src/freedreno/ir3/ir3_depth.c

   1 /*
   2  * Copyright (C) 2014 Rob Clark <robclark@freedesktop.org>
   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 FROM,
  20  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  21  * SOFTWARE.
  22  *
  23  * Authors:
  24  *    Rob Clark <robclark@freedesktop.org>
  25  */
  26
  27 #include "util/u_math.h"
  28
  29 #include "ir3.h"
  30
  31 /*
  32  * Instruction Depth:
  33  *
  34  * Calculates weighted instruction depth, ie. the sum of # of needed
  35  * instructions plus delay slots back to original input (ie INPUT or
  36  * CONST).  That is to say, an instructions depth is:
  37  *
  38  *   depth(instr) {
  39  *     d = 0;
  40  *     // for each src register:
  41  *     foreach (src in instr->regs[1..n])
  42  *       d = max(d, delayslots(src->instr, n) + depth(src->instr));
  43  *     return d + 1;
  44  *   }
  45  *
  46  * After an instruction's depth is calculated, it is inserted into the
  47  * blocks depth sorted list, which is used by the scheduling pass.
  48  */
  49
  50 /* generally don't count false dependencies, since this can just be
  51  * something like a barrier, or SSBO store.  The exception is array
  52  * dependencies if the assigner is an array write and the consumer
  53  * reads the same array.
  54  */
  55 static bool
  56 ignore_dep(struct ir3_instruction *assigner,
  57                 struct ir3_instruction *consumer, unsigned n)
  58 {
  59         if (!__is_false_dep(consumer, n))
  60                 return false;
  61
  62         if (assigner->barrier_class & IR3_BARRIER_ARRAY_W) {
  63                 struct ir3_register *dst = assigner->regs[0];
  64                 struct ir3_register *src;
  65
  66                 debug_assert(dst->flags & IR3_REG_ARRAY);
  67
  68                 foreach_src(src, consumer) {
  69                         if ((src->flags & IR3_REG_ARRAY) &&
  70                                         (dst->array.id == src->array.id)) {
  71                                 return false;
  72                         }
  73                 }
  74         }
  75
  76         return true;
  77 }
  78
  79 /* calculate required # of delay slots between the instruction that
  80  * assigns a value and the one that consumes
  81  */
  82 int ir3_delayslots(struct ir3_instruction *assigner,
  83                 struct ir3_instruction *consumer, unsigned n)
  84 {
  85         if (ignore_dep(assigner, consumer, n))
  86                 return 0;
  87
  88         /* worst case is cat1-3 (alu) -> cat4/5 needing 6 cycles, normal
  89          * alu -> alu needs 3 cycles, cat4 -> alu and texture fetch
  90          * handled with sync bits
  91          */
  92
  93         if (is_meta(assigner))
  94                 return 0;
  95
  96         if (writes_addr(assigner))
  97                 return 6;
  98
  99         /* handled via sync flags: */
 100         if (is_sfu(assigner) || is_tex(assigner) || is_mem(assigner))
 101                 return 0;
 102
 103         /* assigner must be alu: */
 104         if (is_flow(consumer) || is_sfu(consumer) || is_tex(consumer) ||
 105                         is_mem(consumer)) {
 106                 return 6;
 107         } else if ((is_mad(consumer->opc) || is_madsh(consumer->opc)) &&
 108                         (n == 3)) {
 109                 /* special case, 3rd src to cat3 not required on first cycle */
 110                 return 1;
 111         } else {
 112                 return 3;
 113         }
 114 }
 115
 116 void
 117 ir3_insert_by_depth(struct ir3_instruction *instr, struct list_head *list)
 118 {
 119         /* remove from existing spot in list: */
 120         list_delinit(&instr->node);
 121
 122         /* find where to re-insert instruction: */
 123         list_for_each_entry (struct ir3_instruction, pos, list, node) {
 124                 if (pos->depth > instr->depth) {
 125                         list_add(&instr->node, &pos->node);
 126                         return;
 127                 }
 128         }
 129         /* if we get here, we didn't find an insertion spot: */
 130         list_addtail(&instr->node, list);
 131 }
 132
 133 static void
 134 ir3_instr_depth(struct ir3_instruction *instr, unsigned boost, bool falsedep)
 135 {
 136         struct ir3_instruction *src;
 137
 138         /* don't mark falsedep's as used, but otherwise process them normally: */
 139         if (!falsedep)
 140                 instr->flags &= ~IR3_INSTR_UNUSED;
 141
 142         if (ir3_instr_check_mark(instr))
 143                 return;
 144
 145         instr->depth = 0;
 146
 147         foreach_ssa_src_n(src, i, instr) {
 148                 unsigned sd;
 149
 150                 /* visit child to compute it's depth: */
 151                 ir3_instr_depth(src, boost, __is_false_dep(instr, i));
 152
 153                 /* for array writes, no need to delay on previous write: */
 154                 if (i == 0)
 155                         continue;
 156
 157                 sd = ir3_delayslots(src, instr, i) + src->depth;
 158                 sd += boost;
 159
 160                 instr->depth = MAX2(instr->depth, sd);
 161         }
 162
 163         if (!is_meta(instr))
 164                 instr->depth++;
 165
 166         ir3_insert_by_depth(instr, &instr->block->instr_list);
 167 }
 168
 169 static bool
 170 remove_unused_by_block(struct ir3_block *block)
 171 {
 172         bool progress = false;
 173         list_for_each_entry_safe (struct ir3_instruction, instr, &block->instr_list, node) {
 174                 if (instr->opc == OPC_END)
 175                         continue;
 176                 if (instr->flags & IR3_INSTR_UNUSED) {
 177                         list_delinit(&instr->node);
 178                         progress = true;
 179                 }
 180         }
 181         return progress;
 182 }
 183
 184 static bool
 185 compute_depth_and_remove_unused(struct ir3 *ir)
 186 {
 187         unsigned i;
 188         bool progress = false;
 189
 190         ir3_clear_mark(ir);
 191
 192         /* initially mark everything as unused, we'll clear the flag as we
 193          * visit the instructions:
 194          */
 195         list_for_each_entry (struct ir3_block, block, &ir->block_list, node) {
 196                 list_for_each_entry (struct ir3_instruction, instr, &block->instr_list, node) {
 197                         instr->flags |= IR3_INSTR_UNUSED;
 198                 }
 199         }
 200
 201         for (i = 0; i < ir->noutputs; i++)
 202                 if (ir->outputs[i])
 203                         ir3_instr_depth(ir->outputs[i], 0, false);
 204
 205         list_for_each_entry (struct ir3_block, block, &ir->block_list, node) {
 206                 for (i = 0; i < block->keeps_count; i++)
 207                         ir3_instr_depth(block->keeps[i], 0, false);
 208
 209                 /* We also need to account for if-condition: */
 210                 if (block->condition)
 211                         ir3_instr_depth(block->condition, 6, false);
 212         }
 213
 214         /* mark un-used instructions: */
 215         list_for_each_entry (struct ir3_block, block, &ir->block_list, node) {
 216                 progress |= remove_unused_by_block(block);
 217         }
 218
 219         /* note that we can end up with unused indirects, but we should
 220          * not end up with unused predicates.
 221          */
 222         for (i = 0; i < ir->indirects_count; i++) {
 223                 struct ir3_instruction *instr = ir->indirects[i];
 224                 if (instr && (instr->flags & IR3_INSTR_UNUSED))
 225                         ir->indirects[i] = NULL;
 226         }
 227
 228         /* cleanup unused inputs: */
 229         for (i = 0; i < ir->ninputs; i++) {
 230                 struct ir3_instruction *in = ir->inputs[i];
 231                 if (in && (in->flags & IR3_INSTR_UNUSED))
 232                         ir->inputs[i] = NULL;
 233         }
 234
 235         return progress;
 236 }
 237
 238 void
 239 ir3_depth(struct ir3 *ir)
 240 {
 241         bool progress;
 242         do {
 243                 progress = compute_depth_and_remove_unused(ir);
 244         } while (progress);
 245 }