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[NIR_MAX_VEC_COMPONENTS];
  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 && nir_op_infos[instr->op].input_sizes[i] == 0)
  67          bit_size = instr->src[i].src.ssa->bit_size;
  68
  69       nir_instr *src_instr = instr->src[i].src.ssa->parent_instr;
  70
  71       if (src_instr->type != nir_instr_type_load_const)
  72          return false;
  73       nir_load_const_instr* load_const = nir_instr_as_load_const(src_instr);
  74
  75       for (unsigned j = 0; j < nir_ssa_alu_instr_src_components(instr, i);
  76            j++) {
  77          switch(load_const->def.bit_size) {
  78          case 64:
  79             src[i].u64[j] = load_const->value.u64[instr->src[i].swizzle[j]];
  80             break;
  81          case 32:
  82             src[i].u32[j] = load_const->value.u32[instr->src[i].swizzle[j]];
  83             break;
  84          case 16:
  85             src[i].u16[j] = load_const->value.u16[instr->src[i].swizzle[j]];
  86             break;
  87          case 8:
  88             src[i].u8[j] = load_const->value.u8[instr->src[i].swizzle[j]];
  89             break;
  90          default:
  91             unreachable("Invalid bit size");
  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_intrinsic_instr(nir_intrinsic_instr *instr)
 129 {
 130    bool progress = false;
 131
 132    if (instr->intrinsic == nir_intrinsic_discard_if &&
 133        nir_src_is_const(instr->src[0])) {
 134       if (nir_src_as_bool(instr->src[0])) {
 135          /* This method of getting a nir_shader * from a nir_instr is
 136           * admittedly gross, but given the rarity of hitting this case I think
 137           * it's preferable to plumbing an otherwise unused nir_shader *
 138           * parameter through four functions to get here.
 139           */
 140          nir_cf_node *cf_node = &instr->instr.block->cf_node;
 141          nir_function_impl *impl = nir_cf_node_get_function(cf_node);
 142          nir_shader *shader = impl->function->shader;
 143
 144          nir_intrinsic_instr *discard =
 145             nir_intrinsic_instr_create(shader, nir_intrinsic_discard);
 146          nir_instr_insert_before(&instr->instr, &discard->instr);
 147          nir_instr_remove(&instr->instr);
 148          progress = true;
 149       } else {
 150          /* We're not discarding, just delete the instruction */
 151          nir_instr_remove(&instr->instr);
 152          progress = true;
 153       }
 154    }
 155
 156    return progress;
 157 }
 158
 159 static bool
 160 constant_fold_block(nir_block *block, void *mem_ctx)
 161 {
 162    bool progress = false;
 163
 164    nir_foreach_instr_safe(instr, block) {
 165       switch (instr->type) {
 166       case nir_instr_type_alu:
 167          progress |= constant_fold_alu_instr(nir_instr_as_alu(instr), mem_ctx);
 168          break;
 169       case nir_instr_type_intrinsic:
 170          progress |=
 171             constant_fold_intrinsic_instr(nir_instr_as_intrinsic(instr));
 172          break;
 173       default:
 174          /* Don't know how to constant fold */
 175          break;
 176       }
 177    }
 178
 179    return progress;
 180 }
 181
 182 static bool
 183 nir_opt_constant_folding_impl(nir_function_impl *impl)
 184 {
 185    void *mem_ctx = ralloc_parent(impl);
 186    bool progress = false;
 187
 188    nir_foreach_block(block, impl) {
 189       progress |= constant_fold_block(block, mem_ctx);
 190    }
 191
 192    if (progress)
 193       nir_metadata_preserve(impl, nir_metadata_block_index |
 194                                   nir_metadata_dominance);
 195
 196    return progress;
 197 }
 198
 199 bool
 200 nir_opt_constant_folding(nir_shader *shader)
 201 {
 202    bool progress = false;
 203
 204    nir_foreach_function(function, shader) {
 205       if (function->impl)
 206          progress |= nir_opt_constant_folding_impl(function->impl);
 207    }
 208
 209    return progress;
 210 }