dlls/wined3d/shader_sm4.c

   1 /*
   2  * Copyright 2009 Henri Verbeet for CodeWeavers
   3  *
   4  * This library is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU Lesser General Public
   6  * License as published by the Free Software Foundation; either
   7  * version 2.1 of the License, or (at your option) any later version.
   8  *
   9  * This library is distributed in the hope that it will be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  12  * Lesser General Public License for more details.
  13  *
  14  * You should have received a copy of the GNU Lesser General Public
  15  * License along with this library; if not, write to the Free Software
  16  * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
  17  */
  18
  19 #include "config.h"
  20 #include "wine/port.h"
  21
  22 #include "wined3d_private.h"
  23
  24 WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
  25 WINE_DECLARE_DEBUG_CHANNEL(d3d_bytecode);
  26
  27 #define WINED3D_SM4_INSTRUCTION_MODIFIER        (1 << 31)
  28
  29 #define WINED3D_SM4_INSTRUCTION_LENGTH_SHIFT    24
  30 #define WINED3D_SM4_INSTRUCTION_LENGTH_MASK     (0x1f << WINED3D_SM4_INSTRUCTION_LENGTH_SHIFT)
  31
  32 #define WINED3D_SM4_PRIMITIVE_TYPE_SHIFT        11
  33 #define WINED3D_SM4_PRIMITIVE_TYPE_MASK         (0x7 << WINED3D_SM4_PRIMITIVE_TYPE_SHIFT)
  34
  35 #define WINED3D_SM4_INDEX_TYPE_SHIFT            11
  36 #define WINED3D_SM4_INDEX_TYPE_MASK             (0x1 << WINED3D_SM4_INDEX_TYPE_SHIFT)
  37
  38 #define WINED3D_SM4_OPCODE_MASK                 0xff
  39
  40 #define WINED3D_SM4_REGISTER_MODIFIER           (1 << 31)
  41
  42 #define WINED3D_SM4_ADDRESSING_SHIFT1           25
  43 #define WINED3D_SM4_ADDRESSING_MASK1            (0x3 << WINED3D_SM4_ADDRESSING_SHIFT1)
  44
  45 #define WINED3D_SM4_ADDRESSING_SHIFT0           22
  46 #define WINED3D_SM4_ADDRESSING_MASK0            (0x3 << WINED3D_SM4_ADDRESSING_SHIFT0)
  47
  48 #define WINED3D_SM4_REGISTER_ORDER_SHIFT        20
  49 #define WINED3D_SM4_REGISTER_ORDER_MASK         (0x3 << WINED3D_SM4_REGISTER_ORDER_SHIFT)
  50
  51 #define WINED3D_SM4_REGISTER_TYPE_SHIFT         12
  52 #define WINED3D_SM4_REGISTER_TYPE_MASK          (0xf << WINED3D_SM4_REGISTER_TYPE_SHIFT)
  53
  54 #define WINED3D_SM4_SWIZZLE_TYPE_SHIFT          2
  55 #define WINED3D_SM4_SWIZZLE_TYPE_MASK           (0x3 << WINED3D_SM4_SWIZZLE_TYPE_SHIFT)
  56
  57 #define WINED3D_SM4_IMMCONST_TYPE_SHIFT         0
  58 #define WINED3D_SM4_IMMCONST_TYPE_MASK          (0x3 << WINED3D_SM4_IMMCONST_TYPE_SHIFT)
  59
  60 #define WINED3D_SM4_WRITEMASK_SHIFT             4
  61 #define WINED3D_SM4_WRITEMASK_MASK              (0xf << WINED3D_SM4_WRITEMASK_SHIFT)
  62
  63 #define WINED3D_SM4_SWIZZLE_SHIFT               4
  64 #define WINED3D_SM4_SWIZZLE_MASK                (0xff << WINED3D_SM4_SWIZZLE_SHIFT)
  65
  66 #define WINED3D_SM4_VERSION_MAJOR(version)      (((version) >> 4) & 0xf)
  67 #define WINED3D_SM4_VERSION_MINOR(version)      (((version) >> 0) & 0xf)
  68
  69 #define WINED3D_SM4_ADDRESSING_RELATIVE         0x2
  70 #define WINED3D_SM4_ADDRESSING_OFFSET           0x1
  71
  72 enum wined3d_sm4_opcode
  73 {
  74     WINED3D_SM4_OP_ADD                  = 0x00,
  75     WINED3D_SM4_OP_AND                  = 0x01,
  76     WINED3D_SM4_OP_BREAK                = 0x02,
  77     WINED3D_SM4_OP_BREAKC               = 0x03,
  78     WINED3D_SM4_OP_CUT                  = 0x09,
  79     WINED3D_SM4_OP_DERIV_RTX            = 0x0b,
  80     WINED3D_SM4_OP_DERIV_RTY            = 0x0c,
  81     WINED3D_SM4_OP_DISCARD              = 0x0d,
  82     WINED3D_SM4_OP_DIV                  = 0x0e,
  83     WINED3D_SM4_OP_DP2                  = 0x0f,
  84     WINED3D_SM4_OP_DP3                  = 0x10,
  85     WINED3D_SM4_OP_DP4                  = 0x11,
  86     WINED3D_SM4_OP_EMIT                 = 0x13,
  87     WINED3D_SM4_OP_ENDIF                = 0x15,
  88     WINED3D_SM4_OP_ENDLOOP              = 0x16,
  89     WINED3D_SM4_OP_EQ                   = 0x18,
  90     WINED3D_SM4_OP_EXP                  = 0x19,
  91     WINED3D_SM4_OP_FRC                  = 0x1a,
  92     WINED3D_SM4_OP_FTOI                 = 0x1b,
  93     WINED3D_SM4_OP_GE                   = 0x1d,
  94     WINED3D_SM4_OP_IADD                 = 0x1e,
  95     WINED3D_SM4_OP_IF                   = 0x1f,
  96     WINED3D_SM4_OP_IEQ                  = 0x20,
  97     WINED3D_SM4_OP_IGE                  = 0x21,
  98     WINED3D_SM4_OP_IMUL                 = 0x26,
  99     WINED3D_SM4_OP_ISHL                 = 0x29,
 100     WINED3D_SM4_OP_ITOF                 = 0x2b,
 101     WINED3D_SM4_OP_LD                   = 0x2d,
 102     WINED3D_SM4_OP_LOG                  = 0x2f,
 103     WINED3D_SM4_OP_LOOP                 = 0x30,
 104     WINED3D_SM4_OP_LT                   = 0x31,
 105     WINED3D_SM4_OP_MAD                  = 0x32,
 106     WINED3D_SM4_OP_MIN                  = 0x33,
 107     WINED3D_SM4_OP_MAX                  = 0x34,
 108     WINED3D_SM4_OP_MOV                  = 0x36,
 109     WINED3D_SM4_OP_MOVC                 = 0x37,
 110     WINED3D_SM4_OP_MUL                  = 0x38,
 111     WINED3D_SM4_OP_RET                  = 0x3e,
 112     WINED3D_SM4_OP_ROUND_NI             = 0x41,
 113     WINED3D_SM4_OP_RSQ                  = 0x44,
 114     WINED3D_SM4_OP_SAMPLE               = 0x45,
 115     WINED3D_SM4_OP_SAMPLE_LOD           = 0x48,
 116     WINED3D_SM4_OP_SAMPLE_GRAD          = 0x49,
 117     WINED3D_SM4_OP_SQRT                 = 0x4b,
 118     WINED3D_SM4_OP_SINCOS               = 0x4d,
 119     WINED3D_SM4_OP_UDIV                 = 0x4e,
 120     WINED3D_SM4_OP_USHR                 = 0x55,
 121     WINED3D_SM4_OP_UTOF                 = 0x56,
 122     WINED3D_SM4_OP_XOR                  = 0x57,
 123     WINED3D_SM4_OP_DCL_CONSTANT_BUFFER  = 0x59,
 124     WINED3D_SM4_OP_DCL_OUTPUT_TOPOLOGY  = 0x5c,
 125     WINED3D_SM4_OP_DCL_INPUT_PRIMITIVE  = 0x5d,
 126     WINED3D_SM4_OP_DCL_VERTICES_OUT     = 0x5e,
 127 };
 128
 129 enum wined3d_sm4_register_type
 130 {
 131     WINED3D_SM4_RT_TEMP         = 0x0,
 132     WINED3D_SM4_RT_INPUT        = 0x1,
 133     WINED3D_SM4_RT_OUTPUT       = 0x2,
 134     WINED3D_SM4_RT_IMMCONST     = 0x4,
 135     WINED3D_SM4_RT_SAMPLER      = 0x6,
 136     WINED3D_SM4_RT_CONSTBUFFER  = 0x8,
 137     WINED3D_SM4_RT_PRIMID       = 0xb,
 138     WINED3D_SM4_RT_NULL         = 0xd,
 139 };
 140
 141 enum wined3d_sm4_output_primitive_type
 142 {
 143     WINED3D_SM4_OUTPUT_PT_POINTLIST     = 0x1,
 144     WINED3D_SM4_OUTPUT_PT_LINELIST      = 0x3,
 145     WINED3D_SM4_OUTPUT_PT_TRIANGLESTRIP = 0x5,
 146 };
 147
 148 enum wined3d_sm4_input_primitive_type
 149 {
 150     WINED3D_SM4_INPUT_PT_POINT          = 0x1,
 151     WINED3D_SM4_INPUT_PT_LINE           = 0x2,
 152     WINED3D_SM4_INPUT_PT_TRIANGLE       = 0x3,
 153     WINED3D_SM4_INPUT_PT_LINEADJ        = 0x6,
 154     WINED3D_SM4_INPUT_PT_TRIANGLEADJ    = 0x7,
 155 };
 156
 157 enum wined3d_sm4_swizzle_type
 158 {
 159     WINED3D_SM4_SWIZZLE_VEC4            = 0x1,
 160     WINED3D_SM4_SWIZZLE_SCALAR          = 0x2,
 161 };
 162
 163 enum wined3d_sm4_immconst_type
 164 {
 165     WINED3D_SM4_IMMCONST_SCALAR = 0x1,
 166     WINED3D_SM4_IMMCONST_VEC4   = 0x2,
 167 };
 168
 169 struct wined3d_shader_src_param_entry
 170 {
 171     struct list entry;
 172     struct wined3d_shader_src_param param;
 173 };
 174
 175 struct wined3d_sm4_data
 176 {
 177     struct wined3d_shader_version shader_version;
 178     const DWORD *end;
 179     const struct wined3d_shader_signature *output_signature;
 180
 181     struct wined3d_shader_src_param src_param[5];
 182     struct wined3d_shader_dst_param dst_param[2];
 183     struct list src_free;
 184     struct list src;
 185 };
 186
 187 struct wined3d_sm4_opcode_info
 188 {
 189     enum wined3d_sm4_opcode opcode;
 190     enum WINED3D_SHADER_INSTRUCTION_HANDLER handler_idx;
 191     const char *dst_info;
 192     const char *src_info;
 193 };
 194
 195 struct sysval_map
 196 {
 197     enum wined3d_sysval_semantic sysval;
 198     enum wined3d_shader_register_type register_type;
 199     UINT register_idx;
 200 };
 201
 202 /*
 203  * F -> WINED3D_DATA_FLOAT
 204  * I -> WINED3D_DATA_INT
 205  * R -> WINED3D_DATA_RESOURCE
 206  * S -> WINED3D_DATA_SAMPLER
 207  * U -> WINED3D_DATA_UINT
 208  */
 209 static const struct wined3d_sm4_opcode_info opcode_table[] =
 210 {
 211     {WINED3D_SM4_OP_ADD,                    WINED3DSIH_ADD,                 "F",    "FF"},
 212     {WINED3D_SM4_OP_AND,                    WINED3DSIH_AND,                 "U",    "UU"},
 213     {WINED3D_SM4_OP_BREAK,                  WINED3DSIH_BREAK,               "",     ""},
 214     {WINED3D_SM4_OP_BREAKC,                 WINED3DSIH_BREAKP,              "",     "U"},
 215     {WINED3D_SM4_OP_CUT,                    WINED3DSIH_CUT,                 "",     ""},
 216     {WINED3D_SM4_OP_DERIV_RTX,              WINED3DSIH_DSX,                 "F",    "F"},
 217     {WINED3D_SM4_OP_DERIV_RTY,              WINED3DSIH_DSY,                 "F",    "F"},
 218     {WINED3D_SM4_OP_DISCARD,                WINED3DSIH_TEXKILL,             "",     "U"},
 219     {WINED3D_SM4_OP_DIV,                    WINED3DSIH_DIV,                 "F",    "FF"},
 220     {WINED3D_SM4_OP_DP2,                    WINED3DSIH_DP2,                 "F",    "FF"},
 221     {WINED3D_SM4_OP_DP3,                    WINED3DSIH_DP3,                 "F",    "FF"},
 222     {WINED3D_SM4_OP_DP4,                    WINED3DSIH_DP4,                 "F",    "FF"},
 223     {WINED3D_SM4_OP_EMIT,                   WINED3DSIH_EMIT,                "",     ""},
 224     {WINED3D_SM4_OP_ENDIF,                  WINED3DSIH_ENDIF,               "",     ""},
 225     {WINED3D_SM4_OP_ENDLOOP,                WINED3DSIH_ENDLOOP,             "",     ""},
 226     {WINED3D_SM4_OP_EQ,                     WINED3DSIH_EQ,                  "U",    "FF"},
 227     {WINED3D_SM4_OP_EXP,                    WINED3DSIH_EXP,                 "F",    "F"},
 228     {WINED3D_SM4_OP_FRC,                    WINED3DSIH_FRC,                 "F",    "F"},
 229     {WINED3D_SM4_OP_FTOI,                   WINED3DSIH_FTOI,                "I",    "F"},
 230     {WINED3D_SM4_OP_GE,                     WINED3DSIH_GE,                  "U",    "FF"},
 231     {WINED3D_SM4_OP_IADD,                   WINED3DSIH_IADD,                "I",    "II"},
 232     {WINED3D_SM4_OP_IF,                     WINED3DSIH_IF,                  "",     "U"},
 233     {WINED3D_SM4_OP_IEQ,                    WINED3DSIH_IEQ,                 "U",    "II"},
 234     {WINED3D_SM4_OP_IGE,                    WINED3DSIH_IGE,                 "U",    "II"},
 235     {WINED3D_SM4_OP_IMUL,                   WINED3DSIH_IMUL,                "II",   "II"},
 236     {WINED3D_SM4_OP_ISHL,                   WINED3DSIH_ISHL,                "I",    "II"},
 237     {WINED3D_SM4_OP_ITOF,                   WINED3DSIH_ITOF,                "F",    "I"},
 238     {WINED3D_SM4_OP_LD,                     WINED3DSIH_LD,                  "U",    "FR"},
 239     {WINED3D_SM4_OP_LOG,                    WINED3DSIH_LOG,                 "F",    "F"},
 240     {WINED3D_SM4_OP_LOOP,                   WINED3DSIH_LOOP,                "",     ""},
 241     {WINED3D_SM4_OP_LT,                     WINED3DSIH_LT,                  "U",    "FF"},
 242     {WINED3D_SM4_OP_MAD,                    WINED3DSIH_MAD,                 "F",    "FFF"},
 243     {WINED3D_SM4_OP_MIN,                    WINED3DSIH_MIN,                 "F",    "FF"},
 244     {WINED3D_SM4_OP_MAX,                    WINED3DSIH_MAX,                 "F",    "FF"},
 245     {WINED3D_SM4_OP_MOV,                    WINED3DSIH_MOV,                 "F",    "F"},
 246     {WINED3D_SM4_OP_MOVC,                   WINED3DSIH_MOVC,                "F",    "UFF"},
 247     {WINED3D_SM4_OP_MUL,                    WINED3DSIH_MUL,                 "F",    "FF"},
 248     {WINED3D_SM4_OP_RET,                    WINED3DSIH_RET,                 "",     ""},
 249     {WINED3D_SM4_OP_ROUND_NI,               WINED3DSIH_ROUND_NI,            "F",    "F"},
 250     {WINED3D_SM4_OP_RSQ,                    WINED3DSIH_RSQ,                 "F",    "F"},
 251     {WINED3D_SM4_OP_SAMPLE,                 WINED3DSIH_SAMPLE,              "U",    "FRS"},
 252     {WINED3D_SM4_OP_SAMPLE_LOD,             WINED3DSIH_SAMPLE_LOD,          "U",    "FRSF"},
 253     {WINED3D_SM4_OP_SAMPLE_GRAD,            WINED3DSIH_SAMPLE_GRAD,         "U",    "FRSFF"},
 254     {WINED3D_SM4_OP_SQRT,                   WINED3DSIH_SQRT,                "F",    "F"},
 255     {WINED3D_SM4_OP_SINCOS,                 WINED3DSIH_SINCOS,              "FF",   "F"},
 256     {WINED3D_SM4_OP_UDIV,                   WINED3DSIH_UDIV,                "UU",   "UU"},
 257     {WINED3D_SM4_OP_USHR,                   WINED3DSIH_USHR,                "U",    "UU"},
 258     {WINED3D_SM4_OP_UTOF,                   WINED3DSIH_UTOF,                "F",    "U"},
 259     {WINED3D_SM4_OP_XOR,                    WINED3DSIH_XOR,                 "U",    "UU"},
 260     {WINED3D_SM4_OP_DCL_CONSTANT_BUFFER,    WINED3DSIH_DCL_CONSTANT_BUFFER, "",     ""},
 261     {WINED3D_SM4_OP_DCL_OUTPUT_TOPOLOGY,    WINED3DSIH_DCL_OUTPUT_TOPOLOGY, "",     ""},
 262     {WINED3D_SM4_OP_DCL_INPUT_PRIMITIVE,    WINED3DSIH_DCL_INPUT_PRIMITIVE, "",     ""},
 263     {WINED3D_SM4_OP_DCL_VERTICES_OUT,       WINED3DSIH_DCL_VERTICES_OUT,    "",     ""},
 264 };
 265
 266 static const enum wined3d_shader_register_type register_type_table[] =
 267 {
 268     /* WINED3D_SM4_RT_TEMP */           WINED3DSPR_TEMP,
 269     /* WINED3D_SM4_RT_INPUT */          WINED3DSPR_INPUT,
 270     /* WINED3D_SM4_RT_OUTPUT */         WINED3DSPR_OUTPUT,
 271     /* UNKNOWN */                       0,
 272     /* WINED3D_SM4_RT_IMMCONST */       WINED3DSPR_IMMCONST,
 273     /* UNKNOWN */                       0,
 274     /* WINED3D_SM4_RT_SAMPLER */        WINED3DSPR_SAMPLER,
 275     /* WINED3D_SM4_RT_RESOURCE */       WINED3DSPR_RESOURCE,
 276     /* WINED3D_SM4_RT_CONSTBUFFER */    WINED3DSPR_CONSTBUFFER,
 277     /* UNKNOWN */                       0,
 278     /* UNKNOWN */                       0,
 279     /* WINED3D_SM4_RT_PRIMID */         WINED3DSPR_PRIMID,
 280     /* UNKNOWN */                       0,
 281     /* WINED3D_SM4_RT_NULL */           WINED3DSPR_NULL,
 282 };
 283
 284 static const enum wined3d_primitive_type output_primitive_type_table[] =
 285 {
 286     /* UNKNOWN */                               WINED3D_PT_UNDEFINED,
 287     /* WINED3D_SM4_OUTPUT_PT_POINTLIST */       WINED3D_PT_POINTLIST,
 288     /* UNKNOWN */                               WINED3D_PT_UNDEFINED,
 289     /* WINED3D_SM4_OUTPUT_PT_LINELIST */        WINED3D_PT_LINELIST,
 290     /* UNKNOWN */                               WINED3D_PT_UNDEFINED,
 291     /* WINED3D_SM4_OUTPUT_PT_TRIANGLESTRIP */   WINED3D_PT_TRIANGLESTRIP,
 292 };
 293
 294 static const enum wined3d_primitive_type input_primitive_type_table[] =
 295 {
 296     /* UNKNOWN */                               WINED3D_PT_UNDEFINED,
 297     /* WINED3D_SM4_INPUT_PT_POINT */            WINED3D_PT_POINTLIST,
 298     /* WINED3D_SM4_INPUT_PT_LINE */             WINED3D_PT_LINELIST,
 299     /* WINED3D_SM4_INPUT_PT_TRIANGLE */         WINED3D_PT_TRIANGLELIST,
 300     /* UNKNOWN */                               WINED3D_PT_UNDEFINED,
 301     /* UNKNOWN */                               WINED3D_PT_UNDEFINED,
 302     /* WINED3D_SM4_INPUT_PT_LINEADJ */          WINED3D_PT_LINELIST_ADJ,
 303     /* WINED3D_SM4_INPUT_PT_TRIANGLEADJ */      WINED3D_PT_TRIANGLELIST_ADJ,
 304 };
 305
 306 static const struct sysval_map sysval_map[] =
 307 {
 308     {WINED3D_SV_DEPTH,      WINED3DSPR_DEPTHOUT,    0},
 309     {WINED3D_SV_TARGET0,    WINED3DSPR_COLOROUT,    0},
 310     {WINED3D_SV_TARGET1,    WINED3DSPR_COLOROUT,    1},
 311     {WINED3D_SV_TARGET2,    WINED3DSPR_COLOROUT,    2},
 312     {WINED3D_SV_TARGET3,    WINED3DSPR_COLOROUT,    3},
 313     {WINED3D_SV_TARGET4,    WINED3DSPR_COLOROUT,    4},
 314     {WINED3D_SV_TARGET5,    WINED3DSPR_COLOROUT,    5},
 315     {WINED3D_SV_TARGET6,    WINED3DSPR_COLOROUT,    6},
 316     {WINED3D_SV_TARGET7,    WINED3DSPR_COLOROUT,    7},
 317 };
 318
 319 static BOOL shader_sm4_read_src_param(struct wined3d_sm4_data *priv, const DWORD **ptr,
 320         enum wined3d_data_type data_type, struct wined3d_shader_src_param *src_param);
 321
 322 static const struct wined3d_sm4_opcode_info *get_opcode_info(enum wined3d_sm4_opcode opcode)
 323 {
 324     unsigned int i;
 325
 326     for (i = 0; i < sizeof(opcode_table) / sizeof(*opcode_table); ++i)
 327     {
 328         if (opcode == opcode_table[i].opcode) return &opcode_table[i];
 329     }
 330
 331     return NULL;
 332 }
 333
 334 static void map_sysval(enum wined3d_sysval_semantic sysval, struct wined3d_shader_register *reg)
 335 {
 336     unsigned int i;
 337
 338     for (i = 0; i < sizeof(sysval_map) / sizeof(*sysval_map); ++i)
 339     {
 340         if (sysval == sysval_map[i].sysval)
 341         {
 342             reg->type = sysval_map[i].register_type;
 343             reg->idx[0].offset = sysval_map[i].register_idx;
 344         }
 345     }
 346 }
 347
 348 static void map_register(const struct wined3d_sm4_data *priv, struct wined3d_shader_register *reg)
 349 {
 350     switch (priv->shader_version.type)
 351     {
 352         case WINED3D_SHADER_TYPE_PIXEL:
 353             if (reg->type == WINED3DSPR_OUTPUT)
 354             {
 355                 unsigned int i;
 356                 const struct wined3d_shader_signature *s = priv->output_signature;
 357
 358                 if (!s)
 359                 {
 360                     ERR("Shader has no output signature, unable to map register.\n");
 361                     break;
 362                 }
 363
 364                 for (i = 0; i < s->element_count; ++i)
 365                 {
 366                     if (s->elements[i].register_idx == reg->idx[0].offset)
 367                     {
 368                         map_sysval(s->elements[i].sysval_semantic, reg);
 369                         break;
 370                     }
 371                 }
 372             }
 373             break;
 374
 375         default:
 376             break;
 377     }
 378 }
 379
 380 static enum wined3d_data_type map_data_type(char t)
 381 {
 382     switch (t)
 383     {
 384         case 'F':
 385             return WINED3D_DATA_FLOAT;
 386         case 'I':
 387             return WINED3D_DATA_INT;
 388         case 'R':
 389             return WINED3D_DATA_RESOURCE;
 390         case 'S':
 391             return WINED3D_DATA_SAMPLER;
 392         case 'U':
 393             return WINED3D_DATA_UINT;
 394         default:
 395             ERR("Invalid data type '%c'.\n", t);
 396             return WINED3D_DATA_FLOAT;
 397     }
 398 }
 399
 400 static void *shader_sm4_init(const DWORD *byte_code, const struct wined3d_shader_signature *output_signature)
 401 {
 402     struct wined3d_sm4_data *priv = HeapAlloc(GetProcessHeap(), 0, sizeof(*priv));
 403     if (!priv)
 404     {
 405         ERR("Failed to allocate private data\n");
 406         return NULL;
 407     }
 408
 409     priv->output_signature = output_signature;
 410     list_init(&priv->src_free);
 411     list_init(&priv->src);
 412
 413     return priv;
 414 }
 415
 416 static void shader_sm4_free(void *data)
 417 {
 418     struct wined3d_shader_src_param_entry *e1, *e2;
 419     struct wined3d_sm4_data *priv = data;
 420
 421     list_move_head(&priv->src_free, &priv->src);
 422     LIST_FOR_EACH_ENTRY_SAFE(e1, e2, &priv->src_free, struct wined3d_shader_src_param_entry, entry)
 423     {
 424         HeapFree(GetProcessHeap(), 0, e1);
 425     }
 426     HeapFree(GetProcessHeap(), 0, priv);
 427 }
 428
 429 static struct wined3d_shader_src_param *get_src_param(struct wined3d_sm4_data *priv)
 430 {
 431     struct wined3d_shader_src_param_entry *e;
 432     struct list *elem;
 433
 434     if (!list_empty(&priv->src_free))
 435     {
 436         elem = list_head(&priv->src_free);
 437         list_remove(elem);
 438     }
 439     else
 440     {
 441         if (!(e = HeapAlloc(GetProcessHeap(), 0, sizeof(*e))))
 442             return NULL;
 443         elem = &e->entry;
 444     }
 445
 446     list_add_tail(&priv->src, elem);
 447     e = LIST_ENTRY(elem, struct wined3d_shader_src_param_entry, entry);
 448     return &e->param;
 449 }
 450
 451 static void shader_sm4_read_header(void *data, const DWORD **ptr, struct wined3d_shader_version *shader_version)
 452 {
 453     struct wined3d_sm4_data *priv = data;
 454     DWORD version_token;
 455
 456     priv->end = *ptr;
 457
 458     version_token = *(*ptr)++;
 459     TRACE("version: 0x%08x\n", version_token);
 460
 461     TRACE("token count: %u\n", **ptr);
 462     priv->end += *(*ptr)++;
 463
 464     switch (version_token >> 16)
 465     {
 466         case WINED3D_SM4_PS:
 467             priv->shader_version.type = WINED3D_SHADER_TYPE_PIXEL;
 468             break;
 469
 470         case WINED3D_SM4_VS:
 471             priv->shader_version.type = WINED3D_SHADER_TYPE_VERTEX;
 472             break;
 473
 474         case WINED3D_SM4_GS:
 475             priv->shader_version.type = WINED3D_SHADER_TYPE_GEOMETRY;
 476             break;
 477
 478         default:
 479             FIXME("Unrecognized shader type %#x\n", version_token >> 16);
 480     }
 481     priv->shader_version.major = WINED3D_SM4_VERSION_MAJOR(version_token);
 482     priv->shader_version.minor = WINED3D_SM4_VERSION_MINOR(version_token);
 483
 484     *shader_version = priv->shader_version;
 485 }
 486
 487 static BOOL shader_sm4_read_reg_idx(struct wined3d_sm4_data *priv, const DWORD **ptr,
 488         DWORD addressing, struct wined3d_shader_register_index *reg_idx)
 489 {
 490     if (addressing & WINED3D_SM4_ADDRESSING_RELATIVE)
 491     {
 492         struct wined3d_shader_src_param *rel_addr = get_src_param(priv);
 493
 494         if (!(reg_idx->rel_addr = rel_addr))
 495         {
 496             ERR("Failed to get src param for relative addressing.\n");
 497             return FALSE;
 498         }
 499
 500         if (addressing & WINED3D_SM4_ADDRESSING_OFFSET)
 501             reg_idx->offset = *(*ptr)++;
 502         else
 503             reg_idx->offset = 0;
 504         shader_sm4_read_src_param(priv, ptr, WINED3D_DATA_INT, rel_addr);
 505     }
 506     else
 507     {
 508         reg_idx->rel_addr = NULL;
 509         reg_idx->offset = *(*ptr)++;
 510     }
 511
 512     return TRUE;
 513 }
 514
 515 static BOOL shader_sm4_read_param(struct wined3d_sm4_data *priv, const DWORD **ptr,
 516         enum wined3d_data_type data_type, struct wined3d_shader_register *param,
 517         enum wined3d_shader_src_modifier *modifier)
 518 {
 519     enum wined3d_sm4_register_type register_type;
 520     DWORD token = *(*ptr)++;
 521     DWORD order;
 522
 523     register_type = (token & WINED3D_SM4_REGISTER_TYPE_MASK) >> WINED3D_SM4_REGISTER_TYPE_SHIFT;
 524     if (register_type >= sizeof(register_type_table) / sizeof(*register_type_table))
 525     {
 526         FIXME("Unhandled register type %#x.\n", register_type);
 527         param->type = WINED3DSPR_TEMP;
 528     }
 529     else
 530     {
 531         param->type = register_type_table[register_type];
 532     }
 533     param->data_type = data_type;
 534
 535     if (token & WINED3D_SM4_REGISTER_MODIFIER)
 536     {
 537         DWORD m = *(*ptr)++;
 538
 539         /* FIXME: This will probably break down at some point. The SM4
 540          * modifiers look like flags, while wined3d currently has an enum
 541          * with possible combinations, e.g. WINED3DSPSM_ABSNEG. */
 542         switch (m)
 543         {
 544             case 0x41:
 545                 *modifier = WINED3DSPSM_NEG;
 546                 break;
 547
 548             case 0x81:
 549                 *modifier = WINED3DSPSM_ABS;
 550                 break;
 551
 552             default:
 553                 FIXME("Skipping modifier 0x%08x.\n", m);
 554                 *modifier = WINED3DSPSM_NONE;
 555                 break;
 556         }
 557     }
 558     else
 559     {
 560         *modifier = WINED3DSPSM_NONE;
 561     }
 562
 563     order = (token & WINED3D_SM4_REGISTER_ORDER_MASK) >> WINED3D_SM4_REGISTER_ORDER_SHIFT;
 564
 565     if (order < 1)
 566         param->idx[0].offset = ~0U;
 567     else
 568     {
 569         DWORD addressing = (token & WINED3D_SM4_ADDRESSING_MASK0) >> WINED3D_SM4_ADDRESSING_SHIFT0;
 570         if (!(shader_sm4_read_reg_idx(priv, ptr, addressing, &param->idx[0])))
 571         {
 572             ERR("Failed to read register index.\n");
 573             return FALSE;
 574         }
 575     }
 576
 577     if (order < 2)
 578         param->idx[1].offset = ~0U;
 579     else
 580     {
 581         DWORD addressing = (token & WINED3D_SM4_ADDRESSING_MASK1) >> WINED3D_SM4_ADDRESSING_SHIFT1;
 582         if (!(shader_sm4_read_reg_idx(priv, ptr, addressing, &param->idx[1])))
 583         {
 584             ERR("Failed to read register index.\n");
 585             return FALSE;
 586         }
 587     }
 588
 589     if (order > 2)
 590         FIXME("Unhandled order %u.\n", order);
 591
 592     if (register_type == WINED3D_SM4_RT_IMMCONST)
 593     {
 594         enum wined3d_sm4_immconst_type immconst_type =
 595                 (token & WINED3D_SM4_IMMCONST_TYPE_MASK) >> WINED3D_SM4_IMMCONST_TYPE_SHIFT;
 596
 597         switch (immconst_type)
 598         {
 599             case WINED3D_SM4_IMMCONST_SCALAR:
 600                 param->immconst_type = WINED3D_IMMCONST_SCALAR;
 601                 memcpy(param->immconst_data, *ptr, 1 * sizeof(DWORD));
 602                 *ptr += 1;
 603                 break;
 604
 605             case WINED3D_SM4_IMMCONST_VEC4:
 606                 param->immconst_type = WINED3D_IMMCONST_VEC4;
 607                 memcpy(param->immconst_data, *ptr, 4 * sizeof(DWORD));
 608                 *ptr += 4;
 609                 break;
 610
 611             default:
 612                 FIXME("Unhandled immediate constant type %#x.\n", immconst_type);
 613                 break;
 614         }
 615     }
 616
 617     map_register(priv, param);
 618
 619     return TRUE;
 620 }
 621
 622 static BOOL shader_sm4_read_src_param(struct wined3d_sm4_data *priv, const DWORD **ptr,
 623         enum wined3d_data_type data_type, struct wined3d_shader_src_param *src_param)
 624 {
 625     DWORD token = **ptr;
 626
 627     if (!shader_sm4_read_param(priv, ptr, data_type, &src_param->reg, &src_param->modifiers))
 628     {
 629         ERR("Failed to read parameter.\n");
 630         return FALSE;
 631     }
 632
 633     if (src_param->reg.type == WINED3DSPR_IMMCONST)
 634     {
 635         src_param->swizzle = WINED3DSP_NOSWIZZLE;
 636     }
 637     else
 638     {
 639         enum wined3d_sm4_swizzle_type swizzle_type =
 640                 (token & WINED3D_SM4_SWIZZLE_TYPE_MASK) >> WINED3D_SM4_SWIZZLE_TYPE_SHIFT;
 641
 642         switch (swizzle_type)
 643         {
 644             case WINED3D_SM4_SWIZZLE_SCALAR:
 645                 src_param->swizzle = (token & WINED3D_SM4_SWIZZLE_MASK) >> WINED3D_SM4_SWIZZLE_SHIFT;
 646                 src_param->swizzle = (src_param->swizzle & 0x3) * 0x55;
 647                 break;
 648
 649             case WINED3D_SM4_SWIZZLE_VEC4:
 650                 src_param->swizzle = (token & WINED3D_SM4_SWIZZLE_MASK) >> WINED3D_SM4_SWIZZLE_SHIFT;
 651                 break;
 652
 653             default:
 654                 FIXME("Unhandled swizzle type %#x.\n", swizzle_type);
 655                 break;
 656         }
 657     }
 658
 659     return TRUE;
 660 }
 661
 662 static BOOL shader_sm4_read_dst_param(struct wined3d_sm4_data *priv, const DWORD **ptr,
 663         enum wined3d_data_type data_type, struct wined3d_shader_dst_param *dst_param)
 664 {
 665     enum wined3d_shader_src_modifier modifier;
 666     DWORD token = **ptr;
 667
 668     if (!shader_sm4_read_param(priv, ptr, data_type, &dst_param->reg, &modifier))
 669     {
 670         ERR("Failed to read parameter.\n");
 671         return FALSE;
 672     }
 673
 674     if (modifier != WINED3DSPSM_NONE)
 675     {
 676         ERR("Invalid source modifier %#x on destination register.\n", modifier);
 677         return FALSE;
 678     }
 679
 680     dst_param->write_mask = (token & WINED3D_SM4_WRITEMASK_MASK) >> WINED3D_SM4_WRITEMASK_SHIFT;
 681     dst_param->modifiers = 0;
 682     dst_param->shift = 0;
 683
 684     return TRUE;
 685 }
 686
 687 static void shader_sm4_read_instruction(void *data, const DWORD **ptr, struct wined3d_shader_instruction *ins)
 688 {
 689     const struct wined3d_sm4_opcode_info *opcode_info;
 690     struct wined3d_sm4_data *priv = data;
 691     DWORD opcode_token, opcode;
 692     const DWORD *p;
 693     UINT i, len;
 694
 695     list_move_head(&priv->src_free, &priv->src);
 696
 697     opcode_token = *(*ptr)++;
 698     opcode = opcode_token & WINED3D_SM4_OPCODE_MASK;
 699     len = ((opcode_token & WINED3D_SM4_INSTRUCTION_LENGTH_MASK) >> WINED3D_SM4_INSTRUCTION_LENGTH_SHIFT) - 1;
 700
 701     if (TRACE_ON(d3d_bytecode))
 702     {
 703         TRACE_(d3d_bytecode)("[ %08x ", opcode_token);
 704         for (i = 0; i < len; ++i)
 705         {
 706             TRACE_(d3d_bytecode)("%08x ", (*ptr)[i]);
 707         }
 708         TRACE_(d3d_bytecode)("]\n");
 709     }
 710
 711     if (!(opcode_info = get_opcode_info(opcode)))
 712     {
 713         FIXME("Unrecognized opcode %#x, opcode_token 0x%08x.\n", opcode, opcode_token);
 714         ins->handler_idx = WINED3DSIH_TABLE_SIZE;
 715         *ptr += len;
 716         return;
 717     }
 718
 719     ins->handler_idx = opcode_info->handler_idx;
 720     ins->flags = 0;
 721     ins->coissue = 0;
 722     ins->predicate = NULL;
 723     ins->dst_count = strlen(opcode_info->dst_info);
 724     ins->dst = priv->dst_param;
 725     ins->src_count = strlen(opcode_info->src_info);
 726     ins->src = priv->src_param;
 727
 728     p = *ptr;
 729     *ptr += len;
 730
 731     if (opcode_token & WINED3D_SM4_INSTRUCTION_MODIFIER)
 732     {
 733         DWORD modifier = *p++;
 734         FIXME("Skipping modifier 0x%08x.\n", modifier);
 735     }
 736
 737     if (opcode == WINED3D_SM4_OP_DCL_CONSTANT_BUFFER)
 738     {
 739         shader_sm4_read_src_param(priv, &p, WINED3D_DATA_FLOAT, &ins->declaration.src);
 740         if (opcode_token & WINED3D_SM4_INDEX_TYPE_MASK)
 741             ins->flags |= WINED3DSI_INDEXED_DYNAMIC;
 742     }
 743     else if (opcode == WINED3D_SM4_OP_DCL_OUTPUT_TOPOLOGY)
 744     {
 745         enum wined3d_sm4_output_primitive_type primitive_type;
 746
 747         primitive_type = (opcode_token & WINED3D_SM4_PRIMITIVE_TYPE_MASK) >> WINED3D_SM4_PRIMITIVE_TYPE_SHIFT;
 748         if (primitive_type >= sizeof(output_primitive_type_table) / sizeof(*output_primitive_type_table))
 749         {
 750             FIXME("Unhandled output primitive type %#x.\n", primitive_type);
 751             ins->declaration.primitive_type = WINED3D_PT_UNDEFINED;
 752         }
 753         else
 754         {
 755             ins->declaration.primitive_type = output_primitive_type_table[primitive_type];
 756         }
 757     }
 758     else if (opcode == WINED3D_SM4_OP_DCL_INPUT_PRIMITIVE)
 759     {
 760         enum wined3d_sm4_input_primitive_type primitive_type;
 761
 762         primitive_type = (opcode_token & WINED3D_SM4_PRIMITIVE_TYPE_MASK) >> WINED3D_SM4_PRIMITIVE_TYPE_SHIFT;
 763         if (primitive_type >= sizeof(input_primitive_type_table) / sizeof(*input_primitive_type_table))
 764         {
 765             FIXME("Unhandled input primitive type %#x.\n", primitive_type);
 766             ins->declaration.primitive_type = WINED3D_PT_UNDEFINED;
 767         }
 768         else
 769         {
 770             ins->declaration.primitive_type = input_primitive_type_table[primitive_type];
 771         }
 772     }
 773     else if (opcode == WINED3D_SM4_OP_DCL_VERTICES_OUT)
 774     {
 775         ins->declaration.count = *p++;
 776     }
 777     else
 778     {
 779         for (i = 0; i < ins->dst_count; ++i)
 780         {
 781             if (!(shader_sm4_read_dst_param(priv, &p, map_data_type(opcode_info->dst_info[i]), &priv->dst_param[i])))
 782             {
 783                 ins->handler_idx = WINED3DSIH_TABLE_SIZE;
 784                 return;
 785             }
 786         }
 787
 788         for (i = 0; i < ins->src_count; ++i)
 789         {
 790             if (!(shader_sm4_read_src_param(priv, &p, map_data_type(opcode_info->src_info[i]), &priv->src_param[i])))
 791             {
 792                 ins->handler_idx = WINED3DSIH_TABLE_SIZE;
 793                 return;
 794             }
 795         }
 796     }
 797 }
 798
 799 static BOOL shader_sm4_is_end(void *data, const DWORD **ptr)
 800 {
 801     struct wined3d_sm4_data *priv = data;
 802     return *ptr == priv->end;
 803 }
 804
 805 const struct wined3d_shader_frontend sm4_shader_frontend =
 806 {
 807     shader_sm4_init,
 808     shader_sm4_free,
 809     shader_sm4_read_header,
 810     shader_sm4_read_instruction,
 811     shader_sm4_is_end,
 812 };