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 #include "llvm/IR/LegacyPassManager.h"
  31 #pragma pop_macro("DEBUG")
  32
  33 #include "state.h"
  34 #include "gen_state_llvm.h"
  35 #include "builder.h"
  36 #include "functionpasses/passes.h"
  37
  38 #include "tgsi/tgsi_strings.h"
  39 #include "util/u_format.h"
  40 #include "util/u_prim.h"
  41 #include "gallivm/lp_bld_init.h"
  42 #include "gallivm/lp_bld_flow.h"
  43 #include "gallivm/lp_bld_struct.h"
  44 #include "gallivm/lp_bld_tgsi.h"
  45
  46 #include "swr_context.h"
  47 #include "gen_swr_context_llvm.h"
  48 #include "swr_resource.h"
  49 #include "swr_state.h"
  50 #include "swr_screen.h"
  51
  52 using namespace SwrJit;
  53 using namespace llvm;
  54
  55 static unsigned
  56 locate_linkage(ubyte name, ubyte index, struct tgsi_shader_info *info);
  57
  58 bool operator==(const swr_jit_fs_key &lhs, const swr_jit_fs_key &rhs)
  59 {
  60    return !memcmp(&lhs, &rhs, sizeof(lhs));
  61 }
  62
  63 bool operator==(const swr_jit_vs_key &lhs, const swr_jit_vs_key &rhs)
  64 {
  65    return !memcmp(&lhs, &rhs, sizeof(lhs));
  66 }
  67
  68 bool operator==(const swr_jit_fetch_key &lhs, const swr_jit_fetch_key &rhs)
  69 {
  70    return !memcmp(&lhs, &rhs, sizeof(lhs));
  71 }
  72
  73 bool operator==(const swr_jit_gs_key &lhs, const swr_jit_gs_key &rhs)
  74 {
  75    return !memcmp(&lhs, &rhs, sizeof(lhs));
  76 }
  77
  78 static void
  79 swr_generate_sampler_key(const struct lp_tgsi_info &info,
  80                          struct swr_context *ctx,
  81                          enum pipe_shader_type shader_type,
  82                          struct swr_jit_sampler_key &key)
  83 {
  84    key.nr_samplers = info.base.file_max[TGSI_FILE_SAMPLER] + 1;
  85
  86    for (unsigned i = 0; i < key.nr_samplers; i++) {
  87       if (info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
  88          lp_sampler_static_sampler_state(
  89             &key.sampler[i].sampler_state,
  90             ctx->samplers[shader_type][i]);
  91       }
  92    }
  93
  94    /*
  95     * XXX If TGSI_FILE_SAMPLER_VIEW exists assume all texture opcodes
  96     * are dx10-style? Can't really have mixed opcodes, at least not
  97     * if we want to skip the holes here (without rescanning tgsi).
  98     */
  99    if (info.base.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
 100       key.nr_sampler_views =
 101          info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
 102       for (unsigned i = 0; i < key.nr_sampler_views; i++) {
 103          if (info.base.file_mask[TGSI_FILE_SAMPLER_VIEW] & (1u << (i & 31))) {
 104             const struct pipe_sampler_view *view =
 105                ctx->sampler_views[shader_type][i];
 106             lp_sampler_static_texture_state(
 107                &key.sampler[i].texture_state, view);
 108             if (view) {
 109                struct swr_resource *swr_res = swr_resource(view->texture);
 110                const struct util_format_description *desc =
 111                   util_format_description(view->format);
 112                if (swr_res->has_depth && swr_res->has_stencil &&
 113                    !util_format_has_depth(desc))
 114                   key.sampler[i].texture_state.format = PIPE_FORMAT_S8_UINT;
 115             }
 116          }
 117       }
 118    } else {
 119       key.nr_sampler_views = key.nr_samplers;
 120       for (unsigned i = 0; i < key.nr_sampler_views; i++) {
 121          if (info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
 122             const struct pipe_sampler_view *view =
 123                ctx->sampler_views[shader_type][i];
 124             lp_sampler_static_texture_state(
 125                &key.sampler[i].texture_state, view);
 126             if (view) {
 127                struct swr_resource *swr_res = swr_resource(view->texture);
 128                const struct util_format_description *desc =
 129                   util_format_description(view->format);
 130                if (swr_res->has_depth && swr_res->has_stencil &&
 131                    !util_format_has_depth(desc))
 132                   key.sampler[i].texture_state.format = PIPE_FORMAT_S8_UINT;
 133             }
 134          }
 135       }
 136    }
 137 }
 138
 139 void
 140 swr_generate_fs_key(struct swr_jit_fs_key &key,
 141                     struct swr_context *ctx,
 142                     swr_fragment_shader *swr_fs)
 143 {
 144    memset(&key, 0, sizeof(key));
 145
 146    key.nr_cbufs = ctx->framebuffer.nr_cbufs;
 147    key.light_twoside = ctx->rasterizer->light_twoside;
 148    key.sprite_coord_enable = ctx->rasterizer->sprite_coord_enable;
 149
 150    struct tgsi_shader_info *pPrevShader;
 151    if (ctx->gs)
 152       pPrevShader = &ctx->gs->info.base;
 153    else
 154       pPrevShader = &ctx->vs->info.base;
 155
 156    memcpy(&key.vs_output_semantic_name,
 157           &pPrevShader->output_semantic_name,
 158           sizeof(key.vs_output_semantic_name));
 159    memcpy(&key.vs_output_semantic_idx,
 160           &pPrevShader->output_semantic_index,
 161           sizeof(key.vs_output_semantic_idx));
 162
 163    swr_generate_sampler_key(swr_fs->info, ctx, PIPE_SHADER_FRAGMENT, key);
 164
 165    key.poly_stipple_enable = ctx->rasterizer->poly_stipple_enable &&
 166       ctx->poly_stipple.prim_is_poly;
 167 }
 168
 169 void
 170 swr_generate_vs_key(struct swr_jit_vs_key &key,
 171                     struct swr_context *ctx,
 172                     swr_vertex_shader *swr_vs)
 173 {
 174    memset(&key, 0, sizeof(key));
 175
 176    key.clip_plane_mask =
 177       swr_vs->info.base.clipdist_writemask ?
 178       swr_vs->info.base.clipdist_writemask & ctx->rasterizer->clip_plane_enable :
 179       ctx->rasterizer->clip_plane_enable;
 180
 181    swr_generate_sampler_key(swr_vs->info, ctx, PIPE_SHADER_VERTEX, key);
 182 }
 183
 184 void
 185 swr_generate_fetch_key(struct swr_jit_fetch_key &key,
 186                        struct swr_vertex_element_state *velems)
 187 {
 188    memset(&key, 0, sizeof(key));
 189
 190    key.fsState = velems->fsState;
 191 }
 192
 193 void
 194 swr_generate_gs_key(struct swr_jit_gs_key &key,
 195                     struct swr_context *ctx,
 196                     swr_geometry_shader *swr_gs)
 197 {
 198    memset(&key, 0, sizeof(key));
 199
 200    struct tgsi_shader_info *pPrevShader = &ctx->vs->info.base;
 201
 202    memcpy(&key.vs_output_semantic_name,
 203           &pPrevShader->output_semantic_name,
 204           sizeof(key.vs_output_semantic_name));
 205    memcpy(&key.vs_output_semantic_idx,
 206           &pPrevShader->output_semantic_index,
 207           sizeof(key.vs_output_semantic_idx));
 208
 209    swr_generate_sampler_key(swr_gs->info, ctx, PIPE_SHADER_GEOMETRY, key);
 210 }
 211
 212 struct BuilderSWR : public Builder {
 213    BuilderSWR(JitManager *pJitMgr, const char *pName)
 214       : Builder(pJitMgr)
 215    {
 216       pJitMgr->SetupNewModule();
 217       gallivm = gallivm_create(pName, wrap(&JM()->mContext));
 218       pJitMgr->mpCurrentModule = unwrap(gallivm->module);
 219    }
 220
 221    ~BuilderSWR() {
 222       gallivm_free_ir(gallivm);
 223    }
 224
 225    void WriteVS(Value *pVal, Value *pVsContext, Value *pVtxOutput,
 226                 unsigned slot, unsigned channel);
 227
 228    struct gallivm_state *gallivm;
 229    PFN_VERTEX_FUNC CompileVS(struct swr_context *ctx, swr_jit_vs_key &key);
 230    PFN_PIXEL_KERNEL CompileFS(struct swr_context *ctx, swr_jit_fs_key &key);
 231    PFN_GS_FUNC CompileGS(struct swr_context *ctx, swr_jit_gs_key &key);
 232
 233    LLVMValueRef
 234    swr_gs_llvm_fetch_input(const struct lp_build_tgsi_gs_iface *gs_iface,
 235                            struct lp_build_tgsi_context * bld_base,
 236                            boolean is_vindex_indirect,
 237                            LLVMValueRef vertex_index,
 238                            boolean is_aindex_indirect,
 239                            LLVMValueRef attrib_index,
 240                            LLVMValueRef swizzle_index);
 241    void
 242    swr_gs_llvm_emit_vertex(const struct lp_build_tgsi_gs_iface *gs_base,
 243                            struct lp_build_tgsi_context * bld_base,
 244                            LLVMValueRef (*outputs)[4],
 245                            LLVMValueRef emitted_vertices_vec);
 246
 247    void
 248    swr_gs_llvm_end_primitive(const struct lp_build_tgsi_gs_iface *gs_base,
 249                              struct lp_build_tgsi_context * bld_base,
 250                              LLVMValueRef verts_per_prim_vec,
 251                              LLVMValueRef emitted_prims_vec);
 252
 253    void
 254    swr_gs_llvm_epilogue(const struct lp_build_tgsi_gs_iface *gs_base,
 255                         struct lp_build_tgsi_context * bld_base,
 256                         LLVMValueRef total_emitted_vertices_vec,
 257                         LLVMValueRef emitted_prims_vec);
 258
 259 };
 260
 261 struct swr_gs_llvm_iface {
 262    struct lp_build_tgsi_gs_iface base;
 263    struct tgsi_shader_info *info;
 264
 265    BuilderSWR *pBuilder;
 266
 267    Value *pGsCtx;
 268    SWR_GS_STATE *pGsState;
 269    uint32_t num_outputs;
 270    uint32_t num_verts_per_prim;
 271
 272    Value *pVtxAttribMap;
 273 };
 274
 275 // trampoline functions so we can use the builder llvm construction methods
 276 static LLVMValueRef
 277 swr_gs_llvm_fetch_input(const struct lp_build_tgsi_gs_iface *gs_iface,
 278                            struct lp_build_tgsi_context * bld_base,
 279                            boolean is_vindex_indirect,
 280                            LLVMValueRef vertex_index,
 281                            boolean is_aindex_indirect,
 282                            LLVMValueRef attrib_index,
 283                            LLVMValueRef swizzle_index)
 284 {
 285     swr_gs_llvm_iface *iface = (swr_gs_llvm_iface*)gs_iface;
 286
 287     return iface->pBuilder->swr_gs_llvm_fetch_input(gs_iface, bld_base,
 288                                                    is_vindex_indirect,
 289                                                    vertex_index,
 290                                                    is_aindex_indirect,
 291                                                    attrib_index,
 292                                                    swizzle_index);
 293 }
 294
 295 static void
 296 swr_gs_llvm_emit_vertex(const struct lp_build_tgsi_gs_iface *gs_base,
 297                            struct lp_build_tgsi_context * bld_base,
 298                            LLVMValueRef (*outputs)[4],
 299                            LLVMValueRef emitted_vertices_vec)
 300 {
 301     swr_gs_llvm_iface *iface = (swr_gs_llvm_iface*)gs_base;
 302
 303     iface->pBuilder->swr_gs_llvm_emit_vertex(gs_base, bld_base,
 304                                             outputs,
 305                                             emitted_vertices_vec);
 306 }
 307
 308 static void
 309 swr_gs_llvm_end_primitive(const struct lp_build_tgsi_gs_iface *gs_base,
 310                              struct lp_build_tgsi_context * bld_base,
 311                              LLVMValueRef verts_per_prim_vec,
 312                              LLVMValueRef emitted_prims_vec)
 313 {
 314     swr_gs_llvm_iface *iface = (swr_gs_llvm_iface*)gs_base;
 315
 316     iface->pBuilder->swr_gs_llvm_end_primitive(gs_base, bld_base,
 317                                               verts_per_prim_vec,
 318                                               emitted_prims_vec);
 319 }
 320
 321 static void
 322 swr_gs_llvm_epilogue(const struct lp_build_tgsi_gs_iface *gs_base,
 323                         struct lp_build_tgsi_context * bld_base,
 324                         LLVMValueRef total_emitted_vertices_vec,
 325                         LLVMValueRef emitted_prims_vec)
 326 {
 327     swr_gs_llvm_iface *iface = (swr_gs_llvm_iface*)gs_base;
 328
 329     iface->pBuilder->swr_gs_llvm_epilogue(gs_base, bld_base,
 330                                          total_emitted_vertices_vec,
 331                                          emitted_prims_vec);
 332 }
 333
 334 LLVMValueRef
 335 BuilderSWR::swr_gs_llvm_fetch_input(const struct lp_build_tgsi_gs_iface *gs_iface,
 336                            struct lp_build_tgsi_context * bld_base,
 337                            boolean is_vindex_indirect,
 338                            LLVMValueRef vertex_index,
 339                            boolean is_aindex_indirect,
 340                            LLVMValueRef attrib_index,
 341                            LLVMValueRef swizzle_index)
 342 {
 343     swr_gs_llvm_iface *iface = (swr_gs_llvm_iface*)gs_iface;
 344     Value *vert_index = unwrap(vertex_index);
 345     Value *attr_index = unwrap(attrib_index);
 346
 347     IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 348
 349     if (is_vindex_indirect || is_aindex_indirect) {
 350        int i;
 351        Value *res = unwrap(bld_base->base.zero);
 352        struct lp_type type = bld_base->base.type;
 353
 354        for (i = 0; i < type.length; i++) {
 355           Value *vert_chan_index = vert_index;
 356           Value *attr_chan_index = attr_index;
 357
 358           if (is_vindex_indirect) {
 359              vert_chan_index = VEXTRACT(vert_index, C(i));
 360           }
 361           if (is_aindex_indirect) {
 362              attr_chan_index = VEXTRACT(attr_index, C(i));
 363           }
 364
 365           Value *attrib =
 366              LOAD(GEP(iface->pVtxAttribMap, {C(0), attr_chan_index}));
 367
 368           Value *pVertex = LOAD(iface->pGsCtx, {0, SWR_GS_CONTEXT_pVerts});
 369           Value *pInputVertStride = LOAD(iface->pGsCtx, {0, SWR_GS_CONTEXT_inputVertStride});
 370
 371           Value *pVector = ADD(MUL(vert_chan_index, pInputVertStride), attrib);
 372           Value *pInput = LOAD(GEP(pVertex, {pVector, unwrap(swizzle_index)}));
 373
 374           Value *value = VEXTRACT(pInput, C(i));
 375           res = VINSERT(res, value, C(i));
 376        }
 377
 378        return wrap(res);
 379     } else {
 380        Value *attrib = LOAD(GEP(iface->pVtxAttribMap, {C(0), attr_index}));
 381
 382        Value *pVertex = LOAD(iface->pGsCtx, {0, SWR_GS_CONTEXT_pVerts});
 383        Value *pInputVertStride = LOAD(iface->pGsCtx, {0, SWR_GS_CONTEXT_inputVertStride});
 384
 385        Value *pVector = ADD(MUL(vert_index, pInputVertStride), attrib);
 386
 387        Value *pInput = LOAD(GEP(pVertex, {pVector, unwrap(swizzle_index)}));
 388
 389        return wrap(pInput);
 390     }
 391 }
 392
 393 // GS output stream layout
 394 #define VERTEX_COUNT_SIZE 32
 395 #define CONTROL_HEADER_SIZE (8*32)
 396
 397 void
 398 BuilderSWR::swr_gs_llvm_emit_vertex(const struct lp_build_tgsi_gs_iface *gs_base,
 399                            struct lp_build_tgsi_context * bld_base,
 400                            LLVMValueRef (*outputs)[4],
 401                            LLVMValueRef emitted_vertices_vec)
 402 {
 403     swr_gs_llvm_iface *iface = (swr_gs_llvm_iface*)gs_base;
 404
 405     IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 406
 407     const uint32_t headerSize = VERTEX_COUNT_SIZE + CONTROL_HEADER_SIZE;
 408     const uint32_t attribSize = 4 * sizeof(float);
 409     const uint32_t vertSize = attribSize * SWR_VTX_NUM_SLOTS;
 410     Value *pVertexOffset = MUL(unwrap(emitted_vertices_vec), VIMMED1(vertSize));
 411
 412     Value *vMask = LOAD(iface->pGsCtx, {0, SWR_GS_CONTEXT_mask});
 413     Value *vMask1 = TRUNC(vMask, VectorType::get(mInt1Ty, mVWidth));
 414
 415     Value *pStack = STACKSAVE();
 416     Value *pTmpPtr = ALLOCA(mFP32Ty, C(4)); // used for dummy write for lane masking
 417
 418     for (uint32_t attrib = 0; attrib < iface->num_outputs; ++attrib) {
 419        uint32_t attribSlot = attrib;
 420        uint32_t sgvChannel = 0;
 421        if (iface->info->output_semantic_name[attrib] == TGSI_SEMANTIC_PSIZE) {
 422           attribSlot = VERTEX_SGV_SLOT;
 423           sgvChannel = VERTEX_SGV_POINT_SIZE_COMP;
 424        } else if (iface->info->output_semantic_name[attrib] == TGSI_SEMANTIC_LAYER) {
 425           attribSlot = VERTEX_SGV_SLOT;
 426           sgvChannel = VERTEX_SGV_RTAI_COMP;
 427        } else if (iface->info->output_semantic_name[attrib] == TGSI_SEMANTIC_VIEWPORT_INDEX) {
 428           attribSlot = VERTEX_SGV_SLOT;
 429           sgvChannel = VERTEX_SGV_VAI_COMP;
 430        } else if (iface->info->output_semantic_name[attrib] == TGSI_SEMANTIC_POSITION) {
 431           attribSlot = VERTEX_POSITION_SLOT;
 432        } else {
 433           attribSlot = VERTEX_ATTRIB_START_SLOT + attrib;
 434           if (iface->info->writes_position) {
 435              attribSlot--;
 436           }
 437        }
 438
 439        Value *pOutputOffset = ADD(pVertexOffset, VIMMED1(headerSize + attribSize * attribSlot)); // + sgvChannel ?
 440
 441        for (uint32_t lane = 0; lane < mVWidth; ++lane) {
 442           Value *pLaneOffset = VEXTRACT(pOutputOffset, C(lane));
 443           Value *pStream = LOAD(iface->pGsCtx, {0, SWR_GS_CONTEXT_pStreams, lane});
 444           Value *pStreamOffset = GEP(pStream, pLaneOffset);
 445           pStreamOffset = BITCAST(pStreamOffset, mFP32PtrTy);
 446
 447           Value *pLaneMask = VEXTRACT(vMask1, C(lane));
 448           pStreamOffset = SELECT(pLaneMask, pStreamOffset, pTmpPtr);
 449
 450           for (uint32_t channel = 0; channel < 4; ++channel) {
 451              Value *vData;
 452
 453              if (attribSlot == VERTEX_SGV_SLOT)
 454                 vData = LOAD(unwrap(outputs[attrib][0]));
 455              else
 456                 vData = LOAD(unwrap(outputs[attrib][channel]));
 457
 458              if (attribSlot != VERTEX_SGV_SLOT ||
 459                  sgvChannel == channel) {
 460                 vData = VEXTRACT(vData, C(lane));
 461                 STORE(vData, pStreamOffset);
 462              }
 463              pStreamOffset = GEP(pStreamOffset, C(1));
 464           }
 465        }
 466     }
 467
 468     STACKRESTORE(pStack);
 469 }
 470
 471 void
 472 BuilderSWR::swr_gs_llvm_end_primitive(const struct lp_build_tgsi_gs_iface *gs_base,
 473                              struct lp_build_tgsi_context * bld_base,
 474                              LLVMValueRef verts_per_prim_vec,
 475                              LLVMValueRef emitted_prims_vec)
 476 {
 477     swr_gs_llvm_iface *iface = (swr_gs_llvm_iface*)gs_base;
 478
 479     IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 480
 481     Value *vMask = LOAD(iface->pGsCtx, { 0, SWR_GS_CONTEXT_mask });
 482     Value *vMask1 = TRUNC(vMask, VectorType::get(mInt1Ty, 8));
 483
 484     uint32_t vertsPerPrim = iface->num_verts_per_prim;
 485
 486     Value *vCount =
 487        ADD(MUL(unwrap(emitted_prims_vec), VIMMED1(vertsPerPrim)),
 488            unwrap(verts_per_prim_vec));
 489
 490     struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
 491     vCount = LOAD(unwrap(bld->total_emitted_vertices_vec_ptr));
 492
 493     struct lp_exec_mask *exec_mask = &bld->exec_mask;
 494     Value *mask = unwrap(lp_build_mask_value(bld->mask));
 495     if (exec_mask->has_mask)
 496        mask = AND(mask, unwrap(exec_mask->exec_mask));
 497
 498     Value *cmpMask = VMASK(ICMP_NE(unwrap(verts_per_prim_vec), VIMMED1(0)));
 499     mask = AND(mask, cmpMask);
 500     vMask1 = TRUNC(mask, VectorType::get(mInt1Ty, 8));
 501
 502     vCount = SUB(vCount, VIMMED1(1));
 503     Value *vOffset = ADD(UDIV(vCount, VIMMED1(8)), VIMMED1(VERTEX_COUNT_SIZE));
 504     Value *vValue = SHL(VIMMED1(1), UREM(vCount, VIMMED1(8)));
 505
 506     vValue = TRUNC(vValue, VectorType::get(mInt8Ty, 8));
 507
 508     Value *pStack = STACKSAVE();
 509     Value *pTmpPtr = ALLOCA(mInt8Ty, C(4)); // used for dummy read/write for lane masking
 510
 511     for (uint32_t lane = 0; lane < mVWidth; ++lane) {
 512        Value *vLaneOffset = VEXTRACT(vOffset, C(lane));
 513        Value *pStream = LOAD(iface->pGsCtx, {0, SWR_GS_CONTEXT_pStreams, lane});
 514        Value *pStreamOffset = GEP(pStream, vLaneOffset);
 515
 516        Value *pLaneMask = VEXTRACT(vMask1, C(lane));
 517        pStreamOffset = SELECT(pLaneMask, pStreamOffset, pTmpPtr);
 518
 519        Value *vVal = LOAD(pStreamOffset);
 520        vVal = OR(vVal, VEXTRACT(vValue, C(lane)));
 521        STORE(vVal, pStreamOffset);
 522     }
 523
 524     STACKRESTORE(pStack);
 525 }
 526
 527 void
 528 BuilderSWR::swr_gs_llvm_epilogue(const struct lp_build_tgsi_gs_iface *gs_base,
 529                         struct lp_build_tgsi_context * bld_base,
 530                         LLVMValueRef total_emitted_vertices_vec,
 531                         LLVMValueRef emitted_prims_vec)
 532 {
 533    swr_gs_llvm_iface *iface = (swr_gs_llvm_iface*)gs_base;
 534
 535    IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 536
 537    // Store emit count to each output stream in the first DWORD
 538    for (uint32_t lane = 0; lane < mVWidth; ++lane)
 539    {
 540       Value* pStream = LOAD(iface->pGsCtx, {0, SWR_GS_CONTEXT_pStreams, lane});
 541       pStream = BITCAST(pStream, mInt32PtrTy);
 542       Value* pLaneCount = VEXTRACT(unwrap(total_emitted_vertices_vec), C(lane));
 543       STORE(pLaneCount, pStream);
 544    }
 545 }
 546
 547 PFN_GS_FUNC
 548 BuilderSWR::CompileGS(struct swr_context *ctx, swr_jit_gs_key &key)
 549 {
 550    SWR_GS_STATE *pGS = &ctx->gs->gsState;
 551    struct tgsi_shader_info *info = &ctx->gs->info.base;
 552
 553    memset(pGS, 0, sizeof(*pGS));
 554
 555    pGS->gsEnable = true;
 556
 557    pGS->numInputAttribs = info->num_inputs;
 558    pGS->outputTopology =
 559       swr_convert_prim_topology(info->properties[TGSI_PROPERTY_GS_OUTPUT_PRIM]);
 560    pGS->maxNumVerts = info->properties[TGSI_PROPERTY_GS_MAX_OUTPUT_VERTICES];
 561    pGS->instanceCount = info->properties[TGSI_PROPERTY_GS_INVOCATIONS];
 562
 563    // XXX: single stream for now...
 564    pGS->isSingleStream = true;
 565    pGS->singleStreamID = 0;
 566
 567    pGS->vertexAttribOffset = VERTEX_ATTRIB_START_SLOT; // TODO: optimize
 568    pGS->srcVertexAttribOffset = VERTEX_ATTRIB_START_SLOT; // TODO: optimize
 569    pGS->inputVertStride = pGS->numInputAttribs + pGS->vertexAttribOffset;
 570    pGS->outputVertexSize = SWR_VTX_NUM_SLOTS;
 571    pGS->controlDataSize = 8; // GS ouputs max of 8 32B units
 572    pGS->controlDataOffset = VERTEX_COUNT_SIZE;
 573    pGS->outputVertexOffset = pGS->controlDataOffset + CONTROL_HEADER_SIZE;
 574
 575    pGS->allocationSize =
 576       VERTEX_COUNT_SIZE + // vertex count
 577       CONTROL_HEADER_SIZE + // control header
 578       (SWR_VTX_NUM_SLOTS * 16) * // sizeof vertex
 579       pGS->maxNumVerts; // num verts
 580
 581    struct swr_geometry_shader *gs = ctx->gs;
 582
 583    LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
 584    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
 585
 586    memset(outputs, 0, sizeof(outputs));
 587
 588    AttrBuilder attrBuilder;
 589    attrBuilder.addStackAlignmentAttr(JM()->mVWidth * sizeof(float));
 590
 591    std::vector<Type *> gsArgs{PointerType::get(Gen_swr_draw_context(JM()), 0),
 592                               PointerType::get(mInt8Ty, 0),
 593                               PointerType::get(Gen_SWR_GS_CONTEXT(JM()), 0)};
 594    FunctionType *vsFuncType =
 595       FunctionType::get(Type::getVoidTy(JM()->mContext), gsArgs, false);
 596
 597    // create new vertex shader function
 598    auto pFunction = Function::Create(vsFuncType,
 599                                      GlobalValue::ExternalLinkage,
 600                                      "GS",
 601                                      JM()->mpCurrentModule);
 602 #if HAVE_LLVM < 0x0500
 603    AttributeSet attrSet = AttributeSet::get(
 604       JM()->mContext, AttributeSet::FunctionIndex, attrBuilder);
 605    pFunction->addAttributes(AttributeSet::FunctionIndex, attrSet);
 606 #else
 607    pFunction->addAttributes(AttributeList::FunctionIndex, attrBuilder);
 608 #endif
 609
 610    BasicBlock *block = BasicBlock::Create(JM()->mContext, "entry", pFunction);
 611    IRB()->SetInsertPoint(block);
 612    LLVMPositionBuilderAtEnd(gallivm->builder, wrap(block));
 613
 614    auto argitr = pFunction->arg_begin();
 615    Value *hPrivateData = &*argitr++;
 616    hPrivateData->setName("hPrivateData");
 617    Value *pWorkerData = &*argitr++;
 618    pWorkerData->setName("pWorkerData");
 619    Value *pGsCtx = &*argitr++;
 620    pGsCtx->setName("gsCtx");
 621
 622    Value *consts_ptr =
 623       GEP(hPrivateData, {C(0), C(swr_draw_context_constantGS)});
 624    consts_ptr->setName("gs_constants");
 625    Value *const_sizes_ptr =
 626       GEP(hPrivateData, {0, swr_draw_context_num_constantsGS});
 627    const_sizes_ptr->setName("num_gs_constants");
 628
 629    struct lp_build_sampler_soa *sampler =
 630       swr_sampler_soa_create(key.sampler, PIPE_SHADER_GEOMETRY);
 631
 632    struct lp_bld_tgsi_system_values system_values;
 633    memset(&system_values, 0, sizeof(system_values));
 634    system_values.prim_id = wrap(LOAD(pGsCtx, {0, SWR_GS_CONTEXT_PrimitiveID}));
 635    system_values.instance_id = wrap(LOAD(pGsCtx, {0, SWR_GS_CONTEXT_InstanceID}));
 636
 637    std::vector<Constant*> mapConstants;
 638    Value *vtxAttribMap = ALLOCA(ArrayType::get(mInt32Ty, PIPE_MAX_SHADER_INPUTS));
 639    for (unsigned slot = 0; slot < info->num_inputs; slot++) {
 640       ubyte semantic_name = info->input_semantic_name[slot];
 641       ubyte semantic_idx = info->input_semantic_index[slot];
 642
 643       unsigned vs_slot = locate_linkage(semantic_name, semantic_idx, &ctx->vs->info.base);
 644
 645       vs_slot += VERTEX_ATTRIB_START_SLOT;
 646
 647       if (ctx->vs->info.base.output_semantic_name[0] == TGSI_SEMANTIC_POSITION)
 648          vs_slot--;
 649
 650       if (semantic_name == TGSI_SEMANTIC_POSITION)
 651          vs_slot = VERTEX_POSITION_SLOT;
 652
 653       STORE(C(vs_slot), vtxAttribMap, {0, slot});
 654       mapConstants.push_back(C(vs_slot));
 655    }
 656
 657    struct lp_build_mask_context mask;
 658    Value *mask_val = LOAD(pGsCtx, {0, SWR_GS_CONTEXT_mask}, "gsMask");
 659    lp_build_mask_begin(&mask, gallivm,
 660                        lp_type_float_vec(32, 32 * 8), wrap(mask_val));
 661
 662    // zero out cut buffer so we can load/modify/store bits
 663    for (uint32_t lane = 0; lane < mVWidth; ++lane)
 664    {
 665       Value* pStream = LOAD(pGsCtx, {0, SWR_GS_CONTEXT_pStreams, lane});
 666       MEMSET(pStream, C((char)0), VERTEX_COUNT_SIZE + CONTROL_HEADER_SIZE, sizeof(float) * KNOB_SIMD_WIDTH);
 667    }
 668
 669    struct swr_gs_llvm_iface gs_iface;
 670    gs_iface.base.fetch_input = ::swr_gs_llvm_fetch_input;
 671    gs_iface.base.emit_vertex = ::swr_gs_llvm_emit_vertex;
 672    gs_iface.base.end_primitive = ::swr_gs_llvm_end_primitive;
 673    gs_iface.base.gs_epilogue = ::swr_gs_llvm_epilogue;
 674    gs_iface.pBuilder = this;
 675    gs_iface.pGsCtx = pGsCtx;
 676    gs_iface.pGsState = pGS;
 677    gs_iface.num_outputs = gs->info.base.num_outputs;
 678    gs_iface.num_verts_per_prim =
 679       u_vertices_per_prim((pipe_prim_type)info->properties[TGSI_PROPERTY_GS_OUTPUT_PRIM]);
 680    gs_iface.info = info;
 681    gs_iface.pVtxAttribMap = vtxAttribMap;
 682
 683    lp_build_tgsi_soa(gallivm,
 684                      gs->pipe.tokens,
 685                      lp_type_float_vec(32, 32 * 8),
 686                      &mask,
 687                      wrap(consts_ptr),
 688                      wrap(const_sizes_ptr),
 689                      &system_values,
 690                      inputs,
 691                      outputs,
 692                      wrap(hPrivateData), // (sampler context)
 693                      NULL, // thread data
 694                      sampler,
 695                      &gs->info.base,
 696                      &gs_iface.base);
 697
 698    lp_build_mask_end(&mask);
 699
 700    sampler->destroy(sampler);
 701
 702    IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 703
 704    RET_VOID();
 705
 706    gallivm_verify_function(gallivm, wrap(pFunction));
 707    gallivm_compile_module(gallivm);
 708
 709    PFN_GS_FUNC pFunc =
 710       (PFN_GS_FUNC)gallivm_jit_function(gallivm, wrap(pFunction));
 711
 712    debug_printf("geom shader  %p\n", pFunc);
 713    assert(pFunc && "Error: GeomShader = NULL");
 714
 715    JM()->mIsModuleFinalized = true;
 716
 717    return pFunc;
 718 }
 719
 720 PFN_GS_FUNC
 721 swr_compile_gs(struct swr_context *ctx, swr_jit_gs_key &key)
 722 {
 723    BuilderSWR builder(
 724       reinterpret_cast<JitManager *>(swr_screen(ctx->pipe.screen)->hJitMgr),
 725       "GS");
 726    PFN_GS_FUNC func = builder.CompileGS(ctx, key);
 727
 728    ctx->gs->map.insert(std::make_pair(key, make_unique<VariantGS>(builder.gallivm, func)));
 729    return func;
 730 }
 731
 732 void
 733 BuilderSWR::WriteVS(Value *pVal, Value *pVsContext, Value *pVtxOutput, unsigned slot, unsigned channel)
 734 {
 735 #if USE_SIMD16_FRONTEND && !USE_SIMD16_VS
 736    // interleave the simdvertex components into the dest simd16vertex
 737    //   slot16offset = slot8offset * 2
 738    //   comp16offset = comp8offset * 2 + alternateOffset
 739
 740    Value *offset = LOAD(pVsContext, { 0, SWR_VS_CONTEXT_AlternateOffset });
 741    Value *pOut = GEP(pVtxOutput, { C(0), C(0), C(slot * 2), offset } );
 742    STORE(pVal, pOut, {channel * 2});
 743 #else
 744    Value *pOut = GEP(pVtxOutput, {0, 0, slot});
 745    STORE(pVal, pOut, {0, channel});
 746 #endif
 747 }
 748
 749 PFN_VERTEX_FUNC
 750 BuilderSWR::CompileVS(struct swr_context *ctx, swr_jit_vs_key &key)
 751 {
 752    struct swr_vertex_shader *swr_vs = ctx->vs;
 753
 754    LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
 755    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
 756
 757    memset(outputs, 0, sizeof(outputs));
 758
 759    AttrBuilder attrBuilder;
 760    attrBuilder.addStackAlignmentAttr(JM()->mVWidth * sizeof(float));
 761
 762    std::vector<Type *> vsArgs{PointerType::get(Gen_swr_draw_context(JM()), 0),
 763                               PointerType::get(mInt8Ty, 0),
 764                               PointerType::get(Gen_SWR_VS_CONTEXT(JM()), 0)};
 765    FunctionType *vsFuncType =
 766       FunctionType::get(Type::getVoidTy(JM()->mContext), vsArgs, false);
 767
 768    // create new vertex shader function
 769    auto pFunction = Function::Create(vsFuncType,
 770                                      GlobalValue::ExternalLinkage,
 771                                      "VS",
 772                                      JM()->mpCurrentModule);
 773 #if HAVE_LLVM < 0x0500
 774    AttributeSet attrSet = AttributeSet::get(
 775       JM()->mContext, AttributeSet::FunctionIndex, attrBuilder);
 776    pFunction->addAttributes(AttributeSet::FunctionIndex, attrSet);
 777 #else
 778    pFunction->addAttributes(AttributeList::FunctionIndex, attrBuilder);
 779 #endif
 780
 781    BasicBlock *block = BasicBlock::Create(JM()->mContext, "entry", pFunction);
 782    IRB()->SetInsertPoint(block);
 783    LLVMPositionBuilderAtEnd(gallivm->builder, wrap(block));
 784
 785    auto argitr = pFunction->arg_begin();
 786    Value *hPrivateData = &*argitr++;
 787    hPrivateData->setName("hPrivateData");
 788    Value *pWorkerData = &*argitr++;
 789    pWorkerData->setName("pWorkerData");
 790    Value *pVsCtx = &*argitr++;
 791    pVsCtx->setName("vsCtx");
 792
 793    Value *consts_ptr = GEP(hPrivateData, {C(0), C(swr_draw_context_constantVS)});
 794
 795    consts_ptr->setName("vs_constants");
 796    Value *const_sizes_ptr =
 797       GEP(hPrivateData, {0, swr_draw_context_num_constantsVS});
 798    const_sizes_ptr->setName("num_vs_constants");
 799
 800    Value *vtxInput = LOAD(pVsCtx, {0, SWR_VS_CONTEXT_pVin});
 801 #if USE_SIMD16_VS
 802    vtxInput = BITCAST(vtxInput, PointerType::get(Gen_simd16vertex(JM()), 0));
 803 #endif
 804
 805    for (uint32_t attrib = 0; attrib < PIPE_MAX_SHADER_INPUTS; attrib++) {
 806       const unsigned mask = swr_vs->info.base.input_usage_mask[attrib];
 807       for (uint32_t channel = 0; channel < TGSI_NUM_CHANNELS; channel++) {
 808          if (mask & (1 << channel)) {
 809             inputs[attrib][channel] =
 810                wrap(LOAD(vtxInput, {0, 0, attrib, channel}));
 811          }
 812       }
 813    }
 814
 815    struct lp_build_sampler_soa *sampler =
 816       swr_sampler_soa_create(key.sampler, PIPE_SHADER_VERTEX);
 817
 818    struct lp_bld_tgsi_system_values system_values;
 819    memset(&system_values, 0, sizeof(system_values));
 820    system_values.instance_id = wrap(LOAD(pVsCtx, {0, SWR_VS_CONTEXT_InstanceID}));
 821
 822 #if USE_SIMD16_VS
 823    system_values.vertex_id = wrap(LOAD(pVsCtx, {0, SWR_VS_CONTEXT_VertexID16}));
 824 #else
 825    system_values.vertex_id = wrap(LOAD(pVsCtx, {0, SWR_VS_CONTEXT_VertexID}));
 826 #endif
 827
 828 #if USE_SIMD16_VS
 829    uint32_t vectorWidth = mVWidth16;
 830 #else
 831    uint32_t vectorWidth = mVWidth;
 832 #endif
 833
 834    lp_build_tgsi_soa(gallivm,
 835                      swr_vs->pipe.tokens,
 836                      lp_type_float_vec(32, 32 * vectorWidth),
 837                      NULL, // mask
 838                      wrap(consts_ptr),
 839                      wrap(const_sizes_ptr),
 840                      &system_values,
 841                      inputs,
 842                      outputs,
 843                      wrap(hPrivateData), // (sampler context)
 844                      NULL, // thread data
 845                      sampler, // sampler
 846                      &swr_vs->info.base,
 847                      NULL); // geometry shader face
 848
 849    sampler->destroy(sampler);
 850
 851    IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 852
 853    Value *vtxOutput = LOAD(pVsCtx, {0, SWR_VS_CONTEXT_pVout});
 854 #if USE_SIMD16_VS
 855    vtxOutput = BITCAST(vtxOutput, PointerType::get(Gen_simd16vertex(JM()), 0));
 856 #endif
 857
 858    for (uint32_t channel = 0; channel < TGSI_NUM_CHANNELS; channel++) {
 859       for (uint32_t attrib = 0; attrib < PIPE_MAX_SHADER_OUTPUTS; attrib++) {
 860          if (!outputs[attrib][channel])
 861             continue;
 862
 863          Value *val;
 864          uint32_t outSlot;
 865
 866          if (swr_vs->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_PSIZE) {
 867             if (channel != VERTEX_SGV_POINT_SIZE_COMP)
 868                continue;
 869             val = LOAD(unwrap(outputs[attrib][0]));
 870             outSlot = VERTEX_SGV_SLOT;
 871          } else if (swr_vs->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_POSITION) {
 872             val = LOAD(unwrap(outputs[attrib][channel]));
 873             outSlot = VERTEX_POSITION_SLOT;
 874          } else {
 875             val = LOAD(unwrap(outputs[attrib][channel]));
 876             outSlot = VERTEX_ATTRIB_START_SLOT + attrib;
 877             if (swr_vs->info.base.output_semantic_name[0] == TGSI_SEMANTIC_POSITION)
 878                outSlot--;
 879          }
 880
 881          WriteVS(val, pVsCtx, vtxOutput, outSlot, channel);
 882       }
 883    }
 884
 885    if (ctx->rasterizer->clip_plane_enable ||
 886        swr_vs->info.base.culldist_writemask) {
 887       unsigned clip_mask = ctx->rasterizer->clip_plane_enable;
 888
 889       unsigned cv = 0;
 890       if (swr_vs->info.base.writes_clipvertex) {
 891          cv = locate_linkage(TGSI_SEMANTIC_CLIPVERTEX, 0,
 892                              &swr_vs->info.base);
 893       } else {
 894          for (int i = 0; i < PIPE_MAX_SHADER_OUTPUTS; i++) {
 895             if (swr_vs->info.base.output_semantic_name[i] == TGSI_SEMANTIC_POSITION &&
 896                 swr_vs->info.base.output_semantic_index[i] == 0) {
 897                cv = i;
 898                break;
 899             }
 900          }
 901       }
 902       LLVMValueRef cx = LLVMBuildLoad(gallivm->builder, outputs[cv][0], "");
 903       LLVMValueRef cy = LLVMBuildLoad(gallivm->builder, outputs[cv][1], "");
 904       LLVMValueRef cz = LLVMBuildLoad(gallivm->builder, outputs[cv][2], "");
 905       LLVMValueRef cw = LLVMBuildLoad(gallivm->builder, outputs[cv][3], "");
 906
 907       for (unsigned val = 0; val < PIPE_MAX_CLIP_PLANES; val++) {
 908          // clip distance overrides user clip planes
 909          if ((swr_vs->info.base.clipdist_writemask & clip_mask & (1 << val)) ||
 910              ((swr_vs->info.base.culldist_writemask << swr_vs->info.base.num_written_clipdistance) & (1 << val))) {
 911             unsigned cv = locate_linkage(TGSI_SEMANTIC_CLIPDIST, val < 4 ? 0 : 1,
 912                                          &swr_vs->info.base);
 913             if (val < 4) {
 914                LLVMValueRef dist = LLVMBuildLoad(gallivm->builder, outputs[cv][val], "");
 915                WriteVS(unwrap(dist), pVsCtx, vtxOutput, VERTEX_CLIPCULL_DIST_LO_SLOT, val);
 916             } else {
 917                LLVMValueRef dist = LLVMBuildLoad(gallivm->builder, outputs[cv][val - 4], "");
 918                WriteVS(unwrap(dist), pVsCtx, vtxOutput, VERTEX_CLIPCULL_DIST_HI_SLOT, val - 4);
 919             }
 920             continue;
 921          }
 922
 923          if (!(clip_mask & (1 << val)))
 924             continue;
 925
 926          Value *px = LOAD(GEP(hPrivateData, {0, swr_draw_context_userClipPlanes, val, 0}));
 927          Value *py = LOAD(GEP(hPrivateData, {0, swr_draw_context_userClipPlanes, val, 1}));
 928          Value *pz = LOAD(GEP(hPrivateData, {0, swr_draw_context_userClipPlanes, val, 2}));
 929          Value *pw = LOAD(GEP(hPrivateData, {0, swr_draw_context_userClipPlanes, val, 3}));
 930 #if USE_SIMD16_VS
 931          Value *bpx = VBROADCAST_16(px);
 932          Value *bpy = VBROADCAST_16(py);
 933          Value *bpz = VBROADCAST_16(pz);
 934          Value *bpw = VBROADCAST_16(pw);
 935 #else
 936          Value *bpx = VBROADCAST(px);
 937          Value *bpy = VBROADCAST(py);
 938          Value *bpz = VBROADCAST(pz);
 939          Value *bpw = VBROADCAST(pw);
 940 #endif
 941          Value *dist = FADD(FMUL(unwrap(cx), bpx),
 942                             FADD(FMUL(unwrap(cy), bpy),
 943                                  FADD(FMUL(unwrap(cz), bpz),
 944                                       FMUL(unwrap(cw), bpw))));
 945
 946          if (val < 4)
 947             WriteVS(dist, pVsCtx, vtxOutput, VERTEX_CLIPCULL_DIST_LO_SLOT, val);
 948          else
 949             WriteVS(dist, pVsCtx, vtxOutput, VERTEX_CLIPCULL_DIST_HI_SLOT, val - 4);
 950       }
 951    }
 952
 953    RET_VOID();
 954
 955    gallivm_verify_function(gallivm, wrap(pFunction));
 956    gallivm_compile_module(gallivm);
 957
 958    //   lp_debug_dump_value(func);
 959
 960    PFN_VERTEX_FUNC pFunc =
 961       (PFN_VERTEX_FUNC)gallivm_jit_function(gallivm, wrap(pFunction));
 962
 963    debug_printf("vert shader  %p\n", pFunc);
 964    assert(pFunc && "Error: VertShader = NULL");
 965
 966    JM()->mIsModuleFinalized = true;
 967
 968    return pFunc;
 969 }
 970
 971 PFN_VERTEX_FUNC
 972 swr_compile_vs(struct swr_context *ctx, swr_jit_vs_key &key)
 973 {
 974    if (!ctx->vs->pipe.tokens)
 975       return NULL;
 976
 977    BuilderSWR builder(
 978       reinterpret_cast<JitManager *>(swr_screen(ctx->pipe.screen)->hJitMgr),
 979       "VS");
 980    PFN_VERTEX_FUNC func = builder.CompileVS(ctx, key);
 981
 982    ctx->vs->map.insert(std::make_pair(key, make_unique<VariantVS>(builder.gallivm, func)));
 983    return func;
 984 }
 985
 986 unsigned
 987 swr_so_adjust_attrib(unsigned in_attrib,
 988                      swr_vertex_shader *swr_vs)
 989 {
 990    ubyte semantic_name;
 991    unsigned attrib;
 992
 993    attrib = in_attrib + VERTEX_ATTRIB_START_SLOT;
 994
 995    if (swr_vs) {
 996       semantic_name = swr_vs->info.base.output_semantic_name[in_attrib];
 997       if (semantic_name == TGSI_SEMANTIC_POSITION) {
 998          attrib = VERTEX_POSITION_SLOT;
 999       } else if (semantic_name == TGSI_SEMANTIC_PSIZE) {
1000          attrib = VERTEX_SGV_SLOT;
1001       } else if (semantic_name == TGSI_SEMANTIC_LAYER) {
1002          attrib = VERTEX_SGV_SLOT;
1003       } else {
1004          if (swr_vs->info.base.writes_position) {
1005                attrib--;
1006          }
1007       }
1008    }
1009
1010    return attrib;
1011 }
1012
1013 static unsigned
1014 locate_linkage(ubyte name, ubyte index, struct tgsi_shader_info *info)
1015 {
1016    for (int i = 0; i < PIPE_MAX_SHADER_OUTPUTS; i++) {
1017       if ((info->output_semantic_name[i] == name)
1018           && (info->output_semantic_index[i] == index)) {
1019          return i;
1020       }
1021    }
1022
1023    return 0xFFFFFFFF;
1024 }
1025
1026 PFN_PIXEL_KERNEL
1027 BuilderSWR::CompileFS(struct swr_context *ctx, swr_jit_fs_key &key)
1028 {
1029    struct swr_fragment_shader *swr_fs = ctx->fs;
1030
1031    struct tgsi_shader_info *pPrevShader;
1032    if (ctx->gs)
1033       pPrevShader = &ctx->gs->info.base;
1034    else
1035       pPrevShader = &ctx->vs->info.base;
1036
1037    LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
1038    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
1039
1040    memset(inputs, 0, sizeof(inputs));
1041    memset(outputs, 0, sizeof(outputs));
1042
1043    struct lp_build_sampler_soa *sampler = NULL;
1044
1045    AttrBuilder attrBuilder;
1046    attrBuilder.addStackAlignmentAttr(JM()->mVWidth * sizeof(float));
1047
1048    std::vector<Type *> fsArgs{PointerType::get(Gen_swr_draw_context(JM()), 0),
1049                               PointerType::get(mInt8Ty, 0),
1050                               PointerType::get(Gen_SWR_PS_CONTEXT(JM()), 0)};
1051    FunctionType *funcType =
1052       FunctionType::get(Type::getVoidTy(JM()->mContext), fsArgs, false);
1053
1054    auto pFunction = Function::Create(funcType,
1055                                      GlobalValue::ExternalLinkage,
1056                                      "FS",
1057                                      JM()->mpCurrentModule);
1058 #if HAVE_LLVM < 0x0500
1059    AttributeSet attrSet = AttributeSet::get(
1060       JM()->mContext, AttributeSet::FunctionIndex, attrBuilder);
1061    pFunction->addAttributes(AttributeSet::FunctionIndex, attrSet);
1062 #else
1063    pFunction->addAttributes(AttributeList::FunctionIndex, attrBuilder);
1064 #endif
1065
1066    BasicBlock *block = BasicBlock::Create(JM()->mContext, "entry", pFunction);
1067    IRB()->SetInsertPoint(block);
1068    LLVMPositionBuilderAtEnd(gallivm->builder, wrap(block));
1069
1070    auto args = pFunction->arg_begin();
1071    Value *hPrivateData = &*args++;
1072    hPrivateData->setName("hPrivateData");
1073    Value *pWorkerData = &*args++;
1074    pWorkerData->setName("pWorkerData");
1075    Value *pPS = &*args++;
1076    pPS->setName("psCtx");
1077
1078    Value *consts_ptr = GEP(hPrivateData, {0, swr_draw_context_constantFS});
1079    consts_ptr->setName("fs_constants");
1080    Value *const_sizes_ptr =
1081       GEP(hPrivateData, {0, swr_draw_context_num_constantsFS});
1082    const_sizes_ptr->setName("num_fs_constants");
1083
1084    // load *pAttribs, *pPerspAttribs
1085    Value *pRawAttribs = LOAD(pPS, {0, SWR_PS_CONTEXT_pAttribs}, "pRawAttribs");
1086    Value *pPerspAttribs =
1087       LOAD(pPS, {0, SWR_PS_CONTEXT_pPerspAttribs}, "pPerspAttribs");
1088
1089    swr_fs->constantMask = 0;
1090    swr_fs->flatConstantMask = 0;
1091    swr_fs->pointSpriteMask = 0;
1092
1093    for (int attrib = 0; attrib < PIPE_MAX_SHADER_INPUTS; attrib++) {
1094       const unsigned mask = swr_fs->info.base.input_usage_mask[attrib];
1095       const unsigned interpMode = swr_fs->info.base.input_interpolate[attrib];
1096       const unsigned interpLoc = swr_fs->info.base.input_interpolate_loc[attrib];
1097
1098       if (!mask)
1099          continue;
1100
1101       // load i,j
1102       Value *vi = nullptr, *vj = nullptr;
1103       switch (interpLoc) {
1104       case TGSI_INTERPOLATE_LOC_CENTER:
1105          vi = LOAD(pPS, {0, SWR_PS_CONTEXT_vI, PixelPositions_center}, "i");
1106          vj = LOAD(pPS, {0, SWR_PS_CONTEXT_vJ, PixelPositions_center}, "j");
1107          break;
1108       case TGSI_INTERPOLATE_LOC_CENTROID:
1109          vi = LOAD(pPS, {0, SWR_PS_CONTEXT_vI, PixelPositions_centroid}, "i");
1110          vj = LOAD(pPS, {0, SWR_PS_CONTEXT_vJ, PixelPositions_centroid}, "j");
1111          break;
1112       case TGSI_INTERPOLATE_LOC_SAMPLE:
1113          vi = LOAD(pPS, {0, SWR_PS_CONTEXT_vI, PixelPositions_sample}, "i");
1114          vj = LOAD(pPS, {0, SWR_PS_CONTEXT_vJ, PixelPositions_sample}, "j");
1115          break;
1116       }
1117
1118       // load/compute w
1119       Value *vw = nullptr, *pAttribs;
1120       if (interpMode == TGSI_INTERPOLATE_PERSPECTIVE ||
1121           interpMode == TGSI_INTERPOLATE_COLOR) {
1122          pAttribs = pPerspAttribs;
1123          switch (interpLoc) {
1124          case TGSI_INTERPOLATE_LOC_CENTER:
1125             vw = VRCP(LOAD(pPS, {0, SWR_PS_CONTEXT_vOneOverW, PixelPositions_center}));
1126             break;
1127          case TGSI_INTERPOLATE_LOC_CENTROID:
1128             vw = VRCP(LOAD(pPS, {0, SWR_PS_CONTEXT_vOneOverW, PixelPositions_centroid}));
1129             break;
1130          case TGSI_INTERPOLATE_LOC_SAMPLE:
1131             vw = VRCP(LOAD(pPS, {0, SWR_PS_CONTEXT_vOneOverW, PixelPositions_sample}));
1132             break;
1133          }
1134       } else {
1135          pAttribs = pRawAttribs;
1136          vw = VIMMED1(1.f);
1137       }
1138
1139       vw->setName("w");
1140
1141       ubyte semantic_name = swr_fs->info.base.input_semantic_name[attrib];
1142       ubyte semantic_idx = swr_fs->info.base.input_semantic_index[attrib];
1143
1144       if (semantic_name == TGSI_SEMANTIC_FACE) {
1145          Value *ff =
1146             UI_TO_FP(LOAD(pPS, {0, SWR_PS_CONTEXT_frontFace}), mFP32Ty);
1147          ff = FSUB(FMUL(ff, C(2.0f)), C(1.0f));
1148          ff = VECTOR_SPLAT(JM()->mVWidth, ff, "vFrontFace");
1149
1150          inputs[attrib][0] = wrap(ff);
1151          inputs[attrib][1] = wrap(VIMMED1(0.0f));
1152          inputs[attrib][2] = wrap(VIMMED1(0.0f));
1153          inputs[attrib][3] = wrap(VIMMED1(1.0f));
1154          continue;
1155       } else if (semantic_name == TGSI_SEMANTIC_POSITION) { // gl_FragCoord
1156          if (swr_fs->info.base.properties[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER] ==
1157              TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER) {
1158             inputs[attrib][0] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vX, PixelPositions_center}, "vX"));
1159             inputs[attrib][1] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vY, PixelPositions_center}, "vY"));
1160          } else {
1161             inputs[attrib][0] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vX, PixelPositions_UL}, "vX"));
1162             inputs[attrib][1] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vY, PixelPositions_UL}, "vY"));
1163          }
1164          inputs[attrib][2] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vZ}, "vZ"));
1165          inputs[attrib][3] =
1166             wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vOneOverW, PixelPositions_center}, "vOneOverW"));
1167          continue;
1168       }
1169
1170       unsigned linkedAttrib =
1171          locate_linkage(semantic_name, semantic_idx, pPrevShader) - 1;
1172
1173       uint32_t extraAttribs = 0;
1174       if (semantic_name == TGSI_SEMANTIC_PRIMID && !ctx->gs) {
1175          /* non-gs generated primID - need to grab from swizzleMap override */
1176          linkedAttrib = pPrevShader->num_outputs - 1;
1177          swr_fs->constantMask |= 1 << linkedAttrib;
1178          extraAttribs++;
1179       } else if (semantic_name == TGSI_SEMANTIC_GENERIC &&
1180           key.sprite_coord_enable & (1 << semantic_idx)) {
1181          /* we add an extra attrib to the backendState in swr_update_derived. */
1182          linkedAttrib = pPrevShader->num_outputs + extraAttribs - 1;
1183          swr_fs->pointSpriteMask |= (1 << linkedAttrib);
1184          extraAttribs++;
1185       } else if (linkedAttrib == 0xFFFFFFFF) {
1186          inputs[attrib][0] = wrap(VIMMED1(0.0f));
1187          inputs[attrib][1] = wrap(VIMMED1(0.0f));
1188          inputs[attrib][2] = wrap(VIMMED1(0.0f));
1189          inputs[attrib][3] = wrap(VIMMED1(1.0f));
1190          /* If we're reading in color and 2-sided lighting is enabled, we have
1191           * to keep going.
1192           */
1193          if (semantic_name != TGSI_SEMANTIC_COLOR || !key.light_twoside)
1194             continue;
1195       } else {
1196          if (interpMode == TGSI_INTERPOLATE_CONSTANT) {
1197             swr_fs->constantMask |= 1 << linkedAttrib;
1198          } else if (interpMode == TGSI_INTERPOLATE_COLOR) {
1199             swr_fs->flatConstantMask |= 1 << linkedAttrib;
1200          }
1201       }
1202
1203       unsigned bcolorAttrib = 0xFFFFFFFF;
1204       Value *offset = NULL;
1205       if (semantic_name == TGSI_SEMANTIC_COLOR && key.light_twoside) {
1206          bcolorAttrib = locate_linkage(
1207                TGSI_SEMANTIC_BCOLOR, semantic_idx, pPrevShader) - 1;
1208          /* Neither front nor back colors were available. Nothing to load. */
1209          if (bcolorAttrib == 0xFFFFFFFF && linkedAttrib == 0xFFFFFFFF)
1210             continue;
1211          /* If there is no front color, just always use the back color. */
1212          if (linkedAttrib == 0xFFFFFFFF)
1213             linkedAttrib = bcolorAttrib;
1214
1215          if (bcolorAttrib != 0xFFFFFFFF) {
1216             if (interpMode == TGSI_INTERPOLATE_CONSTANT) {
1217                swr_fs->constantMask |= 1 << bcolorAttrib;
1218             } else if (interpMode == TGSI_INTERPOLATE_COLOR) {
1219                swr_fs->flatConstantMask |= 1 << bcolorAttrib;
1220             }
1221
1222             unsigned diff = 12 * (bcolorAttrib - linkedAttrib);
1223
1224             if (diff) {
1225                Value *back =
1226                   XOR(C(1), LOAD(pPS, {0, SWR_PS_CONTEXT_frontFace}), "backFace");
1227
1228                offset = MUL(back, C(diff));
1229                offset->setName("offset");
1230             }
1231          }
1232       }
1233
1234       for (int channel = 0; channel < TGSI_NUM_CHANNELS; channel++) {
1235          if (mask & (1 << channel)) {
1236             Value *indexA = C(linkedAttrib * 12 + channel);
1237             Value *indexB = C(linkedAttrib * 12 + channel + 4);
1238             Value *indexC = C(linkedAttrib * 12 + channel + 8);
1239
1240             if (offset) {
1241                indexA = ADD(indexA, offset);
1242                indexB = ADD(indexB, offset);
1243                indexC = ADD(indexC, offset);
1244             }
1245
1246             Value *va = VBROADCAST(LOAD(GEP(pAttribs, indexA)));
1247             Value *vb = VBROADCAST(LOAD(GEP(pAttribs, indexB)));
1248             Value *vc = VBROADCAST(LOAD(GEP(pAttribs, indexC)));
1249
1250             if (interpMode == TGSI_INTERPOLATE_CONSTANT) {
1251                inputs[attrib][channel] = wrap(va);
1252             } else {
1253                Value *vk = FSUB(FSUB(VIMMED1(1.0f), vi), vj);
1254
1255                vc = FMUL(vk, vc);
1256
1257                Value *interp = FMUL(va, vi);
1258                Value *interp1 = FMUL(vb, vj);
1259                interp = FADD(interp, interp1);
1260                interp = FADD(interp, vc);
1261                if (interpMode == TGSI_INTERPOLATE_PERSPECTIVE ||
1262                    interpMode == TGSI_INTERPOLATE_COLOR)
1263                   interp = FMUL(interp, vw);
1264                inputs[attrib][channel] = wrap(interp);
1265             }
1266          }
1267       }
1268    }
1269
1270    sampler = swr_sampler_soa_create(key.sampler, PIPE_SHADER_FRAGMENT);
1271
1272    struct lp_bld_tgsi_system_values system_values;
1273    memset(&system_values, 0, sizeof(system_values));
1274
1275    struct lp_build_mask_context mask;
1276    bool uses_mask = false;
1277
1278    if (swr_fs->info.base.uses_kill ||
1279        key.poly_stipple_enable) {
1280       Value *vActiveMask = NULL;
1281       if (swr_fs->info.base.uses_kill) {
1282          vActiveMask = LOAD(pPS, {0, SWR_PS_CONTEXT_activeMask}, "activeMask");
1283       }
1284       if (key.poly_stipple_enable) {
1285          // first get fragment xy coords and clip to stipple bounds
1286          Value *vXf = LOAD(pPS, {0, SWR_PS_CONTEXT_vX, PixelPositions_UL});
1287          Value *vYf = LOAD(pPS, {0, SWR_PS_CONTEXT_vY, PixelPositions_UL});
1288          Value *vXu = FP_TO_UI(vXf, mSimdInt32Ty);
1289          Value *vYu = FP_TO_UI(vYf, mSimdInt32Ty);
1290
1291          // stipple pattern is 32x32, which means that one line of stipple
1292          // is stored in one word:
1293          // vXstipple is bit offset inside 32-bit stipple word
1294          // vYstipple is word index is stipple array
1295          Value *vXstipple = AND(vXu, VIMMED1(0x1f)); // & (32-1)
1296          Value *vYstipple = AND(vYu, VIMMED1(0x1f)); // & (32-1)
1297
1298          // grab stipple pattern base address
1299          Value *stipplePtr = GEP(hPrivateData, {0, swr_draw_context_polyStipple, 0});
1300          stipplePtr = BITCAST(stipplePtr, mInt8PtrTy);
1301
1302          // peform a gather to grab stipple words for each lane
1303          Value *vStipple = GATHERDD(VUNDEF_I(), stipplePtr, vYstipple,
1304                                     VIMMED1(0xffffffff), 4);
1305
1306          // create a mask with one bit corresponding to the x stipple
1307          // and AND it with the pattern, to see if we have a bit
1308          Value *vBitMask = LSHR(VIMMED1(0x80000000), vXstipple);
1309          Value *vStippleMask = AND(vStipple, vBitMask);
1310          vStippleMask = ICMP_NE(vStippleMask, VIMMED1(0));
1311          vStippleMask = VMASK(vStippleMask);
1312
1313          if (swr_fs->info.base.uses_kill) {
1314             vActiveMask = AND(vActiveMask, vStippleMask);
1315          } else {
1316             vActiveMask = vStippleMask;
1317          }
1318       }
1319       lp_build_mask_begin(
1320          &mask, gallivm, lp_type_float_vec(32, 32 * 8), wrap(vActiveMask));
1321       uses_mask = true;
1322    }
1323
1324    lp_build_tgsi_soa(gallivm,
1325                      swr_fs->pipe.tokens,
1326                      lp_type_float_vec(32, 32 * 8),
1327                      uses_mask ? &mask : NULL, // mask
1328                      wrap(consts_ptr),
1329                      wrap(const_sizes_ptr),
1330                      &system_values,
1331                      inputs,
1332                      outputs,
1333                      wrap(hPrivateData),
1334                      NULL, // thread data
1335                      sampler, // sampler
1336                      &swr_fs->info.base,
1337                      NULL); // geometry shader face
1338
1339    sampler->destroy(sampler);
1340
1341    IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
1342
1343    for (uint32_t attrib = 0; attrib < swr_fs->info.base.num_outputs;
1344         attrib++) {
1345       switch (swr_fs->info.base.output_semantic_name[attrib]) {
1346       case TGSI_SEMANTIC_POSITION: {
1347          // write z
1348          LLVMValueRef outZ =
1349             LLVMBuildLoad(gallivm->builder, outputs[attrib][2], "");
1350          STORE(unwrap(outZ), pPS, {0, SWR_PS_CONTEXT_vZ});
1351          break;
1352       }
1353       case TGSI_SEMANTIC_COLOR: {
1354          for (uint32_t channel = 0; channel < TGSI_NUM_CHANNELS; channel++) {
1355             if (!outputs[attrib][channel])
1356                continue;
1357
1358             LLVMValueRef out =
1359                LLVMBuildLoad(gallivm->builder, outputs[attrib][channel], "");
1360             if (swr_fs->info.base.properties[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS] &&
1361                 swr_fs->info.base.output_semantic_index[attrib] == 0) {
1362                for (uint32_t rt = 0; rt < key.nr_cbufs; rt++) {
1363                   STORE(unwrap(out),
1364                         pPS,
1365                         {0, SWR_PS_CONTEXT_shaded, rt, channel});
1366                }
1367             } else {
1368                STORE(unwrap(out),
1369                      pPS,
1370                      {0,
1371                            SWR_PS_CONTEXT_shaded,
1372                            swr_fs->info.base.output_semantic_index[attrib],
1373                            channel});
1374             }
1375          }
1376          break;
1377       }
1378       default: {
1379          fprintf(stderr,
1380                  "unknown output from FS %s[%d]\n",
1381                  tgsi_semantic_names[swr_fs->info.base
1382                                         .output_semantic_name[attrib]],
1383                  swr_fs->info.base.output_semantic_index[attrib]);
1384          break;
1385       }
1386       }
1387    }
1388
1389    LLVMValueRef mask_result = 0;
1390    if (uses_mask) {
1391       mask_result = lp_build_mask_end(&mask);
1392    }
1393
1394    IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
1395
1396    if (uses_mask) {
1397       STORE(unwrap(mask_result), pPS, {0, SWR_PS_CONTEXT_activeMask});
1398    }
1399
1400    RET_VOID();
1401
1402    gallivm_verify_function(gallivm, wrap(pFunction));
1403
1404    gallivm_compile_module(gallivm);
1405
1406    // after the gallivm passes, we have to lower the core's intrinsics
1407    llvm::legacy::FunctionPassManager lowerPass(JM()->mpCurrentModule);
1408    lowerPass.add(createLowerX86Pass(this));
1409    lowerPass.run(*pFunction);
1410
1411    PFN_PIXEL_KERNEL kernel =
1412       (PFN_PIXEL_KERNEL)gallivm_jit_function(gallivm, wrap(pFunction));
1413    debug_printf("frag shader  %p\n", kernel);
1414    assert(kernel && "Error: FragShader = NULL");
1415
1416    JM()->mIsModuleFinalized = true;
1417
1418    return kernel;
1419 }
1420
1421 PFN_PIXEL_KERNEL
1422 swr_compile_fs(struct swr_context *ctx, swr_jit_fs_key &key)
1423 {
1424    if (!ctx->fs->pipe.tokens)
1425       return NULL;
1426
1427    BuilderSWR builder(
1428       reinterpret_cast<JitManager *>(swr_screen(ctx->pipe.screen)->hJitMgr),
1429       "FS");
1430    PFN_PIXEL_KERNEL func = builder.CompileFS(ctx, key);
1431
1432    ctx->fs->map.insert(std::make_pair(key, make_unique<VariantFS>(builder.gallivm, func)));
1433    return func;
1434 }