i965/fs: Remove fs_reg::smear.

The same effect can be achieved using a combination of ::stride and
::subreg_offset.  Remove the less flexible ::smear to keep the data
members of fs_reg orthogonal.

Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Paul Berry <stereotype441@gmail.com>
This commit is contained in:
Francisco Jerez 2014-01-15 22:21:30 +01:00
parent 756d37b1d6
commit 019bf6ed8d
7 changed files with 30 additions and 24 deletions

View file

@ -380,7 +380,6 @@ void
fs_reg::init()
{
memset(this, 0, sizeof(*this));
this->smear = -1;
stride = 1;
}
@ -440,7 +439,6 @@ fs_reg::equals(const fs_reg &r) const
!reladdr && !r.reladdr &&
memcmp(&fixed_hw_reg, &r.fixed_hw_reg,
sizeof(fixed_hw_reg)) == 0 &&
smear == r.smear &&
stride == r.stride &&
imm.u == r.imm.u);
}
@ -463,6 +461,15 @@ fs_reg::apply_stride(unsigned stride)
return *this;
}
fs_reg &
fs_reg::set_smear(unsigned subreg)
{
assert(file != HW_REG && file != IMM);
subreg_offset = subreg * type_sz(type);
stride = 0;
return *this;
}
bool
fs_reg::is_contiguous() const
{
@ -567,7 +574,7 @@ fs_visitor::get_timestamp()
* else that might disrupt timing) by setting smear to 2 and checking if
* that field is != 0.
*/
dst.smear = 0;
dst.set_smear(0);
return dst;
}
@ -602,7 +609,7 @@ fs_visitor::emit_shader_time_end()
* were the only two timestamp reads that happened).
*/
fs_reg reset = shader_end_time;
reset.smear = 2;
reset.set_smear(2);
fs_inst *test = emit(AND(reg_null_d, reset, fs_reg(1u)));
test->conditional_mod = BRW_CONDITIONAL_Z;
emit(IF(BRW_PREDICATE_NORMAL));
@ -1973,7 +1980,7 @@ fs_visitor::setup_pull_constants()
inst->src[i].file = GRF;
inst->src[i].reg = dst.reg;
inst->src[i].reg_offset = 0;
inst->src[i].smear = pull_index & 3;
inst->src[i].set_smear(pull_index & 3);
}
}
}
@ -2334,8 +2341,7 @@ fs_visitor::register_coalesce()
inst->src[0].file != GRF ||
inst->src[0].negate ||
inst->src[0].abs ||
inst->src[0].smear != -1 ||
!inst->src[0].is_contiguous() ||
!inst->src[0].is_contiguous() ||
inst->dst.file != GRF ||
inst->dst.type != inst->src[0].type) {
continue;
@ -2496,7 +2502,7 @@ fs_visitor::compute_to_mrf()
inst->dst.file != MRF || inst->src[0].file != GRF ||
inst->dst.type != inst->src[0].type ||
inst->src[0].abs || inst->src[0].negate ||
inst->src[0].smear != -1 || !inst->src[0].is_contiguous() ||
!inst->src[0].is_contiguous() ||
inst->src[0].subreg_offset)
continue;

View file

@ -107,7 +107,9 @@ public:
bool abs;
bool sechalf;
struct brw_reg fixed_hw_reg;
int smear; /* -1, or a channel of the reg to smear to all channels. */
/** Smear a channel of the reg to all channels. */
fs_reg &set_smear(unsigned subreg);
/** Value for file == IMM */
union {

View file

@ -300,7 +300,7 @@ fs_visitor::try_copy_propagate(fs_inst *inst, int arg, acp_entry *entry)
bool has_source_modifiers = entry->src.abs || entry->src.negate;
if ((has_source_modifiers || entry->src.file == UNIFORM ||
entry->src.smear != -1 || !entry->src.is_contiguous()) &&
!entry->src.is_contiguous()) &&
!can_do_source_mods(inst))
return false;
@ -334,8 +334,6 @@ fs_visitor::try_copy_propagate(fs_inst *inst, int arg, acp_entry *entry)
inst->src[arg].file = entry->src.file;
inst->src[arg].reg = entry->src.reg;
inst->src[arg].reg_offset = entry->src.reg_offset;
if (entry->src.smear != -1)
inst->src[arg].smear = entry->src.smear;
inst->src[arg].subreg_offset = entry->src.subreg_offset;
inst->src[arg].stride *= entry->src.stride;

View file

@ -1030,8 +1030,8 @@ brw_reg_from_fs_reg(fs_reg *reg)
switch (reg->file) {
case GRF:
case MRF:
if (reg->stride == 0 || reg->smear >= 0) {
brw_reg = brw_vec1_reg(brw_file_from_reg(reg), reg->reg, reg->smear);
if (reg->stride == 0) {
brw_reg = brw_vec1_reg(brw_file_from_reg(reg), reg->reg, 0);
} else {
brw_reg = brw_vec8_reg(brw_file_from_reg(reg), reg->reg, 0);
brw_reg = stride(brw_reg, 8 * reg->stride, 8, reg->stride);

View file

@ -85,7 +85,7 @@ fs_live_variables::setup_one_read(bblock_t *block, fs_inst *inst,
* would get stomped by the first decode as well.
*/
int end_ip = ip;
if (v->dispatch_width == 16 && (reg.smear != -1 || reg.stride == 0 ||
if (v->dispatch_width == 16 && (reg.stride == 0 ||
(v->pixel_x.reg == reg.reg ||
v->pixel_y.reg == reg.reg))) {
end_ip++;

View file

@ -592,7 +592,7 @@ fs_visitor::choose_spill_reg(struct ra_graph *g)
* loading pull constants, so spilling them is unlikely to reduce
* register pressure anyhow.
*/
if (inst->src[i].smear >= 0 || !inst->src[i].is_contiguous()) {
if (!inst->src[i].is_contiguous()) {
no_spill[inst->src[i].reg] = true;
}
}
@ -601,7 +601,7 @@ fs_visitor::choose_spill_reg(struct ra_graph *g)
if (inst->dst.file == GRF) {
spill_costs[inst->dst.reg] += inst->regs_written * loop_scale;
if (inst->dst.smear >= 0 || !inst->dst.is_contiguous()) {
if (!inst->dst.is_contiguous()) {
no_spill[inst->dst.reg] = true;
}
}

View file

@ -744,8 +744,14 @@ fs_visitor::visit(ir_expression *ir)
emit(fs_inst(FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD,
packed_consts, surf_index, const_offset_reg));
packed_consts.smear = const_offset->value.u[0] % 16 / 4;
for (int i = 0; i < ir->type->vector_elements; i++) {
packed_consts.set_smear(const_offset->value.u[0] % 16 / 4 + i);
/* The std140 packing rules don't allow vectors to cross 16-byte
* boundaries, and a reg is 32 bytes.
*/
assert(packed_consts.subreg_offset < 32);
/* UBO bools are any nonzero value. We consider bools to be
* values with the low bit set to 1. Convert them using CMP.
*/
@ -755,13 +761,7 @@ fs_visitor::visit(ir_expression *ir)
emit(MOV(result, packed_consts));
}
packed_consts.smear++;
result.reg_offset++;
/* The std140 packing rules don't allow vectors to cross 16-byte
* boundaries, and a reg is 32 bytes.
*/
assert(packed_consts.smear < 8);
}
} else {
/* Turn the byte offset into a dword offset. */