aco: allow an extra SGPR with multiple uses to be applied to VOP3

[mesa.git] / src / amd / compiler / aco_optimizer.cpp

diff --git a/src/amd/compiler/aco_optimizer.cpp b/src/amd/compiler/aco_optimizer.cpp
index 387d8c76c0a19c7b351c7220857717061f0cf46f..88075fabfb2df3bb60f6c400609b4a87535bbc2a 100644 (file)
--- a/src/amd/compiler/aco_optimizer.cpp
+++ b/src/amd/compiler/aco_optimizer.cpp
@@ -1394,7 +1394,8 @@ bool combine_constant_comparison_ordering(opt_ctx &ctx, aco_ptr<Instruction>& in
    if (cmp->operands[constant_operand].isConstant()) {
       constant = cmp->operands[constant_operand].constantValue();
    } else if (cmp->operands[constant_operand].isTemp()) {
-      unsigned id = cmp->operands[constant_operand].tempId();
+      Temp tmp = cmp->operands[constant_operand].getTemp();
+      unsigned id = original_temp_id(ctx, tmp);
       if (!ctx.info[id].is_constant() && !ctx.info[id].is_literal())
          return false;
       constant = ctx.info[id].val;
@@ -1678,6 +1679,10 @@ bool combine_salu_n2(opt_ctx& ctx, aco_ptr<Instruction>& instr)
       if (!op2_instr || (op2_instr->opcode != aco_opcode::s_not_b32 && op2_instr->opcode != aco_opcode::s_not_b64))
          continue;
 
+      if (instr->operands[!i].isLiteral() && op2_instr->operands[0].isLiteral() &&
+          instr->operands[!i].constantValue() != op2_instr->operands[0].constantValue())
+         continue;
+
       ctx.uses[instr->operands[i].tempId()]--;
       instr->operands[0] = instr->operands[!i];
       instr->operands[1] = op2_instr->operands[0];
@@ -1720,6 +1725,10 @@ bool combine_salu_lshl_add(opt_ctx& ctx, aco_ptr<Instruction>& instr)
       if (shift < 1 || shift > 4)
          continue;
 
+      if (instr->operands[!i].isLiteral() && op2_instr->operands[0].isLiteral() &&
+          instr->operands[!i].constantValue() != op2_instr->operands[0].constantValue())
+         continue;
+
       ctx.uses[instr->operands[i].tempId()]--;
       instr->operands[1] = instr->operands[!i];
       instr->operands[0] = op2_instr->operands[0];
@@ -1928,7 +1937,7 @@ void apply_sgprs(opt_ctx &ctx, aco_ptr<Instruction>& instr)
       /* Applying two sgprs require making it VOP3, so don't do it unless it's
        * definitively beneficial.
        * TODO: this is too conservative because later the use count could be reduced to 1 */
-      if (num_sgprs && ctx.uses[sgpr_info_id] > 1)
+      if (num_sgprs && ctx.uses[sgpr_info_id] > 1 && !instr->isVOP3())
          break;
 
       Temp sgpr = ctx.info[sgpr_info_id].temp;
@@ -1954,8 +1963,6 @@ void apply_sgprs(opt_ctx &ctx, aco_ptr<Instruction>& instr)
       sgpr_ids[num_sgprs++] = sgpr.id();
       ctx.uses[sgpr_info_id]--;
       ctx.uses[sgpr.id()]++;
-
-      break; /* for testing purposes, only apply 1 new sgpr */
    }
 }