src/gallium/drivers/llvmpipe/lp_state_fs.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2009 VMware, Inc.
   4  * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas.
   5  * All Rights Reserved.
   6  *
   7  * Permission is hereby granted, free of charge, to any person obtaining a
   8  * copy of this software and associated documentation files (the
   9  * "Software"), to deal in the Software without restriction, including
  10  * without limitation the rights to use, copy, modify, merge, publish,
  11  * distribute, sub license, and/or sell copies of the Software, and to
  12  * permit persons to whom the Software is furnished to do so, subject to
  13  * the following conditions:
  14  *
  15  * The above copyright notice and this permission notice (including the
  16  * next paragraph) shall be included in all copies or substantial portions
  17  * of the Software.
  18  *
  19  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  20  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  21  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
  22  * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
  23  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
  24  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
  25  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  26  *
  27  **************************************************************************/
  28
  29 /**
  30  * @file
  31  * Code generate the whole fragment pipeline.
  32  *
  33  * The fragment pipeline consists of the following stages:
  34  * - early depth test
  35  * - fragment shader
  36  * - alpha test
  37  * - depth/stencil test
  38  * - blending
  39  *
  40  * This file has only the glue to assemble the fragment pipeline.  The actual
  41  * plumbing of converting Gallium state into LLVM IR is done elsewhere, in the
  42  * lp_bld_*.[ch] files, and in a complete generic and reusable way. Here we
  43  * muster the LLVM JIT execution engine to create a function that follows an
  44  * established binary interface and that can be called from C directly.
  45  *
  46  * A big source of complexity here is that we often want to run different
  47  * stages with different precisions and data types and precisions. For example,
  48  * the fragment shader needs typically to be done in floats, but the
  49  * depth/stencil test and blending is better done in the type that most closely
  50  * matches the depth/stencil and color buffer respectively.
  51  *
  52  * Since the width of a SIMD vector register stays the same regardless of the
  53  * element type, different types imply different number of elements, so we must
  54  * code generate more instances of the stages with larger types to be able to
  55  * feed/consume the stages with smaller types.
  56  *
  57  * @author Jose Fonseca <jfonseca@vmware.com>
  58  */
  59
  60 #include <limits.h>
  61 #include "pipe/p_defines.h"
  62 #include "util/u_inlines.h"
  63 #include "util/u_memory.h"
  64 #include "util/u_pointer.h"
  65 #include "util/u_format.h"
  66 #include "util/u_dump.h"
  67 #include "util/u_string.h"
  68 #include "util/u_simple_list.h"
  69 #include "os/os_time.h"
  70 #include "pipe/p_shader_tokens.h"
  71 #include "draw/draw_context.h"
  72 #include "tgsi/tgsi_dump.h"
  73 #include "tgsi/tgsi_scan.h"
  74 #include "tgsi/tgsi_parse.h"
  75 #include "gallivm/lp_bld_type.h"
  76 #include "gallivm/lp_bld_const.h"
  77 #include "gallivm/lp_bld_conv.h"
  78 #include "gallivm/lp_bld_init.h"
  79 #include "gallivm/lp_bld_intr.h"
  80 #include "gallivm/lp_bld_logic.h"
  81 #include "gallivm/lp_bld_tgsi.h"
  82 #include "gallivm/lp_bld_swizzle.h"
  83 #include "gallivm/lp_bld_flow.h"
  84 #include "gallivm/lp_bld_debug.h"
  85
  86 #include "lp_bld_alpha.h"
  87 #include "lp_bld_blend.h"
  88 #include "lp_bld_depth.h"
  89 #include "lp_bld_interp.h"
  90 #include "lp_context.h"
  91 #include "lp_debug.h"
  92 #include "lp_perf.h"
  93 #include "lp_setup.h"
  94 #include "lp_state.h"
  95 #include "lp_tex_sample.h"
  96 #include "lp_flush.h"
  97 #include "lp_state_fs.h"
  98
  99
 100 #include <llvm-c/Analysis.h>
 101 #include <llvm-c/BitWriter.h>
 102
 103
 104 /** Fragment shader number (for debugging) */
 105 static unsigned fs_no = 0;
 106
 107
 108 /**
 109  * Expand the relevent bits of mask_input to a 4-dword mask for the
 110  * four pixels in a 2x2 quad.  This will set the four elements of the
 111  * quad mask vector to 0 or ~0.
 112  *
 113  * \param quad  which quad of the quad group to test, in [0,3]
 114  * \param mask_input  bitwise mask for the whole 4x4 stamp
 115  */
 116 static LLVMValueRef
 117 generate_quad_mask(struct gallivm_state *gallivm,
 118                    struct lp_type fs_type,
 119                    unsigned quad,
 120                    LLVMValueRef mask_input) /* int32 */
 121 {
 122    LLVMBuilderRef builder = gallivm->builder;
 123    struct lp_type mask_type;
 124    LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
 125    LLVMValueRef bits[4];
 126    LLVMValueRef mask;
 127    int shift;
 128
 129    /*
 130     * XXX: We'll need a different path for 16 x u8
 131     */
 132    assert(fs_type.width == 32);
 133    assert(fs_type.length == 4);
 134    mask_type = lp_int_type(fs_type);
 135
 136    /*
 137     * mask_input >>= (quad * 4)
 138     */
 139    switch (quad) {
 140    case 0:
 141       shift = 0;
 142       break;
 143    case 1:
 144       shift = 2;
 145       break;
 146    case 2:
 147       shift = 8;
 148       break;
 149    case 3:
 150       shift = 10;
 151       break;
 152    default:
 153       assert(0);
 154       shift = 0;
 155    }
 156
 157    mask_input = LLVMBuildLShr(builder,
 158                               mask_input,
 159                               LLVMConstInt(i32t, shift, 0),
 160                               "");
 161
 162    /*
 163     * mask = { mask_input & (1 << i), for i in [0,3] }
 164     */
 165    mask = lp_build_broadcast(gallivm,
 166                              lp_build_vec_type(gallivm, mask_type),
 167                              mask_input);
 168
 169    bits[0] = LLVMConstInt(i32t, 1 << 0, 0);
 170    bits[1] = LLVMConstInt(i32t, 1 << 1, 0);
 171    bits[2] = LLVMConstInt(i32t, 1 << 4, 0);
 172    bits[3] = LLVMConstInt(i32t, 1 << 5, 0);
 173
 174    mask = LLVMBuildAnd(builder, mask, LLVMConstVector(bits, 4), "");
 175
 176    /*
 177     * mask = mask != 0 ? ~0 : 0
 178     */
 179    mask = lp_build_compare(gallivm,
 180                            mask_type, PIPE_FUNC_NOTEQUAL,
 181                            mask,
 182                            lp_build_const_int_vec(gallivm, mask_type, 0));
 183
 184    return mask;
 185 }
 186
 187
 188 #define EARLY_DEPTH_TEST  0x1
 189 #define LATE_DEPTH_TEST   0x2
 190 #define EARLY_DEPTH_WRITE 0x4
 191 #define LATE_DEPTH_WRITE  0x8
 192
 193 static int
 194 find_output_by_semantic( const struct tgsi_shader_info *info,
 195                          unsigned semantic,
 196                          unsigned index )
 197 {
 198    int i;
 199
 200    for (i = 0; i < info->num_outputs; i++)
 201       if (info->output_semantic_name[i] == semantic &&
 202           info->output_semantic_index[i] == index)
 203          return i;
 204
 205    return -1;
 206 }
 207
 208
 209 /**
 210  * Generate the fragment shader, depth/stencil test, and alpha tests.
 211  * \param i  which quad in the tile, in range [0,3]
 212  * \param partial_mask  if 1, do mask_input testing
 213  */
 214 static void
 215 generate_fs(struct gallivm_state *gallivm,
 216             struct lp_fragment_shader *shader,
 217             const struct lp_fragment_shader_variant_key *key,
 218             LLVMBuilderRef builder,
 219             struct lp_type type,
 220             LLVMValueRef context_ptr,
 221             unsigned i,
 222             struct lp_build_interp_soa_context *interp,
 223             struct lp_build_sampler_soa *sampler,
 224             LLVMValueRef *pmask,
 225             LLVMValueRef (*color)[4],
 226             LLVMValueRef depth_ptr,
 227             LLVMValueRef facing,
 228             unsigned partial_mask,
 229             LLVMValueRef mask_input,
 230             LLVMValueRef counter)
 231 {
 232    const struct util_format_description *zs_format_desc = NULL;
 233    const struct tgsi_token *tokens = shader->base.tokens;
 234    LLVMTypeRef vec_type;
 235    LLVMValueRef consts_ptr;
 236    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][NUM_CHANNELS];
 237    LLVMValueRef z;
 238    LLVMValueRef zs_value = NULL;
 239    LLVMValueRef stencil_refs[2];
 240    struct lp_build_mask_context mask;
 241    boolean simple_shader = (shader->info.base.file_count[TGSI_FILE_SAMPLER] == 0 &&
 242                             shader->info.base.num_inputs < 3 &&
 243                             shader->info.base.num_instructions < 8);
 244    unsigned attrib;
 245    unsigned chan;
 246    unsigned cbuf;
 247    unsigned depth_mode;
 248
 249    if (key->depth.enabled ||
 250        key->stencil[0].enabled ||
 251        key->stencil[1].enabled) {
 252
 253       zs_format_desc = util_format_description(key->zsbuf_format);
 254       assert(zs_format_desc);
 255
 256       if (!shader->info.base.writes_z) {
 257          if (key->alpha.enabled || shader->info.base.uses_kill)
 258             /* With alpha test and kill, can do the depth test early
 259              * and hopefully eliminate some quads.  But need to do a
 260              * special deferred depth write once the final mask value
 261              * is known.
 262              */
 263             depth_mode = EARLY_DEPTH_TEST | LATE_DEPTH_WRITE;
 264          else
 265             depth_mode = EARLY_DEPTH_TEST | EARLY_DEPTH_WRITE;
 266       }
 267       else {
 268          depth_mode = LATE_DEPTH_TEST | LATE_DEPTH_WRITE;
 269       }
 270
 271       if (!(key->depth.enabled && key->depth.writemask) &&
 272           !(key->stencil[0].enabled && key->stencil[0].writemask))
 273          depth_mode &= ~(LATE_DEPTH_WRITE | EARLY_DEPTH_WRITE);
 274    }
 275    else {
 276       depth_mode = 0;
 277    }
 278
 279    assert(i < 4);
 280
 281    stencil_refs[0] = lp_jit_context_stencil_ref_front_value(gallivm, context_ptr);
 282    stencil_refs[1] = lp_jit_context_stencil_ref_back_value(gallivm, context_ptr);
 283
 284    vec_type = lp_build_vec_type(gallivm, type);
 285
 286    consts_ptr = lp_jit_context_constants(gallivm, context_ptr);
 287
 288    memset(outputs, 0, sizeof outputs);
 289
 290    /* Declare the color and z variables */
 291    for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
 292       for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 293          color[cbuf][chan] = lp_build_alloca(gallivm, vec_type, "color");
 294       }
 295    }
 296
 297    /* do triangle edge testing */
 298    if (partial_mask) {
 299       *pmask = generate_quad_mask(gallivm, type,
 300                                   i, mask_input);
 301    }
 302    else {
 303       *pmask = lp_build_const_int_vec(gallivm, type, ~0);
 304    }
 305
 306    /* 'mask' will control execution based on quad's pixel alive/killed state */
 307    lp_build_mask_begin(&mask, gallivm, type, *pmask);
 308
 309    if (!(depth_mode & EARLY_DEPTH_TEST) && !simple_shader)
 310       lp_build_mask_check(&mask);
 311
 312    lp_build_interp_soa_update_pos(interp, gallivm, i);
 313    z = interp->pos[2];
 314
 315    if (depth_mode & EARLY_DEPTH_TEST) {
 316       lp_build_depth_stencil_test(gallivm,
 317                                   &key->depth,
 318                                   key->stencil,
 319                                   type,
 320                                   zs_format_desc,
 321                                   &mask,
 322                                   stencil_refs,
 323                                   z,
 324                                   depth_ptr, facing,
 325                                   &zs_value,
 326                                   !simple_shader);
 327
 328       if (depth_mode & EARLY_DEPTH_WRITE) {
 329          lp_build_depth_write(builder, zs_format_desc, depth_ptr, zs_value);
 330       }
 331    }
 332
 333    lp_build_interp_soa_update_inputs(interp, gallivm, i);
 334
 335    /* Build the actual shader */
 336    lp_build_tgsi_soa(gallivm, tokens, type, &mask,
 337                      consts_ptr, interp->pos, interp->inputs,
 338                      outputs, sampler, &shader->info.base);
 339
 340    /* Alpha test */
 341    if (key->alpha.enabled) {
 342       int color0 = find_output_by_semantic(&shader->info.base,
 343                                            TGSI_SEMANTIC_COLOR,
 344                                            0);
 345
 346       if (color0 != -1 && outputs[color0][3]) {
 347          LLVMValueRef alpha = LLVMBuildLoad(builder, outputs[color0][3], "alpha");
 348          LLVMValueRef alpha_ref_value;
 349
 350          alpha_ref_value = lp_jit_context_alpha_ref_value(gallivm, context_ptr);
 351          alpha_ref_value = lp_build_broadcast(gallivm, vec_type, alpha_ref_value);
 352
 353          lp_build_alpha_test(gallivm, key->alpha.func, type,
 354                              &mask, alpha, alpha_ref_value,
 355                              (depth_mode & LATE_DEPTH_TEST) != 0);
 356       }
 357    }
 358
 359    /* Late Z test */
 360    if (depth_mode & LATE_DEPTH_TEST) {
 361       int pos0 = find_output_by_semantic(&shader->info.base,
 362                                          TGSI_SEMANTIC_POSITION,
 363                                          0);
 364
 365       if (pos0 != -1 && outputs[pos0][2]) {
 366          z = LLVMBuildLoad(builder, outputs[pos0][2], "output.z");
 367       }
 368
 369       lp_build_depth_stencil_test(gallivm,
 370                                   &key->depth,
 371                                   key->stencil,
 372                                   type,
 373                                   zs_format_desc,
 374                                   &mask,
 375                                   stencil_refs,
 376                                   z,
 377                                   depth_ptr, facing,
 378                                   &zs_value,
 379                                   !simple_shader);
 380       /* Late Z write */
 381       if (depth_mode & LATE_DEPTH_WRITE) {
 382          lp_build_depth_write(builder, zs_format_desc, depth_ptr, zs_value);
 383       }
 384    }
 385    else if ((depth_mode & EARLY_DEPTH_TEST) &&
 386             (depth_mode & LATE_DEPTH_WRITE))
 387    {
 388       /* Need to apply a reduced mask to the depth write.  Reload the
 389        * depth value, update from zs_value with the new mask value and
 390        * write that out.
 391        */
 392       lp_build_deferred_depth_write(gallivm,
 393                                     type,
 394                                     zs_format_desc,
 395                                     &mask,
 396                                     depth_ptr,
 397                                     zs_value);
 398    }
 399
 400
 401    /* Color write  */
 402    for (attrib = 0; attrib < shader->info.base.num_outputs; ++attrib)
 403    {
 404       if (shader->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_COLOR &&
 405           shader->info.base.output_semantic_index[attrib] < key->nr_cbufs)
 406       {
 407          unsigned cbuf = shader->info.base.output_semantic_index[attrib];
 408          for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 409             if(outputs[attrib][chan]) {
 410                /* XXX: just initialize outputs to point at colors[] and
 411                 * skip this.
 412                 */
 413                LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
 414                lp_build_name(out, "color%u.%u.%c", i, attrib, "rgba"[chan]);
 415                LLVMBuildStore(builder, out, color[cbuf][chan]);
 416             }
 417          }
 418       }
 419    }
 420
 421    if (counter)
 422       lp_build_occlusion_count(gallivm, type,
 423                                lp_build_mask_value(&mask), counter);
 424
 425    *pmask = lp_build_mask_end(&mask);
 426 }
 427
 428
 429 /**
 430  * Generate color blending and color output.
 431  * \param rt  the render target index (to index blend, colormask state)
 432  * \param type  the pixel color type
 433  * \param context_ptr  pointer to the runtime JIT context
 434  * \param mask  execution mask (active fragment/pixel mask)
 435  * \param src  colors from the fragment shader
 436  * \param dst_ptr  the destination color buffer pointer
 437  */
 438 static void
 439 generate_blend(struct gallivm_state *gallivm,
 440                const struct pipe_blend_state *blend,
 441                unsigned rt,
 442                LLVMBuilderRef builder,
 443                struct lp_type type,
 444                LLVMValueRef context_ptr,
 445                LLVMValueRef mask,
 446                LLVMValueRef *src,
 447                LLVMValueRef dst_ptr,
 448                boolean do_branch)
 449 {
 450    struct lp_build_context bld;
 451    struct lp_build_mask_context mask_ctx;
 452    LLVMTypeRef vec_type;
 453    LLVMValueRef const_ptr;
 454    LLVMValueRef con[4];
 455    LLVMValueRef dst[4];
 456    LLVMValueRef res[4];
 457    unsigned chan;
 458
 459    lp_build_context_init(&bld, gallivm, type);
 460
 461    lp_build_mask_begin(&mask_ctx, gallivm, type, mask);
 462    if (do_branch)
 463       lp_build_mask_check(&mask_ctx);
 464
 465    vec_type = lp_build_vec_type(gallivm, type);
 466
 467    const_ptr = lp_jit_context_blend_color(gallivm, context_ptr);
 468    const_ptr = LLVMBuildBitCast(builder, const_ptr,
 469                                 LLVMPointerType(vec_type, 0), "");
 470
 471    /* load constant blend color and colors from the dest color buffer */
 472    for(chan = 0; chan < 4; ++chan) {
 473       LLVMValueRef index = lp_build_const_int32(gallivm, chan);
 474       con[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, const_ptr, &index, 1, ""), "");
 475
 476       dst[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, dst_ptr, &index, 1, ""), "");
 477
 478       lp_build_name(con[chan], "con.%c", "rgba"[chan]);
 479       lp_build_name(dst[chan], "dst.%c", "rgba"[chan]);
 480    }
 481
 482    /* do blend */
 483    lp_build_blend_soa(gallivm, blend, type, rt, src, dst, con, res);
 484
 485    /* store results to color buffer */
 486    for(chan = 0; chan < 4; ++chan) {
 487       if(blend->rt[rt].colormask & (1 << chan)) {
 488          LLVMValueRef index = lp_build_const_int32(gallivm, chan);
 489          lp_build_name(res[chan], "res.%c", "rgba"[chan]);
 490          res[chan] = lp_build_select(&bld, mask, res[chan], dst[chan]);
 491          LLVMBuildStore(builder, res[chan], LLVMBuildGEP(builder, dst_ptr, &index, 1, ""));
 492       }
 493    }
 494
 495    lp_build_mask_end(&mask_ctx);
 496 }
 497
 498
 499 /**
 500  * Generate the runtime callable function for the whole fragment pipeline.
 501  * Note that the function which we generate operates on a block of 16
 502  * pixels at at time.  The block contains 2x2 quads.  Each quad contains
 503  * 2x2 pixels.
 504  */
 505 static void
 506 generate_fragment(struct llvmpipe_context *lp,
 507                   struct lp_fragment_shader *shader,
 508                   struct lp_fragment_shader_variant *variant,
 509                   unsigned partial_mask)
 510 {
 511    struct gallivm_state *gallivm = lp->gallivm;
 512    const struct lp_fragment_shader_variant_key *key = &variant->key;
 513    struct lp_shader_input inputs[PIPE_MAX_SHADER_INPUTS];
 514    char func_name[256];
 515    struct lp_type fs_type;
 516    struct lp_type blend_type;
 517    LLVMTypeRef fs_elem_type;
 518    LLVMTypeRef fs_int_vec_type;
 519    LLVMTypeRef blend_vec_type;
 520    LLVMTypeRef arg_types[11];
 521    LLVMTypeRef func_type;
 522    LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
 523    LLVMTypeRef int8_type = LLVMInt8TypeInContext(gallivm->context);
 524    LLVMValueRef context_ptr;
 525    LLVMValueRef x;
 526    LLVMValueRef y;
 527    LLVMValueRef a0_ptr;
 528    LLVMValueRef dadx_ptr;
 529    LLVMValueRef dady_ptr;
 530    LLVMValueRef color_ptr_ptr;
 531    LLVMValueRef depth_ptr;
 532    LLVMValueRef mask_input;
 533    LLVMValueRef counter = NULL;
 534    LLVMBasicBlockRef block;
 535    LLVMBuilderRef builder;
 536    struct lp_build_sampler_soa *sampler;
 537    struct lp_build_interp_soa_context interp;
 538    LLVMValueRef fs_mask[LP_MAX_VECTOR_LENGTH];
 539    LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS][LP_MAX_VECTOR_LENGTH];
 540    LLVMValueRef blend_mask;
 541    LLVMValueRef function;
 542    LLVMValueRef facing;
 543    const struct util_format_description *zs_format_desc;
 544    unsigned num_fs;
 545    unsigned i;
 546    unsigned chan;
 547    unsigned cbuf;
 548
 549    /* Adjust color input interpolation according to flatshade state:
 550     */
 551    memcpy(inputs, shader->inputs, shader->info.base.num_inputs * sizeof inputs[0]);
 552    for (i = 0; i < shader->info.base.num_inputs; i++) {
 553       if (inputs[i].interp == LP_INTERP_COLOR) {
 554          if (key->flatshade)
 555             inputs[i].interp = LP_INTERP_CONSTANT;
 556          else
 557             inputs[i].interp = LP_INTERP_LINEAR;
 558       }
 559    }
 560
 561
 562    /* TODO: actually pick these based on the fs and color buffer
 563     * characteristics. */
 564
 565    memset(&fs_type, 0, sizeof fs_type);
 566    fs_type.floating = TRUE; /* floating point values */
 567    fs_type.sign = TRUE;     /* values are signed */
 568    fs_type.norm = FALSE;    /* values are not limited to [0,1] or [-1,1] */
 569    fs_type.width = 32;      /* 32-bit float */
 570    fs_type.length = 4;      /* 4 elements per vector */
 571    num_fs = 4;              /* number of quads per block */
 572
 573    memset(&blend_type, 0, sizeof blend_type);
 574    blend_type.floating = FALSE; /* values are integers */
 575    blend_type.sign = FALSE;     /* values are unsigned */
 576    blend_type.norm = TRUE;      /* values are in [0,1] or [-1,1] */
 577    blend_type.width = 8;        /* 8-bit ubyte values */
 578    blend_type.length = 16;      /* 16 elements per vector */
 579
 580    /*
 581     * Generate the function prototype. Any change here must be reflected in
 582     * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
 583     */
 584
 585    fs_elem_type = lp_build_elem_type(gallivm, fs_type);
 586    fs_int_vec_type = lp_build_int_vec_type(gallivm, fs_type);
 587
 588    blend_vec_type = lp_build_vec_type(gallivm, blend_type);
 589
 590    util_snprintf(func_name, sizeof(func_name), "fs%u_variant%u_%s",
 591                  shader->no, variant->no, partial_mask ? "partial" : "whole");
 592
 593    arg_types[0] = lp_jit_get_context_type(lp);         /* context */
 594    arg_types[1] = int32_type;                          /* x */
 595    arg_types[2] = int32_type;                          /* y */
 596    arg_types[3] = int32_type;                          /* facing */
 597    arg_types[4] = LLVMPointerType(fs_elem_type, 0);    /* a0 */
 598    arg_types[5] = LLVMPointerType(fs_elem_type, 0);    /* dadx */
 599    arg_types[6] = LLVMPointerType(fs_elem_type, 0);    /* dady */
 600    arg_types[7] = LLVMPointerType(LLVMPointerType(blend_vec_type, 0), 0);  /* color */
 601    arg_types[8] = LLVMPointerType(int8_type, 0);       /* depth */
 602    arg_types[9] = int32_type;                          /* mask_input */
 603    arg_types[10] = LLVMPointerType(int32_type, 0);     /* counter */
 604
 605    func_type = LLVMFunctionType(LLVMVoidTypeInContext(gallivm->context),
 606                                 arg_types, Elements(arg_types), 0);
 607
 608    function = LLVMAddFunction(gallivm->module, func_name, func_type);
 609    LLVMSetFunctionCallConv(function, LLVMCCallConv);
 610
 611    variant->function[partial_mask] = function;
 612
 613    /* XXX: need to propagate noalias down into color param now we are
 614     * passing a pointer-to-pointer?
 615     */
 616    for(i = 0; i < Elements(arg_types); ++i)
 617       if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
 618          LLVMAddAttribute(LLVMGetParam(function, i), LLVMNoAliasAttribute);
 619
 620    context_ptr  = LLVMGetParam(function, 0);
 621    x            = LLVMGetParam(function, 1);
 622    y            = LLVMGetParam(function, 2);
 623    facing       = LLVMGetParam(function, 3);
 624    a0_ptr       = LLVMGetParam(function, 4);
 625    dadx_ptr     = LLVMGetParam(function, 5);
 626    dady_ptr     = LLVMGetParam(function, 6);
 627    color_ptr_ptr = LLVMGetParam(function, 7);
 628    depth_ptr    = LLVMGetParam(function, 8);
 629    mask_input   = LLVMGetParam(function, 9);
 630
 631    lp_build_name(context_ptr, "context");
 632    lp_build_name(x, "x");
 633    lp_build_name(y, "y");
 634    lp_build_name(a0_ptr, "a0");
 635    lp_build_name(dadx_ptr, "dadx");
 636    lp_build_name(dady_ptr, "dady");
 637    lp_build_name(color_ptr_ptr, "color_ptr_ptr");
 638    lp_build_name(depth_ptr, "depth");
 639    lp_build_name(mask_input, "mask_input");
 640
 641    if (key->occlusion_count) {
 642       counter = LLVMGetParam(function, 10);
 643       lp_build_name(counter, "counter");
 644    }
 645
 646    /*
 647     * Function body
 648     */
 649
 650    block = LLVMAppendBasicBlockInContext(gallivm->context, function, "entry");
 651    builder = gallivm->builder;
 652    assert(builder);
 653    LLVMPositionBuilderAtEnd(builder, block);
 654
 655    /*
 656     * The shader input interpolation info is not explicitely baked in the
 657     * shader key, but everything it derives from (TGSI, and flatshade) is
 658     * already included in the shader key.
 659     */
 660    lp_build_interp_soa_init(&interp,
 661                             gallivm,
 662                             shader->info.base.num_inputs,
 663                             inputs,
 664                             builder, fs_type,
 665                             a0_ptr, dadx_ptr, dady_ptr,
 666                             x, y);
 667
 668    /* code generated texture sampling */
 669    sampler = lp_llvm_sampler_soa_create(key->sampler, context_ptr);
 670
 671    /* loop over quads in the block */
 672    zs_format_desc = util_format_description(key->zsbuf_format);
 673
 674    for(i = 0; i < num_fs; ++i) {
 675       LLVMValueRef depth_offset = LLVMConstInt(int32_type,
 676                                                i*fs_type.length*zs_format_desc->block.bits/8,
 677                                                0);
 678       LLVMValueRef out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS];
 679       LLVMValueRef depth_ptr_i;
 680
 681       depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &depth_offset, 1, "");
 682
 683       generate_fs(gallivm,
 684                   shader, key,
 685                   builder,
 686                   fs_type,
 687                   context_ptr,
 688                   i,
 689                   &interp,
 690                   sampler,
 691                   &fs_mask[i], /* output */
 692                   out_color,
 693                   depth_ptr_i,
 694                   facing,
 695                   partial_mask,
 696                   mask_input,
 697                   counter);
 698
 699       for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++)
 700          for(chan = 0; chan < NUM_CHANNELS; ++chan)
 701             fs_out_color[cbuf][chan][i] = out_color[cbuf][chan];
 702    }
 703
 704    sampler->destroy(sampler);
 705
 706    /* Loop over color outputs / color buffers to do blending.
 707     */
 708    for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
 709       LLVMValueRef color_ptr;
 710       LLVMValueRef index = lp_build_const_int32(gallivm, cbuf);
 711       LLVMValueRef blend_in_color[NUM_CHANNELS];
 712       unsigned rt;
 713
 714       /*
 715        * Convert the fs's output color and mask to fit to the blending type.
 716        */
 717       for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 718          LLVMValueRef fs_color_vals[LP_MAX_VECTOR_LENGTH];
 719
 720          for (i = 0; i < num_fs; i++) {
 721             fs_color_vals[i] =
 722                LLVMBuildLoad(builder, fs_out_color[cbuf][chan][i], "fs_color_vals");
 723          }
 724
 725          lp_build_conv(gallivm, fs_type, blend_type,
 726                        fs_color_vals,
 727                        num_fs,
 728                        &blend_in_color[chan], 1);
 729
 730          lp_build_name(blend_in_color[chan], "color%d.%c", cbuf, "rgba"[chan]);
 731       }
 732
 733       if (partial_mask || !variant->opaque) {
 734          lp_build_conv_mask(lp->gallivm, fs_type, blend_type,
 735                             fs_mask, num_fs,
 736                             &blend_mask, 1);
 737       } else {
 738          blend_mask = lp_build_const_int_vec(lp->gallivm, blend_type, ~0);
 739       }
 740
 741       color_ptr = LLVMBuildLoad(builder,
 742                                 LLVMBuildGEP(builder, color_ptr_ptr, &index, 1, ""),
 743                                 "");
 744       lp_build_name(color_ptr, "color_ptr%d", cbuf);
 745
 746       /* which blend/colormask state to use */
 747       rt = key->blend.independent_blend_enable ? cbuf : 0;
 748
 749       /*
 750        * Blending.
 751        */
 752       {
 753          /* Could the 4x4 have been killed?
 754           */
 755          boolean do_branch = ((key->depth.enabled || key->stencil[0].enabled) &&
 756                               !key->alpha.enabled &&
 757                               !shader->info.base.uses_kill);
 758
 759          generate_blend(lp->gallivm,
 760                         &key->blend,
 761                         rt,
 762                         builder,
 763                         blend_type,
 764                         context_ptr,
 765                         blend_mask,
 766                         blend_in_color,
 767                         color_ptr,
 768                         do_branch);
 769       }
 770    }
 771
 772    LLVMBuildRetVoid(builder);
 773
 774    /* Verify the LLVM IR.  If invalid, dump and abort */
 775 #ifdef DEBUG
 776    if(LLVMVerifyFunction(function, LLVMPrintMessageAction)) {
 777       if (1)
 778          lp_debug_dump_value(function);
 779       abort();
 780    }
 781 #endif
 782
 783    /* Apply optimizations to LLVM IR */
 784    LLVMRunFunctionPassManager(gallivm->passmgr, function);
 785
 786    if ((gallivm_debug & GALLIVM_DEBUG_IR) || (LP_DEBUG & DEBUG_FS)) {
 787       /* Print the LLVM IR to stderr */
 788       lp_debug_dump_value(function);
 789       debug_printf("\n");
 790    }
 791
 792    /* Dump byte code to a file */
 793    if (0) {
 794       LLVMWriteBitcodeToFile(gallivm->module, "llvmpipe.bc");
 795    }
 796
 797    /*
 798     * Translate the LLVM IR into machine code.
 799     */
 800    {
 801       void *f = LLVMGetPointerToGlobal(gallivm->engine, function);
 802
 803       variant->jit_function[partial_mask] = (lp_jit_frag_func)pointer_to_func(f);
 804
 805       if ((gallivm_debug & GALLIVM_DEBUG_ASM) || (LP_DEBUG & DEBUG_FS)) {
 806          lp_disassemble(f);
 807       }
 808       lp_func_delete_body(function);
 809    }
 810 }
 811
 812
 813 static void
 814 dump_fs_variant_key(const struct lp_fragment_shader_variant_key *key)
 815 {
 816    unsigned i;
 817
 818    debug_printf("fs variant %p:\n", (void *) key);
 819
 820    if (key->flatshade) {
 821       debug_printf("flatshade = 1\n");
 822    }
 823    for (i = 0; i < key->nr_cbufs; ++i) {
 824       debug_printf("cbuf_format[%u] = %s\n", i, util_format_name(key->cbuf_format[i]));
 825    }
 826    if (key->depth.enabled) {
 827       debug_printf("depth.format = %s\n", util_format_name(key->zsbuf_format));
 828       debug_printf("depth.func = %s\n", util_dump_func(key->depth.func, TRUE));
 829       debug_printf("depth.writemask = %u\n", key->depth.writemask);
 830    }
 831
 832    for (i = 0; i < 2; ++i) {
 833       if (key->stencil[i].enabled) {
 834          debug_printf("stencil[%u].func = %s\n", i, util_dump_func(key->stencil[i].func, TRUE));
 835          debug_printf("stencil[%u].fail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].fail_op, TRUE));
 836          debug_printf("stencil[%u].zpass_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zpass_op, TRUE));
 837          debug_printf("stencil[%u].zfail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zfail_op, TRUE));
 838          debug_printf("stencil[%u].valuemask = 0x%x\n", i, key->stencil[i].valuemask);
 839          debug_printf("stencil[%u].writemask = 0x%x\n", i, key->stencil[i].writemask);
 840       }
 841    }
 842
 843    if (key->alpha.enabled) {
 844       debug_printf("alpha.func = %s\n", util_dump_func(key->alpha.func, TRUE));
 845    }
 846
 847    if (key->occlusion_count) {
 848       debug_printf("occlusion_count = 1\n");
 849    }
 850
 851    if (key->blend.logicop_enable) {
 852       debug_printf("blend.logicop_func = %s\n", util_dump_logicop(key->blend.logicop_func, TRUE));
 853    }
 854    else if (key->blend.rt[0].blend_enable) {
 855       debug_printf("blend.rgb_func = %s\n",   util_dump_blend_func  (key->blend.rt[0].rgb_func, TRUE));
 856       debug_printf("blend.rgb_src_factor = %s\n",   util_dump_blend_factor(key->blend.rt[0].rgb_src_factor, TRUE));
 857       debug_printf("blend.rgb_dst_factor = %s\n",   util_dump_blend_factor(key->blend.rt[0].rgb_dst_factor, TRUE));
 858       debug_printf("blend.alpha_func = %s\n",       util_dump_blend_func  (key->blend.rt[0].alpha_func, TRUE));
 859       debug_printf("blend.alpha_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_src_factor, TRUE));
 860       debug_printf("blend.alpha_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_dst_factor, TRUE));
 861    }
 862    debug_printf("blend.colormask = 0x%x\n", key->blend.rt[0].colormask);
 863    for (i = 0; i < key->nr_samplers; ++i) {
 864       debug_printf("sampler[%u] = \n", i);
 865       debug_printf("  .format = %s\n",
 866                    util_format_name(key->sampler[i].format));
 867       debug_printf("  .target = %s\n",
 868                    util_dump_tex_target(key->sampler[i].target, TRUE));
 869       debug_printf("  .pot = %u %u %u\n",
 870                    key->sampler[i].pot_width,
 871                    key->sampler[i].pot_height,
 872                    key->sampler[i].pot_depth);
 873       debug_printf("  .wrap = %s %s %s\n",
 874                    util_dump_tex_wrap(key->sampler[i].wrap_s, TRUE),
 875                    util_dump_tex_wrap(key->sampler[i].wrap_t, TRUE),
 876                    util_dump_tex_wrap(key->sampler[i].wrap_r, TRUE));
 877       debug_printf("  .min_img_filter = %s\n",
 878                    util_dump_tex_filter(key->sampler[i].min_img_filter, TRUE));
 879       debug_printf("  .min_mip_filter = %s\n",
 880                    util_dump_tex_mipfilter(key->sampler[i].min_mip_filter, TRUE));
 881       debug_printf("  .mag_img_filter = %s\n",
 882                    util_dump_tex_filter(key->sampler[i].mag_img_filter, TRUE));
 883       if (key->sampler[i].compare_mode != PIPE_TEX_COMPARE_NONE)
 884          debug_printf("  .compare_func = %s\n", util_dump_func(key->sampler[i].compare_func, TRUE));
 885       debug_printf("  .normalized_coords = %u\n", key->sampler[i].normalized_coords);
 886       debug_printf("  .min_max_lod_equal = %u\n", key->sampler[i].min_max_lod_equal);
 887       debug_printf("  .lod_bias_non_zero = %u\n", key->sampler[i].lod_bias_non_zero);
 888       debug_printf("  .apply_min_lod = %u\n", key->sampler[i].apply_min_lod);
 889       debug_printf("  .apply_max_lod = %u\n", key->sampler[i].apply_max_lod);
 890    }
 891 }
 892
 893
 894 void
 895 lp_debug_fs_variant(const struct lp_fragment_shader_variant *variant)
 896 {
 897    debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
 898                 variant->shader->no, variant->no);
 899    tgsi_dump(variant->shader->base.tokens, 0);
 900    dump_fs_variant_key(&variant->key);
 901    debug_printf("variant->opaque = %u\n", variant->opaque);
 902    debug_printf("\n");
 903 }
 904
 905
 906 /**
 907  * Generate a new fragment shader variant from the shader code and
 908  * other state indicated by the key.
 909  */
 910 static struct lp_fragment_shader_variant *
 911 generate_variant(struct llvmpipe_context *lp,
 912                  struct lp_fragment_shader *shader,
 913                  const struct lp_fragment_shader_variant_key *key)
 914 {
 915    struct lp_fragment_shader_variant *variant;
 916    boolean fullcolormask;
 917
 918    variant = CALLOC_STRUCT(lp_fragment_shader_variant);
 919    if(!variant)
 920       return NULL;
 921
 922    variant->shader = shader;
 923    variant->list_item_global.base = variant;
 924    variant->list_item_local.base = variant;
 925    variant->no = shader->variants_created++;
 926
 927    memcpy(&variant->key, key, shader->variant_key_size);
 928
 929    /*
 930     * Determine whether we are touching all channels in the color buffer.
 931     */
 932    fullcolormask = FALSE;
 933    if (key->nr_cbufs == 1) {
 934       const struct util_format_description *format_desc;
 935       format_desc = util_format_description(key->cbuf_format[0]);
 936       if ((~key->blend.rt[0].colormask &
 937            util_format_colormask(format_desc)) == 0) {
 938          fullcolormask = TRUE;
 939       }
 940    }
 941
 942    variant->opaque =
 943          !key->blend.logicop_enable &&
 944          !key->blend.rt[0].blend_enable &&
 945          fullcolormask &&
 946          !key->stencil[0].enabled &&
 947          !key->alpha.enabled &&
 948          !key->depth.enabled &&
 949          !shader->info.base.uses_kill
 950          ? TRUE : FALSE;
 951
 952
 953    if ((LP_DEBUG & DEBUG_FS) || (gallivm_debug & GALLIVM_DEBUG_IR)) {
 954       lp_debug_fs_variant(variant);
 955    }
 956
 957    generate_fragment(lp, shader, variant, RAST_EDGE_TEST);
 958
 959    if (variant->opaque) {
 960       /* Specialized shader, which doesn't need to read the color buffer. */
 961       generate_fragment(lp, shader, variant, RAST_WHOLE);
 962    } else {
 963       variant->jit_function[RAST_WHOLE] = variant->jit_function[RAST_EDGE_TEST];
 964    }
 965
 966    return variant;
 967 }
 968
 969
 970 static void *
 971 llvmpipe_create_fs_state(struct pipe_context *pipe,
 972                          const struct pipe_shader_state *templ)
 973 {
 974    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
 975    struct lp_fragment_shader *shader;
 976    int nr_samplers;
 977    int i;
 978
 979    shader = CALLOC_STRUCT(lp_fragment_shader);
 980    if (!shader)
 981       return NULL;
 982
 983    shader->no = fs_no++;
 984    make_empty_list(&shader->variants);
 985
 986    /* get/save the summary info for this shader */
 987    lp_build_tgsi_info(templ->tokens, &shader->info);
 988
 989    /* we need to keep a local copy of the tokens */
 990    shader->base.tokens = tgsi_dup_tokens(templ->tokens);
 991
 992    shader->draw_data = draw_create_fragment_shader(llvmpipe->draw, templ);
 993    if (shader->draw_data == NULL) {
 994       FREE((void *) shader->base.tokens);
 995       FREE(shader);
 996       return NULL;
 997    }
 998
 999    nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
1000
1001    shader->variant_key_size = Offset(struct lp_fragment_shader_variant_key,
1002                                      sampler[nr_samplers]);
1003
1004    for (i = 0; i < shader->info.base.num_inputs; i++) {
1005       shader->inputs[i].usage_mask = shader->info.base.input_usage_mask[i];
1006
1007       switch (shader->info.base.input_interpolate[i]) {
1008       case TGSI_INTERPOLATE_CONSTANT:
1009          shader->inputs[i].interp = LP_INTERP_CONSTANT;
1010          break;
1011       case TGSI_INTERPOLATE_LINEAR:
1012          shader->inputs[i].interp = LP_INTERP_LINEAR;
1013          break;
1014       case TGSI_INTERPOLATE_PERSPECTIVE:
1015          shader->inputs[i].interp = LP_INTERP_PERSPECTIVE;
1016          break;
1017       default:
1018          assert(0);
1019          break;
1020       }
1021
1022       switch (shader->info.base.input_semantic_name[i]) {
1023       case TGSI_SEMANTIC_COLOR:
1024          /* Colors may be either linearly or constant interpolated in
1025           * the fragment shader, but that information isn't available
1026           * here.  Mark color inputs and fix them up later.
1027           */
1028          shader->inputs[i].interp = LP_INTERP_COLOR;
1029          break;
1030       case TGSI_SEMANTIC_FACE:
1031          shader->inputs[i].interp = LP_INTERP_FACING;
1032          break;
1033       case TGSI_SEMANTIC_POSITION:
1034          /* Position was already emitted above
1035           */
1036          shader->inputs[i].interp = LP_INTERP_POSITION;
1037          shader->inputs[i].src_index = 0;
1038          continue;
1039       }
1040
1041       shader->inputs[i].src_index = i+1;
1042    }
1043
1044    if (LP_DEBUG & DEBUG_TGSI) {
1045       unsigned attrib;
1046       debug_printf("llvmpipe: Create fragment shader #%u %p:\n",
1047                    shader->no, (void *) shader);
1048       tgsi_dump(templ->tokens, 0);
1049       debug_printf("usage masks:\n");
1050       for (attrib = 0; attrib < shader->info.base.num_inputs; ++attrib) {
1051          unsigned usage_mask = shader->info.base.input_usage_mask[attrib];
1052          debug_printf("  IN[%u].%s%s%s%s\n",
1053                       attrib,
1054                       usage_mask & TGSI_WRITEMASK_X ? "x" : "",
1055                       usage_mask & TGSI_WRITEMASK_Y ? "y" : "",
1056                       usage_mask & TGSI_WRITEMASK_Z ? "z" : "",
1057                       usage_mask & TGSI_WRITEMASK_W ? "w" : "");
1058       }
1059       debug_printf("\n");
1060    }
1061
1062    return shader;
1063 }
1064
1065
1066 static void
1067 llvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs)
1068 {
1069    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
1070
1071    if (llvmpipe->fs == fs)
1072       return;
1073
1074    draw_flush(llvmpipe->draw);
1075
1076    llvmpipe->fs = (struct lp_fragment_shader *) fs;
1077
1078    draw_bind_fragment_shader(llvmpipe->draw,
1079                              (llvmpipe->fs ? llvmpipe->fs->draw_data : NULL));
1080
1081    llvmpipe->dirty |= LP_NEW_FS;
1082 }
1083
1084
1085 /**
1086  * Remove shader variant from two lists: the shader's variant list
1087  * and the context's variant list.
1088  */
1089 void
1090 llvmpipe_remove_shader_variant(struct llvmpipe_context *lp,
1091                                struct lp_fragment_shader_variant *variant)
1092 {
1093    unsigned i;
1094
1095    if (gallivm_debug & GALLIVM_DEBUG_IR) {
1096       debug_printf("llvmpipe: del fs #%u var #%u v created #%u v cached"
1097                    " #%u v total cached #%u\n",
1098                    variant->shader->no,
1099                    variant->no,
1100                    variant->shader->variants_created,
1101                    variant->shader->variants_cached,
1102                    lp->nr_fs_variants);
1103    }
1104
1105    /* free all the variant's JIT'd functions */
1106    for (i = 0; i < Elements(variant->function); i++) {
1107       if (variant->function[i]) {
1108          if (variant->jit_function[i])
1109             LLVMFreeMachineCodeForFunction(lp->gallivm->engine,
1110                                            variant->function[i]);
1111          LLVMDeleteFunction(variant->function[i]);
1112       }
1113    }
1114
1115    /* remove from shader's list */
1116    remove_from_list(&variant->list_item_local);
1117    variant->shader->variants_cached--;
1118
1119    /* remove from context's list */
1120    remove_from_list(&variant->list_item_global);
1121    lp->nr_fs_variants--;
1122
1123    FREE(variant);
1124 }
1125
1126
1127 static void
1128 llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs)
1129 {
1130    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
1131    struct lp_fragment_shader *shader = fs;
1132    struct lp_fs_variant_list_item *li;
1133
1134    assert(fs != llvmpipe->fs);
1135
1136    /*
1137     * XXX: we need to flush the context until we have some sort of reference
1138     * counting in fragment shaders as they may still be binned
1139     * Flushing alone might not sufficient we need to wait on it too.
1140     */
1141    llvmpipe_finish(pipe, __FUNCTION__);
1142
1143    /* Delete all the variants */
1144    li = first_elem(&shader->variants);
1145    while(!at_end(&shader->variants, li)) {
1146       struct lp_fs_variant_list_item *next = next_elem(li);
1147       llvmpipe_remove_shader_variant(llvmpipe, li->base);
1148       li = next;
1149    }
1150
1151    /* Delete draw module's data */
1152    draw_delete_fragment_shader(llvmpipe->draw, shader->draw_data);
1153
1154    assert(shader->variants_cached == 0);
1155    FREE((void *) shader->base.tokens);
1156    FREE(shader);
1157 }
1158
1159
1160
1161 static void
1162 llvmpipe_set_constant_buffer(struct pipe_context *pipe,
1163                              uint shader, uint index,
1164                              struct pipe_resource *constants)
1165 {
1166    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
1167    unsigned size = constants ? constants->width0 : 0;
1168    const void *data = constants ? llvmpipe_resource_data(constants) : NULL;
1169
1170    assert(shader < PIPE_SHADER_TYPES);
1171    assert(index < PIPE_MAX_CONSTANT_BUFFERS);
1172
1173    if(llvmpipe->constants[shader][index] == constants)
1174       return;
1175
1176    draw_flush(llvmpipe->draw);
1177
1178    /* note: reference counting */
1179    pipe_resource_reference(&llvmpipe->constants[shader][index], constants);
1180
1181    if(shader == PIPE_SHADER_VERTEX ||
1182       shader == PIPE_SHADER_GEOMETRY) {
1183       draw_set_mapped_constant_buffer(llvmpipe->draw, shader,
1184                                       index, data, size);
1185    }
1186
1187    llvmpipe->dirty |= LP_NEW_CONSTANTS;
1188 }
1189
1190
1191 /**
1192  * Return the blend factor equivalent to a destination alpha of one.
1193  */
1194 static INLINE unsigned
1195 force_dst_alpha_one(unsigned factor)
1196 {
1197    switch(factor) {
1198    case PIPE_BLENDFACTOR_DST_ALPHA:
1199       return PIPE_BLENDFACTOR_ONE;
1200    case PIPE_BLENDFACTOR_INV_DST_ALPHA:
1201       return PIPE_BLENDFACTOR_ZERO;
1202    case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
1203       return PIPE_BLENDFACTOR_ZERO;
1204    }
1205
1206    return factor;
1207 }
1208
1209
1210 /**
1211  * We need to generate several variants of the fragment pipeline to match
1212  * all the combinations of the contributing state atoms.
1213  *
1214  * TODO: there is actually no reason to tie this to context state -- the
1215  * generated code could be cached globally in the screen.
1216  */
1217 static void
1218 make_variant_key(struct llvmpipe_context *lp,
1219                  struct lp_fragment_shader *shader,
1220                  struct lp_fragment_shader_variant_key *key)
1221 {
1222    unsigned i;
1223
1224    memset(key, 0, shader->variant_key_size);
1225
1226    if (lp->framebuffer.zsbuf) {
1227       if (lp->depth_stencil->depth.enabled) {
1228          key->zsbuf_format = lp->framebuffer.zsbuf->format;
1229          memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth);
1230       }
1231       if (lp->depth_stencil->stencil[0].enabled) {
1232          key->zsbuf_format = lp->framebuffer.zsbuf->format;
1233          memcpy(&key->stencil, &lp->depth_stencil->stencil, sizeof key->stencil);
1234       }
1235    }
1236
1237    key->alpha.enabled = lp->depth_stencil->alpha.enabled;
1238    if(key->alpha.enabled)
1239       key->alpha.func = lp->depth_stencil->alpha.func;
1240    /* alpha.ref_value is passed in jit_context */
1241
1242    key->flatshade = lp->rasterizer->flatshade;
1243    if (lp->active_query_count) {
1244       key->occlusion_count = TRUE;
1245    }
1246
1247    if (lp->framebuffer.nr_cbufs) {
1248       memcpy(&key->blend, lp->blend, sizeof key->blend);
1249    }
1250
1251    key->nr_cbufs = lp->framebuffer.nr_cbufs;
1252    for (i = 0; i < lp->framebuffer.nr_cbufs; i++) {
1253       enum pipe_format format = lp->framebuffer.cbufs[i]->format;
1254       struct pipe_rt_blend_state *blend_rt = &key->blend.rt[i];
1255       const struct util_format_description *format_desc;
1256
1257       key->cbuf_format[i] = format;
1258
1259       format_desc = util_format_description(format);
1260       assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
1261              format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB);
1262
1263       blend_rt->colormask = lp->blend->rt[i].colormask;
1264
1265       /*
1266        * Mask out color channels not present in the color buffer.
1267        */
1268       blend_rt->colormask &= util_format_colormask(format_desc);
1269
1270       /*
1271        * Our swizzled render tiles always have an alpha channel, but the linear
1272        * render target format often does not, so force here the dst alpha to be
1273        * one.
1274        *
1275        * This is not a mere optimization. Wrong results will be produced if the
1276        * dst alpha is used, the dst format does not have alpha, and the previous
1277        * rendering was not flushed from the swizzled to linear buffer. For
1278        * example, NonPowTwo DCT.
1279        *
1280        * TODO: This should be generalized to all channels for better
1281        * performance, but only alpha causes correctness issues.
1282        *
1283        * Also, force rgb/alpha func/factors match, to make AoS blending easier.
1284        */
1285       if (format_desc->swizzle[3] > UTIL_FORMAT_SWIZZLE_W) {
1286          blend_rt->rgb_src_factor   = force_dst_alpha_one(blend_rt->rgb_src_factor);
1287          blend_rt->rgb_dst_factor   = force_dst_alpha_one(blend_rt->rgb_dst_factor);
1288          blend_rt->alpha_func       = blend_rt->rgb_func;
1289          blend_rt->alpha_src_factor = blend_rt->rgb_src_factor;
1290          blend_rt->alpha_dst_factor = blend_rt->rgb_dst_factor;
1291       }
1292    }
1293
1294    /* This value will be the same for all the variants of a given shader:
1295     */
1296    key->nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
1297
1298    for(i = 0; i < key->nr_samplers; ++i) {
1299       if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
1300          lp_sampler_static_state(&key->sampler[i],
1301                                  lp->fragment_sampler_views[i],
1302                                  lp->sampler[i]);
1303       }
1304    }
1305 }
1306
1307
1308
1309 /**
1310  * Update fragment shader state.  This is called just prior to drawing
1311  * something when some fragment-related state has changed.
1312  */
1313 void
1314 llvmpipe_update_fs(struct llvmpipe_context *lp)
1315 {
1316    struct lp_fragment_shader *shader = lp->fs;
1317    struct lp_fragment_shader_variant_key key;
1318    struct lp_fragment_shader_variant *variant = NULL;
1319    struct lp_fs_variant_list_item *li;
1320
1321    make_variant_key(lp, shader, &key);
1322
1323    /* Search the variants for one which matches the key */
1324    li = first_elem(&shader->variants);
1325    while(!at_end(&shader->variants, li)) {
1326       if(memcmp(&li->base->key, &key, shader->variant_key_size) == 0) {
1327          variant = li->base;
1328          break;
1329       }
1330       li = next_elem(li);
1331    }
1332
1333    if (variant) {
1334       /* Move this variant to the head of the list to implement LRU
1335        * deletion of shader's when we have too many.
1336        */
1337       move_to_head(&lp->fs_variants_list, &variant->list_item_global);
1338    }
1339    else {
1340       /* variant not found, create it now */
1341       int64_t t0, t1, dt;
1342       unsigned i;
1343
1344       /* First, check if we've exceeded the max number of shader variants.
1345        * If so, free 25% of them (the least recently used ones).
1346        */
1347       if (lp->nr_fs_variants >= LP_MAX_SHADER_VARIANTS) {
1348          struct pipe_context *pipe = &lp->pipe;
1349
1350          /*
1351           * XXX: we need to flush the context until we have some sort of
1352           * reference counting in fragment shaders as they may still be binned
1353           * Flushing alone might not be sufficient we need to wait on it too.
1354           */
1355          llvmpipe_finish(pipe, __FUNCTION__);
1356
1357          for (i = 0; i < LP_MAX_SHADER_VARIANTS / 4; i++) {
1358             struct lp_fs_variant_list_item *item;
1359             item = last_elem(&lp->fs_variants_list);
1360             llvmpipe_remove_shader_variant(lp, item->base);
1361          }
1362       }
1363
1364       /*
1365        * Generate the new variant.
1366        */
1367       t0 = os_time_get();
1368       variant = generate_variant(lp, shader, &key);
1369       t1 = os_time_get();
1370       dt = t1 - t0;
1371       LP_COUNT_ADD(llvm_compile_time, dt);
1372       LP_COUNT_ADD(nr_llvm_compiles, 2);  /* emit vs. omit in/out test */
1373
1374       llvmpipe_variant_count++;
1375
1376       /* Put the new variant into the list */
1377       if (variant) {
1378          insert_at_head(&shader->variants, &variant->list_item_local);
1379          insert_at_head(&lp->fs_variants_list, &variant->list_item_global);
1380          lp->nr_fs_variants++;
1381          shader->variants_cached++;
1382       }
1383    }
1384
1385    /* Bind this variant */
1386    lp_setup_set_fs_variant(lp->setup, variant);
1387 }
1388
1389
1390
1391
1392
1393
1394
1395 void
1396 llvmpipe_init_fs_funcs(struct llvmpipe_context *llvmpipe)
1397 {
1398    llvmpipe->pipe.create_fs_state = llvmpipe_create_fs_state;
1399    llvmpipe->pipe.bind_fs_state   = llvmpipe_bind_fs_state;
1400    llvmpipe->pipe.delete_fs_state = llvmpipe_delete_fs_state;
1401
1402    llvmpipe->pipe.set_constant_buffer = llvmpipe_set_constant_buffer;
1403 }