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
 180     struct
 181     {
 182         enum wined3d_shader_register_type register_type;
 183         UINT register_idx;
 184     } output_map[MAX_REG_OUTPUT];
 185
 186     struct wined3d_shader_src_param src_param[5];
 187     struct wined3d_shader_dst_param dst_param[2];
 188     struct list src_free;
 189     struct list src;
 190 };
 191
 192 struct wined3d_sm4_opcode_info
 193 {
 194     enum wined3d_sm4_opcode opcode;
 195     enum WINED3D_SHADER_INSTRUCTION_HANDLER handler_idx;
 196     const char *dst_info;
 197     const char *src_info;
 198 };
 199
 200 struct sysval_map
 201 {
 202     enum wined3d_sysval_semantic sysval;
 203     enum wined3d_shader_register_type register_type;
 204     UINT register_idx;
 205 };
 206
 207 /*
 208  * F -> WINED3D_DATA_FLOAT
 209  * I -> WINED3D_DATA_INT
 210  * R -> WINED3D_DATA_RESOURCE
 211  * S -> WINED3D_DATA_SAMPLER
 212  * U -> WINED3D_DATA_UINT
 213  */
 214 static const struct wined3d_sm4_opcode_info opcode_table[] =
 215 {
 216     {WINED3D_SM4_OP_ADD,                    WINED3DSIH_ADD,                 "F",    "FF"},
 217     {WINED3D_SM4_OP_AND,                    WINED3DSIH_AND,                 "U",    "UU"},
 218     {WINED3D_SM4_OP_BREAK,                  WINED3DSIH_BREAK,               "",     ""},
 219     {WINED3D_SM4_OP_BREAKC,                 WINED3DSIH_BREAKP,              "",     "U"},
 220     {WINED3D_SM4_OP_CUT,                    WINED3DSIH_CUT,                 "",     ""},
 221     {WINED3D_SM4_OP_DERIV_RTX,              WINED3DSIH_DSX,                 "F",    "F"},
 222     {WINED3D_SM4_OP_DERIV_RTY,              WINED3DSIH_DSY,                 "F",    "F"},
 223     {WINED3D_SM4_OP_DISCARD,                WINED3DSIH_TEXKILL,             "",     "U"},
 224     {WINED3D_SM4_OP_DIV,                    WINED3DSIH_DIV,                 "F",    "FF"},
 225     {WINED3D_SM4_OP_DP2,                    WINED3DSIH_DP2,                 "F",    "FF"},
 226     {WINED3D_SM4_OP_DP3,                    WINED3DSIH_DP3,                 "F",    "FF"},
 227     {WINED3D_SM4_OP_DP4,                    WINED3DSIH_DP4,                 "F",    "FF"},
 228     {WINED3D_SM4_OP_EMIT,                   WINED3DSIH_EMIT,                "",     ""},
 229     {WINED3D_SM4_OP_ENDIF,                  WINED3DSIH_ENDIF,               "",     ""},
 230     {WINED3D_SM4_OP_ENDLOOP,                WINED3DSIH_ENDLOOP,             "",     ""},
 231     {WINED3D_SM4_OP_EQ,                     WINED3DSIH_EQ,                  "U",    "FF"},
 232     {WINED3D_SM4_OP_EXP,                    WINED3DSIH_EXP,                 "F",    "F"},
 233     {WINED3D_SM4_OP_FRC,                    WINED3DSIH_FRC,                 "F",    "F"},
 234     {WINED3D_SM4_OP_FTOI,                   WINED3DSIH_FTOI,                "I",    "F"},
 235     {WINED3D_SM4_OP_GE,                     WINED3DSIH_GE,                  "U",    "FF"},
 236     {WINED3D_SM4_OP_IADD,                   WINED3DSIH_IADD,                "I",    "II"},
 237     {WINED3D_SM4_OP_IF,                     WINED3DSIH_IF,                  "",     "U"},
 238     {WINED3D_SM4_OP_IEQ,                    WINED3DSIH_IEQ,                 "U",    "II"},
 239     {WINED3D_SM4_OP_IGE,                    WINED3DSIH_IGE,                 "U",    "II"},
 240     {WINED3D_SM4_OP_IMUL,                   WINED3DSIH_IMUL,                "II",   "II"},
 241     {WINED3D_SM4_OP_ISHL,                   WINED3DSIH_ISHL,                "I",    "II"},
 242     {WINED3D_SM4_OP_ITOF,                   WINED3DSIH_ITOF,                "F",    "I"},
 243     {WINED3D_SM4_OP_LD,                     WINED3DSIH_LD,                  "U",    "FR"},
 244     {WINED3D_SM4_OP_LOG,                    WINED3DSIH_LOG,                 "F",    "F"},
 245     {WINED3D_SM4_OP_LOOP,                   WINED3DSIH_LOOP,                "",     ""},
 246     {WINED3D_SM4_OP_LT,                     WINED3DSIH_LT,                  "U",    "FF"},
 247     {WINED3D_SM4_OP_MAD,                    WINED3DSIH_MAD,                 "F",    "FFF"},
 248     {WINED3D_SM4_OP_MIN,                    WINED3DSIH_MIN,                 "F",    "FF"},
 249     {WINED3D_SM4_OP_MAX,                    WINED3DSIH_MAX,                 "F",    "FF"},
 250     {WINED3D_SM4_OP_MOV,                    WINED3DSIH_MOV,                 "F",    "F"},
 251     {WINED3D_SM4_OP_MOVC,                   WINED3DSIH_MOVC,                "F",    "UFF"},
 252     {WINED3D_SM4_OP_MUL,                    WINED3DSIH_MUL,                 "F",    "FF"},
 253     {WINED3D_SM4_OP_RET,                    WINED3DSIH_RET,                 "",     ""},
 254     {WINED3D_SM4_OP_ROUND_NI,               WINED3DSIH_ROUND_NI,            "F",    "F"},
 255     {WINED3D_SM4_OP_RSQ,                    WINED3DSIH_RSQ,                 "F",    "F"},
 256     {WINED3D_SM4_OP_SAMPLE,                 WINED3DSIH_SAMPLE,              "U",    "FRS"},
 257     {WINED3D_SM4_OP_SAMPLE_LOD,             WINED3DSIH_SAMPLE_LOD,          "U",    "FRSF"},
 258     {WINED3D_SM4_OP_SAMPLE_GRAD,            WINED3DSIH_SAMPLE_GRAD,         "U",    "FRSFF"},
 259     {WINED3D_SM4_OP_SQRT,                   WINED3DSIH_SQRT,                "F",    "F"},
 260     {WINED3D_SM4_OP_SINCOS,                 WINED3DSIH_SINCOS,              "FF",   "F"},
 261     {WINED3D_SM4_OP_UDIV,                   WINED3DSIH_UDIV,                "UU",   "UU"},
 262     {WINED3D_SM4_OP_USHR,                   WINED3DSIH_USHR,                "U",    "UU"},
 263     {WINED3D_SM4_OP_UTOF,                   WINED3DSIH_UTOF,                "F",    "U"},
 264     {WINED3D_SM4_OP_XOR,                    WINED3DSIH_XOR,                 "U",    "UU"},
 265     {WINED3D_SM4_OP_DCL_CONSTANT_BUFFER,    WINED3DSIH_DCL_CONSTANT_BUFFER, "",     ""},
 266     {WINED3D_SM4_OP_DCL_OUTPUT_TOPOLOGY,    WINED3DSIH_DCL_OUTPUT_TOPOLOGY, "",     ""},
 267     {WINED3D_SM4_OP_DCL_INPUT_PRIMITIVE,    WINED3DSIH_DCL_INPUT_PRIMITIVE, "",     ""},
 268     {WINED3D_SM4_OP_DCL_VERTICES_OUT,       WINED3DSIH_DCL_VERTICES_OUT,    "",     ""},
 269 };
 270
 271 static const enum wined3d_shader_register_type register_type_table[] =
 272 {
 273     /* WINED3D_SM4_RT_TEMP */           WINED3DSPR_TEMP,
 274     /* WINED3D_SM4_RT_INPUT */          WINED3DSPR_INPUT,
 275     /* WINED3D_SM4_RT_OUTPUT */         WINED3DSPR_OUTPUT,
 276     /* UNKNOWN */                       0,
 277     /* WINED3D_SM4_RT_IMMCONST */       WINED3DSPR_IMMCONST,
 278     /* UNKNOWN */                       0,
 279     /* WINED3D_SM4_RT_SAMPLER */        WINED3DSPR_SAMPLER,
 280     /* WINED3D_SM4_RT_RESOURCE */       WINED3DSPR_RESOURCE,
 281     /* WINED3D_SM4_RT_CONSTBUFFER */    WINED3DSPR_CONSTBUFFER,
 282     /* UNKNOWN */                       0,
 283     /* UNKNOWN */                       0,
 284     /* WINED3D_SM4_RT_PRIMID */         WINED3DSPR_PRIMID,
 285     /* UNKNOWN */                       0,
 286     /* WINED3D_SM4_RT_NULL */           WINED3DSPR_NULL,
 287 };
 288
 289 static const enum wined3d_primitive_type output_primitive_type_table[] =
 290 {
 291     /* UNKNOWN */                               WINED3D_PT_UNDEFINED,
 292     /* WINED3D_SM4_OUTPUT_PT_POINTLIST */       WINED3D_PT_POINTLIST,
 293     /* UNKNOWN */                               WINED3D_PT_UNDEFINED,
 294     /* WINED3D_SM4_OUTPUT_PT_LINELIST */        WINED3D_PT_LINELIST,
 295     /* UNKNOWN */                               WINED3D_PT_UNDEFINED,
 296     /* WINED3D_SM4_OUTPUT_PT_TRIANGLESTRIP */   WINED3D_PT_TRIANGLESTRIP,
 297 };
 298
 299 static const enum wined3d_primitive_type input_primitive_type_table[] =
 300 {
 301     /* UNKNOWN */                               WINED3D_PT_UNDEFINED,
 302     /* WINED3D_SM4_INPUT_PT_POINT */            WINED3D_PT_POINTLIST,
 303     /* WINED3D_SM4_INPUT_PT_LINE */             WINED3D_PT_LINELIST,
 304     /* WINED3D_SM4_INPUT_PT_TRIANGLE */         WINED3D_PT_TRIANGLELIST,
 305     /* UNKNOWN */                               WINED3D_PT_UNDEFINED,
 306     /* UNKNOWN */                               WINED3D_PT_UNDEFINED,
 307     /* WINED3D_SM4_INPUT_PT_LINEADJ */          WINED3D_PT_LINELIST_ADJ,
 308     /* WINED3D_SM4_INPUT_PT_TRIANGLEADJ */      WINED3D_PT_TRIANGLELIST_ADJ,
 309 };
 310
 311 static const struct sysval_map sysval_map[] =
 312 {
 313     {WINED3D_SV_DEPTH,      WINED3DSPR_DEPTHOUT,    0},
 314     {WINED3D_SV_TARGET0,    WINED3DSPR_COLOROUT,    0},
 315     {WINED3D_SV_TARGET1,    WINED3DSPR_COLOROUT,    1},
 316     {WINED3D_SV_TARGET2,    WINED3DSPR_COLOROUT,    2},
 317     {WINED3D_SV_TARGET3,    WINED3DSPR_COLOROUT,    3},
 318     {WINED3D_SV_TARGET4,    WINED3DSPR_COLOROUT,    4},
 319     {WINED3D_SV_TARGET5,    WINED3DSPR_COLOROUT,    5},
 320     {WINED3D_SV_TARGET6,    WINED3DSPR_COLOROUT,    6},
 321     {WINED3D_SV_TARGET7,    WINED3DSPR_COLOROUT,    7},
 322 };
 323
 324 static BOOL shader_sm4_read_src_param(struct wined3d_sm4_data *priv, const DWORD **ptr,
 325         enum wined3d_data_type data_type, struct wined3d_shader_src_param *src_param);
 326
 327 static const struct wined3d_sm4_opcode_info *get_opcode_info(enum wined3d_sm4_opcode opcode)
 328 {
 329     unsigned int i;
 330
 331     for (i = 0; i < sizeof(opcode_table) / sizeof(*opcode_table); ++i)
 332     {
 333         if (opcode == opcode_table[i].opcode) return &opcode_table[i];
 334     }
 335
 336     return NULL;
 337 }
 338
 339 static void map_register(const struct wined3d_sm4_data *priv, struct wined3d_shader_register *reg)
 340 {
 341     switch (priv->shader_version.type)
 342     {
 343         case WINED3D_SHADER_TYPE_PIXEL:
 344             if (reg->type == WINED3DSPR_OUTPUT)
 345             {
 346                 unsigned int reg_idx = reg->idx[0].offset;
 347
 348                 if (reg_idx >= ARRAY_SIZE(priv->output_map))
 349                 {
 350                     ERR("Invalid output index %u.\n", reg_idx);
 351                     break;
 352                 }
 353
 354                 reg->type = priv->output_map[reg_idx].register_type;
 355                 reg->idx[0].offset = priv->output_map[reg_idx].register_idx;
 356             }
 357             break;
 358
 359         default:
 360             break;
 361     }
 362 }
 363
 364 static enum wined3d_data_type map_data_type(char t)
 365 {
 366     switch (t)
 367     {
 368         case 'F':
 369             return WINED3D_DATA_FLOAT;
 370         case 'I':
 371             return WINED3D_DATA_INT;
 372         case 'R':
 373             return WINED3D_DATA_RESOURCE;
 374         case 'S':
 375             return WINED3D_DATA_SAMPLER;
 376         case 'U':
 377             return WINED3D_DATA_UINT;
 378         default:
 379             ERR("Invalid data type '%c'.\n", t);
 380             return WINED3D_DATA_FLOAT;
 381     }
 382 }
 383
 384 static void *shader_sm4_init(const DWORD *byte_code, const struct wined3d_shader_signature *output_signature)
 385 {
 386     struct wined3d_sm4_data *priv;
 387     unsigned int i, j;
 388
 389     if (!(priv = HeapAlloc(GetProcessHeap(), 0, sizeof(*priv))))
 390     {
 391         ERR("Failed to allocate private data\n");
 392         return NULL;
 393     }
 394
 395     memset(priv->output_map, 0xff, sizeof(priv->output_map));
 396     for (i = 0; i < output_signature->element_count; ++i)
 397     {
 398         struct wined3d_shader_signature_element *e = &output_signature->elements[i];
 399
 400         if (e->register_idx >= ARRAY_SIZE(priv->output_map))
 401         {
 402             WARN("Invalid output index %u.\n", e->register_idx);
 403             continue;
 404         }
 405
 406         for (j = 0; j < ARRAY_SIZE(sysval_map); ++j)
 407         {
 408             if (e->sysval_semantic == sysval_map[j].sysval)
 409             {
 410                 priv->output_map[e->register_idx].register_type = sysval_map[j].register_type;
 411                 priv->output_map[e->register_idx].register_idx = sysval_map[j].register_idx;
 412                 break;
 413             }
 414         }
 415     }
 416
 417     list_init(&priv->src_free);
 418     list_init(&priv->src);
 419
 420     return priv;
 421 }
 422
 423 static void shader_sm4_free(void *data)
 424 {
 425     struct wined3d_shader_src_param_entry *e1, *e2;
 426     struct wined3d_sm4_data *priv = data;
 427
 428     list_move_head(&priv->src_free, &priv->src);
 429     LIST_FOR_EACH_ENTRY_SAFE(e1, e2, &priv->src_free, struct wined3d_shader_src_param_entry, entry)
 430     {
 431         HeapFree(GetProcessHeap(), 0, e1);
 432     }
 433     HeapFree(GetProcessHeap(), 0, priv);
 434 }
 435
 436 static struct wined3d_shader_src_param *get_src_param(struct wined3d_sm4_data *priv)
 437 {
 438     struct wined3d_shader_src_param_entry *e;
 439     struct list *elem;
 440
 441     if (!list_empty(&priv->src_free))
 442     {
 443         elem = list_head(&priv->src_free);
 444         list_remove(elem);
 445     }
 446     else
 447     {
 448         if (!(e = HeapAlloc(GetProcessHeap(), 0, sizeof(*e))))
 449             return NULL;
 450         elem = &e->entry;
 451     }
 452
 453     list_add_tail(&priv->src, elem);
 454     e = LIST_ENTRY(elem, struct wined3d_shader_src_param_entry, entry);
 455     return &e->param;
 456 }
 457
 458 static void shader_sm4_read_header(void *data, const DWORD **ptr, struct wined3d_shader_version *shader_version)
 459 {
 460     struct wined3d_sm4_data *priv = data;
 461     DWORD version_token;
 462
 463     priv->end = *ptr;
 464
 465     version_token = *(*ptr)++;
 466     TRACE("version: 0x%08x\n", version_token);
 467
 468     TRACE("token count: %u\n", **ptr);
 469     priv->end += *(*ptr)++;
 470
 471     switch (version_token >> 16)
 472     {
 473         case WINED3D_SM4_PS:
 474             priv->shader_version.type = WINED3D_SHADER_TYPE_PIXEL;
 475             break;
 476
 477         case WINED3D_SM4_VS:
 478             priv->shader_version.type = WINED3D_SHADER_TYPE_VERTEX;
 479             break;
 480
 481         case WINED3D_SM4_GS:
 482             priv->shader_version.type = WINED3D_SHADER_TYPE_GEOMETRY;
 483             break;
 484
 485         default:
 486             FIXME("Unrecognized shader type %#x\n", version_token >> 16);
 487     }
 488     priv->shader_version.major = WINED3D_SM4_VERSION_MAJOR(version_token);
 489     priv->shader_version.minor = WINED3D_SM4_VERSION_MINOR(version_token);
 490
 491     *shader_version = priv->shader_version;
 492 }
 493
 494 static BOOL shader_sm4_read_reg_idx(struct wined3d_sm4_data *priv, const DWORD **ptr,
 495         DWORD addressing, struct wined3d_shader_register_index *reg_idx)
 496 {
 497     if (addressing & WINED3D_SM4_ADDRESSING_RELATIVE)
 498     {
 499         struct wined3d_shader_src_param *rel_addr = get_src_param(priv);
 500
 501         if (!(reg_idx->rel_addr = rel_addr))
 502         {
 503             ERR("Failed to get src param for relative addressing.\n");
 504             return FALSE;
 505         }
 506
 507         if (addressing & WINED3D_SM4_ADDRESSING_OFFSET)
 508             reg_idx->offset = *(*ptr)++;
 509         else
 510             reg_idx->offset = 0;
 511         shader_sm4_read_src_param(priv, ptr, WINED3D_DATA_INT, rel_addr);
 512     }
 513     else
 514     {
 515         reg_idx->rel_addr = NULL;
 516         reg_idx->offset = *(*ptr)++;
 517     }
 518
 519     return TRUE;
 520 }
 521
 522 static BOOL shader_sm4_read_param(struct wined3d_sm4_data *priv, const DWORD **ptr,
 523         enum wined3d_data_type data_type, struct wined3d_shader_register *param,
 524         enum wined3d_shader_src_modifier *modifier)
 525 {
 526     enum wined3d_sm4_register_type register_type;
 527     DWORD token = *(*ptr)++;
 528     DWORD order;
 529
 530     register_type = (token & WINED3D_SM4_REGISTER_TYPE_MASK) >> WINED3D_SM4_REGISTER_TYPE_SHIFT;
 531     if (register_type >= sizeof(register_type_table) / sizeof(*register_type_table))
 532     {
 533         FIXME("Unhandled register type %#x.\n", register_type);
 534         param->type = WINED3DSPR_TEMP;
 535     }
 536     else
 537     {
 538         param->type = register_type_table[register_type];
 539     }
 540     param->data_type = data_type;
 541
 542     if (token & WINED3D_SM4_REGISTER_MODIFIER)
 543     {
 544         DWORD m = *(*ptr)++;
 545
 546         /* FIXME: This will probably break down at some point. The SM4
 547          * modifiers look like flags, while wined3d currently has an enum
 548          * with possible combinations, e.g. WINED3DSPSM_ABSNEG. */
 549         switch (m)
 550         {
 551             case 0x41:
 552                 *modifier = WINED3DSPSM_NEG;
 553                 break;
 554
 555             case 0x81:
 556                 *modifier = WINED3DSPSM_ABS;
 557                 break;
 558
 559             default:
 560                 FIXME("Skipping modifier 0x%08x.\n", m);
 561                 *modifier = WINED3DSPSM_NONE;
 562                 break;
 563         }
 564     }
 565     else
 566     {
 567         *modifier = WINED3DSPSM_NONE;
 568     }
 569
 570     order = (token & WINED3D_SM4_REGISTER_ORDER_MASK) >> WINED3D_SM4_REGISTER_ORDER_SHIFT;
 571
 572     if (order < 1)
 573         param->idx[0].offset = ~0U;
 574     else
 575     {
 576         DWORD addressing = (token & WINED3D_SM4_ADDRESSING_MASK0) >> WINED3D_SM4_ADDRESSING_SHIFT0;
 577         if (!(shader_sm4_read_reg_idx(priv, ptr, addressing, &param->idx[0])))
 578         {
 579             ERR("Failed to read register index.\n");
 580             return FALSE;
 581         }
 582     }
 583
 584     if (order < 2)
 585         param->idx[1].offset = ~0U;
 586     else
 587     {
 588         DWORD addressing = (token & WINED3D_SM4_ADDRESSING_MASK1) >> WINED3D_SM4_ADDRESSING_SHIFT1;
 589         if (!(shader_sm4_read_reg_idx(priv, ptr, addressing, &param->idx[1])))
 590         {
 591             ERR("Failed to read register index.\n");
 592             return FALSE;
 593         }
 594     }
 595
 596     if (order > 2)
 597         FIXME("Unhandled order %u.\n", order);
 598
 599     if (register_type == WINED3D_SM4_RT_IMMCONST)
 600     {
 601         enum wined3d_sm4_immconst_type immconst_type =
 602                 (token & WINED3D_SM4_IMMCONST_TYPE_MASK) >> WINED3D_SM4_IMMCONST_TYPE_SHIFT;
 603
 604         switch (immconst_type)
 605         {
 606             case WINED3D_SM4_IMMCONST_SCALAR:
 607                 param->immconst_type = WINED3D_IMMCONST_SCALAR;
 608                 memcpy(param->immconst_data, *ptr, 1 * sizeof(DWORD));
 609                 *ptr += 1;
 610                 break;
 611
 612             case WINED3D_SM4_IMMCONST_VEC4:
 613                 param->immconst_type = WINED3D_IMMCONST_VEC4;
 614                 memcpy(param->immconst_data, *ptr, 4 * sizeof(DWORD));
 615                 *ptr += 4;
 616                 break;
 617
 618             default:
 619                 FIXME("Unhandled immediate constant type %#x.\n", immconst_type);
 620                 break;
 621         }
 622     }
 623
 624     map_register(priv, param);
 625
 626     return TRUE;
 627 }
 628
 629 static BOOL shader_sm4_read_src_param(struct wined3d_sm4_data *priv, const DWORD **ptr,
 630         enum wined3d_data_type data_type, struct wined3d_shader_src_param *src_param)
 631 {
 632     DWORD token = **ptr;
 633
 634     if (!shader_sm4_read_param(priv, ptr, data_type, &src_param->reg, &src_param->modifiers))
 635     {
 636         ERR("Failed to read parameter.\n");
 637         return FALSE;
 638     }
 639
 640     if (src_param->reg.type == WINED3DSPR_IMMCONST)
 641     {
 642         src_param->swizzle = WINED3DSP_NOSWIZZLE;
 643     }
 644     else
 645     {
 646         enum wined3d_sm4_swizzle_type swizzle_type =
 647                 (token & WINED3D_SM4_SWIZZLE_TYPE_MASK) >> WINED3D_SM4_SWIZZLE_TYPE_SHIFT;
 648
 649         switch (swizzle_type)
 650         {
 651             case WINED3D_SM4_SWIZZLE_SCALAR:
 652                 src_param->swizzle = (token & WINED3D_SM4_SWIZZLE_MASK) >> WINED3D_SM4_SWIZZLE_SHIFT;
 653                 src_param->swizzle = (src_param->swizzle & 0x3) * 0x55;
 654                 break;
 655
 656             case WINED3D_SM4_SWIZZLE_VEC4:
 657                 src_param->swizzle = (token & WINED3D_SM4_SWIZZLE_MASK) >> WINED3D_SM4_SWIZZLE_SHIFT;
 658                 break;
 659
 660             default:
 661                 FIXME("Unhandled swizzle type %#x.\n", swizzle_type);
 662                 break;
 663         }
 664     }
 665
 666     return TRUE;
 667 }
 668
 669 static BOOL shader_sm4_read_dst_param(struct wined3d_sm4_data *priv, const DWORD **ptr,
 670         enum wined3d_data_type data_type, struct wined3d_shader_dst_param *dst_param)
 671 {
 672     enum wined3d_shader_src_modifier modifier;
 673     DWORD token = **ptr;
 674
 675     if (!shader_sm4_read_param(priv, ptr, data_type, &dst_param->reg, &modifier))
 676     {
 677         ERR("Failed to read parameter.\n");
 678         return FALSE;
 679     }
 680
 681     if (modifier != WINED3DSPSM_NONE)
 682     {
 683         ERR("Invalid source modifier %#x on destination register.\n", modifier);
 684         return FALSE;
 685     }
 686
 687     dst_param->write_mask = (token & WINED3D_SM4_WRITEMASK_MASK) >> WINED3D_SM4_WRITEMASK_SHIFT;
 688     dst_param->modifiers = 0;
 689     dst_param->shift = 0;
 690
 691     return TRUE;
 692 }
 693
 694 static void shader_sm4_read_instruction(void *data, const DWORD **ptr, struct wined3d_shader_instruction *ins)
 695 {
 696     const struct wined3d_sm4_opcode_info *opcode_info;
 697     struct wined3d_sm4_data *priv = data;
 698     DWORD opcode_token, opcode;
 699     const DWORD *p;
 700     UINT i, len;
 701
 702     list_move_head(&priv->src_free, &priv->src);
 703
 704     opcode_token = *(*ptr)++;
 705     opcode = opcode_token & WINED3D_SM4_OPCODE_MASK;
 706     len = ((opcode_token & WINED3D_SM4_INSTRUCTION_LENGTH_MASK) >> WINED3D_SM4_INSTRUCTION_LENGTH_SHIFT) - 1;
 707
 708     if (TRACE_ON(d3d_bytecode))
 709     {
 710         TRACE_(d3d_bytecode)("[ %08x ", opcode_token);
 711         for (i = 0; i < len; ++i)
 712         {
 713             TRACE_(d3d_bytecode)("%08x ", (*ptr)[i]);
 714         }
 715         TRACE_(d3d_bytecode)("]\n");
 716     }
 717
 718     if (!(opcode_info = get_opcode_info(opcode)))
 719     {
 720         FIXME("Unrecognized opcode %#x, opcode_token 0x%08x.\n", opcode, opcode_token);
 721         ins->handler_idx = WINED3DSIH_TABLE_SIZE;
 722         *ptr += len;
 723         return;
 724     }
 725
 726     ins->handler_idx = opcode_info->handler_idx;
 727     ins->flags = 0;
 728     ins->coissue = 0;
 729     ins->predicate = NULL;
 730     ins->dst_count = strlen(opcode_info->dst_info);
 731     ins->dst = priv->dst_param;
 732     ins->src_count = strlen(opcode_info->src_info);
 733     ins->src = priv->src_param;
 734
 735     p = *ptr;
 736     *ptr += len;
 737
 738     if (opcode_token & WINED3D_SM4_INSTRUCTION_MODIFIER)
 739     {
 740         DWORD modifier = *p++;
 741         FIXME("Skipping modifier 0x%08x.\n", modifier);
 742     }
 743
 744     if (opcode == WINED3D_SM4_OP_DCL_CONSTANT_BUFFER)
 745     {
 746         shader_sm4_read_src_param(priv, &p, WINED3D_DATA_FLOAT, &ins->declaration.src);
 747         if (opcode_token & WINED3D_SM4_INDEX_TYPE_MASK)
 748             ins->flags |= WINED3DSI_INDEXED_DYNAMIC;
 749     }
 750     else if (opcode == WINED3D_SM4_OP_DCL_OUTPUT_TOPOLOGY)
 751     {
 752         enum wined3d_sm4_output_primitive_type primitive_type;
 753
 754         primitive_type = (opcode_token & WINED3D_SM4_PRIMITIVE_TYPE_MASK) >> WINED3D_SM4_PRIMITIVE_TYPE_SHIFT;
 755         if (primitive_type >= sizeof(output_primitive_type_table) / sizeof(*output_primitive_type_table))
 756         {
 757             FIXME("Unhandled output primitive type %#x.\n", primitive_type);
 758             ins->declaration.primitive_type = WINED3D_PT_UNDEFINED;
 759         }
 760         else
 761         {
 762             ins->declaration.primitive_type = output_primitive_type_table[primitive_type];
 763         }
 764     }
 765     else if (opcode == WINED3D_SM4_OP_DCL_INPUT_PRIMITIVE)
 766     {
 767         enum wined3d_sm4_input_primitive_type primitive_type;
 768
 769         primitive_type = (opcode_token & WINED3D_SM4_PRIMITIVE_TYPE_MASK) >> WINED3D_SM4_PRIMITIVE_TYPE_SHIFT;
 770         if (primitive_type >= sizeof(input_primitive_type_table) / sizeof(*input_primitive_type_table))
 771         {
 772             FIXME("Unhandled input primitive type %#x.\n", primitive_type);
 773             ins->declaration.primitive_type = WINED3D_PT_UNDEFINED;
 774         }
 775         else
 776         {
 777             ins->declaration.primitive_type = input_primitive_type_table[primitive_type];
 778         }
 779     }
 780     else if (opcode == WINED3D_SM4_OP_DCL_VERTICES_OUT)
 781     {
 782         ins->declaration.count = *p++;
 783     }
 784     else
 785     {
 786         for (i = 0; i < ins->dst_count; ++i)
 787         {
 788             if (!(shader_sm4_read_dst_param(priv, &p, map_data_type(opcode_info->dst_info[i]), &priv->dst_param[i])))
 789             {
 790                 ins->handler_idx = WINED3DSIH_TABLE_SIZE;
 791                 return;
 792             }
 793         }
 794
 795         for (i = 0; i < ins->src_count; ++i)
 796         {
 797             if (!(shader_sm4_read_src_param(priv, &p, map_data_type(opcode_info->src_info[i]), &priv->src_param[i])))
 798             {
 799                 ins->handler_idx = WINED3DSIH_TABLE_SIZE;
 800                 return;
 801             }
 802         }
 803     }
 804 }
 805
 806 static BOOL shader_sm4_is_end(void *data, const DWORD **ptr)
 807 {
 808     struct wined3d_sm4_data *priv = data;
 809     return *ptr == priv->end;
 810 }
 811
 812 const struct wined3d_shader_frontend sm4_shader_frontend =
 813 {
 814     shader_sm4_init,
 815     shader_sm4_free,
 816     shader_sm4_read_header,
 817     shader_sm4_read_instruction,
 818     shader_sm4_is_end,
 819 };