label_fcmp = 1 << 20,
label_uniform_bool = 1 << 21,
label_constant_64bit = 1 << 22,
+ label_uniform_bitwise = 1 << 23,
+ label_scc_invert = 1 << 24,
+ label_vcc_hint = 1 << 25,
+ label_scc_needed = 1 << 26,
+ label_b2i = 1 << 27,
};
-static constexpr uint32_t instr_labels = label_vec | label_mul | label_mad | label_omod_success | label_clamp_success | label_add_sub | label_bitwise | label_minmax | label_fcmp;
-static constexpr uint32_t temp_labels = label_abs | label_neg | label_temp | label_vcc | label_b2f | label_uniform_bool | label_omod2 | label_omod4 | label_omod5 | label_clamp;
+static constexpr uint32_t instr_labels = label_vec | label_mul | label_mad | label_omod_success | label_clamp_success |
+ label_add_sub | label_bitwise | label_uniform_bitwise | label_minmax | label_fcmp;
+static constexpr uint32_t temp_labels = label_abs | label_neg | label_temp | label_vcc | label_b2f | label_uniform_bool |
+ label_omod2 | label_omod4 | label_omod5 | label_clamp | label_scc_invert | label_b2i;
static constexpr uint32_t val_labels = label_constant | label_constant_64bit | label_literal | label_mad;
struct ssa_info {
};
uint32_t label;
+ ssa_info() : label(0) {}
+
void add_label(Label new_label)
{
/* Since all labels which use "instr" use it for the same thing
return label & label_bitwise;
}
+ void set_uniform_bitwise()
+ {
+ add_label(label_uniform_bitwise);
+ }
+
+ bool is_uniform_bitwise()
+ {
+ return label & label_uniform_bitwise;
+ }
+
void set_minmax(Instruction *minmax_instr)
{
add_label(label_minmax);
return label & label_fcmp;
}
+ void set_scc_needed()
+ {
+ add_label(label_scc_needed);
+ }
+
+ bool is_scc_needed()
+ {
+ return label & label_scc_needed;
+ }
+
+ void set_scc_invert(Temp scc_inv)
+ {
+ add_label(label_scc_invert);
+ temp = scc_inv;
+ }
+
+ bool is_scc_invert()
+ {
+ return label & label_scc_invert;
+ }
+
void set_uniform_bool(Temp uniform_bool)
{
add_label(label_uniform_bool);
return label & label_uniform_bool;
}
+ void set_vcc_hint()
+ {
+ add_label(label_vcc_hint);
+ }
+
+ bool is_vcc_hint()
+ {
+ return label & label_vcc_hint;
+ }
+
+ void set_b2i(Temp val)
+ {
+ add_label(label_b2i);
+ temp = val;
+ }
+
+ bool is_b2i()
+ {
+ return label & label_b2i;
+ }
+
};
struct opt_ctx {
}
/* only covers special cases */
-bool can_accept_constant(aco_ptr<Instruction>& instr, unsigned operand)
+bool alu_can_accept_constant(aco_opcode opcode, unsigned operand)
{
- switch (instr->opcode) {
+ switch (opcode) {
case aco_opcode::v_interp_p2_f32:
case aco_opcode::v_mac_f32:
case aco_opcode::v_writelane_b32:
case aco_opcode::v_readfirstlane_b32:
return operand != 0;
default:
- if ((instr->format == Format::MUBUF ||
- instr->format == Format::MIMG) &&
- instr->definitions.size() == 1 &&
- instr->operands.size() == 4) {
- return operand != 3;
- }
return true;
}
}
switch (add_instr->opcode) {
case aco_opcode::v_add_u32:
case aco_opcode::v_add_co_u32:
+ case aco_opcode::v_add_co_u32_e64:
case aco_opcode::s_add_i32:
case aco_opcode::s_add_u32:
break;
return op;
}
+bool fixed_to_exec(Operand op)
+{
+ return op.isFixed() && op.physReg() == exec;
+}
+
void label_instruction(opt_ctx &ctx, Block& block, aco_ptr<Instruction>& instr)
{
if (instr->isSALU() || instr->isVALU() || instr->format == Format::PSEUDO) {
/* SALU / PSEUDO: propagate inline constants */
if (instr->isSALU() || instr->format == Format::PSEUDO) {
+ const bool is_subdword = std::any_of(instr->definitions.begin(), instr->definitions.end(),
+ [] (const Definition& def) { return def.regClass().is_subdword();});
+ // TODO: optimize SGPR and constant propagation for subdword pseudo instructions on gfx9+
+ if (is_subdword)
+ continue;
+
if (info.is_temp() && info.temp.type() == RegType::sgpr) {
instr->operands[i].setTemp(info.temp);
info = ctx.info[info.temp.id()];
break;
}
}
- if ((info.is_constant() || info.is_constant_64bit() || (info.is_literal() && instr->format == Format::PSEUDO)) && !instr->operands[i].isFixed() && can_accept_constant(instr, i)) {
+ if ((info.is_constant() || info.is_constant_64bit() || (info.is_literal() && instr->format == Format::PSEUDO)) &&
+ !instr->operands[i].isFixed() && alu_can_accept_constant(instr->opcode, i)) {
instr->operands[i] = get_constant_op(ctx, info.val, info.is_constant_64bit());
continue;
}
static_cast<VOP3A_instruction*>(instr.get())->neg[i] = true;
continue;
}
- if ((info.is_constant() || info.is_constant_64bit()) && can_accept_constant(instr, i)) {
+ if ((info.is_constant() || info.is_constant_64bit()) && alu_can_accept_constant(instr->opcode, i)) {
Operand op = get_constant_op(ctx, info.val, info.is_constant_64bit());
perfwarn(instr->opcode == aco_opcode::v_cndmask_b32 && i == 2, "v_cndmask_b32 with a constant selector", instr.get());
if (i == 0 || instr->opcode == aco_opcode::v_readlane_b32 || instr->opcode == aco_opcode::v_writelane_b32) {
while (info.is_temp())
info = ctx.info[info.temp.id()];
- if (mubuf->offen && i == 0 && info.is_constant_or_literal() && mubuf->offset + info.val < 4096) {
+ if (mubuf->offen && i == 1 && info.is_constant_or_literal() && mubuf->offset + info.val < 4096) {
assert(!mubuf->idxen);
- instr->operands[i] = Operand(v1);
+ instr->operands[1] = Operand(v1);
mubuf->offset += info.val;
mubuf->offen = false;
continue;
instr->operands[2] = Operand((uint32_t) 0);
mubuf->offset += info.val;
continue;
- } else if (mubuf->offen && i == 0 && parse_base_offset(ctx, instr.get(), i, &base, &offset) && base.regClass() == v1 && mubuf->offset + offset < 4096) {
+ } else if (mubuf->offen && i == 1 && parse_base_offset(ctx, instr.get(), i, &base, &offset) && base.regClass() == v1 && mubuf->offset + offset < 4096) {
assert(!mubuf->idxen);
- instr->operands[i].setTemp(base);
+ instr->operands[1].setTemp(base);
mubuf->offset += offset;
continue;
} else if (i == 2 && parse_base_offset(ctx, instr.get(), i, &base, &offset) && base.regClass() == s1 && mubuf->offset + offset < 4096) {
DS_instruction *ds = static_cast<DS_instruction *>(instr.get());
Temp base;
uint32_t offset;
- if (i == 0 && parse_base_offset(ctx, instr.get(), i, &base, &offset) && base.regClass() == instr->operands[i].regClass() && instr->opcode != aco_opcode::ds_swizzle_b32) {
+ bool has_usable_ds_offset = ctx.program->chip_class >= GFX7;
+ if (has_usable_ds_offset &&
+ i == 0 && parse_base_offset(ctx, instr.get(), i, &base, &offset) &&
+ base.regClass() == instr->operands[i].regClass() &&
+ instr->opcode != aco_opcode::ds_swizzle_b32) {
if (instr->opcode == aco_opcode::ds_write2_b32 || instr->opcode == aco_opcode::ds_read2_b32 ||
instr->opcode == aco_opcode::ds_write2_b64 || instr->opcode == aco_opcode::ds_read2_b64) {
- if (offset % 4 == 0 &&
- ds->offset0 + (offset >> 2) <= 255 &&
- ds->offset1 + (offset >> 2) <= 255) {
+ unsigned mask = (instr->opcode == aco_opcode::ds_write2_b64 || instr->opcode == aco_opcode::ds_read2_b64) ? 0x7 : 0x3;
+ unsigned shifts = (instr->opcode == aco_opcode::ds_write2_b64 || instr->opcode == aco_opcode::ds_read2_b64) ? 3 : 2;
+
+ if ((offset & mask) == 0 &&
+ ds->offset0 + (offset >> shifts) <= 255 &&
+ ds->offset1 + (offset >> shifts) <= 255) {
instr->operands[i].setTemp(base);
- ds->offset0 += offset >> 2;
- ds->offset1 += offset >> 2;
+ ds->offset0 += offset >> shifts;
+ ds->offset1 += offset >> shifts;
}
} else {
if (ds->offset0 + offset <= 65535) {
Temp base;
uint32_t offset;
if (i == 1 && info.is_constant_or_literal() &&
- (ctx.program->chip_class < GFX8 || info.val <= 0xFFFFF)) {
+ ((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) && base.regClass() == s1 && offset <= 0xFFFFF && ctx.program->chip_class >= GFX9) {
new_instr->definitions[0] = smem->definitions[0];
new_instr->can_reorder = smem->can_reorder;
new_instr->barrier = smem->barrier;
+ new_instr->glc = smem->glc;
+ new_instr->dlc = smem->dlc;
+ new_instr->nv = smem->nv;
+ new_instr->disable_wqm = smem->disable_wqm;
instr.reset(new_instr);
smem = static_cast<SMEM_instruction *>(instr.get());
}
continue;
}
}
+
+ else if (instr->format == Format::PSEUDO_BRANCH) {
+ if (ctx.info[instr->operands[0].tempId()].is_scc_invert()) {
+ /* Flip the branch instruction to get rid of the scc_invert instruction */
+ instr->opcode = instr->opcode == aco_opcode::p_cbranch_z ? aco_opcode::p_cbranch_nz : aco_opcode::p_cbranch_z;
+ instr->operands[0].setTemp(ctx.info[instr->operands[0].tempId()].temp);
+ }
+ }
}
/* if this instruction doesn't define anything, return */
switch (instr->opcode) {
case aco_opcode::p_create_vector: {
+ bool copy_prop = instr->operands.size() == 1 && instr->operands[0].isTemp() &&
+ instr->operands[0].regClass() == instr->definitions[0].regClass();
+ if (copy_prop) {
+ ctx.info[instr->definitions[0].tempId()].set_temp(instr->operands[0].getTemp());
+ break;
+ }
+
unsigned num_ops = instr->operands.size();
for (const Operand& op : instr->operands) {
if (op.isTemp() && ctx.info[op.tempId()].is_vec())
}
assert(k == num_ops);
}
- if (instr->operands.size() == 1 && instr->operands[0].isTemp())
- ctx.info[instr->definitions[0].tempId()].set_temp(instr->operands[0].getTemp());
- else if (instr->definitions[0].getTemp().size() == instr->operands.size())
- ctx.info[instr->definitions[0].tempId()].set_vec(instr.get());
+
+ ctx.info[instr->definitions[0].tempId()].set_vec(instr.get());
break;
}
case aco_opcode::p_split_vector: {
if (!ctx.info[instr->operands[0].tempId()].is_vec())
break;
Instruction* vec = ctx.info[instr->operands[0].tempId()].instr;
- assert(instr->definitions.size() == vec->operands.size());
- for (unsigned i = 0; i < instr->definitions.size(); i++) {
- Operand vec_op = vec->operands[i];
+ unsigned split_offset = 0;
+ unsigned vec_offset = 0;
+ unsigned vec_index = 0;
+ for (unsigned i = 0; i < instr->definitions.size(); split_offset += instr->definitions[i++].bytes()) {
+ while (vec_offset < split_offset && vec_index < vec->operands.size())
+ vec_offset += vec->operands[vec_index++].bytes();
+
+ if (vec_offset != split_offset || vec->operands[vec_index].bytes() != instr->definitions[i].bytes())
+ continue;
+
+ Operand vec_op = vec->operands[vec_index];
if (vec_op.isConstant()) {
if (vec_op.isLiteral())
ctx.info[instr->definitions[i].tempId()].set_literal(vec_op.constantValue());
ctx.info[instr->definitions[i].tempId()].set_constant(vec_op.constantValue());
else if (vec_op.size() == 2)
ctx.info[instr->definitions[i].tempId()].set_constant_64bit(vec_op.constantValue());
+ } else if (vec_op.isUndefined()) {
+ ctx.info[instr->definitions[i].tempId()].set_undefined();
} else {
assert(vec_op.isTemp());
ctx.info[instr->definitions[i].tempId()].set_temp(vec_op.getTemp());
case aco_opcode::p_extract_vector: { /* mov */
if (!ctx.info[instr->operands[0].tempId()].is_vec())
break;
+
+ /* check if we index directly into a vector element */
Instruction* vec = ctx.info[instr->operands[0].tempId()].instr;
- if (vec->definitions[0].getTemp().size() == vec->operands.size() && /* TODO: what about 64bit or other combinations? */
- vec->operands[0].size() == instr->definitions[0].size()) {
-
- /* convert this extract into a mov instruction */
- Operand vec_op = vec->operands[instr->operands[1].constantValue()];
- bool is_vgpr = instr->definitions[0].getTemp().type() == RegType::vgpr;
- aco_opcode opcode = is_vgpr ? aco_opcode::v_mov_b32 : aco_opcode::s_mov_b32;
- Format format = is_vgpr ? Format::VOP1 : Format::SOP1;
- instr->opcode = opcode;
- instr->format = format;
- while (instr->operands.size() > 1)
- instr->operands.pop_back();
- instr->operands[0] = vec_op;
+ const unsigned index = instr->operands[1].constantValue();
+ const unsigned dst_offset = index * instr->definitions[0].bytes();
+ unsigned offset = 0;
- if (vec_op.isConstant()) {
- if (vec_op.isLiteral())
- ctx.info[instr->definitions[0].tempId()].set_literal(vec_op.constantValue());
- else if (vec_op.size() == 1)
- ctx.info[instr->definitions[0].tempId()].set_constant(vec_op.constantValue());
- else if (vec_op.size() == 2)
- ctx.info[instr->definitions[0].tempId()].set_constant_64bit(vec_op.constantValue());
+ for (const Operand& op : vec->operands) {
+ if (offset < dst_offset) {
+ offset += op.bytes();
+ continue;
+ } else if (offset != dst_offset || op.bytes() != instr->definitions[0].bytes()) {
+ break;
+ }
+ /* convert this extract into a copy instruction */
+ instr->opcode = aco_opcode::p_parallelcopy;
+ instr->operands.pop_back();
+ instr->operands[0] = op;
+
+ if (op.isConstant()) {
+ if (op.isLiteral())
+ ctx.info[instr->definitions[0].tempId()].set_literal(op.constantValue());
+ else if (op.size() == 1)
+ ctx.info[instr->definitions[0].tempId()].set_constant(op.constantValue());
+ else if (op.size() == 2)
+ ctx.info[instr->definitions[0].tempId()].set_constant_64bit(op.constantValue());
+ } else if (op.isUndefined()) {
+ ctx.info[instr->definitions[0].tempId()].set_undefined();
} else {
- assert(vec_op.isTemp());
- ctx.info[instr->definitions[0].tempId()].set_temp(vec_op.getTemp());
+ assert(op.isTemp());
+ ctx.info[instr->definitions[0].tempId()].set_temp(op.getTemp());
}
+ break;
}
break;
}
else
ctx.info[instr->definitions[0].tempId()].set_literal(v);
}
+ break;
}
case aco_opcode::v_mul_f32: { /* omod */
/* TODO: try to move the negate/abs modifier to the consumer instead */
}
case aco_opcode::v_cndmask_b32:
if (instr->operands[0].constantEquals(0) &&
- instr->operands[1].constantEquals(0xFFFFFFFF) &&
- instr->operands[2].isTemp())
+ instr->operands[1].constantEquals(0xFFFFFFFF))
ctx.info[instr->definitions[0].tempId()].set_vcc(instr->operands[2].getTemp());
else if (instr->operands[0].constantEquals(0) &&
- instr->operands[1].constantEquals(0x3f800000u) &&
- instr->operands[2].isTemp())
+ instr->operands[1].constantEquals(0x3f800000u))
ctx.info[instr->definitions[0].tempId()].set_b2f(instr->operands[2].getTemp());
+ else if (instr->operands[0].constantEquals(0) &&
+ instr->operands[1].constantEquals(1))
+ ctx.info[instr->definitions[0].tempId()].set_b2i(instr->operands[2].getTemp());
+
+ ctx.info[instr->operands[2].tempId()].set_vcc_hint();
break;
case aco_opcode::v_cmp_lg_u32:
if (instr->format == Format::VOPC && /* don't optimize VOP3 / SDWA / DPP */
}
case aco_opcode::v_add_u32:
case aco_opcode::v_add_co_u32:
+ case aco_opcode::v_add_co_u32_e64:
case aco_opcode::s_add_i32:
case aco_opcode::s_add_u32:
ctx.info[instr->definitions[0].tempId()].set_add_sub(instr.get());
break;
+ case aco_opcode::s_not_b32:
+ case aco_opcode::s_not_b64:
+ if (ctx.info[instr->operands[0].tempId()].is_uniform_bool()) {
+ ctx.info[instr->definitions[0].tempId()].set_uniform_bitwise();
+ ctx.info[instr->definitions[1].tempId()].set_scc_invert(ctx.info[instr->operands[0].tempId()].temp);
+ } else if (ctx.info[instr->operands[0].tempId()].is_uniform_bitwise()) {
+ ctx.info[instr->definitions[0].tempId()].set_uniform_bitwise();
+ ctx.info[instr->definitions[1].tempId()].set_scc_invert(ctx.info[instr->operands[0].tempId()].instr->definitions[1].getTemp());
+ }
+ ctx.info[instr->definitions[0].tempId()].set_bitwise(instr.get());
+ break;
case aco_opcode::s_and_b32:
case aco_opcode::s_and_b64:
- if (instr->operands[1].isFixed() && instr->operands[1].physReg() == exec &&
- instr->operands[0].isTemp() && ctx.info[instr->operands[0].tempId()].is_uniform_bool()) {
- ctx.info[instr->definitions[1].tempId()].set_temp(ctx.info[instr->operands[0].tempId()].temp);
+ if (fixed_to_exec(instr->operands[1]) && instr->operands[0].isTemp()) {
+ if (ctx.info[instr->operands[0].tempId()].is_uniform_bool()) {
+ /* Try to get rid of the superfluous s_cselect + s_and_b64 that comes from turning a uniform bool into divergent */
+ ctx.info[instr->definitions[1].tempId()].set_temp(ctx.info[instr->operands[0].tempId()].temp);
+ ctx.info[instr->definitions[0].tempId()].set_uniform_bool(ctx.info[instr->operands[0].tempId()].temp);
+ break;
+ } else if (ctx.info[instr->operands[0].tempId()].is_uniform_bitwise()) {
+ /* Try to get rid of the superfluous s_and_b64, since the uniform bitwise instruction already produces the same SCC */
+ ctx.info[instr->definitions[1].tempId()].set_temp(ctx.info[instr->operands[0].tempId()].instr->definitions[1].getTemp());
+ ctx.info[instr->definitions[0].tempId()].set_uniform_bool(ctx.info[instr->operands[0].tempId()].instr->definitions[1].getTemp());
+ break;
+ }
}
/* fallthrough */
- case aco_opcode::s_not_b32:
- case aco_opcode::s_not_b64:
case aco_opcode::s_or_b32:
case aco_opcode::s_or_b64:
case aco_opcode::s_xor_b32:
case aco_opcode::s_xor_b64:
+ if (std::all_of(instr->operands.begin(), instr->operands.end(), [&ctx](const Operand& op) {
+ return op.isTemp() && (ctx.info[op.tempId()].is_uniform_bool() || ctx.info[op.tempId()].is_uniform_bitwise());
+ })) {
+ ctx.info[instr->definitions[0].tempId()].set_uniform_bitwise();
+ }
+ /* fallthrough */
case aco_opcode::s_lshl_b32:
case aco_opcode::v_or_b32:
case aco_opcode::v_lshlrev_b32:
/* Found a cselect that operates on a uniform bool that comes from eg. s_cmp */
ctx.info[instr->definitions[0].tempId()].set_uniform_bool(instr->operands[2].getTemp());
}
+ if (instr->operands[2].isTemp() && ctx.info[instr->operands[2].tempId()].is_scc_invert()) {
+ /* Flip the operands to get rid of the scc_invert instruction */
+ std::swap(instr->operands[0], instr->operands[1]);
+ instr->operands[2].setTemp(ctx.info[instr->operands[2].tempId()].temp);
+ }
+ break;
+ case aco_opcode::p_wqm:
+ if (instr->operands[0].isTemp() &&
+ ctx.info[instr->operands[0].tempId()].is_scc_invert()) {
+ ctx.info[instr->definitions[0].tempId()].set_temp(instr->operands[0].getTemp());
+ }
break;
default:
break;
Instruction *op2_instr = follow_operand(ctx, op1_instr->operands[swap]);
if (!op2_instr || op2_instr->opcode != op2)
return false;
+ if (fixed_to_exec(op2_instr->operands[0]) || fixed_to_exec(op2_instr->operands[1]))
+ return false;
VOP3A_instruction *op1_vop3 = op1_instr->isVOP3() ? static_cast<VOP3A_instruction *>(op1_instr) : NULL;
VOP3A_instruction *op2_vop3 = op2_instr->isVOP3() ? static_cast<VOP3A_instruction *>(op2_instr) : NULL;
/* create instruction */
std::swap(instr->definitions[0], op2_instr->definitions[0]);
+ std::swap(instr->definitions[1], op2_instr->definitions[1]);
ctx.uses[instr->operands[0].tempId()]--;
ctx.info[op2_instr->definitions[0].tempId()].label = 0;
* s_or_b64(a, s_not_b64(b)) -> s_orn2_b64(a, b) */
bool combine_salu_n2(opt_ctx& ctx, aco_ptr<Instruction>& instr)
{
- if (instr->definitions[1].isTemp() && ctx.uses[instr->definitions[1].tempId()])
+ if (instr->definitions[0].isTemp() && ctx.info[instr->definitions[0].tempId()].is_uniform_bool())
return false;
for (unsigned i = 0; i < 2; i++) {
Instruction *op2_instr = follow_operand(ctx, instr->operands[i]);
if (!op2_instr || (op2_instr->opcode != aco_opcode::s_not_b32 && op2_instr->opcode != aco_opcode::s_not_b64))
continue;
+ if (ctx.uses[op2_instr->definitions[1].tempId()] || fixed_to_exec(op2_instr->operands[0]))
+ continue;
if (instr->operands[!i].isLiteral() && op2_instr->operands[0].isLiteral() &&
instr->operands[!i].constantValue() != op2_instr->operands[0].constantValue())
/* s_add_{i32,u32}(a, s_lshl_b32(b, <n>)) -> s_lshl<n>_add_u32(a, b) */
bool combine_salu_lshl_add(opt_ctx& ctx, aco_ptr<Instruction>& instr)
{
- if (instr->definitions[1].isTemp() && ctx.uses[instr->definitions[1].tempId()])
+ if (instr->opcode == aco_opcode::s_add_i32 && ctx.uses[instr->definitions[1].tempId()])
return false;
for (unsigned i = 0; i < 2; i++) {
Instruction *op2_instr = follow_operand(ctx, instr->operands[i]);
- if (!op2_instr || op2_instr->opcode != aco_opcode::s_lshl_b32 || !op2_instr->operands[1].isConstant())
+ if (!op2_instr || op2_instr->opcode != aco_opcode::s_lshl_b32 ||
+ ctx.uses[op2_instr->definitions[1].tempId()])
+ continue;
+ if (!op2_instr->operands[1].isConstant() || fixed_to_exec(op2_instr->operands[0]))
continue;
uint32_t shift = op2_instr->operands[1].constantValue();
return false;
}
+bool combine_add_sub_b2i(opt_ctx& ctx, aco_ptr<Instruction>& instr, aco_opcode new_op, uint8_t ops)
+{
+ if (instr->usesModifiers())
+ return false;
+
+ for (unsigned i = 0; i < 2; i++) {
+ if (!((1 << i) & ops))
+ continue;
+ if (instr->operands[i].isTemp() &&
+ ctx.info[instr->operands[i].tempId()].is_b2i() &&
+ ctx.uses[instr->operands[i].tempId()] == 1) {
+
+ aco_ptr<Instruction> new_instr;
+ if (instr->operands[!i].isTemp() && instr->operands[!i].getTemp().type() == RegType::vgpr) {
+ new_instr.reset(create_instruction<VOP2_instruction>(new_op, Format::VOP2, 3, 2));
+ } else if (ctx.program->chip_class >= GFX10 ||
+ (instr->operands[!i].isConstant() && !instr->operands[!i].isLiteral())) {
+ new_instr.reset(create_instruction<VOP3A_instruction>(new_op, asVOP3(Format::VOP2), 3, 2));
+ } else {
+ return false;
+ }
+ ctx.uses[instr->operands[i].tempId()]--;
+ new_instr->definitions[0] = instr->definitions[0];
+ new_instr->definitions[1] = instr->definitions.size() == 2 ? instr->definitions[1] :
+ Definition(ctx.program->allocateId(), ctx.program->lane_mask);
+ new_instr->definitions[1].setHint(vcc);
+ new_instr->operands[0] = Operand(0u);
+ new_instr->operands[1] = instr->operands[!i];
+ new_instr->operands[2] = Operand(ctx.info[instr->operands[i].tempId()].temp);
+ instr = std::move(new_instr);
+ ctx.info[instr->definitions[0].tempId()].label = 0;
+ return true;
+ }
+ }
+
+ return false;
+}
+
bool get_minmax_info(aco_opcode op, aco_opcode *min, aco_opcode *max, aco_opcode *min3, aco_opcode *max3, aco_opcode *med3, bool *some_gfx9_only)
{
switch (op) {
continue;
}
- sgpr_ids[num_sgprs++] = sgpr.id();
+ if (new_sgpr)
+ sgpr_ids[num_sgprs++] = sgpr.id();
ctx.uses[sgpr_info_id]--;
ctx.uses[sgpr.id()]++;
}
return;
}
+ if (ctx.info[instr->definitions[0].tempId()].is_vcc_hint()) {
+ instr->definitions[0].setHint(vcc);
+ }
+
/* TODO: There are still some peephole optimizations that could be done:
* - abs(a - b) -> s_absdiff_i32
* - various patterns for s_bitcmp{0,1}_b32 and s_bitset{0,1}_b32
}
}
} else if (instr->opcode == aco_opcode::v_or_b32 && ctx.program->chip_class >= GFX9) {
- if (combine_three_valu_op(ctx, instr, aco_opcode::v_or_b32, aco_opcode::v_or3_b32, "012", 1 | 2)) ;
+ if (combine_three_valu_op(ctx, instr, aco_opcode::s_or_b32, aco_opcode::v_or3_b32, "012", 1 | 2)) ;
+ else if (combine_three_valu_op(ctx, instr, aco_opcode::v_or_b32, aco_opcode::v_or3_b32, "012", 1 | 2)) ;
+ else if (combine_three_valu_op(ctx, instr, aco_opcode::s_and_b32, aco_opcode::v_and_or_b32, "120", 1 | 2)) ;
else if (combine_three_valu_op(ctx, instr, aco_opcode::v_and_b32, aco_opcode::v_and_or_b32, "120", 1 | 2)) ;
+ else if (combine_three_valu_op(ctx, instr, aco_opcode::s_lshl_b32, aco_opcode::v_lshl_or_b32, "120", 1 | 2)) ;
else combine_three_valu_op(ctx, instr, aco_opcode::v_lshlrev_b32, aco_opcode::v_lshl_or_b32, "210", 1 | 2);
- } else if (instr->opcode == aco_opcode::v_add_u32 && ctx.program->chip_class >= GFX9) {
- if (combine_three_valu_op(ctx, instr, aco_opcode::v_xor_b32, aco_opcode::v_xad_u32, "120", 1 | 2)) ;
- else if (combine_three_valu_op(ctx, instr, aco_opcode::v_add_u32, aco_opcode::v_add3_u32, "012", 1 | 2)) ;
- else combine_three_valu_op(ctx, instr, aco_opcode::v_lshlrev_b32, aco_opcode::v_lshl_add_u32, "210", 1 | 2);
+ } else if (instr->opcode == aco_opcode::v_add_u32) {
+ if (combine_add_sub_b2i(ctx, instr, aco_opcode::v_addc_co_u32, 1 | 2)) ;
+ else if (ctx.program->chip_class >= GFX9) {
+ if (combine_three_valu_op(ctx, instr, aco_opcode::s_xor_b32, aco_opcode::v_xad_u32, "120", 1 | 2)) ;
+ else if (combine_three_valu_op(ctx, instr, aco_opcode::v_xor_b32, aco_opcode::v_xad_u32, "120", 1 | 2)) ;
+ else if (combine_three_valu_op(ctx, instr, aco_opcode::s_add_i32, aco_opcode::v_add3_u32, "012", 1 | 2)) ;
+ else if (combine_three_valu_op(ctx, instr, aco_opcode::s_add_u32, aco_opcode::v_add3_u32, "012", 1 | 2)) ;
+ else if (combine_three_valu_op(ctx, instr, aco_opcode::v_add_u32, aco_opcode::v_add3_u32, "012", 1 | 2)) ;
+ else if (combine_three_valu_op(ctx, instr, aco_opcode::s_lshl_b32, aco_opcode::v_lshl_add_u32, "120", 1 | 2)) ;
+ else combine_three_valu_op(ctx, instr, aco_opcode::v_lshlrev_b32, aco_opcode::v_lshl_add_u32, "210", 1 | 2);
+ }
+ } else if (instr->opcode == aco_opcode::v_add_co_u32 ||
+ instr->opcode == aco_opcode::v_add_co_u32_e64) {
+ combine_add_sub_b2i(ctx, instr, aco_opcode::v_addc_co_u32, 1 | 2);
+ } else if (instr->opcode == aco_opcode::v_sub_u32 ||
+ instr->opcode == aco_opcode::v_sub_co_u32 ||
+ instr->opcode == aco_opcode::v_sub_co_u32_e64) {
+ combine_add_sub_b2i(ctx, instr, aco_opcode::v_subbrev_co_u32, 2);
+ } else if (instr->opcode == aco_opcode::v_subrev_u32 ||
+ instr->opcode == aco_opcode::v_subrev_co_u32 ||
+ instr->opcode == aco_opcode::v_subrev_co_u32_e64) {
+ combine_add_sub_b2i(ctx, instr, aco_opcode::v_subbrev_co_u32, 1);
} else if (instr->opcode == aco_opcode::v_lshlrev_b32 && ctx.program->chip_class >= GFX9) {
combine_three_valu_op(ctx, instr, aco_opcode::v_add_u32, aco_opcode::v_add_lshl_u32, "120", 2);
} else if ((instr->opcode == aco_opcode::s_add_u32 || instr->opcode == aco_opcode::s_add_i32) && ctx.program->chip_class >= GFX9) {
}
}
+bool to_uniform_bool_instr(opt_ctx &ctx, aco_ptr<Instruction> &instr)
+{
+ switch (instr->opcode) {
+ case aco_opcode::s_and_b32:
+ case aco_opcode::s_and_b64:
+ instr->opcode = aco_opcode::s_and_b32;
+ break;
+ case aco_opcode::s_or_b32:
+ case aco_opcode::s_or_b64:
+ instr->opcode = aco_opcode::s_or_b32;
+ break;
+ case aco_opcode::s_xor_b32:
+ case aco_opcode::s_xor_b64:
+ instr->opcode = aco_opcode::s_absdiff_i32;
+ break;
+ default:
+ /* Don't transform other instructions. They are very unlikely to appear here. */
+ return false;
+ }
+
+ for (Operand &op : instr->operands) {
+ ctx.uses[op.tempId()]--;
+
+ if (ctx.info[op.tempId()].is_uniform_bool()) {
+ /* Just use the uniform boolean temp. */
+ op.setTemp(ctx.info[op.tempId()].temp);
+ } else if (ctx.info[op.tempId()].is_uniform_bitwise()) {
+ /* Use the SCC definition of the predecessor instruction.
+ * This allows the predecessor to get picked up by the same optimization (if it has no divergent users),
+ * and it also makes sure that the current instruction will keep working even if the predecessor won't be transformed.
+ */
+ Instruction *pred_instr = ctx.info[op.tempId()].instr;
+ assert(pred_instr->definitions.size() >= 2);
+ assert(pred_instr->definitions[1].isFixed() && pred_instr->definitions[1].physReg() == scc);
+ op.setTemp(pred_instr->definitions[1].getTemp());
+ } else {
+ unreachable("Invalid operand on uniform bitwise instruction.");
+ }
+
+ ctx.uses[op.tempId()]++;
+ }
+
+ instr->definitions[0].setTemp(Temp(instr->definitions[0].tempId(), s1));
+ assert(instr->operands[0].regClass() == s1);
+ assert(instr->operands[1].regClass() == s1);
+ return true;
+}
void select_instruction(opt_ctx &ctx, aco_ptr<Instruction>& instr)
{
if (instr->opcode == aco_opcode::p_split_vector) {
unsigned num_used = 0;
unsigned idx = 0;
- for (unsigned i = 0; i < instr->definitions.size(); i++) {
+ unsigned split_offset = 0;
+ for (unsigned i = 0, offset = 0; i < instr->definitions.size(); offset += instr->definitions[i++].bytes()) {
if (ctx.uses[instr->definitions[i].tempId()]) {
num_used++;
idx = i;
+ split_offset = offset;
}
}
bool done = false;
unsigned off = 0;
Operand op;
for (Operand& vec_op : vec->operands) {
- if (off == idx * instr->definitions[0].size()) {
+ if (off == split_offset) {
op = vec_op;
break;
}
- off += vec_op.size();
+ off += vec_op.bytes();
}
- if (off != instr->operands[0].size()) {
+ if (off != instr->operands[0].bytes() && op.bytes() == instr->definitions[idx].bytes()) {
ctx.uses[instr->operands[0].tempId()]--;
for (Operand& vec_op : vec->operands) {
if (vec_op.isTemp())
}
}
- if (!done && num_used == 1) {
+ if (!done && num_used == 1 &&
+ instr->operands[0].bytes() % instr->definitions[idx].bytes() == 0 &&
+ split_offset % instr->definitions[idx].bytes() == 0) {
aco_ptr<Pseudo_instruction> extract{create_instruction<Pseudo_instruction>(aco_opcode::p_extract_vector, Format::PSEUDO, 2, 1)};
extract->operands[0] = instr->operands[0];
- extract->operands[1] = Operand((uint32_t) idx);
+ extract->operands[1] = Operand((uint32_t) split_offset / instr->definitions[idx].bytes());
extract->definitions[0] = instr->definitions[idx];
instr.reset(extract.release());
}
continue;
/* if one of the operands is sgpr, we cannot add a literal somewhere else on pre-GFX10 or operands other than the 1st */
if (instr->operands[i].getTemp().type() == RegType::sgpr && (i > 0 || ctx.program->chip_class < GFX10)) {
- if (ctx.info[instr->operands[i].tempId()].is_literal()) {
+ if (!sgpr_used && ctx.info[instr->operands[i].tempId()].is_literal()) {
literal_uses = ctx.uses[instr->operands[i].tempId()];
literal_idx = i;
} else {
literal_idx = i;
}
}
- if (literal_uses < threshold) {
+
+ /* Limit the number of literals to apply to not increase the code
+ * size too much, but always apply literals for v_mad->v_madak
+ * because both instructions are 64-bit and this doesn't increase
+ * code size.
+ * TODO: try to apply the literals earlier to lower the number of
+ * uses below threshold
+ */
+ if (literal_uses < threshold || literal_idx == 2) {
ctx.uses[instr->operands[literal_idx].tempId()]--;
mad_info->check_literal = true;
mad_info->literal_idx = literal_idx;
}
}
+ /* Mark SCC needed, so the uniform boolean transformation won't swap the definitions when it isn't beneficial */
+ if (instr->format == Format::PSEUDO_BRANCH &&
+ instr->operands.size() &&
+ instr->operands[0].isTemp()) {
+ ctx.info[instr->operands[0].tempId()].set_scc_needed();
+ return;
+ } else if ((instr->opcode == aco_opcode::s_cselect_b64 ||
+ instr->opcode == aco_opcode::s_cselect_b32) &&
+ instr->operands[2].isTemp()) {
+ ctx.info[instr->operands[2].tempId()].set_scc_needed();
+ }
+
/* check for literals */
if (!instr->isSALU() && !instr->isVALU())
return;
+ /* Transform uniform bitwise boolean operations to 32-bit when there are no divergent uses. */
+ if (instr->definitions.size() &&
+ ctx.uses[instr->definitions[0].tempId()] == 0 &&
+ ctx.info[instr->definitions[0].tempId()].is_uniform_bitwise()) {
+ bool transform_done = to_uniform_bool_instr(ctx, instr);
+
+ if (transform_done && !ctx.info[instr->definitions[1].tempId()].is_scc_needed()) {
+ /* Swap the two definition IDs in order to avoid overusing the SCC. This reduces extra moves generated by RA. */
+ uint32_t def0_id = instr->definitions[0].getTemp().id();
+ uint32_t def1_id = instr->definitions[1].getTemp().id();
+ instr->definitions[0].setTemp(Temp(def1_id, s1));
+ instr->definitions[1].setTemp(Temp(def0_id, s1));
+ }
+
+ return;
+ }
+
if (instr->isSDWA() || instr->isDPP() || (instr->isVOP3() && ctx.program->chip_class < GFX10))
return; /* some encodings can't ever take literals */
unsigned num_operands = 1;
if (instr->isSALU() || (ctx.program->chip_class >= GFX10 && can_use_VOP3(ctx, instr)))
num_operands = instr->operands.size();
+ /* catch VOP2 with a 3rd SGPR operand (e.g. v_cndmask_b32, v_addc_co_u32) */
+ else if (instr->isVALU() && instr->operands.size() >= 3)
+ return;
unsigned sgpr_ids[2] = {0, 0};
bool is_literal_sgpr = false;
/* choose a literal to apply */
for (unsigned i = 0; i < num_operands; i++) {
Operand op = instr->operands[i];
+
+ if (instr->isVALU() && op.isTemp() && op.getTemp().type() == RegType::sgpr &&
+ op.tempId() != sgpr_ids[0])
+ sgpr_ids[!!sgpr_ids[0]] = op.tempId();
+
if (op.isLiteral()) {
current_literal = op;
continue;
} else if (!op.isTemp() || !ctx.info[op.tempId()].is_literal()) {
- if (instr->isVALU() && op.isTemp() && op.getTemp().type() == RegType::sgpr &&
- op.tempId() != sgpr_ids[0])
- sgpr_ids[!!sgpr_ids[0]] = op.tempId();
continue;
}
- if (!can_accept_constant(instr, i))
+ if (!alu_can_accept_constant(instr->opcode, i))
continue;
if (ctx.uses[op.tempId()] < literal_uses) {
/* apply literals on MAD */
if (instr->opcode == aco_opcode::v_mad_f32 && ctx.info[instr->definitions[0].tempId()].is_mad()) {
mad_info* info = &ctx.mad_infos[ctx.info[instr->definitions[0].tempId()].val];
- if (info->check_literal && ctx.uses[instr->operands[info->literal_idx].tempId()] == 0) {
+ if (info->check_literal &&
+ (ctx.uses[instr->operands[info->literal_idx].tempId()] == 0 || info->literal_idx == 2)) {
aco_ptr<Instruction> new_mad;
if (info->literal_idx == 2) { /* add literal -> madak */
new_mad.reset(create_instruction<VOP2_instruction>(aco_opcode::v_madak_f32, Format::VOP2, 3, 1));
label_instruction(ctx, block, instr);
}
- ctx.uses = std::move(dead_code_analysis(program));
+ ctx.uses = dead_code_analysis(program);
/* 2. Combine v_mad, omod, clamp and propagate sgpr on VALU instructions */
for (Block& block : program->blocks) {