src/gallium/drivers/swr/swr_shader.cpp

   1 /****************************************************************************
   2  * Copyright (C) 2015 Intel Corporation.   All Rights Reserved.
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  ***************************************************************************/
  23
  24 // llvm redefines DEBUG
  25 #pragma push_macro("DEBUG")
  26 #undef DEBUG
  27 #include "JitManager.h"
  28 #include "llvm-c/Core.h"
  29 #include "llvm/Support/CBindingWrapping.h"
  30 #pragma pop_macro("DEBUG")
  31
  32 #include "state.h"
  33 #include "gen_state_llvm.h"
  34 #include "builder.h"
  35
  36 #include "tgsi/tgsi_strings.h"
  37 #include "util/u_format.h"
  38 #include "util/u_prim.h"
  39 #include "gallivm/lp_bld_init.h"
  40 #include "gallivm/lp_bld_flow.h"
  41 #include "gallivm/lp_bld_struct.h"
  42 #include "gallivm/lp_bld_tgsi.h"
  43
  44 #include "swr_context.h"
  45 #include "gen_swr_context_llvm.h"
  46 #include "swr_resource.h"
  47 #include "swr_state.h"
  48 #include "swr_screen.h"
  49
  50 using namespace SwrJit;
  51 using namespace llvm;
  52
  53 static unsigned
  54 locate_linkage(ubyte name, ubyte index, struct tgsi_shader_info *info);
  55
  56 bool operator==(const swr_jit_fs_key &lhs, const swr_jit_fs_key &rhs)
  57 {
  58    return !memcmp(&lhs, &rhs, sizeof(lhs));
  59 }
  60
  61 bool operator==(const swr_jit_vs_key &lhs, const swr_jit_vs_key &rhs)
  62 {
  63    return !memcmp(&lhs, &rhs, sizeof(lhs));
  64 }
  65
  66 bool operator==(const swr_jit_fetch_key &lhs, const swr_jit_fetch_key &rhs)
  67 {
  68    return !memcmp(&lhs, &rhs, sizeof(lhs));
  69 }
  70
  71 bool operator==(const swr_jit_gs_key &lhs, const swr_jit_gs_key &rhs)
  72 {
  73    return !memcmp(&lhs, &rhs, sizeof(lhs));
  74 }
  75
  76 static void
  77 swr_generate_sampler_key(const struct lp_tgsi_info &info,
  78                          struct swr_context *ctx,
  79                          enum pipe_shader_type shader_type,
  80                          struct swr_jit_sampler_key &key)
  81 {
  82    key.nr_samplers = info.base.file_max[TGSI_FILE_SAMPLER] + 1;
  83
  84    for (unsigned i = 0; i < key.nr_samplers; i++) {
  85       if (info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
  86          lp_sampler_static_sampler_state(
  87             &key.sampler[i].sampler_state,
  88             ctx->samplers[shader_type][i]);
  89       }
  90    }
  91
  92    /*
  93     * XXX If TGSI_FILE_SAMPLER_VIEW exists assume all texture opcodes
  94     * are dx10-style? Can't really have mixed opcodes, at least not
  95     * if we want to skip the holes here (without rescanning tgsi).
  96     */
  97    if (info.base.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
  98       key.nr_sampler_views =
  99          info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
 100       for (unsigned i = 0; i < key.nr_sampler_views; i++) {
 101          if (info.base.file_mask[TGSI_FILE_SAMPLER_VIEW] & (1 << i)) {
 102             const struct pipe_sampler_view *view =
 103                ctx->sampler_views[shader_type][i];
 104             lp_sampler_static_texture_state(
 105                &key.sampler[i].texture_state, view);
 106             if (view) {
 107                struct swr_resource *swr_res = swr_resource(view->texture);
 108                const struct util_format_description *desc =
 109                   util_format_description(view->format);
 110                if (swr_res->has_depth && swr_res->has_stencil &&
 111                    !util_format_has_depth(desc))
 112                   key.sampler[i].texture_state.format = PIPE_FORMAT_S8_UINT;
 113             }
 114          }
 115       }
 116    } else {
 117       key.nr_sampler_views = key.nr_samplers;
 118       for (unsigned i = 0; i < key.nr_sampler_views; i++) {
 119          if (info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
 120             const struct pipe_sampler_view *view =
 121                ctx->sampler_views[shader_type][i];
 122             lp_sampler_static_texture_state(
 123                &key.sampler[i].texture_state, view);
 124             if (view) {
 125                struct swr_resource *swr_res = swr_resource(view->texture);
 126                const struct util_format_description *desc =
 127                   util_format_description(view->format);
 128                if (swr_res->has_depth && swr_res->has_stencil &&
 129                    !util_format_has_depth(desc))
 130                   key.sampler[i].texture_state.format = PIPE_FORMAT_S8_UINT;
 131             }
 132          }
 133       }
 134    }
 135 }
 136
 137 void
 138 swr_generate_fs_key(struct swr_jit_fs_key &key,
 139                     struct swr_context *ctx,
 140                     swr_fragment_shader *swr_fs)
 141 {
 142    memset(&key, 0, sizeof(key));
 143
 144    key.nr_cbufs = ctx->framebuffer.nr_cbufs;
 145    key.light_twoside = ctx->rasterizer->light_twoside;
 146    key.sprite_coord_enable = ctx->rasterizer->sprite_coord_enable;
 147
 148    struct tgsi_shader_info *pPrevShader;
 149    if (ctx->gs)
 150       pPrevShader = &ctx->gs->info.base;
 151    else
 152       pPrevShader = &ctx->vs->info.base;
 153
 154    memcpy(&key.vs_output_semantic_name,
 155           &pPrevShader->output_semantic_name,
 156           sizeof(key.vs_output_semantic_name));
 157    memcpy(&key.vs_output_semantic_idx,
 158           &pPrevShader->output_semantic_index,
 159           sizeof(key.vs_output_semantic_idx));
 160
 161    swr_generate_sampler_key(swr_fs->info, ctx, PIPE_SHADER_FRAGMENT, key);
 162
 163    key.poly_stipple_enable = ctx->rasterizer->poly_stipple_enable &&
 164       ctx->poly_stipple.prim_is_poly;
 165 }
 166
 167 void
 168 swr_generate_vs_key(struct swr_jit_vs_key &key,
 169                     struct swr_context *ctx,
 170                     swr_vertex_shader *swr_vs)
 171 {
 172    memset(&key, 0, sizeof(key));
 173
 174    key.clip_plane_mask =
 175       swr_vs->info.base.clipdist_writemask ?
 176       swr_vs->info.base.clipdist_writemask & ctx->rasterizer->clip_plane_enable :
 177       ctx->rasterizer->clip_plane_enable;
 178
 179    swr_generate_sampler_key(swr_vs->info, ctx, PIPE_SHADER_VERTEX, key);
 180 }
 181
 182 void
 183 swr_generate_fetch_key(struct swr_jit_fetch_key &key,
 184                        struct swr_vertex_element_state *velems)
 185 {
 186    memset(&key, 0, sizeof(key));
 187
 188    key.fsState = velems->fsState;
 189 }
 190
 191 void
 192 swr_generate_gs_key(struct swr_jit_gs_key &key,
 193                     struct swr_context *ctx,
 194                     swr_geometry_shader *swr_gs)
 195 {
 196    memset(&key, 0, sizeof(key));
 197
 198    struct tgsi_shader_info *pPrevShader = &ctx->vs->info.base;
 199
 200    memcpy(&key.vs_output_semantic_name,
 201           &pPrevShader->output_semantic_name,
 202           sizeof(key.vs_output_semantic_name));
 203    memcpy(&key.vs_output_semantic_idx,
 204           &pPrevShader->output_semantic_index,
 205           sizeof(key.vs_output_semantic_idx));
 206
 207    swr_generate_sampler_key(swr_gs->info, ctx, PIPE_SHADER_GEOMETRY, key);
 208 }
 209
 210 struct BuilderSWR : public Builder {
 211    BuilderSWR(JitManager *pJitMgr, const char *pName)
 212       : Builder(pJitMgr)
 213    {
 214       pJitMgr->SetupNewModule();
 215       gallivm = gallivm_create(pName, wrap(&JM()->mContext));
 216       pJitMgr->mpCurrentModule = unwrap(gallivm->module);
 217    }
 218
 219    ~BuilderSWR() {
 220       gallivm_free_ir(gallivm);
 221    }
 222
 223    void WriteVS(Value *pVal, Value *pVsContext, Value *pVtxOutput,
 224                 unsigned slot, unsigned channel);
 225
 226    struct gallivm_state *gallivm;
 227    PFN_VERTEX_FUNC CompileVS(struct swr_context *ctx, swr_jit_vs_key &key);
 228    PFN_PIXEL_KERNEL CompileFS(struct swr_context *ctx, swr_jit_fs_key &key);
 229    PFN_GS_FUNC CompileGS(struct swr_context *ctx, swr_jit_gs_key &key);
 230
 231    LLVMValueRef
 232    swr_gs_llvm_fetch_input(const struct lp_build_tgsi_gs_iface *gs_iface,
 233                            struct lp_build_tgsi_context * bld_base,
 234                            boolean is_vindex_indirect,
 235                            LLVMValueRef vertex_index,
 236                            boolean is_aindex_indirect,
 237                            LLVMValueRef attrib_index,
 238                            LLVMValueRef swizzle_index);
 239    void
 240    swr_gs_llvm_emit_vertex(const struct lp_build_tgsi_gs_iface *gs_base,
 241                            struct lp_build_tgsi_context * bld_base,
 242                            LLVMValueRef (*outputs)[4],
 243                            LLVMValueRef emitted_vertices_vec);
 244
 245    void
 246    swr_gs_llvm_end_primitive(const struct lp_build_tgsi_gs_iface *gs_base,
 247                              struct lp_build_tgsi_context * bld_base,
 248                              LLVMValueRef verts_per_prim_vec,
 249                              LLVMValueRef emitted_prims_vec);
 250
 251    void
 252    swr_gs_llvm_epilogue(const struct lp_build_tgsi_gs_iface *gs_base,
 253                         struct lp_build_tgsi_context * bld_base,
 254                         LLVMValueRef total_emitted_vertices_vec,
 255                         LLVMValueRef emitted_prims_vec);
 256
 257 };
 258
 259 struct swr_gs_llvm_iface {
 260    struct lp_build_tgsi_gs_iface base;
 261    struct tgsi_shader_info *info;
 262
 263    BuilderSWR *pBuilder;
 264
 265    Value *pGsCtx;
 266    SWR_GS_STATE *pGsState;
 267    uint32_t num_outputs;
 268    uint32_t num_verts_per_prim;
 269
 270    Value *pVtxAttribMap;
 271 };
 272
 273 // trampoline functions so we can use the builder llvm construction methods
 274 static LLVMValueRef
 275 swr_gs_llvm_fetch_input(const struct lp_build_tgsi_gs_iface *gs_iface,
 276                            struct lp_build_tgsi_context * bld_base,
 277                            boolean is_vindex_indirect,
 278                            LLVMValueRef vertex_index,
 279                            boolean is_aindex_indirect,
 280                            LLVMValueRef attrib_index,
 281                            LLVMValueRef swizzle_index)
 282 {
 283     swr_gs_llvm_iface *iface = (swr_gs_llvm_iface*)gs_iface;
 284
 285     return iface->pBuilder->swr_gs_llvm_fetch_input(gs_iface, bld_base,
 286                                                    is_vindex_indirect,
 287                                                    vertex_index,
 288                                                    is_aindex_indirect,
 289                                                    attrib_index,
 290                                                    swizzle_index);
 291 }
 292
 293 static void
 294 swr_gs_llvm_emit_vertex(const struct lp_build_tgsi_gs_iface *gs_base,
 295                            struct lp_build_tgsi_context * bld_base,
 296                            LLVMValueRef (*outputs)[4],
 297                            LLVMValueRef emitted_vertices_vec)
 298 {
 299     swr_gs_llvm_iface *iface = (swr_gs_llvm_iface*)gs_base;
 300
 301     iface->pBuilder->swr_gs_llvm_emit_vertex(gs_base, bld_base,
 302                                             outputs,
 303                                             emitted_vertices_vec);
 304 }
 305
 306 static void
 307 swr_gs_llvm_end_primitive(const struct lp_build_tgsi_gs_iface *gs_base,
 308                              struct lp_build_tgsi_context * bld_base,
 309                              LLVMValueRef verts_per_prim_vec,
 310                              LLVMValueRef emitted_prims_vec)
 311 {
 312     swr_gs_llvm_iface *iface = (swr_gs_llvm_iface*)gs_base;
 313
 314     iface->pBuilder->swr_gs_llvm_end_primitive(gs_base, bld_base,
 315                                               verts_per_prim_vec,
 316                                               emitted_prims_vec);
 317 }
 318
 319 static void
 320 swr_gs_llvm_epilogue(const struct lp_build_tgsi_gs_iface *gs_base,
 321                         struct lp_build_tgsi_context * bld_base,
 322                         LLVMValueRef total_emitted_vertices_vec,
 323                         LLVMValueRef emitted_prims_vec)
 324 {
 325     swr_gs_llvm_iface *iface = (swr_gs_llvm_iface*)gs_base;
 326
 327     iface->pBuilder->swr_gs_llvm_epilogue(gs_base, bld_base,
 328                                          total_emitted_vertices_vec,
 329                                          emitted_prims_vec);
 330 }
 331
 332 LLVMValueRef
 333 BuilderSWR::swr_gs_llvm_fetch_input(const struct lp_build_tgsi_gs_iface *gs_iface,
 334                            struct lp_build_tgsi_context * bld_base,
 335                            boolean is_vindex_indirect,
 336                            LLVMValueRef vertex_index,
 337                            boolean is_aindex_indirect,
 338                            LLVMValueRef attrib_index,
 339                            LLVMValueRef swizzle_index)
 340 {
 341     swr_gs_llvm_iface *iface = (swr_gs_llvm_iface*)gs_iface;
 342
 343     IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 344
 345     assert(is_vindex_indirect == false && is_aindex_indirect == false);
 346
 347     Value *attrib =
 348        LOAD(GEP(iface->pVtxAttribMap, {C(0), unwrap(attrib_index)}));
 349
 350     Value *pInput =
 351        LOAD(GEP(iface->pGsCtx,
 352                 {C(0),
 353                  C(SWR_GS_CONTEXT_vert),
 354                  unwrap(vertex_index),
 355                  C(0),
 356                  attrib,
 357                  unwrap(swizzle_index)}));
 358
 359     return wrap(pInput);
 360 }
 361
 362 void
 363 BuilderSWR::swr_gs_llvm_emit_vertex(const struct lp_build_tgsi_gs_iface *gs_base,
 364                            struct lp_build_tgsi_context * bld_base,
 365                            LLVMValueRef (*outputs)[4],
 366                            LLVMValueRef emitted_vertices_vec)
 367 {
 368     swr_gs_llvm_iface *iface = (swr_gs_llvm_iface*)gs_base;
 369     SWR_GS_STATE *pGS = iface->pGsState;
 370
 371     IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 372
 373 #if USE_SIMD16_FRONTEND
 374     const uint32_t simdVertexStride = sizeof(simdvertex) * 2;
 375     const uint32_t numSimdBatches = (pGS->maxNumVerts + (mVWidth * 2) - 1) / (mVWidth * 2);
 376 #else
 377     const uint32_t simdVertexStride = sizeof(simdvertex);
 378     const uint32_t numSimdBatches = (pGS->maxNumVerts + mVWidth - 1) / mVWidth;
 379 #endif
 380     const uint32_t inputPrimStride = numSimdBatches * simdVertexStride;
 381
 382     Value *pStream = LOAD(iface->pGsCtx, { 0, SWR_GS_CONTEXT_pStream });
 383     Value *vMask = LOAD(iface->pGsCtx, { 0, SWR_GS_CONTEXT_mask });
 384     Value *vMask1 = TRUNC(vMask, VectorType::get(mInt1Ty, 8));
 385
 386     Value *vOffsets = C({
 387           inputPrimStride * 0,
 388           inputPrimStride * 1,
 389           inputPrimStride * 2,
 390           inputPrimStride * 3,
 391           inputPrimStride * 4,
 392           inputPrimStride * 5,
 393           inputPrimStride * 6,
 394           inputPrimStride * 7 } );
 395
 396 #if USE_SIMD16_FRONTEND
 397     const uint32_t simdShift = log2(mVWidth * 2);
 398     Value *vSimdSlot = AND(unwrap(emitted_vertices_vec), (mVWidth * 2) - 1);
 399 #else
 400     const uint32_t simdShift = log2(mVWidth);
 401     Value *vSimdSlot = AND(unwrap(emitted_vertices_vec), mVWidth - 1);
 402 #endif
 403     Value *vVertexSlot = ASHR(unwrap(emitted_vertices_vec), simdShift);
 404
 405     for (uint32_t attrib = 0; attrib < iface->num_outputs; ++attrib) {
 406        uint32_t attribSlot = attrib;
 407        uint32_t sgvChannel = 0;
 408        if (iface->info->output_semantic_name[attrib] == TGSI_SEMANTIC_PSIZE) {
 409           attribSlot = VERTEX_SGV_SLOT;
 410           sgvChannel = VERTEX_SGV_POINT_SIZE_COMP;
 411        } else if (iface->info->output_semantic_name[attrib] == TGSI_SEMANTIC_LAYER) {
 412           attribSlot = VERTEX_SGV_SLOT;
 413           sgvChannel = VERTEX_SGV_RTAI_COMP;
 414        } else if (iface->info->output_semantic_name[attrib] == TGSI_SEMANTIC_POSITION) {
 415           attribSlot = VERTEX_POSITION_SLOT;
 416        } else {
 417           attribSlot = VERTEX_ATTRIB_START_SLOT + attrib - 1;
 418        }
 419
 420 #if USE_SIMD16_FRONTEND
 421        Value *vOffsetsAttrib =
 422           ADD(vOffsets, MUL(vVertexSlot, VIMMED1((uint32_t)sizeof(simdvertex) * 2)));
 423        vOffsetsAttrib =
 424           ADD(vOffsetsAttrib, VIMMED1((uint32_t)(attribSlot*sizeof(simdvector) * 2)));
 425 #else
 426        Value *vOffsetsAttrib =
 427           ADD(vOffsets, MUL(vVertexSlot, VIMMED1((uint32_t)sizeof(simdvertex))));
 428        vOffsetsAttrib =
 429           ADD(vOffsetsAttrib, VIMMED1((uint32_t)(attribSlot*sizeof(simdvector))));
 430 #endif
 431        vOffsetsAttrib =
 432           ADD(vOffsetsAttrib, MUL(vSimdSlot, VIMMED1((uint32_t)sizeof(float))));
 433
 434        for (uint32_t channel = 0; channel < 4; ++channel) {
 435           Value *vPtrs = GEP(pStream, vOffsetsAttrib);
 436           Value *vData;
 437
 438           if (attribSlot == VERTEX_SGV_SLOT)
 439              vData = LOAD(unwrap(outputs[attrib][0]));
 440           else
 441              vData = LOAD(unwrap(outputs[attrib][channel]));
 442
 443           if (attribSlot != VERTEX_SGV_SLOT ||
 444               sgvChannel == channel) {
 445              vPtrs = BITCAST(vPtrs,
 446                              VectorType::get(PointerType::get(mFP32Ty, 0), 8));
 447
 448              MASKED_SCATTER(vData, vPtrs, 32, vMask1);
 449           }
 450
 451 #if USE_SIMD16_FRONTEND
 452           vOffsetsAttrib =
 453              ADD(vOffsetsAttrib, VIMMED1((uint32_t)sizeof(simdscalar) * 2));
 454 #else
 455           vOffsetsAttrib =
 456              ADD(vOffsetsAttrib, VIMMED1((uint32_t)sizeof(simdscalar)));
 457 #endif
 458        }
 459     }
 460 }
 461
 462 void
 463 BuilderSWR::swr_gs_llvm_end_primitive(const struct lp_build_tgsi_gs_iface *gs_base,
 464                              struct lp_build_tgsi_context * bld_base,
 465                              LLVMValueRef verts_per_prim_vec,
 466                              LLVMValueRef emitted_prims_vec)
 467 {
 468     swr_gs_llvm_iface *iface = (swr_gs_llvm_iface*)gs_base;
 469     SWR_GS_STATE *pGS = iface->pGsState;
 470
 471     IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 472
 473     Value *pCutBuffer =
 474        LOAD(iface->pGsCtx, {0, SWR_GS_CONTEXT_pCutOrStreamIdBuffer});
 475     Value *vMask = LOAD(iface->pGsCtx, { 0, SWR_GS_CONTEXT_mask });
 476     Value *vMask1 = TRUNC(vMask, VectorType::get(mInt1Ty, 8));
 477
 478     uint32_t vertsPerPrim = iface->num_verts_per_prim;
 479
 480     Value *vCount =
 481        ADD(MUL(unwrap(emitted_prims_vec), VIMMED1(vertsPerPrim)),
 482            unwrap(verts_per_prim_vec));
 483
 484     struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
 485     vCount = LOAD(unwrap(bld->total_emitted_vertices_vec_ptr));
 486
 487     struct lp_exec_mask *exec_mask = &bld->exec_mask;
 488     Value *mask = unwrap(lp_build_mask_value(bld->mask));
 489     if (exec_mask->has_mask)
 490        mask = AND(mask, unwrap(exec_mask->exec_mask));
 491
 492     Value *cmpMask = VMASK(ICMP_NE(unwrap(verts_per_prim_vec), VIMMED1(0)));
 493     mask = AND(mask, cmpMask);
 494     vMask1 = TRUNC(mask, VectorType::get(mInt1Ty, 8));
 495
 496     const uint32_t cutPrimStride =
 497        (pGS->maxNumVerts + JM()->mVWidth - 1) / JM()->mVWidth;
 498     Value *vOffsets = C({
 499           (uint32_t)(cutPrimStride * 0),
 500           (uint32_t)(cutPrimStride * 1),
 501           (uint32_t)(cutPrimStride * 2),
 502           (uint32_t)(cutPrimStride * 3),
 503           (uint32_t)(cutPrimStride * 4),
 504           (uint32_t)(cutPrimStride * 5),
 505           (uint32_t)(cutPrimStride * 6),
 506           (uint32_t)(cutPrimStride * 7) } );
 507
 508     vCount = SUB(vCount, VIMMED1(1));
 509     Value *vOffset = ADD(UDIV(vCount, VIMMED1(8)), vOffsets);
 510     Value *vValue = SHL(VIMMED1(1), UREM(vCount, VIMMED1(8)));
 511
 512     vValue = TRUNC(vValue, VectorType::get(mInt8Ty, 8));
 513
 514     Value *vPtrs = GEP(pCutBuffer, vOffset);
 515     vPtrs =
 516        BITCAST(vPtrs, VectorType::get(PointerType::get(mInt8Ty, 0), JM()->mVWidth));
 517
 518     Value *vGather = MASKED_GATHER(vPtrs, 32, vMask1);
 519     vValue = OR(vGather, vValue);
 520     MASKED_SCATTER(vValue, vPtrs, 32, vMask1);
 521 }
 522
 523 void
 524 BuilderSWR::swr_gs_llvm_epilogue(const struct lp_build_tgsi_gs_iface *gs_base,
 525                         struct lp_build_tgsi_context * bld_base,
 526                         LLVMValueRef total_emitted_vertices_vec,
 527                         LLVMValueRef emitted_prims_vec)
 528 {
 529    swr_gs_llvm_iface *iface = (swr_gs_llvm_iface*)gs_base;
 530
 531    IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 532
 533    STORE(unwrap(total_emitted_vertices_vec), iface->pGsCtx, {0, SWR_GS_CONTEXT_vertexCount});
 534 }
 535
 536 PFN_GS_FUNC
 537 BuilderSWR::CompileGS(struct swr_context *ctx, swr_jit_gs_key &key)
 538 {
 539    SWR_GS_STATE *pGS = &ctx->gs->gsState;
 540    struct tgsi_shader_info *info = &ctx->gs->info.base;
 541
 542    pGS->gsEnable = true;
 543
 544    pGS->numInputAttribs = info->num_inputs;
 545    pGS->outputTopology =
 546       swr_convert_prim_topology(info->properties[TGSI_PROPERTY_GS_OUTPUT_PRIM]);
 547    pGS->maxNumVerts = info->properties[TGSI_PROPERTY_GS_MAX_OUTPUT_VERTICES];
 548    pGS->instanceCount = info->properties[TGSI_PROPERTY_GS_INVOCATIONS];
 549
 550    // XXX: single stream for now...
 551    pGS->isSingleStream = true;
 552    pGS->singleStreamID = 0;
 553
 554    struct swr_geometry_shader *gs = ctx->gs;
 555
 556    LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
 557    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
 558
 559    memset(outputs, 0, sizeof(outputs));
 560
 561    AttrBuilder attrBuilder;
 562    attrBuilder.addStackAlignmentAttr(JM()->mVWidth * sizeof(float));
 563
 564    std::vector<Type *> gsArgs{PointerType::get(Gen_swr_draw_context(JM()), 0),
 565                               PointerType::get(Gen_SWR_GS_CONTEXT(JM()), 0)};
 566    FunctionType *vsFuncType =
 567       FunctionType::get(Type::getVoidTy(JM()->mContext), gsArgs, false);
 568
 569    // create new vertex shader function
 570    auto pFunction = Function::Create(vsFuncType,
 571                                      GlobalValue::ExternalLinkage,
 572                                      "GS",
 573                                      JM()->mpCurrentModule);
 574 #if HAVE_LLVM < 0x0500
 575    AttributeSet attrSet = AttributeSet::get(
 576       JM()->mContext, AttributeSet::FunctionIndex, attrBuilder);
 577    pFunction->addAttributes(AttributeSet::FunctionIndex, attrSet);
 578 #else
 579    pFunction->addAttributes(AttributeList::FunctionIndex, attrBuilder);
 580 #endif
 581
 582    BasicBlock *block = BasicBlock::Create(JM()->mContext, "entry", pFunction);
 583    IRB()->SetInsertPoint(block);
 584    LLVMPositionBuilderAtEnd(gallivm->builder, wrap(block));
 585
 586    auto argitr = pFunction->arg_begin();
 587    Value *hPrivateData = &*argitr++;
 588    hPrivateData->setName("hPrivateData");
 589    Value *pGsCtx = &*argitr++;
 590    pGsCtx->setName("gsCtx");
 591
 592    Value *consts_ptr =
 593       GEP(hPrivateData, {C(0), C(swr_draw_context_constantGS)});
 594    consts_ptr->setName("gs_constants");
 595    Value *const_sizes_ptr =
 596       GEP(hPrivateData, {0, swr_draw_context_num_constantsGS});
 597    const_sizes_ptr->setName("num_gs_constants");
 598
 599    struct lp_build_sampler_soa *sampler =
 600       swr_sampler_soa_create(key.sampler, PIPE_SHADER_GEOMETRY);
 601
 602    struct lp_bld_tgsi_system_values system_values;
 603    memset(&system_values, 0, sizeof(system_values));
 604    system_values.prim_id = wrap(LOAD(pGsCtx, {0, SWR_GS_CONTEXT_PrimitiveID}));
 605    system_values.instance_id = wrap(LOAD(pGsCtx, {0, SWR_GS_CONTEXT_InstanceID}));
 606
 607    std::vector<Constant*> mapConstants;
 608    Value *vtxAttribMap = ALLOCA(ArrayType::get(mInt32Ty, PIPE_MAX_SHADER_INPUTS));
 609    for (unsigned slot = 0; slot < info->num_inputs; slot++) {
 610       ubyte semantic_name = info->input_semantic_name[slot];
 611       ubyte semantic_idx = info->input_semantic_index[slot];
 612
 613       unsigned vs_slot = locate_linkage(semantic_name, semantic_idx, &ctx->vs->info.base);
 614
 615       vs_slot += VERTEX_ATTRIB_START_SLOT;
 616
 617       if (ctx->vs->info.base.output_semantic_name[0] == TGSI_SEMANTIC_POSITION)
 618          vs_slot--;
 619
 620       if (semantic_name == TGSI_SEMANTIC_POSITION)
 621          vs_slot = VERTEX_POSITION_SLOT;
 622
 623       STORE(C(vs_slot), vtxAttribMap, {0, slot});
 624       mapConstants.push_back(C(vs_slot));
 625    }
 626
 627    struct lp_build_mask_context mask;
 628    Value *mask_val = LOAD(pGsCtx, {0, SWR_GS_CONTEXT_mask}, "gsMask");
 629    lp_build_mask_begin(&mask, gallivm,
 630                        lp_type_float_vec(32, 32 * 8), wrap(mask_val));
 631
 632    // zero out cut buffer so we can load/modify/store bits
 633    MEMSET(LOAD(pGsCtx, {0, SWR_GS_CONTEXT_pCutOrStreamIdBuffer}),
 634           C((char)0),
 635           pGS->instanceCount * ((pGS->maxNumVerts + 7) / 8) * JM()->mVWidth,
 636           sizeof(float) * KNOB_SIMD_WIDTH);
 637
 638    struct swr_gs_llvm_iface gs_iface;
 639    gs_iface.base.fetch_input = ::swr_gs_llvm_fetch_input;
 640    gs_iface.base.emit_vertex = ::swr_gs_llvm_emit_vertex;
 641    gs_iface.base.end_primitive = ::swr_gs_llvm_end_primitive;
 642    gs_iface.base.gs_epilogue = ::swr_gs_llvm_epilogue;
 643    gs_iface.pBuilder = this;
 644    gs_iface.pGsCtx = pGsCtx;
 645    gs_iface.pGsState = pGS;
 646    gs_iface.num_outputs = gs->info.base.num_outputs;
 647    gs_iface.num_verts_per_prim =
 648       u_vertices_per_prim((pipe_prim_type)info->properties[TGSI_PROPERTY_GS_OUTPUT_PRIM]);
 649    gs_iface.info = info;
 650    gs_iface.pVtxAttribMap = vtxAttribMap;
 651
 652    lp_build_tgsi_soa(gallivm,
 653                      gs->pipe.tokens,
 654                      lp_type_float_vec(32, 32 * 8),
 655                      &mask,
 656                      wrap(consts_ptr),
 657                      wrap(const_sizes_ptr),
 658                      &system_values,
 659                      inputs,
 660                      outputs,
 661                      wrap(hPrivateData), // (sampler context)
 662                      NULL, // thread data
 663                      sampler,
 664                      &gs->info.base,
 665                      &gs_iface.base);
 666
 667    lp_build_mask_end(&mask);
 668
 669    sampler->destroy(sampler);
 670
 671    IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 672
 673    RET_VOID();
 674
 675    gallivm_verify_function(gallivm, wrap(pFunction));
 676    gallivm_compile_module(gallivm);
 677
 678    PFN_GS_FUNC pFunc =
 679       (PFN_GS_FUNC)gallivm_jit_function(gallivm, wrap(pFunction));
 680
 681    debug_printf("geom shader  %p\n", pFunc);
 682    assert(pFunc && "Error: GeomShader = NULL");
 683
 684    JM()->mIsModuleFinalized = true;
 685
 686    return pFunc;
 687 }
 688
 689 PFN_GS_FUNC
 690 swr_compile_gs(struct swr_context *ctx, swr_jit_gs_key &key)
 691 {
 692    BuilderSWR builder(
 693       reinterpret_cast<JitManager *>(swr_screen(ctx->pipe.screen)->hJitMgr),
 694       "GS");
 695    PFN_GS_FUNC func = builder.CompileGS(ctx, key);
 696
 697    ctx->gs->map.insert(std::make_pair(key, make_unique<VariantGS>(builder.gallivm, func)));
 698    return func;
 699 }
 700
 701 void
 702 BuilderSWR::WriteVS(Value *pVal, Value *pVsContext, Value *pVtxOutput, unsigned slot, unsigned channel)
 703 {
 704 #if USE_SIMD16_FRONTEND
 705    // interleave the simdvertex components into the dest simd16vertex
 706    //   slot16offset = slot8offset * 2
 707    //   comp16offset = comp8offset * 2 + alternateOffset
 708
 709    Value *offset = LOAD(pVsContext, { 0, SWR_VS_CONTEXT_AlternateOffset });
 710    Value *pOut = GEP(pVtxOutput, { C(0), C(0), C(slot * 2), offset } );
 711    STORE(pVal, pOut, {channel * 2});
 712 #else
 713    Value *pOut = GEP(pVtxOutput, {0, 0, slot});
 714    STORE(pVal, pOut, {0, channel});
 715 #endif
 716 }
 717
 718 PFN_VERTEX_FUNC
 719 BuilderSWR::CompileVS(struct swr_context *ctx, swr_jit_vs_key &key)
 720 {
 721    struct swr_vertex_shader *swr_vs = ctx->vs;
 722
 723    LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
 724    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
 725
 726    memset(outputs, 0, sizeof(outputs));
 727
 728    AttrBuilder attrBuilder;
 729    attrBuilder.addStackAlignmentAttr(JM()->mVWidth * sizeof(float));
 730
 731    std::vector<Type *> vsArgs{PointerType::get(Gen_swr_draw_context(JM()), 0),
 732                               PointerType::get(Gen_SWR_VS_CONTEXT(JM()), 0)};
 733    FunctionType *vsFuncType =
 734       FunctionType::get(Type::getVoidTy(JM()->mContext), vsArgs, false);
 735
 736    // create new vertex shader function
 737    auto pFunction = Function::Create(vsFuncType,
 738                                      GlobalValue::ExternalLinkage,
 739                                      "VS",
 740                                      JM()->mpCurrentModule);
 741 #if HAVE_LLVM < 0x0500
 742    AttributeSet attrSet = AttributeSet::get(
 743       JM()->mContext, AttributeSet::FunctionIndex, attrBuilder);
 744    pFunction->addAttributes(AttributeSet::FunctionIndex, attrSet);
 745 #else
 746    pFunction->addAttributes(AttributeList::FunctionIndex, attrBuilder);
 747 #endif
 748
 749    BasicBlock *block = BasicBlock::Create(JM()->mContext, "entry", pFunction);
 750    IRB()->SetInsertPoint(block);
 751    LLVMPositionBuilderAtEnd(gallivm->builder, wrap(block));
 752
 753    auto argitr = pFunction->arg_begin();
 754    Value *hPrivateData = &*argitr++;
 755    hPrivateData->setName("hPrivateData");
 756    Value *pVsCtx = &*argitr++;
 757    pVsCtx->setName("vsCtx");
 758
 759    Value *consts_ptr = GEP(hPrivateData, {C(0), C(swr_draw_context_constantVS)});
 760
 761    consts_ptr->setName("vs_constants");
 762    Value *const_sizes_ptr =
 763       GEP(hPrivateData, {0, swr_draw_context_num_constantsVS});
 764    const_sizes_ptr->setName("num_vs_constants");
 765
 766    Value *vtxInput = LOAD(pVsCtx, {0, SWR_VS_CONTEXT_pVin});
 767
 768    for (uint32_t attrib = 0; attrib < PIPE_MAX_SHADER_INPUTS; attrib++) {
 769       const unsigned mask = swr_vs->info.base.input_usage_mask[attrib];
 770       for (uint32_t channel = 0; channel < TGSI_NUM_CHANNELS; channel++) {
 771          if (mask & (1 << channel)) {
 772             inputs[attrib][channel] =
 773                wrap(LOAD(vtxInput, {0, 0, attrib, channel}));
 774          }
 775       }
 776    }
 777
 778    struct lp_build_sampler_soa *sampler =
 779       swr_sampler_soa_create(key.sampler, PIPE_SHADER_VERTEX);
 780
 781    struct lp_bld_tgsi_system_values system_values;
 782    memset(&system_values, 0, sizeof(system_values));
 783    system_values.instance_id = wrap(LOAD(pVsCtx, {0, SWR_VS_CONTEXT_InstanceID}));
 784    system_values.vertex_id = wrap(LOAD(pVsCtx, {0, SWR_VS_CONTEXT_VertexID}));
 785
 786    lp_build_tgsi_soa(gallivm,
 787                      swr_vs->pipe.tokens,
 788                      lp_type_float_vec(32, 32 * 8),
 789                      NULL, // mask
 790                      wrap(consts_ptr),
 791                      wrap(const_sizes_ptr),
 792                      &system_values,
 793                      inputs,
 794                      outputs,
 795                      wrap(hPrivateData), // (sampler context)
 796                      NULL, // thread data
 797                      sampler, // sampler
 798                      &swr_vs->info.base,
 799                      NULL); // geometry shader face
 800
 801    sampler->destroy(sampler);
 802
 803    IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
 804
 805    Value *vtxOutput = LOAD(pVsCtx, {0, SWR_VS_CONTEXT_pVout});
 806
 807    for (uint32_t channel = 0; channel < TGSI_NUM_CHANNELS; channel++) {
 808       for (uint32_t attrib = 0; attrib < PIPE_MAX_SHADER_OUTPUTS; attrib++) {
 809          if (!outputs[attrib][channel])
 810             continue;
 811
 812          Value *val;
 813          uint32_t outSlot;
 814
 815          if (swr_vs->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_PSIZE) {
 816             if (channel != VERTEX_SGV_POINT_SIZE_COMP)
 817                continue;
 818             val = LOAD(unwrap(outputs[attrib][0]));
 819             outSlot = VERTEX_SGV_SLOT;
 820          } else if (swr_vs->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_POSITION) {
 821             val = LOAD(unwrap(outputs[attrib][channel]));
 822             outSlot = VERTEX_POSITION_SLOT;
 823          } else {
 824             val = LOAD(unwrap(outputs[attrib][channel]));
 825             outSlot = VERTEX_ATTRIB_START_SLOT + attrib;
 826             if (swr_vs->info.base.output_semantic_name[0] == TGSI_SEMANTIC_POSITION)
 827                outSlot--;
 828          }
 829
 830          WriteVS(val, pVsCtx, vtxOutput, outSlot, channel);
 831       }
 832    }
 833
 834    if (ctx->rasterizer->clip_plane_enable ||
 835        swr_vs->info.base.culldist_writemask) {
 836       unsigned clip_mask = ctx->rasterizer->clip_plane_enable;
 837
 838       unsigned cv = 0;
 839       if (swr_vs->info.base.writes_clipvertex) {
 840          cv = locate_linkage(TGSI_SEMANTIC_CLIPVERTEX, 0,
 841                              &swr_vs->info.base);
 842       } else {
 843          for (int i = 0; i < PIPE_MAX_SHADER_OUTPUTS; i++) {
 844             if (swr_vs->info.base.output_semantic_name[i] == TGSI_SEMANTIC_POSITION &&
 845                 swr_vs->info.base.output_semantic_index[i] == 0) {
 846                cv = i;
 847                break;
 848             }
 849          }
 850       }
 851       LLVMValueRef cx = LLVMBuildLoad(gallivm->builder, outputs[cv][0], "");
 852       LLVMValueRef cy = LLVMBuildLoad(gallivm->builder, outputs[cv][1], "");
 853       LLVMValueRef cz = LLVMBuildLoad(gallivm->builder, outputs[cv][2], "");
 854       LLVMValueRef cw = LLVMBuildLoad(gallivm->builder, outputs[cv][3], "");
 855
 856       for (unsigned val = 0; val < PIPE_MAX_CLIP_PLANES; val++) {
 857          // clip distance overrides user clip planes
 858          if ((swr_vs->info.base.clipdist_writemask & clip_mask & (1 << val)) ||
 859              ((swr_vs->info.base.culldist_writemask << swr_vs->info.base.num_written_clipdistance) & (1 << val))) {
 860             unsigned cv = locate_linkage(TGSI_SEMANTIC_CLIPDIST, val < 4 ? 0 : 1,
 861                                          &swr_vs->info.base);
 862             if (val < 4) {
 863                LLVMValueRef dist = LLVMBuildLoad(gallivm->builder, outputs[cv][val], "");
 864                WriteVS(unwrap(dist), pVsCtx, vtxOutput, VERTEX_CLIPCULL_DIST_LO_SLOT, val);
 865             } else {
 866                LLVMValueRef dist = LLVMBuildLoad(gallivm->builder, outputs[cv][val - 4], "");
 867                WriteVS(unwrap(dist), pVsCtx, vtxOutput, VERTEX_CLIPCULL_DIST_HI_SLOT, val - 4);
 868             }
 869             continue;
 870          }
 871
 872          if (!(clip_mask & (1 << val)))
 873             continue;
 874
 875          Value *px = LOAD(GEP(hPrivateData, {0, swr_draw_context_userClipPlanes, val, 0}));
 876          Value *py = LOAD(GEP(hPrivateData, {0, swr_draw_context_userClipPlanes, val, 1}));
 877          Value *pz = LOAD(GEP(hPrivateData, {0, swr_draw_context_userClipPlanes, val, 2}));
 878          Value *pw = LOAD(GEP(hPrivateData, {0, swr_draw_context_userClipPlanes, val, 3}));
 879          Value *dist = FADD(FMUL(unwrap(cx), VBROADCAST(px)),
 880                             FADD(FMUL(unwrap(cy), VBROADCAST(py)),
 881                                  FADD(FMUL(unwrap(cz), VBROADCAST(pz)),
 882                                       FMUL(unwrap(cw), VBROADCAST(pw)))));
 883
 884          if (val < 4)
 885             WriteVS(dist, pVsCtx, vtxOutput, VERTEX_CLIPCULL_DIST_LO_SLOT, val);
 886          else
 887             WriteVS(dist, pVsCtx, vtxOutput, VERTEX_CLIPCULL_DIST_HI_SLOT, val - 4);
 888       }
 889    }
 890
 891    RET_VOID();
 892
 893    gallivm_verify_function(gallivm, wrap(pFunction));
 894    gallivm_compile_module(gallivm);
 895
 896    //   lp_debug_dump_value(func);
 897
 898    PFN_VERTEX_FUNC pFunc =
 899       (PFN_VERTEX_FUNC)gallivm_jit_function(gallivm, wrap(pFunction));
 900
 901    debug_printf("vert shader  %p\n", pFunc);
 902    assert(pFunc && "Error: VertShader = NULL");
 903
 904    JM()->mIsModuleFinalized = true;
 905
 906    return pFunc;
 907 }
 908
 909 PFN_VERTEX_FUNC
 910 swr_compile_vs(struct swr_context *ctx, swr_jit_vs_key &key)
 911 {
 912    if (!ctx->vs->pipe.tokens)
 913       return NULL;
 914
 915    BuilderSWR builder(
 916       reinterpret_cast<JitManager *>(swr_screen(ctx->pipe.screen)->hJitMgr),
 917       "VS");
 918    PFN_VERTEX_FUNC func = builder.CompileVS(ctx, key);
 919
 920    ctx->vs->map.insert(std::make_pair(key, make_unique<VariantVS>(builder.gallivm, func)));
 921    return func;
 922 }
 923
 924 static unsigned
 925 locate_linkage(ubyte name, ubyte index, struct tgsi_shader_info *info)
 926 {
 927    for (int i = 0; i < PIPE_MAX_SHADER_OUTPUTS; i++) {
 928       if ((info->output_semantic_name[i] == name)
 929           && (info->output_semantic_index[i] == index)) {
 930          return i;
 931       }
 932    }
 933
 934    return 0xFFFFFFFF;
 935 }
 936
 937 PFN_PIXEL_KERNEL
 938 BuilderSWR::CompileFS(struct swr_context *ctx, swr_jit_fs_key &key)
 939 {
 940    struct swr_fragment_shader *swr_fs = ctx->fs;
 941
 942    struct tgsi_shader_info *pPrevShader;
 943    if (ctx->gs)
 944       pPrevShader = &ctx->gs->info.base;
 945    else
 946       pPrevShader = &ctx->vs->info.base;
 947
 948    LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
 949    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
 950
 951    memset(inputs, 0, sizeof(inputs));
 952    memset(outputs, 0, sizeof(outputs));
 953
 954    struct lp_build_sampler_soa *sampler = NULL;
 955
 956    AttrBuilder attrBuilder;
 957    attrBuilder.addStackAlignmentAttr(JM()->mVWidth * sizeof(float));
 958
 959    std::vector<Type *> fsArgs{PointerType::get(Gen_swr_draw_context(JM()), 0),
 960                               PointerType::get(Gen_SWR_PS_CONTEXT(JM()), 0)};
 961    FunctionType *funcType =
 962       FunctionType::get(Type::getVoidTy(JM()->mContext), fsArgs, false);
 963
 964    auto pFunction = Function::Create(funcType,
 965                                      GlobalValue::ExternalLinkage,
 966                                      "FS",
 967                                      JM()->mpCurrentModule);
 968 #if HAVE_LLVM < 0x0500
 969    AttributeSet attrSet = AttributeSet::get(
 970       JM()->mContext, AttributeSet::FunctionIndex, attrBuilder);
 971    pFunction->addAttributes(AttributeSet::FunctionIndex, attrSet);
 972 #else
 973    pFunction->addAttributes(AttributeList::FunctionIndex, attrBuilder);
 974 #endif
 975
 976    BasicBlock *block = BasicBlock::Create(JM()->mContext, "entry", pFunction);
 977    IRB()->SetInsertPoint(block);
 978    LLVMPositionBuilderAtEnd(gallivm->builder, wrap(block));
 979
 980    auto args = pFunction->arg_begin();
 981    Value *hPrivateData = &*args++;
 982    hPrivateData->setName("hPrivateData");
 983    Value *pPS = &*args++;
 984    pPS->setName("psCtx");
 985
 986    Value *consts_ptr = GEP(hPrivateData, {0, swr_draw_context_constantFS});
 987    consts_ptr->setName("fs_constants");
 988    Value *const_sizes_ptr =
 989       GEP(hPrivateData, {0, swr_draw_context_num_constantsFS});
 990    const_sizes_ptr->setName("num_fs_constants");
 991
 992    // load *pAttribs, *pPerspAttribs
 993    Value *pRawAttribs = LOAD(pPS, {0, SWR_PS_CONTEXT_pAttribs}, "pRawAttribs");
 994    Value *pPerspAttribs =
 995       LOAD(pPS, {0, SWR_PS_CONTEXT_pPerspAttribs}, "pPerspAttribs");
 996
 997    swr_fs->constantMask = 0;
 998    swr_fs->flatConstantMask = 0;
 999    swr_fs->pointSpriteMask = 0;
1000
1001    for (int attrib = 0; attrib < PIPE_MAX_SHADER_INPUTS; attrib++) {
1002       const unsigned mask = swr_fs->info.base.input_usage_mask[attrib];
1003       const unsigned interpMode = swr_fs->info.base.input_interpolate[attrib];
1004       const unsigned interpLoc = swr_fs->info.base.input_interpolate_loc[attrib];
1005
1006       if (!mask)
1007          continue;
1008
1009       // load i,j
1010       Value *vi = nullptr, *vj = nullptr;
1011       switch (interpLoc) {
1012       case TGSI_INTERPOLATE_LOC_CENTER:
1013          vi = LOAD(pPS, {0, SWR_PS_CONTEXT_vI, PixelPositions_center}, "i");
1014          vj = LOAD(pPS, {0, SWR_PS_CONTEXT_vJ, PixelPositions_center}, "j");
1015          break;
1016       case TGSI_INTERPOLATE_LOC_CENTROID:
1017          vi = LOAD(pPS, {0, SWR_PS_CONTEXT_vI, PixelPositions_centroid}, "i");
1018          vj = LOAD(pPS, {0, SWR_PS_CONTEXT_vJ, PixelPositions_centroid}, "j");
1019          break;
1020       case TGSI_INTERPOLATE_LOC_SAMPLE:
1021          vi = LOAD(pPS, {0, SWR_PS_CONTEXT_vI, PixelPositions_sample}, "i");
1022          vj = LOAD(pPS, {0, SWR_PS_CONTEXT_vJ, PixelPositions_sample}, "j");
1023          break;
1024       }
1025
1026       // load/compute w
1027       Value *vw = nullptr, *pAttribs;
1028       if (interpMode == TGSI_INTERPOLATE_PERSPECTIVE ||
1029           interpMode == TGSI_INTERPOLATE_COLOR) {
1030          pAttribs = pPerspAttribs;
1031          switch (interpLoc) {
1032          case TGSI_INTERPOLATE_LOC_CENTER:
1033             vw = VRCP(LOAD(pPS, {0, SWR_PS_CONTEXT_vOneOverW, PixelPositions_center}));
1034             break;
1035          case TGSI_INTERPOLATE_LOC_CENTROID:
1036             vw = VRCP(LOAD(pPS, {0, SWR_PS_CONTEXT_vOneOverW, PixelPositions_centroid}));
1037             break;
1038          case TGSI_INTERPOLATE_LOC_SAMPLE:
1039             vw = VRCP(LOAD(pPS, {0, SWR_PS_CONTEXT_vOneOverW, PixelPositions_sample}));
1040             break;
1041          }
1042       } else {
1043          pAttribs = pRawAttribs;
1044          vw = VIMMED1(1.f);
1045       }
1046
1047       vw->setName("w");
1048
1049       ubyte semantic_name = swr_fs->info.base.input_semantic_name[attrib];
1050       ubyte semantic_idx = swr_fs->info.base.input_semantic_index[attrib];
1051
1052       if (semantic_name == TGSI_SEMANTIC_FACE) {
1053          Value *ff =
1054             UI_TO_FP(LOAD(pPS, {0, SWR_PS_CONTEXT_frontFace}), mFP32Ty);
1055          ff = FSUB(FMUL(ff, C(2.0f)), C(1.0f));
1056          ff = VECTOR_SPLAT(JM()->mVWidth, ff, "vFrontFace");
1057
1058          inputs[attrib][0] = wrap(ff);
1059          inputs[attrib][1] = wrap(VIMMED1(0.0f));
1060          inputs[attrib][2] = wrap(VIMMED1(0.0f));
1061          inputs[attrib][3] = wrap(VIMMED1(1.0f));
1062          continue;
1063       } else if (semantic_name == TGSI_SEMANTIC_POSITION) { // gl_FragCoord
1064          if (swr_fs->info.base.properties[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER] ==
1065              TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER) {
1066             inputs[attrib][0] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vX, PixelPositions_center}, "vX"));
1067             inputs[attrib][1] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vY, PixelPositions_center}, "vY"));
1068          } else {
1069             inputs[attrib][0] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vX, PixelPositions_UL}, "vX"));
1070             inputs[attrib][1] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vY, PixelPositions_UL}, "vY"));
1071          }
1072          inputs[attrib][2] = wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vZ}, "vZ"));
1073          inputs[attrib][3] =
1074             wrap(LOAD(pPS, {0, SWR_PS_CONTEXT_vOneOverW, PixelPositions_center}, "vOneOverW"));
1075          continue;
1076       }
1077
1078       unsigned linkedAttrib =
1079          locate_linkage(semantic_name, semantic_idx, pPrevShader) - 1;
1080
1081       uint32_t extraAttribs = 0;
1082       if (semantic_name == TGSI_SEMANTIC_PRIMID && !ctx->gs) {
1083          /* non-gs generated primID - need to grab from swizzleMap override */
1084          linkedAttrib = pPrevShader->num_outputs - 1;
1085          swr_fs->constantMask |= 1 << linkedAttrib;
1086          extraAttribs++;
1087       } else if (semantic_name == TGSI_SEMANTIC_GENERIC &&
1088           key.sprite_coord_enable & (1 << semantic_idx)) {
1089          /* we add an extra attrib to the backendState in swr_update_derived. */
1090          linkedAttrib = pPrevShader->num_outputs + extraAttribs - 1;
1091          swr_fs->pointSpriteMask |= (1 << linkedAttrib);
1092          extraAttribs++;
1093       } else if (linkedAttrib == 0xFFFFFFFF) {
1094          inputs[attrib][0] = wrap(VIMMED1(0.0f));
1095          inputs[attrib][1] = wrap(VIMMED1(0.0f));
1096          inputs[attrib][2] = wrap(VIMMED1(0.0f));
1097          inputs[attrib][3] = wrap(VIMMED1(1.0f));
1098          /* If we're reading in color and 2-sided lighting is enabled, we have
1099           * to keep going.
1100           */
1101          if (semantic_name != TGSI_SEMANTIC_COLOR || !key.light_twoside)
1102             continue;
1103       } else {
1104          if (interpMode == TGSI_INTERPOLATE_CONSTANT) {
1105             swr_fs->constantMask |= 1 << linkedAttrib;
1106          } else if (interpMode == TGSI_INTERPOLATE_COLOR) {
1107             swr_fs->flatConstantMask |= 1 << linkedAttrib;
1108          }
1109       }
1110
1111       unsigned bcolorAttrib = 0xFFFFFFFF;
1112       Value *offset = NULL;
1113       if (semantic_name == TGSI_SEMANTIC_COLOR && key.light_twoside) {
1114          bcolorAttrib = locate_linkage(
1115                TGSI_SEMANTIC_BCOLOR, semantic_idx, pPrevShader) - 1;
1116          /* Neither front nor back colors were available. Nothing to load. */
1117          if (bcolorAttrib == 0xFFFFFFFF && linkedAttrib == 0xFFFFFFFF)
1118             continue;
1119          /* If there is no front color, just always use the back color. */
1120          if (linkedAttrib == 0xFFFFFFFF)
1121             linkedAttrib = bcolorAttrib;
1122
1123          if (bcolorAttrib != 0xFFFFFFFF) {
1124             if (interpMode == TGSI_INTERPOLATE_CONSTANT) {
1125                swr_fs->constantMask |= 1 << bcolorAttrib;
1126             } else if (interpMode == TGSI_INTERPOLATE_COLOR) {
1127                swr_fs->flatConstantMask |= 1 << bcolorAttrib;
1128             }
1129
1130             unsigned diff = 12 * (bcolorAttrib - linkedAttrib);
1131
1132             if (diff) {
1133                Value *back =
1134                   XOR(C(1), LOAD(pPS, {0, SWR_PS_CONTEXT_frontFace}), "backFace");
1135
1136                offset = MUL(back, C(diff));
1137                offset->setName("offset");
1138             }
1139          }
1140       }
1141
1142       for (int channel = 0; channel < TGSI_NUM_CHANNELS; channel++) {
1143          if (mask & (1 << channel)) {
1144             Value *indexA = C(linkedAttrib * 12 + channel);
1145             Value *indexB = C(linkedAttrib * 12 + channel + 4);
1146             Value *indexC = C(linkedAttrib * 12 + channel + 8);
1147
1148             if (offset) {
1149                indexA = ADD(indexA, offset);
1150                indexB = ADD(indexB, offset);
1151                indexC = ADD(indexC, offset);
1152             }
1153
1154             Value *va = VBROADCAST(LOAD(GEP(pAttribs, indexA)));
1155             Value *vb = VBROADCAST(LOAD(GEP(pAttribs, indexB)));
1156             Value *vc = VBROADCAST(LOAD(GEP(pAttribs, indexC)));
1157
1158             if (interpMode == TGSI_INTERPOLATE_CONSTANT) {
1159                inputs[attrib][channel] = wrap(va);
1160             } else {
1161                Value *vk = FSUB(FSUB(VIMMED1(1.0f), vi), vj);
1162
1163                vc = FMUL(vk, vc);
1164
1165                Value *interp = FMUL(va, vi);
1166                Value *interp1 = FMUL(vb, vj);
1167                interp = FADD(interp, interp1);
1168                interp = FADD(interp, vc);
1169                if (interpMode == TGSI_INTERPOLATE_PERSPECTIVE ||
1170                    interpMode == TGSI_INTERPOLATE_COLOR)
1171                   interp = FMUL(interp, vw);
1172                inputs[attrib][channel] = wrap(interp);
1173             }
1174          }
1175       }
1176    }
1177
1178    sampler = swr_sampler_soa_create(key.sampler, PIPE_SHADER_FRAGMENT);
1179
1180    struct lp_bld_tgsi_system_values system_values;
1181    memset(&system_values, 0, sizeof(system_values));
1182
1183    struct lp_build_mask_context mask;
1184    bool uses_mask = false;
1185
1186    if (swr_fs->info.base.uses_kill ||
1187        key.poly_stipple_enable) {
1188       Value *vActiveMask = NULL;
1189       if (swr_fs->info.base.uses_kill) {
1190          vActiveMask = LOAD(pPS, {0, SWR_PS_CONTEXT_activeMask}, "activeMask");
1191       }
1192       if (key.poly_stipple_enable) {
1193          // first get fragment xy coords and clip to stipple bounds
1194          Value *vXf = LOAD(pPS, {0, SWR_PS_CONTEXT_vX, PixelPositions_UL});
1195          Value *vYf = LOAD(pPS, {0, SWR_PS_CONTEXT_vY, PixelPositions_UL});
1196          Value *vXu = FP_TO_UI(vXf, mSimdInt32Ty);
1197          Value *vYu = FP_TO_UI(vYf, mSimdInt32Ty);
1198
1199          // stipple pattern is 32x32, which means that one line of stipple
1200          // is stored in one word:
1201          // vXstipple is bit offset inside 32-bit stipple word
1202          // vYstipple is word index is stipple array
1203          Value *vXstipple = AND(vXu, VIMMED1(0x1f)); // & (32-1)
1204          Value *vYstipple = AND(vYu, VIMMED1(0x1f)); // & (32-1)
1205
1206          // grab stipple pattern base address
1207          Value *stipplePtr = GEP(hPrivateData, {0, swr_draw_context_polyStipple, 0});
1208          stipplePtr = BITCAST(stipplePtr, mInt8PtrTy);
1209
1210          // peform a gather to grab stipple words for each lane
1211          Value *vStipple = GATHERDD(VUNDEF_I(), stipplePtr, vYstipple,
1212                                     VIMMED1(0xffffffff), C((char)4));
1213
1214          // create a mask with one bit corresponding to the x stipple
1215          // and AND it with the pattern, to see if we have a bit
1216          Value *vBitMask = LSHR(VIMMED1(0x80000000), vXstipple);
1217          Value *vStippleMask = AND(vStipple, vBitMask);
1218          vStippleMask = ICMP_NE(vStippleMask, VIMMED1(0));
1219          vStippleMask = VMASK(vStippleMask);
1220
1221          if (swr_fs->info.base.uses_kill) {
1222             vActiveMask = AND(vActiveMask, vStippleMask);
1223          } else {
1224             vActiveMask = vStippleMask;
1225          }
1226       }
1227       lp_build_mask_begin(
1228          &mask, gallivm, lp_type_float_vec(32, 32 * 8), wrap(vActiveMask));
1229       uses_mask = true;
1230    }
1231
1232    lp_build_tgsi_soa(gallivm,
1233                      swr_fs->pipe.tokens,
1234                      lp_type_float_vec(32, 32 * 8),
1235                      uses_mask ? &mask : NULL, // mask
1236                      wrap(consts_ptr),
1237                      wrap(const_sizes_ptr),
1238                      &system_values,
1239                      inputs,
1240                      outputs,
1241                      wrap(hPrivateData),
1242                      NULL, // thread data
1243                      sampler, // sampler
1244                      &swr_fs->info.base,
1245                      NULL); // geometry shader face
1246
1247    sampler->destroy(sampler);
1248
1249    IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
1250
1251    for (uint32_t attrib = 0; attrib < swr_fs->info.base.num_outputs;
1252         attrib++) {
1253       switch (swr_fs->info.base.output_semantic_name[attrib]) {
1254       case TGSI_SEMANTIC_POSITION: {
1255          // write z
1256          LLVMValueRef outZ =
1257             LLVMBuildLoad(gallivm->builder, outputs[attrib][2], "");
1258          STORE(unwrap(outZ), pPS, {0, SWR_PS_CONTEXT_vZ});
1259          break;
1260       }
1261       case TGSI_SEMANTIC_COLOR: {
1262          for (uint32_t channel = 0; channel < TGSI_NUM_CHANNELS; channel++) {
1263             if (!outputs[attrib][channel])
1264                continue;
1265
1266             LLVMValueRef out =
1267                LLVMBuildLoad(gallivm->builder, outputs[attrib][channel], "");
1268             if (swr_fs->info.base.properties[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS] &&
1269                 swr_fs->info.base.output_semantic_index[attrib] == 0) {
1270                for (uint32_t rt = 0; rt < key.nr_cbufs; rt++) {
1271                   STORE(unwrap(out),
1272                         pPS,
1273                         {0, SWR_PS_CONTEXT_shaded, rt, channel});
1274                }
1275             } else {
1276                STORE(unwrap(out),
1277                      pPS,
1278                      {0,
1279                            SWR_PS_CONTEXT_shaded,
1280                            swr_fs->info.base.output_semantic_index[attrib],
1281                            channel});
1282             }
1283          }
1284          break;
1285       }
1286       default: {
1287          fprintf(stderr,
1288                  "unknown output from FS %s[%d]\n",
1289                  tgsi_semantic_names[swr_fs->info.base
1290                                         .output_semantic_name[attrib]],
1291                  swr_fs->info.base.output_semantic_index[attrib]);
1292          break;
1293       }
1294       }
1295    }
1296
1297    LLVMValueRef mask_result = 0;
1298    if (uses_mask) {
1299       mask_result = lp_build_mask_end(&mask);
1300    }
1301
1302    IRB()->SetInsertPoint(unwrap(LLVMGetInsertBlock(gallivm->builder)));
1303
1304    if (uses_mask) {
1305       STORE(unwrap(mask_result), pPS, {0, SWR_PS_CONTEXT_activeMask});
1306    }
1307
1308    RET_VOID();
1309
1310    gallivm_verify_function(gallivm, wrap(pFunction));
1311
1312    gallivm_compile_module(gallivm);
1313
1314    PFN_PIXEL_KERNEL kernel =
1315       (PFN_PIXEL_KERNEL)gallivm_jit_function(gallivm, wrap(pFunction));
1316    debug_printf("frag shader  %p\n", kernel);
1317    assert(kernel && "Error: FragShader = NULL");
1318
1319    JM()->mIsModuleFinalized = true;
1320
1321    return kernel;
1322 }
1323
1324 PFN_PIXEL_KERNEL
1325 swr_compile_fs(struct swr_context *ctx, swr_jit_fs_key &key)
1326 {
1327    if (!ctx->fs->pipe.tokens)
1328       return NULL;
1329
1330    BuilderSWR builder(
1331       reinterpret_cast<JitManager *>(swr_screen(ctx->pipe.screen)->hJitMgr),
1332       "FS");
1333    PFN_PIXEL_KERNEL func = builder.CompileFS(ctx, key);
1334
1335    ctx->fs->map.insert(std::make_pair(key, make_unique<VariantFS>(builder.gallivm, func)));
1336    return func;
1337 }