for (size_t j = 0; j < block->instructions.size(); j++) {
aco_ptr<Instruction>& instr = block->instructions[j];
aco_ptr<Instruction> mov;
- if (instr->format == Format::PSEUDO) {
+ if (instr->format == Format::PSEUDO && instr->opcode != aco_opcode::p_unit_test) {
Pseudo_instruction *pi = (Pseudo_instruction*)instr.get();
switch (instr->opcode)
//TODO: exec can be zero here with block_kind_discard
assert(instr->operands[0].physReg() == scc);
- bld.sopp(aco_opcode::s_cbranch_scc0, instr->operands[0], discard_block->index);
+ bld.sopp(aco_opcode::s_cbranch_scc0, Definition(exec, s2), instr->operands[0], discard_block->index);
discard_block->linear_preds.push_back(block->index);
block->linear_succs.push_back(discard_block->index);
emit_gfx10_wave64_bpermute(program, instr, bld);
else
unreachable("Current hardware supports ds_bpermute, don't emit p_bpermute.");
+ break;
}
default:
break;
switch (instr->opcode) {
case aco_opcode::p_branch:
assert(block->linear_succs[0] == branch->target[0]);
- bld.sopp(aco_opcode::s_branch, branch->target[0]);
+ bld.sopp(aco_opcode::s_branch, branch->definitions[0], branch->target[0]);
break;
case aco_opcode::p_cbranch_nz:
assert(block->linear_succs[1] == branch->target[0]);
if (branch->operands[0].physReg() == exec)
- bld.sopp(aco_opcode::s_cbranch_execnz, branch->target[0]);
+ bld.sopp(aco_opcode::s_cbranch_execnz, branch->definitions[0], branch->target[0]);
else if (branch->operands[0].physReg() == vcc)
- bld.sopp(aco_opcode::s_cbranch_vccnz, branch->target[0]);
+ bld.sopp(aco_opcode::s_cbranch_vccnz, branch->definitions[0], branch->target[0]);
else {
assert(branch->operands[0].physReg() == scc);
- bld.sopp(aco_opcode::s_cbranch_scc1, branch->target[0]);
+ bld.sopp(aco_opcode::s_cbranch_scc1, branch->definitions[0], branch->target[0]);
}
break;
case aco_opcode::p_cbranch_z:
assert(block->linear_succs[1] == branch->target[0]);
if (branch->operands[0].physReg() == exec)
- bld.sopp(aco_opcode::s_cbranch_execz, branch->target[0]);
+ bld.sopp(aco_opcode::s_cbranch_execz, branch->definitions[0], branch->target[0]);
else if (branch->operands[0].physReg() == vcc)
- bld.sopp(aco_opcode::s_cbranch_vccz, branch->target[0]);
+ bld.sopp(aco_opcode::s_cbranch_vccz, branch->definitions[0], branch->target[0]);
else {
assert(branch->operands[0].physReg() == scc);
- bld.sopp(aco_opcode::s_cbranch_scc0, branch->target[0]);
+ bld.sopp(aco_opcode::s_cbranch_scc0, branch->definitions[0], branch->target[0]);
}
break;
default:
reduce->operands[2].physReg(), // vtmp
reduce->definitions[2].physReg(), // sitmp
reduce->operands[0], reduce->definitions[0]);
+ } else if (instr->format == Format::PSEUDO_BARRIER) {
+ Pseudo_barrier_instruction* barrier = static_cast<Pseudo_barrier_instruction*>(instr.get());
+
+ /* Anything larger than a workgroup isn't possible. Anything
+ * smaller requires no instructions and this pseudo instruction
+ * would only be included to control optimizations. */
+ bool emit_s_barrier = barrier->exec_scope == scope_workgroup &&
+ program->workgroup_size > program->wave_size;
+
+ bld.insert(std::move(instr));
+ if (emit_s_barrier)
+ bld.sopp(aco_opcode::s_barrier);
+ } else if (instr->opcode == aco_opcode::p_cvt_f16_f32_rtne) {
+ float_mode new_mode = block->fp_mode;
+ new_mode.round16_64 = fp_round_ne;
+ bool set_round = new_mode.round != block->fp_mode.round;
+
+ emit_set_mode(bld, new_mode, set_round, false);
+
+ instr->opcode = aco_opcode::v_cvt_f16_f32;
+ ctx.instructions.emplace_back(std::move(instr));
+
+ emit_set_mode(bld, block->fp_mode, set_round, false);
} else {
ctx.instructions.emplace_back(std::move(instr));
}