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          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) {
 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) {
 367          z = LLVMBuildLoad(builder, outputs[pos0][2], "z");
 368          lp_build_name(z, "output%u.%u.%c", i, pos0, "xyzw"[chan]);
 369       }
 370
 371       lp_build_depth_stencil_test(builder,
 372                                   &key->depth,
 373                                   key->stencil,
 374                                   type,
 375                                   zs_format_desc,
 376                                   &mask,
 377                                   stencil_refs,
 378                                   z,
 379                                   depth_ptr, facing,
 380                                   &zs_value,
 381                                   !simple_shader);
 382       /* Late Z write */
 383       if (depth_mode & LATE_DEPTH_WRITE) {
 384          lp_build_depth_write(builder, zs_format_desc, depth_ptr, zs_value);
 385       }
 386    }
 387    else if ((depth_mode & EARLY_DEPTH_TEST) &&
 388             (depth_mode & LATE_DEPTH_WRITE))
 389    {
 390       /* Need to apply a reduced mask to the depth write.  Reload the
 391        * depth value, update from zs_value with the new mask value and
 392        * write that out.
 393        */
 394       lp_build_deferred_depth_write(builder,
 395                                     type,
 396                                     zs_format_desc,
 397                                     &mask,
 398                                     depth_ptr,
 399                                     zs_value);
 400    }
 401
 402
 403    /* Color write  */
 404    for (attrib = 0; attrib < shader->info.base.num_outputs; ++attrib)
 405    {
 406       if (shader->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_COLOR)
 407       {
 408          unsigned cbuf = shader->info.base.output_semantic_index[attrib];
 409          for(chan = 0; chan < NUM_CHANNELS; ++chan)
 410          {
 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    if (counter)
 422       lp_build_occlusion_count(builder, type,
 423                                lp_build_mask_value(&mask), counter);
 424
 425    *pmask = lp_build_mask_end(&mask);
 426 }
 427
 428
 429 /**
 430  * Generate color blending and color output.
 431  * \param rt  the render target index (to index blend, colormask state)
 432  * \param type  the pixel color type
 433  * \param context_ptr  pointer to the runtime JIT context
 434  * \param mask  execution mask (active fragment/pixel mask)
 435  * \param src  colors from the fragment shader
 436  * \param dst_ptr  the destination color buffer pointer
 437  */
 438 static void
 439 generate_blend(const struct pipe_blend_state *blend,
 440                unsigned rt,
 441                LLVMBuilderRef builder,
 442                struct lp_type type,
 443                LLVMValueRef context_ptr,
 444                LLVMValueRef mask,
 445                LLVMValueRef *src,
 446                LLVMValueRef dst_ptr,
 447                boolean do_branch)
 448 {
 449    struct lp_build_context bld;
 450    struct lp_build_mask_context mask_ctx;
 451    LLVMTypeRef vec_type;
 452    LLVMValueRef const_ptr;
 453    LLVMValueRef con[4];
 454    LLVMValueRef dst[4];
 455    LLVMValueRef res[4];
 456    unsigned chan;
 457
 458    lp_build_context_init(&bld, builder, type);
 459
 460    lp_build_mask_begin(&mask_ctx, builder, type, mask);
 461    if (do_branch)
 462       lp_build_mask_check(&mask_ctx);
 463
 464    vec_type = lp_build_vec_type(type);
 465
 466    const_ptr = lp_jit_context_blend_color(builder, context_ptr);
 467    const_ptr = LLVMBuildBitCast(builder, const_ptr,
 468                                 LLVMPointerType(vec_type, 0), "");
 469
 470    /* load constant blend color and colors from the dest color buffer */
 471    for(chan = 0; chan < 4; ++chan) {
 472       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
 473       con[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, const_ptr, &index, 1, ""), "");
 474
 475       dst[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, dst_ptr, &index, 1, ""), "");
 476
 477       lp_build_name(con[chan], "con.%c", "rgba"[chan]);
 478       lp_build_name(dst[chan], "dst.%c", "rgba"[chan]);
 479    }
 480
 481    /* do blend */
 482    lp_build_blend_soa(builder, blend, type, rt, src, dst, con, res);
 483
 484    /* store results to color buffer */
 485    for(chan = 0; chan < 4; ++chan) {
 486       if(blend->rt[rt].colormask & (1 << chan)) {
 487          LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
 488          lp_build_name(res[chan], "res.%c", "rgba"[chan]);
 489          res[chan] = lp_build_select(&bld, mask, res[chan], dst[chan]);
 490          LLVMBuildStore(builder, res[chan], LLVMBuildGEP(builder, dst_ptr, &index, 1, ""));
 491       }
 492    }
 493
 494    lp_build_mask_end(&mask_ctx);
 495 }
 496
 497
 498 /**
 499  * Generate the runtime callable function for the whole fragment pipeline.
 500  * Note that the function which we generate operates on a block of 16
 501  * pixels at at time.  The block contains 2x2 quads.  Each quad contains
 502  * 2x2 pixels.
 503  */
 504 static void
 505 generate_fragment(struct llvmpipe_context *lp,
 506                   struct lp_fragment_shader *shader,
 507                   struct lp_fragment_shader_variant *variant,
 508                   unsigned partial_mask)
 509 {
 510    struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
 511    const struct lp_fragment_shader_variant_key *key = &variant->key;
 512    char func_name[256];
 513    struct lp_type fs_type;
 514    struct lp_type blend_type;
 515    LLVMTypeRef fs_elem_type;
 516    LLVMTypeRef fs_int_vec_type;
 517    LLVMTypeRef blend_vec_type;
 518    LLVMTypeRef arg_types[11];
 519    LLVMTypeRef func_type;
 520    LLVMValueRef context_ptr;
 521    LLVMValueRef x;
 522    LLVMValueRef y;
 523    LLVMValueRef a0_ptr;
 524    LLVMValueRef dadx_ptr;
 525    LLVMValueRef dady_ptr;
 526    LLVMValueRef color_ptr_ptr;
 527    LLVMValueRef depth_ptr;
 528    LLVMValueRef mask_input;
 529    LLVMValueRef counter = NULL;
 530    LLVMBasicBlockRef block;
 531    LLVMBuilderRef builder;
 532    struct lp_build_sampler_soa *sampler;
 533    struct lp_build_interp_soa_context interp;
 534    LLVMValueRef fs_mask[LP_MAX_VECTOR_LENGTH];
 535    LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS][LP_MAX_VECTOR_LENGTH];
 536    LLVMValueRef blend_mask;
 537    LLVMValueRef function;
 538    LLVMValueRef facing;
 539    const struct util_format_description *zs_format_desc;
 540    unsigned num_fs;
 541    unsigned i;
 542    unsigned chan;
 543    unsigned cbuf;
 544
 545
 546    /* TODO: actually pick these based on the fs and color buffer
 547     * characteristics. */
 548
 549    memset(&fs_type, 0, sizeof fs_type);
 550    fs_type.floating = TRUE; /* floating point values */
 551    fs_type.sign = TRUE;     /* values are signed */
 552    fs_type.norm = FALSE;    /* values are not limited to [0,1] or [-1,1] */
 553    fs_type.width = 32;      /* 32-bit float */
 554    fs_type.length = 4;      /* 4 elements per vector */
 555    num_fs = 4;              /* number of quads per block */
 556
 557    memset(&blend_type, 0, sizeof blend_type);
 558    blend_type.floating = FALSE; /* values are integers */
 559    blend_type.sign = FALSE;     /* values are unsigned */
 560    blend_type.norm = TRUE;      /* values are in [0,1] or [-1,1] */
 561    blend_type.width = 8;        /* 8-bit ubyte values */
 562    blend_type.length = 16;      /* 16 elements per vector */
 563
 564    /*
 565     * Generate the function prototype. Any change here must be reflected in
 566     * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
 567     */
 568
 569    fs_elem_type = lp_build_elem_type(fs_type);
 570    fs_int_vec_type = lp_build_int_vec_type(fs_type);
 571
 572    blend_vec_type = lp_build_vec_type(blend_type);
 573
 574    util_snprintf(func_name, sizeof(func_name), "fs%u_variant%u_%s",
 575                  shader->no, variant->no, partial_mask ? "partial" : "whole");
 576
 577    arg_types[0] = screen->context_ptr_type;            /* context */
 578    arg_types[1] = LLVMInt32Type();                     /* x */
 579    arg_types[2] = LLVMInt32Type();                     /* y */
 580    arg_types[3] = LLVMFloatType();                     /* facing */
 581    arg_types[4] = LLVMPointerType(fs_elem_type, 0);    /* a0 */
 582    arg_types[5] = LLVMPointerType(fs_elem_type, 0);    /* dadx */
 583    arg_types[6] = LLVMPointerType(fs_elem_type, 0);    /* dady */
 584    arg_types[7] = LLVMPointerType(LLVMPointerType(blend_vec_type, 0), 0);  /* color */
 585    arg_types[8] = LLVMPointerType(LLVMInt8Type(), 0);  /* depth */
 586    arg_types[9] = LLVMInt32Type();                     /* mask_input */
 587    arg_types[10] = LLVMPointerType(LLVMInt32Type(), 0);/* counter */
 588
 589    func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
 590
 591    function = LLVMAddFunction(screen->module, func_name, func_type);
 592    LLVMSetFunctionCallConv(function, LLVMCCallConv);
 593
 594    variant->function[partial_mask] = function;
 595
 596
 597    /* XXX: need to propagate noalias down into color param now we are
 598     * passing a pointer-to-pointer?
 599     */
 600    for(i = 0; i < Elements(arg_types); ++i)
 601       if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
 602          LLVMAddAttribute(LLVMGetParam(function, i), LLVMNoAliasAttribute);
 603
 604    context_ptr  = LLVMGetParam(function, 0);
 605    x            = LLVMGetParam(function, 1);
 606    y            = LLVMGetParam(function, 2);
 607    facing       = LLVMGetParam(function, 3);
 608    a0_ptr       = LLVMGetParam(function, 4);
 609    dadx_ptr     = LLVMGetParam(function, 5);
 610    dady_ptr     = LLVMGetParam(function, 6);
 611    color_ptr_ptr = LLVMGetParam(function, 7);
 612    depth_ptr    = LLVMGetParam(function, 8);
 613    mask_input   = LLVMGetParam(function, 9);
 614
 615    lp_build_name(context_ptr, "context");
 616    lp_build_name(x, "x");
 617    lp_build_name(y, "y");
 618    lp_build_name(a0_ptr, "a0");
 619    lp_build_name(dadx_ptr, "dadx");
 620    lp_build_name(dady_ptr, "dady");
 621    lp_build_name(color_ptr_ptr, "color_ptr_ptr");
 622    lp_build_name(depth_ptr, "depth");
 623    lp_build_name(mask_input, "mask_input");
 624
 625    if (key->occlusion_count) {
 626       counter = LLVMGetParam(function, 10);
 627       lp_build_name(counter, "counter");
 628    }
 629
 630    /*
 631     * Function body
 632     */
 633
 634    block = LLVMAppendBasicBlock(function, "entry");
 635    builder = LLVMCreateBuilder();
 636    LLVMPositionBuilderAtEnd(builder, block);
 637
 638    /*
 639     * The shader input interpolation info is not explicitely baked in the
 640     * shader key, but everything it derives from (TGSI, and flatshade) is
 641     * already included in the shader key.
 642     */
 643    lp_build_interp_soa_init(&interp,
 644                             lp->num_inputs,
 645                             lp->inputs,
 646                             builder, fs_type,
 647                             a0_ptr, dadx_ptr, dady_ptr,
 648                             x, y);
 649
 650    /* code generated texture sampling */
 651    sampler = lp_llvm_sampler_soa_create(key->sampler, context_ptr);
 652
 653    /* loop over quads in the block */
 654    zs_format_desc = util_format_description(key->zsbuf_format);
 655
 656    for(i = 0; i < num_fs; ++i) {
 657       LLVMValueRef depth_offset = LLVMConstInt(LLVMInt32Type(),
 658                                                i*fs_type.length*zs_format_desc->block.bits/8,
 659                                                0);
 660       LLVMValueRef out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS];
 661       LLVMValueRef depth_ptr_i;
 662
 663       depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &depth_offset, 1, "");
 664
 665       generate_fs(lp, shader, key,
 666                   builder,
 667                   fs_type,
 668                   context_ptr,
 669                   i,
 670                   &interp,
 671                   sampler,
 672                   &fs_mask[i], /* output */
 673                   out_color,
 674                   depth_ptr_i,
 675                   facing,
 676                   partial_mask,
 677                   mask_input,
 678                   counter);
 679
 680       for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++)
 681          for(chan = 0; chan < NUM_CHANNELS; ++chan)
 682             fs_out_color[cbuf][chan][i] = out_color[cbuf][chan];
 683    }
 684
 685    sampler->destroy(sampler);
 686
 687    /* Loop over color outputs / color buffers to do blending.
 688     */
 689    for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
 690       LLVMValueRef color_ptr;
 691       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), cbuf, 0);
 692       LLVMValueRef blend_in_color[NUM_CHANNELS];
 693       unsigned rt;
 694
 695       /*
 696        * Convert the fs's output color and mask to fit to the blending type.
 697        */
 698       for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 699          LLVMValueRef fs_color_vals[LP_MAX_VECTOR_LENGTH];
 700
 701          for (i = 0; i < num_fs; i++) {
 702             fs_color_vals[i] =
 703                LLVMBuildLoad(builder, fs_out_color[cbuf][chan][i], "fs_color_vals");
 704          }
 705
 706          lp_build_conv(builder, fs_type, blend_type,
 707                        fs_color_vals,
 708                        num_fs,
 709                        &blend_in_color[chan], 1);
 710
 711          lp_build_name(blend_in_color[chan], "color%d.%c", cbuf, "rgba"[chan]);
 712       }
 713
 714       if (partial_mask || !variant->opaque) {
 715          lp_build_conv_mask(builder, fs_type, blend_type,
 716                             fs_mask, num_fs,
 717                             &blend_mask, 1);
 718       } else {
 719          blend_mask = lp_build_const_int_vec(blend_type, ~0);
 720       }
 721
 722       color_ptr = LLVMBuildLoad(builder,
 723                                 LLVMBuildGEP(builder, color_ptr_ptr, &index, 1, ""),
 724                                 "");
 725       lp_build_name(color_ptr, "color_ptr%d", cbuf);
 726
 727       /* which blend/colormask state to use */
 728       rt = key->blend.independent_blend_enable ? cbuf : 0;
 729
 730       /*
 731        * Blending.
 732        */
 733       {
 734          /* Could the 4x4 have been killed?
 735           */
 736          boolean do_branch = ((key->depth.enabled || key->stencil[0].enabled) &&
 737                               !key->alpha.enabled &&
 738                               !shader->info.base.uses_kill);
 739
 740          generate_blend(&key->blend,
 741                         rt,
 742                         builder,
 743                         blend_type,
 744                         context_ptr,
 745                         blend_mask,
 746                         blend_in_color,
 747                         color_ptr,
 748                         do_branch);
 749       }
 750    }
 751
 752 #ifdef PIPE_ARCH_X86
 753    /* Avoid corrupting the FPU stack on 32bit OSes. */
 754    lp_build_intrinsic(builder, "llvm.x86.mmx.emms", LLVMVoidType(), NULL, 0);
 755 #endif
 756
 757    LLVMBuildRetVoid(builder);
 758
 759    LLVMDisposeBuilder(builder);
 760
 761
 762    /* Verify the LLVM IR.  If invalid, dump and abort */
 763 #ifdef DEBUG
 764    if(LLVMVerifyFunction(function, LLVMPrintMessageAction)) {
 765       if (1)
 766          lp_debug_dump_value(function);
 767       abort();
 768    }
 769 #endif
 770
 771    /* Apply optimizations to LLVM IR */
 772    LLVMRunFunctionPassManager(screen->pass, function);
 773
 774    if ((gallivm_debug & GALLIVM_DEBUG_IR) || (LP_DEBUG & DEBUG_FS)) {
 775       /* Print the LLVM IR to stderr */
 776       lp_debug_dump_value(function);
 777       debug_printf("\n");
 778    }
 779
 780    /*
 781     * Translate the LLVM IR into machine code.
 782     */
 783    {
 784       void *f = LLVMGetPointerToGlobal(screen->engine, function);
 785
 786       variant->jit_function[partial_mask] = (lp_jit_frag_func)pointer_to_func(f);
 787
 788       if ((gallivm_debug & GALLIVM_DEBUG_ASM) || (LP_DEBUG & DEBUG_FS)) {
 789          lp_disassemble(f);
 790       }
 791       lp_func_delete_body(function);
 792    }
 793 }
 794
 795
 796 static void
 797 dump_fs_variant_key(const struct lp_fragment_shader_variant_key *key)
 798 {
 799    unsigned i;
 800
 801    debug_printf("fs variant %p:\n", (void *) key);
 802
 803    if (key->flatshade) {
 804       debug_printf("flatshade = 1\n");
 805    }
 806    for (i = 0; i < key->nr_cbufs; ++i) {
 807       debug_printf("cbuf_format[%u] = %s\n", i, util_format_name(key->cbuf_format[i]));
 808    }
 809    if (key->depth.enabled) {
 810       debug_printf("depth.format = %s\n", util_format_name(key->zsbuf_format));
 811       debug_printf("depth.func = %s\n", util_dump_func(key->depth.func, TRUE));
 812       debug_printf("depth.writemask = %u\n", key->depth.writemask);
 813    }
 814
 815    for (i = 0; i < 2; ++i) {
 816       if (key->stencil[i].enabled) {
 817          debug_printf("stencil[%u].func = %s\n", i, util_dump_func(key->stencil[i].func, TRUE));
 818          debug_printf("stencil[%u].fail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].fail_op, TRUE));
 819          debug_printf("stencil[%u].zpass_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zpass_op, TRUE));
 820          debug_printf("stencil[%u].zfail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zfail_op, TRUE));
 821          debug_printf("stencil[%u].valuemask = 0x%x\n", i, key->stencil[i].valuemask);
 822          debug_printf("stencil[%u].writemask = 0x%x\n", i, key->stencil[i].writemask);
 823       }
 824    }
 825
 826    if (key->alpha.enabled) {
 827       debug_printf("alpha.func = %s\n", util_dump_func(key->alpha.func, TRUE));
 828    }
 829
 830    if (key->occlusion_count) {
 831       debug_printf("occlusion_count = 1\n");
 832    }
 833
 834    if (key->blend.logicop_enable) {
 835       debug_printf("blend.logicop_func = %s\n", util_dump_logicop(key->blend.logicop_func, TRUE));
 836    }
 837    else if (key->blend.rt[0].blend_enable) {
 838       debug_printf("blend.rgb_func = %s\n",   util_dump_blend_func  (key->blend.rt[0].rgb_func, TRUE));
 839       debug_printf("blend.rgb_src_factor = %s\n",   util_dump_blend_factor(key->blend.rt[0].rgb_src_factor, TRUE));
 840       debug_printf("blend.rgb_dst_factor = %s\n",   util_dump_blend_factor(key->blend.rt[0].rgb_dst_factor, TRUE));
 841       debug_printf("blend.alpha_func = %s\n",       util_dump_blend_func  (key->blend.rt[0].alpha_func, TRUE));
 842       debug_printf("blend.alpha_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_src_factor, TRUE));
 843       debug_printf("blend.alpha_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_dst_factor, TRUE));
 844    }
 845    debug_printf("blend.colormask = 0x%x\n", key->blend.rt[0].colormask);
 846    for (i = 0; i < key->nr_samplers; ++i) {
 847       debug_printf("sampler[%u] = \n", i);
 848       debug_printf("  .format = %s\n",
 849                    util_format_name(key->sampler[i].format));
 850       debug_printf("  .target = %s\n",
 851                    util_dump_tex_target(key->sampler[i].target, TRUE));
 852       debug_printf("  .pot = %u %u %u\n",
 853                    key->sampler[i].pot_width,
 854                    key->sampler[i].pot_height,
 855                    key->sampler[i].pot_depth);
 856       debug_printf("  .wrap = %s %s %s\n",
 857                    util_dump_tex_wrap(key->sampler[i].wrap_s, TRUE),
 858                    util_dump_tex_wrap(key->sampler[i].wrap_t, TRUE),
 859                    util_dump_tex_wrap(key->sampler[i].wrap_r, TRUE));
 860       debug_printf("  .min_img_filter = %s\n",
 861                    util_dump_tex_filter(key->sampler[i].min_img_filter, TRUE));
 862       debug_printf("  .min_mip_filter = %s\n",
 863                    util_dump_tex_mipfilter(key->sampler[i].min_mip_filter, TRUE));
 864       debug_printf("  .mag_img_filter = %s\n",
 865                    util_dump_tex_filter(key->sampler[i].mag_img_filter, TRUE));
 866       if (key->sampler[i].compare_mode != PIPE_TEX_COMPARE_NONE)
 867          debug_printf("  .compare_func = %s\n", util_dump_func(key->sampler[i].compare_func, TRUE));
 868       debug_printf("  .normalized_coords = %u\n", key->sampler[i].normalized_coords);
 869       debug_printf("  .min_max_lod_equal = %u\n", key->sampler[i].min_max_lod_equal);
 870       debug_printf("  .lod_bias_non_zero = %u\n", key->sampler[i].lod_bias_non_zero);
 871       debug_printf("  .apply_min_lod = %u\n", key->sampler[i].apply_min_lod);
 872       debug_printf("  .apply_max_lod = %u\n", key->sampler[i].apply_max_lod);
 873    }
 874 }
 875
 876
 877 void
 878 lp_debug_fs_variant(const struct lp_fragment_shader_variant *variant)
 879 {
 880    debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
 881                 variant->shader->no, variant->no);
 882    tgsi_dump(variant->shader->base.tokens, 0);
 883    dump_fs_variant_key(&variant->key);
 884    debug_printf("variant->opaque = %u\n", variant->opaque);
 885    debug_printf("\n");
 886 }
 887
 888 static struct lp_fragment_shader_variant *
 889 generate_variant(struct llvmpipe_context *lp,
 890                  struct lp_fragment_shader *shader,
 891                  const struct lp_fragment_shader_variant_key *key)
 892 {
 893    struct lp_fragment_shader_variant *variant;
 894    boolean fullcolormask;
 895
 896    variant = CALLOC_STRUCT(lp_fragment_shader_variant);
 897    if(!variant)
 898       return NULL;
 899
 900    variant->shader = shader;
 901    variant->list_item_global.base = variant;
 902    variant->list_item_local.base = variant;
 903    variant->no = shader->variants_created++;
 904
 905    memcpy(&variant->key, key, shader->variant_key_size);
 906
 907    /*
 908     * Determine whether we are touching all channels in the color buffer.
 909     */
 910    fullcolormask = FALSE;
 911    if (key->nr_cbufs == 1) {
 912       const struct util_format_description *format_desc;
 913       format_desc = util_format_description(key->cbuf_format[0]);
 914       if ((~key->blend.rt[0].colormask &
 915            util_format_colormask(format_desc)) == 0) {
 916          fullcolormask = TRUE;
 917       }
 918    }
 919
 920    variant->opaque =
 921          !key->blend.logicop_enable &&
 922          !key->blend.rt[0].blend_enable &&
 923          fullcolormask &&
 924          !key->stencil[0].enabled &&
 925          !key->alpha.enabled &&
 926          !key->depth.enabled &&
 927          !shader->info.base.uses_kill
 928          ? TRUE : FALSE;
 929
 930
 931    if ((LP_DEBUG & DEBUG_FS) || (gallivm_debug & GALLIVM_DEBUG_IR)) {
 932       lp_debug_fs_variant(variant);
 933    }
 934
 935    generate_fragment(lp, shader, variant, RAST_EDGE_TEST);
 936
 937    if (variant->opaque) {
 938       /* Specialized shader, which doesn't need to read the color buffer. */
 939       generate_fragment(lp, shader, variant, RAST_WHOLE);
 940    } else {
 941       variant->jit_function[RAST_WHOLE] = variant->jit_function[RAST_EDGE_TEST];
 942    }
 943
 944    return variant;
 945 }
 946
 947
 948 static void *
 949 llvmpipe_create_fs_state(struct pipe_context *pipe,
 950                          const struct pipe_shader_state *templ)
 951 {
 952    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
 953    struct lp_fragment_shader *shader;
 954    int nr_samplers;
 955
 956    shader = CALLOC_STRUCT(lp_fragment_shader);
 957    if (!shader)
 958       return NULL;
 959
 960    shader->no = fs_no++;
 961    make_empty_list(&shader->variants);
 962
 963    /* get/save the summary info for this shader */
 964    lp_build_tgsi_info(templ->tokens, &shader->info);
 965
 966    /* we need to keep a local copy of the tokens */
 967    shader->base.tokens = tgsi_dup_tokens(templ->tokens);
 968
 969    shader->draw_data = draw_create_fragment_shader(llvmpipe->draw, templ);
 970    if (shader->draw_data == NULL) {
 971       FREE((void *) shader->base.tokens);
 972       FREE(shader);
 973       return NULL;
 974    }
 975
 976    nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
 977
 978    shader->variant_key_size = Offset(struct lp_fragment_shader_variant_key,
 979                                      sampler[nr_samplers]);
 980
 981    if (LP_DEBUG & DEBUG_TGSI) {
 982       unsigned attrib;
 983       debug_printf("llvmpipe: Create fragment shader #%u %p:\n", shader->no, (void *) shader);
 984       tgsi_dump(templ->tokens, 0);
 985       debug_printf("usage masks:\n");
 986       for (attrib = 0; attrib < shader->info.base.num_inputs; ++attrib) {
 987          unsigned usage_mask = shader->info.base.input_usage_mask[attrib];
 988          debug_printf("  IN[%u].%s%s%s%s\n",
 989                       attrib,
 990                       usage_mask & TGSI_WRITEMASK_X ? "x" : "",
 991                       usage_mask & TGSI_WRITEMASK_Y ? "y" : "",
 992                       usage_mask & TGSI_WRITEMASK_Z ? "z" : "",
 993                       usage_mask & TGSI_WRITEMASK_W ? "w" : "");
 994       }
 995       debug_printf("\n");
 996    }
 997
 998    return shader;
 999 }
1000
1001
1002 static void
1003 llvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs)
1004 {
1005    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
1006
1007    if (llvmpipe->fs == fs)
1008       return;
1009
1010    draw_flush(llvmpipe->draw);
1011
1012    draw_bind_fragment_shader(llvmpipe->draw,
1013                              (llvmpipe->fs ? llvmpipe->fs->draw_data : NULL));
1014
1015    llvmpipe->fs = fs;
1016
1017    llvmpipe->dirty |= LP_NEW_FS;
1018 }
1019
1020 static void
1021 remove_shader_variant(struct llvmpipe_context *lp,
1022                       struct lp_fragment_shader_variant *variant)
1023 {
1024    struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
1025    unsigned i;
1026
1027    if (gallivm_debug & GALLIVM_DEBUG_IR) {
1028       debug_printf("llvmpipe: del fs #%u var #%u v created #%u v cached #%u v total cached #%u\n",
1029                     variant->shader->no, variant->no, variant->shader->variants_created,
1030                     variant->shader->variants_cached, lp->nr_fs_variants);
1031    }
1032    for (i = 0; i < Elements(variant->function); i++) {
1033       if (variant->function[i]) {
1034          if (variant->jit_function[i])
1035             LLVMFreeMachineCodeForFunction(screen->engine,
1036                                            variant->function[i]);
1037          LLVMDeleteFunction(variant->function[i]);
1038       }
1039    }
1040    remove_from_list(&variant->list_item_local);
1041    variant->shader->variants_cached--;
1042    remove_from_list(&variant->list_item_global);
1043    lp->nr_fs_variants--;
1044    FREE(variant);
1045 }
1046
1047 static void
1048 llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs)
1049 {
1050    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
1051    struct lp_fragment_shader *shader = fs;
1052    struct lp_fs_variant_list_item *li;
1053
1054    assert(fs != llvmpipe->fs);
1055    (void) llvmpipe;
1056
1057    /*
1058     * XXX: we need to flush the context until we have some sort of reference
1059     * counting in fragment shaders as they may still be binned
1060     * Flushing alone might not sufficient we need to wait on it too.
1061     */
1062
1063    llvmpipe_finish(pipe, __FUNCTION__);
1064
1065    li = first_elem(&shader->variants);
1066    while(!at_end(&shader->variants, li)) {
1067       struct lp_fs_variant_list_item *next = next_elem(li);
1068       remove_shader_variant(llvmpipe, li->base);
1069       li = next;
1070    }
1071
1072    draw_delete_fragment_shader(llvmpipe->draw, shader->draw_data);
1073
1074    assert(shader->variants_cached == 0);
1075    FREE((void *) shader->base.tokens);
1076    FREE(shader);
1077 }
1078
1079
1080
1081 static void
1082 llvmpipe_set_constant_buffer(struct pipe_context *pipe,
1083                              uint shader, uint index,
1084                              struct pipe_resource *constants)
1085 {
1086    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
1087    unsigned size = constants ? constants->width0 : 0;
1088    const void *data = constants ? llvmpipe_resource_data(constants) : NULL;
1089
1090    assert(shader < PIPE_SHADER_TYPES);
1091    assert(index < PIPE_MAX_CONSTANT_BUFFERS);
1092
1093    if(llvmpipe->constants[shader][index] == constants)
1094       return;
1095
1096    draw_flush(llvmpipe->draw);
1097
1098    /* note: reference counting */
1099    pipe_resource_reference(&llvmpipe->constants[shader][index], constants);
1100
1101    if(shader == PIPE_SHADER_VERTEX ||
1102       shader == PIPE_SHADER_GEOMETRY) {
1103       draw_set_mapped_constant_buffer(llvmpipe->draw, shader,
1104                                       index, data, size);
1105    }
1106
1107    llvmpipe->dirty |= LP_NEW_CONSTANTS;
1108 }
1109
1110
1111 /**
1112  * Return the blend factor equivalent to a destination alpha of one.
1113  */
1114 static INLINE unsigned
1115 force_dst_alpha_one(unsigned factor)
1116 {
1117    switch(factor) {
1118    case PIPE_BLENDFACTOR_DST_ALPHA:
1119       return PIPE_BLENDFACTOR_ONE;
1120    case PIPE_BLENDFACTOR_INV_DST_ALPHA:
1121       return PIPE_BLENDFACTOR_ZERO;
1122    case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
1123       return PIPE_BLENDFACTOR_ZERO;
1124    }
1125
1126    return factor;
1127 }
1128
1129
1130 /**
1131  * We need to generate several variants of the fragment pipeline to match
1132  * all the combinations of the contributing state atoms.
1133  *
1134  * TODO: there is actually no reason to tie this to context state -- the
1135  * generated code could be cached globally in the screen.
1136  */
1137 static void
1138 make_variant_key(struct llvmpipe_context *lp,
1139                  struct lp_fragment_shader *shader,
1140                  struct lp_fragment_shader_variant_key *key)
1141 {
1142    unsigned i;
1143
1144    memset(key, 0, shader->variant_key_size);
1145
1146    if (lp->framebuffer.zsbuf) {
1147       if (lp->depth_stencil->depth.enabled) {
1148          key->zsbuf_format = lp->framebuffer.zsbuf->format;
1149          memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth);
1150       }
1151       if (lp->depth_stencil->stencil[0].enabled) {
1152          key->zsbuf_format = lp->framebuffer.zsbuf->format;
1153          memcpy(&key->stencil, &lp->depth_stencil->stencil, sizeof key->stencil);
1154       }
1155    }
1156
1157    key->alpha.enabled = lp->depth_stencil->alpha.enabled;
1158    if(key->alpha.enabled)
1159       key->alpha.func = lp->depth_stencil->alpha.func;
1160    /* alpha.ref_value is passed in jit_context */
1161
1162    key->flatshade = lp->rasterizer->flatshade;
1163    if (lp->active_query_count) {
1164       key->occlusion_count = TRUE;
1165    }
1166
1167    if (lp->framebuffer.nr_cbufs) {
1168       memcpy(&key->blend, lp->blend, sizeof key->blend);
1169    }
1170
1171    key->nr_cbufs = lp->framebuffer.nr_cbufs;
1172    for (i = 0; i < lp->framebuffer.nr_cbufs; i++) {
1173       enum pipe_format format = lp->framebuffer.cbufs[i]->format;
1174       struct pipe_rt_blend_state *blend_rt = &key->blend.rt[i];
1175       const struct util_format_description *format_desc;
1176
1177       key->cbuf_format[i] = format;
1178
1179       format_desc = util_format_description(format);
1180       assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
1181              format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB);
1182
1183       blend_rt->colormask = lp->blend->rt[i].colormask;
1184
1185       /*
1186        * Mask out color channels not present in the color buffer.
1187        */
1188       blend_rt->colormask &= util_format_colormask(format_desc);
1189
1190       /*
1191        * Our swizzled render tiles always have an alpha channel, but the linear
1192        * render target format often does not, so force here the dst alpha to be
1193        * one.
1194        *
1195        * This is not a mere optimization. Wrong results will be produced if the
1196        * dst alpha is used, the dst format does not have alpha, and the previous
1197        * rendering was not flushed from the swizzled to linear buffer. For
1198        * example, NonPowTwo DCT.
1199        *
1200        * TODO: This should be generalized to all channels for better
1201        * performance, but only alpha causes correctness issues.
1202        *
1203        * Also, force rgb/alpha func/factors match, to make AoS blending easier.
1204        */
1205       if (format_desc->swizzle[3] > UTIL_FORMAT_SWIZZLE_W) {
1206          blend_rt->rgb_src_factor   = force_dst_alpha_one(blend_rt->rgb_src_factor);
1207          blend_rt->rgb_dst_factor   = force_dst_alpha_one(blend_rt->rgb_dst_factor);
1208          blend_rt->alpha_func       = blend_rt->rgb_func;
1209          blend_rt->alpha_src_factor = blend_rt->rgb_src_factor;
1210          blend_rt->alpha_dst_factor = blend_rt->rgb_dst_factor;
1211       }
1212    }
1213
1214    /* This value will be the same for all the variants of a given shader:
1215     */
1216    key->nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
1217
1218    for(i = 0; i < key->nr_samplers; ++i) {
1219       if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
1220          lp_sampler_static_state(&key->sampler[i],
1221                                  lp->fragment_sampler_views[i],
1222                                  lp->sampler[i]);
1223       }
1224    }
1225 }
1226
1227 /**
1228  * Update fragment state.  This is called just prior to drawing
1229  * something when some fragment-related state has changed.
1230  */
1231 void
1232 llvmpipe_update_fs(struct llvmpipe_context *lp)
1233 {
1234    struct lp_fragment_shader *shader = lp->fs;
1235    struct lp_fragment_shader_variant_key key;
1236    struct lp_fragment_shader_variant *variant = NULL;
1237    struct lp_fs_variant_list_item *li;
1238
1239    make_variant_key(lp, shader, &key);
1240
1241    li = first_elem(&shader->variants);
1242    while(!at_end(&shader->variants, li)) {
1243       if(memcmp(&li->base->key, &key, shader->variant_key_size) == 0) {
1244          variant = li->base;
1245          break;
1246       }
1247       li = next_elem(li);
1248    }
1249
1250    if (variant) {
1251       move_to_head(&lp->fs_variants_list, &variant->list_item_global);
1252    }
1253    else {
1254       int64_t t0, t1;
1255       int64_t dt;
1256       unsigned i;
1257       if (lp->nr_fs_variants >= LP_MAX_SHADER_VARIANTS) {
1258          struct pipe_context *pipe = &lp->pipe;
1259
1260          /*
1261           * XXX: we need to flush the context until we have some sort of reference
1262           * counting in fragment shaders as they may still be binned
1263           * Flushing alone might not be sufficient we need to wait on it too.
1264           */
1265          llvmpipe_finish(pipe, __FUNCTION__);
1266
1267          for (i = 0; i < LP_MAX_SHADER_VARIANTS / 4; i++) {
1268             struct lp_fs_variant_list_item *item = last_elem(&lp->fs_variants_list);
1269             remove_shader_variant(lp, item->base);
1270          }
1271       }
1272       t0 = os_time_get();
1273
1274       variant = generate_variant(lp, shader, &key);
1275
1276       t1 = os_time_get();
1277       dt = t1 - t0;
1278       LP_COUNT_ADD(llvm_compile_time, dt);
1279       LP_COUNT_ADD(nr_llvm_compiles, 2);  /* emit vs. omit in/out test */
1280
1281       if (variant) {
1282          insert_at_head(&shader->variants, &variant->list_item_local);
1283          insert_at_head(&lp->fs_variants_list, &variant->list_item_global);
1284          lp->nr_fs_variants++;
1285          shader->variants_cached++;
1286       }
1287    }
1288
1289    lp_setup_set_fs_variant(lp->setup, variant);
1290 }
1291
1292
1293
1294 void
1295 llvmpipe_init_fs_funcs(struct llvmpipe_context *llvmpipe)
1296 {
1297    llvmpipe->pipe.create_fs_state = llvmpipe_create_fs_state;
1298    llvmpipe->pipe.bind_fs_state   = llvmpipe_bind_fs_state;
1299    llvmpipe->pipe.delete_fs_state = llvmpipe_delete_fs_state;
1300
1301    llvmpipe->pipe.set_constant_buffer = llvmpipe_set_constant_buffer;
1302 }