src/compiler/nir/nir_intrinsics.h

   1 /*
   2  * Copyright © 2014 Intel Corporation
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  *
  23  * Authors:
  24  *    Connor Abbott (cwabbott0@gmail.com)
  25  *
  26  */
  27
  28 /**
  29  * This header file defines all the available intrinsics in one place. It
  30  * expands to a list of macros of the form:
  31  *
  32  * INTRINSIC(name, num_srcs, src_components, has_dest, dest_components,
  33  *              num_variables, num_indices, idx0, idx1, idx2, flags)
  34  *
  35  * Which should correspond one-to-one with the nir_intrinsic_info structure. It
  36  * is included in both ir.h to create the nir_intrinsic enum (with members of
  37  * the form nir_intrinsic_(name)) and and in opcodes.c to create
  38  * nir_intrinsic_infos, which is a const array of nir_intrinsic_info structures
  39  * for each intrinsic.
  40  */
  41
  42 #define ARR(...) { __VA_ARGS__ }
  43
  44 INTRINSIC(load_var, 0, ARR(0), true, 0, 1, 0, xx, xx, xx, NIR_INTRINSIC_CAN_ELIMINATE)
  45 INTRINSIC(store_var, 1, ARR(0), false, 0, 1, 1, WRMASK, xx, xx, 0)
  46 INTRINSIC(copy_var, 0, ARR(0), false, 0, 2, 0, xx, xx, xx, 0)
  47
  48 /*
  49  * Interpolation of input.  The interp_var_at* intrinsics are similar to the
  50  * load_var intrinsic acting on a shader input except that they interpolate
  51  * the input differently.  The at_sample and at_offset intrinsics take an
  52  * additional source that is an integer sample id or a vec2 position offset
  53  * respectively.
  54  */
  55
  56 INTRINSIC(interp_var_at_centroid, 0, ARR(0), true, 0, 1, 0, xx, xx, xx,
  57           NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
  58 INTRINSIC(interp_var_at_sample, 1, ARR(1), true, 0, 1, 0, xx, xx, xx,
  59           NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
  60 INTRINSIC(interp_var_at_offset, 1, ARR(2), true, 0, 1, 0, xx, xx, xx,
  61           NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
  62
  63 /*
  64  * Ask the driver for the size of a given buffer. It takes the buffer index
  65  * as source.
  66  */
  67 INTRINSIC(get_buffer_size, 1, ARR(1), true, 1, 0, 0, xx, xx, xx,
  68           NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
  69
  70 /*
  71  * a barrier is an intrinsic with no inputs/outputs but which can't be moved
  72  * around/optimized in general
  73  */
  74 #define BARRIER(name) INTRINSIC(name, 0, ARR(0), false, 0, 0, 0, xx, xx, xx, 0)
  75
  76 BARRIER(barrier)
  77 BARRIER(discard)
  78
  79 /*
  80  * Memory barrier with semantics analogous to the memoryBarrier() GLSL
  81  * intrinsic.
  82  */
  83 BARRIER(memory_barrier)
  84
  85 /*
  86  * Shader clock intrinsic with semantics analogous to the clock2x32ARB()
  87  * GLSL intrinsic.
  88  * The latter can be used as code motion barrier, which is currently not
  89  * feasible with NIR.
  90  */
  91 INTRINSIC(shader_clock, 0, ARR(0), true, 1, 0, 0, xx, xx, xx, NIR_INTRINSIC_CAN_ELIMINATE)
  92
  93 /*
  94  * Memory barrier with semantics analogous to the compute shader
  95  * groupMemoryBarrier(), memoryBarrierAtomicCounter(), memoryBarrierBuffer(),
  96  * memoryBarrierImage() and memoryBarrierShared() GLSL intrinsics.
  97  */
  98 BARRIER(group_memory_barrier)
  99 BARRIER(memory_barrier_atomic_counter)
 100 BARRIER(memory_barrier_buffer)
 101 BARRIER(memory_barrier_image)
 102 BARRIER(memory_barrier_shared)
 103
 104 /** A conditional discard, with a single boolean source. */
 105 INTRINSIC(discard_if, 1, ARR(1), false, 0, 0, 0, xx, xx, xx, 0)
 106
 107 /**
 108  * Basic Geometry Shader intrinsics.
 109  *
 110  * emit_vertex implements GLSL's EmitStreamVertex() built-in.  It takes a single
 111  * index, which is the stream ID to write to.
 112  *
 113  * end_primitive implements GLSL's EndPrimitive() built-in.
 114  */
 115 INTRINSIC(emit_vertex,   0, ARR(0), false, 0, 0, 1, STREAM_ID, xx, xx, 0)
 116 INTRINSIC(end_primitive, 0, ARR(0), false, 0, 0, 1, STREAM_ID, xx, xx, 0)
 117
 118 /**
 119  * Geometry Shader intrinsics with a vertex count.
 120  *
 121  * Alternatively, drivers may implement these intrinsics, and use
 122  * nir_lower_gs_intrinsics() to convert from the basic intrinsics.
 123  *
 124  * These maintain a count of the number of vertices emitted, as an additional
 125  * unsigned integer source.
 126  */
 127 INTRINSIC(emit_vertex_with_counter, 1, ARR(1), false, 0, 0, 1, STREAM_ID, xx, xx, 0)
 128 INTRINSIC(end_primitive_with_counter, 1, ARR(1), false, 0, 0, 1, STREAM_ID, xx, xx, 0)
 129 INTRINSIC(set_vertex_count, 1, ARR(1), false, 0, 0, 0, xx, xx, xx, 0)
 130
 131 /*
 132  * Atomic counters
 133  *
 134  * The *_var variants take an atomic_uint nir_variable, while the other,
 135  * lowered, variants take a constant buffer index and register offset.
 136  */
 137
 138 #define ATOMIC(name, flags) \
 139    INTRINSIC(atomic_counter_##name##_var, 0, ARR(0), true, 1, 1, 0, xx, xx, xx, flags) \
 140    INTRINSIC(atomic_counter_##name, 1, ARR(1), true, 1, 0, 1, BASE, xx, xx, flags)
 141
 142 ATOMIC(inc, 0)
 143 ATOMIC(dec, 0)
 144 ATOMIC(read, NIR_INTRINSIC_CAN_ELIMINATE)
 145
 146 /*
 147  * Image load, store and atomic intrinsics.
 148  *
 149  * All image intrinsics take an image target passed as a nir_variable.  Image
 150  * variables contain a number of memory and layout qualifiers that influence
 151  * the semantics of the intrinsic.
 152  *
 153  * All image intrinsics take a four-coordinate vector and a sample index as
 154  * first two sources, determining the location within the image that will be
 155  * accessed by the intrinsic.  Components not applicable to the image target
 156  * in use are undefined.  Image store takes an additional four-component
 157  * argument with the value to be written, and image atomic operations take
 158  * either one or two additional scalar arguments with the same meaning as in
 159  * the ARB_shader_image_load_store specification.
 160  */
 161 INTRINSIC(image_load, 2, ARR(4, 1), true, 4, 1, 0, xx, xx, xx,
 162           NIR_INTRINSIC_CAN_ELIMINATE)
 163 INTRINSIC(image_store, 3, ARR(4, 1, 4), false, 0, 1, 0, xx, xx, xx, 0)
 164 INTRINSIC(image_atomic_add, 3, ARR(4, 1, 1), true, 1, 1, 0, xx, xx, xx, 0)
 165 INTRINSIC(image_atomic_min, 3, ARR(4, 1, 1), true, 1, 1, 0, xx, xx, xx, 0)
 166 INTRINSIC(image_atomic_max, 3, ARR(4, 1, 1), true, 1, 1, 0, xx, xx, xx, 0)
 167 INTRINSIC(image_atomic_and, 3, ARR(4, 1, 1), true, 1, 1, 0, xx, xx, xx, 0)
 168 INTRINSIC(image_atomic_or, 3, ARR(4, 1, 1), true, 1, 1, 0, xx, xx, xx, 0)
 169 INTRINSIC(image_atomic_xor, 3, ARR(4, 1, 1), true, 1, 1, 0, xx, xx, xx, 0)
 170 INTRINSIC(image_atomic_exchange, 3, ARR(4, 1, 1), true, 1, 1, 0, xx, xx, xx, 0)
 171 INTRINSIC(image_atomic_comp_swap, 4, ARR(4, 1, 1, 1), true, 1, 1, 0, xx, xx, xx, 0)
 172 INTRINSIC(image_size, 0, ARR(0), true, 4, 1, 0, xx, xx, xx,
 173           NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
 174 INTRINSIC(image_samples, 0, ARR(0), true, 1, 1, 0, xx, xx, xx,
 175           NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
 176
 177 /*
 178  * Vulkan descriptor set intrinsic
 179  *
 180  * The Vulkan API uses a different binding model from GL.  In the Vulkan
 181  * API, all external resources are represented by a tuple:
 182  *
 183  * (descriptor set, binding, array index)
 184  *
 185  * where the array index is the only thing allowed to be indirect.  The
 186  * vulkan_surface_index intrinsic takes the descriptor set and binding as
 187  * its first two indices and the array index as its source.  The third
 188  * index is a nir_variable_mode in case that's useful to the backend.
 189  *
 190  * The intended usage is that the shader will call vulkan_surface_index to
 191  * get an index and then pass that as the buffer index ubo/ssbo calls.
 192  */
 193 INTRINSIC(vulkan_resource_index, 1, ARR(1), true, 1, 0, 2,
 194           DESC_SET, BINDING, xx,
 195           NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
 196
 197 /*
 198  * variable atomic intrinsics
 199  *
 200  * All of these variable atomic memory operations read a value from memory,
 201  * compute a new value using one of the operations below, write the new value
 202  * to memory, and return the original value read.
 203  *
 204  * All operations take 1 source except CompSwap that takes 2. These sources
 205  * represent:
 206  *
 207  * 0: The data parameter to the atomic function (i.e. the value to add
 208  *    in shared_atomic_add, etc).
 209  * 1: For CompSwap only: the second data parameter.
 210  *
 211  * All operations take 1 variable deref.
 212  */
 213 INTRINSIC(var_atomic_add, 1, ARR(1), true, 1, 1, 0, xx, xx, xx, 0)
 214 INTRINSIC(var_atomic_imin, 1, ARR(1), true, 1, 1, 0, xx, xx, xx, 0)
 215 INTRINSIC(var_atomic_umin, 1, ARR(1), true, 1, 1, 0, xx, xx, xx, 0)
 216 INTRINSIC(var_atomic_imax, 1, ARR(1), true, 1, 1, 0, xx, xx, xx, 0)
 217 INTRINSIC(var_atomic_umax, 1, ARR(1), true, 1, 1, 0, xx, xx, xx, 0)
 218 INTRINSIC(var_atomic_and, 1, ARR(1), true, 1, 1, 0, xx, xx, xx, 0)
 219 INTRINSIC(var_atomic_or, 1, ARR(1), true, 1, 1, 0, xx, xx, xx, 0)
 220 INTRINSIC(var_atomic_xor, 1, ARR(1), true, 1, 1, 0, xx, xx, xx, 0)
 221 INTRINSIC(var_atomic_exchange, 1, ARR(1), true, 1, 1, 0, xx, xx, xx, 0)
 222 INTRINSIC(var_atomic_comp_swap, 2, ARR(1, 1), true, 1, 1, 0, xx, xx, xx, 0)
 223
 224 /*
 225  * SSBO atomic intrinsics
 226  *
 227  * All of the SSBO atomic memory operations read a value from memory,
 228  * compute a new value using one of the operations below, write the new
 229  * value to memory, and return the original value read.
 230  *
 231  * All operations take 3 sources except CompSwap that takes 4. These
 232  * sources represent:
 233  *
 234  * 0: The SSBO buffer index.
 235  * 1: The offset into the SSBO buffer of the variable that the atomic
 236  *    operation will operate on.
 237  * 2: The data parameter to the atomic function (i.e. the value to add
 238  *    in ssbo_atomic_add, etc).
 239  * 3: For CompSwap only: the second data parameter.
 240  */
 241 INTRINSIC(ssbo_atomic_add, 3, ARR(1, 1, 1), true, 1, 0, 0, xx, xx, xx, 0)
 242 INTRINSIC(ssbo_atomic_imin, 3, ARR(1, 1, 1), true, 1, 0, 0, xx, xx, xx, 0)
 243 INTRINSIC(ssbo_atomic_umin, 3, ARR(1, 1, 1), true, 1, 0, 0, xx, xx, xx, 0)
 244 INTRINSIC(ssbo_atomic_imax, 3, ARR(1, 1, 1), true, 1, 0, 0, xx, xx, xx, 0)
 245 INTRINSIC(ssbo_atomic_umax, 3, ARR(1, 1, 1), true, 1, 0, 0, xx, xx, xx, 0)
 246 INTRINSIC(ssbo_atomic_and, 3, ARR(1, 1, 1), true, 1, 0, 0, xx, xx, xx, 0)
 247 INTRINSIC(ssbo_atomic_or, 3, ARR(1, 1, 1), true, 1, 0, 0, xx, xx, xx, 0)
 248 INTRINSIC(ssbo_atomic_xor, 3, ARR(1, 1, 1), true, 1, 0, 0, xx, xx, xx, 0)
 249 INTRINSIC(ssbo_atomic_exchange, 3, ARR(1, 1, 1), true, 1, 0, 0, xx, xx, xx, 0)
 250 INTRINSIC(ssbo_atomic_comp_swap, 4, ARR(1, 1, 1, 1), true, 1, 0, 0, xx, xx, xx, 0)
 251
 252 /*
 253  * CS shared variable atomic intrinsics
 254  *
 255  * All of the shared variable atomic memory operations read a value from
 256  * memory, compute a new value using one of the operations below, write the
 257  * new value to memory, and return the original value read.
 258  *
 259  * All operations take 2 sources except CompSwap that takes 3. These
 260  * sources represent:
 261  *
 262  * 0: The offset into the shared variable storage region that the atomic
 263  *    operation will operate on.
 264  * 1: The data parameter to the atomic function (i.e. the value to add
 265  *    in shared_atomic_add, etc).
 266  * 2: For CompSwap only: the second data parameter.
 267  */
 268 INTRINSIC(shared_atomic_add, 2, ARR(1, 1), true, 1, 0, 1, BASE, xx, xx, 0)
 269 INTRINSIC(shared_atomic_imin, 2, ARR(1, 1), true, 1, 0, 1, BASE, xx, xx, 0)
 270 INTRINSIC(shared_atomic_umin, 2, ARR(1, 1), true, 1, 0, 1, BASE, xx, xx, 0)
 271 INTRINSIC(shared_atomic_imax, 2, ARR(1, 1), true, 1, 0, 1, BASE, xx, xx, 0)
 272 INTRINSIC(shared_atomic_umax, 2, ARR(1, 1), true, 1, 0, 1, BASE, xx, xx, 0)
 273 INTRINSIC(shared_atomic_and, 2, ARR(1, 1), true, 1, 0, 1, BASE, xx, xx, 0)
 274 INTRINSIC(shared_atomic_or, 2, ARR(1, 1), true, 1, 0, 1, BASE, xx, xx, 0)
 275 INTRINSIC(shared_atomic_xor, 2, ARR(1, 1), true, 1, 0, 1, BASE, xx, xx, 0)
 276 INTRINSIC(shared_atomic_exchange, 2, ARR(1, 1), true, 1, 0, 1, BASE, xx, xx, 0)
 277 INTRINSIC(shared_atomic_comp_swap, 3, ARR(1, 1, 1), true, 1, 0, 1, BASE, xx, xx, 0)
 278
 279 /* Used by nir_builder.h to generate loader helpers for the system values. */
 280 #ifndef DEFINE_SYSTEM_VALUE
 281 #define DEFINE_SYSTEM_VALUE(name)
 282 #endif
 283
 284 #define SYSTEM_VALUE(name, components, num_indices, idx0, idx1, idx2) \
 285    DEFINE_SYSTEM_VALUE(name) \
 286    INTRINSIC(load_##name, 0, ARR(0), true, components, 0, num_indices, \
 287    idx0, idx1, idx2, \
 288    NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
 289
 290 SYSTEM_VALUE(front_face, 1, 0, xx, xx, xx)
 291 SYSTEM_VALUE(vertex_id, 1, 0, xx, xx, xx)
 292 SYSTEM_VALUE(vertex_id_zero_base, 1, 0, xx, xx, xx)
 293 SYSTEM_VALUE(base_vertex, 1, 0, xx, xx, xx)
 294 SYSTEM_VALUE(instance_id, 1, 0, xx, xx, xx)
 295 SYSTEM_VALUE(base_instance, 1, 0, xx, xx, xx)
 296 SYSTEM_VALUE(draw_id, 1, 0, xx, xx, xx)
 297 SYSTEM_VALUE(sample_id, 1, 0, xx, xx, xx)
 298 SYSTEM_VALUE(sample_pos, 2, 0, xx, xx, xx)
 299 SYSTEM_VALUE(sample_mask_in, 1, 0, xx, xx, xx)
 300 SYSTEM_VALUE(primitive_id, 1, 0, xx, xx, xx)
 301 SYSTEM_VALUE(invocation_id, 1, 0, xx, xx, xx)
 302 SYSTEM_VALUE(tess_coord, 3, 0, xx, xx, xx)
 303 SYSTEM_VALUE(tess_level_outer, 4, 0, xx, xx, xx)
 304 SYSTEM_VALUE(tess_level_inner, 2, 0, xx, xx, xx)
 305 SYSTEM_VALUE(patch_vertices_in, 1, 0, xx, xx, xx)
 306 SYSTEM_VALUE(local_invocation_id, 3, 0, xx, xx, xx)
 307 SYSTEM_VALUE(local_invocation_index, 1, 0, xx, xx, xx)
 308 SYSTEM_VALUE(work_group_id, 3, 0, xx, xx, xx)
 309 SYSTEM_VALUE(user_clip_plane, 4, 1, UCP_ID, xx, xx)
 310 SYSTEM_VALUE(num_work_groups, 3, 0, xx, xx, xx)
 311 SYSTEM_VALUE(helper_invocation, 1, 0, xx, xx, xx)
 312 SYSTEM_VALUE(channel_num, 1, 0, xx, xx, xx)
 313
 314 /**
 315  * Barycentric coordinate intrinsics.
 316  *
 317  * These set up the barycentric coordinates for a particular interpolation.
 318  * The first three are for the simple cases: pixel, centroid, or per-sample
 319  * (at gl_SampleID).  The next two handle interpolating at a specified
 320  * sample location, or interpolating with a vec2 offset,
 321  *
 322  * The interp_mode index should be either the INTERP_MODE_SMOOTH or
 323  * INTERP_MODE_NOPERSPECTIVE enum values.
 324  *
 325  * The vec2 value produced by these intrinsics is intended for use as the
 326  * barycoord source of a load_interpolated_input intrinsic.
 327  */
 328
 329 #define BARYCENTRIC(name, sources, source_components) \
 330    INTRINSIC(load_barycentric_##name, sources, ARR(source_components), \
 331              true, 2, 0, 1, INTERP_MODE, xx, xx, \
 332              NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
 333
 334 /* no sources.  const_index[] = { interp_mode } */
 335 BARYCENTRIC(pixel, 0, 0)
 336 BARYCENTRIC(centroid, 0, 0)
 337 BARYCENTRIC(sample, 0, 0)
 338 /* src[] = { sample_id }.  const_index[] = { interp_mode } */
 339 BARYCENTRIC(at_sample, 1, 1)
 340 /* src[] = { offset.xy }.  const_index[] = { interp_mode } */
 341 BARYCENTRIC(at_offset, 1, 2)
 342
 343 /*
 344  * Load operations pull data from some piece of GPU memory.  All load
 345  * operations operate in terms of offsets into some piece of theoretical
 346  * memory.  Loads from externally visible memory (UBO and SSBO) simply take a
 347  * byte offset as a source.  Loads from opaque memory (uniforms, inputs, etc.)
 348  * take a base+offset pair where the base (const_index[0]) gives the location
 349  * of the start of the variable being loaded and and the offset source is a
 350  * offset into that variable.
 351  *
 352  * Uniform load operations have a second "range" index that specifies the
 353  * range (starting at base) of the data from which we are loading.  If
 354  * const_index[1] == 0, then the range is unknown.
 355  *
 356  * Some load operations such as UBO/SSBO load and per_vertex loads take an
 357  * additional source to specify which UBO/SSBO/vertex to load from.
 358  *
 359  * The exact address type depends on the lowering pass that generates the
 360  * load/store intrinsics.  Typically, this is vec4 units for things such as
 361  * varying slots and float units for fragment shader inputs.  UBO and SSBO
 362  * offsets are always in bytes.
 363  */
 364
 365 #define LOAD(name, srcs, num_indices, idx0, idx1, idx2, flags) \
 366    INTRINSIC(load_##name, srcs, ARR(1, 1, 1, 1), true, 0, 0, num_indices, idx0, idx1, idx2, flags)
 367
 368 /* src[] = { offset }. const_index[] = { base, range } */
 369 LOAD(uniform, 1, 2, BASE, RANGE, xx, NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
 370 /* src[] = { buffer_index, offset }. No const_index */
 371 LOAD(ubo, 2, 0, xx, xx, xx, NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
 372 /* src[] = { offset }. const_index[] = { base, component } */
 373 LOAD(input, 1, 2, BASE, COMPONENT, xx, NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
 374 /* src[] = { vertex, offset }. const_index[] = { base, component } */
 375 LOAD(per_vertex_input, 2, 2, BASE, COMPONENT, xx, NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
 376 /* src[] = { barycoord, offset }. const_index[] = { base, component } */
 377 LOAD(interpolated_input, 2, 2, BASE, COMPONENT, xx, NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
 378
 379 /* src[] = { buffer_index, offset }. No const_index */
 380 LOAD(ssbo, 2, 0, xx, xx, xx, NIR_INTRINSIC_CAN_ELIMINATE)
 381 /* src[] = { offset }. const_index[] = { base, component } */
 382 LOAD(output, 1, 1, BASE, COMPONENT, xx, NIR_INTRINSIC_CAN_ELIMINATE)
 383 /* src[] = { vertex, offset }. const_index[] = { base, component } */
 384 LOAD(per_vertex_output, 2, 1, BASE, COMPONENT, xx, NIR_INTRINSIC_CAN_ELIMINATE)
 385 /* src[] = { offset }. const_index[] = { base } */
 386 LOAD(shared, 1, 1, BASE, xx, xx, NIR_INTRINSIC_CAN_ELIMINATE)
 387 /* src[] = { offset }. const_index[] = { base, range } */
 388 LOAD(push_constant, 1, 2, BASE, RANGE, xx,
 389      NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER)
 390
 391 /*
 392  * Stores work the same way as loads, except now the first source is the value
 393  * to store and the second (and possibly third) source specify where to store
 394  * the value.  SSBO and shared memory stores also have a write mask as
 395  * const_index[0].
 396  */
 397
 398 #define STORE(name, srcs, num_indices, idx0, idx1, idx2, flags) \
 399    INTRINSIC(store_##name, srcs, ARR(0, 1, 1, 1), false, 0, 0, num_indices, idx0, idx1, idx2, flags)
 400
 401 /* src[] = { value, offset }. const_index[] = { base, write_mask, component } */
 402 STORE(output, 2, 3, BASE, WRMASK, COMPONENT, 0)
 403 /* src[] = { value, vertex, offset }.
 404  * const_index[] = { base, write_mask, component }
 405  */
 406 STORE(per_vertex_output, 3, 3, BASE, WRMASK, COMPONENT, 0)
 407 /* src[] = { value, block_index, offset }. const_index[] = { write_mask } */
 408 STORE(ssbo, 3, 1, WRMASK, xx, xx, 0)
 409 /* src[] = { value, offset }. const_index[] = { base, write_mask } */
 410 STORE(shared, 2, 2, BASE, WRMASK, xx, 0)
 411
 412 LAST_INTRINSIC(store_shared)
 413
 414 #undef DEFINE_SYSTEM_VALUE
 415 #undef INTRINSIC
 416 #undef LAST_INTRINSIC