src/compiler/nir/nir_opt_constant_folding.c

   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  *    Jason Ekstrand (jason@jlekstrand.net)
  25  *
  26  */
  27
  28 #include "nir_constant_expressions.h"
  29 #include <math.h>
  30
  31 /*
  32  * Implements SSA-based constant folding.
  33  */
  34
  35 struct constant_fold_state {
  36    void *mem_ctx;
  37    nir_function_impl *impl;
  38    bool progress;
  39 };
  40
  41 static bool
  42 constant_fold_alu_instr(nir_alu_instr *instr, void *mem_ctx)
  43 {
  44    nir_const_value src[4];
  45
  46    if (!instr->dest.dest.is_ssa)
  47       return false;
  48
  49    /* In the case that any outputs/inputs have unsized types, then we need to
  50     * guess the bit-size. In this case, the validator ensures that all
  51     * bit-sizes match so we can just take the bit-size from first
  52     * output/input with an unsized type. If all the outputs/inputs are sized
  53     * then we don't need to guess the bit-size at all because the code we
  54     * generate for constant opcodes in this case already knows the sizes of
  55     * the types involved and does not need the provided bit-size for anything
  56     * (although it still requires to receive a valid bit-size).
  57     */
  58    unsigned bit_size = 0;
  59    if (!nir_alu_type_get_type_size(nir_op_infos[instr->op].output_type))
  60       bit_size = instr->dest.dest.ssa.bit_size;
  61
  62    for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++) {
  63       if (!instr->src[i].src.is_ssa)
  64          return false;
  65
  66       if (bit_size == 0 &&
  67           !nir_alu_type_get_type_size(nir_op_infos[instr->op].input_sizes[i])) {
  68          bit_size = instr->src[i].src.ssa->bit_size;
  69       }
  70
  71       nir_instr *src_instr = instr->src[i].src.ssa->parent_instr;
  72
  73       if (src_instr->type != nir_instr_type_load_const)
  74          return false;
  75       nir_load_const_instr* load_const = nir_instr_as_load_const(src_instr);
  76
  77       for (unsigned j = 0; j < nir_ssa_alu_instr_src_components(instr, i);
  78            j++) {
  79          switch(load_const->def.bit_size) {
  80          case 64:
  81             src[i].u64[j] = load_const->value.u64[instr->src[i].swizzle[j]];
  82             break;
  83          case 32:
  84             src[i].u32[j] = load_const->value.u32[instr->src[i].swizzle[j]];
  85             break;
  86          case 16:
  87             src[i].u16[j] = load_const->value.u16[instr->src[i].swizzle[j]];
  88             break;
  89          case 8:
  90             src[i].u8[j] = load_const->value.u8[instr->src[i].swizzle[j]];
  91             break;
  92          }
  93       }
  94
  95       /* We shouldn't have any source modifiers in the optimization loop. */
  96       assert(!instr->src[i].abs && !instr->src[i].negate);
  97    }
  98
  99    if (bit_size == 0)
 100       bit_size = 32;
 101
 102    /* We shouldn't have any saturate modifiers in the optimization loop. */
 103    assert(!instr->dest.saturate);
 104
 105    nir_const_value dest =
 106       nir_eval_const_opcode(instr->op, instr->dest.dest.ssa.num_components,
 107                             bit_size, src);
 108
 109    nir_load_const_instr *new_instr =
 110       nir_load_const_instr_create(mem_ctx,
 111                                   instr->dest.dest.ssa.num_components,
 112                                   instr->dest.dest.ssa.bit_size);
 113
 114    new_instr->value = dest;
 115
 116    nir_instr_insert_before(&instr->instr, &new_instr->instr);
 117
 118    nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa,
 119                             nir_src_for_ssa(&new_instr->def));
 120
 121    nir_instr_remove(&instr->instr);
 122    ralloc_free(instr);
 123
 124    return true;
 125 }
 126
 127 static bool
 128 constant_fold_deref(nir_instr *instr, nir_deref_var *deref)
 129 {
 130    bool progress = false;
 131
 132    for (nir_deref *tail = deref->deref.child; tail; tail = tail->child) {
 133       if (tail->deref_type != nir_deref_type_array)
 134          continue;
 135
 136       nir_deref_array *arr = nir_deref_as_array(tail);
 137
 138       if (arr->deref_array_type == nir_deref_array_type_indirect &&
 139           arr->indirect.is_ssa &&
 140           arr->indirect.ssa->parent_instr->type == nir_instr_type_load_const) {
 141          nir_load_const_instr *indirect =
 142             nir_instr_as_load_const(arr->indirect.ssa->parent_instr);
 143
 144          arr->base_offset += indirect->value.u32[0];
 145
 146          /* Clear out the source */
 147          nir_instr_rewrite_src(instr, &arr->indirect, nir_src_for_ssa(NULL));
 148
 149          arr->deref_array_type = nir_deref_array_type_direct;
 150
 151          progress = true;
 152       }
 153    }
 154
 155    return progress;
 156 }
 157
 158 static bool
 159 constant_fold_intrinsic_instr(nir_intrinsic_instr *instr)
 160 {
 161    bool progress = false;
 162
 163    unsigned num_vars = nir_intrinsic_infos[instr->intrinsic].num_variables;
 164    for (unsigned i = 0; i < num_vars; i++) {
 165       progress |= constant_fold_deref(&instr->instr, instr->variables[i]);
 166    }
 167
 168    if (instr->intrinsic == nir_intrinsic_discard_if) {
 169       nir_const_value *src_val = nir_src_as_const_value(instr->src[0]);
 170       if (src_val && src_val->u32[0] == NIR_FALSE) {
 171          nir_instr_remove(&instr->instr);
 172          progress = true;
 173       } else if (src_val && src_val->u32[0] == NIR_TRUE) {
 174          /* This method of getting a nir_shader * from a nir_instr is
 175           * admittedly gross, but given the rarity of hitting this case I think
 176           * it's preferable to plumbing an otherwise unused nir_shader *
 177           * parameter through four functions to get here.
 178           */
 179          nir_cf_node *cf_node = &instr->instr.block->cf_node;
 180          nir_function_impl *impl = nir_cf_node_get_function(cf_node);
 181          nir_shader *shader = impl->function->shader;
 182
 183          nir_intrinsic_instr *discard =
 184             nir_intrinsic_instr_create(shader, nir_intrinsic_discard);
 185          nir_instr_insert_before(&instr->instr, &discard->instr);
 186          nir_instr_remove(&instr->instr);
 187          progress = true;
 188       }
 189    }
 190
 191    return progress;
 192 }
 193
 194 static bool
 195 constant_fold_tex_instr(nir_tex_instr *instr)
 196 {
 197    bool progress = false;
 198
 199    if (instr->texture)
 200       progress |= constant_fold_deref(&instr->instr, instr->texture);
 201
 202    if (instr->sampler)
 203       progress |= constant_fold_deref(&instr->instr, instr->sampler);
 204
 205    return progress;
 206 }
 207
 208 static bool
 209 constant_fold_block(nir_block *block, void *mem_ctx)
 210 {
 211    bool progress = false;
 212
 213    nir_foreach_instr_safe(instr, block) {
 214       switch (instr->type) {
 215       case nir_instr_type_alu:
 216          progress |= constant_fold_alu_instr(nir_instr_as_alu(instr), mem_ctx);
 217          break;
 218       case nir_instr_type_intrinsic:
 219          progress |=
 220             constant_fold_intrinsic_instr(nir_instr_as_intrinsic(instr));
 221          break;
 222       case nir_instr_type_tex:
 223          progress |= constant_fold_tex_instr(nir_instr_as_tex(instr));
 224          break;
 225       default:
 226          /* Don't know how to constant fold */
 227          break;
 228       }
 229    }
 230
 231    return progress;
 232 }
 233
 234 static bool
 235 nir_opt_constant_folding_impl(nir_function_impl *impl)
 236 {
 237    void *mem_ctx = ralloc_parent(impl);
 238    bool progress = false;
 239
 240    nir_foreach_block(block, impl) {
 241       progress |= constant_fold_block(block, mem_ctx);
 242    }
 243
 244    if (progress)
 245       nir_metadata_preserve(impl, nir_metadata_block_index |
 246                                   nir_metadata_dominance);
 247
 248    return progress;
 249 }
 250
 251 bool
 252 nir_opt_constant_folding(nir_shader *shader)
 253 {
 254    bool progress = false;
 255
 256    nir_foreach_function(function, shader) {
 257       if (function->impl)
 258          progress |= nir_opt_constant_folding_impl(function->impl);
 259    }
 260
 261    return progress;
 262 }