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    nir_shader *shader;
  37    unsigned execution_mode;
  38    bool has_load_constant;
  39    bool has_indirect_load_const;
  40 };
  41
  42 static bool
  43 constant_fold_alu_instr(struct constant_fold_state *state, nir_alu_instr *instr)
  44 {
  45    nir_const_value src[NIR_MAX_VEC_COMPONENTS][NIR_MAX_VEC_COMPONENTS];
  46
  47    if (!instr->dest.dest.is_ssa)
  48       return false;
  49
  50    /* In the case that any outputs/inputs have unsized types, then we need to
  51     * guess the bit-size. In this case, the validator ensures that all
  52     * bit-sizes match so we can just take the bit-size from first
  53     * output/input with an unsized type. If all the outputs/inputs are sized
  54     * then we don't need to guess the bit-size at all because the code we
  55     * generate for constant opcodes in this case already knows the sizes of
  56     * the types involved and does not need the provided bit-size for anything
  57     * (although it still requires to receive a valid bit-size).
  58     */
  59    unsigned bit_size = 0;
  60    if (!nir_alu_type_get_type_size(nir_op_infos[instr->op].output_type))
  61       bit_size = instr->dest.dest.ssa.bit_size;
  62
  63    for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++) {
  64       if (!instr->src[i].src.is_ssa)
  65          return false;
  66
  67       if (bit_size == 0 &&
  68           !nir_alu_type_get_type_size(nir_op_infos[instr->op].input_types[i]))
  69          bit_size = instr->src[i].src.ssa->bit_size;
  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          src[i][j] = load_const->value[instr->src[i].swizzle[j]];
  80       }
  81
  82       /* We shouldn't have any source modifiers in the optimization loop. */
  83       assert(!instr->src[i].abs && !instr->src[i].negate);
  84    }
  85
  86    if (bit_size == 0)
  87       bit_size = 32;
  88
  89    /* We shouldn't have any saturate modifiers in the optimization loop. */
  90    assert(!instr->dest.saturate);
  91
  92    nir_const_value dest[NIR_MAX_VEC_COMPONENTS];
  93    nir_const_value *srcs[NIR_MAX_VEC_COMPONENTS];
  94    memset(dest, 0, sizeof(dest));
  95    for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; ++i)
  96       srcs[i] = src[i];
  97    nir_eval_const_opcode(instr->op, dest, instr->dest.dest.ssa.num_components,
  98                          bit_size, srcs, state->execution_mode);
  99
 100    nir_load_const_instr *new_instr =
 101       nir_load_const_instr_create(state->shader,
 102                                   instr->dest.dest.ssa.num_components,
 103                                   instr->dest.dest.ssa.bit_size);
 104
 105    memcpy(new_instr->value, dest, sizeof(*new_instr->value) * new_instr->def.num_components);
 106
 107    nir_instr_insert_before(&instr->instr, &new_instr->instr);
 108
 109    nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa,
 110                             nir_src_for_ssa(&new_instr->def));
 111
 112    nir_instr_remove(&instr->instr);
 113    ralloc_free(instr);
 114
 115    return true;
 116 }
 117
 118 static bool
 119 constant_fold_intrinsic_instr(struct constant_fold_state *state, nir_intrinsic_instr *instr)
 120 {
 121    bool progress = false;
 122
 123    if ((instr->intrinsic == nir_intrinsic_demote_if ||
 124         instr->intrinsic == nir_intrinsic_discard_if) &&
 125        nir_src_is_const(instr->src[0])) {
 126       if (nir_src_as_bool(instr->src[0])) {
 127          nir_intrinsic_op op = instr->intrinsic == nir_intrinsic_discard_if ?
 128                                nir_intrinsic_discard :
 129                                nir_intrinsic_demote;
 130          nir_intrinsic_instr *new_instr = nir_intrinsic_instr_create(state->shader, op);
 131          nir_instr_insert_before(&instr->instr, &new_instr->instr);
 132          nir_instr_remove(&instr->instr);
 133          progress = true;
 134       } else {
 135          /* We're not discarding, just delete the instruction */
 136          nir_instr_remove(&instr->instr);
 137          progress = true;
 138       }
 139    } else if (instr->intrinsic == nir_intrinsic_load_constant) {
 140       state->has_load_constant = true;
 141
 142       if (!nir_src_is_const(instr->src[0])) {
 143          state->has_indirect_load_const = true;
 144          return progress;
 145       }
 146
 147       unsigned offset = nir_src_as_uint(instr->src[0]);
 148       unsigned base = nir_intrinsic_base(instr);
 149       unsigned range = nir_intrinsic_range(instr);
 150       assert(base + range <= state->shader->constant_data_size);
 151
 152       nir_instr *new_instr = NULL;
 153       if (offset >= range) {
 154          nir_ssa_undef_instr *undef =
 155             nir_ssa_undef_instr_create(state->shader,
 156                                        instr->num_components,
 157                                        instr->dest.ssa.bit_size);
 158
 159          nir_ssa_def_rewrite_uses(&instr->dest.ssa, nir_src_for_ssa(&undef->def));
 160          new_instr = &undef->instr;
 161       } else {
 162          nir_load_const_instr *load_const =
 163             nir_load_const_instr_create(state->shader,
 164                                         instr->num_components,
 165                                         instr->dest.ssa.bit_size);
 166
 167          uint8_t *data = (uint8_t*)state->shader->constant_data + base;
 168          for (unsigned i = 0; i < instr->num_components; i++) {
 169             unsigned bytes = instr->dest.ssa.bit_size / 8;
 170             bytes = MIN2(bytes, range - offset);
 171
 172             memcpy(&load_const->value[i].u64, data + offset, bytes);
 173             offset += bytes;
 174          }
 175
 176          nir_ssa_def_rewrite_uses(&instr->dest.ssa, nir_src_for_ssa(&load_const->def));
 177          new_instr = &load_const->instr;
 178       }
 179
 180       nir_instr_insert_before(&instr->instr, new_instr);
 181       nir_instr_remove(&instr->instr);
 182       progress = true;
 183    }
 184
 185    return progress;
 186 }
 187
 188 static bool
 189 constant_fold_block(struct constant_fold_state *state, nir_block *block)
 190 {
 191    bool progress = false;
 192
 193    nir_foreach_instr_safe(instr, block) {
 194       switch (instr->type) {
 195       case nir_instr_type_alu:
 196          progress |= constant_fold_alu_instr(state, nir_instr_as_alu(instr));
 197          break;
 198       case nir_instr_type_intrinsic:
 199          progress |=
 200             constant_fold_intrinsic_instr(state, nir_instr_as_intrinsic(instr));
 201          break;
 202       default:
 203          /* Don't know how to constant fold */
 204          break;
 205       }
 206    }
 207
 208    return progress;
 209 }
 210
 211 static bool
 212 nir_opt_constant_folding_impl(struct constant_fold_state *state, nir_function_impl *impl)
 213 {
 214    bool progress = false;
 215
 216    nir_foreach_block(block, impl) {
 217       progress |= constant_fold_block(state, block);
 218    }
 219
 220    if (progress) {
 221       nir_metadata_preserve(impl, nir_metadata_block_index |
 222                                   nir_metadata_dominance);
 223    } else {
 224       nir_metadata_preserve(impl, nir_metadata_all);
 225    }
 226
 227    return progress;
 228 }
 229
 230 bool
 231 nir_opt_constant_folding(nir_shader *shader)
 232 {
 233    bool progress = false;
 234    struct constant_fold_state state;
 235    state.shader = shader;
 236    state.execution_mode = shader->info.float_controls_execution_mode;
 237    state.has_load_constant = false;
 238    state.has_indirect_load_const = false;
 239
 240    nir_foreach_function(function, shader) {
 241       if (function->impl)
 242          progress |= nir_opt_constant_folding_impl(&state, function->impl);
 243    }
 244
 245    /* This doesn't free the constant data if there are no constant loads because
 246     * the data might still be used but the loads have been lowered to load_ubo
 247     */
 248    if (state.has_load_constant && !state.has_indirect_load_const &&
 249        shader->constant_data_size) {
 250       ralloc_free(shader->constant_data);
 251       shader->constant_data = NULL;
 252       shader->constant_data_size = 0;
 253    }
 254
 255    return progress;
 256 }