static bool constant_fold_alu_instr(nir_alu_instr *instr, void *mem_ctx) { nir_const_value src[4]; if (!instr->dest.dest.is_ssa) return false; for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++) { if (!instr->src[i].src.is_ssa) return false; nir_instr *src_instr = instr->src[i].src.ssa->parent_instr; if (src_instr->type != nir_instr_type_load_const) return false; nir_load_const_instr* load_const = nir_instr_as_load_const(src_instr); for (unsigned j = 0; j < nir_ssa_alu_instr_src_components(instr, i); j++) { src[i].u[j] = load_const->value.u[instr->src[i].swizzle[j]]; } /* We shouldn't have any source modifiers in the optimization loop. */ assert(!instr->src[i].abs && !instr->src[i].negate); } /* We shouldn't have any saturate modifiers in the optimization loop. */ assert(!instr->dest.saturate); nir_const_value dest = nir_eval_const_opcode(instr->op, instr->dest.dest.ssa.num_components, src); nir_load_const_instr *new_instr = nir_load_const_instr_create(mem_ctx, instr->dest.dest.ssa.num_components); new_instr->value = dest; nir_instr_insert_before(&instr->instr, &new_instr->instr); nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa, nir_src_for_ssa(&new_instr->def), mem_ctx); nir_instr_remove(&instr->instr); ralloc_free(instr); return true; }
static bool constant_fold_alu_instr(nir_alu_instr *instr, void *mem_ctx) { nir_const_value src[NIR_MAX_VEC_COMPONENTS]; if (!instr->dest.dest.is_ssa) return false; /* In the case that any outputs/inputs have unsized types, then we need to * guess the bit-size. In this case, the validator ensures that all * bit-sizes match so we can just take the bit-size from first * output/input with an unsized type. If all the outputs/inputs are sized * then we don't need to guess the bit-size at all because the code we * generate for constant opcodes in this case already knows the sizes of * the types involved and does not need the provided bit-size for anything * (although it still requires to receive a valid bit-size). */ unsigned bit_size = 0; if (!nir_alu_type_get_type_size(nir_op_infos[instr->op].output_type)) bit_size = instr->dest.dest.ssa.bit_size; for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++) { if (!instr->src[i].src.is_ssa) return false; if (bit_size == 0 && !nir_alu_type_get_type_size(nir_op_infos[instr->op].input_sizes[i])) { bit_size = instr->src[i].src.ssa->bit_size; } nir_instr *src_instr = instr->src[i].src.ssa->parent_instr; if (src_instr->type != nir_instr_type_load_const) return false; nir_load_const_instr* load_const = nir_instr_as_load_const(src_instr); for (unsigned j = 0; j < nir_ssa_alu_instr_src_components(instr, i); j++) { switch(load_const->def.bit_size) { case 64: src[i].u64[j] = load_const->value.u64[instr->src[i].swizzle[j]]; break; case 32: src[i].u32[j] = load_const->value.u32[instr->src[i].swizzle[j]]; break; case 16: src[i].u16[j] = load_const->value.u16[instr->src[i].swizzle[j]]; break; case 8: src[i].u8[j] = load_const->value.u8[instr->src[i].swizzle[j]]; break; default: unreachable("Invalid bit size"); } } /* We shouldn't have any source modifiers in the optimization loop. */ assert(!instr->src[i].abs && !instr->src[i].negate); } if (bit_size == 0) bit_size = 32; /* We shouldn't have any saturate modifiers in the optimization loop. */ assert(!instr->dest.saturate); nir_const_value dest = nir_eval_const_opcode(instr->op, instr->dest.dest.ssa.num_components, bit_size, src); nir_load_const_instr *new_instr = nir_load_const_instr_create(mem_ctx, instr->dest.dest.ssa.num_components, instr->dest.dest.ssa.bit_size); new_instr->value = dest; nir_instr_insert_before(&instr->instr, &new_instr->instr); nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa, nir_src_for_ssa(&new_instr->def)); nir_instr_remove(&instr->instr); ralloc_free(instr); return true; }