[llvm-dev] Why not do machine instruction scheduling in SSA form?

[Andrew Trick via llvm-dev]

> On Jun 27, 2016, at 4:39 AM, Xing Su via llvm-dev <llvm-dev at lists.llvm.org> wrote:
> 
> Hi LLVM community,
> 
> Currently LLVM backend do pre-RA machine instruction scheduling in non-SSA form, I doubt why not do machine scheduling in SSA machine instruction form? Now LLVM’s machine scheduling uses a list-scheduling algorithm, but if we wang to support more complex scheduling algorithms, for example, modulo scheduling for loops, it seems more easy to accomplish this in SSA form as SSA is more suitable for tracking dependencies and doing code motion.
> 
> I find that LiveIntervals analysis pass, which is required by RegPressureTracker and MachineScheduler, cannot run before phi elimination. Is this one of the reason that machine scheduling is not done in SSA form?
> 
> Any explanation is appreciated. Thanks a lot!
> 
> 
> 
> Cheers,
> Xing


I also find it awkward to schedule and register allocate after eliminating SSA form (it would be possible to coalesce virtual registers while remaining in SSA form). But in theory LLVM has the information you need. LiveIntervals provides the reaching defs after MachineOperands have been renamed to coalesced vregs. The only fundamental complexity is that LiveIntervals needs to be updated during scheduling.

If I were writing a global code motion or cyclic code motion pass, I would run it prior to phi elimination for the reasons you mention. For modulo scheduling, it really depends on whether you want to see the copies and coalesced vregs during scheduling. The machine model and interesting utilities like MachineTraceMetrics are available prior to phi elimination. As Matthias said, LiveIntervals is also available if you really need it, but then you might as well just schedule after phi elimination, since LiveInterval update is the only real downside.

-Andy