A third of the draw path was reference counting

What changed

mtl12's Unix side (the part that turns recorded D3D12 commands into Metal calls) is written in Objective-C with ARC (automatic reference counting: the compiler inserts retain and release calls to keep objects alive). On a Mac, Wine runs as x86_64 code under Rosetta, Apple's translator, and every one of those calls is a call into the Objective-C runtime.

Per draw, the code loaded the Metal encoder and the argument table (the list of buffers a draw binds) into local variables and passed them to three helper functions. At the -O1 optimisation level, ARC retained and released each of them every time. The objects belong to the encoder for the whole draw, so those references are now marked __unsafe_unretained, which tells ARC to leave them alone. Two subtler cases followed. A function that returned the encoder made ARC retain it for the caller. A check run on every draw lived in a function whose object locals ARC released on exit, even when they were still nil. After both were fixed, a counter that intercepts the runtime's retain and release functions found no call per draw left. A review of every unretained reference found none that can outlive its object. It also found one comment that needed tightening.

The interesting part

An earlier session split the Unix side's roughly 0.5 µs per draw by area of code by skipping parts of it, because timing marks disturbed the measurement. They could not see inside the functions, because sample shows no symbols for translated code. Xcode's xctrace does work on a Wine process, but its stacks mislead in one specific way. objc_msgSend (the function behind every Objective-C method call) sets up no stack frame of its own. A sample taken inside it skips the function that made the call, so time spent in cmd_draw's Metal calls showed up under its caller. Reading the disassembly at each return address, not the line numbers, is what turned up the retain calls in the helpers' prologues.

A dead end with a cost: attaching lldb to the Wine process left Rosetta in a state where every newly built x86_64 program hangs at launch and cannot be killed.

Numbers

Three alternating rounds against the previous commit:

• bench-draws (10,000 draws a frame), submit: 4.99 to 5.08 ms before, 3.51 to 3.56 ms after.
• bgfx drawstress, Unix CPU per frame: 31.4 to 32.3 ms before, 21.7 to 22.3 ms after; frame 43.3 to 44.2 ms before, 33.6 to 34.3 ms after.
• ModelViewer: no change outside the run-to-run spread (ab4_modelviewer.txt). It waits for vsync and makes few draws.
• Compiling the Unix side at -O2 instead of -O1 would give another 2 to 3% (ab5_O2.txt). Not adopted yet.
• A single run of the perf history agrees: drawstress frame -25%.
• Tests: 839 of 943 pass, as before; none newly passing or failing.