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  * - triangle edge in/out testing
  35  * - scissor test
  36  * - stipple (TBI)
  37  * - early depth test
  38  * - fragment shader
  39  * - alpha test
  40  * - depth/stencil test
  41  * - blending
  42  *
  43  * This file has only the glue to assemble the fragment pipeline.  The actual
  44  * plumbing of converting Gallium state into LLVM IR is done elsewhere, in the
  45  * lp_bld_*.[ch] files, and in a complete generic and reusable way. Here we
  46  * muster the LLVM JIT execution engine to create a function that follows an
  47  * established binary interface and that can be called from C directly.
  48  *
  49  * A big source of complexity here is that we often want to run different
  50  * stages with different precisions and data types and precisions. For example,
  51  * the fragment shader needs typically to be done in floats, but the
  52  * depth/stencil test and blending is better done in the type that most closely
  53  * matches the depth/stencil and color buffer respectively.
  54  *
  55  * Since the width of a SIMD vector register stays the same regardless of the
  56  * element type, different types imply different number of elements, so we must
  57  * code generate more instances of the stages with larger types to be able to
  58  * feed/consume the stages with smaller types.
  59  *
  60  * @author Jose Fonseca <jfonseca@vmware.com>
  61  */
  62
  63 #include <limits.h>
  64 #include "pipe/p_defines.h"
  65 #include "util/u_inlines.h"
  66 #include "util/u_memory.h"
  67 #include "util/u_format.h"
  68 #include "util/u_dump.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_intr.h"
  79 #include "gallivm/lp_bld_logic.h"
  80 #include "gallivm/lp_bld_tgsi.h"
  81 #include "gallivm/lp_bld_swizzle.h"
  82 #include "gallivm/lp_bld_flow.h"
  83 #include "gallivm/lp_bld_debug.h"
  84
  85 #include "lp_bld_alpha.h"
  86 #include "lp_bld_blend.h"
  87 #include "lp_bld_depth.h"
  88 #include "lp_bld_interp.h"
  89 #include "lp_context.h"
  90 #include "lp_debug.h"
  91 #include "lp_perf.h"
  92 #include "lp_screen.h"
  93 #include "lp_setup.h"
  94 #include "lp_state.h"
  95 #include "lp_tex_sample.h"
  96
  97
  98 #include <llvm-c/Analysis.h>
  99
 100
 101 static const unsigned char quad_offset_x[4] = {0, 1, 0, 1};
 102 static const unsigned char quad_offset_y[4] = {0, 0, 1, 1};
 103
 104
 105 /*
 106  * Derive from the quad's upper left scalar coordinates the coordinates for
 107  * all other quad pixels
 108  */
 109 static void
 110 generate_pos0(LLVMBuilderRef builder,
 111               LLVMValueRef x,
 112               LLVMValueRef y,
 113               LLVMValueRef *x0,
 114               LLVMValueRef *y0)
 115 {
 116    LLVMTypeRef int_elem_type = LLVMInt32Type();
 117    LLVMTypeRef int_vec_type = LLVMVectorType(int_elem_type, QUAD_SIZE);
 118    LLVMTypeRef elem_type = LLVMFloatType();
 119    LLVMTypeRef vec_type = LLVMVectorType(elem_type, QUAD_SIZE);
 120    LLVMValueRef x_offsets[QUAD_SIZE];
 121    LLVMValueRef y_offsets[QUAD_SIZE];
 122    unsigned i;
 123
 124    x = lp_build_broadcast(builder, int_vec_type, x);
 125    y = lp_build_broadcast(builder, int_vec_type, y);
 126
 127    for(i = 0; i < QUAD_SIZE; ++i) {
 128       x_offsets[i] = LLVMConstInt(int_elem_type, quad_offset_x[i], 0);
 129       y_offsets[i] = LLVMConstInt(int_elem_type, quad_offset_y[i], 0);
 130    }
 131
 132    x = LLVMBuildAdd(builder, x, LLVMConstVector(x_offsets, QUAD_SIZE), "");
 133    y = LLVMBuildAdd(builder, y, LLVMConstVector(y_offsets, QUAD_SIZE), "");
 134
 135    *x0 = LLVMBuildSIToFP(builder, x, vec_type, "");
 136    *y0 = LLVMBuildSIToFP(builder, y, vec_type, "");
 137 }
 138
 139
 140 /**
 141  * Generate the depth /stencil test code.
 142  */
 143 static void
 144 generate_depth_stencil(LLVMBuilderRef builder,
 145                        const struct lp_fragment_shader_variant_key *key,
 146                        struct lp_type src_type,
 147                        struct lp_build_mask_context *mask,
 148                        LLVMValueRef stencil_refs[2],
 149                        LLVMValueRef src,
 150                        LLVMValueRef dst_ptr,
 151                        LLVMValueRef facing,
 152                        LLVMValueRef counter)
 153 {
 154    const struct util_format_description *format_desc;
 155    struct lp_type dst_type;
 156
 157    if (!key->depth.enabled && !key->stencil[0].enabled && !key->stencil[1].enabled)
 158       return;
 159
 160    format_desc = util_format_description(key->zsbuf_format);
 161    assert(format_desc);
 162
 163    /*
 164     * Depths are expected to be between 0 and 1, even if they are stored in
 165     * floats. Setting these bits here will ensure that the lp_build_conv() call
 166     * below won't try to unnecessarily clamp the incoming values.
 167     */
 168    if(src_type.floating) {
 169       src_type.sign = FALSE;
 170       src_type.norm = TRUE;
 171    }
 172    else {
 173       assert(!src_type.sign);
 174       assert(src_type.norm);
 175    }
 176
 177    /* Pick the depth type. */
 178    dst_type = lp_depth_type(format_desc, src_type.width*src_type.length);
 179
 180    /* FIXME: Cope with a depth test type with a different bit width. */
 181    assert(dst_type.width == src_type.width);
 182    assert(dst_type.length == src_type.length);
 183
 184    /* Convert fragment Z from float to integer */
 185    lp_build_conv(builder, src_type, dst_type, &src, 1, &src, 1);
 186
 187    dst_ptr = LLVMBuildBitCast(builder,
 188                               dst_ptr,
 189                               LLVMPointerType(lp_build_vec_type(dst_type), 0), "");
 190    lp_build_depth_stencil_test(builder,
 191                                &key->depth,
 192                                key->stencil,
 193                                dst_type,
 194                                format_desc,
 195                                mask,
 196                                stencil_refs,
 197                                src,
 198                                dst_ptr,
 199                                facing,
 200                                counter);
 201 }
 202
 203
 204 /**
 205  * Generate the code to do inside/outside triangle testing for the
 206  * four pixels in a 2x2 quad.  This will set the four elements of the
 207  * quad mask vector to 0 or ~0.
 208  * \param i  which quad of the quad group to test, in [0,3]
 209  */
 210 static void
 211 generate_tri_edge_mask(LLVMBuilderRef builder,
 212                        unsigned i,
 213                        LLVMValueRef *mask,      /* ivec4, out */
 214                        LLVMValueRef c0,         /* int32 */
 215                        LLVMValueRef c1,         /* int32 */
 216                        LLVMValueRef c2,         /* int32 */
 217                        LLVMValueRef step0_ptr,  /* ivec4 */
 218                        LLVMValueRef step1_ptr,  /* ivec4 */
 219                        LLVMValueRef step2_ptr)  /* ivec4 */
 220 {
 221 #define OPTIMIZE_IN_OUT_TEST 0
 222 #if OPTIMIZE_IN_OUT_TEST
 223    struct lp_build_if_state ifctx;
 224    LLVMValueRef not_draw_all;
 225 #endif
 226    struct lp_build_flow_context *flow;
 227    struct lp_type i32_type;
 228    LLVMTypeRef i32vec4_type;
 229    LLVMValueRef c0_vec, c1_vec, c2_vec;
 230    LLVMValueRef in_out_mask;
 231
 232    assert(i < 4);
 233
 234    /* int32 vector type */
 235    memset(&i32_type, 0, sizeof i32_type);
 236    i32_type.floating = FALSE; /* values are integers */
 237    i32_type.sign = TRUE;      /* values are signed */
 238    i32_type.norm = FALSE;     /* values are not normalized */
 239    i32_type.width = 32;       /* 32-bit int values */
 240    i32_type.length = 4;       /* 4 elements per vector */
 241
 242    i32vec4_type = lp_build_int32_vec4_type();
 243
 244    /*
 245     * Use a conditional here to do detailed pixel in/out testing.
 246     * We only have to do this if c0 != INT_MIN.
 247     */
 248    flow = lp_build_flow_create(builder);
 249    lp_build_flow_scope_begin(flow);
 250
 251    {
 252 #if OPTIMIZE_IN_OUT_TEST
 253       /* not_draw_all = (c0 != INT_MIN) */
 254       not_draw_all = LLVMBuildICmp(builder,
 255                                    LLVMIntNE,
 256                                    c0,
 257                                    LLVMConstInt(LLVMInt32Type(), INT_MIN, 0),
 258                                    "");
 259
 260       in_out_mask = lp_build_const_int_vec(i32_type, ~0);
 261
 262
 263       lp_build_flow_scope_declare(flow, &in_out_mask);
 264
 265       /* if (not_draw_all) {... */
 266       lp_build_if(&ifctx, flow, builder, not_draw_all);
 267 #endif
 268       {
 269          LLVMValueRef step0_vec, step1_vec, step2_vec;
 270          LLVMValueRef m0_vec, m1_vec, m2_vec;
 271          LLVMValueRef index, m;
 272
 273          /* c0_vec = {c0, c0, c0, c0}
 274           * Note that we emit this code four times but LLVM optimizes away
 275           * three instances of it.
 276           */
 277          c0_vec = lp_build_broadcast(builder, i32vec4_type, c0);
 278          c1_vec = lp_build_broadcast(builder, i32vec4_type, c1);
 279          c2_vec = lp_build_broadcast(builder, i32vec4_type, c2);
 280          lp_build_name(c0_vec, "edgeconst0vec");
 281          lp_build_name(c1_vec, "edgeconst1vec");
 282          lp_build_name(c2_vec, "edgeconst2vec");
 283
 284          /* load step0vec, step1, step2 vec from memory */
 285          index = LLVMConstInt(LLVMInt32Type(), i, 0);
 286          step0_vec = LLVMBuildLoad(builder, LLVMBuildGEP(builder, step0_ptr, &index, 1, ""), "");
 287          step1_vec = LLVMBuildLoad(builder, LLVMBuildGEP(builder, step1_ptr, &index, 1, ""), "");
 288          step2_vec = LLVMBuildLoad(builder, LLVMBuildGEP(builder, step2_ptr, &index, 1, ""), "");
 289          lp_build_name(step0_vec, "step0vec");
 290          lp_build_name(step1_vec, "step1vec");
 291          lp_build_name(step2_vec, "step2vec");
 292
 293          /* m0_vec = step0_ptr[i] > c0_vec */
 294          m0_vec = lp_build_compare(builder, i32_type, PIPE_FUNC_GREATER, step0_vec, c0_vec);
 295          m1_vec = lp_build_compare(builder, i32_type, PIPE_FUNC_GREATER, step1_vec, c1_vec);
 296          m2_vec = lp_build_compare(builder, i32_type, PIPE_FUNC_GREATER, step2_vec, c2_vec);
 297
 298          /* in_out_mask = m0_vec & m1_vec & m2_vec */
 299          m = LLVMBuildAnd(builder, m0_vec, m1_vec, "");
 300          in_out_mask = LLVMBuildAnd(builder, m, m2_vec, "");
 301          lp_build_name(in_out_mask, "inoutmaskvec");
 302       }
 303 #if OPTIMIZE_IN_OUT_TEST
 304       lp_build_endif(&ifctx);
 305 #endif
 306
 307    }
 308    lp_build_flow_scope_end(flow);
 309    lp_build_flow_destroy(flow);
 310
 311    /* This is the initial alive/dead pixel mask for a quad of four pixels.
 312     * It's an int[4] vector with each word set to 0 or ~0.
 313     * Words will get cleared when pixels faile the Z test, etc.
 314     */
 315    *mask = in_out_mask;
 316 }
 317
 318
 319 static LLVMValueRef
 320 generate_scissor_test(LLVMBuilderRef builder,
 321                       LLVMValueRef context_ptr,
 322                       const struct lp_build_interp_soa_context *interp,
 323                       struct lp_type type)
 324 {
 325    LLVMTypeRef vec_type = lp_build_vec_type(type);
 326    LLVMValueRef xpos = interp->pos[0], ypos = interp->pos[1];
 327    LLVMValueRef xmin, ymin, xmax, ymax;
 328    LLVMValueRef m0, m1, m2, m3, m;
 329
 330    /* xpos, ypos contain the window coords for the four pixels in the quad */
 331    assert(xpos);
 332    assert(ypos);
 333
 334    /* get the current scissor bounds, convert to vectors */
 335    xmin = lp_jit_context_scissor_xmin_value(builder, context_ptr);
 336    xmin = lp_build_broadcast(builder, vec_type, xmin);
 337
 338    ymin = lp_jit_context_scissor_ymin_value(builder, context_ptr);
 339    ymin = lp_build_broadcast(builder, vec_type, ymin);
 340
 341    xmax = lp_jit_context_scissor_xmax_value(builder, context_ptr);
 342    xmax = lp_build_broadcast(builder, vec_type, xmax);
 343
 344    ymax = lp_jit_context_scissor_ymax_value(builder, context_ptr);
 345    ymax = lp_build_broadcast(builder, vec_type, ymax);
 346
 347    /* compare the fragment's position coordinates against the scissor bounds */
 348    m0 = lp_build_compare(builder, type, PIPE_FUNC_GEQUAL, xpos, xmin);
 349    m1 = lp_build_compare(builder, type, PIPE_FUNC_GEQUAL, ypos, ymin);
 350    m2 = lp_build_compare(builder, type, PIPE_FUNC_LESS, xpos, xmax);
 351    m3 = lp_build_compare(builder, type, PIPE_FUNC_LESS, ypos, ymax);
 352
 353    /* AND all the masks together */
 354    m = LLVMBuildAnd(builder, m0, m1, "");
 355    m = LLVMBuildAnd(builder, m, m2, "");
 356    m = LLVMBuildAnd(builder, m, m3, "");
 357
 358    lp_build_name(m, "scissormask");
 359
 360    return m;
 361 }
 362
 363
 364 static LLVMValueRef
 365 build_int32_vec_const(int value)
 366 {
 367    struct lp_type i32_type;
 368
 369    memset(&i32_type, 0, sizeof i32_type);
 370    i32_type.floating = FALSE; /* values are integers */
 371    i32_type.sign = TRUE;      /* values are signed */
 372    i32_type.norm = FALSE;     /* values are not normalized */
 373    i32_type.width = 32;       /* 32-bit int values */
 374    i32_type.length = 4;       /* 4 elements per vector */
 375    return lp_build_const_int_vec(i32_type, value);
 376 }
 377
 378
 379
 380 /**
 381  * Generate the fragment shader, depth/stencil test, and alpha tests.
 382  * \param i  which quad in the tile, in range [0,3]
 383  * \param do_tri_test  if 1, do triangle edge in/out testing
 384  */
 385 static void
 386 generate_fs(struct llvmpipe_context *lp,
 387             struct lp_fragment_shader *shader,
 388             const struct lp_fragment_shader_variant_key *key,
 389             LLVMBuilderRef builder,
 390             struct lp_type type,
 391             LLVMValueRef context_ptr,
 392             unsigned i,
 393             const struct lp_build_interp_soa_context *interp,
 394             struct lp_build_sampler_soa *sampler,
 395             LLVMValueRef *pmask,
 396             LLVMValueRef (*color)[4],
 397             LLVMValueRef depth_ptr,
 398             LLVMValueRef facing,
 399             unsigned do_tri_test,
 400             LLVMValueRef c0,
 401             LLVMValueRef c1,
 402             LLVMValueRef c2,
 403             LLVMValueRef step0_ptr,
 404             LLVMValueRef step1_ptr,
 405             LLVMValueRef step2_ptr,
 406             LLVMValueRef counter)
 407 {
 408    const struct tgsi_token *tokens = shader->base.tokens;
 409    LLVMTypeRef vec_type;
 410    LLVMValueRef consts_ptr;
 411    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][NUM_CHANNELS];
 412    LLVMValueRef z = interp->pos[2];
 413    LLVMValueRef stencil_refs[2];
 414    struct lp_build_flow_context *flow;
 415    struct lp_build_mask_context mask;
 416    boolean early_depth_stencil_test;
 417    unsigned attrib;
 418    unsigned chan;
 419    unsigned cbuf;
 420
 421    assert(i < 4);
 422
 423    stencil_refs[0] = lp_jit_context_stencil_ref_front_value(builder, context_ptr);
 424    stencil_refs[1] = lp_jit_context_stencil_ref_back_value(builder, context_ptr);
 425
 426    vec_type = lp_build_vec_type(type);
 427
 428    consts_ptr = lp_jit_context_constants(builder, context_ptr);
 429
 430    flow = lp_build_flow_create(builder);
 431
 432    memset(outputs, 0, sizeof outputs);
 433
 434    lp_build_flow_scope_begin(flow);
 435
 436    /* Declare the color and z variables */
 437    for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
 438       for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 439          color[cbuf][chan] = LLVMGetUndef(vec_type);
 440          lp_build_flow_scope_declare(flow, &color[cbuf][chan]);
 441       }
 442    }
 443    lp_build_flow_scope_declare(flow, &z);
 444
 445    /* do triangle edge testing */
 446    if (do_tri_test) {
 447       generate_tri_edge_mask(builder, i, pmask,
 448                              c0, c1, c2, step0_ptr, step1_ptr, step2_ptr);
 449    }
 450    else {
 451       *pmask = build_int32_vec_const(~0);
 452    }
 453
 454    /* 'mask' will control execution based on quad's pixel alive/killed state */
 455    lp_build_mask_begin(&mask, flow, type, *pmask);
 456
 457    if (key->scissor) {
 458       LLVMValueRef smask =
 459          generate_scissor_test(builder, context_ptr, interp, type);
 460       lp_build_mask_update(&mask, smask);
 461    }
 462
 463    early_depth_stencil_test =
 464       (key->depth.enabled || key->stencil[0].enabled) &&
 465       !key->alpha.enabled &&
 466       !shader->info.uses_kill &&
 467       !shader->info.writes_z;
 468
 469    if (early_depth_stencil_test)
 470       generate_depth_stencil(builder, key,
 471                              type, &mask,
 472                              stencil_refs, z, depth_ptr, facing, counter);
 473
 474    lp_build_tgsi_soa(builder, tokens, type, &mask,
 475                      consts_ptr, interp->pos, interp->inputs,
 476                      outputs, sampler, &shader->info);
 477
 478    /* loop over fragment shader outputs/results */
 479    for (attrib = 0; attrib < shader->info.num_outputs; ++attrib) {
 480       for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 481          if(outputs[attrib][chan]) {
 482             LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
 483             lp_build_name(out, "output%u.%u.%c", i, attrib, "xyzw"[chan]);
 484
 485             switch (shader->info.output_semantic_name[attrib]) {
 486             case TGSI_SEMANTIC_COLOR:
 487                {
 488                   unsigned cbuf = shader->info.output_semantic_index[attrib];
 489
 490                   lp_build_name(out, "color%u.%u.%c", i, attrib, "rgba"[chan]);
 491
 492                   /* Alpha test */
 493                   /* XXX: should the alpha reference value be passed separately? */
 494                   /* XXX: should only test the final assignment to alpha */
 495                   if(cbuf == 0 && chan == 3) {
 496                      LLVMValueRef alpha = out;
 497                      LLVMValueRef alpha_ref_value;
 498                      alpha_ref_value = lp_jit_context_alpha_ref_value(builder, context_ptr);
 499                      alpha_ref_value = lp_build_broadcast(builder, vec_type, alpha_ref_value);
 500                      lp_build_alpha_test(builder, &key->alpha, type,
 501                                          &mask, alpha, alpha_ref_value);
 502                   }
 503
 504                   color[cbuf][chan] = out;
 505                   break;
 506                }
 507
 508             case TGSI_SEMANTIC_POSITION:
 509                if(chan == 2)
 510                   z = out;
 511                break;
 512             }
 513          }
 514       }
 515    }
 516
 517    if (!early_depth_stencil_test)
 518       generate_depth_stencil(builder, key,
 519                              type, &mask,
 520                              stencil_refs, z, depth_ptr, facing, counter);
 521
 522    lp_build_mask_end(&mask);
 523
 524    lp_build_flow_scope_end(flow);
 525
 526    lp_build_flow_destroy(flow);
 527
 528    *pmask = mask.value;
 529
 530 }
 531
 532
 533 /**
 534  * Generate color blending and color output.
 535  * \param rt  the render target index (to index blend, colormask state)
 536  * \param type  the pixel color type
 537  * \param context_ptr  pointer to the runtime JIT context
 538  * \param mask  execution mask (active fragment/pixel mask)
 539  * \param src  colors from the fragment shader
 540  * \param dst_ptr  the destination color buffer pointer
 541  */
 542 static void
 543 generate_blend(const struct pipe_blend_state *blend,
 544                unsigned rt,
 545                LLVMBuilderRef builder,
 546                struct lp_type type,
 547                LLVMValueRef context_ptr,
 548                LLVMValueRef mask,
 549                LLVMValueRef *src,
 550                LLVMValueRef dst_ptr)
 551 {
 552    struct lp_build_context bld;
 553    struct lp_build_flow_context *flow;
 554    struct lp_build_mask_context mask_ctx;
 555    LLVMTypeRef vec_type;
 556    LLVMValueRef const_ptr;
 557    LLVMValueRef con[4];
 558    LLVMValueRef dst[4];
 559    LLVMValueRef res[4];
 560    unsigned chan;
 561
 562    lp_build_context_init(&bld, builder, type);
 563
 564    flow = lp_build_flow_create(builder);
 565
 566    /* we'll use this mask context to skip blending if all pixels are dead */
 567    lp_build_mask_begin(&mask_ctx, flow, type, mask);
 568
 569    vec_type = lp_build_vec_type(type);
 570
 571    const_ptr = lp_jit_context_blend_color(builder, context_ptr);
 572    const_ptr = LLVMBuildBitCast(builder, const_ptr,
 573                                 LLVMPointerType(vec_type, 0), "");
 574
 575    /* load constant blend color and colors from the dest color buffer */
 576    for(chan = 0; chan < 4; ++chan) {
 577       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
 578       con[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, const_ptr, &index, 1, ""), "");
 579
 580       dst[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, dst_ptr, &index, 1, ""), "");
 581
 582       lp_build_name(con[chan], "con.%c", "rgba"[chan]);
 583       lp_build_name(dst[chan], "dst.%c", "rgba"[chan]);
 584    }
 585
 586    /* do blend */
 587    lp_build_blend_soa(builder, blend, type, rt, src, dst, con, res);
 588
 589    /* store results to color buffer */
 590    for(chan = 0; chan < 4; ++chan) {
 591       if(blend->rt[rt].colormask & (1 << chan)) {
 592          LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
 593          lp_build_name(res[chan], "res.%c", "rgba"[chan]);
 594          res[chan] = lp_build_select(&bld, mask, res[chan], dst[chan]);
 595          LLVMBuildStore(builder, res[chan], LLVMBuildGEP(builder, dst_ptr, &index, 1, ""));
 596       }
 597    }
 598
 599    lp_build_mask_end(&mask_ctx);
 600    lp_build_flow_destroy(flow);
 601 }
 602
 603
 604 /** casting function to avoid compiler warnings */
 605 static lp_jit_frag_func
 606 cast_voidptr_to_lp_jit_frag_func(void *p)
 607 {
 608    union {
 609       void *v;
 610       lp_jit_frag_func f;
 611    } tmp;
 612    assert(sizeof(tmp.v) == sizeof(tmp.f));
 613    tmp.v = p;
 614    return tmp.f;
 615 }
 616
 617
 618 /**
 619  * Generate the runtime callable function for the whole fragment pipeline.
 620  * Note that the function which we generate operates on a block of 16
 621  * pixels at at time.  The block contains 2x2 quads.  Each quad contains
 622  * 2x2 pixels.
 623  */
 624 static void
 625 generate_fragment(struct llvmpipe_context *lp,
 626                   struct lp_fragment_shader *shader,
 627                   struct lp_fragment_shader_variant *variant,
 628                   unsigned do_tri_test)
 629 {
 630    struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
 631    const struct lp_fragment_shader_variant_key *key = &variant->key;
 632    struct lp_type fs_type;
 633    struct lp_type blend_type;
 634    LLVMTypeRef fs_elem_type;
 635    LLVMTypeRef fs_int_vec_type;
 636    LLVMTypeRef blend_vec_type;
 637    LLVMTypeRef arg_types[16];
 638    LLVMTypeRef func_type;
 639    LLVMTypeRef int32_vec4_type = lp_build_int32_vec4_type();
 640    LLVMValueRef context_ptr;
 641    LLVMValueRef x;
 642    LLVMValueRef y;
 643    LLVMValueRef a0_ptr;
 644    LLVMValueRef dadx_ptr;
 645    LLVMValueRef dady_ptr;
 646    LLVMValueRef color_ptr_ptr;
 647    LLVMValueRef depth_ptr;
 648    LLVMValueRef c0, c1, c2, step0_ptr, step1_ptr, step2_ptr, counter = NULL;
 649    LLVMBasicBlockRef block;
 650    LLVMBuilderRef builder;
 651    LLVMValueRef x0;
 652    LLVMValueRef y0;
 653    struct lp_build_sampler_soa *sampler;
 654    struct lp_build_interp_soa_context interp;
 655    LLVMValueRef fs_mask[LP_MAX_VECTOR_LENGTH];
 656    LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS][LP_MAX_VECTOR_LENGTH];
 657    LLVMValueRef blend_mask;
 658    LLVMValueRef function;
 659    LLVMValueRef facing;
 660    unsigned num_fs;
 661    unsigned i;
 662    unsigned chan;
 663    unsigned cbuf;
 664
 665
 666    /* TODO: actually pick these based on the fs and color buffer
 667     * characteristics. */
 668
 669    memset(&fs_type, 0, sizeof fs_type);
 670    fs_type.floating = TRUE; /* floating point values */
 671    fs_type.sign = TRUE;     /* values are signed */
 672    fs_type.norm = FALSE;    /* values are not limited to [0,1] or [-1,1] */
 673    fs_type.width = 32;      /* 32-bit float */
 674    fs_type.length = 4;      /* 4 elements per vector */
 675    num_fs = 4;              /* number of quads per block */
 676
 677    memset(&blend_type, 0, sizeof blend_type);
 678    blend_type.floating = FALSE; /* values are integers */
 679    blend_type.sign = FALSE;     /* values are unsigned */
 680    blend_type.norm = TRUE;      /* values are in [0,1] or [-1,1] */
 681    blend_type.width = 8;        /* 8-bit ubyte values */
 682    blend_type.length = 16;      /* 16 elements per vector */
 683
 684    /*
 685     * Generate the function prototype. Any change here must be reflected in
 686     * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
 687     */
 688
 689    fs_elem_type = lp_build_elem_type(fs_type);
 690    fs_int_vec_type = lp_build_int_vec_type(fs_type);
 691
 692    blend_vec_type = lp_build_vec_type(blend_type);
 693
 694    arg_types[0] = screen->context_ptr_type;            /* context */
 695    arg_types[1] = LLVMInt32Type();                     /* x */
 696    arg_types[2] = LLVMInt32Type();                     /* y */
 697    arg_types[3] = LLVMFloatType();                     /* facing */
 698    arg_types[4] = LLVMPointerType(fs_elem_type, 0);    /* a0 */
 699    arg_types[5] = LLVMPointerType(fs_elem_type, 0);    /* dadx */
 700    arg_types[6] = LLVMPointerType(fs_elem_type, 0);    /* dady */
 701    arg_types[7] = LLVMPointerType(LLVMPointerType(blend_vec_type, 0), 0);  /* color */
 702    arg_types[8] = LLVMPointerType(fs_int_vec_type, 0); /* depth */
 703    arg_types[9] = LLVMInt32Type();                     /* c0 */
 704    arg_types[10] = LLVMInt32Type();                    /* c1 */
 705    arg_types[11] = LLVMInt32Type();                    /* c2 */
 706    /* Note: the step arrays are built as int32[16] but we interpret
 707     * them here as int32_vec4[4].
 708     */
 709    arg_types[12] = LLVMPointerType(int32_vec4_type, 0);/* step0 */
 710    arg_types[13] = LLVMPointerType(int32_vec4_type, 0);/* step1 */
 711    arg_types[14] = LLVMPointerType(int32_vec4_type, 0);/* step2 */
 712    arg_types[15] = LLVMPointerType(LLVMInt32Type(), 0);/* counter */
 713
 714    func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
 715
 716    function = LLVMAddFunction(screen->module, "shader", func_type);
 717    LLVMSetFunctionCallConv(function, LLVMCCallConv);
 718
 719    variant->function[do_tri_test] = function;
 720
 721
 722    /* XXX: need to propagate noalias down into color param now we are
 723     * passing a pointer-to-pointer?
 724     */
 725    for(i = 0; i < Elements(arg_types); ++i)
 726       if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
 727          LLVMAddAttribute(LLVMGetParam(function, i), LLVMNoAliasAttribute);
 728
 729    context_ptr  = LLVMGetParam(function, 0);
 730    x            = LLVMGetParam(function, 1);
 731    y            = LLVMGetParam(function, 2);
 732    facing       = LLVMGetParam(function, 3);
 733    a0_ptr       = LLVMGetParam(function, 4);
 734    dadx_ptr     = LLVMGetParam(function, 5);
 735    dady_ptr     = LLVMGetParam(function, 6);
 736    color_ptr_ptr = LLVMGetParam(function, 7);
 737    depth_ptr    = LLVMGetParam(function, 8);
 738    c0           = LLVMGetParam(function, 9);
 739    c1           = LLVMGetParam(function, 10);
 740    c2           = LLVMGetParam(function, 11);
 741    step0_ptr    = LLVMGetParam(function, 12);
 742    step1_ptr    = LLVMGetParam(function, 13);
 743    step2_ptr    = LLVMGetParam(function, 14);
 744
 745    lp_build_name(context_ptr, "context");
 746    lp_build_name(x, "x");
 747    lp_build_name(y, "y");
 748    lp_build_name(a0_ptr, "a0");
 749    lp_build_name(dadx_ptr, "dadx");
 750    lp_build_name(dady_ptr, "dady");
 751    lp_build_name(color_ptr_ptr, "color_ptr_ptr");
 752    lp_build_name(depth_ptr, "depth");
 753    lp_build_name(c0, "c0");
 754    lp_build_name(c1, "c1");
 755    lp_build_name(c2, "c2");
 756    lp_build_name(step0_ptr, "step0");
 757    lp_build_name(step1_ptr, "step1");
 758    lp_build_name(step2_ptr, "step2");
 759
 760    if (key->occlusion_count) {
 761       counter = LLVMGetParam(function, 15);
 762       lp_build_name(counter, "counter");
 763    }
 764
 765    /*
 766     * Function body
 767     */
 768
 769    block = LLVMAppendBasicBlock(function, "entry");
 770    builder = LLVMCreateBuilder();
 771    LLVMPositionBuilderAtEnd(builder, block);
 772
 773    generate_pos0(builder, x, y, &x0, &y0);
 774
 775    lp_build_interp_soa_init(&interp,
 776                             shader->base.tokens,
 777                             key->flatshade,
 778                             builder, fs_type,
 779                             a0_ptr, dadx_ptr, dady_ptr,
 780                             x0, y0);
 781
 782    /* code generated texture sampling */
 783    sampler = lp_llvm_sampler_soa_create(key->sampler, context_ptr);
 784
 785    /* loop over quads in the block */
 786    for(i = 0; i < num_fs; ++i) {
 787       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
 788       LLVMValueRef out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS];
 789       LLVMValueRef depth_ptr_i;
 790
 791       if(i != 0)
 792          lp_build_interp_soa_update(&interp, i);
 793
 794       depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &index, 1, "");
 795
 796       generate_fs(lp, shader, key,
 797                   builder,
 798                   fs_type,
 799                   context_ptr,
 800                   i,
 801                   &interp,
 802                   sampler,
 803                   &fs_mask[i], /* output */
 804                   out_color,
 805                   depth_ptr_i,
 806                   facing,
 807                   do_tri_test,
 808                   c0, c1, c2,
 809                   step0_ptr, step1_ptr, step2_ptr, counter);
 810
 811       for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++)
 812          for(chan = 0; chan < NUM_CHANNELS; ++chan)
 813             fs_out_color[cbuf][chan][i] = out_color[cbuf][chan];
 814    }
 815
 816    sampler->destroy(sampler);
 817
 818    /* Loop over color outputs / color buffers to do blending.
 819     */
 820    for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
 821       LLVMValueRef color_ptr;
 822       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), cbuf, 0);
 823       LLVMValueRef blend_in_color[NUM_CHANNELS];
 824       unsigned rt;
 825
 826       /*
 827        * Convert the fs's output color and mask to fit to the blending type.
 828        */
 829       for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 830          lp_build_conv(builder, fs_type, blend_type,
 831                        fs_out_color[cbuf][chan], num_fs,
 832                        &blend_in_color[chan], 1);
 833          lp_build_name(blend_in_color[chan], "color%d.%c", cbuf, "rgba"[chan]);
 834       }
 835
 836       lp_build_conv_mask(builder, fs_type, blend_type,
 837                          fs_mask, num_fs,
 838                          &blend_mask, 1);
 839
 840       color_ptr = LLVMBuildLoad(builder,
 841                                 LLVMBuildGEP(builder, color_ptr_ptr, &index, 1, ""),
 842                                 "");
 843       lp_build_name(color_ptr, "color_ptr%d", cbuf);
 844
 845       /* which blend/colormask state to use */
 846       rt = key->blend.independent_blend_enable ? cbuf : 0;
 847
 848       /*
 849        * Blending.
 850        */
 851       generate_blend(&key->blend,
 852                      rt,
 853                      builder,
 854                      blend_type,
 855                      context_ptr,
 856                      blend_mask,
 857                      blend_in_color,
 858                      color_ptr);
 859    }
 860
 861    LLVMBuildRetVoid(builder);
 862
 863    LLVMDisposeBuilder(builder);
 864
 865
 866    /* Verify the LLVM IR.  If invalid, dump and abort */
 867 #ifdef DEBUG
 868    if(LLVMVerifyFunction(function, LLVMPrintMessageAction)) {
 869       if (1)
 870          lp_debug_dump_value(function);
 871       abort();
 872    }
 873 #endif
 874
 875    /* Apply optimizations to LLVM IR */
 876    if (1)
 877       LLVMRunFunctionPassManager(screen->pass, function);
 878
 879    if (LP_DEBUG & DEBUG_JIT) {
 880       /* Print the LLVM IR to stderr */
 881       lp_debug_dump_value(function);
 882       debug_printf("\n");
 883    }
 884
 885    /*
 886     * Translate the LLVM IR into machine code.
 887     */
 888    {
 889       void *f = LLVMGetPointerToGlobal(screen->engine, function);
 890
 891       variant->jit_function[do_tri_test] = cast_voidptr_to_lp_jit_frag_func(f);
 892
 893       if (LP_DEBUG & DEBUG_ASM)
 894          lp_disassemble(f);
 895    }
 896 }
 897
 898
 899 static struct lp_fragment_shader_variant *
 900 generate_variant(struct llvmpipe_context *lp,
 901                  struct lp_fragment_shader *shader,
 902                  const struct lp_fragment_shader_variant_key *key)
 903 {
 904    struct lp_fragment_shader_variant *variant;
 905
 906    if (LP_DEBUG & DEBUG_JIT) {
 907       unsigned i;
 908
 909       tgsi_dump(shader->base.tokens, 0);
 910       if(key->depth.enabled) {
 911          debug_printf("depth.format = %s\n", util_format_name(key->zsbuf_format));
 912          debug_printf("depth.func = %s\n", util_dump_func(key->depth.func, TRUE));
 913          debug_printf("depth.writemask = %u\n", key->depth.writemask);
 914       }
 915       if(key->alpha.enabled) {
 916          debug_printf("alpha.func = %s\n", util_dump_func(key->alpha.func, TRUE));
 917          debug_printf("alpha.ref_value = %f\n", key->alpha.ref_value);
 918       }
 919       if(key->blend.logicop_enable) {
 920          debug_printf("blend.logicop_func = %u\n", key->blend.logicop_func);
 921       }
 922       else if(key->blend.rt[0].blend_enable) {
 923          debug_printf("blend.rgb_func = %s\n",   util_dump_blend_func  (key->blend.rt[0].rgb_func, TRUE));
 924          debug_printf("rgb_src_factor = %s\n",   util_dump_blend_factor(key->blend.rt[0].rgb_src_factor, TRUE));
 925          debug_printf("rgb_dst_factor = %s\n",   util_dump_blend_factor(key->blend.rt[0].rgb_dst_factor, TRUE));
 926          debug_printf("alpha_func = %s\n",       util_dump_blend_func  (key->blend.rt[0].alpha_func, TRUE));
 927          debug_printf("alpha_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_src_factor, TRUE));
 928          debug_printf("alpha_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_dst_factor, TRUE));
 929       }
 930       debug_printf("blend.colormask = 0x%x\n", key->blend.rt[0].colormask);
 931       for(i = 0; i < PIPE_MAX_SAMPLERS; ++i) {
 932          if(key->sampler[i].format) {
 933             debug_printf("sampler[%u] = \n", i);
 934             debug_printf("  .format = %s\n",
 935                          util_format_name(key->sampler[i].format));
 936             debug_printf("  .target = %s\n",
 937                          util_dump_tex_target(key->sampler[i].target, TRUE));
 938             debug_printf("  .pot = %u %u %u\n",
 939                          key->sampler[i].pot_width,
 940                          key->sampler[i].pot_height,
 941                          key->sampler[i].pot_depth);
 942             debug_printf("  .wrap = %s %s %s\n",
 943                          util_dump_tex_wrap(key->sampler[i].wrap_s, TRUE),
 944                          util_dump_tex_wrap(key->sampler[i].wrap_t, TRUE),
 945                          util_dump_tex_wrap(key->sampler[i].wrap_r, TRUE));
 946             debug_printf("  .min_img_filter = %s\n",
 947                          util_dump_tex_filter(key->sampler[i].min_img_filter, TRUE));
 948             debug_printf("  .min_mip_filter = %s\n",
 949                          util_dump_tex_mipfilter(key->sampler[i].min_mip_filter, TRUE));
 950             debug_printf("  .mag_img_filter = %s\n",
 951                          util_dump_tex_filter(key->sampler[i].mag_img_filter, TRUE));
 952             if(key->sampler[i].compare_mode != PIPE_TEX_COMPARE_NONE)
 953                debug_printf("  .compare_func = %s\n", util_dump_func(key->sampler[i].compare_func, TRUE));
 954             debug_printf("  .normalized_coords = %u\n", key->sampler[i].normalized_coords);
 955          }
 956       }
 957    }
 958
 959    variant = CALLOC_STRUCT(lp_fragment_shader_variant);
 960    if(!variant)
 961       return NULL;
 962
 963    memcpy(&variant->key, key, sizeof *key);
 964
 965    generate_fragment(lp, shader, variant, 0);
 966    generate_fragment(lp, shader, variant, 1);
 967
 968    /* TODO: most of these can be relaxed, in particular the colormask */
 969    variant->opaque =
 970          !key->blend.logicop_enable &&
 971          !key->blend.rt[0].blend_enable &&
 972          key->blend.rt[0].colormask == 0xf &&
 973          !key->stencil[0].enabled &&
 974          !key->alpha.enabled &&
 975          !key->depth.enabled &&
 976          !key->scissor &&
 977          !shader->info.uses_kill
 978          ? TRUE : FALSE;
 979
 980    /* insert new variant into linked list */
 981    variant->next = shader->variants;
 982    shader->variants = variant;
 983
 984    return variant;
 985 }
 986
 987
 988 static void *
 989 llvmpipe_create_fs_state(struct pipe_context *pipe,
 990                          const struct pipe_shader_state *templ)
 991 {
 992    struct lp_fragment_shader *shader;
 993
 994    shader = CALLOC_STRUCT(lp_fragment_shader);
 995    if (!shader)
 996       return NULL;
 997
 998    /* get/save the summary info for this shader */
 999    tgsi_scan_shader(templ->tokens, &shader->info);
1000
1001    /* we need to keep a local copy of the tokens */
1002    shader->base.tokens = tgsi_dup_tokens(templ->tokens);
1003
1004    if (LP_DEBUG & DEBUG_TGSI) {
1005       debug_printf("llvmpipe: Create fragment shader %p:\n", (void *) shader);
1006       tgsi_dump(templ->tokens, 0);
1007    }
1008
1009    return shader;
1010 }
1011
1012
1013 static void
1014 llvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs)
1015 {
1016    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
1017
1018    if (llvmpipe->fs == fs)
1019       return;
1020
1021    draw_flush(llvmpipe->draw);
1022
1023    llvmpipe->fs = fs;
1024
1025    llvmpipe->dirty |= LP_NEW_FS;
1026 }
1027
1028
1029 static void
1030 llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs)
1031 {
1032    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
1033    struct llvmpipe_screen *screen = llvmpipe_screen(pipe->screen);
1034    struct lp_fragment_shader *shader = fs;
1035    struct lp_fragment_shader_variant *variant;
1036
1037    assert(fs != llvmpipe->fs);
1038    (void) llvmpipe;
1039
1040    /*
1041     * XXX: we need to flush the context until we have some sort of reference
1042     * counting in fragment shaders as they may still be binned
1043     */
1044    draw_flush(llvmpipe->draw);
1045    lp_setup_flush(llvmpipe->setup, 0);
1046
1047    variant = shader->variants;
1048    while(variant) {
1049       struct lp_fragment_shader_variant *next = variant->next;
1050       unsigned i;
1051
1052       for (i = 0; i < Elements(variant->function); i++) {
1053          if (variant->function[i]) {
1054             if (variant->jit_function[i])
1055                LLVMFreeMachineCodeForFunction(screen->engine,
1056                                               variant->function[i]);
1057             LLVMDeleteFunction(variant->function[i]);
1058          }
1059       }
1060
1061       FREE(variant);
1062
1063       variant = next;
1064    }
1065
1066    FREE((void *) shader->base.tokens);
1067    FREE(shader);
1068 }
1069
1070
1071
1072 static void
1073 llvmpipe_set_constant_buffer(struct pipe_context *pipe,
1074                              uint shader, uint index,
1075                              struct pipe_resource *constants)
1076 {
1077    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
1078    unsigned size = constants ? constants->width0 : 0;
1079    const void *data = constants ? llvmpipe_resource_data(constants) : NULL;
1080
1081    assert(shader < PIPE_SHADER_TYPES);
1082    assert(index == 0);
1083
1084    if(llvmpipe->constants[shader] == constants)
1085       return;
1086
1087    draw_flush(llvmpipe->draw);
1088
1089    /* note: reference counting */
1090    pipe_resource_reference(&llvmpipe->constants[shader], constants);
1091
1092    if(shader == PIPE_SHADER_VERTEX) {
1093       draw_set_mapped_constant_buffer(llvmpipe->draw, PIPE_SHADER_VERTEX, 0,
1094                                       data, size);
1095    }
1096
1097    llvmpipe->dirty |= LP_NEW_CONSTANTS;
1098 }
1099
1100
1101 /**
1102  * Return the blend factor equivalent to a destination alpha of one.
1103  */
1104 static INLINE unsigned
1105 force_dst_alpha_one(unsigned factor, boolean alpha)
1106 {
1107    switch(factor) {
1108    case PIPE_BLENDFACTOR_DST_ALPHA:
1109       return PIPE_BLENDFACTOR_ONE;
1110    case PIPE_BLENDFACTOR_INV_DST_ALPHA:
1111       return PIPE_BLENDFACTOR_ZERO;
1112    case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
1113       return PIPE_BLENDFACTOR_ZERO;
1114    }
1115
1116    if (alpha) {
1117       switch(factor) {
1118       case PIPE_BLENDFACTOR_DST_COLOR:
1119          return PIPE_BLENDFACTOR_ONE;
1120       case PIPE_BLENDFACTOR_INV_DST_COLOR:
1121          return PIPE_BLENDFACTOR_ZERO;
1122       }
1123    }
1124
1125    return factor;
1126 }
1127
1128
1129 /**
1130  * We need to generate several variants of the fragment pipeline to match
1131  * all the combinations of the contributing state atoms.
1132  *
1133  * TODO: there is actually no reason to tie this to context state -- the
1134  * generated code could be cached globally in the screen.
1135  */
1136 static void
1137 make_variant_key(struct llvmpipe_context *lp,
1138                  struct lp_fragment_shader *shader,
1139                  struct lp_fragment_shader_variant_key *key)
1140 {
1141    unsigned i;
1142
1143    memset(key, 0, sizeof *key);
1144
1145    if (lp->framebuffer.zsbuf) {
1146       if (lp->depth_stencil->depth.enabled) {
1147          key->zsbuf_format = lp->framebuffer.zsbuf->format;
1148          memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth);
1149       }
1150       if (lp->depth_stencil->stencil[0].enabled) {
1151          key->zsbuf_format = lp->framebuffer.zsbuf->format;
1152          memcpy(&key->stencil, &lp->depth_stencil->stencil, sizeof key->stencil);
1153       }
1154    }
1155
1156    key->alpha.enabled = lp->depth_stencil->alpha.enabled;
1157    if(key->alpha.enabled)
1158       key->alpha.func = lp->depth_stencil->alpha.func;
1159    /* alpha.ref_value is passed in jit_context */
1160
1161    key->flatshade = lp->rasterizer->flatshade;
1162    key->scissor = lp->rasterizer->scissor;
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       struct pipe_rt_blend_state *blend_rt = &key->blend.rt[i];
1174       const struct util_format_description *format_desc;
1175       unsigned chan;
1176
1177       format_desc = util_format_description(lp->framebuffer.cbufs[i]->format);
1178       assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
1179              format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB);
1180
1181       blend_rt->colormask = lp->blend->rt[i].colormask;
1182
1183       /* mask out color channels not present in the color buffer.
1184        * Should be simple to incorporate per-cbuf writemasks:
1185        */
1186       for(chan = 0; chan < 4; ++chan) {
1187          enum util_format_swizzle swizzle = format_desc->swizzle[chan];
1188
1189          if(swizzle > UTIL_FORMAT_SWIZZLE_W)
1190             blend_rt->colormask &= ~(1 << chan);
1191       }
1192
1193       /*
1194        * Our swizzled render tiles always have an alpha channel, but the linear
1195        * render target format often does not, so force here the dst alpha to be
1196        * one.
1197        *
1198        * This is not a mere optimization. Wrong results will be produced if the
1199        * dst alpha is used, the dst format does not have alpha, and the previous
1200        * rendering was not flushed from the swizzled to linear buffer. For
1201        * example, NonPowTwo DCT.
1202        *
1203        * TODO: This should be generalized to all channels for better
1204        * performance, but only alpha causes correctness issues.
1205        */
1206       if (format_desc->swizzle[3] > UTIL_FORMAT_SWIZZLE_W) {
1207          blend_rt->rgb_src_factor = force_dst_alpha_one(blend_rt->rgb_src_factor, FALSE);
1208          blend_rt->rgb_dst_factor = force_dst_alpha_one(blend_rt->rgb_dst_factor, FALSE);
1209          blend_rt->alpha_src_factor = force_dst_alpha_one(blend_rt->alpha_src_factor, TRUE);
1210          blend_rt->alpha_dst_factor = force_dst_alpha_one(blend_rt->alpha_dst_factor, TRUE);
1211       }
1212    }
1213
1214    for(i = 0; i < PIPE_MAX_SAMPLERS; ++i)
1215       if(shader->info.file_mask[TGSI_FILE_SAMPLER] & (1 << i))
1216          lp_sampler_static_state(&key->sampler[i], lp->fragment_sampler_views[i], lp->sampler[i]);
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;
1230
1231    make_variant_key(lp, shader, &key);
1232
1233    variant = shader->variants;
1234    while(variant) {
1235       if(memcmp(&variant->key, &key, sizeof key) == 0)
1236          break;
1237
1238       variant = variant->next;
1239    }
1240
1241    if (!variant) {
1242       int64_t t0, t1;
1243       int64_t dt;
1244       t0 = os_time_get();
1245
1246       variant = generate_variant(lp, shader, &key);
1247
1248       t1 = os_time_get();
1249       dt = t1 - t0;
1250       LP_COUNT_ADD(llvm_compile_time, dt);
1251       LP_COUNT_ADD(nr_llvm_compiles, 2);  /* emit vs. omit in/out test */
1252    }
1253
1254    lp_setup_set_fs_functions(lp->setup,
1255                              variant->jit_function[RAST_WHOLE],
1256                              variant->jit_function[RAST_EDGE_TEST],
1257                              variant->opaque);
1258 }
1259
1260
1261
1262 void
1263 llvmpipe_init_fs_funcs(struct llvmpipe_context *llvmpipe)
1264 {
1265    llvmpipe->pipe.create_fs_state = llvmpipe_create_fs_state;
1266    llvmpipe->pipe.bind_fs_state   = llvmpipe_bind_fs_state;
1267    llvmpipe->pipe.delete_fs_state = llvmpipe_delete_fs_state;
1268
1269    llvmpipe->pipe.set_constant_buffer = llvmpipe_set_constant_buffer;
1270 }