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