diff options
author | Daniel Schürmann <daniel@schuermann.dev> | 2020-05-27 11:08:31 +0100 |
---|---|---|
committer | Marge Bot <eric+marge@anholt.net> | 2020-06-09 21:25:38 +0000 |
commit | 9e8e12ea6d20763aa1c819a7be4b9a6158df9a37 (patch) | |
tree | ae18db227d974eae8be91c61a6049b64e92a96a2 | |
parent | b083581010d29b2604ad34e0dd7ff6dae4417dd2 (diff) |
aco: adjust GFX6 subdword lowering workarounds for 8bit
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5226>
-rw-r--r-- | src/amd/compiler/aco_lower_to_hw_instr.cpp | 38 |
1 files changed, 21 insertions, 17 deletions
diff --git a/src/amd/compiler/aco_lower_to_hw_instr.cpp b/src/amd/compiler/aco_lower_to_hw_instr.cpp index 8bf57770ce5..f0d6ceecc46 100644 --- a/src/amd/compiler/aco_lower_to_hw_instr.cpp +++ b/src/amd/compiler/aco_lower_to_hw_instr.cpp @@ -1255,39 +1255,43 @@ void handle_operands(std::map<PhysReg, copy_operation>& copy_map, lower_context* /* on GFX6/7, we need some small workarounds as there is no * SDWA instruction to do partial register writes */ - if (ctx->program->chip_class < GFX8 && it->second.bytes == 2) { + if (ctx->program->chip_class < GFX8 && it->second.bytes < 4) { if (it->first.byte() == 0 && it->second.op.physReg().byte() == 0 && !it->second.is_used && pi->opcode == aco_opcode::p_split_vector) { /* Other operations might overwrite the high bits, so change all users * of the high bits to the new target where they are still available. * This mechanism depends on also emitting dead definitions. */ - PhysReg reg_hi = it->second.op.physReg().advance(2); - std::map<PhysReg, copy_operation>::iterator other = copy_map.begin(); - for (other = copy_map.begin(); other != copy_map.end(); other++) { - /* on GFX6/7, if the high bits are used as operand, they cannot be a target */ - if (other->second.op.physReg() == reg_hi) { - other->second.op.setFixed(it->first.advance(2)); - break; /* break because an operand can only be used once */ + PhysReg reg_hi = it->second.op.physReg().advance(it->second.bytes); + while (reg_hi != PhysReg(it->second.op.physReg().reg() + 1)) { + std::map<PhysReg, copy_operation>::iterator other = copy_map.begin(); + for (other = copy_map.begin(); other != copy_map.end(); other++) { + /* on GFX6/7, if the high bits are used as operand, they cannot be a target */ + if (other->second.op.physReg() == reg_hi) { + other->second.op.setFixed(it->first.advance(reg_hi.byte())); + break; /* break because an operand can only be used once */ + } } + reg_hi = reg_hi.advance(it->second.bytes); } - } else if (it->first.byte() == 2) { + } else if (it->first.byte()) { + assert(pi->opcode == aco_opcode::p_create_vector); /* on GFX6/7, if we target an upper half where the lower half hasn't yet been handled, * move to the target operand's high bits. This is save to do as it cannot be an operand */ PhysReg lo = PhysReg(it->first.reg()); std::map<PhysReg, copy_operation>::iterator other = copy_map.find(lo); if (other != copy_map.end()) { - PhysReg new_reg_hi = other->second.op.physReg().advance(2); - assert(other->second.bytes == 2 && new_reg_hi.byte() == 2); - it->second.def = Definition(new_reg_hi, v2b); + assert(other->second.bytes == it->first.byte()); + PhysReg new_reg_hi = other->second.op.physReg().advance(it->first.byte()); + it->second.def = Definition(new_reg_hi, it->second.def.regClass()); it->second.is_used = 0; - other->second.bytes = 4; - other->second.def.setTemp(Temp(other->second.def.tempId(), v1)); - other->second.op.setTemp(Temp(other->second.op.tempId(), v1)); + other->second.bytes += it->second.bytes; + other->second.def.setTemp(Temp(other->second.def.tempId(), RegClass::get(RegType::vgpr, other->second.bytes))); + other->second.op.setTemp(Temp(other->second.op.tempId(), RegClass::get(RegType::vgpr, other->second.bytes))); /* if the new target's high bits are also a target, change uses */ std::map<PhysReg, copy_operation>::iterator target = copy_map.find(new_reg_hi); if (target != copy_map.end()) { - target->second.uses[0]++; - target->second.uses[1]++; + for (unsigned i = 0; i < it->second.bytes; i++) + target->second.uses[i]++; } } } |