src/gallium/auxiliary/gallivm/lp_bld_nir_soa.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2019 Red Hat.
   4  * All Rights Reserved.
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the "Software"),
   8  * to deal in the Software without restriction, including without limitation
   9  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  10  * and/or sell copies of the Software, and to permit persons to whom the
  11  * Software is furnished to do so, subject to the following conditions:
  12  *
  13  * The above copyright notice and this permission notice shall be included
  14  * in all copies or substantial portions of the Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  17  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  19  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  22  * SOFTWARE.
  23  *
  24  **************************************************************************/
  25
  26 #include "lp_bld_nir.h"
  27 #include "lp_bld_init.h"
  28 #include "lp_bld_flow.h"
  29 #include "lp_bld_logic.h"
  30 #include "lp_bld_gather.h"
  31 #include "lp_bld_const.h"
  32 #include "lp_bld_struct.h"
  33 #include "lp_bld_arit.h"
  34 #include "lp_bld_bitarit.h"
  35 #include "lp_bld_coro.h"
  36 #include "lp_bld_printf.h"
  37 #include "util/u_math.h"
  38 /*
  39  * combine the execution mask if there is one with the current mask.
  40  */
  41 static LLVMValueRef
  42 mask_vec(struct lp_build_nir_context *bld_base)
  43 {
  44    struct lp_build_nir_soa_context * bld = (struct lp_build_nir_soa_context *)bld_base;
  45    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
  46    struct lp_exec_mask *exec_mask = &bld->exec_mask;
  47    LLVMValueRef bld_mask = bld->mask ? lp_build_mask_value(bld->mask) : NULL;
  48    if (!exec_mask->has_mask) {
  49       return bld_mask;
  50    }
  51    if (!bld_mask)
  52       return exec_mask->exec_mask;
  53    return LLVMBuildAnd(builder, lp_build_mask_value(bld->mask),
  54                        exec_mask->exec_mask, "");
  55 }
  56
  57 static LLVMValueRef
  58 emit_fetch_64bit(
  59    struct lp_build_nir_context * bld_base,
  60    LLVMValueRef input,
  61    LLVMValueRef input2)
  62 {
  63    struct gallivm_state *gallivm = bld_base->base.gallivm;
  64    LLVMBuilderRef builder = gallivm->builder;
  65    LLVMValueRef res;
  66    int i;
  67    LLVMValueRef shuffles[2 * (LP_MAX_VECTOR_WIDTH/32)];
  68    int len = bld_base->base.type.length * 2;
  69    assert(len <= (2 * (LP_MAX_VECTOR_WIDTH/32)));
  70
  71    for (i = 0; i < bld_base->base.type.length * 2; i+=2) {
  72 #if UTIL_ARCH_LITTLE_ENDIAN
  73       shuffles[i] = lp_build_const_int32(gallivm, i / 2);
  74       shuffles[i + 1] = lp_build_const_int32(gallivm, i / 2 + bld_base->base.type.length);
  75 #else
  76       shuffles[i] = lp_build_const_int32(gallivm, i / 2 + bld_base->base.type.length);
  77       shuffles[i + 1] = lp_build_const_int32(gallivm, i / 2);
  78 #endif
  79    }
  80    res = LLVMBuildShuffleVector(builder, input, input2, LLVMConstVector(shuffles, len), "");
  81
  82    return LLVMBuildBitCast(builder, res, bld_base->dbl_bld.vec_type, "");
  83 }
  84
  85 static void
  86 emit_store_64bit_split(struct lp_build_nir_context *bld_base,
  87                        LLVMValueRef value,
  88                        LLVMValueRef split_values[2])
  89 {
  90    struct gallivm_state *gallivm = bld_base->base.gallivm;
  91    LLVMBuilderRef builder = gallivm->builder;
  92    unsigned i;
  93    LLVMValueRef shuffles[LP_MAX_VECTOR_WIDTH/32];
  94    LLVMValueRef shuffles2[LP_MAX_VECTOR_WIDTH/32];
  95    int len = bld_base->base.type.length * 2;
  96
  97    value = LLVMBuildBitCast(gallivm->builder, value, LLVMVectorType(LLVMFloatTypeInContext(gallivm->context), len), "");
  98    for (i = 0; i < bld_base->base.type.length; i++) {
  99 #if UTIL_ARCH_LITTLE_ENDIAN
 100       shuffles[i] = lp_build_const_int32(gallivm, i * 2);
 101       shuffles2[i] = lp_build_const_int32(gallivm, (i * 2) + 1);
 102 #else
 103       shuffles[i] = lp_build_const_int32(gallivm, i * 2 + 1);
 104       shuffles2[i] = lp_build_const_int32(gallivm, i * 2);
 105 #endif
 106    }
 107
 108    split_values[0] = LLVMBuildShuffleVector(builder, value,
 109                                  LLVMGetUndef(LLVMTypeOf(value)),
 110                                  LLVMConstVector(shuffles,
 111                                                  bld_base->base.type.length),
 112                                  "");
 113    split_values[1] = LLVMBuildShuffleVector(builder, value,
 114                                   LLVMGetUndef(LLVMTypeOf(value)),
 115                                   LLVMConstVector(shuffles2,
 116                                                   bld_base->base.type.length),
 117                                   "");
 118 }
 119
 120 static void
 121 emit_store_64bit_chan(struct lp_build_nir_context *bld_base,
 122                       LLVMValueRef chan_ptr,
 123                       LLVMValueRef chan_ptr2,
 124                       LLVMValueRef value)
 125 {
 126    struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
 127    struct lp_build_context *float_bld = &bld_base->base;
 128    LLVMValueRef split_vals[2];
 129
 130    emit_store_64bit_split(bld_base, value, split_vals);
 131
 132    lp_exec_mask_store(&bld->exec_mask, float_bld, split_vals[0], chan_ptr);
 133    lp_exec_mask_store(&bld->exec_mask, float_bld, split_vals[1], chan_ptr2);
 134 }
 135
 136 static LLVMValueRef
 137 get_soa_array_offsets(struct lp_build_context *uint_bld,
 138                       LLVMValueRef indirect_index,
 139                       int num_components,
 140                       unsigned chan_index,
 141                       bool need_perelement_offset)
 142 {
 143    struct gallivm_state *gallivm = uint_bld->gallivm;
 144    LLVMValueRef chan_vec =
 145       lp_build_const_int_vec(uint_bld->gallivm, uint_bld->type, chan_index);
 146    LLVMValueRef length_vec =
 147       lp_build_const_int_vec(gallivm, uint_bld->type, uint_bld->type.length);
 148    LLVMValueRef index_vec;
 149
 150    /* index_vec = (indirect_index * 4 + chan_index) * length + offsets */
 151    index_vec = lp_build_mul(uint_bld, indirect_index, lp_build_const_int_vec(uint_bld->gallivm, uint_bld->type, num_components));
 152    index_vec = lp_build_add(uint_bld, index_vec, chan_vec);
 153    index_vec = lp_build_mul(uint_bld, index_vec, length_vec);
 154
 155    if (need_perelement_offset) {
 156       LLVMValueRef pixel_offsets;
 157       unsigned i;
 158      /* build pixel offset vector: {0, 1, 2, 3, ...} */
 159       pixel_offsets = uint_bld->undef;
 160       for (i = 0; i < uint_bld->type.length; i++) {
 161          LLVMValueRef ii = lp_build_const_int32(gallivm, i);
 162          pixel_offsets = LLVMBuildInsertElement(gallivm->builder, pixel_offsets,
 163                                                 ii, ii, "");
 164       }
 165       index_vec = lp_build_add(uint_bld, index_vec, pixel_offsets);
 166    }
 167    return index_vec;
 168 }
 169
 170 static LLVMValueRef
 171 build_gather(struct lp_build_nir_context *bld_base,
 172              struct lp_build_context *bld,
 173              LLVMValueRef base_ptr,
 174              LLVMValueRef indexes,
 175              LLVMValueRef overflow_mask,
 176              LLVMValueRef indexes2)
 177 {
 178    struct gallivm_state *gallivm = bld_base->base.gallivm;
 179    LLVMBuilderRef builder = gallivm->builder;
 180    struct lp_build_context *uint_bld = &bld_base->uint_bld;
 181    LLVMValueRef res;
 182    unsigned i;
 183
 184    if (indexes2)
 185       res = LLVMGetUndef(LLVMVectorType(LLVMFloatTypeInContext(gallivm->context), bld_base->base.type.length * 2));
 186    else
 187       res = bld->undef;
 188    /*
 189     * overflow_mask is a vector telling us which channels
 190     * in the vector overflowed. We use the overflow behavior for
 191     * constant buffers which is defined as:
 192     * Out of bounds access to constant buffer returns 0 in all
 193     * components. Out of bounds behavior is always with respect
 194     * to the size of the buffer bound at that slot.
 195     */
 196
 197    if (overflow_mask) {
 198       /*
 199        * We avoid per-element control flow here (also due to llvm going crazy,
 200        * though I suspect it's better anyway since overflow is likely rare).
 201        * Note that since we still fetch from buffers even if num_elements was
 202        * zero (in this case we'll fetch from index zero) the jit func callers
 203        * MUST provide valid fake constant buffers of size 4x32 (the values do
 204        * not matter), otherwise we'd still need (not per element though)
 205        * control flow.
 206        */
 207       indexes = lp_build_select(uint_bld, overflow_mask, uint_bld->zero, indexes);
 208       if (indexes2)
 209          indexes2 = lp_build_select(uint_bld, overflow_mask, uint_bld->zero, indexes2);
 210    }
 211
 212    /*
 213     * Loop over elements of index_vec, load scalar value, insert it into 'res'.
 214     */
 215    for (i = 0; i < bld->type.length * (indexes2 ? 2 : 1); i++) {
 216       LLVMValueRef si, di;
 217       LLVMValueRef index;
 218       LLVMValueRef scalar_ptr, scalar;
 219
 220       di = lp_build_const_int32(gallivm, i);
 221       if (indexes2)
 222          si = lp_build_const_int32(gallivm, i >> 1);
 223       else
 224          si = di;
 225
 226       if (indexes2 && (i & 1)) {
 227          index = LLVMBuildExtractElement(builder,
 228                                          indexes2, si, "");
 229       } else {
 230          index = LLVMBuildExtractElement(builder,
 231                                          indexes, si, "");
 232       }
 233       scalar_ptr = LLVMBuildGEP(builder, base_ptr,
 234                                 &index, 1, "gather_ptr");
 235       scalar = LLVMBuildLoad(builder, scalar_ptr, "");
 236
 237       res = LLVMBuildInsertElement(builder, res, scalar, di, "");
 238    }
 239
 240    if (overflow_mask) {
 241       if (indexes2) {
 242          res = LLVMBuildBitCast(builder, res, bld_base->dbl_bld.vec_type, "");
 243          overflow_mask = LLVMBuildSExt(builder, overflow_mask,
 244                                        bld_base->dbl_bld.int_vec_type, "");
 245          res = lp_build_select(&bld_base->dbl_bld, overflow_mask,
 246                                bld_base->dbl_bld.zero, res);
 247       } else
 248          res = lp_build_select(bld, overflow_mask, bld->zero, res);
 249    }
 250
 251    return res;
 252 }
 253
 254 /**
 255  * Scatter/store vector.
 256  */
 257 static void
 258 emit_mask_scatter(struct lp_build_nir_soa_context *bld,
 259                   LLVMValueRef base_ptr,
 260                   LLVMValueRef indexes,
 261                   LLVMValueRef values,
 262                   struct lp_exec_mask *mask)
 263 {
 264    struct gallivm_state *gallivm = bld->bld_base.base.gallivm;
 265    LLVMBuilderRef builder = gallivm->builder;
 266    unsigned i;
 267    LLVMValueRef pred = mask->has_mask ? mask->exec_mask : NULL;
 268
 269    /*
 270     * Loop over elements of index_vec, store scalar value.
 271     */
 272    for (i = 0; i < bld->bld_base.base.type.length; i++) {
 273       LLVMValueRef ii = lp_build_const_int32(gallivm, i);
 274       LLVMValueRef index = LLVMBuildExtractElement(builder, indexes, ii, "");
 275       LLVMValueRef scalar_ptr = LLVMBuildGEP(builder, base_ptr, &index, 1, "scatter_ptr");
 276       LLVMValueRef val = LLVMBuildExtractElement(builder, values, ii, "scatter_val");
 277       LLVMValueRef scalar_pred = pred ?
 278          LLVMBuildExtractElement(builder, pred, ii, "scatter_pred") : NULL;
 279
 280       if (0)
 281          lp_build_printf(gallivm, "scatter %d: val %f at %d %p\n",
 282                          ii, val, index, scalar_ptr);
 283
 284       if (scalar_pred) {
 285          LLVMValueRef real_val, dst_val;
 286          dst_val = LLVMBuildLoad(builder, scalar_ptr, "");
 287          real_val = lp_build_select(&bld->uint_elem_bld, scalar_pred, val, dst_val);
 288          LLVMBuildStore(builder, real_val, scalar_ptr);
 289       }
 290       else {
 291          LLVMBuildStore(builder, val, scalar_ptr);
 292       }
 293    }
 294 }
 295
 296 static void emit_load_var(struct lp_build_nir_context *bld_base,
 297                            nir_variable_mode deref_mode,
 298                            unsigned num_components,
 299                            unsigned bit_size,
 300                            nir_variable *var,
 301                            unsigned vertex_index,
 302                            LLVMValueRef indir_vertex_index,
 303                            unsigned const_index,
 304                            LLVMValueRef indir_index,
 305                            LLVMValueRef result[NIR_MAX_VEC_COMPONENTS])
 306 {
 307    struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
 308    struct gallivm_state *gallivm = bld_base->base.gallivm;
 309    int dmul = bit_size == 64 ? 2 : 1;
 310    switch (deref_mode) {
 311    case nir_var_shader_in:
 312       for (unsigned i = 0; i < num_components; i++) {
 313          int idx = (i * dmul) + var->data.location_frac;
 314          if (bld->gs_iface) {
 315             LLVMValueRef vertex_index_val = lp_build_const_int32(gallivm, vertex_index);
 316             LLVMValueRef attrib_index_val = lp_build_const_int32(gallivm, const_index + var->data.driver_location);
 317             LLVMValueRef swizzle_index_val = lp_build_const_int32(gallivm, idx);
 318             LLVMValueRef result2;
 319             result[i] = bld->gs_iface->fetch_input(bld->gs_iface, &bld_base->base,
 320                                                    false, vertex_index_val, 0, attrib_index_val, swizzle_index_val);
 321             if (bit_size == 64) {
 322                LLVMValueRef swizzle_index_val = lp_build_const_int32(gallivm, idx + 1);
 323                result2 = bld->gs_iface->fetch_input(bld->gs_iface, &bld_base->base,
 324                                                     false, vertex_index_val, 0, attrib_index_val, swizzle_index_val);
 325                result[i] = emit_fetch_64bit(bld_base, result[i], result2);
 326             }
 327          } else if (bld->tes_iface) {
 328             LLVMValueRef vertex_index_val = lp_build_const_int32(gallivm, vertex_index);
 329             LLVMValueRef attrib_index_val;
 330             LLVMValueRef swizzle_index_val = lp_build_const_int32(gallivm, idx);
 331             LLVMValueRef result2;
 332
 333             if (indir_index)
 334                attrib_index_val = lp_build_add(&bld_base->uint_bld, indir_index, lp_build_const_int_vec(gallivm, bld_base->uint_bld.type, var->data.driver_location));
 335             else
 336                attrib_index_val = lp_build_const_int32(gallivm, const_index + var->data.driver_location);
 337             if (var->data.patch) {
 338                result[i] = bld->tes_iface->fetch_patch_input(bld->tes_iface, &bld_base->base,
 339                                                              indir_index ? true : false, attrib_index_val, swizzle_index_val);
 340                if (bit_size == 64) {
 341                   LLVMValueRef swizzle_index_val = lp_build_const_int32(gallivm, idx + 1);
 342                   result2 = bld->tes_iface->fetch_patch_input(bld->tes_iface, &bld_base->base,
 343                                                               indir_index ? true : false, attrib_index_val, swizzle_index_val);
 344                   result[i] = emit_fetch_64bit(bld_base, result[i], result2);
 345                }
 346             }
 347             else {
 348                result[i] = bld->tes_iface->fetch_vertex_input(bld->tes_iface, &bld_base->base,
 349                                                               indir_vertex_index ? true : false,
 350                                                               indir_vertex_index ? indir_vertex_index : vertex_index_val,
 351                                                               indir_index ? true : false, attrib_index_val, swizzle_index_val);
 352                if (bit_size == 64) {
 353                   LLVMValueRef swizzle_index_val = lp_build_const_int32(gallivm, idx + 1);
 354                   result2 = bld->tes_iface->fetch_vertex_input(bld->tes_iface, &bld_base->base,
 355                                                                indir_vertex_index ? true : false,
 356                                                                indir_vertex_index ? indir_vertex_index : vertex_index_val,
 357                                                                indir_index ? true : false, attrib_index_val, swizzle_index_val);
 358                   result[i] = emit_fetch_64bit(bld_base, result[i], result2);
 359                }
 360             }
 361          } else if (bld->tcs_iface) {
 362             LLVMValueRef vertex_index_val = lp_build_const_int32(gallivm, vertex_index);
 363             LLVMValueRef attrib_index_val;
 364             LLVMValueRef swizzle_index_val = lp_build_const_int32(gallivm, idx);
 365
 366             if (indir_index)
 367                attrib_index_val = lp_build_add(&bld_base->uint_bld, indir_index, lp_build_const_int_vec(gallivm, bld_base->uint_bld.type, var->data.driver_location));
 368             else
 369                attrib_index_val = lp_build_const_int32(gallivm, const_index + var->data.driver_location);
 370             result[i] = bld->tcs_iface->emit_fetch_input(bld->tcs_iface, &bld_base->base,
 371                                                          indir_vertex_index ? true : false, indir_vertex_index ? indir_vertex_index : vertex_index_val,
 372                                                          indir_index ? true : false, attrib_index_val, swizzle_index_val);
 373             if (bit_size == 64) {
 374                LLVMValueRef swizzle_index_val = lp_build_const_int32(gallivm, idx + 1);
 375                LLVMValueRef result2 = bld->tcs_iface->emit_fetch_input(bld->tcs_iface, &bld_base->base,
 376                                                                        indir_vertex_index ? true : false, indir_vertex_index ? indir_vertex_index : vertex_index_val,
 377                                                                        indir_index ? true : false, attrib_index_val, swizzle_index_val);
 378                result[i] = emit_fetch_64bit(bld_base, result[i], result2);
 379             }
 380          } else {
 381             if (indir_index) {
 382                LLVMValueRef attrib_index_val = lp_build_add(&bld_base->uint_bld, indir_index, lp_build_const_int_vec(gallivm, bld_base->uint_bld.type, var->data.driver_location));
 383                LLVMValueRef index_vec = get_soa_array_offsets(&bld_base->uint_bld,
 384                                                               attrib_index_val, 4, idx,
 385                                                               TRUE);
 386                LLVMValueRef index_vec2 = NULL;
 387                LLVMTypeRef fptr_type;
 388                LLVMValueRef inputs_array;
 389                fptr_type = LLVMPointerType(LLVMFloatTypeInContext(gallivm->context), 0);
 390                inputs_array = LLVMBuildBitCast(gallivm->builder, bld->inputs_array, fptr_type, "");
 391
 392                if (bit_size == 64)
 393                   index_vec2 = get_soa_array_offsets(&bld_base->uint_bld,
 394                                                      indir_index, 4, idx + 1, TRUE);
 395
 396                /* Gather values from the input register array */
 397                result[i] = build_gather(bld_base, &bld_base->base, inputs_array, index_vec, NULL, index_vec2);
 398             } else {
 399                if (bld->indirects & nir_var_shader_in) {
 400                   LLVMValueRef lindex = lp_build_const_int32(gallivm,
 401                                                              var->data.driver_location * 4 + idx);
 402                   LLVMValueRef input_ptr = lp_build_pointer_get(gallivm->builder,
 403                                                              bld->inputs_array, lindex);
 404                   if (bit_size == 64) {
 405                      LLVMValueRef lindex2 = lp_build_const_int32(gallivm,
 406                                                                  var->data.driver_location * 4 + (idx + 1));
 407                      LLVMValueRef input_ptr2 = lp_build_pointer_get(gallivm->builder,
 408                                                                     bld->inputs_array, lindex2);
 409                      result[i] = emit_fetch_64bit(bld_base, input_ptr, input_ptr2);
 410                   } else {
 411                      result[i] = input_ptr;
 412                   }
 413                } else {
 414                   if (bit_size == 64) {
 415                      LLVMValueRef tmp[2];
 416                      tmp[0] = bld->inputs[var->data.driver_location + const_index][idx];
 417                      tmp[1] = bld->inputs[var->data.driver_location + const_index][idx + 1];
 418                      result[i] = emit_fetch_64bit(bld_base, tmp[0], tmp[1]);
 419                   } else {
 420                      result[i] = bld->inputs[var->data.driver_location + const_index][idx];
 421                   }
 422                }
 423             }
 424          }
 425       }
 426       break;
 427    case nir_var_shader_out:
 428       for (unsigned i = 0; i < num_components; i++) {
 429          int idx = (i * dmul) + var->data.location_frac;
 430          if (bld->tcs_iface) {
 431             LLVMValueRef vertex_index_val = lp_build_const_int32(gallivm, vertex_index);
 432             LLVMValueRef attrib_index_val;
 433             LLVMValueRef swizzle_index_val = lp_build_const_int32(gallivm, idx);
 434
 435             if (indir_index)
 436                attrib_index_val = lp_build_add(&bld_base->uint_bld, indir_index, lp_build_const_int_vec(gallivm, bld_base->uint_bld.type, var->data.driver_location));
 437             else
 438                attrib_index_val = lp_build_const_int32(gallivm, const_index + var->data.driver_location);
 439
 440             result[i] = bld->tcs_iface->emit_fetch_output(bld->tcs_iface, &bld_base->base,
 441                                                          indir_vertex_index ? true : false, indir_vertex_index ? indir_vertex_index : vertex_index_val,
 442                                                           indir_index ? true : false, attrib_index_val, swizzle_index_val, 0);
 443             if (bit_size == 64) {
 444                LLVMValueRef swizzle_index_val = lp_build_const_int32(gallivm, idx + 1);
 445                LLVMValueRef result2 = bld->tcs_iface->emit_fetch_output(bld->tcs_iface, &bld_base->base,
 446                                                                         indir_vertex_index ? true : false, indir_vertex_index ? indir_vertex_index : vertex_index_val,
 447                                                                         indir_index ? true : false, attrib_index_val, swizzle_index_val, 0);
 448                result[i] = emit_fetch_64bit(bld_base, result[i], result2);
 449             }
 450          }
 451       }
 452       break;
 453    default:
 454       break;
 455    }
 456 }
 457
 458 static void emit_store_chan(struct lp_build_nir_context *bld_base,
 459                             nir_variable_mode deref_mode,
 460                             unsigned bit_size,
 461                             unsigned location, unsigned comp,
 462                             unsigned chan,
 463                             LLVMValueRef dst)
 464 {
 465    struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
 466    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
 467    struct lp_build_context *float_bld = &bld_base->base;
 468
 469    if (bit_size == 64) {
 470       chan *= 2;
 471       chan += comp;
 472       if (chan >= 4) {
 473          chan -= 4;
 474          location++;
 475       }
 476       emit_store_64bit_chan(bld_base, bld->outputs[location][chan],
 477                             bld->outputs[location][chan + 1], dst);
 478    } else {
 479       dst = LLVMBuildBitCast(builder, dst, float_bld->vec_type, "");
 480       lp_exec_mask_store(&bld->exec_mask, float_bld, dst,
 481                          bld->outputs[location][chan + comp]);
 482    }
 483 }
 484
 485 static void emit_store_tcs_chan(struct lp_build_nir_context *bld_base,
 486                                 unsigned bit_size,
 487                                 unsigned location,
 488                                 unsigned const_index,
 489                                 LLVMValueRef indir_vertex_index,
 490                                 LLVMValueRef indir_index,
 491                                 unsigned comp,
 492                                 unsigned chan,
 493                                 LLVMValueRef chan_val)
 494 {
 495    struct gallivm_state *gallivm = bld_base->base.gallivm;
 496    struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
 497    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
 498    unsigned swizzle = chan;
 499    if (bit_size == 64) {
 500       swizzle *= 2;
 501       swizzle += comp;
 502       if (swizzle >= 4) {
 503          swizzle -= 4;
 504          location++;
 505       }
 506    } else
 507       swizzle += comp;
 508    LLVMValueRef attrib_index_val;
 509    LLVMValueRef swizzle_index_val = lp_build_const_int32(gallivm, swizzle);
 510
 511    if (indir_index)
 512       attrib_index_val = lp_build_add(&bld_base->uint_bld, indir_index, lp_build_const_int_vec(gallivm, bld_base->uint_bld.type, location));
 513    else
 514       attrib_index_val = lp_build_const_int32(gallivm, const_index + location);
 515    if (bit_size == 64) {
 516       LLVMValueRef split_vals[2];
 517       LLVMValueRef swizzle_index_val2 = lp_build_const_int32(gallivm, swizzle + 1);
 518       emit_store_64bit_split(bld_base, chan_val, split_vals);
 519       bld->tcs_iface->emit_store_output(bld->tcs_iface, &bld_base->base, 0,
 520                                         indir_vertex_index ? true : false,
 521                                         indir_vertex_index,
 522                                         indir_index ? true : false,
 523                                         attrib_index_val, swizzle_index_val,
 524                                         split_vals[0], mask_vec(bld_base));
 525       bld->tcs_iface->emit_store_output(bld->tcs_iface, &bld_base->base, 0,
 526                                         indir_vertex_index ? true : false,
 527                                         indir_vertex_index,
 528                                         indir_index ? true : false,
 529                                         attrib_index_val, swizzle_index_val2,
 530                                         split_vals[1], mask_vec(bld_base));
 531    } else {
 532       chan_val = LLVMBuildBitCast(builder, chan_val, bld_base->base.vec_type, "");
 533       bld->tcs_iface->emit_store_output(bld->tcs_iface, &bld_base->base, 0,
 534                                         indir_vertex_index ? true : false,
 535                                         indir_vertex_index,
 536                                         indir_index ? true : false,
 537                                         attrib_index_val, swizzle_index_val,
 538                                         chan_val, mask_vec(bld_base));
 539    }
 540 }
 541
 542 static void emit_store_var(struct lp_build_nir_context *bld_base,
 543                            nir_variable_mode deref_mode,
 544                            unsigned num_components,
 545                            unsigned bit_size,
 546                            nir_variable *var,
 547                            unsigned writemask,
 548                            LLVMValueRef indir_vertex_index,
 549                            unsigned const_index,
 550                            LLVMValueRef indir_index,
 551                            LLVMValueRef dst)
 552 {
 553    struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
 554    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
 555    switch (deref_mode) {
 556    case nir_var_shader_out: {
 557       unsigned location = var->data.driver_location;
 558       unsigned comp = var->data.location_frac;
 559       if (bld_base->shader->info.stage == MESA_SHADER_FRAGMENT) {
 560          if (var->data.location == FRAG_RESULT_STENCIL)
 561             comp = 1;
 562          else if (var->data.location == FRAG_RESULT_DEPTH)
 563             comp = 2;
 564       }
 565
 566       for (unsigned chan = 0; chan < num_components; chan++) {
 567          if (writemask & (1u << chan)) {
 568             LLVMValueRef chan_val = (num_components == 1) ? dst : LLVMBuildExtractValue(builder, dst, chan, "");
 569             if (bld->tcs_iface) {
 570                emit_store_tcs_chan(bld_base, bit_size, location, const_index, indir_vertex_index, indir_index, comp, chan, chan_val);
 571             } else
 572                emit_store_chan(bld_base, deref_mode, bit_size, location + const_index, comp, chan, chan_val);
 573          }
 574       }
 575       break;
 576    }
 577    default:
 578       break;
 579    }
 580 }
 581
 582 static LLVMValueRef emit_load_reg(struct lp_build_nir_context *bld_base,
 583                                   struct lp_build_context *reg_bld,
 584                                   const nir_reg_src *reg,
 585                                   LLVMValueRef indir_src,
 586                                   LLVMValueRef reg_storage)
 587 {
 588    struct gallivm_state *gallivm = bld_base->base.gallivm;
 589    LLVMBuilderRef builder = gallivm->builder;
 590    int nc = reg->reg->num_components;
 591    LLVMValueRef vals[NIR_MAX_VEC_COMPONENTS] = { NULL };
 592    struct lp_build_context *uint_bld = &bld_base->uint_bld;
 593    if (reg->reg->num_array_elems) {
 594       LLVMValueRef indirect_val = lp_build_const_int_vec(gallivm, uint_bld->type, reg->base_offset);
 595       if (reg->indirect) {
 596          LLVMValueRef max_index = lp_build_const_int_vec(gallivm, uint_bld->type, reg->reg->num_array_elems - 1);
 597          indirect_val = LLVMBuildAdd(builder, indirect_val, indir_src, "");
 598          indirect_val = lp_build_min(uint_bld, indirect_val, max_index);
 599       }
 600       reg_storage = LLVMBuildBitCast(builder, reg_storage, LLVMPointerType(reg_bld->elem_type, 0), "");
 601       for (unsigned i = 0; i < nc; i++) {
 602          LLVMValueRef indirect_offset = get_soa_array_offsets(uint_bld, indirect_val, nc, i, TRUE);
 603          vals[i] = build_gather(bld_base, reg_bld, reg_storage, indirect_offset, NULL, NULL);
 604       }
 605    } else {
 606       for (unsigned i = 0; i < nc; i++) {
 607          LLVMValueRef this_storage = nc == 1 ? reg_storage : lp_build_array_get_ptr(gallivm, reg_storage,
 608                                                                                     lp_build_const_int32(gallivm, i));
 609          vals[i] = LLVMBuildLoad(builder, this_storage, "");
 610       }
 611    }
 612    return nc == 1 ? vals[0] : lp_nir_array_build_gather_values(builder, vals, nc);
 613 }
 614
 615 static void emit_store_reg(struct lp_build_nir_context *bld_base,
 616                            struct lp_build_context *reg_bld,
 617                            const nir_reg_dest *reg,
 618                            unsigned writemask,
 619                            LLVMValueRef indir_src,
 620                            LLVMValueRef reg_storage,
 621                            LLVMValueRef dst[NIR_MAX_VEC_COMPONENTS])
 622 {
 623    struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
 624    struct gallivm_state *gallivm = bld_base->base.gallivm;
 625    LLVMBuilderRef builder = gallivm->builder;
 626    struct lp_build_context *uint_bld = &bld_base->uint_bld;
 627    int nc = reg->reg->num_components;
 628    if (reg->reg->num_array_elems > 0) {
 629       LLVMValueRef indirect_val = lp_build_const_int_vec(gallivm, uint_bld->type, reg->base_offset);
 630       if (reg->indirect) {
 631          LLVMValueRef max_index = lp_build_const_int_vec(gallivm, uint_bld->type, reg->reg->num_array_elems - 1);
 632          indirect_val = LLVMBuildAdd(builder, indirect_val, indir_src, "");
 633          indirect_val = lp_build_min(uint_bld, indirect_val, max_index);
 634       }
 635       reg_storage = LLVMBuildBitCast(builder, reg_storage, LLVMPointerType(reg_bld->elem_type, 0), "");
 636       for (unsigned i = 0; i < nc; i++) {
 637          if (!(writemask & (1 << i)))
 638             continue;
 639          LLVMValueRef indirect_offset = get_soa_array_offsets(uint_bld, indirect_val, nc, i, TRUE);
 640          dst[i] = LLVMBuildBitCast(builder, dst[i], reg_bld->vec_type, "");
 641          emit_mask_scatter(bld, reg_storage, indirect_offset, dst[i], &bld->exec_mask);
 642       }
 643       return;
 644    }
 645
 646    for (unsigned i = 0; i < nc; i++) {
 647       LLVMValueRef this_storage = nc == 1 ? reg_storage : lp_build_array_get_ptr(gallivm, reg_storage,
 648                                                          lp_build_const_int32(gallivm, i));
 649       dst[i] = LLVMBuildBitCast(builder, dst[i], reg_bld->vec_type, "");
 650       lp_exec_mask_store(&bld->exec_mask, reg_bld, dst[i], this_storage);
 651    }
 652 }
 653
 654 static void emit_load_kernel_arg(struct lp_build_nir_context *bld_base,
 655                                  unsigned nc,
 656                                  unsigned bit_size,
 657                                  unsigned offset_bit_size,
 658                                  bool offset_is_uniform,
 659                                  LLVMValueRef offset,
 660                                  LLVMValueRef result[NIR_MAX_VEC_COMPONENTS])
 661 {
 662    struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
 663    struct gallivm_state *gallivm = bld_base->base.gallivm;
 664    LLVMBuilderRef builder = gallivm->builder;
 665    struct lp_build_context *bld_broad = get_int_bld(bld_base, true, bit_size);
 666    LLVMValueRef kernel_args_ptr = bld->kernel_args_ptr;
 667    unsigned size_shift = 0;
 668    struct lp_build_context *bld_offset = get_int_bld(bld_base, true, offset_bit_size);
 669    if (bit_size == 16)
 670       size_shift = 1;
 671    else if (bit_size == 32)
 672       size_shift = 2;
 673    else if (bit_size == 64)
 674       size_shift = 3;
 675    if (size_shift)
 676       offset = lp_build_shr(bld_offset, offset, lp_build_const_int_vec(gallivm, bld_offset->type, size_shift));
 677
 678    LLVMTypeRef ptr_type = LLVMPointerType(bld_broad->elem_type, 0);
 679    kernel_args_ptr = LLVMBuildBitCast(builder, kernel_args_ptr, ptr_type, "");
 680
 681    if (offset_is_uniform) {
 682       offset = LLVMBuildExtractElement(builder, offset, lp_build_const_int32(gallivm, 0), "");
 683
 684       for (unsigned c = 0; c < nc; c++) {
 685          LLVMValueRef this_offset = LLVMBuildAdd(builder, offset, offset_bit_size == 64 ? lp_build_const_int64(gallivm, c) : lp_build_const_int32(gallivm, c), "");
 686
 687          LLVMValueRef scalar = lp_build_pointer_get(builder, kernel_args_ptr, this_offset);
 688          result[c] = lp_build_broadcast_scalar(bld_broad, scalar);
 689       }
 690    }
 691 }
 692
 693 static LLVMValueRef global_addr_to_ptr(struct gallivm_state *gallivm, LLVMValueRef addr_ptr, unsigned bit_size)
 694 {
 695    LLVMBuilderRef builder = gallivm->builder;
 696    switch (bit_size) {
 697    case 8:
 698       addr_ptr = LLVMBuildIntToPtr(builder, addr_ptr, LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0), "");
 699       break;
 700    case 16:
 701       addr_ptr = LLVMBuildIntToPtr(builder, addr_ptr, LLVMPointerType(LLVMInt16TypeInContext(gallivm->context), 0), "");
 702       break;
 703    case 32:
 704    default:
 705       addr_ptr = LLVMBuildIntToPtr(builder, addr_ptr, LLVMPointerType(LLVMInt32TypeInContext(gallivm->context), 0), "");
 706       break;
 707    case 64:
 708       addr_ptr = LLVMBuildIntToPtr(builder, addr_ptr, LLVMPointerType(LLVMInt64TypeInContext(gallivm->context), 0), "");
 709       break;
 710    }
 711    return addr_ptr;
 712 }
 713
 714 static void emit_load_global(struct lp_build_nir_context *bld_base,
 715                              unsigned nc,
 716                              unsigned bit_size,
 717                              unsigned addr_bit_size,
 718                              LLVMValueRef addr,
 719                              LLVMValueRef outval[NIR_MAX_VEC_COMPONENTS])
 720 {
 721    struct gallivm_state *gallivm = bld_base->base.gallivm;
 722    LLVMBuilderRef builder = gallivm->builder;
 723    struct lp_build_context *uint_bld = &bld_base->uint_bld;
 724    struct lp_build_context *res_bld;
 725
 726    res_bld = get_int_bld(bld_base, true, bit_size);
 727
 728    for (unsigned c = 0; c < nc; c++) {
 729       LLVMValueRef result = lp_build_alloca(gallivm, res_bld->vec_type, "");
 730
 731       struct lp_build_loop_state loop_state;
 732       lp_build_loop_begin(&loop_state, gallivm, lp_build_const_int32(gallivm, 0));
 733
 734       LLVMValueRef addr_ptr = LLVMBuildExtractElement(gallivm->builder, addr,
 735                                                       loop_state.counter, "");
 736       addr_ptr = global_addr_to_ptr(gallivm, addr_ptr, bit_size);
 737
 738       LLVMValueRef value_ptr = lp_build_pointer_get(builder, addr_ptr, lp_build_const_int32(gallivm, c));
 739
 740       LLVMValueRef temp_res;
 741       temp_res = LLVMBuildLoad(builder, result, "");
 742       temp_res = LLVMBuildInsertElement(builder, temp_res, value_ptr, loop_state.counter, "");
 743       LLVMBuildStore(builder, temp_res, result);
 744       lp_build_loop_end_cond(&loop_state, lp_build_const_int32(gallivm, uint_bld->type.length),
 745                              NULL, LLVMIntUGE);
 746       outval[c] = LLVMBuildLoad(builder, result, "");
 747    }
 748 }
 749
 750 static void emit_store_global(struct lp_build_nir_context *bld_base,
 751                               unsigned writemask,
 752                               unsigned nc, unsigned bit_size,
 753                               unsigned addr_bit_size,
 754                               LLVMValueRef addr,
 755                               LLVMValueRef dst)
 756 {
 757    struct gallivm_state *gallivm = bld_base->base.gallivm;
 758    LLVMBuilderRef builder = gallivm->builder;
 759    struct lp_build_context *uint_bld = &bld_base->uint_bld;
 760
 761    for (unsigned c = 0; c < nc; c++) {
 762       if (!(writemask & (1u << c)))
 763          continue;
 764       LLVMValueRef val = (nc == 1) ? dst : LLVMBuildExtractValue(builder, dst, c, "");
 765
 766       LLVMValueRef exec_mask = mask_vec(bld_base);
 767       struct lp_build_loop_state loop_state;
 768       lp_build_loop_begin(&loop_state, gallivm, lp_build_const_int32(gallivm, 0));
 769       LLVMValueRef value_ptr = LLVMBuildExtractElement(gallivm->builder, val,
 770                                                        loop_state.counter, "");
 771
 772       LLVMValueRef addr_ptr = LLVMBuildExtractElement(gallivm->builder, addr,
 773                                                       loop_state.counter, "");
 774       addr_ptr = global_addr_to_ptr(gallivm, addr_ptr, bit_size);
 775       switch (bit_size) {
 776       case 32:
 777          value_ptr = LLVMBuildBitCast(builder, value_ptr, LLVMInt32TypeInContext(gallivm->context), "");
 778          break;
 779       case 64:
 780          value_ptr = LLVMBuildBitCast(builder, value_ptr, LLVMInt64TypeInContext(gallivm->context), "");
 781          break;
 782       default:
 783          break;
 784       }
 785       struct lp_build_if_state ifthen;
 786
 787       LLVMValueRef cond = LLVMBuildICmp(gallivm->builder, LLVMIntNE, exec_mask, uint_bld->zero, "");
 788       cond = LLVMBuildExtractElement(gallivm->builder, cond, loop_state.counter, "");
 789       lp_build_if(&ifthen, gallivm, cond);
 790       lp_build_pointer_set(builder, addr_ptr, lp_build_const_int32(gallivm, c), value_ptr);
 791       lp_build_endif(&ifthen);
 792       lp_build_loop_end_cond(&loop_state, lp_build_const_int32(gallivm, uint_bld->type.length),
 793                              NULL, LLVMIntUGE);
 794    }
 795 }
 796
 797 static void emit_atomic_global(struct lp_build_nir_context *bld_base,
 798                                nir_intrinsic_op nir_op,
 799                                unsigned addr_bit_size,
 800                                LLVMValueRef addr,
 801                                LLVMValueRef val, LLVMValueRef val2,
 802                                LLVMValueRef *result)
 803 {
 804    struct gallivm_state *gallivm = bld_base->base.gallivm;
 805    LLVMBuilderRef builder = gallivm->builder;
 806    struct lp_build_context *uint_bld = &bld_base->uint_bld;
 807
 808    LLVMValueRef atom_res = lp_build_alloca(gallivm,
 809                                            uint_bld->vec_type, "");
 810    LLVMValueRef exec_mask = mask_vec(bld_base);
 811    struct lp_build_loop_state loop_state;
 812    lp_build_loop_begin(&loop_state, gallivm, lp_build_const_int32(gallivm, 0));
 813
 814    LLVMValueRef value_ptr = LLVMBuildExtractElement(gallivm->builder, val,
 815                                                     loop_state.counter, "");
 816
 817    LLVMValueRef addr_ptr = LLVMBuildExtractElement(gallivm->builder, addr,
 818                                                    loop_state.counter, "");
 819    addr_ptr = global_addr_to_ptr(gallivm, addr_ptr, 32);
 820    struct lp_build_if_state ifthen;
 821    LLVMValueRef cond, temp_res;
 822    LLVMValueRef scalar;
 823    cond = LLVMBuildICmp(gallivm->builder, LLVMIntNE, exec_mask, uint_bld->zero, "");
 824    cond = LLVMBuildExtractElement(gallivm->builder, cond, loop_state.counter, "");
 825    lp_build_if(&ifthen, gallivm, cond);
 826
 827    if (nir_op == nir_intrinsic_global_atomic_comp_swap) {
 828       LLVMValueRef cas_src_ptr = LLVMBuildExtractElement(gallivm->builder, val2,
 829                                                          loop_state.counter, "");
 830       cas_src_ptr = LLVMBuildBitCast(gallivm->builder, cas_src_ptr, uint_bld->elem_type, "");
 831       scalar = LLVMBuildAtomicCmpXchg(builder, addr_ptr, value_ptr,
 832                                       cas_src_ptr,
 833                                       LLVMAtomicOrderingSequentiallyConsistent,
 834                                       LLVMAtomicOrderingSequentiallyConsistent,
 835                                       false);
 836       scalar = LLVMBuildExtractValue(gallivm->builder, scalar, 0, "");
 837    } else {
 838       LLVMAtomicRMWBinOp op;
 839       switch (nir_op) {
 840       case nir_intrinsic_global_atomic_add:
 841          op = LLVMAtomicRMWBinOpAdd;
 842          break;
 843       case nir_intrinsic_global_atomic_exchange:
 844          op = LLVMAtomicRMWBinOpXchg;
 845          break;
 846       case nir_intrinsic_global_atomic_and:
 847          op = LLVMAtomicRMWBinOpAnd;
 848          break;
 849       case nir_intrinsic_global_atomic_or:
 850          op = LLVMAtomicRMWBinOpOr;
 851          break;
 852       case nir_intrinsic_global_atomic_xor:
 853          op = LLVMAtomicRMWBinOpXor;
 854          break;
 855       case nir_intrinsic_global_atomic_umin:
 856          op = LLVMAtomicRMWBinOpUMin;
 857          break;
 858       case nir_intrinsic_global_atomic_umax:
 859          op = LLVMAtomicRMWBinOpUMax;
 860          break;
 861       case nir_intrinsic_global_atomic_imin:
 862          op = LLVMAtomicRMWBinOpMin;
 863          break;
 864       case nir_intrinsic_global_atomic_imax:
 865          op = LLVMAtomicRMWBinOpMax;
 866          break;
 867       default:
 868          unreachable("unknown atomic op");
 869       }
 870
 871       scalar = LLVMBuildAtomicRMW(builder, op,
 872                                   addr_ptr, value_ptr,
 873                                   LLVMAtomicOrderingSequentiallyConsistent,
 874                                   false);
 875    }
 876    temp_res = LLVMBuildLoad(builder, atom_res, "");
 877    temp_res = LLVMBuildInsertElement(builder, temp_res, scalar, loop_state.counter, "");
 878    LLVMBuildStore(builder, temp_res, atom_res);
 879    lp_build_else(&ifthen);
 880    temp_res = LLVMBuildLoad(builder, atom_res, "");
 881    temp_res = LLVMBuildInsertElement(builder, temp_res, lp_build_const_int32(gallivm, 0), loop_state.counter, "");
 882    LLVMBuildStore(builder, temp_res, atom_res);
 883    lp_build_endif(&ifthen);
 884    lp_build_loop_end_cond(&loop_state, lp_build_const_int32(gallivm, uint_bld->type.length),
 885                           NULL, LLVMIntUGE);
 886    *result = LLVMBuildLoad(builder, atom_res, "");
 887 }
 888
 889 static void emit_load_ubo(struct lp_build_nir_context *bld_base,
 890                           unsigned nc,
 891                           unsigned bit_size,
 892                           bool offset_is_uniform,
 893                           LLVMValueRef index,
 894                           LLVMValueRef offset,
 895                           LLVMValueRef result[NIR_MAX_VEC_COMPONENTS])
 896 {
 897    struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
 898    struct gallivm_state *gallivm = bld_base->base.gallivm;
 899    LLVMBuilderRef builder = gallivm->builder;
 900    struct lp_build_context *uint_bld = &bld_base->uint_bld;
 901    struct lp_build_context *bld_broad = bit_size == 64 ? &bld_base->dbl_bld : &bld_base->base;
 902    LLVMValueRef consts_ptr = lp_build_array_get(gallivm, bld->consts_ptr, index);
 903    unsigned size_shift = 0;
 904    if (bit_size == 32)
 905       size_shift = 2;
 906    else if (bit_size == 64)
 907       size_shift = 3;
 908    if (size_shift)
 909       offset = lp_build_shr(uint_bld, offset, lp_build_const_int_vec(gallivm, uint_bld->type, size_shift));
 910    if (bit_size == 64) {
 911       LLVMTypeRef dptr_type = LLVMPointerType(bld_base->dbl_bld.elem_type, 0);
 912       consts_ptr = LLVMBuildBitCast(builder, consts_ptr, dptr_type, "");
 913    }
 914
 915    if (offset_is_uniform) {
 916       offset = LLVMBuildExtractElement(builder, offset, lp_build_const_int32(gallivm, 0), "");
 917
 918       for (unsigned c = 0; c < nc; c++) {
 919          LLVMValueRef this_offset = LLVMBuildAdd(builder, offset, lp_build_const_int32(gallivm, c), "");
 920
 921          LLVMValueRef scalar = lp_build_pointer_get(builder, consts_ptr, this_offset);
 922          result[c] = lp_build_broadcast_scalar(bld_broad, scalar);
 923       }
 924    } else {
 925       LLVMValueRef overflow_mask;
 926       LLVMValueRef num_consts = lp_build_array_get(gallivm, bld->const_sizes_ptr, index);
 927
 928       num_consts = LLVMBuildShl(gallivm->builder, num_consts, lp_build_const_int32(gallivm, 4), "");
 929       num_consts = lp_build_broadcast_scalar(uint_bld, num_consts);
 930       for (unsigned c = 0; c < nc; c++) {
 931          LLVMValueRef this_offset = lp_build_add(uint_bld, offset, lp_build_const_int_vec(gallivm, uint_bld->type, c));
 932          overflow_mask = lp_build_compare(gallivm, uint_bld->type, PIPE_FUNC_GEQUAL,
 933                                           this_offset, num_consts);
 934
 935          result[c] = build_gather(bld_base, bld_broad, consts_ptr, this_offset, overflow_mask, NULL);
 936       }
 937    }
 938 }
 939
 940
 941 static void emit_load_mem(struct lp_build_nir_context *bld_base,
 942                           unsigned nc,
 943                           unsigned bit_size,
 944                           LLVMValueRef index,
 945                           LLVMValueRef offset,
 946                           LLVMValueRef outval[NIR_MAX_VEC_COMPONENTS])
 947 {
 948    struct gallivm_state *gallivm = bld_base->base.gallivm;
 949    struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
 950    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
 951    LLVMValueRef ssbo_ptr = NULL;
 952    struct lp_build_context *uint_bld = &bld_base->uint_bld;
 953    struct lp_build_context *uint64_bld = &bld_base->uint64_bld;
 954    LLVMValueRef ssbo_limit = NULL;
 955
 956    if (index) {
 957       LLVMValueRef ssbo_size_ptr = lp_build_array_get(gallivm, bld->ssbo_sizes_ptr, LLVMBuildExtractElement(builder, index, lp_build_const_int32(gallivm, 0), ""));
 958       ssbo_limit = LLVMBuildAShr(gallivm->builder, ssbo_size_ptr, lp_build_const_int32(gallivm, bit_size == 64 ? 3 : 2), "");
 959       ssbo_limit = lp_build_broadcast_scalar(uint_bld, ssbo_limit);
 960
 961       ssbo_ptr = lp_build_array_get(gallivm, bld->ssbo_ptr, LLVMBuildExtractElement(builder, index, lp_build_const_int32(gallivm, 0), ""));
 962    } else
 963       ssbo_ptr = bld->shared_ptr;
 964
 965    offset = LLVMBuildAShr(gallivm->builder, offset, lp_build_const_int_vec(gallivm, uint_bld->type, bit_size == 64 ? 3 : 2), "");
 966    for (unsigned c = 0; c < nc; c++) {
 967       LLVMValueRef loop_index = lp_build_add(uint_bld, offset, lp_build_const_int_vec(gallivm, uint_bld->type, c));
 968       LLVMValueRef exec_mask = mask_vec(bld_base);
 969
 970       if (ssbo_limit) {
 971          LLVMValueRef ssbo_oob_cmp = lp_build_cmp(uint_bld, PIPE_FUNC_LESS, loop_index, ssbo_limit);
 972          exec_mask = LLVMBuildAnd(builder, exec_mask, ssbo_oob_cmp, "");
 973       }
 974
 975       LLVMValueRef result = lp_build_alloca(gallivm, bit_size == 64 ? uint64_bld->vec_type : uint_bld->vec_type, "");
 976       struct lp_build_loop_state loop_state;
 977       lp_build_loop_begin(&loop_state, gallivm, lp_build_const_int32(gallivm, 0));
 978
 979       struct lp_build_if_state ifthen;
 980       LLVMValueRef cond, temp_res;
 981
 982       loop_index = LLVMBuildExtractElement(gallivm->builder, loop_index,
 983                                            loop_state.counter, "");
 984
 985       cond = LLVMBuildICmp(gallivm->builder, LLVMIntNE, exec_mask, uint_bld->zero, "");
 986       cond = LLVMBuildExtractElement(gallivm->builder, cond, loop_state.counter, "");
 987
 988       lp_build_if(&ifthen, gallivm, cond);
 989       LLVMValueRef scalar;
 990       if (bit_size == 64) {
 991          LLVMValueRef ssbo_ptr2 = LLVMBuildBitCast(builder, ssbo_ptr, LLVMPointerType(uint64_bld->elem_type, 0), "");
 992          scalar = lp_build_pointer_get(builder, ssbo_ptr2, loop_index);
 993       } else
 994          scalar = lp_build_pointer_get(builder, ssbo_ptr, loop_index);
 995
 996       temp_res = LLVMBuildLoad(builder, result, "");
 997       temp_res = LLVMBuildInsertElement(builder, temp_res, scalar, loop_state.counter, "");
 998       LLVMBuildStore(builder, temp_res, result);
 999       lp_build_else(&ifthen);
1000       temp_res = LLVMBuildLoad(builder, result, "");
1001       LLVMValueRef zero;
1002       if (bit_size == 64)
1003          zero = LLVMConstInt(LLVMInt64TypeInContext(gallivm->context), 0, 0);
1004       else
1005          zero = lp_build_const_int32(gallivm, 0);
1006       temp_res = LLVMBuildInsertElement(builder, temp_res, zero, loop_state.counter, "");
1007       LLVMBuildStore(builder, temp_res, result);
1008       lp_build_endif(&ifthen);
1009       lp_build_loop_end_cond(&loop_state, lp_build_const_int32(gallivm, uint_bld->type.length),
1010                                 NULL, LLVMIntUGE);
1011       outval[c] = LLVMBuildLoad(gallivm->builder, result, "");
1012    }
1013 }
1014
1015 static void emit_store_mem(struct lp_build_nir_context *bld_base,
1016                            unsigned writemask,
1017                            unsigned nc,
1018                            unsigned bit_size,
1019                            LLVMValueRef index,
1020                            LLVMValueRef offset,
1021                            LLVMValueRef dst)
1022 {
1023    struct gallivm_state *gallivm = bld_base->base.gallivm;
1024    struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
1025    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
1026    LLVMValueRef ssbo_ptr;
1027    struct lp_build_context *uint_bld = &bld_base->uint_bld;
1028    LLVMValueRef ssbo_limit = NULL;
1029
1030    if (index) {
1031       LLVMValueRef ssbo_size_ptr = lp_build_array_get(gallivm, bld->ssbo_sizes_ptr, LLVMBuildExtractElement(builder, index, lp_build_const_int32(gallivm, 0), ""));
1032       ssbo_limit = LLVMBuildAShr(gallivm->builder, ssbo_size_ptr, lp_build_const_int32(gallivm, bit_size == 64 ? 3 : 2), "");
1033       ssbo_limit = lp_build_broadcast_scalar(uint_bld, ssbo_limit);
1034       ssbo_ptr = lp_build_array_get(gallivm, bld->ssbo_ptr, LLVMBuildExtractElement(builder, index, lp_build_const_int32(gallivm, 0), ""));
1035    } else
1036       ssbo_ptr = bld->shared_ptr;
1037
1038    offset = lp_build_shr_imm(uint_bld, offset, bit_size == 64 ? 3 : 2);
1039    for (unsigned c = 0; c < nc; c++) {
1040       if (!(writemask & (1u << c)))
1041          continue;
1042       LLVMValueRef loop_index = lp_build_add(uint_bld, offset, lp_build_const_int_vec(gallivm, uint_bld->type, c));
1043       LLVMValueRef val = (nc == 1) ? dst : LLVMBuildExtractValue(builder, dst, c, "");
1044
1045       LLVMValueRef exec_mask = mask_vec(bld_base);
1046       if (ssbo_limit) {
1047          LLVMValueRef ssbo_oob_cmp = lp_build_cmp(uint_bld, PIPE_FUNC_LESS, loop_index, ssbo_limit);
1048          exec_mask = LLVMBuildAnd(builder, exec_mask, ssbo_oob_cmp, "");
1049       }
1050
1051       struct lp_build_loop_state loop_state;
1052       lp_build_loop_begin(&loop_state, gallivm, lp_build_const_int32(gallivm, 0));
1053       LLVMValueRef value_ptr = LLVMBuildExtractElement(gallivm->builder, val,
1054                                                        loop_state.counter, "");
1055       if (bit_size == 64)
1056          value_ptr = LLVMBuildBitCast(gallivm->builder, value_ptr, bld_base->uint64_bld.elem_type, "");
1057       else
1058          value_ptr = LLVMBuildBitCast(gallivm->builder, value_ptr, uint_bld->elem_type, "");
1059       struct lp_build_if_state ifthen;
1060       LLVMValueRef cond;
1061
1062       loop_index = LLVMBuildExtractElement(gallivm->builder, loop_index,
1063                                            loop_state.counter, "");
1064       cond = LLVMBuildICmp(gallivm->builder, LLVMIntNE, exec_mask, uint_bld->zero, "");
1065       cond = LLVMBuildExtractElement(gallivm->builder, cond, loop_state.counter, "");
1066       lp_build_if(&ifthen, gallivm, cond);
1067       if (bit_size == 64) {
1068          LLVMValueRef ssbo_ptr2 = LLVMBuildBitCast(builder, ssbo_ptr, LLVMPointerType(bld_base->uint64_bld.elem_type, 0), "");
1069          lp_build_pointer_set(builder, ssbo_ptr2, loop_index, value_ptr);
1070       } else
1071          lp_build_pointer_set(builder, ssbo_ptr, loop_index, value_ptr);
1072       lp_build_endif(&ifthen);
1073       lp_build_loop_end_cond(&loop_state, lp_build_const_int32(gallivm, uint_bld->type.length),
1074                              NULL, LLVMIntUGE);
1075    }
1076 }
1077
1078 static void emit_atomic_mem(struct lp_build_nir_context *bld_base,
1079                             nir_intrinsic_op nir_op,
1080                             LLVMValueRef index, LLVMValueRef offset,
1081                             LLVMValueRef val, LLVMValueRef val2,
1082                             LLVMValueRef *result)
1083 {
1084    struct gallivm_state *gallivm = bld_base->base.gallivm;
1085    struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
1086    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
1087    LLVMValueRef ssbo_ptr;
1088    struct lp_build_context *uint_bld = &bld_base->uint_bld;
1089    LLVMValueRef ssbo_limit = NULL;
1090
1091    if (index) {
1092       LLVMValueRef ssbo_size_ptr = lp_build_array_get(gallivm, bld->ssbo_sizes_ptr, LLVMBuildExtractElement(builder, index, lp_build_const_int32(gallivm, 0), ""));
1093       ssbo_limit = LLVMBuildAShr(gallivm->builder, ssbo_size_ptr, lp_build_const_int32(gallivm, 2), "");
1094       ssbo_limit = lp_build_broadcast_scalar(uint_bld, ssbo_limit);
1095       ssbo_ptr = lp_build_array_get(gallivm, bld->ssbo_ptr, LLVMBuildExtractElement(builder, index, lp_build_const_int32(gallivm, 0), ""));
1096    } else
1097       ssbo_ptr = bld->shared_ptr;
1098
1099    offset = lp_build_shr_imm(uint_bld, offset, 2);
1100    LLVMValueRef atom_res = lp_build_alloca(gallivm,
1101                                            uint_bld->vec_type, "");
1102
1103    LLVMValueRef exec_mask = mask_vec(bld_base);
1104    if (ssbo_limit) {
1105       LLVMValueRef ssbo_oob_cmp = lp_build_cmp(uint_bld, PIPE_FUNC_LESS, offset, ssbo_limit);
1106       exec_mask = LLVMBuildAnd(builder, exec_mask, ssbo_oob_cmp, "");
1107    }
1108
1109    struct lp_build_loop_state loop_state;
1110    lp_build_loop_begin(&loop_state, gallivm, lp_build_const_int32(gallivm, 0));
1111
1112    LLVMValueRef value_ptr = LLVMBuildExtractElement(gallivm->builder, val,
1113                                                     loop_state.counter, "");
1114    value_ptr = LLVMBuildBitCast(gallivm->builder, value_ptr, uint_bld->elem_type, "");
1115
1116    offset = LLVMBuildExtractElement(gallivm->builder, offset,
1117                                    loop_state.counter, "");
1118
1119    LLVMValueRef scalar_ptr = LLVMBuildGEP(builder, ssbo_ptr,
1120                                           &offset, 1, "");
1121
1122    struct lp_build_if_state ifthen;
1123    LLVMValueRef cond, temp_res;
1124    LLVMValueRef scalar;
1125    cond = LLVMBuildICmp(gallivm->builder, LLVMIntNE, exec_mask, uint_bld->zero, "");
1126    cond = LLVMBuildExtractElement(gallivm->builder, cond, loop_state.counter, "");
1127    lp_build_if(&ifthen, gallivm, cond);
1128
1129    if (nir_op == nir_intrinsic_ssbo_atomic_comp_swap || nir_op == nir_intrinsic_shared_atomic_comp_swap) {
1130       LLVMValueRef cas_src_ptr = LLVMBuildExtractElement(gallivm->builder, val2,
1131                                                          loop_state.counter, "");
1132       cas_src_ptr = LLVMBuildBitCast(gallivm->builder, cas_src_ptr, uint_bld->elem_type, "");
1133       scalar = LLVMBuildAtomicCmpXchg(builder, scalar_ptr, value_ptr,
1134                                       cas_src_ptr,
1135                                       LLVMAtomicOrderingSequentiallyConsistent,
1136                                       LLVMAtomicOrderingSequentiallyConsistent,
1137                                       false);
1138       scalar = LLVMBuildExtractValue(gallivm->builder, scalar, 0, "");
1139    } else {
1140       LLVMAtomicRMWBinOp op;
1141
1142       switch (nir_op) {
1143       case nir_intrinsic_shared_atomic_add:
1144       case nir_intrinsic_ssbo_atomic_add:
1145          op = LLVMAtomicRMWBinOpAdd;
1146          break;
1147       case nir_intrinsic_shared_atomic_exchange:
1148       case nir_intrinsic_ssbo_atomic_exchange:
1149          op = LLVMAtomicRMWBinOpXchg;
1150          break;
1151       case nir_intrinsic_shared_atomic_and:
1152       case nir_intrinsic_ssbo_atomic_and:
1153          op = LLVMAtomicRMWBinOpAnd;
1154          break;
1155       case nir_intrinsic_shared_atomic_or:
1156       case nir_intrinsic_ssbo_atomic_or:
1157          op = LLVMAtomicRMWBinOpOr;
1158          break;
1159       case nir_intrinsic_shared_atomic_xor:
1160       case nir_intrinsic_ssbo_atomic_xor:
1161          op = LLVMAtomicRMWBinOpXor;
1162          break;
1163       case nir_intrinsic_shared_atomic_umin:
1164       case nir_intrinsic_ssbo_atomic_umin:
1165          op = LLVMAtomicRMWBinOpUMin;
1166          break;
1167       case nir_intrinsic_shared_atomic_umax:
1168       case nir_intrinsic_ssbo_atomic_umax:
1169          op = LLVMAtomicRMWBinOpUMax;
1170          break;
1171       case nir_intrinsic_ssbo_atomic_imin:
1172       case nir_intrinsic_shared_atomic_imin:
1173          op = LLVMAtomicRMWBinOpMin;
1174          break;
1175       case nir_intrinsic_ssbo_atomic_imax:
1176       case nir_intrinsic_shared_atomic_imax:
1177          op = LLVMAtomicRMWBinOpMax;
1178          break;
1179       default:
1180          unreachable("unknown atomic op");
1181       }
1182       scalar = LLVMBuildAtomicRMW(builder, op,
1183                                   scalar_ptr, value_ptr,
1184                                   LLVMAtomicOrderingSequentiallyConsistent,
1185                                   false);
1186    }
1187    temp_res = LLVMBuildLoad(builder, atom_res, "");
1188    temp_res = LLVMBuildInsertElement(builder, temp_res, scalar, loop_state.counter, "");
1189    LLVMBuildStore(builder, temp_res, atom_res);
1190    lp_build_else(&ifthen);
1191    temp_res = LLVMBuildLoad(builder, atom_res, "");
1192    temp_res = LLVMBuildInsertElement(builder, temp_res, lp_build_const_int32(gallivm, 0), loop_state.counter, "");
1193    LLVMBuildStore(builder, temp_res, atom_res);
1194    lp_build_endif(&ifthen);
1195
1196    lp_build_loop_end_cond(&loop_state, lp_build_const_int32(gallivm, uint_bld->type.length),
1197                           NULL, LLVMIntUGE);
1198    *result = LLVMBuildLoad(builder, atom_res, "");
1199 }
1200
1201 static void emit_barrier(struct lp_build_nir_context *bld_base)
1202 {
1203    struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
1204    struct gallivm_state * gallivm = bld_base->base.gallivm;
1205
1206    LLVMBasicBlockRef resume = lp_build_insert_new_block(gallivm, "resume");
1207
1208    lp_build_coro_suspend_switch(gallivm, bld->coro, resume, false);
1209    LLVMPositionBuilderAtEnd(gallivm->builder, resume);
1210 }
1211
1212 static LLVMValueRef emit_get_buffer_size(struct lp_build_nir_context *bld_base,
1213                                          LLVMValueRef index)
1214 {
1215    struct gallivm_state *gallivm = bld_base->base.gallivm;
1216    struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
1217    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
1218    struct lp_build_context *bld_broad = &bld_base->uint_bld;
1219    LLVMValueRef size_ptr = lp_build_array_get(bld_base->base.gallivm, bld->ssbo_sizes_ptr,
1220                                               LLVMBuildExtractElement(builder, index, lp_build_const_int32(gallivm, 0), ""));
1221    return lp_build_broadcast_scalar(bld_broad, size_ptr);
1222 }
1223
1224 static void emit_image_op(struct lp_build_nir_context *bld_base,
1225                           struct lp_img_params *params)
1226 {
1227    struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
1228    struct gallivm_state *gallivm = bld_base->base.gallivm;
1229
1230    params->type = bld_base->base.type;
1231    params->context_ptr = bld->context_ptr;
1232    params->thread_data_ptr = bld->thread_data_ptr;
1233    params->exec_mask = mask_vec(bld_base);
1234
1235    if (params->image_index_offset)
1236       params->image_index_offset = LLVMBuildExtractElement(gallivm->builder, params->image_index_offset,
1237                                                            lp_build_const_int32(gallivm, 0), "");
1238
1239    bld->image->emit_op(bld->image,
1240                        bld->bld_base.base.gallivm,
1241                        params);
1242
1243 }
1244
1245 static void emit_image_size(struct lp_build_nir_context *bld_base,
1246                             struct lp_sampler_size_query_params *params)
1247 {
1248    struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
1249    struct gallivm_state *gallivm = bld_base->base.gallivm;
1250
1251    params->int_type = bld_base->int_bld.type;
1252    params->context_ptr = bld->context_ptr;
1253
1254    if (params->texture_unit_offset)
1255       params->texture_unit_offset = LLVMBuildExtractElement(gallivm->builder, params->texture_unit_offset,
1256                                                             lp_build_const_int32(gallivm, 0), "");
1257    bld->image->emit_size_query(bld->image,
1258                                bld->bld_base.base.gallivm,
1259                                params);
1260
1261 }
1262
1263 static void init_var_slots(struct lp_build_nir_context *bld_base,
1264                            nir_variable *var, unsigned sc)
1265 {
1266    struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
1267    unsigned slots = glsl_count_attribute_slots(var->type, false) * 4;
1268
1269    if (!bld->outputs)
1270      return;
1271    for (unsigned comp = sc; comp < slots + sc; comp++) {
1272       unsigned this_loc = var->data.driver_location + (comp / 4);
1273       unsigned this_chan = comp % 4;
1274
1275       if (!bld->outputs[this_loc][this_chan])
1276          bld->outputs[this_loc][this_chan] = lp_build_alloca(bld_base->base.gallivm,
1277                                                              bld_base->base.vec_type, "output");
1278    }
1279 }
1280
1281 static void emit_var_decl(struct lp_build_nir_context *bld_base,
1282                           nir_variable *var)
1283 {
1284    unsigned sc = var->data.location_frac;
1285    switch (var->data.mode) {
1286    case nir_var_shader_out: {
1287       if (bld_base->shader->info.stage == MESA_SHADER_FRAGMENT) {
1288          if (var->data.location == FRAG_RESULT_STENCIL)
1289             sc = 1;
1290          else if (var->data.location == FRAG_RESULT_DEPTH)
1291             sc = 2;
1292       }
1293       init_var_slots(bld_base, var, sc);
1294       break;
1295    }
1296    default:
1297       break;
1298    }
1299 }
1300
1301 static void emit_tex(struct lp_build_nir_context *bld_base,
1302                      struct lp_sampler_params *params)
1303 {
1304    struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
1305    struct gallivm_state *gallivm = bld_base->base.gallivm;
1306
1307    params->type = bld_base->base.type;
1308    params->context_ptr = bld->context_ptr;
1309    params->thread_data_ptr = bld->thread_data_ptr;
1310
1311    if (params->texture_index_offset && bld_base->shader->info.stage != MESA_SHADER_FRAGMENT) {
1312       /* this is horrible but this can be dynamic */
1313       LLVMValueRef coords[5];
1314       LLVMValueRef *orig_texel_ptr;
1315       struct lp_build_context *uint_bld = &bld_base->uint_bld;
1316       LLVMValueRef result[4] = { LLVMGetUndef(bld_base->base.vec_type),
1317                                  LLVMGetUndef(bld_base->base.vec_type),
1318                                  LLVMGetUndef(bld_base->base.vec_type),
1319                                  LLVMGetUndef(bld_base->base.vec_type) };
1320       LLVMValueRef texel[4], orig_offset;
1321       unsigned i;
1322       orig_texel_ptr = params->texel;
1323
1324       for (i = 0; i < 5; i++) {
1325          coords[i] = params->coords[i];
1326       }
1327       orig_offset = params->texture_index_offset;
1328
1329       for (unsigned v = 0; v < uint_bld->type.length; v++) {
1330          LLVMValueRef idx = lp_build_const_int32(gallivm, v);
1331          LLVMValueRef new_coords[5];
1332          for (i = 0; i < 5; i++) {
1333             new_coords[i] = LLVMBuildExtractElement(gallivm->builder,
1334                                                     coords[i], idx, "");
1335          }
1336          params->coords = new_coords;
1337          params->texture_index_offset = LLVMBuildExtractElement(gallivm->builder,
1338                                                                 orig_offset,
1339                                                                 idx, "");
1340          params->type = lp_elem_type(bld_base->base.type);
1341
1342          params->texel = texel;
1343          bld->sampler->emit_tex_sample(bld->sampler,
1344                                        gallivm,
1345                                        params);
1346
1347          for (i = 0; i < 4; i++) {
1348             result[i] = LLVMBuildInsertElement(gallivm->builder, result[i], texel[i], idx, "");
1349          }
1350       }
1351       for (i = 0; i < 4; i++) {
1352          orig_texel_ptr[i] = result[i];
1353       }
1354       return;
1355    }
1356
1357    if (params->texture_index_offset)
1358       params->texture_index_offset = LLVMBuildExtractElement(bld_base->base.gallivm->builder,
1359                                                              params->texture_index_offset,
1360                                                              lp_build_const_int32(bld_base->base.gallivm, 0), "");
1361
1362    params->type = bld_base->base.type;
1363    bld->sampler->emit_tex_sample(bld->sampler,
1364                                  bld->bld_base.base.gallivm,
1365                                  params);
1366 }
1367
1368 static void emit_tex_size(struct lp_build_nir_context *bld_base,
1369                           struct lp_sampler_size_query_params *params)
1370 {
1371    struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
1372
1373    params->int_type = bld_base->int_bld.type;
1374    params->context_ptr = bld->context_ptr;
1375
1376    if (params->texture_unit_offset)
1377       params->texture_unit_offset = LLVMBuildExtractElement(bld_base->base.gallivm->builder,
1378                                                              params->texture_unit_offset,
1379                                                              lp_build_const_int32(bld_base->base.gallivm, 0), "");
1380    bld->sampler->emit_size_query(bld->sampler,
1381                                  bld->bld_base.base.gallivm,
1382                                  params);
1383 }
1384
1385 static void emit_sysval_intrin(struct lp_build_nir_context *bld_base,
1386                                nir_intrinsic_instr *instr,
1387                                LLVMValueRef result[NIR_MAX_VEC_COMPONENTS])
1388 {
1389    struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
1390    struct gallivm_state *gallivm = bld_base->base.gallivm;
1391    switch (instr->intrinsic) {
1392    case nir_intrinsic_load_instance_id:
1393       result[0] = lp_build_broadcast_scalar(&bld_base->uint_bld, bld->system_values.instance_id);
1394       break;
1395    case nir_intrinsic_load_base_instance:
1396       result[0] = lp_build_broadcast_scalar(&bld_base->uint_bld, bld->system_values.base_instance);
1397       break;
1398    case nir_intrinsic_load_base_vertex:
1399       result[0] = bld->system_values.basevertex;
1400       break;
1401    case nir_intrinsic_load_vertex_id:
1402       result[0] = bld->system_values.vertex_id;
1403       break;
1404    case nir_intrinsic_load_primitive_id:
1405       result[0] = bld->system_values.prim_id;
1406       break;
1407    case nir_intrinsic_load_work_group_id:
1408       for (unsigned i = 0; i < 3; i++)
1409          result[i] = lp_build_broadcast_scalar(&bld_base->uint_bld, LLVMBuildExtractElement(gallivm->builder, bld->system_values.block_id, lp_build_const_int32(gallivm, i), ""));
1410       break;
1411    case nir_intrinsic_load_local_invocation_id:
1412       for (unsigned i = 0; i < 3; i++)
1413          result[i] = LLVMBuildExtractValue(gallivm->builder, bld->system_values.thread_id, i, "");
1414       break;
1415    case nir_intrinsic_load_num_work_groups:
1416       for (unsigned i = 0; i < 3; i++)
1417          result[i] = lp_build_broadcast_scalar(&bld_base->uint_bld, LLVMBuildExtractElement(gallivm->builder, bld->system_values.grid_size, lp_build_const_int32(gallivm, i), ""));
1418       break;
1419    case nir_intrinsic_load_invocation_id:
1420       if (bld_base->shader->info.stage == MESA_SHADER_TESS_CTRL)
1421          result[0] = bld->system_values.invocation_id;
1422       else
1423          result[0] = lp_build_broadcast_scalar(&bld_base->uint_bld, bld->system_values.invocation_id);
1424       break;
1425    case nir_intrinsic_load_front_face:
1426       result[0] = lp_build_broadcast_scalar(&bld_base->uint_bld, bld->system_values.front_facing);
1427       break;
1428    case nir_intrinsic_load_draw_id:
1429       result[0] = lp_build_broadcast_scalar(&bld_base->uint_bld, bld->system_values.draw_id);
1430       break;
1431    default:
1432       break;
1433    case nir_intrinsic_load_local_group_size:
1434      for (unsigned i = 0; i < 3; i++)
1435        result[i] = lp_build_broadcast_scalar(&bld_base->uint_bld, LLVMBuildExtractElement(gallivm->builder, bld->system_values.block_size, lp_build_const_int32(gallivm, i), ""));
1436      break;
1437    case nir_intrinsic_load_work_dim:
1438       result[0] = lp_build_broadcast_scalar(&bld_base->uint_bld, bld->system_values.work_dim);
1439       break;
1440    case nir_intrinsic_load_tess_coord:
1441       for (unsigned i = 0; i < 3; i++) {
1442          result[i] = LLVMBuildExtractValue(gallivm->builder, bld->system_values.tess_coord, i, "");
1443       }
1444       break;
1445    case nir_intrinsic_load_tess_level_outer:
1446       for (unsigned i = 0; i < 4; i++)
1447          result[i] = lp_build_broadcast_scalar(&bld_base->base, LLVMBuildExtractValue(gallivm->builder, bld->system_values.tess_outer, i, ""));
1448       break;
1449    case nir_intrinsic_load_tess_level_inner:
1450       for (unsigned i = 0; i < 2; i++)
1451          result[i] = lp_build_broadcast_scalar(&bld_base->base, LLVMBuildExtractValue(gallivm->builder, bld->system_values.tess_inner, i, ""));
1452       break;
1453    case nir_intrinsic_load_patch_vertices_in:
1454       result[0] = bld->system_values.vertices_in;
1455       break;
1456    case nir_intrinsic_load_sample_id:
1457       result[0] = lp_build_broadcast_scalar(&bld_base->uint_bld, bld->system_values.sample_id);
1458       break;
1459    case nir_intrinsic_load_sample_pos:
1460       for (unsigned i = 0; i < 2; i++) {
1461          LLVMValueRef idx = LLVMBuildMul(gallivm->builder, bld->system_values.sample_id, lp_build_const_int32(gallivm, 2), "");
1462          idx = LLVMBuildAdd(gallivm->builder, idx, lp_build_const_int32(gallivm, i), "");
1463          LLVMValueRef val = lp_build_array_get(gallivm, bld->system_values.sample_pos, idx);
1464          result[i] = lp_build_broadcast_scalar(&bld_base->base, val);
1465       }
1466       break;
1467    case nir_intrinsic_load_sample_mask_in:
1468       result[0] = bld->system_values.sample_mask_in;
1469       break;
1470    }
1471 }
1472
1473 static void emit_helper_invocation(struct lp_build_nir_context *bld_base,
1474                                    LLVMValueRef *dst)
1475 {
1476    struct gallivm_state *gallivm = bld_base->base.gallivm;
1477    struct lp_build_context *uint_bld = &bld_base->uint_bld;
1478    *dst = lp_build_cmp(uint_bld, PIPE_FUNC_NOTEQUAL, mask_vec(bld_base), lp_build_const_int_vec(gallivm, uint_bld->type, -1));
1479 }
1480
1481 static void bgnloop(struct lp_build_nir_context *bld_base)
1482 {
1483    struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
1484    lp_exec_bgnloop(&bld->exec_mask, true);
1485 }
1486
1487 static void endloop(struct lp_build_nir_context *bld_base)
1488 {
1489    struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
1490    lp_exec_endloop(bld_base->base.gallivm, &bld->exec_mask);
1491 }
1492
1493 static void if_cond(struct lp_build_nir_context *bld_base, LLVMValueRef cond)
1494 {
1495    LLVMBuilderRef builder = bld_base->base.gallivm->builder;
1496    struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
1497    lp_exec_mask_cond_push(&bld->exec_mask, LLVMBuildBitCast(builder, cond, bld_base->base.int_vec_type, ""));
1498 }
1499
1500 static void else_stmt(struct lp_build_nir_context *bld_base)
1501 {
1502    struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
1503    lp_exec_mask_cond_invert(&bld->exec_mask);
1504 }
1505
1506 static void endif_stmt(struct lp_build_nir_context *bld_base)
1507 {
1508    struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
1509    lp_exec_mask_cond_pop(&bld->exec_mask);
1510 }
1511
1512 static void break_stmt(struct lp_build_nir_context *bld_base)
1513 {
1514    struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
1515
1516    lp_exec_break(&bld->exec_mask, NULL, false);
1517 }
1518
1519 static void continue_stmt(struct lp_build_nir_context *bld_base)
1520 {
1521    struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
1522    lp_exec_continue(&bld->exec_mask);
1523 }
1524
1525 static void discard(struct lp_build_nir_context *bld_base, LLVMValueRef cond)
1526 {
1527    struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
1528    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
1529    LLVMValueRef mask;
1530
1531    if (!cond) {
1532       if (bld->exec_mask.has_mask) {
1533          mask = LLVMBuildNot(builder, bld->exec_mask.exec_mask, "kilp");
1534       } else {
1535          mask = LLVMConstNull(bld->bld_base.base.int_vec_type);
1536       }
1537    } else {
1538       mask = LLVMBuildNot(builder, cond, "");
1539       if (bld->exec_mask.has_mask) {
1540          LLVMValueRef invmask;
1541          invmask = LLVMBuildNot(builder, bld->exec_mask.exec_mask, "kilp");
1542          mask = LLVMBuildOr(builder, mask, invmask, "");
1543       }
1544    }
1545    lp_build_mask_update(bld->mask, mask);
1546 }
1547
1548 static void
1549 increment_vec_ptr_by_mask(struct lp_build_nir_context * bld_base,
1550                           LLVMValueRef ptr,
1551                           LLVMValueRef mask)
1552 {
1553    LLVMBuilderRef builder = bld_base->base.gallivm->builder;
1554    LLVMValueRef current_vec = LLVMBuildLoad(builder, ptr, "");
1555
1556    current_vec = LLVMBuildSub(builder, current_vec, mask, "");
1557
1558    LLVMBuildStore(builder, current_vec, ptr);
1559 }
1560
1561 static void
1562 clear_uint_vec_ptr_from_mask(struct lp_build_nir_context * bld_base,
1563                              LLVMValueRef ptr,
1564                              LLVMValueRef mask)
1565 {
1566    LLVMBuilderRef builder = bld_base->base.gallivm->builder;
1567    LLVMValueRef current_vec = LLVMBuildLoad(builder, ptr, "");
1568
1569    current_vec = lp_build_select(&bld_base->uint_bld,
1570                                  mask,
1571                                  bld_base->uint_bld.zero,
1572                                  current_vec);
1573
1574    LLVMBuildStore(builder, current_vec, ptr);
1575 }
1576
1577 static LLVMValueRef
1578 clamp_mask_to_max_output_vertices(struct lp_build_nir_soa_context * bld,
1579                                   LLVMValueRef current_mask_vec,
1580                                   LLVMValueRef total_emitted_vertices_vec)
1581 {
1582    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
1583    struct lp_build_context *int_bld = &bld->bld_base.int_bld;
1584    LLVMValueRef max_mask = lp_build_cmp(int_bld, PIPE_FUNC_LESS,
1585                                             total_emitted_vertices_vec,
1586                                             bld->max_output_vertices_vec);
1587
1588    return LLVMBuildAnd(builder, current_mask_vec, max_mask, "");
1589 }
1590
1591 static void emit_vertex(struct lp_build_nir_context *bld_base, uint32_t stream_id)
1592 {
1593    struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
1594    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
1595
1596    assert(bld->gs_iface->emit_vertex);
1597    LLVMValueRef total_emitted_vertices_vec =
1598       LLVMBuildLoad(builder, bld->total_emitted_vertices_vec_ptr[stream_id], "");
1599    LLVMValueRef mask = mask_vec(bld_base);
1600    mask = clamp_mask_to_max_output_vertices(bld, mask,
1601                                             total_emitted_vertices_vec);
1602    bld->gs_iface->emit_vertex(bld->gs_iface, &bld->bld_base.base,
1603                               bld->outputs,
1604                               total_emitted_vertices_vec,
1605                               lp_build_const_int_vec(bld->bld_base.base.gallivm, bld->bld_base.base.type, stream_id));
1606
1607    increment_vec_ptr_by_mask(bld_base, bld->emitted_vertices_vec_ptr[stream_id],
1608                              mask);
1609    increment_vec_ptr_by_mask(bld_base, bld->total_emitted_vertices_vec_ptr[stream_id],
1610                              mask);
1611 }
1612
1613 static void
1614 end_primitive_masked(struct lp_build_nir_context * bld_base,
1615                      LLVMValueRef mask, uint32_t stream_id)
1616 {
1617    struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
1618    LLVMBuilderRef builder = bld->bld_base.base.gallivm->builder;
1619
1620    struct lp_build_context *uint_bld = &bld_base->uint_bld;
1621    LLVMValueRef emitted_vertices_vec =
1622       LLVMBuildLoad(builder, bld->emitted_vertices_vec_ptr[stream_id], "");
1623    LLVMValueRef emitted_prims_vec =
1624       LLVMBuildLoad(builder, bld->emitted_prims_vec_ptr[stream_id], "");
1625    LLVMValueRef total_emitted_vertices_vec =
1626       LLVMBuildLoad(builder, bld->total_emitted_vertices_vec_ptr[stream_id], "");
1627
1628    LLVMValueRef emitted_mask = lp_build_cmp(uint_bld,
1629                                             PIPE_FUNC_NOTEQUAL,
1630                                             emitted_vertices_vec,
1631                                             uint_bld->zero);
1632    mask = LLVMBuildAnd(builder, mask, emitted_mask, "");
1633    if (stream_id == 0)
1634       bld->gs_iface->end_primitive(bld->gs_iface, &bld->bld_base.base,
1635                                    total_emitted_vertices_vec,
1636                                    emitted_vertices_vec, emitted_prims_vec, mask);
1637    increment_vec_ptr_by_mask(bld_base, bld->emitted_prims_vec_ptr[stream_id],
1638                              mask);
1639    clear_uint_vec_ptr_from_mask(bld_base, bld->emitted_vertices_vec_ptr[stream_id],
1640                                 mask);
1641 }
1642
1643 static void end_primitive(struct lp_build_nir_context *bld_base, uint32_t stream_id)
1644 {
1645    ASSERTED struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
1646
1647    assert(bld->gs_iface->end_primitive);
1648
1649    LLVMValueRef mask = mask_vec(bld_base);
1650    end_primitive_masked(bld_base, mask, stream_id);
1651 }
1652
1653 static void
1654 emit_prologue(struct lp_build_nir_soa_context *bld)
1655 {
1656    struct gallivm_state * gallivm = bld->bld_base.base.gallivm;
1657    if (bld->indirects & nir_var_shader_in && !bld->gs_iface && !bld->tcs_iface && !bld->tes_iface) {
1658       uint32_t num_inputs = util_bitcount64(bld->bld_base.shader->info.inputs_read);
1659       unsigned index, chan;
1660       LLVMTypeRef vec_type = bld->bld_base.base.vec_type;
1661       LLVMValueRef array_size = lp_build_const_int32(gallivm, num_inputs * 4);
1662       bld->inputs_array = lp_build_array_alloca(gallivm,
1663                                                vec_type, array_size,
1664                                                "input_array");
1665
1666       for (index = 0; index < num_inputs; ++index) {
1667          for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
1668             LLVMValueRef lindex =
1669                lp_build_const_int32(gallivm, index * 4 + chan);
1670             LLVMValueRef input_ptr =
1671                LLVMBuildGEP(gallivm->builder, bld->inputs_array,
1672                             &lindex, 1, "");
1673             LLVMValueRef value = bld->inputs[index][chan];
1674             if (value)
1675                LLVMBuildStore(gallivm->builder, value, input_ptr);
1676          }
1677       }
1678    }
1679 }
1680
1681 static void emit_vote(struct lp_build_nir_context *bld_base, LLVMValueRef src, nir_intrinsic_instr *instr, LLVMValueRef result[4])
1682 {
1683    struct gallivm_state * gallivm = bld_base->base.gallivm;
1684    LLVMBuilderRef builder = gallivm->builder;
1685
1686    LLVMValueRef exec_mask = mask_vec(bld_base);
1687    struct lp_build_loop_state loop_state;
1688
1689    LLVMValueRef outer_cond = LLVMBuildICmp(builder, LLVMIntNE, exec_mask, bld_base->uint_bld.zero, "");
1690
1691    LLVMValueRef res_store = lp_build_alloca(gallivm, bld_base->int_bld.elem_type, "");
1692    LLVMValueRef init_val = NULL;
1693    if (instr->intrinsic == nir_intrinsic_vote_ieq) {
1694       /* for equal we unfortunately have to loop and find the first valid one. */
1695       lp_build_loop_begin(&loop_state, gallivm, lp_build_const_int32(gallivm, 0));
1696       LLVMValueRef if_cond = LLVMBuildExtractElement(gallivm->builder, outer_cond, loop_state.counter, "");
1697
1698       struct lp_build_if_state ifthen;
1699       lp_build_if(&ifthen, gallivm, if_cond);
1700       LLVMValueRef value_ptr = LLVMBuildExtractElement(gallivm->builder, src,
1701                                                        loop_state.counter, "");
1702       LLVMBuildStore(builder, value_ptr, res_store);
1703       lp_build_endif(&ifthen);
1704       lp_build_loop_end_cond(&loop_state, lp_build_const_int32(gallivm, bld_base->uint_bld.type.length),
1705                              NULL, LLVMIntUGE);
1706       init_val = LLVMBuildLoad(builder, res_store, "");
1707    } else {
1708       LLVMBuildStore(builder, lp_build_const_int32(gallivm, instr->intrinsic == nir_intrinsic_vote_any ? 0 : -1), res_store);
1709    }
1710
1711    LLVMValueRef res;
1712    lp_build_loop_begin(&loop_state, gallivm, lp_build_const_int32(gallivm, 0));
1713    LLVMValueRef value_ptr = LLVMBuildExtractElement(gallivm->builder, src,
1714                                                        loop_state.counter, "");
1715    struct lp_build_if_state ifthen;
1716    LLVMValueRef if_cond;
1717    if_cond = LLVMBuildExtractElement(gallivm->builder, outer_cond, loop_state.counter, "");
1718
1719    lp_build_if(&ifthen, gallivm, if_cond);
1720    res = LLVMBuildLoad(builder, res_store, "");
1721
1722    if (instr->intrinsic == nir_intrinsic_vote_ieq) {
1723       LLVMValueRef tmp = LLVMBuildICmp(builder, LLVMIntEQ, init_val, value_ptr, "");
1724       tmp = LLVMBuildSExt(builder, tmp, bld_base->uint_bld.elem_type, "");
1725       res = LLVMBuildOr(builder, res, tmp, "");
1726    } else if (instr->intrinsic == nir_intrinsic_vote_any)
1727       res = LLVMBuildOr(builder, res, value_ptr, "");
1728    else
1729       res = LLVMBuildAnd(builder, res, value_ptr, "");
1730    LLVMBuildStore(builder, res, res_store);
1731    lp_build_endif(&ifthen);
1732    lp_build_loop_end_cond(&loop_state, lp_build_const_int32(gallivm, bld_base->uint_bld.type.length),
1733                           NULL, LLVMIntUGE);
1734    result[0] = lp_build_broadcast_scalar(&bld_base->uint_bld, LLVMBuildLoad(builder, res_store, ""));
1735 }
1736
1737 static void
1738 emit_interp_at(struct lp_build_nir_context *bld_base,
1739                unsigned num_components,
1740                nir_variable *var,
1741                bool centroid,
1742                bool sample,
1743                unsigned const_index,
1744                LLVMValueRef indir_index,
1745                LLVMValueRef offsets[2],
1746                LLVMValueRef dst[4])
1747 {
1748    struct lp_build_nir_soa_context *bld = (struct lp_build_nir_soa_context *)bld_base;
1749
1750    for (unsigned i = 0; i < num_components; i++) {
1751       dst[i] = bld->fs_iface->interp_fn(bld->fs_iface, &bld_base->base,
1752                                         const_index + var->data.driver_location, i + var->data.location_frac,
1753                                         centroid, sample, indir_index, offsets);
1754    }
1755 }
1756
1757 void lp_build_nir_soa(struct gallivm_state *gallivm,
1758                       struct nir_shader *shader,
1759                       const struct lp_build_tgsi_params *params,
1760                       LLVMValueRef (*outputs)[4])
1761 {
1762    struct lp_build_nir_soa_context bld;
1763    struct lp_type type = params->type;
1764    struct lp_type res_type;
1765
1766    assert(type.length <= LP_MAX_VECTOR_LENGTH);
1767    memset(&res_type, 0, sizeof res_type);
1768    res_type.width = type.width;
1769    res_type.length = type.length;
1770    res_type.sign = 1;
1771
1772    /* Setup build context */
1773    memset(&bld, 0, sizeof bld);
1774    lp_build_context_init(&bld.bld_base.base, gallivm, type);
1775    lp_build_context_init(&bld.bld_base.uint_bld, gallivm, lp_uint_type(type));
1776    lp_build_context_init(&bld.bld_base.int_bld, gallivm, lp_int_type(type));
1777    lp_build_context_init(&bld.elem_bld, gallivm, lp_elem_type(type));
1778    lp_build_context_init(&bld.uint_elem_bld, gallivm, lp_elem_type(lp_uint_type(type)));
1779    {
1780       struct lp_type dbl_type;
1781       dbl_type = type;
1782       dbl_type.width *= 2;
1783       lp_build_context_init(&bld.bld_base.dbl_bld, gallivm, dbl_type);
1784    }
1785    {
1786       struct lp_type uint64_type;
1787       uint64_type = lp_uint_type(type);
1788       uint64_type.width *= 2;
1789       lp_build_context_init(&bld.bld_base.uint64_bld, gallivm, uint64_type);
1790    }
1791    {
1792       struct lp_type int64_type;
1793       int64_type = lp_int_type(type);
1794       int64_type.width *= 2;
1795       lp_build_context_init(&bld.bld_base.int64_bld, gallivm, int64_type);
1796    }
1797    {
1798       struct lp_type uint16_type;
1799       uint16_type = lp_uint_type(type);
1800       uint16_type.width /= 2;
1801       lp_build_context_init(&bld.bld_base.uint16_bld, gallivm, uint16_type);
1802    }
1803    {
1804       struct lp_type int16_type;
1805       int16_type = lp_int_type(type);
1806       int16_type.width /= 2;
1807       lp_build_context_init(&bld.bld_base.int16_bld, gallivm, int16_type);
1808    }
1809    {
1810       struct lp_type uint8_type;
1811       uint8_type = lp_uint_type(type);
1812       uint8_type.width /= 4;
1813       lp_build_context_init(&bld.bld_base.uint8_bld, gallivm, uint8_type);
1814    }
1815    {
1816       struct lp_type int8_type;
1817       int8_type = lp_int_type(type);
1818       int8_type.width /= 4;
1819       lp_build_context_init(&bld.bld_base.int8_bld, gallivm, int8_type);
1820    }
1821    bld.bld_base.load_var = emit_load_var;
1822    bld.bld_base.store_var = emit_store_var;
1823    bld.bld_base.load_reg = emit_load_reg;
1824    bld.bld_base.store_reg = emit_store_reg;
1825    bld.bld_base.emit_var_decl = emit_var_decl;
1826    bld.bld_base.load_ubo = emit_load_ubo;
1827    bld.bld_base.load_kernel_arg = emit_load_kernel_arg;
1828    bld.bld_base.load_global = emit_load_global;
1829    bld.bld_base.store_global = emit_store_global;
1830    bld.bld_base.atomic_global = emit_atomic_global;
1831    bld.bld_base.tex = emit_tex;
1832    bld.bld_base.tex_size = emit_tex_size;
1833    bld.bld_base.bgnloop = bgnloop;
1834    bld.bld_base.endloop = endloop;
1835    bld.bld_base.if_cond = if_cond;
1836    bld.bld_base.else_stmt = else_stmt;
1837    bld.bld_base.endif_stmt = endif_stmt;
1838    bld.bld_base.break_stmt = break_stmt;
1839    bld.bld_base.continue_stmt = continue_stmt;
1840    bld.bld_base.sysval_intrin = emit_sysval_intrin;
1841    bld.bld_base.discard = discard;
1842    bld.bld_base.emit_vertex = emit_vertex;
1843    bld.bld_base.end_primitive = end_primitive;
1844    bld.bld_base.load_mem = emit_load_mem;
1845    bld.bld_base.store_mem = emit_store_mem;
1846    bld.bld_base.get_buffer_size = emit_get_buffer_size;
1847    bld.bld_base.atomic_mem = emit_atomic_mem;
1848    bld.bld_base.barrier = emit_barrier;
1849    bld.bld_base.image_op = emit_image_op;
1850    bld.bld_base.image_size = emit_image_size;
1851    bld.bld_base.vote = emit_vote;
1852    bld.bld_base.helper_invocation = emit_helper_invocation;
1853    bld.bld_base.interp_at = emit_interp_at;
1854
1855    bld.mask = params->mask;
1856    bld.inputs = params->inputs;
1857    bld.outputs = outputs;
1858    bld.consts_ptr = params->consts_ptr;
1859    bld.const_sizes_ptr = params->const_sizes_ptr;
1860    bld.ssbo_ptr = params->ssbo_ptr;
1861    bld.ssbo_sizes_ptr = params->ssbo_sizes_ptr;
1862    bld.sampler = params->sampler;
1863 //   bld.bld_base.info = params->info;
1864
1865    bld.context_ptr = params->context_ptr;
1866    bld.thread_data_ptr = params->thread_data_ptr;
1867    bld.image = params->image;
1868    bld.shared_ptr = params->shared_ptr;
1869    bld.coro = params->coro;
1870    bld.kernel_args_ptr = params->kernel_args;
1871    bld.indirects = 0;
1872    if (params->info->indirect_files & (1 << TGSI_FILE_INPUT))
1873       bld.indirects |= nir_var_shader_in;
1874
1875    bld.gs_iface = params->gs_iface;
1876    bld.tcs_iface = params->tcs_iface;
1877    bld.tes_iface = params->tes_iface;
1878    bld.fs_iface = params->fs_iface;
1879    if (bld.gs_iface) {
1880       struct lp_build_context *uint_bld = &bld.bld_base.uint_bld;
1881
1882       bld.max_output_vertices_vec = lp_build_const_int_vec(gallivm, bld.bld_base.int_bld.type,
1883                                                            shader->info.gs.vertices_out);
1884       for (int i = 0; i < PIPE_MAX_VERTEX_STREAMS; i++) {
1885          bld.emitted_prims_vec_ptr[i] =
1886             lp_build_alloca(gallivm, uint_bld->vec_type, "emitted_prims_ptr");
1887          bld.emitted_vertices_vec_ptr[i] =
1888             lp_build_alloca(gallivm, uint_bld->vec_type, "emitted_vertices_ptr");
1889          bld.total_emitted_vertices_vec_ptr[i] =
1890             lp_build_alloca(gallivm, uint_bld->vec_type, "total_emitted_vertices_ptr");
1891       }
1892    }
1893    lp_exec_mask_init(&bld.exec_mask, &bld.bld_base.int_bld);
1894
1895    bld.system_values = *params->system_values;
1896
1897    bld.bld_base.shader = shader;
1898
1899    emit_prologue(&bld);
1900    lp_build_nir_llvm(&bld.bld_base, shader);
1901
1902    if (bld.gs_iface) {
1903       LLVMBuilderRef builder = bld.bld_base.base.gallivm->builder;
1904       LLVMValueRef total_emitted_vertices_vec;
1905       LLVMValueRef emitted_prims_vec;
1906
1907       end_primitive_masked(&bld.bld_base, lp_build_mask_value(bld.mask), 0);
1908       for (int i = 0; i < PIPE_MAX_VERTEX_STREAMS; i++) {
1909          total_emitted_vertices_vec =
1910             LLVMBuildLoad(builder, bld.total_emitted_vertices_vec_ptr[i], "");
1911
1912          emitted_prims_vec =
1913             LLVMBuildLoad(builder, bld.emitted_prims_vec_ptr[i], "");
1914          bld.gs_iface->gs_epilogue(bld.gs_iface,
1915                                    total_emitted_vertices_vec,
1916                                    emitted_prims_vec, i);
1917       }
1918    }
1919    lp_exec_mask_fini(&bld.exec_mask);
1920 }