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