bc2cpp-array-times-result.md

Array * Integer is a fresh Array (ARRAY_TIMES_RESULT)

Programs often size a table by repeating a literal: @bg_pixels = [0] * 16, @sp_map = [nil] * 264. The class pools (ADR 0296) join the class of every value stored into an ivar. They could not name the class of [0] * 16, so every such ivar lost its pool, and each @sp_map[@hclk] kept the mrb_array_p(...) && ...->c == M->array_class test in front of the element read.

CodeGen#native_array_times_result (tools/bc2cpp/codegen_native_results.rb) answers ARR for a * send when:

• the receiver is exactly an Array and the argument is proven an Integer;
• the native body is the audited mrb_ary_times (NativeCoreDirect::RESULT_ENTRIES). It returns mrb_ary_new for a zero count or its own ary_new_capa, both exactly Array, and raises for a negative count. Nothing in the program defines, installs or prepends * on Array (native_core_entry_safe?);
• nothing can give an Integer a to_str. mrb_ary_times first tries mrb_check_string_type on the argument and joins when that converts, as [1, 2] * "," does. So the registry may hold no to_str on Integer's ancestors, no installer or outside Ruby source may spell it, and no native outside mruby's core may spell it;
• no ancestor of Array defines *.

The last condition lets the fact skip the usual refusal for a name a computed send may remove. Optcarrot's send(*DISPATCH[@opcode]) counts as a possible remove_method of every name. If Array#* were removed, the send would either raise NoMethodError and produce no value, or reach an ancestor's *, and there is none. Either way, every value the send does produce is a fresh Array.

BC2CPP_ARRAY_TIMES_RESULT=0 drops the fact.

Check

MRBC=path/to/mrbc ruby scripts/bc2cpp/array_times_result_check.rb asserts that @cells = [0] * 8 and @cells = [nil] * 4 give @cells an Array pool, in a world with a computed send(*op), and that its index read drops the class test. These keep no pool: a String argument, an argument of unknown class, a program Integer#to_str, a program Object#*, and the opt-out. Interpreted and compiled runs print the same, including the String join and the negative-count ArgumentError.

Measurement

These figures come from the optcarrot probe (Lan_Master, release -O3, callgrind). Instructions per frame are the difference between runs of 21 and 1 frames, divided by 20. Checksum 59662 at 180 frames is unchanged. The probe's ivar proof is only on with the ivar proof scope change, so the build measured here also has that change, resumable locals and HOT_INLINE, on master 479b31ff.

Build	Instructions per frame	Ivar class pools
Before	32,191,709	42
With ARRAY_TIMES_RESULT	30,589,709	48

PPU#@bg_pixels, @sp_map, @sp_buffer, @output_color, @sp_ram and ROM#@prg_ref gain pools. PPU#render_pixel alone drops by 0.62M, and the frame as a whole by 1.60M (5.0%). Startup is unchanged.