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
 190    /*
 191     * XXX: We'll need a different path for 16 x u8
 192     */
 193    assert(fs_type.width == 32);
 194    assert(fs_type.length == 4);
 195    mask_type = lp_int_type(fs_type);
 196
 197    /*
 198     * mask_input >>= (quad * 4)
 199     */
 200
 201    mask_input = LLVMBuildLShr(builder,
 202                               mask_input,
 203                               LLVMConstInt(i32t, quad * 4, 0),
 204                               "");
 205
 206    /*
 207     * mask = { mask_input & (1 << i), for i in [0,3] }
 208     */
 209
 210    mask = lp_build_broadcast(builder, lp_build_vec_type(mask_type), mask_input);
 211
 212    bits[0] = LLVMConstInt(i32t, 1 << 0, 0);
 213    bits[1] = LLVMConstInt(i32t, 1 << 1, 0);
 214    bits[2] = LLVMConstInt(i32t, 1 << 2, 0);
 215    bits[3] = LLVMConstInt(i32t, 1 << 3, 0);
 216
 217    mask = LLVMBuildAnd(builder, mask, LLVMConstVector(bits, 4), "");
 218
 219    /*
 220     * mask = mask != 0 ? ~0 : 0
 221     */
 222
 223    mask = lp_build_compare(builder,
 224                            mask_type, PIPE_FUNC_NOTEQUAL,
 225                            mask,
 226                            lp_build_const_int_vec(mask_type, 0));
 227
 228    return mask;
 229 }
 230
 231
 232
 233 /**
 234  * Generate the fragment shader, depth/stencil test, and alpha tests.
 235  * \param i  which quad in the tile, in range [0,3]
 236  * \param partial_mask  if 1, do mask_input testing
 237  */
 238 static void
 239 generate_fs(struct llvmpipe_context *lp,
 240             struct lp_fragment_shader *shader,
 241             const struct lp_fragment_shader_variant_key *key,
 242             LLVMBuilderRef builder,
 243             struct lp_type type,
 244             LLVMValueRef context_ptr,
 245             unsigned i,
 246             const struct lp_build_interp_soa_context *interp,
 247             struct lp_build_sampler_soa *sampler,
 248             LLVMValueRef *pmask,
 249             LLVMValueRef (*color)[4],
 250             LLVMValueRef depth_ptr,
 251             LLVMValueRef facing,
 252             unsigned partial_mask,
 253             LLVMValueRef mask_input,
 254             LLVMValueRef counter)
 255 {
 256    const struct tgsi_token *tokens = shader->base.tokens;
 257    LLVMTypeRef vec_type;
 258    LLVMValueRef consts_ptr;
 259    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][NUM_CHANNELS];
 260    LLVMValueRef z = interp->pos[2];
 261    LLVMValueRef stencil_refs[2];
 262    struct lp_build_flow_context *flow;
 263    struct lp_build_mask_context mask;
 264    boolean early_depth_stencil_test;
 265    unsigned attrib;
 266    unsigned chan;
 267    unsigned cbuf;
 268
 269    assert(i < 4);
 270
 271    stencil_refs[0] = lp_jit_context_stencil_ref_front_value(builder, context_ptr);
 272    stencil_refs[1] = lp_jit_context_stencil_ref_back_value(builder, context_ptr);
 273
 274    vec_type = lp_build_vec_type(type);
 275
 276    consts_ptr = lp_jit_context_constants(builder, context_ptr);
 277
 278    flow = lp_build_flow_create(builder);
 279
 280    memset(outputs, 0, sizeof outputs);
 281
 282    lp_build_flow_scope_begin(flow);
 283
 284    /* Declare the color and z variables */
 285    for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
 286       for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 287          color[cbuf][chan] = LLVMGetUndef(vec_type);
 288          lp_build_flow_scope_declare(flow, &color[cbuf][chan]);
 289       }
 290    }
 291    lp_build_flow_scope_declare(flow, &z);
 292
 293    /* do triangle edge testing */
 294    if (partial_mask) {
 295       *pmask = generate_quad_mask(builder, type,
 296                                   i, mask_input);
 297    }
 298    else {
 299       *pmask = lp_build_const_int_vec(type, ~0);
 300    }
 301
 302    /* 'mask' will control execution based on quad's pixel alive/killed state */
 303    lp_build_mask_begin(&mask, flow, type, *pmask);
 304
 305    early_depth_stencil_test =
 306       (key->depth.enabled || key->stencil[0].enabled) &&
 307       !key->alpha.enabled &&
 308       !shader->info.uses_kill &&
 309       !shader->info.writes_z;
 310
 311    if (early_depth_stencil_test)
 312       generate_depth_stencil(builder, key,
 313                              type, &mask,
 314                              stencil_refs, z, depth_ptr, facing, counter);
 315
 316    lp_build_tgsi_soa(builder, tokens, type, &mask,
 317                      consts_ptr, interp->pos, interp->inputs,
 318                      outputs, sampler, &shader->info);
 319
 320    /* loop over fragment shader outputs/results */
 321    for (attrib = 0; attrib < shader->info.num_outputs; ++attrib) {
 322       for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 323          if(outputs[attrib][chan]) {
 324             LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
 325             lp_build_name(out, "output%u.%u.%c", i, attrib, "xyzw"[chan]);
 326
 327             switch (shader->info.output_semantic_name[attrib]) {
 328             case TGSI_SEMANTIC_COLOR:
 329                {
 330                   unsigned cbuf = shader->info.output_semantic_index[attrib];
 331
 332                   lp_build_name(out, "color%u.%u.%c", i, attrib, "rgba"[chan]);
 333
 334                   /* Alpha test */
 335                   /* XXX: should only test the final assignment to alpha */
 336                   if (cbuf == 0 && chan == 3 && key->alpha.enabled) {
 337                      LLVMValueRef alpha = out;
 338                      LLVMValueRef alpha_ref_value;
 339                      alpha_ref_value = lp_jit_context_alpha_ref_value(builder, context_ptr);
 340                      alpha_ref_value = lp_build_broadcast(builder, vec_type, alpha_ref_value);
 341                      lp_build_alpha_test(builder, key->alpha.func, type,
 342                                          &mask, alpha, alpha_ref_value);
 343                   }
 344
 345                   color[cbuf][chan] = out;
 346                   break;
 347                }
 348
 349             case TGSI_SEMANTIC_POSITION:
 350                if(chan == 2)
 351                   z = out;
 352                break;
 353             }
 354          }
 355       }
 356    }
 357
 358    if (!early_depth_stencil_test)
 359       generate_depth_stencil(builder, key,
 360                              type, &mask,
 361                              stencil_refs, z, depth_ptr, facing, counter);
 362
 363    lp_build_mask_end(&mask);
 364
 365    lp_build_flow_scope_end(flow);
 366
 367    lp_build_flow_destroy(flow);
 368
 369    *pmask = mask.value;
 370
 371 }
 372
 373
 374 /**
 375  * Generate color blending and color output.
 376  * \param rt  the render target index (to index blend, colormask state)
 377  * \param type  the pixel color type
 378  * \param context_ptr  pointer to the runtime JIT context
 379  * \param mask  execution mask (active fragment/pixel mask)
 380  * \param src  colors from the fragment shader
 381  * \param dst_ptr  the destination color buffer pointer
 382  */
 383 static void
 384 generate_blend(const struct pipe_blend_state *blend,
 385                unsigned rt,
 386                LLVMBuilderRef builder,
 387                struct lp_type type,
 388                LLVMValueRef context_ptr,
 389                LLVMValueRef mask,
 390                LLVMValueRef *src,
 391                LLVMValueRef dst_ptr)
 392 {
 393    struct lp_build_context bld;
 394    struct lp_build_flow_context *flow;
 395    struct lp_build_mask_context mask_ctx;
 396    LLVMTypeRef vec_type;
 397    LLVMValueRef const_ptr;
 398    LLVMValueRef con[4];
 399    LLVMValueRef dst[4];
 400    LLVMValueRef res[4];
 401    unsigned chan;
 402
 403    lp_build_context_init(&bld, builder, type);
 404
 405    flow = lp_build_flow_create(builder);
 406
 407    /* we'll use this mask context to skip blending if all pixels are dead */
 408    lp_build_mask_begin(&mask_ctx, flow, type, mask);
 409
 410    vec_type = lp_build_vec_type(type);
 411
 412    const_ptr = lp_jit_context_blend_color(builder, context_ptr);
 413    const_ptr = LLVMBuildBitCast(builder, const_ptr,
 414                                 LLVMPointerType(vec_type, 0), "");
 415
 416    /* load constant blend color and colors from the dest color buffer */
 417    for(chan = 0; chan < 4; ++chan) {
 418       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
 419       con[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, const_ptr, &index, 1, ""), "");
 420
 421       dst[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, dst_ptr, &index, 1, ""), "");
 422
 423       lp_build_name(con[chan], "con.%c", "rgba"[chan]);
 424       lp_build_name(dst[chan], "dst.%c", "rgba"[chan]);
 425    }
 426
 427    /* do blend */
 428    lp_build_blend_soa(builder, blend, type, rt, src, dst, con, res);
 429
 430    /* store results to color buffer */
 431    for(chan = 0; chan < 4; ++chan) {
 432       if(blend->rt[rt].colormask & (1 << chan)) {
 433          LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
 434          lp_build_name(res[chan], "res.%c", "rgba"[chan]);
 435          res[chan] = lp_build_select(&bld, mask, res[chan], dst[chan]);
 436          LLVMBuildStore(builder, res[chan], LLVMBuildGEP(builder, dst_ptr, &index, 1, ""));
 437       }
 438    }
 439
 440    lp_build_mask_end(&mask_ctx);
 441    lp_build_flow_destroy(flow);
 442 }
 443
 444
 445 /**
 446  * Generate the runtime callable function for the whole fragment pipeline.
 447  * Note that the function which we generate operates on a block of 16
 448  * pixels at at time.  The block contains 2x2 quads.  Each quad contains
 449  * 2x2 pixels.
 450  */
 451 static void
 452 generate_fragment(struct llvmpipe_context *lp,
 453                   struct lp_fragment_shader *shader,
 454                   struct lp_fragment_shader_variant *variant,
 455                   unsigned partial_mask)
 456 {
 457    struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
 458    const struct lp_fragment_shader_variant_key *key = &variant->key;
 459    char func_name[256];
 460    struct lp_type fs_type;
 461    struct lp_type blend_type;
 462    LLVMTypeRef fs_elem_type;
 463    LLVMTypeRef fs_int_vec_type;
 464    LLVMTypeRef blend_vec_type;
 465    LLVMTypeRef arg_types[11];
 466    LLVMTypeRef func_type;
 467    LLVMValueRef context_ptr;
 468    LLVMValueRef x;
 469    LLVMValueRef y;
 470    LLVMValueRef a0_ptr;
 471    LLVMValueRef dadx_ptr;
 472    LLVMValueRef dady_ptr;
 473    LLVMValueRef color_ptr_ptr;
 474    LLVMValueRef depth_ptr;
 475    LLVMValueRef mask_input;
 476    LLVMValueRef counter = NULL;
 477    LLVMBasicBlockRef block;
 478    LLVMBuilderRef builder;
 479    struct lp_build_sampler_soa *sampler;
 480    struct lp_build_interp_soa_context interp;
 481    LLVMValueRef fs_mask[LP_MAX_VECTOR_LENGTH];
 482    LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS][LP_MAX_VECTOR_LENGTH];
 483    LLVMValueRef blend_mask;
 484    LLVMValueRef function;
 485    LLVMValueRef facing;
 486    unsigned num_fs;
 487    unsigned i;
 488    unsigned chan;
 489    unsigned cbuf;
 490
 491
 492    /* TODO: actually pick these based on the fs and color buffer
 493     * characteristics. */
 494
 495    memset(&fs_type, 0, sizeof fs_type);
 496    fs_type.floating = TRUE; /* floating point values */
 497    fs_type.sign = TRUE;     /* values are signed */
 498    fs_type.norm = FALSE;    /* values are not limited to [0,1] or [-1,1] */
 499    fs_type.width = 32;      /* 32-bit float */
 500    fs_type.length = 4;      /* 4 elements per vector */
 501    num_fs = 4;              /* number of quads per block */
 502
 503    memset(&blend_type, 0, sizeof blend_type);
 504    blend_type.floating = FALSE; /* values are integers */
 505    blend_type.sign = FALSE;     /* values are unsigned */
 506    blend_type.norm = TRUE;      /* values are in [0,1] or [-1,1] */
 507    blend_type.width = 8;        /* 8-bit ubyte values */
 508    blend_type.length = 16;      /* 16 elements per vector */
 509
 510    /*
 511     * Generate the function prototype. Any change here must be reflected in
 512     * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
 513     */
 514
 515    fs_elem_type = lp_build_elem_type(fs_type);
 516    fs_int_vec_type = lp_build_int_vec_type(fs_type);
 517
 518    blend_vec_type = lp_build_vec_type(blend_type);
 519
 520    util_snprintf(func_name, sizeof(func_name), "fs%u_variant%u_%s",
 521                  shader->no, variant->no, partial_mask ? "partial" : "whole");
 522
 523    arg_types[0] = screen->context_ptr_type;            /* context */
 524    arg_types[1] = LLVMInt32Type();                     /* x */
 525    arg_types[2] = LLVMInt32Type();                     /* y */
 526    arg_types[3] = LLVMFloatType();                     /* facing */
 527    arg_types[4] = LLVMPointerType(fs_elem_type, 0);    /* a0 */
 528    arg_types[5] = LLVMPointerType(fs_elem_type, 0);    /* dadx */
 529    arg_types[6] = LLVMPointerType(fs_elem_type, 0);    /* dady */
 530    arg_types[7] = LLVMPointerType(LLVMPointerType(blend_vec_type, 0), 0);  /* color */
 531    arg_types[8] = LLVMPointerType(fs_int_vec_type, 0); /* depth */
 532    arg_types[9] = LLVMInt32Type();                     /* mask_input */
 533    arg_types[10] = LLVMPointerType(LLVMInt32Type(), 0);/* counter */
 534
 535    func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
 536
 537    function = LLVMAddFunction(screen->module, func_name, func_type);
 538    LLVMSetFunctionCallConv(function, LLVMCCallConv);
 539
 540    variant->function[partial_mask] = function;
 541
 542
 543    /* XXX: need to propagate noalias down into color param now we are
 544     * passing a pointer-to-pointer?
 545     */
 546    for(i = 0; i < Elements(arg_types); ++i)
 547       if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
 548          LLVMAddAttribute(LLVMGetParam(function, i), LLVMNoAliasAttribute);
 549
 550    context_ptr  = LLVMGetParam(function, 0);
 551    x            = LLVMGetParam(function, 1);
 552    y            = LLVMGetParam(function, 2);
 553    facing       = LLVMGetParam(function, 3);
 554    a0_ptr       = LLVMGetParam(function, 4);
 555    dadx_ptr     = LLVMGetParam(function, 5);
 556    dady_ptr     = LLVMGetParam(function, 6);
 557    color_ptr_ptr = LLVMGetParam(function, 7);
 558    depth_ptr    = LLVMGetParam(function, 8);
 559    mask_input   = LLVMGetParam(function, 9);
 560
 561    lp_build_name(context_ptr, "context");
 562    lp_build_name(x, "x");
 563    lp_build_name(y, "y");
 564    lp_build_name(a0_ptr, "a0");
 565    lp_build_name(dadx_ptr, "dadx");
 566    lp_build_name(dady_ptr, "dady");
 567    lp_build_name(color_ptr_ptr, "color_ptr_ptr");
 568    lp_build_name(depth_ptr, "depth");
 569    lp_build_name(mask_input, "mask_input");
 570
 571    if (key->occlusion_count) {
 572       counter = LLVMGetParam(function, 10);
 573       lp_build_name(counter, "counter");
 574    }
 575
 576    /*
 577     * Function body
 578     */
 579
 580    block = LLVMAppendBasicBlock(function, "entry");
 581    builder = LLVMCreateBuilder();
 582    LLVMPositionBuilderAtEnd(builder, block);
 583
 584    /*
 585     * The shader input interpolation info is not explicitely baked in the
 586     * shader key, but everything it derives from (TGSI, and flatshade) is
 587     * already included in the shader key.
 588     */
 589    lp_build_interp_soa_init(&interp,
 590                             lp->num_inputs,
 591                             lp->inputs,
 592                             builder, fs_type,
 593                             a0_ptr, dadx_ptr, dady_ptr,
 594                             x, y);
 595
 596    /* code generated texture sampling */
 597    sampler = lp_llvm_sampler_soa_create(key->sampler, context_ptr);
 598
 599    /* loop over quads in the block */
 600    for(i = 0; i < num_fs; ++i) {
 601       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
 602       LLVMValueRef out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS];
 603       LLVMValueRef depth_ptr_i;
 604
 605       if(i != 0)
 606          lp_build_interp_soa_update(&interp, i);
 607
 608       depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &index, 1, "");
 609
 610       generate_fs(lp, shader, key,
 611                   builder,
 612                   fs_type,
 613                   context_ptr,
 614                   i,
 615                   &interp,
 616                   sampler,
 617                   &fs_mask[i], /* output */
 618                   out_color,
 619                   depth_ptr_i,
 620                   facing,
 621                   partial_mask,
 622                   mask_input,
 623                   counter);
 624
 625       for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++)
 626          for(chan = 0; chan < NUM_CHANNELS; ++chan)
 627             fs_out_color[cbuf][chan][i] = out_color[cbuf][chan];
 628    }
 629
 630    sampler->destroy(sampler);
 631
 632    /* Loop over color outputs / color buffers to do blending.
 633     */
 634    for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
 635       LLVMValueRef color_ptr;
 636       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), cbuf, 0);
 637       LLVMValueRef blend_in_color[NUM_CHANNELS];
 638       unsigned rt;
 639
 640       /*
 641        * Convert the fs's output color and mask to fit to the blending type.
 642        */
 643       for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 644          lp_build_conv(builder, fs_type, blend_type,
 645                        fs_out_color[cbuf][chan], num_fs,
 646                        &blend_in_color[chan], 1);
 647          lp_build_name(blend_in_color[chan], "color%d.%c", cbuf, "rgba"[chan]);
 648       }
 649
 650       if (partial_mask || !variant->opaque) {
 651          lp_build_conv_mask(builder, fs_type, blend_type,
 652                             fs_mask, num_fs,
 653                             &blend_mask, 1);
 654       } else {
 655          blend_mask = lp_build_const_int_vec(blend_type, ~0);
 656       }
 657
 658       color_ptr = LLVMBuildLoad(builder,
 659                                 LLVMBuildGEP(builder, color_ptr_ptr, &index, 1, ""),
 660                                 "");
 661       lp_build_name(color_ptr, "color_ptr%d", cbuf);
 662
 663       /* which blend/colormask state to use */
 664       rt = key->blend.independent_blend_enable ? cbuf : 0;
 665
 666       /*
 667        * Blending.
 668        */
 669       generate_blend(&key->blend,
 670                      rt,
 671                      builder,
 672                      blend_type,
 673                      context_ptr,
 674                      blend_mask,
 675                      blend_in_color,
 676                      color_ptr);
 677    }
 678
 679 #ifdef PIPE_ARCH_X86
 680    /* Avoid corrupting the FPU stack on 32bit OSes. */
 681    lp_build_intrinsic(builder, "llvm.x86.mmx.emms", LLVMVoidType(), NULL, 0);
 682 #endif
 683
 684    LLVMBuildRetVoid(builder);
 685
 686    LLVMDisposeBuilder(builder);
 687
 688
 689    /* Verify the LLVM IR.  If invalid, dump and abort */
 690 #ifdef DEBUG
 691    if(LLVMVerifyFunction(function, LLVMPrintMessageAction)) {
 692       if (1)
 693          lp_debug_dump_value(function);
 694       abort();
 695    }
 696 #endif
 697
 698    /* Apply optimizations to LLVM IR */
 699    LLVMRunFunctionPassManager(screen->pass, function);
 700
 701    if (gallivm_debug & GALLIVM_DEBUG_IR) {
 702       /* Print the LLVM IR to stderr */
 703       lp_debug_dump_value(function);
 704       debug_printf("\n");
 705    }
 706
 707    /*
 708     * Translate the LLVM IR into machine code.
 709     */
 710    {
 711       void *f = LLVMGetPointerToGlobal(screen->engine, function);
 712
 713       variant->jit_function[partial_mask] = (lp_jit_frag_func)pointer_to_func(f);
 714
 715       if (gallivm_debug & GALLIVM_DEBUG_ASM) {
 716          lp_disassemble(f);
 717       }
 718       lp_func_delete_body(function);
 719    }
 720 }
 721
 722
 723 static void
 724 dump_fs_variant_key(const struct lp_fragment_shader_variant_key *key)
 725 {
 726    unsigned i;
 727
 728    debug_printf("fs variant %p:\n", (void *) key);
 729
 730    for (i = 0; i < key->nr_cbufs; ++i) {
 731       debug_printf("cbuf_format[%u] = %s\n", i, util_format_name(key->cbuf_format[i]));
 732    }
 733    if (key->depth.enabled) {
 734       debug_printf("depth.format = %s\n", util_format_name(key->zsbuf_format));
 735       debug_printf("depth.func = %s\n", util_dump_func(key->depth.func, TRUE));
 736       debug_printf("depth.writemask = %u\n", key->depth.writemask);
 737    }
 738
 739    for (i = 0; i < 2; ++i) {
 740       if (key->stencil[i].enabled) {
 741          debug_printf("stencil[%u].func = %s\n", i, util_dump_func(key->stencil[i].func, TRUE));
 742          debug_printf("stencil[%u].fail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].fail_op, TRUE));
 743          debug_printf("stencil[%u].zpass_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zpass_op, TRUE));
 744          debug_printf("stencil[%u].zfail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zfail_op, TRUE));
 745          debug_printf("stencil[%u].valuemask = 0x%x\n", i, key->stencil[i].valuemask);
 746          debug_printf("stencil[%u].writemask = 0x%x\n", i, key->stencil[i].writemask);
 747       }
 748    }
 749
 750    if (key->alpha.enabled) {
 751       debug_printf("alpha.func = %s\n", util_dump_func(key->alpha.func, TRUE));
 752    }
 753
 754    if (key->blend.logicop_enable) {
 755       debug_printf("blend.logicop_func = %s\n", util_dump_logicop(key->blend.logicop_func, TRUE));
 756    }
 757    else if (key->blend.rt[0].blend_enable) {
 758       debug_printf("blend.rgb_func = %s\n",   util_dump_blend_func  (key->blend.rt[0].rgb_func, TRUE));
 759       debug_printf("blend.rgb_src_factor = %s\n",   util_dump_blend_factor(key->blend.rt[0].rgb_src_factor, TRUE));
 760       debug_printf("blend.rgb_dst_factor = %s\n",   util_dump_blend_factor(key->blend.rt[0].rgb_dst_factor, TRUE));
 761       debug_printf("blend.alpha_func = %s\n",       util_dump_blend_func  (key->blend.rt[0].alpha_func, TRUE));
 762       debug_printf("blend.alpha_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_src_factor, TRUE));
 763       debug_printf("blend.alpha_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_dst_factor, TRUE));
 764    }
 765    debug_printf("blend.colormask = 0x%x\n", key->blend.rt[0].colormask);
 766    for (i = 0; i < PIPE_MAX_SAMPLERS; ++i) {
 767       if (key->sampler[i].format) {
 768          debug_printf("sampler[%u] = \n", i);
 769          debug_printf("  .format = %s\n",
 770                       util_format_name(key->sampler[i].format));
 771          debug_printf("  .target = %s\n",
 772                       util_dump_tex_target(key->sampler[i].target, TRUE));
 773          debug_printf("  .pot = %u %u %u\n",
 774                       key->sampler[i].pot_width,
 775                       key->sampler[i].pot_height,
 776                       key->sampler[i].pot_depth);
 777          debug_printf("  .wrap = %s %s %s\n",
 778                       util_dump_tex_wrap(key->sampler[i].wrap_s, TRUE),
 779                       util_dump_tex_wrap(key->sampler[i].wrap_t, TRUE),
 780                       util_dump_tex_wrap(key->sampler[i].wrap_r, TRUE));
 781          debug_printf("  .min_img_filter = %s\n",
 782                       util_dump_tex_filter(key->sampler[i].min_img_filter, TRUE));
 783          debug_printf("  .min_mip_filter = %s\n",
 784                       util_dump_tex_mipfilter(key->sampler[i].min_mip_filter, TRUE));
 785          debug_printf("  .mag_img_filter = %s\n",
 786                       util_dump_tex_filter(key->sampler[i].mag_img_filter, TRUE));
 787          if (key->sampler[i].compare_mode != PIPE_TEX_COMPARE_NONE)
 788             debug_printf("  .compare_func = %s\n", util_dump_func(key->sampler[i].compare_func, TRUE));
 789          debug_printf("  .normalized_coords = %u\n", key->sampler[i].normalized_coords);
 790       }
 791    }
 792 }
 793
 794
 795 void
 796 lp_debug_fs_variant(const struct lp_fragment_shader_variant *variant)
 797 {
 798    debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
 799                 variant->shader->no, variant->no);
 800    tgsi_dump(variant->shader->base.tokens, 0);
 801    dump_fs_variant_key(&variant->key);
 802    debug_printf("variant->opaque = %u\n", variant->opaque);
 803    debug_printf("\n");
 804 }
 805
 806 static struct lp_fragment_shader_variant *
 807 generate_variant(struct llvmpipe_context *lp,
 808                  struct lp_fragment_shader *shader,
 809                  const struct lp_fragment_shader_variant_key *key)
 810 {
 811    struct lp_fragment_shader_variant *variant;
 812    boolean fullcolormask;
 813
 814    variant = CALLOC_STRUCT(lp_fragment_shader_variant);
 815    if(!variant)
 816       return NULL;
 817
 818    variant->shader = shader;
 819    variant->list_item_global.base = variant;
 820    variant->list_item_local.base = variant;
 821    variant->no = shader->variants_created++;
 822
 823    memcpy(&variant->key, key, shader->variant_key_size);
 824
 825    /*
 826     * Determine whether we are touching all channels in the color buffer.
 827     */
 828    fullcolormask = FALSE;
 829    if (key->nr_cbufs == 1) {
 830       const struct util_format_description *format_desc;
 831       format_desc = util_format_description(key->cbuf_format[0]);
 832       if ((~key->blend.rt[0].colormask &
 833            util_format_colormask(format_desc)) == 0) {
 834          fullcolormask = TRUE;
 835       }
 836    }
 837
 838    variant->opaque =
 839          !key->blend.logicop_enable &&
 840          !key->blend.rt[0].blend_enable &&
 841          fullcolormask &&
 842          !key->stencil[0].enabled &&
 843          !key->alpha.enabled &&
 844          !key->depth.enabled &&
 845          !shader->info.uses_kill
 846          ? TRUE : FALSE;
 847
 848
 849    if (gallivm_debug & GALLIVM_DEBUG_IR) {
 850       lp_debug_fs_variant(variant);
 851    }
 852
 853    generate_fragment(lp, shader, variant, RAST_EDGE_TEST);
 854
 855    if (variant->opaque) {
 856       /* Specialized shader, which doesn't need to read the color buffer. */
 857       generate_fragment(lp, shader, variant, RAST_WHOLE);
 858    } else {
 859       variant->jit_function[RAST_WHOLE] = variant->jit_function[RAST_EDGE_TEST];
 860    }
 861
 862    return variant;
 863 }
 864
 865
 866 static void *
 867 llvmpipe_create_fs_state(struct pipe_context *pipe,
 868                          const struct pipe_shader_state *templ)
 869 {
 870    struct lp_fragment_shader *shader;
 871    int nr_samplers;
 872
 873    shader = CALLOC_STRUCT(lp_fragment_shader);
 874    if (!shader)
 875       return NULL;
 876
 877    shader->no = fs_no++;
 878    make_empty_list(&shader->variants);
 879
 880    /* get/save the summary info for this shader */
 881    tgsi_scan_shader(templ->tokens, &shader->info);
 882
 883    /* we need to keep a local copy of the tokens */
 884    shader->base.tokens = tgsi_dup_tokens(templ->tokens);
 885
 886    nr_samplers = shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
 887
 888    shader->variant_key_size = Offset(struct lp_fragment_shader_variant_key,
 889                                      sampler[nr_samplers]);
 890
 891    if (LP_DEBUG & DEBUG_TGSI) {
 892       unsigned attrib;
 893       debug_printf("llvmpipe: Create fragment shader #%u %p:\n", shader->no, (void *) shader);
 894       tgsi_dump(templ->tokens, 0);
 895       debug_printf("usage masks:\n");
 896       for (attrib = 0; attrib < shader->info.num_inputs; ++attrib) {
 897          unsigned usage_mask = shader->info.input_usage_mask[attrib];
 898          debug_printf("  IN[%u].%s%s%s%s\n",
 899                       attrib,
 900                       usage_mask & TGSI_WRITEMASK_X ? "x" : "",
 901                       usage_mask & TGSI_WRITEMASK_Y ? "y" : "",
 902                       usage_mask & TGSI_WRITEMASK_Z ? "z" : "",
 903                       usage_mask & TGSI_WRITEMASK_W ? "w" : "");
 904       }
 905       debug_printf("\n");
 906    }
 907
 908    return shader;
 909 }
 910
 911
 912 static void
 913 llvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs)
 914 {
 915    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
 916
 917    if (llvmpipe->fs == fs)
 918       return;
 919
 920    draw_flush(llvmpipe->draw);
 921
 922    llvmpipe->fs = fs;
 923
 924    llvmpipe->dirty |= LP_NEW_FS;
 925 }
 926
 927 static void
 928 remove_shader_variant(struct llvmpipe_context *lp,
 929                       struct lp_fragment_shader_variant *variant)
 930 {
 931    struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
 932    unsigned i;
 933
 934    if (gallivm_debug & GALLIVM_DEBUG_IR) {
 935       debug_printf("llvmpipe: del fs #%u var #%u v created #%u v cached #%u v total cached #%u\n",
 936                     variant->shader->no, variant->no, variant->shader->variants_created,
 937                     variant->shader->variants_cached, lp->nr_fs_variants);
 938    }
 939    for (i = 0; i < Elements(variant->function); i++) {
 940       if (variant->function[i]) {
 941          if (variant->jit_function[i])
 942             LLVMFreeMachineCodeForFunction(screen->engine,
 943                                            variant->function[i]);
 944          LLVMDeleteFunction(variant->function[i]);
 945       }
 946    }
 947    remove_from_list(&variant->list_item_local);
 948    variant->shader->variants_cached--;
 949    remove_from_list(&variant->list_item_global);
 950    lp->nr_fs_variants--;
 951    FREE(variant);
 952 }
 953
 954 static void
 955 llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs)
 956 {
 957    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
 958    struct lp_fragment_shader *shader = fs;
 959    struct lp_fs_variant_list_item *li;
 960
 961    assert(fs != llvmpipe->fs);
 962    (void) llvmpipe;
 963
 964    /*
 965     * XXX: we need to flush the context until we have some sort of reference
 966     * counting in fragment shaders as they may still be binned
 967     * Flushing alone might not sufficient we need to wait on it too.
 968     */
 969
 970    llvmpipe_finish(pipe, __FUNCTION__);
 971
 972    li = first_elem(&shader->variants);
 973    while(!at_end(&shader->variants, li)) {
 974       struct lp_fs_variant_list_item *next = next_elem(li);
 975       remove_shader_variant(llvmpipe, li->base);
 976       li = next;
 977    }
 978
 979    assert(shader->variants_cached == 0);
 980    FREE((void *) shader->base.tokens);
 981    FREE(shader);
 982 }
 983
 984
 985
 986 static void
 987 llvmpipe_set_constant_buffer(struct pipe_context *pipe,
 988                              uint shader, uint index,
 989                              struct pipe_resource *constants)
 990 {
 991    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
 992    unsigned size = constants ? constants->width0 : 0;
 993    const void *data = constants ? llvmpipe_resource_data(constants) : NULL;
 994
 995    assert(shader < PIPE_SHADER_TYPES);
 996    assert(index < PIPE_MAX_CONSTANT_BUFFERS);
 997
 998    if(llvmpipe->constants[shader][index] == constants)
 999       return;
1000
1001    draw_flush(llvmpipe->draw);
1002
1003    /* note: reference counting */
1004    pipe_resource_reference(&llvmpipe->constants[shader][index], constants);
1005
1006    if(shader == PIPE_SHADER_VERTEX ||
1007       shader == PIPE_SHADER_GEOMETRY) {
1008       draw_set_mapped_constant_buffer(llvmpipe->draw, shader,
1009                                       index, data, size);
1010    }
1011
1012    llvmpipe->dirty |= LP_NEW_CONSTANTS;
1013 }
1014
1015
1016 /**
1017  * Return the blend factor equivalent to a destination alpha of one.
1018  */
1019 static INLINE unsigned
1020 force_dst_alpha_one(unsigned factor, boolean alpha)
1021 {
1022    switch(factor) {
1023    case PIPE_BLENDFACTOR_DST_ALPHA:
1024       return PIPE_BLENDFACTOR_ONE;
1025    case PIPE_BLENDFACTOR_INV_DST_ALPHA:
1026       return PIPE_BLENDFACTOR_ZERO;
1027    case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
1028       return PIPE_BLENDFACTOR_ZERO;
1029    }
1030
1031    if (alpha) {
1032       switch(factor) {
1033       case PIPE_BLENDFACTOR_DST_COLOR:
1034          return PIPE_BLENDFACTOR_ONE;
1035       case PIPE_BLENDFACTOR_INV_DST_COLOR:
1036          return PIPE_BLENDFACTOR_ZERO;
1037       }
1038    }
1039
1040    return factor;
1041 }
1042
1043
1044 /**
1045  * We need to generate several variants of the fragment pipeline to match
1046  * all the combinations of the contributing state atoms.
1047  *
1048  * TODO: there is actually no reason to tie this to context state -- the
1049  * generated code could be cached globally in the screen.
1050  */
1051 static void
1052 make_variant_key(struct llvmpipe_context *lp,
1053                  struct lp_fragment_shader *shader,
1054                  struct lp_fragment_shader_variant_key *key)
1055 {
1056    unsigned i;
1057
1058    memset(key, 0, shader->variant_key_size);
1059
1060    if (lp->framebuffer.zsbuf) {
1061       if (lp->depth_stencil->depth.enabled) {
1062          key->zsbuf_format = lp->framebuffer.zsbuf->format;
1063          memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth);
1064       }
1065       if (lp->depth_stencil->stencil[0].enabled) {
1066          key->zsbuf_format = lp->framebuffer.zsbuf->format;
1067          memcpy(&key->stencil, &lp->depth_stencil->stencil, sizeof key->stencil);
1068       }
1069    }
1070
1071    key->alpha.enabled = lp->depth_stencil->alpha.enabled;
1072    if(key->alpha.enabled)
1073       key->alpha.func = lp->depth_stencil->alpha.func;
1074    /* alpha.ref_value is passed in jit_context */
1075
1076    key->flatshade = lp->rasterizer->flatshade;
1077    if (lp->active_query_count) {
1078       key->occlusion_count = TRUE;
1079    }
1080
1081    if (lp->framebuffer.nr_cbufs) {
1082       memcpy(&key->blend, lp->blend, sizeof key->blend);
1083    }
1084
1085    key->nr_cbufs = lp->framebuffer.nr_cbufs;
1086    for (i = 0; i < lp->framebuffer.nr_cbufs; i++) {
1087       enum pipe_format format = lp->framebuffer.cbufs[i]->format;
1088       struct pipe_rt_blend_state *blend_rt = &key->blend.rt[i];
1089       const struct util_format_description *format_desc;
1090
1091       key->cbuf_format[i] = format;
1092
1093       format_desc = util_format_description(format);
1094       assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
1095              format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB);
1096
1097       blend_rt->colormask = lp->blend->rt[i].colormask;
1098
1099       /*
1100        * Mask out color channels not present in the color buffer.
1101        */
1102       blend_rt->colormask &= util_format_colormask(format_desc);
1103
1104       /*
1105        * Our swizzled render tiles always have an alpha channel, but the linear
1106        * render target format often does not, so force here the dst alpha to be
1107        * one.
1108        *
1109        * This is not a mere optimization. Wrong results will be produced if the
1110        * dst alpha is used, the dst format does not have alpha, and the previous
1111        * rendering was not flushed from the swizzled to linear buffer. For
1112        * example, NonPowTwo DCT.
1113        *
1114        * TODO: This should be generalized to all channels for better
1115        * performance, but only alpha causes correctness issues.
1116        */
1117       if (format_desc->swizzle[3] > UTIL_FORMAT_SWIZZLE_W) {
1118          blend_rt->rgb_src_factor = force_dst_alpha_one(blend_rt->rgb_src_factor, FALSE);
1119          blend_rt->rgb_dst_factor = force_dst_alpha_one(blend_rt->rgb_dst_factor, FALSE);
1120          blend_rt->alpha_src_factor = force_dst_alpha_one(blend_rt->alpha_src_factor, TRUE);
1121          blend_rt->alpha_dst_factor = force_dst_alpha_one(blend_rt->alpha_dst_factor, TRUE);
1122       }
1123    }
1124
1125    /* This value will be the same for all the variants of a given shader:
1126     */
1127    key->nr_samplers = shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
1128
1129    for(i = 0; i < key->nr_samplers; ++i) {
1130       if(shader->info.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
1131          lp_sampler_static_state(&key->sampler[i],
1132                                  lp->fragment_sampler_views[i],
1133                                  lp->sampler[i]);
1134       }
1135    }
1136 }
1137
1138 /**
1139  * Update fragment state.  This is called just prior to drawing
1140  * something when some fragment-related state has changed.
1141  */
1142 void
1143 llvmpipe_update_fs(struct llvmpipe_context *lp)
1144 {
1145    struct lp_fragment_shader *shader = lp->fs;
1146    struct lp_fragment_shader_variant_key key;
1147    struct lp_fragment_shader_variant *variant = NULL;
1148    struct lp_fs_variant_list_item *li;
1149
1150    make_variant_key(lp, shader, &key);
1151
1152    li = first_elem(&shader->variants);
1153    while(!at_end(&shader->variants, li)) {
1154       if(memcmp(&li->base->key, &key, shader->variant_key_size) == 0) {
1155          variant = li->base;
1156          break;
1157       }
1158       li = next_elem(li);
1159    }
1160
1161    if (variant) {
1162       move_to_head(&lp->fs_variants_list, &variant->list_item_global);
1163    }
1164    else {
1165       int64_t t0, t1;
1166       int64_t dt;
1167       unsigned i;
1168       if (lp->nr_fs_variants >= LP_MAX_SHADER_VARIANTS) {
1169          struct pipe_context *pipe = &lp->pipe;
1170
1171          /*
1172           * XXX: we need to flush the context until we have some sort of reference
1173           * counting in fragment shaders as they may still be binned
1174           * Flushing alone might not be sufficient we need to wait on it too.
1175           */
1176          llvmpipe_finish(pipe, __FUNCTION__);
1177
1178          for (i = 0; i < LP_MAX_SHADER_VARIANTS / 4; i++) {
1179             struct lp_fs_variant_list_item *item = last_elem(&lp->fs_variants_list);
1180             remove_shader_variant(lp, item->base);
1181          }
1182       }
1183       t0 = os_time_get();
1184
1185       variant = generate_variant(lp, shader, &key);
1186
1187       t1 = os_time_get();
1188       dt = t1 - t0;
1189       LP_COUNT_ADD(llvm_compile_time, dt);
1190       LP_COUNT_ADD(nr_llvm_compiles, 2);  /* emit vs. omit in/out test */
1191
1192       if (variant) {
1193          insert_at_head(&shader->variants, &variant->list_item_local);
1194          insert_at_head(&lp->fs_variants_list, &variant->list_item_global);
1195          lp->nr_fs_variants++;
1196          shader->variants_cached++;
1197       }
1198    }
1199
1200    lp_setup_set_fs_variant(lp->setup, variant);
1201 }
1202
1203
1204
1205 void
1206 llvmpipe_init_fs_funcs(struct llvmpipe_context *llvmpipe)
1207 {
1208    llvmpipe->pipe.create_fs_state = llvmpipe_create_fs_state;
1209    llvmpipe->pipe.bind_fs_state   = llvmpipe_bind_fs_state;
1210    llvmpipe->pipe.delete_fs_state = llvmpipe_delete_fs_state;
1211
1212    llvmpipe->pipe.set_constant_buffer = llvmpipe_set_constant_buffer;
1213 }