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, interp->pos, interp->inputs,
 338                      outputs, sampler, &shader->info.base);
 339
 340
 341    /* Alpha test */
 342    if (key->alpha.enabled) {
 343       int color0 = find_output_by_semantic(&shader->info.base,
 344                                            TGSI_SEMANTIC_COLOR,
 345                                            0);
 346
 347       if (color0 != -1 && outputs[color0][3]) {
 348          LLVMValueRef alpha = LLVMBuildLoad(builder, outputs[color0][3], "alpha");
 349          LLVMValueRef alpha_ref_value;
 350
 351          alpha_ref_value = lp_jit_context_alpha_ref_value(builder, context_ptr);
 352          alpha_ref_value = lp_build_broadcast(builder, vec_type, alpha_ref_value);
 353
 354          lp_build_alpha_test(builder, key->alpha.func, type,
 355                              &mask, alpha, alpha_ref_value,
 356                              (depth_mode & LATE_DEPTH_TEST) != 0);
 357       }
 358    }
 359
 360    /* Late Z test */
 361    if (depth_mode & LATE_DEPTH_TEST) {
 362       int pos0 = find_output_by_semantic(&shader->info.base,
 363                                          TGSI_SEMANTIC_POSITION,
 364                                          0);
 365
 366       if (pos0 != -1 && outputs[pos0][2]) {
 367          z = LLVMBuildLoad(builder, outputs[pos0][2], "output.z");
 368       }
 369
 370       lp_build_depth_stencil_test(builder,
 371                                   &key->depth,
 372                                   key->stencil,
 373                                   type,
 374                                   zs_format_desc,
 375                                   &mask,
 376                                   stencil_refs,
 377                                   z,
 378                                   depth_ptr, facing,
 379                                   &zs_value,
 380                                   !simple_shader);
 381       /* Late Z write */
 382       if (depth_mode & LATE_DEPTH_WRITE) {
 383          lp_build_depth_write(builder, zs_format_desc, depth_ptr, zs_value);
 384       }
 385    }
 386    else if ((depth_mode & EARLY_DEPTH_TEST) &&
 387             (depth_mode & LATE_DEPTH_WRITE))
 388    {
 389       /* Need to apply a reduced mask to the depth write.  Reload the
 390        * depth value, update from zs_value with the new mask value and
 391        * write that out.
 392        */
 393       lp_build_deferred_depth_write(builder,
 394                                     type,
 395                                     zs_format_desc,
 396                                     &mask,
 397                                     depth_ptr,
 398                                     zs_value);
 399    }
 400
 401
 402    /* Color write  */
 403    for (attrib = 0; attrib < shader->info.base.num_outputs; ++attrib)
 404    {
 405       if (shader->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_COLOR &&
 406           shader->info.base.output_semantic_index[attrib] < key->nr_cbufs)
 407       {
 408          unsigned cbuf = shader->info.base.output_semantic_index[attrib];
 409          for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 410             if(outputs[attrib][chan]) {
 411                /* XXX: just initialize outputs to point at colors[] and
 412                 * skip this.
 413                 */
 414                LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
 415                lp_build_name(out, "color%u.%u.%c", i, attrib, "rgba"[chan]);
 416                LLVMBuildStore(builder, out, color[cbuf][chan]);
 417             }
 418          }
 419       }
 420    }
 421
 422    if (counter)
 423       lp_build_occlusion_count(builder, type,
 424                                lp_build_mask_value(&mask), counter);
 425
 426    *pmask = lp_build_mask_end(&mask);
 427 }
 428
 429
 430 /**
 431  * Generate color blending and color output.
 432  * \param rt  the render target index (to index blend, colormask state)
 433  * \param type  the pixel color type
 434  * \param context_ptr  pointer to the runtime JIT context
 435  * \param mask  execution mask (active fragment/pixel mask)
 436  * \param src  colors from the fragment shader
 437  * \param dst_ptr  the destination color buffer pointer
 438  */
 439 static void
 440 generate_blend(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, builder, type);
 460
 461    lp_build_mask_begin(&mask_ctx, builder, type, mask);
 462    if (do_branch)
 463       lp_build_mask_check(&mask_ctx);
 464
 465    vec_type = lp_build_vec_type(type);
 466
 467    const_ptr = lp_jit_context_blend_color(builder, 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 = LLVMConstInt(LLVMInt32Type(), chan, 0);
 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(builder, 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 = LLVMConstInt(LLVMInt32Type(), chan, 0);
 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_screen *screen,
 507                   struct lp_fragment_shader *shader,
 508                   struct lp_fragment_shader_variant *variant,
 509                   unsigned partial_mask)
 510 {
 511    const struct lp_fragment_shader_variant_key *key = &variant->key;
 512    struct lp_shader_input inputs[PIPE_MAX_SHADER_INPUTS];
 513    char func_name[256];
 514    struct lp_type fs_type;
 515    struct lp_type blend_type;
 516    LLVMTypeRef fs_elem_type;
 517    LLVMTypeRef fs_int_vec_type;
 518    LLVMTypeRef blend_vec_type;
 519    LLVMTypeRef arg_types[11];
 520    LLVMTypeRef func_type;
 521    LLVMValueRef context_ptr;
 522    LLVMValueRef x;
 523    LLVMValueRef y;
 524    LLVMValueRef a0_ptr;
 525    LLVMValueRef dadx_ptr;
 526    LLVMValueRef dady_ptr;
 527    LLVMValueRef color_ptr_ptr;
 528    LLVMValueRef depth_ptr;
 529    LLVMValueRef mask_input;
 530    LLVMValueRef counter = NULL;
 531    LLVMBasicBlockRef block;
 532    LLVMBuilderRef builder;
 533    struct lp_build_sampler_soa *sampler;
 534    struct lp_build_interp_soa_context interp;
 535    LLVMValueRef fs_mask[LP_MAX_VECTOR_LENGTH];
 536    LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS][LP_MAX_VECTOR_LENGTH];
 537    LLVMValueRef blend_mask;
 538    LLVMValueRef function;
 539    LLVMValueRef facing;
 540    const struct util_format_description *zs_format_desc;
 541    unsigned num_fs;
 542    unsigned i;
 543    unsigned chan;
 544    unsigned cbuf;
 545
 546    /* Adjust color input interpolation according to flatshade state:
 547     */
 548    memcpy(inputs, shader->inputs, shader->info.base.num_inputs * sizeof inputs[0]);
 549    for (i = 0; i < shader->info.base.num_inputs; i++) {
 550       if (inputs[i].interp == LP_INTERP_COLOR) {
 551          if (key->flatshade)
 552             inputs[i].interp = LP_INTERP_CONSTANT;
 553          else
 554             inputs[i].interp = LP_INTERP_LINEAR;
 555       }
 556    }
 557
 558
 559    /* TODO: actually pick these based on the fs and color buffer
 560     * characteristics. */
 561
 562    memset(&fs_type, 0, sizeof fs_type);
 563    fs_type.floating = TRUE; /* floating point values */
 564    fs_type.sign = TRUE;     /* values are signed */
 565    fs_type.norm = FALSE;    /* values are not limited to [0,1] or [-1,1] */
 566    fs_type.width = 32;      /* 32-bit float */
 567    fs_type.length = 4;      /* 4 elements per vector */
 568    num_fs = 4;              /* number of quads per block */
 569
 570    memset(&blend_type, 0, sizeof blend_type);
 571    blend_type.floating = FALSE; /* values are integers */
 572    blend_type.sign = FALSE;     /* values are unsigned */
 573    blend_type.norm = TRUE;      /* values are in [0,1] or [-1,1] */
 574    blend_type.width = 8;        /* 8-bit ubyte values */
 575    blend_type.length = 16;      /* 16 elements per vector */
 576
 577    /*
 578     * Generate the function prototype. Any change here must be reflected in
 579     * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
 580     */
 581
 582    fs_elem_type = lp_build_elem_type(fs_type);
 583    fs_int_vec_type = lp_build_int_vec_type(fs_type);
 584
 585    blend_vec_type = lp_build_vec_type(blend_type);
 586
 587    util_snprintf(func_name, sizeof(func_name), "fs%u_variant%u_%s",
 588                  shader->no, variant->no, partial_mask ? "partial" : "whole");
 589
 590    arg_types[0] = screen->context_ptr_type;            /* context */
 591    arg_types[1] = LLVMInt32Type();                     /* x */
 592    arg_types[2] = LLVMInt32Type();                     /* y */
 593    arg_types[3] = LLVMInt32Type();                     /* facing */
 594    arg_types[4] = LLVMPointerType(fs_elem_type, 0);    /* a0 */
 595    arg_types[5] = LLVMPointerType(fs_elem_type, 0);    /* dadx */
 596    arg_types[6] = LLVMPointerType(fs_elem_type, 0);    /* dady */
 597    arg_types[7] = LLVMPointerType(LLVMPointerType(blend_vec_type, 0), 0);  /* color */
 598    arg_types[8] = LLVMPointerType(LLVMInt8Type(), 0);  /* depth */
 599    arg_types[9] = LLVMInt32Type();                     /* mask_input */
 600    arg_types[10] = LLVMPointerType(LLVMInt32Type(), 0);/* counter */
 601
 602    func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
 603
 604    function = LLVMAddFunction(screen->module, func_name, func_type);
 605    LLVMSetFunctionCallConv(function, LLVMCCallConv);
 606
 607    variant->function[partial_mask] = function;
 608
 609    /* XXX: need to propagate noalias down into color param now we are
 610     * passing a pointer-to-pointer?
 611     */
 612    for(i = 0; i < Elements(arg_types); ++i)
 613       if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
 614          LLVMAddAttribute(LLVMGetParam(function, i), LLVMNoAliasAttribute);
 615
 616    context_ptr  = LLVMGetParam(function, 0);
 617    x            = LLVMGetParam(function, 1);
 618    y            = LLVMGetParam(function, 2);
 619    facing       = LLVMGetParam(function, 3);
 620    a0_ptr       = LLVMGetParam(function, 4);
 621    dadx_ptr     = LLVMGetParam(function, 5);
 622    dady_ptr     = LLVMGetParam(function, 6);
 623    color_ptr_ptr = LLVMGetParam(function, 7);
 624    depth_ptr    = LLVMGetParam(function, 8);
 625    mask_input   = LLVMGetParam(function, 9);
 626
 627    lp_build_name(context_ptr, "context");
 628    lp_build_name(x, "x");
 629    lp_build_name(y, "y");
 630    lp_build_name(a0_ptr, "a0");
 631    lp_build_name(dadx_ptr, "dadx");
 632    lp_build_name(dady_ptr, "dady");
 633    lp_build_name(color_ptr_ptr, "color_ptr_ptr");
 634    lp_build_name(depth_ptr, "depth");
 635    lp_build_name(mask_input, "mask_input");
 636
 637    if (key->occlusion_count) {
 638       counter = LLVMGetParam(function, 10);
 639       lp_build_name(counter, "counter");
 640    }
 641
 642    /*
 643     * Function body
 644     */
 645
 646    block = LLVMAppendBasicBlock(function, "entry");
 647    builder = LLVMCreateBuilder();
 648    LLVMPositionBuilderAtEnd(builder, block);
 649
 650    /*
 651     * The shader input interpolation info is not explicitely baked in the
 652     * shader key, but everything it derives from (TGSI, and flatshade) is
 653     * already included in the shader key.
 654     */
 655    lp_build_interp_soa_init(&interp,
 656                             shader->info.base.num_inputs,
 657                             inputs,
 658                             builder, fs_type,
 659                             a0_ptr, dadx_ptr, dady_ptr,
 660                             x, y);
 661
 662    /* code generated texture sampling */
 663    sampler = lp_llvm_sampler_soa_create(key->sampler, context_ptr);
 664
 665    /* loop over quads in the block */
 666    zs_format_desc = util_format_description(key->zsbuf_format);
 667
 668    for(i = 0; i < num_fs; ++i) {
 669       LLVMValueRef depth_offset = LLVMConstInt(LLVMInt32Type(),
 670                                                i*fs_type.length*zs_format_desc->block.bits/8,
 671                                                0);
 672       LLVMValueRef out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS];
 673       LLVMValueRef depth_ptr_i;
 674
 675       depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &depth_offset, 1, "");
 676
 677       generate_fs(shader, key,
 678                   builder,
 679                   fs_type,
 680                   context_ptr,
 681                   i,
 682                   &interp,
 683                   sampler,
 684                   &fs_mask[i], /* output */
 685                   out_color,
 686                   depth_ptr_i,
 687                   facing,
 688                   partial_mask,
 689                   mask_input,
 690                   counter);
 691
 692       for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++)
 693          for(chan = 0; chan < NUM_CHANNELS; ++chan)
 694             fs_out_color[cbuf][chan][i] = out_color[cbuf][chan];
 695    }
 696
 697    sampler->destroy(sampler);
 698
 699    /* Loop over color outputs / color buffers to do blending.
 700     */
 701    for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
 702       LLVMValueRef color_ptr;
 703       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), cbuf, 0);
 704       LLVMValueRef blend_in_color[NUM_CHANNELS];
 705       unsigned rt;
 706
 707       /*
 708        * Convert the fs's output color and mask to fit to the blending type.
 709        */
 710       for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 711          LLVMValueRef fs_color_vals[LP_MAX_VECTOR_LENGTH];
 712
 713          for (i = 0; i < num_fs; i++) {
 714             fs_color_vals[i] =
 715                LLVMBuildLoad(builder, fs_out_color[cbuf][chan][i], "fs_color_vals");
 716          }
 717
 718          lp_build_conv(builder, fs_type, blend_type,
 719                        fs_color_vals,
 720                        num_fs,
 721                        &blend_in_color[chan], 1);
 722
 723          lp_build_name(blend_in_color[chan], "color%d.%c", cbuf, "rgba"[chan]);
 724       }
 725
 726       if (partial_mask || !variant->opaque) {
 727          lp_build_conv_mask(builder, fs_type, blend_type,
 728                             fs_mask, num_fs,
 729                             &blend_mask, 1);
 730       } else {
 731          blend_mask = lp_build_const_int_vec(blend_type, ~0);
 732       }
 733
 734       color_ptr = LLVMBuildLoad(builder,
 735                                 LLVMBuildGEP(builder, color_ptr_ptr, &index, 1, ""),
 736                                 "");
 737       lp_build_name(color_ptr, "color_ptr%d", cbuf);
 738
 739       /* which blend/colormask state to use */
 740       rt = key->blend.independent_blend_enable ? cbuf : 0;
 741
 742       /*
 743        * Blending.
 744        */
 745       {
 746          /* Could the 4x4 have been killed?
 747           */
 748          boolean do_branch = ((key->depth.enabled || key->stencil[0].enabled) &&
 749                               !key->alpha.enabled &&
 750                               !shader->info.base.uses_kill);
 751
 752          generate_blend(&key->blend,
 753                         rt,
 754                         builder,
 755                         blend_type,
 756                         context_ptr,
 757                         blend_mask,
 758                         blend_in_color,
 759                         color_ptr,
 760                         do_branch);
 761       }
 762    }
 763
 764    LLVMBuildRetVoid(builder);
 765
 766    LLVMDisposeBuilder(builder);
 767
 768
 769    /* Verify the LLVM IR.  If invalid, dump and abort */
 770 #ifdef DEBUG
 771    if(LLVMVerifyFunction(function, LLVMPrintMessageAction)) {
 772       if (1)
 773          lp_debug_dump_value(function);
 774       abort();
 775    }
 776 #endif
 777
 778    /* Apply optimizations to LLVM IR */
 779    LLVMRunFunctionPassManager(screen->pass, function);
 780
 781    if ((gallivm_debug & GALLIVM_DEBUG_IR) || (LP_DEBUG & DEBUG_FS)) {
 782       /* Print the LLVM IR to stderr */
 783       lp_debug_dump_value(function);
 784       debug_printf("\n");
 785    }
 786
 787    /* Dump byte code to a file */
 788    if (0) {
 789       LLVMWriteBitcodeToFile(lp_build_module, "llvmpipe.bc");
 790    }
 791
 792    /*
 793     * Translate the LLVM IR into machine code.
 794     */
 795    {
 796       void *f = LLVMGetPointerToGlobal(screen->engine, function);
 797
 798       variant->jit_function[partial_mask] = (lp_jit_frag_func)pointer_to_func(f);
 799
 800       if ((gallivm_debug & GALLIVM_DEBUG_ASM) || (LP_DEBUG & DEBUG_FS)) {
 801          lp_disassemble(f);
 802       }
 803       lp_func_delete_body(function);
 804    }
 805 }
 806
 807
 808 static void
 809 dump_fs_variant_key(const struct lp_fragment_shader_variant_key *key)
 810 {
 811    unsigned i;
 812
 813    debug_printf("fs variant %p:\n", (void *) key);
 814
 815    if (key->flatshade) {
 816       debug_printf("flatshade = 1\n");
 817    }
 818    for (i = 0; i < key->nr_cbufs; ++i) {
 819       debug_printf("cbuf_format[%u] = %s\n", i, util_format_name(key->cbuf_format[i]));
 820    }
 821    if (key->depth.enabled) {
 822       debug_printf("depth.format = %s\n", util_format_name(key->zsbuf_format));
 823       debug_printf("depth.func = %s\n", util_dump_func(key->depth.func, TRUE));
 824       debug_printf("depth.writemask = %u\n", key->depth.writemask);
 825    }
 826
 827    for (i = 0; i < 2; ++i) {
 828       if (key->stencil[i].enabled) {
 829          debug_printf("stencil[%u].func = %s\n", i, util_dump_func(key->stencil[i].func, TRUE));
 830          debug_printf("stencil[%u].fail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].fail_op, TRUE));
 831          debug_printf("stencil[%u].zpass_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zpass_op, TRUE));
 832          debug_printf("stencil[%u].zfail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zfail_op, TRUE));
 833          debug_printf("stencil[%u].valuemask = 0x%x\n", i, key->stencil[i].valuemask);
 834          debug_printf("stencil[%u].writemask = 0x%x\n", i, key->stencil[i].writemask);
 835       }
 836    }
 837
 838    if (key->alpha.enabled) {
 839       debug_printf("alpha.func = %s\n", util_dump_func(key->alpha.func, TRUE));
 840    }
 841
 842    if (key->occlusion_count) {
 843       debug_printf("occlusion_count = 1\n");
 844    }
 845
 846    if (key->blend.logicop_enable) {
 847       debug_printf("blend.logicop_func = %s\n", util_dump_logicop(key->blend.logicop_func, TRUE));
 848    }
 849    else if (key->blend.rt[0].blend_enable) {
 850       debug_printf("blend.rgb_func = %s\n",   util_dump_blend_func  (key->blend.rt[0].rgb_func, TRUE));
 851       debug_printf("blend.rgb_src_factor = %s\n",   util_dump_blend_factor(key->blend.rt[0].rgb_src_factor, TRUE));
 852       debug_printf("blend.rgb_dst_factor = %s\n",   util_dump_blend_factor(key->blend.rt[0].rgb_dst_factor, TRUE));
 853       debug_printf("blend.alpha_func = %s\n",       util_dump_blend_func  (key->blend.rt[0].alpha_func, TRUE));
 854       debug_printf("blend.alpha_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_src_factor, TRUE));
 855       debug_printf("blend.alpha_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_dst_factor, TRUE));
 856    }
 857    debug_printf("blend.colormask = 0x%x\n", key->blend.rt[0].colormask);
 858    for (i = 0; i < key->nr_samplers; ++i) {
 859       debug_printf("sampler[%u] = \n", i);
 860       debug_printf("  .format = %s\n",
 861                    util_format_name(key->sampler[i].format));
 862       debug_printf("  .target = %s\n",
 863                    util_dump_tex_target(key->sampler[i].target, TRUE));
 864       debug_printf("  .pot = %u %u %u\n",
 865                    key->sampler[i].pot_width,
 866                    key->sampler[i].pot_height,
 867                    key->sampler[i].pot_depth);
 868       debug_printf("  .wrap = %s %s %s\n",
 869                    util_dump_tex_wrap(key->sampler[i].wrap_s, TRUE),
 870                    util_dump_tex_wrap(key->sampler[i].wrap_t, TRUE),
 871                    util_dump_tex_wrap(key->sampler[i].wrap_r, TRUE));
 872       debug_printf("  .min_img_filter = %s\n",
 873                    util_dump_tex_filter(key->sampler[i].min_img_filter, TRUE));
 874       debug_printf("  .min_mip_filter = %s\n",
 875                    util_dump_tex_mipfilter(key->sampler[i].min_mip_filter, TRUE));
 876       debug_printf("  .mag_img_filter = %s\n",
 877                    util_dump_tex_filter(key->sampler[i].mag_img_filter, TRUE));
 878       if (key->sampler[i].compare_mode != PIPE_TEX_COMPARE_NONE)
 879          debug_printf("  .compare_func = %s\n", util_dump_func(key->sampler[i].compare_func, TRUE));
 880       debug_printf("  .normalized_coords = %u\n", key->sampler[i].normalized_coords);
 881       debug_printf("  .min_max_lod_equal = %u\n", key->sampler[i].min_max_lod_equal);
 882       debug_printf("  .lod_bias_non_zero = %u\n", key->sampler[i].lod_bias_non_zero);
 883       debug_printf("  .apply_min_lod = %u\n", key->sampler[i].apply_min_lod);
 884       debug_printf("  .apply_max_lod = %u\n", key->sampler[i].apply_max_lod);
 885    }
 886 }
 887
 888
 889 void
 890 lp_debug_fs_variant(const struct lp_fragment_shader_variant *variant)
 891 {
 892    debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
 893                 variant->shader->no, variant->no);
 894    tgsi_dump(variant->shader->base.tokens, 0);
 895    dump_fs_variant_key(&variant->key);
 896    debug_printf("variant->opaque = %u\n", variant->opaque);
 897    debug_printf("\n");
 898 }
 899
 900 static struct lp_fragment_shader_variant *
 901 generate_variant(struct llvmpipe_screen *screen,
 902                  struct lp_fragment_shader *shader,
 903                  const struct lp_fragment_shader_variant_key *key)
 904 {
 905    struct lp_fragment_shader_variant *variant;
 906    boolean fullcolormask;
 907
 908    variant = CALLOC_STRUCT(lp_fragment_shader_variant);
 909    if(!variant)
 910       return NULL;
 911
 912    variant->shader = shader;
 913    variant->list_item_global.base = variant;
 914    variant->list_item_local.base = variant;
 915    variant->no = shader->variants_created++;
 916
 917    memcpy(&variant->key, key, shader->variant_key_size);
 918
 919    /*
 920     * Determine whether we are touching all channels in the color buffer.
 921     */
 922    fullcolormask = FALSE;
 923    if (key->nr_cbufs == 1) {
 924       const struct util_format_description *format_desc;
 925       format_desc = util_format_description(key->cbuf_format[0]);
 926       if ((~key->blend.rt[0].colormask &
 927            util_format_colormask(format_desc)) == 0) {
 928          fullcolormask = TRUE;
 929       }
 930    }
 931
 932    variant->opaque =
 933          !key->blend.logicop_enable &&
 934          !key->blend.rt[0].blend_enable &&
 935          fullcolormask &&
 936          !key->stencil[0].enabled &&
 937          !key->alpha.enabled &&
 938          !key->depth.enabled &&
 939          !shader->info.base.uses_kill
 940          ? TRUE : FALSE;
 941
 942
 943    if ((LP_DEBUG & DEBUG_FS) || (gallivm_debug & GALLIVM_DEBUG_IR)) {
 944       lp_debug_fs_variant(variant);
 945    }
 946
 947    generate_fragment(screen, shader, variant, RAST_EDGE_TEST);
 948
 949    if (variant->opaque) {
 950       /* Specialized shader, which doesn't need to read the color buffer. */
 951       generate_fragment(screen, shader, variant, RAST_WHOLE);
 952    } else {
 953       variant->jit_function[RAST_WHOLE] = variant->jit_function[RAST_EDGE_TEST];
 954    }
 955
 956    return variant;
 957 }
 958
 959
 960 static void *
 961 llvmpipe_create_fs_state(struct pipe_context *pipe,
 962                          const struct pipe_shader_state *templ)
 963 {
 964    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
 965    struct lp_fragment_shader *shader;
 966    int nr_samplers;
 967    int i;
 968
 969    shader = CALLOC_STRUCT(lp_fragment_shader);
 970    if (!shader)
 971       return NULL;
 972
 973    shader->no = fs_no++;
 974    make_empty_list(&shader->variants);
 975
 976    /* get/save the summary info for this shader */
 977    lp_build_tgsi_info(templ->tokens, &shader->info);
 978
 979    /* we need to keep a local copy of the tokens */
 980    shader->base.tokens = tgsi_dup_tokens(templ->tokens);
 981
 982    shader->draw_data = draw_create_fragment_shader(llvmpipe->draw, templ);
 983    if (shader->draw_data == NULL) {
 984       FREE((void *) shader->base.tokens);
 985       FREE(shader);
 986       return NULL;
 987    }
 988
 989    nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
 990
 991    shader->variant_key_size = Offset(struct lp_fragment_shader_variant_key,
 992                                      sampler[nr_samplers]);
 993
 994    for (i = 0; i < shader->info.base.num_inputs; i++) {
 995       shader->inputs[i].usage_mask = shader->info.base.input_usage_mask[i];
 996
 997       switch (shader->info.base.input_interpolate[i]) {
 998       case TGSI_INTERPOLATE_CONSTANT:
 999          shader->inputs[i].interp = LP_INTERP_CONSTANT;
1000          break;
1001       case TGSI_INTERPOLATE_LINEAR:
1002          shader->inputs[i].interp = LP_INTERP_LINEAR;
1003          break;
1004       case TGSI_INTERPOLATE_PERSPECTIVE:
1005          shader->inputs[i].interp = LP_INTERP_PERSPECTIVE;
1006          break;
1007       default:
1008          assert(0);
1009          break;
1010       }
1011
1012       switch (shader->info.base.input_semantic_name[i]) {
1013       case TGSI_SEMANTIC_COLOR:
1014          /* Colors may be either linearly or constant interpolated in
1015           * the fragment shader, but that information isn't available
1016           * here.  Mark color inputs and fix them up later.
1017           */
1018          shader->inputs[i].interp = LP_INTERP_COLOR;
1019          break;
1020       case TGSI_SEMANTIC_FACE:
1021          shader->inputs[i].interp = LP_INTERP_FACING;
1022          break;
1023       case TGSI_SEMANTIC_POSITION:
1024          /* Position was already emitted above
1025           */
1026          shader->inputs[i].interp = LP_INTERP_POSITION;
1027          shader->inputs[i].src_index = 0;
1028          continue;
1029       }
1030
1031       shader->inputs[i].src_index = i+1;
1032    }
1033
1034    if (LP_DEBUG & DEBUG_TGSI) {
1035       unsigned attrib;
1036       debug_printf("llvmpipe: Create fragment shader #%u %p:\n", shader->no, (void *) shader);
1037       tgsi_dump(templ->tokens, 0);
1038       debug_printf("usage masks:\n");
1039       for (attrib = 0; attrib < shader->info.base.num_inputs; ++attrib) {
1040          unsigned usage_mask = shader->info.base.input_usage_mask[attrib];
1041          debug_printf("  IN[%u].%s%s%s%s\n",
1042                       attrib,
1043                       usage_mask & TGSI_WRITEMASK_X ? "x" : "",
1044                       usage_mask & TGSI_WRITEMASK_Y ? "y" : "",
1045                       usage_mask & TGSI_WRITEMASK_Z ? "z" : "",
1046                       usage_mask & TGSI_WRITEMASK_W ? "w" : "");
1047       }
1048       debug_printf("\n");
1049    }
1050
1051    return shader;
1052 }
1053
1054
1055 static void
1056 llvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs)
1057 {
1058    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
1059
1060    if (llvmpipe->fs == fs)
1061       return;
1062
1063    draw_flush(llvmpipe->draw);
1064
1065    llvmpipe->fs = (struct lp_fragment_shader *) fs;
1066
1067    draw_bind_fragment_shader(llvmpipe->draw,
1068                              (llvmpipe->fs ? llvmpipe->fs->draw_data : NULL));
1069
1070    llvmpipe->dirty |= LP_NEW_FS;
1071 }
1072
1073 static void
1074 remove_shader_variant(struct llvmpipe_context *lp,
1075                       struct lp_fragment_shader_variant *variant)
1076 {
1077    struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
1078    unsigned i;
1079
1080    if (gallivm_debug & GALLIVM_DEBUG_IR) {
1081       debug_printf("llvmpipe: del fs #%u var #%u v created #%u v cached #%u v total cached #%u\n",
1082                     variant->shader->no, variant->no, variant->shader->variants_created,
1083                     variant->shader->variants_cached, lp->nr_fs_variants);
1084    }
1085    for (i = 0; i < Elements(variant->function); i++) {
1086       if (variant->function[i]) {
1087          if (variant->jit_function[i])
1088             LLVMFreeMachineCodeForFunction(screen->engine,
1089                                            variant->function[i]);
1090          LLVMDeleteFunction(variant->function[i]);
1091       }
1092    }
1093    remove_from_list(&variant->list_item_local);
1094    variant->shader->variants_cached--;
1095    remove_from_list(&variant->list_item_global);
1096    lp->nr_fs_variants--;
1097    FREE(variant);
1098 }
1099
1100 static void
1101 llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs)
1102 {
1103    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
1104    struct lp_fragment_shader *shader = fs;
1105    struct lp_fs_variant_list_item *li;
1106
1107    assert(fs != llvmpipe->fs);
1108    (void) llvmpipe;
1109
1110    /*
1111     * XXX: we need to flush the context until we have some sort of reference
1112     * counting in fragment shaders as they may still be binned
1113     * Flushing alone might not sufficient we need to wait on it too.
1114     */
1115
1116    llvmpipe_finish(pipe, __FUNCTION__);
1117
1118    li = first_elem(&shader->variants);
1119    while(!at_end(&shader->variants, li)) {
1120       struct lp_fs_variant_list_item *next = next_elem(li);
1121       remove_shader_variant(llvmpipe, li->base);
1122       li = next;
1123    }
1124
1125    draw_delete_fragment_shader(llvmpipe->draw, shader->draw_data);
1126
1127    assert(shader->variants_cached == 0);
1128    FREE((void *) shader->base.tokens);
1129    FREE(shader);
1130 }
1131
1132
1133
1134 static void
1135 llvmpipe_set_constant_buffer(struct pipe_context *pipe,
1136                              uint shader, uint index,
1137                              struct pipe_resource *constants)
1138 {
1139    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
1140    unsigned size = constants ? constants->width0 : 0;
1141    const void *data = constants ? llvmpipe_resource_data(constants) : NULL;
1142
1143    assert(shader < PIPE_SHADER_TYPES);
1144    assert(index < PIPE_MAX_CONSTANT_BUFFERS);
1145
1146    if(llvmpipe->constants[shader][index] == constants)
1147       return;
1148
1149    draw_flush(llvmpipe->draw);
1150
1151    /* note: reference counting */
1152    pipe_resource_reference(&llvmpipe->constants[shader][index], constants);
1153
1154    if(shader == PIPE_SHADER_VERTEX ||
1155       shader == PIPE_SHADER_GEOMETRY) {
1156       draw_set_mapped_constant_buffer(llvmpipe->draw, shader,
1157                                       index, data, size);
1158    }
1159
1160    llvmpipe->dirty |= LP_NEW_CONSTANTS;
1161 }
1162
1163
1164 /**
1165  * Return the blend factor equivalent to a destination alpha of one.
1166  */
1167 static INLINE unsigned
1168 force_dst_alpha_one(unsigned factor)
1169 {
1170    switch(factor) {
1171    case PIPE_BLENDFACTOR_DST_ALPHA:
1172       return PIPE_BLENDFACTOR_ONE;
1173    case PIPE_BLENDFACTOR_INV_DST_ALPHA:
1174       return PIPE_BLENDFACTOR_ZERO;
1175    case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
1176       return PIPE_BLENDFACTOR_ZERO;
1177    }
1178
1179    return factor;
1180 }
1181
1182
1183 /**
1184  * We need to generate several variants of the fragment pipeline to match
1185  * all the combinations of the contributing state atoms.
1186  *
1187  * TODO: there is actually no reason to tie this to context state -- the
1188  * generated code could be cached globally in the screen.
1189  */
1190 static void
1191 make_variant_key(struct llvmpipe_context *lp,
1192                  struct lp_fragment_shader *shader,
1193                  struct lp_fragment_shader_variant_key *key)
1194 {
1195    unsigned i;
1196
1197    memset(key, 0, shader->variant_key_size);
1198
1199    if (lp->framebuffer.zsbuf) {
1200       if (lp->depth_stencil->depth.enabled) {
1201          key->zsbuf_format = lp->framebuffer.zsbuf->format;
1202          memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth);
1203       }
1204       if (lp->depth_stencil->stencil[0].enabled) {
1205          key->zsbuf_format = lp->framebuffer.zsbuf->format;
1206          memcpy(&key->stencil, &lp->depth_stencil->stencil, sizeof key->stencil);
1207       }
1208    }
1209
1210    key->alpha.enabled = lp->depth_stencil->alpha.enabled;
1211    if(key->alpha.enabled)
1212       key->alpha.func = lp->depth_stencil->alpha.func;
1213    /* alpha.ref_value is passed in jit_context */
1214
1215    key->flatshade = lp->rasterizer->flatshade;
1216    if (lp->active_query_count) {
1217       key->occlusion_count = TRUE;
1218    }
1219
1220    if (lp->framebuffer.nr_cbufs) {
1221       memcpy(&key->blend, lp->blend, sizeof key->blend);
1222    }
1223
1224    key->nr_cbufs = lp->framebuffer.nr_cbufs;
1225    for (i = 0; i < lp->framebuffer.nr_cbufs; i++) {
1226       enum pipe_format format = lp->framebuffer.cbufs[i]->format;
1227       struct pipe_rt_blend_state *blend_rt = &key->blend.rt[i];
1228       const struct util_format_description *format_desc;
1229
1230       key->cbuf_format[i] = format;
1231
1232       format_desc = util_format_description(format);
1233       assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
1234              format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB);
1235
1236       blend_rt->colormask = lp->blend->rt[i].colormask;
1237
1238       /*
1239        * Mask out color channels not present in the color buffer.
1240        */
1241       blend_rt->colormask &= util_format_colormask(format_desc);
1242
1243       /*
1244        * Our swizzled render tiles always have an alpha channel, but the linear
1245        * render target format often does not, so force here the dst alpha to be
1246        * one.
1247        *
1248        * This is not a mere optimization. Wrong results will be produced if the
1249        * dst alpha is used, the dst format does not have alpha, and the previous
1250        * rendering was not flushed from the swizzled to linear buffer. For
1251        * example, NonPowTwo DCT.
1252        *
1253        * TODO: This should be generalized to all channels for better
1254        * performance, but only alpha causes correctness issues.
1255        *
1256        * Also, force rgb/alpha func/factors match, to make AoS blending easier.
1257        */
1258       if (format_desc->swizzle[3] > UTIL_FORMAT_SWIZZLE_W) {
1259          blend_rt->rgb_src_factor   = force_dst_alpha_one(blend_rt->rgb_src_factor);
1260          blend_rt->rgb_dst_factor   = force_dst_alpha_one(blend_rt->rgb_dst_factor);
1261          blend_rt->alpha_func       = blend_rt->rgb_func;
1262          blend_rt->alpha_src_factor = blend_rt->rgb_src_factor;
1263          blend_rt->alpha_dst_factor = blend_rt->rgb_dst_factor;
1264       }
1265    }
1266
1267    /* This value will be the same for all the variants of a given shader:
1268     */
1269    key->nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
1270
1271    for(i = 0; i < key->nr_samplers; ++i) {
1272       if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
1273          lp_sampler_static_state(&key->sampler[i],
1274                                  lp->fragment_sampler_views[i],
1275                                  lp->sampler[i]);
1276       }
1277    }
1278 }
1279
1280 /**
1281  * Update fragment state.  This is called just prior to drawing
1282  * something when some fragment-related state has changed.
1283  */
1284 void
1285 llvmpipe_update_fs(struct llvmpipe_context *lp)
1286 {
1287    struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
1288    struct lp_fragment_shader *shader = lp->fs;
1289    struct lp_fragment_shader_variant_key key;
1290    struct lp_fragment_shader_variant *variant = NULL;
1291    struct lp_fs_variant_list_item *li;
1292
1293    make_variant_key(lp, shader, &key);
1294
1295    li = first_elem(&shader->variants);
1296    while(!at_end(&shader->variants, li)) {
1297       if(memcmp(&li->base->key, &key, shader->variant_key_size) == 0) {
1298          variant = li->base;
1299          break;
1300       }
1301       li = next_elem(li);
1302    }
1303
1304    if (variant) {
1305       move_to_head(&lp->fs_variants_list, &variant->list_item_global);
1306    }
1307    else {
1308       int64_t t0, t1;
1309       int64_t dt;
1310       unsigned i;
1311       if (lp->nr_fs_variants >= LP_MAX_SHADER_VARIANTS) {
1312          struct pipe_context *pipe = &lp->pipe;
1313
1314          /*
1315           * XXX: we need to flush the context until we have some sort of reference
1316           * counting in fragment shaders as they may still be binned
1317           * Flushing alone might not be sufficient we need to wait on it too.
1318           */
1319          llvmpipe_finish(pipe, __FUNCTION__);
1320
1321          for (i = 0; i < LP_MAX_SHADER_VARIANTS / 4; i++) {
1322             struct lp_fs_variant_list_item *item = last_elem(&lp->fs_variants_list);
1323             remove_shader_variant(lp, item->base);
1324          }
1325       }
1326       t0 = os_time_get();
1327
1328       variant = generate_variant(screen, shader, &key);
1329
1330       t1 = os_time_get();
1331       dt = t1 - t0;
1332       LP_COUNT_ADD(llvm_compile_time, dt);
1333       LP_COUNT_ADD(nr_llvm_compiles, 2);  /* emit vs. omit in/out test */
1334
1335       if (variant) {
1336          insert_at_head(&shader->variants, &variant->list_item_local);
1337          insert_at_head(&lp->fs_variants_list, &variant->list_item_global);
1338          lp->nr_fs_variants++;
1339          shader->variants_cached++;
1340       }
1341    }
1342
1343    lp_setup_set_fs_variant(lp->setup, variant);
1344 }
1345
1346
1347
1348
1349
1350
1351
1352 void
1353 llvmpipe_init_fs_funcs(struct llvmpipe_context *llvmpipe)
1354 {
1355    llvmpipe->pipe.create_fs_state = llvmpipe_create_fs_state;
1356    llvmpipe->pipe.bind_fs_state   = llvmpipe_bind_fs_state;
1357    llvmpipe->pipe.delete_fs_state = llvmpipe_delete_fs_state;
1358
1359    llvmpipe->pipe.set_constant_buffer = llvmpipe_set_constant_buffer;
1360 }