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
 103
 104 static unsigned fs_no = 0;
 105
 106
 107
 108 /**
 109  * Expand the relevent bits of mask_input to a 4-dword mask for the
 110  * four pixels in a 2x2 quad.  This will set the four elements of the
 111  * quad mask vector to 0 or ~0.
 112  *
 113  * \param quad  which quad of the quad group to test, in [0,3]
 114  * \param mask_input  bitwise mask for the whole 4x4 stamp
 115  */
 116 static LLVMValueRef
 117 generate_quad_mask(LLVMBuilderRef builder,
 118                    struct lp_type fs_type,
 119                    unsigned quad,
 120                    LLVMValueRef mask_input) /* int32 */
 121 {
 122    struct lp_type mask_type;
 123    LLVMTypeRef i32t = LLVMInt32Type();
 124    LLVMValueRef bits[4];
 125    LLVMValueRef mask;
 126    int shift;
 127
 128    /*
 129     * XXX: We'll need a different path for 16 x u8
 130     */
 131    assert(fs_type.width == 32);
 132    assert(fs_type.length == 4);
 133    mask_type = lp_int_type(fs_type);
 134
 135    /*
 136     * mask_input >>= (quad * 4)
 137     */
 138
 139    switch (quad) {
 140    case 0:
 141       shift = 0;
 142       break;
 143    case 1:
 144       shift = 2;
 145       break;
 146    case 2:
 147       shift = 8;
 148       break;
 149    case 3:
 150       shift = 10;
 151       break;
 152    default:
 153       assert(0);
 154       shift = 0;
 155    }
 156
 157    mask_input = LLVMBuildLShr(builder,
 158                               mask_input,
 159                               LLVMConstInt(i32t, shift, 0),
 160                               "");
 161
 162    /*
 163     * mask = { mask_input & (1 << i), for i in [0,3] }
 164     */
 165
 166    mask = lp_build_broadcast(builder, lp_build_vec_type(mask_type), mask_input);
 167
 168    bits[0] = LLVMConstInt(i32t, 1 << 0, 0);
 169    bits[1] = LLVMConstInt(i32t, 1 << 1, 0);
 170    bits[2] = LLVMConstInt(i32t, 1 << 4, 0);
 171    bits[3] = LLVMConstInt(i32t, 1 << 5, 0);
 172
 173    mask = LLVMBuildAnd(builder, mask, LLVMConstVector(bits, 4), "");
 174
 175    /*
 176     * mask = mask != 0 ? ~0 : 0
 177     */
 178
 179    mask = lp_build_compare(builder,
 180                            mask_type, PIPE_FUNC_NOTEQUAL,
 181                            mask,
 182                            lp_build_const_int_vec(mask_type, 0));
 183
 184    return mask;
 185 }
 186
 187
 188 #define EARLY_DEPTH_TEST  0x1
 189 #define LATE_DEPTH_TEST   0x2
 190 #define EARLY_DEPTH_WRITE 0x4
 191 #define LATE_DEPTH_WRITE  0x8
 192
 193 static int
 194 find_output_by_semantic( const struct tgsi_shader_info *info,
 195                          unsigned semantic,
 196                          unsigned index )
 197 {
 198    int i;
 199
 200    for (i = 0; i < info->num_outputs; i++)
 201       if (info->output_semantic_name[i] == semantic &&
 202           info->output_semantic_index[i] == index)
 203          return i;
 204
 205    return -1;
 206 }
 207
 208
 209 /**
 210  * Generate the fragment shader, depth/stencil test, and alpha tests.
 211  * \param i  which quad in the tile, in range [0,3]
 212  * \param partial_mask  if 1, do mask_input testing
 213  */
 214 static void
 215 generate_fs(struct llvmpipe_context *lp,
 216             struct lp_fragment_shader *shader,
 217             const struct lp_fragment_shader_variant_key *key,
 218             LLVMBuilderRef builder,
 219             struct lp_type type,
 220             LLVMValueRef context_ptr,
 221             unsigned i,
 222             struct lp_build_interp_soa_context *interp,
 223             struct lp_build_sampler_soa *sampler,
 224             LLVMValueRef *pmask,
 225             LLVMValueRef (*color)[4],
 226             LLVMValueRef depth_ptr,
 227             LLVMValueRef facing,
 228             unsigned partial_mask,
 229             LLVMValueRef mask_input,
 230             LLVMValueRef counter)
 231 {
 232    const struct util_format_description *zs_format_desc = NULL;
 233    const struct tgsi_token *tokens = shader->base.tokens;
 234    LLVMTypeRef vec_type;
 235    LLVMValueRef consts_ptr;
 236    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][NUM_CHANNELS];
 237    LLVMValueRef z;
 238    LLVMValueRef zs_value = NULL;
 239    LLVMValueRef stencil_refs[2];
 240    struct lp_build_mask_context mask;
 241    boolean simple_shader = (shader->info.base.file_count[TGSI_FILE_SAMPLER] == 0 &&
 242                             shader->info.base.num_inputs < 3 &&
 243                             shader->info.base.num_instructions < 8);
 244    unsigned attrib;
 245    unsigned chan;
 246    unsigned cbuf;
 247    unsigned depth_mode;
 248
 249    if (key->depth.enabled ||
 250        key->stencil[0].enabled ||
 251        key->stencil[1].enabled) {
 252
 253       zs_format_desc = util_format_description(key->zsbuf_format);
 254       assert(zs_format_desc);
 255
 256       if (!shader->info.base.writes_z) {
 257          if (key->alpha.enabled || shader->info.base.uses_kill)
 258             /* With alpha test and kill, can do the depth test early
 259              * and hopefully eliminate some quads.  But need to do a
 260              * special deferred depth write once the final mask value
 261              * is known.
 262              */
 263             depth_mode = EARLY_DEPTH_TEST | LATE_DEPTH_WRITE;
 264          else
 265             depth_mode = EARLY_DEPTH_TEST | EARLY_DEPTH_WRITE;
 266       }
 267       else {
 268          depth_mode = LATE_DEPTH_TEST | LATE_DEPTH_WRITE;
 269       }
 270
 271       if (!(key->depth.enabled && key->depth.writemask) &&
 272           !(key->stencil[0].enabled && key->stencil[0].writemask))
 273          depth_mode &= ~(LATE_DEPTH_WRITE | EARLY_DEPTH_WRITE);
 274    }
 275    else {
 276       depth_mode = 0;
 277    }
 278
 279    assert(i < 4);
 280
 281    stencil_refs[0] = lp_jit_context_stencil_ref_front_value(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          LLVMBuildStore(builder, zs_value, depth_ptr);
 330    }
 331
 332    lp_build_interp_soa_update_inputs(interp, i);
 333
 334    /* Build the actual shader */
 335    lp_build_tgsi_soa(builder, tokens, type, &mask,
 336                      consts_ptr, interp->pos, interp->inputs,
 337                      outputs, sampler, &shader->info.base);
 338
 339
 340    /* Alpha test */
 341    if (key->alpha.enabled) {
 342       int color0 = find_output_by_semantic(&shader->info.base,
 343                                            TGSI_SEMANTIC_COLOR,
 344                                            0);
 345
 346       if (color0 != -1) {
 347          LLVMValueRef alpha = LLVMBuildLoad(builder, outputs[color0][3], "alpha");
 348          LLVMValueRef alpha_ref_value;
 349
 350          alpha_ref_value = lp_jit_context_alpha_ref_value(builder, context_ptr);
 351          alpha_ref_value = lp_build_broadcast(builder, vec_type, alpha_ref_value);
 352
 353          lp_build_alpha_test(builder, key->alpha.func, type,
 354                              &mask, alpha, alpha_ref_value,
 355                              (depth_mode & LATE_DEPTH_TEST) != 0);
 356       }
 357    }
 358
 359    /* Late Z test */
 360    if (depth_mode & LATE_DEPTH_TEST) {
 361       int pos0 = find_output_by_semantic(&shader->info.base,
 362                                          TGSI_SEMANTIC_POSITION,
 363                                          0);
 364
 365       if (pos0 != -1) {
 366          z = LLVMBuildLoad(builder, outputs[pos0][2], "z");
 367          lp_build_name(z, "output%u.%u.%c", i, pos0, "xyzw"[chan]);
 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          LLVMBuildStore(builder, zs_value, depth_ptr);
 384    }
 385    else if ((depth_mode & EARLY_DEPTH_TEST) &&
 386             (depth_mode & LATE_DEPTH_WRITE))
 387    {
 388       /* Need to apply a reduced mask to the depth write.  Reload the
 389        * depth value, update from zs_value with the new mask value and
 390        * write that out.
 391        */
 392       lp_build_deferred_depth_write(builder,
 393                                     type,
 394                                     zs_format_desc,
 395                                     &mask,
 396                                     depth_ptr,
 397                                     zs_value);
 398    }
 399
 400
 401    /* Color write  */
 402    for (attrib = 0; attrib < shader->info.base.num_outputs; ++attrib)
 403    {
 404       if (shader->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_COLOR)
 405       {
 406          unsigned cbuf = shader->info.base.output_semantic_index[attrib];
 407          for(chan = 0; chan < NUM_CHANNELS; ++chan)
 408          {
 409             /* XXX: just initialize outputs to point at colors[] and
 410              * skip this.
 411              */
 412             LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
 413             lp_build_name(out, "color%u.%u.%c", i, attrib, "rgba"[chan]);
 414             LLVMBuildStore(builder, out, color[cbuf][chan]);
 415          }
 416       }
 417    }
 418
 419    if (counter)
 420       lp_build_occlusion_count(builder, type,
 421                                lp_build_mask_value(&mask), counter);
 422
 423    *pmask = lp_build_mask_end(&mask);
 424 }
 425
 426
 427 /**
 428  * Generate color blending and color output.
 429  * \param rt  the render target index (to index blend, colormask state)
 430  * \param type  the pixel color type
 431  * \param context_ptr  pointer to the runtime JIT context
 432  * \param mask  execution mask (active fragment/pixel mask)
 433  * \param src  colors from the fragment shader
 434  * \param dst_ptr  the destination color buffer pointer
 435  */
 436 static void
 437 generate_blend(const struct pipe_blend_state *blend,
 438                unsigned rt,
 439                LLVMBuilderRef builder,
 440                struct lp_type type,
 441                LLVMValueRef context_ptr,
 442                LLVMValueRef mask,
 443                LLVMValueRef *src,
 444                LLVMValueRef dst_ptr,
 445                boolean do_branch)
 446 {
 447    struct lp_build_context bld;
 448    struct lp_build_mask_context mask_ctx;
 449    LLVMTypeRef vec_type;
 450    LLVMValueRef const_ptr;
 451    LLVMValueRef con[4];
 452    LLVMValueRef dst[4];
 453    LLVMValueRef res[4];
 454    unsigned chan;
 455
 456    lp_build_context_init(&bld, builder, type);
 457
 458    lp_build_mask_begin(&mask_ctx, builder, type, mask);
 459    if (do_branch)
 460       lp_build_mask_check(&mask_ctx);
 461
 462    vec_type = lp_build_vec_type(type);
 463
 464    const_ptr = lp_jit_context_blend_color(builder, context_ptr);
 465    const_ptr = LLVMBuildBitCast(builder, const_ptr,
 466                                 LLVMPointerType(vec_type, 0), "");
 467
 468    /* load constant blend color and colors from the dest color buffer */
 469    for(chan = 0; chan < 4; ++chan) {
 470       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
 471       con[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, const_ptr, &index, 1, ""), "");
 472
 473       dst[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, dst_ptr, &index, 1, ""), "");
 474
 475       lp_build_name(con[chan], "con.%c", "rgba"[chan]);
 476       lp_build_name(dst[chan], "dst.%c", "rgba"[chan]);
 477    }
 478
 479    /* do blend */
 480    lp_build_blend_soa(builder, blend, type, rt, src, dst, con, res);
 481
 482    /* store results to color buffer */
 483    for(chan = 0; chan < 4; ++chan) {
 484       if(blend->rt[rt].colormask & (1 << chan)) {
 485          LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
 486          lp_build_name(res[chan], "res.%c", "rgba"[chan]);
 487          res[chan] = lp_build_select(&bld, mask, res[chan], dst[chan]);
 488          LLVMBuildStore(builder, res[chan], LLVMBuildGEP(builder, dst_ptr, &index, 1, ""));
 489       }
 490    }
 491
 492    lp_build_mask_end(&mask_ctx);
 493 }
 494
 495
 496 /**
 497  * Generate the runtime callable function for the whole fragment pipeline.
 498  * Note that the function which we generate operates on a block of 16
 499  * pixels at at time.  The block contains 2x2 quads.  Each quad contains
 500  * 2x2 pixels.
 501  */
 502 static void
 503 generate_fragment(struct llvmpipe_context *lp,
 504                   struct lp_fragment_shader *shader,
 505                   struct lp_fragment_shader_variant *variant,
 506                   unsigned partial_mask)
 507 {
 508    struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
 509    const struct lp_fragment_shader_variant_key *key = &variant->key;
 510    char func_name[256];
 511    struct lp_type fs_type;
 512    struct lp_type blend_type;
 513    LLVMTypeRef fs_elem_type;
 514    LLVMTypeRef fs_int_vec_type;
 515    LLVMTypeRef blend_vec_type;
 516    LLVMTypeRef arg_types[11];
 517    LLVMTypeRef func_type;
 518    LLVMValueRef context_ptr;
 519    LLVMValueRef x;
 520    LLVMValueRef y;
 521    LLVMValueRef a0_ptr;
 522    LLVMValueRef dadx_ptr;
 523    LLVMValueRef dady_ptr;
 524    LLVMValueRef color_ptr_ptr;
 525    LLVMValueRef depth_ptr;
 526    LLVMValueRef mask_input;
 527    LLVMValueRef counter = NULL;
 528    LLVMBasicBlockRef block;
 529    LLVMBuilderRef builder;
 530    struct lp_build_sampler_soa *sampler;
 531    struct lp_build_interp_soa_context interp;
 532    LLVMValueRef fs_mask[LP_MAX_VECTOR_LENGTH];
 533    LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS][LP_MAX_VECTOR_LENGTH];
 534    LLVMValueRef blend_mask;
 535    LLVMValueRef function;
 536    LLVMValueRef facing;
 537    unsigned num_fs;
 538    unsigned i;
 539    unsigned chan;
 540    unsigned cbuf;
 541
 542
 543    /* TODO: actually pick these based on the fs and color buffer
 544     * characteristics. */
 545
 546    memset(&fs_type, 0, sizeof fs_type);
 547    fs_type.floating = TRUE; /* floating point values */
 548    fs_type.sign = TRUE;     /* values are signed */
 549    fs_type.norm = FALSE;    /* values are not limited to [0,1] or [-1,1] */
 550    fs_type.width = 32;      /* 32-bit float */
 551    fs_type.length = 4;      /* 4 elements per vector */
 552    num_fs = 4;              /* number of quads per block */
 553
 554    memset(&blend_type, 0, sizeof blend_type);
 555    blend_type.floating = FALSE; /* values are integers */
 556    blend_type.sign = FALSE;     /* values are unsigned */
 557    blend_type.norm = TRUE;      /* values are in [0,1] or [-1,1] */
 558    blend_type.width = 8;        /* 8-bit ubyte values */
 559    blend_type.length = 16;      /* 16 elements per vector */
 560
 561    /*
 562     * Generate the function prototype. Any change here must be reflected in
 563     * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
 564     */
 565
 566    fs_elem_type = lp_build_elem_type(fs_type);
 567    fs_int_vec_type = lp_build_int_vec_type(fs_type);
 568
 569    blend_vec_type = lp_build_vec_type(blend_type);
 570
 571    util_snprintf(func_name, sizeof(func_name), "fs%u_variant%u_%s",
 572                  shader->no, variant->no, partial_mask ? "partial" : "whole");
 573
 574    arg_types[0] = screen->context_ptr_type;            /* context */
 575    arg_types[1] = LLVMInt32Type();                     /* x */
 576    arg_types[2] = LLVMInt32Type();                     /* y */
 577    arg_types[3] = LLVMFloatType();                     /* facing */
 578    arg_types[4] = LLVMPointerType(fs_elem_type, 0);    /* a0 */
 579    arg_types[5] = LLVMPointerType(fs_elem_type, 0);    /* dadx */
 580    arg_types[6] = LLVMPointerType(fs_elem_type, 0);    /* dady */
 581    arg_types[7] = LLVMPointerType(LLVMPointerType(blend_vec_type, 0), 0);  /* color */
 582    arg_types[8] = LLVMPointerType(fs_int_vec_type, 0); /* depth */
 583    arg_types[9] = LLVMInt32Type();                     /* mask_input */
 584    arg_types[10] = LLVMPointerType(LLVMInt32Type(), 0);/* counter */
 585
 586    func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
 587
 588    function = LLVMAddFunction(screen->module, func_name, func_type);
 589    LLVMSetFunctionCallConv(function, LLVMCCallConv);
 590
 591    variant->function[partial_mask] = function;
 592
 593
 594    /* XXX: need to propagate noalias down into color param now we are
 595     * passing a pointer-to-pointer?
 596     */
 597    for(i = 0; i < Elements(arg_types); ++i)
 598       if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
 599          LLVMAddAttribute(LLVMGetParam(function, i), LLVMNoAliasAttribute);
 600
 601    context_ptr  = LLVMGetParam(function, 0);
 602    x            = LLVMGetParam(function, 1);
 603    y            = LLVMGetParam(function, 2);
 604    facing       = LLVMGetParam(function, 3);
 605    a0_ptr       = LLVMGetParam(function, 4);
 606    dadx_ptr     = LLVMGetParam(function, 5);
 607    dady_ptr     = LLVMGetParam(function, 6);
 608    color_ptr_ptr = LLVMGetParam(function, 7);
 609    depth_ptr    = LLVMGetParam(function, 8);
 610    mask_input   = LLVMGetParam(function, 9);
 611
 612    lp_build_name(context_ptr, "context");
 613    lp_build_name(x, "x");
 614    lp_build_name(y, "y");
 615    lp_build_name(a0_ptr, "a0");
 616    lp_build_name(dadx_ptr, "dadx");
 617    lp_build_name(dady_ptr, "dady");
 618    lp_build_name(color_ptr_ptr, "color_ptr_ptr");
 619    lp_build_name(depth_ptr, "depth");
 620    lp_build_name(mask_input, "mask_input");
 621
 622    if (key->occlusion_count) {
 623       counter = LLVMGetParam(function, 10);
 624       lp_build_name(counter, "counter");
 625    }
 626
 627    /*
 628     * Function body
 629     */
 630
 631    block = LLVMAppendBasicBlock(function, "entry");
 632    builder = LLVMCreateBuilder();
 633    LLVMPositionBuilderAtEnd(builder, block);
 634
 635    /*
 636     * The shader input interpolation info is not explicitely baked in the
 637     * shader key, but everything it derives from (TGSI, and flatshade) is
 638     * already included in the shader key.
 639     */
 640    lp_build_interp_soa_init(&interp,
 641                             lp->num_inputs,
 642                             lp->inputs,
 643                             builder, fs_type,
 644                             a0_ptr, dadx_ptr, dady_ptr,
 645                             x, y);
 646
 647    /* code generated texture sampling */
 648    sampler = lp_llvm_sampler_soa_create(key->sampler, context_ptr);
 649
 650    /* loop over quads in the block */
 651    for(i = 0; i < num_fs; ++i) {
 652       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
 653       LLVMValueRef out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS];
 654       LLVMValueRef depth_ptr_i;
 655
 656       depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &index, 1, "");
 657
 658       generate_fs(lp, shader, key,
 659                   builder,
 660                   fs_type,
 661                   context_ptr,
 662                   i,
 663                   &interp,
 664                   sampler,
 665                   &fs_mask[i], /* output */
 666                   out_color,
 667                   depth_ptr_i,
 668                   facing,
 669                   partial_mask,
 670                   mask_input,
 671                   counter);
 672
 673       for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++)
 674          for(chan = 0; chan < NUM_CHANNELS; ++chan)
 675             fs_out_color[cbuf][chan][i] = out_color[cbuf][chan];
 676    }
 677
 678    sampler->destroy(sampler);
 679
 680    /* Loop over color outputs / color buffers to do blending.
 681     */
 682    for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
 683       LLVMValueRef color_ptr;
 684       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), cbuf, 0);
 685       LLVMValueRef blend_in_color[NUM_CHANNELS];
 686       unsigned rt;
 687
 688       /*
 689        * Convert the fs's output color and mask to fit to the blending type.
 690        */
 691       for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 692          LLVMValueRef fs_color_vals[LP_MAX_VECTOR_LENGTH];
 693
 694          for (i = 0; i < num_fs; i++) {
 695             fs_color_vals[i] =
 696                LLVMBuildLoad(builder, fs_out_color[cbuf][chan][i], "fs_color_vals");
 697          }
 698
 699          lp_build_conv(builder, fs_type, blend_type,
 700                        fs_color_vals,
 701                        num_fs,
 702                        &blend_in_color[chan], 1);
 703
 704          lp_build_name(blend_in_color[chan], "color%d.%c", cbuf, "rgba"[chan]);
 705       }
 706
 707       if (partial_mask || !variant->opaque) {
 708          lp_build_conv_mask(builder, fs_type, blend_type,
 709                             fs_mask, num_fs,
 710                             &blend_mask, 1);
 711       } else {
 712          blend_mask = lp_build_const_int_vec(blend_type, ~0);
 713       }
 714
 715       color_ptr = LLVMBuildLoad(builder,
 716                                 LLVMBuildGEP(builder, color_ptr_ptr, &index, 1, ""),
 717                                 "");
 718       lp_build_name(color_ptr, "color_ptr%d", cbuf);
 719
 720       /* which blend/colormask state to use */
 721       rt = key->blend.independent_blend_enable ? cbuf : 0;
 722
 723       /*
 724        * Blending.
 725        */
 726       {
 727          /* Could the 4x4 have been killed?
 728           */
 729          boolean do_branch = ((key->depth.enabled || key->stencil[0].enabled) &&
 730                               !key->alpha.enabled &&
 731                               !shader->info.base.uses_kill);
 732
 733          generate_blend(&key->blend,
 734                         rt,
 735                         builder,
 736                         blend_type,
 737                         context_ptr,
 738                         blend_mask,
 739                         blend_in_color,
 740                         color_ptr,
 741                         do_branch);
 742       }
 743    }
 744
 745 #ifdef PIPE_ARCH_X86
 746    /* Avoid corrupting the FPU stack on 32bit OSes. */
 747    lp_build_intrinsic(builder, "llvm.x86.mmx.emms", LLVMVoidType(), NULL, 0);
 748 #endif
 749
 750    LLVMBuildRetVoid(builder);
 751
 752    LLVMDisposeBuilder(builder);
 753
 754
 755    /* Verify the LLVM IR.  If invalid, dump and abort */
 756 #ifdef DEBUG
 757    if(LLVMVerifyFunction(function, LLVMPrintMessageAction)) {
 758       if (1)
 759          lp_debug_dump_value(function);
 760       abort();
 761    }
 762 #endif
 763
 764    /* Apply optimizations to LLVM IR */
 765    LLVMRunFunctionPassManager(screen->pass, function);
 766
 767    if ((gallivm_debug & GALLIVM_DEBUG_IR) || (LP_DEBUG & DEBUG_FS)) {
 768       /* Print the LLVM IR to stderr */
 769       lp_debug_dump_value(function);
 770       debug_printf("\n");
 771    }
 772
 773    /*
 774     * Translate the LLVM IR into machine code.
 775     */
 776    {
 777       void *f = LLVMGetPointerToGlobal(screen->engine, function);
 778
 779       variant->jit_function[partial_mask] = (lp_jit_frag_func)pointer_to_func(f);
 780
 781       if ((gallivm_debug & GALLIVM_DEBUG_ASM) || (LP_DEBUG & DEBUG_FS)) {
 782          lp_disassemble(f);
 783       }
 784       lp_func_delete_body(function);
 785    }
 786 }
 787
 788
 789 static void
 790 dump_fs_variant_key(const struct lp_fragment_shader_variant_key *key)
 791 {
 792    unsigned i;
 793
 794    debug_printf("fs variant %p:\n", (void *) key);
 795
 796    if (key->flatshade) {
 797       debug_printf("flatshade = 1\n");
 798    }
 799    for (i = 0; i < key->nr_cbufs; ++i) {
 800       debug_printf("cbuf_format[%u] = %s\n", i, util_format_name(key->cbuf_format[i]));
 801    }
 802    if (key->depth.enabled) {
 803       debug_printf("depth.format = %s\n", util_format_name(key->zsbuf_format));
 804       debug_printf("depth.func = %s\n", util_dump_func(key->depth.func, TRUE));
 805       debug_printf("depth.writemask = %u\n", key->depth.writemask);
 806    }
 807
 808    for (i = 0; i < 2; ++i) {
 809       if (key->stencil[i].enabled) {
 810          debug_printf("stencil[%u].func = %s\n", i, util_dump_func(key->stencil[i].func, TRUE));
 811          debug_printf("stencil[%u].fail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].fail_op, TRUE));
 812          debug_printf("stencil[%u].zpass_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zpass_op, TRUE));
 813          debug_printf("stencil[%u].zfail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zfail_op, TRUE));
 814          debug_printf("stencil[%u].valuemask = 0x%x\n", i, key->stencil[i].valuemask);
 815          debug_printf("stencil[%u].writemask = 0x%x\n", i, key->stencil[i].writemask);
 816       }
 817    }
 818
 819    if (key->alpha.enabled) {
 820       debug_printf("alpha.func = %s\n", util_dump_func(key->alpha.func, TRUE));
 821    }
 822
 823    if (key->occlusion_count) {
 824       debug_printf("occlusion_count = 1\n");
 825    }
 826
 827    if (key->blend.logicop_enable) {
 828       debug_printf("blend.logicop_func = %s\n", util_dump_logicop(key->blend.logicop_func, TRUE));
 829    }
 830    else if (key->blend.rt[0].blend_enable) {
 831       debug_printf("blend.rgb_func = %s\n",   util_dump_blend_func  (key->blend.rt[0].rgb_func, TRUE));
 832       debug_printf("blend.rgb_src_factor = %s\n",   util_dump_blend_factor(key->blend.rt[0].rgb_src_factor, TRUE));
 833       debug_printf("blend.rgb_dst_factor = %s\n",   util_dump_blend_factor(key->blend.rt[0].rgb_dst_factor, TRUE));
 834       debug_printf("blend.alpha_func = %s\n",       util_dump_blend_func  (key->blend.rt[0].alpha_func, TRUE));
 835       debug_printf("blend.alpha_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_src_factor, TRUE));
 836       debug_printf("blend.alpha_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_dst_factor, TRUE));
 837    }
 838    debug_printf("blend.colormask = 0x%x\n", key->blend.rt[0].colormask);
 839    for (i = 0; i < key->nr_samplers; ++i) {
 840       debug_printf("sampler[%u] = \n", i);
 841       debug_printf("  .format = %s\n",
 842                    util_format_name(key->sampler[i].format));
 843       debug_printf("  .target = %s\n",
 844                    util_dump_tex_target(key->sampler[i].target, TRUE));
 845       debug_printf("  .pot = %u %u %u\n",
 846                    key->sampler[i].pot_width,
 847                    key->sampler[i].pot_height,
 848                    key->sampler[i].pot_depth);
 849       debug_printf("  .wrap = %s %s %s\n",
 850                    util_dump_tex_wrap(key->sampler[i].wrap_s, TRUE),
 851                    util_dump_tex_wrap(key->sampler[i].wrap_t, TRUE),
 852                    util_dump_tex_wrap(key->sampler[i].wrap_r, TRUE));
 853       debug_printf("  .min_img_filter = %s\n",
 854                    util_dump_tex_filter(key->sampler[i].min_img_filter, TRUE));
 855       debug_printf("  .min_mip_filter = %s\n",
 856                    util_dump_tex_mipfilter(key->sampler[i].min_mip_filter, TRUE));
 857       debug_printf("  .mag_img_filter = %s\n",
 858                    util_dump_tex_filter(key->sampler[i].mag_img_filter, TRUE));
 859       if (key->sampler[i].compare_mode != PIPE_TEX_COMPARE_NONE)
 860          debug_printf("  .compare_func = %s\n", util_dump_func(key->sampler[i].compare_func, TRUE));
 861       debug_printf("  .normalized_coords = %u\n", key->sampler[i].normalized_coords);
 862       debug_printf("  .min_max_lod_equal = %u\n", key->sampler[i].min_max_lod_equal);
 863       debug_printf("  .lod_bias_non_zero = %u\n", key->sampler[i].lod_bias_non_zero);
 864       debug_printf("  .apply_min_lod = %u\n", key->sampler[i].apply_min_lod);
 865       debug_printf("  .apply_max_lod = %u\n", key->sampler[i].apply_max_lod);
 866    }
 867 }
 868
 869
 870 void
 871 lp_debug_fs_variant(const struct lp_fragment_shader_variant *variant)
 872 {
 873    debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
 874                 variant->shader->no, variant->no);
 875    tgsi_dump(variant->shader->base.tokens, 0);
 876    dump_fs_variant_key(&variant->key);
 877    debug_printf("variant->opaque = %u\n", variant->opaque);
 878    debug_printf("\n");
 879 }
 880
 881 static struct lp_fragment_shader_variant *
 882 generate_variant(struct llvmpipe_context *lp,
 883                  struct lp_fragment_shader *shader,
 884                  const struct lp_fragment_shader_variant_key *key)
 885 {
 886    struct lp_fragment_shader_variant *variant;
 887    boolean fullcolormask;
 888
 889    variant = CALLOC_STRUCT(lp_fragment_shader_variant);
 890    if(!variant)
 891       return NULL;
 892
 893    variant->shader = shader;
 894    variant->list_item_global.base = variant;
 895    variant->list_item_local.base = variant;
 896    variant->no = shader->variants_created++;
 897
 898    memcpy(&variant->key, key, shader->variant_key_size);
 899
 900    /*
 901     * Determine whether we are touching all channels in the color buffer.
 902     */
 903    fullcolormask = FALSE;
 904    if (key->nr_cbufs == 1) {
 905       const struct util_format_description *format_desc;
 906       format_desc = util_format_description(key->cbuf_format[0]);
 907       if ((~key->blend.rt[0].colormask &
 908            util_format_colormask(format_desc)) == 0) {
 909          fullcolormask = TRUE;
 910       }
 911    }
 912
 913    variant->opaque =
 914          !key->blend.logicop_enable &&
 915          !key->blend.rt[0].blend_enable &&
 916          fullcolormask &&
 917          !key->stencil[0].enabled &&
 918          !key->alpha.enabled &&
 919          !key->depth.enabled &&
 920          !shader->info.base.uses_kill
 921          ? TRUE : FALSE;
 922
 923
 924    if ((LP_DEBUG & DEBUG_FS) || (gallivm_debug & GALLIVM_DEBUG_IR)) {
 925       lp_debug_fs_variant(variant);
 926    }
 927
 928    generate_fragment(lp, shader, variant, RAST_EDGE_TEST);
 929
 930    if (variant->opaque) {
 931       /* Specialized shader, which doesn't need to read the color buffer. */
 932       generate_fragment(lp, shader, variant, RAST_WHOLE);
 933    } else {
 934       variant->jit_function[RAST_WHOLE] = variant->jit_function[RAST_EDGE_TEST];
 935    }
 936
 937    return variant;
 938 }
 939
 940
 941 static void *
 942 llvmpipe_create_fs_state(struct pipe_context *pipe,
 943                          const struct pipe_shader_state *templ)
 944 {
 945    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
 946    struct lp_fragment_shader *shader;
 947    int nr_samplers;
 948
 949    shader = CALLOC_STRUCT(lp_fragment_shader);
 950    if (!shader)
 951       return NULL;
 952
 953    shader->no = fs_no++;
 954    make_empty_list(&shader->variants);
 955
 956    /* get/save the summary info for this shader */
 957    lp_build_tgsi_info(templ->tokens, &shader->info);
 958
 959    /* we need to keep a local copy of the tokens */
 960    shader->base.tokens = tgsi_dup_tokens(templ->tokens);
 961
 962    shader->draw_data = draw_create_fragment_shader(llvmpipe->draw, templ);
 963    if (shader->draw_data == NULL) {
 964       FREE((void *) shader->base.tokens);
 965       FREE(shader);
 966       return NULL;
 967    }
 968
 969    nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
 970
 971    shader->variant_key_size = Offset(struct lp_fragment_shader_variant_key,
 972                                      sampler[nr_samplers]);
 973
 974    if (LP_DEBUG & DEBUG_TGSI) {
 975       unsigned attrib;
 976       debug_printf("llvmpipe: Create fragment shader #%u %p:\n", shader->no, (void *) shader);
 977       tgsi_dump(templ->tokens, 0);
 978       debug_printf("usage masks:\n");
 979       for (attrib = 0; attrib < shader->info.base.num_inputs; ++attrib) {
 980          unsigned usage_mask = shader->info.base.input_usage_mask[attrib];
 981          debug_printf("  IN[%u].%s%s%s%s\n",
 982                       attrib,
 983                       usage_mask & TGSI_WRITEMASK_X ? "x" : "",
 984                       usage_mask & TGSI_WRITEMASK_Y ? "y" : "",
 985                       usage_mask & TGSI_WRITEMASK_Z ? "z" : "",
 986                       usage_mask & TGSI_WRITEMASK_W ? "w" : "");
 987       }
 988       debug_printf("\n");
 989    }
 990
 991    return shader;
 992 }
 993
 994
 995 static void
 996 llvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs)
 997 {
 998    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
 999
1000    if (llvmpipe->fs == fs)
1001       return;
1002
1003    draw_flush(llvmpipe->draw);
1004
1005    draw_bind_fragment_shader(llvmpipe->draw,
1006                              (llvmpipe->fs ? llvmpipe->fs->draw_data : NULL));
1007
1008    llvmpipe->fs = fs;
1009
1010    llvmpipe->dirty |= LP_NEW_FS;
1011 }
1012
1013 static void
1014 remove_shader_variant(struct llvmpipe_context *lp,
1015                       struct lp_fragment_shader_variant *variant)
1016 {
1017    struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
1018    unsigned i;
1019
1020    if (gallivm_debug & GALLIVM_DEBUG_IR) {
1021       debug_printf("llvmpipe: del fs #%u var #%u v created #%u v cached #%u v total cached #%u\n",
1022                     variant->shader->no, variant->no, variant->shader->variants_created,
1023                     variant->shader->variants_cached, lp->nr_fs_variants);
1024    }
1025    for (i = 0; i < Elements(variant->function); i++) {
1026       if (variant->function[i]) {
1027          if (variant->jit_function[i])
1028             LLVMFreeMachineCodeForFunction(screen->engine,
1029                                            variant->function[i]);
1030          LLVMDeleteFunction(variant->function[i]);
1031       }
1032    }
1033    remove_from_list(&variant->list_item_local);
1034    variant->shader->variants_cached--;
1035    remove_from_list(&variant->list_item_global);
1036    lp->nr_fs_variants--;
1037    FREE(variant);
1038 }
1039
1040 static void
1041 llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs)
1042 {
1043    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
1044    struct lp_fragment_shader *shader = fs;
1045    struct lp_fs_variant_list_item *li;
1046
1047    assert(fs != llvmpipe->fs);
1048    (void) llvmpipe;
1049
1050    /*
1051     * XXX: we need to flush the context until we have some sort of reference
1052     * counting in fragment shaders as they may still be binned
1053     * Flushing alone might not sufficient we need to wait on it too.
1054     */
1055
1056    llvmpipe_finish(pipe, __FUNCTION__);
1057
1058    li = first_elem(&shader->variants);
1059    while(!at_end(&shader->variants, li)) {
1060       struct lp_fs_variant_list_item *next = next_elem(li);
1061       remove_shader_variant(llvmpipe, li->base);
1062       li = next;
1063    }
1064
1065    draw_delete_fragment_shader(llvmpipe->draw, shader->draw_data);
1066
1067    assert(shader->variants_cached == 0);
1068    FREE((void *) shader->base.tokens);
1069    FREE(shader);
1070 }
1071
1072
1073
1074 static void
1075 llvmpipe_set_constant_buffer(struct pipe_context *pipe,
1076                              uint shader, uint index,
1077                              struct pipe_resource *constants)
1078 {
1079    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
1080    unsigned size = constants ? constants->width0 : 0;
1081    const void *data = constants ? llvmpipe_resource_data(constants) : NULL;
1082
1083    assert(shader < PIPE_SHADER_TYPES);
1084    assert(index < PIPE_MAX_CONSTANT_BUFFERS);
1085
1086    if(llvmpipe->constants[shader][index] == constants)
1087       return;
1088
1089    draw_flush(llvmpipe->draw);
1090
1091    /* note: reference counting */
1092    pipe_resource_reference(&llvmpipe->constants[shader][index], constants);
1093
1094    if(shader == PIPE_SHADER_VERTEX ||
1095       shader == PIPE_SHADER_GEOMETRY) {
1096       draw_set_mapped_constant_buffer(llvmpipe->draw, shader,
1097                                       index, data, size);
1098    }
1099
1100    llvmpipe->dirty |= LP_NEW_CONSTANTS;
1101 }
1102
1103
1104 /**
1105  * Return the blend factor equivalent to a destination alpha of one.
1106  */
1107 static INLINE unsigned
1108 force_dst_alpha_one(unsigned factor)
1109 {
1110    switch(factor) {
1111    case PIPE_BLENDFACTOR_DST_ALPHA:
1112       return PIPE_BLENDFACTOR_ONE;
1113    case PIPE_BLENDFACTOR_INV_DST_ALPHA:
1114       return PIPE_BLENDFACTOR_ZERO;
1115    case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
1116       return PIPE_BLENDFACTOR_ZERO;
1117    }
1118
1119    return factor;
1120 }
1121
1122
1123 /**
1124  * We need to generate several variants of the fragment pipeline to match
1125  * all the combinations of the contributing state atoms.
1126  *
1127  * TODO: there is actually no reason to tie this to context state -- the
1128  * generated code could be cached globally in the screen.
1129  */
1130 static void
1131 make_variant_key(struct llvmpipe_context *lp,
1132                  struct lp_fragment_shader *shader,
1133                  struct lp_fragment_shader_variant_key *key)
1134 {
1135    unsigned i;
1136
1137    memset(key, 0, shader->variant_key_size);
1138
1139    if (lp->framebuffer.zsbuf) {
1140       if (lp->depth_stencil->depth.enabled) {
1141          key->zsbuf_format = lp->framebuffer.zsbuf->format;
1142          memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth);
1143       }
1144       if (lp->depth_stencil->stencil[0].enabled) {
1145          key->zsbuf_format = lp->framebuffer.zsbuf->format;
1146          memcpy(&key->stencil, &lp->depth_stencil->stencil, sizeof key->stencil);
1147       }
1148    }
1149
1150    key->alpha.enabled = lp->depth_stencil->alpha.enabled;
1151    if(key->alpha.enabled)
1152       key->alpha.func = lp->depth_stencil->alpha.func;
1153    /* alpha.ref_value is passed in jit_context */
1154
1155    key->flatshade = lp->rasterizer->flatshade;
1156    if (lp->active_query_count) {
1157       key->occlusion_count = TRUE;
1158    }
1159
1160    if (lp->framebuffer.nr_cbufs) {
1161       memcpy(&key->blend, lp->blend, sizeof key->blend);
1162    }
1163
1164    key->nr_cbufs = lp->framebuffer.nr_cbufs;
1165    for (i = 0; i < lp->framebuffer.nr_cbufs; i++) {
1166       enum pipe_format format = lp->framebuffer.cbufs[i]->format;
1167       struct pipe_rt_blend_state *blend_rt = &key->blend.rt[i];
1168       const struct util_format_description *format_desc;
1169
1170       key->cbuf_format[i] = format;
1171
1172       format_desc = util_format_description(format);
1173       assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
1174              format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB);
1175
1176       blend_rt->colormask = lp->blend->rt[i].colormask;
1177
1178       /*
1179        * Mask out color channels not present in the color buffer.
1180        */
1181       blend_rt->colormask &= util_format_colormask(format_desc);
1182
1183       /*
1184        * Our swizzled render tiles always have an alpha channel, but the linear
1185        * render target format often does not, so force here the dst alpha to be
1186        * one.
1187        *
1188        * This is not a mere optimization. Wrong results will be produced if the
1189        * dst alpha is used, the dst format does not have alpha, and the previous
1190        * rendering was not flushed from the swizzled to linear buffer. For
1191        * example, NonPowTwo DCT.
1192        *
1193        * TODO: This should be generalized to all channels for better
1194        * performance, but only alpha causes correctness issues.
1195        *
1196        * Also, force rgb/alpha func/factors match, to make AoS blending easier.
1197        */
1198       if (format_desc->swizzle[3] > UTIL_FORMAT_SWIZZLE_W) {
1199          blend_rt->rgb_src_factor   = force_dst_alpha_one(blend_rt->rgb_src_factor);
1200          blend_rt->rgb_dst_factor   = force_dst_alpha_one(blend_rt->rgb_dst_factor);
1201          blend_rt->alpha_func       = blend_rt->rgb_func;
1202          blend_rt->alpha_src_factor = blend_rt->rgb_src_factor;
1203          blend_rt->alpha_dst_factor = blend_rt->rgb_dst_factor;
1204       }
1205    }
1206
1207    /* This value will be the same for all the variants of a given shader:
1208     */
1209    key->nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
1210
1211    for(i = 0; i < key->nr_samplers; ++i) {
1212       if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
1213          lp_sampler_static_state(&key->sampler[i],
1214                                  lp->fragment_sampler_views[i],
1215                                  lp->sampler[i]);
1216       }
1217    }
1218 }
1219
1220 /**
1221  * Update fragment state.  This is called just prior to drawing
1222  * something when some fragment-related state has changed.
1223  */
1224 void
1225 llvmpipe_update_fs(struct llvmpipe_context *lp)
1226 {
1227    struct lp_fragment_shader *shader = lp->fs;
1228    struct lp_fragment_shader_variant_key key;
1229    struct lp_fragment_shader_variant *variant = NULL;
1230    struct lp_fs_variant_list_item *li;
1231
1232    make_variant_key(lp, shader, &key);
1233
1234    li = first_elem(&shader->variants);
1235    while(!at_end(&shader->variants, li)) {
1236       if(memcmp(&li->base->key, &key, shader->variant_key_size) == 0) {
1237          variant = li->base;
1238          break;
1239       }
1240       li = next_elem(li);
1241    }
1242
1243    if (variant) {
1244       move_to_head(&lp->fs_variants_list, &variant->list_item_global);
1245    }
1246    else {
1247       int64_t t0, t1;
1248       int64_t dt;
1249       unsigned i;
1250       if (lp->nr_fs_variants >= LP_MAX_SHADER_VARIANTS) {
1251          struct pipe_context *pipe = &lp->pipe;
1252
1253          /*
1254           * XXX: we need to flush the context until we have some sort of reference
1255           * counting in fragment shaders as they may still be binned
1256           * Flushing alone might not be sufficient we need to wait on it too.
1257           */
1258          llvmpipe_finish(pipe, __FUNCTION__);
1259
1260          for (i = 0; i < LP_MAX_SHADER_VARIANTS / 4; i++) {
1261             struct lp_fs_variant_list_item *item = last_elem(&lp->fs_variants_list);
1262             remove_shader_variant(lp, item->base);
1263          }
1264       }
1265       t0 = os_time_get();
1266
1267       variant = generate_variant(lp, shader, &key);
1268
1269       t1 = os_time_get();
1270       dt = t1 - t0;
1271       LP_COUNT_ADD(llvm_compile_time, dt);
1272       LP_COUNT_ADD(nr_llvm_compiles, 2);  /* emit vs. omit in/out test */
1273
1274       if (variant) {
1275          insert_at_head(&shader->variants, &variant->list_item_local);
1276          insert_at_head(&lp->fs_variants_list, &variant->list_item_global);
1277          lp->nr_fs_variants++;
1278          shader->variants_cached++;
1279       }
1280    }
1281
1282    lp_setup_set_fs_variant(lp->setup, variant);
1283 }
1284
1285
1286
1287 void
1288 llvmpipe_init_fs_funcs(struct llvmpipe_context *llvmpipe)
1289 {
1290    llvmpipe->pipe.create_fs_state = llvmpipe_create_fs_state;
1291    llvmpipe->pipe.bind_fs_state   = llvmpipe_bind_fs_state;
1292    llvmpipe->pipe.delete_fs_state = llvmpipe_delete_fs_state;
1293
1294    llvmpipe->pipe.set_constant_buffer = llvmpipe_set_constant_buffer;
1295 }