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 #include "ir3_shader.h"
  31
  32 /*
  33  * Instruction Depth:
  34  *
  35  * Calculates weighted instruction depth, ie. the sum of # of needed
  36  * instructions plus delay slots back to original input (ie INPUT or
  37  * CONST).  That is to say, an instructions depth is:
  38  *
  39  *   depth(instr) {
  40  *     d = 0;
  41  *     // for each src register:
  42  *     foreach (src in instr->regs[1..n])
  43  *       d = max(d, delayslots(src->instr, n) + depth(src->instr));
  44  *     return d + 1;
  45  *   }
  46  *
  47  * After an instruction's depth is calculated, it is inserted into the
  48  * blocks depth sorted list, which is used by the scheduling pass.
  49  */
  50
  51 void
  52 ir3_insert_by_depth(struct ir3_instruction *instr, struct list_head *list)
  53 {
  54         /* remove from existing spot in list: */
  55         list_delinit(&instr->node);
  56
  57         /* find where to re-insert instruction: */
  58         foreach_instr (pos, list) {
  59                 if (pos->depth > instr->depth) {
  60                         list_add(&instr->node, &pos->node);
  61                         return;
  62                 }
  63         }
  64         /* if we get here, we didn't find an insertion spot: */
  65         list_addtail(&instr->node, list);
  66 }
  67
  68 static void
  69 ir3_instr_depth(struct ir3_instruction *instr, unsigned boost, bool falsedep)
  70 {
  71         struct ir3_instruction *src;
  72
  73         /* don't mark falsedep's as used, but otherwise process them normally: */
  74         if (!falsedep)
  75                 instr->flags &= ~IR3_INSTR_UNUSED;
  76
  77         if (ir3_instr_check_mark(instr))
  78                 return;
  79
  80         instr->depth = 0;
  81
  82         foreach_ssa_src_n (src, i, instr) {
  83                 unsigned sd;
  84
  85                 /* visit child to compute it's depth: */
  86                 ir3_instr_depth(src, boost, __is_false_dep(instr, i));
  87
  88                 /* for array writes, no need to delay on previous write: */
  89                 if (i == 0)
  90                         continue;
  91
  92                 sd = ir3_delayslots(src, instr, i, true) + src->depth;
  93                 sd += boost;
  94
  95                 instr->depth = MAX2(instr->depth, sd);
  96         }
  97
  98         if (!is_meta(instr))
  99                 instr->depth++;
 100
 101         ir3_insert_by_depth(instr, &instr->block->instr_list);
 102 }
 103
 104 static bool
 105 remove_unused_by_block(struct ir3_block *block)
 106 {
 107         bool progress = false;
 108         foreach_instr_safe (instr, &block->instr_list) {
 109                 if (instr->opc == OPC_END || instr->opc == OPC_CHSH || instr->opc == OPC_CHMASK)
 110                         continue;
 111                 if (instr->flags & IR3_INSTR_UNUSED) {
 112                         if (instr->opc == OPC_META_SPLIT) {
 113                                 struct ir3_instruction *src = ssa(instr->regs[1]);
 114                                 /* tex (cat5) instructions have a writemask, so we can
 115                                  * mask off unused components.  Other instructions do not.
 116                                  */
 117                                 if (is_tex_or_prefetch(src) && (src->regs[0]->wrmask > 1)) {
 118                                         src->regs[0]->wrmask &= ~(1 << instr->split.off);
 119
 120                                         /* prune no-longer needed right-neighbors.  We could
 121                                          * probably do the same for left-neighbors (ie. tex
 122                                          * fetch that only need .yw components), but that
 123                                          * makes RA a bit more confusing than it already is
 124                                          */
 125                                         struct ir3_instruction *n = instr;
 126                                         while (n && n->cp.right)
 127                                                 n = n->cp.right;
 128                                         while (n->flags & IR3_INSTR_UNUSED) {
 129                                                 n = n->cp.left;
 130                                                 if (!n)
 131                                                         break;
 132                                                 n->cp.right = NULL;
 133                                         }
 134                                 }
 135                         }
 136                         list_delinit(&instr->node);
 137                         progress = true;
 138                 }
 139         }
 140         return progress;
 141 }
 142
 143 static bool
 144 compute_depth_and_remove_unused(struct ir3 *ir, struct ir3_shader_variant *so)
 145 {
 146         unsigned i;
 147         bool progress = false;
 148
 149         ir3_clear_mark(ir);
 150
 151         /* initially mark everything as unused, we'll clear the flag as we
 152          * visit the instructions:
 153          */
 154         foreach_block (block, &ir->block_list) {
 155                 foreach_instr (instr, &block->instr_list) {
 156                         /* special case, if pre-fs texture fetch used, we cannot
 157                          * eliminate the barycentric i/j input
 158                          */
 159                         if (so->num_sampler_prefetch &&
 160                                         (instr->opc == OPC_META_INPUT) &&
 161                                         (instr->input.sysval == SYSTEM_VALUE_BARYCENTRIC_PERSP_PIXEL))
 162                                 continue;
 163                         instr->flags |= IR3_INSTR_UNUSED;
 164                 }
 165         }
 166
 167         struct ir3_instruction *out;
 168         foreach_output (out, ir)
 169                 ir3_instr_depth(out, 0, false);
 170
 171         foreach_block (block, &ir->block_list) {
 172                 for (i = 0; i < block->keeps_count; i++)
 173                         ir3_instr_depth(block->keeps[i], 0, false);
 174
 175                 /* We also need to account for if-condition: */
 176                 if (block->condition)
 177                         ir3_instr_depth(block->condition, 6, false);
 178         }
 179
 180         /* remove un-used instructions: */
 181         foreach_block (block, &ir->block_list) {
 182                 progress |= remove_unused_by_block(block);
 183         }
 184
 185         /* fixup wrmask of split instructions to account for adjusted tex
 186          * wrmask's:
 187          */
 188         foreach_block (block, &ir->block_list) {
 189                 foreach_instr (instr, &block->instr_list) {
 190                         if (instr->opc != OPC_META_SPLIT)
 191                                 continue;
 192
 193                         struct ir3_instruction *src = ssa(instr->regs[1]);
 194                         if (!is_tex_or_prefetch(src))
 195                                 continue;
 196
 197                         instr->regs[1]->wrmask = src->regs[0]->wrmask;
 198                 }
 199         }
 200
 201         /* note that we can end up with unused indirects, but we should
 202          * not end up with unused predicates.
 203          */
 204         for (i = 0; i < ir->a0_users_count; i++) {
 205                 struct ir3_instruction *instr = ir->a0_users[i];
 206                 if (instr && (instr->flags & IR3_INSTR_UNUSED))
 207                         ir->a0_users[i] = NULL;
 208         }
 209
 210         for (i = 0; i < ir->a1_users_count; i++) {
 211                 struct ir3_instruction *instr = ir->a1_users[i];
 212                 if (instr && (instr->flags & IR3_INSTR_UNUSED))
 213                         ir->a1_users[i] = NULL;
 214         }
 215
 216         /* cleanup unused inputs: */
 217         struct ir3_instruction *in;
 218         foreach_input_n (in, n, ir)
 219                 if (in->flags & IR3_INSTR_UNUSED)
 220                         ir->inputs[n] = NULL;
 221
 222         return progress;
 223 }
 224
 225 void
 226 ir3_depth(struct ir3 *ir, struct ir3_shader_variant *so)
 227 {
 228         bool progress;
 229         do {
 230                 progress = compute_depth_and_remove_unused(ir, so);
 231         } while (progress);
 232 }