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  * - stipple (TBI)
  35  * - early depth test
  36  * - fragment shader
  37  * - alpha test
  38  * - depth/stencil test (stencil TBI)
  39  * - blending
  40  *
  41  * This file has only the glue to assembly the fragment pipeline.  The actual
  42  * plumbing of converting Gallium state into LLVM IR is done elsewhere, in the
  43  * lp_bld_*.[ch] files, and in a complete generic and reusable way. Here we
  44  * muster the LLVM JIT execution engine to create a function that follows an
  45  * established binary interface and that can be called from C directly.
  46  *
  47  * A big source of complexity here is that we often want to run different
  48  * stages with different precisions and data types and precisions. For example,
  49  * the fragment shader needs typically to be done in floats, but the
  50  * depth/stencil test and blending is better done in the type that most closely
  51  * matches the depth/stencil and color buffer respectively.
  52  *
  53  * Since the width of a SIMD vector register stays the same regardless of the
  54  * element type, different types imply different number of elements, so we must
  55  * code generate more instances of the stages with larger types to be able to
  56  * feed/consume the stages with smaller types.
  57  *
  58  * @author Jose Fonseca <jfonseca@vmware.com>
  59  */
  60
  61 #include "pipe/p_defines.h"
  62 #include "util/u_memory.h"
  63 #include "util/u_format.h"
  64 #include "util/u_debug_dump.h"
  65 #include "pipe/internal/p_winsys_screen.h"
  66 #include "pipe/p_shader_tokens.h"
  67 #include "draw/draw_context.h"
  68 #include "tgsi/tgsi_dump.h"
  69 #include "tgsi/tgsi_scan.h"
  70 #include "tgsi/tgsi_parse.h"
  71 #include "lp_bld_type.h"
  72 #include "lp_bld_const.h"
  73 #include "lp_bld_conv.h"
  74 #include "lp_bld_intr.h"
  75 #include "lp_bld_logic.h"
  76 #include "lp_bld_depth.h"
  77 #include "lp_bld_interp.h"
  78 #include "lp_bld_tgsi.h"
  79 #include "lp_bld_alpha.h"
  80 #include "lp_bld_blend.h"
  81 #include "lp_bld_swizzle.h"
  82 #include "lp_bld_flow.h"
  83 #include "lp_bld_debug.h"
  84 #include "lp_screen.h"
  85 #include "lp_context.h"
  86 #include "lp_state.h"
  87 #include "lp_quad.h"
  88 #include "lp_tex_sample.h"
  89
  90
  91 static const unsigned char quad_offset_x[4] = {0, 1, 0, 1};
  92 static const unsigned char quad_offset_y[4] = {0, 0, 1, 1};
  93
  94
  95 /*
  96  * Derive from the quad's upper left scalar coordinates the coordinates for
  97  * all other quad pixels
  98  */
  99 static void
 100 generate_pos0(LLVMBuilderRef builder,
 101               LLVMValueRef x,
 102               LLVMValueRef y,
 103               LLVMValueRef *x0,
 104               LLVMValueRef *y0)
 105 {
 106    LLVMTypeRef int_elem_type = LLVMInt32Type();
 107    LLVMTypeRef int_vec_type = LLVMVectorType(int_elem_type, QUAD_SIZE);
 108    LLVMTypeRef elem_type = LLVMFloatType();
 109    LLVMTypeRef vec_type = LLVMVectorType(elem_type, QUAD_SIZE);
 110    LLVMValueRef x_offsets[QUAD_SIZE];
 111    LLVMValueRef y_offsets[QUAD_SIZE];
 112    unsigned i;
 113
 114    x = lp_build_broadcast(builder, int_vec_type, x);
 115    y = lp_build_broadcast(builder, int_vec_type, y);
 116
 117    for(i = 0; i < QUAD_SIZE; ++i) {
 118       x_offsets[i] = LLVMConstInt(int_elem_type, quad_offset_x[i], 0);
 119       y_offsets[i] = LLVMConstInt(int_elem_type, quad_offset_y[i], 0);
 120    }
 121
 122    x = LLVMBuildAdd(builder, x, LLVMConstVector(x_offsets, QUAD_SIZE), "");
 123    y = LLVMBuildAdd(builder, y, LLVMConstVector(y_offsets, QUAD_SIZE), "");
 124
 125    *x0 = LLVMBuildSIToFP(builder, x, vec_type, "");
 126    *y0 = LLVMBuildSIToFP(builder, y, vec_type, "");
 127 }
 128
 129
 130 /**
 131  * Generate the depth test.
 132  */
 133 static void
 134 generate_depth(LLVMBuilderRef builder,
 135                const struct lp_fragment_shader_variant_key *key,
 136                union lp_type src_type,
 137                struct lp_build_mask_context *mask,
 138                LLVMValueRef src,
 139                LLVMValueRef dst_ptr)
 140 {
 141    const struct util_format_description *format_desc;
 142    union lp_type dst_type;
 143
 144    if(!key->depth.enabled)
 145       return;
 146
 147    format_desc = util_format_description(key->zsbuf_format);
 148    assert(format_desc);
 149
 150    /* Pick the depth type. */
 151    dst_type = lp_depth_type(format_desc, src_type.width*src_type.length);
 152
 153    /* FIXME: Cope with a depth test type with a different bit width. */
 154    assert(dst_type.width == src_type.width);
 155    assert(dst_type.length == src_type.length);
 156
 157 #if 1
 158    src = lp_build_clamped_float_to_unsigned_norm(builder,
 159                                                  src_type,
 160                                                  dst_type.width,
 161                                                  src);
 162 #else
 163    lp_build_conv(builder, src_type, dst_type, &src, 1, &src, 1);
 164 #endif
 165
 166    lp_build_depth_test(builder,
 167                        &key->depth,
 168                        dst_type,
 169                        format_desc,
 170                        mask,
 171                        src,
 172                        dst_ptr);
 173 }
 174
 175
 176 /**
 177  * Generate the fragment shader, depth/stencil test, and alpha tests.
 178  */
 179 static void
 180 generate_fs(struct llvmpipe_context *lp,
 181             struct lp_fragment_shader *shader,
 182             const struct lp_fragment_shader_variant_key *key,
 183             LLVMBuilderRef builder,
 184             union lp_type type,
 185             LLVMValueRef context_ptr,
 186             unsigned i,
 187             const struct lp_build_interp_soa_context *interp,
 188             struct lp_build_sampler_soa *sampler,
 189             LLVMValueRef *pmask,
 190             LLVMValueRef *color,
 191             LLVMValueRef depth_ptr)
 192 {
 193    const struct tgsi_token *tokens = shader->base.tokens;
 194    LLVMTypeRef elem_type;
 195    LLVMTypeRef vec_type;
 196    LLVMTypeRef int_vec_type;
 197    LLVMValueRef consts_ptr;
 198    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][NUM_CHANNELS];
 199    LLVMValueRef z = interp->pos[2];
 200    struct lp_build_mask_context mask;
 201    boolean early_depth_test;
 202    unsigned attrib;
 203    unsigned chan;
 204
 205    elem_type = lp_build_elem_type(type);
 206    vec_type = lp_build_vec_type(type);
 207    int_vec_type = lp_build_int_vec_type(type);
 208
 209    consts_ptr = lp_jit_context_constants(builder, context_ptr);
 210
 211    lp_build_mask_begin(&mask, builder, type, *pmask);
 212
 213    early_depth_test =
 214       key->depth.enabled &&
 215       !key->alpha.enabled &&
 216       !shader->info.uses_kill &&
 217       !shader->info.writes_z;
 218
 219    if(early_depth_test)
 220       generate_depth(builder, key,
 221                      type, &mask,
 222                      z, depth_ptr);
 223
 224    memset(outputs, 0, sizeof outputs);
 225
 226    lp_build_tgsi_soa(builder, tokens, type, &mask,
 227                      consts_ptr, interp->pos, interp->inputs,
 228                      outputs, sampler);
 229
 230    for (attrib = 0; attrib < shader->info.num_outputs; ++attrib) {
 231       for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 232          if(outputs[attrib][chan]) {
 233             lp_build_name(outputs[attrib][chan], "output%u.%u.%c", i, attrib, "xyzw"[chan]);
 234
 235             switch (shader->info.output_semantic_name[attrib]) {
 236             case TGSI_SEMANTIC_COLOR:
 237                {
 238                   unsigned cbuf = shader->info.output_semantic_index[attrib];
 239
 240                   lp_build_name(outputs[attrib][chan], "color%u.%u.%c", i, attrib, "rgba"[chan]);
 241
 242                   /* Alpha test */
 243                   /* XXX: should the alpha reference value be passed separately? */
 244                   if(cbuf == 0 && chan == 3) {
 245                      LLVMValueRef alpha = outputs[attrib][chan];
 246                      LLVMValueRef alpha_ref_value;
 247                      alpha_ref_value = lp_jit_context_alpha_ref_value(builder, context_ptr);
 248                      alpha_ref_value = lp_build_broadcast(builder, vec_type, alpha_ref_value);
 249                      lp_build_alpha_test(builder, &key->alpha, type,
 250                                          &mask, alpha, alpha_ref_value);
 251                   }
 252
 253                   if(cbuf == 0)
 254                      color[chan] = outputs[attrib][chan];
 255
 256                   break;
 257                }
 258
 259             case TGSI_SEMANTIC_POSITION:
 260                if(chan == 2)
 261                   z = outputs[attrib][chan];
 262                break;
 263             }
 264          }
 265       }
 266    }
 267
 268    if(!early_depth_test)
 269       generate_depth(builder, key,
 270                      type, &mask,
 271                      z, depth_ptr);
 272
 273    lp_build_mask_end(&mask);
 274
 275    *pmask = mask.value;
 276
 277 }
 278
 279
 280 /**
 281  * Generate color blending and color output.
 282  */
 283 static void
 284 generate_blend(const struct pipe_blend_state *blend,
 285                LLVMBuilderRef builder,
 286                union lp_type type,
 287                LLVMValueRef context_ptr,
 288                LLVMValueRef mask,
 289                LLVMValueRef *src,
 290                LLVMValueRef dst_ptr)
 291 {
 292    struct lp_build_context bld;
 293    LLVMTypeRef vec_type;
 294    LLVMTypeRef int_vec_type;
 295    LLVMValueRef const_ptr;
 296    LLVMValueRef con[4];
 297    LLVMValueRef dst[4];
 298    LLVMValueRef res[4];
 299    unsigned chan;
 300
 301    vec_type = lp_build_vec_type(type);
 302    int_vec_type = lp_build_int_vec_type(type);
 303
 304    lp_build_context_init(&bld, builder, type);
 305
 306    const_ptr = lp_jit_context_blend_color(builder, context_ptr);
 307    const_ptr = LLVMBuildBitCast(builder, const_ptr,
 308                                 LLVMPointerType(vec_type, 0), "");
 309
 310    for(chan = 0; chan < 4; ++chan) {
 311       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
 312       con[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, const_ptr, &index, 1, ""), "");
 313
 314       dst[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, dst_ptr, &index, 1, ""), "");
 315
 316       lp_build_name(con[chan], "con.%c", "rgba"[chan]);
 317       lp_build_name(dst[chan], "dst.%c", "rgba"[chan]);
 318    }
 319
 320    lp_build_blend_soa(builder, blend, type, src, dst, con, res);
 321
 322    for(chan = 0; chan < 4; ++chan) {
 323       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
 324       lp_build_name(res[chan], "res.%c", "rgba"[chan]);
 325       res[chan] = lp_build_select(&bld, mask, res[chan], dst[chan]);
 326       LLVMBuildStore(builder, res[chan], LLVMBuildGEP(builder, dst_ptr, &index, 1, ""));
 327    }
 328 }
 329
 330
 331 /**
 332  * Generate the runtime callable function for the whole fragment pipeline.
 333  */
 334 static struct lp_fragment_shader_variant *
 335 generate_fragment(struct llvmpipe_context *lp,
 336                   struct lp_fragment_shader *shader,
 337                   const struct lp_fragment_shader_variant_key *key)
 338 {
 339    struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
 340    struct lp_fragment_shader_variant *variant;
 341    union lp_type fs_type;
 342    union lp_type blend_type;
 343    LLVMTypeRef fs_elem_type;
 344    LLVMTypeRef fs_vec_type;
 345    LLVMTypeRef fs_int_vec_type;
 346    LLVMTypeRef blend_vec_type;
 347    LLVMTypeRef blend_int_vec_type;
 348    LLVMTypeRef arg_types[9];
 349    LLVMTypeRef func_type;
 350    LLVMValueRef context_ptr;
 351    LLVMValueRef x;
 352    LLVMValueRef y;
 353    LLVMValueRef a0_ptr;
 354    LLVMValueRef dadx_ptr;
 355    LLVMValueRef dady_ptr;
 356    LLVMValueRef mask_ptr;
 357    LLVMValueRef color_ptr;
 358    LLVMValueRef depth_ptr;
 359    LLVMBasicBlockRef block;
 360    LLVMBuilderRef builder;
 361    LLVMValueRef x0;
 362    LLVMValueRef y0;
 363    struct lp_build_sampler_soa *sampler;
 364    struct lp_build_interp_soa_context interp;
 365    LLVMValueRef fs_mask[LP_MAX_VECTOR_LENGTH];
 366    LLVMValueRef fs_out_color[NUM_CHANNELS][LP_MAX_VECTOR_LENGTH];
 367    LLVMValueRef blend_mask;
 368    LLVMValueRef blend_in_color[NUM_CHANNELS];
 369    unsigned num_fs;
 370    unsigned i;
 371    unsigned chan;
 372
 373 #ifdef DEBUG
 374    tgsi_dump(shader->base.tokens, 0);
 375    if(key->depth.enabled) {
 376       debug_printf("depth.func = %s\n", debug_dump_func(key->depth.func, TRUE));
 377       debug_printf("depth.writemask = %u\n", key->depth.writemask);
 378       debug_printf("depth.occlusion_count = %u\n", key->depth.occlusion_count);
 379    }
 380    if(key->alpha.enabled) {
 381       debug_printf("alpha.func = %s\n", debug_dump_func(key->alpha.func, TRUE));
 382       debug_printf("alpha.ref_value = %f\n", key->alpha.ref_value);
 383    }
 384    if(key->blend.logicop_enable) {
 385       debug_printf("blend.logicop_func = %u\n", key->blend.logicop_func);
 386    }
 387    else if(key->blend.blend_enable) {
 388       debug_printf("blend.rgb_func = %s\n",   debug_dump_blend_func  (key->blend.rgb_func, TRUE));
 389       debug_printf("rgb_src_factor = %s\n",   debug_dump_blend_factor(key->blend.rgb_src_factor, TRUE));
 390       debug_printf("rgb_dst_factor = %s\n",   debug_dump_blend_factor(key->blend.rgb_dst_factor, TRUE));
 391       debug_printf("alpha_func = %s\n",       debug_dump_blend_func  (key->blend.alpha_func, TRUE));
 392       debug_printf("alpha_src_factor = %s\n", debug_dump_blend_factor(key->blend.alpha_src_factor, TRUE));
 393       debug_printf("alpha_dst_factor = %s\n", debug_dump_blend_factor(key->blend.alpha_dst_factor, TRUE));
 394    }
 395    debug_printf("blend.colormask = 0x%x\n", key->blend.colormask);
 396 #endif
 397
 398    variant = CALLOC_STRUCT(lp_fragment_shader_variant);
 399    if(!variant)
 400       return NULL;
 401
 402    variant->shader = shader;
 403    memcpy(&variant->key, key, sizeof *key);
 404
 405    /* TODO: actually pick these based on the fs and color buffer
 406     * characteristics. */
 407
 408    fs_type.value = 0;
 409    fs_type.floating = TRUE; /* floating point values */
 410    fs_type.sign = TRUE;     /* values are signed */
 411    fs_type.norm = FALSE;    /* values are not limited to [0,1] or [-1,1] */
 412    fs_type.width = 32;      /* 32-bit float */
 413    fs_type.length = 4;      /* 4 element per vector */
 414    num_fs = 4;
 415
 416    blend_type.value = 0;
 417    blend_type.floating = FALSE; /* values are integers */
 418    blend_type.sign = FALSE;     /* values are unsigned */
 419    blend_type.norm = TRUE;      /* values are in [0,1] or [-1,1] */
 420    blend_type.width = 8;        /* 8-bit ubyte values */
 421    blend_type.length = 16;      /* 16 elements per vector */
 422
 423    /*
 424     * Generate the function prototype. Any change here must be reflected in
 425     * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
 426     */
 427
 428    fs_elem_type = lp_build_elem_type(fs_type);
 429    fs_vec_type = lp_build_vec_type(fs_type);
 430    fs_int_vec_type = lp_build_int_vec_type(fs_type);
 431
 432    blend_vec_type = lp_build_vec_type(blend_type);
 433    blend_int_vec_type = lp_build_int_vec_type(blend_type);
 434
 435    arg_types[0] = screen->context_ptr_type;            /* context */
 436    arg_types[1] = LLVMInt32Type();                     /* x */
 437    arg_types[2] = LLVMInt32Type();                     /* y */
 438    arg_types[3] = LLVMPointerType(fs_elem_type, 0);    /* a0 */
 439    arg_types[4] = LLVMPointerType(fs_elem_type, 0);    /* dadx */
 440    arg_types[5] = LLVMPointerType(fs_elem_type, 0);    /* dady */
 441    arg_types[6] = LLVMPointerType(fs_int_vec_type, 0); /* mask */
 442    arg_types[7] = LLVMPointerType(blend_vec_type, 0);  /* color */
 443    arg_types[8] = LLVMPointerType(fs_int_vec_type, 0); /* depth */
 444
 445    func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
 446
 447    variant->function = LLVMAddFunction(screen->module, "shader", func_type);
 448    LLVMSetFunctionCallConv(variant->function, LLVMCCallConv);
 449    for(i = 0; i < Elements(arg_types); ++i)
 450       if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
 451          LLVMAddAttribute(LLVMGetParam(variant->function, i), LLVMNoAliasAttribute);
 452
 453    context_ptr  = LLVMGetParam(variant->function, 0);
 454    x            = LLVMGetParam(variant->function, 1);
 455    y            = LLVMGetParam(variant->function, 2);
 456    a0_ptr       = LLVMGetParam(variant->function, 3);
 457    dadx_ptr     = LLVMGetParam(variant->function, 4);
 458    dady_ptr     = LLVMGetParam(variant->function, 5);
 459    mask_ptr     = LLVMGetParam(variant->function, 6);
 460    color_ptr    = LLVMGetParam(variant->function, 7);
 461    depth_ptr    = LLVMGetParam(variant->function, 8);
 462
 463    lp_build_name(context_ptr, "context");
 464    lp_build_name(x, "x");
 465    lp_build_name(y, "y");
 466    lp_build_name(a0_ptr, "a0");
 467    lp_build_name(dadx_ptr, "dadx");
 468    lp_build_name(dady_ptr, "dady");
 469    lp_build_name(mask_ptr, "mask");
 470    lp_build_name(color_ptr, "color");
 471    lp_build_name(depth_ptr, "depth");
 472
 473    /*
 474     * Function body
 475     */
 476
 477    block = LLVMAppendBasicBlock(variant->function, "entry");
 478    builder = LLVMCreateBuilder();
 479    LLVMPositionBuilderAtEnd(builder, block);
 480
 481    generate_pos0(builder, x, y, &x0, &y0);
 482
 483    lp_build_interp_soa_init(&interp, shader->base.tokens, builder, fs_type,
 484                             a0_ptr, dadx_ptr, dady_ptr,
 485                             x0, y0, 2, 0);
 486
 487 #if 0
 488    /* C texture sampling */
 489    sampler = lp_c_sampler_soa_create(context_ptr);
 490 #else
 491    /* code generated texture sampling */
 492    sampler = lp_llvm_sampler_soa_create(key->sampler, context_ptr);
 493 #endif
 494
 495    for(i = 0; i < num_fs; ++i) {
 496       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
 497       LLVMValueRef out_color[NUM_CHANNELS];
 498       LLVMValueRef depth_ptr_i;
 499
 500       if(i != 0)
 501          lp_build_interp_soa_update(&interp);
 502
 503       fs_mask[i] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, mask_ptr, &index, 1, ""), "");
 504       depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &index, 1, "");
 505
 506       generate_fs(lp, shader, key,
 507                   builder,
 508                   fs_type,
 509                   context_ptr,
 510                   i,
 511                   &interp,
 512                   sampler,
 513                   &fs_mask[i],
 514                   out_color,
 515                   depth_ptr_i);
 516
 517       for(chan = 0; chan < NUM_CHANNELS; ++chan)
 518          fs_out_color[chan][i] = out_color[chan];
 519    }
 520
 521    sampler->destroy(sampler);
 522
 523    /*
 524     * Convert the fs's output color and mask to fit to the blending type.
 525     */
 526
 527    for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 528       lp_build_conv(builder, fs_type, blend_type,
 529                     fs_out_color[chan], num_fs,
 530                     &blend_in_color[chan], 1);
 531       lp_build_name(blend_in_color[chan], "color.%c", "rgba"[chan]);
 532
 533    }
 534
 535    lp_build_conv_mask(builder, fs_type, blend_type,
 536                                fs_mask, num_fs,
 537                                &blend_mask, 1);
 538
 539    /*
 540     * Blending.
 541     */
 542
 543    generate_blend(&key->blend,
 544                   builder,
 545                   blend_type,
 546                   context_ptr,
 547                   blend_mask,
 548                   blend_in_color,
 549                   color_ptr);
 550
 551    LLVMBuildRetVoid(builder);
 552
 553    LLVMDisposeBuilder(builder);
 554
 555    /*
 556     * Translate the LLVM IR into machine code.
 557     */
 558
 559    LLVMRunFunctionPassManager(screen->pass, variant->function);
 560
 561 #ifdef DEBUG
 562    LLVMDumpValue(variant->function);
 563    debug_printf("\n");
 564 #endif
 565
 566    if(LLVMVerifyFunction(variant->function, LLVMPrintMessageAction)) {
 567       LLVMDumpValue(variant->function);
 568       abort();
 569    }
 570
 571    variant->jit_function = (lp_jit_frag_func)LLVMGetPointerToGlobal(screen->engine, variant->function);
 572
 573 #ifdef DEBUG
 574    lp_disassemble(variant->jit_function);
 575 #endif
 576
 577    variant->next = shader->variants;
 578    shader->variants = variant;
 579
 580    return variant;
 581 }
 582
 583
 584 void *
 585 llvmpipe_create_fs_state(struct pipe_context *pipe,
 586                          const struct pipe_shader_state *templ)
 587 {
 588    struct lp_fragment_shader *shader;
 589
 590    shader = CALLOC_STRUCT(lp_fragment_shader);
 591    if (!shader)
 592       return NULL;
 593
 594    /* get/save the summary info for this shader */
 595    tgsi_scan_shader(templ->tokens, &shader->info);
 596
 597    /* we need to keep a local copy of the tokens */
 598    shader->base.tokens = tgsi_dup_tokens(templ->tokens);
 599
 600    return shader;
 601 }
 602
 603
 604 void
 605 llvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs)
 606 {
 607    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
 608
 609    llvmpipe->fs = (struct lp_fragment_shader *) fs;
 610
 611    llvmpipe->dirty |= LP_NEW_FS;
 612 }
 613
 614
 615 void
 616 llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs)
 617 {
 618    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
 619    struct llvmpipe_screen *screen = llvmpipe_screen(pipe->screen);
 620    struct lp_fragment_shader *shader = fs;
 621    struct lp_fragment_shader_variant *variant;
 622
 623    assert(fs != llvmpipe->fs);
 624
 625    variant = shader->variants;
 626    while(variant) {
 627       struct lp_fragment_shader_variant *next = variant->next;
 628
 629       if(variant->function) {
 630          if(variant->jit_function)
 631             LLVMFreeMachineCodeForFunction(screen->engine, variant->function);
 632          LLVMDeleteFunction(variant->function);
 633       }
 634
 635       FREE(variant);
 636
 637       variant = next;
 638    }
 639
 640    FREE((void *) shader->base.tokens);
 641    FREE(shader);
 642 }
 643
 644
 645
 646 void
 647 llvmpipe_set_constant_buffer(struct pipe_context *pipe,
 648                              uint shader, uint index,
 649                              const struct pipe_constant_buffer *buf)
 650 {
 651    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
 652
 653    assert(shader < PIPE_SHADER_TYPES);
 654    assert(index == 0);
 655
 656    /* note: reference counting */
 657    pipe_buffer_reference(&llvmpipe->constants[shader].buffer,
 658                          buf ? buf->buffer : NULL);
 659
 660    llvmpipe->dirty |= LP_NEW_CONSTANTS;
 661 }
 662
 663
 664 /**
 665  * We need to generate several variants of the fragment pipeline to match
 666  * all the combinations of the contributing state atoms.
 667  *
 668  * TODO: there is actually no reason to tie this to context state -- the
 669  * generated code could be cached globally in the screen.
 670  */
 671 static void
 672 make_variant_key(struct llvmpipe_context *lp,
 673                  struct lp_fragment_shader *shader,
 674                  struct lp_fragment_shader_variant_key *key)
 675 {
 676    unsigned i;
 677
 678    memset(key, 0, sizeof *key);
 679
 680    if(lp->framebuffer.zsbuf &&
 681       lp->depth_stencil->depth.enabled) {
 682       key->zsbuf_format = lp->framebuffer.zsbuf->format;
 683       memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth);
 684    }
 685
 686    key->alpha.enabled = lp->depth_stencil->alpha.enabled;
 687    if(key->alpha.enabled)
 688       key->alpha.func = lp->depth_stencil->alpha.func;
 689    /* alpha.ref_value is passed in jit_context */
 690
 691    memcpy(&key->blend, lp->blend, sizeof key->blend);
 692
 693    for(i = 0; i < PIPE_MAX_SAMPLERS; ++i)
 694       if(shader->info.file_mask[TGSI_FILE_SAMPLER] & (1 << i))
 695          lp_sampler_static_state(&key->sampler[i], lp->texture[i], lp->sampler[i]);
 696 }
 697
 698
 699 void
 700 llvmpipe_update_fs(struct llvmpipe_context *lp)
 701 {
 702    struct lp_fragment_shader *shader = lp->fs;
 703    struct lp_fragment_shader_variant_key key;
 704    struct lp_fragment_shader_variant *variant;
 705
 706    make_variant_key(lp, shader, &key);
 707
 708    variant = shader->variants;
 709    while(variant) {
 710       if(memcmp(&variant->key, &key, sizeof key) == 0)
 711          break;
 712
 713       variant = variant->next;
 714    }
 715
 716    if(!variant)
 717       variant = generate_fragment(lp, shader, &key);
 718
 719    shader->current = variant;
 720 }