src/gallium/drivers/radeonsi/si_shader_tgsi_alu.c

   1 /*
   2  * Copyright 2016 Advanced Micro Devices, Inc.
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * on the rights to use, copy, modify, merge, publish, distribute, sub
   8  * license, and/or sell copies of the Software, and to permit persons to whom
   9  * the Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
  18  * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
  19  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  20  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  21  * USE OR OTHER DEALINGS IN THE SOFTWARE.
  22  */
  23
  24 #include "si_shader_internal.h"
  25 #include "gallivm/lp_bld_const.h"
  26 #include "gallivm/lp_bld_intr.h"
  27 #include "gallivm/lp_bld_gather.h"
  28 #include "tgsi/tgsi_parse.h"
  29
  30 static void kill_if_fetch_args(struct lp_build_tgsi_context *bld_base,
  31                                struct lp_build_emit_data *emit_data)
  32 {
  33         const struct tgsi_full_instruction *inst = emit_data->inst;
  34         struct gallivm_state *gallivm = bld_base->base.gallivm;
  35         LLVMBuilderRef builder = gallivm->builder;
  36         unsigned i;
  37         LLVMValueRef conds[TGSI_NUM_CHANNELS];
  38
  39         for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
  40                 LLVMValueRef value = lp_build_emit_fetch(bld_base, inst, 0, i);
  41                 conds[i] = LLVMBuildFCmp(builder, LLVMRealOLT, value,
  42                                         bld_base->base.zero, "");
  43         }
  44
  45         /* Or the conditions together */
  46         for (i = TGSI_NUM_CHANNELS - 1; i > 0; i--) {
  47                 conds[i - 1] = LLVMBuildOr(builder, conds[i], conds[i - 1], "");
  48         }
  49
  50         emit_data->dst_type = LLVMVoidTypeInContext(gallivm->context);
  51         emit_data->arg_count = 1;
  52         emit_data->args[0] = LLVMBuildSelect(builder, conds[0],
  53                                         lp_build_const_float(gallivm, -1.0f),
  54                                         bld_base->base.zero, "");
  55 }
  56
  57 static void kil_emit(const struct lp_build_tgsi_action *action,
  58                      struct lp_build_tgsi_context *bld_base,
  59                      struct lp_build_emit_data *emit_data)
  60 {
  61         unsigned i;
  62         for (i = 0; i < emit_data->arg_count; i++) {
  63                 emit_data->output[i] = lp_build_intrinsic_unary(
  64                         bld_base->base.gallivm->builder,
  65                         action->intr_name,
  66                         emit_data->dst_type, emit_data->args[i]);
  67         }
  68 }
  69
  70 static void emit_icmp(const struct lp_build_tgsi_action *action,
  71                       struct lp_build_tgsi_context *bld_base,
  72                       struct lp_build_emit_data *emit_data)
  73 {
  74         unsigned pred;
  75         LLVMBuilderRef builder = bld_base->base.gallivm->builder;
  76         LLVMContextRef context = bld_base->base.gallivm->context;
  77
  78         switch (emit_data->inst->Instruction.Opcode) {
  79         case TGSI_OPCODE_USEQ:
  80         case TGSI_OPCODE_U64SEQ: pred = LLVMIntEQ; break;
  81         case TGSI_OPCODE_USNE:
  82         case TGSI_OPCODE_U64SNE: pred = LLVMIntNE; break;
  83         case TGSI_OPCODE_USGE:
  84         case TGSI_OPCODE_U64SGE: pred = LLVMIntUGE; break;
  85         case TGSI_OPCODE_USLT:
  86         case TGSI_OPCODE_U64SLT: pred = LLVMIntULT; break;
  87         case TGSI_OPCODE_ISGE:
  88         case TGSI_OPCODE_I64SGE: pred = LLVMIntSGE; break;
  89         case TGSI_OPCODE_ISLT:
  90         case TGSI_OPCODE_I64SLT: pred = LLVMIntSLT; break;
  91         default:
  92                 assert(!"unknown instruction");
  93                 pred = 0;
  94                 break;
  95         }
  96
  97         LLVMValueRef v = LLVMBuildICmp(builder, pred,
  98                         emit_data->args[0], emit_data->args[1],"");
  99
 100         v = LLVMBuildSExtOrBitCast(builder, v,
 101                         LLVMInt32TypeInContext(context), "");
 102
 103         emit_data->output[emit_data->chan] = v;
 104 }
 105
 106 static void emit_ucmp(const struct lp_build_tgsi_action *action,
 107                       struct lp_build_tgsi_context *bld_base,
 108                       struct lp_build_emit_data *emit_data)
 109 {
 110         LLVMBuilderRef builder = bld_base->base.gallivm->builder;
 111
 112         LLVMValueRef arg0 = LLVMBuildBitCast(builder, emit_data->args[0],
 113                                              bld_base->uint_bld.elem_type, "");
 114
 115         LLVMValueRef v = LLVMBuildICmp(builder, LLVMIntNE, arg0,
 116                                        bld_base->uint_bld.zero, "");
 117
 118         emit_data->output[emit_data->chan] =
 119                 LLVMBuildSelect(builder, v, emit_data->args[1], emit_data->args[2], "");
 120 }
 121
 122 static void emit_cmp(const struct lp_build_tgsi_action *action,
 123                      struct lp_build_tgsi_context *bld_base,
 124                      struct lp_build_emit_data *emit_data)
 125 {
 126         LLVMBuilderRef builder = bld_base->base.gallivm->builder;
 127         LLVMValueRef cond, *args = emit_data->args;
 128
 129         cond = LLVMBuildFCmp(builder, LLVMRealOLT, args[0],
 130                              bld_base->base.zero, "");
 131
 132         emit_data->output[emit_data->chan] =
 133                 LLVMBuildSelect(builder, cond, args[1], args[2], "");
 134 }
 135
 136 static void emit_set_cond(const struct lp_build_tgsi_action *action,
 137                           struct lp_build_tgsi_context *bld_base,
 138                           struct lp_build_emit_data *emit_data)
 139 {
 140         LLVMBuilderRef builder = bld_base->base.gallivm->builder;
 141         LLVMRealPredicate pred;
 142         LLVMValueRef cond;
 143
 144         /* Use ordered for everything but NE (which is usual for
 145          * float comparisons)
 146          */
 147         switch (emit_data->inst->Instruction.Opcode) {
 148         case TGSI_OPCODE_SGE: pred = LLVMRealOGE; break;
 149         case TGSI_OPCODE_SEQ: pred = LLVMRealOEQ; break;
 150         case TGSI_OPCODE_SLE: pred = LLVMRealOLE; break;
 151         case TGSI_OPCODE_SLT: pred = LLVMRealOLT; break;
 152         case TGSI_OPCODE_SNE: pred = LLVMRealUNE; break;
 153         case TGSI_OPCODE_SGT: pred = LLVMRealOGT; break;
 154         default: assert(!"unknown instruction"); pred = 0; break;
 155         }
 156
 157         cond = LLVMBuildFCmp(builder,
 158                 pred, emit_data->args[0], emit_data->args[1], "");
 159
 160         emit_data->output[emit_data->chan] = LLVMBuildSelect(builder,
 161                 cond, bld_base->base.one, bld_base->base.zero, "");
 162 }
 163
 164 static void emit_fcmp(const struct lp_build_tgsi_action *action,
 165                       struct lp_build_tgsi_context *bld_base,
 166                       struct lp_build_emit_data *emit_data)
 167 {
 168         LLVMBuilderRef builder = bld_base->base.gallivm->builder;
 169         LLVMContextRef context = bld_base->base.gallivm->context;
 170         LLVMRealPredicate pred;
 171
 172         /* Use ordered for everything but NE (which is usual for
 173          * float comparisons)
 174          */
 175         switch (emit_data->inst->Instruction.Opcode) {
 176         case TGSI_OPCODE_FSEQ: pred = LLVMRealOEQ; break;
 177         case TGSI_OPCODE_FSGE: pred = LLVMRealOGE; break;
 178         case TGSI_OPCODE_FSLT: pred = LLVMRealOLT; break;
 179         case TGSI_OPCODE_FSNE: pred = LLVMRealUNE; break;
 180         default: assert(!"unknown instruction"); pred = 0; break;
 181         }
 182
 183         LLVMValueRef v = LLVMBuildFCmp(builder, pred,
 184                         emit_data->args[0], emit_data->args[1],"");
 185
 186         v = LLVMBuildSExtOrBitCast(builder, v,
 187                         LLVMInt32TypeInContext(context), "");
 188
 189         emit_data->output[emit_data->chan] = v;
 190 }
 191
 192 static void emit_dcmp(const struct lp_build_tgsi_action *action,
 193                       struct lp_build_tgsi_context *bld_base,
 194                       struct lp_build_emit_data *emit_data)
 195 {
 196         LLVMBuilderRef builder = bld_base->base.gallivm->builder;
 197         LLVMContextRef context = bld_base->base.gallivm->context;
 198         LLVMRealPredicate pred;
 199
 200         /* Use ordered for everything but NE (which is usual for
 201          * float comparisons)
 202          */
 203         switch (emit_data->inst->Instruction.Opcode) {
 204         case TGSI_OPCODE_DSEQ: pred = LLVMRealOEQ; break;
 205         case TGSI_OPCODE_DSGE: pred = LLVMRealOGE; break;
 206         case TGSI_OPCODE_DSLT: pred = LLVMRealOLT; break;
 207         case TGSI_OPCODE_DSNE: pred = LLVMRealUNE; break;
 208         default: assert(!"unknown instruction"); pred = 0; break;
 209         }
 210
 211         LLVMValueRef v = LLVMBuildFCmp(builder, pred,
 212                         emit_data->args[0], emit_data->args[1],"");
 213
 214         v = LLVMBuildSExtOrBitCast(builder, v,
 215                         LLVMInt32TypeInContext(context), "");
 216
 217         emit_data->output[emit_data->chan] = v;
 218 }
 219
 220 static void emit_not(const struct lp_build_tgsi_action *action,
 221                      struct lp_build_tgsi_context *bld_base,
 222                      struct lp_build_emit_data *emit_data)
 223 {
 224         LLVMBuilderRef builder = bld_base->base.gallivm->builder;
 225         LLVMValueRef v = bitcast(bld_base, TGSI_TYPE_UNSIGNED,
 226                         emit_data->args[0]);
 227         emit_data->output[emit_data->chan] = LLVMBuildNot(builder, v, "");
 228 }
 229
 230 static void emit_arl(const struct lp_build_tgsi_action *action,
 231                      struct lp_build_tgsi_context *bld_base,
 232                      struct lp_build_emit_data *emit_data)
 233 {
 234         LLVMBuilderRef builder = bld_base->base.gallivm->builder;
 235         LLVMValueRef floor_index =  lp_build_emit_llvm_unary(bld_base, TGSI_OPCODE_FLR, emit_data->args[0]);
 236         emit_data->output[emit_data->chan] = LLVMBuildFPToSI(builder,
 237                         floor_index, bld_base->base.int_elem_type , "");
 238 }
 239
 240 static void emit_and(const struct lp_build_tgsi_action *action,
 241                      struct lp_build_tgsi_context *bld_base,
 242                      struct lp_build_emit_data *emit_data)
 243 {
 244         LLVMBuilderRef builder = bld_base->base.gallivm->builder;
 245         emit_data->output[emit_data->chan] = LLVMBuildAnd(builder,
 246                         emit_data->args[0], emit_data->args[1], "");
 247 }
 248
 249 static void emit_or(const struct lp_build_tgsi_action *action,
 250                     struct lp_build_tgsi_context *bld_base,
 251                     struct lp_build_emit_data *emit_data)
 252 {
 253         LLVMBuilderRef builder = bld_base->base.gallivm->builder;
 254         emit_data->output[emit_data->chan] = LLVMBuildOr(builder,
 255                         emit_data->args[0], emit_data->args[1], "");
 256 }
 257
 258 static void emit_uadd(const struct lp_build_tgsi_action *action,
 259                       struct lp_build_tgsi_context *bld_base,
 260                       struct lp_build_emit_data *emit_data)
 261 {
 262         LLVMBuilderRef builder = bld_base->base.gallivm->builder;
 263         emit_data->output[emit_data->chan] = LLVMBuildAdd(builder,
 264                         emit_data->args[0], emit_data->args[1], "");
 265 }
 266
 267 static void emit_udiv(const struct lp_build_tgsi_action *action,
 268                       struct lp_build_tgsi_context *bld_base,
 269                       struct lp_build_emit_data *emit_data)
 270 {
 271         LLVMBuilderRef builder = bld_base->base.gallivm->builder;
 272         emit_data->output[emit_data->chan] = LLVMBuildUDiv(builder,
 273                         emit_data->args[0], emit_data->args[1], "");
 274 }
 275
 276 static void emit_idiv(const struct lp_build_tgsi_action *action,
 277                       struct lp_build_tgsi_context *bld_base,
 278                       struct lp_build_emit_data *emit_data)
 279 {
 280         LLVMBuilderRef builder = bld_base->base.gallivm->builder;
 281         emit_data->output[emit_data->chan] = LLVMBuildSDiv(builder,
 282                         emit_data->args[0], emit_data->args[1], "");
 283 }
 284
 285 static void emit_mod(const struct lp_build_tgsi_action *action,
 286                      struct lp_build_tgsi_context *bld_base,
 287                      struct lp_build_emit_data *emit_data)
 288 {
 289         LLVMBuilderRef builder = bld_base->base.gallivm->builder;
 290         emit_data->output[emit_data->chan] = LLVMBuildSRem(builder,
 291                         emit_data->args[0], emit_data->args[1], "");
 292 }
 293
 294 static void emit_umod(const struct lp_build_tgsi_action *action,
 295                       struct lp_build_tgsi_context *bld_base,
 296                       struct lp_build_emit_data *emit_data)
 297 {
 298         LLVMBuilderRef builder = bld_base->base.gallivm->builder;
 299         emit_data->output[emit_data->chan] = LLVMBuildURem(builder,
 300                         emit_data->args[0], emit_data->args[1], "");
 301 }
 302
 303 static void emit_shl(const struct lp_build_tgsi_action *action,
 304                      struct lp_build_tgsi_context *bld_base,
 305                      struct lp_build_emit_data *emit_data)
 306 {
 307         LLVMBuilderRef builder = bld_base->base.gallivm->builder;
 308         emit_data->output[emit_data->chan] = LLVMBuildShl(builder,
 309                         emit_data->args[0], emit_data->args[1], "");
 310 }
 311
 312 static void emit_ushr(const struct lp_build_tgsi_action *action,
 313                       struct lp_build_tgsi_context *bld_base,
 314                       struct lp_build_emit_data *emit_data)
 315 {
 316         LLVMBuilderRef builder = bld_base->base.gallivm->builder;
 317         emit_data->output[emit_data->chan] = LLVMBuildLShr(builder,
 318                         emit_data->args[0], emit_data->args[1], "");
 319 }
 320 static void emit_ishr(const struct lp_build_tgsi_action *action,
 321                       struct lp_build_tgsi_context *bld_base,
 322                       struct lp_build_emit_data *emit_data)
 323 {
 324         LLVMBuilderRef builder = bld_base->base.gallivm->builder;
 325         emit_data->output[emit_data->chan] = LLVMBuildAShr(builder,
 326                         emit_data->args[0], emit_data->args[1], "");
 327 }
 328
 329 static void emit_xor(const struct lp_build_tgsi_action *action,
 330                      struct lp_build_tgsi_context *bld_base,
 331                      struct lp_build_emit_data *emit_data)
 332 {
 333         LLVMBuilderRef builder = bld_base->base.gallivm->builder;
 334         emit_data->output[emit_data->chan] = LLVMBuildXor(builder,
 335                         emit_data->args[0], emit_data->args[1], "");
 336 }
 337
 338 static void emit_ssg(const struct lp_build_tgsi_action *action,
 339                      struct lp_build_tgsi_context *bld_base,
 340                      struct lp_build_emit_data *emit_data)
 341 {
 342         LLVMBuilderRef builder = bld_base->base.gallivm->builder;
 343
 344         LLVMValueRef cmp, val;
 345
 346         if (emit_data->inst->Instruction.Opcode == TGSI_OPCODE_I64SSG) {
 347                 cmp = LLVMBuildICmp(builder, LLVMIntSGT, emit_data->args[0], bld_base->int64_bld.zero, "");
 348                 val = LLVMBuildSelect(builder, cmp, bld_base->int64_bld.one, emit_data->args[0], "");
 349                 cmp = LLVMBuildICmp(builder, LLVMIntSGE, val, bld_base->int64_bld.zero, "");
 350                 val = LLVMBuildSelect(builder, cmp, val, LLVMConstInt(bld_base->int64_bld.elem_type, -1, true), "");
 351         } else if (emit_data->inst->Instruction.Opcode == TGSI_OPCODE_ISSG) {
 352                 cmp = LLVMBuildICmp(builder, LLVMIntSGT, emit_data->args[0], bld_base->int_bld.zero, "");
 353                 val = LLVMBuildSelect(builder, cmp, bld_base->int_bld.one, emit_data->args[0], "");
 354                 cmp = LLVMBuildICmp(builder, LLVMIntSGE, val, bld_base->int_bld.zero, "");
 355                 val = LLVMBuildSelect(builder, cmp, val, LLVMConstInt(bld_base->int_bld.elem_type, -1, true), "");
 356         } else { // float SSG
 357                 cmp = LLVMBuildFCmp(builder, LLVMRealOGT, emit_data->args[0], bld_base->base.zero, "");
 358                 val = LLVMBuildSelect(builder, cmp, bld_base->base.one, emit_data->args[0], "");
 359                 cmp = LLVMBuildFCmp(builder, LLVMRealOGE, val, bld_base->base.zero, "");
 360                 val = LLVMBuildSelect(builder, cmp, val, LLVMConstReal(bld_base->base.elem_type, -1), "");
 361         }
 362
 363         emit_data->output[emit_data->chan] = val;
 364 }
 365
 366 static void emit_ineg(const struct lp_build_tgsi_action *action,
 367                       struct lp_build_tgsi_context *bld_base,
 368                       struct lp_build_emit_data *emit_data)
 369 {
 370         LLVMBuilderRef builder = bld_base->base.gallivm->builder;
 371         emit_data->output[emit_data->chan] = LLVMBuildNeg(builder,
 372                         emit_data->args[0], "");
 373 }
 374
 375 static void emit_dneg(const struct lp_build_tgsi_action *action,
 376                       struct lp_build_tgsi_context *bld_base,
 377                       struct lp_build_emit_data *emit_data)
 378 {
 379         LLVMBuilderRef builder = bld_base->base.gallivm->builder;
 380         emit_data->output[emit_data->chan] = LLVMBuildFNeg(builder,
 381                         emit_data->args[0], "");
 382 }
 383
 384 static void emit_frac(const struct lp_build_tgsi_action *action,
 385                       struct lp_build_tgsi_context *bld_base,
 386                       struct lp_build_emit_data *emit_data)
 387 {
 388         LLVMBuilderRef builder = bld_base->base.gallivm->builder;
 389         char *intr;
 390
 391         if (emit_data->info->opcode == TGSI_OPCODE_FRC)
 392                 intr = "llvm.floor.f32";
 393         else if (emit_data->info->opcode == TGSI_OPCODE_DFRAC)
 394                 intr = "llvm.floor.f64";
 395         else {
 396                 assert(0);
 397                 return;
 398         }
 399
 400         LLVMValueRef floor = lp_build_intrinsic(builder, intr, emit_data->dst_type,
 401                                                 &emit_data->args[0], 1,
 402                                                 LLVMReadNoneAttribute);
 403         emit_data->output[emit_data->chan] = LLVMBuildFSub(builder,
 404                         emit_data->args[0], floor, "");
 405 }
 406
 407 static void emit_f2i(const struct lp_build_tgsi_action *action,
 408                      struct lp_build_tgsi_context *bld_base,
 409                      struct lp_build_emit_data *emit_data)
 410 {
 411         LLVMBuilderRef builder = bld_base->base.gallivm->builder;
 412         emit_data->output[emit_data->chan] = LLVMBuildFPToSI(builder,
 413                         emit_data->args[0], bld_base->int_bld.elem_type, "");
 414 }
 415
 416 static void emit_f2u(const struct lp_build_tgsi_action *action,
 417                      struct lp_build_tgsi_context *bld_base,
 418                      struct lp_build_emit_data *emit_data)
 419 {
 420         LLVMBuilderRef builder = bld_base->base.gallivm->builder;
 421         emit_data->output[emit_data->chan] = LLVMBuildFPToUI(builder,
 422                         emit_data->args[0], bld_base->uint_bld.elem_type, "");
 423 }
 424
 425 static void emit_i2f(const struct lp_build_tgsi_action *action,
 426                      struct lp_build_tgsi_context *bld_base,
 427                      struct lp_build_emit_data *emit_data)
 428 {
 429         LLVMBuilderRef builder = bld_base->base.gallivm->builder;
 430         emit_data->output[emit_data->chan] = LLVMBuildSIToFP(builder,
 431                         emit_data->args[0], bld_base->base.elem_type, "");
 432 }
 433
 434 static void emit_u2f(const struct lp_build_tgsi_action *action,
 435                      struct lp_build_tgsi_context *bld_base,
 436                      struct lp_build_emit_data *emit_data)
 437 {
 438         LLVMBuilderRef builder = bld_base->base.gallivm->builder;
 439         emit_data->output[emit_data->chan] = LLVMBuildUIToFP(builder,
 440                         emit_data->args[0], bld_base->base.elem_type, "");
 441 }
 442
 443 static void
 444 build_tgsi_intrinsic_nomem(const struct lp_build_tgsi_action *action,
 445                            struct lp_build_tgsi_context *bld_base,
 446                            struct lp_build_emit_data *emit_data)
 447 {
 448         struct lp_build_context *base = &bld_base->base;
 449         emit_data->output[emit_data->chan] =
 450                 lp_build_intrinsic(base->gallivm->builder, action->intr_name,
 451                                    emit_data->dst_type, emit_data->args,
 452                                    emit_data->arg_count, LLVMReadNoneAttribute);
 453 }
 454
 455 static void emit_bfi(const struct lp_build_tgsi_action *action,
 456                      struct lp_build_tgsi_context *bld_base,
 457                      struct lp_build_emit_data *emit_data)
 458 {
 459         struct gallivm_state *gallivm = bld_base->base.gallivm;
 460         LLVMBuilderRef builder = gallivm->builder;
 461         LLVMValueRef bfi_args[3];
 462
 463         // Calculate the bitmask: (((1 << src3) - 1) << src2
 464         bfi_args[0] = LLVMBuildShl(builder,
 465                                    LLVMBuildSub(builder,
 466                                                 LLVMBuildShl(builder,
 467                                                              bld_base->int_bld.one,
 468                                                              emit_data->args[3], ""),
 469                                                 bld_base->int_bld.one, ""),
 470                                    emit_data->args[2], "");
 471
 472         bfi_args[1] = LLVMBuildShl(builder, emit_data->args[1],
 473                                    emit_data->args[2], "");
 474
 475         bfi_args[2] = emit_data->args[0];
 476
 477         /* Calculate:
 478          *   (arg0 & arg1) | (~arg0 & arg2) = arg2 ^ (arg0 & (arg1 ^ arg2)
 479          * Use the right-hand side, which the LLVM backend can convert to V_BFI.
 480          */
 481         emit_data->output[emit_data->chan] =
 482                 LLVMBuildXor(builder, bfi_args[2],
 483                         LLVMBuildAnd(builder, bfi_args[0],
 484                                 LLVMBuildXor(builder, bfi_args[1], bfi_args[2],
 485                                              ""), ""), "");
 486 }
 487
 488 /* this is ffs in C */
 489 static void emit_lsb(const struct lp_build_tgsi_action *action,
 490                      struct lp_build_tgsi_context *bld_base,
 491                      struct lp_build_emit_data *emit_data)
 492 {
 493         struct gallivm_state *gallivm = bld_base->base.gallivm;
 494         LLVMValueRef args[2] = {
 495                 emit_data->args[0],
 496
 497                 /* The value of 1 means that ffs(x=0) = undef, so LLVM won't
 498                  * add special code to check for x=0. The reason is that
 499                  * the LLVM behavior for x=0 is different from what we
 500                  * need here.
 501                  *
 502                  * The hardware already implements the correct behavior.
 503                  */
 504                 LLVMConstInt(LLVMInt1TypeInContext(gallivm->context), 1, 0)
 505         };
 506
 507         emit_data->output[emit_data->chan] =
 508                 lp_build_intrinsic(gallivm->builder, "llvm.cttz.i32",
 509                                 emit_data->dst_type, args, ARRAY_SIZE(args),
 510                                 LLVMReadNoneAttribute);
 511 }
 512
 513 /* Find the last bit set. */
 514 static void emit_umsb(const struct lp_build_tgsi_action *action,
 515                       struct lp_build_tgsi_context *bld_base,
 516                       struct lp_build_emit_data *emit_data)
 517 {
 518         struct gallivm_state *gallivm = bld_base->base.gallivm;
 519         LLVMBuilderRef builder = gallivm->builder;
 520         LLVMValueRef args[2] = {
 521                 emit_data->args[0],
 522                 /* Don't generate code for handling zero: */
 523                 LLVMConstInt(LLVMInt1TypeInContext(gallivm->context), 1, 0)
 524         };
 525
 526         LLVMValueRef msb =
 527                 lp_build_intrinsic(builder, "llvm.ctlz.i32",
 528                                 emit_data->dst_type, args, ARRAY_SIZE(args),
 529                                 LLVMReadNoneAttribute);
 530
 531         /* The HW returns the last bit index from MSB, but TGSI wants
 532          * the index from LSB. Invert it by doing "31 - msb". */
 533         msb = LLVMBuildSub(builder, lp_build_const_int32(gallivm, 31),
 534                            msb, "");
 535
 536         /* Check for zero: */
 537         emit_data->output[emit_data->chan] =
 538                 LLVMBuildSelect(builder,
 539                                 LLVMBuildICmp(builder, LLVMIntEQ, args[0],
 540                                               bld_base->uint_bld.zero, ""),
 541                                 lp_build_const_int32(gallivm, -1), msb, "");
 542 }
 543
 544 /* Find the last bit opposite of the sign bit. */
 545 static void emit_imsb(const struct lp_build_tgsi_action *action,
 546                       struct lp_build_tgsi_context *bld_base,
 547                       struct lp_build_emit_data *emit_data)
 548 {
 549         struct gallivm_state *gallivm = bld_base->base.gallivm;
 550         LLVMBuilderRef builder = gallivm->builder;
 551         LLVMValueRef arg = emit_data->args[0];
 552
 553         LLVMValueRef msb =
 554                 lp_build_intrinsic(builder, "llvm.AMDGPU.flbit.i32",
 555                                 emit_data->dst_type, &arg, 1,
 556                                 LLVMReadNoneAttribute);
 557
 558         /* The HW returns the last bit index from MSB, but TGSI wants
 559          * the index from LSB. Invert it by doing "31 - msb". */
 560         msb = LLVMBuildSub(builder, lp_build_const_int32(gallivm, 31),
 561                            msb, "");
 562
 563         /* If arg == 0 || arg == -1 (0xffffffff), return -1. */
 564         LLVMValueRef all_ones = lp_build_const_int32(gallivm, -1);
 565
 566         LLVMValueRef cond =
 567                 LLVMBuildOr(builder,
 568                             LLVMBuildICmp(builder, LLVMIntEQ, arg,
 569                                           bld_base->uint_bld.zero, ""),
 570                             LLVMBuildICmp(builder, LLVMIntEQ, arg,
 571                                           all_ones, ""), "");
 572
 573         emit_data->output[emit_data->chan] =
 574                 LLVMBuildSelect(builder, cond, all_ones, msb, "");
 575 }
 576
 577 static void emit_iabs(const struct lp_build_tgsi_action *action,
 578                       struct lp_build_tgsi_context *bld_base,
 579                       struct lp_build_emit_data *emit_data)
 580 {
 581         LLVMBuilderRef builder = bld_base->base.gallivm->builder;
 582
 583         emit_data->output[emit_data->chan] =
 584                 lp_build_emit_llvm_binary(bld_base, TGSI_OPCODE_IMAX,
 585                                           emit_data->args[0],
 586                                           LLVMBuildNeg(builder,
 587                                                        emit_data->args[0], ""));
 588 }
 589
 590 static void emit_minmax_int(const struct lp_build_tgsi_action *action,
 591                             struct lp_build_tgsi_context *bld_base,
 592                             struct lp_build_emit_data *emit_data)
 593 {
 594         LLVMBuilderRef builder = bld_base->base.gallivm->builder;
 595         LLVMIntPredicate op;
 596
 597         switch (emit_data->info->opcode) {
 598         default:
 599                 assert(0);
 600         case TGSI_OPCODE_IMAX:
 601         case TGSI_OPCODE_I64MAX:
 602                 op = LLVMIntSGT;
 603                 break;
 604         case TGSI_OPCODE_IMIN:
 605         case TGSI_OPCODE_I64MIN:
 606                 op = LLVMIntSLT;
 607                 break;
 608         case TGSI_OPCODE_UMAX:
 609         case TGSI_OPCODE_U64MAX:
 610                 op = LLVMIntUGT;
 611                 break;
 612         case TGSI_OPCODE_UMIN:
 613         case TGSI_OPCODE_U64MIN:
 614                 op = LLVMIntULT;
 615                 break;
 616         }
 617
 618         emit_data->output[emit_data->chan] =
 619                 LLVMBuildSelect(builder,
 620                                 LLVMBuildICmp(builder, op, emit_data->args[0],
 621                                               emit_data->args[1], ""),
 622                                 emit_data->args[0],
 623                                 emit_data->args[1], "");
 624 }
 625
 626 static void pk2h_fetch_args(struct lp_build_tgsi_context *bld_base,
 627                             struct lp_build_emit_data *emit_data)
 628 {
 629         emit_data->args[0] = lp_build_emit_fetch(bld_base, emit_data->inst,
 630                                                  0, TGSI_CHAN_X);
 631         emit_data->args[1] = lp_build_emit_fetch(bld_base, emit_data->inst,
 632                                                  0, TGSI_CHAN_Y);
 633 }
 634
 635 static void emit_pk2h(const struct lp_build_tgsi_action *action,
 636                       struct lp_build_tgsi_context *bld_base,
 637                       struct lp_build_emit_data *emit_data)
 638 {
 639         LLVMBuilderRef builder = bld_base->base.gallivm->builder;
 640         LLVMContextRef context = bld_base->base.gallivm->context;
 641         struct lp_build_context *uint_bld = &bld_base->uint_bld;
 642         LLVMTypeRef fp16, i16;
 643         LLVMValueRef const16, comp[2];
 644         unsigned i;
 645
 646         fp16 = LLVMHalfTypeInContext(context);
 647         i16 = LLVMInt16TypeInContext(context);
 648         const16 = lp_build_const_int32(uint_bld->gallivm, 16);
 649
 650         for (i = 0; i < 2; i++) {
 651                 comp[i] = LLVMBuildFPTrunc(builder, emit_data->args[i], fp16, "");
 652                 comp[i] = LLVMBuildBitCast(builder, comp[i], i16, "");
 653                 comp[i] = LLVMBuildZExt(builder, comp[i], uint_bld->elem_type, "");
 654         }
 655
 656         comp[1] = LLVMBuildShl(builder, comp[1], const16, "");
 657         comp[0] = LLVMBuildOr(builder, comp[0], comp[1], "");
 658
 659         emit_data->output[emit_data->chan] = comp[0];
 660 }
 661
 662 static void up2h_fetch_args(struct lp_build_tgsi_context *bld_base,
 663                             struct lp_build_emit_data *emit_data)
 664 {
 665         emit_data->args[0] = lp_build_emit_fetch(bld_base, emit_data->inst,
 666                                                  0, TGSI_CHAN_X);
 667 }
 668
 669 static void emit_up2h(const struct lp_build_tgsi_action *action,
 670                       struct lp_build_tgsi_context *bld_base,
 671                       struct lp_build_emit_data *emit_data)
 672 {
 673         LLVMBuilderRef builder = bld_base->base.gallivm->builder;
 674         LLVMContextRef context = bld_base->base.gallivm->context;
 675         struct lp_build_context *uint_bld = &bld_base->uint_bld;
 676         LLVMTypeRef fp16, i16;
 677         LLVMValueRef const16, input, val;
 678         unsigned i;
 679
 680         fp16 = LLVMHalfTypeInContext(context);
 681         i16 = LLVMInt16TypeInContext(context);
 682         const16 = lp_build_const_int32(uint_bld->gallivm, 16);
 683         input = emit_data->args[0];
 684
 685         for (i = 0; i < 2; i++) {
 686                 val = i == 1 ? LLVMBuildLShr(builder, input, const16, "") : input;
 687                 val = LLVMBuildTrunc(builder, val, i16, "");
 688                 val = LLVMBuildBitCast(builder, val, fp16, "");
 689                 emit_data->output[i] =
 690                         LLVMBuildFPExt(builder, val, bld_base->base.elem_type, "");
 691         }
 692 }
 693
 694 static void emit_fdiv(const struct lp_build_tgsi_action *action,
 695                       struct lp_build_tgsi_context *bld_base,
 696                       struct lp_build_emit_data *emit_data)
 697 {
 698         struct si_shader_context *ctx = si_shader_context(bld_base);
 699
 700         emit_data->output[emit_data->chan] =
 701                 LLVMBuildFDiv(bld_base->base.gallivm->builder,
 702                               emit_data->args[0], emit_data->args[1], "");
 703
 704         /* Use v_rcp_f32 instead of precise division. */
 705         if (HAVE_LLVM >= 0x0309 &&
 706             !LLVMIsConstant(emit_data->output[emit_data->chan]))
 707                 LLVMSetMetadata(emit_data->output[emit_data->chan],
 708                                 ctx->fpmath_md_kind, ctx->fpmath_md_2p5_ulp);
 709 }
 710
 711 /* 1/sqrt is translated to rsq for f32 if fp32 denormals are not enabled in
 712  * the target machine. f64 needs global unsafe math flags to get rsq. */
 713 static void emit_rsq(const struct lp_build_tgsi_action *action,
 714                      struct lp_build_tgsi_context *bld_base,
 715                      struct lp_build_emit_data *emit_data)
 716 {
 717         LLVMValueRef sqrt =
 718                 lp_build_emit_llvm_unary(bld_base, TGSI_OPCODE_SQRT,
 719                                          emit_data->args[0]);
 720
 721         emit_data->output[emit_data->chan] =
 722                 lp_build_emit_llvm_binary(bld_base, TGSI_OPCODE_DIV,
 723                                           bld_base->base.one, sqrt);
 724 }
 725
 726 void si_shader_context_init_alu(struct lp_build_tgsi_context *bld_base)
 727 {
 728         lp_set_default_actions(bld_base);
 729
 730         bld_base->op_actions[TGSI_OPCODE_ABS].emit = build_tgsi_intrinsic_nomem;
 731         bld_base->op_actions[TGSI_OPCODE_ABS].intr_name = "llvm.fabs.f32";
 732         bld_base->op_actions[TGSI_OPCODE_AND].emit = emit_and;
 733         bld_base->op_actions[TGSI_OPCODE_ARL].emit = emit_arl;
 734         bld_base->op_actions[TGSI_OPCODE_BFI].emit = emit_bfi;
 735         bld_base->op_actions[TGSI_OPCODE_BREV].emit = build_tgsi_intrinsic_nomem;
 736         bld_base->op_actions[TGSI_OPCODE_BREV].intr_name =
 737                 HAVE_LLVM >= 0x0308 ? "llvm.bitreverse.i32" : "llvm.AMDGPU.brev";
 738         bld_base->op_actions[TGSI_OPCODE_CEIL].emit = build_tgsi_intrinsic_nomem;
 739         bld_base->op_actions[TGSI_OPCODE_CEIL].intr_name = "llvm.ceil.f32";
 740         bld_base->op_actions[TGSI_OPCODE_CLAMP].emit = build_tgsi_intrinsic_nomem;
 741         bld_base->op_actions[TGSI_OPCODE_CLAMP].intr_name =
 742                 HAVE_LLVM >= 0x0308 ? "llvm.AMDGPU.clamp." : "llvm.AMDIL.clamp.";
 743         bld_base->op_actions[TGSI_OPCODE_CMP].emit = emit_cmp;
 744         bld_base->op_actions[TGSI_OPCODE_COS].emit = build_tgsi_intrinsic_nomem;
 745         bld_base->op_actions[TGSI_OPCODE_COS].intr_name = "llvm.cos.f32";
 746         bld_base->op_actions[TGSI_OPCODE_DABS].emit = build_tgsi_intrinsic_nomem;
 747         bld_base->op_actions[TGSI_OPCODE_DABS].intr_name = "llvm.fabs.f64";
 748         bld_base->op_actions[TGSI_OPCODE_DFMA].emit = build_tgsi_intrinsic_nomem;
 749         bld_base->op_actions[TGSI_OPCODE_DFMA].intr_name = "llvm.fma.f64";
 750         bld_base->op_actions[TGSI_OPCODE_DFRAC].emit = emit_frac;
 751         bld_base->op_actions[TGSI_OPCODE_DIV].emit = emit_fdiv;
 752         bld_base->op_actions[TGSI_OPCODE_DNEG].emit = emit_dneg;
 753         bld_base->op_actions[TGSI_OPCODE_DSEQ].emit = emit_dcmp;
 754         bld_base->op_actions[TGSI_OPCODE_DSGE].emit = emit_dcmp;
 755         bld_base->op_actions[TGSI_OPCODE_DSLT].emit = emit_dcmp;
 756         bld_base->op_actions[TGSI_OPCODE_DSNE].emit = emit_dcmp;
 757         bld_base->op_actions[TGSI_OPCODE_DRSQ].emit = build_tgsi_intrinsic_nomem;
 758         bld_base->op_actions[TGSI_OPCODE_DRSQ].intr_name =
 759                 HAVE_LLVM >= 0x0309 ? "llvm.amdgcn.rsq.f64" : "llvm.AMDGPU.rsq.f64";
 760         bld_base->op_actions[TGSI_OPCODE_DSQRT].emit = build_tgsi_intrinsic_nomem;
 761         bld_base->op_actions[TGSI_OPCODE_DSQRT].intr_name = "llvm.sqrt.f64";
 762         bld_base->op_actions[TGSI_OPCODE_EX2].emit = build_tgsi_intrinsic_nomem;
 763         bld_base->op_actions[TGSI_OPCODE_EX2].intr_name =
 764                 HAVE_LLVM >= 0x0308 ? "llvm.exp2.f32" : "llvm.AMDIL.exp.";
 765         bld_base->op_actions[TGSI_OPCODE_FLR].emit = build_tgsi_intrinsic_nomem;
 766         bld_base->op_actions[TGSI_OPCODE_FLR].intr_name = "llvm.floor.f32";
 767         bld_base->op_actions[TGSI_OPCODE_FMA].emit =
 768                 bld_base->op_actions[TGSI_OPCODE_MAD].emit;
 769         bld_base->op_actions[TGSI_OPCODE_FRC].emit = emit_frac;
 770         bld_base->op_actions[TGSI_OPCODE_F2I].emit = emit_f2i;
 771         bld_base->op_actions[TGSI_OPCODE_F2U].emit = emit_f2u;
 772         bld_base->op_actions[TGSI_OPCODE_FSEQ].emit = emit_fcmp;
 773         bld_base->op_actions[TGSI_OPCODE_FSGE].emit = emit_fcmp;
 774         bld_base->op_actions[TGSI_OPCODE_FSLT].emit = emit_fcmp;
 775         bld_base->op_actions[TGSI_OPCODE_FSNE].emit = emit_fcmp;
 776         bld_base->op_actions[TGSI_OPCODE_IABS].emit = emit_iabs;
 777         bld_base->op_actions[TGSI_OPCODE_IBFE].emit = build_tgsi_intrinsic_nomem;
 778         bld_base->op_actions[TGSI_OPCODE_IBFE].intr_name = "llvm.AMDGPU.bfe.i32";
 779         bld_base->op_actions[TGSI_OPCODE_IDIV].emit = emit_idiv;
 780         bld_base->op_actions[TGSI_OPCODE_IMAX].emit = emit_minmax_int;
 781         bld_base->op_actions[TGSI_OPCODE_IMIN].emit = emit_minmax_int;
 782         bld_base->op_actions[TGSI_OPCODE_IMSB].emit = emit_imsb;
 783         bld_base->op_actions[TGSI_OPCODE_INEG].emit = emit_ineg;
 784         bld_base->op_actions[TGSI_OPCODE_ISHR].emit = emit_ishr;
 785         bld_base->op_actions[TGSI_OPCODE_ISGE].emit = emit_icmp;
 786         bld_base->op_actions[TGSI_OPCODE_ISLT].emit = emit_icmp;
 787         bld_base->op_actions[TGSI_OPCODE_ISSG].emit = emit_ssg;
 788         bld_base->op_actions[TGSI_OPCODE_I2F].emit = emit_i2f;
 789         bld_base->op_actions[TGSI_OPCODE_KILL_IF].fetch_args = kill_if_fetch_args;
 790         bld_base->op_actions[TGSI_OPCODE_KILL_IF].emit = kil_emit;
 791         bld_base->op_actions[TGSI_OPCODE_KILL_IF].intr_name = "llvm.AMDGPU.kill";
 792         bld_base->op_actions[TGSI_OPCODE_KILL].emit = lp_build_tgsi_intrinsic;
 793         bld_base->op_actions[TGSI_OPCODE_KILL].intr_name = "llvm.AMDGPU.kilp";
 794         bld_base->op_actions[TGSI_OPCODE_LSB].emit = emit_lsb;
 795         bld_base->op_actions[TGSI_OPCODE_LG2].emit = build_tgsi_intrinsic_nomem;
 796         bld_base->op_actions[TGSI_OPCODE_LG2].intr_name = "llvm.log2.f32";
 797         bld_base->op_actions[TGSI_OPCODE_MAX].emit = build_tgsi_intrinsic_nomem;
 798         bld_base->op_actions[TGSI_OPCODE_MAX].intr_name = "llvm.maxnum.f32";
 799         bld_base->op_actions[TGSI_OPCODE_MIN].emit = build_tgsi_intrinsic_nomem;
 800         bld_base->op_actions[TGSI_OPCODE_MIN].intr_name = "llvm.minnum.f32";
 801         bld_base->op_actions[TGSI_OPCODE_MOD].emit = emit_mod;
 802         bld_base->op_actions[TGSI_OPCODE_UMSB].emit = emit_umsb;
 803         bld_base->op_actions[TGSI_OPCODE_NOT].emit = emit_not;
 804         bld_base->op_actions[TGSI_OPCODE_OR].emit = emit_or;
 805         bld_base->op_actions[TGSI_OPCODE_PK2H].fetch_args = pk2h_fetch_args;
 806         bld_base->op_actions[TGSI_OPCODE_PK2H].emit = emit_pk2h;
 807         bld_base->op_actions[TGSI_OPCODE_POPC].emit = build_tgsi_intrinsic_nomem;
 808         bld_base->op_actions[TGSI_OPCODE_POPC].intr_name = "llvm.ctpop.i32";
 809         bld_base->op_actions[TGSI_OPCODE_POW].emit = build_tgsi_intrinsic_nomem;
 810         bld_base->op_actions[TGSI_OPCODE_POW].intr_name = "llvm.pow.f32";
 811         bld_base->op_actions[TGSI_OPCODE_ROUND].emit = build_tgsi_intrinsic_nomem;
 812         bld_base->op_actions[TGSI_OPCODE_ROUND].intr_name = "llvm.rint.f32";
 813         bld_base->op_actions[TGSI_OPCODE_RSQ].emit = emit_rsq;
 814         bld_base->op_actions[TGSI_OPCODE_SGE].emit = emit_set_cond;
 815         bld_base->op_actions[TGSI_OPCODE_SEQ].emit = emit_set_cond;
 816         bld_base->op_actions[TGSI_OPCODE_SHL].emit = emit_shl;
 817         bld_base->op_actions[TGSI_OPCODE_SLE].emit = emit_set_cond;
 818         bld_base->op_actions[TGSI_OPCODE_SLT].emit = emit_set_cond;
 819         bld_base->op_actions[TGSI_OPCODE_SNE].emit = emit_set_cond;
 820         bld_base->op_actions[TGSI_OPCODE_SGT].emit = emit_set_cond;
 821         bld_base->op_actions[TGSI_OPCODE_SIN].emit = build_tgsi_intrinsic_nomem;
 822         bld_base->op_actions[TGSI_OPCODE_SIN].intr_name = "llvm.sin.f32";
 823         bld_base->op_actions[TGSI_OPCODE_SQRT].emit = build_tgsi_intrinsic_nomem;
 824         bld_base->op_actions[TGSI_OPCODE_SQRT].intr_name = "llvm.sqrt.f32";
 825         bld_base->op_actions[TGSI_OPCODE_SSG].emit = emit_ssg;
 826         bld_base->op_actions[TGSI_OPCODE_TRUNC].emit = build_tgsi_intrinsic_nomem;
 827         bld_base->op_actions[TGSI_OPCODE_TRUNC].intr_name = "llvm.trunc.f32";
 828         bld_base->op_actions[TGSI_OPCODE_UADD].emit = emit_uadd;
 829         bld_base->op_actions[TGSI_OPCODE_UBFE].emit = build_tgsi_intrinsic_nomem;
 830         bld_base->op_actions[TGSI_OPCODE_UBFE].intr_name = "llvm.AMDGPU.bfe.u32";
 831         bld_base->op_actions[TGSI_OPCODE_UDIV].emit = emit_udiv;
 832         bld_base->op_actions[TGSI_OPCODE_UMAX].emit = emit_minmax_int;
 833         bld_base->op_actions[TGSI_OPCODE_UMIN].emit = emit_minmax_int;
 834         bld_base->op_actions[TGSI_OPCODE_UMOD].emit = emit_umod;
 835         bld_base->op_actions[TGSI_OPCODE_USEQ].emit = emit_icmp;
 836         bld_base->op_actions[TGSI_OPCODE_USGE].emit = emit_icmp;
 837         bld_base->op_actions[TGSI_OPCODE_USHR].emit = emit_ushr;
 838         bld_base->op_actions[TGSI_OPCODE_USLT].emit = emit_icmp;
 839         bld_base->op_actions[TGSI_OPCODE_USNE].emit = emit_icmp;
 840         bld_base->op_actions[TGSI_OPCODE_U2F].emit = emit_u2f;
 841         bld_base->op_actions[TGSI_OPCODE_XOR].emit = emit_xor;
 842         bld_base->op_actions[TGSI_OPCODE_UCMP].emit = emit_ucmp;
 843         bld_base->op_actions[TGSI_OPCODE_UP2H].fetch_args = up2h_fetch_args;
 844         bld_base->op_actions[TGSI_OPCODE_UP2H].emit = emit_up2h;
 845
 846         bld_base->op_actions[TGSI_OPCODE_I64MAX].emit = emit_minmax_int;
 847         bld_base->op_actions[TGSI_OPCODE_I64MIN].emit = emit_minmax_int;
 848         bld_base->op_actions[TGSI_OPCODE_U64MAX].emit = emit_minmax_int;
 849         bld_base->op_actions[TGSI_OPCODE_U64MIN].emit = emit_minmax_int;
 850         bld_base->op_actions[TGSI_OPCODE_I64ABS].emit = emit_iabs;
 851         bld_base->op_actions[TGSI_OPCODE_I64SSG].emit = emit_ssg;
 852         bld_base->op_actions[TGSI_OPCODE_I64NEG].emit = emit_ineg;
 853
 854         bld_base->op_actions[TGSI_OPCODE_U64SEQ].emit = emit_icmp;
 855         bld_base->op_actions[TGSI_OPCODE_U64SNE].emit = emit_icmp;
 856         bld_base->op_actions[TGSI_OPCODE_U64SGE].emit = emit_icmp;
 857         bld_base->op_actions[TGSI_OPCODE_U64SLT].emit = emit_icmp;
 858         bld_base->op_actions[TGSI_OPCODE_I64SGE].emit = emit_icmp;
 859         bld_base->op_actions[TGSI_OPCODE_I64SLT].emit = emit_icmp;
 860
 861         bld_base->op_actions[TGSI_OPCODE_U64ADD].emit = emit_uadd;
 862         bld_base->op_actions[TGSI_OPCODE_U64SHL].emit = emit_shl;
 863         bld_base->op_actions[TGSI_OPCODE_U64SHR].emit = emit_ushr;
 864         bld_base->op_actions[TGSI_OPCODE_I64SHR].emit = emit_ishr;
 865
 866         bld_base->op_actions[TGSI_OPCODE_U64MOD].emit = emit_umod;
 867         bld_base->op_actions[TGSI_OPCODE_I64MOD].emit = emit_mod;
 868         bld_base->op_actions[TGSI_OPCODE_U64DIV].emit = emit_udiv;
 869         bld_base->op_actions[TGSI_OPCODE_I64DIV].emit = emit_idiv;
 870 }
 871
 872 static LLVMValueRef build_cube_intrinsic(struct gallivm_state *gallivm,
 873                                          LLVMValueRef in[3])
 874 {
 875         if (HAVE_LLVM >= 0x0309) {
 876                 LLVMTypeRef f32 = LLVMTypeOf(in[0]);
 877                 LLVMValueRef out[4];
 878
 879                 out[0] = lp_build_intrinsic(gallivm->builder, "llvm.amdgcn.cubetc",
 880                                             f32, in, 3, LLVMReadNoneAttribute);
 881                 out[1] = lp_build_intrinsic(gallivm->builder, "llvm.amdgcn.cubesc",
 882                                             f32, in, 3, LLVMReadNoneAttribute);
 883                 out[2] = lp_build_intrinsic(gallivm->builder, "llvm.amdgcn.cubema",
 884                                             f32, in, 3, LLVMReadNoneAttribute);
 885                 out[3] = lp_build_intrinsic(gallivm->builder, "llvm.amdgcn.cubeid",
 886                                             f32, in, 3, LLVMReadNoneAttribute);
 887
 888                 return lp_build_gather_values(gallivm, out, 4);
 889         } else {
 890                 LLVMValueRef c[4] = {
 891                         in[0],
 892                         in[1],
 893                         in[2],
 894                         LLVMGetUndef(LLVMTypeOf(in[0]))
 895                 };
 896                 LLVMValueRef vec = lp_build_gather_values(gallivm, c, 4);
 897
 898                 return lp_build_intrinsic(gallivm->builder, "llvm.AMDGPU.cube",
 899                                           LLVMTypeOf(vec), &vec, 1,
 900                                           LLVMReadNoneAttribute);
 901         }
 902 }
 903
 904 static void si_llvm_cube_to_2d_coords(struct lp_build_tgsi_context *bld_base,
 905                                       LLVMValueRef *in, LLVMValueRef *out)
 906 {
 907         struct gallivm_state *gallivm = bld_base->base.gallivm;
 908         LLVMBuilderRef builder = gallivm->builder;
 909         LLVMTypeRef type = bld_base->base.elem_type;
 910         LLVMValueRef coords[4];
 911         LLVMValueRef mad_args[3];
 912         LLVMValueRef v;
 913         unsigned i;
 914
 915         v = build_cube_intrinsic(gallivm, in);
 916
 917         for (i = 0; i < 4; ++i)
 918                 coords[i] = LLVMBuildExtractElement(builder, v,
 919                                                     lp_build_const_int32(gallivm, i), "");
 920
 921         coords[2] = lp_build_intrinsic(builder, "llvm.fabs.f32",
 922                         type, &coords[2], 1, LLVMReadNoneAttribute);
 923         coords[2] = lp_build_emit_llvm_unary(bld_base, TGSI_OPCODE_RCP, coords[2]);
 924
 925         mad_args[1] = coords[2];
 926         mad_args[2] = LLVMConstReal(type, 1.5);
 927
 928         mad_args[0] = coords[0];
 929         coords[0] = lp_build_emit_llvm_ternary(bld_base, TGSI_OPCODE_MAD,
 930                         mad_args[0], mad_args[1], mad_args[2]);
 931
 932         mad_args[0] = coords[1];
 933         coords[1] = lp_build_emit_llvm_ternary(bld_base, TGSI_OPCODE_MAD,
 934                         mad_args[0], mad_args[1], mad_args[2]);
 935
 936         /* apply xyz = yxw swizzle to cooords */
 937         out[0] = coords[1];
 938         out[1] = coords[0];
 939         out[2] = coords[3];
 940 }
 941
 942 void si_prepare_cube_coords(struct lp_build_tgsi_context *bld_base,
 943                             struct lp_build_emit_data *emit_data,
 944                             LLVMValueRef *coords_arg,
 945                             LLVMValueRef *derivs_arg)
 946 {
 947
 948         unsigned target = emit_data->inst->Texture.Texture;
 949         unsigned opcode = emit_data->inst->Instruction.Opcode;
 950         struct gallivm_state *gallivm = bld_base->base.gallivm;
 951         LLVMBuilderRef builder = gallivm->builder;
 952         LLVMValueRef coords[4];
 953         unsigned i;
 954
 955         si_llvm_cube_to_2d_coords(bld_base, coords_arg, coords);
 956
 957         if (opcode == TGSI_OPCODE_TXD && derivs_arg) {
 958                 LLVMValueRef derivs[4];
 959                 int axis;
 960
 961                 /* Convert cube derivatives to 2D derivatives. */
 962                 for (axis = 0; axis < 2; axis++) {
 963                         LLVMValueRef shifted_cube_coords[4], shifted_coords[4];
 964
 965                         /* Shift the cube coordinates by the derivatives to get
 966                          * the cube coordinates of the "neighboring pixel".
 967                          */
 968                         for (i = 0; i < 3; i++)
 969                                 shifted_cube_coords[i] =
 970                                         LLVMBuildFAdd(builder, coords_arg[i],
 971                                                       derivs_arg[axis*3+i], "");
 972                         shifted_cube_coords[3] = LLVMGetUndef(bld_base->base.elem_type);
 973
 974                         /* Project the shifted cube coordinates onto the face. */
 975                         si_llvm_cube_to_2d_coords(bld_base, shifted_cube_coords,
 976                                                       shifted_coords);
 977
 978                         /* Subtract both sets of 2D coordinates to get 2D derivatives.
 979                          * This won't work if the shifted coordinates ended up
 980                          * in a different face.
 981                          */
 982                         for (i = 0; i < 2; i++)
 983                                 derivs[axis * 2 + i] =
 984                                         LLVMBuildFSub(builder, shifted_coords[i],
 985                                                       coords[i], "");
 986                 }
 987
 988                 memcpy(derivs_arg, derivs, sizeof(derivs));
 989         }
 990
 991         if (target == TGSI_TEXTURE_CUBE_ARRAY ||
 992             target == TGSI_TEXTURE_SHADOWCUBE_ARRAY) {
 993                 /* for cube arrays coord.z = coord.w(array_index) * 8 + face */
 994                 /* coords_arg.w component - array_index for cube arrays */
 995                 coords[2] = lp_build_emit_llvm_ternary(bld_base, TGSI_OPCODE_MAD,
 996                                                        coords_arg[3], lp_build_const_float(gallivm, 8.0), coords[2]);
 997         }
 998
 999         /* Preserve compare/lod/bias. Put it in coords.w. */
1000         if (opcode == TGSI_OPCODE_TEX2 ||
1001             opcode == TGSI_OPCODE_TXB2 ||
1002             opcode == TGSI_OPCODE_TXL2) {
1003                 coords[3] = coords_arg[4];
1004         } else if (opcode == TGSI_OPCODE_TXB ||
1005                    opcode == TGSI_OPCODE_TXL ||
1006                    target == TGSI_TEXTURE_SHADOWCUBE) {
1007                 coords[3] = coords_arg[3];
1008         }
1009
1010         memcpy(coords_arg, coords, sizeof(coords));
1011 }