src/compiler/nir/nir_builder.h

   1 /*
   2  * Copyright © 2014-2015 Broadcom
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  */
  23
  24 #ifndef NIR_BUILDER_H
  25 #define NIR_BUILDER_H
  26
  27 #include "nir_control_flow.h"
  28 #include "util/half_float.h"
  29
  30 struct exec_list;
  31
  32 typedef struct nir_builder {
  33    nir_cursor cursor;
  34
  35    /* Whether new ALU instructions will be marked "exact" */
  36    bool exact;
  37
  38    nir_shader *shader;
  39    nir_function_impl *impl;
  40 } nir_builder;
  41
  42 static inline void
  43 nir_builder_init(nir_builder *build, nir_function_impl *impl)
  44 {
  45    memset(build, 0, sizeof(*build));
  46    build->exact = false;
  47    build->impl = impl;
  48    build->shader = impl->function->shader;
  49 }
  50
  51 static inline void
  52 nir_builder_init_simple_shader(nir_builder *build, void *mem_ctx,
  53                                gl_shader_stage stage,
  54                                const nir_shader_compiler_options *options)
  55 {
  56    build->shader = nir_shader_create(mem_ctx, stage, options, NULL);
  57    nir_function *func = nir_function_create(build->shader, "main");
  58    func->is_entrypoint = true;
  59    build->exact = false;
  60    build->impl = nir_function_impl_create(func);
  61    build->cursor = nir_after_cf_list(&build->impl->body);
  62 }
  63
  64 static inline void
  65 nir_builder_instr_insert(nir_builder *build, nir_instr *instr)
  66 {
  67    nir_instr_insert(build->cursor, instr);
  68
  69    /* Move the cursor forward. */
  70    build->cursor = nir_after_instr(instr);
  71 }
  72
  73 static inline nir_instr *
  74 nir_builder_last_instr(nir_builder *build)
  75 {
  76    assert(build->cursor.option == nir_cursor_after_instr);
  77    return build->cursor.instr;
  78 }
  79
  80 static inline void
  81 nir_builder_cf_insert(nir_builder *build, nir_cf_node *cf)
  82 {
  83    nir_cf_node_insert(build->cursor, cf);
  84 }
  85
  86 static inline bool
  87 nir_builder_is_inside_cf(nir_builder *build, nir_cf_node *cf_node)
  88 {
  89    nir_block *block = nir_cursor_current_block(build->cursor);
  90    for (nir_cf_node *n = &block->cf_node; n; n = n->parent) {
  91       if (n == cf_node)
  92          return true;
  93    }
  94    return false;
  95 }
  96
  97 static inline nir_if *
  98 nir_push_if(nir_builder *build, nir_ssa_def *condition)
  99 {
 100    nir_if *nif = nir_if_create(build->shader);
 101    nif->condition = nir_src_for_ssa(condition);
 102    nir_builder_cf_insert(build, &nif->cf_node);
 103    build->cursor = nir_before_cf_list(&nif->then_list);
 104    return nif;
 105 }
 106
 107 static inline nir_if *
 108 nir_push_else(nir_builder *build, nir_if *nif)
 109 {
 110    if (nif) {
 111       assert(nir_builder_is_inside_cf(build, &nif->cf_node));
 112    } else {
 113       nir_block *block = nir_cursor_current_block(build->cursor);
 114       nif = nir_cf_node_as_if(block->cf_node.parent);
 115    }
 116    build->cursor = nir_before_cf_list(&nif->else_list);
 117    return nif;
 118 }
 119
 120 static inline void
 121 nir_pop_if(nir_builder *build, nir_if *nif)
 122 {
 123    if (nif) {
 124       assert(nir_builder_is_inside_cf(build, &nif->cf_node));
 125    } else {
 126       nir_block *block = nir_cursor_current_block(build->cursor);
 127       nif = nir_cf_node_as_if(block->cf_node.parent);
 128    }
 129    build->cursor = nir_after_cf_node(&nif->cf_node);
 130 }
 131
 132 static inline nir_ssa_def *
 133 nir_if_phi(nir_builder *build, nir_ssa_def *then_def, nir_ssa_def *else_def)
 134 {
 135    nir_block *block = nir_cursor_current_block(build->cursor);
 136    nir_if *nif = nir_cf_node_as_if(nir_cf_node_prev(&block->cf_node));
 137
 138    nir_phi_instr *phi = nir_phi_instr_create(build->shader);
 139
 140    nir_phi_src *src = ralloc(phi, nir_phi_src);
 141    src->pred = nir_if_last_then_block(nif);
 142    src->src = nir_src_for_ssa(then_def);
 143    exec_list_push_tail(&phi->srcs, &src->node);
 144
 145    src = ralloc(phi, nir_phi_src);
 146    src->pred = nir_if_last_else_block(nif);
 147    src->src = nir_src_for_ssa(else_def);
 148    exec_list_push_tail(&phi->srcs, &src->node);
 149
 150    assert(then_def->num_components == else_def->num_components);
 151    assert(then_def->bit_size == else_def->bit_size);
 152    nir_ssa_dest_init(&phi->instr, &phi->dest,
 153                      then_def->num_components, then_def->bit_size, NULL);
 154
 155    nir_builder_instr_insert(build, &phi->instr);
 156
 157    return &phi->dest.ssa;
 158 }
 159
 160 static inline nir_loop *
 161 nir_push_loop(nir_builder *build)
 162 {
 163    nir_loop *loop = nir_loop_create(build->shader);
 164    nir_builder_cf_insert(build, &loop->cf_node);
 165    build->cursor = nir_before_cf_list(&loop->body);
 166    return loop;
 167 }
 168
 169 static inline void
 170 nir_pop_loop(nir_builder *build, nir_loop *loop)
 171 {
 172    if (loop) {
 173       assert(nir_builder_is_inside_cf(build, &loop->cf_node));
 174    } else {
 175       nir_block *block = nir_cursor_current_block(build->cursor);
 176       loop = nir_cf_node_as_loop(block->cf_node.parent);
 177    }
 178    build->cursor = nir_after_cf_node(&loop->cf_node);
 179 }
 180
 181 static inline nir_ssa_def *
 182 nir_ssa_undef(nir_builder *build, unsigned num_components, unsigned bit_size)
 183 {
 184    nir_ssa_undef_instr *undef =
 185       nir_ssa_undef_instr_create(build->shader, num_components, bit_size);
 186    if (!undef)
 187       return NULL;
 188
 189    nir_instr_insert(nir_before_cf_list(&build->impl->body), &undef->instr);
 190
 191    return &undef->def;
 192 }
 193
 194 static inline nir_ssa_def *
 195 nir_build_imm(nir_builder *build, unsigned num_components,
 196               unsigned bit_size, nir_const_value value)
 197 {
 198    nir_load_const_instr *load_const =
 199       nir_load_const_instr_create(build->shader, num_components, bit_size);
 200    if (!load_const)
 201       return NULL;
 202
 203    load_const->value = value;
 204
 205    nir_builder_instr_insert(build, &load_const->instr);
 206
 207    return &load_const->def;
 208 }
 209
 210 static inline nir_ssa_def *
 211 nir_imm_bool(nir_builder *build, bool x)
 212 {
 213    nir_const_value v;
 214
 215    memset(&v, 0, sizeof(v));
 216    v.b[0] = x;
 217
 218    return nir_build_imm(build, 1, 1, v);
 219 }
 220
 221 static inline nir_ssa_def *
 222 nir_imm_true(nir_builder *build)
 223 {
 224    return nir_imm_bool(build, true);
 225 }
 226
 227 static inline nir_ssa_def *
 228 nir_imm_false(nir_builder *build)
 229 {
 230    return nir_imm_bool(build, false);
 231 }
 232
 233 static inline nir_ssa_def *
 234 nir_imm_float16(nir_builder *build, float x)
 235 {
 236    nir_const_value v;
 237
 238    memset(&v, 0, sizeof(v));
 239    v.u16[0] = _mesa_float_to_half(x);
 240
 241    return nir_build_imm(build, 1, 16, v);
 242 }
 243
 244 static inline nir_ssa_def *
 245 nir_imm_float(nir_builder *build, float x)
 246 {
 247    nir_const_value v;
 248
 249    memset(&v, 0, sizeof(v));
 250    v.f32[0] = x;
 251
 252    return nir_build_imm(build, 1, 32, v);
 253 }
 254
 255 static inline nir_ssa_def *
 256 nir_imm_double(nir_builder *build, double x)
 257 {
 258    nir_const_value v;
 259
 260    memset(&v, 0, sizeof(v));
 261    v.f64[0] = x;
 262
 263    return nir_build_imm(build, 1, 64, v);
 264 }
 265
 266 static inline nir_ssa_def *
 267 nir_imm_floatN_t(nir_builder *build, double x, unsigned bit_size)
 268 {
 269    switch (bit_size) {
 270    case 16:
 271       return nir_imm_float16(build, x);
 272    case 32:
 273       return nir_imm_float(build, x);
 274    case 64:
 275       return nir_imm_double(build, x);
 276    }
 277
 278    unreachable("unknown float immediate bit size");
 279 }
 280
 281 static inline nir_ssa_def *
 282 nir_imm_vec4(nir_builder *build, float x, float y, float z, float w)
 283 {
 284    nir_const_value v;
 285
 286    memset(&v, 0, sizeof(v));
 287    v.f32[0] = x;
 288    v.f32[1] = y;
 289    v.f32[2] = z;
 290    v.f32[3] = w;
 291
 292    return nir_build_imm(build, 4, 32, v);
 293 }
 294
 295 static inline nir_ssa_def *
 296 nir_imm_ivec2(nir_builder *build, int x, int y)
 297 {
 298    nir_const_value v;
 299
 300    memset(&v, 0, sizeof(v));
 301    v.i32[0] = x;
 302    v.i32[1] = y;
 303
 304    return nir_build_imm(build, 2, 32, v);
 305 }
 306
 307 static inline nir_ssa_def *
 308 nir_imm_int(nir_builder *build, int x)
 309 {
 310    nir_const_value v;
 311
 312    memset(&v, 0, sizeof(v));
 313    v.i32[0] = x;
 314
 315    return nir_build_imm(build, 1, 32, v);
 316 }
 317
 318 static inline nir_ssa_def *
 319 nir_imm_int64(nir_builder *build, int64_t x)
 320 {
 321    nir_const_value v;
 322
 323    memset(&v, 0, sizeof(v));
 324    v.i64[0] = x;
 325
 326    return nir_build_imm(build, 1, 64, v);
 327 }
 328
 329 static inline nir_ssa_def *
 330 nir_imm_intN_t(nir_builder *build, uint64_t x, unsigned bit_size)
 331 {
 332    nir_const_value v;
 333
 334    memset(&v, 0, sizeof(v));
 335    assert(bit_size <= 64);
 336    if (bit_size == 1)
 337       v.b[0] = x & 1;
 338    else
 339       v.i64[0] = x & (~0ull >> (64 - bit_size));
 340
 341    return nir_build_imm(build, 1, bit_size, v);
 342 }
 343
 344 static inline nir_ssa_def *
 345 nir_imm_ivec4(nir_builder *build, int x, int y, int z, int w)
 346 {
 347    nir_const_value v;
 348
 349    memset(&v, 0, sizeof(v));
 350    v.i32[0] = x;
 351    v.i32[1] = y;
 352    v.i32[2] = z;
 353    v.i32[3] = w;
 354
 355    return nir_build_imm(build, 4, 32, v);
 356 }
 357
 358 static inline nir_ssa_def *
 359 nir_imm_boolN_t(nir_builder *build, bool x, unsigned bit_size)
 360 {
 361    /* We use a 0/-1 convention for all booleans regardless of size */
 362    return nir_imm_intN_t(build, -(int)x, bit_size);
 363 }
 364
 365 static inline nir_ssa_def *
 366 nir_build_alu(nir_builder *build, nir_op op, nir_ssa_def *src0,
 367               nir_ssa_def *src1, nir_ssa_def *src2, nir_ssa_def *src3)
 368 {
 369    const nir_op_info *op_info = &nir_op_infos[op];
 370    nir_alu_instr *instr = nir_alu_instr_create(build->shader, op);
 371    if (!instr)
 372       return NULL;
 373
 374    instr->exact = build->exact;
 375
 376    instr->src[0].src = nir_src_for_ssa(src0);
 377    if (src1)
 378       instr->src[1].src = nir_src_for_ssa(src1);
 379    if (src2)
 380       instr->src[2].src = nir_src_for_ssa(src2);
 381    if (src3)
 382       instr->src[3].src = nir_src_for_ssa(src3);
 383
 384    /* Guess the number of components the destination temporary should have
 385     * based on our input sizes, if it's not fixed for the op.
 386     */
 387    unsigned num_components = op_info->output_size;
 388    if (num_components == 0) {
 389       for (unsigned i = 0; i < op_info->num_inputs; i++) {
 390          if (op_info->input_sizes[i] == 0)
 391             num_components = MAX2(num_components,
 392                                   instr->src[i].src.ssa->num_components);
 393       }
 394    }
 395    assert(num_components != 0);
 396
 397    /* Figure out the bitwidth based on the source bitwidth if the instruction
 398     * is variable-width.
 399     */
 400    unsigned bit_size = nir_alu_type_get_type_size(op_info->output_type);
 401    if (bit_size == 0) {
 402       for (unsigned i = 0; i < op_info->num_inputs; i++) {
 403          unsigned src_bit_size = instr->src[i].src.ssa->bit_size;
 404          if (nir_alu_type_get_type_size(op_info->input_types[i]) == 0) {
 405             if (bit_size)
 406                assert(src_bit_size == bit_size);
 407             else
 408                bit_size = src_bit_size;
 409          } else {
 410             assert(src_bit_size ==
 411                nir_alu_type_get_type_size(op_info->input_types[i]));
 412          }
 413       }
 414    }
 415
 416    /* When in doubt, assume 32. */
 417    if (bit_size == 0)
 418       bit_size = 32;
 419
 420    /* Make sure we don't swizzle from outside of our source vector (like if a
 421     * scalar value was passed into a multiply with a vector).
 422     */
 423    for (unsigned i = 0; i < op_info->num_inputs; i++) {
 424       for (unsigned j = instr->src[i].src.ssa->num_components;
 425            j < NIR_MAX_VEC_COMPONENTS; j++) {
 426          instr->src[i].swizzle[j] = instr->src[i].src.ssa->num_components - 1;
 427       }
 428    }
 429
 430    nir_ssa_dest_init(&instr->instr, &instr->dest.dest, num_components,
 431                      bit_size, NULL);
 432    instr->dest.write_mask = (1 << num_components) - 1;
 433
 434    nir_builder_instr_insert(build, &instr->instr);
 435
 436    return &instr->dest.dest.ssa;
 437 }
 438
 439 #include "nir_builder_opcodes.h"
 440
 441 static inline nir_ssa_def *
 442 nir_vec(nir_builder *build, nir_ssa_def **comp, unsigned num_components)
 443 {
 444    switch (num_components) {
 445    case 4:
 446       return nir_vec4(build, comp[0], comp[1], comp[2], comp[3]);
 447    case 3:
 448       return nir_vec3(build, comp[0], comp[1], comp[2]);
 449    case 2:
 450       return nir_vec2(build, comp[0], comp[1]);
 451    case 1:
 452       return comp[0];
 453    default:
 454       unreachable("bad component count");
 455       return NULL;
 456    }
 457 }
 458
 459 /**
 460  * Similar to nir_fmov, but takes a nir_alu_src instead of a nir_ssa_def.
 461  */
 462 static inline nir_ssa_def *
 463 nir_fmov_alu(nir_builder *build, nir_alu_src src, unsigned num_components)
 464 {
 465    nir_alu_instr *mov = nir_alu_instr_create(build->shader, nir_op_fmov);
 466    nir_ssa_dest_init(&mov->instr, &mov->dest.dest, num_components,
 467                      nir_src_bit_size(src.src), NULL);
 468    mov->exact = build->exact;
 469    mov->dest.write_mask = (1 << num_components) - 1;
 470    mov->src[0] = src;
 471    nir_builder_instr_insert(build, &mov->instr);
 472
 473    return &mov->dest.dest.ssa;
 474 }
 475
 476 static inline nir_ssa_def *
 477 nir_imov_alu(nir_builder *build, nir_alu_src src, unsigned num_components)
 478 {
 479    nir_alu_instr *mov = nir_alu_instr_create(build->shader, nir_op_imov);
 480    nir_ssa_dest_init(&mov->instr, &mov->dest.dest, num_components,
 481                      nir_src_bit_size(src.src), NULL);
 482    mov->exact = build->exact;
 483    mov->dest.write_mask = (1 << num_components) - 1;
 484    mov->src[0] = src;
 485    nir_builder_instr_insert(build, &mov->instr);
 486
 487    return &mov->dest.dest.ssa;
 488 }
 489
 490 /**
 491  * Construct an fmov or imov that reswizzles the source's components.
 492  */
 493 static inline nir_ssa_def *
 494 nir_swizzle(nir_builder *build, nir_ssa_def *src, const unsigned *swiz,
 495             unsigned num_components, bool use_fmov)
 496 {
 497    assert(num_components <= NIR_MAX_VEC_COMPONENTS);
 498    nir_alu_src alu_src = { NIR_SRC_INIT };
 499    alu_src.src = nir_src_for_ssa(src);
 500
 501    bool is_identity_swizzle = true;
 502    for (unsigned i = 0; i < num_components && i < NIR_MAX_VEC_COMPONENTS; i++) {
 503       if (swiz[i] != i)
 504          is_identity_swizzle = false;
 505       alu_src.swizzle[i] = swiz[i];
 506    }
 507
 508    if (num_components == src->num_components && is_identity_swizzle)
 509       return src;
 510
 511    return use_fmov ? nir_fmov_alu(build, alu_src, num_components) :
 512                      nir_imov_alu(build, alu_src, num_components);
 513 }
 514
 515 /* Selects the right fdot given the number of components in each source. */
 516 static inline nir_ssa_def *
 517 nir_fdot(nir_builder *build, nir_ssa_def *src0, nir_ssa_def *src1)
 518 {
 519    assert(src0->num_components == src1->num_components);
 520    switch (src0->num_components) {
 521    case 1: return nir_fmul(build, src0, src1);
 522    case 2: return nir_fdot2(build, src0, src1);
 523    case 3: return nir_fdot3(build, src0, src1);
 524    case 4: return nir_fdot4(build, src0, src1);
 525    default:
 526       unreachable("bad component size");
 527    }
 528
 529    return NULL;
 530 }
 531
 532 static inline nir_ssa_def *
 533 nir_bany_inequal(nir_builder *b, nir_ssa_def *src0, nir_ssa_def *src1)
 534 {
 535    switch (src0->num_components) {
 536    case 1: return nir_ine(b, src0, src1);
 537    case 2: return nir_bany_inequal2(b, src0, src1);
 538    case 3: return nir_bany_inequal3(b, src0, src1);
 539    case 4: return nir_bany_inequal4(b, src0, src1);
 540    default:
 541       unreachable("bad component size");
 542    }
 543 }
 544
 545 static inline nir_ssa_def *
 546 nir_bany(nir_builder *b, nir_ssa_def *src)
 547 {
 548    return nir_bany_inequal(b, src, nir_imm_false(b));
 549 }
 550
 551 static inline nir_ssa_def *
 552 nir_channel(nir_builder *b, nir_ssa_def *def, unsigned c)
 553 {
 554    return nir_swizzle(b, def, &c, 1, false);
 555 }
 556
 557 static inline nir_ssa_def *
 558 nir_channels(nir_builder *b, nir_ssa_def *def, nir_component_mask_t mask)
 559 {
 560    unsigned num_channels = 0, swizzle[NIR_MAX_VEC_COMPONENTS] = { 0 };
 561
 562    for (unsigned i = 0; i < NIR_MAX_VEC_COMPONENTS; i++) {
 563       if ((mask & (1 << i)) == 0)
 564          continue;
 565       swizzle[num_channels++] = i;
 566    }
 567
 568    return nir_swizzle(b, def, swizzle, num_channels, false);
 569 }
 570
 571 static inline nir_ssa_def *
 572 nir_i2i(nir_builder *build, nir_ssa_def *x, unsigned dest_bit_size)
 573 {
 574    if (x->bit_size == dest_bit_size)
 575       return x;
 576
 577    switch (dest_bit_size) {
 578    case 64: return nir_i2i64(build, x);
 579    case 32: return nir_i2i32(build, x);
 580    case 16: return nir_i2i16(build, x);
 581    case 8:  return nir_i2i8(build, x);
 582    default: unreachable("Invalid bit size");
 583    }
 584 }
 585
 586 static inline nir_ssa_def *
 587 nir_u2u(nir_builder *build, nir_ssa_def *x, unsigned dest_bit_size)
 588 {
 589    if (x->bit_size == dest_bit_size)
 590       return x;
 591
 592    switch (dest_bit_size) {
 593    case 64: return nir_u2u64(build, x);
 594    case 32: return nir_u2u32(build, x);
 595    case 16: return nir_u2u16(build, x);
 596    case 8:  return nir_u2u8(build, x);
 597    default: unreachable("Invalid bit size");
 598    }
 599 }
 600
 601 static inline nir_ssa_def *
 602 nir_iadd_imm(nir_builder *build, nir_ssa_def *x, uint64_t y)
 603 {
 604    return nir_iadd(build, x, nir_imm_intN_t(build, y, x->bit_size));
 605 }
 606
 607 static inline nir_ssa_def *
 608 nir_imul_imm(nir_builder *build, nir_ssa_def *x, uint64_t y)
 609 {
 610    return nir_imul(build, x, nir_imm_intN_t(build, y, x->bit_size));
 611 }
 612
 613 static inline nir_ssa_def *
 614 nir_fadd_imm(nir_builder *build, nir_ssa_def *x, double y)
 615 {
 616    return nir_fadd(build, x, nir_imm_floatN_t(build, y, x->bit_size));
 617 }
 618
 619 static inline nir_ssa_def *
 620 nir_fmul_imm(nir_builder *build, nir_ssa_def *x, double y)
 621 {
 622    return nir_fmul(build, x, nir_imm_floatN_t(build, y, x->bit_size));
 623 }
 624
 625 static inline nir_ssa_def *
 626 nir_pack_bits(nir_builder *b, nir_ssa_def *src, unsigned dest_bit_size)
 627 {
 628    assert(src->num_components * src->bit_size == dest_bit_size);
 629
 630    switch (dest_bit_size) {
 631    case 64:
 632       switch (src->bit_size) {
 633       case 32: return nir_pack_64_2x32(b, src);
 634       case 16: return nir_pack_64_4x16(b, src);
 635       default: break;
 636       }
 637       break;
 638
 639    case 32:
 640       if (src->bit_size == 16)
 641          return nir_pack_32_2x16(b, src);
 642       break;
 643
 644    default:
 645       break;
 646    }
 647
 648    /* If we got here, we have no dedicated unpack opcode. */
 649    nir_ssa_def *dest = nir_imm_intN_t(b, 0, dest_bit_size);
 650    for (unsigned i = 0; i < src->num_components; i++) {
 651       nir_ssa_def *val = nir_u2u(b, nir_channel(b, src, i), dest_bit_size);
 652       val = nir_ishl(b, val, nir_imm_int(b, i * src->bit_size));
 653       dest = nir_ior(b, dest, val);
 654    }
 655    return dest;
 656 }
 657
 658 static inline nir_ssa_def *
 659 nir_unpack_bits(nir_builder *b, nir_ssa_def *src, unsigned dest_bit_size)
 660 {
 661    assert(src->num_components == 1);
 662    assert(src->bit_size > dest_bit_size);
 663    const unsigned dest_num_components = src->bit_size / dest_bit_size;
 664    assert(dest_num_components <= NIR_MAX_VEC_COMPONENTS);
 665
 666    switch (src->bit_size) {
 667    case 64:
 668       switch (dest_bit_size) {
 669       case 32: return nir_unpack_64_2x32(b, src);
 670       case 16: return nir_unpack_64_4x16(b, src);
 671       default: break;
 672       }
 673       break;
 674
 675    case 32:
 676       if (dest_bit_size == 16)
 677          return nir_unpack_32_2x16(b, src);
 678       break;
 679
 680    default:
 681       break;
 682    }
 683
 684    /* If we got here, we have no dedicated unpack opcode. */
 685    nir_ssa_def *dest_comps[NIR_MAX_VEC_COMPONENTS];
 686    for (unsigned i = 0; i < dest_num_components; i++) {
 687       nir_ssa_def *val = nir_ushr(b, src, nir_imm_int(b, i * dest_bit_size));
 688       dest_comps[i] = nir_u2u(b, val, dest_bit_size);
 689    }
 690    return nir_vec(b, dest_comps, dest_num_components);
 691 }
 692
 693 static inline nir_ssa_def *
 694 nir_bitcast_vector(nir_builder *b, nir_ssa_def *src, unsigned dest_bit_size)
 695 {
 696    assert((src->bit_size * src->num_components) % dest_bit_size == 0);
 697    const unsigned dest_num_components =
 698       (src->bit_size * src->num_components) / dest_bit_size;
 699    assert(dest_num_components <= NIR_MAX_VEC_COMPONENTS);
 700
 701    if (src->bit_size > dest_bit_size) {
 702       assert(src->bit_size % dest_bit_size == 0);
 703       if (src->num_components == 1) {
 704          return nir_unpack_bits(b, src, dest_bit_size);
 705       } else {
 706          const unsigned divisor = src->bit_size / dest_bit_size;
 707          assert(src->num_components * divisor == dest_num_components);
 708          nir_ssa_def *dest[NIR_MAX_VEC_COMPONENTS];
 709          for (unsigned i = 0; i < src->num_components; i++) {
 710             nir_ssa_def *unpacked =
 711                nir_unpack_bits(b, nir_channel(b, src, i), dest_bit_size);
 712             assert(unpacked->num_components == divisor);
 713             for (unsigned j = 0; j < divisor; j++)
 714                dest[i * divisor + j] = nir_channel(b, unpacked, j);
 715          }
 716          return nir_vec(b, dest, dest_num_components);
 717       }
 718    } else if (src->bit_size < dest_bit_size) {
 719       assert(dest_bit_size % src->bit_size == 0);
 720       if (dest_num_components == 1) {
 721          return nir_pack_bits(b, src, dest_bit_size);
 722       } else {
 723          const unsigned divisor = dest_bit_size / src->bit_size;
 724          assert(src->num_components == dest_num_components * divisor);
 725          nir_ssa_def *dest[NIR_MAX_VEC_COMPONENTS];
 726          for (unsigned i = 0; i < dest_num_components; i++) {
 727             nir_component_mask_t src_mask =
 728                ((1 << divisor) - 1) << (i * divisor);
 729             dest[i] = nir_pack_bits(b, nir_channels(b, src, src_mask),
 730                                        dest_bit_size);
 731          }
 732          return nir_vec(b, dest, dest_num_components);
 733       }
 734    } else {
 735       assert(src->bit_size == dest_bit_size);
 736       return src;
 737    }
 738 }
 739
 740 /**
 741  * Turns a nir_src into a nir_ssa_def * so it can be passed to
 742  * nir_build_alu()-based builder calls.
 743  *
 744  * See nir_ssa_for_alu_src() for alu instructions.
 745  */
 746 static inline nir_ssa_def *
 747 nir_ssa_for_src(nir_builder *build, nir_src src, int num_components)
 748 {
 749    if (src.is_ssa && src.ssa->num_components == num_components)
 750       return src.ssa;
 751
 752    nir_alu_src alu = { NIR_SRC_INIT };
 753    alu.src = src;
 754    for (int j = 0; j < 4; j++)
 755       alu.swizzle[j] = j;
 756
 757    return nir_imov_alu(build, alu, num_components);
 758 }
 759
 760 /**
 761  * Similar to nir_ssa_for_src(), but for alu srcs, respecting the
 762  * nir_alu_src's swizzle.
 763  */
 764 static inline nir_ssa_def *
 765 nir_ssa_for_alu_src(nir_builder *build, nir_alu_instr *instr, unsigned srcn)
 766 {
 767    static uint8_t trivial_swizzle[NIR_MAX_VEC_COMPONENTS];
 768    for (int i = 0; i < NIR_MAX_VEC_COMPONENTS; ++i)
 769       trivial_swizzle[i] = i;
 770    nir_alu_src *src = &instr->src[srcn];
 771    unsigned num_components = nir_ssa_alu_instr_src_components(instr, srcn);
 772
 773    if (src->src.is_ssa && (src->src.ssa->num_components == num_components) &&
 774        !src->abs && !src->negate &&
 775        (memcmp(src->swizzle, trivial_swizzle, num_components) == 0))
 776       return src->src.ssa;
 777
 778    return nir_imov_alu(build, *src, num_components);
 779 }
 780
 781 static inline nir_deref_instr *
 782 nir_build_deref_var(nir_builder *build, nir_variable *var)
 783 {
 784    nir_deref_instr *deref =
 785       nir_deref_instr_create(build->shader, nir_deref_type_var);
 786
 787    deref->mode = var->data.mode;
 788    deref->type = var->type;
 789    deref->var = var;
 790
 791    nir_ssa_dest_init(&deref->instr, &deref->dest, 1, 32, NULL);
 792
 793    nir_builder_instr_insert(build, &deref->instr);
 794
 795    return deref;
 796 }
 797
 798 static inline nir_deref_instr *
 799 nir_build_deref_array(nir_builder *build, nir_deref_instr *parent,
 800                       nir_ssa_def *index)
 801 {
 802    assert(glsl_type_is_array(parent->type) ||
 803           glsl_type_is_matrix(parent->type) ||
 804           glsl_type_is_vector(parent->type));
 805
 806    nir_deref_instr *deref =
 807       nir_deref_instr_create(build->shader, nir_deref_type_array);
 808
 809    deref->mode = parent->mode;
 810    deref->type = glsl_get_array_element(parent->type);
 811    deref->parent = nir_src_for_ssa(&parent->dest.ssa);
 812    deref->arr.index = nir_src_for_ssa(index);
 813
 814    nir_ssa_dest_init(&deref->instr, &deref->dest,
 815                      parent->dest.ssa.num_components,
 816                      parent->dest.ssa.bit_size, NULL);
 817
 818    nir_builder_instr_insert(build, &deref->instr);
 819
 820    return deref;
 821 }
 822
 823 static inline nir_deref_instr *
 824 nir_build_deref_ptr_as_array(nir_builder *build, nir_deref_instr *parent,
 825                              nir_ssa_def *index)
 826 {
 827    assert(parent->deref_type == nir_deref_type_array ||
 828           parent->deref_type == nir_deref_type_ptr_as_array ||
 829           parent->deref_type == nir_deref_type_cast);
 830
 831    nir_deref_instr *deref =
 832       nir_deref_instr_create(build->shader, nir_deref_type_ptr_as_array);
 833
 834    deref->mode = parent->mode;
 835    deref->type = parent->type;
 836    deref->parent = nir_src_for_ssa(&parent->dest.ssa);
 837    deref->arr.index = nir_src_for_ssa(index);
 838
 839    nir_ssa_dest_init(&deref->instr, &deref->dest,
 840                      parent->dest.ssa.num_components,
 841                      parent->dest.ssa.bit_size, NULL);
 842
 843    nir_builder_instr_insert(build, &deref->instr);
 844
 845    return deref;
 846 }
 847
 848 static inline nir_deref_instr *
 849 nir_build_deref_array_wildcard(nir_builder *build, nir_deref_instr *parent)
 850 {
 851    assert(glsl_type_is_array(parent->type) ||
 852           glsl_type_is_matrix(parent->type));
 853
 854    nir_deref_instr *deref =
 855       nir_deref_instr_create(build->shader, nir_deref_type_array_wildcard);
 856
 857    deref->mode = parent->mode;
 858    deref->type = glsl_get_array_element(parent->type);
 859    deref->parent = nir_src_for_ssa(&parent->dest.ssa);
 860
 861    nir_ssa_dest_init(&deref->instr, &deref->dest,
 862                      parent->dest.ssa.num_components,
 863                      parent->dest.ssa.bit_size, NULL);
 864
 865    nir_builder_instr_insert(build, &deref->instr);
 866
 867    return deref;
 868 }
 869
 870 static inline nir_deref_instr *
 871 nir_build_deref_struct(nir_builder *build, nir_deref_instr *parent,
 872                        unsigned index)
 873 {
 874    assert(glsl_type_is_struct_or_ifc(parent->type));
 875
 876    nir_deref_instr *deref =
 877       nir_deref_instr_create(build->shader, nir_deref_type_struct);
 878
 879    deref->mode = parent->mode;
 880    deref->type = glsl_get_struct_field(parent->type, index);
 881    deref->parent = nir_src_for_ssa(&parent->dest.ssa);
 882    deref->strct.index = index;
 883
 884    nir_ssa_dest_init(&deref->instr, &deref->dest,
 885                      parent->dest.ssa.num_components,
 886                      parent->dest.ssa.bit_size, NULL);
 887
 888    nir_builder_instr_insert(build, &deref->instr);
 889
 890    return deref;
 891 }
 892
 893 static inline nir_deref_instr *
 894 nir_build_deref_cast(nir_builder *build, nir_ssa_def *parent,
 895                      nir_variable_mode mode, const struct glsl_type *type,
 896                      unsigned ptr_stride)
 897 {
 898    nir_deref_instr *deref =
 899       nir_deref_instr_create(build->shader, nir_deref_type_cast);
 900
 901    deref->mode = mode;
 902    deref->type = type;
 903    deref->parent = nir_src_for_ssa(parent);
 904    deref->cast.ptr_stride = ptr_stride;
 905
 906    nir_ssa_dest_init(&deref->instr, &deref->dest,
 907                      parent->num_components, parent->bit_size, NULL);
 908
 909    nir_builder_instr_insert(build, &deref->instr);
 910
 911    return deref;
 912 }
 913
 914 /** Returns a deref that follows another but starting from the given parent
 915  *
 916  * The new deref will be the same type and take the same array or struct index
 917  * as the leader deref but it may have a different parent.  This is very
 918  * useful for walking deref paths.
 919  */
 920 static inline nir_deref_instr *
 921 nir_build_deref_follower(nir_builder *b, nir_deref_instr *parent,
 922                          nir_deref_instr *leader)
 923 {
 924    /* If the derefs would have the same parent, don't make a new one */
 925    assert(leader->parent.is_ssa);
 926    if (leader->parent.ssa == &parent->dest.ssa)
 927       return leader;
 928
 929    UNUSED nir_deref_instr *leader_parent = nir_src_as_deref(leader->parent);
 930
 931    switch (leader->deref_type) {
 932    case nir_deref_type_var:
 933       unreachable("A var dereference cannot have a parent");
 934       break;
 935
 936    case nir_deref_type_array:
 937    case nir_deref_type_array_wildcard:
 938       assert(glsl_type_is_matrix(parent->type) ||
 939              glsl_type_is_array(parent->type) ||
 940              (leader->deref_type == nir_deref_type_array &&
 941               glsl_type_is_vector(parent->type)));
 942       assert(glsl_get_length(parent->type) ==
 943              glsl_get_length(leader_parent->type));
 944
 945       if (leader->deref_type == nir_deref_type_array) {
 946          assert(leader->arr.index.is_ssa);
 947          return nir_build_deref_array(b, parent, leader->arr.index.ssa);
 948       } else {
 949          return nir_build_deref_array_wildcard(b, parent);
 950       }
 951
 952    case nir_deref_type_struct:
 953       assert(glsl_type_is_struct_or_ifc(parent->type));
 954       assert(glsl_get_length(parent->type) ==
 955              glsl_get_length(leader_parent->type));
 956
 957       return nir_build_deref_struct(b, parent, leader->strct.index);
 958
 959    default:
 960       unreachable("Invalid deref instruction type");
 961    }
 962 }
 963
 964 static inline nir_ssa_def *
 965 nir_load_reg(nir_builder *build, nir_register *reg)
 966 {
 967    return nir_ssa_for_src(build, nir_src_for_reg(reg), reg->num_components);
 968 }
 969
 970 static inline nir_ssa_def *
 971 nir_load_deref(nir_builder *build, nir_deref_instr *deref)
 972 {
 973    nir_intrinsic_instr *load =
 974       nir_intrinsic_instr_create(build->shader, nir_intrinsic_load_deref);
 975    load->num_components = glsl_get_vector_elements(deref->type);
 976    load->src[0] = nir_src_for_ssa(&deref->dest.ssa);
 977    nir_ssa_dest_init(&load->instr, &load->dest, load->num_components,
 978                      glsl_get_bit_size(deref->type), NULL);
 979    nir_builder_instr_insert(build, &load->instr);
 980    return &load->dest.ssa;
 981 }
 982
 983 static inline void
 984 nir_store_deref(nir_builder *build, nir_deref_instr *deref,
 985                 nir_ssa_def *value, unsigned writemask)
 986 {
 987    nir_intrinsic_instr *store =
 988       nir_intrinsic_instr_create(build->shader, nir_intrinsic_store_deref);
 989    store->num_components = glsl_get_vector_elements(deref->type);
 990    store->src[0] = nir_src_for_ssa(&deref->dest.ssa);
 991    store->src[1] = nir_src_for_ssa(value);
 992    nir_intrinsic_set_write_mask(store,
 993                                 writemask & ((1 << store->num_components) - 1));
 994    nir_builder_instr_insert(build, &store->instr);
 995 }
 996
 997 static inline void
 998 nir_copy_deref(nir_builder *build, nir_deref_instr *dest, nir_deref_instr *src)
 999 {
1000    nir_intrinsic_instr *copy =
1001       nir_intrinsic_instr_create(build->shader, nir_intrinsic_copy_deref);
1002    copy->src[0] = nir_src_for_ssa(&dest->dest.ssa);
1003    copy->src[1] = nir_src_for_ssa(&src->dest.ssa);
1004    nir_builder_instr_insert(build, &copy->instr);
1005 }
1006
1007 static inline nir_ssa_def *
1008 nir_load_var(nir_builder *build, nir_variable *var)
1009 {
1010    return nir_load_deref(build, nir_build_deref_var(build, var));
1011 }
1012
1013 static inline void
1014 nir_store_var(nir_builder *build, nir_variable *var, nir_ssa_def *value,
1015               unsigned writemask)
1016 {
1017    nir_store_deref(build, nir_build_deref_var(build, var), value, writemask);
1018 }
1019
1020 static inline void
1021 nir_copy_var(nir_builder *build, nir_variable *dest, nir_variable *src)
1022 {
1023    nir_copy_deref(build, nir_build_deref_var(build, dest),
1024                          nir_build_deref_var(build, src));
1025 }
1026
1027 static inline nir_ssa_def *
1028 nir_load_param(nir_builder *build, uint32_t param_idx)
1029 {
1030    assert(param_idx < build->impl->function->num_params);
1031    nir_parameter *param = &build->impl->function->params[param_idx];
1032
1033    nir_intrinsic_instr *load =
1034       nir_intrinsic_instr_create(build->shader, nir_intrinsic_load_param);
1035    nir_intrinsic_set_param_idx(load, param_idx);
1036    load->num_components = param->num_components;
1037    nir_ssa_dest_init(&load->instr, &load->dest,
1038                      param->num_components, param->bit_size, NULL);
1039    nir_builder_instr_insert(build, &load->instr);
1040    return &load->dest.ssa;
1041 }
1042
1043 #include "nir_builder_opcodes.h"
1044
1045 static inline nir_ssa_def *
1046 nir_f2b(nir_builder *build, nir_ssa_def *f)
1047 {
1048    return nir_f2b1(build, f);
1049 }
1050
1051 static inline nir_ssa_def *
1052 nir_i2b(nir_builder *build, nir_ssa_def *i)
1053 {
1054    return nir_i2b1(build, i);
1055 }
1056
1057 static inline nir_ssa_def *
1058 nir_b2f(nir_builder *build, nir_ssa_def *b, uint32_t bit_size)
1059 {
1060    switch (bit_size) {
1061    case 64: return nir_b2f64(build, b);
1062    case 32: return nir_b2f32(build, b);
1063    case 16: return nir_b2f16(build, b);
1064    default:
1065       unreachable("Invalid bit-size");
1066    };
1067 }
1068
1069 static inline nir_ssa_def *
1070 nir_load_barycentric(nir_builder *build, nir_intrinsic_op op,
1071                      unsigned interp_mode)
1072 {
1073    nir_intrinsic_instr *bary = nir_intrinsic_instr_create(build->shader, op);
1074    nir_ssa_dest_init(&bary->instr, &bary->dest, 2, 32, NULL);
1075    nir_intrinsic_set_interp_mode(bary, interp_mode);
1076    nir_builder_instr_insert(build, &bary->instr);
1077    return &bary->dest.ssa;
1078 }
1079
1080 static inline void
1081 nir_jump(nir_builder *build, nir_jump_type jump_type)
1082 {
1083    nir_jump_instr *jump = nir_jump_instr_create(build->shader, jump_type);
1084    nir_builder_instr_insert(build, &jump->instr);
1085 }
1086
1087 static inline nir_ssa_def *
1088 nir_compare_func(nir_builder *b, enum compare_func func,
1089                  nir_ssa_def *src0, nir_ssa_def *src1)
1090 {
1091    switch (func) {
1092    case COMPARE_FUNC_NEVER:
1093       return nir_imm_int(b, 0);
1094    case COMPARE_FUNC_ALWAYS:
1095       return nir_imm_int(b, ~0);
1096    case COMPARE_FUNC_EQUAL:
1097       return nir_feq(b, src0, src1);
1098    case COMPARE_FUNC_NOTEQUAL:
1099       return nir_fne(b, src0, src1);
1100    case COMPARE_FUNC_GREATER:
1101       return nir_flt(b, src1, src0);
1102    case COMPARE_FUNC_GEQUAL:
1103       return nir_fge(b, src0, src1);
1104    case COMPARE_FUNC_LESS:
1105       return nir_flt(b, src0, src1);
1106    case COMPARE_FUNC_LEQUAL:
1107       return nir_fge(b, src1, src0);
1108    }
1109    unreachable("bad compare func");
1110 }
1111
1112 #endif /* NIR_BUILDER_H */