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