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