The patch was small. The focused case worked, shader-db produced zero hits, and the optimization never left my branch.
How I found it
I found a TODO in aco_optimizer.cpp about a missing constant-folding case around packed operations and followed it.
The transformation
Before GFX11, the relevant pack_hl form was not directly available in the ISA. If an operand is constant, shifting that constant can allow the operation to use an _ll or _hh form instead.
A focused case is not a workload
I implemented the case and constructed the IR pattern I needed. The transformation worked. Then I ran shader-db over real shaders: no hit.
The focused case showed that the code transformed a synthetic input. It did not show that the pattern reached this optimizer.
Where the pattern went
The likely explanation was an earlier simplification: NIR or another part of the pipeline had already removed the pattern before ACO could see it. I did not isolate the exact pass.
Shader-db only told me that the optimization had no workload in that run. It did not prove the combine was wrong. I stopped because I had not shown that the pattern could reach ACO in a real pipeline.
What survived
The optimizer change did not. Before attempting a similar combine, I would capture the IR at the optimizer boundary first.
I did not send this upstream. The branch was only an experiment; the useful result was finding out why the pass never saw the pattern.