NULL, TGSI_OPCODE_ADD, temp_dst, *indirect, temp_reg);
3987 }
3988 } else
3989 *index += array_index->value.u[0] * *array_elements;
3990
3991 *array_elements *= deref_arr->array->type->length;
3992
3993 calc_deref_offsets(deref_arr->array->as_dereference(), array_elements, index, indirect, location);
3994 break;
3995 }
3996 default:
3997 break;
3998 }
3999 }
4000
4001 void
get_deref_offsets(ir_dereference * ir,unsigned * array_size,unsigned * base,uint16_t * index,st_src_reg * reladdr,bool opaque)4002 glsl_to_tgsi_visitor::get_deref_offsets(ir_dereference *ir,
4003 unsigned *array_size,
4004 unsigned *base,
4005 uint16_t *index,
4006 st_src_reg *reladdr,
4007 bool opaque)
4008 {
4009 GLuint shader = _mesa_program_enum_to_shader_stage(this->prog->Target);
4010 unsigned location = 0;
4011 ir_variable *var = ir->variable_referenced();
4012
4013 memset(reladdr, 0, sizeof(*reladdr));
4014 reladdr->file = PROGRAM_UNDEFINED;
4015
4016 *base = 0;
4017 *array_size = 1;
4018
4019 assert(var);
4020 location = var->data.location;
4021 calc_deref_offsets(ir, array_size, index, reladdr, &location);
4022
4023 /*
4024 * If we end up with no indirect then adjust the base to the index,
4025 * and set the array size to 1.
4026 */
4027 if (reladdr->file == PROGRAM_UNDEFINED) {
4028 *base = *index;
4029 *array_size = 1;
4030 }
4031
4032 if (opaque) {
4033 assert(location != 0xffffffff);
4034 *base += this->shader_program->data->UniformStorage[location].opaque[shader].index;
4035 *index += this->shader_program->data->UniformStorage[location].opaque[shader].index;
4036 }
4037 }
4038
4039 st_src_reg
canonicalize_gather_offset(st_src_reg offset)4040 glsl_to_tgsi_visitor::canonicalize_gather_offset(st_src_reg offset)
4041 {
4042 if (offset.reladdr || offset.reladdr2) {
4043 st_src_reg tmp = get_temp(glsl_type::ivec2_type);
4044 st_dst_reg tmp_dst = st_dst_reg(tmp);
4045 tmp_dst.writemask = WRITEMASK_XY;
4046 emit_asm(NULL, TGSI_OPCODE_MOV, tmp_dst, offset);
4047 return tmp;
4048 }
4049
4050 return offset;
4051 }
4052
4053 void
visit(ir_texture * ir)4054 glsl_to_tgsi_visitor::visit(ir_texture *ir)
4055 {
4056 st_src_reg result_src, coord, cube_sc, lod_info, projector, dx, dy;
4057 st_src_reg offset[MAX_GLSL_TEXTURE_OFFSET], sample_index, component;
4058 st_src_reg levels_src, reladdr;
4059 st_dst_reg result_dst, coord_dst, cube_sc_dst;
4060 glsl_to_tgsi_instruction *inst = NULL;
4061 unsigned opcode = TGSI_OPCODE_NOP;
4062 const glsl_type *sampler_type = ir->sampler->type;
4063 unsigned sampler_array_size = 1, sampler_base = 0;
4064 bool is_cube_array = false, is_cube_shadow = false;
4065 ir_variable *var = ir->sampler->variable_referenced();
4066 unsigned i;
4067
4068 /* if we are a cube array sampler or a cube shadow */
4069 if (sampler_type->sampler_dimensionality == GLSL_SAMPLER_DIM_CUBE) {
4070 is_cube_array = sampler_type->sampler_array;
4071 is_cube_shadow = sampler_type->sampler_shadow;
4072 }
4073
4074 if (ir->coordinate) {
4075 ir->coordinate->accept(this);
4076
4077 /* Put our coords in a temp. We'll need to modify them for shadow,
4078 * projection, or LOD, so the only case we'd use it as-is is if
4079 * we're doing plain old texturing. The optimization passes on
4080 * glsl_to_tgsi_visitor should handle cleaning up our mess in that case.
4081 */
4082 coord = get_temp(glsl_type::vec4_type);
4083 coord_dst = st_dst_reg(coord);
4084 coord_dst.writemask = (1 << ir->coordinate->type->vector_elements) - 1;
4085 emit_asm(ir, TGSI_OPCODE_MOV, coord_dst, this->result);
4086 }
4087
4088 if (ir->projector) {
4089 ir->projector->accept(this);
4090 projector = this->result;
4091 }
4092
4093 /* Storage for our result. Ideally for an assignment we'd be using
4094 * the actual storage for the result here, instead.
4095 */
4096 result_src = get_temp(ir->type);
4097 result_dst = st_dst_reg(result_src);
4098 result_dst.writemask = (1 << ir->type->vector_elements) - 1;
4099
4100 switch (ir->op) {
4101 case ir_tex:
4102 opcode = (is_cube_array && ir->shadow_comparator) ? TGSI_OPCODE_TEX2 : TGSI_OPCODE_TEX;
4103 if (ir->offset) {
4104 ir->offset->accept(this);
4105 offset[0] = this->result;
4106 }
4107 break;
4108 case ir_txb:
4109 if (is_cube_array || is_cube_shadow) {
4110 opcode = TGSI_OPCODE_TXB2;
4111 }
4112 else {
4113 opcode = TGSI_OPCODE_TXB;
4114 }
4115 ir->lod_info.bias->accept(this);
4116 lod_info = this->result;
4117 if (ir->offset) {
4118 ir->offset->accept(this);
4119 offset[0] = this->result;
4120 }
4121 break;
4122 case ir_txl:
4123 if (this->has_tex_txf_lz && ir->lod_info.lod->is_zero()) {
4124 opcode = TGSI_OPCODE_TEX_LZ;
4125 } else {
4126 opcode = is_cube_array ? TGSI_OPCODE_TXL2 : TGSI_OPCODE_TXL;
4127 ir->lod_info.lod->accept(this);
4128 lod_info = this->result;
4129 }
4130 if (ir->offset) {
4131 ir->offset->accept(this);
4132 offset[0] = this->result;
4133 }
4134 break;
4135 case ir_txd:
4136 opcode = TGSI_OPCODE_TXD;
4137 ir->lod_info.grad.dPdx->accept(this);
4138 dx = this->result;
4139 ir->lod_info.grad.dPdy->accept(this);
4140 dy = this->result;
4141 if (ir->offset) {
4142 ir->offset->accept(this);
4143 offset[0] = this->result;
4144 }
4145 break;
4146 case ir_txs:
4147 opcode = TGSI_OPCODE_TXQ;
4148 ir->lod_info.lod->accept(this);
4149 lod_info = this->result;
4150 break;
4151 case ir_query_levels:
4152 opcode = TGSI_OPCODE_TXQ;
4153 lod_info = undef_src;
4154 levels_src = get_temp(ir->type);
4155 break;
4156 case ir_txf:
4157 if (this->has_tex_txf_lz && ir->lod_info.lod->is_zero()) {
4158 opcode = TGSI_OPCODE_TXF_LZ;
4159 } else {
4160 opcode = TGSI_OPCODE_TXF;
4161 ir->lod_info.lod->accept(this);
4162 lod_info = this->result;
4163 }
4164 if (ir->offset) {
4165 ir->offset->accept(this);
4166 offset[0] = this->result;
4167 }
4168 break;
4169 case ir_txf_ms:
4170 opcode = TGSI_OPCODE_TXF;
4171 ir->lod_info.sample_index->accept(this);
4172 sample_index = this->result;
4173 break;
4174 case ir_tg4:
4175 opcode = TGSI_OPCODE_TG4;
4176 ir->lod_info.component->accept(this);
4177 component = this->result;
4178 if (ir->offset) {
4179 ir->offset->accept(this);
4180 if (ir->offset->type->is_array()) {
4181 const glsl_type *elt_type = ir->offset->type->fields.array;
4182 for (i = 0; i < ir->offset->type->length; i++) {
4183 offset[i] = this->result;
4184 offset[i].index += i * type_size(elt_type);
4185 offset[i].type = elt_type->base_type;
4186 offset[i].swizzle = swizzle_for_size(elt_type->vector_elements);
4187 offset[i] = canonicalize_gather_offset(offset[i]);
4188 }
4189 } else {
4190 offset[0] = canonicalize_gather_offset(this->result);
4191 }
4192 }
4193 break;
4194 case ir_lod:
4195 opcode = TGSI_OPCODE_LODQ;
4196 break;
4197 case ir_texture_samples:
4198 opcode = TGSI_OPCODE_TXQS;
4199 break;
4200 case ir_samples_identical:
4201 unreachable("Unexpected ir_samples_identical opcode");
4202 }
4203
4204 if (ir->projector) {
4205 if (opcode == TGSI_OPCODE_TEX) {
4206 /* Slot the projector in as the last component of the coord. */
4207 coord_dst.writemask = WRITEMASK_W;
4208 emit_asm(ir, TGSI_OPCODE_MOV, coord_dst, projector);
4209 coord_dst.writemask = WRITEMASK_XYZW;
4210 opcode = TGSI_OPCODE_TXP;
4211 } else {
4212 st_src_reg coord_w = coord;
4213 coord_w.swizzle = SWIZZLE_WWWW;
4214
4215 /* For the other TEX opcodes there's no projective version
4216 * since the last slot is taken up by LOD info. Do the
4217 * projective divide now.
4218 */
4219 coord_dst.writemask = WRITEMASK_W;
4220 emit_asm(ir, TGSI_OPCODE_RCP, coord_dst, projector);
4221
4222 /* In the case where we have to project the coordinates "by hand,"
4223 * the shadow comparator value must also be projected.
4224 */
4225 st_src_reg tmp_src = coord;
4226 if (ir->shadow_comparator) {
4227 /* Slot the shadow value in as the second to last component of the
4228 * coord.
4229 */
4230 ir->shadow_comparator->accept(this);
4231
4232 tmp_src = get_temp(glsl_type::vec4_type);
4233 st_dst_reg tmp_dst = st_dst_reg(tmp_src);
4234
4235 /* Projective division not allowed for array samplers. */
4236 assert(!sampler_type->sampler_array);
4237
4238 tmp_dst.writemask = WRITEMASK_Z;
4239 emit_asm(ir, TGSI_OPCODE_MOV, tmp_dst, this->result);
4240
4241 tmp_dst.writemask = WRITEMASK_XY;
4242 emit_asm(ir, TGSI_OPCODE_MOV, tmp_dst, coord);
4243 }
4244
4245 coord_dst.writemask = WRITEMASK_XYZ;
4246 emit_asm(ir, TGSI_OPCODE_MUL, coord_dst, tmp_src, coord_w);
4247
4248 coord_dst.writemask = WRITEMASK_XYZW;
4249 coord.swizzle = SWIZZLE_XYZW;
4250 }
4251 }
4252
4253 /* If projection is done and the opcode is not TGSI_OPCODE_TXP, then the shadow
4254 * comparator was put in the correct place (and projected) by the code,
4255 * above, that handles by-hand projection.
4256 */
4257 if (ir->shadow_comparator && (!ir->projector || opcode == TGSI_OPCODE_TXP)) {
4258 /* Slot the shadow value in as the second to last component of the
4259 * coord.
4260 */
4261 ir->shadow_comparator->accept(this);
4262
4263 if (is_cube_array) {
4264 cube_sc = get_temp(glsl_type::float_type);
4265 cube_sc_dst = st_dst_reg(cube_sc);
4266 cube_sc_dst.writemask = WRITEMASK_X;
4267 emit_asm(ir, TGSI_OPCODE_MOV, cube_sc_dst, this->result);
4268 cube_sc_dst.writemask = WRITEMASK_X;
4269 }
4270 else {
4271 if ((sampler_type->sampler_dimensionality == GLSL_SAMPLER_DIM_2D &&
4272 sampler_type->sampler_array) ||
4273 sampler_type->sampler_dimensionality == GLSL_SAMPLER_DIM_CUBE) {
4274 coord_dst.writemask = WRITEMASK_W;
4275 } else {
4276 coord_dst.writemask = WRITEMASK_Z;
4277 }
4278 emit_asm(ir, TGSI_OPCODE_MOV, coord_dst, this->result);
4279 coord_dst.writemask = WRITEMASK_XYZW;
4280 }
4281 }
4282
4283 if (ir->op == ir_txf_ms) {
4284 coord_dst.writemask = WRITEMASK_W;
4285 emit_asm(ir, TGSI_OPCODE_MOV, coord_dst, sample_index);
4286 coord_dst.writemask = WRITEMASK_XYZW;
4287 } else if (opcode == TGSI_OPCODE_TXL || opcode == TGSI_OPCODE_TXB ||
4288 opcode == TGSI_OPCODE_TXF) {
4289 /* TGSI stores LOD or LOD bias in the last channel of the coords. */
4290 coord_dst.writemask = WRITEMASK_W;
4291 emit_asm(ir, TGSI_OPCODE_MOV, coord_dst, lod_info);
4292 coord_dst.writemask = WRITEMASK_XYZW;
4293 }
4294
4295 st_src_reg sampler(PROGRAM_SAMPLER, 0, GLSL_TYPE_UINT);
4296
4297 uint16_t index = 0;
4298 get_deref_offsets(ir->sampler, &sampler_array_size, &sampler_base,
4299 &index, &reladdr, !var->contains_bindless());
4300
4301 sampler.index = index;
4302 if (reladdr.file != PROGRAM_UNDEFINED) {
4303 sampler.reladdr = ralloc(mem_ctx, st_src_reg);
4304 *sampler.reladdr = reladdr;
4305 emit_arl(ir, sampler_reladdr, reladdr);
4306 }
4307
4308 st_src_reg bindless;
4309 if (var->contains_bindless()) {
4310 ir->sampler->accept(this);
4311 bindless = this->result;
4312 }
4313
4314 if (opcode == TGSI_OPCODE_TXD)
4315 inst = emit_asm(ir, opcode, result_dst, coord, dx, dy);
4316 else if (opcode == TGSI_OPCODE_TXQ) {
4317 if (ir->op == ir_query_levels) {
4318 /* the level is stored in W */
4319 inst = emit_asm(ir, opcode, st_dst_reg(levels_src), lod_info);
4320 result_dst.writemask = WRITEMASK_X;
4321 levels_src.swizzle = SWIZZLE_WWWW;
4322 emit_asm(ir, TGSI_OPCODE_MOV, result_dst, levels_src);
4323 } else
4324 inst = emit_asm(ir, opcode, result_dst, lod_info);
4325 } else if (opcode == TGSI_OPCODE_TXQS) {
4326 inst = emit_asm(ir, opcode, result_dst);
4327 } else if (opcode == TGSI_OPCODE_TXL2 || opcode == TGSI_OPCODE_TXB2) {
4328 inst = emit_asm(ir, opcode, result_dst, coord, lod_info);
4329 } else if (opcode == TGSI_OPCODE_TEX2) {
4330 inst = emit_asm(ir, opcode, result_dst, coord, cube_sc);
4331 } else if (opcode == TGSI_OPCODE_TG4) {
4332 if (is_cube_array && ir->shadow_comparator) {
4333 inst = emit_asm(ir, opcode, result_dst, coord, cube_sc);
4334 } else {
4335 inst = emit_asm(ir, opcode, result_dst, coord, component);
4336 }
4337 } else
4338 inst = emit_asm(ir, opcode, result_dst, coord);
4339
4340 if (ir->shadow_comparator)
4341 inst->tex_shadow = GL_TRUE;
4342
4343 if (var->contains_bindless()) {
4344 inst->resource = bindless;
4345 inst->resource.swizzle = MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y,
4346 SWIZZLE_X, SWIZZLE_Y);
4347 } else {
4348 inst->resource = sampler;
4349 inst->sampler_array_size = sampler_array_size;
4350 inst->sampler_base = sampler_base;
4351 }
4352
4353 if (ir->offset) {
4354 if (!inst->tex_offsets)
4355 inst->tex_offsets = rzalloc_array(inst, st_src_reg, MAX_GLSL_TEXTURE_OFFSET);
4356
4357 for (i = 0; i < MAX_GLSL_TEXTURE_OFFSET && offset[i].file != PROGRAM_UNDEFINED; i++)
4358 inst->tex_offsets[i] = offset[i];
4359 inst->tex_offset_num_offset = i;
4360 }
4361
4362 inst->tex_target = sampler_type->sampler_index();
4363 inst->tex_type = ir->type->base_type;
4364
4365 this->result = result_src;
4366 }
4367
4368 void
visit(ir_return * ir)4369 glsl_to_tgsi_visitor::visit(ir_return *ir)
4370 {
4371 assert(!ir->get_value());
4372
4373 emit_asm(ir, TGSI_OPCODE_RET);
4374 }
4375
4376 void
visit(ir_discard * ir)4377 glsl_to_tgsi_visitor::visit(ir_discard *ir)
4378 {
4379 if (ir->condition) {
4380 ir->condition->accept(this);
4381 st_src_reg condition = this->result;
4382
4383 /* Convert the bool condition to a float so we can negate. */
4384 if (native_integers) {
4385 st_src_reg temp = get_temp(ir->condition->type);
4386 emit_asm(ir, TGSI_OPCODE_AND, st_dst_reg(temp),
4387 condition, st_src_reg_for_float(1.0));
4388 condition = temp;
4389 }
4390
4391 condition.negate = ~condition.negate;
4392 emit_asm(ir, TGSI_OPCODE_KILL_IF, undef_dst, condition);
4393 } else {
4394 /* unconditional kil */
4395 emit_asm(ir, TGSI_OPCODE_KILL);
4396 }
4397 }
4398
4399 void
visit(ir_if * ir)4400 glsl_to_tgsi_visitor::visit(ir_if *ir)
4401 {
4402 unsigned if_opcode;
4403 glsl_to_tgsi_instruction *if_inst;
4404
4405 ir->condition->accept(this);
4406 assert(this->result.file != PROGRAM_UNDEFINED);
4407
4408 if_opcode = native_integers ? TGSI_OPCODE_UIF : TGSI_OPCODE_IF;
4409
4410 if_inst = emit_asm(ir->condition, if_opcode, undef_dst, this->result);
4411
4412 this->instructions.push_tail(if_inst);
4413
4414 visit_exec_list(&ir->then_instructions, this);
4415
4416 if (!ir->else_instructions.is_empty()) {
4417 emit_asm(ir->condition, TGSI_OPCODE_ELSE);
4418 visit_exec_list(&ir->else_instructions, this);
4419 }
4420
4421 if_inst = emit_asm(ir->condition, TGSI_OPCODE_ENDIF);
4422 }
4423
4424
4425 void
visit(ir_emit_vertex * ir)4426 glsl_to_tgsi_visitor::visit(ir_emit_vertex *ir)
4427 {
4428 assert(this->prog->Target == GL_GEOMETRY_PROGRAM_NV);
4429
4430 ir->stream->accept(this);
4431 emit_asm(ir, TGSI_OPCODE_EMIT, undef_dst, this->result);
4432 }
4433
4434 void
visit(ir_end_primitive * ir)4435 glsl_to_tgsi_visitor::visit(ir_end_primitive *ir)
4436 {
4437 assert(this->prog->Target == GL_GEOMETRY_PROGRAM_NV);
4438
4439 ir->stream->accept(this);
4440 emit_asm(ir, TGSI_OPCODE_ENDPRIM, undef_dst, this->result);
4441 }
4442
4443 void
visit(ir_barrier * ir)4444 glsl_to_tgsi_visitor::visit(ir_barrier *ir)
4445 {
4446 assert(this->prog->Target == GL_TESS_CONTROL_PROGRAM_NV ||
4447 this->prog->Target == GL_COMPUTE_PROGRAM_NV);
4448
4449 emit_asm(ir, TGSI_OPCODE_BARRIER);
4450 }
4451
glsl_to_tgsi_visitor()4452 glsl_to_tgsi_visitor::glsl_to_tgsi_visitor()
4453 {
4454 STATIC_ASSERT(sizeof(samplers_used) * 8 >= PIPE_MAX_SAMPLERS);
4455
4456 result.file = PROGRAM_UNDEFINED;
4457 next_temp = 1;
4458 array_sizes = NULL;
4459 max_num_arrays = 0;
4460 next_array = 0;
4461 num_inputs = 0;
4462 num_outputs = 0;
4463 num_input_arrays = 0;
4464 num_output_arrays = 0;
4465 num_atomics = 0;
4466 num_atomic_arrays = 0;
4467 num_immediates = 0;
4468 num_address_regs = 0;
4469 samplers_used = 0;
4470 images_used = 0;
4471 indirect_addr_consts = false;
4472 wpos_transform_const = -1;
4473 native_integers = false;
4474 mem_ctx = ralloc_context(NULL);
4475 ctx = NULL;
4476 prog = NULL;
4477 precise = 0;
4478 shader_program = NULL;
4479 shader = NULL;
4480 options = NULL;
4481 have_sqrt = false;
4482 have_fma = false;
4483 use_shared_memory = false;
4484 has_tex_txf_lz = false;
4485 variables = NULL;
4486 }
4487
var_destroy(struct hash_entry * entry)4488 static void var_destroy(struct hash_entry *entry)
4489 {
4490 variable_storage *storage = (variable_storage *)entry->data;
4491
4492 delete storage;
4493 }
4494
~glsl_to_tgsi_visitor()4495 glsl_to_tgsi_visitor::~glsl_to_tgsi_visitor()
4496 {
4497 _mesa_hash_table_destroy(variables, var_destroy);
4498 free(array_sizes);
4499 ralloc_free(mem_ctx);
4500 }
4501
free_glsl_to_tgsi_visitor(glsl_to_tgsi_visitor * v)4502 extern "C" void free_glsl_to_tgsi_visitor(glsl_to_tgsi_visitor *v)
4503 {
4504 delete v;
4505 }
4506
4507
4508 /**
4509 * Count resources used by the given gpu program (number of texture
4510 * samplers, etc).
4511 */
4512 static void
count_resources(glsl_to_tgsi_visitor * v,gl_program * prog)4513 count_resources(glsl_to_tgsi_visitor *v, gl_program *prog)
4514 {
4515 v->samplers_used = 0;
4516 v->images_used = 0;
4517 prog->info.textures_used_by_txf = 0;
4518
4519 foreach_in_list(glsl_to_tgsi_instruction, inst, &v->instructions) {
4520 if (inst->info->is_tex) {
4521 for (int i = 0; i < inst->sampler_array_size; i++) {
4522 unsigned idx = inst->sampler_base + i;
4523 v->samplers_used |= 1u << idx;
4524
4525 debug_assert(idx < (int)ARRAY_SIZE(v->sampler_types));
4526 v->sampler_types[idx] = inst->tex_type;
4527 v->sampler_targets[idx] =
4528 st_translate_texture_target(inst->tex_target, inst->tex_shadow);
4529
4530 if (inst->op == TGSI_OPCODE_TXF || inst->op == TGSI_OPCODE_TXF_LZ) {
4531 prog->info.textures_used_by_txf |= 1u << idx;
4532 }
4533 }
4534 }
4535
4536 if (inst->tex_target == TEXTURE_EXTERNAL_INDEX)
4537 prog->ExternalSamplersUsed |= 1 << inst->resource.index;
4538
4539 if (inst->resource.file != PROGRAM_UNDEFINED && (
4540 is_resource_instruction(inst->op) ||
4541 inst->op == TGSI_OPCODE_STORE)) {
4542 if (inst->resource.file == PROGRAM_MEMORY) {
4543 v->use_shared_memory = true;
4544 } else if (inst->resource.file == PROGRAM_IMAGE) {
4545 for (int i = 0; i < inst->sampler_array_size; i++) {
4546 unsigned idx = inst->sampler_base + i;
4547 v->images_used |= 1 << idx;
4548 v->image_targets[idx] =
4549 st_translate_texture_target(inst->tex_target, false);
4550 v->image_formats[idx] = inst->image_format;
4551 }
4552 }
4553 }
4554 }
4555 prog->SamplersUsed = v->samplers_used;
4556
4557 if (v->shader_program != NULL)
4558 _mesa_update_shader_textures_used(v->shader_program, prog);
4559 }
4560
4561 /**
4562 * Returns the mask of channels (bitmask of WRITEMASK_X,Y,Z,W) which
4563 * are read from the given src in this instruction
4564 */
4565 static int
get_src_arg_mask(st_dst_reg dst,st_src_reg src)4566 get_src_arg_mask(st_dst_reg dst, st_src_reg src)
4567 {
4568 int read_mask = 0, comp;
4569
4570 /* Now, given the src swizzle and the written channels, find which
4571 * components are actually read
4572 */
4573 for (comp = 0; comp < 4; ++comp) {
4574 const unsigned coord = GET_SWZ(src.swizzle, comp);
4575 assert(coord < 4);
4576 if (dst.writemask & (1 << comp) && coord <= SWIZZLE_W)
4577 read_mask |= 1 << coord;
4578 }
4579
4580 return read_mask;
4581 }
4582
4583 /**
4584 * This pass replaces CMP T0, T1 T2 T0 with MOV T0, T2 when the CMP
4585 * instruction is the first instruction to write to register T0. There are
4586 * several lowering passes done in GLSL IR (e.g. branches and
4587 * relative addressing) that create a large number of conditional assignments
4588 * that ir_to_mesa converts to CMP instructions like the one mentioned above.
4589 *
4590 * Here is why this conversion is safe:
4591 * CMP T0, T1 T2 T0 can be expanded to:
4592 * if (T1 < 0.0)
4593 * MOV T0, T2;
4594 * else
4595 * MOV T0, T0;
4596 *
4597 * If (T1 < 0.0) evaluates to true then our replacement MOV T0, T2 is the same
4598 * as the original program. If (T1 < 0.0) evaluates to false, executing
4599 * MOV T0, T0 will store a garbage value in T0 since T0 is uninitialized.
4600 * Therefore, it doesn't matter that we are replacing MOV T0, T0 with MOV T0, T2
4601 * because any instruction that was going to read from T0 after this was going
4602 * to read a garbage value anyway.
4603 */
4604 void
simplify_cmp(void)4605 glsl_to_tgsi_visitor::simplify_cmp(void)
4606 {
4607 int tempWritesSize = 0;
4608 unsigned *tempWrites = NULL;
4609 unsigned outputWrites[VARYING_SLOT_TESS_MAX];
4610
4611 memset(outputWrites, 0, sizeof(outputWrites));
4612
4613 foreach_in_list(glsl_to_tgsi_instruction, inst, &this->instructions) {
4614 unsigned prevWriteMask = 0;
4615
4616 /* Give up if we encounter relative addressing or flow control. */
4617 if (inst->dst[0].reladdr || inst->dst[0].reladdr2 ||
4618 inst->dst[1].reladdr || inst->dst[1].reladdr2 ||
4619 inst->info->is_branch ||
4620 inst->op == TGSI_OPCODE_CONT ||
4621 inst->op == TGSI_OPCODE_END ||
4622 inst->op == TGSI_OPCODE_RET) {
4623 break;
4624 }
4625
4626 if (inst->dst[0].file == PROGRAM_OUTPUT) {
4627 assert(inst->dst[0].index < (signed)ARRAY_SIZE(outputWrites));
4628 prevWriteMask = outputWrites[inst->dst[0].index];
4629 outputWrites[inst->dst[0].index] |= inst->dst[0].writemask;
4630 } else if (inst->dst[0].file == PROGRAM_TEMPORARY) {
4631 if (inst->dst[0].index >= tempWritesSize) {
4632 const int inc = 4096;
4633
4634 tempWrites = (unsigned*)
4635 realloc(tempWrites,
4636 (tempWritesSize + inc) * sizeof(unsigned));
4637 if (!tempWrites)
4638 return;
4639
4640 memset(tempWrites + tempWritesSize, 0, inc * sizeof(unsigned));
4641 tempWritesSize += inc;
4642 }
4643
4644 prevWriteMask = tempWrites[inst->dst[0].index];
4645 tempWrites[inst->dst[0].index] |= inst->dst[0].writemask;
4646 } else
4647 continue;
4648
4649 /* For a CMP to be considered a conditional write, the destination
4650 * register and source register two must be the same. */
4651 if (inst->op == TGSI_OPCODE_CMP
4652 && !(inst->dst[0].writemask & prevWriteMask)
4653 && inst->src[2].file == inst->dst[0].file
4654 && inst->src[2].index == inst->dst[0].index
4655 && inst->dst[0].writemask == get_src_arg_mask(inst->dst[0], inst->src[2])) {
4656
4657 inst->op = TGSI_OPCODE_MOV;
4658 inst->info = tgsi_get_opcode_info(inst->op);
4659 inst->src[0] = inst->src[1];
4660 }
4661 }
4662
4663 free(tempWrites);
4664 }
4665
4666 static void
rename_temp_handle_src(struct rename_reg_pair * renames,st_src_reg * src)4667 rename_temp_handle_src(struct rename_reg_pair *renames, st_src_reg *src)
4668 {
4669 if (src && src->file == PROGRAM_TEMPORARY) {
4670 int old_idx = src->index;
4671 if (renames[old_idx].valid)
4672 src->index = renames[old_idx].new_reg;
4673 }
4674 }
4675
4676 /* Replaces all references to a temporary register index with another index. */
4677 void
rename_temp_registers(struct rename_reg_pair * renames)4678 glsl_to_tgsi_visitor::rename_temp_registers(struct rename_reg_pair *renames)
4679 {
4680 foreach_in_list(glsl_to_tgsi_instruction, inst, &this->instructions) {
4681 unsigned j;
4682 for (j = 0; j < num_inst_src_regs(inst); j++) {
4683 rename_temp_handle_src(renames, &inst->src[j]);
4684 rename_temp_handle_src(renames, inst->src[j].reladdr);
4685 rename_temp_handle_src(renames, inst->src[j].reladdr2);
4686 }
4687
4688 for (j = 0; j < inst->tex_offset_num_offset; j++) {
4689 rename_temp_handle_src(renames, &inst->tex_offsets[j]);
4690 rename_temp_handle_src(renames, inst->tex_offsets[j].reladdr);
4691 rename_temp_handle_src(renames, inst->tex_offsets[j].reladdr2);
4692 }
4693
4694 rename_temp_handle_src(renames, &inst->resource);
4695 rename_temp_handle_src(renames, inst->resource.reladdr);
4696 rename_temp_handle_src(renames, inst->resource.reladdr2);
4697
4698 for (j = 0; j < num_inst_dst_regs(inst); j++) {
4699 if (inst->dst[j].file == PROGRAM_TEMPORARY) {
4700 int old_idx = inst->dst[j].index;
4701 if (renames[old_idx].valid)
4702 inst->dst[j].index = renames[old_idx].new_reg;
4703 }
4704 rename_temp_handle_src(renames, inst->dst[j].reladdr);
4705 rename_temp_handle_src(renames, inst->dst[j].reladdr2);
4706 }
4707 }
4708 }
4709
4710 void
get_first_temp_write(int * first_writes)4711 glsl_to_tgsi_visitor::get_first_temp_write(int *first_writes)
4712 {
4713 int depth = 0; /* loop depth */
4714 int loop_start = -1; /* index of the first active BGNLOOP (if any) */
4715 unsigned i = 0, j;
4716
4717 foreach_in_list(glsl_to_tgsi_instruction, inst, &this->instructions) {
4718 for (j = 0; j < num_inst_dst_regs(inst); j++) {
4719 if (inst->dst[j].file == PROGRAM_TEMPORARY) {
4720 if (first_writes[inst->dst[j].index] == -1)
4721 first_writes[inst->dst[j].index] = (depth == 0) ? i : loop_start;
4722 }
4723 }
4724
4725 if (inst->op == TGSI_OPCODE_BGNLOOP) {
4726 if(depth++ == 0)
4727 loop_start = i;
4728 } else if (inst->op == TGSI_OPCODE_ENDLOOP) {
4729 if (--depth == 0)
4730 loop_start = -1;
4731 }
4732 assert(depth >= 0);
4733 i++;
4734 }
4735 }
4736
4737 void
get_first_temp_read(int * first_reads)4738 glsl_to_tgsi_visitor::get_first_temp_read(int *first_reads)
4739 {
4740 int depth = 0; /* loop depth */
4741 int loop_start = -1; /* index of the first active BGNLOOP (if any) */
4742 unsigned i = 0, j;
4743
4744 foreach_in_list(glsl_to_tgsi_instruction, inst, &this->instructions) {
4745 for (j = 0; j < num_inst_src_regs(inst); j++) {
4746 if (inst->src[j].file == PROGRAM_TEMPORARY) {
4747 if (first_reads[inst->src[j].index] == -1)
4748 first_reads[inst->src[j].index] = (depth == 0) ? i : loop_start;
4749 }
4750 }
4751 for (j = 0; j < inst->tex_offset_num_offset; j++) {
4752 if (inst->tex_offsets[j].file == PROGRAM_TEMPORARY) {
4753 if (first_reads[inst->tex_offsets[j].index] == -1)
4754 first_reads[inst->tex_offsets[j].index] = (depth == 0) ? i : loop_start;
4755 }
4756 }
4757 if (inst->op == TGSI_OPCODE_BGNLOOP) {
4758 if(depth++ == 0)
4759 loop_start = i;
4760 } else if (inst->op == TGSI_OPCODE_ENDLOOP) {
4761 if (--depth == 0)
4762 loop_start = -1;
4763 }
4764 assert(depth >= 0);
4765 i++;
4766 }
4767 }
4768
4769 void
get_last_temp_read_first_temp_write(int * last_reads,int * first_writes)4770 glsl_to_tgsi_visitor::get_last_temp_read_first_temp_write(int *last_reads, int *first_writes)
4771 {
4772 int depth = 0; /* loop depth */
4773 int loop_start = -1; /* index of the first active BGNLOOP (if any) */
4774 unsigned i = 0, j;
4775 int k;
4776 foreach_in_list(glsl_to_tgsi_instruction, inst, &this->instructions) {
4777 for (j = 0; j < num_inst_src_regs(inst); j++) {
4778 if (inst->src[j].file == PROGRAM_TEMPORARY)
4779 last_reads[inst->src[j].index] = (depth == 0) ? i : -2;
4780 }
4781 for (j = 0; j < num_inst_dst_regs(inst); j++) {
4782 if (inst->dst[j].file == PROGRAM_TEMPORARY) {
4783 if (first_writes[inst->dst[j].index] == -1)
4784 first_writes[inst->dst[j].index] = (depth == 0) ? i : loop_start;
4785 last_reads[inst->dst[j].index] = (depth == 0) ? i : -2;
4786 }
4787 }
4788 for (j = 0; j < inst->tex_offset_num_offset; j++) {
4789 if (inst->tex_offsets[j].file == PROGRAM_TEMPORARY)
4790 last_reads[inst->tex_offsets[j].index] = (depth == 0) ? i : -2;
4791 }
4792 if (inst->op == TGSI_OPCODE_BGNLOOP) {
4793 if(depth++ == 0)
4794 loop_start = i;
4795 } else if (inst->op == TGSI_OPCODE_ENDLOOP) {
4796 if (--depth == 0) {
4797 loop_start = -1;
4798 for (k = 0; k < this->next_temp; k++) {
4799 if (last_reads[k] == -2) {
4800 last_reads[k] = i;
4801 }
4802 }
4803 }
4804 }
4805 assert(depth >= 0);
4806 i++;
4807 }
4808 }
4809
4810 void
get_last_temp_write(int * last_writes)4811 glsl_to_tgsi_visitor::get_last_temp_write(int *last_writes)
4812 {
4813 int depth = 0; /* loop depth */
4814 int i = 0, k;
4815 unsigned j;
4816
4817 foreach_in_list(glsl_to_tgsi_instruction, inst, &this->instructions) {
4818 for (j = 0; j < num_inst_dst_regs(inst); j++) {
4819 if (inst->dst[j].file == PROGRAM_TEMPORARY)
4820 last_writes[inst->dst[j].index] = (depth == 0) ? i : -2;
4821 }
4822
4823 if (inst->op == TGSI_OPCODE_BGNLOOP)
4824 depth++;
4825 else if (inst->op == TGSI_OPCODE_ENDLOOP)
4826 if (--depth == 0) {
4827 for (k = 0; k < this->next_temp; k++) {
4828 if (last_writes[k] == -2) {
4829 last_writes[k] = i;
4830 }
4831 }
4832 }
4833 assert(depth >= 0);
4834 i++;
4835 }
4836 }
4837
4838 /*
4839 * On a basic block basis, tracks available PROGRAM_TEMPORARY register
4840 * channels for copy propagation and updates following instructions to
4841 * use the original versions.
4842 *
4843 * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
4844 * will occur. As an example, a TXP production before this pass:
4845 *
4846 * 0: MOV TEMP[1], INPUT[4].xyyy;
4847 * 1: MOV TEMP[1].w, INPUT[4].wwww;
4848 * 2: TXP TEMP[2], TEMP[1], texture[0], 2D;
4849 *
4850 * and after:
4851 *
4852 * 0: MOV TEMP[1], INPUT[4].xyyy;
4853 * 1: MOV TEMP[1].w, INPUT[4].wwww;
4854 * 2: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
4855 *
4856 * which allows for dead code elimination on TEMP[1]'s writes.
4857 */
4858 void
copy_propagate(void)4859 glsl_to_tgsi_visitor::copy_propagate(void)
4860 {
4861 glsl_to_tgsi_instruction **acp = rzalloc_array(mem_ctx,
4862 glsl_to_tgsi_instruction *,
4863 this->next_temp * 4);
4864 int *acp_level = rzalloc_array(mem_ctx, int, this->next_temp * 4);
4865 int level = 0;
4866
4867 foreach_in_list(glsl_to_tgsi_instruction, inst, &this->instructions) {
4868 assert(inst->dst[0].file != PROGRAM_TEMPORARY
4869 || inst->dst[0].index < this->next_temp);
4870
4871 /* First, do any copy propagation possible into the src regs. */
4872 for (int r = 0; r < 3; r++) {
4873 glsl_to_tgsi_instruction *first = NULL;
4874 bool good = true;
4875 int acp_base = inst->src[r].index * 4;
4876
4877 if (inst->src[r].file != PROGRAM_TEMPORARY ||
4878 inst->src[r].reladdr ||
4879 inst->src[r].reladdr2)
4880 continue;
4881
4882 /* See if we can find entries in the ACP consisting of MOVs
4883 * from the same src register for all the swizzled channels
4884 * of this src register reference.
4885 */
4886 for (int i = 0; i < 4; i++) {
4887 int src_chan = GET_SWZ(inst->src[r].swizzle, i);
4888 glsl_to_tgsi_instruction *copy_chan = acp[acp_base + src_chan];
4889
4890 if (!copy_chan) {
4891 good = false;
4892 break;
4893 }
4894
4895 assert(acp_level[acp_base + src_chan] <= level);
4896
4897 if (!first) {
4898 first = copy_chan;
4899 } else {
4900 if (first->src[0].file != copy_chan->src[0].file ||
4901 first->src[0].index != copy_chan->src[0].index ||
4902 first->src[0].double_reg2 != copy_chan->src[0].double_reg2 ||
4903 first->src[0].index2D != copy_chan->src[0].index2D) {
4904 good = false;
4905 break;
4906 }
4907 }
4908 }
4909
4910 if (good) {
4911 /* We've now validated that we can copy-propagate to
4912 * replace this src register reference. Do it.
4913 */
4914 inst->src[r].file = first->src[0].file;
4915 inst->src[r].index = first->src[0].index;
4916 inst->src[r].index2D = first->src[0].index2D;
4917 inst->src[r].has_index2 = first->src[0].has_index2;
4918 inst->src[r].double_reg2 = first->src[0].double_reg2;
4919 inst->src[r].array_id = first->src[0].array_id;
4920
4921 int swizzle = 0;
4922 for (int i = 0; i < 4; i++) {
4923 int src_chan = GET_SWZ(inst->src[r].swizzle, i);
4924 glsl_to_tgsi_instruction *copy_inst = acp[acp_base + src_chan];
4925 swizzle |= (GET_SWZ(copy_inst->src[0].swizzle, src_chan) << (3 * i));
4926 }
4927 inst->src[r].swizzle = swizzle;
4928 }
4929 }
4930
4931 switch (inst->op) {
4932 case TGSI_OPCODE_BGNLOOP:
4933 case TGSI_OPCODE_ENDLOOP:
4934 /* End of a basic block, clear the ACP entirely. */
4935 memset(acp, 0, sizeof(*acp) * this->next_temp * 4);
4936 break;
4937
4938 case TGSI_OPCODE_IF:
4939 case TGSI_OPCODE_UIF:
4940 ++level;
4941 break;
4942
4943 case TGSI_OPCODE_ENDIF:
4944 case TGSI_OPCODE_ELSE:
4945 /* Clear all channels written inside the block from the ACP, but
4946 * leaving those that were not touched.
4947 */
4948 for (int r = 0; r < this->next_temp; r++) {
4949 for (int c = 0; c < 4; c++) {
4950 if (!acp[4 * r + c])
4951 continue;
4952
4953 if (acp_level[4 * r + c] >= level)
4954 acp[4 * r + c] = NULL;
4955 }
4956 }
4957 if (inst->op == TGSI_OPCODE_ENDIF)
4958 --level;
4959 break;
4960
4961 default:
4962 /* Continuing the block, clear any written channels from
4963 * the ACP.
4964 */
4965 for (int d = 0; d < 2; d++) {
4966 if (inst->dst[d].file == PROGRAM_TEMPORARY && inst->dst[d].reladdr) {
4967 /* Any temporary might be written, so no copy propagation
4968 * across this instruction.
4969 */
4970 memset(acp, 0, sizeof(*acp) * this->next_temp * 4);
4971 } else if (inst->dst[d].file == PROGRAM_OUTPUT &&
4972 inst->dst[d].reladdr) {
4973 /* Any output might be written, so no copy propagation
4974 * from outputs across this instruction.
4975 */
4976 for (int r = 0; r < this->next_temp; r++) {
4977 for (int c = 0; c < 4; c++) {
4978 if (!acp[4 * r + c])
4979 continue;
4980
4981 if (acp[4 * r + c]->src[0].file == PROGRAM_OUTPUT)
4982 acp[4 * r + c] = NULL;
4983 }
4984 }
4985 } else if (inst->dst[d].file == PROGRAM_TEMPORARY ||
4986 inst->dst[d].file == PROGRAM_OUTPUT) {
4987 /* Clear where it's used as dst. */
4988 if (inst->dst[d].file == PROGRAM_TEMPORARY) {
4989 for (int c = 0; c < 4; c++) {
4990 if (inst->dst[d].writemask & (1 << c))
4991 acp[4 * inst->dst[d].index + c] = NULL;
4992 }
4993 }
4994
4995 /* Clear where it's used as src. */
4996 for (int r = 0; r < this->next_temp; r++) {
4997 for (int c = 0; c < 4; c++) {
4998 if (!acp[4 * r + c])
4999 continue;
5000
5001 int src_chan = GET_SWZ(acp[4 * r + c]->src[0].swizzle, c);
5002
5003 if (acp[4 * r + c]->src[0].file == inst->dst[d].file &&
5004 acp[4 * r + c]->src[0].index == inst->dst[d].index &&
5005 inst->dst[d].writemask & (1 << src_chan)) {
5006 acp[4 * r + c] = NULL;
5007 }
5008 }
5009 }
5010 }
5011 }
5012 break;
5013 }
5014
5015 /* If this is a copy, add it to the ACP. */
5016 if (inst->op == TGSI_OPCODE_MOV &&
5017 inst->dst[0].file == PROGRAM_TEMPORARY &&
5018 !(inst->dst[0].file == inst->src[0].file &&
5019 inst->dst[0].index == inst->src[0].index) &&
5020 !inst->dst[0].reladdr &&
5021 !inst->dst[0].reladdr2 &&
5022 !inst->saturate &&
5023 inst->src[0].file != PROGRAM_ARRAY &&
5024 (inst->src[0].file != PROGRAM_OUTPUT ||
5025 this->shader->Stage != MESA_SHADER_TESS_CTRL) &&
5026 !inst->src[0].reladdr &&
5027 !inst->src[0].reladdr2 &&
5028 !inst->src[0].negate &&
5029 !inst->src[0].abs) {
5030 for (int i = 0; i < 4; i++) {
5031 if (inst->dst[0].writemask & (1 << i)) {
5032 acp[4 * inst->dst[0].index + i] = inst;
5033 acp_level[4 * inst->dst[0].index + i] = level;
5034 }
5035 }
5036 }
5037 }
5038
5039 ralloc_free(acp_level);
5040 ralloc_free(acp);
5041 }
5042
5043 static void
dead_code_handle_reladdr(glsl_to_tgsi_instruction ** writes,st_src_reg * reladdr)5044 dead_code_handle_reladdr(glsl_to_tgsi_instruction **writes, st_src_reg *reladdr)
5045 {
5046 if (reladdr && reladdr->file == PROGRAM_TEMPORARY) {
5047 /* Clear where it's used as src. */
5048 int swz = GET_SWZ(reladdr->swizzle, 0);
5049 writes[4 * reladdr->index + swz] = NULL;
5050 }
5051 }
5052
5053 /*
5054 * On a basic block basis, tracks available PROGRAM_TEMPORARY registers for dead
5055 * code elimination.
5056 *
5057 * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
5058 * will occur. As an example, a TXP production after copy propagation but
5059 * before this pass:
5060 *
5061 * 0: MOV TEMP[1], INPUT[4].xyyy;
5062 * 1: MOV TEMP[1].w, INPUT[4].wwww;
5063 * 2: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
5064 *
5065 * and after this pass:
5066 *
5067 * 0: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
5068 */
5069 int
eliminate_dead_code(void)5070 glsl_to_tgsi_visitor::eliminate_dead_code(void)
5071 {
5072 glsl_to_tgsi_instruction **writes = rzalloc_array(mem_ctx,
5073 glsl_to_tgsi_instruction *,
5074 this->next_temp * 4);
5075 int *write_level = rzalloc_array(mem_ctx, int, this->next_temp * 4);
5076 int level = 0;
5077 int removed = 0;
5078
5079 foreach_in_list(glsl_to_tgsi_instruction, inst, &this->instructions) {
5080 assert(inst->dst[0].file != PROGRAM_TEMPORARY
5081 || inst->dst[0].index < this->next_temp);
5082
5083 switch (inst->op) {
5084 case TGSI_OPCODE_BGNLOOP:
5085 case TGSI_OPCODE_ENDLOOP:
5086 case TGSI_OPCODE_CONT:
5087 case TGSI_OPCODE_BRK:
5088 /* End of a basic block, clear the write array entirely.
5089 *
5090 * This keeps us from killing dead code when the writes are
5091 * on either side of a loop, even when the register isn't touched
5092 * inside the loop. However, glsl_to_tgsi_visitor doesn't seem to emit
5093 * dead code of this type, so it shouldn't make a difference as long as
5094 * the dead code elimination pass in the GLSL compiler does its job.
5095 */
5096 memset(writes, 0, sizeof(*writes) * this->next_temp * 4);
5097 break;
5098
5099 case TGSI_OPCODE_ENDIF:
5100 case TGSI_OPCODE_ELSE:
5101 /* Promote the recorded level of all channels written inside the
5102 * preceding if or else block to the level above the if/else block.
5103 */
5104 for (int r = 0; r < this->next_temp; r++) {
5105 for (int c = 0; c < 4; c++) {
5106 if (!writes[4 * r + c])
5107 continue;
5108
5109 if (write_level[4 * r + c] == level)
5110 write_level[4 * r + c] = level-1;
5111 }
5112 }
5113 if(inst->op == TGSI_OPCODE_ENDIF)
5114 --level;
5115 break;
5116
5117 case TGSI_OPCODE_IF:
5118 case TGSI_OPCODE_UIF:
5119 ++level;
5120 /* fallthrough to default case to mark the condition as read */
5121 default:
5122 /* Continuing the block, clear any channels from the write array that
5123 * are read by this instruction.
5124 */
5125 for (unsigned i = 0; i < ARRAY_SIZE(inst->src); i++) {
5126 if (inst->src[i].file == PROGRAM_TEMPORARY && inst->src[i].reladdr){
5127 /* Any temporary might be read, so no dead code elimination
5128 * across this instruction.
5129 */
5130 memset(writes, 0, sizeof(*writes) * this->next_temp * 4);
5131 } else if (inst->src[i].file == PROGRAM_TEMPORARY) {
5132 /* Clear where it's used as src. */
5133 int src_chans = 1 << GET_SWZ(inst->src[i].swizzle, 0);
5134 src_chans |= 1 << GET_SWZ(inst->src[i].swizzle, 1);
5135 src_chans |= 1 << GET_SWZ(inst->src[i].swizzle, 2);
5136 src_chans |= 1 << GET_SWZ(inst->src[i].swizzle, 3);
5137
5138 for (int c = 0; c < 4; c++) {
5139 if (src_chans & (1 << c))
5140 writes[4 * inst->src[i].index + c] = NULL;
5141 }
5142 }
5143 dead_code_handle_reladdr(writes, inst->src[i].reladdr);
5144 dead_code_handle_reladdr(writes, inst->src[i].reladdr2);
5145 }
5146 for (unsigned i = 0; i < inst->tex_offset_num_offset; i++) {
5147 if (inst->tex_offsets[i].file == PROGRAM_TEMPORARY && inst->tex_offsets[i].reladdr){
5148 /* Any temporary might be read, so no dead code elimination
5149 * across this instruction.
5150 */
5151 memset(writes, 0, sizeof(*writes) * this->next_temp * 4);
5152 } else if (inst->tex_offsets[i].file == PROGRAM_TEMPORARY) {
5153 /* Clear where it's used as src. */
5154 int src_chans = 1 << GET_SWZ(inst->tex_offsets[i].swizzle, 0);
5155 src_chans |= 1 << GET_SWZ(inst->tex_offsets[i].swizzle, 1);
5156 src_chans |= 1 << GET_SWZ(inst->tex_offsets[i].swizzle, 2);
5157 src_chans |= 1 << GET_SWZ(inst->tex_offsets[i].swizzle, 3);
5158
5159 for (int c = 0; c < 4; c++) {
5160 if (src_chans & (1 << c))
5161 writes[4 * inst->tex_offsets[i].index + c] = NULL;
5162 }
5163 }
5164 dead_code_handle_reladdr(writes, inst->tex_offsets[i].reladdr);
5165 dead_code_handle_reladdr(writes, inst->tex_offsets[i].reladdr2);
5166 }
5167
5168 if (inst->resource.file == PROGRAM_TEMPORARY) {
5169 int src_chans;
5170
5171 src_chans = 1 << GET_SWZ(inst->resource.swizzle, 0);
5172 src_chans |= 1 << GET_SWZ(inst->resource.swizzle, 1);
5173 src_chans |= 1 << GET_SWZ(inst->resource.swizzle, 2);
5174 src_chans |= 1 << GET_SWZ(inst->resource.swizzle, 3);
5175
5176 for (int c = 0; c < 4; c++) {
5177 if (src_chans & (1 << c))
5178 writes[4 * inst->resource.index + c] = NULL;
5179 }
5180 }
5181 dead_code_handle_reladdr(writes, inst->resource.reladdr);
5182 dead_code_handle_reladdr(writes, inst->resource.reladdr2);
5183
5184 for (unsigned i = 0; i < ARRAY_SIZE(inst->dst); i++) {
5185 dead_code_handle_reladdr(writes, inst->dst[i].reladdr);
5186 dead_code_handle_reladdr(writes, inst->dst[i].reladdr2);
5187 }
5188 break;
5189 }
5190
5191 /* If this instruction writes to a temporary, add it to the write array.
5192 * If there is already an instruction in the write array for one or more
5193 * of the channels, flag that channel write as dead.
5194 */
5195 for (unsigned i = 0; i < ARRAY_SIZE(inst->dst); i++) {
5196 if (inst->dst[i].file == PROGRAM_TEMPORARY &&
5197 !inst->dst[i].reladdr) {
5198 for (int c = 0; c < 4; c++) {
5199 if (inst->dst[i].writemask & (1 << c)) {
5200 if (writes[4 * inst->dst[i].index + c]) {
5201 if (write_level[4 * inst->dst[i].index + c] < level)
5202 continue;
5203 else
5204 writes[4 * inst->dst[i].index + c]->dead_mask |= (1 << c);
5205 }
5206 writes[4 * inst->dst[i].index + c] = inst;
5207 write_level[4 * inst->dst[i].index + c] = level;
5208 }
5209 }
5210 }
5211 }
5212 }
5213
5214 /* Anything still in the write array at this point is dead code. */
5215 for (int r = 0; r < this->next_temp; r++) {
5216 for (int c = 0; c < 4; c++) {
5217 glsl_to_tgsi_instruction *inst = writes[4 * r + c];
5218 if (inst)
5219 inst->dead_mask |= (1 << c);
5220 }
5221 }
5222
5223 /* Now actually remove the instructions that are completely dead and update
5224 * the writemask of other instructions with dead channels.
5225 */
5226 foreach_in_list_safe(glsl_to_tgsi_instruction, inst, &this->instructions) {
5227 if (!inst->dead_mask || !inst->dst[0].writemask)
5228 continue;
5229 /* No amount of dead masks should remove memory stores */
5230 if (inst->info->is_store)
5231 continue;
5232
5233 if ((inst->dst[0].writemask & ~inst->dead_mask) == 0) {
5234 inst->remove();
5235 delete inst;
5236 removed++;
5237 } else {
5238 if (glsl_base_type_is_64bit(inst->dst[0].type)) {
5239 if (inst->dead_mask == WRITEMASK_XY ||
5240 inst->dead_mask == WRITEMASK_ZW)
5241 inst->dst[0].writemask &= ~(inst->dead_mask);
5242 } else
5243 inst->dst[0].writemask &= ~(inst->dead_mask);
5244 }
5245 }
5246
5247 ralloc_free(write_level);
5248 ralloc_free(writes);
5249
5250 return removed;
5251 }
5252
5253 /* merge DFRACEXP instructions into one. */
5254 void
merge_two_dsts(void)5255 glsl_to_tgsi_visitor::merge_two_dsts(void)
5256 {
5257 /* We never delete inst, but we may delete its successor. */
5258 foreach_in_list(glsl_to_tgsi_instruction, inst, &this->instructions) {
5259 glsl_to_tgsi_instruction *inst2;
5260 unsigned defined;
5261
5262 if (num_inst_dst_regs(inst) != 2)
5263 continue;
5264
5265 if (inst->dst[0].file != PROGRAM_UNDEFINED &&
5266 inst->dst[1].file != PROGRAM_UNDEFINED)
5267 continue;
5268
5269 assert(inst->dst[0].file != PROGRAM_UNDEFINED ||
5270 inst->dst[1].file != PROGRAM_UNDEFINED);
5271
5272 if (inst->dst[0].file == PROGRAM_UNDEFINED)
5273 defined = 1;
5274 else
5275 defined = 0;
5276
5277 inst2 = (glsl_to_tgsi_instruction *) inst->next;
5278 while (!inst2->is_tail_sentinel()) {
5279 if (inst->op == inst2->op &&
5280 inst2->dst[defined].file == PROGRAM_UNDEFINED &&
5281 inst->src[0].file == inst2->src[0].file &&
5282 inst->src[0].index == inst2->src[0].index &&
5283 inst->src[0].type == inst2->src[0].type &&
5284 inst->src[0].swizzle == inst2->src[0].swizzle)
5285 break;
5286 inst2 = (glsl_to_tgsi_instruction *) inst2->next;
5287 }
5288
5289 if (inst2->is_tail_sentinel()) {
5290 /* Undefined destinations are not allowed, substitute with an unused
5291 * temporary register.
5292 */
5293 st_src_reg tmp = get_temp(glsl_type::vec4_type);
5294 inst->dst[defined ^ 1] = st_dst_reg(tmp);
5295 inst->dst[defined ^ 1].writemask = 0;
5296 continue;
5297 }
5298
5299 inst->dst[defined ^ 1] = inst2->dst[defined ^ 1];
5300 inst2->remove();
5301 delete inst2;
5302 }
5303 }
5304
5305 /* Merges temporary registers together where possible to reduce the number of
5306 * registers needed to run a program.
5307 *
5308 * Produces optimal code only after copy propagation and dead code elimination
5309 * have been run. */
5310 void
merge_registers(void)5311 glsl_to_tgsi_visitor::merge_registers(void)
5312 {
5313 assert(need_uarl);
5314 struct lifetime *lifetimes =
5315 rzalloc_array(mem_ctx, struct lifetime, this->next_temp);
5316
5317 if (get_temp_registers_required_lifetimes(mem_ctx, &this->instructions,
5318 this->next_temp, lifetimes)) {
5319 struct rename_reg_pair *renames =
5320 rzalloc_array(mem_ctx, struct rename_reg_pair, this->next_temp);
5321 get_temp_registers_remapping(mem_ctx, this->next_temp, lifetimes, renames);
5322 rename_temp_registers(renames);
5323 ralloc_free(renames);
5324 }
5325
5326 ralloc_free(lifetimes);
5327 }
5328
5329 /* Reassign indices to temporary registers by reusing unused indices created
5330 * by optimization passes. */
5331 void
renumber_registers(void)5332 glsl_to_tgsi_visitor::renumber_registers(void)
5333 {
5334 int i = 0;
5335 int new_index = 0;
5336 int *first_writes = ralloc_array(mem_ctx, int, this->next_temp);
5337 struct rename_reg_pair *renames = rzalloc_array(mem_ctx, struct rename_reg_pair, this->next_temp);
5338
5339 for (i = 0; i < this->next_temp; i++) {
5340 first_writes[i] = -1;
5341 }
5342 get_first_temp_write(first_writes);
5343
5344 for (i = 0; i < this->next_temp; i++) {
5345 if (first_writes[i] < 0) continue;
5346 if (i != new_index) {
5347 renames[i].new_reg = new_index;
5348 renames[i].valid = true;
5349 }
5350 new_index++;
5351 }
5352
5353 rename_temp_registers(renames);
5354 this->next_temp = new_index;
5355 ralloc_free(renames);
5356 ralloc_free(first_writes);
5357 }
5358
5359 /* ------------------------- TGSI conversion stuff -------------------------- */
5360
5361 /**
5362 * Intermediate state used during shader translation.
5363 */
5364 struct st_translate {
5365 struct ureg_program *ureg;
5366
5367 unsigned temps_size;
5368 struct ureg_dst *temps;
5369
5370 struct ureg_dst *arrays;
5371 unsigned num_temp_arrays;
5372 struct ureg_src *constants;
5373 int num_constants;
5374 struct ureg_src *immediates;
5375 int num_immediates;
5376 struct ureg_dst outputs[PIPE_MAX_SHADER_OUTPUTS];
5377 struct ureg_src inputs[PIPE_MAX_SHADER_INPUTS];
5378 struct ureg_dst address[3];
5379 struct ureg_src samplers[PIPE_MAX_SAMPLERS];
5380 struct ureg_src buffers[PIPE_MAX_SHADER_BUFFERS];
5381 struct ureg_src images[PIPE_MAX_SHADER_IMAGES];
5382 struct ureg_src systemValues[SYSTEM_VALUE_MAX];
5383 struct ureg_src hw_atomics[PIPE_MAX_HW_ATOMIC_BUFFERS];
5384 struct ureg_src shared_memory;
5385 unsigned *array_sizes;
5386 struct inout_decl *input_decls;
5387 unsigned num_input_decls;
5388 struct inout_decl *output_decls;
5389 unsigned num_output_decls;
5390
5391 const ubyte *inputMapping;
5392 const ubyte *outputMapping;
5393
5394 unsigned procType; /**< PIPE_SHADER_VERTEX/FRAGMENT */
5395 bool need_uarl;
5396 };
5397
5398 /** Map Mesa's SYSTEM_VALUE_x to TGSI_SEMANTIC_x */
5399 unsigned
_mesa_sysval_to_semantic(unsigned sysval)5400 _mesa_sysval_to_semantic(unsigned sysval)
5401 {
5402 switch (sysval) {
5403 /* Vertex shader */
5404 case SYSTEM_VALUE_VERTEX_ID:
5405 return TGSI_SEMANTIC_VERTEXID;
5406 case SYSTEM_VALUE_INSTANCE_ID:
5407 return TGSI_SEMANTIC_INSTANCEID;
5408 case SYSTEM_VALUE_VERTEX_ID_ZERO_BASE:
5409 return TGSI_SEMANTIC_VERTEXID_NOBASE;
5410 case SYSTEM_VALUE_BASE_VERTEX:
5411 return TGSI_SEMANTIC_BASEVERTEX;
5412 case SYSTEM_VALUE_BASE_INSTANCE:
5413 return TGSI_SEMANTIC_BASEINSTANCE;
5414 case SYSTEM_VALUE_DRAW_ID:
5415 return TGSI_SEMANTIC_DRAWID;
5416
5417 /* Geometry shader */
5418 case SYSTEM_VALUE_INVOCATION_ID:
5419 return TGSI_SEMANTIC_INVOCATIONID;
5420
5421 /* Fragment shader */
5422 case SYSTEM_VALUE_FRAG_COORD:
5423 return TGSI_SEMANTIC_POSITION;
5424 case SYSTEM_VALUE_FRONT_FACE:
5425 return TGSI_SEMANTIC_FACE;
5426 case SYSTEM_VALUE_SAMPLE_ID:
5427 return TGSI_SEMANTIC_SAMPLEID;
5428 case SYSTEM_VALUE_SAMPLE_POS:
5429 return TGSI_SEMANTIC_SAMPLEPOS;
5430 case SYSTEM_VALUE_SAMPLE_MASK_IN:
5431 return TGSI_SEMANTIC_SAMPLEMASK;
5432 case SYSTEM_VALUE_HELPER_INVOCATION:
5433 return TGSI_SEMANTIC_HELPER_INVOCATION;
5434
5435 /* Tessellation shader */
5436 case SYSTEM_VALUE_TESS_COORD:
5437 return TGSI_SEMANTIC_TESSCOORD;
5438 case SYSTEM_VALUE_VERTICES_IN:
5439 return TGSI_SEMANTIC_VERTICESIN;
5440 case SYSTEM_VALUE_PRIMITIVE_ID:
5441 return TGSI_SEMANTIC_PRIMID;
5442 case SYSTEM_VALUE_TESS_LEVEL_OUTER:
5443 return TGSI_SEMANTIC_TESSOUTER;
5444 case SYSTEM_VALUE_TESS_LEVEL_INNER:
5445 return TGSI_SEMANTIC_TESSINNER;
5446
5447 /* Compute shader */
5448 case SYSTEM_VALUE_LOCAL_INVOCATION_ID:
5449 return TGSI_SEMANTIC_THREAD_ID;
5450 case SYSTEM_VALUE_WORK_GROUP_ID:
5451 return TGSI_SEMANTIC_BLOCK_ID;
5452 case SYSTEM_VALUE_NUM_WORK_GROUPS:
5453 return TGSI_SEMANTIC_GRID_SIZE;
5454 case SYSTEM_VALUE_LOCAL_GROUP_SIZE:
5455 return TGSI_SEMANTIC_BLOCK_SIZE;
5456
5457 /* ARB_shader_ballot */
5458 case SYSTEM_VALUE_SUBGROUP_SIZE:
5459 return TGSI_SEMANTIC_SUBGROUP_SIZE;
5460 case SYSTEM_VALUE_SUBGROUP_INVOCATION:
5461 return TGSI_SEMANTIC_SUBGROUP_INVOCATION;
5462 case SYSTEM_VALUE_SUBGROUP_EQ_MASK:
5463 return TGSI_SEMANTIC_SUBGROUP_EQ_MASK;
5464 case SYSTEM_VALUE_SUBGROUP_GE_MASK:
5465 return TGSI_SEMANTIC_SUBGROUP_GE_MASK;
5466 case SYSTEM_VALUE_SUBGROUP_GT_MASK:
5467 return TGSI_SEMANTIC_SUBGROUP_GT_MASK;
5468 case SYSTEM_VALUE_SUBGROUP_LE_MASK:
5469 return TGSI_SEMANTIC_SUBGROUP_LE_MASK;
5470 case SYSTEM_VALUE_SUBGROUP_LT_MASK:
5471 return TGSI_SEMANTIC_SUBGROUP_LT_MASK;
5472
5473 /* Unhandled */
5474 case SYSTEM_VALUE_LOCAL_INVOCATION_INDEX:
5475 case SYSTEM_VALUE_GLOBAL_INVOCATION_ID:
5476 case SYSTEM_VALUE_VERTEX_CNT:
5477 default:
5478 assert(!"Unexpected SYSTEM_VALUE_ enum");
5479 return TGSI_SEMANTIC_COUNT;
5480 }
5481 }
5482
5483 /**
5484 * Map a glsl_to_tgsi constant/immediate to a TGSI immediate.
5485 */
5486 static struct ureg_src
emit_immediate(struct st_translate * t,gl_constant_value values[4],int type,int size)5487 emit_immediate(struct st_translate *t,
5488 gl_constant_value values[4],
5489 int type, int size)
5490 {
5491 struct ureg_program *ureg = t->ureg;
5492
5493 switch(type)
5494 {
5495 case GL_FLOAT:
5496 return ureg_DECL_immediate(ureg, &values[0].f, size);
5497 case GL_DOUBLE:
5498 return ureg_DECL_immediate_f64(ureg, (double *)&values[0].f, size);
5499 case GL_INT64_ARB:
5500 return ureg_DECL_immediate_int64(ureg, (int64_t *)&values[0].f, size);
5501 case GL_UNSIGNED_INT64_ARB:
5502 return ureg_DECL_immediate_uint64(ureg, (uint64_t *)&values[0].f, size);
5503 case GL_INT:
5504 return ureg_DECL_immediate_int(ureg, &values[0].i, size);
5505 case GL_UNSIGNED_INT:
5506 case GL_BOOL:
5507 return ureg_DECL_immediate_uint(ureg, &values[0].u, size);
5508 default:
5509 assert(!"should not get here - type must be float, int, uint, or bool");
5510 return ureg_src_undef();
5511 }
5512 }
5513
5514 /**
5515 * Map a glsl_to_tgsi dst register to a TGSI ureg_dst register.
5516 */
5517 static struct ureg_dst
dst_register(struct st_translate * t,gl_register_file file,unsigned index,unsigned array_id)5518 dst_register(struct st_translate *t, gl_register_file file, unsigned index,
5519 unsigned array_id)
5520 {
5521 unsigned array;
5522
5523 switch(file) {
5524 case PROGRAM_UNDEFINED:
5525 return ureg_dst_undef();
5526
5527 case PROGRAM_TEMPORARY:
5528 /* Allocate space for temporaries on demand. */
5529 if (index >= t->temps_size) {
5530 const int inc = align(index - t->temps_size + 1, 4096);
5531
5532 t->temps = (struct ureg_dst*)
5533 realloc(t->temps,
5534 (t->temps_size + inc) * sizeof(struct ureg_dst));
5535 if (!t->temps)
5536 return ureg_dst_undef();
5537
5538 memset(t->temps + t->temps_size, 0, inc * sizeof(struct ureg_dst));
5539 t->temps_size += inc;
5540 }
5541
5542 if (ureg_dst_is_undef(t->temps[index]))
5543 t->temps[index] = ureg_DECL_local_temporary(t->ureg);
5544
5545 return t->temps[index];
5546
5547 case PROGRAM_ARRAY:
5548 assert(array_id && array_id <= t->num_temp_arrays);
5549 array = array_id - 1;
5550
5551 if (ureg_dst_is_undef(t->arrays[array]))
5552 t->arrays[array] = ureg_DECL_array_temporary(
5553 t->ureg, t->array_sizes[array], TRUE);
5554
5555 return ureg_dst_array_offset(t->arrays[array], index);
5556
5557 case PROGRAM_OUTPUT:
5558 if (!array_id) {
5559 if (t->procType == PIPE_SHADER_FRAGMENT)
5560 assert(index < 2 * FRAG_RESULT_MAX);
5561 else if (t->procType == PIPE_SHADER_TESS_CTRL ||
5562 t->procType == PIPE_SHADER_TESS_EVAL)
5563 assert(index < VARYING_SLOT_TESS_MAX);
5564 else
5565 assert(index < VARYING_SLOT_MAX);
5566
5567 assert(t->outputMapping[index] < ARRAY_SIZE(t->outputs));
5568 assert(t->outputs[t->outputMapping[index]].File != TGSI_FILE_NULL);
5569 return t->outputs[t->outputMapping[index]];
5570 }
5571 else {
5572 struct inout_decl *decl = find_inout_array(t->output_decls, t->num_output_decls, array_id);
5573 unsigned mesa_index = decl->mesa_index;
5574 int slot = t->outputMapping[mesa_index];
5575
5576 assert(slot != -1 && t->outputs[slot].File == TGSI_FILE_OUTPUT);
5577
5578 struct ureg_dst dst = t->outputs[slot];
5579 dst.ArrayID = array_id;
5580 return ureg_dst_array_offset(dst, index - mesa_index);
5581 }
5582
5583 case PROGRAM_ADDRESS:
5584 return t->address[index];
5585
5586 default:
5587 assert(!"unknown dst register file");
5588 return ureg_dst_undef();
5589 }
5590 }
5591
5592 static struct ureg_src
5593 translate_src(struct st_translate *t, const st_src_reg *src_reg);
5594
5595 static struct ureg_src
translate_addr(struct st_translate * t,const st_src_reg * reladdr,unsigned addr_index)5596 translate_addr(struct st_translate *t, const st_src_reg *reladdr,
5597 unsigned addr_index)
5598 {
5599 if (t->need_uarl || !reladdr->is_legal_tgsi_address_operand())
5600 return ureg_src(t->address[addr_index]);
5601
5602 return translate_src(t, reladdr);
5603 }
5604
5605 /**
5606 * Create a TGSI ureg_dst register from an st_dst_reg.
5607 */
5608 static struct ureg_dst
translate_dst(struct st_translate * t,const st_dst_reg * dst_reg,bool saturate)5609 translate_dst(struct st_translate *t,
5610 const st_dst_reg *dst_reg,
5611 bool saturate)
5612 {
5613 struct ureg_dst dst = dst_register(t, dst_reg->file, dst_reg->index,
5614 dst_reg->array_id);
5615
5616 if (dst.File == TGSI_FILE_NULL)
5617 return dst;
5618
5619 dst = ureg_writemask(dst, dst_reg->writemask);
5620
5621 if (saturate)
5622 dst = ureg_saturate(dst);
5623
5624 if (dst_reg->reladdr != NULL) {
5625 assert(dst_reg->file != PROGRAM_TEMPORARY);
5626 dst = ureg_dst_indirect(dst, translate_addr(t, dst_reg->reladdr, 0));
5627 }
5628
5629 if (dst_reg->has_index2) {
5630 if (dst_reg->reladdr2)
5631 dst = ureg_dst_dimension_indirect(dst,
5632 translate_addr(t, dst_reg->reladdr2, 1),
5633 dst_reg->index2D);
5634 else
5635 dst = ureg_dst_dimension(dst, dst_reg->index2D);
5636 }
5637
5638 return dst;
5639 }
5640
5641 /**
5642 * Create a TGSI ureg_src register from an st_src_reg.
5643 */
5644 static struct ureg_src
translate_src(struct st_translate * t,const st_src_reg * src_reg)5645 translate_src(struct st_translate *t, const st_src_reg *src_reg)
5646 {
5647 struct ureg_src src;
5648 int index = src_reg->index;
5649 int double_reg2 = src_reg->double_reg2 ? 1 : 0;
5650
5651 switch(src_reg->file) {
5652 case PROGRAM_UNDEFINED:
5653 src = ureg_imm4f(t->ureg, 0, 0, 0, 0);
5654 break;
5655
5656 case PROGRAM_TEMPORARY:
5657 case PROGRAM_ARRAY:
5658 src = ureg_src(dst_register(t, src_reg->file, src_reg->index, src_reg->array_id));
5659 break;
5660
5661 case PROGRAM_OUTPUT: {
5662 struct ureg_dst dst = dst_register(t, src_reg->file, src_reg->index, src_reg->array_id);
5663 assert(dst.WriteMask != 0);
5664 unsigned shift = ffs(dst.WriteMask) - 1;
5665 src = ureg_swizzle(ureg_src(dst),
5666 shift,
5667 MIN2(shift + 1, 3),
5668 MIN2(shift + 2, 3),
5669 MIN2(shift + 3, 3));
5670 break;
5671 }
5672
5673 case PROGRAM_UNIFORM:
5674 assert(src_reg->index >= 0);
5675 src = src_reg->index < t->num_constants ?
5676 t->constants[src_reg->index] : ureg_imm4f(t->ureg, 0, 0, 0, 0);
5677 break;
5678 case PROGRAM_STATE_VAR:
5679 case PROGRAM_CONSTANT: /* ie, immediate */
5680 if (src_reg->has_index2)
5681 src = ureg_src_register(TGSI_FILE_CONSTANT, src_reg->index);
5682 else
5683 src = src_reg->index >= 0 && src_reg->index < t->num_constants ?
5684 t->constants[src_reg->index] : ureg_imm4f(t->ureg, 0, 0, 0, 0);
5685 break;
5686
5687 case PROGRAM_IMMEDIATE:
5688 assert(src_reg->index >= 0 && src_reg->index < t->num_immediates);
5689 src = t->immediates[src_reg->index];
5690 break;
5691
5692 case PROGRAM_INPUT:
5693 /* GLSL inputs are 64-bit containers, so we have to
5694 * map back to the original index and add the offset after
5695 * mapping. */
5696 index -= double_reg2;
5697 if (!src_reg->array_id) {
5698 assert(t->inputMapping[index] < ARRAY_SIZE(t->inputs));
5699 assert(t->inputs[t->inputMapping[index]].File != TGSI_FILE_NULL);
5700 src = t->inputs[t->inputMapping[index] + double_reg2];
5701 }
5702 else {
5703 struct inout_decl *decl = find_inout_array(t->input_decls, t->num_input_decls,
5704 src_reg->array_id);
5705 unsigned mesa_index = decl->mesa_index;
5706 int slot = t->inputMapping[mesa_index];
5707
5708 assert(slot != -1 && t->inputs[slot].File == TGSI_FILE_INPUT);
5709
5710 src = t->inputs[slot];
5711 src.ArrayID = src_reg->array_id;
5712 src = ureg_src_array_offset(src, index + double_reg2 - mesa_index);
5713 }
5714 break;
5715
5716 case PROGRAM_ADDRESS:
5717 src = ureg_src(t->address[src_reg->index]);
5718 break;
5719
5720 case PROGRAM_SYSTEM_VALUE:
5721 assert(src_reg->index < (int) ARRAY_SIZE(t->systemValues));
5722 src = t->systemValues[src_reg->index];
5723 break;
5724
5725 case PROGRAM_HW_ATOMIC:
5726 src = ureg_src_array_register(TGSI_FILE_HW_ATOMIC, src_reg->index,
5727 src_reg->array_id);
5728 break;
5729
5730 default:
5731 assert(!"unknown src register file");
5732 return ureg_src_undef();
5733 }
5734
5735 if (src_reg->has_index2) {
5736 /* 2D indexes occur with geometry shader inputs (attrib, vertex)
5737 * and UBO constant buffers (buffer, position).
5738 */
5739 if (src_reg->reladdr2)
5740 src = ureg_src_dimension_indirect(src,
5741 translate_addr(t, src_reg->reladdr2, 1),
5742 src_reg->index2D);
5743 else
5744 src = ureg_src_dimension(src, src_reg->index2D);
5745 }
5746
5747 src = ureg_swizzle(src,
5748 GET_SWZ(src_reg->swizzle, 0) & 0x3,
5749 GET_SWZ(src_reg->swizzle, 1) & 0x3,
5750 GET_SWZ(src_reg->swizzle, 2) & 0x3,
5751 GET_SWZ(src_reg->swizzle, 3) & 0x3);
5752
5753 if (src_reg->abs)
5754 src = ureg_abs(src);
5755
5756 if ((src_reg->negate & 0xf) == NEGATE_XYZW)
5757 src = ureg_negate(src);
5758
5759 if (src_reg->reladdr != NULL) {
5760 assert(src_reg->file != PROGRAM_TEMPORARY);
5761 src = ureg_src_indirect(src, translate_addr(t, src_reg->reladdr, 0));
5762 }
5763
5764 return src;
5765 }
5766
5767 static struct tgsi_texture_offset
translate_tex_offset(struct st_translate * t,const st_src_reg * in_offset)5768 translate_tex_offset(struct st_translate *t,
5769 const st_src_reg *in_offset)
5770 {
5771 struct tgsi_texture_offset offset;
5772 struct ureg_src src = translate_src(t, in_offset);
5773
5774 offset.File = src.File;
5775 offset.Index = src.Index;
5776 offset.SwizzleX = src.SwizzleX;
5777 offset.SwizzleY = src.SwizzleY;
5778 offset.SwizzleZ = src.SwizzleZ;
5779 offset.Padding = 0;
5780
5781 assert(!src.Indirect);
5782 assert(!src.DimIndirect);
5783 assert(!src.Dimension);
5784 assert(!src.Absolute); /* those shouldn't be used with integers anyway */
5785 assert(!src.Negate);
5786
5787 return offset;
5788 }
5789
5790 static void
compile_tgsi_instruction(struct st_translate * t,const glsl_to_tgsi_instruction * inst)5791 compile_tgsi_instruction(struct st_translate *t,
5792 const glsl_to_tgsi_instruction *inst)
5793 {
5794 struct ureg_program *ureg = t->ureg;
5795 int i;
5796 struct ureg_dst dst[2];
5797 struct ureg_src src[4];
5798 struct tgsi_texture_offset texoffsets[MAX_GLSL_TEXTURE_OFFSET];
5799
5800 int num_dst;
5801 int num_src;
5802 unsigned tex_target = 0;
5803
5804 num_dst = num_inst_dst_regs(inst);
5805 num_src = num_inst_src_regs(inst);
5806
5807 for (i = 0; i < num_dst; i++)
5808 dst[i] = translate_dst(t,
5809 &inst->dst[i],
5810 inst->saturate);
5811
5812 for (i = 0; i < num_src; i++)
5813 src[i] = translate_src(t, &inst->src[i]);
5814
5815 switch(inst->op) {
5816 case TGSI_OPCODE_BGNLOOP:
5817 case TGSI_OPCODE_ELSE:
5818 case TGSI_OPCODE_ENDLOOP:
5819 case TGSI_OPCODE_IF:
5820 case TGSI_OPCODE_UIF:
5821 assert(num_dst == 0);
5822 ureg_insn(ureg, inst->op, NULL, 0, src, num_src, inst->precise);
5823 return;
5824
5825 case TGSI_OPCODE_TEX:
5826 case TGSI_OPCODE_TEX_LZ:
5827 case TGSI_OPCODE_TXB:
5828 case TGSI_OPCODE_TXD:
5829 case TGSI_OPCODE_TXL:
5830 case TGSI_OPCODE_TXP:
5831 case TGSI_OPCODE_TXQ:
5832 case TGSI_OPCODE_TXQS:
5833 case TGSI_OPCODE_TXF:
5834 case TGSI_OPCODE_TXF_LZ:
5835 case TGSI_OPCODE_TEX2:
5836 case TGSI_OPCODE_TXB2:
5837 case TGSI_OPCODE_TXL2:
5838 case TGSI_OPCODE_TG4:
5839 case TGSI_OPCODE_LODQ:
5840 if (inst->resource.file == PROGRAM_SAMPLER) {
5841 src[num_src] = t->samplers[inst->resource.index];
5842 } else {
5843 /* Bindless samplers. */
5844 src[num_src] = translate_src(t, &inst->resource);
5845 }
5846 assert(src[num_src].File != TGSI_FILE_NULL);
5847 if (inst->resource.reladdr)
5848 src[num_src] =
5849 ureg_src_indirect(src[num_src],
5850 translate_addr(t, inst->resource.reladdr, 2));
5851 num_src++;
5852 for (i = 0; i < (int)inst->tex_offset_num_offset; i++) {
5853 texoffsets[i] = translate_tex_offset(t, &inst->tex_offsets[i]);
5854 }
5855 tex_target = st_translate_texture_target(inst->tex_target, inst->tex_shadow);
5856
5857 ureg_tex_insn(ureg,
5858 inst->op,
5859 dst, num_dst,
5860 tex_target,
5861 st_translate_texture_type(inst->tex_type),
5862 texoffsets, inst->tex_offset_num_offset,
5863 src, num_src);
5864 return;
5865
5866 case TGSI_OPCODE_RESQ:
5867 case TGSI_OPCODE_LOAD:
5868 case TGSI_OPCODE_ATOMUADD:
5869 case TGSI_OPCODE_ATOMXCHG:
5870 case TGSI_OPCODE_ATOMCAS:
5871 case TGSI_OPCODE_ATOMAND:
5872 case TGSI_OPCODE_ATOMOR:
5873 case TGSI_OPCODE_ATOMXOR:
5874 case TGSI_OPCODE_ATOMUMIN:
5875 case TGSI_OPCODE_ATOMUMAX:
5876 case TGSI_OPCODE_ATOMIMIN:
5877 case TGSI_OPCODE_ATOMIMAX:
5878 for (i = num_src - 1; i >= 0; i--)
5879 src[i + 1] = src[i];
5880 num_src++;
5881 if (inst->resource.file == PROGRAM_MEMORY) {
5882 src[0] = t->shared_memory;
5883 } else if (inst->resource.file == PROGRAM_BUFFER) {
5884 src[0] = t->buffers[inst->resource.index];
5885 } else if (inst->resource.file == PROGRAM_HW_ATOMIC) {
5886 src[0] = translate_src(t, &inst->resource);
5887 } else if (inst->resource.file == PROGRAM_CONSTANT) {
5888 assert(inst->resource.has_index2);
5889 src[0] = ureg_src_register(TGSI_FILE_CONSTBUF, inst->resource.index);
5890 } else {
5891 assert(inst->resource.file != PROGRAM_UNDEFINED);
5892 if (inst->resource.file == PROGRAM_IMAGE) {
5893 src[0] = t->images[inst->resource.index];
5894 } else {
5895 /* Bindless images. */
5896 src[0] = translate_src(t, &inst->resource);
5897 }
5898 tex_target = st_translate_texture_target(inst->tex_target, inst->tex_shadow);
5899 }
5900 if (inst->resource.reladdr)
5901 src[0] = ureg_src_indirect(src[0],
5902 translate_addr(t, inst->resource.reladdr, 2));
5903 assert(src[0].File != TGSI_FILE_NULL);
5904 ureg_memory_insn(ureg, inst->op, dst, num_dst, src, num_src,
5905 inst->buffer_access,
5906 tex_target, inst->image_format);
5907 break;
5908
5909 case TGSI_OPCODE_STORE:
5910 if (inst->resource.file == PROGRAM_MEMORY) {
5911 dst[0] = ureg_dst(t->shared_memory);
5912 } else if (inst->resource.file == PROGRAM_BUFFER) {
5913 dst[0] = ureg_dst(t->buffers[inst->resource.index]);
5914 } else {
5915 if (inst->resource.file == PROGRAM_IMAGE) {
5916 dst[0] = ureg_dst(t->images[inst->resource.index]);
5917 } else {
5918 /* Bindless images. */
5919 dst[0] = ureg_dst(translate_src(t, &inst->resource));
5920 }
5921 tex_target = st_translate_texture_target(inst->tex_target, inst->tex_shadow);
5922 }
5923 dst[0] = ureg_writemask(dst[0], inst->dst[0].writemask);
5924 if (inst->resource.reladdr)
5925 dst[0] = ureg_dst_indirect(dst[0],
5926 translate_addr(t, inst->resource.reladdr, 2));
5927 assert(dst[0].File != TGSI_FILE_NULL);
5928 ureg_memory_insn(ureg, inst->op, dst, num_dst, src, num_src,
5929 inst->buffer_access,
5930 tex_target, inst->image_format);
5931 break;
5932
5933 default:
5934 ureg_insn(ureg,
5935 inst->op,
5936 dst, num_dst,
5937 src, num_src, inst->precise);
5938 break;
5939 }
5940 }
5941
5942 /**
5943 * Emit the TGSI instructions for inverting and adjusting WPOS.
5944 * This code is unavoidable because it also depends on whether
5945 * a FBO is bound (STATE_FB_WPOS_Y_TRANSFORM).
5946 */
5947 static void
emit_wpos_adjustment(struct gl_context * ctx,struct st_translate * t,int wpos_transform_const,boolean invert,GLfloat adjX,GLfloat adjY[2])5948 emit_wpos_adjustment(struct gl_context *ctx,
5949 struct st_translate *t,
5950 int wpos_transform_const,
5951 boolean invert,
5952 GLfloat adjX, GLfloat adjY[2])
5953 {
5954 struct ureg_program *ureg = t->ureg;
5955
5956 assert(wpos_transform_const >= 0);
5957
5958 /* Fragment program uses fragment position input.
5959 * Need to replace instances of INPUT[WPOS] with temp T
5960 * where T = INPUT[WPOS] is inverted by Y.
5961 */
5962 struct ureg_src wpostrans = ureg_DECL_constant(ureg, wpos_transform_const);
5963 struct ureg_dst wpos_temp = ureg_DECL_temporary( ureg );
5964 struct ureg_src *wpos =
5965 ctx->Const.GLSLFragCoordIsSysVal ?
5966 &t->systemValues[SYSTEM_VALUE_FRAG_COORD] :
5967 &t->inputs[t->inputMapping[VARYING_SLOT_POS]];
5968 struct ureg_src wpos_input = *wpos;
5969
5970 /* First, apply the coordinate shift: */
5971 if (adjX || adjY[0] || adjY[1]) {
5972 if (adjY[0] != adjY[1]) {
5973 /* Adjust the y coordinate by adjY[1] or adjY[0] respectively
5974 * depending on whether inversion is actually going to be applied
5975 * or not, which is determined by testing against the inversion
5976 * state variable used below, which will be either +1 or -1.
5977 */
5978 struct ureg_dst adj_temp = ureg_DECL_local_temporary(ureg);
5979
5980 ureg_CMP(ureg, adj_temp,
5981 ureg_scalar(wpostrans, invert ? 2 : 0),
5982 ureg_imm4f(ureg, adjX, adjY[0], 0.0f, 0.0f),
5983 ureg_imm4f(ureg, adjX, adjY[1], 0.0f, 0.0f));
5984 ureg_ADD(ureg, wpos_temp, wpos_input, ureg_src(adj_temp));
5985 } else {
5986 ureg_ADD(ureg, wpos_temp, wpos_input,
5987 ureg_imm4f(ureg, adjX, adjY[0], 0.0f, 0.0f));
5988 }
5989 wpos_input = ureg_src(wpos_temp);
5990 } else {
5991 /* MOV wpos_temp, input[wpos]
5992 */
5993 ureg_MOV( ureg, wpos_temp, wpos_input );
5994 }
5995
5996 /* Now the conditional y flip: STATE_FB_WPOS_Y_TRANSFORM.xy/zw will be
5997 * inversion/identity, or the other way around if we're drawing to an FBO.
5998 */
5999 if (invert) {
6000 /* MAD wpos_temp.y, wpos_input, wpostrans.xxxx, wpostrans.yyyy
6001 */
6002 ureg_MAD( ureg,
6003 ureg_writemask(wpos_temp, TGSI_WRITEMASK_Y ),
6004 wpos_input,
6005 ureg_scalar(wpostrans, 0),
6006 ureg_scalar(wpostrans, 1));
6007 } else {
6008 /* MAD wpos_temp.y, wpos_input, wpostrans.zzzz, wpostrans.wwww
6009 */
6010 ureg_MAD( ureg,
6011 ureg_writemask(wpos_temp, TGSI_WRITEMASK_Y ),
6012 wpos_input,
6013 ureg_scalar(wpostrans, 2),
6014 ureg_scalar(wpostrans, 3));
6015 }
6016
6017 /* Use wpos_temp as position input from here on:
6018 */
6019 *wpos = ureg_src(wpos_temp);
6020 }
6021
6022
6023 /**
6024 * Emit fragment position/ooordinate code.
6025 */
6026 static void
emit_wpos(struct st_context * st,struct st_translate * t,const struct gl_program * program,struct ureg_program * ureg,int wpos_transform_const)6027 emit_wpos(struct st_context *st,
6028 struct st_translate *t,
6029 const struct gl_program *program,
6030 struct ureg_program *ureg,
6031 int wpos_transform_const)
6032 {
6033 struct pipe_screen *pscreen = st->pipe->screen;
6034 GLfloat adjX = 0.0f;
6035 GLfloat adjY[2] = { 0.0f, 0.0f };
6036 boolean invert = FALSE;
6037
6038 /* Query the pixel center conventions supported by the pipe driver and set
6039 * adjX, adjY to help out if it cannot handle the requested one internally.
6040 *
6041 * The bias of the y-coordinate depends on whether y-inversion takes place
6042 * (adjY[1]) or not (adjY[0]), which is in turn dependent on whether we are
6043 * drawing to an FBO (causes additional inversion), and whether the pipe
6044 * driver origin and the requested origin differ (the latter condition is
6045 * stored in the 'invert' variable).
6046 *
6047 * For height = 100 (i = integer, h = half-integer, l = lower, u = upper):
6048 *
6049 * center shift only:
6050 * i -> h: +0.5
6051 * h -> i: -0.5
6052 *
6053 * inversion only:
6054 * l,i -> u,i: ( 0.0 + 1.0) * -1 + 100 = 99
6055 * l,h -> u,h: ( 0.5 + 0.0) * -1 + 100 = 99.5
6056 * u,i -> l,i: (99.0 + 1.0) * -1 + 100 = 0
6057 * u,h -> l,h: (99.5 + 0.0) * -1 + 100 = 0.5
6058 *
6059 * inversion and center shift:
6060 * l,i -> u,h: ( 0.0 + 0.5) * -1 + 100 = 99.5
6061 * l,h -> u,i: ( 0.5 + 0.5) * -1 + 100 = 99
6062 * u,i -> l,h: (99.0 + 0.5) * -1 + 100 = 0.5
6063 * u,h -> l,i: (99.5 + 0.5) * -1 + 100 = 0
6064 */
6065 if (program->OriginUpperLeft) {
6066 /* Fragment shader wants origin in upper-left */
6067 if (pscreen->get_param(pscreen, PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT)) {
6068 /* the driver supports upper-left origin */
6069 }
6070 else if (pscreen->get_param(pscreen, PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT)) {
6071 /* the driver supports lower-left origin, need to invert Y */
6072 ureg_property(ureg, TGSI_PROPERTY_FS_COORD_ORIGIN,
6073 TGSI_FS_COORD_ORIGIN_LOWER_LEFT);
6074 invert = TRUE;
6075 }
6076 else
6077 assert(0);
6078 }
6079 else {
6080 /* Fragment shader wants origin in lower-left */
6081 if (pscreen->get_param(pscreen, PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT))
6082 /* the driver supports lower-left origin */
6083 ureg_property(ureg, TGSI_PROPERTY_FS_COORD_ORIGIN,
6084 TGSI_FS_COORD_ORIGIN_LOWER_LEFT);
6085 else if (pscreen->get_param(pscreen, PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT))
6086 /* the driver supports upper-left origin, need to invert Y */
6087 invert = TRUE;
6088 else
6089 assert(0);
6090 }
6091
6092 if (program->PixelCenterInteger) {
6093 /* Fragment shader wants pixel center integer */
6094 if (pscreen->get_param(pscreen, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER)) {
6095 /* the driver supports pixel center integer */
6096 adjY[1] = 1.0f;
6097 ureg_property(ureg, TGSI_PROPERTY_FS_COORD_PIXEL_CENTER,
6098 TGSI_FS_COORD_PIXEL_CENTER_INTEGER);
6099 }
6100 else if (pscreen->get_param(pscreen, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER)) {
6101 /* the driver supports pixel center half integer, need to bias X,Y */
6102 adjX = -0.5f;
6103 adjY[0] = -0.5f;
6104 adjY[1] = 0.5f;
6105 }
6106 else
6107 assert(0);
6108 }
6109 else {
6110 /* Fragment shader wants pixel center half integer */
6111 if (pscreen->get_param(pscreen, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER)) {
6112 /* the driver supports pixel center half integer */
6113 }
6114 else if (pscreen->get_param(pscreen, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER)) {
6115 /* the driver supports pixel center integer, need to bias X,Y */
6116 adjX = adjY[0] = adjY[1] = 0.5f;
6117 ureg_property(ureg, TGSI_PROPERTY_FS_COORD_PIXEL_CENTER,
6118 TGSI_FS_COORD_PIXEL_CENTER_INTEGER);
6119 }
6120 else
6121 assert(0);
6122 }
6123
6124 /* we invert after adjustment so that we avoid the MOV to temporary,
6125 * and reuse the adjustment ADD instead */
6126 emit_wpos_adjustment(st->ctx, t, wpos_transform_const, invert, adjX, adjY);
6127 }
6128
6129 /**
6130 * OpenGL's fragment gl_FrontFace input is 1 for front-facing, 0 for back.
6131 * TGSI uses +1 for front, -1 for back.
6132 * This function converts the TGSI value to the GL value. Simply clamping/
6133 * saturating the value to [0,1] does the job.
6134 */
6135 static void
emit_face_var(struct gl_context * ctx,struct st_translate * t)6136 emit_face_var(struct gl_context *ctx, struct st_translate *t)
6137 {
6138 struct ureg_program *ureg = t->ureg;
6139 struct ureg_dst face_temp = ureg_DECL_temporary(ureg);
6140 struct ureg_src face_input = t->inputs[t->inputMapping[VARYING_SLOT_FACE]];
6141
6142 if (ctx->Const.NativeIntegers) {
6143 ureg_FSGE(ureg, face_temp, face_input, ureg_imm1f(ureg, 0));
6144 }
6145 else {
6146 /* MOV_SAT face_temp, input[face] */
6147 ureg_MOV(ureg, ureg_saturate(face_temp), face_input);
6148 }
6149
6150 /* Use face_temp as face input from here on: */
6151 t->inputs[t->inputMapping[VARYING_SLOT_FACE]] = ureg_src(face_temp);
6152 }
6153
6154 static void
emit_compute_block_size(const struct gl_program * prog,struct ureg_program * ureg)6155 emit_compute_block_size(const struct gl_program *prog,
6156 struct ureg_program *ureg) {
6157 ureg_property(ureg, TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH,
6158 prog->info.cs.local_size[0]);
6159 ureg_property(ureg, TGSI_PROPERTY_CS_FIXED_BLOCK_HEIGHT,
6160 prog->info.cs.local_size[1]);
6161 ureg_property(ureg, TGSI_PROPERTY_CS_FIXED_BLOCK_DEPTH,
6162 prog->info.cs.local_size[2]);
6163 }
6164
6165 struct sort_inout_decls {
operator ()sort_inout_decls6166 bool operator()(const struct inout_decl &a, const struct inout_decl &b) const {
6167 return mapping[a.mesa_index] < mapping[b.mesa_index];
6168 }
6169
6170 const ubyte *mapping;
6171 };
6172
6173 /* Sort the given array of decls by the corresponding slot (TGSI file index).
6174 *
6175 * This is for the benefit of older drivers which are broken when the
6176 * declarations aren't sorted in this way.
6177 */
6178 static void
sort_inout_decls_by_slot(struct inout_decl * decls,unsigned count,const ubyte mapping[])6179 sort_inout_decls_by_slot(struct inout_decl *decls,
6180 unsigned count,
6181 const ubyte mapping[])
6182 {
6183 sort_inout_decls sorter;
6184 sorter.mapping = mapping;
6185 std::sort(decls, decls + count, sorter);
6186 }
6187
6188 static unsigned
st_translate_interp(enum glsl_interp_mode glsl_qual,GLuint varying)6189 st_translate_interp(enum glsl_interp_mode glsl_qual, GLuint varying)
6190 {
6191 switch (glsl_qual) {
6192 case INTERP_MODE_NONE:
6193 if (varying == VARYING_SLOT_COL0 || varying == VARYING_SLOT_COL1)
6194 return TGSI_INTERPOLATE_COLOR;
6195 return TGSI_INTERPOLATE_PERSPECTIVE;
6196 case INTERP_MODE_SMOOTH:
6197 return TGSI_INTERPOLATE_PERSPECTIVE;
6198 case INTERP_MODE_FLAT:
6199 return TGSI_INTERPOLATE_CONSTANT;
6200 case INTERP_MODE_NOPERSPECTIVE:
6201 return TGSI_INTERPOLATE_LINEAR;
6202 default:
6203 assert(0 && "unexpected interp mode in st_translate_interp()");
6204 return TGSI_INTERPOLATE_PERSPECTIVE;
6205 }
6206 }
6207
6208 /**
6209 * Translate intermediate IR (glsl_to_tgsi_instruction) to TGSI format.
6210 * \param program the program to translate
6211 * \param numInputs number of input registers used
6212 * \param inputMapping maps Mesa fragment program inputs to TGSI generic
6213 * input indexes
6214 * \param inputSemanticName the TGSI_SEMANTIC flag for each input
6215 * \param inputSemanticIndex the semantic index (ex: which texcoord) for
6216 * each input
6217 * \param interpMode the TGSI_INTERPOLATE_LINEAR/PERSP mode for each input
6218 * \param numOutputs number of output registers used
6219 * \param outputMapping maps Mesa fragment program outputs to TGSI
6220 * generic outputs
6221 * \param outputSemanticName the TGSI_SEMANTIC flag for each output
6222 * \param outputSemanticIndex the semantic index (ex: which texcoord) for
6223 * each output
6224 *
6225 * \return PIPE_OK or PIPE_ERROR_OUT_OF_MEMORY
6226 */
6227 extern "C" enum pipe_error
st_translate_program(struct gl_context * ctx,uint procType,struct ureg_program * ureg,glsl_to_tgsi_visitor * program,const struct gl_program * proginfo,GLuint numInputs,const ubyte inputMapping[],const ubyte inputSlotToAttr[],const ubyte inputSemanticName[],const ubyte inputSemanticIndex[],const ubyte interpMode[],GLuint numOutputs,const ubyte outputMapping[],const ubyte outputSemanticName[],const ubyte outputSemanticIndex[])6228 st_translate_program(
6229 struct gl_context *ctx,
6230 uint procType,
6231 struct ureg_program *ureg,
6232 glsl_to_tgsi_visitor *program,
6233 const struct gl_program *proginfo,
6234 GLuint numInputs,
6235 const ubyte inputMapping[],
6236 const ubyte inputSlotToAttr[],
6237 const ubyte inputSemanticName[],
6238 const ubyte inputSemanticIndex[],
6239 const ubyte interpMode[],
6240 GLuint numOutputs,
6241 const ubyte outputMapping[],
6242 const ubyte outputSemanticName[],
6243 const ubyte outputSemanticIndex[])
6244 {
6245 struct pipe_screen *screen = st_context(ctx)->pipe->screen;
6246 struct st_translate *t;
6247 unsigned i;
6248 struct gl_program_constants *frag_const =
6249 &ctx->Const.Program[MESA_SHADER_FRAGMENT];
6250 enum pipe_error ret = PIPE_OK;
6251
6252 assert(numInputs <= ARRAY_SIZE(t->inputs));
6253 assert(numOutputs <= ARRAY_SIZE(t->outputs));
6254
6255 ASSERT_BITFIELD_SIZE(st_src_reg, type, GLSL_TYPE_ERROR);
6256 ASSERT_BITFIELD_SIZE(st_dst_reg, type, GLSL_TYPE_ERROR);
6257 ASSERT_BITFIELD_SIZE(glsl_to_tgsi_instruction, tex_type, GLSL_TYPE_ERROR);
6258 ASSERT_BITFIELD_SIZE(glsl_to_tgsi_instruction, image_format, PIPE_FORMAT_COUNT);
6259 ASSERT_BITFIELD_SIZE(glsl_to_tgsi_instruction, tex_target,
6260 (gl_texture_index) (NUM_TEXTURE_TARGETS - 1));
6261 ASSERT_BITFIELD_SIZE(glsl_to_tgsi_instruction, image_format,
6262 (enum pipe_format) (PIPE_FORMAT_COUNT - 1));
6263 ASSERT_BITFIELD_SIZE(glsl_to_tgsi_instruction, op, TGSI_OPCODE_LAST - 1);
6264
6265 t = CALLOC_STRUCT(st_translate);
6266 if (!t) {
6267 ret = PIPE_ERROR_OUT_OF_MEMORY;
6268 goto out;
6269 }
6270
6271 t->procType = procType;
6272 t->need_uarl = !screen->get_param(screen, PIPE_CAP_TGSI_ANY_REG_AS_ADDRESS);
6273 t->inputMapping = inputMapping;
6274 t->outputMapping = outputMapping;
6275 t->ureg = ureg;
6276 t->num_temp_arrays = program->next_array;
6277 if (t->num_temp_arrays)
6278 t->arrays = (struct ureg_dst*)
6279 calloc(t->num_temp_arrays, sizeof(t->arrays[0]));
6280
6281 /*
6282 * Declare input attributes.
6283 */
6284 switch (procType) {
6285 case PIPE_SHADER_FRAGMENT:
6286 case PIPE_SHADER_GEOMETRY:
6287 case PIPE_SHADER_TESS_EVAL:
6288 case PIPE_SHADER_TESS_CTRL:
6289 sort_inout_decls_by_slot(program->inputs, program->num_inputs, inputMapping);
6290
6291 for (i = 0; i < program->num_inputs; ++i) {
6292 struct inout_decl *decl = &program->inputs[i];
6293 unsigned slot = inputMapping[decl->mesa_index];
6294 struct ureg_src src;
6295 ubyte tgsi_usage_mask = decl->usage_mask;
6296
6297 if (glsl_base_type_is_64bit(decl->base_type)) {
6298 if (tgsi_usage_mask == 1)
6299 tgsi_usage_mask = TGSI_WRITEMASK_XY;
6300 else if (tgsi_usage_mask == 2)
6301 tgsi_usage_mask = TGSI_WRITEMASK_ZW;
6302 else
6303 tgsi_usage_mask = TGSI_WRITEMASK_XYZW;
6304 }
6305
6306 unsigned interp_mode = 0;
6307 unsigned interp_location = 0;
6308 if (procType == PIPE_SHADER_FRAGMENT) {
6309 assert(interpMode);
6310 interp_mode = interpMode[slot] != TGSI_INTERPOLATE_COUNT ?
6311 interpMode[slot] :
6312 st_translate_interp(decl->interp, inputSlotToAttr[slot]);
6313
6314 interp_location = decl->interp_loc;
6315 }
6316
6317 src = ureg_DECL_fs_input_cyl_centroid_layout(ureg,
6318 inputSemanticName[slot], inputSemanticIndex[slot],
6319 interp_mode, 0, interp_location, slot, tgsi_usage_mask,
6320 decl->array_id, decl->size);
6321
6322 for (unsigned j = 0; j < decl->size; ++j) {
6323 if (t->inputs[slot + j].File != TGSI_FILE_INPUT) {
6324 /* The ArrayID is set up in dst_register */
6325 t->inputs[slot + j] = src;
6326 t->inputs[slot + j].ArrayID = 0;
6327 t->inputs[slot + j].Index += j;
6328 }
6329 }
6330 }
6331 break;
6332 case PIPE_SHADER_VERTEX:
6333 for (i = 0; i < numInputs; i++) {
6334 t->inputs[i] = ureg_DECL_vs_input(ureg, i);
6335 }
6336 break;
6337 case PIPE_SHADER_COMPUTE:
6338 break;
6339 default:
6340 assert(0);
6341 }
6342
6343 /*
6344 * Declare output attributes.
6345 */
6346 switch (procType) {
6347 case PIPE_SHADER_FRAGMENT:
6348 case PIPE_SHADER_COMPUTE:
6349 break;
6350 case PIPE_SHADER_GEOMETRY:
6351 case PIPE_SHADER_TESS_EVAL:
6352 case PIPE_SHADER_TESS_CTRL:
6353 case PIPE_SHADER_VERTEX:
6354 sort_inout_decls_by_slot(program->outputs, program->num_outputs, outputMapping);
6355
6356 for (i = 0; i < program->num_outputs; ++i) {
6357 struct inout_decl *decl = &program->outputs[i];
6358 unsigned slot = outputMapping[decl->mesa_index];
6359 struct ureg_dst dst;
6360 ubyte tgsi_usage_mask = decl->usage_mask;
6361
6362 if (glsl_base_type_is_64bit(decl->base_type)) {
6363 if (tgsi_usage_mask == 1)
6364 tgsi_usage_mask = TGSI_WRITEMASK_XY;
6365 else if (tgsi_usage_mask == 2)
6366 tgsi_usage_mask = TGSI_WRITEMASK_ZW;
6367 else
6368 tgsi_usage_mask = TGSI_WRITEMASK_XYZW;
6369 }
6370
6371 dst = ureg_DECL_output_layout(ureg,
6372 outputSemanticName[slot], outputSemanticIndex[slot],
6373 decl->gs_out_streams,
6374 slot, tgsi_usage_mask, decl->array_id, decl->size);
6375
6376 for (unsigned j = 0; j < decl->size; ++j) {
6377 if (t->outputs[slot + j].File != TGSI_FILE_OUTPUT) {
6378 /* The ArrayID is set up in dst_register */
6379 t->outputs[slot + j] = dst;
6380 t->outputs[slot + j].ArrayID = 0;
6381 t->outputs[slot + j].Index += j;
6382 }
6383 }
6384 }
6385 break;
6386 default:
6387 assert(0);
6388 }
6389
6390 if (procType == PIPE_SHADER_FRAGMENT) {
6391 if (program->shader->Program->info.fs.early_fragment_tests ||
6392 program->shader->Program->info.fs.post_depth_coverage) {
6393 ureg_property(ureg, TGSI_PROPERTY_FS_EARLY_DEPTH_STENCIL, 1);
6394
6395 if (program->shader->Program->info.fs.post_depth_coverage)
6396 ureg_property(ureg, TGSI_PROPERTY_FS_POST_DEPTH_COVERAGE, 1);
6397 }
6398
6399 if (proginfo->info.inputs_read & VARYING_BIT_POS) {
6400 /* Must do this after setting up t->inputs. */
6401 emit_wpos(st_context(ctx), t, proginfo, ureg,
6402 program->wpos_transform_const);
6403 }
6404
6405 if (proginfo->info.inputs_read & VARYING_BIT_FACE)
6406 emit_face_var(ctx, t);
6407
6408 for (i = 0; i < numOutputs; i++) {
6409 switch (outputSemanticName[i]) {
6410 case TGSI_SEMANTIC_POSITION:
6411 t->outputs[i] = ureg_DECL_output(ureg,
6412 TGSI_SEMANTIC_POSITION, /* Z/Depth */
6413 outputSemanticIndex[i]);
6414 t->outputs[i] = ureg_writemask(t->outputs[i], TGSI_WRITEMASK_Z);
6415 break;
6416 case TGSI_SEMANTIC_STENCIL:
6417 t->outputs[i] = ureg_DECL_output(ureg,
6418 TGSI_SEMANTIC_STENCIL, /* Stencil */
6419 outputSemanticIndex[i]);
6420 t->outputs[i] = ureg_writemask(t->outputs[i], TGSI_WRITEMASK_Y);
6421 break;
6422 case TGSI_SEMANTIC_COLOR:
6423 t->outputs[i] = ureg_DECL_output(ureg,
6424 TGSI_SEMANTIC_COLOR,
6425 outputSemanticIndex[i]);
6426 break;
6427 case TGSI_SEMANTIC_SAMPLEMASK:
6428 t->outputs[i] = ureg_DECL_output(ureg,
6429 TGSI_SEMANTIC_SAMPLEMASK,
6430 outputSemanticIndex[i]);
6431 /* TODO: If we ever support more than 32 samples, this will have
6432 * to become an array.
6433 */
6434 t->outputs[i] = ureg_writemask(t->outputs[i], TGSI_WRITEMASK_X);
6435 break;
6436 default:
6437 assert(!"fragment shader outputs must be POSITION/STENCIL/COLOR");
6438 ret = PIPE_ERROR_BAD_INPUT;
6439 goto out;
6440 }
6441 }
6442 }
6443 else if (procType == PIPE_SHADER_VERTEX) {
6444 for (i = 0; i < numOutputs; i++) {
6445 if (outputSemanticName[i] == TGSI_SEMANTIC_FOG) {
6446 /* force register to contain a fog coordinate in the form (F, 0, 0, 1). */
6447 ureg_MOV(ureg,
6448 ureg_writemask(t->outputs[i], TGSI_WRITEMASK_YZW),
6449 ureg_imm4f(ureg, 0.0f, 0.0f, 0.0f, 1.0f));
6450 t->outputs[i] = ureg_writemask(t->outputs[i], TGSI_WRITEMASK_X);
6451 }
6452 }
6453 }
6454
6455 if (procType == PIPE_SHADER_COMPUTE) {
6456 emit_compute_block_size(proginfo, ureg);
6457 }
6458
6459 /* Declare address register.
6460 */
6461 if (program->num_address_regs > 0) {
6462 assert(program->num_address_regs <= 3);
6463 for (int i = 0; i < program->num_address_regs; i++)
6464 t->address[i] = ureg_DECL_address(ureg);
6465 }
6466
6467 /* Declare misc input registers
6468 */
6469 {
6470 GLbitfield sysInputs = proginfo->info.system_values_read;
6471
6472 for (i = 0; sysInputs; i++) {
6473 if (sysInputs & (1 << i)) {
6474 unsigned semName = _mesa_sysval_to_semantic(i);
6475
6476 t->systemValues[i] = ureg_DECL_system_value(ureg, semName, 0);
6477
6478 if (semName == TGSI_SEMANTIC_INSTANCEID ||
6479 semName == TGSI_SEMANTIC_VERTEXID) {
6480 /* From Gallium perspective, these system values are always
6481 * integer, and require native integer support. However, if
6482 * native integer is supported on the vertex stage but not the
6483 * pixel stage (e.g, i915g + draw), Mesa will generate IR that
6484 * assumes these system values are floats. To resolve the
6485 * inconsistency, we insert a U2F.
6486 */
6487 struct st_context *st = st_context(ctx);
6488 struct pipe_screen *pscreen = st->pipe->screen;
6489 assert(procType == PIPE_SHADER_VERTEX);
6490 assert(pscreen->get_shader_param(pscreen, PIPE_SHADER_VERTEX, PIPE_SHADER_CAP_INTEGERS));
6491 (void) pscreen;
6492 if (!ctx->Const.NativeIntegers) {
6493 struct ureg_dst temp = ureg_DECL_local_temporary(t->ureg);
6494 ureg_U2F( t->ureg, ureg_writemask(temp, TGSI_WRITEMASK_X), t->systemValues[i]);
6495 t->systemValues[i] = ureg_scalar(ureg_src(temp), 0);
6496 }
6497 }
6498
6499 if (procType == PIPE_SHADER_FRAGMENT &&
6500 semName == TGSI_SEMANTIC_POSITION)
6501 emit_wpos(st_context(ctx), t, proginfo, ureg,
6502 program->wpos_transform_const);
6503
6504 sysInputs &= ~(1 << i);
6505 }
6506 }
6507 }
6508
6509 t->array_sizes = program->array_sizes;
6510 t->input_decls = program->inputs;
6511 t->num_input_decls = program->num_inputs;
6512 t->output_decls = program->outputs;
6513 t->num_output_decls = program->num_outputs;
6514
6515 /* Emit constants and uniforms. TGSI uses a single index space for these,
6516 * so we put all the translated regs in t->constants.
6517 */
6518 if (proginfo->Parameters) {
6519 t->constants = (struct ureg_src *)
6520 calloc(proginfo->Parameters->NumParameters, sizeof(t->constants[0]));
6521 if (t->constants == NULL) {
6522 ret = PIPE_ERROR_OUT_OF_MEMORY;
6523 goto out;
6524 }
6525 t->num_constants = proginfo->Parameters->NumParameters;
6526
6527 for (i = 0; i < proginfo->Parameters->NumParameters; i++) {
6528 switch (proginfo->Parameters->Parameters[i].Type) {
6529 case PROGRAM_STATE_VAR:
6530 case PROGRAM_UNIFORM:
6531 t->constants[i] = ureg_DECL_constant(ureg, i);
6532 break;
6533
6534 /* Emit immediates for PROGRAM_CONSTANT only when there's no indirect
6535 * addressing of the const buffer.
6536 * FIXME: Be smarter and recognize param arrays:
6537 * indirect addressing is only valid within the referenced
6538 * array.
6539 */
6540 case PROGRAM_CONSTANT:
6541 if (program->indirect_addr_consts)
6542 t->constants[i] = ureg_DECL_constant(ureg, i);
6543 else
6544 t->constants[i] = emit_immediate(t,
6545 proginfo->Parameters->ParameterValues[i],
6546 proginfo->Parameters->Parameters[i].DataType,
6547 4);
6548 break;
6549 default:
6550 break;
6551 }
6552 }
6553 }
6554
6555 for (i = 0; i < proginfo->info.num_ubos; i++) {
6556 unsigned size = proginfo->sh.UniformBlocks[i]->UniformBufferSize;
6557 unsigned num_const_vecs = (size + 15) / 16;
6558 unsigned first, last;
6559 assert(num_const_vecs > 0);
6560 first = 0;
6561 last = num_const_vecs > 0 ? num_const_vecs - 1 : 0;
6562 ureg_DECL_constant2D(t->ureg, first, last, i + 1);
6563 }
6564
6565 /* Emit immediate values.
6566 */
6567 t->immediates = (struct ureg_src *)
6568 calloc(program->num_immediates, sizeof(struct ureg_src));
6569 if (t->immediates == NULL) {
6570 ret = PIPE_ERROR_OUT_OF_MEMORY;
6571 goto out;
6572 }
6573 t->num_immediates = program->num_immediates;
6574
6575 i = 0;
6576 foreach_in_list(immediate_storage, imm, &program->immediates) {
6577 assert(i < program->num_immediates);
6578 t->immediates[i++] = emit_immediate(t, imm->values, imm->type, imm->size32);
6579 }
6580 assert(i == program->num_immediates);
6581
6582 /* texture samplers */
6583 for (i = 0; i < frag_const->MaxTextureImageUnits; i++) {
6584 if (program->samplers_used & (1u << i)) {
6585 enum tgsi_return_type type =
6586 st_translate_texture_type(program->sampler_types[i]);
6587
6588 t->samplers[i] = ureg_DECL_sampler(ureg, i);
6589
6590 ureg_DECL_sampler_view( ureg, i, program->sampler_targets[i],
6591 type, type, type, type );
6592 }
6593 }
6594
6595 /* Declare atomic and shader storage buffers. */
6596 {
6597 struct gl_program *prog = program->prog;
6598
6599 if (!st_context(ctx)->has_hw_atomics) {
6600 for (i = 0; i < prog->info.num_abos; i++) {
6601 unsigned index = prog->sh.AtomicBuffers[i]->Binding;
6602 assert(index < frag_const->MaxAtomicBuffers);
6603 t->buffers[index] = ureg_DECL_buffer(ureg, index, true);
6604 }
6605 } else {
6606 for (i = 0; i < program->num_atomics; i++) {
6607 struct hwatomic_decl *ainfo = &program->atomic_info[i];
6608 gl_uniform_storage *uni_storage = &prog->sh.data->UniformStorage[ainfo->location];
6609 int base = uni_storage->offset / ATOMIC_COUNTER_SIZE;
6610 ureg_DECL_hw_atomic(ureg, base, base + ainfo->size - 1, ainfo->binding,
6611 ainfo->array_id);
6612 }
6613 }
6614
6615 assert(prog->info.num_ssbos <= frag_const->MaxShaderStorageBlocks);
6616 for (i = 0; i < prog->info.num_ssbos; i++) {
6617 unsigned index = i;
6618 if (!st_context(ctx)->has_hw_atomics)
6619 index += frag_const->MaxAtomicBuffers;
6620
6621 t->buffers[index] = ureg_DECL_buffer(ureg, index, false);
6622 }
6623 }
6624
6625 if (program->use_shared_memory)
6626 t->shared_memory = ureg_DECL_memory(ureg, TGSI_MEMORY_TYPE_SHARED);
6627
6628 for (i = 0; i < program->shader->Program->info.num_images; i++) {
6629 if (program->images_used & (1 << i)) {
6630 t->images[i] = ureg_DECL_image(ureg, i,
6631 program->image_targets[i],
6632 program->image_formats[i],
6633 true, false);
6634 }
6635 }
6636
6637 /* Emit each instruction in turn:
6638 */
6639 foreach_in_list(glsl_to_tgsi_instruction, inst, &program->instructions)
6640 compile_tgsi_instruction(t, inst);
6641
6642 /* Set the next shader stage hint for VS and TES. */
6643 switch (procType) {
6644 case PIPE_SHADER_VERTEX:
6645 case PIPE_SHADER_TESS_EVAL:
6646 if (program->shader_program->SeparateShader)
6647 break;
6648
6649 for (i = program->shader->Stage+1; i <= MESA_SHADER_FRAGMENT; i++) {
6650 if (program->shader_program->_LinkedShaders[i]) {
6651 ureg_set_next_shader_processor(
6652 ureg, pipe_shader_type_from_mesa((gl_shader_stage)i));
6653 break;
6654 }
6655 }
6656 break;
6657 }
6658
6659 out:
6660 if (t) {
6661 free(t->arrays);
6662 free(t->temps);
6663 free(t->constants);
6664 t->num_constants = 0;
6665 free(t->immediates);
6666 t->num_immediates = 0;
6667 FREE(t);
6668 }
6669
6670 return ret;
6671 }
6672 /* ----------------------------- End TGSI code ------------------------------ */
6673
6674
6675 /**
6676 * Convert a shader's GLSL IR into a Mesa gl_program, although without
6677 * generating Mesa IR.
6678 */
6679 static struct gl_program *
get_mesa_program_tgsi(struct gl_context * ctx,struct gl_shader_program * shader_program,struct gl_linked_shader * shader)6680 get_mesa_program_tgsi(struct gl_context *ctx,
6681 struct gl_shader_program *shader_program,
6682 struct gl_linked_shader *shader)
6683 {
6684 glsl_to_tgsi_visitor* v;
6685 struct gl_program *prog;
6686 struct gl_shader_compiler_options *options =
6687 &ctx->Const.ShaderCompilerOptions[shader->Stage];
6688 struct pipe_screen *pscreen = ctx->st->pipe->screen;
6689 enum pipe_shader_type ptarget = pipe_shader_type_from_mesa(shader->Stage);
6690 unsigned skip_merge_registers;
6691
6692 validate_ir_tree(shader->ir);
6693
6694 prog = shader->Program;
6695
6696 prog->Parameters = _mesa_new_parameter_list();
6697 v = new glsl_to_tgsi_visitor();
6698 v->ctx = ctx;
6699 v->prog = prog;
6700 v->shader_program = shader_program;
6701 v->shader = shader;
6702 v->options = options;
6703 v->native_integers = ctx->Const.NativeIntegers;
6704
6705 v->have_sqrt = pscreen->get_shader_param(pscreen, ptarget,
6706 PIPE_SHADER_CAP_TGSI_SQRT_SUPPORTED);
6707 v->have_fma = pscreen->get_shader_param(pscreen, ptarget,
6708 PIPE_SHADER_CAP_TGSI_FMA_SUPPORTED);
6709 v->has_tex_txf_lz = pscreen->get_param(pscreen,
6710 PIPE_CAP_TGSI_TEX_TXF_LZ);
6711 v->need_uarl = !pscreen->get_param(pscreen, PIPE_CAP_TGSI_ANY_REG_AS_ADDRESS);
6712
6713 v->variables = _mesa_hash_table_create(v->mem_ctx, _mesa_hash_pointer,
6714 _mesa_key_pointer_equal);
6715 skip_merge_registers =
6716 pscreen->get_shader_param(pscreen, ptarget,
6717 PIPE_SHADER_CAP_TGSI_SKIP_MERGE_REGISTERS);
6718
6719 _mesa_generate_parameters_list_for_uniforms(ctx, shader_program, shader,
6720 prog->Parameters);
6721
6722 /* Remove reads from output registers. */
6723 if (!pscreen->get_param(pscreen, PIPE_CAP_TGSI_CAN_READ_OUTPUTS))
6724 lower_output_reads(shader->Stage, shader->ir);
6725
6726 /* Emit intermediate IR for main(). */
6727 visit_exec_list(shader->ir, v);
6728
6729 #if 0
6730 /* Print out some information (for debugging purposes) used by the
6731 * optimization passes. */
6732 {
6733 int i;
6734 int *first_writes = ralloc_array(v->mem_ctx, int, v->next_temp);
6735 int *first_reads = ralloc_array(v->mem_ctx, int, v->next_temp);
6736 int *last_writes = ralloc_array(v->mem_ctx, int, v->next_temp);
6737 int *last_reads = ralloc_array(v->mem_ctx, int, v->next_temp);
6738
6739 for (i = 0; i < v->next_temp; i++) {
6740 first_writes[i] = -1;
6741 first_reads[i] = -1;
6742 last_writes[i] = -1;
6743 last_reads[i] = -1;
6744 }
6745 v->get_first_temp_read(first_reads);
6746 v->get_last_temp_read_first_temp_write(last_reads, first_writes);
6747 v->get_last_temp_write(last_writes);
6748 for (i = 0; i < v->next_temp; i++)
6749 printf("Temp %d: FR=%3d FW=%3d LR=%3d LW=%3d\n", i, first_reads[i],
6750 first_writes[i],
6751 last_reads[i],
6752 last_writes[i]);
6753 ralloc_free(first_writes);
6754 ralloc_free(first_reads);
6755 ralloc_free(last_writes);
6756 ralloc_free(last_reads);
6757 }
6758 #endif
6759
6760 /* Perform optimizations on the instructions in the glsl_to_tgsi_visitor. */
6761 v->simplify_cmp();
6762 v->copy_propagate();
6763
6764 while (v->eliminate_dead_code());
6765
6766 v->merge_two_dsts();
6767 if (!skip_merge_registers)
6768 v->merge_registers();
6769 v->renumber_registers();
6770
6771 /* Write the END instruction. */
6772 v->emit_asm(NULL, TGSI_OPCODE_END);
6773
6774 if (ctx->_Shader->Flags & GLSL_DUMP) {
6775 _mesa_log("\n");
6776 _mesa_log("GLSL IR for linked %s program %d:\n",
6777 _mesa_shader_stage_to_string(shader->Stage),
6778 shader_program->Name);
6779 _mesa_print_ir(_mesa_get_log_file(), shader->ir, NULL);
6780 _mesa_log("\n\n");
6781 }
6782
6783 do_set_program_inouts(shader->ir, prog, shader->Stage);
6784 _mesa_copy_linked_program_data(shader_program, shader);
6785 shrink_array_declarations(v->inputs, v->num_inputs,
6786 &prog->info.inputs_read,
6787 prog->info.double_inputs_read,
6788 &prog->info.patch_inputs_read);
6789 shrink_array_declarations(v->outputs, v->num_outputs,
6790 &prog->info.outputs_written, 0ULL,
6791 &prog->info.patch_outputs_written);
6792 count_resources(v, prog);
6793
6794 /* The GLSL IR won't be needed anymore. */
6795 ralloc_free(shader->ir);
6796 shader->ir = NULL;
6797
6798 /* This must be done before the uniform storage is associated. */
6799 if (shader->Stage == MESA_SHADER_FRAGMENT &&
6800 (prog->info.inputs_read & VARYING_BIT_POS ||
6801 prog->info.system_values_read & (1 << SYSTEM_VALUE_FRAG_COORD))) {
6802 static const gl_state_index wposTransformState[STATE_LENGTH] = {
6803 STATE_INTERNAL, STATE_FB_WPOS_Y_TRANSFORM
6804 };
6805
6806 v->wpos_transform_const = _mesa_add_state_reference(prog->Parameters,
6807 wposTransformState);
6808 }
6809
6810 /* Avoid reallocation of the program parameter list, because the uniform
6811 * storage is only associated with the original parameter list.
6812 * This should be enough for Bitmap and DrawPixels constants.
6813 */
6814 _mesa_reserve_parameter_storage(prog->Parameters, 8);
6815
6816 /* This has to be done last. Any operation the can cause
6817 * prog->ParameterValues to get reallocated (e.g., anything that adds a
6818 * program constant) has to happen before creating this linkage.
6819 */
6820 _mesa_associate_uniform_storage(ctx, shader_program, prog, true);
6821 if (!shader_program->data->LinkStatus) {
6822 free_glsl_to_tgsi_visitor(v);
6823 _mesa_reference_program(ctx, &shader->Program, NULL);
6824 return NULL;
6825 }
6826
6827 struct st_vertex_program *stvp;
6828 struct st_fragment_program *stfp;
6829 struct st_common_program *stp;
6830 struct st_compute_program *stcp;
6831
6832 switch (shader->Stage) {
6833 case MESA_SHADER_VERTEX:
6834 stvp = (struct st_vertex_program *)prog;
6835 stvp->glsl_to_tgsi = v;
6836 break;
6837 case MESA_SHADER_FRAGMENT:
6838 stfp = (struct st_fragment_program *)prog;
6839 stfp->glsl_to_tgsi = v;
6840 break;
6841 case MESA_SHADER_TESS_CTRL:
6842 case MESA_SHADER_TESS_EVAL:
6843 case MESA_SHADER_GEOMETRY:
6844 stp = st_common_program(prog);
6845 stp->glsl_to_tgsi = v;
6846 break;
6847 case MESA_SHADER_COMPUTE:
6848 stcp = (struct st_compute_program *)prog;
6849 stcp->glsl_to_tgsi = v;
6850 break;
6851 default:
6852 assert(!"should not be reached");
6853 return NULL;
6854 }
6855
6856 return prog;
6857 }
6858
6859 /* See if there are unsupported control flow statements. */
6860 class ir_control_flow_info_visitor : public ir_hierarchical_visitor {
6861 private:
6862 const struct gl_shader_compiler_options *options;
6863 public:
ir_control_flow_info_visitor(const struct gl_shader_compiler_options * options)6864 ir_control_flow_info_visitor(const struct gl_shader_compiler_options *options)
6865 : options(options),
6866 unsupported(false)
6867 {
6868 }
6869
visit_enter(ir_function * ir)6870 virtual ir_visitor_status visit_enter(ir_function *ir)
6871 {
6872 /* Other functions are skipped (same as glsl_to_tgsi). */
6873 if (strcmp(ir->name, "main") == 0)
6874 return visit_continue;
6875
6876 return visit_continue_with_parent;
6877 }
6878
visit_enter(ir_call * ir)6879 virtual ir_visitor_status visit_enter(ir_call *ir)
6880 {
6881 if (!ir->callee->is_intrinsic()) {
6882 unsupported = true; /* it's a function call */
6883 return visit_stop;
6884 }
6885 return visit_continue;
6886 }
6887
visit_enter(ir_return * ir)6888 virtual ir_visitor_status visit_enter(ir_return *ir)
6889 {
6890 if (options->EmitNoMainReturn) {
6891 unsupported = true;
6892 return visit_stop;
6893 }
6894 return visit_continue;
6895 }
6896
6897 bool unsupported;
6898 };
6899
6900 static bool
has_unsupported_control_flow(exec_list * ir,const struct gl_shader_compiler_options * options)6901 has_unsupported_control_flow(exec_list *ir,
6902 const struct gl_shader_compiler_options *options)
6903 {
6904 ir_control_flow_info_visitor visitor(options);
6905 visit_list_elements(&visitor, ir);
6906 return visitor.unsupported;
6907 }
6908
6909 extern "C" {
6910
6911 /**
6912 * Link a shader.
6913 * Called via ctx->Driver.LinkShader()
6914 * This actually involves converting GLSL IR into an intermediate TGSI-like IR
6915 * with code lowering and other optimizations.
6916 */
6917 GLboolean
st_link_shader(struct gl_context * ctx,struct gl_shader_program * prog)6918 st_link_shader(struct gl_context *ctx, struct gl_shader_program *prog)
6919 {
6920 /* Return early if we are loading the shader from on-disk cache */
6921 if (st_load_tgsi_from_disk_cache(ctx, prog)) {
6922 return GL_TRUE;
6923 }
6924
6925 struct pipe_screen *pscreen = ctx->st->pipe->screen;
6926 assert(prog->data->LinkStatus);
6927
6928 bool use_nir = false;
6929 for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
6930 if (prog->_LinkedShaders[i] == NULL)
6931 continue;
6932
6933 struct gl_linked_shader *shader = prog->_LinkedShaders[i];
6934 exec_list *ir = shader->ir;
6935 gl_shader_stage stage = shader->Stage;
6936 const struct gl_shader_compiler_options *options =
6937 &ctx->Const.ShaderCompilerOptions[stage];
6938 enum pipe_shader_type ptarget = pipe_shader_type_from_mesa(stage);
6939 bool have_dround = pscreen->get_shader_param(pscreen, ptarget,
6940 PIPE_SHADER_CAP_TGSI_DROUND_SUPPORTED);
6941 bool have_dfrexp = pscreen->get_shader_param(pscreen, ptarget,
6942 PIPE_SHADER_CAP_TGSI_DFRACEXP_DLDEXP_SUPPORTED);
6943 bool have_ldexp = pscreen->get_shader_param(pscreen, ptarget,
6944 PIPE_SHADER_CAP_TGSI_LDEXP_SUPPORTED);
6945 unsigned if_threshold = pscreen->get_shader_param(pscreen, ptarget,
6946 PIPE_SHADER_CAP_LOWER_IF_THRESHOLD);
6947
6948 enum pipe_shader_ir preferred_ir = (enum pipe_shader_ir)
6949 pscreen->get_shader_param(pscreen, ptarget,
6950 PIPE_SHADER_CAP_PREFERRED_IR);
6951 if (preferred_ir == PIPE_SHADER_IR_NIR)
6952 use_nir = true;
6953
6954 /* If there are forms of indirect addressing that the driver
6955 * cannot handle, perform the lowering pass.
6956 */
6957 if (options->EmitNoIndirectInput || options->EmitNoIndirectOutput ||
6958 options->EmitNoIndirectTemp || options->EmitNoIndirectUniform) {
6959 lower_variable_index_to_cond_assign(stage, ir,
6960 options->EmitNoIndirectInput,
6961 options->EmitNoIndirectOutput,
6962 options->EmitNoIndirectTemp,
6963 options->EmitNoIndirectUniform);
6964 }
6965
6966 if (!pscreen->get_param(pscreen, PIPE_CAP_INT64_DIVMOD))
6967 lower_64bit_integer_instructions(ir, DIV64 | MOD64);
6968
6969 if (ctx->Extensions.ARB_shading_language_packing) {
6970 unsigned lower_inst = LOWER_PACK_SNORM_2x16 |
6971 LOWER_UNPACK_SNORM_2x16 |
6972 LOWER_PACK_UNORM_2x16 |
6973 LOWER_UNPACK_UNORM_2x16 |
6974 LOWER_PACK_SNORM_4x8 |
6975 LOWER_UNPACK_SNORM_4x8 |
6976 LOWER_UNPACK_UNORM_4x8 |
6977 LOWER_PACK_UNORM_4x8;
6978
6979 if (ctx->Extensions.ARB_gpu_shader5)
6980 lower_inst |= LOWER_PACK_USE_BFI |
6981 LOWER_PACK_USE_BFE;
6982 if (!ctx->st->has_half_float_packing)
6983 lower_inst |= LOWER_PACK_HALF_2x16 |
6984 LOWER_UNPACK_HALF_2x16;
6985
6986 lower_packing_builtins(ir, lower_inst);
6987 }
6988
6989 if (!pscreen->get_param(pscreen, PIPE_CAP_TEXTURE_GATHER_OFFSETS))
6990 lower_offset_arrays(ir);
6991 do_mat_op_to_vec(ir);
6992
6993 if (stage == MESA_SHADER_FRAGMENT)
6994 lower_blend_equation_advanced(shader);
6995
6996 lower_instructions(ir,
6997 MOD_TO_FLOOR |
6998 FDIV_TO_MUL_RCP |
6999 EXP_TO_EXP2 |
7000 LOG_TO_LOG2 |
7001 (have_ldexp ? 0 : LDEXP_TO_ARITH) |
7002 (have_dfrexp ? 0 : DFREXP_DLDEXP_TO_ARITH) |
7003 CARRY_TO_ARITH |
7004 BORROW_TO_ARITH |
7005 (have_dround ? 0 : DOPS_TO_DFRAC) |
7006 (options->EmitNoPow ? POW_TO_EXP2 : 0) |
7007 (!ctx->Const.NativeIntegers ? INT_DIV_TO_MUL_RCP : 0) |
7008 (options->EmitNoSat ? SAT_TO_CLAMP : 0) |
7009 (ctx->Const.ForceGLSLAbsSqrt ? SQRT_TO_ABS_SQRT : 0) |
7010 /* Assume that if ARB_gpu_shader5 is not supported
7011 * then all of the extended integer functions need
7012 * lowering. It may be necessary to add some caps
7013 * for individual instructions.
7014 */
7015 (!ctx->Extensions.ARB_gpu_shader5
7016 ? BIT_COUNT_TO_MATH |
7017 EXTRACT_TO_SHIFTS |
7018 INSERT_TO_SHIFTS |
7019 REVERSE_TO_SHIFTS |
7020 FIND_LSB_TO_FLOAT_CAST |
7021 FIND_MSB_TO_FLOAT_CAST |
7022 IMUL_HIGH_TO_MUL
7023 : 0));
7024
7025 do_vec_index_to_cond_assign(ir);
7026 lower_vector_insert(ir, true);
7027 lower_quadop_vector(ir, false);
7028 lower_noise(ir);
7029 if (options->MaxIfDepth == 0) {
7030 lower_discard(ir);
7031 }
7032
7033 if (ctx->Const.GLSLOptimizeConservatively) {
7034 /* Do it once and repeat only if there's unsupported control flow. */
7035 do {
7036 do_common_optimization(ir, true, true, options,
7037 ctx->Const.NativeIntegers);
7038 lower_if_to_cond_assign((gl_shader_stage)i, ir,
7039 options->MaxIfDepth, if_threshold);
7040 } while (has_unsupported_control_flow(ir, options));
7041 } else {
7042 /* Repeat it until it stops making changes. */
7043 bool progress;
7044 do {
7045 progress = do_common_optimization(ir, true, true, options,
7046 ctx->Const.NativeIntegers);
7047 progress |= lower_if_to_cond_assign((gl_shader_stage)i, ir,
7048 options->MaxIfDepth, if_threshold);
7049 } while (progress);
7050 }
7051
7052 /* Do this again to lower ir_binop_vector_extract introduced
7053 * by optimization passes.
7054 */
7055 do_vec_index_to_cond_assign(ir);
7056
7057 validate_ir_tree(ir);
7058 }
7059
7060 build_program_resource_list(ctx, prog);
7061
7062 if (use_nir)
7063 return st_link_nir(ctx, prog);
7064
7065 for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
7066 struct gl_linked_shader *shader = prog->_LinkedShaders[i];
7067 if (shader == NULL)
7068 continue;
7069
7070 struct gl_program *linked_prog =
7071 get_mesa_program_tgsi(ctx, prog, shader);
7072 st_set_prog_affected_state_flags(linked_prog);
7073
7074 if (linked_prog) {
7075 if (!ctx->Driver.ProgramStringNotify(ctx,
7076 _mesa_shader_stage_to_program(i),
7077 linked_prog)) {
7078 _mesa_reference_program(ctx, &shader->Program, NULL);
7079 return GL_FALSE;
7080 }
7081 }
7082 }
7083
7084 return GL_TRUE;
7085 }
7086
7087 void
st_translate_stream_output_info(glsl_to_tgsi_visitor * glsl_to_tgsi,const ubyte outputMapping[],struct pipe_stream_output_info * so)7088 st_translate_stream_output_info(glsl_to_tgsi_visitor *glsl_to_tgsi,
7089 const ubyte outputMapping[],
7090 struct pipe_stream_output_info *so)
7091 {
7092 if (!glsl_to_tgsi->shader_program->last_vert_prog)
7093 return;
7094
7095 struct gl_transform_feedback_info *info =
7096 glsl_to_tgsi->shader_program->last_vert_prog->sh.LinkedTransformFeedback;
7097 st_translate_stream_output_info2(info, outputMapping, so);
7098 }
7099
7100 void
st_translate_stream_output_info2(struct gl_transform_feedback_info * info,const ubyte outputMapping[],struct pipe_stream_output_info * so)7101 st_translate_stream_output_info2(struct gl_transform_feedback_info *info,
7102 const ubyte outputMapping[],
7103 struct pipe_stream_output_info *so)
7104 {
7105 unsigned i;
7106
7107 for (i = 0; i < info->NumOutputs; i++) {
7108 so->output[i].register_index =
7109 outputMapping[info->Outputs[i].OutputRegister];
7110 so->output[i].start_component = info->Outputs[i].ComponentOffset;
7111 so->output[i].num_components = info->Outputs[i].NumComponents;
7112 so->output[i].output_buffer = info->Outputs[i].OutputBuffer;
7113 so->output[i].dst_offset = info->Outputs[i].DstOffset;
7114 so->output[i].stream = info->Outputs[i].StreamId;
7115 }
7116
7117 for (i = 0; i < PIPE_MAX_SO_BUFFERS; i++) {
7118 so->stride[i] = info->Buffers[i].Stride;
7119 }
7120 so->num_outputs = info->NumOutputs;
7121 }
7122
7123 } /* extern "C" */
7124