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     Value *vert_index = unwrap(vertex_index);
 343     Value *attr_index = unwrap(attrib_index);
 344
 345     IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 346
 347     if (is_vindex_indirect || is_aindex_indirect) {
 348        int i;
 349        Value *res = unwrap(bld_base->base.zero);
 350        struct lp_type type = bld_base->base.type;
 351
 352        for (i = 0; i < type.length; i++) {
 353           Value *vert_chan_index = vert_index;
 354           Value *attr_chan_index = attr_index;
 355
 356           if (is_vindex_indirect) {
 357              vert_chan_index = VEXTRACT(vert_index, C(i));
 358           }
 359           if (is_aindex_indirect) {
 360              attr_chan_index = VEXTRACT(attr_index, C(i));
 361           }
 362
 363           Value *attrib =
 364              LOAD(GEP(iface->pVtxAttribMap, {C(0), attr_chan_index}));
 365
 366           Value *pVertex = LOAD(iface->pGsCtx, {0, SWR_GS_CONTEXT_pVerts});
 367           Value *pInputVertStride = LOAD(iface->pGsCtx, {0, SWR_GS_CONTEXT_inputVertStride});
 368
 369           Value *pVector = ADD(MUL(vert_chan_index, pInputVertStride), attrib);
 370           Value *pInput = LOAD(GEP(pVertex, {pVector, unwrap(swizzle_index)}));
 371
 372           Value *value = VEXTRACT(pInput, C(i));
 373           res = VINSERT(res, value, C(i));
 374        }
 375
 376        return wrap(res);
 377     } else {
 378        Value *attrib = LOAD(GEP(iface->pVtxAttribMap, {C(0), attr_index}));
 379
 380        Value *pVertex = LOAD(iface->pGsCtx, {0, SWR_GS_CONTEXT_pVerts});
 381        Value *pInputVertStride = LOAD(iface->pGsCtx, {0, SWR_GS_CONTEXT_inputVertStride});
 382
 383        Value *pVector = ADD(MUL(vert_index, pInputVertStride), attrib);
 384
 385        Value *pInput = LOAD(GEP(pVertex, {pVector, unwrap(swizzle_index)}));
 386
 387        return wrap(pInput);
 388     }
 389 }
 390
 391 // GS output stream layout
 392 #define VERTEX_COUNT_SIZE 32
 393 #define CONTROL_HEADER_SIZE (8*32)
 394
 395 void
 396 BuilderSWR::swr_gs_llvm_emit_vertex(const struct lp_build_tgsi_gs_iface *gs_base,
 397                            struct lp_build_tgsi_context * bld_base,
 398                            LLVMValueRef (*outputs)[4],
 399                            LLVMValueRef emitted_vertices_vec)
 400 {
 401     swr_gs_llvm_iface *iface = (swr_gs_llvm_iface*)gs_base;
 402
 403     IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 404
 405     const uint32_t headerSize = VERTEX_COUNT_SIZE + CONTROL_HEADER_SIZE;
 406     const uint32_t attribSize = 4 * sizeof(float);
 407     const uint32_t vertSize = attribSize * SWR_VTX_NUM_SLOTS;
 408     Value *pVertexOffset = MUL(unwrap(emitted_vertices_vec), VIMMED1(vertSize));
 409
 410     Value *vMask = LOAD(iface->pGsCtx, {0, SWR_GS_CONTEXT_mask});
 411     Value *vMask1 = TRUNC(vMask, VectorType::get(mInt1Ty, mVWidth));
 412
 413     Value *pStack = STACKSAVE();
 414     Value *pTmpPtr = ALLOCA(mFP32Ty, C(4)); // used for dummy write for lane masking
 415
 416     for (uint32_t attrib = 0; attrib < iface->num_outputs; ++attrib) {
 417        uint32_t attribSlot = attrib;
 418        uint32_t sgvChannel = 0;
 419        if (iface->info->output_semantic_name[attrib] == TGSI_SEMANTIC_PSIZE) {
 420           attribSlot = VERTEX_SGV_SLOT;
 421           sgvChannel = VERTEX_SGV_POINT_SIZE_COMP;
 422        } else if (iface->info->output_semantic_name[attrib] == TGSI_SEMANTIC_LAYER) {
 423           attribSlot = VERTEX_SGV_SLOT;
 424           sgvChannel = VERTEX_SGV_RTAI_COMP;
 425        } else if (iface->info->output_semantic_name[attrib] == TGSI_SEMANTIC_POSITION) {
 426           attribSlot = VERTEX_POSITION_SLOT;
 427        } else {
 428           attribSlot = VERTEX_ATTRIB_START_SLOT + attrib;
 429           if (iface->info->writes_position) {
 430              attribSlot--;
 431           }
 432        }
 433
 434        Value *pOutputOffset = ADD(pVertexOffset, VIMMED1(headerSize + attribSize * attribSlot)); // + sgvChannel ?
 435
 436        for (uint32_t lane = 0; lane < mVWidth; ++lane) {
 437           Value *pLaneOffset = VEXTRACT(pOutputOffset, C(lane));
 438           Value *pStream = LOAD(iface->pGsCtx, {0, SWR_GS_CONTEXT_pStreams, lane});
 439           Value *pStreamOffset = GEP(pStream, pLaneOffset);
 440           pStreamOffset = BITCAST(pStreamOffset, mFP32PtrTy);
 441
 442           Value *pLaneMask = VEXTRACT(vMask1, C(lane));
 443           pStreamOffset = SELECT(pLaneMask, pStreamOffset, pTmpPtr);
 444
 445           for (uint32_t channel = 0; channel < 4; ++channel) {
 446              Value *vData;
 447
 448              if (attribSlot == VERTEX_SGV_SLOT)
 449                 vData = LOAD(unwrap(outputs[attrib][0]));
 450              else
 451                 vData = LOAD(unwrap(outputs[attrib][channel]));
 452
 453              if (attribSlot != VERTEX_SGV_SLOT ||
 454                  sgvChannel == channel) {
 455                 vData = VEXTRACT(vData, C(lane));
 456                 STORE(vData, pStreamOffset);
 457              }
 458              pStreamOffset = GEP(pStreamOffset, C(1));
 459           }
 460        }
 461     }
 462
 463     STACKRESTORE(pStack);
 464 }
 465
 466 void
 467 BuilderSWR::swr_gs_llvm_end_primitive(const struct lp_build_tgsi_gs_iface *gs_base,
 468                              struct lp_build_tgsi_context * bld_base,
 469                              LLVMValueRef verts_per_prim_vec,
 470                              LLVMValueRef emitted_prims_vec)
 471 {
 472     swr_gs_llvm_iface *iface = (swr_gs_llvm_iface*)gs_base;
 473
 474     IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 475
 476     Value *vMask = LOAD(iface->pGsCtx, { 0, SWR_GS_CONTEXT_mask });
 477     Value *vMask1 = TRUNC(vMask, VectorType::get(mInt1Ty, 8));
 478
 479     uint32_t vertsPerPrim = iface->num_verts_per_prim;
 480
 481     Value *vCount =
 482        ADD(MUL(unwrap(emitted_prims_vec), VIMMED1(vertsPerPrim)),
 483            unwrap(verts_per_prim_vec));
 484
 485     struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
 486     vCount = LOAD(unwrap(bld->total_emitted_vertices_vec_ptr));
 487
 488     struct lp_exec_mask *exec_mask = &bld->exec_mask;
 489     Value *mask = unwrap(lp_build_mask_value(bld->mask));
 490     if (exec_mask->has_mask)
 491        mask = AND(mask, unwrap(exec_mask->exec_mask));
 492
 493     Value *cmpMask = VMASK(ICMP_NE(unwrap(verts_per_prim_vec), VIMMED1(0)));
 494     mask = AND(mask, cmpMask);
 495     vMask1 = TRUNC(mask, VectorType::get(mInt1Ty, 8));
 496
 497     vCount = SUB(vCount, VIMMED1(1));
 498     Value *vOffset = ADD(UDIV(vCount, VIMMED1(8)), VIMMED1(VERTEX_COUNT_SIZE));
 499     Value *vValue = SHL(VIMMED1(1), UREM(vCount, VIMMED1(8)));
 500
 501     vValue = TRUNC(vValue, VectorType::get(mInt8Ty, 8));
 502
 503     Value *pStack = STACKSAVE();
 504     Value *pTmpPtr = ALLOCA(mInt8Ty, C(4)); // used for dummy read/write for lane masking
 505
 506     for (uint32_t lane = 0; lane < mVWidth; ++lane) {
 507        Value *vLaneOffset = VEXTRACT(vOffset, C(lane));
 508        Value *pStream = LOAD(iface->pGsCtx, {0, SWR_GS_CONTEXT_pStreams, lane});
 509        Value *pStreamOffset = GEP(pStream, vLaneOffset);
 510
 511        Value *pLaneMask = VEXTRACT(vMask1, C(lane));
 512        pStreamOffset = SELECT(pLaneMask, pStreamOffset, pTmpPtr);
 513
 514        Value *vVal = LOAD(pStreamOffset);
 515        vVal = OR(vVal, VEXTRACT(vValue, C(lane)));
 516        STORE(vVal, pStreamOffset);
 517     }
 518
 519     STACKRESTORE(pStack);
 520 }
 521
 522 void
 523 BuilderSWR::swr_gs_llvm_epilogue(const struct lp_build_tgsi_gs_iface *gs_base,
 524                         struct lp_build_tgsi_context * bld_base,
 525                         LLVMValueRef total_emitted_vertices_vec,
 526                         LLVMValueRef emitted_prims_vec)
 527 {
 528    swr_gs_llvm_iface *iface = (swr_gs_llvm_iface*)gs_base;
 529
 530    IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 531
 532    // Store emit count to each output stream in the first DWORD
 533    for (uint32_t lane = 0; lane < mVWidth; ++lane)
 534    {
 535       Value* pStream = LOAD(iface->pGsCtx, {0, SWR_GS_CONTEXT_pStreams, lane});
 536       pStream = BITCAST(pStream, mInt32PtrTy);
 537       Value* pLaneCount = VEXTRACT(unwrap(total_emitted_vertices_vec), C(lane));
 538       STORE(pLaneCount, pStream);
 539    }
 540 }
 541
 542 PFN_GS_FUNC
 543 BuilderSWR::CompileGS(struct swr_context *ctx, swr_jit_gs_key &key)
 544 {
 545    SWR_GS_STATE *pGS = &ctx->gs->gsState;
 546    struct tgsi_shader_info *info = &ctx->gs->info.base;
 547
 548    memset(pGS, 0, sizeof(*pGS));
 549
 550    pGS->gsEnable = true;
 551
 552    pGS->numInputAttribs = info->num_inputs;
 553    pGS->outputTopology =
 554       swr_convert_prim_topology(info->properties[TGSI_PROPERTY_GS_OUTPUT_PRIM]);
 555    pGS->maxNumVerts = info->properties[TGSI_PROPERTY_GS_MAX_OUTPUT_VERTICES];
 556    pGS->instanceCount = info->properties[TGSI_PROPERTY_GS_INVOCATIONS];
 557
 558    // XXX: single stream for now...
 559    pGS->isSingleStream = true;
 560    pGS->singleStreamID = 0;
 561
 562    pGS->vertexAttribOffset = VERTEX_ATTRIB_START_SLOT; // TODO: optimize
 563    pGS->srcVertexAttribOffset = VERTEX_ATTRIB_START_SLOT; // TODO: optimize
 564    pGS->inputVertStride = pGS->numInputAttribs + pGS->vertexAttribOffset;
 565    pGS->outputVertexSize = SWR_VTX_NUM_SLOTS;
 566    pGS->controlDataSize = 8; // GS ouputs max of 8 32B units
 567    pGS->controlDataOffset = VERTEX_COUNT_SIZE;
 568    pGS->outputVertexOffset = pGS->controlDataOffset + CONTROL_HEADER_SIZE;
 569
 570    pGS->allocationSize =
 571       VERTEX_COUNT_SIZE + // vertex count
 572       CONTROL_HEADER_SIZE + // control header
 573       (SWR_VTX_NUM_SLOTS * 16) * // sizeof vertex
 574       pGS->maxNumVerts; // num verts
 575
 576    struct swr_geometry_shader *gs = ctx->gs;
 577
 578    LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
 579    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
 580
 581    memset(outputs, 0, sizeof(outputs));
 582
 583    AttrBuilder attrBuilder;
 584    attrBuilder.addStackAlignmentAttr(JM()->mVWidth * sizeof(float));
 585
 586    std::vector<Type *> gsArgs{PointerType::get(Gen_swr_draw_context(JM()), 0),
 587                               PointerType::get(Gen_SWR_GS_CONTEXT(JM()), 0)};
 588    FunctionType *vsFuncType =
 589       FunctionType::get(Type::getVoidTy(JM()->mContext), gsArgs, false);
 590
 591    // create new vertex shader function
 592    auto pFunction = Function::Create(vsFuncType,
 593                                      GlobalValue::ExternalLinkage,
 594                                      "GS",
 595                                      JM()->mpCurrentModule);
 596 #if HAVE_LLVM < 0x0500
 597    AttributeSet attrSet = AttributeSet::get(
 598       JM()->mContext, AttributeSet::FunctionIndex, attrBuilder);
 599    pFunction->addAttributes(AttributeSet::FunctionIndex, attrSet);
 600 #else
 601    pFunction->addAttributes(AttributeList::FunctionIndex, attrBuilder);
 602 #endif
 603
 604    BasicBlock *block = BasicBlock::Create(JM()->mContext, "entry", pFunction);
 605    IRB()->SetInsertPoint(block);
 606    LLVMPositionBuilderAtEnd(gallivm->builder, wrap(block));
 607
 608    auto argitr = pFunction->arg_begin();
 609    Value *hPrivateData = &*argitr++;
 610    hPrivateData->setName("hPrivateData");
 611    Value *pGsCtx = &*argitr++;
 612    pGsCtx->setName("gsCtx");
 613
 614    Value *consts_ptr =
 615       GEP(hPrivateData, {C(0), C(swr_draw_context_constantGS)});
 616    consts_ptr->setName("gs_constants");
 617    Value *const_sizes_ptr =
 618       GEP(hPrivateData, {0, swr_draw_context_num_constantsGS});
 619    const_sizes_ptr->setName("num_gs_constants");
 620
 621    struct lp_build_sampler_soa *sampler =
 622       swr_sampler_soa_create(key.sampler, PIPE_SHADER_GEOMETRY);
 623
 624    struct lp_bld_tgsi_system_values system_values;
 625    memset(&system_values, 0, sizeof(system_values));
 626    system_values.prim_id = wrap(LOAD(pGsCtx, {0, SWR_GS_CONTEXT_PrimitiveID}));
 627    system_values.instance_id = wrap(LOAD(pGsCtx, {0, SWR_GS_CONTEXT_InstanceID}));
 628
 629    std::vector<Constant*> mapConstants;
 630    Value *vtxAttribMap = ALLOCA(ArrayType::get(mInt32Ty, PIPE_MAX_SHADER_INPUTS));
 631    for (unsigned slot = 0; slot < info->num_inputs; slot++) {
 632       ubyte semantic_name = info->input_semantic_name[slot];
 633       ubyte semantic_idx = info->input_semantic_index[slot];
 634
 635       unsigned vs_slot = locate_linkage(semantic_name, semantic_idx, &ctx->vs->info.base);
 636
 637       vs_slot += VERTEX_ATTRIB_START_SLOT;
 638
 639       if (ctx->vs->info.base.output_semantic_name[0] == TGSI_SEMANTIC_POSITION)
 640          vs_slot--;
 641
 642       if (semantic_name == TGSI_SEMANTIC_POSITION)
 643          vs_slot = VERTEX_POSITION_SLOT;
 644
 645       STORE(C(vs_slot), vtxAttribMap, {0, slot});
 646       mapConstants.push_back(C(vs_slot));
 647    }
 648
 649    struct lp_build_mask_context mask;
 650    Value *mask_val = LOAD(pGsCtx, {0, SWR_GS_CONTEXT_mask}, "gsMask");
 651    lp_build_mask_begin(&mask, gallivm,
 652                        lp_type_float_vec(32, 32 * 8), wrap(mask_val));
 653
 654    // zero out cut buffer so we can load/modify/store bits
 655    for (uint32_t lane = 0; lane < mVWidth; ++lane)
 656    {
 657       Value* pStream = LOAD(pGsCtx, {0, SWR_GS_CONTEXT_pStreams, lane});
 658       MEMSET(pStream, C((char)0), VERTEX_COUNT_SIZE + CONTROL_HEADER_SIZE, sizeof(float) * KNOB_SIMD_WIDTH);
 659    }
 660
 661    struct swr_gs_llvm_iface gs_iface;
 662    gs_iface.base.fetch_input = ::swr_gs_llvm_fetch_input;
 663    gs_iface.base.emit_vertex = ::swr_gs_llvm_emit_vertex;
 664    gs_iface.base.end_primitive = ::swr_gs_llvm_end_primitive;
 665    gs_iface.base.gs_epilogue = ::swr_gs_llvm_epilogue;
 666    gs_iface.pBuilder = this;
 667    gs_iface.pGsCtx = pGsCtx;
 668    gs_iface.pGsState = pGS;
 669    gs_iface.num_outputs = gs->info.base.num_outputs;
 670    gs_iface.num_verts_per_prim =
 671       u_vertices_per_prim((pipe_prim_type)info->properties[TGSI_PROPERTY_GS_OUTPUT_PRIM]);
 672    gs_iface.info = info;
 673    gs_iface.pVtxAttribMap = vtxAttribMap;
 674
 675    lp_build_tgsi_soa(gallivm,
 676                      gs->pipe.tokens,
 677                      lp_type_float_vec(32, 32 * 8),
 678                      &mask,
 679                      wrap(consts_ptr),
 680                      wrap(const_sizes_ptr),
 681                      &system_values,
 682                      inputs,
 683                      outputs,
 684                      wrap(hPrivateData), // (sampler context)
 685                      NULL, // thread data
 686                      sampler,
 687                      &gs->info.base,
 688                      &gs_iface.base);
 689
 690    lp_build_mask_end(&mask);
 691
 692    sampler->destroy(sampler);
 693
 694    IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 695
 696    RET_VOID();
 697
 698    gallivm_verify_function(gallivm, wrap(pFunction));
 699    gallivm_compile_module(gallivm);
 700
 701    PFN_GS_FUNC pFunc =
 702       (PFN_GS_FUNC)gallivm_jit_function(gallivm, wrap(pFunction));
 703
 704    debug_printf("geom shader  %p\n", pFunc);
 705    assert(pFunc && "Error: GeomShader = NULL");
 706
 707    JM()->mIsModuleFinalized = true;
 708
 709    return pFunc;
 710 }
 711
 712 PFN_GS_FUNC
 713 swr_compile_gs(struct swr_context *ctx, swr_jit_gs_key &key)
 714 {
 715    BuilderSWR builder(
 716       reinterpret_cast<JitManager *>(swr_screen(ctx->pipe.screen)->hJitMgr),
 717       "GS");
 718    PFN_GS_FUNC func = builder.CompileGS(ctx, key);
 719
 720    ctx->gs->map.insert(std::make_pair(key, make_unique<VariantGS>(builder.gallivm, func)));
 721    return func;
 722 }
 723
 724 void
 725 BuilderSWR::WriteVS(Value *pVal, Value *pVsContext, Value *pVtxOutput, unsigned slot, unsigned channel)
 726 {
 727 #if USE_SIMD16_FRONTEND && !USE_SIMD16_VS
 728    // interleave the simdvertex components into the dest simd16vertex
 729    //   slot16offset = slot8offset * 2
 730    //   comp16offset = comp8offset * 2 + alternateOffset
 731
 732    Value *offset = LOAD(pVsContext, { 0, SWR_VS_CONTEXT_AlternateOffset });
 733    Value *pOut = GEP(pVtxOutput, { C(0), C(0), C(slot * 2), offset } );
 734    STORE(pVal, pOut, {channel * 2});
 735 #else
 736    Value *pOut = GEP(pVtxOutput, {0, 0, slot});
 737    STORE(pVal, pOut, {0, channel});
 738 #endif
 739 }
 740
 741 PFN_VERTEX_FUNC
 742 BuilderSWR::CompileVS(struct swr_context *ctx, swr_jit_vs_key &key)
 743 {
 744    struct swr_vertex_shader *swr_vs = ctx->vs;
 745
 746    LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
 747    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
 748
 749    memset(outputs, 0, sizeof(outputs));
 750
 751    AttrBuilder attrBuilder;
 752    attrBuilder.addStackAlignmentAttr(JM()->mVWidth * sizeof(float));
 753
 754    std::vector<Type *> vsArgs{PointerType::get(Gen_swr_draw_context(JM()), 0),
 755                               PointerType::get(Gen_SWR_VS_CONTEXT(JM()), 0)};
 756    FunctionType *vsFuncType =
 757       FunctionType::get(Type::getVoidTy(JM()->mContext), vsArgs, false);
 758
 759    // create new vertex shader function
 760    auto pFunction = Function::Create(vsFuncType,
 761                                      GlobalValue::ExternalLinkage,
 762                                      "VS",
 763                                      JM()->mpCurrentModule);
 764 #if HAVE_LLVM < 0x0500
 765    AttributeSet attrSet = AttributeSet::get(
 766       JM()->mContext, AttributeSet::FunctionIndex, attrBuilder);
 767    pFunction->addAttributes(AttributeSet::FunctionIndex, attrSet);
 768 #else
 769    pFunction->addAttributes(AttributeList::FunctionIndex, attrBuilder);
 770 #endif
 771
 772    BasicBlock *block = BasicBlock::Create(JM()->mContext, "entry", pFunction);
 773    IRB()->SetInsertPoint(block);
 774    LLVMPositionBuilderAtEnd(gallivm->builder, wrap(block));
 775
 776    auto argitr = pFunction->arg_begin();
 777    Value *hPrivateData = &*argitr++;
 778    hPrivateData->setName("hPrivateData");
 779    Value *pVsCtx = &*argitr++;
 780    pVsCtx->setName("vsCtx");
 781
 782    Value *consts_ptr = GEP(hPrivateData, {C(0), C(swr_draw_context_constantVS)});
 783
 784    consts_ptr->setName("vs_constants");
 785    Value *const_sizes_ptr =
 786       GEP(hPrivateData, {0, swr_draw_context_num_constantsVS});
 787    const_sizes_ptr->setName("num_vs_constants");
 788
 789    Value *vtxInput = LOAD(pVsCtx, {0, SWR_VS_CONTEXT_pVin});
 790 #if USE_SIMD16_VS
 791    vtxInput = BITCAST(vtxInput, PointerType::get(Gen_simd16vertex(JM()), 0));
 792 #endif
 793
 794    for (uint32_t attrib = 0; attrib < PIPE_MAX_SHADER_INPUTS; attrib++) {
 795       const unsigned mask = swr_vs->info.base.input_usage_mask[attrib];
 796       for (uint32_t channel = 0; channel < TGSI_NUM_CHANNELS; channel++) {
 797          if (mask & (1 << channel)) {
 798             inputs[attrib][channel] =
 799                wrap(LOAD(vtxInput, {0, 0, attrib, channel}));
 800          }
 801       }
 802    }
 803
 804    struct lp_build_sampler_soa *sampler =
 805       swr_sampler_soa_create(key.sampler, PIPE_SHADER_VERTEX);
 806
 807    struct lp_bld_tgsi_system_values system_values;
 808    memset(&system_values, 0, sizeof(system_values));
 809    system_values.instance_id = wrap(LOAD(pVsCtx, {0, SWR_VS_CONTEXT_InstanceID}));
 810
 811 #if USE_SIMD16_VS
 812    system_values.vertex_id = wrap(LOAD(pVsCtx, {0, SWR_VS_CONTEXT_VertexID16}));
 813 #else
 814    system_values.vertex_id = wrap(LOAD(pVsCtx, {0, SWR_VS_CONTEXT_VertexID}));
 815 #endif
 816
 817 #if USE_SIMD16_VS
 818    uint32_t vectorWidth = mVWidth16;
 819 #else
 820    uint32_t vectorWidth = mVWidth;
 821 #endif
 822
 823    lp_build_tgsi_soa(gallivm,
 824                      swr_vs->pipe.tokens,
 825                      lp_type_float_vec(32, 32 * vectorWidth),
 826                      NULL, // mask
 827                      wrap(consts_ptr),
 828                      wrap(const_sizes_ptr),
 829                      &system_values,
 830                      inputs,
 831                      outputs,
 832                      wrap(hPrivateData), // (sampler context)
 833                      NULL, // thread data
 834                      sampler, // sampler
 835                      &swr_vs->info.base,
 836                      NULL); // geometry shader face
 837
 838    sampler->destroy(sampler);
 839
 840    IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 841
 842    Value *vtxOutput = LOAD(pVsCtx, {0, SWR_VS_CONTEXT_pVout});
 843 #if USE_SIMD16_VS
 844    vtxOutput = BITCAST(vtxOutput, PointerType::get(Gen_simd16vertex(JM()), 0));
 845 #endif
 846
 847    for (uint32_t channel = 0; channel < TGSI_NUM_CHANNELS; channel++) {
 848       for (uint32_t attrib = 0; attrib < PIPE_MAX_SHADER_OUTPUTS; attrib++) {
 849          if (!outputs[attrib][channel])
 850             continue;
 851
 852          Value *val;
 853          uint32_t outSlot;
 854
 855          if (swr_vs->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_PSIZE) {
 856             if (channel != VERTEX_SGV_POINT_SIZE_COMP)
 857                continue;
 858             val = LOAD(unwrap(outputs[attrib][0]));
 859             outSlot = VERTEX_SGV_SLOT;
 860          } else if (swr_vs->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_POSITION) {
 861             val = LOAD(unwrap(outputs[attrib][channel]));
 862             outSlot = VERTEX_POSITION_SLOT;
 863          } else {
 864             val = LOAD(unwrap(outputs[attrib][channel]));
 865             outSlot = VERTEX_ATTRIB_START_SLOT + attrib;
 866             if (swr_vs->info.base.output_semantic_name[0] == TGSI_SEMANTIC_POSITION)
 867                outSlot--;
 868          }
 869
 870          WriteVS(val, pVsCtx, vtxOutput, outSlot, channel);
 871       }
 872    }
 873
 874    if (ctx->rasterizer->clip_plane_enable ||
 875        swr_vs->info.base.culldist_writemask) {
 876       unsigned clip_mask = ctx->rasterizer->clip_plane_enable;
 877
 878       unsigned cv = 0;
 879       if (swr_vs->info.base.writes_clipvertex) {
 880          cv = locate_linkage(TGSI_SEMANTIC_CLIPVERTEX, 0,
 881                              &swr_vs->info.base);
 882       } else {
 883          for (int i = 0; i < PIPE_MAX_SHADER_OUTPUTS; i++) {
 884             if (swr_vs->info.base.output_semantic_name[i] == TGSI_SEMANTIC_POSITION &&
 885                 swr_vs->info.base.output_semantic_index[i] == 0) {
 886                cv = i;
 887                break;
 888             }
 889          }
 890       }
 891       LLVMValueRef cx = LLVMBuildLoad(gallivm->builder, outputs[cv][0], "");
 892       LLVMValueRef cy = LLVMBuildLoad(gallivm->builder, outputs[cv][1], "");
 893       LLVMValueRef cz = LLVMBuildLoad(gallivm->builder, outputs[cv][2], "");
 894       LLVMValueRef cw = LLVMBuildLoad(gallivm->builder, outputs[cv][3], "");
 895
 896       for (unsigned val = 0; val < PIPE_MAX_CLIP_PLANES; val++) {
 897          // clip distance overrides user clip planes
 898          if ((swr_vs->info.base.clipdist_writemask & clip_mask & (1 << val)) ||
 899              ((swr_vs->info.base.culldist_writemask << swr_vs->info.base.num_written_clipdistance) & (1 << val))) {
 900             unsigned cv = locate_linkage(TGSI_SEMANTIC_CLIPDIST, val < 4 ? 0 : 1,
 901                                          &swr_vs->info.base);
 902             if (val < 4) {
 903                LLVMValueRef dist = LLVMBuildLoad(gallivm->builder, outputs[cv][val], "");
 904                WriteVS(unwrap(dist), pVsCtx, vtxOutput, VERTEX_CLIPCULL_DIST_LO_SLOT, val);
 905             } else {
 906                LLVMValueRef dist = LLVMBuildLoad(gallivm->builder, outputs[cv][val - 4], "");
 907                WriteVS(unwrap(dist), pVsCtx, vtxOutput, VERTEX_CLIPCULL_DIST_HI_SLOT, val - 4);
 908             }
 909             continue;
 910          }
 911
 912          if (!(clip_mask & (1 << val)))
 913             continue;
 914
 915          Value *px = LOAD(GEP(hPrivateData, {0, swr_draw_context_userClipPlanes, val, 0}));
 916          Value *py = LOAD(GEP(hPrivateData, {0, swr_draw_context_userClipPlanes, val, 1}));
 917          Value *pz = LOAD(GEP(hPrivateData, {0, swr_draw_context_userClipPlanes, val, 2}));
 918          Value *pw = LOAD(GEP(hPrivateData, {0, swr_draw_context_userClipPlanes, val, 3}));
 919 #if USE_SIMD16_VS
 920          Value *bpx = VBROADCAST_16(px);
 921          Value *bpy = VBROADCAST_16(py);
 922          Value *bpz = VBROADCAST_16(pz);
 923          Value *bpw = VBROADCAST_16(pw);
 924 #else
 925          Value *bpx = VBROADCAST(px);
 926          Value *bpy = VBROADCAST(py);
 927          Value *bpz = VBROADCAST(pz);
 928          Value *bpw = VBROADCAST(pw);
 929 #endif
 930          Value *dist = FADD(FMUL(unwrap(cx), bpx),
 931                             FADD(FMUL(unwrap(cy), bpy),
 932                                  FADD(FMUL(unwrap(cz), bpz),
 933                                       FMUL(unwrap(cw), bpw))));
 934
 935          if (val < 4)
 936             WriteVS(dist, pVsCtx, vtxOutput, VERTEX_CLIPCULL_DIST_LO_SLOT, val);
 937          else
 938             WriteVS(dist, pVsCtx, vtxOutput, VERTEX_CLIPCULL_DIST_HI_SLOT, val - 4);
 939       }
 940    }
 941
 942    RET_VOID();
 943
 944    gallivm_verify_function(gallivm, wrap(pFunction));
 945    gallivm_compile_module(gallivm);
 946
 947    //   lp_debug_dump_value(func);
 948
 949    PFN_VERTEX_FUNC pFunc =
 950       (PFN_VERTEX_FUNC)gallivm_jit_function(gallivm, wrap(pFunction));
 951
 952    debug_printf("vert shader  %p\n", pFunc);
 953    assert(pFunc && "Error: VertShader = NULL");
 954
 955    JM()->mIsModuleFinalized = true;
 956
 957    return pFunc;
 958 }
 959
 960 PFN_VERTEX_FUNC
 961 swr_compile_vs(struct swr_context *ctx, swr_jit_vs_key &key)
 962 {
 963    if (!ctx->vs->pipe.tokens)
 964       return NULL;
 965
 966    BuilderSWR builder(
 967       reinterpret_cast<JitManager *>(swr_screen(ctx->pipe.screen)->hJitMgr),
 968       "VS");
 969    PFN_VERTEX_FUNC func = builder.CompileVS(ctx, key);
 970
 971    ctx->vs->map.insert(std::make_pair(key, make_unique<VariantVS>(builder.gallivm, func)));
 972    return func;
 973 }
 974
 975 unsigned
 976 swr_so_adjust_attrib(unsigned in_attrib,
 977                      swr_vertex_shader *swr_vs)
 978 {
 979    ubyte semantic_name;
 980    unsigned attrib;
 981
 982    attrib = in_attrib + VERTEX_ATTRIB_START_SLOT;
 983
 984    if (swr_vs) {
 985       semantic_name = swr_vs->info.base.output_semantic_name[in_attrib];
 986       if (semantic_name == TGSI_SEMANTIC_POSITION) {
 987          attrib = VERTEX_POSITION_SLOT;
 988       } else if (semantic_name == TGSI_SEMANTIC_PSIZE) {
 989          attrib = VERTEX_SGV_SLOT;
 990       } else if (semantic_name == TGSI_SEMANTIC_LAYER) {
 991          attrib = VERTEX_SGV_SLOT;
 992       } else {
 993          if (swr_vs->info.base.writes_position) {
 994                attrib--;
 995          }
 996       }
 997    }
 998
 999    return attrib;
1000 }
1001
1002 static unsigned
1003 locate_linkage(ubyte name, ubyte index, struct tgsi_shader_info *info)
1004 {
1005    for (int i = 0; i < PIPE_MAX_SHADER_OUTPUTS; i++) {
1006       if ((info->output_semantic_name[i] == name)
1007           && (info->output_semantic_index[i] == index)) {
1008          return i;
1009       }
1010    }
1011
1012    return 0xFFFFFFFF;
1013 }
1014
1015 PFN_PIXEL_KERNEL
1016 BuilderSWR::CompileFS(struct swr_context *ctx, swr_jit_fs_key &key)
1017 {
1018    struct swr_fragment_shader *swr_fs = ctx->fs;
1019
1020    struct tgsi_shader_info *pPrevShader;
1021    if (ctx->gs)
1022       pPrevShader = &ctx->gs->info.base;
1023    else
1024       pPrevShader = &ctx->vs->info.base;
1025
1026    LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
1027    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
1028
1029    memset(inputs, 0, sizeof(inputs));
1030    memset(outputs, 0, sizeof(outputs));
1031
1032    struct lp_build_sampler_soa *sampler = NULL;
1033
1034    AttrBuilder attrBuilder;
1035    attrBuilder.addStackAlignmentAttr(JM()->mVWidth * sizeof(float));
1036
1037    std::vector<Type *> fsArgs{PointerType::get(Gen_swr_draw_context(JM()), 0),
1038                               PointerType::get(Gen_SWR_PS_CONTEXT(JM()), 0)};
1039    FunctionType *funcType =
1040       FunctionType::get(Type::getVoidTy(JM()->mContext), fsArgs, false);
1041
1042    auto pFunction = Function::Create(funcType,
1043                                      GlobalValue::ExternalLinkage,
1044                                      "FS",
1045                                      JM()->mpCurrentModule);
1046 #if HAVE_LLVM < 0x0500
1047    AttributeSet attrSet = AttributeSet::get(
1048       JM()->mContext, AttributeSet::FunctionIndex, attrBuilder);
1049    pFunction->addAttributes(AttributeSet::FunctionIndex, attrSet);
1050 #else
1051    pFunction->addAttributes(AttributeList::FunctionIndex, attrBuilder);
1052 #endif
1053
1054    BasicBlock *block = BasicBlock::Create(JM()->mContext, "entry", pFunction);
1055    IRB()->SetInsertPoint(block);
1056    LLVMPositionBuilderAtEnd(gallivm->builder, wrap(block));
1057
1058    auto args = pFunction->arg_begin();
1059    Value *hPrivateData = &*args++;
1060    hPrivateData->setName("hPrivateData");
1061    Value *pPS = &*args++;
1062    pPS->setName("psCtx");
1063
1064    Value *consts_ptr = GEP(hPrivateData, {0, swr_draw_context_constantFS});
1065    consts_ptr->setName("fs_constants");
1066    Value *const_sizes_ptr =
1067       GEP(hPrivateData, {0, swr_draw_context_num_constantsFS});
1068    const_sizes_ptr->setName("num_fs_constants");
1069
1070    // load *pAttribs, *pPerspAttribs
1071    Value *pRawAttribs = LOAD(pPS, {0, SWR_PS_CONTEXT_pAttribs}, "pRawAttribs");
1072    Value *pPerspAttribs =
1073       LOAD(pPS, {0, SWR_PS_CONTEXT_pPerspAttribs}, "pPerspAttribs");
1074
1075    swr_fs->constantMask = 0;
1076    swr_fs->flatConstantMask = 0;
1077    swr_fs->pointSpriteMask = 0;
1078
1079    for (int attrib = 0; attrib < PIPE_MAX_SHADER_INPUTS; attrib++) {
1080       const unsigned mask = swr_fs->info.base.input_usage_mask[attrib];
1081       const unsigned interpMode = swr_fs->info.base.input_interpolate[attrib];
1082       const unsigned interpLoc = swr_fs->info.base.input_interpolate_loc[attrib];
1083
1084       if (!mask)
1085          continue;
1086
1087       // load i,j
1088       Value *vi = nullptr, *vj = nullptr;
1089       switch (interpLoc) {
1090       case TGSI_INTERPOLATE_LOC_CENTER:
1091          vi = LOAD(pPS, {0, SWR_PS_CONTEXT_vI, PixelPositions_center}, "i");
1092          vj = LOAD(pPS, {0, SWR_PS_CONTEXT_vJ, PixelPositions_center}, "j");
1093          break;
1094       case TGSI_INTERPOLATE_LOC_CENTROID:
1095          vi = LOAD(pPS, {0, SWR_PS_CONTEXT_vI, PixelPositions_centroid}, "i");
1096          vj = LOAD(pPS, {0, SWR_PS_CONTEXT_vJ, PixelPositions_centroid}, "j");
1097          break;
1098       case TGSI_INTERPOLATE_LOC_SAMPLE:
1099          vi = LOAD(pPS, {0, SWR_PS_CONTEXT_vI, PixelPositions_sample}, "i");
1100          vj = LOAD(pPS, {0, SWR_PS_CONTEXT_vJ, PixelPositions_sample}, "j");
1101          break;
1102       }
1103
1104       // load/compute w
1105       Value *vw = nullptr, *pAttribs;
1106       if (interpMode == TGSI_INTERPOLATE_PERSPECTIVE ||
1107           interpMode == TGSI_INTERPOLATE_COLOR) {
1108          pAttribs = pPerspAttribs;
1109          switch (interpLoc) {
1110          case TGSI_INTERPOLATE_LOC_CENTER:
1111             vw = VRCP(LOAD(pPS, {0, SWR_PS_CONTEXT_vOneOverW, PixelPositions_center}));
1112             break;
1113          case TGSI_INTERPOLATE_LOC_CENTROID:
1114             vw = VRCP(LOAD(pPS, {0, SWR_PS_CONTEXT_vOneOverW, PixelPositions_centroid}));
1115             break;
1116          case TGSI_INTERPOLATE_LOC_SAMPLE:
1117             vw = VRCP(LOAD(pPS, {0, SWR_PS_CONTEXT_vOneOverW, PixelPositions_sample}));
1118             break;
1119          }
1120       } else {
1121          pAttribs = pRawAttribs;
1122          vw = VIMMED1(1.f);
1123       }
1124
1125       vw->setName("w");
1126
1127       ubyte semantic_name = swr_fs->info.base.input_semantic_name[attrib];
1128       ubyte semantic_idx = swr_fs->info.base.input_semantic_index[attrib];
1129
1130       if (semantic_name == TGSI_SEMANTIC_FACE) {
1131          Value *ff =
1132             UI_TO_FP(LOAD(pPS, {0, SWR_PS_CONTEXT_frontFace}), mFP32Ty);
1133          ff = FSUB(FMUL(ff, C(2.0f)), C(1.0f));
1134          ff = VECTOR_SPLAT(JM()->mVWidth, ff, "vFrontFace");
1135
1136          inputs[attrib][0] = wrap(ff);
1137          inputs[attrib][1] = wrap(VIMMED1(0.0f));
1138          inputs[attrib][2] = wrap(VIMMED1(0.0f));
1139          inputs[attrib][3] = wrap(VIMMED1(1.0f));
1140          continue;
1141       } else if (semantic_name == TGSI_SEMANTIC_POSITION) { // gl_FragCoord
1142          if (swr_fs->info.base.properties[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER] ==
1143              TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER) {
1144             inputs[attrib][0] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vX, PixelPositions_center}, "vX"));
1145             inputs[attrib][1] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vY, PixelPositions_center}, "vY"));
1146          } else {
1147             inputs[attrib][0] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vX, PixelPositions_UL}, "vX"));
1148             inputs[attrib][1] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vY, PixelPositions_UL}, "vY"));
1149          }
1150          inputs[attrib][2] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vZ}, "vZ"));
1151          inputs[attrib][3] =
1152             wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vOneOverW, PixelPositions_center}, "vOneOverW"));
1153          continue;
1154       }
1155
1156       unsigned linkedAttrib =
1157          locate_linkage(semantic_name, semantic_idx, pPrevShader) - 1;
1158
1159       uint32_t extraAttribs = 0;
1160       if (semantic_name == TGSI_SEMANTIC_PRIMID && !ctx->gs) {
1161          /* non-gs generated primID - need to grab from swizzleMap override */
1162          linkedAttrib = pPrevShader->num_outputs - 1;
1163          swr_fs->constantMask |= 1 << linkedAttrib;
1164          extraAttribs++;
1165       } else if (semantic_name == TGSI_SEMANTIC_GENERIC &&
1166           key.sprite_coord_enable & (1 << semantic_idx)) {
1167          /* we add an extra attrib to the backendState in swr_update_derived. */
1168          linkedAttrib = pPrevShader->num_outputs + extraAttribs - 1;
1169          swr_fs->pointSpriteMask |= (1 << linkedAttrib);
1170          extraAttribs++;
1171       } else if (linkedAttrib == 0xFFFFFFFF) {
1172          inputs[attrib][0] = wrap(VIMMED1(0.0f));
1173          inputs[attrib][1] = wrap(VIMMED1(0.0f));
1174          inputs[attrib][2] = wrap(VIMMED1(0.0f));
1175          inputs[attrib][3] = wrap(VIMMED1(1.0f));
1176          /* If we're reading in color and 2-sided lighting is enabled, we have
1177           * to keep going.
1178           */
1179          if (semantic_name != TGSI_SEMANTIC_COLOR || !key.light_twoside)
1180             continue;
1181       } else {
1182          if (interpMode == TGSI_INTERPOLATE_CONSTANT) {
1183             swr_fs->constantMask |= 1 << linkedAttrib;
1184          } else if (interpMode == TGSI_INTERPOLATE_COLOR) {
1185             swr_fs->flatConstantMask |= 1 << linkedAttrib;
1186          }
1187       }
1188
1189       unsigned bcolorAttrib = 0xFFFFFFFF;
1190       Value *offset = NULL;
1191       if (semantic_name == TGSI_SEMANTIC_COLOR && key.light_twoside) {
1192          bcolorAttrib = locate_linkage(
1193                TGSI_SEMANTIC_BCOLOR, semantic_idx, pPrevShader) - 1;
1194          /* Neither front nor back colors were available. Nothing to load. */
1195          if (bcolorAttrib == 0xFFFFFFFF && linkedAttrib == 0xFFFFFFFF)
1196             continue;
1197          /* If there is no front color, just always use the back color. */
1198          if (linkedAttrib == 0xFFFFFFFF)
1199             linkedAttrib = bcolorAttrib;
1200
1201          if (bcolorAttrib != 0xFFFFFFFF) {
1202             if (interpMode == TGSI_INTERPOLATE_CONSTANT) {
1203                swr_fs->constantMask |= 1 << bcolorAttrib;
1204             } else if (interpMode == TGSI_INTERPOLATE_COLOR) {
1205                swr_fs->flatConstantMask |= 1 << bcolorAttrib;
1206             }
1207
1208             unsigned diff = 12 * (bcolorAttrib - linkedAttrib);
1209
1210             if (diff) {
1211                Value *back =
1212                   XOR(C(1), LOAD(pPS, {0, SWR_PS_CONTEXT_frontFace}), "backFace");
1213
1214                offset = MUL(back, C(diff));
1215                offset->setName("offset");
1216             }
1217          }
1218       }
1219
1220       for (int channel = 0; channel < TGSI_NUM_CHANNELS; channel++) {
1221          if (mask & (1 << channel)) {
1222             Value *indexA = C(linkedAttrib * 12 + channel);
1223             Value *indexB = C(linkedAttrib * 12 + channel + 4);
1224             Value *indexC = C(linkedAttrib * 12 + channel + 8);
1225
1226             if (offset) {
1227                indexA = ADD(indexA, offset);
1228                indexB = ADD(indexB, offset);
1229                indexC = ADD(indexC, offset);
1230             }
1231
1232             Value *va = VBROADCAST(LOAD(GEP(pAttribs, indexA)));
1233             Value *vb = VBROADCAST(LOAD(GEP(pAttribs, indexB)));
1234             Value *vc = VBROADCAST(LOAD(GEP(pAttribs, indexC)));
1235
1236             if (interpMode == TGSI_INTERPOLATE_CONSTANT) {
1237                inputs[attrib][channel] = wrap(va);
1238             } else {
1239                Value *vk = FSUB(FSUB(VIMMED1(1.0f), vi), vj);
1240
1241                vc = FMUL(vk, vc);
1242
1243                Value *interp = FMUL(va, vi);
1244                Value *interp1 = FMUL(vb, vj);
1245                interp = FADD(interp, interp1);
1246                interp = FADD(interp, vc);
1247                if (interpMode == TGSI_INTERPOLATE_PERSPECTIVE ||
1248                    interpMode == TGSI_INTERPOLATE_COLOR)
1249                   interp = FMUL(interp, vw);
1250                inputs[attrib][channel] = wrap(interp);
1251             }
1252          }
1253       }
1254    }
1255
1256    sampler = swr_sampler_soa_create(key.sampler, PIPE_SHADER_FRAGMENT);
1257
1258    struct lp_bld_tgsi_system_values system_values;
1259    memset(&system_values, 0, sizeof(system_values));
1260
1261    struct lp_build_mask_context mask;
1262    bool uses_mask = false;
1263
1264    if (swr_fs->info.base.uses_kill ||
1265        key.poly_stipple_enable) {
1266       Value *vActiveMask = NULL;
1267       if (swr_fs->info.base.uses_kill) {
1268          vActiveMask = LOAD(pPS, {0, SWR_PS_CONTEXT_activeMask}, "activeMask");
1269       }
1270       if (key.poly_stipple_enable) {
1271          // first get fragment xy coords and clip to stipple bounds
1272          Value *vXf = LOAD(pPS, {0, SWR_PS_CONTEXT_vX, PixelPositions_UL});
1273          Value *vYf = LOAD(pPS, {0, SWR_PS_CONTEXT_vY, PixelPositions_UL});
1274          Value *vXu = FP_TO_UI(vXf, mSimdInt32Ty);
1275          Value *vYu = FP_TO_UI(vYf, mSimdInt32Ty);
1276
1277          // stipple pattern is 32x32, which means that one line of stipple
1278          // is stored in one word:
1279          // vXstipple is bit offset inside 32-bit stipple word
1280          // vYstipple is word index is stipple array
1281          Value *vXstipple = AND(vXu, VIMMED1(0x1f)); // & (32-1)
1282          Value *vYstipple = AND(vYu, VIMMED1(0x1f)); // & (32-1)
1283
1284          // grab stipple pattern base address
1285          Value *stipplePtr = GEP(hPrivateData, {0, swr_draw_context_polyStipple, 0});
1286          stipplePtr = BITCAST(stipplePtr, mInt8PtrTy);
1287
1288          // peform a gather to grab stipple words for each lane
1289          Value *vStipple = GATHERDD(VUNDEF_I(), stipplePtr, vYstipple,
1290                                     VIMMED1(0xffffffff), 4);
1291
1292          // create a mask with one bit corresponding to the x stipple
1293          // and AND it with the pattern, to see if we have a bit
1294          Value *vBitMask = LSHR(VIMMED1(0x80000000), vXstipple);
1295          Value *vStippleMask = AND(vStipple, vBitMask);
1296          vStippleMask = ICMP_NE(vStippleMask, VIMMED1(0));
1297          vStippleMask = VMASK(vStippleMask);
1298
1299          if (swr_fs->info.base.uses_kill) {
1300             vActiveMask = AND(vActiveMask, vStippleMask);
1301          } else {
1302             vActiveMask = vStippleMask;
1303          }
1304       }
1305       lp_build_mask_begin(
1306          &mask, gallivm, lp_type_float_vec(32, 32 * 8), wrap(vActiveMask));
1307       uses_mask = true;
1308    }
1309
1310    lp_build_tgsi_soa(gallivm,
1311                      swr_fs->pipe.tokens,
1312                      lp_type_float_vec(32, 32 * 8),
1313                      uses_mask ? &mask : NULL, // mask
1314                      wrap(consts_ptr),
1315                      wrap(const_sizes_ptr),
1316                      &system_values,
1317                      inputs,
1318                      outputs,
1319                      wrap(hPrivateData),
1320                      NULL, // thread data
1321                      sampler, // sampler
1322                      &swr_fs->info.base,
1323                      NULL); // geometry shader face
1324
1325    sampler->destroy(sampler);
1326
1327    IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
1328
1329    for (uint32_t attrib = 0; attrib < swr_fs->info.base.num_outputs;
1330         attrib++) {
1331       switch (swr_fs->info.base.output_semantic_name[attrib]) {
1332       case TGSI_SEMANTIC_POSITION: {
1333          // write z
1334          LLVMValueRef outZ =
1335             LLVMBuildLoad(gallivm->builder, outputs[attrib][2], "");
1336          STORE(unwrap(outZ), pPS, {0, SWR_PS_CONTEXT_vZ});
1337          break;
1338       }
1339       case TGSI_SEMANTIC_COLOR: {
1340          for (uint32_t channel = 0; channel < TGSI_NUM_CHANNELS; channel++) {
1341             if (!outputs[attrib][channel])
1342                continue;
1343
1344             LLVMValueRef out =
1345                LLVMBuildLoad(gallivm->builder, outputs[attrib][channel], "");
1346             if (swr_fs->info.base.properties[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS] &&
1347                 swr_fs->info.base.output_semantic_index[attrib] == 0) {
1348                for (uint32_t rt = 0; rt < key.nr_cbufs; rt++) {
1349                   STORE(unwrap(out),
1350                         pPS,
1351                         {0, SWR_PS_CONTEXT_shaded, rt, channel});
1352                }
1353             } else {
1354                STORE(unwrap(out),
1355                      pPS,
1356                      {0,
1357                            SWR_PS_CONTEXT_shaded,
1358                            swr_fs->info.base.output_semantic_index[attrib],
1359                            channel});
1360             }
1361          }
1362          break;
1363       }
1364       default: {
1365          fprintf(stderr,
1366                  "unknown output from FS %s[%d]\n",
1367                  tgsi_semantic_names[swr_fs->info.base
1368                                         .output_semantic_name[attrib]],
1369                  swr_fs->info.base.output_semantic_index[attrib]);
1370          break;
1371       }
1372       }
1373    }
1374
1375    LLVMValueRef mask_result = 0;
1376    if (uses_mask) {
1377       mask_result = lp_build_mask_end(&mask);
1378    }
1379
1380    IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
1381
1382    if (uses_mask) {
1383       STORE(unwrap(mask_result), pPS, {0, SWR_PS_CONTEXT_activeMask});
1384    }
1385
1386    RET_VOID();
1387
1388    gallivm_verify_function(gallivm, wrap(pFunction));
1389
1390    gallivm_compile_module(gallivm);
1391
1392    PFN_PIXEL_KERNEL kernel =
1393       (PFN_PIXEL_KERNEL)gallivm_jit_function(gallivm, wrap(pFunction));
1394    debug_printf("frag shader  %p\n", kernel);
1395    assert(kernel && "Error: FragShader = NULL");
1396
1397    JM()->mIsModuleFinalized = true;
1398
1399    return kernel;
1400 }
1401
1402 PFN_PIXEL_KERNEL
1403 swr_compile_fs(struct swr_context *ctx, swr_jit_fs_key &key)
1404 {
1405    if (!ctx->fs->pipe.tokens)
1406       return NULL;
1407
1408    BuilderSWR builder(
1409       reinterpret_cast<JitManager *>(swr_screen(ctx->pipe.screen)->hJitMgr),
1410       "FS");
1411    PFN_PIXEL_KERNEL func = builder.CompileFS(ctx, key);
1412
1413    ctx->fs->map.insert(std::make_pair(key, make_unique<VariantFS>(builder.gallivm, func)));
1414    return func;
1415 }