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 &&
  67           !nir_alu_type_get_type_size(nir_op_infos[instr->op].input_types[i]))
  68          bit_size = instr->src[i].src.ssa->bit_size;
  69
  70       nir_instr *src_instr = instr->src[i].src.ssa->parent_instr;
  71
  72       if (src_instr->type != nir_instr_type_load_const)
  73          return false;
  74       nir_load_const_instr* load_const = nir_instr_as_load_const(src_instr);
  75
  76       for (unsigned j = 0; j < nir_ssa_alu_instr_src_components(instr, i);
  77            j++) {
  78          switch(load_const->def.bit_size) {
  79          case 64:
  80             src[i].u64[j] = load_const->value.u64[instr->src[i].swizzle[j]];
  81             break;
  82          case 32:
  83             src[i].u32[j] = load_const->value.u32[instr->src[i].swizzle[j]];
  84             break;
  85          case 16:
  86             src[i].u16[j] = load_const->value.u16[instr->src[i].swizzle[j]];
  87             break;
  88          case 8:
  89             src[i].u8[j] = load_const->value.u8[instr->src[i].swizzle[j]];
  90             break;
  91          default:
  92             unreachable("Invalid bit size");
  93          }
  94       }
  95
  96       /* We shouldn't have any source modifiers in the optimization loop. */
  97       assert(!instr->src[i].abs && !instr->src[i].negate);
  98    }
  99
 100    if (bit_size == 0)
 101       bit_size = 32;
 102
 103    /* We shouldn't have any saturate modifiers in the optimization loop. */
 104    assert(!instr->dest.saturate);
 105
 106    nir_const_value dest =
 107       nir_eval_const_opcode(instr->op, instr->dest.dest.ssa.num_components,
 108                             bit_size, src);
 109
 110    nir_load_const_instr *new_instr =
 111       nir_load_const_instr_create(mem_ctx,
 112                                   instr->dest.dest.ssa.num_components,
 113                                   instr->dest.dest.ssa.bit_size);
 114
 115    new_instr->value = dest;
 116
 117    nir_instr_insert_before(&instr->instr, &new_instr->instr);
 118
 119    nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa,
 120                             nir_src_for_ssa(&new_instr->def));
 121
 122    nir_instr_remove(&instr->instr);
 123    ralloc_free(instr);
 124
 125    return true;
 126 }
 127
 128 static bool
 129 constant_fold_intrinsic_instr(nir_intrinsic_instr *instr)
 130 {
 131    bool progress = false;
 132
 133    if (instr->intrinsic == nir_intrinsic_discard_if &&
 134        nir_src_is_const(instr->src[0])) {
 135       if (nir_src_as_bool(instr->src[0])) {
 136          /* This method of getting a nir_shader * from a nir_instr is
 137           * admittedly gross, but given the rarity of hitting this case I think
 138           * it's preferable to plumbing an otherwise unused nir_shader *
 139           * parameter through four functions to get here.
 140           */
 141          nir_cf_node *cf_node = &instr->instr.block->cf_node;
 142          nir_function_impl *impl = nir_cf_node_get_function(cf_node);
 143          nir_shader *shader = impl->function->shader;
 144
 145          nir_intrinsic_instr *discard =
 146             nir_intrinsic_instr_create(shader, nir_intrinsic_discard);
 147          nir_instr_insert_before(&instr->instr, &discard->instr);
 148          nir_instr_remove(&instr->instr);
 149          progress = true;
 150       } else {
 151          /* We're not discarding, just delete the instruction */
 152          nir_instr_remove(&instr->instr);
 153          progress = true;
 154       }
 155    }
 156
 157    return progress;
 158 }
 159
 160 static bool
 161 constant_fold_block(nir_block *block, void *mem_ctx)
 162 {
 163    bool progress = false;
 164
 165    nir_foreach_instr_safe(instr, block) {
 166       switch (instr->type) {
 167       case nir_instr_type_alu:
 168          progress |= constant_fold_alu_instr(nir_instr_as_alu(instr), mem_ctx);
 169          break;
 170       case nir_instr_type_intrinsic:
 171          progress |=
 172             constant_fold_intrinsic_instr(nir_instr_as_intrinsic(instr));
 173          break;
 174       default:
 175          /* Don't know how to constant fold */
 176          break;
 177       }
 178    }
 179
 180    return progress;
 181 }
 182
 183 static bool
 184 nir_opt_constant_folding_impl(nir_function_impl *impl)
 185 {
 186    void *mem_ctx = ralloc_parent(impl);
 187    bool progress = false;
 188
 189    nir_foreach_block(block, impl) {
 190       progress |= constant_fold_block(block, mem_ctx);
 191    }
 192
 193    if (progress)
 194       nir_metadata_preserve(impl, nir_metadata_block_index |
 195                                   nir_metadata_dominance);
 196
 197    return progress;
 198 }
 199
 200 bool
 201 nir_opt_constant_folding(nir_shader *shader)
 202 {
 203    bool progress = false;
 204
 205    nir_foreach_function(function, shader) {
 206       if (function->impl)
 207          progress |= nir_opt_constant_folding_impl(function->impl);
 208    }
 209
 210    return progress;
 211 }