nir: define behavior of nir_op_bfm and nir_op_u/ibfe according to SM5 spec.

[mesa.git] / src / compiler / nir / nir_opt_gcm.c

diff --git a/src/compiler/nir/nir_opt_gcm.c b/src/compiler/nir/nir_opt_gcm.c
index 33278d8b1fcca89ee4094231d9204adb6667ed40..aeae2ad6401074146cbc4c1285706e6cad657425 100644 (file)
--- a/src/compiler/nir/nir_opt_gcm.c
+++ b/src/compiler/nir/nir_opt_gcm.c
@@ -26,6 +26,7 @@
  */
 
 #include "nir.h"
+#include "nir_instr_set.h"
 
 /*
  * Implements Global Code Motion.  A description of GCM can be found in
@@ -107,7 +108,7 @@ gcm_build_block_info(struct exec_list *cf_list, struct gcm_state *state,
 static bool
 gcm_pin_instructions_block(nir_block *block, struct gcm_state *state)
 {
-   nir_foreach_instr_safe(block, instr) {
+   nir_foreach_instr_safe(instr, block) {
       switch (instr->type) {
       case nir_instr_type_alu:
          switch (nir_instr_as_alu(instr)->op) {
@@ -127,6 +128,10 @@ gcm_pin_instructions_block(nir_block *block, struct gcm_state *state)
          }
          break;
 
+      case nir_instr_type_deref:
+         instr->pass_flags = 0;
+         break;
+
       case nir_instr_type_tex:
          switch (nir_instr_as_tex(instr)->op) {
          case nir_texop_tex:
@@ -147,11 +152,7 @@ gcm_pin_instructions_block(nir_block *block, struct gcm_state *state)
          break;
 
       case nir_instr_type_intrinsic: {
-         const nir_intrinsic_info *info =
-            &nir_intrinsic_infos[nir_instr_as_intrinsic(instr)->intrinsic];
-
-         if ((info->flags & NIR_INTRINSIC_CAN_ELIMINATE) &&
-             (info->flags & NIR_INTRINSIC_CAN_REORDER)) {
+         if (nir_intrinsic_can_reorder(nir_instr_as_intrinsic(instr))) {
             instr->pass_flags = 0;
          } else {
             instr->pass_flags = GCM_INSTR_PINNED;
@@ -277,7 +278,7 @@ gcm_schedule_late_def(nir_ssa_def *def, void *void_state)
 
    nir_block *lca = NULL;
 
-   nir_foreach_use(def, use_src) {
+   nir_foreach_use(use_src, def) {
       nir_instr *use_instr = use_src->parent_instr;
 
       gcm_schedule_late_instr(use_instr, state);
@@ -292,7 +293,7 @@ gcm_schedule_late_def(nir_ssa_def *def, void *void_state)
       if (use_instr->type == nir_instr_type_phi) {
          nir_phi_instr *phi = nir_instr_as_phi(use_instr);
 
-         nir_foreach_phi_src(phi, phi_src) {
+         nir_foreach_phi_src(phi_src, phi) {
             if (phi_src->src.ssa == def)
                lca = nir_dominance_lca(lca, phi_src->pred);
          }
@@ -301,7 +302,7 @@ gcm_schedule_late_def(nir_ssa_def *def, void *void_state)
       }
    }
 
-   nir_foreach_if_use(def, use_src) {
+   nir_foreach_if_use(use_src, def) {
       nir_if *if_stmt = use_src->parent_if;
 
       /* For if statements, we consider the block to be the one immediately
@@ -319,18 +320,19 @@ gcm_schedule_late_def(nir_ssa_def *def, void *void_state)
    if (lca == NULL)
       return true;
 
-   /* We know have the LCA of all of the uses.  If our invariants hold,
+   /* We now have the LCA of all of the uses.  If our invariants hold,
     * this is dominated by the block that we chose when scheduling early.
     * We now walk up the dominance tree and pick the lowest block that is
     * as far outside loops as we can get.
     */
    nir_block *best = lca;
-   while (lca != def->parent_instr->block) {
-      assert(lca);
-      if (state->blocks[lca->index].loop_depth <
+   for (nir_block *block = lca; block != NULL; block = block->imm_dom) {
+      if (state->blocks[block->index].loop_depth <
           state->blocks[best->index].loop_depth)
-         best = lca;
-      lca = lca->imm_dom;
+         best = block;
+
+      if (block == def->parent_instr->block)
+         break;
    }
    def->parent_instr->block = best;
 
@@ -374,7 +376,7 @@ gcm_place_instr(nir_instr *instr, struct gcm_state *state);
 static bool
 gcm_place_instr_def(nir_ssa_def *def, void *state)
 {
-   nir_foreach_use(def, use_src)
+   nir_foreach_use(use_src, def)
       gcm_place_instr(use_src->parent_instr, state);
 
    return false;
@@ -451,9 +453,12 @@ gcm_place_instr(nir_instr *instr, struct gcm_state *state)
    block_info->last_instr = instr;
 }
 
-static void
-opt_gcm_impl(nir_function_impl *impl)
+static bool
+opt_gcm_impl(nir_function_impl *impl, bool value_number)
 {
+   nir_metadata_require(impl, nir_metadata_block_index |
+                              nir_metadata_dominance);
+
    struct gcm_state state;
 
    state.impl = impl;
@@ -461,15 +466,24 @@ opt_gcm_impl(nir_function_impl *impl)
    exec_list_make_empty(&state.instrs);
    state.blocks = rzalloc_array(NULL, struct gcm_block_info, impl->num_blocks);
 
-   nir_metadata_require(impl, nir_metadata_block_index |
-                              nir_metadata_dominance);
-
    gcm_build_block_info(&impl->body, &state, 0);
 
    nir_foreach_block(block, impl) {
       gcm_pin_instructions_block(block, &state);
    }
 
+   bool progress = false;
+   if (value_number) {
+      struct set *gvn_set = nir_instr_set_create(NULL);
+      foreach_list_typed_safe(nir_instr, instr, node, &state.instrs) {
+         if (nir_instr_set_add_or_rewrite(gvn_set, instr)) {
+            nir_instr_remove(instr);
+            progress = true;
+         }
+      }
+      nir_instr_set_destroy(gvn_set);
+   }
+
    foreach_list_typed(nir_instr, instr, node, &state.instrs)
       gcm_schedule_early_instr(instr, &state);
 
@@ -478,18 +492,27 @@ opt_gcm_impl(nir_function_impl *impl)
 
    while (!exec_list_is_empty(&state.instrs)) {
       nir_instr *instr = exec_node_data(nir_instr,
-                                        state.instrs.tail_pred, node);
+                                        state.instrs.tail_sentinel.prev, node);
       gcm_place_instr(instr, &state);
    }
 
    ralloc_free(state.blocks);
+
+   nir_metadata_preserve(impl, nir_metadata_block_index |
+                               nir_metadata_dominance);
+
+   return progress;
 }
 
-void
-nir_opt_gcm(nir_shader *shader)
+bool
+nir_opt_gcm(nir_shader *shader, bool value_number)
 {
-   nir_foreach_function(shader, function) {
+   bool progress = false;
+
+   nir_foreach_function(function, shader) {
       if (function->impl)
-         opt_gcm_impl(function->impl);
+         progress |= opt_gcm_impl(function->impl, value_number);
    }
+
+   return progress;
 }