aco: rework barriers and replace can_reorder

[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 f66521399ca6523220ac91d8abca52201575cc34..c1f0bc20f9e41a38acd841cda76952346c1d235c 100644 (file)
--- a/src/amd/compiler/aco_optimizer.cpp
+++ b/src/amd/compiler/aco_optimizer.cpp
@@ -717,9 +717,6 @@ bool check_vop3_operands(opt_ctx& ctx, unsigned num_operands, Operand *operands)
 
 bool parse_base_offset(opt_ctx &ctx, Instruction* instr, unsigned op_index, Temp *base, uint32_t *offset, bool prevent_overflow)
 {
-   if (prevent_overflow)
-      return false; //TODO
-
    Operand op = instr->operands[op_index];
 
    if (!op.isTemp())
@@ -740,6 +737,8 @@ bool parse_base_offset(opt_ctx &ctx, Instruction* instr, unsigned op_index, Temp
    default:
       return false;
    }
+   if (prevent_overflow && !add_instr->definitions[0].isNUW())
+      return false;
 
    if (add_instr->usesModifiers())
       return false;
@@ -1001,13 +1000,14 @@ void label_instruction(opt_ctx &ctx, Block& block, aco_ptr<Instruction>& instr)
          SMEM_instruction *smem = static_cast<SMEM_instruction *>(instr.get());
          Temp base;
          uint32_t offset;
+         bool prevent_overflow = smem->operands[0].size() > 2 || smem->prevent_overflow;
          if (i == 1 && info.is_constant_or_literal(32) &&
              ((ctx.program->chip_class == GFX6 && info.val <= 0x3FF) ||
               (ctx.program->chip_class == GFX7 && info.val <= 0xFFFFFFFF) ||
               (ctx.program->chip_class >= GFX8 && info.val <= 0xFFFFF))) {
             instr->operands[i] = Operand(info.val);
             continue;
-         } else if (i == 1 && parse_base_offset(ctx, instr.get(), i, &base, &offset, true) && base.regClass() == s1 && offset <= 0xFFFFF && ctx.program->chip_class >= GFX9) {
+         } else if (i == 1 && parse_base_offset(ctx, instr.get(), i, &base, &offset, prevent_overflow) && base.regClass() == s1 && offset <= 0xFFFFF && ctx.program->chip_class >= GFX9) {
             bool soe = smem->operands.size() >= (!smem->definitions.empty() ? 3 : 4);
             if (soe &&
                 (!ctx.info[smem->operands.back().tempId()].is_constant_or_literal(32) ||
@@ -1026,8 +1026,7 @@ void label_instruction(opt_ctx &ctx, Block& block, aco_ptr<Instruction>& instr)
                new_instr->operands.back() = Operand(base);
                if (!smem->definitions.empty())
                   new_instr->definitions[0] = smem->definitions[0];
-               new_instr->can_reorder = smem->can_reorder;
-               new_instr->barrier = smem->barrier;
+               new_instr->sync = smem->sync;
                new_instr->glc = smem->glc;
                new_instr->dlc = smem->dlc;
                new_instr->nv = smem->nv;