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
  29 struct exec_list;
  30
  31 typedef struct nir_builder {
  32    nir_cursor cursor;
  33
  34    /* Whether new ALU instructions will be marked "exact" */
  35    bool exact;
  36
  37    nir_shader *shader;
  38    nir_function_impl *impl;
  39 } nir_builder;
  40
  41 static inline void
  42 nir_builder_init(nir_builder *build, nir_function_impl *impl)
  43 {
  44    memset(build, 0, sizeof(*build));
  45    build->exact = false;
  46    build->impl = impl;
  47    build->shader = impl->function->shader;
  48 }
  49
  50 static inline void
  51 nir_builder_init_simple_shader(nir_builder *build, void *mem_ctx,
  52                                gl_shader_stage stage,
  53                                const nir_shader_compiler_options *options)
  54 {
  55    build->shader = nir_shader_create(mem_ctx, stage, options);
  56    nir_function *func = nir_function_create(build->shader, "main");
  57    build->exact = false;
  58    build->impl = nir_function_impl_create(func);
  59    build->cursor = nir_after_cf_list(&build->impl->body);
  60 }
  61
  62 static inline void
  63 nir_builder_instr_insert(nir_builder *build, nir_instr *instr)
  64 {
  65    nir_instr_insert(build->cursor, instr);
  66
  67    /* Move the cursor forward. */
  68    build->cursor = nir_after_instr(instr);
  69 }
  70
  71 static inline void
  72 nir_builder_cf_insert(nir_builder *build, nir_cf_node *cf)
  73 {
  74    nir_cf_node_insert(build->cursor, cf);
  75 }
  76
  77 static inline nir_ssa_def *
  78 nir_ssa_undef(nir_builder *build, unsigned num_components, unsigned bit_size)
  79 {
  80    nir_ssa_undef_instr *undef =
  81       nir_ssa_undef_instr_create(build->shader, num_components);
  82    undef->def.bit_size = bit_size;
  83    if (!undef)
  84       return NULL;
  85
  86    nir_instr_insert(nir_before_block(nir_start_block(build->impl)),
  87                     &undef->instr);
  88
  89    return &undef->def;
  90 }
  91
  92 static inline nir_ssa_def *
  93 nir_build_imm(nir_builder *build, unsigned num_components, nir_const_value value)
  94 {
  95    nir_load_const_instr *load_const =
  96       nir_load_const_instr_create(build->shader, num_components);
  97    if (!load_const)
  98       return NULL;
  99
 100    load_const->value = value;
 101
 102    nir_builder_instr_insert(build, &load_const->instr);
 103
 104    return &load_const->def;
 105 }
 106
 107 static inline nir_ssa_def *
 108 nir_imm_float(nir_builder *build, float x)
 109 {
 110    nir_const_value v;
 111
 112    memset(&v, 0, sizeof(v));
 113    v.f32[0] = x;
 114
 115    return nir_build_imm(build, 1, v);
 116 }
 117
 118 static inline nir_ssa_def *
 119 nir_imm_vec4(nir_builder *build, float x, float y, float z, float w)
 120 {
 121    nir_const_value v;
 122
 123    memset(&v, 0, sizeof(v));
 124    v.f32[0] = x;
 125    v.f32[1] = y;
 126    v.f32[2] = z;
 127    v.f32[3] = w;
 128
 129    return nir_build_imm(build, 4, v);
 130 }
 131
 132 static inline nir_ssa_def *
 133 nir_imm_int(nir_builder *build, int x)
 134 {
 135    nir_const_value v;
 136
 137    memset(&v, 0, sizeof(v));
 138    v.i32[0] = x;
 139
 140    return nir_build_imm(build, 1, v);
 141 }
 142
 143 static inline nir_ssa_def *
 144 nir_imm_ivec4(nir_builder *build, int x, int y, int z, int w)
 145 {
 146    nir_const_value v;
 147
 148    memset(&v, 0, sizeof(v));
 149    v.i32[0] = x;
 150    v.i32[1] = y;
 151    v.i32[2] = z;
 152    v.i32[3] = w;
 153
 154    return nir_build_imm(build, 4, v);
 155 }
 156
 157 static inline nir_ssa_def *
 158 nir_build_alu(nir_builder *build, nir_op op, nir_ssa_def *src0,
 159               nir_ssa_def *src1, nir_ssa_def *src2, nir_ssa_def *src3)
 160 {
 161    const nir_op_info *op_info = &nir_op_infos[op];
 162    nir_alu_instr *instr = nir_alu_instr_create(build->shader, op);
 163    if (!instr)
 164       return NULL;
 165
 166    instr->exact = build->exact;
 167
 168    instr->src[0].src = nir_src_for_ssa(src0);
 169    if (src1)
 170       instr->src[1].src = nir_src_for_ssa(src1);
 171    if (src2)
 172       instr->src[2].src = nir_src_for_ssa(src2);
 173    if (src3)
 174       instr->src[3].src = nir_src_for_ssa(src3);
 175
 176    /* Guess the number of components the destination temporary should have
 177     * based on our input sizes, if it's not fixed for the op.
 178     */
 179    unsigned num_components = op_info->output_size;
 180    if (num_components == 0) {
 181       for (unsigned i = 0; i < op_info->num_inputs; i++) {
 182          if (op_info->input_sizes[i] == 0)
 183             num_components = MAX2(num_components,
 184                                   instr->src[i].src.ssa->num_components);
 185       }
 186    }
 187    assert(num_components != 0);
 188
 189    /* Figure out the bitwidth based on the source bitwidth if the instruction
 190     * is variable-width.
 191     */
 192    unsigned bit_size = nir_alu_type_get_type_size(op_info->output_type);
 193    if (bit_size == 0) {
 194       for (unsigned i = 0; i < op_info->num_inputs; i++) {
 195          unsigned src_bit_size = instr->src[i].src.ssa->bit_size;
 196          if (nir_alu_type_get_type_size(op_info->input_types[i]) == 0) {
 197             if (bit_size)
 198                assert(src_bit_size == bit_size);
 199             else
 200                bit_size = src_bit_size;
 201          } else {
 202             assert(src_bit_size ==
 203                nir_alu_type_get_type_size(op_info->input_types[i]));
 204          }
 205       }
 206    }
 207
 208    /* Make sure we don't swizzle from outside of our source vector (like if a
 209     * scalar value was passed into a multiply with a vector).
 210     */
 211    for (unsigned i = 0; i < op_info->num_inputs; i++) {
 212       for (unsigned j = instr->src[i].src.ssa->num_components; j < 4; j++) {
 213          instr->src[i].swizzle[j] = instr->src[i].src.ssa->num_components - 1;
 214       }
 215    }
 216
 217    nir_ssa_dest_init(&instr->instr, &instr->dest.dest, num_components,
 218                      bit_size, NULL);
 219    instr->dest.write_mask = (1 << num_components) - 1;
 220
 221    nir_builder_instr_insert(build, &instr->instr);
 222
 223    return &instr->dest.dest.ssa;
 224 }
 225
 226 #define ALU1(op)                                                          \
 227 static inline nir_ssa_def *                                               \
 228 nir_##op(nir_builder *build, nir_ssa_def *src0)                           \
 229 {                                                                         \
 230    return nir_build_alu(build, nir_op_##op, src0, NULL, NULL, NULL);      \
 231 }
 232
 233 #define ALU2(op)                                                          \
 234 static inline nir_ssa_def *                                               \
 235 nir_##op(nir_builder *build, nir_ssa_def *src0, nir_ssa_def *src1)        \
 236 {                                                                         \
 237    return nir_build_alu(build, nir_op_##op, src0, src1, NULL, NULL);      \
 238 }
 239
 240 #define ALU3(op)                                                          \
 241 static inline nir_ssa_def *                                               \
 242 nir_##op(nir_builder *build, nir_ssa_def *src0,                           \
 243          nir_ssa_def *src1, nir_ssa_def *src2)                            \
 244 {                                                                         \
 245    return nir_build_alu(build, nir_op_##op, src0, src1, src2, NULL);      \
 246 }
 247
 248 #define ALU4(op)                                                          \
 249 static inline nir_ssa_def *                                               \
 250 nir_##op(nir_builder *build, nir_ssa_def *src0,                           \
 251          nir_ssa_def *src1, nir_ssa_def *src2, nir_ssa_def *src3)         \
 252 {                                                                         \
 253    return nir_build_alu(build, nir_op_##op, src0, src1, src2, src3);      \
 254 }
 255
 256 #include "nir_builder_opcodes.h"
 257
 258 static inline nir_ssa_def *
 259 nir_vec(nir_builder *build, nir_ssa_def **comp, unsigned num_components)
 260 {
 261    switch (num_components) {
 262    case 4:
 263       return nir_vec4(build, comp[0], comp[1], comp[2], comp[3]);
 264    case 3:
 265       return nir_vec3(build, comp[0], comp[1], comp[2]);
 266    case 2:
 267       return nir_vec2(build, comp[0], comp[1]);
 268    case 1:
 269       return comp[0];
 270    default:
 271       unreachable("bad component count");
 272       return NULL;
 273    }
 274 }
 275
 276 /**
 277  * Similar to nir_fmov, but takes a nir_alu_src instead of a nir_ssa_def.
 278  */
 279 static inline nir_ssa_def *
 280 nir_fmov_alu(nir_builder *build, nir_alu_src src, unsigned num_components)
 281 {
 282    nir_alu_instr *mov = nir_alu_instr_create(build->shader, nir_op_fmov);
 283    nir_ssa_dest_init(&mov->instr, &mov->dest.dest, num_components,
 284                      nir_src_bit_size(src.src), NULL);
 285    mov->exact = build->exact;
 286    mov->dest.write_mask = (1 << num_components) - 1;
 287    mov->src[0] = src;
 288    nir_builder_instr_insert(build, &mov->instr);
 289
 290    return &mov->dest.dest.ssa;
 291 }
 292
 293 static inline nir_ssa_def *
 294 nir_imov_alu(nir_builder *build, nir_alu_src src, unsigned num_components)
 295 {
 296    nir_alu_instr *mov = nir_alu_instr_create(build->shader, nir_op_imov);
 297    nir_ssa_dest_init(&mov->instr, &mov->dest.dest, num_components,
 298                      nir_src_bit_size(src.src), NULL);
 299    mov->exact = build->exact;
 300    mov->dest.write_mask = (1 << num_components) - 1;
 301    mov->src[0] = src;
 302    nir_builder_instr_insert(build, &mov->instr);
 303
 304    return &mov->dest.dest.ssa;
 305 }
 306
 307 /**
 308  * Construct an fmov or imov that reswizzles the source's components.
 309  */
 310 static inline nir_ssa_def *
 311 nir_swizzle(nir_builder *build, nir_ssa_def *src, unsigned swiz[4],
 312             unsigned num_components, bool use_fmov)
 313 {
 314    nir_alu_src alu_src = { NIR_SRC_INIT };
 315    alu_src.src = nir_src_for_ssa(src);
 316    for (unsigned i = 0; i < num_components; i++)
 317       alu_src.swizzle[i] = swiz[i];
 318
 319    return use_fmov ? nir_fmov_alu(build, alu_src, num_components) :
 320                      nir_imov_alu(build, alu_src, num_components);
 321 }
 322
 323 /* Selects the right fdot given the number of components in each source. */
 324 static inline nir_ssa_def *
 325 nir_fdot(nir_builder *build, nir_ssa_def *src0, nir_ssa_def *src1)
 326 {
 327    assert(src0->num_components == src1->num_components);
 328    switch (src0->num_components) {
 329    case 1: return nir_fmul(build, src0, src1);
 330    case 2: return nir_fdot2(build, src0, src1);
 331    case 3: return nir_fdot3(build, src0, src1);
 332    case 4: return nir_fdot4(build, src0, src1);
 333    default:
 334       unreachable("bad component size");
 335    }
 336
 337    return NULL;
 338 }
 339
 340 static inline nir_ssa_def *
 341 nir_channel(nir_builder *b, nir_ssa_def *def, unsigned c)
 342 {
 343    unsigned swizzle[4] = {c, c, c, c};
 344    return nir_swizzle(b, def, swizzle, 1, false);
 345 }
 346
 347 /**
 348  * Turns a nir_src into a nir_ssa_def * so it can be passed to
 349  * nir_build_alu()-based builder calls.
 350  *
 351  * See nir_ssa_for_alu_src() for alu instructions.
 352  */
 353 static inline nir_ssa_def *
 354 nir_ssa_for_src(nir_builder *build, nir_src src, int num_components)
 355 {
 356    if (src.is_ssa && src.ssa->num_components == num_components)
 357       return src.ssa;
 358
 359    nir_alu_src alu = { NIR_SRC_INIT };
 360    alu.src = src;
 361    for (int j = 0; j < 4; j++)
 362       alu.swizzle[j] = j;
 363
 364    return nir_imov_alu(build, alu, num_components);
 365 }
 366
 367 /**
 368  * Similar to nir_ssa_for_src(), but for alu src's, respecting the
 369  * nir_alu_src's swizzle.
 370  */
 371 static inline nir_ssa_def *
 372 nir_ssa_for_alu_src(nir_builder *build, nir_alu_instr *instr, unsigned srcn)
 373 {
 374    static uint8_t trivial_swizzle[4] = { 0, 1, 2, 3 };
 375    nir_alu_src *src = &instr->src[srcn];
 376    unsigned num_components = nir_ssa_alu_instr_src_components(instr, srcn);
 377
 378    if (src->src.is_ssa && (src->src.ssa->num_components == num_components) &&
 379        !src->abs && !src->negate &&
 380        (memcmp(src->swizzle, trivial_swizzle, num_components) == 0))
 381       return src->src.ssa;
 382
 383    return nir_imov_alu(build, *src, num_components);
 384 }
 385
 386 static inline nir_ssa_def *
 387 nir_load_var(nir_builder *build, nir_variable *var)
 388 {
 389    const unsigned num_components = glsl_get_vector_elements(var->type);
 390
 391    nir_intrinsic_instr *load =
 392       nir_intrinsic_instr_create(build->shader, nir_intrinsic_load_var);
 393    load->num_components = num_components;
 394    load->variables[0] = nir_deref_var_create(load, var);
 395    nir_ssa_dest_init(&load->instr, &load->dest, num_components,
 396                      glsl_get_bit_size(glsl_get_base_type(var->type)), NULL);
 397    nir_builder_instr_insert(build, &load->instr);
 398    return &load->dest.ssa;
 399 }
 400
 401 static inline void
 402 nir_store_var(nir_builder *build, nir_variable *var, nir_ssa_def *value,
 403               unsigned writemask)
 404 {
 405    const unsigned num_components = glsl_get_vector_elements(var->type);
 406
 407    nir_intrinsic_instr *store =
 408       nir_intrinsic_instr_create(build->shader, nir_intrinsic_store_var);
 409    store->num_components = num_components;
 410    nir_intrinsic_set_write_mask(store, writemask);
 411    store->variables[0] = nir_deref_var_create(store, var);
 412    store->src[0] = nir_src_for_ssa(value);
 413    nir_builder_instr_insert(build, &store->instr);
 414 }
 415
 416 static inline void
 417 nir_store_deref_var(nir_builder *build, nir_deref_var *deref,
 418                     nir_ssa_def *value, unsigned writemask)
 419 {
 420    const unsigned num_components =
 421       glsl_get_vector_elements(nir_deref_tail(&deref->deref)->type);
 422
 423    nir_intrinsic_instr *store =
 424       nir_intrinsic_instr_create(build->shader, nir_intrinsic_store_var);
 425    store->num_components = num_components;
 426    store->const_index[0] = writemask & ((1 << num_components) - 1);
 427    store->variables[0] = nir_deref_as_var(nir_copy_deref(store, &deref->deref));
 428    store->src[0] = nir_src_for_ssa(value);
 429    nir_builder_instr_insert(build, &store->instr);
 430 }
 431
 432 static inline void
 433 nir_copy_deref_var(nir_builder *build, nir_deref_var *dest, nir_deref_var *src)
 434 {
 435    assert(nir_deref_tail(&dest->deref)->type ==
 436           nir_deref_tail(&src->deref)->type);
 437
 438    nir_intrinsic_instr *copy =
 439       nir_intrinsic_instr_create(build->shader, nir_intrinsic_copy_var);
 440    copy->variables[0] = nir_deref_as_var(nir_copy_deref(copy, &dest->deref));
 441    copy->variables[1] = nir_deref_as_var(nir_copy_deref(copy, &src->deref));
 442    nir_builder_instr_insert(build, &copy->instr);
 443 }
 444
 445 static inline void
 446 nir_copy_var(nir_builder *build, nir_variable *dest, nir_variable *src)
 447 {
 448    nir_intrinsic_instr *copy =
 449       nir_intrinsic_instr_create(build->shader, nir_intrinsic_copy_var);
 450    copy->variables[0] = nir_deref_var_create(copy, dest);
 451    copy->variables[1] = nir_deref_var_create(copy, src);
 452    nir_builder_instr_insert(build, &copy->instr);
 453 }
 454
 455 static inline nir_ssa_def *
 456 nir_load_system_value(nir_builder *build, nir_intrinsic_op op, int index)
 457 {
 458    nir_intrinsic_instr *load = nir_intrinsic_instr_create(build->shader, op);
 459    load->num_components = nir_intrinsic_infos[op].dest_components;
 460    load->const_index[0] = index;
 461    nir_ssa_dest_init(&load->instr, &load->dest,
 462                      nir_intrinsic_infos[op].dest_components, 32, NULL);
 463    nir_builder_instr_insert(build, &load->instr);
 464    return &load->dest.ssa;
 465 }
 466
 467 static inline void
 468 nir_jump(nir_builder *build, nir_jump_type jump_type)
 469 {
 470    nir_jump_instr *jump = nir_jump_instr_create(build->shader, jump_type);
 471    nir_builder_instr_insert(build, &jump->instr);
 472 }
 473
 474 #endif /* NIR_BUILDER_H */