/* Otherwise, check mode hazards */
bool could_scalar = true;
+ unsigned sz0 = nir_alu_type_get_type_size(ains->src_types[0]);
+ unsigned sz1 = nir_alu_type_get_type_size(ains->src_types[1]);
/* Only 16/32-bit can run on a scalar unit */
could_scalar &= ains->alu.reg_mode != midgard_reg_mode_8;
could_scalar &= ains->alu.reg_mode != midgard_reg_mode_64;
- could_scalar &= ains->alu.dest_override == midgard_dest_override_none;
- if (ains->alu.reg_mode == midgard_reg_mode_16) {
- /* If we're running in 16-bit mode, we
- * can't have any 8-bit sources on the
- * scalar unit (since the scalar unit
- * doesn't understand 8-bit) */
+ if (ains->src[0] != ~0)
+ could_scalar &= (sz0 == 16) || (sz0 == 32);
- midgard_vector_alu_src s1 =
- vector_alu_from_unsigned(ains->alu.src1);
-
- could_scalar &= !s1.half;
-
- midgard_vector_alu_src s2 =
- vector_alu_from_unsigned(ains->alu.src2);
-
- could_scalar &= !s2.half;
- }
+ if (ains->src[1] != ~0)
+ could_scalar &= (sz1 == 16) || (sz1 == 32);
return could_scalar;
}
unsigned pipeline_count;
};
+static bool
+mir_adjust_constant(midgard_instruction *ins, unsigned src,
+ unsigned *bundle_constant_mask,
+ unsigned *comp_mapping,
+ uint8_t *bundle_constants,
+ bool upper)
+{
+ unsigned type_size = nir_alu_type_get_type_size(ins->src_types[src]) / 8;
+ unsigned type_shift = util_logbase2(type_size);
+ unsigned max_comp = mir_components_for_type(ins->src_types[src]);
+ unsigned comp_mask = mir_from_bytemask(mir_round_bytemask_up(
+ mir_bytemask_of_read_components_index(ins, src),
+ type_size * 8),
+ type_size * 8);
+ unsigned type_mask = (1 << type_size) - 1;
+
+ /* Upper only makes sense for 16-bit */
+ if (type_size != 16 && upper)
+ return false;
+
+ /* For 16-bit, we need to stay on either upper or lower halves to avoid
+ * disrupting the swizzle */
+ unsigned start = upper ? 8 : 0;
+ unsigned length = (type_size == 2) ? 8 : 16;
+
+ for (unsigned comp = 0; comp < max_comp; comp++) {
+ if (!(comp_mask & (1 << comp)))
+ continue;
+
+ uint8_t *constantp = ins->constants.u8 + (type_size * comp);
+ unsigned best_reuse_bytes = 0;
+ signed best_place = -1;
+ unsigned i, j;
+
+ for (i = start; i < (start + length); i += type_size) {
+ unsigned reuse_bytes = 0;
+
+ for (j = 0; j < type_size; j++) {
+ if (!(*bundle_constant_mask & (1 << (i + j))))
+ continue;
+ if (constantp[j] != bundle_constants[i + j])
+ break;
+ if ((i + j) > (start + length))
+ break;
+
+ reuse_bytes++;
+ }
+
+ /* Select the place where existing bytes can be
+ * reused so we leave empty slots to others
+ */
+ if (j == type_size &&
+ (reuse_bytes > best_reuse_bytes || best_place < 0)) {
+ best_reuse_bytes = reuse_bytes;
+ best_place = i;
+ break;
+ }
+ }
+
+ /* This component couldn't fit in the remaining constant slot,
+ * no need check the remaining components, bail out now
+ */
+ if (best_place < 0)
+ return false;
+
+ memcpy(&bundle_constants[i], constantp, type_size);
+ *bundle_constant_mask |= type_mask << best_place;
+ comp_mapping[comp] = best_place >> type_shift;
+ }
+
+ return true;
+}
+
/* For an instruction that can fit, adjust it to fit and update the constants
* array, in destructive mode. Returns whether the fitting was successful. */
return true;
unsigned r_constant = SSA_FIXED_REGISTER(REGISTER_CONSTANT);
- midgard_reg_mode dst_mode = mir_typesize(ins);
-
unsigned bundle_constant_mask = pred->constant_mask;
unsigned comp_mapping[2][16] = { };
uint8_t bundle_constants[16];
if (ins->src[src] != SSA_FIXED_REGISTER(REGISTER_CONSTANT))
continue;
- midgard_reg_mode src_mode = mir_srcsize(ins, src);
- unsigned type_size = mir_bytes_for_mode(src_mode);
- unsigned max_comp = 16 / type_size;
- unsigned comp_mask = mir_from_bytemask(mir_bytemask_of_read_components_index(ins, src),
- dst_mode);
- unsigned type_mask = (1 << type_size) - 1;
-
- for (unsigned comp = 0; comp < max_comp; comp++) {
- if (!(comp_mask & (1 << comp)))
- continue;
-
- uint8_t *constantp = ins->constants.u8 + (type_size * comp);
- unsigned best_reuse_bytes = 0;
- signed best_place = -1;
- unsigned i, j;
-
- for (i = 0; i < 16; i += type_size) {
- unsigned reuse_bytes = 0;
-
- for (j = 0; j < type_size; j++) {
- if (!(bundle_constant_mask & (1 << (i + j))))
- continue;
- if (constantp[j] != bundle_constants[i + j])
- break;
-
- reuse_bytes++;
- }
-
- /* Select the place where existing bytes can be
- * reused so we leave empty slots to others
- */
- if (j == type_size &&
- (reuse_bytes > best_reuse_bytes || best_place < 0)) {
- best_reuse_bytes = reuse_bytes;
- best_place = i;
- break;
- }
- }
+ /* First, try lower half (or whole for !16) */
+ if (mir_adjust_constant(ins, src, &bundle_constant_mask,
+ comp_mapping[src], bundle_constants, false))
+ continue;
- /* This component couldn't fit in the remaining constant slot,
- * no need check the remaining components, bail out now
- */
- if (best_place < 0)
- return false;
+ /* Next, try upper half */
+ if (mir_adjust_constant(ins, src, &bundle_constant_mask,
+ comp_mapping[src], bundle_constants, true))
+ continue;
- memcpy(&bundle_constants[i], constantp, type_size);
- bundle_constant_mask |= type_mask << best_place;
- comp_mapping[src][comp] = best_place / type_size;
- }
+ /* Otherwise bail */
+ return false;
}
/* If non-destructive, we're done */