src/gallium/auxiliary/draw/draw_llvm.c

   1 #include "draw_llvm.h"
   2
   3 #include "draw_context.h"
   4 #include "draw_vs.h"
   5
   6 #include "gallivm/lp_bld_arit.h"
   7 #include "gallivm/lp_bld_struct.h"
   8 #include "gallivm/lp_bld_type.h"
   9 #include "gallivm/lp_bld_flow.h"
  10 #include "gallivm/lp_bld_debug.h"
  11 #include "gallivm/lp_bld_tgsi.h"
  12 #include "gallivm/lp_bld_printf.h"
  13
  14 #include "tgsi/tgsi_exec.h"
  15 #include "tgsi/tgsi_dump.h"
  16
  17 #include "util/u_cpu_detect.h"
  18 #include "util/u_string.h"
  19 #include "util/u_pointer.h"
  20
  21 #include <llvm-c/Transforms/Scalar.h>
  22
  23 #define DEBUG_STORE 0
  24
  25
  26 /* generates the draw jit function */
  27 static void
  28 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *var);
  29 static void
  30 draw_llvm_generate_elts(struct draw_llvm *llvm, struct draw_llvm_variant *var);
  31
  32 static void
  33 init_globals(struct draw_llvm *llvm)
  34 {
  35    LLVMTypeRef texture_type;
  36
  37    /* struct draw_jit_texture */
  38    {
  39       LLVMTypeRef elem_types[4];
  40
  41       elem_types[DRAW_JIT_TEXTURE_WIDTH]  = LLVMInt32Type();
  42       elem_types[DRAW_JIT_TEXTURE_HEIGHT] = LLVMInt32Type();
  43       elem_types[DRAW_JIT_TEXTURE_STRIDE] = LLVMInt32Type();
  44       elem_types[DRAW_JIT_TEXTURE_DATA]   = LLVMPointerType(LLVMInt8Type(), 0);
  45
  46       texture_type = LLVMStructType(elem_types, Elements(elem_types), 0);
  47
  48       LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, width,
  49                              llvm->target, texture_type,
  50                              DRAW_JIT_TEXTURE_WIDTH);
  51       LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, height,
  52                              llvm->target, texture_type,
  53                              DRAW_JIT_TEXTURE_HEIGHT);
  54       LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, stride,
  55                              llvm->target, texture_type,
  56                              DRAW_JIT_TEXTURE_STRIDE);
  57       LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, data,
  58                              llvm->target, texture_type,
  59                              DRAW_JIT_TEXTURE_DATA);
  60       LP_CHECK_STRUCT_SIZE(struct draw_jit_texture,
  61                            llvm->target, texture_type);
  62
  63       LLVMAddTypeName(llvm->module, "texture", texture_type);
  64    }
  65
  66
  67    /* struct draw_jit_context */
  68    {
  69       LLVMTypeRef elem_types[3];
  70       LLVMTypeRef context_type;
  71
  72       elem_types[0] = LLVMPointerType(LLVMFloatType(), 0); /* vs_constants */
  73       elem_types[1] = LLVMPointerType(LLVMFloatType(), 0); /* vs_constants */
  74       elem_types[2] = LLVMArrayType(texture_type, PIPE_MAX_SAMPLERS); /* textures */
  75
  76       context_type = LLVMStructType(elem_types, Elements(elem_types), 0);
  77
  78       LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, vs_constants,
  79                              llvm->target, context_type, 0);
  80       LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, gs_constants,
  81                              llvm->target, context_type, 1);
  82       LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, textures,
  83                              llvm->target, context_type,
  84                              DRAW_JIT_CONTEXT_TEXTURES_INDEX);
  85       LP_CHECK_STRUCT_SIZE(struct draw_jit_context,
  86                            llvm->target, context_type);
  87
  88       LLVMAddTypeName(llvm->module, "draw_jit_context", context_type);
  89
  90       llvm->context_ptr_type = LLVMPointerType(context_type, 0);
  91    }
  92    {
  93       LLVMTypeRef buffer_ptr = LLVMPointerType(LLVMIntType(8), 0);
  94       llvm->buffer_ptr_type = LLVMPointerType(buffer_ptr, 0);
  95    }
  96    /* struct pipe_vertex_buffer */
  97    {
  98       LLVMTypeRef elem_types[4];
  99       LLVMTypeRef vb_type;
 100
 101       elem_types[0] = LLVMInt32Type();
 102       elem_types[1] = LLVMInt32Type();
 103       elem_types[2] = LLVMInt32Type();
 104       elem_types[3] = LLVMPointerType(LLVMOpaqueType(), 0); /* vs_constants */
 105
 106       vb_type = LLVMStructType(elem_types, Elements(elem_types), 0);
 107
 108       LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, stride,
 109                              llvm->target, vb_type, 0);
 110       LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, buffer_offset,
 111                              llvm->target, vb_type, 2);
 112       LP_CHECK_STRUCT_SIZE(struct pipe_vertex_buffer,
 113                            llvm->target, vb_type);
 114
 115       LLVMAddTypeName(llvm->module, "pipe_vertex_buffer", vb_type);
 116
 117       llvm->vb_ptr_type = LLVMPointerType(vb_type, 0);
 118    }
 119 }
 120
 121 static LLVMTypeRef
 122 create_vertex_header(struct draw_llvm *llvm, int data_elems)
 123 {
 124    /* struct vertex_header */
 125    LLVMTypeRef elem_types[3];
 126    LLVMTypeRef vertex_header;
 127    char struct_name[24];
 128
 129    util_snprintf(struct_name, 23, "vertex_header%d", data_elems);
 130
 131    elem_types[0]  = LLVMIntType(32);
 132    elem_types[1]  = LLVMArrayType(LLVMFloatType(), 4);
 133    elem_types[2]  = LLVMArrayType(elem_types[1], data_elems);
 134
 135    vertex_header = LLVMStructType(elem_types, Elements(elem_types), 0);
 136
 137    /* these are bit-fields and we can't take address of them
 138       LP_CHECK_MEMBER_OFFSET(struct vertex_header, clipmask,
 139       llvm->target, vertex_header,
 140       DRAW_JIT_VERTEX_CLIPMASK);
 141       LP_CHECK_MEMBER_OFFSET(struct vertex_header, edgeflag,
 142       llvm->target, vertex_header,
 143       DRAW_JIT_VERTEX_EDGEFLAG);
 144       LP_CHECK_MEMBER_OFFSET(struct vertex_header, pad,
 145       llvm->target, vertex_header,
 146       DRAW_JIT_VERTEX_PAD);
 147       LP_CHECK_MEMBER_OFFSET(struct vertex_header, vertex_id,
 148       llvm->target, vertex_header,
 149       DRAW_JIT_VERTEX_VERTEX_ID);
 150    */
 151    LP_CHECK_MEMBER_OFFSET(struct vertex_header, clip,
 152                           llvm->target, vertex_header,
 153                           DRAW_JIT_VERTEX_CLIP);
 154    LP_CHECK_MEMBER_OFFSET(struct vertex_header, data,
 155                           llvm->target, vertex_header,
 156                           DRAW_JIT_VERTEX_DATA);
 157
 158    LLVMAddTypeName(llvm->module, struct_name, vertex_header);
 159
 160    return LLVMPointerType(vertex_header, 0);
 161 }
 162
 163 struct draw_llvm *
 164 draw_llvm_create(struct draw_context *draw)
 165 {
 166    struct draw_llvm *llvm;
 167
 168 #ifdef PIPE_ARCH_X86
 169    util_cpu_detect();
 170    /* require SSE2 due to LLVM PR6960. */
 171    if (!util_cpu_caps.has_sse2)
 172        return NULL;
 173 #endif
 174
 175    llvm = CALLOC_STRUCT( draw_llvm );
 176    if (!llvm)
 177       return NULL;
 178
 179    llvm->draw = draw;
 180    llvm->engine = draw->engine;
 181
 182    debug_assert(llvm->engine);
 183
 184    llvm->module = LLVMModuleCreateWithName("draw_llvm");
 185    llvm->provider = LLVMCreateModuleProviderForExistingModule(llvm->module);
 186
 187    LLVMAddModuleProvider(llvm->engine, llvm->provider);
 188
 189    llvm->target = LLVMGetExecutionEngineTargetData(llvm->engine);
 190
 191    llvm->pass = LLVMCreateFunctionPassManager(llvm->provider);
 192    LLVMAddTargetData(llvm->target, llvm->pass);
 193
 194    if ((gallivm_debug & GALLIVM_DEBUG_NO_OPT) == 0) {
 195       /* These are the passes currently listed in llvm-c/Transforms/Scalar.h,
 196        * but there are more on SVN. */
 197       /* TODO: Add more passes */
 198       LLVMAddCFGSimplificationPass(llvm->pass);
 199       LLVMAddPromoteMemoryToRegisterPass(llvm->pass);
 200       LLVMAddConstantPropagationPass(llvm->pass);
 201       if(util_cpu_caps.has_sse4_1) {
 202          /* FIXME: There is a bug in this pass, whereby the combination of fptosi
 203           * and sitofp (necessary for trunc/floor/ceil/round implementation)
 204           * somehow becomes invalid code.
 205           */
 206          LLVMAddInstructionCombiningPass(llvm->pass);
 207       }
 208       LLVMAddGVNPass(llvm->pass);
 209    } else {
 210       /* We need at least this pass to prevent the backends to fail in
 211        * unexpected ways.
 212        */
 213       LLVMAddPromoteMemoryToRegisterPass(llvm->pass);
 214    }
 215
 216    init_globals(llvm);
 217
 218    if (gallivm_debug & GALLIVM_DEBUG_IR) {
 219       LLVMDumpModule(llvm->module);
 220    }
 221
 222    return llvm;
 223 }
 224
 225 void
 226 draw_llvm_destroy(struct draw_llvm *llvm)
 227 {
 228    LLVMDisposePassManager(llvm->pass);
 229
 230    FREE(llvm);
 231 }
 232
 233 struct draw_llvm_variant *
 234 draw_llvm_prepare(struct draw_llvm *llvm, int num_inputs)
 235 {
 236    struct draw_llvm_variant *variant = MALLOC(sizeof(struct draw_llvm_variant));
 237
 238    draw_llvm_make_variant_key(llvm, &variant->key);
 239
 240    llvm->vertex_header_ptr_type = create_vertex_header(llvm, num_inputs);
 241
 242    draw_llvm_generate(llvm, variant);
 243    draw_llvm_generate_elts(llvm, variant);
 244
 245    return variant;
 246 }
 247
 248 static void
 249 generate_vs(struct draw_llvm *llvm,
 250             LLVMBuilderRef builder,
 251             LLVMValueRef (*outputs)[NUM_CHANNELS],
 252             const LLVMValueRef (*inputs)[NUM_CHANNELS],
 253             LLVMValueRef context_ptr)
 254 {
 255    const struct tgsi_token *tokens = llvm->draw->vs.vertex_shader->state.tokens;
 256    struct lp_type vs_type;
 257    LLVMValueRef consts_ptr = draw_jit_context_vs_constants(builder, context_ptr);
 258
 259    memset(&vs_type, 0, sizeof vs_type);
 260    vs_type.floating = TRUE; /* floating point values */
 261    vs_type.sign = TRUE;     /* values are signed */
 262    vs_type.norm = FALSE;    /* values are not limited to [0,1] or [-1,1] */
 263    vs_type.width = 32;      /* 32-bit float */
 264    vs_type.length = 4;      /* 4 elements per vector */
 265 #if 0
 266    num_vs = 4;              /* number of vertices per block */
 267 #endif
 268
 269    if (gallivm_debug & GALLIVM_DEBUG_IR) {
 270       tgsi_dump(tokens, 0);
 271    }
 272
 273    lp_build_tgsi_soa(builder,
 274                      tokens,
 275                      vs_type,
 276                      NULL /*struct lp_build_mask_context *mask*/,
 277                      consts_ptr,
 278                      NULL /*pos*/,
 279                      inputs,
 280                      outputs,
 281                      NULL/*sampler*/,
 282                      &llvm->draw->vs.vertex_shader->info);
 283 }
 284
 285 #if DEBUG_STORE
 286 static void print_vectorf(LLVMBuilderRef builder,
 287                          LLVMValueRef vec)
 288 {
 289    LLVMValueRef val[4];
 290    val[0] = LLVMBuildExtractElement(builder, vec,
 291                                     LLVMConstInt(LLVMInt32Type(), 0, 0), "");
 292    val[1] = LLVMBuildExtractElement(builder, vec,
 293                                     LLVMConstInt(LLVMInt32Type(), 1, 0), "");
 294    val[2] = LLVMBuildExtractElement(builder, vec,
 295                                     LLVMConstInt(LLVMInt32Type(), 2, 0), "");
 296    val[3] = LLVMBuildExtractElement(builder, vec,
 297                                     LLVMConstInt(LLVMInt32Type(), 3, 0), "");
 298    lp_build_printf(builder, "vector = [%f, %f, %f, %f]\n",
 299                    val[0], val[1], val[2], val[3]);
 300 }
 301 #endif
 302
 303 static void
 304 generate_fetch(LLVMBuilderRef builder,
 305                LLVMValueRef vbuffers_ptr,
 306                LLVMValueRef *res,
 307                struct pipe_vertex_element *velem,
 308                LLVMValueRef vbuf,
 309                LLVMValueRef index)
 310 {
 311    LLVMValueRef indices = LLVMConstInt(LLVMInt64Type(), velem->vertex_buffer_index, 0);
 312    LLVMValueRef vbuffer_ptr = LLVMBuildGEP(builder, vbuffers_ptr,
 313                                            &indices, 1, "");
 314    LLVMValueRef vb_stride = draw_jit_vbuffer_stride(builder, vbuf);
 315    LLVMValueRef vb_max_index = draw_jit_vbuffer_max_index(builder, vbuf);
 316    LLVMValueRef vb_buffer_offset = draw_jit_vbuffer_offset(builder, vbuf);
 317    LLVMValueRef cond;
 318    LLVMValueRef stride;
 319
 320    cond = LLVMBuildICmp(builder, LLVMIntULE, index, vb_max_index, "");
 321
 322    index = LLVMBuildSelect(builder, cond, index, vb_max_index, "");
 323
 324    stride = LLVMBuildMul(builder, vb_stride, index, "");
 325
 326    vbuffer_ptr = LLVMBuildLoad(builder, vbuffer_ptr, "vbuffer");
 327
 328    stride = LLVMBuildAdd(builder, stride,
 329                          vb_buffer_offset,
 330                          "");
 331    stride = LLVMBuildAdd(builder, stride,
 332                          LLVMConstInt(LLVMInt32Type(), velem->src_offset, 0),
 333                          "");
 334
 335    /*lp_build_printf(builder, "vbuf index = %d, stride is %d\n", indices, stride);*/
 336    vbuffer_ptr = LLVMBuildGEP(builder, vbuffer_ptr, &stride, 1, "");
 337
 338    *res = draw_llvm_translate_from(builder, vbuffer_ptr, velem->src_format);
 339 }
 340
 341 static LLVMValueRef
 342 aos_to_soa(LLVMBuilderRef builder,
 343            LLVMValueRef val0,
 344            LLVMValueRef val1,
 345            LLVMValueRef val2,
 346            LLVMValueRef val3,
 347            LLVMValueRef channel)
 348 {
 349    LLVMValueRef ex, res;
 350
 351    ex = LLVMBuildExtractElement(builder, val0,
 352                                 channel, "");
 353    res = LLVMBuildInsertElement(builder,
 354                                 LLVMConstNull(LLVMTypeOf(val0)),
 355                                 ex,
 356                                 LLVMConstInt(LLVMInt32Type(), 0, 0),
 357                                 "");
 358
 359    ex = LLVMBuildExtractElement(builder, val1,
 360                                 channel, "");
 361    res = LLVMBuildInsertElement(builder,
 362                                 res, ex,
 363                                 LLVMConstInt(LLVMInt32Type(), 1, 0),
 364                                 "");
 365
 366    ex = LLVMBuildExtractElement(builder, val2,
 367                                 channel, "");
 368    res = LLVMBuildInsertElement(builder,
 369                                 res, ex,
 370                                 LLVMConstInt(LLVMInt32Type(), 2, 0),
 371                                 "");
 372
 373    ex = LLVMBuildExtractElement(builder, val3,
 374                                 channel, "");
 375    res = LLVMBuildInsertElement(builder,
 376                                 res, ex,
 377                                 LLVMConstInt(LLVMInt32Type(), 3, 0),
 378                                 "");
 379
 380    return res;
 381 }
 382
 383 static void
 384 soa_to_aos(LLVMBuilderRef builder,
 385            LLVMValueRef soa[NUM_CHANNELS],
 386            LLVMValueRef aos[NUM_CHANNELS])
 387 {
 388    LLVMValueRef comp;
 389    int i = 0;
 390
 391    debug_assert(NUM_CHANNELS == 4);
 392
 393    aos[0] = LLVMConstNull(LLVMTypeOf(soa[0]));
 394    aos[1] = aos[2] = aos[3] = aos[0];
 395
 396    for (i = 0; i < NUM_CHANNELS; ++i) {
 397       LLVMValueRef channel = LLVMConstInt(LLVMInt32Type(), i, 0);
 398
 399       comp = LLVMBuildExtractElement(builder, soa[i],
 400                                      LLVMConstInt(LLVMInt32Type(), 0, 0), "");
 401       aos[0] = LLVMBuildInsertElement(builder, aos[0], comp, channel, "");
 402
 403       comp = LLVMBuildExtractElement(builder, soa[i],
 404                                      LLVMConstInt(LLVMInt32Type(), 1, 0), "");
 405       aos[1] = LLVMBuildInsertElement(builder, aos[1], comp, channel, "");
 406
 407       comp = LLVMBuildExtractElement(builder, soa[i],
 408                                      LLVMConstInt(LLVMInt32Type(), 2, 0), "");
 409       aos[2] = LLVMBuildInsertElement(builder, aos[2], comp, channel, "");
 410
 411       comp = LLVMBuildExtractElement(builder, soa[i],
 412                                      LLVMConstInt(LLVMInt32Type(), 3, 0), "");
 413       aos[3] = LLVMBuildInsertElement(builder, aos[3], comp, channel, "");
 414
 415    }
 416 }
 417
 418 static void
 419 convert_to_soa(LLVMBuilderRef builder,
 420                LLVMValueRef (*aos)[NUM_CHANNELS],
 421                LLVMValueRef (*soa)[NUM_CHANNELS],
 422                int num_attribs)
 423 {
 424    int i;
 425
 426    debug_assert(NUM_CHANNELS == 4);
 427
 428    for (i = 0; i < num_attribs; ++i) {
 429       LLVMValueRef val0 = aos[i][0];
 430       LLVMValueRef val1 = aos[i][1];
 431       LLVMValueRef val2 = aos[i][2];
 432       LLVMValueRef val3 = aos[i][3];
 433
 434       soa[i][0] = aos_to_soa(builder, val0, val1, val2, val3,
 435                              LLVMConstInt(LLVMInt32Type(), 0, 0));
 436       soa[i][1] = aos_to_soa(builder, val0, val1, val2, val3,
 437                              LLVMConstInt(LLVMInt32Type(), 1, 0));
 438       soa[i][2] = aos_to_soa(builder, val0, val1, val2, val3,
 439                              LLVMConstInt(LLVMInt32Type(), 2, 0));
 440       soa[i][3] = aos_to_soa(builder, val0, val1, val2, val3,
 441                              LLVMConstInt(LLVMInt32Type(), 3, 0));
 442    }
 443 }
 444
 445 static void
 446 store_aos(LLVMBuilderRef builder,
 447           LLVMValueRef io_ptr,
 448           LLVMValueRef index,
 449           LLVMValueRef value)
 450 {
 451    LLVMValueRef id_ptr = draw_jit_header_id(builder, io_ptr);
 452    LLVMValueRef data_ptr = draw_jit_header_data(builder, io_ptr);
 453    LLVMValueRef indices[3];
 454
 455    indices[0] = LLVMConstInt(LLVMInt32Type(), 0, 0);
 456    indices[1] = index;
 457    indices[2] = LLVMConstInt(LLVMInt32Type(), 0, 0);
 458
 459    /* undefined vertex */
 460    LLVMBuildStore(builder, LLVMConstInt(LLVMInt32Type(),
 461                                         0xffff, 0), id_ptr);
 462
 463 #if DEBUG_STORE
 464    lp_build_printf(builder, "    ---- %p storing attribute %d (io = %p)\n", data_ptr, index, io_ptr);
 465 #endif
 466 #if 0
 467    /*lp_build_printf(builder, " ---- %p storing at %d (%p)  ", io_ptr, index, data_ptr);
 468      print_vectorf(builder, value);*/
 469    data_ptr = LLVMBuildBitCast(builder, data_ptr,
 470                                LLVMPointerType(LLVMArrayType(LLVMVectorType(LLVMFloatType(), 4), 0), 0),
 471                                "datavec");
 472    data_ptr = LLVMBuildGEP(builder, data_ptr, indices, 2, "");
 473
 474    LLVMBuildStore(builder, value, data_ptr);
 475 #else
 476    {
 477       LLVMValueRef x, y, z, w;
 478       LLVMValueRef idx0, idx1, idx2, idx3;
 479       LLVMValueRef gep0, gep1, gep2, gep3;
 480       data_ptr = LLVMBuildGEP(builder, data_ptr, indices, 3, "");
 481
 482       idx0 = LLVMConstInt(LLVMInt32Type(), 0, 0);
 483       idx1 = LLVMConstInt(LLVMInt32Type(), 1, 0);
 484       idx2 = LLVMConstInt(LLVMInt32Type(), 2, 0);
 485       idx3 = LLVMConstInt(LLVMInt32Type(), 3, 0);
 486
 487       x = LLVMBuildExtractElement(builder, value,
 488                                   idx0, "");
 489       y = LLVMBuildExtractElement(builder, value,
 490                                   idx1, "");
 491       z = LLVMBuildExtractElement(builder, value,
 492                                   idx2, "");
 493       w = LLVMBuildExtractElement(builder, value,
 494                                   idx3, "");
 495
 496       gep0 = LLVMBuildGEP(builder, data_ptr, &idx0, 1, "");
 497       gep1 = LLVMBuildGEP(builder, data_ptr, &idx1, 1, "");
 498       gep2 = LLVMBuildGEP(builder, data_ptr, &idx2, 1, "");
 499       gep3 = LLVMBuildGEP(builder, data_ptr, &idx3, 1, "");
 500
 501       /*lp_build_printf(builder, "##### x = %f (%p), y = %f (%p), z = %f (%p), w = %f (%p)\n",
 502         x, gep0, y, gep1, z, gep2, w, gep3);*/
 503       LLVMBuildStore(builder, x, gep0);
 504       LLVMBuildStore(builder, y, gep1);
 505       LLVMBuildStore(builder, z, gep2);
 506       LLVMBuildStore(builder, w, gep3);
 507    }
 508 #endif
 509 }
 510
 511 static void
 512 store_aos_array(LLVMBuilderRef builder,
 513                 LLVMValueRef io_ptr,
 514                 LLVMValueRef aos[NUM_CHANNELS],
 515                 int attrib,
 516                 int num_outputs)
 517 {
 518    LLVMValueRef attr_index = LLVMConstInt(LLVMInt32Type(), attrib, 0);
 519    LLVMValueRef ind0 = LLVMConstInt(LLVMInt32Type(), 0, 0);
 520    LLVMValueRef ind1 = LLVMConstInt(LLVMInt32Type(), 1, 0);
 521    LLVMValueRef ind2 = LLVMConstInt(LLVMInt32Type(), 2, 0);
 522    LLVMValueRef ind3 = LLVMConstInt(LLVMInt32Type(), 3, 0);
 523    LLVMValueRef io0_ptr, io1_ptr, io2_ptr, io3_ptr;
 524
 525    debug_assert(NUM_CHANNELS == 4);
 526
 527    io0_ptr = LLVMBuildGEP(builder, io_ptr,
 528                           &ind0, 1, "");
 529    io1_ptr = LLVMBuildGEP(builder, io_ptr,
 530                           &ind1, 1, "");
 531    io2_ptr = LLVMBuildGEP(builder, io_ptr,
 532                           &ind2, 1, "");
 533    io3_ptr = LLVMBuildGEP(builder, io_ptr,
 534                           &ind3, 1, "");
 535
 536 #if DEBUG_STORE
 537    lp_build_printf(builder, "   io = %p, indexes[%d, %d, %d, %d]\n",
 538                    io_ptr, ind0, ind1, ind2, ind3);
 539 #endif
 540
 541    store_aos(builder, io0_ptr, attr_index, aos[0]);
 542    store_aos(builder, io1_ptr, attr_index, aos[1]);
 543    store_aos(builder, io2_ptr, attr_index, aos[2]);
 544    store_aos(builder, io3_ptr, attr_index, aos[3]);
 545 }
 546
 547 static void
 548 convert_to_aos(LLVMBuilderRef builder,
 549                LLVMValueRef io,
 550                LLVMValueRef (*outputs)[NUM_CHANNELS],
 551                int num_outputs,
 552                int max_vertices)
 553 {
 554    unsigned chan, attrib;
 555
 556 #if DEBUG_STORE
 557    lp_build_printf(builder, "   # storing begin\n");
 558 #endif
 559    for (attrib = 0; attrib < num_outputs; ++attrib) {
 560       LLVMValueRef soa[4];
 561       LLVMValueRef aos[4];
 562       for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 563          if(outputs[attrib][chan]) {
 564             LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
 565             lp_build_name(out, "output%u.%c", attrib, "xyzw"[chan]);
 566             /*lp_build_printf(builder, "output %d : %d ",
 567                             LLVMConstInt(LLVMInt32Type(), attrib, 0),
 568                             LLVMConstInt(LLVMInt32Type(), chan, 0));
 569               print_vectorf(builder, out);*/
 570             soa[chan] = out;
 571          } else
 572             soa[chan] = 0;
 573       }
 574       soa_to_aos(builder, soa, aos);
 575       store_aos_array(builder,
 576                       io,
 577                       aos,
 578                       attrib,
 579                       num_outputs);
 580    }
 581 #if DEBUG_STORE
 582    lp_build_printf(builder, "   # storing end\n");
 583 #endif
 584 }
 585
 586 static void
 587 draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *variant)
 588 {
 589    LLVMTypeRef arg_types[7];
 590    LLVMTypeRef func_type;
 591    LLVMValueRef context_ptr;
 592    LLVMBasicBlockRef block;
 593    LLVMBuilderRef builder;
 594    LLVMValueRef start, end, count, stride, step, io_itr;
 595    LLVMValueRef io_ptr, vbuffers_ptr, vb_ptr;
 596    struct draw_context *draw = llvm->draw;
 597    unsigned i, j;
 598    struct lp_build_context bld;
 599    struct lp_build_loop_state lp_loop;
 600    struct lp_type vs_type = lp_type_float_vec(32);
 601    const int max_vertices = 4;
 602    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][NUM_CHANNELS];
 603    void *code;
 604
 605    arg_types[0] = llvm->context_ptr_type;           /* context */
 606    arg_types[1] = llvm->vertex_header_ptr_type;     /* vertex_header */
 607    arg_types[2] = llvm->buffer_ptr_type;            /* vbuffers */
 608    arg_types[3] = LLVMInt32Type();                  /* start */
 609    arg_types[4] = LLVMInt32Type();                  /* count */
 610    arg_types[5] = LLVMInt32Type();                  /* stride */
 611    arg_types[6] = llvm->vb_ptr_type;                /* pipe_vertex_buffer's */
 612
 613    func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
 614
 615    variant->function = LLVMAddFunction(llvm->module, "draw_llvm_shader", func_type);
 616    LLVMSetFunctionCallConv(variant->function, LLVMCCallConv);
 617    for(i = 0; i < Elements(arg_types); ++i)
 618       if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
 619          LLVMAddAttribute(LLVMGetParam(variant->function, i), LLVMNoAliasAttribute);
 620
 621    context_ptr  = LLVMGetParam(variant->function, 0);
 622    io_ptr       = LLVMGetParam(variant->function, 1);
 623    vbuffers_ptr = LLVMGetParam(variant->function, 2);
 624    start        = LLVMGetParam(variant->function, 3);
 625    count        = LLVMGetParam(variant->function, 4);
 626    stride       = LLVMGetParam(variant->function, 5);
 627    vb_ptr       = LLVMGetParam(variant->function, 6);
 628
 629    lp_build_name(context_ptr, "context");
 630    lp_build_name(io_ptr, "io");
 631    lp_build_name(vbuffers_ptr, "vbuffers");
 632    lp_build_name(start, "start");
 633    lp_build_name(count, "count");
 634    lp_build_name(stride, "stride");
 635    lp_build_name(vb_ptr, "vb");
 636
 637    /*
 638     * Function body
 639     */
 640
 641    block = LLVMAppendBasicBlock(variant->function, "entry");
 642    builder = LLVMCreateBuilder();
 643    LLVMPositionBuilderAtEnd(builder, block);
 644
 645    lp_build_context_init(&bld, builder, vs_type);
 646
 647    end = lp_build_add(&bld, start, count);
 648
 649    step = LLVMConstInt(LLVMInt32Type(), max_vertices, 0);
 650
 651 #if DEBUG_STORE
 652    lp_build_printf(builder, "start = %d, end = %d, step = %d\n",
 653                    start, end, step);
 654 #endif
 655    lp_build_loop_begin(builder, start, &lp_loop);
 656    {
 657       LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][NUM_CHANNELS];
 658       LLVMValueRef aos_attribs[PIPE_MAX_SHADER_INPUTS][NUM_CHANNELS] = { { 0 } };
 659       LLVMValueRef io;
 660       const LLVMValueRef (*ptr_aos)[NUM_CHANNELS];
 661
 662       io_itr = LLVMBuildSub(builder, lp_loop.counter, start, "");
 663       io = LLVMBuildGEP(builder, io_ptr, &io_itr, 1, "");
 664 #if DEBUG_STORE
 665       lp_build_printf(builder, " --- io %d = %p, loop counter %d\n",
 666                       io_itr, io, lp_loop.counter);
 667 #endif
 668       for (i = 0; i < NUM_CHANNELS; ++i) {
 669          LLVMValueRef true_index = LLVMBuildAdd(
 670             builder,
 671             lp_loop.counter,
 672             LLVMConstInt(LLVMInt32Type(), i, 0), "");
 673          for (j = 0; j < draw->pt.nr_vertex_elements; ++j) {
 674             struct pipe_vertex_element *velem = &draw->pt.vertex_element[j];
 675             LLVMValueRef vb_index = LLVMConstInt(LLVMInt32Type(),
 676                                                  velem->vertex_buffer_index,
 677                                                  0);
 678             LLVMValueRef vb = LLVMBuildGEP(builder, vb_ptr,
 679                                            &vb_index, 1, "");
 680             generate_fetch(builder, vbuffers_ptr,
 681                            &aos_attribs[j][i], velem, vb, true_index);
 682          }
 683       }
 684       convert_to_soa(builder, aos_attribs, inputs,
 685                      draw->pt.nr_vertex_elements);
 686
 687       ptr_aos = (const LLVMValueRef (*)[NUM_CHANNELS]) inputs;
 688       generate_vs(llvm,
 689                   builder,
 690                   outputs,
 691                   ptr_aos,
 692                   context_ptr);
 693
 694       convert_to_aos(builder, io, outputs,
 695                      draw->vs.vertex_shader->info.num_outputs,
 696                      max_vertices);
 697    }
 698    lp_build_loop_end_cond(builder, end, step, LLVMIntUGE, &lp_loop);
 699
 700    LLVMBuildRetVoid(builder);
 701
 702    LLVMDisposeBuilder(builder);
 703
 704    /*
 705     * Translate the LLVM IR into machine code.
 706     */
 707 #ifdef DEBUG
 708    if(LLVMVerifyFunction(variant->function, LLVMPrintMessageAction)) {
 709       lp_debug_dump_value(variant->function);
 710       assert(0);
 711    }
 712 #endif
 713
 714    LLVMRunFunctionPassManager(llvm->pass, variant->function);
 715
 716    if (gallivm_debug & GALLIVM_DEBUG_IR) {
 717       lp_debug_dump_value(variant->function);
 718       debug_printf("\n");
 719    }
 720
 721    code = LLVMGetPointerToGlobal(llvm->draw->engine, variant->function);
 722    variant->jit_func = (draw_jit_vert_func)pointer_to_func(code);
 723
 724    if (gallivm_debug & GALLIVM_DEBUG_ASM) {
 725       lp_disassemble(code);
 726    }
 727 }
 728
 729
 730 static void
 731 draw_llvm_generate_elts(struct draw_llvm *llvm, struct draw_llvm_variant *variant)
 732 {
 733    LLVMTypeRef arg_types[7];
 734    LLVMTypeRef func_type;
 735    LLVMValueRef context_ptr;
 736    LLVMBasicBlockRef block;
 737    LLVMBuilderRef builder;
 738    LLVMValueRef fetch_elts, fetch_count, stride, step, io_itr;
 739    LLVMValueRef io_ptr, vbuffers_ptr, vb_ptr;
 740    struct draw_context *draw = llvm->draw;
 741    unsigned i, j;
 742    struct lp_build_context bld;
 743    struct lp_build_context bld_int;
 744    struct lp_build_loop_state lp_loop;
 745    struct lp_type vs_type = lp_type_float_vec(32);
 746    const int max_vertices = 4;
 747    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][NUM_CHANNELS];
 748    LLVMValueRef fetch_max;
 749    void *code;
 750
 751    arg_types[0] = llvm->context_ptr_type;               /* context */
 752    arg_types[1] = llvm->vertex_header_ptr_type;         /* vertex_header */
 753    arg_types[2] = llvm->buffer_ptr_type;                /* vbuffers */
 754    arg_types[3] = LLVMPointerType(LLVMInt32Type(), 0);  /* fetch_elts * */
 755    arg_types[4] = LLVMInt32Type();                      /* fetch_count */
 756    arg_types[5] = LLVMInt32Type();                      /* stride */
 757    arg_types[6] = llvm->vb_ptr_type;                    /* pipe_vertex_buffer's */
 758
 759    func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
 760
 761    variant->function_elts = LLVMAddFunction(llvm->module, "draw_llvm_shader_elts", func_type);
 762    LLVMSetFunctionCallConv(variant->function_elts, LLVMCCallConv);
 763    for(i = 0; i < Elements(arg_types); ++i)
 764       if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
 765          LLVMAddAttribute(LLVMGetParam(variant->function_elts, i), LLVMNoAliasAttribute);
 766
 767    context_ptr  = LLVMGetParam(variant->function_elts, 0);
 768    io_ptr       = LLVMGetParam(variant->function_elts, 1);
 769    vbuffers_ptr = LLVMGetParam(variant->function_elts, 2);
 770    fetch_elts   = LLVMGetParam(variant->function_elts, 3);
 771    fetch_count  = LLVMGetParam(variant->function_elts, 4);
 772    stride       = LLVMGetParam(variant->function_elts, 5);
 773    vb_ptr       = LLVMGetParam(variant->function_elts, 6);
 774
 775    lp_build_name(context_ptr, "context");
 776    lp_build_name(io_ptr, "io");
 777    lp_build_name(vbuffers_ptr, "vbuffers");
 778    lp_build_name(fetch_elts, "fetch_elts");
 779    lp_build_name(fetch_count, "fetch_count");
 780    lp_build_name(stride, "stride");
 781    lp_build_name(vb_ptr, "vb");
 782
 783    /*
 784     * Function body
 785     */
 786
 787    block = LLVMAppendBasicBlock(variant->function_elts, "entry");
 788    builder = LLVMCreateBuilder();
 789    LLVMPositionBuilderAtEnd(builder, block);
 790
 791    lp_build_context_init(&bld, builder, vs_type);
 792    lp_build_context_init(&bld_int, builder, lp_type_int(32));
 793
 794    step = LLVMConstInt(LLVMInt32Type(), max_vertices, 0);
 795
 796    fetch_max = LLVMBuildSub(builder, fetch_count,
 797                             LLVMConstInt(LLVMInt32Type(), 1, 0),
 798                             "fetch_max");
 799
 800    lp_build_loop_begin(builder, LLVMConstInt(LLVMInt32Type(), 0, 0), &lp_loop);
 801    {
 802       LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][NUM_CHANNELS];
 803       LLVMValueRef aos_attribs[PIPE_MAX_SHADER_INPUTS][NUM_CHANNELS] = { { 0 } };
 804       LLVMValueRef io;
 805       const LLVMValueRef (*ptr_aos)[NUM_CHANNELS];
 806
 807       io_itr = lp_loop.counter;
 808       io = LLVMBuildGEP(builder, io_ptr, &io_itr, 1, "");
 809 #if DEBUG_STORE
 810       lp_build_printf(builder, " --- io %d = %p, loop counter %d\n",
 811                       io_itr, io, lp_loop.counter);
 812 #endif
 813       for (i = 0; i < NUM_CHANNELS; ++i) {
 814          LLVMValueRef true_index = LLVMBuildAdd(
 815             builder,
 816             lp_loop.counter,
 817             LLVMConstInt(LLVMInt32Type(), i, 0), "");
 818          LLVMValueRef fetch_ptr;
 819
 820          /* make sure we're not out of bounds which can happen
 821           * if fetch_count % 4 != 0, because on the last iteration
 822           * a few of the 4 vertex fetches will be out of bounds */
 823          true_index = lp_build_min(&bld_int, true_index, fetch_max);
 824
 825          fetch_ptr = LLVMBuildGEP(builder, fetch_elts,
 826                                   &true_index, 1, "");
 827          true_index = LLVMBuildLoad(builder, fetch_ptr, "fetch_elt");
 828          for (j = 0; j < draw->pt.nr_vertex_elements; ++j) {
 829             struct pipe_vertex_element *velem = &draw->pt.vertex_element[j];
 830             LLVMValueRef vb_index = LLVMConstInt(LLVMInt32Type(),
 831                                                  velem->vertex_buffer_index,
 832                                                  0);
 833             LLVMValueRef vb = LLVMBuildGEP(builder, vb_ptr,
 834                                            &vb_index, 1, "");
 835             generate_fetch(builder, vbuffers_ptr,
 836                            &aos_attribs[j][i], velem, vb, true_index);
 837          }
 838       }
 839       convert_to_soa(builder, aos_attribs, inputs,
 840                      draw->pt.nr_vertex_elements);
 841
 842       ptr_aos = (const LLVMValueRef (*)[NUM_CHANNELS]) inputs;
 843       generate_vs(llvm,
 844                   builder,
 845                   outputs,
 846                   ptr_aos,
 847                   context_ptr);
 848
 849       convert_to_aos(builder, io, outputs,
 850                      draw->vs.vertex_shader->info.num_outputs,
 851                      max_vertices);
 852    }
 853    lp_build_loop_end_cond(builder, fetch_count, step, LLVMIntUGE, &lp_loop);
 854
 855    LLVMBuildRetVoid(builder);
 856
 857    LLVMDisposeBuilder(builder);
 858
 859    /*
 860     * Translate the LLVM IR into machine code.
 861     */
 862 #ifdef DEBUG
 863    if(LLVMVerifyFunction(variant->function_elts, LLVMPrintMessageAction)) {
 864       lp_debug_dump_value(variant->function_elts);
 865       assert(0);
 866    }
 867 #endif
 868
 869    LLVMRunFunctionPassManager(llvm->pass, variant->function_elts);
 870
 871    if (gallivm_debug & GALLIVM_DEBUG_IR) {
 872       lp_debug_dump_value(variant->function_elts);
 873       debug_printf("\n");
 874    }
 875
 876    code = LLVMGetPointerToGlobal(llvm->draw->engine, variant->function_elts);
 877    variant->jit_func_elts = (draw_jit_vert_func_elts)pointer_to_func(code);
 878
 879    if (gallivm_debug & GALLIVM_DEBUG_ASM) {
 880       lp_disassemble(code);
 881    }
 882 }
 883
 884 void
 885 draw_llvm_make_variant_key(struct draw_llvm *llvm,
 886                            struct draw_llvm_variant_key *key)
 887 {
 888    memset(key, 0, sizeof(struct draw_llvm_variant_key));
 889
 890    key->nr_vertex_elements = llvm->draw->pt.nr_vertex_elements;
 891
 892    memcpy(key->vertex_element,
 893           llvm->draw->pt.vertex_element,
 894           sizeof(struct pipe_vertex_element) * key->nr_vertex_elements);
 895
 896    memcpy(&key->vs,
 897           &llvm->draw->vs.vertex_shader->state,
 898           sizeof(struct pipe_shader_state));
 899 }