aco: ignore parallelcopies to the same register on jump threading

The more conservative lowering to CSSA inserts unnecessary parallelcopies
which might get coalesced and can be ignored on jump threading.

v2: outline is_empty_block() check.

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Tested-by: Marge Bot <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3385>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3385>

diff --git a/src/amd/compiler/aco_ssa_elimination.cpp b/src/amd/compiler/aco_ssa_elimination.cpp
index 95e2e842a9c6ae64c4ec35d015ec4af978ab8298..94f5a1ade8384ba0816eb9ef0464375b939c0b2c 100644 (file)
--- a/src/amd/compiler/aco_ssa_elimination.cpp
+++ b/src/amd/compiler/aco_ssa_elimination.cpp
@@ -111,6 +111,35 @@ void insert_parallelcopies(ssa_elimination_ctx& ctx)
    }
 }
 
+bool is_empty_block(Block* block, bool ignore_exec_writes)
+{
+   /* check if this block is empty and the exec mask is not needed */
+   for (aco_ptr<Instruction>& instr : block->instructions) {
+      switch (instr->opcode) {
+         case aco_opcode::p_linear_phi:
+         case aco_opcode::p_phi:
+         case aco_opcode::p_logical_start:
+         case aco_opcode::p_logical_end:
+         case aco_opcode::p_branch:
+            break;
+         case aco_opcode::p_parallelcopy:
+            for (unsigned i = 0; i < instr->definitions.size(); i++) {
+               if (ignore_exec_writes && instr->definitions[i].physReg() == exec)
+                  continue;
+               if (instr->definitions[i].physReg() != instr->operands[i].physReg())
+                  return false;
+            }
+            break;
+         case aco_opcode::s_andn2_b64:
+         case aco_opcode::s_andn2_b32:
+            if (ignore_exec_writes && instr->definitions[0].physReg() == exec)
+               break;
+         default:
+            return false;
+      }
+   }
+   return true;
+}
 
 void try_remove_merge_block(ssa_elimination_ctx& ctx, Block* block)
 {
@@ -120,22 +149,9 @@ void try_remove_merge_block(ssa_elimination_ctx& ctx, Block* block)
        !(ctx.program->blocks[block->linear_succs[0]].kind & block_kind_merge))
       return;
 
-   /* check if this block is empty and the exec mask is not needed */
-   for (aco_ptr<Instruction>& instr : block->instructions) {
-      if (instr->opcode == aco_opcode::p_parallelcopy) {
-         if (instr->definitions[0].physReg() == exec)
-            continue;
-         else
-            return;
-      }
-
-      if (instr->opcode != aco_opcode::p_linear_phi &&
-          instr->opcode != aco_opcode::p_phi &&
-          instr->opcode != aco_opcode::p_logical_start &&
-          instr->opcode != aco_opcode::p_logical_end &&
-          instr->opcode != aco_opcode::p_branch)
-         return;
-   }
+   /* check if this block is empty */
+   if (!is_empty_block(block, true))
+      return;
 
    /* keep the branch instruction and remove the rest */
    aco_ptr<Instruction> branch = std::move(block->instructions.back());
@@ -146,18 +162,13 @@ void try_remove_merge_block(ssa_elimination_ctx& ctx, Block* block)
 void try_remove_invert_block(ssa_elimination_ctx& ctx, Block* block)
 {
    assert(block->linear_succs.size() == 2);
+   /* only remove this block if the successor got removed as well */
    if (block->linear_succs[0] != block->linear_succs[1])
       return;
 
-   /* check if we can remove this block */
-   for (aco_ptr<Instruction>& instr : block->instructions) {
-      if (instr->opcode != aco_opcode::p_linear_phi &&
-          instr->opcode != aco_opcode::p_phi &&
-          (instr->opcode != aco_opcode::s_andn2_b64 || ctx.program->wave_size != 64) &&
-          (instr->opcode != aco_opcode::s_andn2_b32 || ctx.program->wave_size != 32) &&
-          instr->opcode != aco_opcode::p_branch)
-         return;
-   }
+   /* check if block is otherwise empty */
+   if (!is_empty_block(block, true))
+      return;
 
    unsigned succ_idx = block->linear_succs[0];
    assert(block->linear_preds.size() == 2);
@@ -179,12 +190,8 @@ void try_remove_invert_block(ssa_elimination_ctx& ctx, Block* block)
 
 void try_remove_simple_block(ssa_elimination_ctx& ctx, Block* block)
 {
-   for (aco_ptr<Instruction>& instr : block->instructions) {
-      if (instr->opcode != aco_opcode::p_logical_start &&
-          instr->opcode != aco_opcode::p_logical_end &&
-          instr->opcode != aco_opcode::p_branch)
-         return;
-   }
+   if (!is_empty_block(block, false))
+      return;
 
    Block& pred = ctx.program->blocks[block->linear_preds[0]];
    Block& succ = ctx.program->blocks[block->linear_succs[0]];