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_screen.h"
  94 #include "lp_setup.h"
  95 #include "lp_state.h"
  96 #include "lp_tex_sample.h"
  97 #include "lp_flush.h"
  98 #include "lp_state_fs.h"
  99
 100
 101 #include <llvm-c/Analysis.h>
 102 #include <llvm-c/BitWriter.h>
 103
 104
 105 static unsigned fs_no = 0;
 106
 107
 108
 109 /**
 110  * Expand the relevent bits of mask_input to a 4-dword mask for the
 111  * four pixels in a 2x2 quad.  This will set the four elements of the
 112  * quad mask vector to 0 or ~0.
 113  *
 114  * \param quad  which quad of the quad group to test, in [0,3]
 115  * \param mask_input  bitwise mask for the whole 4x4 stamp
 116  */
 117 static LLVMValueRef
 118 generate_quad_mask(LLVMBuilderRef builder,
 119                    struct lp_type fs_type,
 120                    unsigned quad,
 121                    LLVMValueRef mask_input) /* int32 */
 122 {
 123    struct lp_type mask_type;
 124    LLVMTypeRef i32t = LLVMInt32Type();
 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
 140    switch (quad) {
 141    case 0:
 142       shift = 0;
 143       break;
 144    case 1:
 145       shift = 2;
 146       break;
 147    case 2:
 148       shift = 8;
 149       break;
 150    case 3:
 151       shift = 10;
 152       break;
 153    default:
 154       assert(0);
 155       shift = 0;
 156    }
 157
 158    mask_input = LLVMBuildLShr(builder,
 159                               mask_input,
 160                               LLVMConstInt(i32t, shift, 0),
 161                               "");
 162
 163    /*
 164     * mask = { mask_input & (1 << i), for i in [0,3] }
 165     */
 166
 167    mask = lp_build_broadcast(builder, lp_build_vec_type(mask_type), 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
 180    mask = lp_build_compare(builder,
 181                            mask_type, PIPE_FUNC_NOTEQUAL,
 182                            mask,
 183                            lp_build_const_int_vec(mask_type, 0));
 184
 185    return mask;
 186 }
 187
 188
 189 #define EARLY_DEPTH_TEST  0x1
 190 #define LATE_DEPTH_TEST   0x2
 191 #define EARLY_DEPTH_WRITE 0x4
 192 #define LATE_DEPTH_WRITE  0x8
 193
 194 static int
 195 find_output_by_semantic( const struct tgsi_shader_info *info,
 196                          unsigned semantic,
 197                          unsigned index )
 198 {
 199    int i;
 200
 201    for (i = 0; i < info->num_outputs; i++)
 202       if (info->output_semantic_name[i] == semantic &&
 203           info->output_semantic_index[i] == index)
 204          return i;
 205
 206    return -1;
 207 }
 208
 209
 210 /**
 211  * Generate the fragment shader, depth/stencil test, and alpha tests.
 212  * \param i  which quad in the tile, in range [0,3]
 213  * \param partial_mask  if 1, do mask_input testing
 214  */
 215 static void
 216 generate_fs(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(builder, context_ptr);
 282    stencil_refs[1] = lp_jit_context_stencil_ref_back_value(builder, context_ptr);
 283
 284    vec_type = lp_build_vec_type(type);
 285
 286    consts_ptr = lp_jit_context_constants(builder, 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(builder, vec_type, "color");
 294       }
 295    }
 296
 297    /* do triangle edge testing */
 298    if (partial_mask) {
 299       *pmask = generate_quad_mask(builder, type,
 300                                   i, mask_input);
 301    }
 302    else {
 303       *pmask = lp_build_const_int_vec(type, ~0);
 304    }
 305
 306    /* 'mask' will control execution based on quad's pixel alive/killed state */
 307    lp_build_mask_begin(&mask, builder, 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, i);
 313    z = interp->pos[2];
 314
 315    if (depth_mode & EARLY_DEPTH_TEST) {
 316       lp_build_depth_stencil_test(builder,
 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, i);
 334
 335    /* Build the actual shader */
 336    lp_build_tgsi_soa(builder, tokens, type, &mask,
 337                      consts_ptr, NULL, /* sys values array */
 338                      interp->pos, interp->inputs,
 339                      outputs, sampler, &shader->info.base);
 340
 341
 342    /* Alpha test */
 343    if (key->alpha.enabled) {
 344       int color0 = find_output_by_semantic(&shader->info.base,
 345                                            TGSI_SEMANTIC_COLOR,
 346                                            0);
 347
 348       if (color0 != -1 && outputs[color0][3]) {
 349          LLVMValueRef alpha = LLVMBuildLoad(builder, outputs[color0][3], "alpha");
 350          LLVMValueRef alpha_ref_value;
 351
 352          alpha_ref_value = lp_jit_context_alpha_ref_value(builder, context_ptr);
 353          alpha_ref_value = lp_build_broadcast(builder, vec_type, alpha_ref_value);
 354
 355          lp_build_alpha_test(builder, key->alpha.func, type,
 356                              &mask, alpha, alpha_ref_value,
 357                              (depth_mode & LATE_DEPTH_TEST) != 0);
 358       }
 359    }
 360
 361    /* Late Z test */
 362    if (depth_mode & LATE_DEPTH_TEST) {
 363       int pos0 = find_output_by_semantic(&shader->info.base,
 364                                          TGSI_SEMANTIC_POSITION,
 365                                          0);
 366
 367       if (pos0 != -1 && outputs[pos0][2]) {
 368          z = LLVMBuildLoad(builder, outputs[pos0][2], "output.z");
 369       }
 370
 371       lp_build_depth_stencil_test(builder,
 372                                   &key->depth,
 373                                   key->stencil,
 374                                   type,
 375                                   zs_format_desc,
 376                                   &mask,
 377                                   stencil_refs,
 378                                   z,
 379                                   depth_ptr, facing,
 380                                   &zs_value,
 381                                   !simple_shader);
 382       /* Late Z write */
 383       if (depth_mode & LATE_DEPTH_WRITE) {
 384          lp_build_depth_write(builder, zs_format_desc, depth_ptr, zs_value);
 385       }
 386    }
 387    else if ((depth_mode & EARLY_DEPTH_TEST) &&
 388             (depth_mode & LATE_DEPTH_WRITE))
 389    {
 390       /* Need to apply a reduced mask to the depth write.  Reload the
 391        * depth value, update from zs_value with the new mask value and
 392        * write that out.
 393        */
 394       lp_build_deferred_depth_write(builder,
 395                                     type,
 396                                     zs_format_desc,
 397                                     &mask,
 398                                     depth_ptr,
 399                                     zs_value);
 400    }
 401
 402
 403    /* Color write  */
 404    for (attrib = 0; attrib < shader->info.base.num_outputs; ++attrib)
 405    {
 406       if (shader->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_COLOR &&
 407           shader->info.base.output_semantic_index[attrib] < key->nr_cbufs)
 408       {
 409          unsigned cbuf = shader->info.base.output_semantic_index[attrib];
 410          for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 411             if(outputs[attrib][chan]) {
 412                /* XXX: just initialize outputs to point at colors[] and
 413                 * skip this.
 414                 */
 415                LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
 416                lp_build_name(out, "color%u.%u.%c", i, attrib, "rgba"[chan]);
 417                LLVMBuildStore(builder, out, color[cbuf][chan]);
 418             }
 419          }
 420       }
 421    }
 422
 423    if (counter)
 424       lp_build_occlusion_count(builder, type,
 425                                lp_build_mask_value(&mask), counter);
 426
 427    *pmask = lp_build_mask_end(&mask);
 428 }
 429
 430
 431 /**
 432  * Generate color blending and color output.
 433  * \param rt  the render target index (to index blend, colormask state)
 434  * \param type  the pixel color type
 435  * \param context_ptr  pointer to the runtime JIT context
 436  * \param mask  execution mask (active fragment/pixel mask)
 437  * \param src  colors from the fragment shader
 438  * \param dst_ptr  the destination color buffer pointer
 439  */
 440 static void
 441 generate_blend(const struct pipe_blend_state *blend,
 442                unsigned rt,
 443                LLVMBuilderRef builder,
 444                struct lp_type type,
 445                LLVMValueRef context_ptr,
 446                LLVMValueRef mask,
 447                LLVMValueRef *src,
 448                LLVMValueRef dst_ptr,
 449                boolean do_branch)
 450 {
 451    struct lp_build_context bld;
 452    struct lp_build_mask_context mask_ctx;
 453    LLVMTypeRef vec_type;
 454    LLVMValueRef const_ptr;
 455    LLVMValueRef con[4];
 456    LLVMValueRef dst[4];
 457    LLVMValueRef res[4];
 458    unsigned chan;
 459
 460    lp_build_context_init(&bld, builder, type);
 461
 462    lp_build_mask_begin(&mask_ctx, builder, type, mask);
 463    if (do_branch)
 464       lp_build_mask_check(&mask_ctx);
 465
 466    vec_type = lp_build_vec_type(type);
 467
 468    const_ptr = lp_jit_context_blend_color(builder, context_ptr);
 469    const_ptr = LLVMBuildBitCast(builder, const_ptr,
 470                                 LLVMPointerType(vec_type, 0), "");
 471
 472    /* load constant blend color and colors from the dest color buffer */
 473    for(chan = 0; chan < 4; ++chan) {
 474       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
 475       con[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, const_ptr, &index, 1, ""), "");
 476
 477       dst[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, dst_ptr, &index, 1, ""), "");
 478
 479       lp_build_name(con[chan], "con.%c", "rgba"[chan]);
 480       lp_build_name(dst[chan], "dst.%c", "rgba"[chan]);
 481    }
 482
 483    /* do blend */
 484    lp_build_blend_soa(builder, blend, type, rt, src, dst, con, res);
 485
 486    /* store results to color buffer */
 487    for(chan = 0; chan < 4; ++chan) {
 488       if(blend->rt[rt].colormask & (1 << chan)) {
 489          LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
 490          lp_build_name(res[chan], "res.%c", "rgba"[chan]);
 491          res[chan] = lp_build_select(&bld, mask, res[chan], dst[chan]);
 492          LLVMBuildStore(builder, res[chan], LLVMBuildGEP(builder, dst_ptr, &index, 1, ""));
 493       }
 494    }
 495
 496    lp_build_mask_end(&mask_ctx);
 497 }
 498
 499
 500 /**
 501  * Generate the runtime callable function for the whole fragment pipeline.
 502  * Note that the function which we generate operates on a block of 16
 503  * pixels at at time.  The block contains 2x2 quads.  Each quad contains
 504  * 2x2 pixels.
 505  */
 506 static void
 507 generate_fragment(struct llvmpipe_screen *screen,
 508                   struct lp_fragment_shader *shader,
 509                   struct lp_fragment_shader_variant *variant,
 510                   unsigned partial_mask)
 511 {
 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    LLVMValueRef context_ptr;
 523    LLVMValueRef x;
 524    LLVMValueRef y;
 525    LLVMValueRef a0_ptr;
 526    LLVMValueRef dadx_ptr;
 527    LLVMValueRef dady_ptr;
 528    LLVMValueRef color_ptr_ptr;
 529    LLVMValueRef depth_ptr;
 530    LLVMValueRef mask_input;
 531    LLVMValueRef counter = NULL;
 532    LLVMBasicBlockRef block;
 533    LLVMBuilderRef builder;
 534    struct lp_build_sampler_soa *sampler;
 535    struct lp_build_interp_soa_context interp;
 536    LLVMValueRef fs_mask[LP_MAX_VECTOR_LENGTH];
 537    LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS][LP_MAX_VECTOR_LENGTH];
 538    LLVMValueRef blend_mask;
 539    LLVMValueRef function;
 540    LLVMValueRef facing;
 541    const struct util_format_description *zs_format_desc;
 542    unsigned num_fs;
 543    unsigned i;
 544    unsigned chan;
 545    unsigned cbuf;
 546
 547    /* Adjust color input interpolation according to flatshade state:
 548     */
 549    memcpy(inputs, shader->inputs, shader->info.base.num_inputs * sizeof inputs[0]);
 550    for (i = 0; i < shader->info.base.num_inputs; i++) {
 551       if (inputs[i].interp == LP_INTERP_COLOR) {
 552          if (key->flatshade)
 553             inputs[i].interp = LP_INTERP_CONSTANT;
 554          else
 555             inputs[i].interp = LP_INTERP_LINEAR;
 556       }
 557    }
 558
 559
 560    /* TODO: actually pick these based on the fs and color buffer
 561     * characteristics. */
 562
 563    memset(&fs_type, 0, sizeof fs_type);
 564    fs_type.floating = TRUE; /* floating point values */
 565    fs_type.sign = TRUE;     /* values are signed */
 566    fs_type.norm = FALSE;    /* values are not limited to [0,1] or [-1,1] */
 567    fs_type.width = 32;      /* 32-bit float */
 568    fs_type.length = 4;      /* 4 elements per vector */
 569    num_fs = 4;              /* number of quads per block */
 570
 571    memset(&blend_type, 0, sizeof blend_type);
 572    blend_type.floating = FALSE; /* values are integers */
 573    blend_type.sign = FALSE;     /* values are unsigned */
 574    blend_type.norm = TRUE;      /* values are in [0,1] or [-1,1] */
 575    blend_type.width = 8;        /* 8-bit ubyte values */
 576    blend_type.length = 16;      /* 16 elements per vector */
 577
 578    /*
 579     * Generate the function prototype. Any change here must be reflected in
 580     * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
 581     */
 582
 583    fs_elem_type = lp_build_elem_type(fs_type);
 584    fs_int_vec_type = lp_build_int_vec_type(fs_type);
 585
 586    blend_vec_type = lp_build_vec_type(blend_type);
 587
 588    util_snprintf(func_name, sizeof(func_name), "fs%u_variant%u_%s",
 589                  shader->no, variant->no, partial_mask ? "partial" : "whole");
 590
 591    arg_types[0] = screen->context_ptr_type;            /* context */
 592    arg_types[1] = LLVMInt32Type();                     /* x */
 593    arg_types[2] = LLVMInt32Type();                     /* y */
 594    arg_types[3] = LLVMInt32Type();                     /* facing */
 595    arg_types[4] = LLVMPointerType(fs_elem_type, 0);    /* a0 */
 596    arg_types[5] = LLVMPointerType(fs_elem_type, 0);    /* dadx */
 597    arg_types[6] = LLVMPointerType(fs_elem_type, 0);    /* dady */
 598    arg_types[7] = LLVMPointerType(LLVMPointerType(blend_vec_type, 0), 0);  /* color */
 599    arg_types[8] = LLVMPointerType(LLVMInt8Type(), 0);  /* depth */
 600    arg_types[9] = LLVMInt32Type();                     /* mask_input */
 601    arg_types[10] = LLVMPointerType(LLVMInt32Type(), 0);/* counter */
 602
 603    func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
 604
 605    function = LLVMAddFunction(screen->module, func_name, func_type);
 606    LLVMSetFunctionCallConv(function, LLVMCCallConv);
 607
 608    variant->function[partial_mask] = function;
 609
 610    /* XXX: need to propagate noalias down into color param now we are
 611     * passing a pointer-to-pointer?
 612     */
 613    for(i = 0; i < Elements(arg_types); ++i)
 614       if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
 615          LLVMAddAttribute(LLVMGetParam(function, i), LLVMNoAliasAttribute);
 616
 617    context_ptr  = LLVMGetParam(function, 0);
 618    x            = LLVMGetParam(function, 1);
 619    y            = LLVMGetParam(function, 2);
 620    facing       = LLVMGetParam(function, 3);
 621    a0_ptr       = LLVMGetParam(function, 4);
 622    dadx_ptr     = LLVMGetParam(function, 5);
 623    dady_ptr     = LLVMGetParam(function, 6);
 624    color_ptr_ptr = LLVMGetParam(function, 7);
 625    depth_ptr    = LLVMGetParam(function, 8);
 626    mask_input   = LLVMGetParam(function, 9);
 627
 628    lp_build_name(context_ptr, "context");
 629    lp_build_name(x, "x");
 630    lp_build_name(y, "y");
 631    lp_build_name(a0_ptr, "a0");
 632    lp_build_name(dadx_ptr, "dadx");
 633    lp_build_name(dady_ptr, "dady");
 634    lp_build_name(color_ptr_ptr, "color_ptr_ptr");
 635    lp_build_name(depth_ptr, "depth");
 636    lp_build_name(mask_input, "mask_input");
 637
 638    if (key->occlusion_count) {
 639       counter = LLVMGetParam(function, 10);
 640       lp_build_name(counter, "counter");
 641    }
 642
 643    /*
 644     * Function body
 645     */
 646
 647    block = LLVMAppendBasicBlock(function, "entry");
 648    builder = LLVMCreateBuilder();
 649    LLVMPositionBuilderAtEnd(builder, block);
 650
 651    /*
 652     * The shader input interpolation info is not explicitely baked in the
 653     * shader key, but everything it derives from (TGSI, and flatshade) is
 654     * already included in the shader key.
 655     */
 656    lp_build_interp_soa_init(&interp,
 657                             shader->info.base.num_inputs,
 658                             inputs,
 659                             builder, fs_type,
 660                             a0_ptr, dadx_ptr, dady_ptr,
 661                             x, y);
 662
 663    /* code generated texture sampling */
 664    sampler = lp_llvm_sampler_soa_create(key->sampler, context_ptr);
 665
 666    /* loop over quads in the block */
 667    zs_format_desc = util_format_description(key->zsbuf_format);
 668
 669    for(i = 0; i < num_fs; ++i) {
 670       LLVMValueRef depth_offset = LLVMConstInt(LLVMInt32Type(),
 671                                                i*fs_type.length*zs_format_desc->block.bits/8,
 672                                                0);
 673       LLVMValueRef out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS];
 674       LLVMValueRef depth_ptr_i;
 675
 676       depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &depth_offset, 1, "");
 677
 678       generate_fs(shader, key,
 679                   builder,
 680                   fs_type,
 681                   context_ptr,
 682                   i,
 683                   &interp,
 684                   sampler,
 685                   &fs_mask[i], /* output */
 686                   out_color,
 687                   depth_ptr_i,
 688                   facing,
 689                   partial_mask,
 690                   mask_input,
 691                   counter);
 692
 693       for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++)
 694          for(chan = 0; chan < NUM_CHANNELS; ++chan)
 695             fs_out_color[cbuf][chan][i] = out_color[cbuf][chan];
 696    }
 697
 698    sampler->destroy(sampler);
 699
 700    /* Loop over color outputs / color buffers to do blending.
 701     */
 702    for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
 703       LLVMValueRef color_ptr;
 704       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), cbuf, 0);
 705       LLVMValueRef blend_in_color[NUM_CHANNELS];
 706       unsigned rt;
 707
 708       /*
 709        * Convert the fs's output color and mask to fit to the blending type.
 710        */
 711       for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 712          LLVMValueRef fs_color_vals[LP_MAX_VECTOR_LENGTH];
 713
 714          for (i = 0; i < num_fs; i++) {
 715             fs_color_vals[i] =
 716                LLVMBuildLoad(builder, fs_out_color[cbuf][chan][i], "fs_color_vals");
 717          }
 718
 719          lp_build_conv(builder, fs_type, blend_type,
 720                        fs_color_vals,
 721                        num_fs,
 722                        &blend_in_color[chan], 1);
 723
 724          lp_build_name(blend_in_color[chan], "color%d.%c", cbuf, "rgba"[chan]);
 725       }
 726
 727       if (partial_mask || !variant->opaque) {
 728          lp_build_conv_mask(builder, fs_type, blend_type,
 729                             fs_mask, num_fs,
 730                             &blend_mask, 1);
 731       } else {
 732          blend_mask = lp_build_const_int_vec(blend_type, ~0);
 733       }
 734
 735       color_ptr = LLVMBuildLoad(builder,
 736                                 LLVMBuildGEP(builder, color_ptr_ptr, &index, 1, ""),
 737                                 "");
 738       lp_build_name(color_ptr, "color_ptr%d", cbuf);
 739
 740       /* which blend/colormask state to use */
 741       rt = key->blend.independent_blend_enable ? cbuf : 0;
 742
 743       /*
 744        * Blending.
 745        */
 746       {
 747          /* Could the 4x4 have been killed?
 748           */
 749          boolean do_branch = ((key->depth.enabled || key->stencil[0].enabled) &&
 750                               !key->alpha.enabled &&
 751                               !shader->info.base.uses_kill);
 752
 753          generate_blend(&key->blend,
 754                         rt,
 755                         builder,
 756                         blend_type,
 757                         context_ptr,
 758                         blend_mask,
 759                         blend_in_color,
 760                         color_ptr,
 761                         do_branch);
 762       }
 763    }
 764
 765    LLVMBuildRetVoid(builder);
 766
 767    LLVMDisposeBuilder(builder);
 768
 769
 770    /* Verify the LLVM IR.  If invalid, dump and abort */
 771 #ifdef DEBUG
 772    if(LLVMVerifyFunction(function, LLVMPrintMessageAction)) {
 773       if (1)
 774          lp_debug_dump_value(function);
 775       abort();
 776    }
 777 #endif
 778
 779    /* Apply optimizations to LLVM IR */
 780    LLVMRunFunctionPassManager(screen->pass, function);
 781
 782    if ((gallivm_debug & GALLIVM_DEBUG_IR) || (LP_DEBUG & DEBUG_FS)) {
 783       /* Print the LLVM IR to stderr */
 784       lp_debug_dump_value(function);
 785       debug_printf("\n");
 786    }
 787
 788    /* Dump byte code to a file */
 789    if (0) {
 790       LLVMWriteBitcodeToFile(lp_build_module, "llvmpipe.bc");
 791    }
 792
 793    /*
 794     * Translate the LLVM IR into machine code.
 795     */
 796    {
 797       void *f = LLVMGetPointerToGlobal(screen->engine, function);
 798
 799       variant->jit_function[partial_mask] = (lp_jit_frag_func)pointer_to_func(f);
 800
 801       if ((gallivm_debug & GALLIVM_DEBUG_ASM) || (LP_DEBUG & DEBUG_FS)) {
 802          lp_disassemble(f);
 803       }
 804       lp_func_delete_body(function);
 805    }
 806 }
 807
 808
 809 static void
 810 dump_fs_variant_key(const struct lp_fragment_shader_variant_key *key)
 811 {
 812    unsigned i;
 813
 814    debug_printf("fs variant %p:\n", (void *) key);
 815
 816    if (key->flatshade) {
 817       debug_printf("flatshade = 1\n");
 818    }
 819    for (i = 0; i < key->nr_cbufs; ++i) {
 820       debug_printf("cbuf_format[%u] = %s\n", i, util_format_name(key->cbuf_format[i]));
 821    }
 822    if (key->depth.enabled) {
 823       debug_printf("depth.format = %s\n", util_format_name(key->zsbuf_format));
 824       debug_printf("depth.func = %s\n", util_dump_func(key->depth.func, TRUE));
 825       debug_printf("depth.writemask = %u\n", key->depth.writemask);
 826    }
 827
 828    for (i = 0; i < 2; ++i) {
 829       if (key->stencil[i].enabled) {
 830          debug_printf("stencil[%u].func = %s\n", i, util_dump_func(key->stencil[i].func, TRUE));
 831          debug_printf("stencil[%u].fail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].fail_op, TRUE));
 832          debug_printf("stencil[%u].zpass_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zpass_op, TRUE));
 833          debug_printf("stencil[%u].zfail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zfail_op, TRUE));
 834          debug_printf("stencil[%u].valuemask = 0x%x\n", i, key->stencil[i].valuemask);
 835          debug_printf("stencil[%u].writemask = 0x%x\n", i, key->stencil[i].writemask);
 836       }
 837    }
 838
 839    if (key->alpha.enabled) {
 840       debug_printf("alpha.func = %s\n", util_dump_func(key->alpha.func, TRUE));
 841    }
 842
 843    if (key->occlusion_count) {
 844       debug_printf("occlusion_count = 1\n");
 845    }
 846
 847    if (key->blend.logicop_enable) {
 848       debug_printf("blend.logicop_func = %s\n", util_dump_logicop(key->blend.logicop_func, TRUE));
 849    }
 850    else if (key->blend.rt[0].blend_enable) {
 851       debug_printf("blend.rgb_func = %s\n",   util_dump_blend_func  (key->blend.rt[0].rgb_func, TRUE));
 852       debug_printf("blend.rgb_src_factor = %s\n",   util_dump_blend_factor(key->blend.rt[0].rgb_src_factor, TRUE));
 853       debug_printf("blend.rgb_dst_factor = %s\n",   util_dump_blend_factor(key->blend.rt[0].rgb_dst_factor, TRUE));
 854       debug_printf("blend.alpha_func = %s\n",       util_dump_blend_func  (key->blend.rt[0].alpha_func, TRUE));
 855       debug_printf("blend.alpha_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_src_factor, TRUE));
 856       debug_printf("blend.alpha_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_dst_factor, TRUE));
 857    }
 858    debug_printf("blend.colormask = 0x%x\n", key->blend.rt[0].colormask);
 859    for (i = 0; i < key->nr_samplers; ++i) {
 860       debug_printf("sampler[%u] = \n", i);
 861       debug_printf("  .format = %s\n",
 862                    util_format_name(key->sampler[i].format));
 863       debug_printf("  .target = %s\n",
 864                    util_dump_tex_target(key->sampler[i].target, TRUE));
 865       debug_printf("  .pot = %u %u %u\n",
 866                    key->sampler[i].pot_width,
 867                    key->sampler[i].pot_height,
 868                    key->sampler[i].pot_depth);
 869       debug_printf("  .wrap = %s %s %s\n",
 870                    util_dump_tex_wrap(key->sampler[i].wrap_s, TRUE),
 871                    util_dump_tex_wrap(key->sampler[i].wrap_t, TRUE),
 872                    util_dump_tex_wrap(key->sampler[i].wrap_r, TRUE));
 873       debug_printf("  .min_img_filter = %s\n",
 874                    util_dump_tex_filter(key->sampler[i].min_img_filter, TRUE));
 875       debug_printf("  .min_mip_filter = %s\n",
 876                    util_dump_tex_mipfilter(key->sampler[i].min_mip_filter, TRUE));
 877       debug_printf("  .mag_img_filter = %s\n",
 878                    util_dump_tex_filter(key->sampler[i].mag_img_filter, TRUE));
 879       if (key->sampler[i].compare_mode != PIPE_TEX_COMPARE_NONE)
 880          debug_printf("  .compare_func = %s\n", util_dump_func(key->sampler[i].compare_func, TRUE));
 881       debug_printf("  .normalized_coords = %u\n", key->sampler[i].normalized_coords);
 882       debug_printf("  .min_max_lod_equal = %u\n", key->sampler[i].min_max_lod_equal);
 883       debug_printf("  .lod_bias_non_zero = %u\n", key->sampler[i].lod_bias_non_zero);
 884       debug_printf("  .apply_min_lod = %u\n", key->sampler[i].apply_min_lod);
 885       debug_printf("  .apply_max_lod = %u\n", key->sampler[i].apply_max_lod);
 886    }
 887 }
 888
 889
 890 void
 891 lp_debug_fs_variant(const struct lp_fragment_shader_variant *variant)
 892 {
 893    debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
 894                 variant->shader->no, variant->no);
 895    tgsi_dump(variant->shader->base.tokens, 0);
 896    dump_fs_variant_key(&variant->key);
 897    debug_printf("variant->opaque = %u\n", variant->opaque);
 898    debug_printf("\n");
 899 }
 900
 901 static struct lp_fragment_shader_variant *
 902 generate_variant(struct llvmpipe_screen *screen,
 903                  struct lp_fragment_shader *shader,
 904                  const struct lp_fragment_shader_variant_key *key)
 905 {
 906    struct lp_fragment_shader_variant *variant;
 907    boolean fullcolormask;
 908
 909    variant = CALLOC_STRUCT(lp_fragment_shader_variant);
 910    if(!variant)
 911       return NULL;
 912
 913    variant->shader = shader;
 914    variant->list_item_global.base = variant;
 915    variant->list_item_local.base = variant;
 916    variant->no = shader->variants_created++;
 917
 918    memcpy(&variant->key, key, shader->variant_key_size);
 919
 920    /*
 921     * Determine whether we are touching all channels in the color buffer.
 922     */
 923    fullcolormask = FALSE;
 924    if (key->nr_cbufs == 1) {
 925       const struct util_format_description *format_desc;
 926       format_desc = util_format_description(key->cbuf_format[0]);
 927       if ((~key->blend.rt[0].colormask &
 928            util_format_colormask(format_desc)) == 0) {
 929          fullcolormask = TRUE;
 930       }
 931    }
 932
 933    variant->opaque =
 934          !key->blend.logicop_enable &&
 935          !key->blend.rt[0].blend_enable &&
 936          fullcolormask &&
 937          !key->stencil[0].enabled &&
 938          !key->alpha.enabled &&
 939          !key->depth.enabled &&
 940          !shader->info.base.uses_kill
 941          ? TRUE : FALSE;
 942
 943
 944    if ((LP_DEBUG & DEBUG_FS) || (gallivm_debug & GALLIVM_DEBUG_IR)) {
 945       lp_debug_fs_variant(variant);
 946    }
 947
 948    generate_fragment(screen, shader, variant, RAST_EDGE_TEST);
 949
 950    if (variant->opaque) {
 951       /* Specialized shader, which doesn't need to read the color buffer. */
 952       generate_fragment(screen, shader, variant, RAST_WHOLE);
 953    } else {
 954       variant->jit_function[RAST_WHOLE] = variant->jit_function[RAST_EDGE_TEST];
 955    }
 956
 957    return variant;
 958 }
 959
 960
 961 static void *
 962 llvmpipe_create_fs_state(struct pipe_context *pipe,
 963                          const struct pipe_shader_state *templ)
 964 {
 965    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
 966    struct lp_fragment_shader *shader;
 967    int nr_samplers;
 968    int i;
 969
 970    shader = CALLOC_STRUCT(lp_fragment_shader);
 971    if (!shader)
 972       return NULL;
 973
 974    shader->no = fs_no++;
 975    make_empty_list(&shader->variants);
 976
 977    /* get/save the summary info for this shader */
 978    lp_build_tgsi_info(templ->tokens, &shader->info);
 979
 980    /* we need to keep a local copy of the tokens */
 981    shader->base.tokens = tgsi_dup_tokens(templ->tokens);
 982
 983    shader->draw_data = draw_create_fragment_shader(llvmpipe->draw, templ);
 984    if (shader->draw_data == NULL) {
 985       FREE((void *) shader->base.tokens);
 986       FREE(shader);
 987       return NULL;
 988    }
 989
 990    nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
 991
 992    shader->variant_key_size = Offset(struct lp_fragment_shader_variant_key,
 993                                      sampler[nr_samplers]);
 994
 995    for (i = 0; i < shader->info.base.num_inputs; i++) {
 996       shader->inputs[i].usage_mask = shader->info.base.input_usage_mask[i];
 997
 998       switch (shader->info.base.input_interpolate[i]) {
 999       case TGSI_INTERPOLATE_CONSTANT:
1000          shader->inputs[i].interp = LP_INTERP_CONSTANT;
1001          break;
1002       case TGSI_INTERPOLATE_LINEAR:
1003          shader->inputs[i].interp = LP_INTERP_LINEAR;
1004          break;
1005       case TGSI_INTERPOLATE_PERSPECTIVE:
1006          shader->inputs[i].interp = LP_INTERP_PERSPECTIVE;
1007          break;
1008       default:
1009          assert(0);
1010          break;
1011       }
1012
1013       switch (shader->info.base.input_semantic_name[i]) {
1014       case TGSI_SEMANTIC_COLOR:
1015          /* Colors may be either linearly or constant interpolated in
1016           * the fragment shader, but that information isn't available
1017           * here.  Mark color inputs and fix them up later.
1018           */
1019          shader->inputs[i].interp = LP_INTERP_COLOR;
1020          break;
1021       case TGSI_SEMANTIC_FACE:
1022          shader->inputs[i].interp = LP_INTERP_FACING;
1023          break;
1024       case TGSI_SEMANTIC_POSITION:
1025          /* Position was already emitted above
1026           */
1027          shader->inputs[i].interp = LP_INTERP_POSITION;
1028          shader->inputs[i].src_index = 0;
1029          continue;
1030       }
1031
1032       shader->inputs[i].src_index = i+1;
1033    }
1034
1035    if (LP_DEBUG & DEBUG_TGSI) {
1036       unsigned attrib;
1037       debug_printf("llvmpipe: Create fragment shader #%u %p:\n", shader->no, (void *) shader);
1038       tgsi_dump(templ->tokens, 0);
1039       debug_printf("usage masks:\n");
1040       for (attrib = 0; attrib < shader->info.base.num_inputs; ++attrib) {
1041          unsigned usage_mask = shader->info.base.input_usage_mask[attrib];
1042          debug_printf("  IN[%u].%s%s%s%s\n",
1043                       attrib,
1044                       usage_mask & TGSI_WRITEMASK_X ? "x" : "",
1045                       usage_mask & TGSI_WRITEMASK_Y ? "y" : "",
1046                       usage_mask & TGSI_WRITEMASK_Z ? "z" : "",
1047                       usage_mask & TGSI_WRITEMASK_W ? "w" : "");
1048       }
1049       debug_printf("\n");
1050    }
1051
1052    return shader;
1053 }
1054
1055
1056 static void
1057 llvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs)
1058 {
1059    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
1060
1061    if (llvmpipe->fs == fs)
1062       return;
1063
1064    draw_flush(llvmpipe->draw);
1065
1066    llvmpipe->fs = (struct lp_fragment_shader *) fs;
1067
1068    draw_bind_fragment_shader(llvmpipe->draw,
1069                              (llvmpipe->fs ? llvmpipe->fs->draw_data : NULL));
1070
1071    llvmpipe->dirty |= LP_NEW_FS;
1072 }
1073
1074 static void
1075 remove_shader_variant(struct llvmpipe_context *lp,
1076                       struct lp_fragment_shader_variant *variant)
1077 {
1078    struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
1079    unsigned i;
1080
1081    if (gallivm_debug & GALLIVM_DEBUG_IR) {
1082       debug_printf("llvmpipe: del fs #%u var #%u v created #%u v cached #%u v total cached #%u\n",
1083                     variant->shader->no, variant->no, variant->shader->variants_created,
1084                     variant->shader->variants_cached, lp->nr_fs_variants);
1085    }
1086    for (i = 0; i < Elements(variant->function); i++) {
1087       if (variant->function[i]) {
1088          if (variant->jit_function[i])
1089             LLVMFreeMachineCodeForFunction(screen->engine,
1090                                            variant->function[i]);
1091          LLVMDeleteFunction(variant->function[i]);
1092       }
1093    }
1094    remove_from_list(&variant->list_item_local);
1095    variant->shader->variants_cached--;
1096    remove_from_list(&variant->list_item_global);
1097    lp->nr_fs_variants--;
1098    FREE(variant);
1099 }
1100
1101 static void
1102 llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs)
1103 {
1104    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
1105    struct lp_fragment_shader *shader = fs;
1106    struct lp_fs_variant_list_item *li;
1107
1108    assert(fs != llvmpipe->fs);
1109    (void) llvmpipe;
1110
1111    /*
1112     * XXX: we need to flush the context until we have some sort of reference
1113     * counting in fragment shaders as they may still be binned
1114     * Flushing alone might not sufficient we need to wait on it too.
1115     */
1116
1117    llvmpipe_finish(pipe, __FUNCTION__);
1118
1119    li = first_elem(&shader->variants);
1120    while(!at_end(&shader->variants, li)) {
1121       struct lp_fs_variant_list_item *next = next_elem(li);
1122       remove_shader_variant(llvmpipe, li->base);
1123       li = next;
1124    }
1125
1126    draw_delete_fragment_shader(llvmpipe->draw, shader->draw_data);
1127
1128    assert(shader->variants_cached == 0);
1129    FREE((void *) shader->base.tokens);
1130    FREE(shader);
1131 }
1132
1133
1134
1135 static void
1136 llvmpipe_set_constant_buffer(struct pipe_context *pipe,
1137                              uint shader, uint index,
1138                              struct pipe_resource *constants)
1139 {
1140    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
1141    unsigned size = constants ? constants->width0 : 0;
1142    const void *data = constants ? llvmpipe_resource_data(constants) : NULL;
1143
1144    assert(shader < PIPE_SHADER_TYPES);
1145    assert(index < PIPE_MAX_CONSTANT_BUFFERS);
1146
1147    if(llvmpipe->constants[shader][index] == constants)
1148       return;
1149
1150    draw_flush(llvmpipe->draw);
1151
1152    /* note: reference counting */
1153    pipe_resource_reference(&llvmpipe->constants[shader][index], constants);
1154
1155    if(shader == PIPE_SHADER_VERTEX ||
1156       shader == PIPE_SHADER_GEOMETRY) {
1157       draw_set_mapped_constant_buffer(llvmpipe->draw, shader,
1158                                       index, data, size);
1159    }
1160
1161    llvmpipe->dirty |= LP_NEW_CONSTANTS;
1162 }
1163
1164
1165 /**
1166  * Return the blend factor equivalent to a destination alpha of one.
1167  */
1168 static INLINE unsigned
1169 force_dst_alpha_one(unsigned factor)
1170 {
1171    switch(factor) {
1172    case PIPE_BLENDFACTOR_DST_ALPHA:
1173       return PIPE_BLENDFACTOR_ONE;
1174    case PIPE_BLENDFACTOR_INV_DST_ALPHA:
1175       return PIPE_BLENDFACTOR_ZERO;
1176    case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
1177       return PIPE_BLENDFACTOR_ZERO;
1178    }
1179
1180    return factor;
1181 }
1182
1183
1184 /**
1185  * We need to generate several variants of the fragment pipeline to match
1186  * all the combinations of the contributing state atoms.
1187  *
1188  * TODO: there is actually no reason to tie this to context state -- the
1189  * generated code could be cached globally in the screen.
1190  */
1191 static void
1192 make_variant_key(struct llvmpipe_context *lp,
1193                  struct lp_fragment_shader *shader,
1194                  struct lp_fragment_shader_variant_key *key)
1195 {
1196    unsigned i;
1197
1198    memset(key, 0, shader->variant_key_size);
1199
1200    if (lp->framebuffer.zsbuf) {
1201       if (lp->depth_stencil->depth.enabled) {
1202          key->zsbuf_format = lp->framebuffer.zsbuf->format;
1203          memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth);
1204       }
1205       if (lp->depth_stencil->stencil[0].enabled) {
1206          key->zsbuf_format = lp->framebuffer.zsbuf->format;
1207          memcpy(&key->stencil, &lp->depth_stencil->stencil, sizeof key->stencil);
1208       }
1209    }
1210
1211    key->alpha.enabled = lp->depth_stencil->alpha.enabled;
1212    if(key->alpha.enabled)
1213       key->alpha.func = lp->depth_stencil->alpha.func;
1214    /* alpha.ref_value is passed in jit_context */
1215
1216    key->flatshade = lp->rasterizer->flatshade;
1217    if (lp->active_query_count) {
1218       key->occlusion_count = TRUE;
1219    }
1220
1221    if (lp->framebuffer.nr_cbufs) {
1222       memcpy(&key->blend, lp->blend, sizeof key->blend);
1223    }
1224
1225    key->nr_cbufs = lp->framebuffer.nr_cbufs;
1226    for (i = 0; i < lp->framebuffer.nr_cbufs; i++) {
1227       enum pipe_format format = lp->framebuffer.cbufs[i]->format;
1228       struct pipe_rt_blend_state *blend_rt = &key->blend.rt[i];
1229       const struct util_format_description *format_desc;
1230
1231       key->cbuf_format[i] = format;
1232
1233       format_desc = util_format_description(format);
1234       assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
1235              format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB);
1236
1237       blend_rt->colormask = lp->blend->rt[i].colormask;
1238
1239       /*
1240        * Mask out color channels not present in the color buffer.
1241        */
1242       blend_rt->colormask &= util_format_colormask(format_desc);
1243
1244       /*
1245        * Our swizzled render tiles always have an alpha channel, but the linear
1246        * render target format often does not, so force here the dst alpha to be
1247        * one.
1248        *
1249        * This is not a mere optimization. Wrong results will be produced if the
1250        * dst alpha is used, the dst format does not have alpha, and the previous
1251        * rendering was not flushed from the swizzled to linear buffer. For
1252        * example, NonPowTwo DCT.
1253        *
1254        * TODO: This should be generalized to all channels for better
1255        * performance, but only alpha causes correctness issues.
1256        *
1257        * Also, force rgb/alpha func/factors match, to make AoS blending easier.
1258        */
1259       if (format_desc->swizzle[3] > UTIL_FORMAT_SWIZZLE_W) {
1260          blend_rt->rgb_src_factor   = force_dst_alpha_one(blend_rt->rgb_src_factor);
1261          blend_rt->rgb_dst_factor   = force_dst_alpha_one(blend_rt->rgb_dst_factor);
1262          blend_rt->alpha_func       = blend_rt->rgb_func;
1263          blend_rt->alpha_src_factor = blend_rt->rgb_src_factor;
1264          blend_rt->alpha_dst_factor = blend_rt->rgb_dst_factor;
1265       }
1266    }
1267
1268    /* This value will be the same for all the variants of a given shader:
1269     */
1270    key->nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
1271
1272    for(i = 0; i < key->nr_samplers; ++i) {
1273       if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
1274          lp_sampler_static_state(&key->sampler[i],
1275                                  lp->fragment_sampler_views[i],
1276                                  lp->sampler[i]);
1277       }
1278    }
1279 }
1280
1281 /**
1282  * Update fragment state.  This is called just prior to drawing
1283  * something when some fragment-related state has changed.
1284  */
1285 void
1286 llvmpipe_update_fs(struct llvmpipe_context *lp)
1287 {
1288    struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
1289    struct lp_fragment_shader *shader = lp->fs;
1290    struct lp_fragment_shader_variant_key key;
1291    struct lp_fragment_shader_variant *variant = NULL;
1292    struct lp_fs_variant_list_item *li;
1293
1294    make_variant_key(lp, shader, &key);
1295
1296    li = first_elem(&shader->variants);
1297    while(!at_end(&shader->variants, li)) {
1298       if(memcmp(&li->base->key, &key, shader->variant_key_size) == 0) {
1299          variant = li->base;
1300          break;
1301       }
1302       li = next_elem(li);
1303    }
1304
1305    if (variant) {
1306       move_to_head(&lp->fs_variants_list, &variant->list_item_global);
1307    }
1308    else {
1309       int64_t t0, t1;
1310       int64_t dt;
1311       unsigned i;
1312       if (lp->nr_fs_variants >= LP_MAX_SHADER_VARIANTS) {
1313          struct pipe_context *pipe = &lp->pipe;
1314
1315          /*
1316           * XXX: we need to flush the context until we have some sort of reference
1317           * counting in fragment shaders as they may still be binned
1318           * Flushing alone might not be sufficient we need to wait on it too.
1319           */
1320          llvmpipe_finish(pipe, __FUNCTION__);
1321
1322          for (i = 0; i < LP_MAX_SHADER_VARIANTS / 4; i++) {
1323             struct lp_fs_variant_list_item *item = last_elem(&lp->fs_variants_list);
1324             remove_shader_variant(lp, item->base);
1325          }
1326       }
1327       t0 = os_time_get();
1328
1329       variant = generate_variant(screen, shader, &key);
1330
1331       t1 = os_time_get();
1332       dt = t1 - t0;
1333       LP_COUNT_ADD(llvm_compile_time, dt);
1334       LP_COUNT_ADD(nr_llvm_compiles, 2);  /* emit vs. omit in/out test */
1335
1336       if (variant) {
1337          insert_at_head(&shader->variants, &variant->list_item_local);
1338          insert_at_head(&lp->fs_variants_list, &variant->list_item_global);
1339          lp->nr_fs_variants++;
1340          shader->variants_cached++;
1341       }
1342    }
1343
1344    lp_setup_set_fs_variant(lp->setup, variant);
1345 }
1346
1347
1348
1349
1350
1351
1352
1353 void
1354 llvmpipe_init_fs_funcs(struct llvmpipe_context *llvmpipe)
1355 {
1356    llvmpipe->pipe.create_fs_state = llvmpipe_create_fs_state;
1357    llvmpipe->pipe.bind_fs_state   = llvmpipe_bind_fs_state;
1358    llvmpipe->pipe.delete_fs_state = llvmpipe_delete_fs_state;
1359
1360    llvmpipe->pipe.set_constant_buffer = llvmpipe_set_constant_buffer;
1361 }