ed/Odin
1
0
Fork 0
mirror of https://github.com/Ed94/Odin.git synced 2026-08-06 15:48:51 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge pull request #4893 from Barinzaya/for-bitset-ctz

Browse Source 
Use CTZ in `for in bit_set` codegen
This commit is contained in:
gingerBill
2025-03-04 07:51:09 +00:00
committed by GitHub
parent 71eb26ef15 ef83f2115a
commit 8214892b5b
1 changed files with 34 additions and 53 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/llvm_backend_stmt.cpp
+34 -53
View File
@@ -1114,62 +1114,43 @@ gb_internal void lb_build_range_stmt(lbProcedure *p, AstRangeStmt *rs, Scope *sc
} }
Type *elem = et->BitSet.elem; Type *elem = et->BitSet.elem;
if (is_type_enum(elem)) { Type *mask = bit_set_to_int(et);
i64 enum_count = 0;
lbValue values = lb_enum_values_slice(p, elem, &enum_count);
lbValue values_data = lb_slice_elem(p, values);
lbValue max_count = lb_const_int(m, t_int, enum_count);
lbAddr offset_ = lb_add_local_generated(p, t_int, false); lbValue all_mask = lb_const_value(p->module, mask, exact_bit_set_all_set_mask(et));
lb_addr_store(p, offset_, lb_const_int(m, t_int, 0)); lbValue initial_mask = lb_emit_arith(p, Token_And, the_set, all_mask, mask);
loop = lb_create_block(p, "for.bit_set.enum.loop"); if (rs->reverse) {
lb_emit_jump(p, loop); initial_mask = lb_emit_reverse_bits(p, initial_mask, mask);
lb_start_block(p, loop);
lbBlock *body_check = lb_create_block(p, "for.bit_set.enum.body-check");
lbBlock *body = lb_create_block(p, "for.bit_set.enum.body");
done = lb_create_block(p, "for.bit_set.enum.done");
lbValue offset = lb_addr_load(p, offset_);
lbValue cond = lb_emit_comp(p, Token_Lt, offset, max_count);
lb_emit_if(p, cond, body_check, done);
lb_start_block(p, body_check);
lbValue val_ptr = lb_emit_ptr_offset(p, values_data, offset);
lb_emit_increment(p, offset_.addr);
val = lb_emit_load(p, val_ptr);
val = lb_emit_conv(p, val, elem);
lbValue check = lb_build_binary_in(p, val, the_set, Token_in);
lb_emit_if(p, check, body, loop);
lb_start_block(p, body);
} else {
lbAddr offset_ = lb_add_local_generated(p, t_int, false);
lb_addr_store(p, offset_, lb_const_int(m, t_int, et->BitSet.lower));
lbValue max_count = lb_const_int(m, t_int, et->BitSet.upper);
loop = lb_create_block(p, "for.bit_set.range.loop");
lb_emit_jump(p, loop);
lb_start_block(p, loop);
lbBlock *body_check = lb_create_block(p, "for.bit_set.range.body-check");
lbBlock *body = lb_create_block(p, "for.bit_set.range.body");
done = lb_create_block(p, "for.bit_set.range.done");
lbValue offset = lb_addr_load(p, offset_);
lbValue cond = lb_emit_comp(p, Token_LtEq, offset, max_count);
lb_emit_if(p, cond, body_check, done);
lb_start_block(p, body_check);
val = lb_emit_conv(p, offset, elem);
lb_emit_increment(p, offset_.addr);
lbValue check = lb_build_binary_in(p, val, the_set, Token_in);
lb_emit_if(p, check, body, loop);
lb_start_block(p, body);
} }
lbAddr remaining = lb_add_local_generated(p, mask, false);
lb_addr_store(p, remaining, initial_mask);
loop = lb_create_block(p, "for.bit_set.loop");
lbBlock *body = lb_create_block(p, "for.bit_set.body");
done = lb_create_block(p, "for.bit_set.done");
lb_emit_jump(p, loop);
lb_start_block(p, loop);
lbValue remaining_val = lb_addr_load(p, remaining);
lbValue cond = lb_emit_comp(p, Token_NotEq, remaining_val, lb_zero(m, mask));
lb_emit_if(p, cond, body, done);
lb_start_block(p, body);
val = lb_emit_count_trailing_zeros(p, remaining_val, mask);
val = lb_emit_conv(p, val, elem);
if (rs->reverse) {
val = lb_emit_arith(p, Token_Sub, lb_const_int(m, elem, et->BitSet.lower + 8*type_size_of(mask) - 1), val, elem);
} else {
val = lb_emit_arith(p, Token_Add, val, lb_const_int(m, elem, et->BitSet.lower), elem);
}
lbValue reduce_val = lb_emit_arith(p, Token_Sub, remaining_val, lb_const_int(m, mask, 1), mask);
remaining_val = lb_emit_arith(p, Token_And, remaining_val, reduce_val, mask);
lb_addr_store(p, remaining, remaining_val);
break; break;
} }
default: default: