Killing what can't be handled – putting a spaghetti demon to eternal rest

The DXIL-to-SPIR-V control-flow structurizer had become a fragile, roughly 9,000-line system that repeatedly broke on extreme game shaders. Its author argues that trying to handle every case while preserving readable output made the implementation too complex to reason about; a particularly difficult shader and a second validation failure prompted a redesign focused on correctness, with performance retained where possible. The replacement borrows Mesa’s `goto_if` approach, used for unstructured OpenCL control flow. A test mode first sent unstructured SPIR-V directly to drivers: games worked, and RADV performance matched the old path, making the two-week gamble on the new strategy viable.

The new method recursively organizes control flow into levels. When a branch targets a block that cannot be reached directly, it records the intended path in selector variables and routes execution through an outer merge; helper blocks handle divergent branches, while skip analysis records when entire levels can be bypassed. Switches are split into conditional branches rather than preserving complicated fall-through behavior. The rewrite also abandons manual PHI-node maintenance: PHIs become variables and stores before structurization, with occasional extra loads and stores repairing SSA dominance afterward. Fuzzing passed highly unusual forward-flow patterns, and the design guarantees maximal reconvergence - the largest possible group of threads follows any executed block. The account ends as it turns to the remaining challenge of handling loops.