nir/opt_remove_phis: optimize out phis with undef

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

diff --git a/src/compiler/nir/nir_opt_remove_phis.c b/src/compiler/nir/nir_opt_remove_phis.c
index 5234c0f667ad9154be5451e561c3bb6e7a1b2176..db58c618f6bd2070fecea18055dd47d9bb6144eb 100644 (file)
--- a/src/compiler/nir/nir_opt_remove_phis.c
+++ b/src/compiler/nir/nir_opt_remove_phis.c
@@ -26,12 +26,33 @@
  */
 
 #include "nir.h"
+#include "nir_builder.h"
+
+static nir_alu_instr *
+get_parent_mov(nir_ssa_def *ssa)
+{
+   if (ssa->parent_instr->type != nir_instr_type_alu)
+      return NULL;
+
+   nir_alu_instr *alu = nir_instr_as_alu(ssa->parent_instr);
+   return (alu->op == nir_op_mov) ? alu : NULL;
+}
+
+static bool
+matching_mov(nir_alu_instr *mov1, nir_ssa_def *ssa)
+{
+   if (!mov1)
+      return false;
+
+   nir_alu_instr *mov2 = get_parent_mov(ssa);
+   return mov2 && nir_alu_srcs_equal(mov1, mov2, 0, 0);
+}
 
 /*
  * This is a pass for removing phi nodes that look like:
  * a = phi(b, b, b, ...)
  *
- * Note that we can't ignore undef sources here, or else we may create a
+ * Note that we can't always ignore undef sources here, or else we may create a
  * situation where the definition of b isn't dominated by its uses. We're
  * allowed to do this since the definition of b must dominate all of the
  * phi node's predecessors, which means it must dominate the phi node as well
@@ -43,7 +64,7 @@
  */
 
 static bool
-remove_phis_block(nir_block *block)
+remove_phis_block(nir_block *block, nir_builder *b)
 {
    bool progress = false;
 
@@ -54,9 +75,10 @@ remove_phis_block(nir_block *block)
       nir_phi_instr *phi = nir_instr_as_phi(instr);
 
       nir_ssa_def *def = NULL;
+      nir_alu_instr *mov = NULL;
       bool srcs_same = true;
 
-      nir_foreach_phi_src(phi, src) {
+      nir_foreach_phi_src(src, phi) {
          assert(src->src.is_ssa);
 
          /* For phi nodes at the beginning of loops, we may encounter some
@@ -75,8 +97,12 @@ remove_phis_block(nir_block *block)
          
          if (def == NULL) {
             def  = src->src.ssa;
+            mov = get_parent_mov(def);
+         } else if (src->src.ssa->parent_instr->type == nir_instr_type_ssa_undef &&
+                    nir_block_dominates(def->parent_instr->block, src->pred)) {
+            /* Ignore this undef source. */
          } else {
-            if (src->src.ssa != def) {
+            if (src->src.ssa != def && !matching_mov(mov, src->src.ssa)) {
                srcs_same = false;
                break;
             }
@@ -86,10 +112,24 @@ remove_phis_block(nir_block *block)
       if (!srcs_same)
          continue;
 
-      /* We must have found at least one definition, since there must be at
-       * least one forward edge.
-       */
-      assert(def != NULL);
+      if (!def) {
+         /* In this case, the phi had no sources. So turn it into an undef. */
+
+         b->cursor = nir_after_phis(block);
+         def = nir_ssa_undef(b, phi->dest.ssa.num_components,
+                             phi->dest.ssa.bit_size);
+      } else if (mov) {
+         /* If the sources were all movs from the same source with the same
+          * swizzle, then we can't just pick a random move because it may not
+          * dominate the phi node. Instead, we need to emit our own move after
+          * the phi which uses the shared source, and rewrite uses of the phi
+          * to use the move instead. This is ok, because while the movs may
+          * not all dominate the phi node, their shared source does.
+          */
+
+         b->cursor = nir_after_phis(block);
+         def = nir_mov_alu(b, mov->src[0], def->num_components);
+      }
 
       assert(phi->dest.is_ssa);
       nir_ssa_def_rewrite_uses(&phi->dest.ssa, nir_src_for_ssa(def));
@@ -101,18 +141,32 @@ remove_phis_block(nir_block *block)
    return progress;
 }
 
+bool
+nir_opt_remove_phis_block(nir_block *block)
+{
+   nir_builder b;
+   nir_builder_init(&b, nir_cf_node_get_function(&block->cf_node));
+   return remove_phis_block(block, &b);
+}
+
 static bool
-remove_phis_impl(nir_function_impl *impl)
+nir_opt_remove_phis_impl(nir_function_impl *impl)
 {
    bool progress = false;
+   nir_builder bld;
+   nir_builder_init(&bld, impl);
+
+   nir_metadata_require(impl, nir_metadata_dominance);
 
    nir_foreach_block(block, impl) {
-      progress |= remove_phis_block(block);
+      progress |= remove_phis_block(block, &bld);
    }
 
    if (progress) {
       nir_metadata_preserve(impl, nir_metadata_block_index |
                                   nir_metadata_dominance);
+   } else {
+      nir_metadata_preserve(impl, nir_metadata_all);
    }
 
    return progress;
@@ -123,9 +177,9 @@ nir_opt_remove_phis(nir_shader *shader)
 {
    bool progress = false;
 
-   nir_foreach_function(shader, function)
+   nir_foreach_function(function, shader)
       if (function->impl)
-         progress = remove_phis_impl(function->impl) || progress;
+         progress = nir_opt_remove_phis_impl(function->impl) || progress;
 
    return progress;
 }