nir/vtn: Add type constant to image intrinsics
[mesa.git] / src / compiler / nir / nir_opt_if.c
index 75e4244cb2f56a360e048b77e22f288d2b9c8ce8..8b35b167013c1f87f3e29049f3eb6d644b403412 100644 (file)
@@ -58,25 +58,23 @@ find_continue_block(nir_loop *loop)
 static bool
 phi_has_constant_from_outside_and_one_from_inside_loop(nir_phi_instr *phi,
                                                        const nir_block *entry_block,
-                                                       uint32_t *entry_val,
-                                                       uint32_t *continue_val)
+                                                       bool *entry_val,
+                                                       bool *continue_val)
 {
    /* We already know we have exactly one continue */
    assert(exec_list_length(&phi->srcs) == 2);
 
-   *entry_val = 0;
-   *continue_val = 0;
+   *entry_val = false;
+   *continue_val = false;
 
     nir_foreach_phi_src(src, phi) {
-       assert(src->src.is_ssa);
-       nir_const_value *const_src = nir_src_as_const_value(src->src);
-       if (!const_src)
-          return false;
+       if (!nir_src_is_const(src->src))
+         return false;
 
        if (src->pred != entry_block) {
-          *continue_val = const_src[0].u32;
+          *continue_val = nir_src_as_bool(src->src);
        } else {
-          *entry_val = const_src[0].u32;
+          *entry_val = nir_src_as_bool(src->src);
        }
     }
 
@@ -169,7 +167,7 @@ opt_peel_loop_initial_if(nir_loop *loop)
    if (cond->parent_instr->block != header_block)
       return false;
 
-   uint32_t entry_val = 0, continue_val = 0;
+   bool entry_val = false, continue_val = false;
    if (!phi_has_constant_from_outside_and_one_from_inside_loop(cond_phi,
                                                                prev_block,
                                                                &entry_val,
@@ -277,7 +275,7 @@ alu_instr_is_comparison(const nir_alu_instr *alu)
    case nir_op_flt32:
    case nir_op_fge32:
    case nir_op_feq32:
-   case nir_op_fne32:
+   case nir_op_fneu32:
    case nir_op_ilt32:
    case nir_op_ult32:
    case nir_op_ige32:
@@ -294,8 +292,7 @@ static bool
 alu_instr_is_type_conversion(const nir_alu_instr *alu)
 {
    return nir_op_infos[alu->op].num_inputs == 1 &&
-          nir_alu_type_get_base_type(nir_op_infos[alu->op].output_type) !=
-          nir_alu_type_get_base_type(nir_op_infos[alu->op].input_types[0]);
+          nir_op_infos[alu->op].output_type != nir_op_infos[alu->op].input_types[0];
 }
 
 /**
@@ -632,7 +629,7 @@ opt_simplify_bcsel_of_phi(nir_builder *b, nir_loop *loop)
     * bcsel that must come before any break.
     *
     * For more details, see
-    * https://gitlab.freedesktop.org/mesa/mesa/merge_requests/170#note_110305
+    * https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/170#note_110305
     */
    nir_foreach_instr_safe(instr, header_block) {
       if (instr->type != nir_instr_type_alu)
@@ -665,7 +662,7 @@ opt_simplify_bcsel_of_phi(nir_builder *b, nir_loop *loop)
       nir_phi_instr *const cond_phi =
          nir_instr_as_phi(bcsel->src[0].src.ssa->parent_instr);
 
-      uint32_t entry_val = 0, continue_val = 0;
+      bool entry_val = false, continue_val = false;
       if (!phi_has_constant_from_outside_and_one_from_inside_loop(cond_phi,
                                                                   prev_block,
                                                                   &entry_val,
@@ -933,6 +930,17 @@ opt_if_simplification(nir_builder *b, nir_if *nif)
    nir_block *then_block = nir_if_last_then_block(nif);
    nir_block *else_block = nir_if_last_else_block(nif);
 
+   if (nir_block_ends_in_jump(else_block)) {
+      /* Even though this if statement has a jump on one side, we may still have
+       * phis afterwards.  Single-source phis can be produced by loop unrolling
+       * or dead control-flow passes and are perfectly legal.  Run a quick phi
+       * removal on the block after the if to clean up any such phis.
+       */
+      nir_block *const next_block =
+         nir_cf_node_as_block(nir_cf_node_next(&nif->cf_node));
+      nir_opt_remove_phis_block(next_block);
+   }
+
    rewrite_phi_predecessor_blocks(nif, then_block, else_block, else_block,
                                   then_block);
 
@@ -998,7 +1006,7 @@ opt_if_loop_terminator(nir_if *nif)
    if (is_block_empty(first_continue_from_blk))
       return false;
 
-   if (!nir_is_trivial_loop_if(nif, break_blk))
+   if (nir_block_ends_in_jump(continue_from_blk))
       return false;
 
    /* Even though this if statement has a jump on one side, we may still have
@@ -1334,7 +1342,6 @@ opt_if_cf_list(nir_builder *b, struct exec_list *cf_list,
          progress |= opt_if_cf_list(b, &loop->body,
                                     aggressive_last_continue);
          progress |= opt_simplify_bcsel_of_phi(b, loop);
-         progress |= opt_peel_loop_initial_if(loop);
          progress |= opt_if_loop_last_continue(loop,
                                                aggressive_last_continue);
          break;
@@ -1348,6 +1355,37 @@ opt_if_cf_list(nir_builder *b, struct exec_list *cf_list,
    return progress;
 }
 
+static bool
+opt_peel_loop_initial_if_cf_list(struct exec_list *cf_list)
+{
+   bool progress = false;
+   foreach_list_typed(nir_cf_node, cf_node, node, cf_list) {
+      switch (cf_node->type) {
+      case nir_cf_node_block:
+         break;
+
+      case nir_cf_node_if: {
+         nir_if *nif = nir_cf_node_as_if(cf_node);
+         progress |= opt_peel_loop_initial_if_cf_list(&nif->then_list);
+         progress |= opt_peel_loop_initial_if_cf_list(&nif->else_list);
+         break;
+      }
+
+      case nir_cf_node_loop: {
+         nir_loop *loop = nir_cf_node_as_loop(cf_node);
+         progress |= opt_peel_loop_initial_if_cf_list(&loop->body);
+         progress |= opt_peel_loop_initial_if(loop);
+         break;
+      }
+
+      case nir_cf_node_function:
+         unreachable("Invalid cf type");
+      }
+   }
+
+   return progress;
+}
+
 /**
  * These optimisations depend on nir_metadata_block_index and therefore must
  * not do anything to cause the metadata to become invalid.
@@ -1402,21 +1440,28 @@ nir_opt_if(nir_shader *shader, bool aggressive_last_continue)
       nir_metadata_preserve(function->impl, nir_metadata_block_index |
                             nir_metadata_dominance);
 
-      if (opt_if_cf_list(&b, &function->impl->body,
-                         aggressive_last_continue)) {
-         nir_metadata_preserve(function->impl, nir_metadata_none);
+      bool preserve = true;
+
+      if (opt_if_cf_list(&b, &function->impl->body, aggressive_last_continue)) {
+         preserve = false;
+         progress = true;
+      }
+
+      if (opt_peel_loop_initial_if_cf_list(&function->impl->body)) {
+         preserve = false;
+         progress = true;
 
          /* If that made progress, we're no longer really in SSA form.  We
           * need to convert registers back into SSA defs and clean up SSA defs
           * that don't dominate their uses.
           */
          nir_lower_regs_to_ssa_impl(function->impl);
+      }
 
-         progress = true;
+      if (preserve) {
+         nir_metadata_preserve(function->impl, nir_metadata_none);
       } else {
-   #ifndef NDEBUG
-         function->impl->valid_metadata &= ~nir_metadata_not_properly_reset;
-   #endif
+         nir_metadata_preserve(function->impl, nir_metadata_all);
       }
    }