#include "aco_ir.h"
#include "aco_builder.h"
+#include "util/u_math.h"
namespace aco {
std::vector<WQMState> instr_needs;
uint8_t block_needs;
uint8_t ever_again_needs;
+ bool logical_end_wqm;
/* more... */
};
if (instr->format == Format::PSEUDO) {
switch (instr->opcode) {
case aco_opcode::p_create_vector:
- return instr->definitions[0].getTemp().type() == RegType::vgpr;
case aco_opcode::p_extract_vector:
case aco_opcode::p_split_vector:
- return instr->operands[0].getTemp().type() == RegType::vgpr;
+ for (Definition def : instr->definitions) {
+ if (def.getTemp().type() == RegType::vgpr)
+ return true;
+ }
+ return false;
case aco_opcode::p_spill:
case aco_opcode::p_reload:
return false;
ctx.branch_wqm[block_idx] = true;
Block& block = ctx.program->blocks[block_idx];
- aco_ptr<Instruction>& branch = block.instructions.back();
-
- if (branch->opcode != aco_opcode::p_branch) {
- assert(!branch->operands.empty() && branch->operands[0].isTemp());
- set_needs_wqm(ctx, branch->operands[0].getTemp());
- }
/* TODO: this sets more branch conditions to WQM than it needs to
* it should be enough to stop at the "exec mask top level" */
if (block->kind & block_kind_top_level) {
if (ctx.loop && ctx.wqm) {
- /* mark all break conditions as WQM */
unsigned block_idx = block->index + 1;
while (!(ctx.program->blocks[block_idx].kind & block_kind_top_level)) {
+ /* flag all break conditions as WQM:
+ * the conditions might be computed outside the nested CF */
if (ctx.program->blocks[block_idx].kind & block_kind_break)
mark_block_wqm(ctx, block_idx);
+ /* flag all blocks as WQM to ensure we enter all (nested) loops in WQM */
+ exec_ctx.info[block_idx].block_needs |= WQM;
block_idx++;
}
} else if (ctx.loop && !ctx.wqm) {
}
}
+ if (instr->format == Format::PSEUDO_BRANCH && ctx.branch_wqm[block->index]) {
+ needs = WQM;
+ propagate_wqm = true;
+ }
+
if (propagate_wqm) {
for (const Operand& op : instr->operands) {
if (op.isTemp()) {
/* ensure the condition controlling the control flow for this phi is in WQM */
if (needs == WQM && instr->opcode == aco_opcode::p_phi) {
- for (unsigned pred_idx : block->logical_preds)
+ for (unsigned pred_idx : block->logical_preds) {
mark_block_wqm(ctx, pred_idx);
+ exec_ctx.info[pred_idx].logical_end_wqm = true;
+ ctx.worklist.insert(pred_idx);
+ }
+ }
+
+ if ((instr->opcode == aco_opcode::p_logical_end && info.logical_end_wqm) ||
+ instr->opcode == aco_opcode::p_wqm) {
+ assert(needs != Exact);
+ needs = WQM;
}
instr_needs[i] = needs;
return;
if (ctx.info[idx].exec.back().second & mask_type_global) {
Temp exec_mask = ctx.info[idx].exec.back().first;
+ /* TODO: we might generate better code if we pass the uncopied "exec_mask"
+ * directly to the s_wqm (we still need to keep this parallelcopy for
+ * potential later uses of exec_mask though). We currently can't do this
+ * because of a RA bug. */
+ exec_mask = bld.pseudo(aco_opcode::p_parallelcopy, bld.def(bld.lm), bld.exec(exec_mask));
+ ctx.info[idx].exec.back().first = exec_mask;
+
exec_mask = bld.sop1(Builder::s_wqm, bld.def(bld.lm, exec), bld.def(s1, scc), exec_mask);
ctx.info[idx].exec.emplace_back(exec_mask, mask_type_global | mask_type_wqm);
return;
Temp exec_mask = startpgm->definitions.back().getTemp();
bld.insert(std::move(startpgm));
+ /* exec seems to need to be manually initialized with combined shaders */
+ if (util_bitcount(ctx.program->stage & sw_mask) > 1 || (ctx.program->stage & hw_ngg_gs)) {
+ bld.sop1(Builder::s_mov, bld.exec(Definition(exec_mask)), bld.lm == s2 ? Operand(UINT64_MAX) : Operand(UINT32_MAX));
+ instructions[0]->definitions.pop_back();
+ }
+
if (ctx.handle_wqm) {
ctx.info[0].exec.emplace_back(exec_mask, mask_type_global | mask_type_exact | mask_type_initial);
/* if this block only needs WQM, initialize already */
assert(!(block->kind & block_kind_top_level) || info.num_exec_masks <= 2);
/* create the loop exit phis if not trivial */
+ bool need_parallelcopy = false;
for (unsigned k = 0; k < info.num_exec_masks; k++) {
Temp same = ctx.info[preds[0]].exec[k].first;
uint8_t type = ctx.info[header_preds[0]].exec[k].second;
trivial = false;
}
+ if (k == info.num_exec_masks - 1u) {
+ bool all_liveout_exec = true;
+ bool all_not_liveout_exec = true;
+ for (unsigned pred : preds) {
+ all_liveout_exec = all_liveout_exec && same == ctx.program->blocks[pred].live_out_exec;
+ all_not_liveout_exec = all_not_liveout_exec && same != ctx.program->blocks[pred].live_out_exec;
+ }
+ if (!all_liveout_exec && !all_not_liveout_exec)
+ trivial = false;
+ else if (all_not_liveout_exec)
+ need_parallelcopy = true;
+
+ need_parallelcopy |= !trivial;
+ }
+
if (trivial) {
ctx.info[idx].exec.emplace_back(same, type);
} else {
/* create phi for loop footer */
aco_ptr<Pseudo_instruction> phi{create_instruction<Pseudo_instruction>(aco_opcode::p_linear_phi, Format::PSEUDO, preds.size(), 1)};
phi->definitions[0] = bld.def(bld.lm);
+ if (k == info.num_exec_masks - 1u) {
+ phi->definitions[0].setFixed(exec);
+ need_parallelcopy = false;
+ }
for (unsigned i = 0; i < phi->operands.size(); i++)
phi->operands[i] = Operand(ctx.info[preds[i]].exec[k].first);
ctx.info[idx].exec.emplace_back(bld.insert(std::move(phi)), type);
}
assert(ctx.info[idx].exec.back().first.size() == bld.lm.size());
- ctx.info[idx].exec.back().first = bld.pseudo(aco_opcode::p_parallelcopy, bld.def(bld.lm, exec),
- ctx.info[idx].exec.back().first);
+ if (need_parallelcopy) {
+ /* only create this parallelcopy is needed, since the operand isn't
+ * fixed to exec which causes the spiller to miscalculate register demand */
+ /* TODO: Fix register_demand calculation for spilling on loop exits.
+ * The problem is only mitigated because the register demand could be
+ * higher if the exec phi doesn't get assigned to exec. */
+ ctx.info[idx].exec.back().first = bld.pseudo(aco_opcode::p_parallelcopy, bld.def(bld.lm, exec),
+ ctx.info[idx].exec.back().first);
+ }
ctx.loop.pop_back();
return i;
has_discard);
}
+ /* For normal breaks, this is the exec mask. For discard+break, it's the
+ * old exec mask before it was zero'd.
+ */
+ Operand break_cond = bld.exec(ctx.info[idx].exec.back().first);
+
if (block->kind & block_kind_discard) {
assert(block->instructions.back()->format == Format::PSEUDO_BRANCH);
}
assert(!ctx.handle_wqm || (ctx.info[block->index].exec[0].second & mask_type_wqm) == 0);
- if ((block->kind & (block_kind_break | block_kind_uniform)) == block_kind_break)
- ctx.info[idx].exec.back().first = cond;
+ break_cond = Operand(cond);
bld.insert(std::move(branch));
/* no return here as it can be followed by a divergent break */
}
assert(block->instructions.back()->opcode == aco_opcode::p_branch);
block->instructions.pop_back();
- while (!(ctx.info[idx].exec.back().second & mask_type_loop))
+ bool need_parallelcopy = false;
+ while (!(ctx.info[idx].exec.back().second & mask_type_loop)) {
ctx.info[idx].exec.pop_back();
+ need_parallelcopy = true;
+ }
- ctx.info[idx].exec.back().first = bld.pseudo(aco_opcode::p_parallelcopy, bld.def(bld.lm, exec), ctx.info[idx].exec.back().first);
- bld.branch(aco_opcode::p_cbranch_nz, bld.exec(ctx.info[idx].exec.back().first), block->linear_succs[1], block->linear_succs[0]);
+ if (need_parallelcopy)
+ ctx.info[idx].exec.back().first = bld.pseudo(aco_opcode::p_parallelcopy, bld.def(bld.lm, exec), ctx.info[idx].exec.back().first);
+ bld.branch(aco_opcode::p_cbranch_nz, bld.hint_vcc(bld.def(s2)), bld.exec(ctx.info[idx].exec.back().first), block->linear_succs[1], block->linear_succs[0]);
return;
}
/* add next current exec to the stack */
ctx.info[idx].exec.emplace_back(then_mask, mask_type);
- bld.branch(aco_opcode::p_cbranch_z, bld.exec(then_mask), block->linear_succs[1], block->linear_succs[0]);
+ bld.branch(aco_opcode::p_cbranch_z, bld.hint_vcc(bld.def(s2)), bld.exec(then_mask), block->linear_succs[1], block->linear_succs[0]);
return;
}
/* add next current exec to the stack */
ctx.info[idx].exec.emplace_back(else_mask, mask_type);
- bld.branch(aco_opcode::p_cbranch_z, bld.exec(else_mask), block->linear_succs[1], block->linear_succs[0]);
+ bld.branch(aco_opcode::p_cbranch_z, bld.hint_vcc(bld.def(s2)), bld.exec(else_mask), block->linear_succs[1], block->linear_succs[0]);
return;
}
assert(block->instructions.back()->opcode == aco_opcode::p_branch);
block->instructions.pop_back();
- Temp current_exec = ctx.info[idx].exec.back().first;
Temp cond = Temp();
for (int exec_idx = ctx.info[idx].exec.size() - 2; exec_idx >= 0; exec_idx--) {
cond = bld.tmp(s1);
Temp exec_mask = ctx.info[idx].exec[exec_idx].first;
exec_mask = bld.sop2(Builder::s_andn2, bld.def(bld.lm), bld.scc(Definition(cond)),
- exec_mask, current_exec);
+ exec_mask, break_cond);
ctx.info[idx].exec[exec_idx].first = exec_mask;
if (ctx.info[idx].exec[exec_idx].second & mask_type_loop)
break;
ctx.info[idx].exec.back().first = bld.sop1(Builder::s_mov, bld.def(bld.lm, exec), Operand(0u));
}
- bld.branch(aco_opcode::p_cbranch_nz, bld.scc(cond), block->linear_succs[1], block->linear_succs[0]);
+ bld.branch(aco_opcode::p_cbranch_nz, bld.hint_vcc(bld.def(s2)), bld.scc(cond), block->linear_succs[1], block->linear_succs[0]);
return;
}
ctx.info[idx].exec.back().first = bld.sop1(Builder::s_mov, bld.def(bld.lm, exec), Operand(0u));
}
- bld.branch(aco_opcode::p_cbranch_nz, bld.scc(cond), block->linear_succs[1], block->linear_succs[0]);
+ bld.branch(aco_opcode::p_cbranch_nz, bld.hint_vcc(bld.def(s2)), bld.scc(cond), block->linear_succs[1], block->linear_succs[0]);
return;
}
}
projects / mesa.git / blobdiff