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][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][j].u64 = load_const->value[instr->src[i].swizzle[j]].u64;
  81             break;
  82          case 32:
  83             src[i][j].u32 = load_const->value[instr->src[i].swizzle[j]].u32;
  84             break;
  85          case 16:
  86             src[i][j].u16 = load_const->value[instr->src[i].swizzle[j]].u16;
  87             break;
  88          case 8:
  89             src[i][j].u8 = load_const->value[instr->src[i].swizzle[j]].u8;
  90             break;
  91          case 1:
  92             src[i][j].b = load_const->value[instr->src[i].swizzle[j]].b;
  93             break;
  94          default:
  95             unreachable("Invalid bit size");
  96          }
  97       }
  98
  99       /* We shouldn't have any source modifiers in the optimization loop. */
 100       assert(!instr->src[i].abs && !instr->src[i].negate);
 101    }
 102
 103    if (bit_size == 0)
 104       bit_size = 32;
 105
 106    /* We shouldn't have any saturate modifiers in the optimization loop. */
 107    assert(!instr->dest.saturate);
 108
 109    nir_const_value dest[NIR_MAX_VEC_COMPONENTS];
 110    nir_const_value *srcs[NIR_MAX_VEC_COMPONENTS];
 111    memset(dest, 0, sizeof(dest));
 112    for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; ++i)
 113       srcs[i] = src[i];
 114    nir_eval_const_opcode(instr->op, dest, instr->dest.dest.ssa.num_components,
 115                          bit_size, srcs);
 116
 117    nir_load_const_instr *new_instr =
 118       nir_load_const_instr_create(mem_ctx,
 119                                   instr->dest.dest.ssa.num_components,
 120                                   instr->dest.dest.ssa.bit_size);
 121
 122    memcpy(new_instr->value, dest, sizeof(*new_instr->value) * new_instr->def.num_components);
 123
 124    nir_instr_insert_before(&instr->instr, &new_instr->instr);
 125
 126    nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa,
 127                             nir_src_for_ssa(&new_instr->def));
 128
 129    nir_instr_remove(&instr->instr);
 130    ralloc_free(instr);
 131
 132    return true;
 133 }
 134
 135 static bool
 136 constant_fold_intrinsic_instr(nir_intrinsic_instr *instr)
 137 {
 138    bool progress = false;
 139
 140    if (instr->intrinsic == nir_intrinsic_discard_if &&
 141        nir_src_is_const(instr->src[0])) {
 142       if (nir_src_as_bool(instr->src[0])) {
 143          /* This method of getting a nir_shader * from a nir_instr is
 144           * admittedly gross, but given the rarity of hitting this case I think
 145           * it's preferable to plumbing an otherwise unused nir_shader *
 146           * parameter through four functions to get here.
 147           */
 148          nir_cf_node *cf_node = &instr->instr.block->cf_node;
 149          nir_function_impl *impl = nir_cf_node_get_function(cf_node);
 150          nir_shader *shader = impl->function->shader;
 151
 152          nir_intrinsic_instr *discard =
 153             nir_intrinsic_instr_create(shader, nir_intrinsic_discard);
 154          nir_instr_insert_before(&instr->instr, &discard->instr);
 155          nir_instr_remove(&instr->instr);
 156          progress = true;
 157       } else {
 158          /* We're not discarding, just delete the instruction */
 159          nir_instr_remove(&instr->instr);
 160          progress = true;
 161       }
 162    }
 163
 164    return progress;
 165 }
 166
 167 static bool
 168 constant_fold_block(nir_block *block, void *mem_ctx)
 169 {
 170    bool progress = false;
 171
 172    nir_foreach_instr_safe(instr, block) {
 173       switch (instr->type) {
 174       case nir_instr_type_alu:
 175          progress |= constant_fold_alu_instr(nir_instr_as_alu(instr), mem_ctx);
 176          break;
 177       case nir_instr_type_intrinsic:
 178          progress |=
 179             constant_fold_intrinsic_instr(nir_instr_as_intrinsic(instr));
 180          break;
 181       default:
 182          /* Don't know how to constant fold */
 183          break;
 184       }
 185    }
 186
 187    return progress;
 188 }
 189
 190 static bool
 191 nir_opt_constant_folding_impl(nir_function_impl *impl)
 192 {
 193    void *mem_ctx = ralloc_parent(impl);
 194    bool progress = false;
 195
 196    nir_foreach_block(block, impl) {
 197       progress |= constant_fold_block(block, mem_ctx);
 198    }
 199
 200    if (progress) {
 201       nir_metadata_preserve(impl, nir_metadata_block_index |
 202                                   nir_metadata_dominance);
 203    } else {
 204 #ifndef NDEBUG
 205       impl->valid_metadata &= ~nir_metadata_not_properly_reset;
 206 #endif
 207    }
 208
 209    return progress;
 210 }
 211
 212 bool
 213 nir_opt_constant_folding(nir_shader *shader)
 214 {
 215    bool progress = false;
 216
 217    nir_foreach_function(function, shader) {
 218       if (function->impl)
 219          progress |= nir_opt_constant_folding_impl(function->impl);
 220    }
 221
 222    return progress;
 223 }