src/gallium/drivers/llvmpipe/lp_state_fs.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2009 VMware, Inc.
   4  * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas.
   5  * All Rights Reserved.
   6  *
   7  * Permission is hereby granted, free of charge, to any person obtaining a
   8  * copy of this software and associated documentation files (the
   9  * "Software"), to deal in the Software without restriction, including
  10  * without limitation the rights to use, copy, modify, merge, publish,
  11  * distribute, sub license, and/or sell copies of the Software, and to
  12  * permit persons to whom the Software is furnished to do so, subject to
  13  * the following conditions:
  14  *
  15  * The above copyright notice and this permission notice (including the
  16  * next paragraph) shall be included in all copies or substantial portions
  17  * of the Software.
  18  *
  19  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  20  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  21  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
  22  * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
  23  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
  24  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
  25  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  26  *
  27  **************************************************************************/
  28
  29 /**
  30  * @file
  31  * Code generate the whole fragment pipeline.
  32  *
  33  * The fragment pipeline consists of the following stages:
  34  * - early depth test
  35  * - fragment shader
  36  * - alpha test
  37  * - depth/stencil test
  38  * - blending
  39  *
  40  * This file has only the glue to assemble the fragment pipeline.  The actual
  41  * plumbing of converting Gallium state into LLVM IR is done elsewhere, in the
  42  * lp_bld_*.[ch] files, and in a complete generic and reusable way. Here we
  43  * muster the LLVM JIT execution engine to create a function that follows an
  44  * established binary interface and that can be called from C directly.
  45  *
  46  * A big source of complexity here is that we often want to run different
  47  * stages with different precisions and data types and precisions. For example,
  48  * the fragment shader needs typically to be done in floats, but the
  49  * depth/stencil test and blending is better done in the type that most closely
  50  * matches the depth/stencil and color buffer respectively.
  51  *
  52  * Since the width of a SIMD vector register stays the same regardless of the
  53  * element type, different types imply different number of elements, so we must
  54  * code generate more instances of the stages with larger types to be able to
  55  * feed/consume the stages with smaller types.
  56  *
  57  * @author Jose Fonseca <jfonseca@vmware.com>
  58  */
  59
  60 #include <limits.h>
  61 #include "pipe/p_defines.h"
  62 #include "util/u_inlines.h"
  63 #include "util/u_memory.h"
  64 #include "util/u_pointer.h"
  65 #include "util/u_format.h"
  66 #include "util/u_dump.h"
  67 #include "util/u_string.h"
  68 #include "util/u_simple_list.h"
  69 #include "os/os_time.h"
  70 #include "pipe/p_shader_tokens.h"
  71 #include "draw/draw_context.h"
  72 #include "tgsi/tgsi_dump.h"
  73 #include "tgsi/tgsi_scan.h"
  74 #include "tgsi/tgsi_parse.h"
  75 #include "gallivm/lp_bld_type.h"
  76 #include "gallivm/lp_bld_const.h"
  77 #include "gallivm/lp_bld_conv.h"
  78 #include "gallivm/lp_bld_init.h"
  79 #include "gallivm/lp_bld_intr.h"
  80 #include "gallivm/lp_bld_logic.h"
  81 #include "gallivm/lp_bld_tgsi.h"
  82 #include "gallivm/lp_bld_swizzle.h"
  83 #include "gallivm/lp_bld_flow.h"
  84 #include "gallivm/lp_bld_debug.h"
  85
  86 #include "lp_bld_alpha.h"
  87 #include "lp_bld_blend.h"
  88 #include "lp_bld_depth.h"
  89 #include "lp_bld_interp.h"
  90 #include "lp_context.h"
  91 #include "lp_debug.h"
  92 #include "lp_perf.h"
  93 #include "lp_screen.h"
  94 #include "lp_setup.h"
  95 #include "lp_state.h"
  96 #include "lp_tex_sample.h"
  97 #include "lp_flush.h"
  98 #include "lp_state_fs.h"
  99
 100
 101 #include <llvm-c/Analysis.h>
 102
 103
 104 static unsigned fs_no = 0;
 105
 106
 107 /**
 108  * Generate the depth /stencil test code.
 109  */
 110 static void
 111 generate_depth_stencil(LLVMBuilderRef builder,
 112                        const struct lp_fragment_shader_variant_key *key,
 113                        struct lp_type src_type,
 114                        struct lp_build_mask_context *mask,
 115                        LLVMValueRef stencil_refs[2],
 116                        LLVMValueRef src,
 117                        LLVMValueRef dst_ptr,
 118                        LLVMValueRef facing,
 119                        LLVMValueRef counter)
 120 {
 121    const struct util_format_description *format_desc;
 122    struct lp_type dst_type;
 123
 124    if (!key->depth.enabled && !key->stencil[0].enabled && !key->stencil[1].enabled)
 125       return;
 126
 127    format_desc = util_format_description(key->zsbuf_format);
 128    assert(format_desc);
 129
 130    /*
 131     * Depths are expected to be between 0 and 1, even if they are stored in
 132     * floats. Setting these bits here will ensure that the lp_build_conv() call
 133     * below won't try to unnecessarily clamp the incoming values.
 134     */
 135    if(src_type.floating) {
 136       src_type.sign = FALSE;
 137       src_type.norm = TRUE;
 138    }
 139    else {
 140       assert(!src_type.sign);
 141       assert(src_type.norm);
 142    }
 143
 144    /* Pick the depth type. */
 145    dst_type = lp_depth_type(format_desc, src_type.width*src_type.length);
 146
 147    /* FIXME: Cope with a depth test type with a different bit width. */
 148    assert(dst_type.width == src_type.width);
 149    assert(dst_type.length == src_type.length);
 150
 151    /* Convert fragment Z from float to integer */
 152    lp_build_conv(builder, src_type, dst_type, &src, 1, &src, 1);
 153
 154    dst_ptr = LLVMBuildBitCast(builder,
 155                               dst_ptr,
 156                               LLVMPointerType(lp_build_vec_type(dst_type), 0), "");
 157    lp_build_depth_stencil_test(builder,
 158                                &key->depth,
 159                                key->stencil,
 160                                dst_type,
 161                                format_desc,
 162                                mask,
 163                                stencil_refs,
 164                                src,
 165                                dst_ptr,
 166                                facing,
 167                                counter);
 168 }
 169
 170
 171 /**
 172  * Expand the relevent bits of mask_input to a 4-dword mask for the
 173  * four pixels in a 2x2 quad.  This will set the four elements of the
 174  * quad mask vector to 0 or ~0.
 175  *
 176  * \param quad  which quad of the quad group to test, in [0,3]
 177  * \param mask_input  bitwise mask for the whole 4x4 stamp
 178  */
 179 static LLVMValueRef
 180 generate_quad_mask(LLVMBuilderRef builder,
 181                    struct lp_type fs_type,
 182                    unsigned quad,
 183                    LLVMValueRef mask_input) /* int32 */
 184 {
 185    struct lp_type mask_type;
 186    LLVMTypeRef i32t = LLVMInt32Type();
 187    LLVMValueRef bits[4];
 188    LLVMValueRef mask;
 189    int shift;
 190
 191    /*
 192     * XXX: We'll need a different path for 16 x u8
 193     */
 194    assert(fs_type.width == 32);
 195    assert(fs_type.length == 4);
 196    mask_type = lp_int_type(fs_type);
 197
 198    /*
 199     * mask_input >>= (quad * 4)
 200     */
 201
 202    switch (quad) {
 203    case 0:
 204       shift = 0;
 205       break;
 206    case 1:
 207       shift = 2;
 208       break;
 209    case 2:
 210       shift = 8;
 211       break;
 212    case 3:
 213       shift = 10;
 214       break;
 215    default:
 216       assert(0);
 217       shift = 0;
 218    }
 219
 220    mask_input = LLVMBuildLShr(builder,
 221                               mask_input,
 222                               LLVMConstInt(i32t, shift, 0),
 223                               "");
 224
 225    /*
 226     * mask = { mask_input & (1 << i), for i in [0,3] }
 227     */
 228
 229    mask = lp_build_broadcast(builder, lp_build_vec_type(mask_type), mask_input);
 230
 231    bits[0] = LLVMConstInt(i32t, 1 << 0, 0);
 232    bits[1] = LLVMConstInt(i32t, 1 << 1, 0);
 233    bits[2] = LLVMConstInt(i32t, 1 << 4, 0);
 234    bits[3] = LLVMConstInt(i32t, 1 << 5, 0);
 235
 236    mask = LLVMBuildAnd(builder, mask, LLVMConstVector(bits, 4), "");
 237
 238    /*
 239     * mask = mask != 0 ? ~0 : 0
 240     */
 241
 242    mask = lp_build_compare(builder,
 243                            mask_type, PIPE_FUNC_NOTEQUAL,
 244                            mask,
 245                            lp_build_const_int_vec(mask_type, 0));
 246
 247    return mask;
 248 }
 249
 250
 251
 252 /**
 253  * Generate the fragment shader, depth/stencil test, and alpha tests.
 254  * \param i  which quad in the tile, in range [0,3]
 255  * \param partial_mask  if 1, do mask_input testing
 256  */
 257 static void
 258 generate_fs(struct llvmpipe_context *lp,
 259             struct lp_fragment_shader *shader,
 260             const struct lp_fragment_shader_variant_key *key,
 261             LLVMBuilderRef builder,
 262             struct lp_type type,
 263             LLVMValueRef context_ptr,
 264             unsigned i,
 265             const struct lp_build_interp_soa_context *interp,
 266             struct lp_build_sampler_soa *sampler,
 267             LLVMValueRef *pmask,
 268             LLVMValueRef (*color)[4],
 269             LLVMValueRef depth_ptr,
 270             LLVMValueRef facing,
 271             unsigned partial_mask,
 272             LLVMValueRef mask_input,
 273             LLVMValueRef counter)
 274 {
 275    const struct tgsi_token *tokens = shader->base.tokens;
 276    LLVMTypeRef vec_type;
 277    LLVMValueRef consts_ptr;
 278    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][NUM_CHANNELS];
 279    LLVMValueRef z = interp->pos[2];
 280    LLVMValueRef stencil_refs[2];
 281    struct lp_build_flow_context *flow;
 282    struct lp_build_mask_context mask;
 283    boolean early_depth_stencil_test;
 284    unsigned attrib;
 285    unsigned chan;
 286    unsigned cbuf;
 287
 288    assert(i < 4);
 289
 290    stencil_refs[0] = lp_jit_context_stencil_ref_front_value(builder, context_ptr);
 291    stencil_refs[1] = lp_jit_context_stencil_ref_back_value(builder, context_ptr);
 292
 293    vec_type = lp_build_vec_type(type);
 294
 295    consts_ptr = lp_jit_context_constants(builder, context_ptr);
 296
 297    flow = lp_build_flow_create(builder);
 298
 299    memset(outputs, 0, sizeof outputs);
 300
 301    lp_build_flow_scope_begin(flow);
 302
 303    /* Declare the color and z variables */
 304    for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
 305       for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 306          color[cbuf][chan] = LLVMGetUndef(vec_type);
 307          lp_build_flow_scope_declare(flow, &color[cbuf][chan]);
 308       }
 309    }
 310    lp_build_flow_scope_declare(flow, &z);
 311
 312    /* do triangle edge testing */
 313    if (partial_mask) {
 314       *pmask = generate_quad_mask(builder, type,
 315                                   i, mask_input);
 316    }
 317    else {
 318       *pmask = lp_build_const_int_vec(type, ~0);
 319    }
 320
 321    /* 'mask' will control execution based on quad's pixel alive/killed state */
 322    lp_build_mask_begin(&mask, flow, type, *pmask);
 323
 324    early_depth_stencil_test =
 325       (key->depth.enabled || key->stencil[0].enabled) &&
 326       !key->alpha.enabled &&
 327       !shader->info.uses_kill &&
 328       !shader->info.writes_z;
 329
 330    if (early_depth_stencil_test)
 331       generate_depth_stencil(builder, key,
 332                              type, &mask,
 333                              stencil_refs, z, depth_ptr, facing, counter);
 334
 335    lp_build_tgsi_soa(builder, tokens, type, &mask,
 336                      consts_ptr, interp->pos, interp->inputs,
 337                      outputs, sampler, &shader->info);
 338
 339    /* loop over fragment shader outputs/results */
 340    for (attrib = 0; attrib < shader->info.num_outputs; ++attrib) {
 341       for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 342          if(outputs[attrib][chan]) {
 343             LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
 344             lp_build_name(out, "output%u.%u.%c", i, attrib, "xyzw"[chan]);
 345
 346             switch (shader->info.output_semantic_name[attrib]) {
 347             case TGSI_SEMANTIC_COLOR:
 348                {
 349                   unsigned cbuf = shader->info.output_semantic_index[attrib];
 350
 351                   lp_build_name(out, "color%u.%u.%c", i, attrib, "rgba"[chan]);
 352
 353                   /* Alpha test */
 354                   /* XXX: should only test the final assignment to alpha */
 355                   if (cbuf == 0 && chan == 3 && key->alpha.enabled) {
 356                      LLVMValueRef alpha = out;
 357                      LLVMValueRef alpha_ref_value;
 358                      alpha_ref_value = lp_jit_context_alpha_ref_value(builder, context_ptr);
 359                      alpha_ref_value = lp_build_broadcast(builder, vec_type, alpha_ref_value);
 360                      lp_build_alpha_test(builder, key->alpha.func, type,
 361                                          &mask, alpha, alpha_ref_value);
 362                   }
 363
 364                   color[cbuf][chan] = out;
 365                   break;
 366                }
 367
 368             case TGSI_SEMANTIC_POSITION:
 369                if(chan == 2)
 370                   z = out;
 371                break;
 372             }
 373          }
 374       }
 375    }
 376
 377    if (!early_depth_stencil_test)
 378       generate_depth_stencil(builder, key,
 379                              type, &mask,
 380                              stencil_refs, z, depth_ptr, facing, counter);
 381
 382    lp_build_mask_end(&mask);
 383
 384    lp_build_flow_scope_end(flow);
 385
 386    lp_build_flow_destroy(flow);
 387
 388    *pmask = mask.value;
 389
 390 }
 391
 392
 393 /**
 394  * Generate color blending and color output.
 395  * \param rt  the render target index (to index blend, colormask state)
 396  * \param type  the pixel color type
 397  * \param context_ptr  pointer to the runtime JIT context
 398  * \param mask  execution mask (active fragment/pixel mask)
 399  * \param src  colors from the fragment shader
 400  * \param dst_ptr  the destination color buffer pointer
 401  */
 402 static void
 403 generate_blend(const struct pipe_blend_state *blend,
 404                unsigned rt,
 405                LLVMBuilderRef builder,
 406                struct lp_type type,
 407                LLVMValueRef context_ptr,
 408                LLVMValueRef mask,
 409                LLVMValueRef *src,
 410                LLVMValueRef dst_ptr)
 411 {
 412    struct lp_build_context bld;
 413    struct lp_build_flow_context *flow;
 414    struct lp_build_mask_context mask_ctx;
 415    LLVMTypeRef vec_type;
 416    LLVMValueRef const_ptr;
 417    LLVMValueRef con[4];
 418    LLVMValueRef dst[4];
 419    LLVMValueRef res[4];
 420    unsigned chan;
 421
 422    lp_build_context_init(&bld, builder, type);
 423
 424    flow = lp_build_flow_create(builder);
 425
 426    /* we'll use this mask context to skip blending if all pixels are dead */
 427    lp_build_mask_begin(&mask_ctx, flow, type, mask);
 428
 429    vec_type = lp_build_vec_type(type);
 430
 431    const_ptr = lp_jit_context_blend_color(builder, context_ptr);
 432    const_ptr = LLVMBuildBitCast(builder, const_ptr,
 433                                 LLVMPointerType(vec_type, 0), "");
 434
 435    /* load constant blend color and colors from the dest color buffer */
 436    for(chan = 0; chan < 4; ++chan) {
 437       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
 438       con[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, const_ptr, &index, 1, ""), "");
 439
 440       dst[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, dst_ptr, &index, 1, ""), "");
 441
 442       lp_build_name(con[chan], "con.%c", "rgba"[chan]);
 443       lp_build_name(dst[chan], "dst.%c", "rgba"[chan]);
 444    }
 445
 446    /* do blend */
 447    lp_build_blend_soa(builder, blend, type, rt, src, dst, con, res);
 448
 449    /* store results to color buffer */
 450    for(chan = 0; chan < 4; ++chan) {
 451       if(blend->rt[rt].colormask & (1 << chan)) {
 452          LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
 453          lp_build_name(res[chan], "res.%c", "rgba"[chan]);
 454          res[chan] = lp_build_select(&bld, mask, res[chan], dst[chan]);
 455          LLVMBuildStore(builder, res[chan], LLVMBuildGEP(builder, dst_ptr, &index, 1, ""));
 456       }
 457    }
 458
 459    lp_build_mask_end(&mask_ctx);
 460    lp_build_flow_destroy(flow);
 461 }
 462
 463
 464 /**
 465  * Generate the runtime callable function for the whole fragment pipeline.
 466  * Note that the function which we generate operates on a block of 16
 467  * pixels at at time.  The block contains 2x2 quads.  Each quad contains
 468  * 2x2 pixels.
 469  */
 470 static void
 471 generate_fragment(struct llvmpipe_context *lp,
 472                   struct lp_fragment_shader *shader,
 473                   struct lp_fragment_shader_variant *variant,
 474                   unsigned partial_mask)
 475 {
 476    struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
 477    const struct lp_fragment_shader_variant_key *key = &variant->key;
 478    char func_name[256];
 479    struct lp_type fs_type;
 480    struct lp_type blend_type;
 481    LLVMTypeRef fs_elem_type;
 482    LLVMTypeRef fs_int_vec_type;
 483    LLVMTypeRef blend_vec_type;
 484    LLVMTypeRef arg_types[11];
 485    LLVMTypeRef func_type;
 486    LLVMValueRef context_ptr;
 487    LLVMValueRef x;
 488    LLVMValueRef y;
 489    LLVMValueRef a0_ptr;
 490    LLVMValueRef dadx_ptr;
 491    LLVMValueRef dady_ptr;
 492    LLVMValueRef color_ptr_ptr;
 493    LLVMValueRef depth_ptr;
 494    LLVMValueRef mask_input;
 495    LLVMValueRef counter = NULL;
 496    LLVMBasicBlockRef block;
 497    LLVMBuilderRef builder;
 498    struct lp_build_sampler_soa *sampler;
 499    struct lp_build_interp_soa_context interp;
 500    LLVMValueRef fs_mask[LP_MAX_VECTOR_LENGTH];
 501    LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS][LP_MAX_VECTOR_LENGTH];
 502    LLVMValueRef blend_mask;
 503    LLVMValueRef function;
 504    LLVMValueRef facing;
 505    unsigned num_fs;
 506    unsigned i;
 507    unsigned chan;
 508    unsigned cbuf;
 509
 510
 511    /* TODO: actually pick these based on the fs and color buffer
 512     * characteristics. */
 513
 514    memset(&fs_type, 0, sizeof fs_type);
 515    fs_type.floating = TRUE; /* floating point values */
 516    fs_type.sign = TRUE;     /* values are signed */
 517    fs_type.norm = FALSE;    /* values are not limited to [0,1] or [-1,1] */
 518    fs_type.width = 32;      /* 32-bit float */
 519    fs_type.length = 4;      /* 4 elements per vector */
 520    num_fs = 4;              /* number of quads per block */
 521
 522    memset(&blend_type, 0, sizeof blend_type);
 523    blend_type.floating = FALSE; /* values are integers */
 524    blend_type.sign = FALSE;     /* values are unsigned */
 525    blend_type.norm = TRUE;      /* values are in [0,1] or [-1,1] */
 526    blend_type.width = 8;        /* 8-bit ubyte values */
 527    blend_type.length = 16;      /* 16 elements per vector */
 528
 529    /*
 530     * Generate the function prototype. Any change here must be reflected in
 531     * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
 532     */
 533
 534    fs_elem_type = lp_build_elem_type(fs_type);
 535    fs_int_vec_type = lp_build_int_vec_type(fs_type);
 536
 537    blend_vec_type = lp_build_vec_type(blend_type);
 538
 539    util_snprintf(func_name, sizeof(func_name), "fs%u_variant%u_%s",
 540                  shader->no, variant->no, partial_mask ? "partial" : "whole");
 541
 542    arg_types[0] = screen->context_ptr_type;            /* context */
 543    arg_types[1] = LLVMInt32Type();                     /* x */
 544    arg_types[2] = LLVMInt32Type();                     /* y */
 545    arg_types[3] = LLVMFloatType();                     /* facing */
 546    arg_types[4] = LLVMPointerType(fs_elem_type, 0);    /* a0 */
 547    arg_types[5] = LLVMPointerType(fs_elem_type, 0);    /* dadx */
 548    arg_types[6] = LLVMPointerType(fs_elem_type, 0);    /* dady */
 549    arg_types[7] = LLVMPointerType(LLVMPointerType(blend_vec_type, 0), 0);  /* color */
 550    arg_types[8] = LLVMPointerType(fs_int_vec_type, 0); /* depth */
 551    arg_types[9] = LLVMInt32Type();                     /* mask_input */
 552    arg_types[10] = LLVMPointerType(LLVMInt32Type(), 0);/* counter */
 553
 554    func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
 555
 556    function = LLVMAddFunction(screen->module, func_name, func_type);
 557    LLVMSetFunctionCallConv(function, LLVMCCallConv);
 558
 559    variant->function[partial_mask] = function;
 560
 561
 562    /* XXX: need to propagate noalias down into color param now we are
 563     * passing a pointer-to-pointer?
 564     */
 565    for(i = 0; i < Elements(arg_types); ++i)
 566       if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
 567          LLVMAddAttribute(LLVMGetParam(function, i), LLVMNoAliasAttribute);
 568
 569    context_ptr  = LLVMGetParam(function, 0);
 570    x            = LLVMGetParam(function, 1);
 571    y            = LLVMGetParam(function, 2);
 572    facing       = LLVMGetParam(function, 3);
 573    a0_ptr       = LLVMGetParam(function, 4);
 574    dadx_ptr     = LLVMGetParam(function, 5);
 575    dady_ptr     = LLVMGetParam(function, 6);
 576    color_ptr_ptr = LLVMGetParam(function, 7);
 577    depth_ptr    = LLVMGetParam(function, 8);
 578    mask_input   = LLVMGetParam(function, 9);
 579
 580    lp_build_name(context_ptr, "context");
 581    lp_build_name(x, "x");
 582    lp_build_name(y, "y");
 583    lp_build_name(a0_ptr, "a0");
 584    lp_build_name(dadx_ptr, "dadx");
 585    lp_build_name(dady_ptr, "dady");
 586    lp_build_name(color_ptr_ptr, "color_ptr_ptr");
 587    lp_build_name(depth_ptr, "depth");
 588    lp_build_name(mask_input, "mask_input");
 589
 590    if (key->occlusion_count) {
 591       counter = LLVMGetParam(function, 10);
 592       lp_build_name(counter, "counter");
 593    }
 594
 595    /*
 596     * Function body
 597     */
 598
 599    block = LLVMAppendBasicBlock(function, "entry");
 600    builder = LLVMCreateBuilder();
 601    LLVMPositionBuilderAtEnd(builder, block);
 602
 603    /*
 604     * The shader input interpolation info is not explicitely baked in the
 605     * shader key, but everything it derives from (TGSI, and flatshade) is
 606     * already included in the shader key.
 607     */
 608    lp_build_interp_soa_init(&interp,
 609                             lp->num_inputs,
 610                             lp->inputs,
 611                             builder, fs_type,
 612                             a0_ptr, dadx_ptr, dady_ptr,
 613                             x, y);
 614
 615    /* code generated texture sampling */
 616    sampler = lp_llvm_sampler_soa_create(key->sampler, context_ptr);
 617
 618    /* loop over quads in the block */
 619    for(i = 0; i < num_fs; ++i) {
 620       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
 621       LLVMValueRef out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS];
 622       LLVMValueRef depth_ptr_i;
 623
 624       if(i != 0)
 625          lp_build_interp_soa_update(&interp, i);
 626
 627       depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &index, 1, "");
 628
 629       generate_fs(lp, shader, key,
 630                   builder,
 631                   fs_type,
 632                   context_ptr,
 633                   i,
 634                   &interp,
 635                   sampler,
 636                   &fs_mask[i], /* output */
 637                   out_color,
 638                   depth_ptr_i,
 639                   facing,
 640                   partial_mask,
 641                   mask_input,
 642                   counter);
 643
 644       for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++)
 645          for(chan = 0; chan < NUM_CHANNELS; ++chan)
 646             fs_out_color[cbuf][chan][i] = out_color[cbuf][chan];
 647    }
 648
 649    sampler->destroy(sampler);
 650
 651    /* Loop over color outputs / color buffers to do blending.
 652     */
 653    for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
 654       LLVMValueRef color_ptr;
 655       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), cbuf, 0);
 656       LLVMValueRef blend_in_color[NUM_CHANNELS];
 657       unsigned rt;
 658
 659       /*
 660        * Convert the fs's output color and mask to fit to the blending type.
 661        */
 662       for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 663          lp_build_conv(builder, fs_type, blend_type,
 664                        fs_out_color[cbuf][chan], num_fs,
 665                        &blend_in_color[chan], 1);
 666          lp_build_name(blend_in_color[chan], "color%d.%c", cbuf, "rgba"[chan]);
 667       }
 668
 669       if (partial_mask || !variant->opaque) {
 670          lp_build_conv_mask(builder, fs_type, blend_type,
 671                             fs_mask, num_fs,
 672                             &blend_mask, 1);
 673       } else {
 674          blend_mask = lp_build_const_int_vec(blend_type, ~0);
 675       }
 676
 677       color_ptr = LLVMBuildLoad(builder,
 678                                 LLVMBuildGEP(builder, color_ptr_ptr, &index, 1, ""),
 679                                 "");
 680       lp_build_name(color_ptr, "color_ptr%d", cbuf);
 681
 682       /* which blend/colormask state to use */
 683       rt = key->blend.independent_blend_enable ? cbuf : 0;
 684
 685       /*
 686        * Blending.
 687        */
 688       generate_blend(&key->blend,
 689                      rt,
 690                      builder,
 691                      blend_type,
 692                      context_ptr,
 693                      blend_mask,
 694                      blend_in_color,
 695                      color_ptr);
 696    }
 697
 698 #ifdef PIPE_ARCH_X86
 699    /* Avoid corrupting the FPU stack on 32bit OSes. */
 700    lp_build_intrinsic(builder, "llvm.x86.mmx.emms", LLVMVoidType(), NULL, 0);
 701 #endif
 702
 703    LLVMBuildRetVoid(builder);
 704
 705    LLVMDisposeBuilder(builder);
 706
 707
 708    /* Verify the LLVM IR.  If invalid, dump and abort */
 709 #ifdef DEBUG
 710    if(LLVMVerifyFunction(function, LLVMPrintMessageAction)) {
 711       if (1)
 712          lp_debug_dump_value(function);
 713       abort();
 714    }
 715 #endif
 716
 717    /* Apply optimizations to LLVM IR */
 718    LLVMRunFunctionPassManager(screen->pass, function);
 719
 720    if (gallivm_debug & GALLIVM_DEBUG_IR) {
 721       /* Print the LLVM IR to stderr */
 722       lp_debug_dump_value(function);
 723       debug_printf("\n");
 724    }
 725
 726    /*
 727     * Translate the LLVM IR into machine code.
 728     */
 729    {
 730       void *f = LLVMGetPointerToGlobal(screen->engine, function);
 731
 732       variant->jit_function[partial_mask] = (lp_jit_frag_func)pointer_to_func(f);
 733
 734       if (gallivm_debug & GALLIVM_DEBUG_ASM) {
 735          lp_disassemble(f);
 736       }
 737       lp_func_delete_body(function);
 738    }
 739 }
 740
 741
 742 static void
 743 dump_fs_variant_key(const struct lp_fragment_shader_variant_key *key)
 744 {
 745    unsigned i;
 746
 747    debug_printf("fs variant %p:\n", (void *) key);
 748
 749    for (i = 0; i < key->nr_cbufs; ++i) {
 750       debug_printf("cbuf_format[%u] = %s\n", i, util_format_name(key->cbuf_format[i]));
 751    }
 752    if (key->depth.enabled) {
 753       debug_printf("depth.format = %s\n", util_format_name(key->zsbuf_format));
 754       debug_printf("depth.func = %s\n", util_dump_func(key->depth.func, TRUE));
 755       debug_printf("depth.writemask = %u\n", key->depth.writemask);
 756    }
 757
 758    for (i = 0; i < 2; ++i) {
 759       if (key->stencil[i].enabled) {
 760          debug_printf("stencil[%u].func = %s\n", i, util_dump_func(key->stencil[i].func, TRUE));
 761          debug_printf("stencil[%u].fail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].fail_op, TRUE));
 762          debug_printf("stencil[%u].zpass_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zpass_op, TRUE));
 763          debug_printf("stencil[%u].zfail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zfail_op, TRUE));
 764          debug_printf("stencil[%u].valuemask = 0x%x\n", i, key->stencil[i].valuemask);
 765          debug_printf("stencil[%u].writemask = 0x%x\n", i, key->stencil[i].writemask);
 766       }
 767    }
 768
 769    if (key->alpha.enabled) {
 770       debug_printf("alpha.func = %s\n", util_dump_func(key->alpha.func, TRUE));
 771    }
 772
 773    if (key->blend.logicop_enable) {
 774       debug_printf("blend.logicop_func = %s\n", util_dump_logicop(key->blend.logicop_func, TRUE));
 775    }
 776    else if (key->blend.rt[0].blend_enable) {
 777       debug_printf("blend.rgb_func = %s\n",   util_dump_blend_func  (key->blend.rt[0].rgb_func, TRUE));
 778       debug_printf("blend.rgb_src_factor = %s\n",   util_dump_blend_factor(key->blend.rt[0].rgb_src_factor, TRUE));
 779       debug_printf("blend.rgb_dst_factor = %s\n",   util_dump_blend_factor(key->blend.rt[0].rgb_dst_factor, TRUE));
 780       debug_printf("blend.alpha_func = %s\n",       util_dump_blend_func  (key->blend.rt[0].alpha_func, TRUE));
 781       debug_printf("blend.alpha_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_src_factor, TRUE));
 782       debug_printf("blend.alpha_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_dst_factor, TRUE));
 783    }
 784    debug_printf("blend.colormask = 0x%x\n", key->blend.rt[0].colormask);
 785    for (i = 0; i < PIPE_MAX_SAMPLERS; ++i) {
 786       if (key->sampler[i].format) {
 787          debug_printf("sampler[%u] = \n", i);
 788          debug_printf("  .format = %s\n",
 789                       util_format_name(key->sampler[i].format));
 790          debug_printf("  .target = %s\n",
 791                       util_dump_tex_target(key->sampler[i].target, TRUE));
 792          debug_printf("  .pot = %u %u %u\n",
 793                       key->sampler[i].pot_width,
 794                       key->sampler[i].pot_height,
 795                       key->sampler[i].pot_depth);
 796          debug_printf("  .wrap = %s %s %s\n",
 797                       util_dump_tex_wrap(key->sampler[i].wrap_s, TRUE),
 798                       util_dump_tex_wrap(key->sampler[i].wrap_t, TRUE),
 799                       util_dump_tex_wrap(key->sampler[i].wrap_r, TRUE));
 800          debug_printf("  .min_img_filter = %s\n",
 801                       util_dump_tex_filter(key->sampler[i].min_img_filter, TRUE));
 802          debug_printf("  .min_mip_filter = %s\n",
 803                       util_dump_tex_mipfilter(key->sampler[i].min_mip_filter, TRUE));
 804          debug_printf("  .mag_img_filter = %s\n",
 805                       util_dump_tex_filter(key->sampler[i].mag_img_filter, TRUE));
 806          if (key->sampler[i].compare_mode != PIPE_TEX_COMPARE_NONE)
 807             debug_printf("  .compare_func = %s\n", util_dump_func(key->sampler[i].compare_func, TRUE));
 808          debug_printf("  .normalized_coords = %u\n", key->sampler[i].normalized_coords);
 809       }
 810    }
 811 }
 812
 813
 814 void
 815 lp_debug_fs_variant(const struct lp_fragment_shader_variant *variant)
 816 {
 817    debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
 818                 variant->shader->no, variant->no);
 819    tgsi_dump(variant->shader->base.tokens, 0);
 820    dump_fs_variant_key(&variant->key);
 821    debug_printf("variant->opaque = %u\n", variant->opaque);
 822    debug_printf("\n");
 823 }
 824
 825 static struct lp_fragment_shader_variant *
 826 generate_variant(struct llvmpipe_context *lp,
 827                  struct lp_fragment_shader *shader,
 828                  const struct lp_fragment_shader_variant_key *key)
 829 {
 830    struct lp_fragment_shader_variant *variant;
 831    boolean fullcolormask;
 832
 833    variant = CALLOC_STRUCT(lp_fragment_shader_variant);
 834    if(!variant)
 835       return NULL;
 836
 837    variant->shader = shader;
 838    variant->list_item_global.base = variant;
 839    variant->list_item_local.base = variant;
 840    variant->no = shader->variants_created++;
 841
 842    memcpy(&variant->key, key, shader->variant_key_size);
 843
 844    /*
 845     * Determine whether we are touching all channels in the color buffer.
 846     */
 847    fullcolormask = FALSE;
 848    if (key->nr_cbufs == 1) {
 849       const struct util_format_description *format_desc;
 850       format_desc = util_format_description(key->cbuf_format[0]);
 851       if ((~key->blend.rt[0].colormask &
 852            util_format_colormask(format_desc)) == 0) {
 853          fullcolormask = TRUE;
 854       }
 855    }
 856
 857    variant->opaque =
 858          !key->blend.logicop_enable &&
 859          !key->blend.rt[0].blend_enable &&
 860          fullcolormask &&
 861          !key->stencil[0].enabled &&
 862          !key->alpha.enabled &&
 863          !key->depth.enabled &&
 864          !shader->info.uses_kill
 865          ? TRUE : FALSE;
 866
 867
 868    if ((LP_DEBUG & DEBUG_FS) || (gallivm_debug & GALLIVM_DEBUG_IR)) {
 869       lp_debug_fs_variant(variant);
 870    }
 871
 872    generate_fragment(lp, shader, variant, RAST_EDGE_TEST);
 873
 874    if (variant->opaque) {
 875       /* Specialized shader, which doesn't need to read the color buffer. */
 876       generate_fragment(lp, shader, variant, RAST_WHOLE);
 877    } else {
 878       variant->jit_function[RAST_WHOLE] = variant->jit_function[RAST_EDGE_TEST];
 879    }
 880
 881    return variant;
 882 }
 883
 884
 885 static void *
 886 llvmpipe_create_fs_state(struct pipe_context *pipe,
 887                          const struct pipe_shader_state *templ)
 888 {
 889    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
 890    struct lp_fragment_shader *shader;
 891    int nr_samplers;
 892
 893    shader = CALLOC_STRUCT(lp_fragment_shader);
 894    if (!shader)
 895       return NULL;
 896
 897    shader->no = fs_no++;
 898    make_empty_list(&shader->variants);
 899
 900    /* get/save the summary info for this shader */
 901    tgsi_scan_shader(templ->tokens, &shader->info);
 902
 903    /* we need to keep a local copy of the tokens */
 904    shader->base.tokens = tgsi_dup_tokens(templ->tokens);
 905
 906    shader->draw_data = draw_create_fragment_shader(llvmpipe->draw, templ);
 907    if (shader->draw_data == NULL) {
 908       FREE((void *) shader->base.tokens);
 909       return NULL;
 910    }
 911
 912    nr_samplers = shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
 913
 914    shader->variant_key_size = Offset(struct lp_fragment_shader_variant_key,
 915                                      sampler[nr_samplers]);
 916
 917    if (LP_DEBUG & DEBUG_TGSI) {
 918       unsigned attrib;
 919       debug_printf("llvmpipe: Create fragment shader #%u %p:\n", shader->no, (void *) shader);
 920       tgsi_dump(templ->tokens, 0);
 921       debug_printf("usage masks:\n");
 922       for (attrib = 0; attrib < shader->info.num_inputs; ++attrib) {
 923          unsigned usage_mask = shader->info.input_usage_mask[attrib];
 924          debug_printf("  IN[%u].%s%s%s%s\n",
 925                       attrib,
 926                       usage_mask & TGSI_WRITEMASK_X ? "x" : "",
 927                       usage_mask & TGSI_WRITEMASK_Y ? "y" : "",
 928                       usage_mask & TGSI_WRITEMASK_Z ? "z" : "",
 929                       usage_mask & TGSI_WRITEMASK_W ? "w" : "");
 930       }
 931       debug_printf("\n");
 932    }
 933
 934    return shader;
 935 }
 936
 937
 938 static void
 939 llvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs)
 940 {
 941    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
 942
 943    if (llvmpipe->fs == fs)
 944       return;
 945
 946    draw_flush(llvmpipe->draw);
 947
 948    draw_bind_fragment_shader(llvmpipe->draw,
 949                              (llvmpipe->fs ? llvmpipe->fs->draw_data : NULL));
 950
 951    llvmpipe->fs = fs;
 952
 953    llvmpipe->dirty |= LP_NEW_FS;
 954 }
 955
 956 static void
 957 remove_shader_variant(struct llvmpipe_context *lp,
 958                       struct lp_fragment_shader_variant *variant)
 959 {
 960    struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
 961    unsigned i;
 962
 963    if (gallivm_debug & GALLIVM_DEBUG_IR) {
 964       debug_printf("llvmpipe: del fs #%u var #%u v created #%u v cached #%u v total cached #%u\n",
 965                     variant->shader->no, variant->no, variant->shader->variants_created,
 966                     variant->shader->variants_cached, lp->nr_fs_variants);
 967    }
 968    for (i = 0; i < Elements(variant->function); i++) {
 969       if (variant->function[i]) {
 970          if (variant->jit_function[i])
 971             LLVMFreeMachineCodeForFunction(screen->engine,
 972                                            variant->function[i]);
 973          LLVMDeleteFunction(variant->function[i]);
 974       }
 975    }
 976    remove_from_list(&variant->list_item_local);
 977    variant->shader->variants_cached--;
 978    remove_from_list(&variant->list_item_global);
 979    lp->nr_fs_variants--;
 980    FREE(variant);
 981 }
 982
 983 static void
 984 llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs)
 985 {
 986    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
 987    struct lp_fragment_shader *shader = fs;
 988    struct lp_fs_variant_list_item *li;
 989
 990    assert(fs != llvmpipe->fs);
 991    (void) llvmpipe;
 992
 993    /*
 994     * XXX: we need to flush the context until we have some sort of reference
 995     * counting in fragment shaders as they may still be binned
 996     * Flushing alone might not sufficient we need to wait on it too.
 997     */
 998
 999    llvmpipe_finish(pipe, __FUNCTION__);
1000
1001    li = first_elem(&shader->variants);
1002    while(!at_end(&shader->variants, li)) {
1003       struct lp_fs_variant_list_item *next = next_elem(li);
1004       remove_shader_variant(llvmpipe, li->base);
1005       li = next;
1006    }
1007
1008    draw_delete_fragment_shader(llvmpipe->draw, shader->draw_data);
1009
1010    assert(shader->variants_cached == 0);
1011    FREE((void *) shader->base.tokens);
1012    FREE(shader);
1013 }
1014
1015
1016
1017 static void
1018 llvmpipe_set_constant_buffer(struct pipe_context *pipe,
1019                              uint shader, uint index,
1020                              struct pipe_resource *constants)
1021 {
1022    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
1023    unsigned size = constants ? constants->width0 : 0;
1024    const void *data = constants ? llvmpipe_resource_data(constants) : NULL;
1025
1026    assert(shader < PIPE_SHADER_TYPES);
1027    assert(index < PIPE_MAX_CONSTANT_BUFFERS);
1028
1029    if(llvmpipe->constants[shader][index] == constants)
1030       return;
1031
1032    draw_flush(llvmpipe->draw);
1033
1034    /* note: reference counting */
1035    pipe_resource_reference(&llvmpipe->constants[shader][index], constants);
1036
1037    if(shader == PIPE_SHADER_VERTEX ||
1038       shader == PIPE_SHADER_GEOMETRY) {
1039       draw_set_mapped_constant_buffer(llvmpipe->draw, shader,
1040                                       index, data, size);
1041    }
1042
1043    llvmpipe->dirty |= LP_NEW_CONSTANTS;
1044 }
1045
1046
1047 /**
1048  * Return the blend factor equivalent to a destination alpha of one.
1049  */
1050 static INLINE unsigned
1051 force_dst_alpha_one(unsigned factor, boolean alpha)
1052 {
1053    switch(factor) {
1054    case PIPE_BLENDFACTOR_DST_ALPHA:
1055       return PIPE_BLENDFACTOR_ONE;
1056    case PIPE_BLENDFACTOR_INV_DST_ALPHA:
1057       return PIPE_BLENDFACTOR_ZERO;
1058    case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
1059       return PIPE_BLENDFACTOR_ZERO;
1060    }
1061
1062    if (alpha) {
1063       switch(factor) {
1064       case PIPE_BLENDFACTOR_DST_COLOR:
1065          return PIPE_BLENDFACTOR_ONE;
1066       case PIPE_BLENDFACTOR_INV_DST_COLOR:
1067          return PIPE_BLENDFACTOR_ZERO;
1068       }
1069    }
1070
1071    return factor;
1072 }
1073
1074
1075 /**
1076  * We need to generate several variants of the fragment pipeline to match
1077  * all the combinations of the contributing state atoms.
1078  *
1079  * TODO: there is actually no reason to tie this to context state -- the
1080  * generated code could be cached globally in the screen.
1081  */
1082 static void
1083 make_variant_key(struct llvmpipe_context *lp,
1084                  struct lp_fragment_shader *shader,
1085                  struct lp_fragment_shader_variant_key *key)
1086 {
1087    unsigned i;
1088
1089    memset(key, 0, shader->variant_key_size);
1090
1091    if (lp->framebuffer.zsbuf) {
1092       if (lp->depth_stencil->depth.enabled) {
1093          key->zsbuf_format = lp->framebuffer.zsbuf->format;
1094          memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth);
1095       }
1096       if (lp->depth_stencil->stencil[0].enabled) {
1097          key->zsbuf_format = lp->framebuffer.zsbuf->format;
1098          memcpy(&key->stencil, &lp->depth_stencil->stencil, sizeof key->stencil);
1099       }
1100    }
1101
1102    key->alpha.enabled = lp->depth_stencil->alpha.enabled;
1103    if(key->alpha.enabled)
1104       key->alpha.func = lp->depth_stencil->alpha.func;
1105    /* alpha.ref_value is passed in jit_context */
1106
1107    key->flatshade = lp->rasterizer->flatshade;
1108    if (lp->active_query_count) {
1109       key->occlusion_count = TRUE;
1110    }
1111
1112    if (lp->framebuffer.nr_cbufs) {
1113       memcpy(&key->blend, lp->blend, sizeof key->blend);
1114    }
1115
1116    key->nr_cbufs = lp->framebuffer.nr_cbufs;
1117    for (i = 0; i < lp->framebuffer.nr_cbufs; i++) {
1118       enum pipe_format format = lp->framebuffer.cbufs[i]->format;
1119       struct pipe_rt_blend_state *blend_rt = &key->blend.rt[i];
1120       const struct util_format_description *format_desc;
1121
1122       key->cbuf_format[i] = format;
1123
1124       format_desc = util_format_description(format);
1125       assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
1126              format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB);
1127
1128       blend_rt->colormask = lp->blend->rt[i].colormask;
1129
1130       /*
1131        * Mask out color channels not present in the color buffer.
1132        */
1133       blend_rt->colormask &= util_format_colormask(format_desc);
1134
1135       /*
1136        * Our swizzled render tiles always have an alpha channel, but the linear
1137        * render target format often does not, so force here the dst alpha to be
1138        * one.
1139        *
1140        * This is not a mere optimization. Wrong results will be produced if the
1141        * dst alpha is used, the dst format does not have alpha, and the previous
1142        * rendering was not flushed from the swizzled to linear buffer. For
1143        * example, NonPowTwo DCT.
1144        *
1145        * TODO: This should be generalized to all channels for better
1146        * performance, but only alpha causes correctness issues.
1147        */
1148       if (format_desc->swizzle[3] > UTIL_FORMAT_SWIZZLE_W) {
1149          blend_rt->rgb_src_factor = force_dst_alpha_one(blend_rt->rgb_src_factor, FALSE);
1150          blend_rt->rgb_dst_factor = force_dst_alpha_one(blend_rt->rgb_dst_factor, FALSE);
1151          blend_rt->alpha_src_factor = force_dst_alpha_one(blend_rt->alpha_src_factor, TRUE);
1152          blend_rt->alpha_dst_factor = force_dst_alpha_one(blend_rt->alpha_dst_factor, TRUE);
1153       }
1154    }
1155
1156    /* This value will be the same for all the variants of a given shader:
1157     */
1158    key->nr_samplers = shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
1159
1160    for(i = 0; i < key->nr_samplers; ++i) {
1161       if(shader->info.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
1162          lp_sampler_static_state(&key->sampler[i],
1163                                  lp->fragment_sampler_views[i],
1164                                  lp->sampler[i]);
1165       }
1166    }
1167 }
1168
1169 /**
1170  * Update fragment state.  This is called just prior to drawing
1171  * something when some fragment-related state has changed.
1172  */
1173 void
1174 llvmpipe_update_fs(struct llvmpipe_context *lp)
1175 {
1176    struct lp_fragment_shader *shader = lp->fs;
1177    struct lp_fragment_shader_variant_key key;
1178    struct lp_fragment_shader_variant *variant = NULL;
1179    struct lp_fs_variant_list_item *li;
1180
1181    make_variant_key(lp, shader, &key);
1182
1183    li = first_elem(&shader->variants);
1184    while(!at_end(&shader->variants, li)) {
1185       if(memcmp(&li->base->key, &key, shader->variant_key_size) == 0) {
1186          variant = li->base;
1187          break;
1188       }
1189       li = next_elem(li);
1190    }
1191
1192    if (variant) {
1193       move_to_head(&lp->fs_variants_list, &variant->list_item_global);
1194    }
1195    else {
1196       int64_t t0, t1;
1197       int64_t dt;
1198       unsigned i;
1199       if (lp->nr_fs_variants >= LP_MAX_SHADER_VARIANTS) {
1200          struct pipe_context *pipe = &lp->pipe;
1201
1202          /*
1203           * XXX: we need to flush the context until we have some sort of reference
1204           * counting in fragment shaders as they may still be binned
1205           * Flushing alone might not be sufficient we need to wait on it too.
1206           */
1207          llvmpipe_finish(pipe, __FUNCTION__);
1208
1209          for (i = 0; i < LP_MAX_SHADER_VARIANTS / 4; i++) {
1210             struct lp_fs_variant_list_item *item = last_elem(&lp->fs_variants_list);
1211             remove_shader_variant(lp, item->base);
1212          }
1213       }
1214       t0 = os_time_get();
1215
1216       variant = generate_variant(lp, shader, &key);
1217
1218       t1 = os_time_get();
1219       dt = t1 - t0;
1220       LP_COUNT_ADD(llvm_compile_time, dt);
1221       LP_COUNT_ADD(nr_llvm_compiles, 2);  /* emit vs. omit in/out test */
1222
1223       if (variant) {
1224          insert_at_head(&shader->variants, &variant->list_item_local);
1225          insert_at_head(&lp->fs_variants_list, &variant->list_item_global);
1226          lp->nr_fs_variants++;
1227          shader->variants_cached++;
1228       }
1229    }
1230
1231    lp_setup_set_fs_variant(lp->setup, variant);
1232 }
1233
1234
1235
1236 void
1237 llvmpipe_init_fs_funcs(struct llvmpipe_context *llvmpipe)
1238 {
1239    llvmpipe->pipe.create_fs_state = llvmpipe_create_fs_state;
1240    llvmpipe->pipe.bind_fs_state   = llvmpipe_bind_fs_state;
1241    llvmpipe->pipe.delete_fs_state = llvmpipe_delete_fs_state;
1242
1243    llvmpipe->pipe.set_constant_buffer = llvmpipe_set_constant_buffer;
1244 }