static bool
try_copy_propagate(const struct gen_device_info *devinfo,
vec4_instruction *inst, int arg,
const copy_entry *entry, int attributes_per_reg)
{
/* Build up the value we are propagating as if it were the source of a
* single MOV
*/
src_reg value =
get_copy_value(*entry,
brw_apply_inv_swizzle_to_mask(inst->src[arg].swizzle,
WRITEMASK_XYZW));
/* Check that we can propagate that value */
if (value.file != UNIFORM &&
value.file != VGRF &&
value.file != ATTR)
return false;
if (devinfo->gen >= 8 && (value.negate || value.abs) &&
is_logic_op(inst->opcode)) {
return false;
}
bool has_source_modifiers = value.negate || value.abs;
/* gen6 math and gen7+ SENDs from GRFs ignore source modifiers on
* instructions.
*/
if ((has_source_modifiers || value.file == UNIFORM ||
value.swizzle != BRW_SWIZZLE_XYZW) && !inst->can_do_source_mods(devinfo))
return false;
if (has_source_modifiers &&
value.type != inst->src[arg].type &&
!inst->can_change_types())
return false;
if (has_source_modifiers &&
inst->opcode == SHADER_OPCODE_GEN4_SCRATCH_WRITE)
return false;
unsigned composed_swizzle = brw_compose_swizzle(inst->src[arg].swizzle,
value.swizzle);
if (inst->is_3src(devinfo) &&
(value.file == UNIFORM ||
(value.file == ATTR && attributes_per_reg != 1)) &&
!brw_is_single_value_swizzle(composed_swizzle))
return false;
if (inst->is_send_from_grf())
return false;
/* we can't generally copy-propagate UD negations becuse we
* end up accessing the resulting values as signed integers
* instead. See also resolve_ud_negate().
*/
if (value.negate &&
value.type == BRW_REGISTER_TYPE_UD)
return false;
/* Don't report progress if this is a noop. */
if (value.equals(inst->src[arg]))
return false;
const unsigned dst_saturate_mask = inst->dst.writemask &
brw_apply_swizzle_to_mask(inst->src[arg].swizzle, entry->saturatemask);
if (dst_saturate_mask) {
/* We either saturate all or nothing. */
if (dst_saturate_mask != inst->dst.writemask)
return false;
/* Limit saturate propagation only to SEL with src1 bounded within 0.0
* and 1.0, otherwise skip copy propagate altogether.
*/
switch(inst->opcode) {
case BRW_OPCODE_SEL:
if (arg != 0 ||
inst->src[0].type != BRW_REGISTER_TYPE_F ||
inst->src[1].file != IMM ||
inst->src[1].type != BRW_REGISTER_TYPE_F ||
inst->src[1].f < 0.0 ||
inst->src[1].f > 1.0) {
return false;
}
if (!inst->saturate)
inst->saturate = true;
break;
default:
return false;
}
}
/* Build the final value */
if (inst->src[arg].abs) {
value.negate = false;
value.abs = true;
}
if (inst->src[arg].negate)
value.negate = !value.negate;
value.swizzle = composed_swizzle;
if (has_source_modifiers &&
value.type != inst->src[arg].type) {
assert(inst->can_change_types());
for (int i = 0; i < 3; i++) {
inst->src[i].type = value.type;
}
inst->dst.type = value.type;
} else {
value.type = inst->src[arg].type;
}
inst->src[arg] = value;
return true;
}
static bool
try_copy_propagate(const struct brw_device_info *devinfo,
vec4_instruction *inst,
int arg, struct copy_entry *entry)
{
/* For constant propagation, we only handle the same constant
* across all 4 channels. Some day, we should handle the 8-bit
* float vector format, which would let us constant propagate
* vectors better.
*/
src_reg value = *entry->value[0];
for (int i = 1; i < 4; i++) {
/* This is equals() except we don't care about the swizzle. */
if (value.file != entry->value[i]->file ||
value.reg != entry->value[i]->reg ||
value.reg_offset != entry->value[i]->reg_offset ||
value.type != entry->value[i]->type ||
value.negate != entry->value[i]->negate ||
value.abs != entry->value[i]->abs) {
return false;
}
}
/* Compute the swizzle of the original register by swizzling the
* component loaded from each value according to the swizzle of
* operand we're going to change.
*/
int s[4];
for (int i = 0; i < 4; i++) {
s[i] = BRW_GET_SWZ(entry->value[i]->swizzle, i);
}
value.swizzle = brw_compose_swizzle(inst->src[arg].swizzle,
BRW_SWIZZLE4(s[0], s[1], s[2], s[3]));
if (value.file != UNIFORM &&
value.file != GRF &&
value.file != ATTR)
return false;
if (devinfo->gen >= 8 && (value.negate || value.abs) &&
is_logic_op(inst->opcode)) {
return false;
}
if (inst->src[arg].abs) {
value.negate = false;
value.abs = true;
}
if (inst->src[arg].negate)
value.negate = !value.negate;
bool has_source_modifiers = value.negate || value.abs;
/* gen6 math and gen7+ SENDs from GRFs ignore source modifiers on
* instructions.
*/
if ((has_source_modifiers || value.file == UNIFORM ||
value.swizzle != BRW_SWIZZLE_XYZW) && !inst->can_do_source_mods(devinfo))
return false;
if (has_source_modifiers && value.type != inst->src[arg].type)
return false;
if (has_source_modifiers &&
inst->opcode == SHADER_OPCODE_GEN4_SCRATCH_WRITE)
return false;
if (inst->is_3src() && value.file == UNIFORM)
return false;
if (inst->is_send_from_grf())
return false;
/* we can't generally copy-propagate UD negations becuse we
* end up accessing the resulting values as signed integers
* instead. See also resolve_ud_negate().
*/
if (value.negate &&
value.type == BRW_REGISTER_TYPE_UD)
return false;
/* Don't report progress if this is a noop. */
if (value.equals(inst->src[arg]))
return false;
const unsigned dst_saturate_mask = inst->dst.writemask &
brw_apply_swizzle_to_mask(inst->src[arg].swizzle, entry->saturatemask);
if (dst_saturate_mask) {
/* We either saturate all or nothing. */
if (dst_saturate_mask != inst->dst.writemask)
return false;
/* Limit saturate propagation only to SEL with src1 bounded within 0.0
* and 1.0, otherwise skip copy propagate altogether.
*/
switch(inst->opcode) {
case BRW_OPCODE_SEL:
if (arg != 0 ||
inst->src[0].type != BRW_REGISTER_TYPE_F ||
inst->src[1].file != IMM ||
inst->src[1].type != BRW_REGISTER_TYPE_F ||
inst->src[1].fixed_hw_reg.dw1.f < 0.0 ||
inst->src[1].fixed_hw_reg.dw1.f > 1.0) {
return false;
}
if (!inst->saturate)
inst->saturate = true;
break;
default:
return false;
}
}
value.type = inst->src[arg].type;
inst->src[arg] = value;
return true;
}
static bool
try_constant_propagate(const struct gen_device_info *devinfo,
vec4_instruction *inst,
int arg, const copy_entry *entry)
{
/* For constant propagation, we only handle the same constant
* across all 4 channels. Some day, we should handle the 8-bit
* float vector format, which would let us constant propagate
* vectors better.
* We could be more aggressive here -- some channels might not get used
* based on the destination writemask.
*/
src_reg value =
get_copy_value(*entry,
brw_apply_inv_swizzle_to_mask(inst->src[arg].swizzle,
WRITEMASK_XYZW));
if (value.file != IMM)
return false;
if (value.type == BRW_REGISTER_TYPE_VF) {
/* The result of bit-casting the component values of a vector float
* cannot in general be represented as an immediate.
*/
if (inst->src[arg].type != BRW_REGISTER_TYPE_F)
return false;
} else {
value.type = inst->src[arg].type;
}
if (inst->src[arg].abs) {
if ((devinfo->gen >= 8 && is_logic_op(inst->opcode)) ||
!brw_abs_immediate(value.type, &value.as_brw_reg())) {
return false;
}
}
if (inst->src[arg].negate) {
if ((devinfo->gen >= 8 && is_logic_op(inst->opcode)) ||
!brw_negate_immediate(value.type, &value.as_brw_reg())) {
return false;
}
}
value = swizzle(value, inst->src[arg].swizzle);
switch (inst->opcode) {
case BRW_OPCODE_MOV:
case SHADER_OPCODE_BROADCAST:
inst->src[arg] = value;
return true;
case SHADER_OPCODE_POW:
case SHADER_OPCODE_INT_QUOTIENT:
case SHADER_OPCODE_INT_REMAINDER:
if (devinfo->gen < 8)
break;
/* fallthrough */
case BRW_OPCODE_DP2:
case BRW_OPCODE_DP3:
case BRW_OPCODE_DP4:
case BRW_OPCODE_DPH:
case BRW_OPCODE_BFI1:
case BRW_OPCODE_ASR:
case BRW_OPCODE_SHL:
case BRW_OPCODE_SHR:
case BRW_OPCODE_SUBB:
if (arg == 1) {
inst->src[arg] = value;
return true;
}
break;
case BRW_OPCODE_MACH:
case BRW_OPCODE_MUL:
case SHADER_OPCODE_MULH:
case BRW_OPCODE_ADD:
case BRW_OPCODE_OR:
case BRW_OPCODE_AND:
case BRW_OPCODE_XOR:
case BRW_OPCODE_ADDC:
if (arg == 1) {
inst->src[arg] = value;
return true;
} else if (arg == 0 && inst->src[1].file != IMM) {
/* Fit this constant in by commuting the operands. Exception: we
* can't do this for 32-bit integer MUL/MACH because it's asymmetric.
*/
if ((inst->opcode == BRW_OPCODE_MUL ||
inst->opcode == BRW_OPCODE_MACH) &&
(inst->src[1].type == BRW_REGISTER_TYPE_D ||
inst->src[1].type == BRW_REGISTER_TYPE_UD))
break;
inst->src[0] = inst->src[1];
inst->src[1] = value;
return true;
}
break;
case GS_OPCODE_SET_WRITE_OFFSET:
/* This is just a multiply by a constant with special strides.
* The generator will handle immediates in both arguments (generating
* a single MOV of the product). So feel free to propagate in src0.
*/
inst->src[arg] = value;
return true;
case BRW_OPCODE_CMP:
if (arg == 1) {
inst->src[arg] = value;
return true;
} else if (arg == 0 && inst->src[1].file != IMM) {
enum brw_conditional_mod new_cmod;
new_cmod = brw_swap_cmod(inst->conditional_mod);
if (new_cmod != BRW_CONDITIONAL_NONE) {
/* Fit this constant in by swapping the operands and
* flipping the test.
*/
inst->src[0] = inst->src[1];
inst->src[1] = value;
inst->conditional_mod = new_cmod;
return true;
}
}
break;
case BRW_OPCODE_SEL:
if (arg == 1) {
inst->src[arg] = value;
return true;
} else if (arg == 0 && inst->src[1].file != IMM) {
inst->src[0] = inst->src[1];
inst->src[1] = value;
/* If this was predicated, flipping operands means
* we also need to flip the predicate.
*/
if (inst->conditional_mod == BRW_CONDITIONAL_NONE) {
inst->predicate_inverse = !inst->predicate_inverse;
}
return true;
}
break;
default:
break;
}
return false;
}