Subversion Repositories Kolibri OS

Rev

Blame | Last modification | View Log | RSS feed

  1. /**************************************************************************
  2.  *
  3.  * Copyright 2009 VMware, Inc.
  4.  * Copyright 2007 VMware, Inc.
  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 VMWARE 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/simple_list.h"
  69. #include "util/u_dual_blend.h"
  70. #include "os/os_time.h"
  71. #include "pipe/p_shader_tokens.h"
  72. #include "draw/draw_context.h"
  73. #include "tgsi/tgsi_dump.h"
  74. #include "tgsi/tgsi_scan.h"
  75. #include "tgsi/tgsi_parse.h"
  76. #include "gallivm/lp_bld_type.h"
  77. #include "gallivm/lp_bld_const.h"
  78. #include "gallivm/lp_bld_conv.h"
  79. #include "gallivm/lp_bld_init.h"
  80. #include "gallivm/lp_bld_intr.h"
  81. #include "gallivm/lp_bld_logic.h"
  82. #include "gallivm/lp_bld_tgsi.h"
  83. #include "gallivm/lp_bld_swizzle.h"
  84. #include "gallivm/lp_bld_flow.h"
  85. #include "gallivm/lp_bld_debug.h"
  86. #include "gallivm/lp_bld_arit.h"
  87. #include "gallivm/lp_bld_pack.h"
  88. #include "gallivm/lp_bld_format.h"
  89. #include "gallivm/lp_bld_quad.h"
  90.  
  91. #include "lp_bld_alpha.h"
  92. #include "lp_bld_blend.h"
  93. #include "lp_bld_depth.h"
  94. #include "lp_bld_interp.h"
  95. #include "lp_context.h"
  96. #include "lp_debug.h"
  97. #include "lp_perf.h"
  98. #include "lp_setup.h"
  99. #include "lp_state.h"
  100. #include "lp_tex_sample.h"
  101. #include "lp_flush.h"
  102. #include "lp_state_fs.h"
  103. #include "lp_rast.h"
  104.  
  105.  
  106. /** Fragment shader number (for debugging) */
  107. static unsigned fs_no = 0;
  108.  
  109.  
  110. /**
  111.  * Expand the relevant bits of mask_input to a n*4-dword mask for the
  112.  * n*four pixels in n 2x2 quads.  This will set the n*four elements of the
  113.  * quad mask vector to 0 or ~0.
  114.  * Grouping is 01, 23 for 2 quad mode hence only 0 and 2 are valid
  115.  * quad arguments with fs length 8.
  116.  *
  117.  * \param first_quad  which quad(s) of the quad group to test, in [0,3]
  118.  * \param mask_input  bitwise mask for the whole 4x4 stamp
  119.  */
  120. static LLVMValueRef
  121. generate_quad_mask(struct gallivm_state *gallivm,
  122.                    struct lp_type fs_type,
  123.                    unsigned first_quad,
  124.                    LLVMValueRef mask_input) /* int32 */
  125. {
  126.    LLVMBuilderRef builder = gallivm->builder;
  127.    struct lp_type mask_type;
  128.    LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
  129.    LLVMValueRef bits[16];
  130.    LLVMValueRef mask;
  131.    int shift, i;
  132.  
  133.    /*
  134.     * XXX: We'll need a different path for 16 x u8
  135.     */
  136.    assert(fs_type.width == 32);
  137.    assert(fs_type.length <= Elements(bits));
  138.    mask_type = lp_int_type(fs_type);
  139.  
  140.    /*
  141.     * mask_input >>= (quad * 4)
  142.     */
  143.    switch (first_quad) {
  144.    case 0:
  145.       shift = 0;
  146.       break;
  147.    case 1:
  148.       assert(fs_type.length == 4);
  149.       shift = 2;
  150.       break;
  151.    case 2:
  152.       shift = 8;
  153.       break;
  154.    case 3:
  155.       assert(fs_type.length == 4);
  156.       shift = 10;
  157.       break;
  158.    default:
  159.       assert(0);
  160.       shift = 0;
  161.    }
  162.  
  163.    mask_input = LLVMBuildLShr(builder,
  164.                               mask_input,
  165.                               LLVMConstInt(i32t, shift, 0),
  166.                               "");
  167.  
  168.    /*
  169.     * mask = { mask_input & (1 << i), for i in [0,3] }
  170.     */
  171.    mask = lp_build_broadcast(gallivm,
  172.                              lp_build_vec_type(gallivm, mask_type),
  173.                              mask_input);
  174.  
  175.    for (i = 0; i < fs_type.length / 4; i++) {
  176.       unsigned j = 2 * (i % 2) + (i / 2) * 8;
  177.       bits[4*i + 0] = LLVMConstInt(i32t, 1ULL << (j + 0), 0);
  178.       bits[4*i + 1] = LLVMConstInt(i32t, 1ULL << (j + 1), 0);
  179.       bits[4*i + 2] = LLVMConstInt(i32t, 1ULL << (j + 4), 0);
  180.       bits[4*i + 3] = LLVMConstInt(i32t, 1ULL << (j + 5), 0);
  181.    }
  182.    mask = LLVMBuildAnd(builder, mask, LLVMConstVector(bits, fs_type.length), "");
  183.  
  184.    /*
  185.     * mask = mask != 0 ? ~0 : 0
  186.     */
  187.    mask = lp_build_compare(gallivm,
  188.                            mask_type, PIPE_FUNC_NOTEQUAL,
  189.                            mask,
  190.                            lp_build_const_int_vec(gallivm, mask_type, 0));
  191.  
  192.    return mask;
  193. }
  194.  
  195.  
  196. #define EARLY_DEPTH_TEST  0x1
  197. #define LATE_DEPTH_TEST   0x2
  198. #define EARLY_DEPTH_WRITE 0x4
  199. #define LATE_DEPTH_WRITE  0x8
  200.  
  201. static int
  202. find_output_by_semantic( const struct tgsi_shader_info *info,
  203.                          unsigned semantic,
  204.                          unsigned index )
  205. {
  206.    int i;
  207.  
  208.    for (i = 0; i < info->num_outputs; i++)
  209.       if (info->output_semantic_name[i] == semantic &&
  210.           info->output_semantic_index[i] == index)
  211.          return i;
  212.  
  213.    return -1;
  214. }
  215.  
  216.  
  217. /**
  218.  * Fetch the specified lp_jit_viewport structure for a given viewport_index.
  219.  */
  220. static LLVMValueRef
  221. lp_llvm_viewport(LLVMValueRef context_ptr,
  222.                  struct gallivm_state *gallivm,
  223.                  LLVMValueRef viewport_index)
  224. {
  225.    LLVMBuilderRef builder = gallivm->builder;
  226.    LLVMValueRef ptr;
  227.    LLVMValueRef res;
  228.    struct lp_type viewport_type =
  229.       lp_type_float_vec(32, 32 * LP_JIT_VIEWPORT_NUM_FIELDS);
  230.  
  231.    ptr = lp_jit_context_viewports(gallivm, context_ptr);
  232.    ptr = LLVMBuildPointerCast(builder, ptr,
  233.             LLVMPointerType(lp_build_vec_type(gallivm, viewport_type), 0), "");
  234.  
  235.    res = lp_build_pointer_get(builder, ptr, viewport_index);
  236.  
  237.    return res;
  238. }
  239.  
  240.  
  241. /**
  242.  * Generate the fragment shader, depth/stencil test, and alpha tests.
  243.  */
  244. static void
  245. generate_fs_loop(struct gallivm_state *gallivm,
  246.                  struct lp_fragment_shader *shader,
  247.                  const struct lp_fragment_shader_variant_key *key,
  248.                  LLVMBuilderRef builder,
  249.                  struct lp_type type,
  250.                  LLVMValueRef context_ptr,
  251.                  LLVMValueRef num_loop,
  252.                  struct lp_build_interp_soa_context *interp,
  253.                  struct lp_build_sampler_soa *sampler,
  254.                  LLVMValueRef mask_store,
  255.                  LLVMValueRef (*out_color)[4],
  256.                  LLVMValueRef depth_ptr,
  257.                  LLVMValueRef depth_stride,
  258.                  LLVMValueRef facing,
  259.                  LLVMValueRef thread_data_ptr)
  260. {
  261.    const struct util_format_description *zs_format_desc = NULL;
  262.    const struct tgsi_token *tokens = shader->base.tokens;
  263.    LLVMTypeRef vec_type;
  264.    LLVMValueRef mask_ptr, mask_val;
  265.    LLVMValueRef consts_ptr, num_consts_ptr;
  266.    LLVMValueRef z;
  267.    LLVMValueRef z_value, s_value;
  268.    LLVMValueRef z_fb, s_fb;
  269.    LLVMValueRef stencil_refs[2];
  270.    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
  271.    struct lp_build_for_loop_state loop_state;
  272.    struct lp_build_mask_context mask;
  273.    /*
  274.     * TODO: figure out if simple_shader optimization is really worthwile to
  275.     * keep. Disabled because it may hide some real bugs in the (depth/stencil)
  276.     * code since tests tend to take another codepath than real shaders.
  277.     */
  278.    boolean simple_shader = (shader->info.base.file_count[TGSI_FILE_SAMPLER] == 0 &&
  279.                             shader->info.base.num_inputs < 3 &&
  280.                             shader->info.base.num_instructions < 8) && 0;
  281.    const boolean dual_source_blend = key->blend.rt[0].blend_enable &&
  282.                                      util_blend_state_is_dual(&key->blend, 0);
  283.    unsigned attrib;
  284.    unsigned chan;
  285.    unsigned cbuf;
  286.    unsigned depth_mode;
  287.  
  288.    struct lp_bld_tgsi_system_values system_values;
  289.  
  290.    memset(&system_values, 0, sizeof(system_values));
  291.  
  292.    if (key->depth.enabled ||
  293.        key->stencil[0].enabled) {
  294.  
  295.       zs_format_desc = util_format_description(key->zsbuf_format);
  296.       assert(zs_format_desc);
  297.  
  298.       if (!shader->info.base.writes_z) {
  299.          if (key->alpha.enabled ||
  300.              key->blend.alpha_to_coverage ||
  301.              shader->info.base.uses_kill) {
  302.             /* With alpha test and kill, can do the depth test early
  303.              * and hopefully eliminate some quads.  But need to do a
  304.              * special deferred depth write once the final mask value
  305.              * is known. This only works though if there's either no
  306.              * stencil test or the stencil value isn't written.
  307.              */
  308.             if (key->stencil[0].enabled && (key->stencil[0].writemask ||
  309.                                             (key->stencil[1].enabled &&
  310.                                              key->stencil[1].writemask)))
  311.                depth_mode = LATE_DEPTH_TEST | LATE_DEPTH_WRITE;
  312.             else
  313.                depth_mode = EARLY_DEPTH_TEST | LATE_DEPTH_WRITE;
  314.          }
  315.          else
  316.             depth_mode = EARLY_DEPTH_TEST | EARLY_DEPTH_WRITE;
  317.       }
  318.       else {
  319.          depth_mode = LATE_DEPTH_TEST | LATE_DEPTH_WRITE;
  320.       }
  321.  
  322.       if (!(key->depth.enabled && key->depth.writemask) &&
  323.           !(key->stencil[0].enabled && (key->stencil[0].writemask ||
  324.                                         (key->stencil[1].enabled &&
  325.                                          key->stencil[1].writemask))))
  326.          depth_mode &= ~(LATE_DEPTH_WRITE | EARLY_DEPTH_WRITE);
  327.    }
  328.    else {
  329.       depth_mode = 0;
  330.    }
  331.  
  332.  
  333.    stencil_refs[0] = lp_jit_context_stencil_ref_front_value(gallivm, context_ptr);
  334.    stencil_refs[1] = lp_jit_context_stencil_ref_back_value(gallivm, context_ptr);
  335.  
  336.    vec_type = lp_build_vec_type(gallivm, type);
  337.  
  338.    consts_ptr = lp_jit_context_constants(gallivm, context_ptr);
  339.    num_consts_ptr = lp_jit_context_num_constants(gallivm, context_ptr);
  340.  
  341.    lp_build_for_loop_begin(&loop_state, gallivm,
  342.                            lp_build_const_int32(gallivm, 0),
  343.                            LLVMIntULT,
  344.                            num_loop,
  345.                            lp_build_const_int32(gallivm, 1));
  346.  
  347.    mask_ptr = LLVMBuildGEP(builder, mask_store,
  348.                            &loop_state.counter, 1, "mask_ptr");
  349.    mask_val = LLVMBuildLoad(builder, mask_ptr, "");
  350.  
  351.    memset(outputs, 0, sizeof outputs);
  352.  
  353.    for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
  354.       for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
  355.          out_color[cbuf][chan] = lp_build_array_alloca(gallivm,
  356.                                                        lp_build_vec_type(gallivm,
  357.                                                                          type),
  358.                                                        num_loop, "color");
  359.       }
  360.    }
  361.    if (dual_source_blend) {
  362.       assert(key->nr_cbufs <= 1);
  363.       for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
  364.          out_color[1][chan] = lp_build_array_alloca(gallivm,
  365.                                                     lp_build_vec_type(gallivm,
  366.                                                                       type),
  367.                                                     num_loop, "color1");
  368.       }
  369.    }
  370.  
  371.  
  372.    /* 'mask' will control execution based on quad's pixel alive/killed state */
  373.    lp_build_mask_begin(&mask, gallivm, type, mask_val);
  374.  
  375.    if (!(depth_mode & EARLY_DEPTH_TEST) && !simple_shader)
  376.       lp_build_mask_check(&mask);
  377.  
  378.    lp_build_interp_soa_update_pos_dyn(interp, gallivm, loop_state.counter);
  379.    z = interp->pos[2];
  380.  
  381.    if (depth_mode & EARLY_DEPTH_TEST) {
  382.       lp_build_depth_stencil_load_swizzled(gallivm, type,
  383.                                            zs_format_desc, key->resource_1d,
  384.                                            depth_ptr, depth_stride,
  385.                                            &z_fb, &s_fb, loop_state.counter);
  386.       lp_build_depth_stencil_test(gallivm,
  387.                                   &key->depth,
  388.                                   key->stencil,
  389.                                   type,
  390.                                   zs_format_desc,
  391.                                   &mask,
  392.                                   stencil_refs,
  393.                                   z, z_fb, s_fb,
  394.                                   facing,
  395.                                   &z_value, &s_value,
  396.                                   !simple_shader);
  397.  
  398.       if (depth_mode & EARLY_DEPTH_WRITE) {
  399.          lp_build_depth_stencil_write_swizzled(gallivm, type,
  400.                                                zs_format_desc, key->resource_1d,
  401.                                                NULL, NULL, NULL, loop_state.counter,
  402.                                                depth_ptr, depth_stride,
  403.                                                z_value, s_value);
  404.       }
  405.       /*
  406.        * Note mask check if stencil is enabled must be after ds write not after
  407.        * stencil test otherwise new stencil values may not get written if all
  408.        * fragments got killed by depth/stencil test.
  409.        */
  410.       if (!simple_shader && key->stencil[0].enabled)
  411.          lp_build_mask_check(&mask);
  412.    }
  413.  
  414.    lp_build_interp_soa_update_inputs_dyn(interp, gallivm, loop_state.counter);
  415.  
  416.    /* Build the actual shader */
  417.    lp_build_tgsi_soa(gallivm, tokens, type, &mask,
  418.                      consts_ptr, num_consts_ptr, &system_values,
  419.                      interp->inputs,
  420.                      outputs, context_ptr,
  421.                      sampler, &shader->info.base, NULL);
  422.  
  423.    /* Alpha test */
  424.    if (key->alpha.enabled) {
  425.       int color0 = find_output_by_semantic(&shader->info.base,
  426.                                            TGSI_SEMANTIC_COLOR,
  427.                                            0);
  428.  
  429.       if (color0 != -1 && outputs[color0][3]) {
  430.          const struct util_format_description *cbuf_format_desc;
  431.          LLVMValueRef alpha = LLVMBuildLoad(builder, outputs[color0][3], "alpha");
  432.          LLVMValueRef alpha_ref_value;
  433.  
  434.          alpha_ref_value = lp_jit_context_alpha_ref_value(gallivm, context_ptr);
  435.          alpha_ref_value = lp_build_broadcast(gallivm, vec_type, alpha_ref_value);
  436.  
  437.          cbuf_format_desc = util_format_description(key->cbuf_format[0]);
  438.  
  439.          lp_build_alpha_test(gallivm, key->alpha.func, type, cbuf_format_desc,
  440.                              &mask, alpha, alpha_ref_value,
  441.                              (depth_mode & LATE_DEPTH_TEST) != 0);
  442.       }
  443.    }
  444.  
  445.    /* Emulate Alpha to Coverage with Alpha test */
  446.    if (key->blend.alpha_to_coverage) {
  447.       int color0 = find_output_by_semantic(&shader->info.base,
  448.                                            TGSI_SEMANTIC_COLOR,
  449.                                            0);
  450.  
  451.       if (color0 != -1 && outputs[color0][3]) {
  452.          LLVMValueRef alpha = LLVMBuildLoad(builder, outputs[color0][3], "alpha");
  453.  
  454.          lp_build_alpha_to_coverage(gallivm, type,
  455.                                     &mask, alpha,
  456.                                     (depth_mode & LATE_DEPTH_TEST) != 0);
  457.       }
  458.    }
  459.  
  460.    /* Late Z test */
  461.    if (depth_mode & LATE_DEPTH_TEST) {
  462.       int pos0 = find_output_by_semantic(&shader->info.base,
  463.                                          TGSI_SEMANTIC_POSITION,
  464.                                          0);
  465.  
  466.       if (pos0 != -1 && outputs[pos0][2]) {
  467.          z = LLVMBuildLoad(builder, outputs[pos0][2], "output.z");
  468.  
  469.          /*
  470.           * Clamp according to ARB_depth_clamp semantics.
  471.           */
  472.          if (key->depth_clamp) {
  473.             LLVMValueRef viewport, min_depth, max_depth;
  474.             LLVMValueRef viewport_index;
  475.             struct lp_build_context f32_bld;
  476.  
  477.             assert(type.floating);
  478.             lp_build_context_init(&f32_bld, gallivm, type);
  479.  
  480.             /*
  481.              * Assumes clamping of the viewport index will occur in setup/gs. Value
  482.              * is passed through the rasterization stage via lp_rast_shader_inputs.
  483.              *
  484.              * See: draw_clamp_viewport_idx and lp_clamp_viewport_idx for clamping
  485.              *      semantics.
  486.              */
  487.             viewport_index = lp_jit_thread_data_raster_state_viewport_index(gallivm,
  488.                                 thread_data_ptr);
  489.  
  490.             /*
  491.              * Load the min and max depth from the lp_jit_context.viewports
  492.              * array of lp_jit_viewport structures.
  493.              */
  494.             viewport = lp_llvm_viewport(context_ptr, gallivm, viewport_index);
  495.  
  496.             /* viewports[viewport_index].min_depth */
  497.             min_depth = LLVMBuildExtractElement(builder, viewport,
  498.                            lp_build_const_int32(gallivm, LP_JIT_VIEWPORT_MIN_DEPTH),
  499.                            "");
  500.             min_depth = lp_build_broadcast_scalar(&f32_bld, min_depth);
  501.  
  502.             /* viewports[viewport_index].max_depth */
  503.             max_depth = LLVMBuildExtractElement(builder, viewport,
  504.                            lp_build_const_int32(gallivm, LP_JIT_VIEWPORT_MAX_DEPTH),
  505.                            "");
  506.             max_depth = lp_build_broadcast_scalar(&f32_bld, max_depth);
  507.  
  508.             /*
  509.              * Clamp to the min and max depth values for the given viewport.
  510.              */
  511.             z = lp_build_clamp(&f32_bld, z, min_depth, max_depth);
  512.          }
  513.       }
  514.  
  515.       lp_build_depth_stencil_load_swizzled(gallivm, type,
  516.                                            zs_format_desc, key->resource_1d,
  517.                                            depth_ptr, depth_stride,
  518.                                            &z_fb, &s_fb, loop_state.counter);
  519.  
  520.       lp_build_depth_stencil_test(gallivm,
  521.                                   &key->depth,
  522.                                   key->stencil,
  523.                                   type,
  524.                                   zs_format_desc,
  525.                                   &mask,
  526.                                   stencil_refs,
  527.                                   z, z_fb, s_fb,
  528.                                   facing,
  529.                                   &z_value, &s_value,
  530.                                   !simple_shader);
  531.       /* Late Z write */
  532.       if (depth_mode & LATE_DEPTH_WRITE) {
  533.          lp_build_depth_stencil_write_swizzled(gallivm, type,
  534.                                                zs_format_desc, key->resource_1d,
  535.                                                NULL, NULL, NULL, loop_state.counter,
  536.                                                depth_ptr, depth_stride,
  537.                                                z_value, s_value);
  538.       }
  539.    }
  540.    else if ((depth_mode & EARLY_DEPTH_TEST) &&
  541.             (depth_mode & LATE_DEPTH_WRITE))
  542.    {
  543.       /* Need to apply a reduced mask to the depth write.  Reload the
  544.        * depth value, update from zs_value with the new mask value and
  545.        * write that out.
  546.        */
  547.       lp_build_depth_stencil_write_swizzled(gallivm, type,
  548.                                             zs_format_desc, key->resource_1d,
  549.                                             &mask, z_fb, s_fb, loop_state.counter,
  550.                                             depth_ptr, depth_stride,
  551.                                             z_value, s_value);
  552.    }
  553.  
  554.  
  555.    /* Color write  */
  556.    for (attrib = 0; attrib < shader->info.base.num_outputs; ++attrib)
  557.    {
  558.       unsigned cbuf = shader->info.base.output_semantic_index[attrib];
  559.       if ((shader->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_COLOR) &&
  560.            ((cbuf < key->nr_cbufs) || (cbuf == 1 && dual_source_blend)))
  561.       {
  562.          for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
  563.             if(outputs[attrib][chan]) {
  564.                /* XXX: just initialize outputs to point at colors[] and
  565.                 * skip this.
  566.                 */
  567.                LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
  568.                LLVMValueRef color_ptr;
  569.                color_ptr = LLVMBuildGEP(builder, out_color[cbuf][chan],
  570.                                         &loop_state.counter, 1, "");
  571.                lp_build_name(out, "color%u.%c", attrib, "rgba"[chan]);
  572.                LLVMBuildStore(builder, out, color_ptr);
  573.             }
  574.          }
  575.       }
  576.    }
  577.  
  578.    if (key->occlusion_count) {
  579.       LLVMValueRef counter = lp_jit_thread_data_counter(gallivm, thread_data_ptr);
  580.       lp_build_name(counter, "counter");
  581.       lp_build_occlusion_count(gallivm, type,
  582.                                lp_build_mask_value(&mask), counter);
  583.    }
  584.  
  585.    mask_val = lp_build_mask_end(&mask);
  586.    LLVMBuildStore(builder, mask_val, mask_ptr);
  587.    lp_build_for_loop_end(&loop_state);
  588. }
  589.  
  590.  
  591. /**
  592.  * This function will reorder pixels from the fragment shader SoA to memory layout AoS
  593.  *
  594.  * Fragment Shader outputs pixels in small 2x2 blocks
  595.  *  e.g. (0, 0), (1, 0), (0, 1), (1, 1) ; (2, 0) ...
  596.  *
  597.  * However in memory pixels are stored in rows
  598.  *  e.g. (0, 0), (1, 0), (2, 0), (3, 0) ; (0, 1) ...
  599.  *
  600.  * @param type            fragment shader type (4x or 8x float)
  601.  * @param num_fs          number of fs_src
  602.  * @param is_1d           whether we're outputting to a 1d resource
  603.  * @param dst_channels    number of output channels
  604.  * @param fs_src          output from fragment shader
  605.  * @param dst             pointer to store result
  606.  * @param pad_inline      is channel padding inline or at end of row
  607.  * @return                the number of dsts
  608.  */
  609. static int
  610. generate_fs_twiddle(struct gallivm_state *gallivm,
  611.                     struct lp_type type,
  612.                     unsigned num_fs,
  613.                     unsigned dst_channels,
  614.                     LLVMValueRef fs_src[][4],
  615.                     LLVMValueRef* dst,
  616.                     bool pad_inline)
  617. {
  618.    LLVMValueRef src[16];
  619.  
  620.    bool swizzle_pad;
  621.    bool twiddle;
  622.    bool split;
  623.  
  624.    unsigned pixels = type.length / 4;
  625.    unsigned reorder_group;
  626.    unsigned src_channels;
  627.    unsigned src_count;
  628.    unsigned i;
  629.  
  630.    src_channels = dst_channels < 3 ? dst_channels : 4;
  631.    src_count = num_fs * src_channels;
  632.  
  633.    assert(pixels == 2 || pixels == 1);
  634.    assert(num_fs * src_channels <= Elements(src));
  635.  
  636.    /*
  637.     * Transpose from SoA -> AoS
  638.     */
  639.    for (i = 0; i < num_fs; ++i) {
  640.       lp_build_transpose_aos_n(gallivm, type, &fs_src[i][0], src_channels, &src[i * src_channels]);
  641.    }
  642.  
  643.    /*
  644.     * Pick transformation options
  645.     */
  646.    swizzle_pad = false;
  647.    twiddle = false;
  648.    split = false;
  649.    reorder_group = 0;
  650.  
  651.    if (dst_channels == 1) {
  652.       twiddle = true;
  653.  
  654.       if (pixels == 2) {
  655.          split = true;
  656.       }
  657.    } else if (dst_channels == 2) {
  658.       if (pixels == 1) {
  659.          reorder_group = 1;
  660.       }
  661.    } else if (dst_channels > 2) {
  662.       if (pixels == 1) {
  663.          reorder_group = 2;
  664.       } else {
  665.          twiddle = true;
  666.       }
  667.  
  668.       if (!pad_inline && dst_channels == 3 && pixels > 1) {
  669.          swizzle_pad = true;
  670.       }
  671.    }
  672.  
  673.    /*
  674.     * Split the src in half
  675.     */
  676.    if (split) {
  677.       for (i = num_fs; i > 0; --i) {
  678.          src[(i - 1)*2 + 1] = lp_build_extract_range(gallivm, src[i - 1], 4, 4);
  679.          src[(i - 1)*2 + 0] = lp_build_extract_range(gallivm, src[i - 1], 0, 4);
  680.       }
  681.  
  682.       src_count *= 2;
  683.       type.length = 4;
  684.    }
  685.  
  686.    /*
  687.     * Ensure pixels are in memory order
  688.     */
  689.    if (reorder_group) {
  690.       /* Twiddle pixels by reordering the array, e.g.:
  691.        *
  692.        * src_count =  8 -> 0 2 1 3 4 6 5 7
  693.        * src_count = 16 -> 0 1 4 5 2 3 6 7 8 9 12 13 10 11 14 15
  694.        */
  695.       const unsigned reorder_sw[] = { 0, 2, 1, 3 };
  696.  
  697.       for (i = 0; i < src_count; ++i) {
  698.          unsigned group = i / reorder_group;
  699.          unsigned block = (group / 4) * 4 * reorder_group;
  700.          unsigned j = block + (reorder_sw[group % 4] * reorder_group) + (i % reorder_group);
  701.          dst[i] = src[j];
  702.       }
  703.    } else if (twiddle) {
  704.       /* Twiddle pixels across elements of array */
  705.       lp_bld_quad_twiddle(gallivm, type, src, src_count, dst);
  706.    } else {
  707.       /* Do nothing */
  708.       memcpy(dst, src, sizeof(LLVMValueRef) * src_count);
  709.    }
  710.  
  711.    /*
  712.     * Moves any padding between pixels to the end
  713.     * e.g. RGBXRGBX -> RGBRGBXX
  714.     */
  715.    if (swizzle_pad) {
  716.       unsigned char swizzles[16];
  717.       unsigned elems = pixels * dst_channels;
  718.  
  719.       for (i = 0; i < type.length; ++i) {
  720.          if (i < elems)
  721.             swizzles[i] = i % dst_channels + (i / dst_channels) * 4;
  722.          else
  723.             swizzles[i] = LP_BLD_SWIZZLE_DONTCARE;
  724.       }
  725.  
  726.       for (i = 0; i < src_count; ++i) {
  727.          dst[i] = lp_build_swizzle_aos_n(gallivm, dst[i], swizzles, type.length, type.length);
  728.       }
  729.    }
  730.  
  731.    return src_count;
  732. }
  733.  
  734.  
  735. /**
  736.  * Load an unswizzled block of pixels from memory
  737.  */
  738. static void
  739. load_unswizzled_block(struct gallivm_state *gallivm,
  740.                       LLVMValueRef base_ptr,
  741.                       LLVMValueRef stride,
  742.                       unsigned block_width,
  743.                       unsigned block_height,
  744.                       LLVMValueRef* dst,
  745.                       struct lp_type dst_type,
  746.                       unsigned dst_count,
  747.                       unsigned dst_alignment)
  748. {
  749.    LLVMBuilderRef builder = gallivm->builder;
  750.    unsigned row_size = dst_count / block_height;
  751.    unsigned i;
  752.  
  753.    /* Ensure block exactly fits into dst */
  754.    assert((block_width * block_height) % dst_count == 0);
  755.  
  756.    for (i = 0; i < dst_count; ++i) {
  757.       unsigned x = i % row_size;
  758.       unsigned y = i / row_size;
  759.  
  760.       LLVMValueRef bx = lp_build_const_int32(gallivm, x * (dst_type.width / 8) * dst_type.length);
  761.       LLVMValueRef by = LLVMBuildMul(builder, lp_build_const_int32(gallivm, y), stride, "");
  762.  
  763.       LLVMValueRef gep[2];
  764.       LLVMValueRef dst_ptr;
  765.  
  766.       gep[0] = lp_build_const_int32(gallivm, 0);
  767.       gep[1] = LLVMBuildAdd(builder, bx, by, "");
  768.  
  769.       dst_ptr = LLVMBuildGEP(builder, base_ptr, gep, 2, "");
  770.       dst_ptr = LLVMBuildBitCast(builder, dst_ptr, LLVMPointerType(lp_build_vec_type(gallivm, dst_type), 0), "");
  771.  
  772.       dst[i] = LLVMBuildLoad(builder, dst_ptr, "");
  773.  
  774.       lp_set_load_alignment(dst[i], dst_alignment);
  775.    }
  776. }
  777.  
  778.  
  779. /**
  780.  * Store an unswizzled block of pixels to memory
  781.  */
  782. static void
  783. store_unswizzled_block(struct gallivm_state *gallivm,
  784.                        LLVMValueRef base_ptr,
  785.                        LLVMValueRef stride,
  786.                        unsigned block_width,
  787.                        unsigned block_height,
  788.                        LLVMValueRef* src,
  789.                        struct lp_type src_type,
  790.                        unsigned src_count,
  791.                        unsigned src_alignment)
  792. {
  793.    LLVMBuilderRef builder = gallivm->builder;
  794.    unsigned row_size = src_count / block_height;
  795.    unsigned i;
  796.  
  797.    /* Ensure src exactly fits into block */
  798.    assert((block_width * block_height) % src_count == 0);
  799.  
  800.    for (i = 0; i < src_count; ++i) {
  801.       unsigned x = i % row_size;
  802.       unsigned y = i / row_size;
  803.  
  804.       LLVMValueRef bx = lp_build_const_int32(gallivm, x * (src_type.width / 8) * src_type.length);
  805.       LLVMValueRef by = LLVMBuildMul(builder, lp_build_const_int32(gallivm, y), stride, "");
  806.  
  807.       LLVMValueRef gep[2];
  808.       LLVMValueRef src_ptr;
  809.  
  810.       gep[0] = lp_build_const_int32(gallivm, 0);
  811.       gep[1] = LLVMBuildAdd(builder, bx, by, "");
  812.  
  813.       src_ptr = LLVMBuildGEP(builder, base_ptr, gep, 2, "");
  814.       src_ptr = LLVMBuildBitCast(builder, src_ptr, LLVMPointerType(lp_build_vec_type(gallivm, src_type), 0), "");
  815.  
  816.       src_ptr = LLVMBuildStore(builder, src[i], src_ptr);
  817.  
  818.       lp_set_store_alignment(src_ptr, src_alignment);
  819.    }
  820. }
  821.  
  822.  
  823. /**
  824.  * Checks if a format description is an arithmetic format
  825.  *
  826.  * A format which has irregular channel sizes such as R3_G3_B2 or R5_G6_B5.
  827.  */
  828. static INLINE boolean
  829. is_arithmetic_format(const struct util_format_description *format_desc)
  830. {
  831.    boolean arith = false;
  832.    unsigned i;
  833.  
  834.    for (i = 0; i < format_desc->nr_channels; ++i) {
  835.       arith |= format_desc->channel[i].size != format_desc->channel[0].size;
  836.       arith |= (format_desc->channel[i].size % 8) != 0;
  837.    }
  838.  
  839.    return arith;
  840. }
  841.  
  842.  
  843. /**
  844.  * Checks if this format requires special handling due to required expansion
  845.  * to floats for blending, and furthermore has "natural" packed AoS -> unpacked
  846.  * SoA conversion.
  847.  */
  848. static INLINE boolean
  849. format_expands_to_float_soa(const struct util_format_description *format_desc)
  850. {
  851.    if (format_desc->format == PIPE_FORMAT_R11G11B10_FLOAT ||
  852.        format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB) {
  853.       return true;
  854.    }
  855.    return false;
  856. }
  857.  
  858.  
  859. /**
  860.  * Retrieves the type representing the memory layout for a format
  861.  *
  862.  * e.g. RGBA16F = 4x half-float and R3G3B2 = 1x byte
  863.  */
  864. static INLINE void
  865. lp_mem_type_from_format_desc(const struct util_format_description *format_desc,
  866.                              struct lp_type* type)
  867. {
  868.    unsigned i;
  869.    unsigned chan;
  870.  
  871.    if (format_expands_to_float_soa(format_desc)) {
  872.       /* just make this a uint with width of block */
  873.       type->floating = false;
  874.       type->fixed = false;
  875.       type->sign = false;
  876.       type->norm = false;
  877.       type->width = format_desc->block.bits;
  878.       type->length = 1;
  879.       return;
  880.    }
  881.  
  882.    for (i = 0; i < 4; i++)
  883.       if (format_desc->channel[i].type != UTIL_FORMAT_TYPE_VOID)
  884.          break;
  885.    chan = i;
  886.  
  887.    memset(type, 0, sizeof(struct lp_type));
  888.    type->floating = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FLOAT;
  889.    type->fixed    = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FIXED;
  890.    type->sign     = format_desc->channel[chan].type != UTIL_FORMAT_TYPE_UNSIGNED;
  891.    type->norm     = format_desc->channel[chan].normalized;
  892.  
  893.    if (is_arithmetic_format(format_desc)) {
  894.       type->width = 0;
  895.       type->length = 1;
  896.  
  897.       for (i = 0; i < format_desc->nr_channels; ++i) {
  898.          type->width += format_desc->channel[i].size;
  899.       }
  900.    } else {
  901.       type->width = format_desc->channel[chan].size;
  902.       type->length = format_desc->nr_channels;
  903.    }
  904. }
  905.  
  906.  
  907. /**
  908.  * Retrieves the type for a format which is usable in the blending code.
  909.  *
  910.  * e.g. RGBA16F = 4x float, R3G3B2 = 3x byte
  911.  */
  912. static INLINE void
  913. lp_blend_type_from_format_desc(const struct util_format_description *format_desc,
  914.                                struct lp_type* type)
  915. {
  916.    unsigned i;
  917.    unsigned chan;
  918.  
  919.    if (format_expands_to_float_soa(format_desc)) {
  920.       /* always use ordinary floats for blending */
  921.       type->floating = true;
  922.       type->fixed = false;
  923.       type->sign = true;
  924.       type->norm = false;
  925.       type->width = 32;
  926.       type->length = 4;
  927.       return;
  928.    }
  929.  
  930.    for (i = 0; i < 4; i++)
  931.       if (format_desc->channel[i].type != UTIL_FORMAT_TYPE_VOID)
  932.          break;
  933.    chan = i;
  934.  
  935.    memset(type, 0, sizeof(struct lp_type));
  936.    type->floating = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FLOAT;
  937.    type->fixed    = format_desc->channel[chan].type == UTIL_FORMAT_TYPE_FIXED;
  938.    type->sign     = format_desc->channel[chan].type != UTIL_FORMAT_TYPE_UNSIGNED;
  939.    type->norm     = format_desc->channel[chan].normalized;
  940.    type->width    = format_desc->channel[chan].size;
  941.    type->length   = format_desc->nr_channels;
  942.  
  943.    for (i = 1; i < format_desc->nr_channels; ++i) {
  944.       if (format_desc->channel[i].size > type->width)
  945.          type->width = format_desc->channel[i].size;
  946.    }
  947.  
  948.    if (type->floating) {
  949.       type->width = 32;
  950.    } else {
  951.       if (type->width <= 8) {
  952.          type->width = 8;
  953.       } else if (type->width <= 16) {
  954.          type->width = 16;
  955.       } else {
  956.          type->width = 32;
  957.       }
  958.    }
  959.  
  960.    if (is_arithmetic_format(format_desc) && type->length == 3) {
  961.       type->length = 4;
  962.    }
  963. }
  964.  
  965.  
  966. /**
  967.  * Scale a normalized value from src_bits to dst_bits.
  968.  *
  969.  * The exact calculation is
  970.  *
  971.  *    dst = iround(src * dst_mask / src_mask)
  972.  *
  973.  *  or with integer rounding
  974.  *
  975.  *    dst = src * (2*dst_mask + sign(src)*src_mask) / (2*src_mask)
  976.  *
  977.  *  where
  978.  *
  979.  *    src_mask = (1 << src_bits) - 1
  980.  *    dst_mask = (1 << dst_bits) - 1
  981.  *
  982.  * but we try to avoid division and multiplication through shifts.
  983.  */
  984. static INLINE LLVMValueRef
  985. scale_bits(struct gallivm_state *gallivm,
  986.            int src_bits,
  987.            int dst_bits,
  988.            LLVMValueRef src,
  989.            struct lp_type src_type)
  990. {
  991.    LLVMBuilderRef builder = gallivm->builder;
  992.    LLVMValueRef result = src;
  993.  
  994.    if (dst_bits < src_bits) {
  995.       int delta_bits = src_bits - dst_bits;
  996.  
  997.       if (delta_bits <= dst_bits) {
  998.          /*
  999.           * Approximate the rescaling with a single shift.
  1000.           *
  1001.           * This gives the wrong rounding.
  1002.           */
  1003.  
  1004.          result = LLVMBuildLShr(builder,
  1005.                                 src,
  1006.                                 lp_build_const_int_vec(gallivm, src_type, delta_bits),
  1007.                                 "");
  1008.  
  1009.       } else {
  1010.          /*
  1011.           * Try more accurate rescaling.
  1012.           */
  1013.  
  1014.          /*
  1015.           * Drop the least significant bits to make space for the multiplication.
  1016.           *
  1017.           * XXX: A better approach would be to use a wider integer type as intermediate.  But
  1018.           * this is enough to convert alpha from 16bits -> 2 when rendering to
  1019.           * PIPE_FORMAT_R10G10B10A2_UNORM.
  1020.           */
  1021.          result = LLVMBuildLShr(builder,
  1022.                                 src,
  1023.                                 lp_build_const_int_vec(gallivm, src_type, dst_bits),
  1024.                                 "");
  1025.  
  1026.  
  1027.          result = LLVMBuildMul(builder,
  1028.                                result,
  1029.                                lp_build_const_int_vec(gallivm, src_type, (1LL << dst_bits) - 1),
  1030.                                "");
  1031.  
  1032.          /*
  1033.           * Add a rounding term before the division.
  1034.           *
  1035.           * TODO: Handle signed integers too.
  1036.           */
  1037.          if (!src_type.sign) {
  1038.             result = LLVMBuildAdd(builder,
  1039.                                   result,
  1040.                                   lp_build_const_int_vec(gallivm, src_type, (1LL << (delta_bits - 1))),
  1041.                                   "");
  1042.          }
  1043.  
  1044.          /*
  1045.           * Approximate the division by src_mask with a src_bits shift.
  1046.           *
  1047.           * Given the src has already been shifted by dst_bits, all we need
  1048.           * to do is to shift by the difference.
  1049.           */
  1050.  
  1051.          result = LLVMBuildLShr(builder,
  1052.                                 result,
  1053.                                 lp_build_const_int_vec(gallivm, src_type, delta_bits),
  1054.                                 "");
  1055.       }
  1056.  
  1057.    } else if (dst_bits > src_bits) {
  1058.       /* Scale up bits */
  1059.       int db = dst_bits - src_bits;
  1060.  
  1061.       /* Shift left by difference in bits */
  1062.       result = LLVMBuildShl(builder,
  1063.                             src,
  1064.                             lp_build_const_int_vec(gallivm, src_type, db),
  1065.                             "");
  1066.  
  1067.       if (db < src_bits) {
  1068.          /* Enough bits in src to fill the remainder */
  1069.          LLVMValueRef lower = LLVMBuildLShr(builder,
  1070.                                             src,
  1071.                                             lp_build_const_int_vec(gallivm, src_type, src_bits - db),
  1072.                                             "");
  1073.  
  1074.          result = LLVMBuildOr(builder, result, lower, "");
  1075.       } else if (db > src_bits) {
  1076.          /* Need to repeatedly copy src bits to fill remainder in dst */
  1077.          unsigned n;
  1078.  
  1079.          for (n = src_bits; n < dst_bits; n *= 2) {
  1080.             LLVMValueRef shuv = lp_build_const_int_vec(gallivm, src_type, n);
  1081.  
  1082.             result = LLVMBuildOr(builder,
  1083.                                  result,
  1084.                                  LLVMBuildLShr(builder, result, shuv, ""),
  1085.                                  "");
  1086.          }
  1087.       }
  1088.    }
  1089.  
  1090.    return result;
  1091. }
  1092.  
  1093. /**
  1094.  * If RT is a smallfloat (needing denorms) format
  1095.  */
  1096. static INLINE int
  1097. have_smallfloat_format(struct lp_type dst_type,
  1098.                        enum pipe_format format)
  1099. {
  1100.    return ((dst_type.floating && dst_type.width != 32) ||
  1101.     /* due to format handling hacks this format doesn't have floating set
  1102.      * here (and actually has width set to 32 too) so special case this. */
  1103.     (format == PIPE_FORMAT_R11G11B10_FLOAT));
  1104. }
  1105.  
  1106.  
  1107. /**
  1108.  * Convert from memory format to blending format
  1109.  *
  1110.  * e.g. GL_R3G3B2 is 1 byte in memory but 3 bytes for blending
  1111.  */
  1112. static void
  1113. convert_to_blend_type(struct gallivm_state *gallivm,
  1114.                       unsigned block_size,
  1115.                       const struct util_format_description *src_fmt,
  1116.                       struct lp_type src_type,
  1117.                       struct lp_type dst_type,
  1118.                       LLVMValueRef* src, // and dst
  1119.                       unsigned num_srcs)
  1120. {
  1121.    LLVMValueRef *dst = src;
  1122.    LLVMBuilderRef builder = gallivm->builder;
  1123.    struct lp_type blend_type;
  1124.    struct lp_type mem_type;
  1125.    unsigned i, j, k;
  1126.    unsigned pixels = block_size / num_srcs;
  1127.    bool is_arith;
  1128.  
  1129.    /*
  1130.     * full custom path for packed floats and srgb formats - none of the later
  1131.     * functions would do anything useful, and given the lp_type representation they
  1132.     * can't be fixed. Should really have some SoA blend path for these kind of
  1133.     * formats rather than hacking them in here.
  1134.     */
  1135.    if (format_expands_to_float_soa(src_fmt)) {
  1136.       LLVMValueRef tmpsrc[4];
  1137.       /*
  1138.        * This is pretty suboptimal for this case blending in SoA would be much
  1139.        * better, since conversion gets us SoA values so need to convert back.
  1140.        */
  1141.       assert(src_type.width == 32 || src_type.width == 16);
  1142.       assert(dst_type.floating);
  1143.       assert(dst_type.width == 32);
  1144.       assert(dst_type.length % 4 == 0);
  1145.       assert(num_srcs % 4 == 0);
  1146.  
  1147.       if (src_type.width == 16) {
  1148.          /* expand 4x16bit values to 4x32bit */
  1149.          struct lp_type type32x4 = src_type;
  1150.          LLVMTypeRef ltype32x4;
  1151.          unsigned num_fetch = dst_type.length == 8 ? num_srcs / 2 : num_srcs / 4;
  1152.          type32x4.width = 32;
  1153.          ltype32x4 = lp_build_vec_type(gallivm, type32x4);
  1154.          for (i = 0; i < num_fetch; i++) {
  1155.             src[i] = LLVMBuildZExt(builder, src[i], ltype32x4, "");
  1156.          }
  1157.          src_type.width = 32;
  1158.       }
  1159.       for (i = 0; i < 4; i++) {
  1160.          tmpsrc[i] = src[i];
  1161.       }
  1162.       for (i = 0; i < num_srcs / 4; i++) {
  1163.          LLVMValueRef tmpsoa[4];
  1164.          LLVMValueRef tmps = tmpsrc[i];
  1165.          if (dst_type.length == 8) {
  1166.             LLVMValueRef shuffles[8];
  1167.             unsigned j;
  1168.             /* fetch was 4 values but need 8-wide output values */
  1169.             tmps = lp_build_concat(gallivm, &tmpsrc[i * 2], src_type, 2);
  1170.             /*
  1171.              * for 8-wide aos transpose would give us wrong order not matching
  1172.              * incoming converted fs values and mask. ARGH.
  1173.              */
  1174.             for (j = 0; j < 4; j++) {
  1175.                shuffles[j] = lp_build_const_int32(gallivm, j * 2);
  1176.                shuffles[j + 4] = lp_build_const_int32(gallivm, j * 2 + 1);
  1177.             }
  1178.             tmps = LLVMBuildShuffleVector(builder, tmps, tmps,
  1179.                                           LLVMConstVector(shuffles, 8), "");
  1180.          }
  1181.          if (src_fmt->format == PIPE_FORMAT_R11G11B10_FLOAT) {
  1182.             lp_build_r11g11b10_to_float(gallivm, tmps, tmpsoa);
  1183.          }
  1184.          else {
  1185.             lp_build_unpack_rgba_soa(gallivm, src_fmt, dst_type, tmps, tmpsoa);
  1186.          }
  1187.          lp_build_transpose_aos(gallivm, dst_type, tmpsoa, &src[i * 4]);
  1188.       }
  1189.       return;
  1190.    }
  1191.  
  1192.    lp_mem_type_from_format_desc(src_fmt, &mem_type);
  1193.    lp_blend_type_from_format_desc(src_fmt, &blend_type);
  1194.  
  1195.    /* Is the format arithmetic */
  1196.    is_arith = blend_type.length * blend_type.width != mem_type.width * mem_type.length;
  1197.    is_arith &= !(mem_type.width == 16 && mem_type.floating);
  1198.  
  1199.    /* Pad if necessary */
  1200.    if (!is_arith && src_type.length < dst_type.length) {
  1201.       for (i = 0; i < num_srcs; ++i) {
  1202.          dst[i] = lp_build_pad_vector(gallivm, src[i], dst_type.length);
  1203.       }
  1204.  
  1205.       src_type.length = dst_type.length;
  1206.    }
  1207.  
  1208.    /* Special case for half-floats */
  1209.    if (mem_type.width == 16 && mem_type.floating) {
  1210.       assert(blend_type.width == 32 && blend_type.floating);
  1211.       lp_build_conv_auto(gallivm, src_type, &dst_type, dst, num_srcs, dst);
  1212.       is_arith = false;
  1213.    }
  1214.  
  1215.    if (!is_arith) {
  1216.       return;
  1217.    }
  1218.  
  1219.    src_type.width = blend_type.width * blend_type.length;
  1220.    blend_type.length *= pixels;
  1221.    src_type.length *= pixels / (src_type.length / mem_type.length);
  1222.  
  1223.    for (i = 0; i < num_srcs; ++i) {
  1224.       LLVMValueRef chans[4];
  1225.       LLVMValueRef res = NULL;
  1226.  
  1227.       dst[i] = LLVMBuildZExt(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
  1228.  
  1229.       for (j = 0; j < src_fmt->nr_channels; ++j) {
  1230.          unsigned mask = 0;
  1231.          unsigned sa = src_fmt->channel[j].shift;
  1232. #ifdef PIPE_ARCH_LITTLE_ENDIAN
  1233.          unsigned from_lsb = j;
  1234. #else
  1235.          unsigned from_lsb = src_fmt->nr_channels - j - 1;
  1236. #endif
  1237.  
  1238.          for (k = 0; k < src_fmt->channel[j].size; ++k) {
  1239.             mask |= 1 << k;
  1240.          }
  1241.  
  1242.          /* Extract bits from source */
  1243.          chans[j] = LLVMBuildLShr(builder,
  1244.                                   dst[i],
  1245.                                   lp_build_const_int_vec(gallivm, src_type, sa),
  1246.                                   "");
  1247.  
  1248.          chans[j] = LLVMBuildAnd(builder,
  1249.                                  chans[j],
  1250.                                  lp_build_const_int_vec(gallivm, src_type, mask),
  1251.                                  "");
  1252.  
  1253.          /* Scale bits */
  1254.          if (src_type.norm) {
  1255.             chans[j] = scale_bits(gallivm, src_fmt->channel[j].size,
  1256.                                   blend_type.width, chans[j], src_type);
  1257.          }
  1258.  
  1259.          /* Insert bits into correct position */
  1260.          chans[j] = LLVMBuildShl(builder,
  1261.                                  chans[j],
  1262.                                  lp_build_const_int_vec(gallivm, src_type, from_lsb * blend_type.width),
  1263.                                  "");
  1264.  
  1265.          if (j == 0) {
  1266.             res = chans[j];
  1267.          } else {
  1268.             res = LLVMBuildOr(builder, res, chans[j], "");
  1269.          }
  1270.       }
  1271.  
  1272.       dst[i] = LLVMBuildBitCast(builder, res, lp_build_vec_type(gallivm, blend_type), "");
  1273.    }
  1274. }
  1275.  
  1276.  
  1277. /**
  1278.  * Convert from blending format to memory format
  1279.  *
  1280.  * e.g. GL_R3G3B2 is 3 bytes for blending but 1 byte in memory
  1281.  */
  1282. static void
  1283. convert_from_blend_type(struct gallivm_state *gallivm,
  1284.                         unsigned block_size,
  1285.                         const struct util_format_description *src_fmt,
  1286.                         struct lp_type src_type,
  1287.                         struct lp_type dst_type,
  1288.                         LLVMValueRef* src, // and dst
  1289.                         unsigned num_srcs)
  1290. {
  1291.    LLVMValueRef* dst = src;
  1292.    unsigned i, j, k;
  1293.    struct lp_type mem_type;
  1294.    struct lp_type blend_type;
  1295.    LLVMBuilderRef builder = gallivm->builder;
  1296.    unsigned pixels = block_size / num_srcs;
  1297.    bool is_arith;
  1298.  
  1299.    /*
  1300.     * full custom path for packed floats and srgb formats - none of the later
  1301.     * functions would do anything useful, and given the lp_type representation they
  1302.     * can't be fixed. Should really have some SoA blend path for these kind of
  1303.     * formats rather than hacking them in here.
  1304.     */
  1305.    if (format_expands_to_float_soa(src_fmt)) {
  1306.       /*
  1307.        * This is pretty suboptimal for this case blending in SoA would be much
  1308.        * better - we need to transpose the AoS values back to SoA values for
  1309.        * conversion/packing.
  1310.        */
  1311.       assert(src_type.floating);
  1312.       assert(src_type.width == 32);
  1313.       assert(src_type.length % 4 == 0);
  1314.       assert(dst_type.width == 32 || dst_type.width == 16);
  1315.  
  1316.       for (i = 0; i < num_srcs / 4; i++) {
  1317.          LLVMValueRef tmpsoa[4], tmpdst;
  1318.          lp_build_transpose_aos(gallivm, src_type, &src[i * 4], tmpsoa);
  1319.          /* really really need SoA here */
  1320.  
  1321.          if (src_fmt->format == PIPE_FORMAT_R11G11B10_FLOAT) {
  1322.             tmpdst = lp_build_float_to_r11g11b10(gallivm, tmpsoa);
  1323.          }
  1324.          else {
  1325.             tmpdst = lp_build_float_to_srgb_packed(gallivm, src_fmt,
  1326.                                                    src_type, tmpsoa);
  1327.          }
  1328.  
  1329.          if (src_type.length == 8) {
  1330.             LLVMValueRef tmpaos, shuffles[8];
  1331.             unsigned j;
  1332.             /*
  1333.              * for 8-wide aos transpose has given us wrong order not matching
  1334.              * output order. HMPF. Also need to split the output values manually.
  1335.              */
  1336.             for (j = 0; j < 4; j++) {
  1337.                shuffles[j * 2] = lp_build_const_int32(gallivm, j);
  1338.                shuffles[j * 2 + 1] = lp_build_const_int32(gallivm, j + 4);
  1339.             }
  1340.             tmpaos = LLVMBuildShuffleVector(builder, tmpdst, tmpdst,
  1341.                                             LLVMConstVector(shuffles, 8), "");
  1342.             src[i * 2] = lp_build_extract_range(gallivm, tmpaos, 0, 4);
  1343.             src[i * 2 + 1] = lp_build_extract_range(gallivm, tmpaos, 4, 4);
  1344.          }
  1345.          else {
  1346.             src[i] = tmpdst;
  1347.          }
  1348.       }
  1349.       if (dst_type.width == 16) {
  1350.          struct lp_type type16x8 = dst_type;
  1351.          struct lp_type type32x4 = dst_type;
  1352.          LLVMTypeRef ltype16x4, ltypei64, ltypei128;
  1353.          unsigned num_fetch = src_type.length == 8 ? num_srcs / 2 : num_srcs / 4;
  1354.          type16x8.length = 8;
  1355.          type32x4.width = 32;
  1356.          ltypei128 = LLVMIntTypeInContext(gallivm->context, 128);
  1357.          ltypei64 = LLVMIntTypeInContext(gallivm->context, 64);
  1358.          ltype16x4 = lp_build_vec_type(gallivm, dst_type);
  1359.          /* We could do vector truncation but it doesn't generate very good code */
  1360.          for (i = 0; i < num_fetch; i++) {
  1361.             src[i] = lp_build_pack2(gallivm, type32x4, type16x8,
  1362.                                     src[i], lp_build_zero(gallivm, type32x4));
  1363.             src[i] = LLVMBuildBitCast(builder, src[i], ltypei128, "");
  1364.             src[i] = LLVMBuildTrunc(builder, src[i], ltypei64, "");
  1365.             src[i] = LLVMBuildBitCast(builder, src[i], ltype16x4, "");
  1366.          }
  1367.       }
  1368.       return;
  1369.    }
  1370.  
  1371.    lp_mem_type_from_format_desc(src_fmt, &mem_type);
  1372.    lp_blend_type_from_format_desc(src_fmt, &blend_type);
  1373.  
  1374.    is_arith = (blend_type.length * blend_type.width != mem_type.width * mem_type.length);
  1375.  
  1376.    /* Special case for half-floats */
  1377.    if (mem_type.width == 16 && mem_type.floating) {
  1378.       int length = dst_type.length;
  1379.       assert(blend_type.width == 32 && blend_type.floating);
  1380.  
  1381.       dst_type.length = src_type.length;
  1382.  
  1383.       lp_build_conv_auto(gallivm, src_type, &dst_type, dst, num_srcs, dst);
  1384.  
  1385.       dst_type.length = length;
  1386.       is_arith = false;
  1387.    }
  1388.  
  1389.    /* Remove any padding */
  1390.    if (!is_arith && (src_type.length % mem_type.length)) {
  1391.       src_type.length -= (src_type.length % mem_type.length);
  1392.  
  1393.       for (i = 0; i < num_srcs; ++i) {
  1394.          dst[i] = lp_build_extract_range(gallivm, dst[i], 0, src_type.length);
  1395.       }
  1396.    }
  1397.  
  1398.    /* No bit arithmetic to do */
  1399.    if (!is_arith) {
  1400.       return;
  1401.    }
  1402.  
  1403.    src_type.length = pixels;
  1404.    src_type.width = blend_type.length * blend_type.width;
  1405.    dst_type.length = pixels;
  1406.  
  1407.    for (i = 0; i < num_srcs; ++i) {
  1408.       LLVMValueRef chans[4];
  1409.       LLVMValueRef res = NULL;
  1410.  
  1411.       dst[i] = LLVMBuildBitCast(builder, src[i], lp_build_vec_type(gallivm, src_type), "");
  1412.  
  1413.       for (j = 0; j < src_fmt->nr_channels; ++j) {
  1414.          unsigned mask = 0;
  1415.          unsigned sa = src_fmt->channel[j].shift;
  1416. #ifdef PIPE_ARCH_LITTLE_ENDIAN
  1417.          unsigned from_lsb = j;
  1418. #else
  1419.          unsigned from_lsb = src_fmt->nr_channels - j - 1;
  1420. #endif
  1421.  
  1422.          assert(blend_type.width > src_fmt->channel[j].size);
  1423.  
  1424.          for (k = 0; k < blend_type.width; ++k) {
  1425.             mask |= 1 << k;
  1426.          }
  1427.  
  1428.          /* Extract bits */
  1429.          chans[j] = LLVMBuildLShr(builder,
  1430.                                   dst[i],
  1431.                                   lp_build_const_int_vec(gallivm, src_type, from_lsb * blend_type.width),
  1432.                                   "");
  1433.  
  1434.          chans[j] = LLVMBuildAnd(builder,
  1435.                                  chans[j],
  1436.                                  lp_build_const_int_vec(gallivm, src_type, mask),
  1437.                                  "");
  1438.  
  1439.          /* Scale down bits */
  1440.          if (src_type.norm) {
  1441.             chans[j] = scale_bits(gallivm, blend_type.width,
  1442.                                   src_fmt->channel[j].size, chans[j], src_type);
  1443.          }
  1444.  
  1445.          /* Insert bits */
  1446.          chans[j] = LLVMBuildShl(builder,
  1447.                                  chans[j],
  1448.                                  lp_build_const_int_vec(gallivm, src_type, sa),
  1449.                                  "");
  1450.  
  1451.          sa += src_fmt->channel[j].size;
  1452.  
  1453.          if (j == 0) {
  1454.             res = chans[j];
  1455.          } else {
  1456.             res = LLVMBuildOr(builder, res, chans[j], "");
  1457.          }
  1458.       }
  1459.  
  1460.       assert (dst_type.width != 24);
  1461.  
  1462.       dst[i] = LLVMBuildTrunc(builder, res, lp_build_vec_type(gallivm, dst_type), "");
  1463.    }
  1464. }
  1465.  
  1466.  
  1467. /**
  1468.  * Convert alpha to same blend type as src
  1469.  */
  1470. static void
  1471. convert_alpha(struct gallivm_state *gallivm,
  1472.               struct lp_type row_type,
  1473.               struct lp_type alpha_type,
  1474.               const unsigned block_size,
  1475.               const unsigned block_height,
  1476.               const unsigned src_count,
  1477.               const unsigned dst_channels,
  1478.               const bool pad_inline,
  1479.               LLVMValueRef* src_alpha)
  1480. {
  1481.    LLVMBuilderRef builder = gallivm->builder;
  1482.    unsigned i, j;
  1483.    unsigned length = row_type.length;
  1484.    row_type.length = alpha_type.length;
  1485.  
  1486.    /* Twiddle the alpha to match pixels */
  1487.    lp_bld_quad_twiddle(gallivm, alpha_type, src_alpha, block_height, src_alpha);
  1488.  
  1489.    /*
  1490.     * TODO this should use single lp_build_conv call for
  1491.     * src_count == 1 && dst_channels == 1 case (dropping the concat below)
  1492.     */
  1493.    for (i = 0; i < block_height; ++i) {
  1494.       lp_build_conv(gallivm, alpha_type, row_type, &src_alpha[i], 1, &src_alpha[i], 1);
  1495.    }
  1496.  
  1497.    alpha_type = row_type;
  1498.    row_type.length = length;
  1499.  
  1500.    /* If only one channel we can only need the single alpha value per pixel */
  1501.    if (src_count == 1 && dst_channels == 1) {
  1502.  
  1503.       lp_build_concat_n(gallivm, alpha_type, src_alpha, block_height, src_alpha, src_count);
  1504.    } else {
  1505.       /* If there are more srcs than rows then we need to split alpha up */
  1506.       if (src_count > block_height) {
  1507.          for (i = src_count; i > 0; --i) {
  1508.             unsigned pixels = block_size / src_count;
  1509.             unsigned idx = i - 1;
  1510.  
  1511.             src_alpha[idx] = lp_build_extract_range(gallivm, src_alpha[(idx * pixels) / 4],
  1512.                                                     (idx * pixels) % 4, pixels);
  1513.          }
  1514.       }
  1515.  
  1516.       /* If there is a src for each pixel broadcast the alpha across whole row */
  1517.       if (src_count == block_size) {
  1518.          for (i = 0; i < src_count; ++i) {
  1519.             src_alpha[i] = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, row_type), src_alpha[i]);
  1520.          }
  1521.       } else {
  1522.          unsigned pixels = block_size / src_count;
  1523.          unsigned channels = pad_inline ? TGSI_NUM_CHANNELS : dst_channels;
  1524.          unsigned alpha_span = 1;
  1525.          LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
  1526.  
  1527.          /* Check if we need 2 src_alphas for our shuffles */
  1528.          if (pixels > alpha_type.length) {
  1529.             alpha_span = 2;
  1530.          }
  1531.  
  1532.          /* Broadcast alpha across all channels, e.g. a1a2 to a1a1a1a1a2a2a2a2 */
  1533.          for (j = 0; j < row_type.length; ++j) {
  1534.             if (j < pixels * channels) {
  1535.                shuffles[j] = lp_build_const_int32(gallivm, j / channels);
  1536.             } else {
  1537.                shuffles[j] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
  1538.             }
  1539.          }
  1540.  
  1541.          for (i = 0; i < src_count; ++i) {
  1542.             unsigned idx1 = i, idx2 = i;
  1543.  
  1544.             if (alpha_span > 1){
  1545.                idx1 *= alpha_span;
  1546.                idx2 = idx1 + 1;
  1547.             }
  1548.  
  1549.             src_alpha[i] = LLVMBuildShuffleVector(builder,
  1550.                                                   src_alpha[idx1],
  1551.                                                   src_alpha[idx2],
  1552.                                                   LLVMConstVector(shuffles, row_type.length),
  1553.                                                   "");
  1554.          }
  1555.       }
  1556.    }
  1557. }
  1558.  
  1559.  
  1560. /**
  1561.  * Generates the blend function for unswizzled colour buffers
  1562.  * Also generates the read & write from colour buffer
  1563.  */
  1564. static void
  1565. generate_unswizzled_blend(struct gallivm_state *gallivm,
  1566.                           unsigned rt,
  1567.                           struct lp_fragment_shader_variant *variant,
  1568.                           enum pipe_format out_format,
  1569.                           unsigned int num_fs,
  1570.                           struct lp_type fs_type,
  1571.                           LLVMValueRef* fs_mask,
  1572.                           LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][4],
  1573.                           LLVMValueRef context_ptr,
  1574.                           LLVMValueRef color_ptr,
  1575.                           LLVMValueRef stride,
  1576.                           unsigned partial_mask,
  1577.                           boolean do_branch)
  1578. {
  1579.    const unsigned alpha_channel = 3;
  1580.    const unsigned block_width = LP_RASTER_BLOCK_SIZE;
  1581.    const unsigned block_height = LP_RASTER_BLOCK_SIZE;
  1582.    const unsigned block_size = block_width * block_height;
  1583.    const unsigned lp_integer_vector_width = 128;
  1584.  
  1585.    LLVMBuilderRef builder = gallivm->builder;
  1586.    LLVMValueRef fs_src[4][TGSI_NUM_CHANNELS];
  1587.    LLVMValueRef fs_src1[4][TGSI_NUM_CHANNELS];
  1588.    LLVMValueRef src_alpha[4 * 4];
  1589.    LLVMValueRef src1_alpha[4 * 4];
  1590.    LLVMValueRef src_mask[4 * 4];
  1591.    LLVMValueRef src[4 * 4];
  1592.    LLVMValueRef src1[4 * 4];
  1593.    LLVMValueRef dst[4 * 4];
  1594.    LLVMValueRef blend_color;
  1595.    LLVMValueRef blend_alpha;
  1596.    LLVMValueRef i32_zero;
  1597.    LLVMValueRef check_mask;
  1598.    LLVMValueRef undef_src_val;
  1599.  
  1600.    struct lp_build_mask_context mask_ctx;
  1601.    struct lp_type mask_type;
  1602.    struct lp_type blend_type;
  1603.    struct lp_type row_type;
  1604.    struct lp_type dst_type;
  1605.  
  1606.    unsigned char swizzle[TGSI_NUM_CHANNELS];
  1607.    unsigned vector_width;
  1608.    unsigned src_channels = TGSI_NUM_CHANNELS;
  1609.    unsigned dst_channels;
  1610.    unsigned dst_count;
  1611.    unsigned src_count;
  1612.    unsigned i, j;
  1613.  
  1614.    const struct util_format_description* out_format_desc = util_format_description(out_format);
  1615.  
  1616.    unsigned dst_alignment;
  1617.  
  1618.    bool pad_inline = is_arithmetic_format(out_format_desc);
  1619.    bool has_alpha = false;
  1620.    const boolean dual_source_blend = variant->key.blend.rt[0].blend_enable &&
  1621.                                      util_blend_state_is_dual(&variant->key.blend, 0);
  1622.  
  1623.    const boolean is_1d = variant->key.resource_1d;
  1624.    unsigned num_fullblock_fs = is_1d ? 2 * num_fs : num_fs;
  1625.    LLVMValueRef fpstate = 0;
  1626.  
  1627.    /* Get type from output format */
  1628.    lp_blend_type_from_format_desc(out_format_desc, &row_type);
  1629.    lp_mem_type_from_format_desc(out_format_desc, &dst_type);
  1630.  
  1631.    /*
  1632.     * Technically this code should go into lp_build_smallfloat_to_float
  1633.     * and lp_build_float_to_smallfloat but due to the
  1634.     * http://llvm.org/bugs/show_bug.cgi?id=6393
  1635.     * llvm reorders the mxcsr intrinsics in a way that breaks the code.
  1636.     * So the ordering is important here and there shouldn't be any
  1637.     * llvm ir instrunctions in this function before
  1638.     * this, otherwise half-float format conversions won't work
  1639.     * (again due to llvm bug #6393).
  1640.     */
  1641.    if (have_smallfloat_format(dst_type, out_format)) {
  1642.       /* We need to make sure that denorms are ok for half float
  1643.          conversions */
  1644.       fpstate = lp_build_fpstate_get(gallivm);
  1645.       lp_build_fpstate_set_denorms_zero(gallivm, FALSE);
  1646.    }
  1647.  
  1648.    mask_type = lp_int32_vec4_type();
  1649.    mask_type.length = fs_type.length;
  1650.  
  1651.    for (i = num_fs; i < num_fullblock_fs; i++) {
  1652.       fs_mask[i] = lp_build_zero(gallivm, mask_type);
  1653.    }
  1654.  
  1655.    /* Do not bother executing code when mask is empty.. */
  1656.    if (do_branch) {
  1657.       check_mask = LLVMConstNull(lp_build_int_vec_type(gallivm, mask_type));
  1658.  
  1659.       for (i = 0; i < num_fullblock_fs; ++i) {
  1660.          check_mask = LLVMBuildOr(builder, check_mask, fs_mask[i], "");
  1661.       }
  1662.  
  1663.       lp_build_mask_begin(&mask_ctx, gallivm, mask_type, check_mask);
  1664.       lp_build_mask_check(&mask_ctx);
  1665.    }
  1666.  
  1667.    partial_mask |= !variant->opaque;
  1668.    i32_zero = lp_build_const_int32(gallivm, 0);
  1669.  
  1670.    undef_src_val = lp_build_undef(gallivm, fs_type);
  1671.  
  1672.    row_type.length = fs_type.length;
  1673.    vector_width    = dst_type.floating ? lp_native_vector_width : lp_integer_vector_width;
  1674.  
  1675.    /* Compute correct swizzle and count channels */
  1676.    memset(swizzle, LP_BLD_SWIZZLE_DONTCARE, TGSI_NUM_CHANNELS);
  1677.    dst_channels = 0;
  1678.  
  1679.    for (i = 0; i < TGSI_NUM_CHANNELS; ++i) {
  1680.       /* Ensure channel is used */
  1681.       if (out_format_desc->swizzle[i] >= TGSI_NUM_CHANNELS) {
  1682.          continue;
  1683.       }
  1684.  
  1685.       /* Ensure not already written to (happens in case with GL_ALPHA) */
  1686.       if (swizzle[out_format_desc->swizzle[i]] < TGSI_NUM_CHANNELS) {
  1687.          continue;
  1688.       }
  1689.  
  1690.       /* Ensure we havn't already found all channels */
  1691.       if (dst_channels >= out_format_desc->nr_channels) {
  1692.          continue;
  1693.       }
  1694.  
  1695.       swizzle[out_format_desc->swizzle[i]] = i;
  1696.       ++dst_channels;
  1697.  
  1698.       if (i == alpha_channel) {
  1699.          has_alpha = true;
  1700.       }
  1701.    }
  1702.  
  1703.    if (format_expands_to_float_soa(out_format_desc)) {
  1704.       /*
  1705.        * the code above can't work for layout_other
  1706.        * for srgb it would sort of work but we short-circuit swizzles, etc.
  1707.        * as that is done as part of unpack / pack.
  1708.        */
  1709.       dst_channels = 4; /* HACK: this is fake 4 really but need it due to transpose stuff later */
  1710.       has_alpha = true;
  1711.       swizzle[0] = 0;
  1712.       swizzle[1] = 1;
  1713.       swizzle[2] = 2;
  1714.       swizzle[3] = 3;
  1715.       pad_inline = true; /* HACK: prevent rgbxrgbx->rgbrgbxx conversion later */
  1716.    }
  1717.  
  1718.    /* If 3 channels then pad to include alpha for 4 element transpose */
  1719.    if (dst_channels == 3 && !has_alpha) {
  1720.       for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
  1721.          if (swizzle[i] > TGSI_NUM_CHANNELS)
  1722.             swizzle[i] = 3;
  1723.       }
  1724.       if (out_format_desc->nr_channels == 4) {
  1725.          dst_channels = 4;
  1726.       }
  1727.    }
  1728.  
  1729.    /*
  1730.     * Load shader output
  1731.     */
  1732.    for (i = 0; i < num_fullblock_fs; ++i) {
  1733.       /* Always load alpha for use in blending */
  1734.       LLVMValueRef alpha;
  1735.       if (i < num_fs) {
  1736.          alpha = LLVMBuildLoad(builder, fs_out_color[rt][alpha_channel][i], "");
  1737.       }
  1738.       else {
  1739.          alpha = undef_src_val;
  1740.       }
  1741.  
  1742.       /* Load each channel */
  1743.       for (j = 0; j < dst_channels; ++j) {
  1744.          assert(swizzle[j] < 4);
  1745.          if (i < num_fs) {
  1746.             fs_src[i][j] = LLVMBuildLoad(builder, fs_out_color[rt][swizzle[j]][i], "");
  1747.          }
  1748.          else {
  1749.             fs_src[i][j] = undef_src_val;
  1750.          }
  1751.       }
  1752.  
  1753.       /* If 3 channels then pad to include alpha for 4 element transpose */
  1754.       /*
  1755.        * XXX If we include that here maybe could actually use it instead of
  1756.        * separate alpha for blending?
  1757.        */
  1758.       if (dst_channels == 3 && !has_alpha) {
  1759.          fs_src[i][3] = alpha;
  1760.       }
  1761.  
  1762.       /* We split the row_mask and row_alpha as we want 128bit interleave */
  1763.       if (fs_type.length == 8) {
  1764.          src_mask[i*2 + 0]  = lp_build_extract_range(gallivm, fs_mask[i], 0, src_channels);
  1765.          src_mask[i*2 + 1]  = lp_build_extract_range(gallivm, fs_mask[i], src_channels, src_channels);
  1766.  
  1767.          src_alpha[i*2 + 0] = lp_build_extract_range(gallivm, alpha, 0, src_channels);
  1768.          src_alpha[i*2 + 1] = lp_build_extract_range(gallivm, alpha, src_channels, src_channels);
  1769.       } else {
  1770.          src_mask[i] = fs_mask[i];
  1771.          src_alpha[i] = alpha;
  1772.       }
  1773.    }
  1774.    if (dual_source_blend) {
  1775.       /* same as above except different src/dst, skip masks and comments... */
  1776.       for (i = 0; i < num_fullblock_fs; ++i) {
  1777.          LLVMValueRef alpha;
  1778.          if (i < num_fs) {
  1779.             alpha = LLVMBuildLoad(builder, fs_out_color[1][alpha_channel][i], "");
  1780.          }
  1781.          else {
  1782.             alpha = undef_src_val;
  1783.          }
  1784.  
  1785.          for (j = 0; j < dst_channels; ++j) {
  1786.             assert(swizzle[j] < 4);
  1787.             if (i < num_fs) {
  1788.                fs_src1[i][j] = LLVMBuildLoad(builder, fs_out_color[1][swizzle[j]][i], "");
  1789.             }
  1790.             else {
  1791.                fs_src1[i][j] = undef_src_val;
  1792.             }
  1793.          }
  1794.          if (dst_channels == 3 && !has_alpha) {
  1795.             fs_src1[i][3] = alpha;
  1796.          }
  1797.          if (fs_type.length == 8) {
  1798.             src1_alpha[i*2 + 0] = lp_build_extract_range(gallivm, alpha, 0, src_channels);
  1799.             src1_alpha[i*2 + 1] = lp_build_extract_range(gallivm, alpha, src_channels, src_channels);
  1800.          } else {
  1801.             src1_alpha[i] = alpha;
  1802.          }
  1803.       }
  1804.    }
  1805.  
  1806.    if (util_format_is_pure_integer(out_format)) {
  1807.       /*
  1808.        * In this case fs_type was really ints or uints disguised as floats,
  1809.        * fix that up now.
  1810.        */
  1811.       fs_type.floating = 0;
  1812.       fs_type.sign = dst_type.sign;
  1813.       for (i = 0; i < num_fullblock_fs; ++i) {
  1814.          for (j = 0; j < dst_channels; ++j) {
  1815.             fs_src[i][j] = LLVMBuildBitCast(builder, fs_src[i][j],
  1816.                                             lp_build_vec_type(gallivm, fs_type), "");
  1817.          }
  1818.          if (dst_channels == 3 && !has_alpha) {
  1819.             fs_src[i][3] = LLVMBuildBitCast(builder, fs_src[i][3],
  1820.                                             lp_build_vec_type(gallivm, fs_type), "");
  1821.          }
  1822.       }
  1823.    }
  1824.  
  1825.    /*
  1826.     * Pixel twiddle from fragment shader order to memory order
  1827.     */
  1828.    src_count = generate_fs_twiddle(gallivm, fs_type, num_fullblock_fs,
  1829.                                    dst_channels, fs_src, src, pad_inline);
  1830.    if (dual_source_blend) {
  1831.       generate_fs_twiddle(gallivm, fs_type, num_fullblock_fs, dst_channels,
  1832.                           fs_src1, src1, pad_inline);
  1833.    }
  1834.  
  1835.    src_channels = dst_channels < 3 ? dst_channels : 4;
  1836.    if (src_count != num_fullblock_fs * src_channels) {
  1837.       unsigned ds = src_count / (num_fullblock_fs * src_channels);
  1838.       row_type.length /= ds;
  1839.       fs_type.length = row_type.length;
  1840.    }
  1841.  
  1842.    blend_type = row_type;
  1843.    mask_type.length = 4;
  1844.  
  1845.    /* Convert src to row_type */
  1846.    if (dual_source_blend) {
  1847.       struct lp_type old_row_type = row_type;
  1848.       lp_build_conv_auto(gallivm, fs_type, &row_type, src, src_count, src);
  1849.       src_count = lp_build_conv_auto(gallivm, fs_type, &old_row_type, src1, src_count, src1);
  1850.    }
  1851.    else {
  1852.       src_count = lp_build_conv_auto(gallivm, fs_type, &row_type, src, src_count, src);
  1853.    }
  1854.  
  1855.    /* If the rows are not an SSE vector, combine them to become SSE size! */
  1856.    if ((row_type.width * row_type.length) % 128) {
  1857.       unsigned bits = row_type.width * row_type.length;
  1858.       unsigned combined;
  1859.  
  1860.       assert(src_count >= (vector_width / bits));
  1861.  
  1862.       dst_count = src_count / (vector_width / bits);
  1863.  
  1864.       combined = lp_build_concat_n(gallivm, row_type, src, src_count, src, dst_count);
  1865.       if (dual_source_blend) {
  1866.          lp_build_concat_n(gallivm, row_type, src1, src_count, src1, dst_count);
  1867.       }
  1868.  
  1869.       row_type.length *= combined;
  1870.       src_count /= combined;
  1871.  
  1872.       bits = row_type.width * row_type.length;
  1873.       assert(bits == 128 || bits == 256);
  1874.    }
  1875.  
  1876.  
  1877.    /*
  1878.     * Blend Colour conversion
  1879.     */
  1880.    blend_color = lp_jit_context_f_blend_color(gallivm, context_ptr);
  1881.    blend_color = LLVMBuildPointerCast(builder, blend_color, LLVMPointerType(lp_build_vec_type(gallivm, fs_type), 0), "");
  1882.    blend_color = LLVMBuildLoad(builder, LLVMBuildGEP(builder, blend_color, &i32_zero, 1, ""), "");
  1883.  
  1884.    /* Convert */
  1885.    lp_build_conv(gallivm, fs_type, blend_type, &blend_color, 1, &blend_color, 1);
  1886.  
  1887.    if (out_format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB) {
  1888.       /*
  1889.        * since blending is done with floats, there was no conversion.
  1890.        * However, the rules according to fixed point renderbuffers still
  1891.        * apply, that is we must clamp inputs to 0.0/1.0.
  1892.        * (This would apply to separate alpha conversion too but we currently
  1893.        * force has_alpha to be true.)
  1894.        * TODO: should skip this with "fake" blend, since post-blend conversion
  1895.        * will clamp anyway.
  1896.        * TODO: could also skip this if fragment color clamping is enabled. We
  1897.        * don't support it natively so it gets baked into the shader however, so
  1898.        * can't really tell here.
  1899.        */
  1900.       struct lp_build_context f32_bld;
  1901.       assert(row_type.floating);
  1902.       lp_build_context_init(&f32_bld, gallivm, row_type);
  1903.       for (i = 0; i < src_count; i++) {
  1904.          src[i] = lp_build_clamp_zero_one_nanzero(&f32_bld, src[i]);
  1905.       }
  1906.       if (dual_source_blend) {
  1907.          for (i = 0; i < src_count; i++) {
  1908.             src1[i] = lp_build_clamp_zero_one_nanzero(&f32_bld, src1[i]);
  1909.          }
  1910.       }
  1911.       /* probably can't be different than row_type but better safe than sorry... */
  1912.       lp_build_context_init(&f32_bld, gallivm, blend_type);
  1913.       blend_color = lp_build_clamp(&f32_bld, blend_color, f32_bld.zero, f32_bld.one);
  1914.    }
  1915.  
  1916.    /* Extract alpha */
  1917.    blend_alpha = lp_build_extract_broadcast(gallivm, blend_type, row_type, blend_color, lp_build_const_int32(gallivm, 3));
  1918.  
  1919.    /* Swizzle to appropriate channels, e.g. from RGBA to BGRA BGRA */
  1920.    pad_inline &= (dst_channels * (block_size / src_count) * row_type.width) != vector_width;
  1921.    if (pad_inline) {
  1922.       /* Use all 4 channels e.g. from RGBA RGBA to RGxx RGxx */
  1923.       blend_color = lp_build_swizzle_aos_n(gallivm, blend_color, swizzle, TGSI_NUM_CHANNELS, row_type.length);
  1924.    } else {
  1925.       /* Only use dst_channels e.g. RGBA RGBA to RG RG xxxx */
  1926.       blend_color = lp_build_swizzle_aos_n(gallivm, blend_color, swizzle, dst_channels, row_type.length);
  1927.    }
  1928.  
  1929.    /*
  1930.     * Mask conversion
  1931.     */
  1932.    lp_bld_quad_twiddle(gallivm, mask_type, &src_mask[0], block_height, &src_mask[0]);
  1933.  
  1934.    if (src_count < block_height) {
  1935.       lp_build_concat_n(gallivm, mask_type, src_mask, 4, src_mask, src_count);
  1936.    } else if (src_count > block_height) {
  1937.       for (i = src_count; i > 0; --i) {
  1938.          unsigned pixels = block_size / src_count;
  1939.          unsigned idx = i - 1;
  1940.  
  1941.          src_mask[idx] = lp_build_extract_range(gallivm, src_mask[(idx * pixels) / 4],
  1942.                                                 (idx * pixels) % 4, pixels);
  1943.       }
  1944.    }
  1945.  
  1946.    assert(mask_type.width == 32);
  1947.  
  1948.    for (i = 0; i < src_count; ++i) {
  1949.       unsigned pixels = block_size / src_count;
  1950.       unsigned pixel_width = row_type.width * dst_channels;
  1951.  
  1952.       if (pixel_width == 24) {
  1953.          mask_type.width = 8;
  1954.          mask_type.length = vector_width / mask_type.width;
  1955.       } else {
  1956.          mask_type.length = pixels;
  1957.          mask_type.width = row_type.width * dst_channels;
  1958.  
  1959.          src_mask[i] = LLVMBuildIntCast(builder, src_mask[i], lp_build_int_vec_type(gallivm, mask_type), "");
  1960.  
  1961.          mask_type.length *= dst_channels;
  1962.          mask_type.width /= dst_channels;
  1963.       }
  1964.  
  1965.       src_mask[i] = LLVMBuildBitCast(builder, src_mask[i], lp_build_int_vec_type(gallivm, mask_type), "");
  1966.       src_mask[i] = lp_build_pad_vector(gallivm, src_mask[i], row_type.length);
  1967.    }
  1968.  
  1969.    /*
  1970.     * Alpha conversion
  1971.     */
  1972.    if (!has_alpha) {
  1973.       struct lp_type alpha_type = fs_type;
  1974.       alpha_type.length = 4;
  1975.       convert_alpha(gallivm, row_type, alpha_type,
  1976.                     block_size, block_height,
  1977.                     src_count, dst_channels,
  1978.                     pad_inline, src_alpha);
  1979.       if (dual_source_blend) {
  1980.          convert_alpha(gallivm, row_type, alpha_type,
  1981.                        block_size, block_height,
  1982.                        src_count, dst_channels,
  1983.                        pad_inline, src1_alpha);
  1984.       }
  1985.    }
  1986.  
  1987.  
  1988.    /*
  1989.     * Load dst from memory
  1990.     */
  1991.    if (src_count < block_height) {
  1992.       dst_count = block_height;
  1993.    } else {
  1994.       dst_count = src_count;
  1995.    }
  1996.  
  1997.    dst_type.length *= block_size / dst_count;
  1998.  
  1999.    if (format_expands_to_float_soa(out_format_desc)) {
  2000.       /*
  2001.        * we need multiple values at once for the conversion, so can as well
  2002.        * load them vectorized here too instead of concatenating later.
  2003.        * (Still need concatenation later for 8-wide vectors).
  2004.        */
  2005.       dst_count = block_height;
  2006.       dst_type.length = block_width;
  2007.    }
  2008.  
  2009.    /*
  2010.     * Compute the alignment of the destination pointer in bytes
  2011.     * We fetch 1-4 pixels, if the format has pot alignment then those fetches
  2012.     * are always aligned by MIN2(16, fetch_width) except for buffers (not
  2013.     * 1d tex but can't distinguish here) so need to stick with per-pixel
  2014.     * alignment in this case.
  2015.     */
  2016.    if (is_1d) {
  2017.       dst_alignment = (out_format_desc->block.bits + 7)/(out_format_desc->block.width * 8);
  2018.    }
  2019.    else {
  2020.       dst_alignment = dst_type.length * dst_type.width / 8;
  2021.    }
  2022.    /* Force power-of-two alignment by extracting only the least-significant-bit */
  2023.    dst_alignment = 1 << (ffs(dst_alignment) - 1);
  2024.    /*
  2025.     * Resource base and stride pointers are aligned to 16 bytes, so that's
  2026.     * the maximum alignment we can guarantee
  2027.     */
  2028.    dst_alignment = MIN2(16, dst_alignment);
  2029.  
  2030.    if (is_1d) {
  2031.       load_unswizzled_block(gallivm, color_ptr, stride, block_width, 1,
  2032.                             dst, dst_type, dst_count / 4, dst_alignment);
  2033.       for (i = dst_count / 4; i < dst_count; i++) {
  2034.          dst[i] = lp_build_undef(gallivm, dst_type);
  2035.       }
  2036.  
  2037.    }
  2038.    else {
  2039.       load_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
  2040.                             dst, dst_type, dst_count, dst_alignment);
  2041.    }
  2042.  
  2043.  
  2044.    /*
  2045.     * Convert from dst/output format to src/blending format.
  2046.     *
  2047.     * This is necessary as we can only read 1 row from memory at a time,
  2048.     * so the minimum dst_count will ever be at this point is 4.
  2049.     *
  2050.     * With, for example, R8 format you can have all 16 pixels in a 128 bit vector,
  2051.     * this will take the 4 dsts and combine them into 1 src so we can perform blending
  2052.     * on all 16 pixels in that single vector at once.
  2053.     */
  2054.    if (dst_count > src_count) {
  2055.       lp_build_concat_n(gallivm, dst_type, dst, 4, dst, src_count);
  2056.    }
  2057.  
  2058.    /*
  2059.     * Blending
  2060.     */
  2061.    /* XXX this is broken for RGB8 formats -
  2062.     * they get expanded from 12 to 16 elements (to include alpha)
  2063.     * by convert_to_blend_type then reduced to 15 instead of 12
  2064.     * by convert_from_blend_type (a simple fix though breaks A8...).
  2065.     * R16G16B16 also crashes differently however something going wrong
  2066.     * inside llvm handling npot vector sizes seemingly.
  2067.     * It seems some cleanup could be done here (like skipping conversion/blend
  2068.     * when not needed).
  2069.     */
  2070.    convert_to_blend_type(gallivm, block_size, out_format_desc, dst_type, row_type, dst, src_count);
  2071.  
  2072.    /*
  2073.     * FIXME: Really should get logic ops / masks out of generic blend / row
  2074.     * format. Logic ops will definitely not work on the blend float format
  2075.     * used for SRGB here and I think OpenGL expects this to work as expected
  2076.     * (that is incoming values converted to srgb then logic op applied).
  2077.     */
  2078.    for (i = 0; i < src_count; ++i) {
  2079.       dst[i] = lp_build_blend_aos(gallivm,
  2080.                                   &variant->key.blend,
  2081.                                   out_format,
  2082.                                   row_type,
  2083.                                   rt,
  2084.                                   src[i],
  2085.                                   has_alpha ? NULL : src_alpha[i],
  2086.                                   src1[i],
  2087.                                   has_alpha ? NULL : src1_alpha[i],
  2088.                                   dst[i],
  2089.                                   partial_mask ? src_mask[i] : NULL,
  2090.                                   blend_color,
  2091.                                   has_alpha ? NULL : blend_alpha,
  2092.                                   swizzle,
  2093.                                   pad_inline ? 4 : dst_channels);
  2094.    }
  2095.  
  2096.    convert_from_blend_type(gallivm, block_size, out_format_desc, row_type, dst_type, dst, src_count);
  2097.  
  2098.    /* Split the blend rows back to memory rows */
  2099.    if (dst_count > src_count) {
  2100.       row_type.length = dst_type.length * (dst_count / src_count);
  2101.  
  2102.       if (src_count == 1) {
  2103.          dst[1] = lp_build_extract_range(gallivm, dst[0], row_type.length / 2, row_type.length / 2);
  2104.          dst[0] = lp_build_extract_range(gallivm, dst[0], 0, row_type.length / 2);
  2105.  
  2106.          row_type.length /= 2;
  2107.          src_count *= 2;
  2108.       }
  2109.  
  2110.       dst[3] = lp_build_extract_range(gallivm, dst[1], row_type.length / 2, row_type.length / 2);
  2111.       dst[2] = lp_build_extract_range(gallivm, dst[1], 0, row_type.length / 2);
  2112.       dst[1] = lp_build_extract_range(gallivm, dst[0], row_type.length / 2, row_type.length / 2);
  2113.       dst[0] = lp_build_extract_range(gallivm, dst[0], 0, row_type.length / 2);
  2114.  
  2115.       row_type.length /= 2;
  2116.       src_count *= 2;
  2117.    }
  2118.  
  2119.    /*
  2120.     * Store blend result to memory
  2121.     */
  2122.    if (is_1d) {
  2123.       store_unswizzled_block(gallivm, color_ptr, stride, block_width, 1,
  2124.                              dst, dst_type, dst_count / 4, dst_alignment);
  2125.    }
  2126.    else {
  2127.       store_unswizzled_block(gallivm, color_ptr, stride, block_width, block_height,
  2128.                              dst, dst_type, dst_count, dst_alignment);
  2129.    }
  2130.  
  2131.    if (have_smallfloat_format(dst_type, out_format)) {
  2132.       lp_build_fpstate_set(gallivm, fpstate);
  2133.    }
  2134.  
  2135.    if (do_branch) {
  2136.       lp_build_mask_end(&mask_ctx);
  2137.    }
  2138. }
  2139.  
  2140.  
  2141. /**
  2142.  * Generate the runtime callable function for the whole fragment pipeline.
  2143.  * Note that the function which we generate operates on a block of 16
  2144.  * pixels at at time.  The block contains 2x2 quads.  Each quad contains
  2145.  * 2x2 pixels.
  2146.  */
  2147. static void
  2148. generate_fragment(struct llvmpipe_context *lp,
  2149.                   struct lp_fragment_shader *shader,
  2150.                   struct lp_fragment_shader_variant *variant,
  2151.                   unsigned partial_mask)
  2152. {
  2153.    struct gallivm_state *gallivm = variant->gallivm;
  2154.    const struct lp_fragment_shader_variant_key *key = &variant->key;
  2155.    struct lp_shader_input inputs[PIPE_MAX_SHADER_INPUTS];
  2156.    char func_name[64];
  2157.    struct lp_type fs_type;
  2158.    struct lp_type blend_type;
  2159.    LLVMTypeRef fs_elem_type;
  2160.    LLVMTypeRef blend_vec_type;
  2161.    LLVMTypeRef arg_types[13];
  2162.    LLVMTypeRef func_type;
  2163.    LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
  2164.    LLVMTypeRef int8_type = LLVMInt8TypeInContext(gallivm->context);
  2165.    LLVMValueRef context_ptr;
  2166.    LLVMValueRef x;
  2167.    LLVMValueRef y;
  2168.    LLVMValueRef a0_ptr;
  2169.    LLVMValueRef dadx_ptr;
  2170.    LLVMValueRef dady_ptr;
  2171.    LLVMValueRef color_ptr_ptr;
  2172.    LLVMValueRef stride_ptr;
  2173.    LLVMValueRef depth_ptr;
  2174.    LLVMValueRef depth_stride;
  2175.    LLVMValueRef mask_input;
  2176.    LLVMValueRef thread_data_ptr;
  2177.    LLVMBasicBlockRef block;
  2178.    LLVMBuilderRef builder;
  2179.    struct lp_build_sampler_soa *sampler;
  2180.    struct lp_build_interp_soa_context interp;
  2181.    LLVMValueRef fs_mask[16 / 4];
  2182.    LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS][16 / 4];
  2183.    LLVMValueRef function;
  2184.    LLVMValueRef facing;
  2185.    unsigned num_fs;
  2186.    unsigned i;
  2187.    unsigned chan;
  2188.    unsigned cbuf;
  2189.    boolean cbuf0_write_all;
  2190.    const boolean dual_source_blend = key->blend.rt[0].blend_enable &&
  2191.                                      util_blend_state_is_dual(&key->blend, 0);
  2192.  
  2193.    assert(lp_native_vector_width / 32 >= 4);
  2194.  
  2195.    /* Adjust color input interpolation according to flatshade state:
  2196.     */
  2197.    memcpy(inputs, shader->inputs, shader->info.base.num_inputs * sizeof inputs[0]);
  2198.    for (i = 0; i < shader->info.base.num_inputs; i++) {
  2199.       if (inputs[i].interp == LP_INTERP_COLOR) {
  2200.          if (key->flatshade)
  2201.             inputs[i].interp = LP_INTERP_CONSTANT;
  2202.          else
  2203.             inputs[i].interp = LP_INTERP_PERSPECTIVE;
  2204.       }
  2205.    }
  2206.  
  2207.    /* check if writes to cbuf[0] are to be copied to all cbufs */
  2208.    cbuf0_write_all =
  2209.      shader->info.base.properties[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS];
  2210.  
  2211.    /* TODO: actually pick these based on the fs and color buffer
  2212.     * characteristics. */
  2213.  
  2214.    memset(&fs_type, 0, sizeof fs_type);
  2215.    fs_type.floating = TRUE;      /* floating point values */
  2216.    fs_type.sign = TRUE;          /* values are signed */
  2217.    fs_type.norm = FALSE;         /* values are not limited to [0,1] or [-1,1] */
  2218.    fs_type.width = 32;           /* 32-bit float */
  2219.    fs_type.length = MIN2(lp_native_vector_width / 32, 16); /* n*4 elements per vector */
  2220.  
  2221.    memset(&blend_type, 0, sizeof blend_type);
  2222.    blend_type.floating = FALSE; /* values are integers */
  2223.    blend_type.sign = FALSE;     /* values are unsigned */
  2224.    blend_type.norm = TRUE;      /* values are in [0,1] or [-1,1] */
  2225.    blend_type.width = 8;        /* 8-bit ubyte values */
  2226.    blend_type.length = 16;      /* 16 elements per vector */
  2227.  
  2228.    /*
  2229.     * Generate the function prototype. Any change here must be reflected in
  2230.     * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
  2231.     */
  2232.  
  2233.    fs_elem_type = lp_build_elem_type(gallivm, fs_type);
  2234.  
  2235.    blend_vec_type = lp_build_vec_type(gallivm, blend_type);
  2236.  
  2237.    util_snprintf(func_name, sizeof(func_name), "fs%u_variant%u_%s",
  2238.                  shader->no, variant->no, partial_mask ? "partial" : "whole");
  2239.  
  2240.    arg_types[0] = variant->jit_context_ptr_type;       /* context */
  2241.    arg_types[1] = int32_type;                          /* x */
  2242.    arg_types[2] = int32_type;                          /* y */
  2243.    arg_types[3] = int32_type;                          /* facing */
  2244.    arg_types[4] = LLVMPointerType(fs_elem_type, 0);    /* a0 */
  2245.    arg_types[5] = LLVMPointerType(fs_elem_type, 0);    /* dadx */
  2246.    arg_types[6] = LLVMPointerType(fs_elem_type, 0);    /* dady */
  2247.    arg_types[7] = LLVMPointerType(LLVMPointerType(blend_vec_type, 0), 0);  /* color */
  2248.    arg_types[8] = LLVMPointerType(int8_type, 0);       /* depth */
  2249.    arg_types[9] = int32_type;                          /* mask_input */
  2250.    arg_types[10] = variant->jit_thread_data_ptr_type;  /* per thread data */
  2251.    arg_types[11] = LLVMPointerType(int32_type, 0);     /* stride */
  2252.    arg_types[12] = int32_type;                         /* depth_stride */
  2253.  
  2254.    func_type = LLVMFunctionType(LLVMVoidTypeInContext(gallivm->context),
  2255.                                 arg_types, Elements(arg_types), 0);
  2256.  
  2257.    function = LLVMAddFunction(gallivm->module, func_name, func_type);
  2258.    LLVMSetFunctionCallConv(function, LLVMCCallConv);
  2259.  
  2260.    variant->function[partial_mask] = function;
  2261.  
  2262.    /* XXX: need to propagate noalias down into color param now we are
  2263.     * passing a pointer-to-pointer?
  2264.     */
  2265.    for(i = 0; i < Elements(arg_types); ++i)
  2266.       if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
  2267.          LLVMAddAttribute(LLVMGetParam(function, i), LLVMNoAliasAttribute);
  2268.  
  2269.    context_ptr  = LLVMGetParam(function, 0);
  2270.    x            = LLVMGetParam(function, 1);
  2271.    y            = LLVMGetParam(function, 2);
  2272.    facing       = LLVMGetParam(function, 3);
  2273.    a0_ptr       = LLVMGetParam(function, 4);
  2274.    dadx_ptr     = LLVMGetParam(function, 5);
  2275.    dady_ptr     = LLVMGetParam(function, 6);
  2276.    color_ptr_ptr = LLVMGetParam(function, 7);
  2277.    depth_ptr    = LLVMGetParam(function, 8);
  2278.    mask_input   = LLVMGetParam(function, 9);
  2279.    thread_data_ptr  = LLVMGetParam(function, 10);
  2280.    stride_ptr   = LLVMGetParam(function, 11);
  2281.    depth_stride = LLVMGetParam(function, 12);
  2282.  
  2283.    lp_build_name(context_ptr, "context");
  2284.    lp_build_name(x, "x");
  2285.    lp_build_name(y, "y");
  2286.    lp_build_name(a0_ptr, "a0");
  2287.    lp_build_name(dadx_ptr, "dadx");
  2288.    lp_build_name(dady_ptr, "dady");
  2289.    lp_build_name(color_ptr_ptr, "color_ptr_ptr");
  2290.    lp_build_name(depth_ptr, "depth");
  2291.    lp_build_name(thread_data_ptr, "thread_data");
  2292.    lp_build_name(mask_input, "mask_input");
  2293.    lp_build_name(stride_ptr, "stride_ptr");
  2294.    lp_build_name(depth_stride, "depth_stride");
  2295.  
  2296.    /*
  2297.     * Function body
  2298.     */
  2299.  
  2300.    block = LLVMAppendBasicBlockInContext(gallivm->context, function, "entry");
  2301.    builder = gallivm->builder;
  2302.    assert(builder);
  2303.    LLVMPositionBuilderAtEnd(builder, block);
  2304.  
  2305.    /* code generated texture sampling */
  2306.    sampler = lp_llvm_sampler_soa_create(key->state);
  2307.  
  2308.    num_fs = 16 / fs_type.length; /* number of loops per 4x4 stamp */
  2309.    /* for 1d resources only run "upper half" of stamp */
  2310.    if (key->resource_1d)
  2311.       num_fs /= 2;
  2312.  
  2313.    {
  2314.       LLVMValueRef num_loop = lp_build_const_int32(gallivm, num_fs);
  2315.       LLVMTypeRef mask_type = lp_build_int_vec_type(gallivm, fs_type);
  2316.       LLVMValueRef mask_store = lp_build_array_alloca(gallivm, mask_type,
  2317.                                                       num_loop, "mask_store");
  2318.       LLVMValueRef color_store[PIPE_MAX_COLOR_BUFS][TGSI_NUM_CHANNELS];
  2319.       boolean pixel_center_integer =
  2320.          shader->info.base.properties[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER];
  2321.  
  2322.       /*
  2323.        * The shader input interpolation info is not explicitely baked in the
  2324.        * shader key, but everything it derives from (TGSI, and flatshade) is
  2325.        * already included in the shader key.
  2326.        */
  2327.       lp_build_interp_soa_init(&interp,
  2328.                                gallivm,
  2329.                                shader->info.base.num_inputs,
  2330.                                inputs,
  2331.                                pixel_center_integer,
  2332.                                builder, fs_type,
  2333.                                a0_ptr, dadx_ptr, dady_ptr,
  2334.                                x, y);
  2335.  
  2336.       for (i = 0; i < num_fs; i++) {
  2337.          LLVMValueRef mask;
  2338.          LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
  2339.          LLVMValueRef mask_ptr = LLVMBuildGEP(builder, mask_store,
  2340.                                               &indexi, 1, "mask_ptr");
  2341.  
  2342.          if (partial_mask) {
  2343.             mask = generate_quad_mask(gallivm, fs_type,
  2344.                                       i*fs_type.length/4, mask_input);
  2345.          }
  2346.          else {
  2347.             mask = lp_build_const_int_vec(gallivm, fs_type, ~0);
  2348.          }
  2349.          LLVMBuildStore(builder, mask, mask_ptr);
  2350.       }
  2351.  
  2352.       generate_fs_loop(gallivm,
  2353.                        shader, key,
  2354.                        builder,
  2355.                        fs_type,
  2356.                        context_ptr,
  2357.                        num_loop,
  2358.                        &interp,
  2359.                        sampler,
  2360.                        mask_store, /* output */
  2361.                        color_store,
  2362.                        depth_ptr,
  2363.                        depth_stride,
  2364.                        facing,
  2365.                        thread_data_ptr);
  2366.  
  2367.       for (i = 0; i < num_fs; i++) {
  2368.          LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
  2369.          LLVMValueRef ptr = LLVMBuildGEP(builder, mask_store,
  2370.                                          &indexi, 1, "");
  2371.          fs_mask[i] = LLVMBuildLoad(builder, ptr, "mask");
  2372.          /* This is fucked up need to reorganize things */
  2373.          for (cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
  2374.             for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
  2375.                ptr = LLVMBuildGEP(builder,
  2376.                                   color_store[cbuf * !cbuf0_write_all][chan],
  2377.                                   &indexi, 1, "");
  2378.                fs_out_color[cbuf][chan][i] = ptr;
  2379.             }
  2380.          }
  2381.          if (dual_source_blend) {
  2382.             /* only support one dual source blend target hence always use output 1 */
  2383.             for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
  2384.                ptr = LLVMBuildGEP(builder,
  2385.                                   color_store[1][chan],
  2386.                                   &indexi, 1, "");
  2387.                fs_out_color[1][chan][i] = ptr;
  2388.             }
  2389.          }
  2390.       }
  2391.    }
  2392.  
  2393.    sampler->destroy(sampler);
  2394.  
  2395.    /* Loop over color outputs / color buffers to do blending.
  2396.     */
  2397.    for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
  2398.       if (key->cbuf_format[cbuf] != PIPE_FORMAT_NONE) {
  2399.          LLVMValueRef color_ptr;
  2400.          LLVMValueRef stride;
  2401.          LLVMValueRef index = lp_build_const_int32(gallivm, cbuf);
  2402.  
  2403.          boolean do_branch = ((key->depth.enabled
  2404.                                || key->stencil[0].enabled
  2405.                                || key->alpha.enabled)
  2406.                               && !shader->info.base.uses_kill);
  2407.  
  2408.          color_ptr = LLVMBuildLoad(builder,
  2409.                                    LLVMBuildGEP(builder, color_ptr_ptr,
  2410.                                                 &index, 1, ""),
  2411.                                    "");
  2412.  
  2413.          lp_build_name(color_ptr, "color_ptr%d", cbuf);
  2414.  
  2415.          stride = LLVMBuildLoad(builder,
  2416.                                 LLVMBuildGEP(builder, stride_ptr, &index, 1, ""),
  2417.                                 "");
  2418.  
  2419.          generate_unswizzled_blend(gallivm, cbuf, variant,
  2420.                                    key->cbuf_format[cbuf],
  2421.                                    num_fs, fs_type, fs_mask, fs_out_color,
  2422.                                    context_ptr, color_ptr, stride,
  2423.                                    partial_mask, do_branch);
  2424.       }
  2425.    }
  2426.  
  2427.    LLVMBuildRetVoid(builder);
  2428.  
  2429.    gallivm_verify_function(gallivm, function);
  2430. }
  2431.  
  2432.  
  2433. static void
  2434. dump_fs_variant_key(const struct lp_fragment_shader_variant_key *key)
  2435. {
  2436.    unsigned i;
  2437.  
  2438.    debug_printf("fs variant %p:\n", (void *) key);
  2439.  
  2440.    if (key->flatshade) {
  2441.       debug_printf("flatshade = 1\n");
  2442.    }
  2443.    for (i = 0; i < key->nr_cbufs; ++i) {
  2444.       debug_printf("cbuf_format[%u] = %s\n", i, util_format_name(key->cbuf_format[i]));
  2445.    }
  2446.    if (key->depth.enabled) {
  2447.       debug_printf("depth.format = %s\n", util_format_name(key->zsbuf_format));
  2448.       debug_printf("depth.func = %s\n", util_dump_func(key->depth.func, TRUE));
  2449.       debug_printf("depth.writemask = %u\n", key->depth.writemask);
  2450.    }
  2451.  
  2452.    for (i = 0; i < 2; ++i) {
  2453.       if (key->stencil[i].enabled) {
  2454.          debug_printf("stencil[%u].func = %s\n", i, util_dump_func(key->stencil[i].func, TRUE));
  2455.          debug_printf("stencil[%u].fail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].fail_op, TRUE));
  2456.          debug_printf("stencil[%u].zpass_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zpass_op, TRUE));
  2457.          debug_printf("stencil[%u].zfail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zfail_op, TRUE));
  2458.          debug_printf("stencil[%u].valuemask = 0x%x\n", i, key->stencil[i].valuemask);
  2459.          debug_printf("stencil[%u].writemask = 0x%x\n", i, key->stencil[i].writemask);
  2460.       }
  2461.    }
  2462.  
  2463.    if (key->alpha.enabled) {
  2464.       debug_printf("alpha.func = %s\n", util_dump_func(key->alpha.func, TRUE));
  2465.    }
  2466.  
  2467.    if (key->occlusion_count) {
  2468.       debug_printf("occlusion_count = 1\n");
  2469.    }
  2470.  
  2471.    if (key->blend.logicop_enable) {
  2472.       debug_printf("blend.logicop_func = %s\n", util_dump_logicop(key->blend.logicop_func, TRUE));
  2473.    }
  2474.    else if (key->blend.rt[0].blend_enable) {
  2475.       debug_printf("blend.rgb_func = %s\n",   util_dump_blend_func  (key->blend.rt[0].rgb_func, TRUE));
  2476.       debug_printf("blend.rgb_src_factor = %s\n",   util_dump_blend_factor(key->blend.rt[0].rgb_src_factor, TRUE));
  2477.       debug_printf("blend.rgb_dst_factor = %s\n",   util_dump_blend_factor(key->blend.rt[0].rgb_dst_factor, TRUE));
  2478.       debug_printf("blend.alpha_func = %s\n",       util_dump_blend_func  (key->blend.rt[0].alpha_func, TRUE));
  2479.       debug_printf("blend.alpha_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_src_factor, TRUE));
  2480.       debug_printf("blend.alpha_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_dst_factor, TRUE));
  2481.    }
  2482.    debug_printf("blend.colormask = 0x%x\n", key->blend.rt[0].colormask);
  2483.    if (key->blend.alpha_to_coverage) {
  2484.       debug_printf("blend.alpha_to_coverage is enabled\n");
  2485.    }
  2486.    for (i = 0; i < key->nr_samplers; ++i) {
  2487.       const struct lp_static_sampler_state *sampler = &key->state[i].sampler_state;
  2488.       debug_printf("sampler[%u] = \n", i);
  2489.       debug_printf("  .wrap = %s %s %s\n",
  2490.                    util_dump_tex_wrap(sampler->wrap_s, TRUE),
  2491.                    util_dump_tex_wrap(sampler->wrap_t, TRUE),
  2492.                    util_dump_tex_wrap(sampler->wrap_r, TRUE));
  2493.       debug_printf("  .min_img_filter = %s\n",
  2494.                    util_dump_tex_filter(sampler->min_img_filter, TRUE));
  2495.       debug_printf("  .min_mip_filter = %s\n",
  2496.                    util_dump_tex_mipfilter(sampler->min_mip_filter, TRUE));
  2497.       debug_printf("  .mag_img_filter = %s\n",
  2498.                    util_dump_tex_filter(sampler->mag_img_filter, TRUE));
  2499.       if (sampler->compare_mode != PIPE_TEX_COMPARE_NONE)
  2500.          debug_printf("  .compare_func = %s\n", util_dump_func(sampler->compare_func, TRUE));
  2501.       debug_printf("  .normalized_coords = %u\n", sampler->normalized_coords);
  2502.       debug_printf("  .min_max_lod_equal = %u\n", sampler->min_max_lod_equal);
  2503.       debug_printf("  .lod_bias_non_zero = %u\n", sampler->lod_bias_non_zero);
  2504.       debug_printf("  .apply_min_lod = %u\n", sampler->apply_min_lod);
  2505.       debug_printf("  .apply_max_lod = %u\n", sampler->apply_max_lod);
  2506.    }
  2507.    for (i = 0; i < key->nr_sampler_views; ++i) {
  2508.       const struct lp_static_texture_state *texture = &key->state[i].texture_state;
  2509.       debug_printf("texture[%u] = \n", i);
  2510.       debug_printf("  .format = %s\n",
  2511.                    util_format_name(texture->format));
  2512.       debug_printf("  .target = %s\n",
  2513.                    util_dump_tex_target(texture->target, TRUE));
  2514.       debug_printf("  .level_zero_only = %u\n",
  2515.                    texture->level_zero_only);
  2516.       debug_printf("  .pot = %u %u %u\n",
  2517.                    texture->pot_width,
  2518.                    texture->pot_height,
  2519.                    texture->pot_depth);
  2520.    }
  2521. }
  2522.  
  2523.  
  2524. void
  2525. lp_debug_fs_variant(const struct lp_fragment_shader_variant *variant)
  2526. {
  2527.    debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
  2528.                 variant->shader->no, variant->no);
  2529.    tgsi_dump(variant->shader->base.tokens, 0);
  2530.    dump_fs_variant_key(&variant->key);
  2531.    debug_printf("variant->opaque = %u\n", variant->opaque);
  2532.    debug_printf("\n");
  2533. }
  2534.  
  2535.  
  2536. /**
  2537.  * Generate a new fragment shader variant from the shader code and
  2538.  * other state indicated by the key.
  2539.  */
  2540. static struct lp_fragment_shader_variant *
  2541. generate_variant(struct llvmpipe_context *lp,
  2542.                  struct lp_fragment_shader *shader,
  2543.                  const struct lp_fragment_shader_variant_key *key)
  2544. {
  2545.    struct lp_fragment_shader_variant *variant;
  2546.    const struct util_format_description *cbuf0_format_desc;
  2547.    boolean fullcolormask;
  2548.    char module_name[64];
  2549.  
  2550.    variant = CALLOC_STRUCT(lp_fragment_shader_variant);
  2551.    if(!variant)
  2552.       return NULL;
  2553.  
  2554.    util_snprintf(module_name, sizeof(module_name), "fs%u_variant%u",
  2555.                  shader->no, shader->variants_created);
  2556.  
  2557.    variant->gallivm = gallivm_create(module_name, lp->context);
  2558.    if (!variant->gallivm) {
  2559.       FREE(variant);
  2560.       return NULL;
  2561.    }
  2562.  
  2563.    variant->shader = shader;
  2564.    variant->list_item_global.base = variant;
  2565.    variant->list_item_local.base = variant;
  2566.    variant->no = shader->variants_created++;
  2567.  
  2568.    memcpy(&variant->key, key, shader->variant_key_size);
  2569.  
  2570.    /*
  2571.     * Determine whether we are touching all channels in the color buffer.
  2572.     */
  2573.    fullcolormask = FALSE;
  2574.    if (key->nr_cbufs == 1) {
  2575.       cbuf0_format_desc = util_format_description(key->cbuf_format[0]);
  2576.       fullcolormask = util_format_colormask_full(cbuf0_format_desc, key->blend.rt[0].colormask);
  2577.    }
  2578.  
  2579.    variant->opaque =
  2580.          !key->blend.logicop_enable &&
  2581.          !key->blend.rt[0].blend_enable &&
  2582.          fullcolormask &&
  2583.          !key->stencil[0].enabled &&
  2584.          !key->alpha.enabled &&
  2585.          !key->blend.alpha_to_coverage &&
  2586.          !key->depth.enabled &&
  2587.          !shader->info.base.uses_kill
  2588.       ? TRUE : FALSE;
  2589.  
  2590.    if ((shader->info.base.num_tokens <= 1) &&
  2591.        !key->depth.enabled && !key->stencil[0].enabled) {
  2592.       variant->ps_inv_multiplier = 0;
  2593.    } else {
  2594.       variant->ps_inv_multiplier = 1;
  2595.    }
  2596.  
  2597.    if ((LP_DEBUG & DEBUG_FS) || (gallivm_debug & GALLIVM_DEBUG_IR)) {
  2598.       lp_debug_fs_variant(variant);
  2599.    }
  2600.  
  2601.    lp_jit_init_types(variant);
  2602.    
  2603.    if (variant->jit_function[RAST_EDGE_TEST] == NULL)
  2604.       generate_fragment(lp, shader, variant, RAST_EDGE_TEST);
  2605.  
  2606.    if (variant->jit_function[RAST_WHOLE] == NULL) {
  2607.       if (variant->opaque) {
  2608.          /* Specialized shader, which doesn't need to read the color buffer. */
  2609.          generate_fragment(lp, shader, variant, RAST_WHOLE);
  2610.       }
  2611.    }
  2612.  
  2613.    /*
  2614.     * Compile everything
  2615.     */
  2616.  
  2617.    gallivm_compile_module(variant->gallivm);
  2618.  
  2619.    variant->nr_instrs += lp_build_count_ir_module(variant->gallivm->module);
  2620.  
  2621.    if (variant->function[RAST_EDGE_TEST]) {
  2622.       variant->jit_function[RAST_EDGE_TEST] = (lp_jit_frag_func)
  2623.             gallivm_jit_function(variant->gallivm,
  2624.                                  variant->function[RAST_EDGE_TEST]);
  2625.    }
  2626.  
  2627.    if (variant->function[RAST_WHOLE]) {
  2628.          variant->jit_function[RAST_WHOLE] = (lp_jit_frag_func)
  2629.                gallivm_jit_function(variant->gallivm,
  2630.                                     variant->function[RAST_WHOLE]);
  2631.    } else if (!variant->jit_function[RAST_WHOLE]) {
  2632.       variant->jit_function[RAST_WHOLE] = variant->jit_function[RAST_EDGE_TEST];
  2633.    }
  2634.  
  2635.    gallivm_free_ir(variant->gallivm);
  2636.  
  2637.    return variant;
  2638. }
  2639.  
  2640.  
  2641. static void *
  2642. llvmpipe_create_fs_state(struct pipe_context *pipe,
  2643.                          const struct pipe_shader_state *templ)
  2644. {
  2645.    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
  2646.    struct lp_fragment_shader *shader;
  2647.    int nr_samplers;
  2648.    int nr_sampler_views;
  2649.    int i;
  2650.  
  2651.    shader = CALLOC_STRUCT(lp_fragment_shader);
  2652.    if (!shader)
  2653.       return NULL;
  2654.  
  2655.    shader->no = fs_no++;
  2656.    make_empty_list(&shader->variants);
  2657.  
  2658.    /* get/save the summary info for this shader */
  2659.    lp_build_tgsi_info(templ->tokens, &shader->info);
  2660.  
  2661.    /* we need to keep a local copy of the tokens */
  2662.    shader->base.tokens = tgsi_dup_tokens(templ->tokens);
  2663.  
  2664.    shader->draw_data = draw_create_fragment_shader(llvmpipe->draw, templ);
  2665.    if (shader->draw_data == NULL) {
  2666.       FREE((void *) shader->base.tokens);
  2667.       FREE(shader);
  2668.       return NULL;
  2669.    }
  2670.  
  2671.    nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
  2672.    nr_sampler_views = shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
  2673.  
  2674.    shader->variant_key_size = Offset(struct lp_fragment_shader_variant_key,
  2675.                                      state[MAX2(nr_samplers, nr_sampler_views)]);
  2676.  
  2677.    for (i = 0; i < shader->info.base.num_inputs; i++) {
  2678.       shader->inputs[i].usage_mask = shader->info.base.input_usage_mask[i];
  2679.       shader->inputs[i].cyl_wrap = shader->info.base.input_cylindrical_wrap[i];
  2680.  
  2681.       switch (shader->info.base.input_interpolate[i]) {
  2682.       case TGSI_INTERPOLATE_CONSTANT:
  2683.          shader->inputs[i].interp = LP_INTERP_CONSTANT;
  2684.          break;
  2685.       case TGSI_INTERPOLATE_LINEAR:
  2686.          shader->inputs[i].interp = LP_INTERP_LINEAR;
  2687.          break;
  2688.       case TGSI_INTERPOLATE_PERSPECTIVE:
  2689.          shader->inputs[i].interp = LP_INTERP_PERSPECTIVE;
  2690.          break;
  2691.       case TGSI_INTERPOLATE_COLOR:
  2692.          shader->inputs[i].interp = LP_INTERP_COLOR;
  2693.          break;
  2694.       default:
  2695.          assert(0);
  2696.          break;
  2697.       }
  2698.  
  2699.       switch (shader->info.base.input_semantic_name[i]) {
  2700.       case TGSI_SEMANTIC_FACE:
  2701.          shader->inputs[i].interp = LP_INTERP_FACING;
  2702.          break;
  2703.       case TGSI_SEMANTIC_POSITION:
  2704.          /* Position was already emitted above
  2705.           */
  2706.          shader->inputs[i].interp = LP_INTERP_POSITION;
  2707.          shader->inputs[i].src_index = 0;
  2708.          continue;
  2709.       }
  2710.  
  2711.       shader->inputs[i].src_index = i+1;
  2712.    }
  2713.  
  2714.    if (LP_DEBUG & DEBUG_TGSI) {
  2715.       unsigned attrib;
  2716.       debug_printf("llvmpipe: Create fragment shader #%u %p:\n",
  2717.                    shader->no, (void *) shader);
  2718.       tgsi_dump(templ->tokens, 0);
  2719.       debug_printf("usage masks:\n");
  2720.       for (attrib = 0; attrib < shader->info.base.num_inputs; ++attrib) {
  2721.          unsigned usage_mask = shader->info.base.input_usage_mask[attrib];
  2722.          debug_printf("  IN[%u].%s%s%s%s\n",
  2723.                       attrib,
  2724.                       usage_mask & TGSI_WRITEMASK_X ? "x" : "",
  2725.                       usage_mask & TGSI_WRITEMASK_Y ? "y" : "",
  2726.                       usage_mask & TGSI_WRITEMASK_Z ? "z" : "",
  2727.                       usage_mask & TGSI_WRITEMASK_W ? "w" : "");
  2728.       }
  2729.       debug_printf("\n");
  2730.    }
  2731.  
  2732.    return shader;
  2733. }
  2734.  
  2735.  
  2736. static void
  2737. llvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs)
  2738. {
  2739.    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
  2740.  
  2741.    if (llvmpipe->fs == fs)
  2742.       return;
  2743.  
  2744.    llvmpipe->fs = (struct lp_fragment_shader *) fs;
  2745.  
  2746.    draw_bind_fragment_shader(llvmpipe->draw,
  2747.                              (llvmpipe->fs ? llvmpipe->fs->draw_data : NULL));
  2748.  
  2749.    llvmpipe->dirty |= LP_NEW_FS;
  2750. }
  2751.  
  2752.  
  2753. /**
  2754.  * Remove shader variant from two lists: the shader's variant list
  2755.  * and the context's variant list.
  2756.  */
  2757. void
  2758. llvmpipe_remove_shader_variant(struct llvmpipe_context *lp,
  2759.                                struct lp_fragment_shader_variant *variant)
  2760. {
  2761.    if (gallivm_debug & GALLIVM_DEBUG_IR) {
  2762.       debug_printf("llvmpipe: del fs #%u var #%u v created #%u v cached"
  2763.                    " #%u v total cached #%u\n",
  2764.                    variant->shader->no,
  2765.                    variant->no,
  2766.                    variant->shader->variants_created,
  2767.                    variant->shader->variants_cached,
  2768.                    lp->nr_fs_variants);
  2769.    }
  2770.  
  2771.    gallivm_destroy(variant->gallivm);
  2772.  
  2773.    /* remove from shader's list */
  2774.    remove_from_list(&variant->list_item_local);
  2775.    variant->shader->variants_cached--;
  2776.  
  2777.    /* remove from context's list */
  2778.    remove_from_list(&variant->list_item_global);
  2779.    lp->nr_fs_variants--;
  2780.    lp->nr_fs_instrs -= variant->nr_instrs;
  2781.  
  2782.    FREE(variant);
  2783. }
  2784.  
  2785.  
  2786. static void
  2787. llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs)
  2788. {
  2789.    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
  2790.    struct lp_fragment_shader *shader = fs;
  2791.    struct lp_fs_variant_list_item *li;
  2792.  
  2793.    assert(fs != llvmpipe->fs);
  2794.  
  2795.    /*
  2796.     * XXX: we need to flush the context until we have some sort of reference
  2797.     * counting in fragment shaders as they may still be binned
  2798.     * Flushing alone might not sufficient we need to wait on it too.
  2799.     */
  2800.    llvmpipe_finish(pipe, __FUNCTION__);
  2801.  
  2802.    /* Delete all the variants */
  2803.    li = first_elem(&shader->variants);
  2804.    while(!at_end(&shader->variants, li)) {
  2805.       struct lp_fs_variant_list_item *next = next_elem(li);
  2806.       llvmpipe_remove_shader_variant(llvmpipe, li->base);
  2807.       li = next;
  2808.    }
  2809.  
  2810.    /* Delete draw module's data */
  2811.    draw_delete_fragment_shader(llvmpipe->draw, shader->draw_data);
  2812.  
  2813.    assert(shader->variants_cached == 0);
  2814.    FREE((void *) shader->base.tokens);
  2815.    FREE(shader);
  2816. }
  2817.  
  2818.  
  2819.  
  2820. static void
  2821. llvmpipe_set_constant_buffer(struct pipe_context *pipe,
  2822.                              uint shader, uint index,
  2823.                              struct pipe_constant_buffer *cb)
  2824. {
  2825.    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
  2826.    struct pipe_resource *constants = cb ? cb->buffer : NULL;
  2827.  
  2828.    assert(shader < PIPE_SHADER_TYPES);
  2829.    assert(index < Elements(llvmpipe->constants[shader]));
  2830.  
  2831.    /* note: reference counting */
  2832.    util_copy_constant_buffer(&llvmpipe->constants[shader][index], cb);
  2833.  
  2834.    if (shader == PIPE_SHADER_VERTEX ||
  2835.        shader == PIPE_SHADER_GEOMETRY) {
  2836.       /* Pass the constants to the 'draw' module */
  2837.       const unsigned size = cb ? cb->buffer_size : 0;
  2838.       const ubyte *data;
  2839.  
  2840.       if (constants) {
  2841.          data = (ubyte *) llvmpipe_resource_data(constants);
  2842.       }
  2843.       else if (cb && cb->user_buffer) {
  2844.          data = (ubyte *) cb->user_buffer;
  2845.       }
  2846.       else {
  2847.          data = NULL;
  2848.       }
  2849.  
  2850.       if (data)
  2851.          data += cb->buffer_offset;
  2852.  
  2853.       draw_set_mapped_constant_buffer(llvmpipe->draw, shader,
  2854.                                       index, data, size);
  2855.    }
  2856.  
  2857.    llvmpipe->dirty |= LP_NEW_CONSTANTS;
  2858.  
  2859.    if (cb && cb->user_buffer) {
  2860.       pipe_resource_reference(&constants, NULL);
  2861.    }
  2862. }
  2863.  
  2864.  
  2865. /**
  2866.  * Return the blend factor equivalent to a destination alpha of one.
  2867.  */
  2868. static INLINE unsigned
  2869. force_dst_alpha_one(unsigned factor, boolean clamped_zero)
  2870. {
  2871.    switch(factor) {
  2872.    case PIPE_BLENDFACTOR_DST_ALPHA:
  2873.       return PIPE_BLENDFACTOR_ONE;
  2874.    case PIPE_BLENDFACTOR_INV_DST_ALPHA:
  2875.       return PIPE_BLENDFACTOR_ZERO;
  2876.    case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
  2877.       if (clamped_zero)
  2878.          return PIPE_BLENDFACTOR_ZERO;
  2879.       else
  2880.          return PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE;
  2881.    }
  2882.  
  2883.    return factor;
  2884. }
  2885.  
  2886.  
  2887. /**
  2888.  * We need to generate several variants of the fragment pipeline to match
  2889.  * all the combinations of the contributing state atoms.
  2890.  *
  2891.  * TODO: there is actually no reason to tie this to context state -- the
  2892.  * generated code could be cached globally in the screen.
  2893.  */
  2894. static void
  2895. make_variant_key(struct llvmpipe_context *lp,
  2896.                  struct lp_fragment_shader *shader,
  2897.                  struct lp_fragment_shader_variant_key *key)
  2898. {
  2899.    unsigned i;
  2900.  
  2901.    memset(key, 0, shader->variant_key_size);
  2902.  
  2903.    if (lp->framebuffer.zsbuf) {
  2904.       enum pipe_format zsbuf_format = lp->framebuffer.zsbuf->format;
  2905.       const struct util_format_description *zsbuf_desc =
  2906.          util_format_description(zsbuf_format);
  2907.  
  2908.       if (lp->depth_stencil->depth.enabled &&
  2909.           util_format_has_depth(zsbuf_desc)) {
  2910.          key->zsbuf_format = zsbuf_format;
  2911.          memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth);
  2912.       }
  2913.       if (lp->depth_stencil->stencil[0].enabled &&
  2914.           util_format_has_stencil(zsbuf_desc)) {
  2915.          key->zsbuf_format = zsbuf_format;
  2916.          memcpy(&key->stencil, &lp->depth_stencil->stencil, sizeof key->stencil);
  2917.       }
  2918.       if (llvmpipe_resource_is_1d(lp->framebuffer.zsbuf->texture)) {
  2919.          key->resource_1d = TRUE;
  2920.       }
  2921.    }
  2922.  
  2923.    /*
  2924.     * Propagate the depth clamp setting from the rasterizer state.
  2925.     * depth_clip == 0 implies depth clamping is enabled.
  2926.     *
  2927.     * When clip_halfz is enabled, then always clamp the depth values.
  2928.     */
  2929.    if (lp->rasterizer->clip_halfz) {
  2930.       key->depth_clamp = 1;
  2931.    } else {
  2932.       key->depth_clamp = (lp->rasterizer->depth_clip == 0) ? 1 : 0;
  2933.    }
  2934.  
  2935.    /* alpha test only applies if render buffer 0 is non-integer (or does not exist) */
  2936.    if (!lp->framebuffer.nr_cbufs ||
  2937.        !lp->framebuffer.cbufs[0] ||
  2938.        !util_format_is_pure_integer(lp->framebuffer.cbufs[0]->format)) {
  2939.       key->alpha.enabled = lp->depth_stencil->alpha.enabled;
  2940.    }
  2941.    if(key->alpha.enabled)
  2942.       key->alpha.func = lp->depth_stencil->alpha.func;
  2943.    /* alpha.ref_value is passed in jit_context */
  2944.  
  2945.    key->flatshade = lp->rasterizer->flatshade;
  2946.    if (lp->active_occlusion_queries) {
  2947.       key->occlusion_count = TRUE;
  2948.    }
  2949.  
  2950.    if (lp->framebuffer.nr_cbufs) {
  2951.       memcpy(&key->blend, lp->blend, sizeof key->blend);
  2952.    }
  2953.  
  2954.    key->nr_cbufs = lp->framebuffer.nr_cbufs;
  2955.  
  2956.    if (!key->blend.independent_blend_enable) {
  2957.       /* we always need independent blend otherwise the fixups below won't work */
  2958.       for (i = 1; i < key->nr_cbufs; i++) {
  2959.          memcpy(&key->blend.rt[i], &key->blend.rt[0], sizeof(key->blend.rt[0]));
  2960.       }
  2961.       key->blend.independent_blend_enable = 1;
  2962.    }
  2963.  
  2964.    for (i = 0; i < lp->framebuffer.nr_cbufs; i++) {
  2965.       struct pipe_rt_blend_state *blend_rt = &key->blend.rt[i];
  2966.  
  2967.       if (lp->framebuffer.cbufs[i]) {
  2968.          enum pipe_format format = lp->framebuffer.cbufs[i]->format;
  2969.          const struct util_format_description *format_desc;
  2970.  
  2971.          key->cbuf_format[i] = format;
  2972.  
  2973.          /*
  2974.           * Figure out if this is a 1d resource. Note that OpenGL allows crazy
  2975.           * mixing of 2d textures with height 1 and 1d textures, so make sure
  2976.           * we pick 1d if any cbuf or zsbuf is 1d.
  2977.           */
  2978.          if (llvmpipe_resource_is_1d(lp->framebuffer.cbufs[i]->texture)) {
  2979.             key->resource_1d = TRUE;
  2980.          }
  2981.  
  2982.          format_desc = util_format_description(format);
  2983.          assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
  2984.                 format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB);
  2985.  
  2986.          /*
  2987.           * Mask out color channels not present in the color buffer.
  2988.           */
  2989.          blend_rt->colormask &= util_format_colormask(format_desc);
  2990.  
  2991.          /*
  2992.           * Disable blend for integer formats.
  2993.           */
  2994.          if (util_format_is_pure_integer(format)) {
  2995.             blend_rt->blend_enable = 0;
  2996.          }
  2997.  
  2998.          /*
  2999.           * Our swizzled render tiles always have an alpha channel, but the
  3000.           * linear render target format often does not, so force here the dst
  3001.           * alpha to be one.
  3002.           *
  3003.           * This is not a mere optimization. Wrong results will be produced if
  3004.           * the dst alpha is used, the dst format does not have alpha, and the
  3005.           * previous rendering was not flushed from the swizzled to linear
  3006.           * buffer. For example, NonPowTwo DCT.
  3007.           *
  3008.           * TODO: This should be generalized to all channels for better
  3009.           * performance, but only alpha causes correctness issues.
  3010.           *
  3011.           * Also, force rgb/alpha func/factors match, to make AoS blending
  3012.           * easier.
  3013.           */
  3014.          if (format_desc->swizzle[3] > UTIL_FORMAT_SWIZZLE_W ||
  3015.              format_desc->swizzle[3] == format_desc->swizzle[0]) {
  3016.             /* Doesn't cover mixed snorm/unorm but can't render to them anyway */
  3017.             boolean clamped_zero = !util_format_is_float(format) &&
  3018.                                    !util_format_is_snorm(format);
  3019.             blend_rt->rgb_src_factor =
  3020.                force_dst_alpha_one(blend_rt->rgb_src_factor, clamped_zero);
  3021.             blend_rt->rgb_dst_factor =
  3022.                force_dst_alpha_one(blend_rt->rgb_dst_factor, clamped_zero);
  3023.             blend_rt->alpha_func       = blend_rt->rgb_func;
  3024.             blend_rt->alpha_src_factor = blend_rt->rgb_src_factor;
  3025.             blend_rt->alpha_dst_factor = blend_rt->rgb_dst_factor;
  3026.          }
  3027.       }
  3028.       else {
  3029.          /* no color buffer for this fragment output */
  3030.          key->cbuf_format[i] = PIPE_FORMAT_NONE;
  3031.          blend_rt->colormask = 0x0;
  3032.          blend_rt->blend_enable = 0;
  3033.       }
  3034.    }
  3035.  
  3036.    /* This value will be the same for all the variants of a given shader:
  3037.     */
  3038.    key->nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
  3039.  
  3040.    for(i = 0; i < key->nr_samplers; ++i) {
  3041.       if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
  3042.          lp_sampler_static_sampler_state(&key->state[i].sampler_state,
  3043.                                          lp->samplers[PIPE_SHADER_FRAGMENT][i]);
  3044.       }
  3045.    }
  3046.  
  3047.    /*
  3048.     * XXX If TGSI_FILE_SAMPLER_VIEW exists assume all texture opcodes
  3049.     * are dx10-style? Can't really have mixed opcodes, at least not
  3050.     * if we want to skip the holes here (without rescanning tgsi).
  3051.     */
  3052.    if (shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] != -1) {
  3053.       key->nr_sampler_views = shader->info.base.file_max[TGSI_FILE_SAMPLER_VIEW] + 1;
  3054.       for(i = 0; i < key->nr_sampler_views; ++i) {
  3055.          if(shader->info.base.file_mask[TGSI_FILE_SAMPLER_VIEW] & (1 << i)) {
  3056.             lp_sampler_static_texture_state(&key->state[i].texture_state,
  3057.                                             lp->sampler_views[PIPE_SHADER_FRAGMENT][i]);
  3058.          }
  3059.       }
  3060.    }
  3061.    else {
  3062.       key->nr_sampler_views = key->nr_samplers;
  3063.       for(i = 0; i < key->nr_sampler_views; ++i) {
  3064.          if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
  3065.             lp_sampler_static_texture_state(&key->state[i].texture_state,
  3066.                                             lp->sampler_views[PIPE_SHADER_FRAGMENT][i]);
  3067.          }
  3068.       }
  3069.    }
  3070. }
  3071.  
  3072.  
  3073.  
  3074. /**
  3075.  * Update fragment shader state.  This is called just prior to drawing
  3076.  * something when some fragment-related state has changed.
  3077.  */
  3078. void
  3079. llvmpipe_update_fs(struct llvmpipe_context *lp)
  3080. {
  3081.    struct lp_fragment_shader *shader = lp->fs;
  3082.    struct lp_fragment_shader_variant_key key;
  3083.    struct lp_fragment_shader_variant *variant = NULL;
  3084.    struct lp_fs_variant_list_item *li;
  3085.  
  3086.    make_variant_key(lp, shader, &key);
  3087.  
  3088.    /* Search the variants for one which matches the key */
  3089.    li = first_elem(&shader->variants);
  3090.    while(!at_end(&shader->variants, li)) {
  3091.       if(memcmp(&li->base->key, &key, shader->variant_key_size) == 0) {
  3092.          variant = li->base;
  3093.          break;
  3094.       }
  3095.       li = next_elem(li);
  3096.    }
  3097.  
  3098.    if (variant) {
  3099.       /* Move this variant to the head of the list to implement LRU
  3100.        * deletion of shader's when we have too many.
  3101.        */
  3102.       move_to_head(&lp->fs_variants_list, &variant->list_item_global);
  3103.    }
  3104.    else {
  3105.       /* variant not found, create it now */
  3106.       int64_t t0, t1, dt;
  3107.       unsigned i;
  3108.       unsigned variants_to_cull;
  3109.  
  3110.       if (0) {
  3111.          debug_printf("%u variants,\t%u instrs,\t%u instrs/variant\n",
  3112.                       lp->nr_fs_variants,
  3113.                       lp->nr_fs_instrs,
  3114.                       lp->nr_fs_variants ? lp->nr_fs_instrs / lp->nr_fs_variants : 0);
  3115.       }
  3116.  
  3117.       /* First, check if we've exceeded the max number of shader variants.
  3118.        * If so, free 25% of them (the least recently used ones).
  3119.        */
  3120.       variants_to_cull = lp->nr_fs_variants >= LP_MAX_SHADER_VARIANTS ? LP_MAX_SHADER_VARIANTS / 4 : 0;
  3121.  
  3122.       if (variants_to_cull ||
  3123.           lp->nr_fs_instrs >= LP_MAX_SHADER_INSTRUCTIONS) {
  3124.          struct pipe_context *pipe = &lp->pipe;
  3125.  
  3126.          /*
  3127.           * XXX: we need to flush the context until we have some sort of
  3128.           * reference counting in fragment shaders as they may still be binned
  3129.           * Flushing alone might not be sufficient we need to wait on it too.
  3130.           */
  3131.          llvmpipe_finish(pipe, __FUNCTION__);
  3132.  
  3133.          /*
  3134.           * We need to re-check lp->nr_fs_variants because an arbitrarliy large
  3135.           * number of shader variants (potentially all of them) could be
  3136.           * pending for destruction on flush.
  3137.           */
  3138.  
  3139.          for (i = 0; i < variants_to_cull || lp->nr_fs_instrs >= LP_MAX_SHADER_INSTRUCTIONS; i++) {
  3140.             struct lp_fs_variant_list_item *item;
  3141.             if (is_empty_list(&lp->fs_variants_list)) {
  3142.                break;
  3143.             }
  3144.             item = last_elem(&lp->fs_variants_list);
  3145.             assert(item);
  3146.             assert(item->base);
  3147.             llvmpipe_remove_shader_variant(lp, item->base);
  3148.          }
  3149.       }
  3150.  
  3151.       /*
  3152.        * Generate the new variant.
  3153.        */
  3154.       t0 = os_time_get();
  3155.       variant = generate_variant(lp, shader, &key);
  3156.       t1 = os_time_get();
  3157.       dt = t1 - t0;
  3158.       LP_COUNT_ADD(llvm_compile_time, dt);
  3159.       LP_COUNT_ADD(nr_llvm_compiles, 2);  /* emit vs. omit in/out test */
  3160.  
  3161.       /* Put the new variant into the list */
  3162.       if (variant) {
  3163.          insert_at_head(&shader->variants, &variant->list_item_local);
  3164.          insert_at_head(&lp->fs_variants_list, &variant->list_item_global);
  3165.          lp->nr_fs_variants++;
  3166.          lp->nr_fs_instrs += variant->nr_instrs;
  3167.          shader->variants_cached++;
  3168.       }
  3169.    }
  3170.  
  3171.    /* Bind this variant */
  3172.    lp_setup_set_fs_variant(lp->setup, variant);
  3173. }
  3174.  
  3175.  
  3176.  
  3177.  
  3178.  
  3179. void
  3180. llvmpipe_init_fs_funcs(struct llvmpipe_context *llvmpipe)
  3181. {
  3182.    llvmpipe->pipe.create_fs_state = llvmpipe_create_fs_state;
  3183.    llvmpipe->pipe.bind_fs_state   = llvmpipe_bind_fs_state;
  3184.    llvmpipe->pipe.delete_fs_state = llvmpipe_delete_fs_state;
  3185.  
  3186.    llvmpipe->pipe.set_constant_buffer = llvmpipe_set_constant_buffer;
  3187. }
  3188.  
  3189. /*
  3190.  * Rasterization is disabled if there is no pixel shader and
  3191.  * both depth and stencil testing are disabled:
  3192.  * http://msdn.microsoft.com/en-us/library/windows/desktop/bb205125
  3193.  */
  3194. boolean
  3195. llvmpipe_rasterization_disabled(struct llvmpipe_context *lp)
  3196. {
  3197.    boolean null_fs = !lp->fs || lp->fs->info.base.num_tokens <= 1;
  3198.  
  3199.    return (null_fs &&
  3200.            !lp->depth_stencil->depth.enabled &&
  3201.            !lp->depth_stencil->stencil[0].enabled);
  3202. }
  3203.