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