src/gallium/drivers/llvmpipe/lp_state_fs.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2009 VMware, Inc.
   4  * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas.
   5  * All Rights Reserved.
   6  *
   7  * Permission is hereby granted, free of charge, to any person obtaining a
   8  * copy of this software and associated documentation files (the
   9  * "Software"), to deal in the Software without restriction, including
  10  * without limitation the rights to use, copy, modify, merge, publish,
  11  * distribute, sub license, and/or sell copies of the Software, and to
  12  * permit persons to whom the Software is furnished to do so, subject to
  13  * the following conditions:
  14  *
  15  * The above copyright notice and this permission notice (including the
  16  * next paragraph) shall be included in all copies or substantial portions
  17  * of the Software.
  18  *
  19  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  20  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  21  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
  22  * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
  23  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
  24  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
  25  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  26  *
  27  **************************************************************************/
  28
  29 /**
  30  * @file
  31  * Code generate the whole fragment pipeline.
  32  *
  33  * The fragment pipeline consists of the following stages:
  34  * - triangle edge in/out testing
  35  * - scissor test
  36  * - stipple (TBI)
  37  * - early depth test
  38  * - fragment shader
  39  * - alpha test
  40  * - depth/stencil test
  41  * - blending
  42  *
  43  * This file has only the glue to assemble the fragment pipeline.  The actual
  44  * plumbing of converting Gallium state into LLVM IR is done elsewhere, in the
  45  * lp_bld_*.[ch] files, and in a complete generic and reusable way. Here we
  46  * muster the LLVM JIT execution engine to create a function that follows an
  47  * established binary interface and that can be called from C directly.
  48  *
  49  * A big source of complexity here is that we often want to run different
  50  * stages with different precisions and data types and precisions. For example,
  51  * the fragment shader needs typically to be done in floats, but the
  52  * depth/stencil test and blending is better done in the type that most closely
  53  * matches the depth/stencil and color buffer respectively.
  54  *
  55  * Since the width of a SIMD vector register stays the same regardless of the
  56  * element type, different types imply different number of elements, so we must
  57  * code generate more instances of the stages with larger types to be able to
  58  * feed/consume the stages with smaller types.
  59  *
  60  * @author Jose Fonseca <jfonseca@vmware.com>
  61  */
  62
  63 #include <limits.h>
  64 #include "pipe/p_defines.h"
  65 #include "util/u_inlines.h"
  66 #include "util/u_memory.h"
  67 #include "util/u_pointer.h"
  68 #include "util/u_format.h"
  69 #include "util/u_dump.h"
  70 #include "util/u_string.h"
  71 #include "util/u_simple_list.h"
  72 #include "os/os_time.h"
  73 #include "pipe/p_shader_tokens.h"
  74 #include "draw/draw_context.h"
  75 #include "tgsi/tgsi_dump.h"
  76 #include "tgsi/tgsi_scan.h"
  77 #include "tgsi/tgsi_parse.h"
  78 #include "gallivm/lp_bld_type.h"
  79 #include "gallivm/lp_bld_const.h"
  80 #include "gallivm/lp_bld_conv.h"
  81 #include "gallivm/lp_bld_intr.h"
  82 #include "gallivm/lp_bld_logic.h"
  83 #include "gallivm/lp_bld_tgsi.h"
  84 #include "gallivm/lp_bld_swizzle.h"
  85 #include "gallivm/lp_bld_flow.h"
  86 #include "gallivm/lp_bld_debug.h"
  87
  88 #include "lp_bld_alpha.h"
  89 #include "lp_bld_blend.h"
  90 #include "lp_bld_depth.h"
  91 #include "lp_bld_interp.h"
  92 #include "lp_context.h"
  93 #include "lp_debug.h"
  94 #include "lp_perf.h"
  95 #include "lp_screen.h"
  96 #include "lp_setup.h"
  97 #include "lp_state.h"
  98 #include "lp_tex_sample.h"
  99 #include "lp_flush.h"
 100
 101
 102 #include <llvm-c/Analysis.h>
 103
 104
 105 static unsigned fs_no = 0;
 106
 107
 108 /**
 109  * Generate the depth /stencil test code.
 110  */
 111 static void
 112 generate_depth_stencil(LLVMBuilderRef builder,
 113                        const struct lp_fragment_shader_variant_key *key,
 114                        struct lp_type src_type,
 115                        struct lp_build_mask_context *mask,
 116                        LLVMValueRef stencil_refs[2],
 117                        LLVMValueRef src,
 118                        LLVMValueRef dst_ptr,
 119                        LLVMValueRef facing,
 120                        LLVMValueRef counter)
 121 {
 122    const struct util_format_description *format_desc;
 123    struct lp_type dst_type;
 124
 125    if (!key->depth.enabled && !key->stencil[0].enabled && !key->stencil[1].enabled)
 126       return;
 127
 128    format_desc = util_format_description(key->zsbuf_format);
 129    assert(format_desc);
 130
 131    /*
 132     * Depths are expected to be between 0 and 1, even if they are stored in
 133     * floats. Setting these bits here will ensure that the lp_build_conv() call
 134     * below won't try to unnecessarily clamp the incoming values.
 135     */
 136    if(src_type.floating) {
 137       src_type.sign = FALSE;
 138       src_type.norm = TRUE;
 139    }
 140    else {
 141       assert(!src_type.sign);
 142       assert(src_type.norm);
 143    }
 144
 145    /* Pick the depth type. */
 146    dst_type = lp_depth_type(format_desc, src_type.width*src_type.length);
 147
 148    /* FIXME: Cope with a depth test type with a different bit width. */
 149    assert(dst_type.width == src_type.width);
 150    assert(dst_type.length == src_type.length);
 151
 152    /* Convert fragment Z from float to integer */
 153    lp_build_conv(builder, src_type, dst_type, &src, 1, &src, 1);
 154
 155    dst_ptr = LLVMBuildBitCast(builder,
 156                               dst_ptr,
 157                               LLVMPointerType(lp_build_vec_type(dst_type), 0), "");
 158    lp_build_depth_stencil_test(builder,
 159                                &key->depth,
 160                                key->stencil,
 161                                dst_type,
 162                                format_desc,
 163                                mask,
 164                                stencil_refs,
 165                                src,
 166                                dst_ptr,
 167                                facing,
 168                                counter);
 169 }
 170
 171
 172 /**
 173  * Generate the code to do inside/outside triangle testing for the
 174  * four pixels in a 2x2 quad.  This will set the four elements of the
 175  * quad mask vector to 0 or ~0.
 176  * \param i  which quad of the quad group to test, in [0,3]
 177  */
 178 static void
 179 generate_tri_edge_mask(LLVMBuilderRef builder,
 180                        unsigned i,
 181                        LLVMValueRef *mask,      /* ivec4, out */
 182                        LLVMValueRef c0,         /* int32 */
 183                        LLVMValueRef c1,         /* int32 */
 184                        LLVMValueRef c2,         /* int32 */
 185                        LLVMValueRef step0_ptr,  /* ivec4 */
 186                        LLVMValueRef step1_ptr,  /* ivec4 */
 187                        LLVMValueRef step2_ptr)  /* ivec4 */
 188 {
 189 #define OPTIMIZE_IN_OUT_TEST 0
 190 #if OPTIMIZE_IN_OUT_TEST
 191    struct lp_build_if_state ifctx;
 192    LLVMValueRef not_draw_all;
 193 #endif
 194    struct lp_build_flow_context *flow;
 195    struct lp_type i32_type;
 196    LLVMTypeRef i32vec4_type;
 197    LLVMValueRef c0_vec, c1_vec, c2_vec;
 198    LLVMValueRef in_out_mask;
 199
 200    assert(i < 4);
 201
 202    /* int32 vector type */
 203    memset(&i32_type, 0, sizeof i32_type);
 204    i32_type.floating = FALSE; /* values are integers */
 205    i32_type.sign = TRUE;      /* values are signed */
 206    i32_type.norm = FALSE;     /* values are not normalized */
 207    i32_type.width = 32;       /* 32-bit int values */
 208    i32_type.length = 4;       /* 4 elements per vector */
 209
 210    i32vec4_type = lp_build_int32_vec4_type();
 211
 212    /*
 213     * Use a conditional here to do detailed pixel in/out testing.
 214     * We only have to do this if c0 != INT_MIN.
 215     */
 216    flow = lp_build_flow_create(builder);
 217    lp_build_flow_scope_begin(flow);
 218
 219    {
 220 #if OPTIMIZE_IN_OUT_TEST
 221       /* not_draw_all = (c0 != INT_MIN) */
 222       not_draw_all = LLVMBuildICmp(builder,
 223                                    LLVMIntNE,
 224                                    c0,
 225                                    LLVMConstInt(LLVMInt32Type(), INT_MIN, 0),
 226                                    "");
 227
 228       in_out_mask = lp_build_const_int_vec(i32_type, ~0);
 229
 230
 231       lp_build_flow_scope_declare(flow, &in_out_mask);
 232
 233       /* if (not_draw_all) {... */
 234       lp_build_if(&ifctx, flow, builder, not_draw_all);
 235 #endif
 236       {
 237          LLVMValueRef step0_vec, step1_vec, step2_vec;
 238          LLVMValueRef m0_vec, m1_vec, m2_vec;
 239          LLVMValueRef index, m;
 240
 241          /* c0_vec = {c0, c0, c0, c0}
 242           * Note that we emit this code four times but LLVM optimizes away
 243           * three instances of it.
 244           */
 245          c0_vec = lp_build_broadcast(builder, i32vec4_type, c0);
 246          c1_vec = lp_build_broadcast(builder, i32vec4_type, c1);
 247          c2_vec = lp_build_broadcast(builder, i32vec4_type, c2);
 248          lp_build_name(c0_vec, "edgeconst0vec");
 249          lp_build_name(c1_vec, "edgeconst1vec");
 250          lp_build_name(c2_vec, "edgeconst2vec");
 251
 252          /* load step0vec, step1, step2 vec from memory */
 253          index = LLVMConstInt(LLVMInt32Type(), i, 0);
 254          step0_vec = LLVMBuildLoad(builder, LLVMBuildGEP(builder, step0_ptr, &index, 1, ""), "");
 255          step1_vec = LLVMBuildLoad(builder, LLVMBuildGEP(builder, step1_ptr, &index, 1, ""), "");
 256          step2_vec = LLVMBuildLoad(builder, LLVMBuildGEP(builder, step2_ptr, &index, 1, ""), "");
 257          lp_build_name(step0_vec, "step0vec");
 258          lp_build_name(step1_vec, "step1vec");
 259          lp_build_name(step2_vec, "step2vec");
 260
 261          /* m0_vec = step0_ptr[i] > c0_vec */
 262          m0_vec = lp_build_compare(builder, i32_type, PIPE_FUNC_GREATER, step0_vec, c0_vec);
 263          m1_vec = lp_build_compare(builder, i32_type, PIPE_FUNC_GREATER, step1_vec, c1_vec);
 264          m2_vec = lp_build_compare(builder, i32_type, PIPE_FUNC_GREATER, step2_vec, c2_vec);
 265
 266          /* in_out_mask = m0_vec & m1_vec & m2_vec */
 267          m = LLVMBuildAnd(builder, m0_vec, m1_vec, "");
 268          in_out_mask = LLVMBuildAnd(builder, m, m2_vec, "");
 269          lp_build_name(in_out_mask, "inoutmaskvec");
 270       }
 271 #if OPTIMIZE_IN_OUT_TEST
 272       lp_build_endif(&ifctx);
 273 #endif
 274
 275    }
 276    lp_build_flow_scope_end(flow);
 277    lp_build_flow_destroy(flow);
 278
 279    /* This is the initial alive/dead pixel mask for a quad of four pixels.
 280     * It's an int[4] vector with each word set to 0 or ~0.
 281     * Words will get cleared when pixels faile the Z test, etc.
 282     */
 283    *mask = in_out_mask;
 284 }
 285
 286
 287 static LLVMValueRef
 288 generate_scissor_test(LLVMBuilderRef builder,
 289                       LLVMValueRef context_ptr,
 290                       const struct lp_build_interp_soa_context *interp,
 291                       struct lp_type type)
 292 {
 293    LLVMTypeRef vec_type = lp_build_vec_type(type);
 294    LLVMValueRef xpos = interp->pos[0], ypos = interp->pos[1];
 295    LLVMValueRef xmin, ymin, xmax, ymax;
 296    LLVMValueRef m0, m1, m2, m3, m;
 297
 298    /* xpos, ypos contain the window coords for the four pixels in the quad */
 299    assert(xpos);
 300    assert(ypos);
 301
 302    /* get the current scissor bounds, convert to vectors */
 303    xmin = lp_jit_context_scissor_xmin_value(builder, context_ptr);
 304    xmin = lp_build_broadcast(builder, vec_type, xmin);
 305
 306    ymin = lp_jit_context_scissor_ymin_value(builder, context_ptr);
 307    ymin = lp_build_broadcast(builder, vec_type, ymin);
 308
 309    xmax = lp_jit_context_scissor_xmax_value(builder, context_ptr);
 310    xmax = lp_build_broadcast(builder, vec_type, xmax);
 311
 312    ymax = lp_jit_context_scissor_ymax_value(builder, context_ptr);
 313    ymax = lp_build_broadcast(builder, vec_type, ymax);
 314
 315    /* compare the fragment's position coordinates against the scissor bounds */
 316    m0 = lp_build_compare(builder, type, PIPE_FUNC_GEQUAL, xpos, xmin);
 317    m1 = lp_build_compare(builder, type, PIPE_FUNC_GEQUAL, ypos, ymin);
 318    m2 = lp_build_compare(builder, type, PIPE_FUNC_LESS, xpos, xmax);
 319    m3 = lp_build_compare(builder, type, PIPE_FUNC_LESS, ypos, ymax);
 320
 321    /* AND all the masks together */
 322    m = LLVMBuildAnd(builder, m0, m1, "");
 323    m = LLVMBuildAnd(builder, m, m2, "");
 324    m = LLVMBuildAnd(builder, m, m3, "");
 325
 326    lp_build_name(m, "scissormask");
 327
 328    return m;
 329 }
 330
 331
 332 static LLVMValueRef
 333 build_int32_vec_const(int value)
 334 {
 335    struct lp_type i32_type;
 336
 337    memset(&i32_type, 0, sizeof i32_type);
 338    i32_type.floating = FALSE; /* values are integers */
 339    i32_type.sign = TRUE;      /* values are signed */
 340    i32_type.norm = FALSE;     /* values are not normalized */
 341    i32_type.width = 32;       /* 32-bit int values */
 342    i32_type.length = 4;       /* 4 elements per vector */
 343    return lp_build_const_int_vec(i32_type, value);
 344 }
 345
 346
 347
 348 /**
 349  * Generate the fragment shader, depth/stencil test, and alpha tests.
 350  * \param i  which quad in the tile, in range [0,3]
 351  * \param do_tri_test  if 1, do triangle edge in/out testing
 352  */
 353 static void
 354 generate_fs(struct llvmpipe_context *lp,
 355             struct lp_fragment_shader *shader,
 356             const struct lp_fragment_shader_variant_key *key,
 357             LLVMBuilderRef builder,
 358             struct lp_type type,
 359             LLVMValueRef context_ptr,
 360             unsigned i,
 361             const struct lp_build_interp_soa_context *interp,
 362             struct lp_build_sampler_soa *sampler,
 363             LLVMValueRef *pmask,
 364             LLVMValueRef (*color)[4],
 365             LLVMValueRef depth_ptr,
 366             LLVMValueRef facing,
 367             unsigned do_tri_test,
 368             LLVMValueRef c0,
 369             LLVMValueRef c1,
 370             LLVMValueRef c2,
 371             LLVMValueRef step0_ptr,
 372             LLVMValueRef step1_ptr,
 373             LLVMValueRef step2_ptr,
 374             LLVMValueRef counter)
 375 {
 376    const struct tgsi_token *tokens = shader->base.tokens;
 377    LLVMTypeRef vec_type;
 378    LLVMValueRef consts_ptr;
 379    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][NUM_CHANNELS];
 380    LLVMValueRef z = interp->pos[2];
 381    LLVMValueRef stencil_refs[2];
 382    struct lp_build_flow_context *flow;
 383    struct lp_build_mask_context mask;
 384    boolean early_depth_stencil_test;
 385    unsigned attrib;
 386    unsigned chan;
 387    unsigned cbuf;
 388
 389    assert(i < 4);
 390
 391    stencil_refs[0] = lp_jit_context_stencil_ref_front_value(builder, context_ptr);
 392    stencil_refs[1] = lp_jit_context_stencil_ref_back_value(builder, context_ptr);
 393
 394    vec_type = lp_build_vec_type(type);
 395
 396    consts_ptr = lp_jit_context_constants(builder, context_ptr);
 397
 398    flow = lp_build_flow_create(builder);
 399
 400    memset(outputs, 0, sizeof outputs);
 401
 402    lp_build_flow_scope_begin(flow);
 403
 404    /* Declare the color and z variables */
 405    for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
 406       for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 407          color[cbuf][chan] = LLVMGetUndef(vec_type);
 408          lp_build_flow_scope_declare(flow, &color[cbuf][chan]);
 409       }
 410    }
 411    lp_build_flow_scope_declare(flow, &z);
 412
 413    /* do triangle edge testing */
 414    if (do_tri_test) {
 415       generate_tri_edge_mask(builder, i, pmask,
 416                              c0, c1, c2, step0_ptr, step1_ptr, step2_ptr);
 417    }
 418    else {
 419       *pmask = build_int32_vec_const(~0);
 420    }
 421
 422    /* 'mask' will control execution based on quad's pixel alive/killed state */
 423    lp_build_mask_begin(&mask, flow, type, *pmask);
 424
 425    if (key->scissor) {
 426       LLVMValueRef smask =
 427          generate_scissor_test(builder, context_ptr, interp, type);
 428       lp_build_mask_update(&mask, smask);
 429    }
 430
 431    early_depth_stencil_test =
 432       (key->depth.enabled || key->stencil[0].enabled) &&
 433       !key->alpha.enabled &&
 434       !shader->info.uses_kill &&
 435       !shader->info.writes_z;
 436
 437    if (early_depth_stencil_test)
 438       generate_depth_stencil(builder, key,
 439                              type, &mask,
 440                              stencil_refs, z, depth_ptr, facing, counter);
 441
 442    lp_build_tgsi_soa(builder, tokens, type, &mask,
 443                      consts_ptr, interp->pos, interp->inputs,
 444                      outputs, sampler, &shader->info);
 445
 446    /* loop over fragment shader outputs/results */
 447    for (attrib = 0; attrib < shader->info.num_outputs; ++attrib) {
 448       for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 449          if(outputs[attrib][chan]) {
 450             LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
 451             lp_build_name(out, "output%u.%u.%c", i, attrib, "xyzw"[chan]);
 452
 453             switch (shader->info.output_semantic_name[attrib]) {
 454             case TGSI_SEMANTIC_COLOR:
 455                {
 456                   unsigned cbuf = shader->info.output_semantic_index[attrib];
 457
 458                   lp_build_name(out, "color%u.%u.%c", i, attrib, "rgba"[chan]);
 459
 460                   /* Alpha test */
 461                   /* XXX: should the alpha reference value be passed separately? */
 462                   /* XXX: should only test the final assignment to alpha */
 463                   if(cbuf == 0 && chan == 3) {
 464                      LLVMValueRef alpha = out;
 465                      LLVMValueRef alpha_ref_value;
 466                      alpha_ref_value = lp_jit_context_alpha_ref_value(builder, context_ptr);
 467                      alpha_ref_value = lp_build_broadcast(builder, vec_type, alpha_ref_value);
 468                      lp_build_alpha_test(builder, &key->alpha, type,
 469                                          &mask, alpha, alpha_ref_value);
 470                   }
 471
 472                   color[cbuf][chan] = out;
 473                   break;
 474                }
 475
 476             case TGSI_SEMANTIC_POSITION:
 477                if(chan == 2)
 478                   z = out;
 479                break;
 480             }
 481          }
 482       }
 483    }
 484
 485    if (!early_depth_stencil_test)
 486       generate_depth_stencil(builder, key,
 487                              type, &mask,
 488                              stencil_refs, z, depth_ptr, facing, counter);
 489
 490    lp_build_mask_end(&mask);
 491
 492    lp_build_flow_scope_end(flow);
 493
 494    lp_build_flow_destroy(flow);
 495
 496    *pmask = mask.value;
 497
 498 }
 499
 500
 501 /**
 502  * Generate color blending and color output.
 503  * \param rt  the render target index (to index blend, colormask state)
 504  * \param type  the pixel color type
 505  * \param context_ptr  pointer to the runtime JIT context
 506  * \param mask  execution mask (active fragment/pixel mask)
 507  * \param src  colors from the fragment shader
 508  * \param dst_ptr  the destination color buffer pointer
 509  */
 510 static void
 511 generate_blend(const struct pipe_blend_state *blend,
 512                unsigned rt,
 513                LLVMBuilderRef builder,
 514                struct lp_type type,
 515                LLVMValueRef context_ptr,
 516                LLVMValueRef mask,
 517                LLVMValueRef *src,
 518                LLVMValueRef dst_ptr)
 519 {
 520    struct lp_build_context bld;
 521    struct lp_build_flow_context *flow;
 522    struct lp_build_mask_context mask_ctx;
 523    LLVMTypeRef vec_type;
 524    LLVMValueRef const_ptr;
 525    LLVMValueRef con[4];
 526    LLVMValueRef dst[4];
 527    LLVMValueRef res[4];
 528    unsigned chan;
 529
 530    lp_build_context_init(&bld, builder, type);
 531
 532    flow = lp_build_flow_create(builder);
 533
 534    /* we'll use this mask context to skip blending if all pixels are dead */
 535    lp_build_mask_begin(&mask_ctx, flow, type, mask);
 536
 537    vec_type = lp_build_vec_type(type);
 538
 539    const_ptr = lp_jit_context_blend_color(builder, context_ptr);
 540    const_ptr = LLVMBuildBitCast(builder, const_ptr,
 541                                 LLVMPointerType(vec_type, 0), "");
 542
 543    /* load constant blend color and colors from the dest color buffer */
 544    for(chan = 0; chan < 4; ++chan) {
 545       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
 546       con[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, const_ptr, &index, 1, ""), "");
 547
 548       dst[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, dst_ptr, &index, 1, ""), "");
 549
 550       lp_build_name(con[chan], "con.%c", "rgba"[chan]);
 551       lp_build_name(dst[chan], "dst.%c", "rgba"[chan]);
 552    }
 553
 554    /* do blend */
 555    lp_build_blend_soa(builder, blend, type, rt, src, dst, con, res);
 556
 557    /* store results to color buffer */
 558    for(chan = 0; chan < 4; ++chan) {
 559       if(blend->rt[rt].colormask & (1 << chan)) {
 560          LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
 561          lp_build_name(res[chan], "res.%c", "rgba"[chan]);
 562          res[chan] = lp_build_select(&bld, mask, res[chan], dst[chan]);
 563          LLVMBuildStore(builder, res[chan], LLVMBuildGEP(builder, dst_ptr, &index, 1, ""));
 564       }
 565    }
 566
 567    lp_build_mask_end(&mask_ctx);
 568    lp_build_flow_destroy(flow);
 569 }
 570
 571
 572 /**
 573  * Generate the runtime callable function for the whole fragment pipeline.
 574  * Note that the function which we generate operates on a block of 16
 575  * pixels at at time.  The block contains 2x2 quads.  Each quad contains
 576  * 2x2 pixels.
 577  */
 578 static void
 579 generate_fragment(struct llvmpipe_context *lp,
 580                   struct lp_fragment_shader *shader,
 581                   struct lp_fragment_shader_variant *variant,
 582                   unsigned do_tri_test)
 583 {
 584    struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
 585    const struct lp_fragment_shader_variant_key *key = &variant->key;
 586    char func_name[256];
 587    struct lp_type fs_type;
 588    struct lp_type blend_type;
 589    LLVMTypeRef fs_elem_type;
 590    LLVMTypeRef fs_int_vec_type;
 591    LLVMTypeRef blend_vec_type;
 592    LLVMTypeRef arg_types[16];
 593    LLVMTypeRef func_type;
 594    LLVMTypeRef int32_vec4_type = lp_build_int32_vec4_type();
 595    LLVMValueRef context_ptr;
 596    LLVMValueRef x;
 597    LLVMValueRef y;
 598    LLVMValueRef a0_ptr;
 599    LLVMValueRef dadx_ptr;
 600    LLVMValueRef dady_ptr;
 601    LLVMValueRef color_ptr_ptr;
 602    LLVMValueRef depth_ptr;
 603    LLVMValueRef c0, c1, c2, step0_ptr, step1_ptr, step2_ptr, counter = NULL;
 604    LLVMBasicBlockRef block;
 605    LLVMBuilderRef builder;
 606    struct lp_build_sampler_soa *sampler;
 607    struct lp_build_interp_soa_context interp;
 608    LLVMValueRef fs_mask[LP_MAX_VECTOR_LENGTH];
 609    LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS][LP_MAX_VECTOR_LENGTH];
 610    LLVMValueRef blend_mask;
 611    LLVMValueRef function;
 612    LLVMValueRef facing;
 613    unsigned num_fs;
 614    unsigned i;
 615    unsigned chan;
 616    unsigned cbuf;
 617
 618
 619    /* TODO: actually pick these based on the fs and color buffer
 620     * characteristics. */
 621
 622    memset(&fs_type, 0, sizeof fs_type);
 623    fs_type.floating = TRUE; /* floating point values */
 624    fs_type.sign = TRUE;     /* values are signed */
 625    fs_type.norm = FALSE;    /* values are not limited to [0,1] or [-1,1] */
 626    fs_type.width = 32;      /* 32-bit float */
 627    fs_type.length = 4;      /* 4 elements per vector */
 628    num_fs = 4;              /* number of quads per block */
 629
 630    memset(&blend_type, 0, sizeof blend_type);
 631    blend_type.floating = FALSE; /* values are integers */
 632    blend_type.sign = FALSE;     /* values are unsigned */
 633    blend_type.norm = TRUE;      /* values are in [0,1] or [-1,1] */
 634    blend_type.width = 8;        /* 8-bit ubyte values */
 635    blend_type.length = 16;      /* 16 elements per vector */
 636
 637    /*
 638     * Generate the function prototype. Any change here must be reflected in
 639     * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
 640     */
 641
 642    fs_elem_type = lp_build_elem_type(fs_type);
 643    fs_int_vec_type = lp_build_int_vec_type(fs_type);
 644
 645    blend_vec_type = lp_build_vec_type(blend_type);
 646
 647    util_snprintf(func_name, sizeof(func_name), "fs%u_variant%u_%s",
 648                  shader->no, variant->no, do_tri_test ? "edge" : "whole");
 649
 650    arg_types[0] = screen->context_ptr_type;            /* context */
 651    arg_types[1] = LLVMInt32Type();                     /* x */
 652    arg_types[2] = LLVMInt32Type();                     /* y */
 653    arg_types[3] = LLVMFloatType();                     /* facing */
 654    arg_types[4] = LLVMPointerType(fs_elem_type, 0);    /* a0 */
 655    arg_types[5] = LLVMPointerType(fs_elem_type, 0);    /* dadx */
 656    arg_types[6] = LLVMPointerType(fs_elem_type, 0);    /* dady */
 657    arg_types[7] = LLVMPointerType(LLVMPointerType(blend_vec_type, 0), 0);  /* color */
 658    arg_types[8] = LLVMPointerType(fs_int_vec_type, 0); /* depth */
 659    arg_types[9] = LLVMInt32Type();                     /* c0 */
 660    arg_types[10] = LLVMInt32Type();                    /* c1 */
 661    arg_types[11] = LLVMInt32Type();                    /* c2 */
 662    /* Note: the step arrays are built as int32[16] but we interpret
 663     * them here as int32_vec4[4].
 664     */
 665    arg_types[12] = LLVMPointerType(int32_vec4_type, 0);/* step0 */
 666    arg_types[13] = LLVMPointerType(int32_vec4_type, 0);/* step1 */
 667    arg_types[14] = LLVMPointerType(int32_vec4_type, 0);/* step2 */
 668    arg_types[15] = LLVMPointerType(LLVMInt32Type(), 0);/* counter */
 669
 670    func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
 671
 672    function = LLVMAddFunction(screen->module, func_name, func_type);
 673    LLVMSetFunctionCallConv(function, LLVMCCallConv);
 674
 675    variant->function[do_tri_test] = function;
 676
 677
 678    /* XXX: need to propagate noalias down into color param now we are
 679     * passing a pointer-to-pointer?
 680     */
 681    for(i = 0; i < Elements(arg_types); ++i)
 682       if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
 683          LLVMAddAttribute(LLVMGetParam(function, i), LLVMNoAliasAttribute);
 684
 685    context_ptr  = LLVMGetParam(function, 0);
 686    x            = LLVMGetParam(function, 1);
 687    y            = LLVMGetParam(function, 2);
 688    facing       = LLVMGetParam(function, 3);
 689    a0_ptr       = LLVMGetParam(function, 4);
 690    dadx_ptr     = LLVMGetParam(function, 5);
 691    dady_ptr     = LLVMGetParam(function, 6);
 692    color_ptr_ptr = LLVMGetParam(function, 7);
 693    depth_ptr    = LLVMGetParam(function, 8);
 694    c0           = LLVMGetParam(function, 9);
 695    c1           = LLVMGetParam(function, 10);
 696    c2           = LLVMGetParam(function, 11);
 697    step0_ptr    = LLVMGetParam(function, 12);
 698    step1_ptr    = LLVMGetParam(function, 13);
 699    step2_ptr    = LLVMGetParam(function, 14);
 700
 701    lp_build_name(context_ptr, "context");
 702    lp_build_name(x, "x");
 703    lp_build_name(y, "y");
 704    lp_build_name(a0_ptr, "a0");
 705    lp_build_name(dadx_ptr, "dadx");
 706    lp_build_name(dady_ptr, "dady");
 707    lp_build_name(color_ptr_ptr, "color_ptr_ptr");
 708    lp_build_name(depth_ptr, "depth");
 709    lp_build_name(c0, "c0");
 710    lp_build_name(c1, "c1");
 711    lp_build_name(c2, "c2");
 712    lp_build_name(step0_ptr, "step0");
 713    lp_build_name(step1_ptr, "step1");
 714    lp_build_name(step2_ptr, "step2");
 715
 716    if (key->occlusion_count) {
 717       counter = LLVMGetParam(function, 15);
 718       lp_build_name(counter, "counter");
 719    }
 720
 721    /*
 722     * Function body
 723     */
 724
 725    block = LLVMAppendBasicBlock(function, "entry");
 726    builder = LLVMCreateBuilder();
 727    LLVMPositionBuilderAtEnd(builder, block);
 728
 729    /*
 730     * The shader input interpolation info is not explicitely baked in the
 731     * shader key, but everything it derives from (TGSI, and flatshade) is
 732     * already included in the shader key.
 733     */
 734    lp_build_interp_soa_init(&interp,
 735                             lp->num_inputs,
 736                             lp->inputs,
 737                             builder, fs_type,
 738                             a0_ptr, dadx_ptr, dady_ptr,
 739                             x, y);
 740
 741    /* code generated texture sampling */
 742    sampler = lp_llvm_sampler_soa_create(key->sampler, context_ptr);
 743
 744    /* loop over quads in the block */
 745    for(i = 0; i < num_fs; ++i) {
 746       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
 747       LLVMValueRef out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS];
 748       LLVMValueRef depth_ptr_i;
 749
 750       if(i != 0)
 751          lp_build_interp_soa_update(&interp, i);
 752
 753       depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &index, 1, "");
 754
 755       generate_fs(lp, shader, key,
 756                   builder,
 757                   fs_type,
 758                   context_ptr,
 759                   i,
 760                   &interp,
 761                   sampler,
 762                   &fs_mask[i], /* output */
 763                   out_color,
 764                   depth_ptr_i,
 765                   facing,
 766                   do_tri_test,
 767                   c0, c1, c2,
 768                   step0_ptr, step1_ptr, step2_ptr, counter);
 769
 770       for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++)
 771          for(chan = 0; chan < NUM_CHANNELS; ++chan)
 772             fs_out_color[cbuf][chan][i] = out_color[cbuf][chan];
 773    }
 774
 775    sampler->destroy(sampler);
 776
 777    /* Loop over color outputs / color buffers to do blending.
 778     */
 779    for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
 780       LLVMValueRef color_ptr;
 781       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), cbuf, 0);
 782       LLVMValueRef blend_in_color[NUM_CHANNELS];
 783       unsigned rt;
 784
 785       /*
 786        * Convert the fs's output color and mask to fit to the blending type.
 787        */
 788       for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 789          lp_build_conv(builder, fs_type, blend_type,
 790                        fs_out_color[cbuf][chan], num_fs,
 791                        &blend_in_color[chan], 1);
 792          lp_build_name(blend_in_color[chan], "color%d.%c", cbuf, "rgba"[chan]);
 793       }
 794
 795       lp_build_conv_mask(builder, fs_type, blend_type,
 796                          fs_mask, num_fs,
 797                          &blend_mask, 1);
 798
 799       color_ptr = LLVMBuildLoad(builder,
 800                                 LLVMBuildGEP(builder, color_ptr_ptr, &index, 1, ""),
 801                                 "");
 802       lp_build_name(color_ptr, "color_ptr%d", cbuf);
 803
 804       /* which blend/colormask state to use */
 805       rt = key->blend.independent_blend_enable ? cbuf : 0;
 806
 807       /*
 808        * Blending.
 809        */
 810       generate_blend(&key->blend,
 811                      rt,
 812                      builder,
 813                      blend_type,
 814                      context_ptr,
 815                      blend_mask,
 816                      blend_in_color,
 817                      color_ptr);
 818    }
 819
 820    LLVMBuildRetVoid(builder);
 821
 822    LLVMDisposeBuilder(builder);
 823
 824
 825    /* Verify the LLVM IR.  If invalid, dump and abort */
 826 #ifdef DEBUG
 827    if(LLVMVerifyFunction(function, LLVMPrintMessageAction)) {
 828       if (1)
 829          lp_debug_dump_value(function);
 830       abort();
 831    }
 832 #endif
 833
 834    /* Apply optimizations to LLVM IR */
 835    if (1)
 836       LLVMRunFunctionPassManager(screen->pass, function);
 837
 838    if (gallivm_debug & GALLIVM_DEBUG_IR) {
 839       /* Print the LLVM IR to stderr */
 840       lp_debug_dump_value(function);
 841       debug_printf("\n");
 842    }
 843
 844    /*
 845     * Translate the LLVM IR into machine code.
 846     */
 847    {
 848       void *f = LLVMGetPointerToGlobal(screen->engine, function);
 849
 850       variant->jit_function[do_tri_test] = (lp_jit_frag_func)pointer_to_func(f);
 851
 852       if (gallivm_debug & GALLIVM_DEBUG_ASM) {
 853          lp_disassemble(f);
 854       }
 855    }
 856 }
 857
 858
 859 static void
 860 dump_fs_variant_key(const struct lp_fragment_shader_variant_key *key)
 861 {
 862    unsigned i;
 863
 864    debug_printf("fs variant %p:\n", (void *) key);
 865
 866    if (key->depth.enabled) {
 867       debug_printf("depth.format = %s\n", util_format_name(key->zsbuf_format));
 868       debug_printf("depth.func = %s\n", util_dump_func(key->depth.func, TRUE));
 869       debug_printf("depth.writemask = %u\n", key->depth.writemask);
 870    }
 871
 872    for (i = 0; i < 2; ++i) {
 873       if (key->stencil[i].enabled) {
 874          debug_printf("stencil[%u].func = %s\n", i, util_dump_func(key->stencil[i].func, TRUE));
 875          debug_printf("stencil[%u].fail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].fail_op, TRUE));
 876          debug_printf("stencil[%u].zpass_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zpass_op, TRUE));
 877          debug_printf("stencil[%u].zfail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zfail_op, TRUE));
 878          debug_printf("stencil[%u].valuemask = 0x%x\n", i, key->stencil[i].valuemask);
 879          debug_printf("stencil[%u].writemask = 0x%x\n", i, key->stencil[i].writemask);
 880       }
 881    }
 882
 883    if (key->alpha.enabled) {
 884       debug_printf("alpha.func = %s\n", util_dump_func(key->alpha.func, TRUE));
 885       debug_printf("alpha.ref_value = %f\n", key->alpha.ref_value);
 886    }
 887
 888    if (key->blend.logicop_enable) {
 889       debug_printf("blend.logicop_func = %s\n", util_dump_logicop(key->blend.logicop_func, TRUE));
 890    }
 891    else if (key->blend.rt[0].blend_enable) {
 892       debug_printf("blend.rgb_func = %s\n",   util_dump_blend_func  (key->blend.rt[0].rgb_func, TRUE));
 893       debug_printf("blend.rgb_src_factor = %s\n",   util_dump_blend_factor(key->blend.rt[0].rgb_src_factor, TRUE));
 894       debug_printf("blend.rgb_dst_factor = %s\n",   util_dump_blend_factor(key->blend.rt[0].rgb_dst_factor, TRUE));
 895       debug_printf("blend.alpha_func = %s\n",       util_dump_blend_func  (key->blend.rt[0].alpha_func, TRUE));
 896       debug_printf("blend.alpha_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_src_factor, TRUE));
 897       debug_printf("blend.alpha_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_dst_factor, TRUE));
 898    }
 899    debug_printf("blend.colormask = 0x%x\n", key->blend.rt[0].colormask);
 900    for (i = 0; i < PIPE_MAX_SAMPLERS; ++i) {
 901       if (key->sampler[i].format) {
 902          debug_printf("sampler[%u] = \n", i);
 903          debug_printf("  .format = %s\n",
 904                       util_format_name(key->sampler[i].format));
 905          debug_printf("  .target = %s\n",
 906                       util_dump_tex_target(key->sampler[i].target, TRUE));
 907          debug_printf("  .pot = %u %u %u\n",
 908                       key->sampler[i].pot_width,
 909                       key->sampler[i].pot_height,
 910                       key->sampler[i].pot_depth);
 911          debug_printf("  .wrap = %s %s %s\n",
 912                       util_dump_tex_wrap(key->sampler[i].wrap_s, TRUE),
 913                       util_dump_tex_wrap(key->sampler[i].wrap_t, TRUE),
 914                       util_dump_tex_wrap(key->sampler[i].wrap_r, TRUE));
 915          debug_printf("  .min_img_filter = %s\n",
 916                       util_dump_tex_filter(key->sampler[i].min_img_filter, TRUE));
 917          debug_printf("  .min_mip_filter = %s\n",
 918                       util_dump_tex_mipfilter(key->sampler[i].min_mip_filter, TRUE));
 919          debug_printf("  .mag_img_filter = %s\n",
 920                       util_dump_tex_filter(key->sampler[i].mag_img_filter, TRUE));
 921          if (key->sampler[i].compare_mode != PIPE_TEX_COMPARE_NONE)
 922             debug_printf("  .compare_func = %s\n", util_dump_func(key->sampler[i].compare_func, TRUE));
 923          debug_printf("  .normalized_coords = %u\n", key->sampler[i].normalized_coords);
 924       }
 925    }
 926 }
 927
 928
 929
 930 static struct lp_fragment_shader_variant *
 931 generate_variant(struct llvmpipe_context *lp,
 932                  struct lp_fragment_shader *shader,
 933                  const struct lp_fragment_shader_variant_key *key)
 934 {
 935    struct lp_fragment_shader_variant *variant;
 936
 937    variant = CALLOC_STRUCT(lp_fragment_shader_variant);
 938    if(!variant)
 939       return NULL;
 940
 941    variant->lpfs = shader;
 942    variant->list_item_global.base = variant;
 943    variant->list_item_local.base = variant;
 944    variant->no = shader->variants_created++;
 945
 946    memcpy(&variant->key, key, sizeof *key);
 947
 948    if (gallivm_debug & GALLIVM_DEBUG_IR) {
 949       debug_printf("llvmpipe: Creating fragment shader #%u variant #%u:\n",
 950                    shader->no, variant->no);
 951       tgsi_dump(shader->base.tokens, 0);
 952       dump_fs_variant_key(key);
 953    }
 954
 955    generate_fragment(lp, shader, variant, RAST_WHOLE);
 956    generate_fragment(lp, shader, variant, RAST_EDGE_TEST);
 957
 958    /* TODO: most of these can be relaxed, in particular the colormask */
 959    variant->opaque =
 960          !key->blend.logicop_enable &&
 961          !key->blend.rt[0].blend_enable &&
 962          key->blend.rt[0].colormask == 0xf &&
 963          !key->stencil[0].enabled &&
 964          !key->alpha.enabled &&
 965          !key->depth.enabled &&
 966          !key->scissor &&
 967          !shader->info.uses_kill
 968          ? TRUE : FALSE;
 969
 970    return variant;
 971 }
 972
 973
 974 static void *
 975 llvmpipe_create_fs_state(struct pipe_context *pipe,
 976                          const struct pipe_shader_state *templ)
 977 {
 978    struct lp_fragment_shader *shader;
 979
 980    shader = CALLOC_STRUCT(lp_fragment_shader);
 981    if (!shader)
 982       return NULL;
 983
 984    shader->no = fs_no++;
 985    make_empty_list(&shader->variants);
 986
 987    /* get/save the summary info for this shader */
 988    tgsi_scan_shader(templ->tokens, &shader->info);
 989
 990    /* we need to keep a local copy of the tokens */
 991    shader->base.tokens = tgsi_dup_tokens(templ->tokens);
 992
 993    if (LP_DEBUG & DEBUG_TGSI) {
 994       unsigned attrib;
 995       debug_printf("llvmpipe: Create fragment shader #%u %p:\n", shader->no, (void *) shader);
 996       tgsi_dump(templ->tokens, 0);
 997       debug_printf("usage masks:\n");
 998       for (attrib = 0; attrib < shader->info.num_inputs; ++attrib) {
 999          unsigned usage_mask = shader->info.input_usage_mask[attrib];
1000          debug_printf("  IN[%u].%s%s%s%s\n",
1001                       attrib,
1002                       usage_mask & TGSI_WRITEMASK_X ? "x" : "",
1003                       usage_mask & TGSI_WRITEMASK_Y ? "y" : "",
1004                       usage_mask & TGSI_WRITEMASK_Z ? "z" : "",
1005                       usage_mask & TGSI_WRITEMASK_W ? "w" : "");
1006       }
1007       debug_printf("\n");
1008    }
1009
1010    return shader;
1011 }
1012
1013
1014 static void
1015 llvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs)
1016 {
1017    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
1018
1019    if (llvmpipe->fs == fs)
1020       return;
1021
1022    draw_flush(llvmpipe->draw);
1023
1024    llvmpipe->fs = fs;
1025
1026    llvmpipe->dirty |= LP_NEW_FS;
1027 }
1028
1029 static void
1030 remove_shader_variant(struct llvmpipe_context *lp,
1031                       struct lp_fragment_shader_variant *variant)
1032 {
1033    struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
1034    unsigned i;
1035
1036    if (gallivm_debug & GALLIVM_DEBUG_IR) {
1037       debug_printf("llvmpipe: del fs #%u var #%u v created #%u v cached #%u v total cached #%u\n",
1038                     variant->lpfs->no, variant->no, variant->lpfs->variants_created,
1039                     variant->lpfs->variants_cached, lp->nr_fs_variants);
1040    }
1041    for (i = 0; i < Elements(variant->function); i++) {
1042       if (variant->function[i]) {
1043          if (variant->jit_function[i])
1044             LLVMFreeMachineCodeForFunction(screen->engine,
1045                                            variant->function[i]);
1046          LLVMDeleteFunction(variant->function[i]);
1047       }
1048    }
1049    remove_from_list(&variant->list_item_local);
1050    variant->lpfs->variants_cached--;
1051    remove_from_list(&variant->list_item_global);
1052    lp->nr_fs_variants--;
1053    FREE(variant);
1054 }
1055
1056 static void
1057 llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs)
1058 {
1059    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
1060    struct pipe_fence_handle *fence = NULL;
1061    struct lp_fragment_shader *shader = fs;
1062    struct lp_fs_variant_list_item *li;
1063
1064    assert(fs != llvmpipe->fs);
1065    (void) llvmpipe;
1066
1067    /*
1068     * XXX: we need to flush the context until we have some sort of reference
1069     * counting in fragment shaders as they may still be binned
1070     * Flushing alone might not sufficient we need to wait on it too.
1071     */
1072
1073    llvmpipe_flush(pipe, 0, &fence);
1074
1075    if (fence) {
1076       pipe->screen->fence_finish(pipe->screen, fence, 0);
1077       pipe->screen->fence_reference(pipe->screen, &fence, NULL);
1078    }
1079
1080    li = first_elem(&shader->variants);
1081    while(!at_end(&shader->variants, li)) {
1082       remove_shader_variant(llvmpipe, li->base);
1083       li = next_elem(li);
1084    }
1085
1086    assert(shader->variants_cached == 0);
1087    FREE((void *) shader->base.tokens);
1088    FREE(shader);
1089 }
1090
1091
1092
1093 static void
1094 llvmpipe_set_constant_buffer(struct pipe_context *pipe,
1095                              uint shader, uint index,
1096                              struct pipe_resource *constants)
1097 {
1098    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
1099    unsigned size = constants ? constants->width0 : 0;
1100    const void *data = constants ? llvmpipe_resource_data(constants) : NULL;
1101
1102    assert(shader < PIPE_SHADER_TYPES);
1103    assert(index < PIPE_MAX_CONSTANT_BUFFERS);
1104
1105    if(llvmpipe->constants[shader][index] == constants)
1106       return;
1107
1108    draw_flush(llvmpipe->draw);
1109
1110    /* note: reference counting */
1111    pipe_resource_reference(&llvmpipe->constants[shader][index], constants);
1112
1113    if(shader == PIPE_SHADER_VERTEX) {
1114       draw_set_mapped_constant_buffer(llvmpipe->draw, PIPE_SHADER_VERTEX, index,
1115                                       data, size);
1116    }
1117
1118    llvmpipe->dirty |= LP_NEW_CONSTANTS;
1119 }
1120
1121
1122 /**
1123  * Return the blend factor equivalent to a destination alpha of one.
1124  */
1125 static INLINE unsigned
1126 force_dst_alpha_one(unsigned factor, boolean alpha)
1127 {
1128    switch(factor) {
1129    case PIPE_BLENDFACTOR_DST_ALPHA:
1130       return PIPE_BLENDFACTOR_ONE;
1131    case PIPE_BLENDFACTOR_INV_DST_ALPHA:
1132       return PIPE_BLENDFACTOR_ZERO;
1133    case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
1134       return PIPE_BLENDFACTOR_ZERO;
1135    }
1136
1137    if (alpha) {
1138       switch(factor) {
1139       case PIPE_BLENDFACTOR_DST_COLOR:
1140          return PIPE_BLENDFACTOR_ONE;
1141       case PIPE_BLENDFACTOR_INV_DST_COLOR:
1142          return PIPE_BLENDFACTOR_ZERO;
1143       }
1144    }
1145
1146    return factor;
1147 }
1148
1149
1150 /**
1151  * We need to generate several variants of the fragment pipeline to match
1152  * all the combinations of the contributing state atoms.
1153  *
1154  * TODO: there is actually no reason to tie this to context state -- the
1155  * generated code could be cached globally in the screen.
1156  */
1157 static void
1158 make_variant_key(struct llvmpipe_context *lp,
1159                  struct lp_fragment_shader *shader,
1160                  struct lp_fragment_shader_variant_key *key)
1161 {
1162    unsigned i;
1163
1164    memset(key, 0, sizeof *key);
1165
1166    if (lp->framebuffer.zsbuf) {
1167       if (lp->depth_stencil->depth.enabled) {
1168          key->zsbuf_format = lp->framebuffer.zsbuf->format;
1169          memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth);
1170       }
1171       if (lp->depth_stencil->stencil[0].enabled) {
1172          key->zsbuf_format = lp->framebuffer.zsbuf->format;
1173          memcpy(&key->stencil, &lp->depth_stencil->stencil, sizeof key->stencil);
1174       }
1175    }
1176
1177    key->alpha.enabled = lp->depth_stencil->alpha.enabled;
1178    if(key->alpha.enabled)
1179       key->alpha.func = lp->depth_stencil->alpha.func;
1180    /* alpha.ref_value is passed in jit_context */
1181
1182    key->flatshade = lp->rasterizer->flatshade;
1183    key->scissor = lp->rasterizer->scissor;
1184    if (lp->active_query_count) {
1185       key->occlusion_count = TRUE;
1186    }
1187
1188    if (lp->framebuffer.nr_cbufs) {
1189       memcpy(&key->blend, lp->blend, sizeof key->blend);
1190    }
1191
1192    key->nr_cbufs = lp->framebuffer.nr_cbufs;
1193    for (i = 0; i < lp->framebuffer.nr_cbufs; i++) {
1194       struct pipe_rt_blend_state *blend_rt = &key->blend.rt[i];
1195       const struct util_format_description *format_desc;
1196       unsigned chan;
1197
1198       format_desc = util_format_description(lp->framebuffer.cbufs[i]->format);
1199       assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
1200              format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB);
1201
1202       blend_rt->colormask = lp->blend->rt[i].colormask;
1203
1204       /* mask out color channels not present in the color buffer.
1205        * Should be simple to incorporate per-cbuf writemasks:
1206        */
1207       for(chan = 0; chan < 4; ++chan) {
1208          enum util_format_swizzle swizzle = format_desc->swizzle[chan];
1209
1210          if(swizzle > UTIL_FORMAT_SWIZZLE_W)
1211             blend_rt->colormask &= ~(1 << chan);
1212       }
1213
1214       /*
1215        * Our swizzled render tiles always have an alpha channel, but the linear
1216        * render target format often does not, so force here the dst alpha to be
1217        * one.
1218        *
1219        * This is not a mere optimization. Wrong results will be produced if the
1220        * dst alpha is used, the dst format does not have alpha, and the previous
1221        * rendering was not flushed from the swizzled to linear buffer. For
1222        * example, NonPowTwo DCT.
1223        *
1224        * TODO: This should be generalized to all channels for better
1225        * performance, but only alpha causes correctness issues.
1226        */
1227       if (format_desc->swizzle[3] > UTIL_FORMAT_SWIZZLE_W) {
1228          blend_rt->rgb_src_factor = force_dst_alpha_one(blend_rt->rgb_src_factor, FALSE);
1229          blend_rt->rgb_dst_factor = force_dst_alpha_one(blend_rt->rgb_dst_factor, FALSE);
1230          blend_rt->alpha_src_factor = force_dst_alpha_one(blend_rt->alpha_src_factor, TRUE);
1231          blend_rt->alpha_dst_factor = force_dst_alpha_one(blend_rt->alpha_dst_factor, TRUE);
1232       }
1233    }
1234
1235    for(i = 0; i < PIPE_MAX_SAMPLERS; ++i)
1236       if(shader->info.file_mask[TGSI_FILE_SAMPLER] & (1 << i))
1237          lp_sampler_static_state(&key->sampler[i], lp->fragment_sampler_views[i], lp->sampler[i]);
1238 }
1239
1240 /**
1241  * Update fragment state.  This is called just prior to drawing
1242  * something when some fragment-related state has changed.
1243  */
1244 void
1245 llvmpipe_update_fs(struct llvmpipe_context *lp)
1246 {
1247    struct lp_fragment_shader *shader = lp->fs;
1248    struct lp_fragment_shader_variant_key key;
1249    struct lp_fragment_shader_variant *variant = NULL;
1250    struct lp_fs_variant_list_item *li;
1251
1252    make_variant_key(lp, shader, &key);
1253
1254    li = first_elem(&shader->variants);
1255    while(!at_end(&shader->variants, li)) {
1256       if(memcmp(&li->base->key, &key, sizeof key) == 0) {
1257          variant = li->base;
1258          break;
1259       }
1260       li = next_elem(li);
1261    }
1262
1263    if (variant) {
1264       move_to_head(&lp->fs_variants_list, &variant->list_item_global);
1265    }
1266    else {
1267       int64_t t0, t1;
1268       int64_t dt;
1269       unsigned i;
1270       if (lp->nr_fs_variants >= LP_MAX_SHADER_VARIANTS) {
1271          struct pipe_context *pipe = &lp->pipe;
1272          struct pipe_fence_handle *fence = NULL;
1273
1274          /*
1275           * XXX: we need to flush the context until we have some sort of reference
1276           * counting in fragment shaders as they may still be binned
1277           * Flushing alone might not be sufficient we need to wait on it too.
1278           */
1279          llvmpipe_flush(pipe, 0, &fence);
1280
1281          if (fence) {
1282             pipe->screen->fence_finish(pipe->screen, fence, 0);
1283             pipe->screen->fence_reference(pipe->screen, &fence, NULL);
1284          }
1285          for (i = 0; i < LP_MAX_SHADER_VARIANTS / 4; i++) {
1286             struct lp_fs_variant_list_item *item = last_elem(&lp->fs_variants_list);
1287             remove_shader_variant(lp, item->base);
1288          }
1289       }
1290       t0 = os_time_get();
1291
1292       variant = generate_variant(lp, shader, &key);
1293
1294       t1 = os_time_get();
1295       dt = t1 - t0;
1296       LP_COUNT_ADD(llvm_compile_time, dt);
1297       LP_COUNT_ADD(nr_llvm_compiles, 2);  /* emit vs. omit in/out test */
1298
1299       if (variant) {
1300          insert_at_head(&shader->variants, &variant->list_item_local);
1301          insert_at_head(&lp->fs_variants_list, &variant->list_item_global);
1302          lp->nr_fs_variants++;
1303          shader->variants_cached++;
1304       }
1305    }
1306
1307    lp_setup_set_fs_variant(lp->setup, variant);
1308 }
1309
1310
1311
1312 void
1313 llvmpipe_init_fs_funcs(struct llvmpipe_context *llvmpipe)
1314 {
1315    llvmpipe->pipe.create_fs_state = llvmpipe_create_fs_state;
1316    llvmpipe->pipe.bind_fs_state   = llvmpipe_bind_fs_state;
1317    llvmpipe->pipe.delete_fs_state = llvmpipe_delete_fs_state;
1318
1319    llvmpipe->pipe.set_constant_buffer = llvmpipe_set_constant_buffer;
1320 }