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_POSITION) {
 428           attribSlot = VERTEX_POSITION_SLOT;
 429        } else {
 430           attribSlot = VERTEX_ATTRIB_START_SLOT + attrib;
 431           if (iface->info->writes_position) {
 432              attribSlot--;
 433           }
 434        }
 435
 436        Value *pOutputOffset = ADD(pVertexOffset, VIMMED1(headerSize + attribSize * attribSlot)); // + sgvChannel ?
 437
 438        for (uint32_t lane = 0; lane < mVWidth; ++lane) {
 439           Value *pLaneOffset = VEXTRACT(pOutputOffset, C(lane));
 440           Value *pStream = LOAD(iface->pGsCtx, {0, SWR_GS_CONTEXT_pStreams, lane});
 441           Value *pStreamOffset = GEP(pStream, pLaneOffset);
 442           pStreamOffset = BITCAST(pStreamOffset, mFP32PtrTy);
 443
 444           Value *pLaneMask = VEXTRACT(vMask1, C(lane));
 445           pStreamOffset = SELECT(pLaneMask, pStreamOffset, pTmpPtr);
 446
 447           for (uint32_t channel = 0; channel < 4; ++channel) {
 448              Value *vData;
 449
 450              if (attribSlot == VERTEX_SGV_SLOT)
 451                 vData = LOAD(unwrap(outputs[attrib][0]));
 452              else
 453                 vData = LOAD(unwrap(outputs[attrib][channel]));
 454
 455              if (attribSlot != VERTEX_SGV_SLOT ||
 456                  sgvChannel == channel) {
 457                 vData = VEXTRACT(vData, C(lane));
 458                 STORE(vData, pStreamOffset);
 459              }
 460              pStreamOffset = GEP(pStreamOffset, C(1));
 461           }
 462        }
 463     }
 464
 465     STACKRESTORE(pStack);
 466 }
 467
 468 void
 469 BuilderSWR::swr_gs_llvm_end_primitive(const struct lp_build_tgsi_gs_iface *gs_base,
 470                              struct lp_build_tgsi_context * bld_base,
 471                              LLVMValueRef verts_per_prim_vec,
 472                              LLVMValueRef emitted_prims_vec)
 473 {
 474     swr_gs_llvm_iface *iface = (swr_gs_llvm_iface*)gs_base;
 475
 476     IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 477
 478     Value *vMask = LOAD(iface->pGsCtx, { 0, SWR_GS_CONTEXT_mask });
 479     Value *vMask1 = TRUNC(vMask, VectorType::get(mInt1Ty, 8));
 480
 481     uint32_t vertsPerPrim = iface->num_verts_per_prim;
 482
 483     Value *vCount =
 484        ADD(MUL(unwrap(emitted_prims_vec), VIMMED1(vertsPerPrim)),
 485            unwrap(verts_per_prim_vec));
 486
 487     struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
 488     vCount = LOAD(unwrap(bld->total_emitted_vertices_vec_ptr));
 489
 490     struct lp_exec_mask *exec_mask = &bld->exec_mask;
 491     Value *mask = unwrap(lp_build_mask_value(bld->mask));
 492     if (exec_mask->has_mask)
 493        mask = AND(mask, unwrap(exec_mask->exec_mask));
 494
 495     Value *cmpMask = VMASK(ICMP_NE(unwrap(verts_per_prim_vec), VIMMED1(0)));
 496     mask = AND(mask, cmpMask);
 497     vMask1 = TRUNC(mask, VectorType::get(mInt1Ty, 8));
 498
 499     vCount = SUB(vCount, VIMMED1(1));
 500     Value *vOffset = ADD(UDIV(vCount, VIMMED1(8)), VIMMED1(VERTEX_COUNT_SIZE));
 501     Value *vValue = SHL(VIMMED1(1), UREM(vCount, VIMMED1(8)));
 502
 503     vValue = TRUNC(vValue, VectorType::get(mInt8Ty, 8));
 504
 505     Value *pStack = STACKSAVE();
 506     Value *pTmpPtr = ALLOCA(mInt8Ty, C(4)); // used for dummy read/write for lane masking
 507
 508     for (uint32_t lane = 0; lane < mVWidth; ++lane) {
 509        Value *vLaneOffset = VEXTRACT(vOffset, C(lane));
 510        Value *pStream = LOAD(iface->pGsCtx, {0, SWR_GS_CONTEXT_pStreams, lane});
 511        Value *pStreamOffset = GEP(pStream, vLaneOffset);
 512
 513        Value *pLaneMask = VEXTRACT(vMask1, C(lane));
 514        pStreamOffset = SELECT(pLaneMask, pStreamOffset, pTmpPtr);
 515
 516        Value *vVal = LOAD(pStreamOffset);
 517        vVal = OR(vVal, VEXTRACT(vValue, C(lane)));
 518        STORE(vVal, pStreamOffset);
 519     }
 520
 521     STACKRESTORE(pStack);
 522 }
 523
 524 void
 525 BuilderSWR::swr_gs_llvm_epilogue(const struct lp_build_tgsi_gs_iface *gs_base,
 526                         struct lp_build_tgsi_context * bld_base,
 527                         LLVMValueRef total_emitted_vertices_vec,
 528                         LLVMValueRef emitted_prims_vec)
 529 {
 530    swr_gs_llvm_iface *iface = (swr_gs_llvm_iface*)gs_base;
 531
 532    IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 533
 534    // Store emit count to each output stream in the first DWORD
 535    for (uint32_t lane = 0; lane < mVWidth; ++lane)
 536    {
 537       Value* pStream = LOAD(iface->pGsCtx, {0, SWR_GS_CONTEXT_pStreams, lane});
 538       pStream = BITCAST(pStream, mInt32PtrTy);
 539       Value* pLaneCount = VEXTRACT(unwrap(total_emitted_vertices_vec), C(lane));
 540       STORE(pLaneCount, pStream);
 541    }
 542 }
 543
 544 PFN_GS_FUNC
 545 BuilderSWR::CompileGS(struct swr_context *ctx, swr_jit_gs_key &key)
 546 {
 547    SWR_GS_STATE *pGS = &ctx->gs->gsState;
 548    struct tgsi_shader_info *info = &ctx->gs->info.base;
 549
 550    memset(pGS, 0, sizeof(*pGS));
 551
 552    pGS->gsEnable = true;
 553
 554    pGS->numInputAttribs = info->num_inputs;
 555    pGS->outputTopology =
 556       swr_convert_prim_topology(info->properties[TGSI_PROPERTY_GS_OUTPUT_PRIM]);
 557    pGS->maxNumVerts = info->properties[TGSI_PROPERTY_GS_MAX_OUTPUT_VERTICES];
 558    pGS->instanceCount = info->properties[TGSI_PROPERTY_GS_INVOCATIONS];
 559
 560    // XXX: single stream for now...
 561    pGS->isSingleStream = true;
 562    pGS->singleStreamID = 0;
 563
 564    pGS->vertexAttribOffset = VERTEX_ATTRIB_START_SLOT; // TODO: optimize
 565    pGS->srcVertexAttribOffset = VERTEX_ATTRIB_START_SLOT; // TODO: optimize
 566    pGS->inputVertStride = pGS->numInputAttribs + pGS->vertexAttribOffset;
 567    pGS->outputVertexSize = SWR_VTX_NUM_SLOTS;
 568    pGS->controlDataSize = 8; // GS ouputs max of 8 32B units
 569    pGS->controlDataOffset = VERTEX_COUNT_SIZE;
 570    pGS->outputVertexOffset = pGS->controlDataOffset + CONTROL_HEADER_SIZE;
 571
 572    pGS->allocationSize =
 573       VERTEX_COUNT_SIZE + // vertex count
 574       CONTROL_HEADER_SIZE + // control header
 575       (SWR_VTX_NUM_SLOTS * 16) * // sizeof vertex
 576       pGS->maxNumVerts; // num verts
 577
 578    struct swr_geometry_shader *gs = ctx->gs;
 579
 580    LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
 581    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
 582
 583    memset(outputs, 0, sizeof(outputs));
 584
 585    AttrBuilder attrBuilder;
 586    attrBuilder.addStackAlignmentAttr(JM()->mVWidth * sizeof(float));
 587
 588    std::vector<Type *> gsArgs{PointerType::get(Gen_swr_draw_context(JM()), 0),
 589                               PointerType::get(Gen_SWR_GS_CONTEXT(JM()), 0)};
 590    FunctionType *vsFuncType =
 591       FunctionType::get(Type::getVoidTy(JM()->mContext), gsArgs, false);
 592
 593    // create new vertex shader function
 594    auto pFunction = Function::Create(vsFuncType,
 595                                      GlobalValue::ExternalLinkage,
 596                                      "GS",
 597                                      JM()->mpCurrentModule);
 598 #if HAVE_LLVM < 0x0500
 599    AttributeSet attrSet = AttributeSet::get(
 600       JM()->mContext, AttributeSet::FunctionIndex, attrBuilder);
 601    pFunction->addAttributes(AttributeSet::FunctionIndex, attrSet);
 602 #else
 603    pFunction->addAttributes(AttributeList::FunctionIndex, attrBuilder);
 604 #endif
 605
 606    BasicBlock *block = BasicBlock::Create(JM()->mContext, "entry", pFunction);
 607    IRB()->SetInsertPoint(block);
 608    LLVMPositionBuilderAtEnd(gallivm->builder, wrap(block));
 609
 610    auto argitr = pFunction->arg_begin();
 611    Value *hPrivateData = &*argitr++;
 612    hPrivateData->setName("hPrivateData");
 613    Value *pGsCtx = &*argitr++;
 614    pGsCtx->setName("gsCtx");
 615
 616    Value *consts_ptr =
 617       GEP(hPrivateData, {C(0), C(swr_draw_context_constantGS)});
 618    consts_ptr->setName("gs_constants");
 619    Value *const_sizes_ptr =
 620       GEP(hPrivateData, {0, swr_draw_context_num_constantsGS});
 621    const_sizes_ptr->setName("num_gs_constants");
 622
 623    struct lp_build_sampler_soa *sampler =
 624       swr_sampler_soa_create(key.sampler, PIPE_SHADER_GEOMETRY);
 625
 626    struct lp_bld_tgsi_system_values system_values;
 627    memset(&system_values, 0, sizeof(system_values));
 628    system_values.prim_id = wrap(LOAD(pGsCtx, {0, SWR_GS_CONTEXT_PrimitiveID}));
 629    system_values.instance_id = wrap(LOAD(pGsCtx, {0, SWR_GS_CONTEXT_InstanceID}));
 630
 631    std::vector<Constant*> mapConstants;
 632    Value *vtxAttribMap = ALLOCA(ArrayType::get(mInt32Ty, PIPE_MAX_SHADER_INPUTS));
 633    for (unsigned slot = 0; slot < info->num_inputs; slot++) {
 634       ubyte semantic_name = info->input_semantic_name[slot];
 635       ubyte semantic_idx = info->input_semantic_index[slot];
 636
 637       unsigned vs_slot = locate_linkage(semantic_name, semantic_idx, &ctx->vs->info.base);
 638
 639       vs_slot += VERTEX_ATTRIB_START_SLOT;
 640
 641       if (ctx->vs->info.base.output_semantic_name[0] == TGSI_SEMANTIC_POSITION)
 642          vs_slot--;
 643
 644       if (semantic_name == TGSI_SEMANTIC_POSITION)
 645          vs_slot = VERTEX_POSITION_SLOT;
 646
 647       STORE(C(vs_slot), vtxAttribMap, {0, slot});
 648       mapConstants.push_back(C(vs_slot));
 649    }
 650
 651    struct lp_build_mask_context mask;
 652    Value *mask_val = LOAD(pGsCtx, {0, SWR_GS_CONTEXT_mask}, "gsMask");
 653    lp_build_mask_begin(&mask, gallivm,
 654                        lp_type_float_vec(32, 32 * 8), wrap(mask_val));
 655
 656    // zero out cut buffer so we can load/modify/store bits
 657    for (uint32_t lane = 0; lane < mVWidth; ++lane)
 658    {
 659       Value* pStream = LOAD(pGsCtx, {0, SWR_GS_CONTEXT_pStreams, lane});
 660       MEMSET(pStream, C((char)0), VERTEX_COUNT_SIZE + CONTROL_HEADER_SIZE, sizeof(float) * KNOB_SIMD_WIDTH);
 661    }
 662
 663    struct swr_gs_llvm_iface gs_iface;
 664    gs_iface.base.fetch_input = ::swr_gs_llvm_fetch_input;
 665    gs_iface.base.emit_vertex = ::swr_gs_llvm_emit_vertex;
 666    gs_iface.base.end_primitive = ::swr_gs_llvm_end_primitive;
 667    gs_iface.base.gs_epilogue = ::swr_gs_llvm_epilogue;
 668    gs_iface.pBuilder = this;
 669    gs_iface.pGsCtx = pGsCtx;
 670    gs_iface.pGsState = pGS;
 671    gs_iface.num_outputs = gs->info.base.num_outputs;
 672    gs_iface.num_verts_per_prim =
 673       u_vertices_per_prim((pipe_prim_type)info->properties[TGSI_PROPERTY_GS_OUTPUT_PRIM]);
 674    gs_iface.info = info;
 675    gs_iface.pVtxAttribMap = vtxAttribMap;
 676
 677    lp_build_tgsi_soa(gallivm,
 678                      gs->pipe.tokens,
 679                      lp_type_float_vec(32, 32 * 8),
 680                      &mask,
 681                      wrap(consts_ptr),
 682                      wrap(const_sizes_ptr),
 683                      &system_values,
 684                      inputs,
 685                      outputs,
 686                      wrap(hPrivateData), // (sampler context)
 687                      NULL, // thread data
 688                      sampler,
 689                      &gs->info.base,
 690                      &gs_iface.base);
 691
 692    lp_build_mask_end(&mask);
 693
 694    sampler->destroy(sampler);
 695
 696    IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 697
 698    RET_VOID();
 699
 700    gallivm_verify_function(gallivm, wrap(pFunction));
 701    gallivm_compile_module(gallivm);
 702
 703    PFN_GS_FUNC pFunc =
 704       (PFN_GS_FUNC)gallivm_jit_function(gallivm, wrap(pFunction));
 705
 706    debug_printf("geom shader  %p\n", pFunc);
 707    assert(pFunc && "Error: GeomShader = NULL");
 708
 709    JM()->mIsModuleFinalized = true;
 710
 711    return pFunc;
 712 }
 713
 714 PFN_GS_FUNC
 715 swr_compile_gs(struct swr_context *ctx, swr_jit_gs_key &key)
 716 {
 717    BuilderSWR builder(
 718       reinterpret_cast<JitManager *>(swr_screen(ctx->pipe.screen)->hJitMgr),
 719       "GS");
 720    PFN_GS_FUNC func = builder.CompileGS(ctx, key);
 721
 722    ctx->gs->map.insert(std::make_pair(key, make_unique<VariantGS>(builder.gallivm, func)));
 723    return func;
 724 }
 725
 726 void
 727 BuilderSWR::WriteVS(Value *pVal, Value *pVsContext, Value *pVtxOutput, unsigned slot, unsigned channel)
 728 {
 729 #if USE_SIMD16_FRONTEND && !USE_SIMD16_VS
 730    // interleave the simdvertex components into the dest simd16vertex
 731    //   slot16offset = slot8offset * 2
 732    //   comp16offset = comp8offset * 2 + alternateOffset
 733
 734    Value *offset = LOAD(pVsContext, { 0, SWR_VS_CONTEXT_AlternateOffset });
 735    Value *pOut = GEP(pVtxOutput, { C(0), C(0), C(slot * 2), offset } );
 736    STORE(pVal, pOut, {channel * 2});
 737 #else
 738    Value *pOut = GEP(pVtxOutput, {0, 0, slot});
 739    STORE(pVal, pOut, {0, channel});
 740 #endif
 741 }
 742
 743 PFN_VERTEX_FUNC
 744 BuilderSWR::CompileVS(struct swr_context *ctx, swr_jit_vs_key &key)
 745 {
 746    struct swr_vertex_shader *swr_vs = ctx->vs;
 747
 748    LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
 749    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
 750
 751    memset(outputs, 0, sizeof(outputs));
 752
 753    AttrBuilder attrBuilder;
 754    attrBuilder.addStackAlignmentAttr(JM()->mVWidth * sizeof(float));
 755
 756    std::vector<Type *> vsArgs{PointerType::get(Gen_swr_draw_context(JM()), 0),
 757                               PointerType::get(Gen_SWR_VS_CONTEXT(JM()), 0)};
 758    FunctionType *vsFuncType =
 759       FunctionType::get(Type::getVoidTy(JM()->mContext), vsArgs, false);
 760
 761    // create new vertex shader function
 762    auto pFunction = Function::Create(vsFuncType,
 763                                      GlobalValue::ExternalLinkage,
 764                                      "VS",
 765                                      JM()->mpCurrentModule);
 766 #if HAVE_LLVM < 0x0500
 767    AttributeSet attrSet = AttributeSet::get(
 768       JM()->mContext, AttributeSet::FunctionIndex, attrBuilder);
 769    pFunction->addAttributes(AttributeSet::FunctionIndex, attrSet);
 770 #else
 771    pFunction->addAttributes(AttributeList::FunctionIndex, attrBuilder);
 772 #endif
 773
 774    BasicBlock *block = BasicBlock::Create(JM()->mContext, "entry", pFunction);
 775    IRB()->SetInsertPoint(block);
 776    LLVMPositionBuilderAtEnd(gallivm->builder, wrap(block));
 777
 778    auto argitr = pFunction->arg_begin();
 779    Value *hPrivateData = &*argitr++;
 780    hPrivateData->setName("hPrivateData");
 781    Value *pVsCtx = &*argitr++;
 782    pVsCtx->setName("vsCtx");
 783
 784    Value *consts_ptr = GEP(hPrivateData, {C(0), C(swr_draw_context_constantVS)});
 785
 786    consts_ptr->setName("vs_constants");
 787    Value *const_sizes_ptr =
 788       GEP(hPrivateData, {0, swr_draw_context_num_constantsVS});
 789    const_sizes_ptr->setName("num_vs_constants");
 790
 791    Value *vtxInput = LOAD(pVsCtx, {0, SWR_VS_CONTEXT_pVin});
 792 #if USE_SIMD16_VS
 793    vtxInput = BITCAST(vtxInput, PointerType::get(Gen_simd16vertex(JM()), 0));
 794 #endif
 795
 796    for (uint32_t attrib = 0; attrib < PIPE_MAX_SHADER_INPUTS; attrib++) {
 797       const unsigned mask = swr_vs->info.base.input_usage_mask[attrib];
 798       for (uint32_t channel = 0; channel < TGSI_NUM_CHANNELS; channel++) {
 799          if (mask & (1 << channel)) {
 800             inputs[attrib][channel] =
 801                wrap(LOAD(vtxInput, {0, 0, attrib, channel}));
 802          }
 803       }
 804    }
 805
 806    struct lp_build_sampler_soa *sampler =
 807       swr_sampler_soa_create(key.sampler, PIPE_SHADER_VERTEX);
 808
 809    struct lp_bld_tgsi_system_values system_values;
 810    memset(&system_values, 0, sizeof(system_values));
 811    system_values.instance_id = wrap(LOAD(pVsCtx, {0, SWR_VS_CONTEXT_InstanceID}));
 812
 813 #if USE_SIMD16_VS
 814    system_values.vertex_id = wrap(LOAD(pVsCtx, {0, SWR_VS_CONTEXT_VertexID16}));
 815 #else
 816    system_values.vertex_id = wrap(LOAD(pVsCtx, {0, SWR_VS_CONTEXT_VertexID}));
 817 #endif
 818
 819 #if USE_SIMD16_VS
 820    uint32_t vectorWidth = mVWidth16;
 821 #else
 822    uint32_t vectorWidth = mVWidth;
 823 #endif
 824
 825    lp_build_tgsi_soa(gallivm,
 826                      swr_vs->pipe.tokens,
 827                      lp_type_float_vec(32, 32 * vectorWidth),
 828                      NULL, // mask
 829                      wrap(consts_ptr),
 830                      wrap(const_sizes_ptr),
 831                      &system_values,
 832                      inputs,
 833                      outputs,
 834                      wrap(hPrivateData), // (sampler context)
 835                      NULL, // thread data
 836                      sampler, // sampler
 837                      &swr_vs->info.base,
 838                      NULL); // geometry shader face
 839
 840    sampler->destroy(sampler);
 841
 842    IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 843
 844    Value *vtxOutput = LOAD(pVsCtx, {0, SWR_VS_CONTEXT_pVout});
 845 #if USE_SIMD16_VS
 846    vtxOutput = BITCAST(vtxOutput, PointerType::get(Gen_simd16vertex(JM()), 0));
 847 #endif
 848
 849    for (uint32_t channel = 0; channel < TGSI_NUM_CHANNELS; channel++) {
 850       for (uint32_t attrib = 0; attrib < PIPE_MAX_SHADER_OUTPUTS; attrib++) {
 851          if (!outputs[attrib][channel])
 852             continue;
 853
 854          Value *val;
 855          uint32_t outSlot;
 856
 857          if (swr_vs->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_PSIZE) {
 858             if (channel != VERTEX_SGV_POINT_SIZE_COMP)
 859                continue;
 860             val = LOAD(unwrap(outputs[attrib][0]));
 861             outSlot = VERTEX_SGV_SLOT;
 862          } else if (swr_vs->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_POSITION) {
 863             val = LOAD(unwrap(outputs[attrib][channel]));
 864             outSlot = VERTEX_POSITION_SLOT;
 865          } else {
 866             val = LOAD(unwrap(outputs[attrib][channel]));
 867             outSlot = VERTEX_ATTRIB_START_SLOT + attrib;
 868             if (swr_vs->info.base.output_semantic_name[0] == TGSI_SEMANTIC_POSITION)
 869                outSlot--;
 870          }
 871
 872          WriteVS(val, pVsCtx, vtxOutput, outSlot, channel);
 873       }
 874    }
 875
 876    if (ctx->rasterizer->clip_plane_enable ||
 877        swr_vs->info.base.culldist_writemask) {
 878       unsigned clip_mask = ctx->rasterizer->clip_plane_enable;
 879
 880       unsigned cv = 0;
 881       if (swr_vs->info.base.writes_clipvertex) {
 882          cv = locate_linkage(TGSI_SEMANTIC_CLIPVERTEX, 0,
 883                              &swr_vs->info.base);
 884       } else {
 885          for (int i = 0; i < PIPE_MAX_SHADER_OUTPUTS; i++) {
 886             if (swr_vs->info.base.output_semantic_name[i] == TGSI_SEMANTIC_POSITION &&
 887                 swr_vs->info.base.output_semantic_index[i] == 0) {
 888                cv = i;
 889                break;
 890             }
 891          }
 892       }
 893       LLVMValueRef cx = LLVMBuildLoad(gallivm->builder, outputs[cv][0], "");
 894       LLVMValueRef cy = LLVMBuildLoad(gallivm->builder, outputs[cv][1], "");
 895       LLVMValueRef cz = LLVMBuildLoad(gallivm->builder, outputs[cv][2], "");
 896       LLVMValueRef cw = LLVMBuildLoad(gallivm->builder, outputs[cv][3], "");
 897
 898       for (unsigned val = 0; val < PIPE_MAX_CLIP_PLANES; val++) {
 899          // clip distance overrides user clip planes
 900          if ((swr_vs->info.base.clipdist_writemask & clip_mask & (1 << val)) ||
 901              ((swr_vs->info.base.culldist_writemask << swr_vs->info.base.num_written_clipdistance) & (1 << val))) {
 902             unsigned cv = locate_linkage(TGSI_SEMANTIC_CLIPDIST, val < 4 ? 0 : 1,
 903                                          &swr_vs->info.base);
 904             if (val < 4) {
 905                LLVMValueRef dist = LLVMBuildLoad(gallivm->builder, outputs[cv][val], "");
 906                WriteVS(unwrap(dist), pVsCtx, vtxOutput, VERTEX_CLIPCULL_DIST_LO_SLOT, val);
 907             } else {
 908                LLVMValueRef dist = LLVMBuildLoad(gallivm->builder, outputs[cv][val - 4], "");
 909                WriteVS(unwrap(dist), pVsCtx, vtxOutput, VERTEX_CLIPCULL_DIST_HI_SLOT, val - 4);
 910             }
 911             continue;
 912          }
 913
 914          if (!(clip_mask & (1 << val)))
 915             continue;
 916
 917          Value *px = LOAD(GEP(hPrivateData, {0, swr_draw_context_userClipPlanes, val, 0}));
 918          Value *py = LOAD(GEP(hPrivateData, {0, swr_draw_context_userClipPlanes, val, 1}));
 919          Value *pz = LOAD(GEP(hPrivateData, {0, swr_draw_context_userClipPlanes, val, 2}));
 920          Value *pw = LOAD(GEP(hPrivateData, {0, swr_draw_context_userClipPlanes, val, 3}));
 921 #if USE_SIMD16_VS
 922          Value *bpx = VBROADCAST_16(px);
 923          Value *bpy = VBROADCAST_16(py);
 924          Value *bpz = VBROADCAST_16(pz);
 925          Value *bpw = VBROADCAST_16(pw);
 926 #else
 927          Value *bpx = VBROADCAST(px);
 928          Value *bpy = VBROADCAST(py);
 929          Value *bpz = VBROADCAST(pz);
 930          Value *bpw = VBROADCAST(pw);
 931 #endif
 932          Value *dist = FADD(FMUL(unwrap(cx), bpx),
 933                             FADD(FMUL(unwrap(cy), bpy),
 934                                  FADD(FMUL(unwrap(cz), bpz),
 935                                       FMUL(unwrap(cw), bpw))));
 936
 937          if (val < 4)
 938             WriteVS(dist, pVsCtx, vtxOutput, VERTEX_CLIPCULL_DIST_LO_SLOT, val);
 939          else
 940             WriteVS(dist, pVsCtx, vtxOutput, VERTEX_CLIPCULL_DIST_HI_SLOT, val - 4);
 941       }
 942    }
 943
 944    RET_VOID();
 945
 946    gallivm_verify_function(gallivm, wrap(pFunction));
 947    gallivm_compile_module(gallivm);
 948
 949    //   lp_debug_dump_value(func);
 950
 951    PFN_VERTEX_FUNC pFunc =
 952       (PFN_VERTEX_FUNC)gallivm_jit_function(gallivm, wrap(pFunction));
 953
 954    debug_printf("vert shader  %p\n", pFunc);
 955    assert(pFunc && "Error: VertShader = NULL");
 956
 957    JM()->mIsModuleFinalized = true;
 958
 959    return pFunc;
 960 }
 961
 962 PFN_VERTEX_FUNC
 963 swr_compile_vs(struct swr_context *ctx, swr_jit_vs_key &key)
 964 {
 965    if (!ctx->vs->pipe.tokens)
 966       return NULL;
 967
 968    BuilderSWR builder(
 969       reinterpret_cast<JitManager *>(swr_screen(ctx->pipe.screen)->hJitMgr),
 970       "VS");
 971    PFN_VERTEX_FUNC func = builder.CompileVS(ctx, key);
 972
 973    ctx->vs->map.insert(std::make_pair(key, make_unique<VariantVS>(builder.gallivm, func)));
 974    return func;
 975 }
 976
 977 unsigned
 978 swr_so_adjust_attrib(unsigned in_attrib,
 979                      swr_vertex_shader *swr_vs)
 980 {
 981    ubyte semantic_name;
 982    unsigned attrib;
 983
 984    attrib = in_attrib + VERTEX_ATTRIB_START_SLOT;
 985
 986    if (swr_vs) {
 987       semantic_name = swr_vs->info.base.output_semantic_name[in_attrib];
 988       if (semantic_name == TGSI_SEMANTIC_POSITION) {
 989          attrib = VERTEX_POSITION_SLOT;
 990       } else if (semantic_name == TGSI_SEMANTIC_PSIZE) {
 991          attrib = VERTEX_SGV_SLOT;
 992       } else if (semantic_name == TGSI_SEMANTIC_LAYER) {
 993          attrib = VERTEX_SGV_SLOT;
 994       } else {
 995          if (swr_vs->info.base.writes_position) {
 996                attrib--;
 997          }
 998       }
 999    }
1000
1001    return attrib;
1002 }
1003
1004 static unsigned
1005 locate_linkage(ubyte name, ubyte index, struct tgsi_shader_info *info)
1006 {
1007    for (int i = 0; i < PIPE_MAX_SHADER_OUTPUTS; i++) {
1008       if ((info->output_semantic_name[i] == name)
1009           && (info->output_semantic_index[i] == index)) {
1010          return i;
1011       }
1012    }
1013
1014    return 0xFFFFFFFF;
1015 }
1016
1017 PFN_PIXEL_KERNEL
1018 BuilderSWR::CompileFS(struct swr_context *ctx, swr_jit_fs_key &key)
1019 {
1020    struct swr_fragment_shader *swr_fs = ctx->fs;
1021
1022    struct tgsi_shader_info *pPrevShader;
1023    if (ctx->gs)
1024       pPrevShader = &ctx->gs->info.base;
1025    else
1026       pPrevShader = &ctx->vs->info.base;
1027
1028    LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
1029    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
1030
1031    memset(inputs, 0, sizeof(inputs));
1032    memset(outputs, 0, sizeof(outputs));
1033
1034    struct lp_build_sampler_soa *sampler = NULL;
1035
1036    AttrBuilder attrBuilder;
1037    attrBuilder.addStackAlignmentAttr(JM()->mVWidth * sizeof(float));
1038
1039    std::vector<Type *> fsArgs{PointerType::get(Gen_swr_draw_context(JM()), 0),
1040                               PointerType::get(Gen_SWR_PS_CONTEXT(JM()), 0)};
1041    FunctionType *funcType =
1042       FunctionType::get(Type::getVoidTy(JM()->mContext), fsArgs, false);
1043
1044    auto pFunction = Function::Create(funcType,
1045                                      GlobalValue::ExternalLinkage,
1046                                      "FS",
1047                                      JM()->mpCurrentModule);
1048 #if HAVE_LLVM < 0x0500
1049    AttributeSet attrSet = AttributeSet::get(
1050       JM()->mContext, AttributeSet::FunctionIndex, attrBuilder);
1051    pFunction->addAttributes(AttributeSet::FunctionIndex, attrSet);
1052 #else
1053    pFunction->addAttributes(AttributeList::FunctionIndex, attrBuilder);
1054 #endif
1055
1056    BasicBlock *block = BasicBlock::Create(JM()->mContext, "entry", pFunction);
1057    IRB()->SetInsertPoint(block);
1058    LLVMPositionBuilderAtEnd(gallivm->builder, wrap(block));
1059
1060    auto args = pFunction->arg_begin();
1061    Value *hPrivateData = &*args++;
1062    hPrivateData->setName("hPrivateData");
1063    Value *pPS = &*args++;
1064    pPS->setName("psCtx");
1065
1066    Value *consts_ptr = GEP(hPrivateData, {0, swr_draw_context_constantFS});
1067    consts_ptr->setName("fs_constants");
1068    Value *const_sizes_ptr =
1069       GEP(hPrivateData, {0, swr_draw_context_num_constantsFS});
1070    const_sizes_ptr->setName("num_fs_constants");
1071
1072    // load *pAttribs, *pPerspAttribs
1073    Value *pRawAttribs = LOAD(pPS, {0, SWR_PS_CONTEXT_pAttribs}, "pRawAttribs");
1074    Value *pPerspAttribs =
1075       LOAD(pPS, {0, SWR_PS_CONTEXT_pPerspAttribs}, "pPerspAttribs");
1076
1077    swr_fs->constantMask = 0;
1078    swr_fs->flatConstantMask = 0;
1079    swr_fs->pointSpriteMask = 0;
1080
1081    for (int attrib = 0; attrib < PIPE_MAX_SHADER_INPUTS; attrib++) {
1082       const unsigned mask = swr_fs->info.base.input_usage_mask[attrib];
1083       const unsigned interpMode = swr_fs->info.base.input_interpolate[attrib];
1084       const unsigned interpLoc = swr_fs->info.base.input_interpolate_loc[attrib];
1085
1086       if (!mask)
1087          continue;
1088
1089       // load i,j
1090       Value *vi = nullptr, *vj = nullptr;
1091       switch (interpLoc) {
1092       case TGSI_INTERPOLATE_LOC_CENTER:
1093          vi = LOAD(pPS, {0, SWR_PS_CONTEXT_vI, PixelPositions_center}, "i");
1094          vj = LOAD(pPS, {0, SWR_PS_CONTEXT_vJ, PixelPositions_center}, "j");
1095          break;
1096       case TGSI_INTERPOLATE_LOC_CENTROID:
1097          vi = LOAD(pPS, {0, SWR_PS_CONTEXT_vI, PixelPositions_centroid}, "i");
1098          vj = LOAD(pPS, {0, SWR_PS_CONTEXT_vJ, PixelPositions_centroid}, "j");
1099          break;
1100       case TGSI_INTERPOLATE_LOC_SAMPLE:
1101          vi = LOAD(pPS, {0, SWR_PS_CONTEXT_vI, PixelPositions_sample}, "i");
1102          vj = LOAD(pPS, {0, SWR_PS_CONTEXT_vJ, PixelPositions_sample}, "j");
1103          break;
1104       }
1105
1106       // load/compute w
1107       Value *vw = nullptr, *pAttribs;
1108       if (interpMode == TGSI_INTERPOLATE_PERSPECTIVE ||
1109           interpMode == TGSI_INTERPOLATE_COLOR) {
1110          pAttribs = pPerspAttribs;
1111          switch (interpLoc) {
1112          case TGSI_INTERPOLATE_LOC_CENTER:
1113             vw = VRCP(LOAD(pPS, {0, SWR_PS_CONTEXT_vOneOverW, PixelPositions_center}));
1114             break;
1115          case TGSI_INTERPOLATE_LOC_CENTROID:
1116             vw = VRCP(LOAD(pPS, {0, SWR_PS_CONTEXT_vOneOverW, PixelPositions_centroid}));
1117             break;
1118          case TGSI_INTERPOLATE_LOC_SAMPLE:
1119             vw = VRCP(LOAD(pPS, {0, SWR_PS_CONTEXT_vOneOverW, PixelPositions_sample}));
1120             break;
1121          }
1122       } else {
1123          pAttribs = pRawAttribs;
1124          vw = VIMMED1(1.f);
1125       }
1126
1127       vw->setName("w");
1128
1129       ubyte semantic_name = swr_fs->info.base.input_semantic_name[attrib];
1130       ubyte semantic_idx = swr_fs->info.base.input_semantic_index[attrib];
1131
1132       if (semantic_name == TGSI_SEMANTIC_FACE) {
1133          Value *ff =
1134             UI_TO_FP(LOAD(pPS, {0, SWR_PS_CONTEXT_frontFace}), mFP32Ty);
1135          ff = FSUB(FMUL(ff, C(2.0f)), C(1.0f));
1136          ff = VECTOR_SPLAT(JM()->mVWidth, ff, "vFrontFace");
1137
1138          inputs[attrib][0] = wrap(ff);
1139          inputs[attrib][1] = wrap(VIMMED1(0.0f));
1140          inputs[attrib][2] = wrap(VIMMED1(0.0f));
1141          inputs[attrib][3] = wrap(VIMMED1(1.0f));
1142          continue;
1143       } else if (semantic_name == TGSI_SEMANTIC_POSITION) { // gl_FragCoord
1144          if (swr_fs->info.base.properties[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER] ==
1145              TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER) {
1146             inputs[attrib][0] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vX, PixelPositions_center}, "vX"));
1147             inputs[attrib][1] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vY, PixelPositions_center}, "vY"));
1148          } else {
1149             inputs[attrib][0] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vX, PixelPositions_UL}, "vX"));
1150             inputs[attrib][1] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vY, PixelPositions_UL}, "vY"));
1151          }
1152          inputs[attrib][2] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vZ}, "vZ"));
1153          inputs[attrib][3] =
1154             wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vOneOverW, PixelPositions_center}, "vOneOverW"));
1155          continue;
1156       }
1157
1158       unsigned linkedAttrib =
1159          locate_linkage(semantic_name, semantic_idx, pPrevShader) - 1;
1160
1161       uint32_t extraAttribs = 0;
1162       if (semantic_name == TGSI_SEMANTIC_PRIMID && !ctx->gs) {
1163          /* non-gs generated primID - need to grab from swizzleMap override */
1164          linkedAttrib = pPrevShader->num_outputs - 1;
1165          swr_fs->constantMask |= 1 << linkedAttrib;
1166          extraAttribs++;
1167       } else if (semantic_name == TGSI_SEMANTIC_GENERIC &&
1168           key.sprite_coord_enable & (1 << semantic_idx)) {
1169          /* we add an extra attrib to the backendState in swr_update_derived. */
1170          linkedAttrib = pPrevShader->num_outputs + extraAttribs - 1;
1171          swr_fs->pointSpriteMask |= (1 << linkedAttrib);
1172          extraAttribs++;
1173       } else if (linkedAttrib == 0xFFFFFFFF) {
1174          inputs[attrib][0] = wrap(VIMMED1(0.0f));
1175          inputs[attrib][1] = wrap(VIMMED1(0.0f));
1176          inputs[attrib][2] = wrap(VIMMED1(0.0f));
1177          inputs[attrib][3] = wrap(VIMMED1(1.0f));
1178          /* If we're reading in color and 2-sided lighting is enabled, we have
1179           * to keep going.
1180           */
1181          if (semantic_name != TGSI_SEMANTIC_COLOR || !key.light_twoside)
1182             continue;
1183       } else {
1184          if (interpMode == TGSI_INTERPOLATE_CONSTANT) {
1185             swr_fs->constantMask |= 1 << linkedAttrib;
1186          } else if (interpMode == TGSI_INTERPOLATE_COLOR) {
1187             swr_fs->flatConstantMask |= 1 << linkedAttrib;
1188          }
1189       }
1190
1191       unsigned bcolorAttrib = 0xFFFFFFFF;
1192       Value *offset = NULL;
1193       if (semantic_name == TGSI_SEMANTIC_COLOR && key.light_twoside) {
1194          bcolorAttrib = locate_linkage(
1195                TGSI_SEMANTIC_BCOLOR, semantic_idx, pPrevShader) - 1;
1196          /* Neither front nor back colors were available. Nothing to load. */
1197          if (bcolorAttrib == 0xFFFFFFFF && linkedAttrib == 0xFFFFFFFF)
1198             continue;
1199          /* If there is no front color, just always use the back color. */
1200          if (linkedAttrib == 0xFFFFFFFF)
1201             linkedAttrib = bcolorAttrib;
1202
1203          if (bcolorAttrib != 0xFFFFFFFF) {
1204             if (interpMode == TGSI_INTERPOLATE_CONSTANT) {
1205                swr_fs->constantMask |= 1 << bcolorAttrib;
1206             } else if (interpMode == TGSI_INTERPOLATE_COLOR) {
1207                swr_fs->flatConstantMask |= 1 << bcolorAttrib;
1208             }
1209
1210             unsigned diff = 12 * (bcolorAttrib - linkedAttrib);
1211
1212             if (diff) {
1213                Value *back =
1214                   XOR(C(1), LOAD(pPS, {0, SWR_PS_CONTEXT_frontFace}), "backFace");
1215
1216                offset = MUL(back, C(diff));
1217                offset->setName("offset");
1218             }
1219          }
1220       }
1221
1222       for (int channel = 0; channel < TGSI_NUM_CHANNELS; channel++) {
1223          if (mask & (1 << channel)) {
1224             Value *indexA = C(linkedAttrib * 12 + channel);
1225             Value *indexB = C(linkedAttrib * 12 + channel + 4);
1226             Value *indexC = C(linkedAttrib * 12 + channel + 8);
1227
1228             if (offset) {
1229                indexA = ADD(indexA, offset);
1230                indexB = ADD(indexB, offset);
1231                indexC = ADD(indexC, offset);
1232             }
1233
1234             Value *va = VBROADCAST(LOAD(GEP(pAttribs, indexA)));
1235             Value *vb = VBROADCAST(LOAD(GEP(pAttribs, indexB)));
1236             Value *vc = VBROADCAST(LOAD(GEP(pAttribs, indexC)));
1237
1238             if (interpMode == TGSI_INTERPOLATE_CONSTANT) {
1239                inputs[attrib][channel] = wrap(va);
1240             } else {
1241                Value *vk = FSUB(FSUB(VIMMED1(1.0f), vi), vj);
1242
1243                vc = FMUL(vk, vc);
1244
1245                Value *interp = FMUL(va, vi);
1246                Value *interp1 = FMUL(vb, vj);
1247                interp = FADD(interp, interp1);
1248                interp = FADD(interp, vc);
1249                if (interpMode == TGSI_INTERPOLATE_PERSPECTIVE ||
1250                    interpMode == TGSI_INTERPOLATE_COLOR)
1251                   interp = FMUL(interp, vw);
1252                inputs[attrib][channel] = wrap(interp);
1253             }
1254          }
1255       }
1256    }
1257
1258    sampler = swr_sampler_soa_create(key.sampler, PIPE_SHADER_FRAGMENT);
1259
1260    struct lp_bld_tgsi_system_values system_values;
1261    memset(&system_values, 0, sizeof(system_values));
1262
1263    struct lp_build_mask_context mask;
1264    bool uses_mask = false;
1265
1266    if (swr_fs->info.base.uses_kill ||
1267        key.poly_stipple_enable) {
1268       Value *vActiveMask = NULL;
1269       if (swr_fs->info.base.uses_kill) {
1270          vActiveMask = LOAD(pPS, {0, SWR_PS_CONTEXT_activeMask}, "activeMask");
1271       }
1272       if (key.poly_stipple_enable) {
1273          // first get fragment xy coords and clip to stipple bounds
1274          Value *vXf = LOAD(pPS, {0, SWR_PS_CONTEXT_vX, PixelPositions_UL});
1275          Value *vYf = LOAD(pPS, {0, SWR_PS_CONTEXT_vY, PixelPositions_UL});
1276          Value *vXu = FP_TO_UI(vXf, mSimdInt32Ty);
1277          Value *vYu = FP_TO_UI(vYf, mSimdInt32Ty);
1278
1279          // stipple pattern is 32x32, which means that one line of stipple
1280          // is stored in one word:
1281          // vXstipple is bit offset inside 32-bit stipple word
1282          // vYstipple is word index is stipple array
1283          Value *vXstipple = AND(vXu, VIMMED1(0x1f)); // & (32-1)
1284          Value *vYstipple = AND(vYu, VIMMED1(0x1f)); // & (32-1)
1285
1286          // grab stipple pattern base address
1287          Value *stipplePtr = GEP(hPrivateData, {0, swr_draw_context_polyStipple, 0});
1288          stipplePtr = BITCAST(stipplePtr, mInt8PtrTy);
1289
1290          // peform a gather to grab stipple words for each lane
1291          Value *vStipple = GATHERDD(VUNDEF_I(), stipplePtr, vYstipple,
1292                                     VIMMED1(0xffffffff), 4);
1293
1294          // create a mask with one bit corresponding to the x stipple
1295          // and AND it with the pattern, to see if we have a bit
1296          Value *vBitMask = LSHR(VIMMED1(0x80000000), vXstipple);
1297          Value *vStippleMask = AND(vStipple, vBitMask);
1298          vStippleMask = ICMP_NE(vStippleMask, VIMMED1(0));
1299          vStippleMask = VMASK(vStippleMask);
1300
1301          if (swr_fs->info.base.uses_kill) {
1302             vActiveMask = AND(vActiveMask, vStippleMask);
1303          } else {
1304             vActiveMask = vStippleMask;
1305          }
1306       }
1307       lp_build_mask_begin(
1308          &mask, gallivm, lp_type_float_vec(32, 32 * 8), wrap(vActiveMask));
1309       uses_mask = true;
1310    }
1311
1312    lp_build_tgsi_soa(gallivm,
1313                      swr_fs->pipe.tokens,
1314                      lp_type_float_vec(32, 32 * 8),
1315                      uses_mask ? &mask : NULL, // mask
1316                      wrap(consts_ptr),
1317                      wrap(const_sizes_ptr),
1318                      &system_values,
1319                      inputs,
1320                      outputs,
1321                      wrap(hPrivateData),
1322                      NULL, // thread data
1323                      sampler, // sampler
1324                      &swr_fs->info.base,
1325                      NULL); // geometry shader face
1326
1327    sampler->destroy(sampler);
1328
1329    IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
1330
1331    for (uint32_t attrib = 0; attrib < swr_fs->info.base.num_outputs;
1332         attrib++) {
1333       switch (swr_fs->info.base.output_semantic_name[attrib]) {
1334       case TGSI_SEMANTIC_POSITION: {
1335          // write z
1336          LLVMValueRef outZ =
1337             LLVMBuildLoad(gallivm->builder, outputs[attrib][2], "");
1338          STORE(unwrap(outZ), pPS, {0, SWR_PS_CONTEXT_vZ});
1339          break;
1340       }
1341       case TGSI_SEMANTIC_COLOR: {
1342          for (uint32_t channel = 0; channel < TGSI_NUM_CHANNELS; channel++) {
1343             if (!outputs[attrib][channel])
1344                continue;
1345
1346             LLVMValueRef out =
1347                LLVMBuildLoad(gallivm->builder, outputs[attrib][channel], "");
1348             if (swr_fs->info.base.properties[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS] &&
1349                 swr_fs->info.base.output_semantic_index[attrib] == 0) {
1350                for (uint32_t rt = 0; rt < key.nr_cbufs; rt++) {
1351                   STORE(unwrap(out),
1352                         pPS,
1353                         {0, SWR_PS_CONTEXT_shaded, rt, channel});
1354                }
1355             } else {
1356                STORE(unwrap(out),
1357                      pPS,
1358                      {0,
1359                            SWR_PS_CONTEXT_shaded,
1360                            swr_fs->info.base.output_semantic_index[attrib],
1361                            channel});
1362             }
1363          }
1364          break;
1365       }
1366       default: {
1367          fprintf(stderr,
1368                  "unknown output from FS %s[%d]\n",
1369                  tgsi_semantic_names[swr_fs->info.base
1370                                         .output_semantic_name[attrib]],
1371                  swr_fs->info.base.output_semantic_index[attrib]);
1372          break;
1373       }
1374       }
1375    }
1376
1377    LLVMValueRef mask_result = 0;
1378    if (uses_mask) {
1379       mask_result = lp_build_mask_end(&mask);
1380    }
1381
1382    IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
1383
1384    if (uses_mask) {
1385       STORE(unwrap(mask_result), pPS, {0, SWR_PS_CONTEXT_activeMask});
1386    }
1387
1388    RET_VOID();
1389
1390    gallivm_verify_function(gallivm, wrap(pFunction));
1391
1392    gallivm_compile_module(gallivm);
1393
1394    // after the gallivm passes, we have to lower the core's intrinsics
1395    llvm::legacy::FunctionPassManager lowerPass(JM()->mpCurrentModule);
1396    lowerPass.add(createLowerX86Pass(mpJitMgr, this));
1397    lowerPass.run(*pFunction);
1398
1399    PFN_PIXEL_KERNEL kernel =
1400       (PFN_PIXEL_KERNEL)gallivm_jit_function(gallivm, wrap(pFunction));
1401    debug_printf("frag shader  %p\n", kernel);
1402    assert(kernel && "Error: FragShader = NULL");
1403
1404    JM()->mIsModuleFinalized = true;
1405
1406    return kernel;
1407 }
1408
1409 PFN_PIXEL_KERNEL
1410 swr_compile_fs(struct swr_context *ctx, swr_jit_fs_key &key)
1411 {
1412    if (!ctx->fs->pipe.tokens)
1413       return NULL;
1414
1415    BuilderSWR builder(
1416       reinterpret_cast<JitManager *>(swr_screen(ctx->pipe.screen)->hJitMgr),
1417       "FS");
1418    PFN_PIXEL_KERNEL func = builder.CompileFS(ctx, key);
1419
1420    ctx->fs->map.insert(std::make_pair(key, make_unique<VariantFS>(builder.gallivm, func)));
1421    return func;
1422 }