src/gallium/drivers/swr/swr_shader.cpp

   1 /****************************************************************************
   2  * Copyright (C) 2015 Intel Corporation.   All Rights Reserved.
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  ***************************************************************************/
  23
  24 // llvm redefines DEBUG
  25 #pragma push_macro("DEBUG")
  26 #undef DEBUG
  27 #include "JitManager.h"
  28 #include "llvm-c/Core.h"
  29 #include "llvm/Support/CBindingWrapping.h"
  30 #pragma pop_macro("DEBUG")
  31
  32 #include "state.h"
  33 #include "gen_state_llvm.h"
  34 #include "builder.h"
  35
  36 #include "tgsi/tgsi_strings.h"
  37 #include "util/u_format.h"
  38 #include "util/u_prim.h"
  39 #include "gallivm/lp_bld_init.h"
  40 #include "gallivm/lp_bld_flow.h"
  41 #include "gallivm/lp_bld_struct.h"
  42 #include "gallivm/lp_bld_tgsi.h"
  43
  44 #include "swr_context.h"
  45 #include "gen_swr_context_llvm.h"
  46 #include "swr_resource.h"
  47 #include "swr_state.h"
  48 #include "swr_screen.h"
  49
  50 using namespace SwrJit;
  51 using namespace llvm;
  52
  53 static unsigned
  54 locate_linkage(ubyte name, ubyte index, struct tgsi_shader_info *info);
  55
  56 bool operator==(const swr_jit_fs_key &lhs, const swr_jit_fs_key &rhs)
  57 {
  58    return !memcmp(&lhs, &rhs, sizeof(lhs));
  59 }
  60
  61 bool operator==(const swr_jit_vs_key &lhs, const swr_jit_vs_key &rhs)
  62 {
  63    return !memcmp(&lhs, &rhs, sizeof(lhs));
  64 }
  65
  66 bool operator==(const swr_jit_fetch_key &lhs, const swr_jit_fetch_key &rhs)
  67 {
  68    return !memcmp(&lhs, &rhs, sizeof(lhs));
  69 }
  70
  71 bool operator==(const swr_jit_gs_key &lhs, const swr_jit_gs_key &rhs)
  72 {
  73    return !memcmp(&lhs, &rhs, sizeof(lhs));
  74 }
  75
  76 static void
  77 swr_generate_sampler_key(const struct lp_tgsi_info &info,
  78                          struct swr_context *ctx,
  79                          enum pipe_shader_type shader_type,
  80                          struct swr_jit_sampler_key &key)
  81 {
  82    key.nr_samplers = info.base.file_max[TGSI_FILE_SAMPLER] + 1;
  83
  84    for (unsigned i = 0; i < key.nr_samplers; i++) {
  85       if (info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
  86          lp_sampler_static_sampler_state(
  87             &key.sampler[i].sampler_state,
  88             ctx->samplers[shader_type][i]);
  89       }
  90    }
  91
  92    /*
  93     * XXX If TGSI_FILE_SAMPLER_VIEW exists assume all texture opcodes
  94     * are dx10-style? Can't really have mixed opcodes, at least not
  95     * if we want to skip the holes here (without rescanning tgsi).
  96     */
  97    if (info.base.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
  98       key.nr_sampler_views =
  99          info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
 100       for (unsigned i = 0; i < key.nr_sampler_views; i++) {
 101          if (info.base.file_mask[TGSI_FILE_SAMPLER_VIEW] & (1 << i)) {
 102             const struct pipe_sampler_view *view =
 103                ctx->sampler_views[shader_type][i];
 104             lp_sampler_static_texture_state(
 105                &key.sampler[i].texture_state, view);
 106             if (view) {
 107                struct swr_resource *swr_res = swr_resource(view->texture);
 108                const struct util_format_description *desc =
 109                   util_format_description(view->format);
 110                if (swr_res->has_depth && swr_res->has_stencil &&
 111                    !util_format_has_depth(desc))
 112                   key.sampler[i].texture_state.format = PIPE_FORMAT_S8_UINT;
 113             }
 114          }
 115       }
 116    } else {
 117       key.nr_sampler_views = key.nr_samplers;
 118       for (unsigned i = 0; i < key.nr_sampler_views; i++) {
 119          if (info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
 120             const struct pipe_sampler_view *view =
 121                ctx->sampler_views[shader_type][i];
 122             lp_sampler_static_texture_state(
 123                &key.sampler[i].texture_state, view);
 124             if (view) {
 125                struct swr_resource *swr_res = swr_resource(view->texture);
 126                const struct util_format_description *desc =
 127                   util_format_description(view->format);
 128                if (swr_res->has_depth && swr_res->has_stencil &&
 129                    !util_format_has_depth(desc))
 130                   key.sampler[i].texture_state.format = PIPE_FORMAT_S8_UINT;
 131             }
 132          }
 133       }
 134    }
 135 }
 136
 137 void
 138 swr_generate_fs_key(struct swr_jit_fs_key &key,
 139                     struct swr_context *ctx,
 140                     swr_fragment_shader *swr_fs)
 141 {
 142    memset(&key, 0, sizeof(key));
 143
 144    key.nr_cbufs = ctx->framebuffer.nr_cbufs;
 145    key.light_twoside = ctx->rasterizer->light_twoside;
 146    key.sprite_coord_enable = ctx->rasterizer->sprite_coord_enable;
 147
 148    struct tgsi_shader_info *pPrevShader;
 149    if (ctx->gs)
 150       pPrevShader = &ctx->gs->info.base;
 151    else
 152       pPrevShader = &ctx->vs->info.base;
 153
 154    memcpy(&key.vs_output_semantic_name,
 155           &pPrevShader->output_semantic_name,
 156           sizeof(key.vs_output_semantic_name));
 157    memcpy(&key.vs_output_semantic_idx,
 158           &pPrevShader->output_semantic_index,
 159           sizeof(key.vs_output_semantic_idx));
 160
 161    swr_generate_sampler_key(swr_fs->info, ctx, PIPE_SHADER_FRAGMENT, key);
 162
 163    key.poly_stipple_enable = ctx->rasterizer->poly_stipple_enable &&
 164       ctx->poly_stipple.prim_is_poly;
 165 }
 166
 167 void
 168 swr_generate_vs_key(struct swr_jit_vs_key &key,
 169                     struct swr_context *ctx,
 170                     swr_vertex_shader *swr_vs)
 171 {
 172    memset(&key, 0, sizeof(key));
 173
 174    key.clip_plane_mask =
 175       swr_vs->info.base.clipdist_writemask ?
 176       swr_vs->info.base.clipdist_writemask & ctx->rasterizer->clip_plane_enable :
 177       ctx->rasterizer->clip_plane_enable;
 178
 179    swr_generate_sampler_key(swr_vs->info, ctx, PIPE_SHADER_VERTEX, key);
 180 }
 181
 182 void
 183 swr_generate_fetch_key(struct swr_jit_fetch_key &key,
 184                        struct swr_vertex_element_state *velems)
 185 {
 186    memset(&key, 0, sizeof(key));
 187
 188    key.fsState = velems->fsState;
 189 }
 190
 191 void
 192 swr_generate_gs_key(struct swr_jit_gs_key &key,
 193                     struct swr_context *ctx,
 194                     swr_geometry_shader *swr_gs)
 195 {
 196    memset(&key, 0, sizeof(key));
 197
 198    struct tgsi_shader_info *pPrevShader = &ctx->vs->info.base;
 199
 200    memcpy(&key.vs_output_semantic_name,
 201           &pPrevShader->output_semantic_name,
 202           sizeof(key.vs_output_semantic_name));
 203    memcpy(&key.vs_output_semantic_idx,
 204           &pPrevShader->output_semantic_index,
 205           sizeof(key.vs_output_semantic_idx));
 206
 207    swr_generate_sampler_key(swr_gs->info, ctx, PIPE_SHADER_GEOMETRY, key);
 208 }
 209
 210 struct BuilderSWR : public Builder {
 211    BuilderSWR(JitManager *pJitMgr, const char *pName)
 212       : Builder(pJitMgr)
 213    {
 214       pJitMgr->SetupNewModule();
 215       gallivm = gallivm_create(pName, wrap(&JM()->mContext));
 216       pJitMgr->mpCurrentModule = unwrap(gallivm->module);
 217    }
 218
 219    ~BuilderSWR() {
 220       gallivm_free_ir(gallivm);
 221    }
 222
 223    void WriteVS(Value *pVal, Value *pVsContext, Value *pVtxOutput,
 224                 unsigned slot, unsigned channel);
 225
 226    struct gallivm_state *gallivm;
 227    PFN_VERTEX_FUNC CompileVS(struct swr_context *ctx, swr_jit_vs_key &key);
 228    PFN_PIXEL_KERNEL CompileFS(struct swr_context *ctx, swr_jit_fs_key &key);
 229    PFN_GS_FUNC CompileGS(struct swr_context *ctx, swr_jit_gs_key &key);
 230
 231    LLVMValueRef
 232    swr_gs_llvm_fetch_input(const struct lp_build_tgsi_gs_iface *gs_iface,
 233                            struct lp_build_tgsi_context * bld_base,
 234                            boolean is_vindex_indirect,
 235                            LLVMValueRef vertex_index,
 236                            boolean is_aindex_indirect,
 237                            LLVMValueRef attrib_index,
 238                            LLVMValueRef swizzle_index);
 239    void
 240    swr_gs_llvm_emit_vertex(const struct lp_build_tgsi_gs_iface *gs_base,
 241                            struct lp_build_tgsi_context * bld_base,
 242                            LLVMValueRef (*outputs)[4],
 243                            LLVMValueRef emitted_vertices_vec);
 244
 245    void
 246    swr_gs_llvm_end_primitive(const struct lp_build_tgsi_gs_iface *gs_base,
 247                              struct lp_build_tgsi_context * bld_base,
 248                              LLVMValueRef verts_per_prim_vec,
 249                              LLVMValueRef emitted_prims_vec);
 250
 251    void
 252    swr_gs_llvm_epilogue(const struct lp_build_tgsi_gs_iface *gs_base,
 253                         struct lp_build_tgsi_context * bld_base,
 254                         LLVMValueRef total_emitted_vertices_vec,
 255                         LLVMValueRef emitted_prims_vec);
 256
 257 };
 258
 259 struct swr_gs_llvm_iface {
 260    struct lp_build_tgsi_gs_iface base;
 261    struct tgsi_shader_info *info;
 262
 263    BuilderSWR *pBuilder;
 264
 265    Value *pGsCtx;
 266    SWR_GS_STATE *pGsState;
 267    uint32_t num_outputs;
 268    uint32_t num_verts_per_prim;
 269
 270    Value *pVtxAttribMap;
 271 };
 272
 273 // trampoline functions so we can use the builder llvm construction methods
 274 static LLVMValueRef
 275 swr_gs_llvm_fetch_input(const struct lp_build_tgsi_gs_iface *gs_iface,
 276                            struct lp_build_tgsi_context * bld_base,
 277                            boolean is_vindex_indirect,
 278                            LLVMValueRef vertex_index,
 279                            boolean is_aindex_indirect,
 280                            LLVMValueRef attrib_index,
 281                            LLVMValueRef swizzle_index)
 282 {
 283     swr_gs_llvm_iface *iface = (swr_gs_llvm_iface*)gs_iface;
 284
 285     return iface->pBuilder->swr_gs_llvm_fetch_input(gs_iface, bld_base,
 286                                                    is_vindex_indirect,
 287                                                    vertex_index,
 288                                                    is_aindex_indirect,
 289                                                    attrib_index,
 290                                                    swizzle_index);
 291 }
 292
 293 static void
 294 swr_gs_llvm_emit_vertex(const struct lp_build_tgsi_gs_iface *gs_base,
 295                            struct lp_build_tgsi_context * bld_base,
 296                            LLVMValueRef (*outputs)[4],
 297                            LLVMValueRef emitted_vertices_vec)
 298 {
 299     swr_gs_llvm_iface *iface = (swr_gs_llvm_iface*)gs_base;
 300
 301     iface->pBuilder->swr_gs_llvm_emit_vertex(gs_base, bld_base,
 302                                             outputs,
 303                                             emitted_vertices_vec);
 304 }
 305
 306 static void
 307 swr_gs_llvm_end_primitive(const struct lp_build_tgsi_gs_iface *gs_base,
 308                              struct lp_build_tgsi_context * bld_base,
 309                              LLVMValueRef verts_per_prim_vec,
 310                              LLVMValueRef emitted_prims_vec)
 311 {
 312     swr_gs_llvm_iface *iface = (swr_gs_llvm_iface*)gs_base;
 313
 314     iface->pBuilder->swr_gs_llvm_end_primitive(gs_base, bld_base,
 315                                               verts_per_prim_vec,
 316                                               emitted_prims_vec);
 317 }
 318
 319 static void
 320 swr_gs_llvm_epilogue(const struct lp_build_tgsi_gs_iface *gs_base,
 321                         struct lp_build_tgsi_context * bld_base,
 322                         LLVMValueRef total_emitted_vertices_vec,
 323                         LLVMValueRef emitted_prims_vec)
 324 {
 325     swr_gs_llvm_iface *iface = (swr_gs_llvm_iface*)gs_base;
 326
 327     iface->pBuilder->swr_gs_llvm_epilogue(gs_base, bld_base,
 328                                          total_emitted_vertices_vec,
 329                                          emitted_prims_vec);
 330 }
 331
 332 LLVMValueRef
 333 BuilderSWR::swr_gs_llvm_fetch_input(const struct lp_build_tgsi_gs_iface *gs_iface,
 334                            struct lp_build_tgsi_context * bld_base,
 335                            boolean is_vindex_indirect,
 336                            LLVMValueRef vertex_index,
 337                            boolean is_aindex_indirect,
 338                            LLVMValueRef attrib_index,
 339                            LLVMValueRef swizzle_index)
 340 {
 341     swr_gs_llvm_iface *iface = (swr_gs_llvm_iface*)gs_iface;
 342
 343     IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 344
 345     assert(is_vindex_indirect == false && is_aindex_indirect == false);
 346
 347     Value *attrib =
 348        LOAD(GEP(iface->pVtxAttribMap, {C(0), unwrap(attrib_index)}));
 349
 350     Value *pVertex = LOAD(iface->pGsCtx, {0, SWR_GS_CONTEXT_pVerts});
 351     Value *pInputVertStride = LOAD(iface->pGsCtx, {0, SWR_GS_CONTEXT_inputVertStride});
 352
 353     Value *pVector = ADD(MUL(unwrap(vertex_index), pInputVertStride), attrib);
 354
 355     Value *pInput = LOAD(GEP(pVertex, {pVector, unwrap(swizzle_index)}));
 356
 357     return wrap(pInput);
 358 }
 359
 360 // GS output stream layout
 361 #define VERTEX_COUNT_SIZE 32
 362 #define CONTROL_HEADER_SIZE (8*32)
 363
 364 void
 365 BuilderSWR::swr_gs_llvm_emit_vertex(const struct lp_build_tgsi_gs_iface *gs_base,
 366                            struct lp_build_tgsi_context * bld_base,
 367                            LLVMValueRef (*outputs)[4],
 368                            LLVMValueRef emitted_vertices_vec)
 369 {
 370     swr_gs_llvm_iface *iface = (swr_gs_llvm_iface*)gs_base;
 371
 372     IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 373
 374     const uint32_t headerSize = VERTEX_COUNT_SIZE + CONTROL_HEADER_SIZE;
 375     const uint32_t attribSize = 4 * sizeof(float);
 376     const uint32_t vertSize = attribSize * SWR_VTX_NUM_SLOTS;
 377     Value *pVertexOffset = MUL(unwrap(emitted_vertices_vec), VIMMED1(vertSize));
 378
 379     Value *vMask = LOAD(iface->pGsCtx, {0, SWR_GS_CONTEXT_mask});
 380     Value *vMask1 = TRUNC(vMask, VectorType::get(mInt1Ty, mVWidth));
 381
 382     Value *pStack = STACKSAVE();
 383     Value *pTmpPtr = ALLOCA(mFP32Ty, C(4)); // used for dummy write for lane masking
 384
 385     for (uint32_t attrib = 0; attrib < iface->num_outputs; ++attrib) {
 386        uint32_t attribSlot = attrib;
 387        uint32_t sgvChannel = 0;
 388        if (iface->info->output_semantic_name[attrib] == TGSI_SEMANTIC_PSIZE) {
 389           attribSlot = VERTEX_SGV_SLOT;
 390           sgvChannel = VERTEX_SGV_POINT_SIZE_COMP;
 391        } else if (iface->info->output_semantic_name[attrib] == TGSI_SEMANTIC_LAYER) {
 392           attribSlot = VERTEX_SGV_SLOT;
 393           sgvChannel = VERTEX_SGV_RTAI_COMP;
 394        } else if (iface->info->output_semantic_name[attrib] == TGSI_SEMANTIC_POSITION) {
 395           attribSlot = VERTEX_POSITION_SLOT;
 396        } else {
 397           attribSlot = VERTEX_ATTRIB_START_SLOT + attrib;
 398           if (iface->info->writes_position) {
 399              attribSlot--;
 400           }
 401        }
 402
 403        Value *pOutputOffset = ADD(pVertexOffset, VIMMED1(headerSize + attribSize * attribSlot)); // + sgvChannel ?
 404
 405        for (uint32_t lane = 0; lane < mVWidth; ++lane) {
 406           Value *pLaneOffset = VEXTRACT(pOutputOffset, C(lane));
 407           Value *pStream = LOAD(iface->pGsCtx, {0, SWR_GS_CONTEXT_pStreams, lane});
 408           Value *pStreamOffset = GEP(pStream, pLaneOffset);
 409           pStreamOffset = BITCAST(pStreamOffset, mFP32PtrTy);
 410
 411           Value *pLaneMask = VEXTRACT(vMask1, C(lane));
 412           pStreamOffset = SELECT(pLaneMask, pStreamOffset, pTmpPtr);
 413
 414           for (uint32_t channel = 0; channel < 4; ++channel) {
 415              Value *vData;
 416
 417              if (attribSlot == VERTEX_SGV_SLOT)
 418                 vData = LOAD(unwrap(outputs[attrib][0]));
 419              else
 420                 vData = LOAD(unwrap(outputs[attrib][channel]));
 421
 422              if (attribSlot != VERTEX_SGV_SLOT ||
 423                  sgvChannel == channel) {
 424                 vData = VEXTRACT(vData, C(lane));
 425                 STORE(vData, pStreamOffset);
 426              }
 427              pStreamOffset = GEP(pStreamOffset, C(1));
 428           }
 429        }
 430     }
 431
 432     STACKRESTORE(pStack);
 433 }
 434
 435 void
 436 BuilderSWR::swr_gs_llvm_end_primitive(const struct lp_build_tgsi_gs_iface *gs_base,
 437                              struct lp_build_tgsi_context * bld_base,
 438                              LLVMValueRef verts_per_prim_vec,
 439                              LLVMValueRef emitted_prims_vec)
 440 {
 441     swr_gs_llvm_iface *iface = (swr_gs_llvm_iface*)gs_base;
 442
 443     IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 444
 445     Value *vMask = LOAD(iface->pGsCtx, { 0, SWR_GS_CONTEXT_mask });
 446     Value *vMask1 = TRUNC(vMask, VectorType::get(mInt1Ty, 8));
 447
 448     uint32_t vertsPerPrim = iface->num_verts_per_prim;
 449
 450     Value *vCount =
 451        ADD(MUL(unwrap(emitted_prims_vec), VIMMED1(vertsPerPrim)),
 452            unwrap(verts_per_prim_vec));
 453
 454     struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
 455     vCount = LOAD(unwrap(bld->total_emitted_vertices_vec_ptr));
 456
 457     struct lp_exec_mask *exec_mask = &bld->exec_mask;
 458     Value *mask = unwrap(lp_build_mask_value(bld->mask));
 459     if (exec_mask->has_mask)
 460        mask = AND(mask, unwrap(exec_mask->exec_mask));
 461
 462     Value *cmpMask = VMASK(ICMP_NE(unwrap(verts_per_prim_vec), VIMMED1(0)));
 463     mask = AND(mask, cmpMask);
 464     vMask1 = TRUNC(mask, VectorType::get(mInt1Ty, 8));
 465
 466     vCount = SUB(vCount, VIMMED1(1));
 467     Value *vOffset = ADD(UDIV(vCount, VIMMED1(8)), VIMMED1(VERTEX_COUNT_SIZE));
 468     Value *vValue = SHL(VIMMED1(1), UREM(vCount, VIMMED1(8)));
 469
 470     vValue = TRUNC(vValue, VectorType::get(mInt8Ty, 8));
 471
 472     Value *pStack = STACKSAVE();
 473     Value *pTmpPtr = ALLOCA(mInt8Ty, C(4)); // used for dummy read/write for lane masking
 474
 475     for (uint32_t lane = 0; lane < mVWidth; ++lane) {
 476        Value *vLaneOffset = VEXTRACT(vOffset, C(lane));
 477        Value *pStream = LOAD(iface->pGsCtx, {0, SWR_GS_CONTEXT_pStreams, lane});
 478        Value *pStreamOffset = GEP(pStream, vLaneOffset);
 479
 480        Value *pLaneMask = VEXTRACT(vMask1, C(lane));
 481        pStreamOffset = SELECT(pLaneMask, pStreamOffset, pTmpPtr);
 482
 483        Value *vVal = LOAD(pStreamOffset);
 484        vVal = OR(vVal, VEXTRACT(vValue, C(lane)));
 485        STORE(vVal, pStreamOffset);
 486     }
 487
 488     STACKRESTORE(pStack);
 489 }
 490
 491 void
 492 BuilderSWR::swr_gs_llvm_epilogue(const struct lp_build_tgsi_gs_iface *gs_base,
 493                         struct lp_build_tgsi_context * bld_base,
 494                         LLVMValueRef total_emitted_vertices_vec,
 495                         LLVMValueRef emitted_prims_vec)
 496 {
 497    swr_gs_llvm_iface *iface = (swr_gs_llvm_iface*)gs_base;
 498
 499    IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 500
 501    // Store emit count to each output stream in the first DWORD
 502    for (uint32_t lane = 0; lane < mVWidth; ++lane)
 503    {
 504       Value* pStream = LOAD(iface->pGsCtx, {0, SWR_GS_CONTEXT_pStreams, lane});
 505       pStream = BITCAST(pStream, mInt32PtrTy);
 506       Value* pLaneCount = VEXTRACT(unwrap(total_emitted_vertices_vec), C(lane));
 507       STORE(pLaneCount, pStream);
 508    }
 509 }
 510
 511 PFN_GS_FUNC
 512 BuilderSWR::CompileGS(struct swr_context *ctx, swr_jit_gs_key &key)
 513 {
 514    SWR_GS_STATE *pGS = &ctx->gs->gsState;
 515    struct tgsi_shader_info *info = &ctx->gs->info.base;
 516
 517    memset(pGS, 0, sizeof(*pGS));
 518
 519    pGS->gsEnable = true;
 520
 521    pGS->numInputAttribs = info->num_inputs;
 522    pGS->outputTopology =
 523       swr_convert_prim_topology(info->properties[TGSI_PROPERTY_GS_OUTPUT_PRIM]);
 524    pGS->maxNumVerts = info->properties[TGSI_PROPERTY_GS_MAX_OUTPUT_VERTICES];
 525    pGS->instanceCount = info->properties[TGSI_PROPERTY_GS_INVOCATIONS];
 526
 527    // XXX: single stream for now...
 528    pGS->isSingleStream = true;
 529    pGS->singleStreamID = 0;
 530
 531    pGS->vertexAttribOffset = VERTEX_ATTRIB_START_SLOT; // TODO: optimize
 532    pGS->srcVertexAttribOffset = VERTEX_ATTRIB_START_SLOT; // TODO: optimize
 533    pGS->inputVertStride = pGS->numInputAttribs + pGS->vertexAttribOffset;
 534    pGS->outputVertexSize = SWR_VTX_NUM_SLOTS;
 535    pGS->controlDataSize = 8; // GS ouputs max of 8 32B units
 536    pGS->controlDataOffset = VERTEX_COUNT_SIZE;
 537    pGS->outputVertexOffset = pGS->controlDataOffset + CONTROL_HEADER_SIZE;
 538
 539    pGS->allocationSize =
 540       VERTEX_COUNT_SIZE + // vertex count
 541       CONTROL_HEADER_SIZE + // control header
 542       (SWR_VTX_NUM_SLOTS * 16) * // sizeof vertex
 543       pGS->maxNumVerts; // num verts
 544
 545    struct swr_geometry_shader *gs = ctx->gs;
 546
 547    LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
 548    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
 549
 550    memset(outputs, 0, sizeof(outputs));
 551
 552    AttrBuilder attrBuilder;
 553    attrBuilder.addStackAlignmentAttr(JM()->mVWidth * sizeof(float));
 554
 555    std::vector<Type *> gsArgs{PointerType::get(Gen_swr_draw_context(JM()), 0),
 556                               PointerType::get(Gen_SWR_GS_CONTEXT(JM()), 0)};
 557    FunctionType *vsFuncType =
 558       FunctionType::get(Type::getVoidTy(JM()->mContext), gsArgs, false);
 559
 560    // create new vertex shader function
 561    auto pFunction = Function::Create(vsFuncType,
 562                                      GlobalValue::ExternalLinkage,
 563                                      "GS",
 564                                      JM()->mpCurrentModule);
 565 #if HAVE_LLVM < 0x0500
 566    AttributeSet attrSet = AttributeSet::get(
 567       JM()->mContext, AttributeSet::FunctionIndex, attrBuilder);
 568    pFunction->addAttributes(AttributeSet::FunctionIndex, attrSet);
 569 #else
 570    pFunction->addAttributes(AttributeList::FunctionIndex, attrBuilder);
 571 #endif
 572
 573    BasicBlock *block = BasicBlock::Create(JM()->mContext, "entry", pFunction);
 574    IRB()->SetInsertPoint(block);
 575    LLVMPositionBuilderAtEnd(gallivm->builder, wrap(block));
 576
 577    auto argitr = pFunction->arg_begin();
 578    Value *hPrivateData = &*argitr++;
 579    hPrivateData->setName("hPrivateData");
 580    Value *pGsCtx = &*argitr++;
 581    pGsCtx->setName("gsCtx");
 582
 583    Value *consts_ptr =
 584       GEP(hPrivateData, {C(0), C(swr_draw_context_constantGS)});
 585    consts_ptr->setName("gs_constants");
 586    Value *const_sizes_ptr =
 587       GEP(hPrivateData, {0, swr_draw_context_num_constantsGS});
 588    const_sizes_ptr->setName("num_gs_constants");
 589
 590    struct lp_build_sampler_soa *sampler =
 591       swr_sampler_soa_create(key.sampler, PIPE_SHADER_GEOMETRY);
 592
 593    struct lp_bld_tgsi_system_values system_values;
 594    memset(&system_values, 0, sizeof(system_values));
 595    system_values.prim_id = wrap(LOAD(pGsCtx, {0, SWR_GS_CONTEXT_PrimitiveID}));
 596    system_values.instance_id = wrap(LOAD(pGsCtx, {0, SWR_GS_CONTEXT_InstanceID}));
 597
 598    std::vector<Constant*> mapConstants;
 599    Value *vtxAttribMap = ALLOCA(ArrayType::get(mInt32Ty, PIPE_MAX_SHADER_INPUTS));
 600    for (unsigned slot = 0; slot < info->num_inputs; slot++) {
 601       ubyte semantic_name = info->input_semantic_name[slot];
 602       ubyte semantic_idx = info->input_semantic_index[slot];
 603
 604       unsigned vs_slot = locate_linkage(semantic_name, semantic_idx, &ctx->vs->info.base);
 605
 606       vs_slot += VERTEX_ATTRIB_START_SLOT;
 607
 608       if (ctx->vs->info.base.output_semantic_name[0] == TGSI_SEMANTIC_POSITION)
 609          vs_slot--;
 610
 611       if (semantic_name == TGSI_SEMANTIC_POSITION)
 612          vs_slot = VERTEX_POSITION_SLOT;
 613
 614       STORE(C(vs_slot), vtxAttribMap, {0, slot});
 615       mapConstants.push_back(C(vs_slot));
 616    }
 617
 618    struct lp_build_mask_context mask;
 619    Value *mask_val = LOAD(pGsCtx, {0, SWR_GS_CONTEXT_mask}, "gsMask");
 620    lp_build_mask_begin(&mask, gallivm,
 621                        lp_type_float_vec(32, 32 * 8), wrap(mask_val));
 622
 623    // zero out cut buffer so we can load/modify/store bits
 624    for (uint32_t lane = 0; lane < mVWidth; ++lane)
 625    {
 626       Value* pStream = LOAD(pGsCtx, {0, SWR_GS_CONTEXT_pStreams, lane});
 627       MEMSET(pStream, C((char)0), VERTEX_COUNT_SIZE + CONTROL_HEADER_SIZE, sizeof(float) * KNOB_SIMD_WIDTH);
 628    }
 629
 630    struct swr_gs_llvm_iface gs_iface;
 631    gs_iface.base.fetch_input = ::swr_gs_llvm_fetch_input;
 632    gs_iface.base.emit_vertex = ::swr_gs_llvm_emit_vertex;
 633    gs_iface.base.end_primitive = ::swr_gs_llvm_end_primitive;
 634    gs_iface.base.gs_epilogue = ::swr_gs_llvm_epilogue;
 635    gs_iface.pBuilder = this;
 636    gs_iface.pGsCtx = pGsCtx;
 637    gs_iface.pGsState = pGS;
 638    gs_iface.num_outputs = gs->info.base.num_outputs;
 639    gs_iface.num_verts_per_prim =
 640       u_vertices_per_prim((pipe_prim_type)info->properties[TGSI_PROPERTY_GS_OUTPUT_PRIM]);
 641    gs_iface.info = info;
 642    gs_iface.pVtxAttribMap = vtxAttribMap;
 643
 644    lp_build_tgsi_soa(gallivm,
 645                      gs->pipe.tokens,
 646                      lp_type_float_vec(32, 32 * 8),
 647                      &mask,
 648                      wrap(consts_ptr),
 649                      wrap(const_sizes_ptr),
 650                      &system_values,
 651                      inputs,
 652                      outputs,
 653                      wrap(hPrivateData), // (sampler context)
 654                      NULL, // thread data
 655                      sampler,
 656                      &gs->info.base,
 657                      &gs_iface.base);
 658
 659    lp_build_mask_end(&mask);
 660
 661    sampler->destroy(sampler);
 662
 663    IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 664
 665    RET_VOID();
 666
 667    gallivm_verify_function(gallivm, wrap(pFunction));
 668    gallivm_compile_module(gallivm);
 669
 670    PFN_GS_FUNC pFunc =
 671       (PFN_GS_FUNC)gallivm_jit_function(gallivm, wrap(pFunction));
 672
 673    debug_printf("geom shader  %p\n", pFunc);
 674    assert(pFunc && "Error: GeomShader = NULL");
 675
 676    JM()->mIsModuleFinalized = true;
 677
 678    return pFunc;
 679 }
 680
 681 PFN_GS_FUNC
 682 swr_compile_gs(struct swr_context *ctx, swr_jit_gs_key &key)
 683 {
 684    BuilderSWR builder(
 685       reinterpret_cast<JitManager *>(swr_screen(ctx->pipe.screen)->hJitMgr),
 686       "GS");
 687    PFN_GS_FUNC func = builder.CompileGS(ctx, key);
 688
 689    ctx->gs->map.insert(std::make_pair(key, make_unique<VariantGS>(builder.gallivm, func)));
 690    return func;
 691 }
 692
 693 void
 694 BuilderSWR::WriteVS(Value *pVal, Value *pVsContext, Value *pVtxOutput, unsigned slot, unsigned channel)
 695 {
 696 #if USE_SIMD16_FRONTEND && !USE_SIMD16_SHADERS
 697    // interleave the simdvertex components into the dest simd16vertex
 698    //   slot16offset = slot8offset * 2
 699    //   comp16offset = comp8offset * 2 + alternateOffset
 700
 701    Value *offset = LOAD(pVsContext, { 0, SWR_VS_CONTEXT_AlternateOffset });
 702    Value *pOut = GEP(pVtxOutput, { C(0), C(0), C(slot * 2), offset } );
 703    STORE(pVal, pOut, {channel * 2});
 704 #else
 705    Value *pOut = GEP(pVtxOutput, {0, 0, slot});
 706    STORE(pVal, pOut, {0, channel});
 707 #endif
 708 }
 709
 710 PFN_VERTEX_FUNC
 711 BuilderSWR::CompileVS(struct swr_context *ctx, swr_jit_vs_key &key)
 712 {
 713    struct swr_vertex_shader *swr_vs = ctx->vs;
 714
 715    LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
 716    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
 717
 718    memset(outputs, 0, sizeof(outputs));
 719
 720    AttrBuilder attrBuilder;
 721    attrBuilder.addStackAlignmentAttr(JM()->mVWidth * sizeof(float));
 722
 723    std::vector<Type *> vsArgs{PointerType::get(Gen_swr_draw_context(JM()), 0),
 724                               PointerType::get(Gen_SWR_VS_CONTEXT(JM()), 0)};
 725    FunctionType *vsFuncType =
 726       FunctionType::get(Type::getVoidTy(JM()->mContext), vsArgs, false);
 727
 728    // create new vertex shader function
 729    auto pFunction = Function::Create(vsFuncType,
 730                                      GlobalValue::ExternalLinkage,
 731                                      "VS",
 732                                      JM()->mpCurrentModule);
 733 #if HAVE_LLVM < 0x0500
 734    AttributeSet attrSet = AttributeSet::get(
 735       JM()->mContext, AttributeSet::FunctionIndex, attrBuilder);
 736    pFunction->addAttributes(AttributeSet::FunctionIndex, attrSet);
 737 #else
 738    pFunction->addAttributes(AttributeList::FunctionIndex, attrBuilder);
 739 #endif
 740
 741    BasicBlock *block = BasicBlock::Create(JM()->mContext, "entry", pFunction);
 742    IRB()->SetInsertPoint(block);
 743    LLVMPositionBuilderAtEnd(gallivm->builder, wrap(block));
 744
 745    auto argitr = pFunction->arg_begin();
 746    Value *hPrivateData = &*argitr++;
 747    hPrivateData->setName("hPrivateData");
 748    Value *pVsCtx = &*argitr++;
 749    pVsCtx->setName("vsCtx");
 750
 751    Value *consts_ptr = GEP(hPrivateData, {C(0), C(swr_draw_context_constantVS)});
 752
 753    consts_ptr->setName("vs_constants");
 754    Value *const_sizes_ptr =
 755       GEP(hPrivateData, {0, swr_draw_context_num_constantsVS});
 756    const_sizes_ptr->setName("num_vs_constants");
 757
 758    Value *vtxInput = LOAD(pVsCtx, {0, SWR_VS_CONTEXT_pVin});
 759 #if USE_SIMD16_SHADERS
 760    vtxInput = BITCAST(vtxInput, PointerType::get(Gen_simd16vertex(JM()), 0));
 761 #endif
 762
 763    for (uint32_t attrib = 0; attrib < PIPE_MAX_SHADER_INPUTS; attrib++) {
 764       const unsigned mask = swr_vs->info.base.input_usage_mask[attrib];
 765       for (uint32_t channel = 0; channel < TGSI_NUM_CHANNELS; channel++) {
 766          if (mask & (1 << channel)) {
 767             inputs[attrib][channel] =
 768                wrap(LOAD(vtxInput, {0, 0, attrib, channel}));
 769          }
 770       }
 771    }
 772
 773    struct lp_build_sampler_soa *sampler =
 774       swr_sampler_soa_create(key.sampler, PIPE_SHADER_VERTEX);
 775
 776    struct lp_bld_tgsi_system_values system_values;
 777    memset(&system_values, 0, sizeof(system_values));
 778    system_values.instance_id = wrap(LOAD(pVsCtx, {0, SWR_VS_CONTEXT_InstanceID}));
 779    system_values.vertex_id = wrap(LOAD(pVsCtx, {0, SWR_VS_CONTEXT_VertexID}));
 780
 781    lp_build_tgsi_soa(gallivm,
 782                      swr_vs->pipe.tokens,
 783                      lp_type_float_vec(32, 32 * mVWidth),
 784                      NULL, // mask
 785                      wrap(consts_ptr),
 786                      wrap(const_sizes_ptr),
 787                      &system_values,
 788                      inputs,
 789                      outputs,
 790                      wrap(hPrivateData), // (sampler context)
 791                      NULL, // thread data
 792                      sampler, // sampler
 793                      &swr_vs->info.base,
 794                      NULL); // geometry shader face
 795
 796    sampler->destroy(sampler);
 797
 798    IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 799
 800    Value *vtxOutput = LOAD(pVsCtx, {0, SWR_VS_CONTEXT_pVout});
 801 #if USE_SIMD16_SHADERS
 802    vtxOutput = BITCAST(vtxOutput, PointerType::get(Gen_simd16vertex(JM()), 0));
 803 #endif
 804
 805    for (uint32_t channel = 0; channel < TGSI_NUM_CHANNELS; channel++) {
 806       for (uint32_t attrib = 0; attrib < PIPE_MAX_SHADER_OUTPUTS; attrib++) {
 807          if (!outputs[attrib][channel])
 808             continue;
 809
 810          Value *val;
 811          uint32_t outSlot;
 812
 813          if (swr_vs->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_PSIZE) {
 814             if (channel != VERTEX_SGV_POINT_SIZE_COMP)
 815                continue;
 816             val = LOAD(unwrap(outputs[attrib][0]));
 817             outSlot = VERTEX_SGV_SLOT;
 818          } else if (swr_vs->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_POSITION) {
 819             val = LOAD(unwrap(outputs[attrib][channel]));
 820             outSlot = VERTEX_POSITION_SLOT;
 821          } else {
 822             val = LOAD(unwrap(outputs[attrib][channel]));
 823             outSlot = VERTEX_ATTRIB_START_SLOT + attrib;
 824             if (swr_vs->info.base.output_semantic_name[0] == TGSI_SEMANTIC_POSITION)
 825                outSlot--;
 826          }
 827
 828          WriteVS(val, pVsCtx, vtxOutput, outSlot, channel);
 829       }
 830    }
 831
 832    if (ctx->rasterizer->clip_plane_enable ||
 833        swr_vs->info.base.culldist_writemask) {
 834       unsigned clip_mask = ctx->rasterizer->clip_plane_enable;
 835
 836       unsigned cv = 0;
 837       if (swr_vs->info.base.writes_clipvertex) {
 838          cv = locate_linkage(TGSI_SEMANTIC_CLIPVERTEX, 0,
 839                              &swr_vs->info.base);
 840       } else {
 841          for (int i = 0; i < PIPE_MAX_SHADER_OUTPUTS; i++) {
 842             if (swr_vs->info.base.output_semantic_name[i] == TGSI_SEMANTIC_POSITION &&
 843                 swr_vs->info.base.output_semantic_index[i] == 0) {
 844                cv = i;
 845                break;
 846             }
 847          }
 848       }
 849       LLVMValueRef cx = LLVMBuildLoad(gallivm->builder, outputs[cv][0], "");
 850       LLVMValueRef cy = LLVMBuildLoad(gallivm->builder, outputs[cv][1], "");
 851       LLVMValueRef cz = LLVMBuildLoad(gallivm->builder, outputs[cv][2], "");
 852       LLVMValueRef cw = LLVMBuildLoad(gallivm->builder, outputs[cv][3], "");
 853
 854       for (unsigned val = 0; val < PIPE_MAX_CLIP_PLANES; val++) {
 855          // clip distance overrides user clip planes
 856          if ((swr_vs->info.base.clipdist_writemask & clip_mask & (1 << val)) ||
 857              ((swr_vs->info.base.culldist_writemask << swr_vs->info.base.num_written_clipdistance) & (1 << val))) {
 858             unsigned cv = locate_linkage(TGSI_SEMANTIC_CLIPDIST, val < 4 ? 0 : 1,
 859                                          &swr_vs->info.base);
 860             if (val < 4) {
 861                LLVMValueRef dist = LLVMBuildLoad(gallivm->builder, outputs[cv][val], "");
 862                WriteVS(unwrap(dist), pVsCtx, vtxOutput, VERTEX_CLIPCULL_DIST_LO_SLOT, val);
 863             } else {
 864                LLVMValueRef dist = LLVMBuildLoad(gallivm->builder, outputs[cv][val - 4], "");
 865                WriteVS(unwrap(dist), pVsCtx, vtxOutput, VERTEX_CLIPCULL_DIST_HI_SLOT, val - 4);
 866             }
 867             continue;
 868          }
 869
 870          if (!(clip_mask & (1 << val)))
 871             continue;
 872
 873          Value *px = LOAD(GEP(hPrivateData, {0, swr_draw_context_userClipPlanes, val, 0}));
 874          Value *py = LOAD(GEP(hPrivateData, {0, swr_draw_context_userClipPlanes, val, 1}));
 875          Value *pz = LOAD(GEP(hPrivateData, {0, swr_draw_context_userClipPlanes, val, 2}));
 876          Value *pw = LOAD(GEP(hPrivateData, {0, swr_draw_context_userClipPlanes, val, 3}));
 877          Value *dist = FADD(FMUL(unwrap(cx), VBROADCAST(px)),
 878                             FADD(FMUL(unwrap(cy), VBROADCAST(py)),
 879                                  FADD(FMUL(unwrap(cz), VBROADCAST(pz)),
 880                                       FMUL(unwrap(cw), VBROADCAST(pw)))));
 881
 882          if (val < 4)
 883             WriteVS(dist, pVsCtx, vtxOutput, VERTEX_CLIPCULL_DIST_LO_SLOT, val);
 884          else
 885             WriteVS(dist, pVsCtx, vtxOutput, VERTEX_CLIPCULL_DIST_HI_SLOT, val - 4);
 886       }
 887    }
 888
 889    RET_VOID();
 890
 891    gallivm_verify_function(gallivm, wrap(pFunction));
 892    gallivm_compile_module(gallivm);
 893
 894    //   lp_debug_dump_value(func);
 895
 896    PFN_VERTEX_FUNC pFunc =
 897       (PFN_VERTEX_FUNC)gallivm_jit_function(gallivm, wrap(pFunction));
 898
 899    debug_printf("vert shader  %p\n", pFunc);
 900    assert(pFunc && "Error: VertShader = NULL");
 901
 902    JM()->mIsModuleFinalized = true;
 903
 904    return pFunc;
 905 }
 906
 907 PFN_VERTEX_FUNC
 908 swr_compile_vs(struct swr_context *ctx, swr_jit_vs_key &key)
 909 {
 910    if (!ctx->vs->pipe.tokens)
 911       return NULL;
 912
 913    BuilderSWR builder(
 914 #if USE_SIMD16_SHADERS
 915       reinterpret_cast<JitManager *>(swr_screen(ctx->pipe.screen)->hJitMgr16),
 916 #else
 917       reinterpret_cast<JitManager *>(swr_screen(ctx->pipe.screen)->hJitMgr),
 918 #endif
 919       "VS");
 920    PFN_VERTEX_FUNC func = builder.CompileVS(ctx, key);
 921
 922    ctx->vs->map.insert(std::make_pair(key, make_unique<VariantVS>(builder.gallivm, func)));
 923    return func;
 924 }
 925
 926 unsigned
 927 swr_so_adjust_attrib(unsigned in_attrib,
 928                      swr_vertex_shader *swr_vs)
 929 {
 930    ubyte semantic_name;
 931    unsigned attrib;
 932
 933    attrib = in_attrib + VERTEX_ATTRIB_START_SLOT;
 934
 935    if (swr_vs) {
 936       semantic_name = swr_vs->info.base.output_semantic_name[in_attrib];
 937       if (semantic_name == TGSI_SEMANTIC_POSITION) {
 938          attrib = VERTEX_POSITION_SLOT;
 939       } else if (semantic_name == TGSI_SEMANTIC_PSIZE) {
 940          attrib = VERTEX_SGV_SLOT;
 941       } else if (semantic_name == TGSI_SEMANTIC_LAYER) {
 942          attrib = VERTEX_SGV_SLOT;
 943       } else {
 944          if (swr_vs->info.base.writes_position) {
 945                attrib--;
 946          }
 947       }
 948    }
 949
 950    return attrib;
 951 }
 952
 953 static unsigned
 954 locate_linkage(ubyte name, ubyte index, struct tgsi_shader_info *info)
 955 {
 956    for (int i = 0; i < PIPE_MAX_SHADER_OUTPUTS; i++) {
 957       if ((info->output_semantic_name[i] == name)
 958           && (info->output_semantic_index[i] == index)) {
 959          return i;
 960       }
 961    }
 962
 963    return 0xFFFFFFFF;
 964 }
 965
 966 PFN_PIXEL_KERNEL
 967 BuilderSWR::CompileFS(struct swr_context *ctx, swr_jit_fs_key &key)
 968 {
 969    struct swr_fragment_shader *swr_fs = ctx->fs;
 970
 971    struct tgsi_shader_info *pPrevShader;
 972    if (ctx->gs)
 973       pPrevShader = &ctx->gs->info.base;
 974    else
 975       pPrevShader = &ctx->vs->info.base;
 976
 977    LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
 978    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
 979
 980    memset(inputs, 0, sizeof(inputs));
 981    memset(outputs, 0, sizeof(outputs));
 982
 983    struct lp_build_sampler_soa *sampler = NULL;
 984
 985    AttrBuilder attrBuilder;
 986    attrBuilder.addStackAlignmentAttr(JM()->mVWidth * sizeof(float));
 987
 988    std::vector<Type *> fsArgs{PointerType::get(Gen_swr_draw_context(JM()), 0),
 989                               PointerType::get(Gen_SWR_PS_CONTEXT(JM()), 0)};
 990    FunctionType *funcType =
 991       FunctionType::get(Type::getVoidTy(JM()->mContext), fsArgs, false);
 992
 993    auto pFunction = Function::Create(funcType,
 994                                      GlobalValue::ExternalLinkage,
 995                                      "FS",
 996                                      JM()->mpCurrentModule);
 997 #if HAVE_LLVM < 0x0500
 998    AttributeSet attrSet = AttributeSet::get(
 999       JM()->mContext, AttributeSet::FunctionIndex, attrBuilder);
1000    pFunction->addAttributes(AttributeSet::FunctionIndex, attrSet);
1001 #else
1002    pFunction->addAttributes(AttributeList::FunctionIndex, attrBuilder);
1003 #endif
1004
1005    BasicBlock *block = BasicBlock::Create(JM()->mContext, "entry", pFunction);
1006    IRB()->SetInsertPoint(block);
1007    LLVMPositionBuilderAtEnd(gallivm->builder, wrap(block));
1008
1009    auto args = pFunction->arg_begin();
1010    Value *hPrivateData = &*args++;
1011    hPrivateData->setName("hPrivateData");
1012    Value *pPS = &*args++;
1013    pPS->setName("psCtx");
1014
1015    Value *consts_ptr = GEP(hPrivateData, {0, swr_draw_context_constantFS});
1016    consts_ptr->setName("fs_constants");
1017    Value *const_sizes_ptr =
1018       GEP(hPrivateData, {0, swr_draw_context_num_constantsFS});
1019    const_sizes_ptr->setName("num_fs_constants");
1020
1021    // load *pAttribs, *pPerspAttribs
1022    Value *pRawAttribs = LOAD(pPS, {0, SWR_PS_CONTEXT_pAttribs}, "pRawAttribs");
1023    Value *pPerspAttribs =
1024       LOAD(pPS, {0, SWR_PS_CONTEXT_pPerspAttribs}, "pPerspAttribs");
1025
1026    swr_fs->constantMask = 0;
1027    swr_fs->flatConstantMask = 0;
1028    swr_fs->pointSpriteMask = 0;
1029
1030    for (int attrib = 0; attrib < PIPE_MAX_SHADER_INPUTS; attrib++) {
1031       const unsigned mask = swr_fs->info.base.input_usage_mask[attrib];
1032       const unsigned interpMode = swr_fs->info.base.input_interpolate[attrib];
1033       const unsigned interpLoc = swr_fs->info.base.input_interpolate_loc[attrib];
1034
1035       if (!mask)
1036          continue;
1037
1038       // load i,j
1039       Value *vi = nullptr, *vj = nullptr;
1040       switch (interpLoc) {
1041       case TGSI_INTERPOLATE_LOC_CENTER:
1042          vi = LOAD(pPS, {0, SWR_PS_CONTEXT_vI, PixelPositions_center}, "i");
1043          vj = LOAD(pPS, {0, SWR_PS_CONTEXT_vJ, PixelPositions_center}, "j");
1044          break;
1045       case TGSI_INTERPOLATE_LOC_CENTROID:
1046          vi = LOAD(pPS, {0, SWR_PS_CONTEXT_vI, PixelPositions_centroid}, "i");
1047          vj = LOAD(pPS, {0, SWR_PS_CONTEXT_vJ, PixelPositions_centroid}, "j");
1048          break;
1049       case TGSI_INTERPOLATE_LOC_SAMPLE:
1050          vi = LOAD(pPS, {0, SWR_PS_CONTEXT_vI, PixelPositions_sample}, "i");
1051          vj = LOAD(pPS, {0, SWR_PS_CONTEXT_vJ, PixelPositions_sample}, "j");
1052          break;
1053       }
1054
1055       // load/compute w
1056       Value *vw = nullptr, *pAttribs;
1057       if (interpMode == TGSI_INTERPOLATE_PERSPECTIVE ||
1058           interpMode == TGSI_INTERPOLATE_COLOR) {
1059          pAttribs = pPerspAttribs;
1060          switch (interpLoc) {
1061          case TGSI_INTERPOLATE_LOC_CENTER:
1062             vw = VRCP(LOAD(pPS, {0, SWR_PS_CONTEXT_vOneOverW, PixelPositions_center}));
1063             break;
1064          case TGSI_INTERPOLATE_LOC_CENTROID:
1065             vw = VRCP(LOAD(pPS, {0, SWR_PS_CONTEXT_vOneOverW, PixelPositions_centroid}));
1066             break;
1067          case TGSI_INTERPOLATE_LOC_SAMPLE:
1068             vw = VRCP(LOAD(pPS, {0, SWR_PS_CONTEXT_vOneOverW, PixelPositions_sample}));
1069             break;
1070          }
1071       } else {
1072          pAttribs = pRawAttribs;
1073          vw = VIMMED1(1.f);
1074       }
1075
1076       vw->setName("w");
1077
1078       ubyte semantic_name = swr_fs->info.base.input_semantic_name[attrib];
1079       ubyte semantic_idx = swr_fs->info.base.input_semantic_index[attrib];
1080
1081       if (semantic_name == TGSI_SEMANTIC_FACE) {
1082          Value *ff =
1083             UI_TO_FP(LOAD(pPS, {0, SWR_PS_CONTEXT_frontFace}), mFP32Ty);
1084          ff = FSUB(FMUL(ff, C(2.0f)), C(1.0f));
1085          ff = VECTOR_SPLAT(JM()->mVWidth, ff, "vFrontFace");
1086
1087          inputs[attrib][0] = wrap(ff);
1088          inputs[attrib][1] = wrap(VIMMED1(0.0f));
1089          inputs[attrib][2] = wrap(VIMMED1(0.0f));
1090          inputs[attrib][3] = wrap(VIMMED1(1.0f));
1091          continue;
1092       } else if (semantic_name == TGSI_SEMANTIC_POSITION) { // gl_FragCoord
1093          if (swr_fs->info.base.properties[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER] ==
1094              TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER) {
1095             inputs[attrib][0] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vX, PixelPositions_center}, "vX"));
1096             inputs[attrib][1] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vY, PixelPositions_center}, "vY"));
1097          } else {
1098             inputs[attrib][0] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vX, PixelPositions_UL}, "vX"));
1099             inputs[attrib][1] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vY, PixelPositions_UL}, "vY"));
1100          }
1101          inputs[attrib][2] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vZ}, "vZ"));
1102          inputs[attrib][3] =
1103             wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vOneOverW, PixelPositions_center}, "vOneOverW"));
1104          continue;
1105       }
1106
1107       unsigned linkedAttrib =
1108          locate_linkage(semantic_name, semantic_idx, pPrevShader) - 1;
1109
1110       uint32_t extraAttribs = 0;
1111       if (semantic_name == TGSI_SEMANTIC_PRIMID && !ctx->gs) {
1112          /* non-gs generated primID - need to grab from swizzleMap override */
1113          linkedAttrib = pPrevShader->num_outputs - 1;
1114          swr_fs->constantMask |= 1 << linkedAttrib;
1115          extraAttribs++;
1116       } else if (semantic_name == TGSI_SEMANTIC_GENERIC &&
1117           key.sprite_coord_enable & (1 << semantic_idx)) {
1118          /* we add an extra attrib to the backendState in swr_update_derived. */
1119          linkedAttrib = pPrevShader->num_outputs + extraAttribs - 1;
1120          swr_fs->pointSpriteMask |= (1 << linkedAttrib);
1121          extraAttribs++;
1122       } else if (linkedAttrib == 0xFFFFFFFF) {
1123          inputs[attrib][0] = wrap(VIMMED1(0.0f));
1124          inputs[attrib][1] = wrap(VIMMED1(0.0f));
1125          inputs[attrib][2] = wrap(VIMMED1(0.0f));
1126          inputs[attrib][3] = wrap(VIMMED1(1.0f));
1127          /* If we're reading in color and 2-sided lighting is enabled, we have
1128           * to keep going.
1129           */
1130          if (semantic_name != TGSI_SEMANTIC_COLOR || !key.light_twoside)
1131             continue;
1132       } else {
1133          if (interpMode == TGSI_INTERPOLATE_CONSTANT) {
1134             swr_fs->constantMask |= 1 << linkedAttrib;
1135          } else if (interpMode == TGSI_INTERPOLATE_COLOR) {
1136             swr_fs->flatConstantMask |= 1 << linkedAttrib;
1137          }
1138       }
1139
1140       unsigned bcolorAttrib = 0xFFFFFFFF;
1141       Value *offset = NULL;
1142       if (semantic_name == TGSI_SEMANTIC_COLOR && key.light_twoside) {
1143          bcolorAttrib = locate_linkage(
1144                TGSI_SEMANTIC_BCOLOR, semantic_idx, pPrevShader) - 1;
1145          /* Neither front nor back colors were available. Nothing to load. */
1146          if (bcolorAttrib == 0xFFFFFFFF && linkedAttrib == 0xFFFFFFFF)
1147             continue;
1148          /* If there is no front color, just always use the back color. */
1149          if (linkedAttrib == 0xFFFFFFFF)
1150             linkedAttrib = bcolorAttrib;
1151
1152          if (bcolorAttrib != 0xFFFFFFFF) {
1153             if (interpMode == TGSI_INTERPOLATE_CONSTANT) {
1154                swr_fs->constantMask |= 1 << bcolorAttrib;
1155             } else if (interpMode == TGSI_INTERPOLATE_COLOR) {
1156                swr_fs->flatConstantMask |= 1 << bcolorAttrib;
1157             }
1158
1159             unsigned diff = 12 * (bcolorAttrib - linkedAttrib);
1160
1161             if (diff) {
1162                Value *back =
1163                   XOR(C(1), LOAD(pPS, {0, SWR_PS_CONTEXT_frontFace}), "backFace");
1164
1165                offset = MUL(back, C(diff));
1166                offset->setName("offset");
1167             }
1168          }
1169       }
1170
1171       for (int channel = 0; channel < TGSI_NUM_CHANNELS; channel++) {
1172          if (mask & (1 << channel)) {
1173             Value *indexA = C(linkedAttrib * 12 + channel);
1174             Value *indexB = C(linkedAttrib * 12 + channel + 4);
1175             Value *indexC = C(linkedAttrib * 12 + channel + 8);
1176
1177             if (offset) {
1178                indexA = ADD(indexA, offset);
1179                indexB = ADD(indexB, offset);
1180                indexC = ADD(indexC, offset);
1181             }
1182
1183             Value *va = VBROADCAST(LOAD(GEP(pAttribs, indexA)));
1184             Value *vb = VBROADCAST(LOAD(GEP(pAttribs, indexB)));
1185             Value *vc = VBROADCAST(LOAD(GEP(pAttribs, indexC)));
1186
1187             if (interpMode == TGSI_INTERPOLATE_CONSTANT) {
1188                inputs[attrib][channel] = wrap(va);
1189             } else {
1190                Value *vk = FSUB(FSUB(VIMMED1(1.0f), vi), vj);
1191
1192                vc = FMUL(vk, vc);
1193
1194                Value *interp = FMUL(va, vi);
1195                Value *interp1 = FMUL(vb, vj);
1196                interp = FADD(interp, interp1);
1197                interp = FADD(interp, vc);
1198                if (interpMode == TGSI_INTERPOLATE_PERSPECTIVE ||
1199                    interpMode == TGSI_INTERPOLATE_COLOR)
1200                   interp = FMUL(interp, vw);
1201                inputs[attrib][channel] = wrap(interp);
1202             }
1203          }
1204       }
1205    }
1206
1207    sampler = swr_sampler_soa_create(key.sampler, PIPE_SHADER_FRAGMENT);
1208
1209    struct lp_bld_tgsi_system_values system_values;
1210    memset(&system_values, 0, sizeof(system_values));
1211
1212    struct lp_build_mask_context mask;
1213    bool uses_mask = false;
1214
1215    if (swr_fs->info.base.uses_kill ||
1216        key.poly_stipple_enable) {
1217       Value *vActiveMask = NULL;
1218       if (swr_fs->info.base.uses_kill) {
1219          vActiveMask = LOAD(pPS, {0, SWR_PS_CONTEXT_activeMask}, "activeMask");
1220       }
1221       if (key.poly_stipple_enable) {
1222          // first get fragment xy coords and clip to stipple bounds
1223          Value *vXf = LOAD(pPS, {0, SWR_PS_CONTEXT_vX, PixelPositions_UL});
1224          Value *vYf = LOAD(pPS, {0, SWR_PS_CONTEXT_vY, PixelPositions_UL});
1225          Value *vXu = FP_TO_UI(vXf, mSimdInt32Ty);
1226          Value *vYu = FP_TO_UI(vYf, mSimdInt32Ty);
1227
1228          // stipple pattern is 32x32, which means that one line of stipple
1229          // is stored in one word:
1230          // vXstipple is bit offset inside 32-bit stipple word
1231          // vYstipple is word index is stipple array
1232          Value *vXstipple = AND(vXu, VIMMED1(0x1f)); // & (32-1)
1233          Value *vYstipple = AND(vYu, VIMMED1(0x1f)); // & (32-1)
1234
1235          // grab stipple pattern base address
1236          Value *stipplePtr = GEP(hPrivateData, {0, swr_draw_context_polyStipple, 0});
1237          stipplePtr = BITCAST(stipplePtr, mInt8PtrTy);
1238
1239          // peform a gather to grab stipple words for each lane
1240          Value *vStipple = GATHERDD(VUNDEF_I(), stipplePtr, vYstipple,
1241                                     VIMMED1(0xffffffff), 4);
1242
1243          // create a mask with one bit corresponding to the x stipple
1244          // and AND it with the pattern, to see if we have a bit
1245          Value *vBitMask = LSHR(VIMMED1(0x80000000), vXstipple);
1246          Value *vStippleMask = AND(vStipple, vBitMask);
1247          vStippleMask = ICMP_NE(vStippleMask, VIMMED1(0));
1248          vStippleMask = VMASK(vStippleMask);
1249
1250          if (swr_fs->info.base.uses_kill) {
1251             vActiveMask = AND(vActiveMask, vStippleMask);
1252          } else {
1253             vActiveMask = vStippleMask;
1254          }
1255       }
1256       lp_build_mask_begin(
1257          &mask, gallivm, lp_type_float_vec(32, 32 * 8), wrap(vActiveMask));
1258       uses_mask = true;
1259    }
1260
1261    lp_build_tgsi_soa(gallivm,
1262                      swr_fs->pipe.tokens,
1263                      lp_type_float_vec(32, 32 * 8),
1264                      uses_mask ? &mask : NULL, // mask
1265                      wrap(consts_ptr),
1266                      wrap(const_sizes_ptr),
1267                      &system_values,
1268                      inputs,
1269                      outputs,
1270                      wrap(hPrivateData),
1271                      NULL, // thread data
1272                      sampler, // sampler
1273                      &swr_fs->info.base,
1274                      NULL); // geometry shader face
1275
1276    sampler->destroy(sampler);
1277
1278    IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
1279
1280    for (uint32_t attrib = 0; attrib < swr_fs->info.base.num_outputs;
1281         attrib++) {
1282       switch (swr_fs->info.base.output_semantic_name[attrib]) {
1283       case TGSI_SEMANTIC_POSITION: {
1284          // write z
1285          LLVMValueRef outZ =
1286             LLVMBuildLoad(gallivm->builder, outputs[attrib][2], "");
1287          STORE(unwrap(outZ), pPS, {0, SWR_PS_CONTEXT_vZ});
1288          break;
1289       }
1290       case TGSI_SEMANTIC_COLOR: {
1291          for (uint32_t channel = 0; channel < TGSI_NUM_CHANNELS; channel++) {
1292             if (!outputs[attrib][channel])
1293                continue;
1294
1295             LLVMValueRef out =
1296                LLVMBuildLoad(gallivm->builder, outputs[attrib][channel], "");
1297             if (swr_fs->info.base.properties[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS] &&
1298                 swr_fs->info.base.output_semantic_index[attrib] == 0) {
1299                for (uint32_t rt = 0; rt < key.nr_cbufs; rt++) {
1300                   STORE(unwrap(out),
1301                         pPS,
1302                         {0, SWR_PS_CONTEXT_shaded, rt, channel});
1303                }
1304             } else {
1305                STORE(unwrap(out),
1306                      pPS,
1307                      {0,
1308                            SWR_PS_CONTEXT_shaded,
1309                            swr_fs->info.base.output_semantic_index[attrib],
1310                            channel});
1311             }
1312          }
1313          break;
1314       }
1315       default: {
1316          fprintf(stderr,
1317                  "unknown output from FS %s[%d]\n",
1318                  tgsi_semantic_names[swr_fs->info.base
1319                                         .output_semantic_name[attrib]],
1320                  swr_fs->info.base.output_semantic_index[attrib]);
1321          break;
1322       }
1323       }
1324    }
1325
1326    LLVMValueRef mask_result = 0;
1327    if (uses_mask) {
1328       mask_result = lp_build_mask_end(&mask);
1329    }
1330
1331    IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
1332
1333    if (uses_mask) {
1334       STORE(unwrap(mask_result), pPS, {0, SWR_PS_CONTEXT_activeMask});
1335    }
1336
1337    RET_VOID();
1338
1339    gallivm_verify_function(gallivm, wrap(pFunction));
1340
1341    gallivm_compile_module(gallivm);
1342
1343    PFN_PIXEL_KERNEL kernel =
1344       (PFN_PIXEL_KERNEL)gallivm_jit_function(gallivm, wrap(pFunction));
1345    debug_printf("frag shader  %p\n", kernel);
1346    assert(kernel && "Error: FragShader = NULL");
1347
1348    JM()->mIsModuleFinalized = true;
1349
1350    return kernel;
1351 }
1352
1353 PFN_PIXEL_KERNEL
1354 swr_compile_fs(struct swr_context *ctx, swr_jit_fs_key &key)
1355 {
1356    if (!ctx->fs->pipe.tokens)
1357       return NULL;
1358
1359    BuilderSWR builder(
1360       reinterpret_cast<JitManager *>(swr_screen(ctx->pipe.screen)->hJitMgr),
1361       "FS");
1362    PFN_PIXEL_KERNEL func = builder.CompileFS(ctx, key);
1363
1364    ctx->fs->map.insert(std::make_pair(key, make_unique<VariantFS>(builder.gallivm, func)));
1365    return func;
1366 }