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    build->exact = false;
  59    build->impl = nir_function_impl_create(func);
  60    build->cursor = nir_after_cf_list(&build->impl->body);
  61 }
  62
  63 static inline void
  64 nir_builder_instr_insert(nir_builder *build, nir_instr *instr)
  65 {
  66    nir_instr_insert(build->cursor, instr);
  67
  68    /* Move the cursor forward. */
  69    build->cursor = nir_after_instr(instr);
  70 }
  71
  72 static inline nir_instr *
  73 nir_builder_last_instr(nir_builder *build)
  74 {
  75    assert(build->cursor.option == nir_cursor_after_instr);
  76    return build->cursor.instr;
  77 }
  78
  79 static inline void
  80 nir_builder_cf_insert(nir_builder *build, nir_cf_node *cf)
  81 {
  82    nir_cf_node_insert(build->cursor, cf);
  83 }
  84
  85 static inline bool
  86 nir_builder_is_inside_cf(nir_builder *build, nir_cf_node *cf_node)
  87 {
  88    nir_block *block = nir_cursor_current_block(build->cursor);
  89    for (nir_cf_node *n = &block->cf_node; n; n = n->parent) {
  90       if (n == cf_node)
  91          return true;
  92    }
  93    return false;
  94 }
  95
  96 static inline nir_if *
  97 nir_push_if(nir_builder *build, nir_ssa_def *condition)
  98 {
  99    nir_if *nif = nir_if_create(build->shader);
 100    nif->condition = nir_src_for_ssa(condition);
 101    nir_builder_cf_insert(build, &nif->cf_node);
 102    build->cursor = nir_before_cf_list(&nif->then_list);
 103    return nif;
 104 }
 105
 106 static inline nir_if *
 107 nir_push_else(nir_builder *build, nir_if *nif)
 108 {
 109    if (nif) {
 110       assert(nir_builder_is_inside_cf(build, &nif->cf_node));
 111    } else {
 112       nir_block *block = nir_cursor_current_block(build->cursor);
 113       nif = nir_cf_node_as_if(block->cf_node.parent);
 114    }
 115    build->cursor = nir_before_cf_list(&nif->else_list);
 116    return nif;
 117 }
 118
 119 static inline void
 120 nir_pop_if(nir_builder *build, nir_if *nif)
 121 {
 122    if (nif) {
 123       assert(nir_builder_is_inside_cf(build, &nif->cf_node));
 124    } else {
 125       nir_block *block = nir_cursor_current_block(build->cursor);
 126       nif = nir_cf_node_as_if(block->cf_node.parent);
 127    }
 128    build->cursor = nir_after_cf_node(&nif->cf_node);
 129 }
 130
 131 static inline nir_ssa_def *
 132 nir_if_phi(nir_builder *build, nir_ssa_def *then_def, nir_ssa_def *else_def)
 133 {
 134    nir_block *block = nir_cursor_current_block(build->cursor);
 135    nir_if *nif = nir_cf_node_as_if(nir_cf_node_prev(&block->cf_node));
 136
 137    nir_phi_instr *phi = nir_phi_instr_create(build->shader);
 138
 139    nir_phi_src *src = ralloc(phi, nir_phi_src);
 140    src->pred = nir_if_last_then_block(nif);
 141    src->src = nir_src_for_ssa(then_def);
 142    exec_list_push_tail(&phi->srcs, &src->node);
 143
 144    src = ralloc(phi, nir_phi_src);
 145    src->pred = nir_if_last_else_block(nif);
 146    src->src = nir_src_for_ssa(else_def);
 147    exec_list_push_tail(&phi->srcs, &src->node);
 148
 149    assert(then_def->num_components == else_def->num_components);
 150    assert(then_def->bit_size == else_def->bit_size);
 151    nir_ssa_dest_init(&phi->instr, &phi->dest,
 152                      then_def->num_components, then_def->bit_size, NULL);
 153
 154    nir_builder_instr_insert(build, &phi->instr);
 155
 156    return &phi->dest.ssa;
 157 }
 158
 159 static inline nir_loop *
 160 nir_push_loop(nir_builder *build)
 161 {
 162    nir_loop *loop = nir_loop_create(build->shader);
 163    nir_builder_cf_insert(build, &loop->cf_node);
 164    build->cursor = nir_before_cf_list(&loop->body);
 165    return loop;
 166 }
 167
 168 static inline void
 169 nir_pop_loop(nir_builder *build, nir_loop *loop)
 170 {
 171    if (loop) {
 172       assert(nir_builder_is_inside_cf(build, &loop->cf_node));
 173    } else {
 174       nir_block *block = nir_cursor_current_block(build->cursor);
 175       loop = nir_cf_node_as_loop(block->cf_node.parent);
 176    }
 177    build->cursor = nir_after_cf_node(&loop->cf_node);
 178 }
 179
 180 static inline nir_ssa_def *
 181 nir_ssa_undef(nir_builder *build, unsigned num_components, unsigned bit_size)
 182 {
 183    nir_ssa_undef_instr *undef =
 184       nir_ssa_undef_instr_create(build->shader, num_components, bit_size);
 185    if (!undef)
 186       return NULL;
 187
 188    nir_instr_insert(nir_before_cf_list(&build->impl->body), &undef->instr);
 189
 190    return &undef->def;
 191 }
 192
 193 static inline nir_ssa_def *
 194 nir_build_imm(nir_builder *build, unsigned num_components,
 195               unsigned bit_size, nir_const_value value)
 196 {
 197    nir_load_const_instr *load_const =
 198       nir_load_const_instr_create(build->shader, num_components, bit_size);
 199    if (!load_const)
 200       return NULL;
 201
 202    load_const->value = value;
 203
 204    nir_builder_instr_insert(build, &load_const->instr);
 205
 206    return &load_const->def;
 207 }
 208
 209 static inline nir_ssa_def *
 210 nir_imm_bool(nir_builder *build, bool x)
 211 {
 212    nir_const_value v;
 213
 214    memset(&v, 0, sizeof(v));
 215    v.u32[0] = x ? NIR_TRUE : NIR_FALSE;
 216
 217    return nir_build_imm(build, 1, 32, v);
 218 }
 219
 220 static inline nir_ssa_def *
 221 nir_imm_true(nir_builder *build)
 222 {
 223    return nir_imm_bool(build, true);
 224 }
 225
 226 static inline nir_ssa_def *
 227 nir_imm_false(nir_builder *build)
 228 {
 229    return nir_imm_bool(build, false);
 230 }
 231
 232 static inline nir_ssa_def *
 233 nir_imm_float16(nir_builder *build, float x)
 234 {
 235    nir_const_value v;
 236
 237    memset(&v, 0, sizeof(v));
 238    v.u16[0] = _mesa_float_to_half(x);
 239
 240    return nir_build_imm(build, 1, 16, v);
 241 }
 242
 243 static inline nir_ssa_def *
 244 nir_imm_float(nir_builder *build, float x)
 245 {
 246    nir_const_value v;
 247
 248    memset(&v, 0, sizeof(v));
 249    v.f32[0] = x;
 250
 251    return nir_build_imm(build, 1, 32, v);
 252 }
 253
 254 static inline nir_ssa_def *
 255 nir_imm_double(nir_builder *build, double x)
 256 {
 257    nir_const_value v;
 258
 259    memset(&v, 0, sizeof(v));
 260    v.f64[0] = x;
 261
 262    return nir_build_imm(build, 1, 64, v);
 263 }
 264
 265 static inline nir_ssa_def *
 266 nir_imm_floatN_t(nir_builder *build, double x, unsigned bit_size)
 267 {
 268    switch (bit_size) {
 269    case 16:
 270       return nir_imm_float16(build, x);
 271    case 32:
 272       return nir_imm_float(build, x);
 273    case 64:
 274       return nir_imm_double(build, x);
 275    }
 276
 277    unreachable("unknown float immediate bit size");
 278 }
 279
 280 static inline nir_ssa_def *
 281 nir_imm_vec4(nir_builder *build, float x, float y, float z, float w)
 282 {
 283    nir_const_value v;
 284
 285    memset(&v, 0, sizeof(v));
 286    v.f32[0] = x;
 287    v.f32[1] = y;
 288    v.f32[2] = z;
 289    v.f32[3] = w;
 290
 291    return nir_build_imm(build, 4, 32, v);
 292 }
 293
 294 static inline nir_ssa_def *
 295 nir_imm_ivec2(nir_builder *build, int x, int y)
 296 {
 297    nir_const_value v;
 298
 299    memset(&v, 0, sizeof(v));
 300    v.i32[0] = x;
 301    v.i32[1] = y;
 302
 303    return nir_build_imm(build, 2, 32, v);
 304 }
 305
 306 static inline nir_ssa_def *
 307 nir_imm_int(nir_builder *build, int x)
 308 {
 309    nir_const_value v;
 310
 311    memset(&v, 0, sizeof(v));
 312    v.i32[0] = x;
 313
 314    return nir_build_imm(build, 1, 32, v);
 315 }
 316
 317 static inline nir_ssa_def *
 318 nir_imm_int64(nir_builder *build, int64_t x)
 319 {
 320    nir_const_value v;
 321
 322    memset(&v, 0, sizeof(v));
 323    v.i64[0] = x;
 324
 325    return nir_build_imm(build, 1, 64, v);
 326 }
 327
 328 static inline nir_ssa_def *
 329 nir_imm_intN_t(nir_builder *build, uint64_t x, unsigned bit_size)
 330 {
 331    nir_const_value v;
 332
 333    assert(bit_size == 64 ||
 334           (int64_t)x >> bit_size == 0 ||
 335           (int64_t)x >> bit_size == -1);
 336
 337    memset(&v, 0, sizeof(v));
 338    assert(bit_size <= 64);
 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_build_alu(nir_builder *build, nir_op op, nir_ssa_def *src0,
 360               nir_ssa_def *src1, nir_ssa_def *src2, nir_ssa_def *src3)
 361 {
 362    const nir_op_info *op_info = &nir_op_infos[op];
 363    nir_alu_instr *instr = nir_alu_instr_create(build->shader, op);
 364    if (!instr)
 365       return NULL;
 366
 367    instr->exact = build->exact;
 368
 369    instr->src[0].src = nir_src_for_ssa(src0);
 370    if (src1)
 371       instr->src[1].src = nir_src_for_ssa(src1);
 372    if (src2)
 373       instr->src[2].src = nir_src_for_ssa(src2);
 374    if (src3)
 375       instr->src[3].src = nir_src_for_ssa(src3);
 376
 377    /* Guess the number of components the destination temporary should have
 378     * based on our input sizes, if it's not fixed for the op.
 379     */
 380    unsigned num_components = op_info->output_size;
 381    if (num_components == 0) {
 382       for (unsigned i = 0; i < op_info->num_inputs; i++) {
 383          if (op_info->input_sizes[i] == 0)
 384             num_components = MAX2(num_components,
 385                                   instr->src[i].src.ssa->num_components);
 386       }
 387    }
 388    assert(num_components != 0);
 389
 390    /* Figure out the bitwidth based on the source bitwidth if the instruction
 391     * is variable-width.
 392     */
 393    unsigned bit_size = nir_alu_type_get_type_size(op_info->output_type);
 394    if (bit_size == 0) {
 395       for (unsigned i = 0; i < op_info->num_inputs; i++) {
 396          unsigned src_bit_size = instr->src[i].src.ssa->bit_size;
 397          if (nir_alu_type_get_type_size(op_info->input_types[i]) == 0) {
 398             if (bit_size)
 399                assert(src_bit_size == bit_size);
 400             else
 401                bit_size = src_bit_size;
 402          } else {
 403             assert(src_bit_size ==
 404                nir_alu_type_get_type_size(op_info->input_types[i]));
 405          }
 406       }
 407    }
 408
 409    /* When in doubt, assume 32. */
 410    if (bit_size == 0)
 411       bit_size = 32;
 412
 413    /* Make sure we don't swizzle from outside of our source vector (like if a
 414     * scalar value was passed into a multiply with a vector).
 415     */
 416    for (unsigned i = 0; i < op_info->num_inputs; i++) {
 417       for (unsigned j = instr->src[i].src.ssa->num_components;
 418            j < NIR_MAX_VEC_COMPONENTS; j++) {
 419          instr->src[i].swizzle[j] = instr->src[i].src.ssa->num_components - 1;
 420       }
 421    }
 422
 423    nir_ssa_dest_init(&instr->instr, &instr->dest.dest, num_components,
 424                      bit_size, NULL);
 425    instr->dest.write_mask = (1 << num_components) - 1;
 426
 427    nir_builder_instr_insert(build, &instr->instr);
 428
 429    return &instr->dest.dest.ssa;
 430 }
 431
 432 #include "nir_builder_opcodes.h"
 433
 434 static inline nir_ssa_def *
 435 nir_vec(nir_builder *build, nir_ssa_def **comp, unsigned num_components)
 436 {
 437    switch (num_components) {
 438    case 4:
 439       return nir_vec4(build, comp[0], comp[1], comp[2], comp[3]);
 440    case 3:
 441       return nir_vec3(build, comp[0], comp[1], comp[2]);
 442    case 2:
 443       return nir_vec2(build, comp[0], comp[1]);
 444    case 1:
 445       return comp[0];
 446    default:
 447       unreachable("bad component count");
 448       return NULL;
 449    }
 450 }
 451
 452 /**
 453  * Similar to nir_fmov, but takes a nir_alu_src instead of a nir_ssa_def.
 454  */
 455 static inline nir_ssa_def *
 456 nir_fmov_alu(nir_builder *build, nir_alu_src src, unsigned num_components)
 457 {
 458    nir_alu_instr *mov = nir_alu_instr_create(build->shader, nir_op_fmov);
 459    nir_ssa_dest_init(&mov->instr, &mov->dest.dest, num_components,
 460                      nir_src_bit_size(src.src), NULL);
 461    mov->exact = build->exact;
 462    mov->dest.write_mask = (1 << num_components) - 1;
 463    mov->src[0] = src;
 464    nir_builder_instr_insert(build, &mov->instr);
 465
 466    return &mov->dest.dest.ssa;
 467 }
 468
 469 static inline nir_ssa_def *
 470 nir_imov_alu(nir_builder *build, nir_alu_src src, unsigned num_components)
 471 {
 472    nir_alu_instr *mov = nir_alu_instr_create(build->shader, nir_op_imov);
 473    nir_ssa_dest_init(&mov->instr, &mov->dest.dest, num_components,
 474                      nir_src_bit_size(src.src), NULL);
 475    mov->exact = build->exact;
 476    mov->dest.write_mask = (1 << num_components) - 1;
 477    mov->src[0] = src;
 478    nir_builder_instr_insert(build, &mov->instr);
 479
 480    return &mov->dest.dest.ssa;
 481 }
 482
 483 /**
 484  * Construct an fmov or imov that reswizzles the source's components.
 485  */
 486 static inline nir_ssa_def *
 487 nir_swizzle(nir_builder *build, nir_ssa_def *src, const unsigned *swiz,
 488             unsigned num_components, bool use_fmov)
 489 {
 490    assert(num_components <= NIR_MAX_VEC_COMPONENTS);
 491    nir_alu_src alu_src = { NIR_SRC_INIT };
 492    alu_src.src = nir_src_for_ssa(src);
 493    for (unsigned i = 0; i < num_components && i < NIR_MAX_VEC_COMPONENTS; i++)
 494       alu_src.swizzle[i] = swiz[i];
 495
 496    return use_fmov ? nir_fmov_alu(build, alu_src, num_components) :
 497                      nir_imov_alu(build, alu_src, num_components);
 498 }
 499
 500 /* Selects the right fdot given the number of components in each source. */
 501 static inline nir_ssa_def *
 502 nir_fdot(nir_builder *build, nir_ssa_def *src0, nir_ssa_def *src1)
 503 {
 504    assert(src0->num_components == src1->num_components);
 505    switch (src0->num_components) {
 506    case 1: return nir_fmul(build, src0, src1);
 507    case 2: return nir_fdot2(build, src0, src1);
 508    case 3: return nir_fdot3(build, src0, src1);
 509    case 4: return nir_fdot4(build, src0, src1);
 510    default:
 511       unreachable("bad component size");
 512    }
 513
 514    return NULL;
 515 }
 516
 517 static inline nir_ssa_def *
 518 nir_bany_inequal(nir_builder *b, nir_ssa_def *src0, nir_ssa_def *src1)
 519 {
 520    switch (src0->num_components) {
 521    case 1: return nir_ine(b, src0, src1);
 522    case 2: return nir_bany_inequal2(b, src0, src1);
 523    case 3: return nir_bany_inequal3(b, src0, src1);
 524    case 4: return nir_bany_inequal4(b, src0, src1);
 525    default:
 526       unreachable("bad component size");
 527    }
 528 }
 529
 530 static inline nir_ssa_def *
 531 nir_bany(nir_builder *b, nir_ssa_def *src)
 532 {
 533    return nir_bany_inequal(b, src, nir_imm_false(b));
 534 }
 535
 536 static inline nir_ssa_def *
 537 nir_channel(nir_builder *b, nir_ssa_def *def, unsigned c)
 538 {
 539    return nir_swizzle(b, def, &c, 1, false);
 540 }
 541
 542 static inline nir_ssa_def *
 543 nir_channels(nir_builder *b, nir_ssa_def *def, nir_component_mask_t mask)
 544 {
 545    unsigned num_channels = 0, swizzle[NIR_MAX_VEC_COMPONENTS] = { 0 };
 546
 547    for (unsigned i = 0; i < NIR_MAX_VEC_COMPONENTS; i++) {
 548       if ((mask & (1 << i)) == 0)
 549          continue;
 550       swizzle[num_channels++] = i;
 551    }
 552
 553    return nir_swizzle(b, def, swizzle, num_channels, false);
 554 }
 555
 556 static inline nir_ssa_def *
 557 nir_iadd_imm(nir_builder *build, nir_ssa_def *x, uint64_t y)
 558 {
 559    return nir_iadd(build, x, nir_imm_intN_t(build, y, x->bit_size));
 560 }
 561
 562 static inline nir_ssa_def *
 563 nir_imul_imm(nir_builder *build, nir_ssa_def *x, uint64_t y)
 564 {
 565    return nir_imul(build, x, nir_imm_intN_t(build, y, x->bit_size));
 566 }
 567
 568 static inline nir_ssa_def *
 569 nir_pack_bits(nir_builder *b, nir_ssa_def *src, unsigned dest_bit_size)
 570 {
 571    assert(src->num_components * src->bit_size == dest_bit_size);
 572
 573    switch (dest_bit_size) {
 574    case 64:
 575       switch (src->bit_size) {
 576       case 32: return nir_pack_64_2x32(b, src);
 577       case 16: return nir_pack_64_4x16(b, src);
 578       default: break;
 579       }
 580       break;
 581
 582    case 32:
 583       if (src->bit_size == 16)
 584          return nir_pack_32_2x16(b, src);
 585       break;
 586
 587    default:
 588       break;
 589    }
 590
 591    /* If we got here, we have no dedicated unpack opcode. */
 592    nir_ssa_def *dest = nir_imm_intN_t(b, 0, dest_bit_size);
 593    for (unsigned i = 0; i < src->num_components; i++) {
 594       nir_ssa_def *val;
 595       switch (dest_bit_size) {
 596       case 64: val = nir_u2u64(b, nir_channel(b, src, i));  break;
 597       case 32: val = nir_u2u32(b, nir_channel(b, src, i));  break;
 598       case 16: val = nir_u2u16(b, nir_channel(b, src, i));  break;
 599       default: unreachable("Invalid bit size");
 600       }
 601       val = nir_ishl(b, val, nir_imm_int(b, i * src->bit_size));
 602       dest = nir_ior(b, dest, val);
 603    }
 604    return dest;
 605 }
 606
 607 static inline nir_ssa_def *
 608 nir_unpack_bits(nir_builder *b, nir_ssa_def *src, unsigned dest_bit_size)
 609 {
 610    assert(src->num_components == 1);
 611    assert(src->bit_size > dest_bit_size);
 612    const unsigned dest_num_components = src->bit_size / dest_bit_size;
 613    assert(dest_num_components <= NIR_MAX_VEC_COMPONENTS);
 614
 615    switch (src->bit_size) {
 616    case 64:
 617       switch (dest_bit_size) {
 618       case 32: return nir_unpack_64_2x32(b, src);
 619       case 16: return nir_unpack_64_4x16(b, src);
 620       default: break;
 621       }
 622       break;
 623
 624    case 32:
 625       if (dest_bit_size == 16)
 626          return nir_unpack_32_2x16(b, src);
 627       break;
 628
 629    default:
 630       break;
 631    }
 632
 633    /* If we got here, we have no dedicated unpack opcode. */
 634    nir_ssa_def *dest_comps[NIR_MAX_VEC_COMPONENTS];
 635    for (unsigned i = 0; i < dest_num_components; i++) {
 636       nir_ssa_def *val = nir_ushr(b, src, nir_imm_int(b, i * dest_bit_size));
 637       switch (dest_bit_size) {
 638       case 32: dest_comps[i] = nir_u2u32(b, val);  break;
 639       case 16: dest_comps[i] = nir_u2u16(b, val);  break;
 640       case 8:  dest_comps[i] = nir_u2u8(b, val);   break;
 641       default: unreachable("Invalid bit size");
 642       }
 643    }
 644    return nir_vec(b, dest_comps, dest_num_components);
 645 }
 646
 647 static inline nir_ssa_def *
 648 nir_bitcast_vector(nir_builder *b, nir_ssa_def *src, unsigned dest_bit_size)
 649 {
 650    assert((src->bit_size * src->num_components) % dest_bit_size == 0);
 651    const unsigned dest_num_components =
 652       (src->bit_size * src->num_components) / dest_bit_size;
 653    assert(dest_num_components <= NIR_MAX_VEC_COMPONENTS);
 654
 655    if (src->bit_size > dest_bit_size) {
 656       assert(src->bit_size % dest_bit_size == 0);
 657       if (src->num_components == 1) {
 658          return nir_unpack_bits(b, src, dest_bit_size);
 659       } else {
 660          const unsigned divisor = src->bit_size / dest_bit_size;
 661          assert(src->num_components * divisor == dest_num_components);
 662          nir_ssa_def *dest[NIR_MAX_VEC_COMPONENTS];
 663          for (unsigned i = 0; i < src->num_components; i++) {
 664             nir_ssa_def *unpacked =
 665                nir_unpack_bits(b, nir_channel(b, src, i), dest_bit_size);
 666             assert(unpacked->num_components == divisor);
 667             for (unsigned j = 0; j < divisor; j++)
 668                dest[i * divisor + j] = nir_channel(b, unpacked, j);
 669          }
 670          return nir_vec(b, dest, dest_num_components);
 671       }
 672    } else if (src->bit_size < dest_bit_size) {
 673       assert(dest_bit_size % src->bit_size == 0);
 674       if (dest_num_components == 1) {
 675          return nir_pack_bits(b, src, dest_bit_size);
 676       } else {
 677          const unsigned divisor = dest_bit_size / src->bit_size;
 678          assert(src->num_components == dest_num_components * divisor);
 679          nir_ssa_def *dest[NIR_MAX_VEC_COMPONENTS];
 680          for (unsigned i = 0; i < dest_num_components; i++) {
 681             nir_component_mask_t src_mask =
 682                ((1 << divisor) - 1) << (i * divisor);
 683             dest[i] = nir_pack_bits(b, nir_channels(b, src, src_mask),
 684                                        dest_bit_size);
 685          }
 686          return nir_vec(b, dest, dest_num_components);
 687       }
 688    } else {
 689       assert(src->bit_size == dest_bit_size);
 690       return src;
 691    }
 692 }
 693
 694 /**
 695  * Turns a nir_src into a nir_ssa_def * so it can be passed to
 696  * nir_build_alu()-based builder calls.
 697  *
 698  * See nir_ssa_for_alu_src() for alu instructions.
 699  */
 700 static inline nir_ssa_def *
 701 nir_ssa_for_src(nir_builder *build, nir_src src, int num_components)
 702 {
 703    if (src.is_ssa && src.ssa->num_components == num_components)
 704       return src.ssa;
 705
 706    nir_alu_src alu = { NIR_SRC_INIT };
 707    alu.src = src;
 708    for (int j = 0; j < 4; j++)
 709       alu.swizzle[j] = j;
 710
 711    return nir_imov_alu(build, alu, num_components);
 712 }
 713
 714 /**
 715  * Similar to nir_ssa_for_src(), but for alu srcs, respecting the
 716  * nir_alu_src's swizzle.
 717  */
 718 static inline nir_ssa_def *
 719 nir_ssa_for_alu_src(nir_builder *build, nir_alu_instr *instr, unsigned srcn)
 720 {
 721    static uint8_t trivial_swizzle[NIR_MAX_VEC_COMPONENTS];
 722    for (int i = 0; i < NIR_MAX_VEC_COMPONENTS; ++i)
 723       trivial_swizzle[i] = i;
 724    nir_alu_src *src = &instr->src[srcn];
 725    unsigned num_components = nir_ssa_alu_instr_src_components(instr, srcn);
 726
 727    if (src->src.is_ssa && (src->src.ssa->num_components == num_components) &&
 728        !src->abs && !src->negate &&
 729        (memcmp(src->swizzle, trivial_swizzle, num_components) == 0))
 730       return src->src.ssa;
 731
 732    return nir_imov_alu(build, *src, num_components);
 733 }
 734
 735 static inline nir_deref_instr *
 736 nir_build_deref_var(nir_builder *build, nir_variable *var)
 737 {
 738    nir_deref_instr *deref =
 739       nir_deref_instr_create(build->shader, nir_deref_type_var);
 740
 741    deref->mode = var->data.mode;
 742    deref->type = var->type;
 743    deref->var = var;
 744
 745    nir_ssa_dest_init(&deref->instr, &deref->dest, 1, 32, NULL);
 746
 747    nir_builder_instr_insert(build, &deref->instr);
 748
 749    return deref;
 750 }
 751
 752 static inline nir_deref_instr *
 753 nir_build_deref_array(nir_builder *build, nir_deref_instr *parent,
 754                       nir_ssa_def *index)
 755 {
 756    assert(glsl_type_is_array(parent->type) ||
 757           glsl_type_is_matrix(parent->type) ||
 758           glsl_type_is_vector(parent->type));
 759
 760    nir_deref_instr *deref =
 761       nir_deref_instr_create(build->shader, nir_deref_type_array);
 762
 763    deref->mode = parent->mode;
 764    deref->type = glsl_get_array_element(parent->type);
 765    deref->parent = nir_src_for_ssa(&parent->dest.ssa);
 766    deref->arr.index = nir_src_for_ssa(index);
 767
 768    nir_ssa_dest_init(&deref->instr, &deref->dest,
 769                      parent->dest.ssa.num_components,
 770                      parent->dest.ssa.bit_size, NULL);
 771
 772    nir_builder_instr_insert(build, &deref->instr);
 773
 774    return deref;
 775 }
 776
 777 static inline nir_deref_instr *
 778 nir_build_deref_array_wildcard(nir_builder *build, nir_deref_instr *parent)
 779 {
 780    assert(glsl_type_is_array(parent->type) ||
 781           glsl_type_is_matrix(parent->type));
 782
 783    nir_deref_instr *deref =
 784       nir_deref_instr_create(build->shader, nir_deref_type_array_wildcard);
 785
 786    deref->mode = parent->mode;
 787    deref->type = glsl_get_array_element(parent->type);
 788    deref->parent = nir_src_for_ssa(&parent->dest.ssa);
 789
 790    nir_ssa_dest_init(&deref->instr, &deref->dest,
 791                      parent->dest.ssa.num_components,
 792                      parent->dest.ssa.bit_size, NULL);
 793
 794    nir_builder_instr_insert(build, &deref->instr);
 795
 796    return deref;
 797 }
 798
 799 static inline nir_deref_instr *
 800 nir_build_deref_struct(nir_builder *build, nir_deref_instr *parent,
 801                        unsigned index)
 802 {
 803    assert(glsl_type_is_struct(parent->type));
 804
 805    nir_deref_instr *deref =
 806       nir_deref_instr_create(build->shader, nir_deref_type_struct);
 807
 808    deref->mode = parent->mode;
 809    deref->type = glsl_get_struct_field(parent->type, index);
 810    deref->parent = nir_src_for_ssa(&parent->dest.ssa);
 811    deref->strct.index = index;
 812
 813    nir_ssa_dest_init(&deref->instr, &deref->dest,
 814                      parent->dest.ssa.num_components,
 815                      parent->dest.ssa.bit_size, NULL);
 816
 817    nir_builder_instr_insert(build, &deref->instr);
 818
 819    return deref;
 820 }
 821
 822 static inline nir_deref_instr *
 823 nir_build_deref_cast(nir_builder *build, nir_ssa_def *parent,
 824                      nir_variable_mode mode, const struct glsl_type *type)
 825 {
 826    nir_deref_instr *deref =
 827       nir_deref_instr_create(build->shader, nir_deref_type_cast);
 828
 829    deref->mode = mode;
 830    deref->type = type;
 831    deref->parent = nir_src_for_ssa(parent);
 832
 833    nir_ssa_dest_init(&deref->instr, &deref->dest,
 834                      parent->num_components, parent->bit_size, NULL);
 835
 836    nir_builder_instr_insert(build, &deref->instr);
 837
 838    return deref;
 839 }
 840
 841 /** Returns a deref that follows another but starting from the given parent
 842  *
 843  * The new deref will be the same type and take the same array or struct index
 844  * as the leader deref but it may have a different parent.  This is very
 845  * useful for walking deref paths.
 846  */
 847 static inline nir_deref_instr *
 848 nir_build_deref_follower(nir_builder *b, nir_deref_instr *parent,
 849                          nir_deref_instr *leader)
 850 {
 851    /* If the derefs would have the same parent, don't make a new one */
 852    assert(leader->parent.is_ssa);
 853    if (leader->parent.ssa == &parent->dest.ssa)
 854       return leader;
 855
 856    UNUSED nir_deref_instr *leader_parent = nir_src_as_deref(leader->parent);
 857
 858    switch (leader->deref_type) {
 859    case nir_deref_type_var:
 860       unreachable("A var dereference cannot have a parent");
 861       break;
 862
 863    case nir_deref_type_array:
 864    case nir_deref_type_array_wildcard:
 865       assert(glsl_type_is_matrix(parent->type) ||
 866              glsl_type_is_array(parent->type));
 867       assert(glsl_get_length(parent->type) ==
 868              glsl_get_length(leader_parent->type));
 869
 870       if (leader->deref_type == nir_deref_type_array) {
 871          assert(leader->arr.index.is_ssa);
 872          return nir_build_deref_array(b, parent, leader->arr.index.ssa);
 873       } else {
 874          return nir_build_deref_array_wildcard(b, parent);
 875       }
 876
 877    case nir_deref_type_struct:
 878       assert(glsl_type_is_struct(parent->type));
 879       assert(glsl_get_length(parent->type) ==
 880              glsl_get_length(leader_parent->type));
 881
 882       return nir_build_deref_struct(b, parent, leader->strct.index);
 883
 884    default:
 885       unreachable("Invalid deref instruction type");
 886    }
 887 }
 888
 889 static inline nir_ssa_def *
 890 nir_load_reg(nir_builder *build, nir_register *reg)
 891 {
 892    return nir_ssa_for_src(build, nir_src_for_reg(reg), reg->num_components);
 893 }
 894
 895 static inline nir_ssa_def *
 896 nir_load_deref(nir_builder *build, nir_deref_instr *deref)
 897 {
 898    nir_intrinsic_instr *load =
 899       nir_intrinsic_instr_create(build->shader, nir_intrinsic_load_deref);
 900    load->num_components = glsl_get_vector_elements(deref->type);
 901    load->src[0] = nir_src_for_ssa(&deref->dest.ssa);
 902    nir_ssa_dest_init(&load->instr, &load->dest, load->num_components,
 903                      glsl_get_bit_size(deref->type), NULL);
 904    nir_builder_instr_insert(build, &load->instr);
 905    return &load->dest.ssa;
 906 }
 907
 908 static inline void
 909 nir_store_deref(nir_builder *build, nir_deref_instr *deref,
 910                 nir_ssa_def *value, unsigned writemask)
 911 {
 912    nir_intrinsic_instr *store =
 913       nir_intrinsic_instr_create(build->shader, nir_intrinsic_store_deref);
 914    store->num_components = glsl_get_vector_elements(deref->type);
 915    store->src[0] = nir_src_for_ssa(&deref->dest.ssa);
 916    store->src[1] = nir_src_for_ssa(value);
 917    nir_intrinsic_set_write_mask(store,
 918                                 writemask & ((1 << store->num_components) - 1));
 919    nir_builder_instr_insert(build, &store->instr);
 920 }
 921
 922 static inline void
 923 nir_copy_deref(nir_builder *build, nir_deref_instr *dest, nir_deref_instr *src)
 924 {
 925    nir_intrinsic_instr *copy =
 926       nir_intrinsic_instr_create(build->shader, nir_intrinsic_copy_deref);
 927    copy->src[0] = nir_src_for_ssa(&dest->dest.ssa);
 928    copy->src[1] = nir_src_for_ssa(&src->dest.ssa);
 929    nir_builder_instr_insert(build, &copy->instr);
 930 }
 931
 932 static inline nir_ssa_def *
 933 nir_load_var(nir_builder *build, nir_variable *var)
 934 {
 935    return nir_load_deref(build, nir_build_deref_var(build, var));
 936 }
 937
 938 static inline void
 939 nir_store_var(nir_builder *build, nir_variable *var, nir_ssa_def *value,
 940               unsigned writemask)
 941 {
 942    nir_store_deref(build, nir_build_deref_var(build, var), value, writemask);
 943 }
 944
 945 static inline void
 946 nir_copy_var(nir_builder *build, nir_variable *dest, nir_variable *src)
 947 {
 948    nir_copy_deref(build, nir_build_deref_var(build, dest),
 949                          nir_build_deref_var(build, src));
 950 }
 951
 952 static inline nir_ssa_def *
 953 nir_load_param(nir_builder *build, uint32_t param_idx)
 954 {
 955    assert(param_idx < build->impl->function->num_params);
 956    nir_parameter *param = &build->impl->function->params[param_idx];
 957
 958    nir_intrinsic_instr *load =
 959       nir_intrinsic_instr_create(build->shader, nir_intrinsic_load_param);
 960    nir_intrinsic_set_param_idx(load, param_idx);
 961    load->num_components = param->num_components;
 962    nir_ssa_dest_init(&load->instr, &load->dest,
 963                      param->num_components, param->bit_size, NULL);
 964    nir_builder_instr_insert(build, &load->instr);
 965    return &load->dest.ssa;
 966 }
 967
 968 #include "nir_builder_opcodes.h"
 969
 970 static inline nir_ssa_def *
 971 nir_load_barycentric(nir_builder *build, nir_intrinsic_op op,
 972                      unsigned interp_mode)
 973 {
 974    nir_intrinsic_instr *bary = nir_intrinsic_instr_create(build->shader, op);
 975    nir_ssa_dest_init(&bary->instr, &bary->dest, 2, 32, NULL);
 976    nir_intrinsic_set_interp_mode(bary, interp_mode);
 977    nir_builder_instr_insert(build, &bary->instr);
 978    return &bary->dest.ssa;
 979 }
 980
 981 static inline void
 982 nir_jump(nir_builder *build, nir_jump_type jump_type)
 983 {
 984    nir_jump_instr *jump = nir_jump_instr_create(build->shader, jump_type);
 985    nir_builder_instr_insert(build, &jump->instr);
 986 }
 987
 988 static inline nir_ssa_def *
 989 nir_compare_func(nir_builder *b, enum compare_func func,
 990                  nir_ssa_def *src0, nir_ssa_def *src1)
 991 {
 992    switch (func) {
 993    case COMPARE_FUNC_NEVER:
 994       return nir_imm_int(b, 0);
 995    case COMPARE_FUNC_ALWAYS:
 996       return nir_imm_int(b, ~0);
 997    case COMPARE_FUNC_EQUAL:
 998       return nir_feq(b, src0, src1);
 999    case COMPARE_FUNC_NOTEQUAL:
1000       return nir_fne(b, src0, src1);
1001    case COMPARE_FUNC_GREATER:
1002       return nir_flt(b, src1, src0);
1003    case COMPARE_FUNC_GEQUAL:
1004       return nir_fge(b, src0, src1);
1005    case COMPARE_FUNC_LESS:
1006       return nir_flt(b, src0, src1);
1007    case COMPARE_FUNC_LEQUAL:
1008       return nir_fge(b, src1, src0);
1009    }
1010    unreachable("bad compare func");
1011 }
1012
1013 #endif /* NIR_BUILDER_H */